From a3420fd7b722d1ff14ffe7c588ae2225a37f2763 Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Sun, 6 Aug 2023 01:12:39 -0400
Subject: [PATCH 01/19] Update to v0.3.0 (#237)
* Checkpointing, more input formats, HTML report, test suite, documentation, new posterior and estimators, fix readthedocs, improve prior heuristics, Github actions for tests, address memory issues
---
.dockerignore | 10 +
.dockstore.yml | 6 +
.github/workflows/miniwdl_check.yml | 17 +
.github/workflows/run_packaging_check.yml | 45 +
.github/workflows/run_pytest.yml | 32 +
.gitignore | 6 +
MANIFEST.in | 5 +-
README.rst | 138 +-
REQUIREMENTS-DOCKER.txt | 9 -
REQUIREMENTS.txt | 13 -
build_docker_local.sh | 8 +
build_docker_release.sh | 10 +
cellbender/__init__.py | 1 +
cellbender/base_cli.py | 40 +-
cellbender/monitor.py | 41 +
cellbender/remove_background/argparse.py | 166 --
cellbender/remove_background/argparser.py | 284 +++
cellbender/remove_background/checkpoint.py | 368 ++++
cellbender/remove_background/cli.py | 231 +-
cellbender/remove_background/consts.py | 68 +-
cellbender/remove_background/data/dataprep.py | 106 +-
cellbender/remove_background/data/dataset.py | 1763 ++++-----------
.../remove_background/data/extras/simulate.py | 1200 +++++++----
cellbender/remove_background/data/io.py | 1207 +++++++++++
cellbender/remove_background/data/priors.py | 391 ++++
.../NegativeBinomialPoissonConvApprox.py | 2 +-
cellbender/remove_background/downstream.py | 479 +++++
cellbender/remove_background/estimation.py | 902 ++++++++
cellbender/remove_background/exceptions.py | 13 +-
cellbender/remove_background/infer.py | 781 -------
cellbender/remove_background/model.py | 452 ++--
cellbender/remove_background/posterior.py | 1708 +++++++++++++++
cellbender/remove_background/report.ipynb | 133 ++
cellbender/remove_background/report.py | 1888 +++++++++++++++++
cellbender/remove_background/run.py | 779 +++++++
cellbender/remove_background/sparse_utils.py | 101 +
.../tests/benchmarking/README.md | 62 +
.../tests/benchmarking/benchmark.wdl | 29 +
.../tests/benchmarking/benchmark_inputs.json | 19 +
.../tests/benchmarking/cuda_check_inputs.json | 4 +
.../docker_image_check_cuda_status.wdl | 38 +
.../tests/benchmarking/run_benchmark.py | 225 ++
.../run_benchmark_result_tabulation.py | 146 ++
.../benchmarking/run_usual_benchmarks.sh | 74 +
.../remove_background/tests/conftest.py | 202 ++
.../tests/mckp_memory_profiling.py | 146 ++
.../tests/miniwdl_check_wdl.sh | 13 +
cellbender/remove_background/tests/test.py | 193 --
.../tests/test_checkpoint.py | 626 ++++++
.../remove_background/tests/test_dataprep.py | 101 +
.../remove_background/tests/test_dataset.py | 40 +
.../tests/test_downstream.py | 119 ++
.../tests/test_estimation.py | 358 ++++
.../remove_background/tests/test_infer.py | 301 +++
.../tests/test_integration.py | 52 +
cellbender/remove_background/tests/test_io.py | 223 ++
.../remove_background/tests/test_monitor.py | 21 +
.../remove_background/tests/test_posterior.py | 428 ++++
.../tests/test_sparse_utils.py | 203 ++
.../remove_background/tests/test_train.py | 135 ++
cellbender/remove_background/train.py | 362 ++--
cellbender/remove_background/vae/base.py | 164 ++
cellbender/remove_background/vae/decoder.py | 42 +-
cellbender/remove_background/vae/encoder.py | 258 ++-
docker/Dockerfile | 65 +-
docker/DockerfileGit | 35 +
.../PCL_rat_A_LA6_learning_curve.png | Bin 0 -> 23433 bytes
.../remove_background/bad_learning_curves.png | Bin 0 -> 214054 bytes
.../simulated_s6_learning_curve.png | Bin 0 -> 30054 bytes
docs/source/_static/theme_overrides.css | 15 -
docs/source/changelog/index.rst | 85 +
docs/source/citation/index.rst | 13 +-
docs/source/conf.py | 7 +-
docs/source/contributing/index.rst | 10 +-
docs/source/getting_started/index.rst | 11 -
.../remove_background/index.rst | 138 --
docs/source/help_and_reference/index.rst | 10 -
.../remove_background/index.rst | 81 -
docs/source/index.rst | 25 +-
docs/source/installation/index.rst | 76 +-
docs/source/introduction/index.rst | 20 +-
docs/source/reference/index.rst | 188 ++
.../license/index.rst | 0
docs/source/troubleshooting/index.rst | 426 ++++
docs/source/tutorial/index.rst | 224 ++
docs/source/usage/index.rst | 49 +-
examples/remove_background/.gitignore | 8 +-
.../generate_tiny_10x_dataset.py | 79 +
.../generate_tiny_10x_pbmc.py | 103 -
readthedocs.yml | 2 +-
REQUIREMENTS-RTD.txt => requirements-rtd.txt | 0
requirements.txt | 15 +
setup.py | 56 +-
wdl/README.rst | 86 +-
wdl/cellbender_remove_background.wdl | 309 ++-
95 files changed, 15843 insertions(+), 4280 deletions(-)
create mode 100644 .dockerignore
create mode 100644 .dockstore.yml
create mode 100644 .github/workflows/miniwdl_check.yml
create mode 100644 .github/workflows/run_packaging_check.yml
create mode 100644 .github/workflows/run_pytest.yml
delete mode 100644 REQUIREMENTS-DOCKER.txt
delete mode 100644 REQUIREMENTS.txt
create mode 100755 build_docker_local.sh
create mode 100755 build_docker_release.sh
create mode 100644 cellbender/monitor.py
delete mode 100644 cellbender/remove_background/argparse.py
create mode 100644 cellbender/remove_background/argparser.py
create mode 100644 cellbender/remove_background/checkpoint.py
create mode 100644 cellbender/remove_background/data/io.py
create mode 100644 cellbender/remove_background/data/priors.py
create mode 100644 cellbender/remove_background/downstream.py
create mode 100644 cellbender/remove_background/estimation.py
delete mode 100644 cellbender/remove_background/infer.py
create mode 100644 cellbender/remove_background/posterior.py
create mode 100644 cellbender/remove_background/report.ipynb
create mode 100644 cellbender/remove_background/report.py
create mode 100644 cellbender/remove_background/run.py
create mode 100644 cellbender/remove_background/sparse_utils.py
create mode 100644 cellbender/remove_background/tests/benchmarking/README.md
create mode 100644 cellbender/remove_background/tests/benchmarking/benchmark.wdl
create mode 100644 cellbender/remove_background/tests/benchmarking/benchmark_inputs.json
create mode 100644 cellbender/remove_background/tests/benchmarking/cuda_check_inputs.json
create mode 100644 cellbender/remove_background/tests/benchmarking/docker_image_check_cuda_status.wdl
create mode 100644 cellbender/remove_background/tests/benchmarking/run_benchmark.py
create mode 100644 cellbender/remove_background/tests/benchmarking/run_benchmark_result_tabulation.py
create mode 100755 cellbender/remove_background/tests/benchmarking/run_usual_benchmarks.sh
create mode 100644 cellbender/remove_background/tests/conftest.py
create mode 100644 cellbender/remove_background/tests/mckp_memory_profiling.py
create mode 100755 cellbender/remove_background/tests/miniwdl_check_wdl.sh
delete mode 100644 cellbender/remove_background/tests/test.py
create mode 100644 cellbender/remove_background/tests/test_checkpoint.py
create mode 100644 cellbender/remove_background/tests/test_dataprep.py
create mode 100644 cellbender/remove_background/tests/test_dataset.py
create mode 100644 cellbender/remove_background/tests/test_downstream.py
create mode 100644 cellbender/remove_background/tests/test_estimation.py
create mode 100644 cellbender/remove_background/tests/test_infer.py
create mode 100644 cellbender/remove_background/tests/test_integration.py
create mode 100644 cellbender/remove_background/tests/test_io.py
create mode 100644 cellbender/remove_background/tests/test_monitor.py
create mode 100644 cellbender/remove_background/tests/test_posterior.py
create mode 100644 cellbender/remove_background/tests/test_sparse_utils.py
create mode 100644 cellbender/remove_background/tests/test_train.py
create mode 100644 cellbender/remove_background/vae/base.py
create mode 100644 docker/DockerfileGit
create mode 100644 docs/source/_static/remove_background/PCL_rat_A_LA6_learning_curve.png
create mode 100644 docs/source/_static/remove_background/bad_learning_curves.png
create mode 100644 docs/source/_static/remove_background/simulated_s6_learning_curve.png
delete mode 100644 docs/source/_static/theme_overrides.css
create mode 100644 docs/source/changelog/index.rst
delete mode 100644 docs/source/getting_started/index.rst
delete mode 100644 docs/source/getting_started/remove_background/index.rst
delete mode 100644 docs/source/help_and_reference/index.rst
delete mode 100644 docs/source/help_and_reference/remove_background/index.rst
create mode 100644 docs/source/reference/index.rst
rename docs/source/{help_and_reference => reference}/license/index.rst (100%)
create mode 100644 docs/source/troubleshooting/index.rst
create mode 100644 docs/source/tutorial/index.rst
create mode 100755 examples/remove_background/generate_tiny_10x_dataset.py
delete mode 100755 examples/remove_background/generate_tiny_10x_pbmc.py
rename REQUIREMENTS-RTD.txt => requirements-rtd.txt (100%)
create mode 100644 requirements.txt
diff --git a/.dockerignore b/.dockerignore
new file mode 100644
index 00000000..186f305f
--- /dev/null
+++ b/.dockerignore
@@ -0,0 +1,10 @@
+**/.git
+.github
+examples
+docs
+wdl
+dist
+cellbender.egg-info
+.pytest_cache
+*.h5
+*.tar.gz
\ No newline at end of file
diff --git a/.dockstore.yml b/.dockstore.yml
new file mode 100644
index 00000000..0f2192a5
--- /dev/null
+++ b/.dockstore.yml
@@ -0,0 +1,6 @@
+version: 1.2
+
+workflows:
+ - subclass: WDL
+ primaryDescriptorPath: /wdl/cellbender_remove_background.wdl
+ name: cellbender_remove_background
diff --git a/.github/workflows/miniwdl_check.yml b/.github/workflows/miniwdl_check.yml
new file mode 100644
index 00000000..8a8de2c8
--- /dev/null
+++ b/.github/workflows/miniwdl_check.yml
@@ -0,0 +1,17 @@
+name: 'validate WDL'
+on: [pull_request]
+env:
+ MINIWDL_VERSION: 1.8.0
+jobs:
+ miniwdl-check:
+ runs-on: ubuntu-latest
+ steps:
+ - uses: actions/checkout@v3
+ - uses: actions/setup-python@v4
+ with:
+ python-version: '3.7'
+ - name: 'Install miniwdl and Run Validation Test'
+ run: |
+ pip install miniwdl==$MINIWDL_VERSION;
+ ./cellbender/remove_background/tests/miniwdl_check_wdl.sh;
+ shell: bash
diff --git a/.github/workflows/run_packaging_check.yml b/.github/workflows/run_packaging_check.yml
new file mode 100644
index 00000000..8a9fc1cd
--- /dev/null
+++ b/.github/workflows/run_packaging_check.yml
@@ -0,0 +1,45 @@
+# Package for PyPI
+
+name: 'packaging'
+
+on: pull_request
+
+jobs:
+ build:
+
+ runs-on: 'ubuntu-latest'
+ strategy:
+ matrix:
+ python-version: ['3.7']
+
+ steps:
+ - name: 'Checkout repo'
+ uses: actions/checkout@v3
+
+ - name: 'Set up Python ${{ matrix.python-version }}'
+ uses: actions/setup-python@v4
+ with:
+ python-version: ${{ matrix.python-version }}
+ cache: 'pip'
+
+ - name: 'Install packaging software'
+ run: pip install --upgrade setuptools build twine
+
+ - name: 'Print package versions'
+ run: pip list
+
+ - name: 'Package code using build'
+ run: python -m build
+
+ - name: 'Check package using twine'
+ run: python -m twine check dist/*
+
+ - name: 'Extract branch name'
+ shell: bash
+ run: echo "branch=$(echo ${GITHUB_REF#refs/heads/})" >>$GITHUB_OUTPUT
+ id: extract_branch
+
+ - name: 'Install from github branch and run a test'
+ run: |
+ pip install pytest git+https://github.com/broadinstitute/CellBender@${{ steps.extract_branch.outputs.branch }}
+ cellbender -v
diff --git a/.github/workflows/run_pytest.yml b/.github/workflows/run_pytest.yml
new file mode 100644
index 00000000..37e652b0
--- /dev/null
+++ b/.github/workflows/run_pytest.yml
@@ -0,0 +1,32 @@
+# Run cellbender's tests
+
+name: 'pytest'
+
+on: pull_request
+
+jobs:
+ build:
+
+ runs-on: 'ubuntu-latest'
+ strategy:
+ matrix:
+ python-version: ['3.7']
+
+ steps:
+ - name: 'Checkout repo'
+ uses: actions/checkout@v3
+
+ - name: 'Set up Python ${{ matrix.python-version }}'
+ uses: actions/setup-python@v4
+ with:
+ python-version: ${{ matrix.python-version }}
+ cache: 'pip'
+
+ - name: 'Install package including pytest'
+ run: pip install .[dev]
+
+ - name: 'Print package versions'
+ run: pip list
+
+ - name: 'Test with pytest'
+ run: pytest -v
diff --git a/.gitignore b/.gitignore
index b1e889ac..80b6c0c2 100644
--- a/.gitignore
+++ b/.gitignore
@@ -12,3 +12,9 @@ dist/
.eggs/
.idea/
*.iml
+*.h5
+*.h5ad
+*.tsv
+*.csv
+*.npz
+*.tar.gz
diff --git a/MANIFEST.in b/MANIFEST.in
index c4bf4561..66e89a17 100644
--- a/MANIFEST.in
+++ b/MANIFEST.in
@@ -1 +1,4 @@
-include README.rst
\ No newline at end of file
+include README.rst
+include LICENSE
+include requirements.txt
+include requirements-rtd.txt
\ No newline at end of file
diff --git a/README.rst b/README.rst
index a12cd22e..ea39e2ca 100644
--- a/README.rst
+++ b/README.rst
@@ -1,11 +1,27 @@
CellBender
==========
+.. image:: https://img.shields.io/github/license/broadinstitute/CellBender?color=white
+ :target: LICENSE
+ :alt: License
+
.. image:: https://readthedocs.org/projects/cellbender/badge/?version=latest
:target: https://cellbender.readthedocs.io/en/latest/?badge=latest
:alt: Documentation Status
-.. image:: https://github.com/broadinstitute/CellBender/blob/master/docs/source/_static/design/logo_250_185.png
+.. image:: https://img.shields.io/pypi/v/CellBender.svg
+ :target: https://pypi.org/project/CellBender
+ :alt: PyPI
+
+.. image:: https://static.pepy.tech/personalized-badge/cellbender?period=total&units=international_system&left_color=grey&right_color=blue&left_text=pypi%20downloads
+ :target: https://pepy.tech/project/CellBender
+ :alt: Downloads
+
+.. image:: https://img.shields.io/github/stars/broadinstitute/CellBender?color=yellow&logoColor=yellow)
+ :target: https://github.com/broadinstitute/CellBender/stargazers
+ :alt: Stars
+
+.. image:: docs/source/_static/design/logo_250_185.png
:alt: CellBender Logo
CellBender is a software package for eliminating technical artifacts from
@@ -16,67 +32,123 @@ The current release contains the following modules. More modules will be added i
* ``remove-background``:
This module removes counts due to ambient RNA molecules and random barcode swapping from (raw)
- UMI-based scRNA-seq count matrices. At the moment, only the count matrices produced by the
- CellRanger ``count`` pipeline is supported. Support for additional tools and protocols will be
- added in the future. A quick start tutorial can be found
- `here <https://cellbender.readthedocs.io/en/latest/getting_started/remove_background/index.html>`_.
+ UMI-based scRNA-seq count matrices. Also works for snRNA-seq and CITE-seq.
-Please refer to the `documentation <https://cellbender.readthedocs.io/en/latest/>`_ for a quick start tutorial on using CellBender.
+Please refer to `the documentation <https://cellbender.readthedocs.io/en/latest/>`_ for a quick start tutorial.
Installation and Usage
----------------------
-Manual installation
-~~~~~~~~~~~~~~~~~~~
+CellBender can be installed via
+
+.. code-block:: console
+
+ $ pip install cellbender
+
+(and we recommend installing in its own ``conda`` environment to prevent
+conflicts with other software).
+
+CellBender is run as a command-line tool, as in
+
+.. code-block:: console
+
+ (cellbender) $ cellbender remove-background \
+ --cuda \
+ --input my_raw_count_matrix_file.h5 \
+ --output my_cellbender_output_file.h5
+
+See `the usage documentation <https://cellbender.readthedocs.io/en/latest/usage/index.html>`_
+for details.
+
+
+Using The Official Docker Image
+-------------------------------
+
+A GPU-enabled docker image is available from the Google Container Registry (GCR) as:
+
+``us.gcr.io/broad-dsde-methods/cellbender:latest``
+
+Available image tags track release tags in GitHub, and include ``latest``,
+``0.1.0``, ``0.2.0``, ``0.2.1``, ``0.2.2``, and ``0.3.0``.
+
+
+WDL Users
+---------
+
+A workflow written in the
+`workflow description language (WDL) <https://github.com/openwdl/wdl>`_
+is available for CellBender remove-background.
+
+For `Terra <https://app.terra.bio>`_ users, a workflow called
+``cellbender/remove-background`` is
+`available from the Broad Methods repository
+<https://portal.firecloud.org/#methods/cellbender/remove-background/>`_.
+
+There is also a `version available on Dockstore
+<https://dockstore.org/workflows/github.com/broadinstitute/CellBender>`_.
-The recommended installation is as follows. Create a conda environment and activate it:
-.. code-block:: bash
+Advanced installation
+---------------------
- $ conda create -n cellbender python=3.7
- $ source activate cellbender
+From source for development
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Create a conda environment and activate it:
+
+.. code-block:: console
+
+ $ conda create -n cellbender python=3.7
+ $ conda activate cellbender
Install the `pytables <https://www.pytables.org>`_ module:
-.. code-block:: bash
+.. code-block:: console
+
+ (cellbender) $ conda install -c anaconda pytables
- (cellbender) $ conda install -c anaconda pytables
+Install `pytorch <https://pytorch.org>`_ via
+`these instructions <https://pytorch.org/get-started/locally/>`_, for example:
-Install `pytorch <https://pytorch.org>`_ (shown below for CPU; if you have a CUDA-ready GPU, please skip
-this part and follow `these <https://pytorch.org/get-started/locally/>`_ instructions instead):
+.. code-block:: console
-.. code-block:: bash
+ (cellbender) $ pip install torch
- (cellbender) $ conda install pytorch torchvision -c pytorch
+and ensure that your installation is appropriate for your hardware (i.e. that
+the relevant CUDA drivers get installed and that ``torch.cuda.is_available()``
+returns ``True`` if you have a GPU available.
-Clone this repository and install CellBender:
+Clone this repository and install CellBender (in editable ``-e`` mode):
-.. code-block:: bash
+.. code-block:: console
+ (cellbender) $ git clone https://github.com/broadinstitute/CellBender.git
(cellbender) $ pip install -e CellBender
-Using The Official Docker Image
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-A GPU-enabled docker image is available from the Google Container Registry (GCR) as:
+From a specific commit
+~~~~~~~~~~~~~~~~~~~~~~
-``us.gcr.io/broad-dsde-methods/cellbender:latest``
+This can be achieved via
-Terra Users
-~~~~~~~~~~~
+.. code-block:: console
-For `Terra <https://app.terra.bio>`_ users, a `workflow <https://portal.firecloud.org/#methods/cellbender/remove-background/>`_
-is available as:
+ (cellbender) $ pip install --no-cache-dir -U git+https://github.com/broadinstitute/CellBender.git@<SHA>
-``cellbender/remove-background``
+where ``<SHA>`` must be replaced by any reference to a particular git commit,
+such as a tag, a branch name, or a commit sha.
Citing CellBender
-----------------
If you use CellBender in your research (and we hope you will), please consider
-citing `our paper on bioRxiv <https://doi.org/10.1101/791699>`_.
+citing our paper in Nature Methods:
+
+Stephen J Fleming, Mark D Chaffin, Alessandro Arduini, Amer-Denis Akkad,
+Eric Banks, John C Marioni, Anthony A Phillipakis, Patrick T Ellinor,
+and Mehrtash Babadi. Unsupervised removal of systematic background noise from
+droplet-based single-cell experiments using CellBender.
+`Nature Methods` (in press), 2023.
-Stephen J Fleming, John C Marioni, and Mehrtash Babadi. CellBender remove-background:
-a deep generative model for unsupervised removal of background noise from scRNA-seq
-datasets. bioRxiv 791699; doi: `https://doi.org/10.1101/791699 <https://doi.org/10.1101/791699>`_
+See also `our preprint on bioRxiv <https://doi.org/10.1101/791699>`_.
diff --git a/REQUIREMENTS-DOCKER.txt b/REQUIREMENTS-DOCKER.txt
deleted file mode 100644
index 3884f08e..00000000
--- a/REQUIREMENTS-DOCKER.txt
+++ /dev/null
@@ -1,9 +0,0 @@
-anndata>=0.7
-numpy
-scipy
-tables
-pandas
-pyro-ppl>=0.3.2
-torch>=1.9.0
-scikit-learn
-matplotlib
diff --git a/REQUIREMENTS.txt b/REQUIREMENTS.txt
deleted file mode 100644
index 535e3e78..00000000
--- a/REQUIREMENTS.txt
+++ /dev/null
@@ -1,13 +0,0 @@
-anndata>=0.7
-numpy
-scipy
-tables
-pandas
-pyro-ppl>=0.3.2
-scikit-learn
-matplotlib
-sphinx>=2.1
-sphinx-rtd-theme
-sphinx-autodoc-typehints
-sphinxcontrib-programoutput
-sphinx-argparse
diff --git a/build_docker_local.sh b/build_docker_local.sh
new file mode 100755
index 00000000..c89367ff
--- /dev/null
+++ b/build_docker_local.sh
@@ -0,0 +1,8 @@
+#!/bin/bash
+
+tag=$(cat cellbender/__init__.py | sed -e 's?__version__ = ??' | sed "s/^'\(.*\)'$/\1/")
+
+docker build \
+ -t us.gcr.io/broad-dsde-methods/cellbender:${tag} \
+ -f docker/Dockerfile \
+ .
diff --git a/build_docker_release.sh b/build_docker_release.sh
new file mode 100755
index 00000000..ff0511fc
--- /dev/null
+++ b/build_docker_release.sh
@@ -0,0 +1,10 @@
+#!/bin/bash
+
+tag=$(cat cellbender/__init__.py | sed -e 's?__version__ = ??' | sed "s/^'\(.*\)'$/\1/")
+release=v${tag}
+
+docker build \
+ -t us.gcr.io/broad-dsde-methods/cellbender:${tag} \
+ --build-arg GIT_SHA=${release} \
+ -f docker/DockerfileGit \
+ .
diff --git a/cellbender/__init__.py b/cellbender/__init__.py
index e69de29b..0404d810 100644
--- a/cellbender/__init__.py
+++ b/cellbender/__init__.py
@@ -0,0 +1 @@
+__version__ = '0.3.0'
diff --git a/cellbender/base_cli.py b/cellbender/base_cli.py
index 1489772c..bf5a71b3 100644
--- a/cellbender/base_cli.py
+++ b/cellbender/base_cli.py
@@ -5,16 +5,32 @@
"""
import sys
+import os
import argparse
from abc import ABC, abstractmethod
from typing import Dict
import importlib
-
+import codecs
# New tools should be added to this list.
TOOL_NAME_LIST = ['remove-background']
+def read(rel_path):
+ here = os.path.abspath(os.path.dirname(__file__))
+ with codecs.open(os.path.join(here, rel_path), 'r') as fp:
+ return fp.read()
+
+
+def get_version() -> str:
+ for line in read('__init__.py').splitlines():
+ if line.startswith('__version__'):
+ delim = '"' if '"' in line else "'"
+ return line.split(delim)[1]
+ else:
+ raise RuntimeError("Unable to find version string.")
+
+
class AbstractCLI(ABC):
"""Abstract class for cellbender command-line interface tools.
@@ -29,13 +45,15 @@ def get_name(self) -> str:
"""Return the command-line name of the tool."""
pass
+ @staticmethod
@abstractmethod
- def validate_args(self, parser: argparse):
+ def validate_args(parser: argparse) -> argparse.Namespace:
"""Do tool-specific argument validation, returning args."""
pass
+ @staticmethod
@abstractmethod
- def run(self, args):
+ def run(args):
"""Run the tool using the parsed arguments."""
pass
@@ -62,8 +80,12 @@ def get_populated_argparser() -> argparse.ArgumentParser:
# Set up argument parser.
parser = argparse.ArgumentParser(
prog="cellbender",
- description="CellBender is a software package for eliminating technical artifacts from high-throughput "
- "single-cell RNA sequencing (scRNA-seq) data.")
+ description="CellBender is a software package for eliminating technical "
+ "artifacts from high-throughput single-cell RNA sequencing "
+ "(scRNA-seq) data.")
+
+ # Add the ability to display the version.
+ parser.add_argument('-v', '--version', action='version', version=get_version())
# Declare the existence of sub-parsers.
subparsers = parser.add_subparsers(
@@ -72,7 +94,7 @@ def get_populated_argparser() -> argparse.ArgumentParser:
dest="tool")
for tool_name in TOOL_NAME_LIST:
- module_argparse_str_list = ["cellbender", tool_name.replace("-", "_"), "argparse"]
+ module_argparse_str_list = ["cellbender", tool_name.replace("-", "_"), "argparser"]
module_argparse = importlib.import_module('.'.join(module_argparse_str_list))
subparsers = module_argparse.add_subparser_args(subparsers)
@@ -103,3 +125,9 @@ def main():
else:
parser.print_help()
+
+ return 0
+
+
+if __name__ == '__main__':
+ sys.exit(main())
diff --git a/cellbender/monitor.py b/cellbender/monitor.py
new file mode 100644
index 00000000..e7dca617
--- /dev/null
+++ b/cellbender/monitor.py
@@ -0,0 +1,41 @@
+"""Utility functions for hardware monitoring"""
+
+# Inspiration for the nvidia-smi command comes from here:
+# https://pytorch-lightning.readthedocs.io/en/latest/_modules/pytorch_lightning/callbacks/gpu_stats_monitor.html#GPUStatsMonitor
+# but here it is stripped down to the absolute minimum
+
+import torch
+import psutil
+from psutil._common import bytes2human
+import shutil
+import subprocess
+
+
+def get_hardware_usage(use_cuda: bool) -> str:
+ """Get a current snapshot of RAM, CPU, GPU memory, and GPU utilization as a string"""
+
+ mem = psutil.virtual_memory()
+
+ if use_cuda:
+ # Run nvidia-smi to get GPU utilization
+ gpu_query = 'utilization.gpu'
+ format = 'csv,nounits,noheader'
+ result = subprocess.run(
+ [shutil.which("nvidia-smi"), f"--query-gpu={gpu_query}", f"--format={format}"],
+ encoding="utf-8",
+ stdout=subprocess.PIPE,
+ stderr=subprocess.PIPE, # for backward compatibility with python version 3.6
+ check=True,
+ )
+ pct_gpu_util = result.stdout.strip()
+ gpu_string = (f'Volatile GPU utilization: {pct_gpu_util} %\n'
+ f'GPU memory reserved: {torch.cuda.memory_reserved() / 1e9} GB\n'
+ f'GPU memory allocated: {torch.cuda.memory_allocated() / 1e9} GB\n')
+ else:
+ gpu_string = ''
+
+ cpu_string = (f'Avg CPU load over past minute: '
+ f'{psutil.getloadavg()[0] / psutil.cpu_count() * 100:.1f} %\n'
+ f'RAM in use: {bytes2human(mem.used)} ({mem.percent} %)')
+
+ return gpu_string + cpu_string
diff --git a/cellbender/remove_background/argparse.py b/cellbender/remove_background/argparse.py
deleted file mode 100644
index a52a3eb8..00000000
--- a/cellbender/remove_background/argparse.py
+++ /dev/null
@@ -1,166 +0,0 @@
-import argparse
-from cellbender.remove_background import consts
-
-
-def add_subparser_args(subparsers: argparse) -> argparse:
- """Add tool-specific arguments for remove-background.
-
- Args:
- subparsers: Parser object before addition of arguments specific to
- remove-background.
-
- Returns:
- parser: Parser object with additional parameters.
-
- """
-
- subparser = subparsers.add_parser("remove-background",
- description="Remove background RNA "
- "from count matrix.",
- help="Remove background ambient RNA "
- "and barcode-swapped reads from "
- "a count matrix, producing a "
- "new count matrix and "
- "determining which barcodes "
- "contain real cells.")
-
- subparser.add_argument("--input", nargs=None, type=str,
- dest='input_file', default=None,
- required=True,
- help="Data file on which to run tool, as either "
- "_raw_gene_barcode_matrices_h5.h5 file, or "
- "as the path containing the raw "
- "matrix.mtx, barcodes.tsv, and genes.tsv "
- "files. Supported for outputs of "
- "CellRanger v2 and v3.")
- subparser.add_argument("--output", nargs=None, type=str,
- dest='output_file', default=None,
- required=True,
- help="Output file location (the path must "
- "exist, and the file name must have .h5 "
- "extension).")
- subparser.add_argument("--expected-cells", nargs=None, type=int,
- default=None,
- dest="expected_cell_count",
- help="Number of cells expected in the dataset "
- "(a rough estimate within a factor of 2 "
- "is sufficient).")
- subparser.add_argument("--total-droplets-included",
- nargs=None, type=int,
- default=consts.TOTAL_DROPLET_DEFAULT,
- dest="total_droplets",
- help="The number of droplets from the "
- "rank-ordered UMI plot that will be "
- "analyzed. The largest 'total_droplets' "
- "droplets will have their cell "
- "probabilities inferred as an output. (default: %(default)s)")
- subparser.add_argument("--model", nargs=None, type=str,
- default="full",
- choices=["simple", "ambient", "swapping", "full"],
- dest="model",
- help="Which model is being used for count data. "
- " 'simple' does not model either ambient "
- "RNA or random barcode swapping (for "
- "debugging purposes -- not recommended). "
- "'ambient' assumes background RNA is "
- "incorporated into droplets. 'swapping' "
- "assumes background RNA comes from random "
- "barcode swapping. 'full' uses a "
- "combined ambient and swapping model. "
- "(default: %(default)s)")
- subparser.add_argument("--epochs", nargs=None, type=int, default=150,
- dest="epochs",
- help="Number of epochs to train. (default: %(default)s)")
- subparser.add_argument("--cuda",
- dest="use_cuda", action="store_true",
- help="Including the flag --cuda will run the "
- "inference on a GPU.")
- subparser.add_argument("--low-count-threshold", type=int,
- default=consts.LOW_UMI_CUTOFF,
- dest="low_count_threshold",
- help="Droplets with UMI counts below this "
- "number are completely excluded from the "
- "analysis. This can help identify the "
- "correct prior for empty droplet counts "
- "in the rare case where empty counts "
- "are extremely high (over 200). (default: %(default)s)")
- subparser.add_argument("--z-dim", type=int, default=100,
- dest="z_dim",
- help="Dimension of latent variable z. (default: %(default)s)")
- subparser.add_argument("--z-layers", nargs="+", type=int, default=[500],
- dest="z_hidden_dims",
- help="Dimension of hidden layers in the encoder "
- "for z. (default: %(default)s)")
- subparser.add_argument("--training-fraction",
- type=float, nargs=None,
- default=consts.TRAINING_FRACTION,
- dest="training_fraction",
- help="Training detail: the fraction of the "
- "data used for training. The rest is never "
- "seen by the inference algorithm. Speeds up "
- "learning. (default: %(default)s)")
- subparser.add_argument("--empty-drop-training-fraction",
- type=float, nargs=None,
- default=consts.FRACTION_EMPTIES,
- dest="fraction_empties",
- help="Training detail: the fraction of the "
- "training data each epoch "
- "that is drawn (randomly sampled) from "
- "surely empty droplets. (default: %(default)s)")
- subparser.add_argument("--blacklist-genes", nargs="+", type=int,
- default=[],
- dest="blacklisted_genes",
- help="Integer indices of genes to ignore "
- "entirely. In the output count matrix, "
- "the counts for these genes will be set "
- "to zero.")
- subparser.add_argument("--fpr", nargs="+",
- type=float, default=[0.01],
- dest="fpr",
- help="Target false positive rate in (0, 1). A false "
- "positive is a true signal count that is "
- "erroneously removed. More background removal "
- "is accompanied by more signal removal "
- "at high values of FPR. You can specify "
- "multiple values, which will create multiple "
- "output files. (default: %(default)s)")
- subparser.add_argument("--exclude-antibody-capture",
- dest="exclude_antibodies", action="store_true",
- help="Including the flag --exclude-antibody-capture "
- "will cause remove-background to operate on "
- "gene counts only, ignoring other features.")
- subparser.add_argument("--learning-rate", nargs=None,
- type=float, default=1e-4,
- dest="learning_rate",
- help="Training detail: lower learning rate for "
- "inference. A OneCycle learning rate schedule "
- "is used, where the upper learning rate is ten "
- "times this value. (For this value, probably "
- "do not exceed 1e-3). (default: %(default)s)")
- subparser.add_argument("--final-elbo-fail-fraction", type=float,
- help="Training is considered to have failed if "
- "(best_test_ELBO - final_test_ELBO)/(best_test_DLBO - initial_train_ELBO) > FINAL_ELBO_FAIL_FRACTION."
- "(default: do not fail training based on final_training_ELBO)")
- subparser.add_argument("--epoch-elbo-fail-fraction", type=float,
- help="Training is considered to have failed if "
- "(previous_epoch_test_ELBO - current_epoch_test_ELBO)/(previous_epoch_test_ELBO - initial_train_ELBO) > EPOCH_ELBO_FAIL_FRACTION."
- "(default: do not fail training based on epoch_training_ELBO)")
- subparser.add_argument("--num-training-tries", type=int, default=1,
- help="Number of times to attempt to train the model. Each subsequent "
- "attempt the learning rate is multiplied by LEARNING_RATE_RETRY_MULT. "
- "(default: %(default)s)")
- subparser.add_argument("--learning-rate-retry-mult", type=float, default=0.5,
- help="Learning rate is multiplied by this amount each time "
- "(default: %(default)s)")
-
- subparser.add_argument("--posterior-batch-size", type=int, default=consts.PROP_POSTERIOR_BATCH_SIZE,
- dest="posterior_batch_size",
- help="Size of batches when creating the posterior. Reduce this to avoid "
- "running out of GPU memory creating the posterior (will be slower). "
- "(default: %(default)s)")
- subparser.add_argument("--cells-posterior-reg-calc", type=int, default=consts.CELLS_POSTERIOR_REG_CALC,
- dest="cells_posterior_reg_calc",
- help="Number of cells used to estimate posterior regularization lambda. "
- "(default: %(default)s)")
-
- return subparsers
diff --git a/cellbender/remove_background/argparser.py b/cellbender/remove_background/argparser.py
new file mode 100644
index 00000000..d53aca3c
--- /dev/null
+++ b/cellbender/remove_background/argparser.py
@@ -0,0 +1,284 @@
+import argparse
+from cellbender.remove_background import consts
+
+
+def add_subparser_args(subparsers: argparse) -> argparse:
+ """Add tool-specific arguments for remove-background.
+
+ Args:
+ subparsers: Parser object before addition of arguments specific to
+ remove-background.
+
+ Returns:
+ parser: Parser object with additional parameters.
+
+ """
+
+ subparser = subparsers.add_parser("remove-background",
+ description="Remove background RNA "
+ "from count matrix.",
+ help="Remove background ambient RNA "
+ "and barcode-swapped reads from "
+ "a count matrix, producing a "
+ "new count matrix and "
+ "determining which barcodes "
+ "contain real cells.",
+ formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+
+ subparser.add_argument("--input", nargs=None, type=str,
+ dest='input_file',
+ required=True,
+ help="Data file on which to run tool. Data must be "
+ "un-filtered: it should include empty droplets. "
+ "The following input formats are supported: "
+ "CellRanger v2 and v3 (.h5 or the directory that "
+ "contains the .mtx file), Dropseq "
+ "DGE (.txt or .txt.gz), BD Rhapsody (.csv or "
+ ".csv.gz), AnnData (.h5ad), and Loom (.loom).")
+ subparser.add_argument("--output", nargs=None, type=str,
+ dest='output_file',
+ required=True,
+ help="Output file location (the path must "
+ "exist, and the file name must have .h5 "
+ "extension).")
+ subparser.add_argument("--cuda",
+ dest="use_cuda", action="store_true",
+ help="Including the flag --cuda will run the "
+ "inference on a GPU.")
+ subparser.add_argument("--checkpoint", nargs=None, type=str,
+ dest='input_checkpoint_tarball',
+ required=False, default=consts.CHECKPOINT_FILE_NAME,
+ help="Checkpoint tarball produced by the same version "
+ "of CellBender remove-background. If present, "
+ "and the workflow hashes match, training will "
+ "restart from this checkpoint.")
+ subparser.add_argument("--expected-cells", nargs=None, type=int,
+ default=None,
+ dest="expected_cell_count",
+ help="Number of cells expected in the dataset "
+ "(a rough estimate within a factor of 2 "
+ "is sufficient).")
+ subparser.add_argument("--total-droplets-included",
+ nargs=None, type=int, default=None,
+ dest="total_droplets",
+ help="The number of droplets from the "
+ "rank-ordered UMI plot that will have their "
+ "cell probabilities inferred as an output. "
+ "Include the droplets which might contain "
+ "cells. Droplets beyond TOTAL_DROPLETS_INCLUDED "
+ "should be 'surely empty' droplets.")
+ subparser.add_argument("--force-cell-umi-prior",
+ nargs=None, type=float, default=None,
+ dest="force_cell_umi_prior",
+ help="Ignore CellBender's heuristic prior estimation, "
+ "and use this prior for UMI counts in cells.")
+ subparser.add_argument("--force-empty-umi-prior",
+ nargs=None, type=float, default=None,
+ dest="force_empty_umi_prior",
+ help="Ignore CellBender's heuristic prior estimation, "
+ "and use this prior for UMI counts in empty droplets.")
+ subparser.add_argument("--model", nargs=None, type=str,
+ default="full",
+ choices=["naive", "simple", "ambient", "swapping", "full"],
+ dest="model",
+ help="Which model is being used for count data. "
+ "'naive' subtracts the estimated ambient profile."
+ " 'simple' does not model either ambient "
+ "RNA or random barcode swapping (for "
+ "debugging purposes -- not recommended). "
+ "'ambient' assumes background RNA is "
+ "incorporated into droplets. 'swapping' "
+ "assumes background RNA comes from random "
+ "barcode swapping (via PCR chimeras). 'full' "
+ "uses a combined ambient and swapping model.")
+ subparser.add_argument("--epochs", nargs=None, type=int, default=150,
+ dest="epochs",
+ help="Number of epochs to train.")
+ subparser.add_argument("--low-count-threshold", type=int,
+ default=consts.LOW_UMI_CUTOFF,
+ dest="low_count_threshold",
+ help="Droplets with UMI counts below this "
+ "number are completely excluded from the "
+ "analysis. This can help identify the "
+ "correct prior for empty droplet counts "
+ "in the rare case where empty counts "
+ "are extremely high (over 200).")
+ subparser.add_argument("--z-dim", type=int, default=64,
+ dest="z_dim",
+ help="Dimension of latent variable z.")
+ subparser.add_argument("--z-layers", nargs="+", type=int, default=[512],
+ dest="z_hidden_dims",
+ help="Dimension of hidden layers in the encoder for z.")
+ subparser.add_argument("--training-fraction",
+ type=float, nargs=None,
+ default=consts.TRAINING_FRACTION,
+ dest="training_fraction",
+ help="Training detail: the fraction of the "
+ "data used for training. The rest is never "
+ "seen by the inference algorithm. Speeds up "
+ "learning.")
+ subparser.add_argument("--empty-drop-training-fraction",
+ type=float, nargs=None,
+ default=consts.FRACTION_EMPTIES,
+ dest="fraction_empties",
+ help="Training detail: the fraction of the "
+ "training data each epoch "
+ "that is drawn (randomly sampled) from "
+ "surely empty droplets.")
+ subparser.add_argument("--ignore-features", nargs="+", type=int,
+ default=[],
+ dest="blacklisted_genes",
+ help="Integer indices of features to ignore "
+ "entirely. In the output count matrix, the "
+ "counts for these features will be unchanged.")
+ subparser.add_argument("--fpr", nargs="+",
+ default=[0.01],
+ dest="fpr",
+ help="Target 'delta' false positive rate in [0, 1). "
+ "Use 0 for a cohort of samples which will be "
+ "jointly analyzed for differential expression. "
+ "A false positive is a true signal count that is "
+ "erroneously removed. More background removal "
+ "is accompanied by more signal removal "
+ "at high values of FPR. You can specify "
+ "multiple values, which will create multiple "
+ "output files.")
+ subparser.add_argument("--exclude-feature-types",
+ type=str, nargs="+", default=[],
+ dest="exclude_features",
+ help="Feature types to ignore during the analysis. "
+ "These features will be left unchanged in the "
+ "output file.")
+ subparser.add_argument("--projected-ambient-count-threshold",
+ type=float, nargs=None,
+ default=consts.AMBIENT_COUNTS_IN_CELLS_LOW_LIMIT,
+ dest="ambient_counts_in_cells_low_limit",
+ help="Controls how many features are included in the "
+ "analysis, which can lead to a large speedup. "
+ "If a feature is expected to have less than "
+ "PROJECTED_AMBIENT_COUNT_THRESHOLD counts total "
+ "in all cells (summed), then that gene is "
+ "excluded, and it will be unchanged in the "
+ "output count matrix. For example, "
+ "PROJECTED_AMBIENT_COUNT_THRESHOLD = 0 will "
+ "include all features which have even a single "
+ "count in any empty droplet.")
+ subparser.add_argument("--learning-rate",
+ type=float, default=1e-4,
+ dest="learning_rate",
+ help="Training detail: lower learning rate for "
+ "inference. A OneCycle learning rate schedule "
+ "is used, where the upper learning rate is ten "
+ "times this value. (For this value, probably "
+ "do not exceed 1e-3).")
+ subparser.add_argument("--checkpoint-mins",
+ type=float, default=7.,
+ dest="checkpoint_min",
+ help="Checkpoint file will be saved periodically, "
+ "with this many minutes between each checkpoint.")
+ subparser.add_argument("--final-elbo-fail-fraction", type=float,
+ dest="final_elbo_fail_fraction",
+ help="Training is considered to have failed if "
+ "(best_test_ELBO - final_test_ELBO)/(best_test_ELBO "
+ "- initial_test_ELBO) > FINAL_ELBO_FAIL_FRACTION. Training will "
+ "automatically re-run if --num-training-tries > "
+ "1. By default, will not fail training based "
+ "on final_training_ELBO.")
+ subparser.add_argument("--epoch-elbo-fail-fraction", type=float,
+ dest="epoch_elbo_fail_fraction",
+ help="Training is considered to have failed if "
+ "(previous_epoch_test_ELBO - current_epoch_test_ELBO)"
+ "/(previous_epoch_test_ELBO - initial_train_ELBO) "
+ "> EPOCH_ELBO_FAIL_FRACTION. Training will "
+ "automatically re-run if --num-training-tries > "
+ "1. By default, will not fail training based "
+ "on epoch_training_ELBO.")
+ subparser.add_argument("--num-training-tries", type=int, default=1,
+ dest="num_training_tries",
+ help="Number of times to attempt to train the model. "
+ "At each subsequent attempt, the learning rate is "
+ "multiplied by LEARNING_RATE_RETRY_MULT.")
+ subparser.add_argument("--learning-rate-retry-mult", type=float, default=0.2,
+ dest="learning_rate_retry_mult",
+ help="Learning rate is multiplied by this amount each "
+ "time a new training attempt is made. (This "
+ "parameter is only used if training fails based "
+ "on EPOCH_ELBO_FAIL_FRACTION or "
+ "FINAL_ELBO_FAIL_FRACTION and NUM_TRAINING_TRIES"
+ " is > 1.)")
+ subparser.add_argument("--posterior-batch-size", type=int,
+ default=consts.PROB_POSTERIOR_BATCH_SIZE,
+ dest="posterior_batch_size",
+ help="Training detail: size of batches when creating "
+ "the posterior. Reduce this to avoid running "
+ "out of GPU memory creating the posterior "
+ "(will be slower).")
+ subparser.add_argument("--posterior-regularization", type=str,
+ default=None,
+ choices=["PRq", "PRmu", "PRmu_gene"],
+ dest="posterior_regularization",
+ help="Posterior regularization method. (For experts: "
+ "not required for normal usage, see "
+ "documentation). PRq is approximate quantile-"
+ "targeting. PRmu is approximate mean-targeting "
+ "aggregated over genes (behavior of v0.2.0). "
+ "PRmu_gene is approximate mean-targeting per "
+ "gene.")
+ subparser.add_argument("--alpha",
+ type=float, default=None,
+ dest="prq_alpha",
+ help="Tunable parameter alpha for the PRq posterior "
+ "regularization method (not normally used: see "
+ "documentation).")
+ subparser.add_argument("--q",
+ type=float, default=None,
+ dest="cdf_threshold_q",
+ help="Tunable parameter q for the CDF threshold "
+ "estimation method (not normally used: see "
+ "documentation).")
+ subparser.add_argument("--estimator", type=str,
+ default="mckp",
+ choices=["map", "mean", "cdf", "sample", "mckp"],
+ dest="estimator",
+ help="Output denoised count estimation method. (For "
+ "experts: not required for normal usage, see "
+ "documentation).")
+ subparser.add_argument("--estimator-multiple-cpu",
+ dest="use_multiprocessing_estimation", action="store_true",
+ default=False,
+ help="Including the flag --estimator-multiple-cpu will "
+ "use more than one CPU to compute the MCKP "
+ "output count estimator in parallel (does "
+ "nothing for other estimators).")
+ subparser.add_argument("--constant-learning-rate",
+ dest="constant_learning_rate", action="store_true",
+ default=False,
+ help="Including the flag --constant-learning-rate will "
+ "use the ClippedAdam optimizer instead of the "
+ "OneCycleLR learning rate schedule, which is "
+ "the default. Learning is faster with the "
+ "OneCycleLR schedule. However, training can "
+ "easily be continued from a checkpoint for more "
+ "epochs than the initial command specified when "
+ "using ClippedAdam. On the other hand, if using "
+ "the OneCycleLR schedule with 150 epochs "
+ "specified, it is not possible to pick up from "
+ "that final checkpoint and continue training "
+ "until 250 epochs.")
+ subparser.add_argument("--cpu-threads",
+ type=int, default=None,
+ dest="n_threads",
+ help="Number of threads to use when pytorch is run "
+ "on CPU. Defaults to the number of logical cores.")
+ subparser.add_argument("--debug",
+ dest="debug", action="store_true",
+ help="Including the flag --debug will log "
+ "extra messages useful for debugging.")
+ subparser.add_argument("--truth",
+ type=str, default=None,
+ dest="truth_file",
+ help="This is only used by developers for report "
+ "generation. Truth h5 file (for "
+ "simulated data only).")
+
+ return subparsers
diff --git a/cellbender/remove_background/checkpoint.py b/cellbender/remove_background/checkpoint.py
new file mode 100644
index 00000000..5b0903de
--- /dev/null
+++ b/cellbender/remove_background/checkpoint.py
@@ -0,0 +1,368 @@
+"""Handle the saving and loading of models from checkpoint files."""
+
+from cellbender.remove_background import consts
+from cellbender.remove_background.data.dataprep import DataLoader
+
+import torch
+import numpy as np
+from typing import Tuple, List, Optional, Union, Dict
+import logging
+import argparse
+import hashlib
+import os
+import tarfile
+import glob
+import random
+import pickle
+import tempfile
+import shutil
+import traceback
+
+
+logger = logging.getLogger('cellbender')
+
+USE_PYRO = True
+try:
+ import pyro
+except ImportError:
+ USE_PYRO = False
+USE_CUDA = torch.cuda.is_available()
+
+
+def save_random_state(filebase: str) -> List[str]:
+ """Write states of various random number generators to files.
+
+ NOTE: the pyro.util.get_rng_state() is a compilation of python, numpy, and
+ torch random states. Here, we save the three explicitly ourselves.
+ This is useful for potential future usages outside of pyro.
+ """
+ # https://stackoverflow.com/questions/32808686/storing-a-random-state/32809283
+
+ file_dict = {}
+
+ # Random state information
+ if USE_PYRO:
+ pyro_random_state = pyro.util.get_rng_state() # this is shorthand for the following
+ file_dict.update({filebase + '_random.pyro': pyro_random_state})
+ else:
+ python_random_state = random.getstate()
+ numpy_random_state = np.random.get_state()
+ torch_random_state = torch.get_rng_state()
+ file_dict.update({filebase + '_random.python': python_random_state,
+ filebase + '_random.numpy': numpy_random_state,
+ filebase + '_random.torch': torch_random_state})
+ if USE_CUDA:
+ cuda_random_state = torch.cuda.get_rng_state_all()
+ file_dict.update({filebase + '_random.cuda': cuda_random_state})
+
+ # Save it
+ for file, state in file_dict.items():
+ with open(file, 'wb') as f:
+ pickle.dump(state, f)
+
+ return list(file_dict.keys())
+
+
+def load_random_state(filebase: str):
+ """Load random states from files and update generators with them."""
+
+ if USE_PYRO:
+ with open(filebase + '_random.pyro', 'rb') as f:
+ pyro_random_state = pickle.load(f)
+ pyro.util.set_rng_state(pyro_random_state)
+
+ else:
+ with open(filebase + '_random.python', 'rb') as f:
+ python_random_state = pickle.load(f)
+ random.setstate(python_random_state)
+
+ with open(filebase + '_random.numpy', 'rb') as f:
+ numpy_random_state = pickle.load(f)
+ np.random.set_state(numpy_random_state)
+
+ with open(filebase + '_random.torch', 'rb') as f:
+ torch_random_state = pickle.load(f)
+ torch.set_rng_state(torch_random_state)
+
+ if USE_CUDA:
+ with open(filebase + '_random.cuda', 'rb') as f:
+ cuda_random_state = pickle.load(f)
+ torch.cuda.set_rng_state_all(cuda_random_state)
+
+
+def save_checkpoint(filebase: str,
+ model_obj: 'RemoveBackgroundPyroModel',
+ scheduler: pyro.optim.PyroOptim,
+ args: argparse.Namespace,
+ train_loader: Optional[DataLoader] = None,
+ test_loader: Optional[DataLoader] = None,
+ tarball_name: str = consts.CHECKPOINT_FILE_NAME) -> bool:
+ """Save trained model and optimizer state in a checkpoint file.
+ Any hyperparameters or metadata should be part of the model object."""
+
+ logger.info(f'Saving a checkpoint...')
+
+ try:
+
+ # Work in a temporary directory.
+ with tempfile.TemporaryDirectory() as tmp_dir:
+
+ basename = os.path.basename(filebase)
+ filebase = os.path.join(tmp_dir, basename)
+
+ file_list = save_random_state(filebase=filebase)
+
+ torch.save(model_obj, filebase + '_model.torch')
+ torch.save(scheduler, filebase + '_optim.torch')
+ scheduler.save(filebase + '_optim.pyro') # use PyroOptim method
+ pyro.get_param_store().save(filebase + '_params.pyro')
+ file_list = file_list + [filebase + '_model.torch',
+ filebase + '_optim.torch',
+ filebase + '_optim.pyro',
+ filebase + '_params.pyro']
+
+ if train_loader is not None:
+ # train_loader_file = save_dataloader_state(filebase=filebase,
+ # data_loader_state=train_loader.get_state(),
+ # name='train')
+ torch.save(train_loader, filebase + '_train.loaderstate')
+ file_list.append(filebase + '_train.loaderstate')
+ if test_loader is not None:
+ # test_loader_file = save_dataloader_state(filebase=filebase,
+ # data_loader_state=test_loader.get_state(),
+ # name='test')
+ torch.save(test_loader, filebase + '_test.loaderstate')
+ file_list.append(filebase + '_test.loaderstate')
+
+ np.save(filebase + '_args.npy', args)
+ file_list.append(filebase + '_args.npy')
+
+ make_tarball(files=file_list, tarball_name=tarball_name)
+
+ logger.info(f'Saved checkpoint as {os.path.abspath(tarball_name)}')
+ return True
+
+ except KeyboardInterrupt:
+ logger.warning('Keyboard interrupt: will not save checkpoint')
+ return False
+
+ except Exception:
+ logger.warning('Could not save checkpoint')
+ logger.warning(traceback.format_exc())
+ return False
+
+
+def load_checkpoint(filebase: Optional[str],
+ tarball_name: str = consts.CHECKPOINT_FILE_NAME,
+ force_device: Optional[str] = None)\
+ -> Dict[str, Union['RemoveBackgroundPyroModel', pyro.optim.PyroOptim, DataLoader, bool]]:
+ """Load checkpoint and prepare a RemoveBackgroundPyroModel and optimizer."""
+
+ out = load_from_checkpoint(
+ filebase=filebase,
+ tarball_name=tarball_name,
+ to_load=['model', 'optim', 'param_store', 'dataloader', 'args', 'random_state'],
+ force_device=force_device,
+ )
+ out.update({'loaded': True})
+ logger.info(f'Loaded partially-trained checkpoint from {tarball_name}')
+ return out
+
+
+def load_from_checkpoint(filebase: Optional[str],
+ tarball_name: str = consts.CHECKPOINT_FILE_NAME,
+ to_load: List[str] = ['model'],
+ force_device: Optional[str] = None) -> Dict:
+ """Load specific files from a checkpoint tarball."""
+
+ load_kwargs = {}
+ if force_device is not None:
+ load_kwargs.update({'map_location': torch.device(force_device)})
+
+ # Work in a temporary directory.
+ with tempfile.TemporaryDirectory() as tmp_dir:
+
+ # Unpack the checkpoint tarball.
+ logger.info(f'Attempting to unpack tarball "{tarball_name}" to {tmp_dir}')
+ success = unpack_tarball(tarball_name=tarball_name, directory=tmp_dir)
+ if success:
+ unpacked_files = '\n'.join(glob.glob(os.path.join(tmp_dir, "*")))
+ logger.info(f'Successfully unpacked tarball to {tmp_dir}\n'
+ f'{unpacked_files}')
+ else:
+ # no tarball loaded, so do not continue trying to load files
+ raise FileNotFoundError
+
+ # See if files have a hash matching input filebase.
+ if filebase is not None:
+ basename = os.path.basename(filebase)
+ filebase = os.path.join(tmp_dir, basename)
+ logger.debug(f'Looking for files with base name matching {filebase}*')
+ if not os.path.exists(filebase + '_model.torch'):
+ logger.info('Workflow hash does not match that of checkpoint.')
+ raise ValueError('Workflow hash does not match that of checkpoint.')
+ else:
+ filebase = (glob.glob(os.path.join(tmp_dir, '*_model.torch'))[0]
+ .replace('_model.torch', ''))
+ logger.debug(f'Accepting any file hash, so loading {filebase}*')
+
+ out = {}
+
+ # Load the saved model.
+ if 'model' in to_load:
+ model_obj = torch.load(filebase + '_model.torch', **load_kwargs)
+ logger.debug('Model loaded from ' + filebase + '_model.torch')
+ out.update({'model': model_obj})
+
+ # Load the saved optimizer.
+ if 'optim' in to_load:
+ scheduler = torch.load(filebase + '_optim.torch', **load_kwargs)
+ scheduler.load(filebase + '_optim.pyro', **load_kwargs) # use PyroOptim method
+ logger.debug('Optimizer loaded from ' + filebase + '_optim.*')
+ out.update({'optim': scheduler})
+
+ # Load the pyro param store.
+ if 'param_store' in to_load:
+ pyro.get_param_store().load(filebase + '_params.pyro', map_location=force_device)
+ logger.debug('Pyro param store loaded from ' + filebase + '_params.pyro')
+
+ # Load dataloader states.
+ if 'dataloader' in to_load:
+ # load_dataloader_state(data_loader=train_loader, file=filebase + '_train.loaderstate')
+ # load_dataloader_state(data_loader=test_loader, file=filebase + '_test.loaderstate')
+ train_loader = None
+ test_loader = None
+ if os.path.exists(filebase + '_train.loaderstate'):
+ train_loader = torch.load(filebase + '_train.loaderstate', **load_kwargs)
+ logger.debug('Train loader loaded from ' + filebase + '_train.loaderstate')
+ out.update({'train_loader': train_loader})
+ if os.path.exists(filebase + '_test.loaderstate'):
+ test_loader = torch.load(filebase + '_test.loaderstate', **load_kwargs)
+ logger.debug('Test loader loaded from ' + filebase + '_test.loaderstate')
+ out.update({'test_loader': test_loader})
+
+ # Load args, which can be modified in the case of auto-learning-rate updates.
+ if 'args' in to_load:
+ args = np.load(filebase + '_args.npy', allow_pickle=True).item()
+ out.update({'args': args})
+
+ # Update states of random number generators across the board.
+ if 'random_state' in to_load:
+ load_random_state(filebase=filebase)
+ logger.debug('Loaded random state globally for python, numpy, pytorch, and cuda')
+
+ # Copy the posterior file outside the temp dir so it can be loaded later.
+ if 'posterior' in to_load:
+ posterior_file = os.path.join(os.path.dirname(filebase), 'posterior.h5')
+ if os.path.exists(posterior_file):
+ shutil.copyfile(posterior_file, 'posterior.h5')
+ out.update({'posterior_file': 'posterior.h5'})
+ logger.debug(f'Copied posterior file from {posterior_file} to posterior.h5')
+ else:
+ logger.debug('Trying to load posterior in load_from_checkpoint(), but posterior '
+ 'is not present in checkpoint tarball.')
+
+ return out
+
+
+def attempt_load_checkpoint(filebase: str,
+ tarball_name: str = consts.CHECKPOINT_FILE_NAME,
+ force_device: Optional[str] = None)\
+ -> Dict[str, Union['RemoveBackgroundPyroModel', pyro.optim.PyroOptim, DataLoader, bool]]:
+ """Load checkpoint and prepare a RemoveBackgroundPyroModel and optimizer,
+ or return the inputs if loading fails."""
+
+ try:
+ logger.debug('Attempting to load checkpoint from ' + tarball_name)
+ return load_checkpoint(filebase=filebase,
+ tarball_name=tarball_name,
+ force_device=force_device)
+
+ except FileNotFoundError:
+ logger.debug('No tarball found')
+ return {'loaded': False}
+
+ except ValueError:
+ logger.debug('Unpacked tarball files have a different workflow hash: will not load')
+ return {'loaded': False}
+
+
+def make_tarball(files: List[str], tarball_name: str) -> bool:
+ """Tar and gzip a list of files as a checkpoint tarball.
+ NOTE: used by automated checkpoint handling in Cromwell
+ NOTE2: .tmp file is used so that incomplete checkpoint files do not exist
+ even temporarily
+ """
+ with tarfile.open(tarball_name + '.tmp', 'w:gz', compresslevel=1) as tar:
+ for file in files:
+ # without arcname, unpacking results in unpredictable file locations!
+ tar.add(file, arcname=os.path.basename(file))
+ os.rename(tarball_name + '.tmp', tarball_name)
+ return True
+
+
+def unpack_tarball(tarball_name: str, directory: str) -> bool:
+ """Untar a checkpoint tarball and put the files in directory."""
+ if not os.path.exists(tarball_name):
+ logger.info("No saved checkpoint.")
+ return False
+
+ try:
+ with tarfile.open(tarball_name, 'r:gz') as tar:
+ tar.extractall(path=directory)
+ return True
+
+ except Exception:
+ logger.warning("Failed to unpack existing tarball.")
+ return False
+
+
+def create_workflow_hashcode(module_path: str,
+ args: argparse.Namespace,
+ args_to_remove: List[str] = ['epochs', 'fpr'],
+ name: str = 'md5',
+ verbose: bool = False) -> str:
+ """Create a hash blob from cellbender python code plus input arguments."""
+
+ hasher = hashlib.new(name=name)
+
+ files_safe_to_ignore = ['infer.py', 'simulate.py', 'report.py',
+ 'downstream.py', 'monitor.py', 'fpr.py',
+ 'posterior.py', 'estimation.py', 'sparse_utils.py']
+
+ if not os.path.exists(module_path):
+ return ''
+
+ try:
+
+ # files
+ for root, _, files in os.walk(module_path):
+ for file in files:
+ if not file.endswith('.py'):
+ continue
+ if file in files_safe_to_ignore:
+ continue
+ if 'test' in file:
+ continue
+ if verbose:
+ print(file)
+ with open(os.path.join(root, file), 'rb') as f:
+ buf = b'\n'.join(f.readlines())
+ hasher.update(buf) # encode python files
+
+ # inputs
+ args_dict = vars(args).copy()
+ for arg in args_to_remove:
+ args_dict.pop(arg, None)
+ hasher.update(str(args_dict).encode('utf-8')) # encode parsed input args
+
+ # input file
+ # TODO
+ # this is probably not necessary for real data... why would two different
+ # files have the same name?
+ # but it's useful for development where simulated datasets change
+
+ except Exception:
+ return ''
+
+ return hasher.hexdigest()
diff --git a/cellbender/remove_background/cli.py b/cellbender/remove_background/cli.py
index 16e93e88..ad49c018 100644
--- a/cellbender/remove_background/cli.py
+++ b/cellbender/remove_background/cli.py
@@ -1,15 +1,17 @@
"""Command-line tool functionality for remove-background."""
import torch
-from cellbender.remove_background.data.dataset import SingleCellRNACountsDataset
-from cellbender.remove_background.train import run_inference
-from cellbender.base_cli import AbstractCLI
-from cellbender.remove_background import consts
+import cellbender
+
+from cellbender.base_cli import AbstractCLI, get_version
+from cellbender.remove_background.checkpoint import create_workflow_hashcode
+from cellbender.remove_background.run import run_remove_background
+from cellbender.remove_background.posterior import Posterior
import logging
import os
import sys
-from datetime import datetime
+import argparse
class CLI(AbstractCLI):
@@ -22,16 +24,25 @@ def __init__(self):
def get_name(self) -> str:
return self.name
- def validate_args(self, args):
+ @staticmethod
+ def validate_args(args) -> argparse.Namespace:
"""Validate parsed arguments."""
# Ensure that if there's a tilde for $HOME in the file path, it works.
try:
args.input_file = os.path.expanduser(args.input_file)
args.output_file = os.path.expanduser(args.output_file)
+ if args.truth_file is not None:
+ args.truth_file = os.path.expanduser(args.truth_file)
except TypeError:
raise ValueError("Problem with provided input and output paths.")
+ # Ensure that if truth data is specified, it is accessible
+ if args.truth_file is not None:
+ assert os.access(args.truth_file, os.R_OK), \
+ f"Cannot read specified simulated truth file {args.truth_file}. " \
+ f"Ensure the file exists and is read accessible."
+
# Ensure write access to the save directory.
file_dir, _ = os.path.split(args.output_file) # Get the output directory
if file_dir:
@@ -76,6 +87,10 @@ def validate_args(self, args):
"significant speed-ups.\n\n")
sys.stdout.flush() # Write immediately
+ # Make sure n_threads makes sense.
+ if args.n_threads is not None:
+ assert args.n_threads > 0, "--cpu-threads must be an integer >= 1"
+
# Ensure all network layer dimensions are positive.
for n in args.z_hidden_dims:
assert n > 0, "--z-layers must be all positive integers."
@@ -88,117 +103,131 @@ def validate_args(self, args):
args.use_jit = False
# Ensure false positive rate is between zero and one.
+ fpr_list_correct_dtypes = [] # for a mix of floats and strings later on
for fpr in args.fpr:
- assert (fpr > 0.) and (fpr < 1.), \
- "False positive rate --fpr must be between 0 and 1."
-
- assert args.num_training_tries > 0, "num-training-tries must be > 0"
- assert args.epoch_elbo_fail_fraction is None or args.epoch_elbo_fail_fraction > 0., \
- "--epoch-elbo-fail-fraction must be > 0"
- assert args.final_elbo_fail_fraction is None or args.final_elbo_fail_fraction > 0., \
- "--epoch-elbo-fail-fraction must be > 0"
-
- self.args = args
+ try:
+ fpr = float(fpr)
+ assert (fpr >= 0.) and (fpr < 1.), \
+ "False positive rate --fpr must be in [0, 1)"
+ except ValueError:
+ # the input is not a float
+ assert fpr == 'cohort', \
+ "The only allowed non-float value for FPR is the word 'cohort'."
+ fpr_list_correct_dtypes.append(fpr)
+ args.fpr = fpr_list_correct_dtypes
+
+ # Ensure that "exclude_features" specifies allowed features.
+ # As of CellRanger 6.0, the possible features are:
+ # Gene Expression
+ # Antibody Capture
+ # CRISPR Guide Capture
+ # Custom
+ # Peaks
+ allowed_features = ['Gene Expression', 'Antibody Capture',
+ 'CRISPR Guide Capture', 'Custom', 'Peaks']
+ for feature in args.exclude_features:
+ if feature not in allowed_features:
+ sys.stdout.write(f"Specified '{feature}' using --exclude-feature-types, "
+ f"but this is not a valid CellRanger feature "
+ f"designation: {allowed_features}. Ensure that "
+ f"this feature appears in your dataset, and "
+ f"ensure that this log file makes note of the "
+ f"exclusion of the appropriate features below.\n\n")
+ sys.stdout.flush() # Write immediately
+ if 'Gene Expression' in args.exclude_features:
+ sys.stdout.write("WARNING: Excluding 'Gene Expression' features from the analysis "
+ "is not recommended, since other features alone are typically "
+ "too sparse to form a good prior on cell type, and CellBender "
+ "relies on being able to construct this sort of prior\n\n")
+ sys.stdout.flush() # Write immediately
+
+ # Automatic training failures and restarts.
+ assert args.num_training_tries > 0, "--num-training-tries must be > 0 (default 1)"
+ if args.epoch_elbo_fail_fraction is not None:
+ assert (args.epoch_elbo_fail_fraction > 0.), \
+ "--epoch-elbo-fail-fraction must be in > 0"
+ if args.final_elbo_fail_fraction is not None:
+ assert (args.final_elbo_fail_fraction > 0.), \
+ "--final-elbo-fail-fraction must be in > 0"
+
+ # Ensure timing of checkpoints is within bounds.
+ assert args.checkpoint_min > 0, "--checkpoint-min must be > 0"
+ if args.checkpoint_min > 15:
+ sys.stdout.write(f"Warning: Timing between checkpoints is specified as "
+ f"{args.checkpoint_min} minutes. Consider reducing "
+ f"this number if you are concerned about the "
+ f"possibility of lost work upon preemption.\n\n")
+ sys.stdout.flush() # Write immediately
+
+ # Posterior regularization checking.
+ if args.cdf_threshold_q is not None:
+ assert (args.cdf_threshold_q >= 0.) and (args.cdf_threshold_q <= 1.), \
+ f"Argument --q must be in range [0, 1] since it is a CDF threshold."
+ if args.posterior_regularization == 'PRq':
+ # We need q for the CDF threshold estimator.
+ assert args.prq_alpha is not None, \
+ 'Input argument --alpha must be specified when using ' \
+ '--posterior-regularization PRq'
+
+ # Estimator checking.
+ if args.estimator == 'cdf':
+ # We need q for the CDF threshold estimator.
+ assert args.cdf_threshold_q is not None, \
+ 'Input argument --q must be specified when using --estimator cdf'
return args
- def run(self, args):
+ @staticmethod
+ def run(args) -> Posterior:
"""Run the main tool functionality on parsed arguments."""
# Run the tool.
- main(args)
-
+ return main(args)
-def run_remove_background(args):
- """The full script for the command line tool to remove background RNA.
- Args:
- args: Inputs from the command line, already parsed using argparse.
-
- Note: Returns nothing, but writes output to a file(s) specified from
- command line.
-
- """
-
- # Load dataset, run inference, and write the output to a file.
+def setup_and_logging(args):
+ """Take command-line input, parse arguments, and run tests or tool."""
# Send logging messages to stdout as well as a log file.
file_dir, file_base = os.path.split(args.output_file)
file_name = os.path.splitext(os.path.basename(file_base))[0]
log_file = os.path.join(file_dir, file_name + ".log")
- logging.basicConfig(level=logging.INFO,
- format="cellbender:remove-background: %(message)s",
- filename=log_file,
- filemode="w")
- console = logging.StreamHandler()
- formatter = logging.Formatter("cellbender:remove-background: "
- "%(message)s")
- console.setFormatter(formatter) # Use the same format for stdout.
- logging.getLogger('').addHandler(console) # Log to stdout and a file.
+ logger = logging.getLogger('cellbender') # name of the logger
+ logger.setLevel(logging.INFO if not args.debug else logging.DEBUG)
+ formatter = logging.Formatter('cellbender:remove-background: %(message)s')
+ file_handler = logging.FileHandler(filename=log_file, mode='w', encoding='UTF-8')
+ console_handler = logging.StreamHandler()
+ file_handler.setFormatter(formatter) # set the file format
+ console_handler.setFormatter(formatter) # use the same format for stdout
+ logger.addHandler(file_handler) # log to file
+ logger.addHandler(console_handler) # log to stdout
# Log the command as typed by user.
- logging.info("Command:\n" + ' '.join(['cellbender', 'remove-background']
- + sys.argv[2:]))
-
- # Log the start time.
- logging.info(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
-
- logging.info("Running remove-background")
-
- # Load data from file and choose barcodes and genes to analyze.
- try:
- dataset_obj = \
- SingleCellRNACountsDataset(input_file=args.input_file,
- expected_cell_count=args.expected_cell_count,
- total_droplet_barcodes=args.total_droplets,
- fraction_empties=args.fraction_empties,
- model_name=args.model,
- gene_blacklist=args.blacklisted_genes,
- exclude_antibodies=args.exclude_antibodies,
- low_count_threshold=args.low_count_threshold,
- fpr=args.fpr)
-
- except OSError:
- logging.error(f"OSError: Unable to open file {args.input_file}.")
- sys.exit(1)
-
- # Instantiate latent variable model and run full inference procedure.
- inferred_model = run_inference(dataset_obj, args)
-
- # Write outputs to file.
- try:
- dataset_obj.save_to_output_file(args.output_file,
- inferred_model,
- posterior_batch_size=args.posterior_batch_size,
- cells_posterior_reg_calc=args.cells_posterior_reg_calc,
- save_plots=True)
-
- logging.info("Completed remove-background.")
- logging.info(datetime.now().strftime('%Y-%m-%d %H:%M:%S\n'))
-
- # The exception allows user to end inference prematurely with CTRL-C.
- except KeyboardInterrupt:
-
- # If partial output has been saved, delete it.
- full_file = args.output_file
-
- # Name of the filtered (cells only) file.
- file_dir, file_base = os.path.split(full_file)
- file_name = os.path.splitext(os.path.basename(file_base))[0]
- filtered_file = os.path.join(file_dir,
- file_name + "_filtered.h5")
-
- if os.path.exists(full_file):
- os.remove(full_file)
-
- if os.path.exists(filtered_file):
- os.remove(filtered_file)
-
- logging.info("Keyboard interrupt. Terminated without saving.\n")
-
-
-def main(args):
+ logger.info("Command:\n"
+ + ' '.join(['cellbender', 'remove-background'] + sys.argv[2:]))
+ logger.info("CellBender " + get_version())
+
+ # Set up checkpointing by creating a unique workflow hash.
+ hashcode = create_workflow_hashcode(
+ module_path=os.path.dirname(cellbender.__file__),
+ args_to_remove=(['output_file', 'fpr', 'input_checkpoint_tarball', 'debug',
+ 'posterior_batch_size', 'checkpoint_min', 'truth_file',
+ 'posterior_regularization', 'cdf_threshold_q', 'prq_alpha',
+ 'estimator', 'use_multiprocessing_estimation', 'cpu_threads']
+ + (['epochs'] if args.constant_learning_rate else [])),
+ args=args)[:10]
+ args.checkpoint_filename = hashcode # store this in args
+ logger.info(f'(Workflow hash {hashcode})')
+ return args, file_handler
+
+
+def main(args) -> Posterior:
"""Take command-line input, parse arguments, and run tests or tool."""
+ args, file_handler = setup_and_logging(args)
+
# Run the tool.
- run_remove_background(args)
+ posterior = run_remove_background(args)
+ file_handler.close()
+
+ return posterior
diff --git a/cellbender/remove_background/consts.py b/cellbender/remove_background/consts.py
index 2874913a..ee5fc11c 100644
--- a/cellbender/remove_background/consts.py
+++ b/cellbender/remove_background/consts.py
@@ -1,13 +1,21 @@
"""Constant numbers used in remove-background."""
+# Seed for random number generators.
+RANDOM_SEED = 1234
+
# Factor by which the mode UMI count of the empty droplet plateau is
# multiplied to come up with a UMI cutoff below which no barcodes are used.
EMPIRICAL_LOW_UMI_TO_EMPTY_DROPLET_THRESHOLD = 0.5
+# Features with fewer than this many counts summed over all empty droplets
+# are ignored during analysis, and the output count matrix for these features
+# is identical to the input.
+COUNTS_IN_EMPTIES_LOW_LIMIT = 0
+AMBIENT_COUNTS_IN_CELLS_LOW_LIMIT = 0.1
+
# Default prior for the standard deviation of the LogNormal distribution for
# cell size, used only in the case of the 'simple' model.
SIMPLE_MODEL_D_STD_PRIOR = 0.2
-D_STD_PRIOR = 0.02
# Probability cutoff for determining which droplets contain cells and which
# are empty. The droplets n with inferred probability q_n > CELL_PROB_CUTOFF
@@ -16,7 +24,7 @@
# Default UMI cutoff. Droplets with UMI less than this will be completely
# excluded from the analysis.
-LOW_UMI_CUTOFF = 15
+LOW_UMI_CUTOFF = 5
# Default total number of droplets to include in the analysis.
TOTAL_DROPLET_DEFAULT = 25000
@@ -28,32 +36,34 @@
DEFAULT_BATCH_SIZE = 128
# Fraction of totally empty droplets that makes up each minibatch, by default.
-FRACTION_EMPTIES = 0.5
+FRACTION_EMPTIES = 0.2
# Prior on rho, the swapping fraction: the two concentration parameters alpha and beta.
-RHO_ALPHA_PRIOR = 18.
-RHO_BETA_PRIOR = 200.
+RHO_ALPHA_PRIOR = 1.5
+RHO_BETA_PRIOR = 50.
# Constraints on rho posterior latents.
-RHO_PARAM_MIN = 1.
+RHO_PARAM_MIN = 0.001
RHO_PARAM_MAX = 1000.
# Prior on epsilon, the RT efficiency concentration parameter [Gamma(alpha, alpha)].
-EPSILON_PRIOR = 500.
+EPSILON_PRIOR = 50.
# Prior used for the global overdispersion parameter.
PHI_LOC_PRIOR = 0.2
PHI_SCALE_PRIOR = 0.2
+# Prior for mixture weights in Gaussian mixture model.
+GMM_ALPHA_PRIOR = 1e-1
+GMM_COMPONENTS = 10
+GMM_EPOCHS = 500
+
# Initial value of global latent scale for d_cell.
D_CELL_SCALE_INIT = 0.02
# Scale used to regularize values of logit cell probability (mean zero).
P_LOGIT_SCALE = 2.
-# Hidden layer sizes of non-z latent encoder neural network.
-ENC_HIDDEN_DIMS = [100, 50]
-
# False to use an approximate log_prob computation which is much faster.
USE_EXACT_LOG_PROB = False
@@ -64,23 +74,43 @@
NBPC_MU_EPS_SAFEGAURD = 1e-10
NBPC_ALPHA_EPS_SAFEGAURD = 1e-10
NBPC_LAM_EPS_SAFEGAURD = 1e-10
+POISSON_EPS_SAFEGAURD = 1e-10
# Scale factors for loss function regularization terms: semi-supervision.
REG_SCALE_AMBIENT_EXPRESSION = 0.01
-REG_SCALE_EMPTY_PROB = 1.0
-REG_SCALE_CELL_PROB = 10.0
+REG_SCALE_SOFT_SUPERVISION = 10.0
+
+# Regularize logit probabilities toward this lognormal distribution.
+REG_LOGIT_MEAN = 10.0
+REG_LOGIT_SCALE = 0.2
+REG_LOGIT_SOFT_SCALE = 1.0
-# Number of cells used to estimate posterior regularization lambda. Memory hungry.
+# Number of cells used to esitmate posterior regularization lambda. Memory hungry.
CELLS_POSTERIOR_REG_CALC = 100
# Posterior regularization constant's upper and lower bounds.
-POSTERIOR_REG_MIN = 0.1
+POSTERIOR_REG_MIN = 0.01
POSTERIOR_REG_MAX = 500
-POSTERIOR_REG_SEARCH_MAX_ITER = 20
+POSTERIOR_REG_SEARCH_MAX_ITER = 50
-# Minimum number of barcodes we expect in an unfiltered `h5ad` input file.
-# Throws a warning if the input has fewer than this number.
+# For AnnData h5ad files, fewer than this many barcodes will trigger a warning,
+# since it indicates that the file might be "filtered" to cells only.
MINIMUM_BARCODES_H5AD = 1e5
-# reduce this if running out of GPU memory https://github.com/broadinstitute/CellBender/issues/67
-PROP_POSTERIOR_BATCH_SIZE = 20
+# Batch size for posterior inference.
+PROB_POSTERIOR_BATCH_SIZE = 128
+
+# Name of checkpoint file.
+CHECKPOINT_FILE_NAME = 'ckpt.tar.gz'
+
+# Whether to create an extended report (for development purposes).
+EXTENDED_REPORT = False
+
+# Maximum batch size
+MAX_BATCH_SIZE = 256
+SMALLEST_ALLOWED_BATCH = 4 # BatchNorm chokes if there is only 1 cell in last batch
+
+# Guesses during prior estimation
+MAX_TOTAL_DROPLETS_GUESSED = 70000
+MAX_EMPTIES_TO_INCLUDE = 20000
+NUM_EMPTIES_INCREMENT = 20000 # if input expected_cells > heuristic prior total_drops
diff --git a/cellbender/remove_background/data/dataprep.py b/cellbender/remove_background/data/dataprep.py
index b53003fd..00e1d5d8 100644
--- a/cellbender/remove_background/data/dataprep.py
+++ b/cellbender/remove_background/data/dataprep.py
@@ -1,5 +1,8 @@
"""Helper functions for preparing dataloaders, as well as
a class to implement loading data from a sparse matrix in mini-batches.
+
+Intentionally uses global random state, and does not keep its own random number
+generator, in order to facilitate checkpointing.
"""
import numpy as np
@@ -10,7 +13,11 @@
import torch
import torch.utils.data
-from typing import Tuple, List, Optional
+import logging
+from typing import Tuple, List, Optional, Callable
+
+
+logger = logging.getLogger('cellbender')
class SparseDataset(torch.utils.data.Dataset):
@@ -51,9 +58,37 @@ def __init__(self,
batch_size: int = consts.DEFAULT_BATCH_SIZE,
fraction_empties: float = consts.FRACTION_EMPTIES,
shuffle: bool = True,
+ sort_by: Optional[Callable[[sp.csr_matrix], np.ndarray]] = None,
use_cuda: bool = True):
+ """
+ Args:
+ dataset: Droplet count matrix [cell, gene]
+ empty_drop_dataset: Surely empty droplets count matrix
+ batch_size: Number of droplets in minibatch
+ fraction_empties: Fraction of each minibatch that will consist of
+ surely empty droplets
+ shuffle: True to shuffle data. Incompatible with sort_by.
+ sort_by: Lambda function which, when applied to the sparse matrix,
+ will return values that can be sorted to give a sort order to
+ the dataset. Dataloader will load data in order of increasing
+ values. Object attributes sort_order and unsort_order will be
+ made available.
+ use_cuda: True to load data to GPU
+ """
+ if shuffle:
+ assert sort_by is None, 'Cannot sort_by and shuffle at the same time'
+ self.sort_fn = sort_by
self.dataset = dataset
- self.ind_list = np.arange(self.dataset.shape[0])
+ if self.sort_fn is not None:
+ sort_values = self.sort_fn(self.dataset)
+ sort_order = np.argsort(sort_values)
+ self.ind_list = sort_order
+ self.sort_order = sort_order.copy()
+ self._unsort_dict = {i: val for i, val in enumerate(self.sort_order)}
+ else:
+ self.ind_list = np.arange(self.dataset.shape[0])
+ self.sort_order = self.ind_list.copy()
+ self._unsort_dict = {i: i for i in self.ind_list}
self.empty_drop_dataset = empty_drop_dataset
if self.empty_drop_dataset is None:
self.empty_ind_list = np.array([])
@@ -63,27 +98,64 @@ def __init__(self,
self.fraction_empties = fraction_empties
self.cell_batch_size = int(batch_size * (1. - fraction_empties))
self.shuffle = shuffle
- self.random = np.random.RandomState(seed=1234)
self.device = 'cpu'
self.use_cuda = use_cuda
if self.use_cuda:
self.device = 'cuda'
+ self._length = None
self._reset()
+ @torch.no_grad()
+ def unsort_inds(self, bcs):
+ if self.sort_fn is None:
+ return bcs # just for speed
+ else:
+ return torch.tensor([self._unsort_dict[bc.item()] for bc in bcs], device='cpu')
+
def _reset(self):
if self.shuffle:
- self.random.shuffle(self.ind_list) # Shuffle cell inds in place
+ np.random.shuffle(self.ind_list) # Shuffle cell inds in place
self.ptr = 0
+ def get_state(self):
+ """Internal state of the data loader, used for checkpointing"""
+ return {'ind_list': self.ind_list, 'ptr': self.ptr}
+
+ def set_state(self, ind_list: np.ndarray, ptr: int):
+ self.ind_list = ind_list
+ self.ptr = ptr
+ assert self.ptr <= len(self.ind_list), \
+ f'Problem setting dataloader state: pointer ({ptr}) is outside the ' \
+ f'length of the ind_list ({len(ind_list)})'
+
+ def reset_ptr(self):
+ self.ptr = 0
+
+ @property
+ def length(self):
+ if self._length is None:
+ self._length = self._get_length()
+ return self._length
+
+ def _get_length(self):
+ # avoid the potential for an off-by-one error by just going through it
+ i = 0
+ for _ in self:
+ i += 1
+ return i
+
def __len__(self):
- return int(self.ind_list.size *
- (1 + (self.fraction_empties / (1 - self.fraction_empties)))) # ...ish
+ return self.length
def __iter__(self):
return self
def __next__(self):
- if self.ptr == self.ind_list.size:
+ # Skip last batch if the size is < smallest allowed batch
+ remaining_cells = self.ind_list.size - self.ptr
+ if remaining_cells < consts.SMALLEST_ALLOWED_BATCH:
+ if remaining_cells > 0:
+ logger.debug(f'Dropped last minibatch of {remaining_cells} cells')
self._reset()
raise StopIteration()
@@ -101,9 +173,9 @@ def __next__(self):
(1 - self.fraction_empties)))
if self.empty_ind_list.size > 0:
# This does not happen for 'simple' model.
- empty_inds = self.random.choice(self.empty_ind_list,
- size=n_empties,
- replace=True)
+ empty_inds = np.random.choice(self.empty_ind_list,
+ size=n_empties,
+ replace=True)
if empty_inds.size > 0:
csr_list = [self.dataset[cell_inds, :],
@@ -121,9 +193,8 @@ def __next__(self):
return dense_tensor.to(device=self.device)
-def prep_sparse_data_for_training(dataset: sp.csr.csr_matrix,
- empty_drop_dataset: sp.csr.csr_matrix,
- random_state: np.random.RandomState,
+def prep_sparse_data_for_training(dataset: sp.csr_matrix,
+ empty_drop_dataset: sp.csr_matrix,
training_fraction: float = consts.TRAINING_FRACTION,
fraction_empties: float = consts.FRACTION_EMPTIES,
batch_size: int = consts.DEFAULT_BATCH_SIZE,
@@ -144,8 +215,6 @@ def prep_sparse_data_for_training(dataset: sp.csr.csr_matrix,
columns are genes.
empty_drop_dataset: Matrix of gene counts, where rows are surely-empty
droplet barcodes and columns are genes.
- random_state: numpy.random.RandomState from the Dataset object, for
- deterministic outputs.
training_fraction: Fraction of data to use as the training set. The
rest becomes the test set.
fraction_empties: Fraction of each minibatch to be composed of empty
@@ -167,15 +236,14 @@ def prep_sparse_data_for_training(dataset: sp.csr.csr_matrix,
"""
# Choose train and test indices from analysis dataset.
- training_mask = random_state.rand(dataset.shape[0]) < training_fraction
+ training_mask = np.random.rand(dataset.shape[0]) < training_fraction
training_indices = [idx for idx in range(dataset.shape[0])
if training_mask[idx]]
test_indices = [idx for idx in range(dataset.shape[0])
if not training_mask[idx]]
# Choose train and test indices from empty drop dataset.
- training_mask_empty = (random_state.rand(empty_drop_dataset.shape[0])
- < training_fraction)
+ training_mask_empty = (np.random.rand(empty_drop_dataset.shape[0]) < training_fraction)
training_indices_empty = [idx for idx in range(empty_drop_dataset.shape[0])
if training_mask_empty[idx]]
test_indices_empty = [idx for idx in range(empty_drop_dataset.shape[0])
@@ -204,7 +272,7 @@ def prep_sparse_data_for_training(dataset: sp.csr.csr_matrix,
return train_loader, test_loader
-def sparse_collate(batch: List[Tuple[sp.csr.csr_matrix]]) -> torch.Tensor:
+def sparse_collate(batch: List[Tuple[sp.csr_matrix]]) -> torch.Tensor:
"""Load a minibatch of sparse data as a dense torch.Tensor in memory.
Puts each data field into a tensor with leftmost dimension batch size.
diff --git a/cellbender/remove_background/data/dataset.py b/cellbender/remove_background/data/dataset.py
index a4eb84b2..4a9f97e7 100644
--- a/cellbender/remove_background/data/dataset.py
+++ b/cellbender/remove_background/data/dataset.py
@@ -1,26 +1,25 @@
"""Class and functions for working with a count matrix dataset."""
-import tables
import numpy as np
import scipy.sparse as sp
-import scipy.io as io
-from scipy.stats import mode
-from sklearn.decomposition import PCA
import torch
-import anndata
-import cellbender.remove_background.model
import cellbender.remove_background.consts as consts
-from cellbender.remove_background.infer import ProbPosterior
from cellbender.remove_background.data.dataprep import DataLoader
-
-from typing import Dict, List, Union, Tuple, Optional
+from cellbender.remove_background.data.io import load_data
+from cellbender.remove_background.data.priors import get_priors, \
+ get_cell_count_given_expected_cells, \
+ get_empty_count_given_expected_cells_and_total_droplets, \
+ compute_crossover_surely_empty_and_stds
+from cellbender.remove_background.sparse_utils import csr_set_rows_to_zero, \
+ overwrite_matrix_with_columns_from_another
+
+from typing import Dict, List, Optional, Iterable, Callable
import logging
-import os
+import argparse
+
-import matplotlib
-matplotlib.use('Agg')
-import matplotlib.pyplot as plt # This needs to be after matplotlib.use('Agg')
+logger = logging.getLogger('cellbender')
class SingleCellRNACountsDataset:
@@ -32,12 +31,18 @@ class SingleCellRNACountsDataset:
expected_cell_count: Expected number of real cells a priori.
total_droplet_barcodes: Total number of droplets to include in the
cell-calling analysis.
+ force_cell_umi_prior: User wants to force cell UMI prior to be this
+ force_empty_umi_prior: User wants to force empty UMI prior to be this
+ ambient_counts_in_cells_low_limit: Limit to determine how many features
+ are included in the analysis
model_name: Model to use.
gene_blacklist: List of integer indices of genes to exclude entirely.
+ exclude_features: List of feature types to exclude from the analysis.
+ Must be in ['Gene Expression', 'Antibody Capture',
+ 'CRISPR Guide Capture', 'Custom']
low_count_threshold: Droplets with UMI counts below this number are
excluded entirely from the analysis.
- lambda_multiplier: Float that multiplies background estimate before
- MAP estimation.
+ fpr: Target expected false positive rate.
Attributes:
input_file: Name of data source file.
@@ -52,7 +57,7 @@ class SingleCellRNACountsDataset:
trim_dataset_for_analysis().
model_name: Name of model being run.
priors: Priors estimated from the data useful for modelling.
- posterior: Posterior estimated after inference.
+ posterior: BasePosterior estimated after inference.
empty_UMI_threshold: This UMI count is the maximum UMI count in the
user-defined surely empty droplets.
@@ -64,397 +69,388 @@ class SingleCellRNACountsDataset:
def __init__(self,
input_file: str,
model_name: str,
- exclude_antibodies: bool,
+ exclude_features: List[str],
low_count_threshold: int,
fpr: List[float],
expected_cell_count: Optional[int] = None,
- total_droplet_barcodes: int = consts.TOTAL_DROPLET_DEFAULT,
+ total_droplet_barcodes: Optional[int] = None,
+ force_cell_umi_prior: Optional[float] = None,
+ force_empty_umi_prior: Optional[float] = None,
fraction_empties: Optional[float] = None,
+ ambient_counts_in_cells_low_limit: float = consts.AMBIENT_COUNTS_IN_CELLS_LOW_LIMIT,
gene_blacklist: List[int] = []):
assert input_file is not None, "Attempting to load data, but no " \
"input file was specified."
self.input_file = input_file
self.analyzed_barcode_inds = np.array([]) # Barcodes trained each epoch
- self.analyzed_gene_inds = np.array([])
- self.empty_barcode_inds = np.array([]) # Barcodes randomized each epoch
+ self.empty_barcode_inds = np.array([]) # Barcodes sampled randomly each epoch
+ self.low_count_cutoff = low_count_threshold
self.data = None
- self.exclude_antibodies = exclude_antibodies
+ self.gmm = None
+ self.exclude_features = exclude_features
self.model_name = model_name
self.fraction_empties = fraction_empties
self.is_trimmed = False
self.low_count_threshold = low_count_threshold
- self.priors = {'n_cells': expected_cell_count} # Expected cells could be None.
+ self.ambient_counts_in_cells_low_limit = ambient_counts_in_cells_low_limit
+ self.priors = {}
self.posterior = None
self.fpr = fpr
- self.random = np.random.RandomState(seed=1234)
- self.EMPIRICAL_LOW_UMI_TO_EMPTY_DROPLET_THRESHOLD = \
- consts.EMPIRICAL_LOW_UMI_TO_EMPTY_DROPLET_THRESHOLD
- self.SIMPLE_MODEL_D_STD_PRIOR = consts.SIMPLE_MODEL_D_STD_PRIOR
- self.CELL_PROB_CUTOFF = consts.CELL_PROB_CUTOFF
-
- # Load the dataset.
- self._load_data()
-
- # Trim the dataset.
- self._trim_dataset_for_analysis(total_droplet_barcodes=total_droplet_barcodes,
- low_UMI_count_cutoff=low_count_threshold,
- gene_blacklist=gene_blacklist)
-
- # Estimate priors.
- self._estimate_priors()
- def _detect_input_data_type(self) -> str:
- """Detect the type of input data."""
-
- # Error if no input data file has been specified.
- assert self.input_file is not None, \
- "Attempting to load data, but no input file was specified."
-
- # Detect type.
- if os.path.isdir(self.input_file):
- return 'cellranger_mtx'
-
- elif os.path.splitext(self.input_file)[1] == '.h5ad':
- # assume unfiltered AnnData object, e.g. from kallisto | bustools
- return 'anndata'
-
- else:
- return 'cellranger_h5'
-
- def _load_data(self):
- """Load a dataset into the SingleCellRNACountsDataset object from
- the self.input_file"""
-
- # Detect input data type.
- data_type = self._detect_input_data_type()
+ # Are empty droplets included in this model?
+ self.include_empties = False if (self.model_name in ['simple']) else True
# Load the dataset.
- if data_type == 'cellranger_mtx':
+ self.data = load_data(self.input_file)
+ self.analyzed_gene_logic = True
- logging.info(f"Loading data from directory {self.input_file}")
- self.data = get_matrix_from_cellranger_mtx(self.input_file)
+ # Eliminate feature types not used in the analysis.
+ self._exclude_feature_types()
- elif data_type == 'cellranger_h5':
+ # Eliminate zero-count features and blacklisted features.
+ self._clean_features(gene_blacklist=gene_blacklist)
- logging.info(f"Loading data from file {self.input_file}")
- self.data = get_matrix_from_cellranger_h5(self.input_file)
-
- elif data_type == 'anndata':
-
- logging.info(f"Loading data from file {self.input_file}")
- self.data = get_matrix_from_anndata(self.input_file)
+ # Estimate priors.
+ counts = np.array(self.data['matrix']
+ [:, self.analyzed_gene_logic].sum(axis=1)).squeeze()
+ self.priors = get_priors(umi_counts=counts, low_count_threshold=low_count_threshold)
+
+ # Overwrite heuristic priors with user inputs.
+ if expected_cell_count is not None:
+ logger.debug(f'Fixing expected_cells at {expected_cell_count}')
+ self.priors['expected_cells'] = expected_cell_count
+ self.priors.update(
+ get_cell_count_given_expected_cells(
+ umi_counts=counts, expected_cells=expected_cell_count,
+ )
+ )
+ if (expected_cell_count + consts.NUM_EMPTIES_INCREMENT) > self.priors['total_droplets']:
+ # Bump this up to avoid an immediate error
+ total_drops = expected_cell_count + consts.NUM_EMPTIES_INCREMENT
+ self.priors['total_droplets'] = total_drops
+ logger.debug(f'Incrementing total_droplets to be {total_drops}')
+ if total_droplet_barcodes is None:
+ # If this isn't getting recomputed next, recompute now
+ self.priors.update(
+ get_empty_count_given_expected_cells_and_total_droplets(
+ umi_counts=counts, expected_cells=self.priors['expected_cells'],
+ total_droplets=total_drops,
+ )
+ )
+ if total_droplet_barcodes is not None:
+ logger.debug(f'Fixing total_droplets at {total_droplet_barcodes}')
+ self.priors['total_droplets'] = total_droplet_barcodes
+ self.priors.update(
+ get_empty_count_given_expected_cells_and_total_droplets(
+ umi_counts=counts, expected_cells=self.priors['expected_cells'],
+ total_droplets=total_droplet_barcodes,
+ )
+ )
+
+ # Force priors if user elects to do so.
+ if force_cell_umi_prior is not None:
+ logger.debug(f'Forcing cell UMI count prior to be {force_cell_umi_prior}')
+ self.priors['cell_counts'] = force_cell_umi_prior
+ if force_empty_umi_prior is not None:
+ logger.debug(f'Forcing empty droplet UMI count prior to be {force_empty_umi_prior}')
+ self.priors['empty_counts'] = force_empty_umi_prior
+ middle = np.sqrt(self.priors['cell_counts'] * force_empty_umi_prior)
+ self.priors['empty_count_upper_limit'] = min(middle, 2 * force_empty_umi_prior)
+
+ # Recompute a few quantities if some things were replaced by user input.
+ compute_crossover_surely_empty_and_stds(umi_counts=counts, priors=self.priors)
+ logger.info(f"Prior on counts for cells is {int(self.priors['cell_counts'])}")
+ logger.info(f"Prior on counts for empty droplets is {int(self.priors['empty_counts'])}")
+ logger.debug('\n'.join(['Priors:'] + [f'{k}: {v}' for k, v in self.priors.items()]))
+
+ # Do not analyze features which are not expected to contribute to noise.
+ self._trim_noiseless_features()
+ self.analyzed_gene_inds = np.where(self.analyzed_gene_logic)[0].astype(dtype=int)
+ logger.info(f"Including {len(self.analyzed_gene_inds)} features in the analysis.")
+
+ # Determine barcodes to be analyzed.
+ self._trim_droplets()
+ self.is_trimmed = True
- else:
- raise NotImplementedError
+ # Estimate gene expression priors.
+ self.priors.update(self._estimate_chi_ambient())
- def _trim_dataset_for_analysis(
- self,
- total_droplet_barcodes: int = consts.TOTAL_DROPLET_DEFAULT,
- low_UMI_count_cutoff: int = consts.LOW_UMI_CUTOFF,
- gene_blacklist: List[int] = []):
- """Trim the dataset for inference, choosing barcodes and genes to use.
+ logger.debug('\n'.join(['Priors:'] + [f'{k}: {v}' for k, v in self.priors.items()]))
- Sets the values of self.analyzed_barcode_inds, and
- self.empty_barcode_inds, which are used throughout training.
+ def _exclude_feature_types(self):
+ """Exclude feature types as specified by user.
+ Sets the value of self.analyzed_gene_logic
+ """
+
+ # Skip if feature types are not specified.
+ if self.data['feature_types'] is None:
+ if len(self.exclude_features) > 0:
+ logger.warning(f"WARNING: specified --exclude-feature-types "
+ f"{self.exclude_features} but no feature_type "
+ f"information was found in the input file. Proceeding "
+ f"without excluding any features. If this is not the "
+ f"intended outcome, please check your input file "
+ f"format.")
+ else:
+ logger.info('No feature type information specified in the input '
+ 'file. Including all features in the analysis.')
+ return
+
+ # Feature types are specified.
+ dtype = type(self.data['feature_types'][0])
+ is_bytestring = (dtype != str) and (dtype != np.str_)
+
+ def convert(s):
+ if is_bytestring:
+ return str(s, encoding='utf-8')
+ return s
+
+ feature_type_array = np.array([convert(f) for f in self.data['feature_types']], dtype=str)
+ feature_types = np.unique(feature_type_array)
+ feature_info = [f"{(feature_type_array == f).sum()} {f}" for f in feature_types]
+ logger.info(f"Features in dataset: {', '.join(feature_info)}")
+
+ # Keep track of a logical array for which features to include.
+ inclusion_logic = np.ones(len(feature_type_array), dtype=bool) # all True
+
+ for feature in self.exclude_features:
+ logger.info(f"Excluding {feature} features (output will equal input).")
+ logic = np.array([f not in self.exclude_features
+ for f in feature_type_array], dtype=bool)
+ inclusion_logic = np.logical_and(inclusion_logic, logic)
+ logger.info(f" - This results in the exclusion of "
+ f"{len(feature_type_array) - logic.sum()} "
+ f"features.")
+
+ # Update self.analyzed_gene_logic
+ self.analyzed_gene_logic = np.logical_and(
+ self.analyzed_gene_logic,
+ inclusion_logic,
+ )
+
+ def _clean_features(self, gene_blacklist: List[int] = []):
+ """Trim the dataset by removing zero-count and blacklisted features.
+ Sets the value of self.analyzed_gene_logic
Args:
- total_droplet_barcodes: Number of total droplets to include
- during inference each epoch, and for which to make cell calls.
- low_UMI_count_cutoff: Barcodes with total UMI counts below
- this number are excluded.
gene_blacklist: List of gene indices to trim out and exclude.
-
- Note:
- self.priors['n_cells'] is only used to choose which barcodes to
- include in the analysis, as well as to estimate the sizes of cells
- and empty droplets (elsewhere). It need only be a reasonable guess.
- The full analysis makes inferences about which barcodes contain
- cells, and ultimately will determine this number.
- However, if running only the 'simple' model, the expected_cell_count
- is taken to be the true number of cells, since empty droplets are
- not part of the 'simple' model.
-
"""
- logging.info("Trimming dataset for inference.")
+ logger.info("Trimming features for inference.")
# Get data matrix and barcode order that sorts barcodes by UMI count.
matrix = self.data['matrix']
- umi_counts = np.array(matrix.sum(axis=1)).squeeze()
- umi_count_order = np.argsort(umi_counts)[::-1]
-
- # Initially set the default to be the whole dataset.
- self.analyzed_barcode_inds = np.arange(start=0, stop=matrix.shape[0])
- self.analyzed_gene_inds = np.arange(start=0, stop=matrix.shape[1])
-
- # Expected cells must not exceed nonzero count barcodes.
- num_nonzero_barcodes = np.sum(umi_counts > 0).item()
# Choose which genes to use based on their having nonzero counts.
# (All barcodes must be included so that inference can generalize.)
- gene_counts_per_barcode = np.array(matrix.sum(axis=0)).squeeze()
- if self.exclude_antibodies:
- antibody_logic = (self.data['feature_types'] == b'Antibody Capture')
- # Exclude these by setting their counts to zero
- logging.info(f"Excluding {antibody_logic.sum()} features that "
- f"correspond to antibody capture.")
- gene_counts_per_barcode[antibody_logic] = 0
- nonzero_gene_counts = (gene_counts_per_barcode > 0)
- self.analyzed_gene_inds = np.where(nonzero_gene_counts)[0].astype(dtype=int)
-
- if self.analyzed_gene_inds.size == 0:
- logging.warning("During data loading, found no genes with > 0 "
- "counts. Terminating analysis.")
- raise AssertionError("No genes with nonzero counts. Check the "
- "input data file.")
+ feature_counts_per_barcode = np.array(matrix.sum(axis=0)).squeeze()
+ inclusion_logic = (feature_counts_per_barcode > 0)
+
+ # Update self.analyzed_gene_inds
+ self.analyzed_gene_logic = np.logical_and(self.analyzed_gene_logic, inclusion_logic)
# Remove blacklisted genes.
if len(gene_blacklist) > 0:
-
- # Make gene_blacklist a set for fast inclusion testing.
gene_blacklist = set(gene_blacklist)
+ inclusion_logic = np.array([g not in gene_blacklist
+ for g in range(len(self.analyzed_gene_logic))])
+ logger.info(f"Ignoring {np.logical_not(inclusion_logic).sum()} "
+ f"specified features.")
+ self.analyzed_gene_logic = np.logical_and(
+ self.analyzed_gene_logic,
+ inclusion_logic,
+ )
+
+ if self.analyzed_gene_logic.sum() == 0:
+ logger.warning("No features remain after eliminating zero-count "
+ "and ignored features. Terminating analysis.")
+ raise AssertionError("No features with nonzero counts remain. Check "
+ "the input data file and check "
+ "--exclude-feature-types and --ignore-features")
+
+ logger.info(f"{self.analyzed_gene_logic.sum()} features have nonzero counts.")
+
+ def _trim_noiseless_features(self):
+ """Trim the dataset for inference, choosing genes to use.
+
+ To be run after trimming features first, then trimming barcodes.
+
+ Sets the value of self.analyzed_gene_inds
+ """
- # Ensure genes on the blacklist are excluded.
- self.analyzed_gene_inds = np.array([g for g in
- self.analyzed_gene_inds
- if g not in gene_blacklist])
-
- if self.analyzed_gene_inds.size == 0:
- logging.warning("All nonzero genes have been blacklisted. "
- "Terminating analysis.")
- raise AssertionError("All genes with nonzero counts have been "
- "blacklisted. Examine the dataset and "
- "reduce the blacklist.")
-
- logging.info(f"Including {self.analyzed_gene_inds.size} genes that have"
- f" nonzero counts.")
-
- # Estimate priors on cell size and 'empty' droplet size.
- self.priors['cell_counts'], self.priors['empty_counts'] = \
- get_d_priors_from_dataset(self) # After gene trimming
+ assert len(self.priors.keys()) > 0, 'Run self.priors = get_priors() before ' \
+ 'self._trim_noiseless_features()'
+
+ # Find average counts per gene in empty droplets.
+ count_matrix = self.data['matrix'][:, self.analyzed_gene_logic]
+ counts = np.array(count_matrix.sum(axis=1)).squeeze()
+ cutoff = self.priors['empty_count_upper_limit']
+ count_matrix_empties = count_matrix[(counts < cutoff)
+ & (counts > self.low_count_threshold), :]
+ mean_counts_per_empty_g = np.array(count_matrix_empties.mean(axis=0)).squeeze()
+
+ # Estimate counts in cells, but use a minimum number of cells just as a floor.
+ ambient_counts_in_cells_g = (max(self.priors['expected_cells'], 5000)
+ * mean_counts_per_empty_g)
+
+ # Include features that are projected to have noise above the low limit.
+ inclusion_logic_subset = (ambient_counts_in_cells_g
+ > self.ambient_counts_in_cells_low_limit)
+ inclusion_logic = np.ones(len(self.analyzed_gene_logic), dtype=bool)
+ inclusion_logic[self.analyzed_gene_logic] = inclusion_logic_subset
+
+ # Record in self.analyzed_gene_logic
+ self.analyzed_gene_logic = np.logical_and(
+ self.analyzed_gene_logic,
+ inclusion_logic,
+ )
+ logger.info(f"Excluding {np.logical_not(inclusion_logic).sum()} features that "
+ f"are estimated to have <= {self.ambient_counts_in_cells_low_limit} "
+ f"background counts in cells.")
+
+ def _trim_droplets(self):
+ """Trim the dataset for inference, choosing barcodes to use.
- # Estimate the number of real cells if it was not specified.
- if self.priors['n_cells'] is None:
- self.priors['n_cells'] = estimate_cell_count_from_dataset(self)
+ Sets the values of self.analyzed_barcode_inds, and
+ self.empty_barcode_inds, which are used throughout training.
+ """
- # n_cells is at most the number of nonzero barcodes.
- n_cells = min(self.priors['n_cells'], num_nonzero_barcodes)
+ logger.info("Trimming barcodes for inference.")
- assert n_cells > 0, "No cells identified. Try to use the option " \
- "--expected-cells to specify a prior on cell " \
- "count. Also ensure that --low-count-threshold " \
- "is not too large (less than empty droplets)."
+ # Get data matrix and barcode order that sorts barcodes by UMI count.
+ matrix = self.get_count_matrix_all_barcodes() # self.data['matrix']
+ umi_counts = np.array(matrix.sum(axis=1)).squeeze()
+ # umi_counts_features_trimmed = self.get_count_matrix_all_barcodes()
+ umi_count_order = np.argsort(umi_counts)[::-1]
+ n_cells = self.priors['expected_cells']
+ total_droplet_barcodes = self.priors['total_droplets']
# If running the simple model, just use the expected cells, no more.
- if self.model_name == "simple":
+ if not self.include_empties:
self.analyzed_barcode_inds = np.array(umi_count_order[:n_cells],
dtype=int)
- logging.info(f"Simple model: using "
- f"{self.analyzed_barcode_inds.size} cell barcodes.")
+ logger.info(f"Excluding empty droplets due to '{self.model_name}' model: "
+ f"using {self.analyzed_barcode_inds.size} cell barcodes.")
# If not using the simple model, include empty droplets.
else:
- # Get the cell barcodes.
- cell_barcodes = umi_count_order[:n_cells]
+ assert total_droplet_barcodes - n_cells > 0, \
+ f"The number of cells is {n_cells}, but the " \
+ f"number of total droplets included is " \
+ f"{total_droplet_barcodes}. Increase " \
+ f"--total_droplet_barcodes above {n_cells}, or " \
+ f"specify a different number of expected cells using " \
+ f"--expected-cells"
# Set the low UMI count cutoff to be the greater of either
# the user input value, or an empirically-derived value.
- empirical_low_UMI = int(self.priors['empty_counts'] *
- self.EMPIRICAL_LOW_UMI_TO_EMPTY_DROPLET_THRESHOLD)
- low_UMI_count_cutoff = max(low_UMI_count_cutoff,
- empirical_low_UMI)
- logging.info(f"Excluding barcodes with counts below "
- f"{low_UMI_count_cutoff}")
+ factor = consts.EMPIRICAL_LOW_UMI_TO_EMPTY_DROPLET_THRESHOLD
+ empirical_low_count_cutoff = int(self.priors['empty_counts'] * factor)
+ low_count_cutoff = max(self.low_count_threshold, empirical_low_count_cutoff)
+ self.low_count_cutoff = low_count_cutoff
+ logger.info(f"Excluding barcodes with counts below {low_count_cutoff}")
# See how many barcodes there are to work with total.
- num_barcodes_above_umi_cutoff = \
- np.sum(umi_counts > low_UMI_count_cutoff).item()
+ num_barcodes_above_umi_cutoff = np.sum(umi_counts > low_count_cutoff).item()
assert num_barcodes_above_umi_cutoff > 0, \
f"There are no barcodes with UMI counts over the lower " \
- f"cutoff of {low_UMI_count_cutoff}"
+ f"cutoff of {low_count_cutoff}"
assert num_barcodes_above_umi_cutoff > n_cells, \
f"There are no empty droplets with UMI counts over the lower " \
- f"cutoff of {low_UMI_count_cutoff}. Some empty droplets are " \
+ f"cutoff of {low_count_cutoff}. Some empty droplets are " \
f"necessary for the analysis. Reduce the " \
f"--low-count-threshold parameter."
- # Get a number of transition-region barcodes.
- num_transition_barcodes = (total_droplet_barcodes
- - cell_barcodes.size)
-
- assert num_transition_barcodes > 0, \
- f"The number of cells is {cell_barcodes.size}, but the " \
- f"number of total droplets included is " \
- f"{total_droplet_barcodes}. Increase " \
- f"--total_droplet_barcodes above {cell_barcodes.size}, or " \
- f"specify a different number of expected cells using " \
- f"--expected-cells"
-
- num = min(num_transition_barcodes,
- num_barcodes_above_umi_cutoff - cell_barcodes.size)
- num = max(0, num)
- transition_barcodes = umi_count_order[n_cells:
- (n_cells + num)]
-
- assert transition_barcodes.size > 0, \
- f"There are no barcodes identified from the transition " \
- f"region between cell and empty. The intended number of " \
- f"transition barcodes was {num_transition_barcodes}. " \
- f"This indicates that the low UMI count cutoff, " \
- f"{low_UMI_count_cutoff}, was likely too high. Try to " \
- f"reduce --low-count-threshold"
-
# Use the cell barcodes and transition barcodes for analysis.
- self.analyzed_barcode_inds = np.concatenate((
- cell_barcodes,
- transition_barcodes)).astype(dtype=int)
-
- # Identify probable empty droplet barcodes.
- if num < num_transition_barcodes:
+ self.analyzed_barcode_inds = umi_count_order[:total_droplet_barcodes]
- # This means we already used all the barcodes.
- empty_droplet_barcodes = np.array([])
-
- else:
-
- # Decide which empty barcodes to include.
- empty_droplet_sorted_barcode_inds = \
- np.arange(n_cells + num, num_barcodes_above_umi_cutoff,
- dtype=int) # The entire range
- empty_droplet_barcodes = \
- umi_count_order[empty_droplet_sorted_barcode_inds]
-
- self.empty_barcode_inds = empty_droplet_barcodes\
- .astype(dtype=int)
+ # Decide which empty barcodes to include.
+ empty_droplet_sorted_barcode_inds = \
+ np.arange(total_droplet_barcodes, num_barcodes_above_umi_cutoff,
+ dtype=int) # The entire range
+ self.empty_barcode_inds = umi_count_order[empty_droplet_sorted_barcode_inds]
# Find the UMI threshold for surely empty droplets.
- last_analyzed_bc = min(cell_barcodes.size + transition_barcodes.size,
- umi_count_order.size)
- self.empty_UMI_threshold = (umi_counts[umi_count_order]
- [last_analyzed_bc])
+ self.empty_UMI_threshold = umi_counts[umi_count_order][total_droplet_barcodes]
# Find the max UMI count for any cell.
self.max_UMI_count = umi_counts.max()
- logging.info(f"Using {cell_barcodes.size} probable cell "
- f"barcodes, plus an additional "
- f"{transition_barcodes.size} barcodes, "
- f"and {empty_droplet_barcodes.size} empty "
- f"droplets.")
- logging.info(f"Largest surely-empty droplet has "
- f"{self.empty_UMI_threshold} UMI counts.")
-
- if ((low_UMI_count_cutoff == self.low_count_threshold)
- and (empty_droplet_barcodes.size == 0)):
- logging.warning("Warning: few empty droplets identified. Low "
- "UMI cutoff may be too high. Check the UMI "
- "decay curve, and decrease the "
- "--low-count-threshold parameter if necessary.")
-
- self.is_trimmed = True
-
- def _estimate_priors(self):
- """Estimate relevant priors, populating fields in the self.priors dict.
- """
-
- # Estimate the log UMI count turning point between cells and 'empties'.
- self.priors['log_counts_crossover'] = \
- np.mean(np.log1p([self.priors['cell_counts'],
- self.priors['empty_counts']])).item()
-
- # TODO: overhaul estimation of d_std. add estimate of d_empty_std
-
- # Estimate prior for the scale param of LogNormal for d.
- if self.model_name != "simple":
- self.priors['d_std'] = (np.log1p(self.priors['cell_counts'])
- - self.priors['log_counts_crossover']) / 5
- else:
- # Use a reasonable prior in log space.
- self.priors['d_std'] = self.SIMPLE_MODEL_D_STD_PRIOR
-
- # Priors for models that include empty droplets:
- if self.model_name != "simple":
- # Estimate fraction of trimmed dataset that contains cells.
- # cell_prob = self.priors['n_cells']
- # / self.analyzed_barcode_inds.size
- cell_prob = (1 - self.fraction_empties) \
- * (self.priors['n_cells']
- / self.analyzed_barcode_inds.size)
- self.priors['cell_prob'] = cell_prob
-
- assert cell_prob > 0, f"Fraction of trimmed dataset " \
- f"containing cells should be > 0, " \
- f"but is {cell_prob}."
-
- assert cell_prob <= 1, f"Fraction of trimmed dataset " \
- f"containing cells should be at most 1, " \
- f"but is {cell_prob}."
-
- # Turn cell probability into logit.
- self.priors['cell_logit'] = np.log(cell_prob
- / (1 - cell_prob)).item()
-
- # Estimate the ambient gene expression profile.
- self.priors['chi_ambient'], self.priors['chi_bar'] = \
- estimate_chi_ambient_from_dataset(self)
-
- def get_count_matrix(self) -> sp.csr.csr_matrix:
+ # Estimate cell logit probability prior.
+ cell_prob = (n_cells / total_droplet_barcodes) * (1. - self.fraction_empties)
+ self.priors['cell_logit'] = np.log(cell_prob) - np.log(1. - cell_prob)
+
+ logger.info(f"Using {n_cells} probable cell "
+ f"barcodes, plus an additional "
+ f"{total_droplet_barcodes - n_cells} barcodes, "
+ f"and {len(self.empty_barcode_inds)} empty "
+ f"droplets.")
+ logger.info(f"Largest surely-empty droplet has "
+ f"{int(self.empty_UMI_threshold)} UMI counts.")
+
+ if ((low_count_cutoff == self.low_count_threshold)
+ and (len(self.empty_barcode_inds) == 0)):
+ logger.warning("Warning: few empty droplets identified. Low "
+ "UMI cutoff may be too high. Check the UMI "
+ "decay curve, and decrease the "
+ "--low-count-threshold parameter if necessary.")
+
+ def _estimate_chi_ambient(self) -> Dict[str, torch.Tensor]:
+ """Estimate chi_ambient and chi_bar"""
+
+ matrix = self.data['matrix'].tocsc()
+ count_matrix = matrix[:, self.analyzed_gene_inds].tocsr()
+ umi_counts = np.array(count_matrix.sum(axis=1)).squeeze()
+
+ # Estimate the ambient gene expression profile.
+ ep = np.finfo(np.float32).eps.item() # small value
+ empty_droplet_logic = ((umi_counts < self.priors['surely_empty_counts'])
+ & (umi_counts > self.low_count_cutoff))
+ chi_ambient = np.array(count_matrix[empty_droplet_logic, :].sum(axis=0)).squeeze() + ep
+ chi_ambient = torch.tensor(chi_ambient / chi_ambient.sum()).float()
+ chi_bar = np.array(count_matrix.sum(axis=0)).squeeze() + ep
+ chi_bar = torch.tensor(chi_bar / chi_bar.sum()).float()
+
+ return {'chi_ambient': chi_ambient, 'chi_bar': chi_bar}
+
+ def get_count_matrix(self) -> sp.csr_matrix:
"""Get the count matrix, trimmed if trimming has occurred."""
- if self.is_trimmed:
-
- # Return the count matrix for selected barcodes and genes.
- trimmed_bc_matrix = self.data['matrix'][self.analyzed_barcode_inds,
- :].tocsc()
- trimmed_matrix = trimmed_bc_matrix[:, self.analyzed_gene_inds].tocsr()
- return trimmed_matrix
+ # Return the count matrix for selected barcodes and genes.
+ trimmed_bc_matrix = self.data['matrix'][self.analyzed_barcode_inds,
+ :].tocsc()
+ trimmed_matrix = trimmed_bc_matrix[:, self.analyzed_gene_inds].tocsr()
+ return trimmed_matrix
- else:
- logging.warning("Using full count matrix, without any trimming. "
- "Could be slow.")
- return self.data['matrix']
-
- def get_count_matrix_empties(self) -> sp.csr.csr_matrix:
+ def get_count_matrix_empties(self) -> sp.csr_matrix:
"""Get the count matrix for empty drops, trimmed if possible."""
- if self.is_trimmed:
-
- # Return the count matrix for selected barcodes and genes.
- trimmed_bc_matrix = self.data['matrix'][self.empty_barcode_inds,
- :].tocsc()
- trimmed_matrix = trimmed_bc_matrix[:, self.analyzed_gene_inds].tocsr()
- return trimmed_matrix
-
- else:
- logging.error("Trying to get empty count matrix without "
- "trimmed data.")
- return self.data['matrix']
+ # Return the count matrix for selected barcodes and genes.
+ trimmed_bc_matrix = self.data['matrix'][self.empty_barcode_inds,
+ :].tocsc()
+ trimmed_matrix = trimmed_bc_matrix[:, self.analyzed_gene_inds].tocsr()
+ return trimmed_matrix
- def get_count_matrix_all_barcodes(self) -> sp.csr.csr_matrix:
+ def get_count_matrix_all_barcodes(self) -> sp.csr_matrix:
"""Get the count matrix, trimming only genes, not barcodes."""
- if self.is_trimmed:
-
- # Return the count matrix for selected barcodes and genes.
- trimmed_bc_matrix = self.data['matrix'].tocsc()
- trimmed_matrix = trimmed_bc_matrix[:, self.analyzed_gene_inds].tocsr()
- return trimmed_matrix
-
- else:
- logging.warning("Using full count matrix, without any trimming. "
- "Could be slow.")
- return self.data['matrix']
+ # Return the count matrix for selected barcodes and genes.
+ trimmed_bc_matrix = self.data['matrix'].tocsc()
+ trimmed_matrix = trimmed_bc_matrix[:, self.analyzed_gene_inds].tocsr()
+ return trimmed_matrix
def get_dataloader(self,
use_cuda: bool = True,
batch_size: int = 200,
shuffle: bool = False,
- analyzed_bcs_only: bool = True) -> DataLoader:
+ analyzed_bcs_only: bool = True,
+ sort_by: Optional[Callable[[sp.csr_matrix], float]] = None,
+ ) -> DataLoader:
"""Return a dataloader for the count matrix.
Args:
@@ -463,6 +459,10 @@ def get_dataloader(self,
shuffle: Whether dataloader should shuffle the data.
analyzed_bcs_only: Only include the barcodes that have been
analyzed, not the surely empty droplets.
+ sort_by: Lambda function which, when applied to the sparse matrix,
+ will return values that can be sorted to give a sort order to
+ the dataset. Dataloader will load data in order of increasing
+ values.
Returns:
data_loader: A dataloader that yields the entire dataset in batches.
@@ -474,1083 +474,66 @@ def get_dataloader(self,
else:
count_matrix = self.get_count_matrix_all_barcodes()
- return DataLoader(count_matrix,
- empty_drop_dataset=None,
- batch_size=batch_size,
- fraction_empties=0.,
- shuffle=shuffle,
- use_cuda=use_cuda)
-
- def save_to_output_file(
+ data_loader = DataLoader(
+ count_matrix,
+ empty_drop_dataset=None,
+ batch_size=batch_size,
+ fraction_empties=0.,
+ shuffle=shuffle,
+ sort_by=sort_by,
+ use_cuda=use_cuda,
+ )
+ return data_loader
+
+ def restore_eliminated_features_in_cells(
self,
- output_file: str,
- inferred_model: 'RemoveBackgroundPyroModel',
- posterior_batch_size: int,
- cells_posterior_reg_calc: int,
- save_plots: bool = False) -> bool:
- """Write the results of an inference procedure to an output file.
-
- Output is an HDF5 file. To be written:
- Inferred ambient-subtracted UMI count matrix.
- Inferred probabilities that each barcode contains a real cell.
- Inferred cell size scale factors.
- Inferred ambient gene expression count vector for droplets without
- cells.
- Inferred contamination fraction hyperparameters.
- Embeddings of gene expression of cells into a low-dimensional latent
- space.
+ inferred_count_matrix: sp.csc_matrix,
+ cell_probabilities_analyzed_bcs: np.ndarray
+ ) -> sp.csc_matrix:
+ """Add the data back in for any features not used during inference,
+ (the used features being defined in self.analyzed_gene_inds),
+ so that the output for these features exactly matches the input.
Args:
- inferred_model: RemoveBackgroundPyroModel which has
- already had the inference procedure run.
- output_file: Name of output .h5 file
- save_plots: Setting this to True will save plots of outputs.
+ inferred_count_matrix: A sparse count matrix
+ cell_probabilities_analyzed_bcs: Cell probabilities for the subset of
+ barcodes that were analyzed.
Returns:
- True if the output was written to file successfully.
+ output: A sparse count matrix, with the unused feature values filled
+ in with the raw data.
"""
- logging.info("Preparing to write outputs to file...")
-
- # Create posterior.
- self.posterior = ProbPosterior(dataset_obj=self,
- vi_model=inferred_model,
- fpr=self.fpr[0],
- batch_size=posterior_batch_size,
- cells_posterior_reg_calc=cells_posterior_reg_calc)
-
- # Encoded values of latent variables.
- enc = self.posterior.latents
- z = enc['z']
- d = enc['d']
- p = enc['p']
- epsilon = enc['epsilon']
- phi_params = enc['phi_loc_scale']
-
- # Estimate the ambient-background-subtracted UMI count matrix.
- if self.model_name != "simple":
-
- inferred_count_matrix = self.posterior.mean
-
- else:
-
- # No need to generate a new count matrix for simple model.
- inferred_count_matrix = self.data['matrix'].tocsc()
- logging.info("Simple model: outputting un-altered count matrix.")
-
- # Inferred ambient gene expression vector.
- ambient_expression_trimmed = cellbender.remove_background.model.\
- get_ambient_expression()
-
- # Convert the indices from trimmed gene set to original gene indices.
- ambient_expression = np.zeros(self.data['matrix'].shape[1])
- ambient_expression[self.analyzed_gene_inds] = ambient_expression_trimmed
-
- # Figure out the indices of barcodes that have cells.
- if p is not None:
- p[np.isnan(p)] = 0.
- cell_barcode_inds = self.analyzed_barcode_inds
- if np.sum(p > 0.5) == 0:
- logging.warning("Warning: Found no cells!")
- filtered_inds_of_analyzed_barcodes = p > self.CELL_PROB_CUTOFF
- else:
- cell_barcode_inds = self.analyzed_barcode_inds
- filtered_inds_of_analyzed_barcodes = np.arange(0, d.size)
-
- # CellRanger version (format output like input).
- if self._detect_input_data_type() == 'cellranger_h5':
- cellranger_version = detect_cellranger_version_h5(self.input_file)
- elif self._detect_input_data_type() == 'anndata':
- # Arbitrarily peg output format for anndata inputs to be CellRanger v3 format
- cellranger_version = 3
- else:
- cellranger_version = detect_cellranger_version_mtx(self.input_file)
-
- # Write to output file, for each lambda specified by user.
- if len(self.fpr) == 1:
- write_succeeded = write_matrix_to_cellranger_h5(
- cellranger_version=cellranger_version,
- output_file=output_file,
- gene_names=self.data['gene_names'],
- gene_ids=self.data['gene_ids'],
- feature_types=self.data['feature_types'],
- genomes=self.data['genomes'],
- barcodes=self.data['barcodes'],
- inferred_count_matrix=inferred_count_matrix,
- cell_barcode_inds=cell_barcode_inds,
- ambient_expression=ambient_expression,
- z=z, d=d, p=p, phi=phi_params, epsilon=epsilon,
- rho=cellbender.remove_background.model.get_rho(),
- fpr=self.fpr[0],
- lambda_multiplier=self.posterior.lambda_multiplier,
- loss=inferred_model.loss)
-
- # Generate filename for filtered matrix output.
- file_dir, file_base = os.path.split(output_file)
- file_name = os.path.splitext(os.path.basename(file_base))[0]
- filtered_output_file = os.path.join(file_dir, file_name + "_filtered.h5")
-
- # Write filtered matrix (cells only) to output file.
- if self.model_name != "simple":
- cell_barcode_inds = \
- self.analyzed_barcode_inds[filtered_inds_of_analyzed_barcodes]
-
- cell_barcodes = self.data['barcodes'][cell_barcode_inds]
-
- write_matrix_to_cellranger_h5(
- cellranger_version=cellranger_version,
- output_file=filtered_output_file,
- gene_names=self.data['gene_names'],
- gene_ids=self.data['gene_ids'],
- feature_types=self.data['feature_types'],
- genomes=self.data['genomes'],
- barcodes=cell_barcodes,
- inferred_count_matrix=inferred_count_matrix[cell_barcode_inds, :],
- cell_barcode_inds=None,
- ambient_expression=ambient_expression,
- z=z[filtered_inds_of_analyzed_barcodes, :],
- d=d[filtered_inds_of_analyzed_barcodes],
- p=p[filtered_inds_of_analyzed_barcodes],
- phi=phi_params,
- epsilon=epsilon[filtered_inds_of_analyzed_barcodes],
- rho=cellbender.remove_background.model.get_rho(),
- fpr=self.fpr[0],
- lambda_multiplier=self.posterior.lambda_multiplier,
- loss=inferred_model.loss)
-
- else:
-
- file_dir, file_base = os.path.split(output_file)
- file_name = os.path.splitext(os.path.basename(file_base))[0]
-
- for i, fpr in enumerate(self.fpr):
-
- if i > 0: # first FPR has already been computed
-
- # Re-compute posterior counts for each new lambda.
- self.posterior.fpr = fpr # reach in and change the FPR
- self.posterior._get_mean() # force re-computation of posterior
- inferred_count_matrix = self.posterior.mean
-
- # Create an output filename for this lambda value.
- temp_output_filename = os.path.join(file_dir, file_name + f"_FPR_{fpr}.h5")
-
- write_succeeded = write_matrix_to_cellranger_h5(
- cellranger_version=cellranger_version,
- output_file=temp_output_filename,
- gene_names=self.data['gene_names'],
- gene_ids=self.data['gene_ids'],
- feature_types=self.data['feature_types'],
- genomes=self.data['genomes'],
- barcodes=self.data['barcodes'],
- inferred_count_matrix=inferred_count_matrix,
- cell_barcode_inds=self.analyzed_barcode_inds,
- ambient_expression=ambient_expression,
- z=z, d=d, p=p, phi=phi_params, epsilon=epsilon,
- rho=cellbender.remove_background.model.get_rho(),
- fpr=fpr,
- lambda_multiplier=self.posterior.lambda_multiplier,
- loss=inferred_model.loss)
-
- # Generate filename for filtered matrix output.
- filtered_output_file = os.path.join(file_dir, file_name + f"_FPR_{fpr}_filtered.h5")
-
- # Write filtered matrix (cells only) to output file.
- if self.model_name != "simple":
- cell_barcode_inds = \
- self.analyzed_barcode_inds[filtered_inds_of_analyzed_barcodes]
-
- cell_barcodes = self.data['barcodes'][cell_barcode_inds]
-
- write_matrix_to_cellranger_h5(
- cellranger_version=cellranger_version,
- output_file=filtered_output_file,
- gene_names=self.data['gene_names'],
- gene_ids=self.data['gene_ids'],
- feature_types=self.data['feature_types'],
- genomes=self.data['genomes'],
- barcodes=cell_barcodes,
- inferred_count_matrix=inferred_count_matrix[cell_barcode_inds, :],
- cell_barcode_inds=None,
- ambient_expression=ambient_expression,
- z=z[filtered_inds_of_analyzed_barcodes, :],
- d=d[filtered_inds_of_analyzed_barcodes],
- p=p[filtered_inds_of_analyzed_barcodes],
- phi=phi_params,
- epsilon=epsilon[filtered_inds_of_analyzed_barcodes],
- rho=cellbender.remove_background.model.get_rho(),
- fpr=fpr,
- lambda_multiplier=self.posterior.lambda_multiplier,
- loss=inferred_model.loss)
-
- # Save barcodes determined to contain cells as _cell_barcodes.csv
- try:
- barcode_names = np.array([str(cell_barcodes[i],
- encoding='UTF-8')
- for i in range(cell_barcodes.size)])
- except UnicodeDecodeError:
- # necessary if barcodes are ints
- barcode_names = cell_barcodes
- except TypeError:
- # necessary if barcodes are already decoded
- barcode_names = cell_barcodes
- bc_file_name = os.path.join(file_dir,
- file_name + "_cell_barcodes.csv")
- np.savetxt(bc_file_name, barcode_names, delimiter=',', fmt='%s')
- logging.info(f"Saved cell barcodes in {bc_file_name}")
-
- try:
- # Save plots, if called for.
- if save_plots:
- plt.figure(figsize=(6, 18))
-
- # Plot the train error.
- plt.subplot(3, 1, 1)
- plt.plot(inferred_model.loss['train']['elbo'], '.--',
- label='Train')
-
- # Plot the test error, if there was held-out test data.
- if 'test' in inferred_model.loss.keys():
- if len(inferred_model.loss['test']['epoch']) > 0:
- plt.plot(inferred_model.loss['test']['epoch'],
- inferred_model.loss['test']['elbo'], 'o:',
- label='Test')
- plt.legend()
-
- plt.gca().set_ylim(bottom=max(inferred_model.loss['train']
- ['elbo'][0],
- inferred_model.loss['train']
- ['elbo'][-1] - 2000))
- plt.xlabel('Epoch')
- plt.ylabel('ELBO')
- plt.title('Progress of the training procedure')
-
- # Plot the barcodes used, along with the inferred
- # cell probabilities.
- plt.subplot(3, 1, 2)
- count_mat = self.get_count_matrix()
- counts = np.array(count_mat.sum(axis=1)).squeeze()
- count_order = np.argsort(counts)[::-1]
- plt.semilogy(counts[count_order], color='black')
- plt.ylabel('UMI counts')
- plt.xlabel('Barcode index, sorted by UMI count')
- if p is not None: # The case of a simple model.
- plt.gca().twinx()
- plt.plot(p[count_order], '.:', color='red', alpha=0.3)
- plt.ylabel('Cell probability', color='red')
- plt.ylim([-0.05, 1.05])
- plt.title('Determination of which barcodes contain cells')
- else:
- plt.title('The subset of barcodes used for training')
-
- # Plot the latent encoding via PCA.
- plt.subplot(3, 1, 3)
- pca = PCA(n_components=2)
- if p is None:
- p = np.ones_like(d)
- z_pca = pca.fit_transform(z[p >= 0.5])
- plt.plot(z_pca[:, 0], z_pca[:, 1],
- '.', ms=3, color='black', alpha=0.3)
- plt.ylabel('PC 1')
- plt.xlabel('PC 0')
- plt.title('PCA of latent encoding of cell gene expression')
-
- file_dir, file_base = os.path.split(output_file)
- file_name = os.path.splitext(os.path.basename(file_base))[0]
- fig_name = os.path.join(file_dir, file_name + ".pdf")
- plt.savefig(fig_name, bbox_inches='tight', format='pdf')
- logging.info(f"Saved summary plots as {fig_name}")
-
- except Exception:
- logging.warning("Unable to save plot.")
-
- return write_succeeded
-
-
-def detect_cellranger_version_mtx(filedir: str) -> int:
- """Detect which version of CellRanger (2 or 3) created this mtx directory.
-
- Args:
- filedir: string path to .mtx file that contains the raw gene
- barcode matrix in a sparse coo text format.
-
- Returns:
- CellRanger version, either 2 or 3, as an integer.
-
- """
-
- assert os.path.isdir(filedir), f"The directory {filedir} is not accessible."
-
- if os.path.isfile(os.path.join(filedir, 'features.tsv.gz')):
- return 3
-
- else:
- return 2
-
-
-def get_matrix_from_cellranger_mtx(filedir: str) \
- -> Dict[str, Union[sp.csr.csr_matrix, List[np.ndarray], np.ndarray]]:
- """Load a count matrix from an mtx directory from CellRanger's output.
-
- For CellRanger v2:
- The directory must contain three files:
- matrix.mtx
- barcodes.tsv
- genes.tsv
-
- For CellRanger v3:
- The directory must contain three files:
- matrix.mtx.gz
- barcodes.tsv.gz
- features.tsv.gz
-
- This function returns a dictionary that includes the count matrix, the gene
- names (which correspond to columns of the count matrix), and the barcodes
- (which correspond to rows of the count matrix).
-
- Args:
- filedir: string path to .mtx file that contains the raw gene
- barcode matrix in a sparse coo text format.
-
- Returns:
- out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
- barcodes as rows and genes as columns
- out['barcodes']: numpy array of strings which are the nucleotide
- sequences of the barcodes that correspond to the rows in
- the out['matrix']
- out['gene_names']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string names of genes in the genome, which
- correspond to the columns in the out['matrix'].
- out['gene_ids']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string Ensembl ID of genes in the genome, which
- also correspond to the columns in the out['matrix'].
- out['feature_types']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string feature types of genes (or possibly
- antibody capture reads), which also correspond to the columns
- in the out['matrix'].
-
- """
-
- assert os.path.isdir(filedir), "The directory {filedir} is not accessible."
-
- # Decide whether data is CellRanger v2 or v3.
- cellranger_version = detect_cellranger_version_mtx(filedir=filedir)
- logging.info(f"CellRanger v{cellranger_version} format")
-
- # CellRanger version 3
- if cellranger_version == 3:
-
- matrix_file = os.path.join(filedir, 'matrix.mtx.gz')
- gene_file = os.path.join(filedir, 'features.tsv.gz')
- barcode_file = os.path.join(filedir, 'barcodes.tsv.gz')
-
- # Read in feature names.
- features = np.genfromtxt(fname=gene_file,
- delimiter="\t", skip_header=0,
- dtype='<U100')
-
- # Read in gene expression and feature data.
- gene_ids = features[:, 0].squeeze() # first column
- gene_names = features[:, 1].squeeze() # second column
- feature_types = features[:, 2].squeeze() # third column
-
- # CellRanger version 2
- elif cellranger_version == 2:
-
- # Read in the count matrix using scipy.
- matrix_file = os.path.join(filedir, 'matrix.mtx')
- gene_file = os.path.join(filedir, "genes.tsv")
- barcode_file = os.path.join(filedir, "barcodes.tsv")
-
- # Read in gene names.
- gene_data = np.genfromtxt(fname=gene_file,
- delimiter="\t", skip_header=0,
- dtype='<U100')
- if len(gene_data.shape) == 1: # custom file format with just gene names
- gene_names = gene_data.squeeze()
- gene_ids = None
- else: # the 10x CellRanger v2 format with two columns
- gene_names = gene_data[:, 1].squeeze() # second column
- gene_ids = gene_data[:, 0].squeeze() # first column
- feature_types = None
-
- else:
- raise NotImplementedError('MTX format was not identifiable as CellRanger '
- 'v2 or v3. Please check 10x Genomics formatting.')
-
- # For both versions:
-
- # Read in sparse count matrix.
- count_matrix = io.mmread(matrix_file).tocsr().transpose()
-
- # Read in barcode names.
- barcodes = np.genfromtxt(fname=barcode_file,
- delimiter="\t", skip_header=0, dtype='<U20')
-
- # Issue warnings if necessary, based on dimensions matching.
- if count_matrix.shape[1] != len(gene_names):
- logging.warning(f"Number of gene names in {filedir}/genes.tsv "
- f"does not match the number expected from the "
- f"count matrix.")
- if count_matrix.shape[0] != len(barcodes):
- logging.warning(f"Number of barcodes in {filedir}/barcodes.tsv "
- f"does not match the number expected from the "
- f"count matrix.")
-
- return {'matrix': count_matrix,
- 'gene_names': gene_names,
- 'feature_types': feature_types,
- 'gene_ids': gene_ids,
- 'genomes': None, # TODO: check if this info is available in either version
- 'barcodes': barcodes}
-
-
-def detect_cellranger_version_h5(filename: str) -> int:
- """Detect which version of CellRanger (2 or 3) created this h5 file.
-
- Args:
- filename: string path to .mtx file that contains the raw gene
- barcode matrix in a sparse coo text format.
-
- Returns:
- version: CellRanger version, either 2 or 3, as an integer.
-
- """
-
- with tables.open_file(filename, 'r') as f:
-
- # For CellRanger v2, each group in the table (other than root)
- # contains a genome.
- # For CellRanger v3, there is a 'matrix' group that contains 'features'.
-
- version = 2
-
- try:
-
- # This works for version 3 but not for version 2.
- getattr(f.root.matrix, 'features')
- version = 3
-
- except tables.NoSuchNodeError:
- pass
-
- return version
-
-
-def get_matrix_from_cellranger_h5(filename: str) \
- -> Dict[str, Union[sp.csr.csr_matrix, List[np.ndarray], np.ndarray]]:
- """Load a count matrix from an h5 file from CellRanger's output.
-
- The file needs to be a _raw_gene_bc_matrices_h5.h5 file. This function
- returns a dictionary that includes the count matrix, the gene names (which
- correspond to columns of the count matrix), and the barcodes (which
- correspond to rows of the count matrix).
-
- This function works for CellRanger v2 and v3 HDF5 formats.
-
- Args:
- filename: string path to .h5 file that contains the raw gene
- barcode matrices
-
- Returns:
- out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
- barcodes as rows and genes as columns
- out['barcodes']: numpy array of strings which are the nucleotide
- sequences of the barcodes that correspond to the rows in
- the out['matrix']
- out['gene_names']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string names of genes in the genome, which
- correspond to the columns in the out['matrix'].
- out['gene_ids']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string Ensembl ID of genes in the genome, which
- also correspond to the columns in the out['matrix'].
- out['feature_types']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string feature types of genes (or possibly
- antibody capture reads), which also correspond to the columns
- in the out['matrix'].
-
- """
-
- # Detect CellRanger version.
- cellranger_version = detect_cellranger_version_h5(filename=filename)
- logging.info(f"CellRanger v{cellranger_version} format")
-
- with tables.open_file(filename, 'r') as f:
- # Initialize empty lists.
- csc_list = []
- barcodes = None
- feature_ids = None
- feature_types = None
- genomes = None
-
- # CellRanger v2:
- # Each group in the table (other than root) contains a genome,
- # so walk through the groups to get data for each genome.
- if cellranger_version == 2:
-
- feature_names = []
- feature_ids = []
-
- for group in f.walk_groups():
- try:
- # Read in data for this genome, and put it into a
- # scipy.sparse.csc.csc_matrix
- barcodes = getattr(group, 'barcodes').read()
- data = getattr(group, 'data').read()
- indices = getattr(group, 'indices').read()
- indptr = getattr(group, 'indptr').read()
- shape = getattr(group, 'shape').read()
- csc_list.append(sp.csc_matrix((data, indices, indptr),
- shape=shape))
- feature_names.extend(getattr(group, 'gene_names').read())
- feature_ids.extend(getattr(group, 'genes').read())
-
- except tables.NoSuchNodeError:
- # This exists to bypass the root node, which has no data.
- pass
-
- # Create numpy arrays.
- feature_names = np.array(feature_names)
- if len(feature_ids) > 0:
- feature_ids = np.array(feature_ids)
- else:
- feature_ids = None
-
- # CellRanger v3:
- # There is only the 'matrix' group.
- elif cellranger_version == 3:
-
- # Read in data for this genome, and put it into a
- # scipy.sparse.csc.csc_matrix
- barcodes = getattr(f.root.matrix, 'barcodes').read()
- data = getattr(f.root.matrix, 'data').read()
- indices = getattr(f.root.matrix, 'indices').read()
- indptr = getattr(f.root.matrix, 'indptr').read()
- shape = getattr(f.root.matrix, 'shape').read()
- csc_list.append(sp.csc_matrix((data, indices, indptr),
- shape=shape))
-
- # Read in 'feature' information
- feature_group = f.get_node(f.root.matrix, 'features')
- feature_names = getattr(feature_group, 'name').read()
-
- try:
- feature_types = getattr(feature_group, 'feature_type').read()
- except tables.NoSuchNodeError:
- # This exists in case someone produced a file without feature_type.
- pass
- try:
- feature_ids = getattr(feature_group, 'id').read()
- except tables.NoSuchNodeError:
- # This exists in case someone produced a file without feature id.
- pass
- try:
- genomes = getattr(feature_group, 'genome').read()
- except tables.NoSuchNodeError:
- # This exists in case someone produced a file without feature genome.
- pass
-
- # Put the data together (possibly from several genomes for v2 datasets).
- count_matrix = sp.vstack(csc_list, format='csc')
- count_matrix = count_matrix.transpose().tocsr()
-
- # Issue warnings if necessary, based on dimensions matching.
- if count_matrix.shape[1] != feature_names.size:
- logging.warning(f"Number of gene names ({feature_names.size}) in {filename} "
- f"does not match the number expected from the count "
- f"matrix ({count_matrix.shape[1]}).")
- if count_matrix.shape[0] != barcodes.size:
- logging.warning(f"Number of barcodes ({barcodes.size}) in {filename} "
- f"does not match the number expected from the count "
- f"matrix ({count_matrix.shape[0]}).")
-
- return {'matrix': count_matrix,
- 'gene_names': feature_names,
- 'gene_ids': feature_ids,
- 'genomes': genomes,
- 'feature_types': feature_types,
- 'barcodes': barcodes}
-
-
-def get_matrix_from_anndata(filename: str) \
- -> Dict[str, Union[sp.csr.csr_matrix, List[np.ndarray], np.ndarray]]:
- """Load a count matrix from an h5ad AnnData file.
-
- The file needs to contain raw counts for all measured barcodes in the
- `.X` attribute or a `.layer[{'counts', 'spliced'}]` attribute. This function
- returns a dictionary that includes the count matrix, the gene names (which
- correspond to columns of the count matrix), and the barcodes (which
- correspond to rows of the count matrix).
-
- This function works for any AnnData object meeting the above requirements,
- as generated by alignment methods like `kallisto | bustools`.
-
- Args:
- filename: string path to .h5ad file that contains the raw gene
- barcode matrices
-
- Returns:
- out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
- barcodes as rows and genes as columns
- out['barcodes']: numpy array of strings which are the nucleotide
- sequences of the barcodes that correspond to the rows in
- the out['matrix']
- out['gene_names']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string names of genes in the genome, which
- correspond to the columns in the out['matrix'].
- out['gene_ids']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string Ensembl ID of genes in the genome, which
- also correspond to the columns in the out['matrix'].
- out['feature_types']: List of numpy arrays, where the number of elements
- in the list is the number of genomes in the dataset. Each numpy
- array contains the string feature types of genes (or possibly
- antibody capture reads), which also correspond to the columns
- in the out['matrix'].
-
- """
- logging.info(f"Detected AnnData format")
-
- adata = anndata.read_h5ad(filename)
-
- if "counts" in adata.layers.keys():
- # this is a common manual setting for users of scVI
- # given the manual convention, we prefer this matrix to
- # .X since it is less likely to represent something other
- # than counts
- logging.info("Found `.layers['counts']`. Using for count data.")
- count_matrix = adata.layers["counts"]
- elif "spliced" in adata.layers.keys() and adata.X is None:
- # alignment using kallisto | bustools with intronic counts
- # does not populate `.X` by default, but does populate
- # `.layers['spliced'], .layers['unspliced']`.
- # we use spliced counts for analysis
- logging.info("Found `.layers['spliced']`. Using for count data.")
- count_matrix = adata.layers["spliced"]
- else:
- logging.info("Using `.X` for count data.")
- count_matrix = adata.X
-
- # check that `count_matrix` contains a large number of barcodes,
- # consistent with a raw single cell experiment
- if count_matrix.shape[0] < consts.MINIMUM_BARCODES_H5AD:
- # this experiment might be prefiltered
- msg = f"Only {count_matrix.shape[0]} barcodes were found.\n"
- msg += "This suggests the matrix was prefiltered.\n"
- msg += "CellBender requires a raw, unfiltered [Barcodes, Genes] matrix."
- logging.warning(msg)
-
- # AnnData is [Cells, Genes], no need to transpose
- # we typecast explicitly in the off chance `count_matrix` was dense.
- count_matrix = sp.csr_matrix(count_matrix)
- # feature names and ids are not consistently delineated in AnnData objects
- # so we attempt to find relevant features using common values.
- feature_names = np.array(adata.var_names, dtype=str)
- barcodes = np.array(adata.obs_names, dtype=str)
-
- # Make an attempt to find feature_IDs if they are present.
- feature_ids = None
- if 'gene_ids' in adata.var.keys():
- feature_ids = np.array(adata.var['gene_ids'].values, dtype=str)
- elif 'gene_id' in adata.var.keys():
- feature_ids = np.array(adata.var['gene_id'].values, dtype=str)
-
- # Make an attempt to find feature_types if they are present.
- feature_types = None
- if 'feature_types' in adata.var.keys():
- feature_types = np.array(adata.var['feature_types'].values, dtype=str)
- elif 'feature_type' in adata.var.keys():
- feature_types = np.array(adata.var['feature_type'].values, dtype=str)
-
- # Make an attempt to find genomes if they are present.
- genomes = None
- if 'genomes' in adata.var.keys():
- genomes = np.array(adata.var['genomes'].values, dtype=str)
- elif 'genome' in adata.var.keys():
- genomes = np.array(adata.var['genome'].values, dtype=str)
-
- # Issue warnings if necessary, based on dimensions matching.
- if count_matrix.shape[1] != feature_names.size:
- logging.warning(f"Number of gene names ({feature_names.size}) in {filename} "
- f"does not match the number expected from the count "
- f"matrix ({count_matrix.shape[1]}).")
- if count_matrix.shape[0] != barcodes.size:
- logging.warning(f"Number of barcodes ({barcodes.size}) in {filename} "
- f"does not match the number expected from the count "
- f"matrix ({count_matrix.shape[0]}).")
-
- return {'matrix': count_matrix,
- 'gene_names': feature_names,
- 'gene_ids': feature_ids,
- 'genomes': genomes,
- 'feature_types': feature_types,
- 'barcodes': barcodes}
-
-
-def write_matrix_to_cellranger_h5(
- cellranger_version: int,
- output_file: str,
- gene_names: np.ndarray,
- barcodes: np.ndarray,
- inferred_count_matrix: sp.csc.csc_matrix,
- feature_types: Optional[np.ndarray] = None,
- gene_ids: Optional[np.ndarray] = None,
- genomes: Optional[np.ndarray] = None,
- cell_barcode_inds: Optional[np.ndarray] = None,
- ambient_expression: Optional[np.ndarray] = None,
- rho: Optional[np.ndarray] = None,
- z: Optional[np.ndarray] = None,
- d: Optional[np.ndarray] = None,
- p: Optional[np.ndarray] = None,
- phi: Optional[np.ndarray] = None,
- epsilon: Optional[np.ndarray] = None,
- fpr: Optional[float] = None,
- lambda_multiplier: Optional[float] = None,
- loss: Optional[Dict] = None) -> bool:
- """Write count matrix data to output HDF5 file using CellRanger format.
-
- Args:
- cellranger_version: Either 2 or 3. Determines the format of the output
- h5 file.
- output_file: Path to output .h5 file (e.g., 'output.h5').
- gene_names: Name of each gene (column of count matrix).
- gene_ids: Ensembl ID of each gene (column of count matrix).
- genomes: Name of the genome that each gene comes from.
- feature_types: Type of each feature (column of count matrix).
- barcodes: Name of each barcode (row of count matrix).
- inferred_count_matrix: Count matrix to be written to file, in sparse
- format. Rows are barcodes, columns are genes.
- cell_barcode_inds: Indices into the original cell barcode array that
- were found to contain cells.
- ambient_expression: Vector of gene expression of the ambient RNA
- background counts that contaminate cell counts.
- rho: Hyperparameters for the contamination fraction distribution.
- epsilon: Latent encoding of droplet RT efficiency.
- z: Latent encoding of gene expression.
- d: Latent cell size scale factor.
- p: Latent probability that a barcode contains a cell.
- phi: Latent global overdispersion mean and scale.
- fpr: Target false positive rate for the regularized posterior denoised
- counts, where false positives are true counts that are (erroneously)
- removed.
- lambda_multiplier: The lambda multiplier value used to achieve the
- targeted false positive rate.
- loss: Training and test error, as ELBO, for each epoch.
-
- Note:
- To match the CellRanger .h5 files, the matrix is stored as its
- transpose, with rows as genes and cell barcodes as columns.
-
- """
-
- assert isinstance(inferred_count_matrix,
- sp.csc_matrix), "The count matrix must be csc_matrix " \
- "format in order to write to HDF5."
-
- assert gene_names.size == inferred_count_matrix.shape[1], \
- "The number of gene names must match the number of columns in the " \
- "count matrix."
-
- if gene_ids is not None:
- assert gene_names.size == gene_ids.size, \
- f"The number of gene_names {gene_names.shape} must match " \
- f"the number of gene_ids {gene_ids.shape}."
-
- if feature_types is not None:
- assert gene_names.size == feature_types.size, \
- f"The number of gene_names {gene_names.shape} must match " \
- f"the number of feature_types {feature_types.shape}."
-
- if genomes is not None:
- assert gene_names.size == genomes.size, \
- "The number of gene_names must match the number of genome designations."
-
- assert barcodes.size == inferred_count_matrix.shape[0], \
- "The number of barcodes must match the number of rows in the count" \
- "matrix."
-
- # This reverses the role of rows and columns, to match CellRanger format.
- inferred_count_matrix = inferred_count_matrix.transpose().tocsc()
-
- # Write to output file.
- try:
- with tables.open_file(output_file, "w",
- title="CellBender remove-background output") as f:
-
- if cellranger_version == 2:
-
- # Create the group where data will be stored.
- group = f.create_group("/", "background_removed",
- "Counts after background correction")
-
- # Create arrays within that group for gene info.
- f.create_array(group, "gene_names", gene_names)
- if gene_ids is not None:
- f.create_array(group, "genes", gene_ids)
-
- elif cellranger_version == 3:
-
- # Create the group where data will be stored: name is "matrix".
- group = f.create_group("/", "matrix",
- "Counts after background correction")
-
- # Create a sub-group called "features"
- feature_group = f.create_group(group, "features",
- "Genes and other features measured")
-
- # Create arrays within that group for feature info.
- f.create_array(feature_group, "name", gene_names)
- if gene_ids is not None:
- f.create_array(feature_group, "id", gene_ids)
- if feature_types is not None:
- f.create_array(feature_group, "feature_type", feature_types)
- if genomes is not None:
- f.create_array(feature_group, "genome", genomes)
-
- # Copy the other extraneous information from the input file.
- # (Some user might need it for some reason.)
- # TODO
-
- else:
- raise NotImplementedError(f'Trying to save to CellRanger '
- f'v{cellranger_version} format, which '
- f'is not implemented.')
-
- # Code for both versions.
- f.create_array(group, "barcodes", barcodes)
-
- # Create arrays to store the count data.
- f.create_array(group, "data", inferred_count_matrix.data)
- f.create_array(group, "indices", inferred_count_matrix.indices)
- f.create_array(group, "indptr", inferred_count_matrix.indptr)
- f.create_array(group, "shape", inferred_count_matrix.shape)
-
- # Store background gene expression, barcode_inds, z, d, and p.
- if cell_barcode_inds is not None:
- f.create_array(group, "barcode_indices_for_latents",
- cell_barcode_inds)
- if ambient_expression is not None:
- f.create_array(group, "ambient_expression", ambient_expression)
- if z is not None:
- f.create_array(group, "latent_gene_encoding", z)
- if d is not None:
- f.create_array(group, "latent_scale", d)
- if p is not None:
- f.create_array(group, "latent_cell_probability", p)
- if phi is not None:
- f.create_array(group, "overdispersion_mean_and_scale", phi)
- if rho is not None:
- f.create_array(group, "contamination_fraction_params", rho)
- if epsilon is not None:
- f.create_array(group, "latent_RT_efficiency", epsilon)
- if fpr is not None:
- f.create_array(group, "target_false_positive_rate", fpr)
- if lambda_multiplier is not None:
- f.create_array(group, "lambda_multiplier", lambda_multiplier)
- if loss is not None:
- f.create_array(group, "training_elbo_per_epoch",
- np.array(loss['train']['elbo']))
- if 'test' in loss.keys():
- f.create_array(group, "test_elbo",
- np.array(loss['test']['elbo']))
- f.create_array(group, "test_epoch",
- np.array(loss['test']['epoch']))
- f.create_array(group, "fraction_data_used_for_testing",
- 1. - consts.TRAINING_FRACTION)
-
- logging.info(f"Succeeded in writing CellRanger v{cellranger_version} "
- f"format output to file {output_file}")
-
- return True
-
- except Exception:
- logging.warning(f"Encountered an error writing output to file "
- f"{output_file}. "
- f"Output may be incomplete.")
-
- return False
-
-
-def get_d_priors_from_dataset(dataset: SingleCellRNACountsDataset) \
- -> Tuple[float, float]:
- """Compute an estimate of reasonable priors on cell size and ambient size.
-
- Given a dataset (scipy.sparse.csr matrix of counts where
- rows are barcodes and columns are genes), and an expected
- cell count, compute an estimate of reasonable priors on cell size
- and ambient count size. This is done by a series of heuristics.
-
- Args:
- dataset: Dataset object containing a matrix of unique UMI counts,
- where rows are barcodes and columns are genes.
-
- Returns:
- cell_counts: Estimated mean number of UMI counts per real cell, in
- terms of transformed count data.
- empty_counts: Estimated mean number of UMI counts per 'empty'
- droplet, in terms of transformed count data.
-
- NOTE: Picks barcodes using cutoffs in untransformed count data. The output
- is in terms of transformed counts.
-
- """
-
- # Count the total unique UMIs per barcode (summing after transforming).
- counts = np.array(dataset.data['matrix']
- [:, dataset.analyzed_gene_inds].sum(axis=1)).squeeze()
-
- # If it's a model that does not model empty droplets, the dataset is cells.
- if dataset.model_name == 'simple':
-
- if dataset.priors['n_cells'] is None:
- # No prior on number of cells. Assume all are cells.
- dataset.priors['n_cells'] = int(np.sum(counts > 0).item())
-
- # Sort order the cells by counts.
- sort_order = np.argsort(counts)[::-1]
-
- # Estimate cell count by median, taking 'cells' to be largest counts.
- cell_counts = int(np.median(counts[sort_order]
- [:dataset.priors['n_cells']]).item())
-
- empty_counts = 0
-
- # Models that include both cells and empty droplets.
- else:
-
- # Cutoff for original data. Empirical.
- cut = dataset.low_count_threshold
-
- # Estimate the number of UMI counts in empty droplets.
-
- # Mode of (rounded) log counts (for counts > cut) is a robust
- # empty estimator.
- empty_log_counts = mode(np.round(np.log1p(counts[counts > cut]),
- decimals=1))[0]
- empty_counts = int(np.expm1(empty_log_counts).item())
-
- # Estimate the number of UMI counts in cells.
-
- # Use expected cells if it is available.
- if dataset.priors['n_cells'] is not None:
-
- # Sort order the cells by counts.
- sort_order = np.argsort(counts)[::-1]
-
- cell_counts = int(np.median(counts[sort_order]
- [:dataset.priors['n_cells']]).item())
-
- else:
-
- # Median of log counts above 5 * empty counts is a robust
- # cell estimator.
- cell_log_counts = np.median(np.log1p(counts[counts > 5 * empty_counts]))
- cell_counts = int(np.expm1(cell_log_counts).item())
-
- logging.info(f"Prior on counts in empty droplets is {empty_counts}")
-
- logging.info(f"Prior on counts for cells is {cell_counts}")
-
- return cell_counts, empty_counts
-
-
-def estimate_cell_count_from_dataset(dataset: SingleCellRNACountsDataset) \
- -> int:
- """Compute an estimate of number of real cells in a dataset.
-
- Given a Dataset, compute an estimate of the number of real cells.
-
- Args:
- dataset: Dataset object containing a matrix of unique UMI counts,
- where rows are barcodes and columns are genes.
-
- Returns:
- cell_count_est: Estimated number of real cells.
-
- """
-
- # If it's a model that does not model empty droplets, the dataset is cells.
- # NOTE: this is overridden if --expected_cells is specified.
- if dataset.model_name == 'simple':
- return dataset.data['matrix'].shape[0]
-
- # Count number of UMIs in each barcode.
- counts = np.array(dataset.data['matrix'].sum(axis=1),
- dtype=int).squeeze()
-
- # Find mid-way between cell_counts and empty_counts in log space.
- midway = np.mean([np.log1p(dataset.priors['cell_counts']),
- np.log1p(dataset.priors['empty_counts'])])
- umi_cutoff = np.expm1(midway)
-
- # Count the number of barcodes with UMI counts above the cutoff.
- cell_count_est = int(np.sum(counts > umi_cutoff).item())
-
- return cell_count_est
-
-
-def estimate_chi_ambient_from_dataset(dataset: SingleCellRNACountsDataset) \
- -> Tuple[torch.Tensor, torch.Tensor]:
- """Compute an estimate of ambient RNA levels.
-
- Given a Dataset, compute an estimate of the ambient gene expression and
- compute the average gene expression.
-
- Args:
- dataset: Dataset object containing a matrix of unique UMI counts,
- where rows are barcodes and columns are genes.
-
- Returns:
- chi_ambient_init: Estimated number of real cells.
- chi_bar: Average gene expression over dataset.
-
- NOTE: This must be done on transformed data.
-
- """
-
- # Ensure that an estimate of the log count crossover point between cells
- # and empty droplets has already been calculated.
- try:
- log_crossover = dataset.priors['log_counts_crossover']
- except KeyError:
- raise AssertionError("Could not find dataset parameter "
- "log_counts_crossover.")
-
- ep = np.finfo(np.float32).eps.item() # Small value
-
- # Trimmed and appropriately transformed count matrix.
- count_matrix = dataset.get_count_matrix()
-
- # Empty droplets have log counts < log_crossover.
- empty_barcodes = (np.log1p(np.array(count_matrix.sum(axis=1)).squeeze())
- < log_crossover)
-
- # Sum gene expression for the empty droplets.
- gene_expression = np.array(count_matrix[empty_barcodes, :].sum(axis=0))\
- .squeeze()
-
- # As a vector on a simplex.
- gene_expression = gene_expression + ep
- chi_ambient_init = \
- torch.Tensor(gene_expression / np.sum(gene_expression))
-
- # Full count matrix, appropriately transformed.
- full_count_matrix = dataset.get_count_matrix_all_barcodes()
-
- # Sum all gene expression.
- gene_expression_total = np.array(full_count_matrix.sum(axis=0)).squeeze()
-
- # As a vector on a simplex.
- gene_expression_total = gene_expression_total + ep
- chi_bar = \
- torch.Tensor(gene_expression_total / np.sum(gene_expression_total))
-
- return chi_ambient_init, chi_bar
+ # Rescue the raw data for ignored features.
+ out = overwrite_matrix_with_columns_from_another(
+ mat1=inferred_count_matrix,
+ mat2=self.data['matrix'],
+ column_inds=self.analyzed_gene_inds)
+
+ # But ensure that empty droplets are empty in the output.
+ cell_probabilities_all_bcs = np.zeros(out.shape[0])
+ cell_probabilities_all_bcs[self.analyzed_barcode_inds] = cell_probabilities_analyzed_bcs
+ empty_inds = np.where(cell_probabilities_all_bcs <= consts.CELL_PROB_CUTOFF)[0]
+ out = csr_set_rows_to_zero(csr=out.tocsr(), row_inds=empty_inds)
+
+ return out.tocsc()
+
+
+def get_dataset_obj(args: argparse.Namespace) -> SingleCellRNACountsDataset:
+ """Helper function that uses the argparse namespace"""
+
+ return SingleCellRNACountsDataset(
+ input_file=args.input_file,
+ expected_cell_count=args.expected_cell_count,
+ total_droplet_barcodes=args.total_droplets,
+ force_cell_umi_prior=args.force_cell_umi_prior,
+ force_empty_umi_prior=args.force_empty_umi_prior,
+ fraction_empties=args.fraction_empties,
+ model_name=args.model,
+ gene_blacklist=args.blacklisted_genes,
+ exclude_features=args.exclude_features,
+ low_count_threshold=args.low_count_threshold,
+ ambient_counts_in_cells_low_limit=args.ambient_counts_in_cells_low_limit,
+ fpr=args.fpr,
+ )
diff --git a/cellbender/remove_background/data/extras/simulate.py b/cellbender/remove_background/data/extras/simulate.py
index da6d27f1..1654c66f 100644
--- a/cellbender/remove_background/data/extras/simulate.py
+++ b/cellbender/remove_background/data/extras/simulate.py
@@ -1,382 +1,704 @@
-"""Simulate a basic scRNA-seq count matrix dataset, for unit tests."""
+"""Simulate a basic scRNA-seq count matrix dataset, for tests."""
+
+from cellbender.remove_background.model import calculate_mu, calculate_lambda
+from cellbender.remove_background.data.io import write_matrix_to_cellranger_h5
+from cellbender.remove_background.checkpoint import load_from_checkpoint
import numpy as np
import scipy.sparse as sp
+from sklearn.cluster import DBSCAN
+from sklearn.decomposition import PCA
import torch
-from typing import Tuple, List, Union
+import pyro
+import random
+import matplotlib.pyplot as plt
+from typing import List, Union, Dict, Optional
-def simulate_dataset_without_ambient_rna(
- n_cells: int = 100,
- clusters: int = 1,
- n_genes: int = 10000,
- cells_in_clusters: Union[List[int], None] = None,
- d_cell: int = 5000) -> Tuple[sp.csr.csr_matrix,
- np.ndarray,
- np.ndarray,
- np.ndarray]:
- """Simulate a dataset with ambient background RNA counts.
+if torch.cuda.is_available():
+ USE_CUDA = True
+ DEVICE = 'cuda'
+else:
+ USE_CUDA = False
+ DEVICE = 'cpu'
+
- Empty drops have ambient RNA only, while barcodes with cells have cell RNA
- plus some amount of ambient background RNA (in proportion to the sizes of
- cell and droplet).
+def comprehensive_random_seed(seed, use_cuda=USE_CUDA):
+ """Establish a base random state
+ https://pytorch.org/docs/stable/notes/randomness.html
+ """
+ random.seed(seed)
+ np.random.seed(seed)
+ torch.manual_seed(seed)
+ pyro.util.set_rng_seed(seed)
+ if use_cuda:
+ torch.cuda.manual_seed_all(seed)
+
+
+def generate_sample_inferred_model_dataset(
+ checkpoint_file: str,
+ cells_of_each_type: List[int] = [100, 100, 100],
+ n_droplets: int = 10000,
+ model_type: str = 'full',
+ plot_pca: bool = True,
+ dbscan_eps: float = 0.3,
+ dbscan_min_samples: int = 10,
+ random_seed: int = 0) -> Dict[str, Union[float, np.ndarray, sp.csr_matrix]]:
+ """Create a sample dataset for use as 'truth' data, based on a trained checkpoint.
Args:
- n_cells: Number of cells.
- clusters: Number of distinct cell types to simulate.
- n_genes: Number of genes.
- d_cell: Cell size scale factor.
- cells_in_clusters: Number of cells of each cell type. If specified,
- the number of ints in this list must be equal to clusters.
+ checkpoint_file: ckpt.tar.gz file from a remove-background run
+ cells_of_each_type: Number of cells of each cell type to simulate
+ n_droplets: Total number of droplets to simulate
+ model_type: The --model argument to specify remove-background's model
+ plot_pca: True to plot the PCA plot used to define "cell types"
+ dbscan_eps: Smaller makes the clusters finer
- Returns:
- csr_barcode_gene_synthetic: The simulated barcode by gene matrix of UMI
- counts, as a scipy.sparse.csr.csr_matrix.
- z: The simulated cell type identities. A numpy array of integers, one
- for each barcode. The number 0 is used to denote barcodes
- without a cell present.
- chi: The simulated gene expression, one corresponding to each z.
- Access the vector of gene expression for a given z using chi[z, :].
- d: The simulated size scale factors, one for each barcode.
"""
- assert d_cell > 0, "Location parameter, d_cell, of LogNormal " \
- "distribution must be greater than zero."
- assert clusters > 0, "clusters must be a positive integer."
- assert n_cells > 0, "n_cells must be a positive integer."
- assert n_genes > 0, "n_genes must be a positive integer."
-
- # Figure out how many cells are in each cell cluster.
- if cells_in_clusters is None:
- # No user input: make equal numbers of each cell type
- cells_in_clusters = np.ones(clusters) * int(n_cells / clusters)
+ # Input checking.
+ assert len(cells_of_each_type) > 0, 'cells_of_each_type must be a List of ints'
+ n_cell_types = len(cells_of_each_type)
+
+ # Load a trained model from a checkpoint file.
+ ckpt = load_from_checkpoint(
+ filebase=None,
+ tarball_name=checkpoint_file,
+ to_load=['model', 'dataloader', 'param_store'],
+ force_device='cpu' if not torch.cuda.is_available() else None)
+ model = ckpt['model']
+ n_genes = model.n_genes
+
+ # Reach in and set model type.
+ model.model_type = model_type
+
+ # Seed random number generators.
+ comprehensive_random_seed(seed=random_seed)
+
+ # Find z values for cells.
+ data_loader = ckpt['train_loader']
+ if torch.cuda.is_available():
+ data_loader.use_cuda = True
+ data_loader.device = 'cuda'
+ model.use_cuda = True
+ model.device = 'cuda'
else:
- assert len(cells_in_clusters) == clusters, "len(cells_in_clusters) " \
- "must equal clusters."
- assert sum(cells_in_clusters) == n_cells, "sum(cells_in_clusters) " \
- "must equal n_cells."
-
- # Initialize arrays and lists.
- chi = np.zeros((clusters + 1, n_genes))
- csr_list = []
- z = []
- d = []
-
- # Get chi for cell expression.
- for i in range(clusters):
- chi[i, :] = generate_chi(alpha=1.0, n_genes=n_genes)
- csr, d_n = sample_expression_from(chi[i, :],
- n=int(cells_in_clusters[i]),
- d_mu=np.log(d_cell).item())
- csr_list.append(csr)
- z = z + [i for _ in range(csr.shape[0])]
- d = d + [j for j in d_n]
-
- # Package the results.
- csr_barcode_gene_synthetic = sp.vstack(csr_list)
- z = np.array(z)
- d = np.array(d)
-
- # Permute the barcode order and return results.
- order = np.random.permutation(z.size)
- csr_barcode_gene_synthetic = csr_barcode_gene_synthetic[order, ...]
- z = z[order]
- d = d[order]
-
- return csr_barcode_gene_synthetic, z, chi, d
-
-
-def simulate_dataset_with_ambient_rna(
- n_cells: int = 150,
- n_empty: int = 300,
- clusters: int = 3,
+ data_loader.use_cuda = False
+ data_loader.device = 'cpu'
+ model.use_cuda = False
+ model.device = 'cpu'
+ z = np.zeros((len(data_loader), model.encoder['z'].output_dim))
+ p = np.zeros(len(data_loader))
+ chi_ambient = pyro.param('chi_ambient').detach()
+ for i, data in enumerate(data_loader):
+ enc = model.encoder(x=data,
+ chi_ambient=chi_ambient,
+ cell_prior_log=model.d_cell_loc_prior)
+ ind = i * data_loader.batch_size
+ z[ind:(ind + data.shape[0]), :] = enc['z']['loc'].detach().cpu().numpy()
+ p[ind:(ind + data.shape[0])] = enc['p_y'].sigmoid().detach().cpu().numpy()
+ z = z[p > 0.5, :] # select cells only
+
+ # Cluster cells based on embeddings.
+ db = DBSCAN(eps=dbscan_eps, min_samples=dbscan_min_samples).fit(z)
+ core_samples_mask = np.zeros_like(db.labels_, dtype=bool)
+ core_samples_mask[db.core_sample_indices_] = True
+ labels = db.labels_
+ n_clusters_ = len(set(labels)) - (1 if -1 in labels else 0)
+ if n_clusters_ < len(cells_of_each_type):
+ print(f'WARNING: DBSCAN found only {n_clusters_} clusters, but the input '
+ f'cells_of_each_type dictates {len(cells_of_each_type)} cell types '
+ f'are required.')
+
+ # Show a plot.
+ if plot_pca:
+ pca = PCA(n_components=2).fit_transform(z)
+ for label in np.unique(labels):
+ if label == -1:
+ color = 'lightgray'
+ else:
+ color = None
+ plt.plot(pca[labels == label, 0], pca[labels == label, 1],
+ '.', color=color, label=label)
+ plt.title('Embedding PCA')
+ plt.xlabel('PCA 0')
+ plt.ylabel('PCA 1')
+ plt.legend()
+ plt.show()
+
+ unique_labels = set(np.unique(labels)) - {-1}
+ print({label: (labels == label).sum() for label in unique_labels})
+
+ # Sample z values for cells of each cluster based on this posterior.
+ z_cluster = []
+ for label, n_cells in zip(np.unique(labels)[:len(cells_of_each_type)], cells_of_each_type):
+ if n_cells == 0:
+ continue
+ logic = (labels == label)
+ z_tmp = z[logic][torch.randperm(logic.sum())][:n_cells]
+ while len(z_tmp) < n_cells:
+ z_tmp = np.concatenate((z_tmp, z[logic][torch.randperm(logic.sum())][:n_cells]), axis=0)
+ z_cluster.append(z_tmp[:n_cells])
+
+ # Account for the possibility that one cell type has zero cells (useful for selecting populations of interest)
+ cells_of_each_type = [c for c in cells_of_each_type if c > 0]
+ n_cell_types = len(cells_of_each_type)
+
+ # Prep variables.
+ c_real = np.zeros((n_droplets, n_genes))
+ c_bkg = np.zeros((n_droplets, n_genes))
+ labels = np.zeros(n_droplets)
+ epsilon = np.zeros(n_droplets)
+ rho = np.zeros(n_droplets)
+ d_cell = np.zeros(n_droplets)
+ d_empty = np.zeros(n_droplets)
+ chi = np.zeros((n_cell_types, n_genes))
+ d_cell_type = np.zeros(n_cell_types)
+
+ # Sample counts from cell-containing droplets.
+ i = 0
+ for t, (num_cells, z_chunk) in enumerate(zip(cells_of_each_type, z_cluster)):
+ sim = sample_from_inferred_model(
+ num=num_cells,
+ model=model,
+ z=torch.tensor(z_chunk).to('cuda' if torch.cuda.is_available() else 'cpu').float(),
+ y=torch.ones(num_cells).to('cuda' if torch.cuda.is_available() else 'cpu').float(),
+ )
+ c_real[i:(i + num_cells), :] = sim['counts_real']
+ c_bkg[i:(i + num_cells), :] = sim['counts_bkg']
+ epsilon[i:(i + num_cells)] = sim['epsilon']
+ rho[i:(i + num_cells)] = sim['rho']
+ d_cell[i:(i + num_cells)] = sim['cell_counts']
+ d_empty[i:(i + num_cells)] = sim['empty_counts']
+ labels[i:(i + num_cells)] = t + 1 # label cell types with integers > 0
+ chi[t, :] = sim['chi'].sum(axis=0) / sim['chi'].sum()
+ d_cell_type[t] = sim['cell_counts'].mean()
+ i = i + num_cells
+
+ # Sample counts from empty droplets.
+ sim = sample_from_inferred_model(
+ num=n_droplets - i,
+ model=model,
+ y=torch.zeros(n_droplets - i).to('cuda' if torch.cuda.is_available() else 'cpu').float(),
+ )
+ c_real[i:, :] = sim['counts_real']
+ c_bkg[i:, :] = sim['counts_bkg']
+ epsilon[i:] = sim['epsilon']
+ rho[i:] = sim['rho']
+
+ # Get sparse matrices from dense.
+ counts_true = sp.csr_matrix(c_real, shape=c_real.shape, dtype=int)
+ counts_bkg = sp.csr_matrix(c_bkg, shape=c_bkg.shape, dtype=int)
+
+ return {'counts_true': counts_true,
+ 'counts_bkg': counts_bkg,
+ 'barcodes': np.array([f'bc{n:06d}' for n in np.arange(n_droplets)]),
+ 'gene_names': model.analyzed_gene_names,
+ 'chi': {i + 1: chi[i, :] for i in range(chi.shape[0])},
+ 'chi_ambient': chi_ambient.detach().cpu().numpy(),
+ 'droplet_labels': labels.astype(int),
+ 'd_cell': d_cell,
+ 'd_empty': d_empty,
+ 'epsilon': epsilon,
+ 'rho': rho,
+ 'cell_mean_umi': {i + 1: u for i, u in enumerate(d_cell_type)},
+ 'cell_lognormal_sigma': model.d_cell_scale_prior.item(),
+ 'empty_mean_umi': model.d_empty_loc_prior.exp().item(),
+ 'empty_lognormal_sigma': model.d_empty_scale_prior.item(),
+ 'epsilon_param': model.epsilon_prior.item(),
+ 'model': [model_type]}
+
+
+def generate_sample_dirichlet_dataset(
n_genes: int = 10000,
- d_cell: int = 5000,
- d_empty: int = 100,
- cells_in_clusters: Union[List[int], None] = None,
- ambient_different: bool = False,
- chi_input: Union[np.ndarray, None] = None) \
- -> Tuple[sp.csr.csr_matrix, np.ndarray, np.ndarray, np.ndarray]:
- """Simulate a dataset with ambient background RNA counts.
-
- Empty drops have ambient RNA only, while barcodes with cells have cell
- RNA plus some amount of ambient background RNA (in proportion to the
- sizes of cell and droplet).
-
- Args:
- n_cells: Number of cells.
- n_empty: Number of empty droplets with only ambient RNA.
- clusters: Number of distinct cell types to simulate.
- n_genes: Number of genes.
- d_cell: Cell size scale factor.
- d_empty: Empty droplet size scale factor.
- cells_in_clusters: Number of cells of each cell type. If specified,
- the number of ints in this list must be equal to clusters.
- ambient_different: If False, the gene expression profile of ambient
- RNA is drawn from the sum of cellular gene expression. If True,
- the ambient RNA expression is completely different from cellular
- gene expression.
- chi_input: Gene expression arrays in a matrix, with rows as clusters and
- columns as genes. Expression should add to one for each row.
- Setting chi=None will generate new chi randomly according to a
- Dirichlet distribution.
-
- Returns:
- csr_barcode_gene_synthetic: The simulated barcode by gene matrix of
- UMI counts, as a scipy.sparse.csr.csr_matrix.
- z: The simulated cell type identities. A numpy array of integers,
- one for each barcode. The number 0 is used to denote barcodes
- without a cell present.
- chi: The simulated gene expression, one corresponding to each z.
- Access the vector of gene expression for a given z using chi[z, :].
- d: The simulated size scale factors, one for each barcode.
-
+ cells_of_each_type: List[int] = [100, 100, 100],
+ n_droplets: int = 5000,
+ model_type: str = 'full',
+ dirichlet_alpha: Union[float, np.ndarray] = 0.05,
+ chi: Optional[np.ndarray] = None,
+ chi_artificial_similarity: float = 0,
+ vector_to_add_to_chi_ambient: Optional[np.ndarray] = None,
+ chi_ambient: Optional[np.ndarray] = None,
+ cell_mean_umi: List[int] = [5000],
+ cell_lognormal_sigma: float = 0.2,
+ empty_mean_umi: int = 200,
+ empty_lognormal_sigma: float = 0.4,
+ alpha0: float = 5000,
+ epsilon_param: float = 100,
+ rho_alpha: float = 3,
+ rho_beta: float = 80,
+ random_seed: int = 0) -> Dict[str, Union[float, np.ndarray, sp.csr_matrix]]:
+ """Create a sample dataset for use as 'truth' data.
"""
- assert d_cell > 0, "Location parameter, d_cell, of LogNormal " \
- "distribution must be greater than zero."
- assert d_empty > 0, "Location parameter, d_cell, of LogNormal " \
- "distribution must be greater than zero."
- assert clusters > 0, "clusters must be a positive integer."
- assert n_cells > 0, "n_cells must be a positive integer."
- assert n_empty > 0, "n_empty must be a positive integer."
- assert n_genes > 0, "n_genes must be a positive integer."
- if chi_input is not None:
- assert chi_input.shape[0] == clusters, "Chi was specified, but the " \
- "number of rows must match " \
- "the number of clusters."
- assert chi_input.shape[1] == n_genes, "Chi was specified, but the " \
- "number of columns must match " \
- "the number of genes."
-
- # Figure out how many cells are in each cell cluster.
- if cells_in_clusters is None:
- # No user input: make equal numbers of each cell type
- cells_in_clusters = (np.ones(clusters, dtype=int)
- * int(n_cells/clusters))
+ # Input checking.
+ assert len(cells_of_each_type) > 0, 'cells_of_each_type must be a List of ints'
+ n_cell_types = len(cells_of_each_type)
+ if vector_to_add_to_chi_ambient is not None:
+ assert len(vector_to_add_to_chi_ambient) == n_genes, \
+ f'vector_to_add_to_chi_ambient {vector_to_add_to_chi_ambient.shape} must ' \
+ f'be of length n_genes ({n_genes})'
+ if type(dirichlet_alpha) == np.ndarray:
+ assert dirichlet_alpha.size == n_genes, \
+ f'If you input a vector for dirichlet_alpha, its length ' \
+ f'({dirichlet_alpha.size}) must be n_genes ({n_genes})'
+ assert cell_lognormal_sigma > 0, 'cell_lognormal_sigma must be > 0'
+ assert empty_lognormal_sigma > 0, 'empty_lognormal_sigma must be > 0'
+ cell_mean_umi = list(cell_mean_umi)
+ if len(cell_mean_umi) == 1:
+ cell_mean_umi = cell_mean_umi * n_cell_types # repeat the entry
+ for umi in cell_mean_umi:
+ assert umi > empty_mean_umi, \
+ f'cell_size_mean_umi ({umi}) must be > empty_mean_umi ({empty_mean_umi})'
+ assert len(cell_mean_umi) == len(cells_of_each_type), \
+ 'cells_of_each_type and cell_mean_umi (if a list) must be of the same length'
+ assert (chi_artificial_similarity >= 0) and (chi_artificial_similarity <= 1), \
+ 'chi_artificial_similarity must be in the range [0, 1]'
+
+ # Seed random number generators.
+ comprehensive_random_seed(seed=random_seed)
+
+ # Draw gene expression profiles for each cell type.
+ if chi is None:
+ chi = np.zeros((n_cell_types, n_genes))
+ for i in range(chi.shape[0]):
+ chi[i, :] = draw_random_chi(alpha=dirichlet_alpha, n_genes=n_genes).cpu().numpy()
+ if i > 0:
+ # Make expression profiles similar (if chi_artificial_similarity > 0) via gating.
+ chi[i, :] = chi[i, :] * (1 - chi_artificial_similarity) + chi[0, :] * chi_artificial_similarity
else:
- assert len(cells_in_clusters) == clusters, "len(cells_in_clusters) " \
- "must equal clusters."
- assert sum(cells_in_clusters) == n_cells, "sum(cells_in_clusters) " \
- "must equal n_cells."
-
- # Initialize arrays and lists.
- chi = np.zeros((clusters+1, n_genes))
- csr_list = []
- z = []
- d = []
-
- if chi_input is not None:
-
- # Go with the chi that was input.
- chi[1:, :] = chi_input
-
+ dirichlet_alpha = ['None: chi was input to generate_sample_dataset()']
+
+ # Get chi_ambient: a weighted average of expression, possibly with extra.
+ if chi_ambient is None:
+ chi_ambient = np.zeros(n_genes)
+ for i in range(n_cell_types):
+ chi_ambient = chi_ambient + chi[i, :] * cells_of_each_type[i] * cell_mean_umi[i]
+ if vector_to_add_to_chi_ambient is None:
+ vector_to_add_to_chi_ambient = np.zeros(n_genes)
+ chi_ambient = chi_ambient + vector_to_add_to_chi_ambient
else:
-
- # Get chi for cell expression.
- for i in range(1, clusters+1):
- chi[i, :] = generate_chi(alpha=0.01, n_genes=n_genes)
-
- # Get chi for ambient expression. This becomes chi[0, :].
- if ambient_different:
-
- # Ambient expression is unrelated to cells, and is itself random.
- chi[0, :] = generate_chi(alpha=0.001, n_genes=n_genes) # Sparse
+ if vector_to_add_to_chi_ambient is not None:
+ print('You specified both `chi_ambient` and `vector_to_add_to_chi_ambient`. '
+ 'Ignoring `vector_to_add_to_chi_ambient` and using `chi_ambient` '
+ 'as provided.')
+ chi_ambient = chi_ambient / chi_ambient.sum()
+
+ c_real = np.zeros((n_droplets, n_genes))
+ c_bkg = np.zeros((n_droplets, n_genes))
+ labels = np.zeros(n_droplets)
+ epsilon = np.zeros(n_droplets)
+ rho = np.zeros(n_droplets)
+ d_cell = np.zeros(n_droplets)
+ d_empty = np.zeros(n_droplets)
+
+ # Sample counts from cell-containing droplets.
+ i = 0
+ for t, (num_cells, celltype_umi) in enumerate(zip(cells_of_each_type, cell_mean_umi)):
+ sim = sample_from_dirichlet_model(
+ num=num_cells,
+ alpha=chi[t, :] * alpha0 + 1e-20, # must be > 0
+ d_mu=np.log(celltype_umi),
+ d_sigma=cell_lognormal_sigma,
+ v_mu=np.log(empty_mean_umi),
+ v_sigma=empty_lognormal_sigma,
+ y=1, # cell present
+ chi_ambient=chi_ambient,
+ eps_param=epsilon_param,
+ rho_alpha=rho_alpha,
+ rho_beta=rho_beta,
+ model_type=model_type,
+ )
+ c_real[i:(i + num_cells), :] = sim['counts_real']
+ c_bkg[i:(i + num_cells), :] = sim['counts_bkg']
+ epsilon[i:(i + num_cells)] = sim['epsilon']
+ rho[i:(i + num_cells)] = sim['rho']
+ d_cell[i:(i + num_cells)] = sim['cell_counts']
+ d_empty[i:(i + num_cells)] = sim['empty_counts']
+ labels[i:(i + num_cells)] = t + 1 # label cell types with integers > 0
+ i = i + num_cells
+
+ # Sample counts from empty droplets.
+ sim = sample_from_dirichlet_model(
+ num=n_droplets - i,
+ alpha=np.ones(n_genes), # this doesn't get used since y=0
+ d_mu=1, # this doesn't get used since y=0
+ d_sigma=1, # this doesn't get used since y=0
+ v_mu=np.log(empty_mean_umi),
+ v_sigma=empty_lognormal_sigma,
+ y=0, # no cell present
+ chi_ambient=chi_ambient,
+ eps_param=epsilon_param,
+ rho_alpha=rho_alpha,
+ rho_beta=rho_beta,
+ model_type=model_type,
+ )
+ c_real[i:, :] = sim['counts_real']
+ c_bkg[i:, :] = sim['counts_bkg']
+ epsilon[i:] = sim['epsilon']
+ rho[i:] = sim['rho']
+ d_empty[i:] = sim['empty_counts']
+
+ # Get sparse matrices from dense.
+ counts_true = sp.csr_matrix(c_real, shape=c_real.shape, dtype=int)
+ counts_bkg = sp.csr_matrix(c_bkg, shape=c_bkg.shape, dtype=int)
+
+ return {'counts_true': counts_true,
+ 'counts_bkg': counts_bkg,
+ 'barcodes': np.array([f'bc{n:06d}' for n in np.arange(n_droplets)]),
+ 'gene_names': np.array([f'g{n:05d}' for n in np.arange(n_genes)]),
+ 'chi': {str(i + 1): chi[i, :] for i in range(chi.shape[0])},
+ 'chi_ambient': chi_ambient,
+ 'droplet_labels': labels.astype(int),
+ 'd_cell': d_cell,
+ 'd_empty': d_empty,
+ 'epsilon': epsilon,
+ 'rho': rho,
+ 'cell_mean_umi': {str(i + 1): u for i, u in enumerate(cell_mean_umi)},
+ 'cell_lognormal_sigma': cell_lognormal_sigma,
+ 'empty_mean_umi': empty_mean_umi,
+ 'empty_lognormal_sigma': empty_lognormal_sigma,
+ 'alpha0': alpha0,
+ 'epsilon_param': epsilon_param,
+ 'dirichlet_alpha': dirichlet_alpha,
+ 'model': [model_type]}
+
+
+@torch.no_grad()
+def generate_sample_model_dataset(
+ n_genes: int = 10000,
+ cells_of_each_type: List[int] = [100, 100, 100],
+ n_droplets: int = 5000,
+ model_type: str = 'full',
+ dirichlet_alpha: Union[float, np.ndarray] = 0.05,
+ chi: Optional[np.ndarray] = None,
+ chi_artificial_similarity: float = 0,
+ vector_to_add_to_chi_ambient: Optional[np.ndarray] = None,
+ cell_mean_umi: List[int] = [5000],
+ cell_lognormal_sigma: float = 0.2,
+ empty_mean_umi: int = 200,
+ empty_lognormal_sigma: float = 0.4,
+ epsilon_param: float = 100,
+ rho_alpha: float = 3,
+ rho_beta: float = 80,
+ phi: float = 0.1,
+ random_seed: int = 0) -> Dict[str, Union[float, np.ndarray, sp.csr_matrix]]:
+ """Create a sample dataset for use as 'truth' data.
+ """
+ # Input checking.
+ assert len(cells_of_each_type) > 0, 'cells_of_each_type must be a List of ints'
+ n_cell_types = len(cells_of_each_type)
+ if vector_to_add_to_chi_ambient is not None:
+ assert len(vector_to_add_to_chi_ambient) == n_genes, \
+ f'vector_to_add_to_chi_ambient {vector_to_add_to_chi_ambient.shape} must ' \
+ f'be of length n_genes ({n_genes})'
+ if type(dirichlet_alpha) == np.ndarray:
+ assert dirichlet_alpha.size == n_genes, \
+ f'If you input a vector for dirichlet_alpha, its length ' \
+ f'({dirichlet_alpha.size}) must be n_genes ({n_genes})'
+ assert cell_lognormal_sigma > 0, 'cell_lognormal_sigma must be > 0'
+ assert empty_lognormal_sigma > 0, 'empty_lognormal_sigma must be > 0'
+ cell_mean_umi = list(cell_mean_umi)
+ if len(cell_mean_umi) == 1:
+ cell_mean_umi = cell_mean_umi * n_cell_types # repeat the entry
+ for umi in cell_mean_umi:
+ assert umi > empty_mean_umi, \
+ f'cell_size_mean_umi ({umi}) must be > empty_mean_umi ({empty_mean_umi})'
+ assert len(cell_mean_umi) == len(cells_of_each_type), \
+ 'cells_of_each_type and cell_mean_umi (if a list) must be of the same length'
+ assert (chi_artificial_similarity >= 0) and (chi_artificial_similarity <= 1), \
+ 'chi_artificial_similarity must be in the range [0, 1]'
+ assert phi > 0, 'phi must be > 0'
+
+ # Seed random number generators.
+ comprehensive_random_seed(seed=random_seed)
+
+ # Draw gene expression profiles for each cell type.
+ if chi is None:
+ chi = torch.zeros((n_cell_types, n_genes)).to(DEVICE)
+ for i in range(chi.shape[0]):
+ chi[i, :] = draw_random_chi(alpha=dirichlet_alpha, n_genes=n_genes)
+ if i > 0:
+ # Make expression profiles similar (if chi_artificial_similarity > 0) via gating.
+ chi[i, :] = chi[i, :] * (1 - chi_artificial_similarity) + chi[0, :] * chi_artificial_similarity
else:
-
- # Ambient gene expression comes from the sum of cell expression.
- for i in range(1, clusters+1):
-
- chi[0, :] += cells_in_clusters[i-1] * chi[i, :] # Weighted sum
-
- chi[0, :] = chi[0, :] / np.sum(chi[0, :]) # Normalize
-
- # Sample gene expression for ambient.
- csr, d_n = sample_expression_from(chi[0, :],
- n=n_empty,
- d_mu=np.log(d_empty).item())
-
- # Add data to lists.
- csr_list.append(csr)
- z = z + [0 for _ in range(csr.shape[0])]
- d = d + [i for i in d_n]
-
- # Sample gene expression for cells.
- for i in range(1, clusters+1):
-
- # Get chi for cells once ambient expression is added.
- chi_tilde = chi[i, :] * d_cell + chi[0, :] * d_empty
- chi_tilde = chi_tilde / np.sum(chi_tilde) # Normalize
- csr, d_n = sample_expression_from(chi_tilde,
- n=cells_in_clusters[i-1],
- d_mu=np.log(d_cell).item())
-
- # Add data to lists.
- csr_list.append(csr)
- z = z + [i for _ in range(csr.shape[0])]
- d = d + [j for j in d_n]
-
- # Package the results.
- csr_barcode_gene_synthetic = sp.vstack(csr_list)
- z = np.array(z)
- d = np.array(d)
-
- # Permute the barcode order and return results.
- order = np.random.permutation(z.size)
- csr_barcode_gene_synthetic = csr_barcode_gene_synthetic[order, ...]
- z = z[order]
- d = d[order]
-
- return csr_barcode_gene_synthetic, z, chi, d
-
-
-def generate_chi(alpha: float = 1., n_genes: int = 10000) -> np.ndarray:
+ dirichlet_alpha = ['None: chi was input to generate_sample_model_dataset()']
+
+ # Get chi_ambient: a weighted average of expression, possibly with extra.
+ chi_ambient = torch.zeros(n_genes).to(DEVICE)
+ for i in range(n_cell_types):
+ chi_ambient = chi_ambient + chi[i, :] * cells_of_each_type[i] * cell_mean_umi[i]
+ if vector_to_add_to_chi_ambient is None:
+ vector_to_add_to_chi_ambient = torch.zeros(n_genes)
+ chi_ambient = chi_ambient + vector_to_add_to_chi_ambient
+ chi_ambient = chi_ambient / chi_ambient.sum()
+
+ c_real = torch.zeros((n_droplets, n_genes)).to(DEVICE)
+ c_bkg = torch.zeros((n_droplets, n_genes)).to(DEVICE)
+ labels = torch.zeros(n_droplets).to(DEVICE)
+ epsilon = torch.zeros(n_droplets).to(DEVICE)
+ rho = torch.zeros(n_droplets).to(DEVICE)
+ d_cell = torch.zeros(n_droplets).to(DEVICE)
+ d_empty = torch.zeros(n_droplets).to(DEVICE)
+
+ # Sample counts from cell-containing droplets.
+ i = 0
+ for t, (num_cells, celltype_umi) in enumerate(zip(cells_of_each_type, cell_mean_umi)):
+ sim = sample_from_model(
+ num=num_cells,
+ chi=chi[t, :],
+ d_mu=np.log(celltype_umi),
+ d_sigma=cell_lognormal_sigma,
+ v_mu=np.log(empty_mean_umi),
+ v_sigma=empty_lognormal_sigma,
+ y=1, # cell present
+ chi_ambient=chi_ambient,
+ eps_param=epsilon_param,
+ rho_alpha=rho_alpha,
+ rho_beta=rho_beta,
+ phi=phi,
+ model_type=model_type,
+ )
+ c_real[i:(i + num_cells), :] = sim['counts_real']
+ c_bkg[i:(i + num_cells), :] = sim['counts_bkg']
+ epsilon[i:(i + num_cells)] = sim['epsilon']
+ rho[i:(i + num_cells)] = sim['rho']
+ d_cell[i:(i + num_cells)] = sim['cell_counts']
+ d_empty[i:(i + num_cells)] = sim['empty_counts']
+ labels[i:(i + num_cells)] = t + 1 # label cell types with integers > 0
+ i = i + num_cells
+
+ # Sample counts from empty droplets.
+ sim = sample_from_model(
+ num=n_droplets - i,
+ chi=chi[0, :], # this doesn't get used since y=0
+ d_mu=1, # this doesn't get used since y=0
+ d_sigma=1, # this doesn't get used since y=0
+ v_mu=np.log(empty_mean_umi),
+ v_sigma=empty_lognormal_sigma,
+ y=0, # no cell present
+ chi_ambient=chi_ambient,
+ eps_param=epsilon_param,
+ rho_alpha=rho_alpha,
+ rho_beta=rho_beta,
+ phi=phi,
+ model_type=model_type,
+ )
+ c_real[i:, :] = sim['counts_real']
+ c_bkg[i:, :] = sim['counts_bkg']
+ epsilon[i:] = sim['epsilon']
+ rho[i:] = sim['rho']
+ d_empty[i:] = sim['empty_counts']
+
+ # Get sparse matrices from dense.
+ counts_true = sp.csr_matrix(c_real.cpu().numpy(), shape=c_real.shape, dtype=int)
+ counts_bkg = sp.csr_matrix(c_bkg.cpu().numpy(), shape=c_bkg.shape, dtype=int)
+
+ return {'counts_true': counts_true,
+ 'counts_bkg': counts_bkg,
+ 'barcodes': np.array([f'bc{n:06d}' for n in np.arange(n_droplets)]),
+ 'gene_names': np.array([f'g{n:05d}' for n in np.arange(n_genes)]),
+ 'chi': {str(i + 1): chi[i, :].cpu().numpy() for i in range(chi.shape[0])},
+ 'chi_ambient': chi_ambient.cpu().numpy(),
+ 'droplet_labels': labels.int().cpu().numpy(),
+ 'd_cell': d_cell.cpu().numpy(),
+ 'd_empty': d_empty.cpu().numpy(),
+ 'epsilon': epsilon.cpu().numpy(),
+ 'rho': rho.cpu().numpy(),
+ 'cell_mean_umi': {str(i + 1): u for i, u in enumerate(cell_mean_umi)},
+ 'cell_lognormal_sigma': cell_lognormal_sigma,
+ 'empty_mean_umi': empty_mean_umi,
+ 'empty_lognormal_sigma': empty_lognormal_sigma,
+ 'epsilon_param': epsilon_param,
+ 'dirichlet_alpha': dirichlet_alpha,
+ 'phi': phi,
+ 'model': [model_type]}
+
+
+def draw_random_chi(alpha: float = 1.,
+ n_genes: int = 10000) -> torch.Tensor:
"""Sample a gene expression vector, chi, from a Dirichlet prior.
-
Args:
alpha: Concentration parameter for Dirichlet distribution, to be
expanded into a vector to use as the Dirichlet concentration
parameter.
n_genes: Number of genes.
-
Returns:
chi: Vector of fractional gene expression, drawn from a Dirichlet
distribution.
-
"""
assert alpha > 0, "Concentration parameter, alpha, must be > 0."
assert n_genes > 0, "Number of genes, n_genes, must be > 0."
# Draw gene expression from a Dirichlet distribution.
- chi = np.random.dirichlet(alpha * np.ones(n_genes), size=1).squeeze()
-
- # Normalize gene expression and return result.
- chi = chi / np.sum(chi)
-
+ chi = torch.distributions.Dirichlet(alpha * torch.ones(n_genes).to(DEVICE)).sample().squeeze()
+
return chi
-def sample_expression_from(chi: np.ndarray,
- n: int = 100,
- d_mu: float = np.log(5000).item(),
- d_sigma: float = 0.2,
- phi: float = 0.3) -> Tuple[sp.csr.csr_matrix,
- np.ndarray]:
- """Generate a count matrix given a mean expression distribution.
-
+def sample_from_inferred_model(num: int,
+ model: 'RemoveBackgroundPyroModel',
+ z: Optional[torch.Tensor] = None,
+ y: Optional[torch.Tensor] = None) -> Dict[str, np.ndarray]:
+ """Sample data from the model, where the model is not completely naive, but
+ uses a decoder trained on real data."""
+
+ # Run the model, .
+ data = torch.zeros(size=(num, model.n_genes)) # model uses shape only
+ tracing_model = model.model
+ do_list = []
+ if z is not None:
+ tracing_model = pyro.poutine.do(tracing_model, data={'z': z})
+ do_list.append('z')
+ if y is not None:
+ tracing_model = pyro.poutine.do(tracing_model, data={'y': y})
+ do_list.append('z')
+
+ model_trace = pyro.poutine.trace(tracing_model).get_trace(data)
+
+ # Get outputs of model (mu, alpha, lambda).
+ model_output = model_trace.nodes['_RETURN']['value']
+
+ # Get traced parameters.
+ params = {name: (model_trace.nodes[name]['value'].unconstrained()
+ if (model_trace.nodes[name]['type'] == 'param')
+ else model_trace.nodes[name]['value'])
+ for name in ['chi'] + model_trace.param_nodes + model_trace.stochastic_nodes
+ if (not ('$$$' in name) and name not in do_list)} # the do-samples are dummies
+
+ # Extract latent values.
+ epsilon = params['epsilon']
+ rho = params['rho']
+ d = params['d_cell']
+ d_empty = params['d_empty']
+ y = y if (y is not None) else params['y'] # the "do" does not show up in the trace, apparently
+ chi = params['chi']
+ chi_bar = model.avg_gene_expression
+ chi_ambient = params['chi_ambient'] # pull from the (loaded) param store
+
+ # Because the model integrates out real counts and noise counts, we sample here.
+ logit = torch.log(model_output['mu']) - torch.log(model_output['alpha'])
+ c_real = pyro.distributions.NegativeBinomial(total_count=model_output['alpha'],
+ logits=logit).to_event(1).sample()
+ c_bkg = pyro.distributions.Poisson(model_output['lam']).to_event(1).sample()
+
+ # Output observed counts are the sum, but return them separately.
+ return {'counts_real': c_real.detach().cpu().numpy(),
+ 'counts_bkg': c_bkg.detach().cpu().numpy(),
+ 'cell_counts': (d * y).detach().cpu().numpy(),
+ 'empty_counts': d_empty.detach().cpu().numpy(),
+ 'epsilon': epsilon.detach().cpu().numpy(),
+ 'rho': rho.detach().cpu().numpy(),
+ 'chi': chi.detach().cpu().numpy()}
+
+
+def sample_from_model(
+ num: int,
+ chi: torch.Tensor,
+ d_mu: float,
+ d_sigma: float,
+ v_mu: float,
+ v_sigma: float,
+ y: int,
+ chi_ambient: torch.Tensor,
+ eps_param: float,
+ model_type: str = 'full',
+ rho_alpha: float = 3,
+ rho_beta: float = 80,
+ phi: float = 0.1,
+) -> Dict[str, torch.Tensor]:
+ """Draw samples of cell expression profiles using the negative binomial -
+ Poisson sum model.
+
+ NOTE: Single cell type with expression chi.
+ NOTE: Random number seeding should be done before this function call.
+
Args:
- chi: Normalized gene expression vector (sums to one).
- n: Number of desired cells to simulate.
- d_mu: Log mean number of UMI counts per cell.
- d_sigma: Standard deviation of a normal in log space for the number
- of UMI counts per cell.
- phi: The overdispersion parameter of a negative binomial,
- i.e., variance = mean + phi * mean^2
-
+ num: Number of expression profiles to draw.
+ chi: Cell expression profile, size (n_gene).
+ d_mu: Mean of LogNormal cell size distribution.
+ d_sigma: Scale parameter of LogNormal cell size distribution.
+ v_mu: Mean of LogNormal empty size distribution.
+ v_sigma: Scale parameter of LogNormal empty size distribution.
+ y: 1 for cell(s), 0 for empties.
+ chi_ambient: Ambient gene expression profile (sums to one).
+ eps_param: Parameter for gamma distribution of the epsilon droplet
+ efficiency factor ~ Gamma(eps_param, 1/eps_param), i.e. mean is 1.
+ model_type: ['ambient', 'swapping', 'full']
+ rho_alpha: Beta distribution param alpha for swapping fraction.
+ rho_beta: Beta distribution param beta for swapping fraction.
+ phi: The negative binomial overdispersion parameter.
Returns:
- csr_cell_gene: scipy.sparse.csr_matrix of gene expression
- counts per cell, with cells in axis=0 and genes in axis=1.
-
- Note:
- Draw gene expression from a negative binomial distribution
- counts ~ NB(d*chi, phi)
-
+ Tuple of (c_real, c_bkg)
+ c_real: Count matrix (cells, genes) for real cell counts.
+ c_bkg: Count matrix (cells, genes) for background RNA.
"""
- assert phi > 0, "Phi must be greater than zero in the negative binomial."
- assert d_sigma > 0, "Scale parameter, d_sigma, of LogNormal distribution " \
- " must be greater than zero."
- assert d_mu > 0, "Location parameter, d_mu, of LogNormal distribution " \
- " must be greater than zero."
- assert n > 0, "Number of cells to simulate, n, must be a positive integer."
- assert chi.min() >= 0, "Minimum allowed value in chi vector is zero."
-
- n_genes = chi.size # Number of genes
-
- # Initialize arrays.
- barcodes = np.arange(n)
- genes = np.arange(n_genes)
- predicted_reads = int(np.exp(d_mu) * n * 2) # Guess array sizes
- coo_bc_list = np.zeros(predicted_reads, dtype=np.uint32)
- coo_gene_list = np.zeros(predicted_reads, dtype=np.uint32)
- coo_count_list = np.zeros(predicted_reads, dtype=np.uint32)
- d = np.zeros(n)
- a = 0
-
- # Go barcode by barcode, sampling UMI counts per gene.
- for i in range(n):
-
- # Sample cell size parameter from a LogNormal distribution.
- d[i] = np.exp(np.random.normal(loc=d_mu, scale=d_sigma, size=1))
-
- # Sample counts from a negative binomial distribution.
- gene_counts = neg_binom(d[i] * chi, phi, size=n_genes)
-
- # Keep only the non-zero counts to populate the sparse matrix.
- num_nonzeros = np.sum(gene_counts > 0)
-
- # Check whether arrays need to be re-sized to accommodate more entries.
- if (a + num_nonzeros) < coo_count_list.size:
- # Fill in.
- coo_bc_list[a:a+num_nonzeros] = barcodes[i]
- coo_gene_list[a:a+num_nonzeros] = genes[gene_counts > 0]
- coo_count_list[a:a+num_nonzeros] = gene_counts[gene_counts > 0]
- else:
- # Resize arrays by doubling.
- coo_bc_list = np.resize(coo_bc_list, coo_bc_list.size * 2)
- coo_gene_list = np.resize(coo_gene_list, coo_gene_list.size * 2)
- coo_count_list = np.resize(coo_count_list, coo_count_list.size * 2)
- # Fill in.
- coo_bc_list[a:a+num_nonzeros] = barcodes[i]
- coo_gene_list[a:a+num_nonzeros] = genes[gene_counts > 0]
- coo_count_list[a:a+num_nonzeros] = gene_counts[gene_counts > 0]
-
- a += num_nonzeros
-
- # Lop off any unused zero entries at the end of the arrays.
- coo_bc_list = coo_bc_list[coo_count_list > 0]
- coo_gene_list = coo_gene_list[coo_count_list > 0]
- coo_count_list = coo_count_list[coo_count_list > 0]
-
- # Package data into a scipy.sparse.coo.coo_matrix.
- count_matrix = sp.coo_matrix((coo_count_list, (coo_bc_list, coo_gene_list)),
- shape=(barcodes.size, n_genes),
- dtype=np.uint32)
-
- # Convert to a scipy.sparse.csr.csr_matrix and return.
- count_matrix = count_matrix.tocsr()
-
- return count_matrix, d
-
-
-def neg_binom(mu: float, phi: float, size: int = 1) -> np.ndarray:
- """Parameterize numpy's negative binomial distribution
- in terms of the mean and the overdispersion.
+ # Check inputs.
+ assert y in [0, 1], f'y must be 0 or 1, but was {y}'
+ assert d_mu > 0, f'd_mu must be > 0, but was {d_mu}'
+ assert d_sigma > 0, f'd_sigma must be > 0, but was {d_sigma}'
+ assert v_mu > 0, f'v_mu must be > 0, but was {v_mu}'
+ assert v_sigma > 0, f'v_sigma must be > 0, but was {v_sigma}'
+ assert (chi >= 0).all(), 'all chi values must be >= 0.'
+ assert (chi_ambient >= 0).all(), 'all chi_ambient must be >= 0.'
+ assert (1. - chi_ambient.sum()).abs() < 1e-6, f'chi_ambient must sum ' \
+ f'to 1, but it sums to {chi_ambient.sum()}'
+ assert len(chi_ambient.shape) == 1, 'chi_ambient should be 1-dimensional.'
+ assert chi.shape[-1] == chi_ambient.shape[-1], 'chi and chi_ambient must ' \
+ 'be the same size in the rightmost dimension.'
+ assert (1. - chi.sum()).abs() < 1e-6, f'chi must sum ' \
+ f'to 1, but it sums to {chi.sum()}'
+ assert num > 0, f'num must be > 0, but was {num}'
+ assert eps_param > 1, f'eps_param must be > 1, but was {eps_param}'
- Args:
- mu: Mean of the distribution
- phi: Overdispersion, such that variance = mean + phi * mean^2
- size: How many numbers to return
+ # Draw epsilon ~ Gamma(eps_param, 1 / eps_param)
+ epsilon = torch.distributions.Gamma(concentration=eps_param, rate=eps_param).sample([num])
- Returns:
- 'size' number of random draws from a negative binomial distribution.
+ # Draw d ~ LogNormal(d_mu, d_sigma)
+ d = torch.distributions.LogNormal(loc=d_mu, scale=d_sigma).sample([num])
- Note:
- Setting phi=0 turns the negative binomial distribution into a
- Poisson distribution.
+ # Draw d ~ LogNormal(d_mu, d_sigma)
+ v = torch.distributions.LogNormal(loc=v_mu, scale=v_sigma).sample([num])
- """
+ # Draw rho ~ Beta(rho_alpha, rho_beta)
+ rho = torch.distributions.Beta(rho_alpha, rho_beta).sample([num])
+
+ mu = calculate_mu(epsilon=epsilon,
+ d_cell=d,
+ chi=chi,
+ y=torch.ones(d.shape).to(DEVICE) * y,
+ rho=rho,
+ model_type=model_type) + 1e-30
+
+ # Draw cell counts ~ NB(mean = y * epsilon * d * chi, overdispersion = phi)
+ alpha = 1. / phi
+ logits = (mu.log() - np.log(alpha))
+ c_real = torch.distributions.NegativeBinomial(total_count=alpha,
+ logits=logits).sample()
+
+ lam = calculate_lambda(epsilon=epsilon,
+ chi_ambient=chi_ambient,
+ d_cell=d,
+ d_empty=v,
+ y=torch.ones(d.shape).to(DEVICE) * y,
+ rho=rho,
+ chi_bar=chi_ambient,
+ model_type=model_type) + 1e-30
- assert phi > 0, "Phi must be greater than zero in the negative binomial."
- assert size > 0, "Number of draws from negative binomial, size, must " \
- "be a positive integer."
+ # Draw empty counts ~ Poisson(epsilon * v * chi_ambient)
+ c_bkg = torch.distributions.Poisson(rate=lam).sample()
- n = 1. / phi
- p = n / (mu + n)
- return np.random.negative_binomial(n, p, size=size)
+ # Output observed counts are the sum, but return them separately.
+ return {'counts_real': c_real,
+ 'counts_bkg': c_bkg,
+ 'cell_counts': d * y,
+ 'empty_counts': v,
+ 'epsilon': epsilon,
+ 'rho': rho}
def sample_from_dirichlet_model(num: int,
@@ -388,14 +710,12 @@ def sample_from_dirichlet_model(num: int,
y: int,
chi_ambient: np.ndarray,
eps_param: float,
- random_seed: int = 0,
- include_swapping: bool = False,
+ rng: Optional[np.random.RandomState] = None,
+ model_type: str = 'full',
rho_alpha: float = 3,
- rho_beta: float = 80) -> Tuple[np.ndarray,
- np.ndarray]:
+ rho_beta: float = 80) -> Dict[str, np.ndarray]:
"""Draw samples of cell expression profiles using the Dirichlet-Poisson,
Poisson sum model.
-
Args:
num: Number of expression profiles to draw.
alpha: Dirichlet concentration parameters for cell expression profile,
@@ -408,16 +728,14 @@ def sample_from_dirichlet_model(num: int,
chi_ambient: Ambient gene expression profile (sums to one).
eps_param: Parameter for gamma distribution of the epsilon droplet
efficiency factor ~ Gamma(eps_param, 1/eps_param), i.e. mean is 1.
- random_seed: Seed a random number generator.
- include_swapping: Whether to include swapping in the model.
+ rng: A random number generator.
+ model_type: ['ambient', 'swapping', 'full']
rho_alpha: Beta distribution param alpha for swapping fraction.
rho_beta: Beta distribution param beta for swapping fraction.
-
Returns:
Tuple of (c_real, c_bkg)
c_real: Count matrix (cells, genes) for real cell counts.
c_bkg: Count matrix (cells, genes) for background RNA.
-
"""
# Check inputs.
@@ -427,9 +745,9 @@ def sample_from_dirichlet_model(num: int,
assert v_mu > 0, f'v_mu must be > 0, but was {v_mu}'
assert v_sigma > 0, f'v_sigma must be > 0, but was {v_sigma}'
assert np.all(alpha > 0), 'all alphas must be > 0.'
- assert np.all(chi_ambient > 0), 'all chi_ambient must be > 0.'
- assert np.abs(1. - chi_ambient.sum()) < 1e-10, f'chi_ambient must sum to 1, but it sums ' \
- f'to {chi_ambient.sum()}'
+ assert np.all(chi_ambient >= 0), 'all chi_ambient must be >= 0.'
+ assert np.abs(1. - chi_ambient.sum()) < 1e-10, f'chi_ambient must sum ' \
+ f'to 1, but it sums to {chi_ambient.sum()}'
assert len(chi_ambient.shape) == 1, 'chi_ambient should be 1-dimensional.'
assert alpha.shape[0] == chi_ambient.size, 'alpha and chi_ambient must ' \
'be the same size in the rightmost dimension.'
@@ -437,7 +755,8 @@ def sample_from_dirichlet_model(num: int,
assert eps_param > 1, f'eps_param must be > 1, but was {eps_param}'
# Seed random number generator.
- rng = np.random.RandomState(seed=random_seed)
+ if rng is None:
+ rng = np.random.RandomState(seed=0)
# Draw chi ~ Dir(alpha)
chi = rng.dirichlet(alpha=alpha, size=num)
@@ -454,91 +773,138 @@ def sample_from_dirichlet_model(num: int,
# Draw rho ~ Beta(rho_alpha, rho_beta)
rho = rng.beta(a=rho_alpha, b=rho_beta, size=num)
- # print(f'eps.shape is {epsilon.shape}')
- # print(f'd.shape is {d.shape}')
- # print(f'chi.shape is {chi.shape}')
- # print(f'rho.shape is {rho.shape}')
-
- mu = _calculate_mu(model_type='ambient' if not include_swapping else 'full',
- epsilon=torch.Tensor(epsilon),
- d_cell=torch.Tensor(d),
- chi=torch.Tensor(chi),
- y=torch.ones(d.shape) * y,
- rho=torch.Tensor(rho)).numpy()
+ mu = calculate_mu(epsilon=torch.tensor(epsilon),
+ d_cell=torch.tensor(d),
+ chi=torch.tensor(chi),
+ y=torch.ones(d.shape) * y,
+ rho=torch.tensor(rho),
+ model_type=model_type).numpy()
# Draw cell counts ~ Poisson(y * epsilon * d * chi)
- c_real = rng.poisson(lam=mu,
- size=(num, chi_ambient.size))
-
- lam = _calculate_lambda(model_type='ambient' if not include_swapping else 'full',
- epsilon=torch.Tensor(epsilon),
- chi_ambient=torch.Tensor(chi_ambient),
- d_cell=torch.Tensor(d),
- d_empty=torch.Tensor(v),
- y=torch.ones(d.shape) * y,
- rho=torch.Tensor(rho),
- chi_bar=torch.Tensor(chi_ambient)).numpy()
+ c_real = rng.poisson(lam=mu, size=(num, chi_ambient.size))
+
+ lam = calculate_lambda(epsilon=torch.tensor(epsilon),
+ chi_ambient=torch.tensor(chi_ambient),
+ d_cell=torch.tensor(d),
+ d_empty=torch.tensor(v),
+ y=torch.ones(d.shape) * y,
+ rho=torch.tensor(rho),
+ chi_bar=torch.tensor(chi_ambient),
+ model_type=model_type).numpy()
# Draw empty counts ~ Poisson(epsilon * v * chi_ambient)
- c_bkg = rng.poisson(lam=lam,
- size=(num, chi_ambient.size))
+ c_bkg = rng.poisson(lam=lam, size=(num, chi_ambient.size))
# Output observed counts are the sum, but return them separately.
- return c_real, c_bkg
-
-
-def _calculate_lambda(model_type: str,
- epsilon: torch.Tensor,
- chi_ambient: torch.Tensor,
- d_empty: torch.Tensor,
- y: Union[torch.Tensor, None] = None,
- d_cell: Union[torch.Tensor, None] = None,
- rho: Union[torch.Tensor, None] = None,
- chi_bar: Union[torch.Tensor, None] = None):
- """Calculate noise rate based on the model."""
-
- if model_type == "simple" or model_type == "ambient":
- lam = epsilon.unsqueeze(-1) * d_empty.unsqueeze(-1) * chi_ambient
-
- elif model_type == "swapping":
- lam = (rho.unsqueeze(-1) * y.unsqueeze(-1)
- * epsilon.unsqueeze(-1) * d_cell.unsqueeze(-1)
- + d_empty.unsqueeze(-1)) * chi_bar
-
- elif model_type == "full":
- lam = ((1 - rho.unsqueeze(-1)) * d_empty.unsqueeze(-1) * chi_ambient.unsqueeze(0)
- + rho.unsqueeze(-1)
- * (y.unsqueeze(-1) * epsilon.unsqueeze(-1) * d_cell.unsqueeze(-1)
- + d_empty.unsqueeze(-1)) * chi_bar)
- else:
- raise NotImplementedError(f"model_type was set to {model_type}, "
- f"which is not implemented.")
+ return {'counts_real': c_real,
+ 'counts_bkg': c_bkg,
+ 'cell_counts': d * y,
+ 'empty_counts': v,
+ 'epsilon': epsilon,
+ 'rho': rho}
- return lam
+def get_dataset_dict_as_anndata(
+ sample_dataset: Dict[str, Union[float, np.ndarray, sp.csr_matrix]]
+) -> 'anndata.AnnData':
+ """Return a simulated dataset as an AnnData object."""
-def _calculate_mu(model_type: str,
- epsilon: torch.Tensor,
- d_cell: torch.Tensor,
- chi: torch.Tensor,
- y: Union[torch.Tensor, None] = None,
- rho: Union[torch.Tensor, None] = None):
- """Calculate mean expression based on the model."""
+ import anndata
- if model_type == 'simple':
- mu = epsilon.unsqueeze(-1) * d_cell.unsqueeze(-1) * chi
+ d = sample_dataset.copy()
- elif model_type == 'ambient':
- mu = (y.unsqueeze(-1) * epsilon.unsqueeze(-1)
- * d_cell.unsqueeze(-1) * chi)
+ # counts
+ if 'counts_true' in d.keys() and 'counts_bkg' in d.keys():
+ adata = anndata.AnnData(X=(d['counts_true'] + d['counts_bkg']).astype(np.float32),
+ obs={'barcode': d.pop('barcodes')},
+ var={'gene': d.pop('gene_names')})
+ adata.layers['counts_true'] = d.pop('counts_true')
+ adata.layers['counts'] = adata.layers['counts_true'] + d.pop('counts_bkg')
+ elif 'matrix' in d.keys():
+ adata = anndata.AnnData(X=d.pop('matrix').astype(np.float32),
+ obs={'barcode': d.pop('barcodes')},
+ var={'gene': d.pop('gene_names')})
- elif model_type == 'swapping' or model_type == 'full':
- mu = ((1 - rho.unsqueeze(-1))
- * y.unsqueeze(-1) * epsilon.unsqueeze(-1)
- * d_cell.unsqueeze(-1) * chi)
+ # obs
+ obs_keys = ['droplet_labels', 'd_cell', 'epsilon']
+ for key in obs_keys:
+ adata.obs[key] = d.pop(key, None)
- else:
- raise NotImplementedError(f"model_type was set to {model_type}, "
- f"which is not implemented.")
+ # uns
+ for key, value in d.items():
+ adata.uns[key] = value
+
+ return adata
+
+
+def write_simulated_data_to_h5(output_file: str,
+ d: Dict[str, Union[float, np.ndarray, sp.csr_matrix]],
+ cellranger_version: int = 3) -> bool:
+ """Helper function to write the full (noisy) simulate dataset to an H5 file.
+
+ Args:
+ output_file: File name
+ d: Resulting dict from `generate_sample_dataset`
+
+ Returns:
+ True if save was successful.
+ """
+
+ assert cellranger_version in [2, 3], 'cellranger_version must be 2 or 3'
+
+ return write_matrix_to_cellranger_h5(
+ output_file=output_file,
+ gene_names=d['gene_names'],
+ feature_types=np.array(['Gene Expression'] * d['gene_names'].size),
+ genomes=(np.array(['simulated'] * d['gene_names'].size)
+ if ('genome' not in d.keys()) else d['genome']),
+ barcodes=d['barcodes'],
+ count_matrix=(d['counts_true'] + d['counts_bkg']).tocsc(),
+ cellranger_version=cellranger_version,
+ )
+
+
+def write_simulated_truth_to_h5(output_file: str,
+ d: Dict[str, Union[float, np.ndarray, sp.csr_matrix]],
+ cellranger_version: int = 3) -> bool:
+ """Helper function to write the full (noisy) simulate dataset to an H5 file.
+
+ Args:
+ output_file: File name
+ d: Resulting dict from `generate_sample_dataset`
+
+ Returns:
+ True if save was successful.
+ """
- return mu
+ assert cellranger_version in [2, 3], 'cellranger_version must be 2 or 3'
+
+ latents = set(d.keys())
+ extra_latents = latents - {'gene_names', 'genome', 'barcodes', 'counts_true',
+ 'droplet_labels', 'd_cell', 'd_empty', 'epsilon', 'rho',
+ 'chi', 'cell_mean_umi', 'counts_bkg'}
+ global_latents = {f'truth_{key.replace("chi_ambient", "ambient_expression")}': d[key]
+ for key in extra_latents}
+
+ for i, v in d['chi'].items():
+ global_latents.update({f'truth_gene_expression_cell_label_{i}': v})
+ for i, v in d['cell_mean_umi'].items():
+ global_latents.update({f'truth_mean_umi_cell_label_{i}': v})
+
+ return write_matrix_to_cellranger_h5(
+ output_file=output_file,
+ gene_names=d['gene_names'],
+ feature_types=np.array(['Gene Expression'] * d['gene_names'].size),
+ genomes=(np.array(['simulated'] * d['gene_names'].size)
+ if ('genome' not in d.keys()) else d['genome']),
+ barcodes=d['barcodes'],
+ count_matrix=d['counts_true'].tocsc(), # just the truth, no background
+ cellranger_version=cellranger_version,
+ local_latents={'truth_cell_label': d['droplet_labels'],
+ 'truth_cell_size': d['d_cell'],
+ 'truth_empty_droplet_size': d['d_empty'],
+ 'truth_droplet_efficiency': d['epsilon'],
+ 'truth_cell_probability': (d['droplet_labels'] != 0).astype(float),
+ 'truth_swapping_fraction': d['rho']},
+ global_latents=global_latents,
+ )
diff --git a/cellbender/remove_background/data/io.py b/cellbender/remove_background/data/io.py
new file mode 100644
index 00000000..8c1fec7a
--- /dev/null
+++ b/cellbender/remove_background/data/io.py
@@ -0,0 +1,1207 @@
+"""Handle input parsing and output writing."""
+
+import tables
+import anndata
+import numpy as np
+import scipy.sparse as sp
+import scipy.io as io
+
+from cellbender.remove_background import consts
+
+from typing import Dict, Union, List, Optional, Callable
+import logging
+import os
+import gzip
+import traceback
+
+
+logger = logging.getLogger('cellbender')
+
+
+class IngestedData(dict):
+ """Small container object for the results of file loading. This is a way to
+ ensure that all filetypes are loaded into the same general format that can
+ be used in dataset.py
+
+ NOTE: This really exists to ensure all these fields are present, and to
+ force each loader to specify each field
+ """
+
+ def __init__(self, matrix, barcodes,
+ gene_names, gene_ids, feature_types, genomes,
+ **kwargs):
+ # Fill in some fields no matter the input source (for loading in scanpy)
+ blank_array = np.array(['NA'] * len(gene_names))
+ if genomes is None:
+ genomes = blank_array
+ if gene_ids is None:
+ gene_ids = blank_array
+ if feature_types is None:
+ feature_types = blank_array
+
+ # Warn if file looks filtered.
+ if len(barcodes) < consts.MINIMUM_BARCODES_H5AD:
+ logger.warning(f'WARNING: Only {len(barcodes)} barcodes in the input file. '
+ f'Ensure this is a raw (unfiltered) file with all barcodes, '
+ f'including the empty droplets.')
+
+ # Required values, some of which can be None
+ super().__init__([('matrix', matrix),
+ ('barcodes', barcodes),
+ ('gene_names', gene_names),
+ ('gene_ids', gene_ids),
+ ('feature_types', feature_types),
+ ('genomes', genomes)])
+ self.update(**kwargs) # cellranger version, for example, is optional
+
+
+class FileLoader:
+ """Make explicit guarantees about what a file-loading method yields."""
+
+ def __init__(self, load_fn):
+ self.load_fn = load_fn
+
+ def load(self, file) -> IngestedData:
+ data = self.load_fn(file)
+ return IngestedData(**data)
+
+
+def write_matrix_to_cellranger_h5(
+ cellranger_version: int,
+ output_file: str,
+ gene_names: np.ndarray,
+ barcodes: np.ndarray,
+ count_matrix: sp.csc_matrix,
+ feature_types: Optional[np.ndarray] = None,
+ gene_ids: Optional[np.ndarray] = None,
+ genomes: Optional[np.ndarray] = None,
+ local_latents: Dict[str, Optional[np.ndarray]] = {},
+ global_latents: Dict[str, Optional[np.ndarray]] = {},
+ metadata: Dict[str, Optional[Union[np.ndarray, int, str, Dict]]] = {}) -> bool:
+ """Write count matrix data to output HDF5 file using CellRanger format.
+
+ Args:
+ cellranger_version: Either 2 or 3. Determines the format of the output
+ h5 file.
+ output_file: Path to output .h5 file (e.g., 'output.h5').
+ gene_names: Name of each gene (column of count matrix).
+ gene_ids: Ensembl ID of each gene (column of count matrix).
+ genomes: Name of the genome that each gene comes from.
+ feature_types: Type of each feature (column of count matrix).
+ barcodes: Name of each barcode (row of count matrix).
+ count_matrix: Count matrix to be written to file, in sparse
+ format. Rows are barcodes, columns are genes.
+ local_latents: Local latent variables. Should include one key called
+ 'barcodes' which specifies the droplets being referred to.
+ global_latents: Global latent variables.
+ metadata: Other metadata like loss per epoch and FPR, etc.
+
+ Note:
+ To match the CellRanger .h5 files, the matrix is stored as its
+ transpose, with rows as genes and cell barcodes as columns.
+
+ """
+
+ assert isinstance(count_matrix, sp.csc_matrix), \
+ "The count matrix must be csc_matrix format in order to write to HDF5."
+
+ assert gene_names.size == count_matrix.shape[1], \
+ "The number of gene names must match the number of columns in the count matrix."
+
+ if gene_ids is not None:
+ assert gene_names.size == gene_ids.size, \
+ f"The number of gene_names {gene_names.shape} must match " \
+ f"the number of gene_ids {gene_ids.shape}."
+
+ if feature_types is not None:
+ assert gene_names.size == feature_types.size, \
+ f"The number of gene_names {gene_names.shape} must match " \
+ f"the number of feature_types {feature_types.shape}."
+
+ if genomes is not None:
+ assert gene_names.size == genomes.size, \
+ "The number of gene_names must match the number of genome designations."
+
+ assert barcodes.size == count_matrix.shape[0], \
+ "The number of barcodes must match the number of rows in the count matrix."
+
+ # This reverses the role of rows and columns, to match CellRanger format.
+ count_matrix = count_matrix.transpose().tocsc()
+
+ # Write to output file.
+ filters = tables.Filters(complevel=1, complib='zlib', shuffle=True)
+ filter_noshuffle = tables.Filters(complevel=1, complib='zlib', shuffle=False)
+ with tables.open_file(output_file, "w",
+ title="CellBender remove-background output") as f:
+
+ if cellranger_version == 2:
+
+ # Create the group where count data will be stored
+ group = f.create_group("/", "matrix_v2", "Counts after background correction")
+
+ # Create arrays within that group for gene info.
+ f.create_carray(group, "gene_names", obj=gene_names, filters=filters)
+ if gene_ids is None:
+ # some R loaders require unique values here
+ gene_ids = np.array([f'NA_{i}' for i in range(gene_names.size)])
+ f.create_carray(group, "genes", obj=gene_ids, filters=filters)
+ if genomes is None:
+ genomes = np.array(['NA'] * gene_names.size)
+ f.create_carray(group, "genome", obj=genomes, filters=filters)
+
+ elif cellranger_version == 3:
+
+ # Create the group where count data will be stored
+ group = f.create_group("/", "matrix", "Counts after background correction")
+
+ # Create a sub-group called "features"
+ feature_group = f.create_group(group, "features",
+ "Genes and other features measured")
+
+ # Create arrays within that group for feature info.
+ f.create_carray(feature_group, "name", obj=gene_names, filters=filters)
+ if gene_ids is None:
+ # some R loaders require unique values here
+ gene_ids = np.array([f'NA_{i}' for i in range(gene_names.size)])
+ f.create_carray(feature_group, "id", obj=gene_ids, filters=filters)
+ if feature_types is None:
+ feature_types = np.array(['Gene Expression'] * gene_names.size)
+ f.create_carray(feature_group, "feature_type", obj=feature_types, filters=filters)
+ if genomes is None:
+ genomes = np.array(['NA'] * gene_names.size)
+ f.create_carray(feature_group, "genome", obj=genomes, filters=filters)
+
+ # TODO: Copy the other extraneous information from the input file.
+ # (Some user might need it for some reason.)
+
+ else:
+ raise ValueError(f'Trying to save to CellRanger v{cellranger_version} '
+ f'format, which is not implemented.')
+
+ # Code for both versions.
+ f.create_carray(group, "barcodes", obj=barcodes, filters=filter_noshuffle)
+
+ # Create arrays to store the count data.
+ f.create_carray(group, "data", obj=count_matrix.data, filters=filters)
+ f.create_carray(group, "indices", obj=count_matrix.indices, filters=filters)
+ f.create_carray(group, "indptr", obj=count_matrix.indptr, filters=filters)
+ f.create_carray(group, "shape", atom=tables.Int32Atom(),
+ obj=np.array(count_matrix.shape, dtype=np.int32), filters=filters)
+
+ # Store local latent variables.
+ droplet_latent_group = f.create_group("/", "droplet_latents", "Latent variables per droplet")
+ for key, value in local_latents.items():
+ if value is not None:
+ f.create_carray(droplet_latent_group, key, obj=value, filters=filters)
+
+ # Store global latent variables.
+ global_group = f.create_group("/", "global_latents", "Global latent variables")
+ for key, value in global_latents.items():
+ if value is not None:
+ f.create_array(global_group, key, value)
+
+ def create_nonscalar_metadata_array(f, group, k, v):
+ """Wrap scalar or string values in lists"""
+ if v is None:
+ return
+ if (type(v) == list) or (type(v) == np.ndarray):
+ f.create_array(group, k, v)
+ else:
+ f.create_array(group, k, [v])
+
+ # Store metadata.
+ metadata_group = f.create_group("/", "metadata", "Metadata")
+ for key, value in metadata.items():
+ for k, v in unravel_dict(key, value).items():
+ create_nonscalar_metadata_array(f, metadata_group, k, v)
+
+ logger.info(f"Succeeded in writing CellRanger "
+ f"format output to file {output_file}")
+
+ return True
+
+
+def write_posterior_coo_to_h5(
+ output_file: str,
+ posterior_coo: sp.coo_matrix,
+ noise_count_offsets: Dict[int, int],
+ latents: Dict[str, np.ndarray],
+ feature_inds: np.ndarray,
+ barcode_inds: np.ndarray,
+ regularized_posterior_coo: Optional[sp.coo_matrix] = None,
+ posterior_kwargs: Optional[Dict] = None,
+ regularized_posterior_kwargs: Optional[Dict] = None) -> bool:
+ """Write sparse COO matrix to an HDF5 file, using compression.
+
+ NOTE: COO matrix is indexed by rows 'm' which each map to a unique
+ (cell, feature). The cell and feature are denoted in the barcode_inds
+ and feature_inds arrays. The column indices for the COO matrix are the
+ number of noise counts for each entry in count matrix, starting with zero,
+ except these noise count values get added to noise_count_offsets, which is
+ length m.
+
+ Args:
+ output_file: Path to output .h5 file (e.g., 'output.h5').
+ posterior_coo: Posterior to be written to file, in sparse COO [m, c]
+ format. Rows are 'm'-index, columns are number of noise counts.
+ noise_count_offsets: The number of noise counts at which each 'm' starts.
+ Absence of an 'm'-index from the keys of this dict means that the
+ corresponding 'm'-index starts at 0 noise counts.
+ latents: MAP values of latent variables for each analyzed barcode.
+ barcode_inds: Index of each barcode (row of input count matrix).
+ feature_inds: Index of each feature (column of input count matrix).
+ regularized_posterior_coo: Regularized posterior.
+ posterior_kwargs: Keyword arguments used to generate posterior (for
+ caching)
+ regularized_posterior_kwargs: Keyword arguments used to generate
+ posterior (for caching)
+
+ """
+
+ assert isinstance(posterior_coo, sp.coo_matrix), \
+ "The posterior must be coo_matrix format in order to write to HDF5."
+
+ assert barcode_inds.size == posterior_coo.row.size, \
+ "len(barcode_inds) must match the number of entries in the posterior COO"
+
+ assert feature_inds.size == posterior_coo.row.size, \
+ "len(feature_inds) must match the number of entries in the posterior COO"
+
+ # Write to output file.
+ filters = tables.Filters(complevel=1, complib='zlib', shuffle=True)
+ with tables.open_file(
+ output_file,
+ "w",
+ title="CellBender remove-background posterior noise count probabilities"
+ ) as f:
+
+ # metadata
+ extras = f.create_group("/", "metadata", "Posterior metadata")
+ f.create_carray(extras, "barcode_inds", obj=barcode_inds, filters=filters)
+ f.create_carray(extras, "feature_inds", obj=feature_inds, filters=filters)
+ if noise_count_offsets != {}:
+ f.create_carray(extras, "noise_count_offsets_keys",
+ obj=list(noise_count_offsets.keys()), filters=filters)
+ f.create_carray(extras, "noise_count_offsets_values",
+ obj=list(noise_count_offsets.values()), filters=filters)
+
+ # posterior COO
+ group = f.create_group("/", "posterior_noise_log_prob", "Posterior noise count log probabilities")
+ f.create_carray(group, "log_prob", obj=posterior_coo.data, filters=filters)
+ f.create_carray(group, "m_index", obj=posterior_coo.row, filters=filters)
+ f.create_carray(group, "noise_count", obj=posterior_coo.col, filters=filters)
+ f.create_carray(group, "shape", atom=tables.Int64Atom(),
+ obj=np.array(posterior_coo.shape, dtype=np.int64), filters=filters)
+
+ # regularized posterior COO
+ if regularized_posterior_coo is not None:
+ group = f.create_group("/", "regularized_posterior_noise_log_prob",
+ "Regularized posterior noise count log probabilities")
+ f.create_carray(group, "log_prob", obj=regularized_posterior_coo.data, filters=filters)
+ f.create_carray(group, "m_index", obj=regularized_posterior_coo.row, filters=filters)
+ f.create_carray(group, "noise_count", obj=regularized_posterior_coo.col, filters=filters)
+ f.create_carray(group, "shape", atom=tables.Int64Atom(),
+ obj=np.array(regularized_posterior_coo.shape, dtype=np.int64), filters=filters)
+
+ # latents
+ droplet_latent_group = f.create_group("/", "droplet_latents_map", "Latent variables per droplet")
+ for key, value in latents.items():
+ if value is not None:
+ f.create_carray(droplet_latent_group, key, obj=value, filters=filters)
+
+ # kwargs
+ if posterior_kwargs is not None:
+ kwargs_group = f.create_group("/", "kwargs", "Function arguments for posterior")
+ for key, value in posterior_kwargs.items():
+ for k, v in unravel_dict(key, value).items():
+ if type(v) == str:
+ v = np.array([v], dtype=str)
+ f.create_array(kwargs_group, k, v)
+ reg_kwargs_group = f.create_group("/", "kwargs_regularized",
+ "Function arguments for regularized posterior")
+ if regularized_posterior_kwargs is not None:
+ for key, value in regularized_posterior_kwargs.items():
+ for k, v in unravel_dict(key, value).items():
+ if type(v) == str:
+ v = np.array([v], dtype=str)
+ f.create_array(reg_kwargs_group, k, v)
+
+ logger.info(f"Succeeded in writing posterior to file {output_file}")
+
+ return True
+
+
+def load_posterior_from_h5(filename: str) -> Dict[str, Union[sp.coo_matrix, np.ndarray]]:
+ """Load a posterior noise count COO from an h5 file.
+
+ Args:
+ filename: string path to .h5 file that contains the raw gene
+ barcode matrices
+
+ Returns:
+ Dict with ['coo', 'noise_count_offsets', 'barcode_inds', 'feature_inds']
+ Posterior noise count COO
+ Noise count offsets for COO rows
+ Droplet indices for COO rows
+ Feature indices for COO rows
+ """
+
+ with tables.open_file(filename, 'r') as f:
+
+ # read metadata
+ barcode_inds = getattr(f.root.metadata, 'barcode_inds').read()
+ feature_inds = getattr(f.root.metadata, 'barcode_inds').read()
+ if hasattr(f.root.metadata, 'noise_count_offsets_keys'):
+ noise_count_offsets_keys = getattr(f.root.metadata, 'noise_count_offsets_keys').read()
+ noise_count_offsets_values = getattr(f.root.metadata, 'noise_count_offsets_values').read()
+ noise_count_offsets = dict(zip(noise_count_offsets_keys, noise_count_offsets_values))
+ else:
+ noise_count_offsets = {}
+
+ def _read_coo(group: tables.Group) -> sp.coo_matrix:
+ data = getattr(group, 'log_prob').read()
+ row = getattr(group, 'm_index').read()
+ col = getattr(group, 'noise_count').read()
+ shape = getattr(group, 'shape').read()
+ return sp.coo_matrix((data, (row, col)), shape=shape)
+
+ # read coo
+ posterior_coo = _read_coo(group=f.root.posterior_noise_log_prob)
+
+ # read regularized coo
+ if hasattr(f.root, 'regularized_posterior_noise_log_prob'):
+ regularized_posterior_coo = _read_coo(group=f.root.regularized_posterior_noise_log_prob)
+ else:
+ regularized_posterior_coo = None
+
+ def _read_as_dict(group: tables.Group) -> Dict:
+ d = {}
+ for n in group._f_walknodes('Leaf'):
+ val = n.read()
+ if (type(val) == np.ndarray) and ('S' in val.dtype.kind):
+ val = val.item().decode()
+ d.update({n.name: val})
+ return d
+
+ # read latents
+ latents = _read_as_dict(group=f.root.droplet_latents_map)
+
+ # read kwargs
+ if hasattr(f.root, 'kwargs'):
+ kwargs = _read_as_dict(group=f.root.kwargs)
+ else:
+ kwargs = None
+
+ if hasattr(f.root, 'kwargs_regularized'):
+ kwargs_regularized = _read_as_dict(group=f.root.kwargs_regularized)
+ else:
+ kwargs_regularized = None
+
+ # Issue warnings if necessary, based on dimensions matching.
+ if posterior_coo.row.size != barcode_inds.size:
+ logger.warning(f"Number of barcode_inds ({barcode_inds.size}) "
+ f"in {filename} does not match the number expected from "
+ f"the sparse COO matrix ({posterior_coo.shape[0]}).")
+ if posterior_coo.row.size != feature_inds.size:
+ logger.warning(f"Number of feature_inds ({feature_inds.size}) "
+ f"in {filename} does not match the number expected from "
+ f"the sparse COO matrix ({posterior_coo.shape[0]}).")
+
+ return {'coo': posterior_coo,
+ 'kwargs': kwargs,
+ 'regularized_coo': regularized_posterior_coo,
+ 'kwargs_regularized': kwargs_regularized,
+ 'latents': latents,
+ 'noise_count_offsets': noise_count_offsets,
+ 'feature_inds': feature_inds,
+ 'barcode_inds': barcode_inds}
+
+
+def unravel_dict(pref: str, d: Dict) -> Dict:
+ """Unravel a nested dict, returning a dict with values that are not dicts"""
+
+ if type(d) != dict:
+ return {pref: d}
+ out_d = {}
+ for k, v in d.items():
+ out_d.update({pref + '_' + key: val for key, val in unravel_dict(k, v).items()})
+ return out_d
+
+
+def load_data(input_file: str)\
+ -> Dict[str, Union[sp.csr_matrix, List[np.ndarray], np.ndarray]]:
+ """Load a dataset into the SingleCellRNACountsDataset object from
+ the self.input_file"""
+
+ # Detect input data type.
+ load_fn = choose_data_loader(input_file=input_file)
+
+ # Load data using the appropriate loader.
+ logger.info(f"Loading data from {input_file}")
+ data = FileLoader(load_fn).load(input_file)
+
+ return data
+
+
+def choose_data_loader(input_file: str) -> Callable:
+ """Detect the type of input data and return the relevant load function."""
+
+ # Error if no input data file has been specified.
+ assert input_file is not None, \
+ 'Attempting to load data, but no input file was specified.'
+
+ file_ext = os.path.splitext(input_file)[1]
+
+ # Detect type.
+ if os.path.isdir(input_file):
+ return get_matrix_from_cellranger_mtx
+
+ elif file_ext == '.h5':
+ return get_matrix_from_cellranger_h5
+
+ elif input_file.endswith('.txt.gz') or input_file.endswith('.txt'):
+ return get_matrix_from_dropseq_dge
+
+ elif input_file.endswith('.csv.gz') or input_file.endswith('.csv'):
+ return get_matrix_from_bd_rhapsody
+
+ elif file_ext == '.h5ad':
+ return get_matrix_from_anndata
+
+ elif file_ext == '.loom':
+ return get_matrix_from_loom
+
+ elif file_ext == '.npz':
+ return get_matrix_from_npz
+
+ else:
+ raise ValueError('Failed to determine input file type for '
+ + input_file + '\n'
+ + 'This must either be: a directory that contains '
+ 'CellRanger-format MTX outputs; a single CellRanger '
+ '".h5" file; a DropSeq-format DGE ".txt.gz" file; '
+ 'a BD-Rhapsody-format ".csv" file; a ".h5ad" file '
+ 'produced by anndata (include all barcodes); a '
+ '".loom" file (include all barcodes); or a ".npz" '
+ 'sparse matrix file')
+
+
+def detect_cellranger_version_mtx(filedir: str) -> int:
+ """Detect which version of CellRanger (2 or 3) created this mtx directory.
+
+ Args:
+ filedir: string path to .mtx file that contains the raw gene
+ barcode matrix in a sparse coo text format.
+
+ Returns:
+ CellRanger version, either 2 or 3, as an integer.
+
+ """
+
+ assert os.path.isdir(filedir), f"The directory {filedir} is not accessible."
+
+ if os.path.isfile(os.path.join(filedir, 'features.tsv.gz')):
+ return 3
+
+ else:
+ return 2
+
+
+def detect_cellranger_version_h5(filename: str) -> int:
+ """Detect which version of CellRanger (2 or 3) created this h5 file.
+
+ Args:
+ filename: string path to .mtx file that contains the raw gene
+ barcode matrix in a sparse coo text format.
+
+ Returns:
+ version: CellRanger version, either 2 or 3, as an integer.
+
+ """
+
+ with tables.open_file(filename, 'r') as f:
+
+ # For CellRanger v2, each group in the table (other than root)
+ # contains a genome.
+ # For CellRanger v3, there is a 'matrix' group that contains 'features'.
+
+ version = 2
+
+ try:
+
+ # This works for version 3 but not for version 2.
+ getattr(f.root.matrix, 'features')
+ version = 3
+
+ except tables.NoSuchNodeError:
+ pass
+
+ return version
+
+
+def get_matrix_from_cellranger_mtx(filedir: str) \
+ -> Dict[str, Union[sp.csr_matrix, List[np.ndarray], np.ndarray]]:
+ """Load a count matrix from an mtx directory from CellRanger's output.
+
+ For CellRanger v2:
+ The directory must contain three files:
+ matrix.mtx
+ barcodes.tsv
+ genes.tsv
+
+ For CellRanger v3:
+ The directory must contain three files:
+ matrix.mtx.gz
+ barcodes.tsv.gz
+ features.tsv.gz
+
+ This function returns a dictionary that includes the count matrix, the gene
+ names (which correspond to columns of the count matrix), and the barcodes
+ (which correspond to rows of the count matrix).
+
+ Args:
+ filedir: string path to .mtx file that contains the raw gene
+ barcode matrix in a sparse coo text format.
+
+ Returns:
+ out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
+ barcodes as rows and genes as columns
+ out['barcodes']: numpy array of strings which are the nucleotide
+ sequences of the barcodes that correspond to the rows in
+ the out['matrix']
+ out['gene_names']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string names of genes in the genome, which
+ correspond to the columns in the out['matrix'].
+ out['gene_ids']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string Ensembl ID of genes in the genome, which
+ also correspond to the columns in the out['matrix'].
+ out['feature_types']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string feature types of genes (or possibly
+ antibody capture reads), which also correspond to the columns
+ in the out['matrix'].
+
+ """
+
+ assert os.path.isdir(filedir), "The directory {filedir} is not accessible."
+
+ # Decide whether data is CellRanger v2 or v3.
+ cellranger_version = detect_cellranger_version_mtx(filedir=filedir)
+ logger.info(f"CellRanger v{cellranger_version} format")
+
+ # CellRanger version 3
+ if cellranger_version == 3:
+
+ matrix_file = os.path.join(filedir, 'matrix.mtx.gz')
+ gene_file = os.path.join(filedir, 'features.tsv.gz')
+ barcode_file = os.path.join(filedir, 'barcodes.tsv.gz')
+
+ # Read in feature names.
+ features = np.genfromtxt(fname=gene_file,
+ delimiter="\t",
+ skip_header=0,
+ dtype=str)
+
+ # Read in gene expression and feature data.
+ gene_ids = features[:, 0].squeeze() # first column
+ gene_names = features[:, 1].squeeze() # second column
+ feature_types = features[:, 2].squeeze() # third column
+
+ # CellRanger version 2
+ elif cellranger_version == 2:
+
+ # Read in the count matrix using scipy.
+ matrix_file = os.path.join(filedir, 'matrix.mtx')
+ gene_file = os.path.join(filedir, 'genes.tsv')
+ barcode_file = os.path.join(filedir, 'barcodes.tsv')
+
+ # Read in gene names.
+ gene_data = np.genfromtxt(fname=gene_file,
+ delimiter="\t",
+ skip_header=0,
+ dtype=str)
+ if len(gene_data.shape) == 1: # custom file format with just gene names
+ gene_names = gene_data.squeeze()
+ gene_ids = None
+ else: # the 10x CellRanger v2 format with two columns
+ gene_names = gene_data[:, 1].squeeze() # second column
+ gene_ids = gene_data[:, 0].squeeze() # first column
+ feature_types = None
+
+ else:
+ raise NotImplementedError('MTX format was not identifiable as CellRanger '
+ 'v2 or v3. Please check 10x Genomics formatting.')
+
+ # For both versions:
+
+ # Read in sparse count matrix.
+ count_matrix = io.mmread(matrix_file).tocsr().transpose()
+
+ # Read in barcode names.
+ barcodes = np.genfromtxt(fname=barcode_file,
+ delimiter="\t",
+ skip_header=0,
+ dtype=str)
+
+ # Issue warnings if necessary, based on dimensions matching.
+ if count_matrix.shape[1] != len(gene_names):
+ logger.warning(f"Number of gene names in {filedir}/genes.tsv "
+ f"does not match the number expected from the "
+ f"count matrix.")
+ if count_matrix.shape[0] != len(barcodes):
+ logger.warning(f"Number of barcodes in {filedir}/barcodes.tsv "
+ f"does not match the number expected from the "
+ f"count matrix.")
+
+ return {'matrix': count_matrix,
+ 'gene_names': gene_names,
+ 'feature_types': feature_types,
+ 'gene_ids': gene_ids,
+ 'genomes': None,
+ 'barcodes': barcodes,
+ 'cellranger_version': cellranger_version}
+
+
+def get_matrix_from_cellranger_h5(filename: str) \
+ -> Dict[str, Union[sp.csr_matrix, np.ndarray]]:
+ """Load a count matrix from an h5 file from CellRanger's output.
+
+ The file needs to be a _raw_gene_bc_matrices_h5.h5 file. This function
+ returns a dictionary that includes the count matrix, the gene names (which
+ correspond to columns of the count matrix), and the barcodes (which
+ correspond to rows of the count matrix).
+
+ This function works for CellRanger v2 and v3 HDF5 formats.
+
+ Args:
+ filename: string path to .h5 file that contains the raw gene
+ barcode matrices
+
+ Returns:
+ out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
+ barcodes as rows and genes as columns
+ out['barcodes']: numpy array of strings which are the nucleotide
+ sequences of the barcodes that correspond to the rows in
+ the out['matrix']
+ out['gene_names']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string names of genes in the genome, which
+ correspond to the columns in the out['matrix'].
+ out['gene_ids']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string Ensembl ID of genes in the genome, which
+ also correspond to the columns in the out['matrix'].
+ out['feature_types']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string feature types of genes (or possibly
+ antibody capture reads), which also correspond to the columns
+ in the out['matrix'].
+
+ """
+
+ # Detect CellRanger version.
+ cellranger_version = detect_cellranger_version_h5(filename=filename)
+ logger.info(f"CellRanger v{cellranger_version} format")
+
+ with tables.open_file(filename, 'r') as f:
+ # Initialize empty lists.
+ csc_list = []
+ barcodes = None
+ feature_ids = None
+ feature_types = None
+ genomes = None
+
+ # CellRanger v2:
+ # Each group in the table (other than root) contains a genome,
+ # so walk through the groups to get data for each genome.
+ if cellranger_version == 2:
+
+ feature_names = []
+ feature_ids = []
+ genomes = []
+
+ for group in f.walk_groups():
+ try:
+ # Read in data for this genome, and put it into a
+ # scipy.sparse.csc.csc_matrix
+ barcodes = getattr(group, 'barcodes').read()
+ data = getattr(group, 'data').read()
+ indices = getattr(group, 'indices').read()
+ indptr = getattr(group, 'indptr').read()
+ shape = getattr(group, 'shape').read()
+ csc_list.append(sp.csc_matrix((data, indices, indptr),
+ shape=shape))
+ fnames_this_genome = getattr(group, 'gene_names').read()
+ feature_names.extend(fnames_this_genome)
+ feature_ids.extend(getattr(group, 'genes').read())
+ genomes.extend([group._g_gettitle()] * fnames_this_genome.size)
+
+ except tables.NoSuchNodeError:
+ # This exists to bypass the root node, which has no data.
+ pass
+
+ # Create numpy arrays.
+ feature_names = np.array(feature_names, dtype=str)
+ genomes = np.array(genomes, dtype=str)
+ if len(feature_ids) > 0:
+ feature_ids = np.array(feature_ids)
+ else:
+ feature_ids = None
+
+ # CellRanger v3:
+ # There is only the 'matrix' group.
+ elif cellranger_version == 3:
+
+ # Read in data for this genome, and put it into a
+ # scipy.sparse.csc.csc_matrix
+ barcodes = getattr(f.root.matrix, 'barcodes').read()
+ data = getattr(f.root.matrix, 'data').read()
+ indices = getattr(f.root.matrix, 'indices').read()
+ indptr = getattr(f.root.matrix, 'indptr').read()
+ shape = getattr(f.root.matrix, 'shape').read()
+ csc_list.append(sp.csc_matrix((data, indices, indptr),
+ shape=shape))
+
+ # Read in 'feature' information
+ feature_group = f.get_node(f.root.matrix, 'features')
+ feature_names = getattr(feature_group, 'name').read()
+
+ try:
+ feature_types = getattr(feature_group, 'feature_type').read()
+ except tables.NoSuchNodeError:
+ # This exists in case someone produced a file without feature_type.
+ pass
+ try:
+ feature_ids = getattr(feature_group, 'id').read()
+ except tables.NoSuchNodeError:
+ # This exists in case someone produced a file without feature id.
+ pass
+ try:
+ genomes = getattr(feature_group, 'genome').read()
+ except tables.NoSuchNodeError:
+ # This exists in case someone produced a file without feature genome.
+ pass
+
+ # Put the data together (possibly from several genomes for v2 datasets).
+ count_matrix = sp.vstack(csc_list, format='csc')
+ count_matrix = count_matrix.transpose().tocsr()
+
+ # Issue warnings if necessary, based on dimensions matching.
+ if count_matrix.shape[1] != feature_names.size:
+ logger.warning(f"Number of gene names ({feature_names.size}) in {filename} "
+ f"does not match the number expected from the count "
+ f"matrix ({count_matrix.shape[1]}).")
+ if count_matrix.shape[0] != barcodes.size:
+ logger.warning(f"Number of barcodes ({barcodes.size}) in {filename} "
+ f"does not match the number expected from the count "
+ f"matrix ({count_matrix.shape[0]}).")
+
+ return {'matrix': count_matrix,
+ 'gene_names': feature_names,
+ 'gene_ids': feature_ids,
+ 'genomes': genomes,
+ 'feature_types': feature_types,
+ 'barcodes': barcodes,
+ 'cellranger_version': cellranger_version}
+
+
+def get_matrix_from_dropseq_dge(filename: str) \
+ -> Dict[str, Union[sp.csr_matrix, np.ndarray]]:
+ """Load a count matrix from a DropSeq DGE matrix file.
+
+ The file needs to be a gzipped text file in DGE format. This function
+ returns a dictionary that includes the count matrix, the gene names (which
+ correspond to columns of the count matrix), and the barcodes (which
+ correspond to rows of the count matrix). Reads in the file line by line
+ instead of trying to read in an entire dense matrix at once, which might
+ require quite a bit of memory.
+
+ Args:
+ filename: string path to .txt.gz file that contains the raw gene
+ barcode matrix
+
+ Returns:
+ out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
+ barcodes as rows and genes as columns
+ out['barcodes']: numpy array of strings which are the nucleotide
+ sequences of the barcodes that correspond to the rows in
+ the out['matrix']
+ out['gene_names']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string names of genes in the genome, which
+ correspond to the columns in the out['matrix'].
+ out['gene_ids']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string Ensembl ID of genes in the genome, which
+ also correspond to the columns in the out['matrix'].
+
+ """
+
+ logger.info(f"DropSeq DGE format")
+
+ load_fcn = gzip.open if filename.endswith('.gz') else open
+
+ with load_fcn(filename, 'rt') as f:
+
+ # Skip the comment '#' lines in header
+ for header in f:
+ if header[0] == '#':
+ continue
+ else:
+ break
+
+ # Read in first row with droplet barcodes
+ barcodes = header.split('\n')[0].split('\t')[1:]
+
+ # Gene names are first entry per row
+ gene_names = []
+
+ # Arrays used to construct a sparse matrix
+ row = []
+ col = []
+ data = []
+
+ # Read in rest of file row by row
+ for i, line in enumerate(f):
+ # Parse row into gene name and count data
+ parsed_line = line.split('\n')[0].split('\t')
+ gene_names.append(parsed_line[0])
+ counts = np.array(parsed_line[1:], dtype=int)
+
+ # Create sparse version of data and add to arrays
+ nonzero_col_inds = np.nonzero(counts)[0]
+ row.extend([i] * nonzero_col_inds.size)
+ col.extend(nonzero_col_inds)
+ data.extend(counts[nonzero_col_inds])
+
+ count_matrix = sp.csc_matrix((data, (row, col)),
+ shape=(len(gene_names), len(barcodes)),
+ dtype=float).transpose()
+
+ return {'matrix': count_matrix,
+ 'gene_names': np.array(gene_names),
+ 'gene_ids': None,
+ 'genomes': None,
+ 'feature_types': None,
+ 'barcodes': np.array(barcodes)}
+
+
+def get_matrix_from_bd_rhapsody(filename: str) \
+ -> Dict[str, Union[sp.csr_matrix, np.ndarray]]:
+ """Load a count matrix from a BD Rhapsody MolsPerCell.csv file.
+
+ The file needs to be in MolsPerCell_Unfiltered format, which is comma
+ separated, where rows are barcodes and columns are genes. Can be gzipped
+ or not. This function returns a dictionary that includes the count matrix,
+ the gene names (which correspond to columns of the count matrix), and the
+ barcodes (which correspond to rows of the count matrix). Reads in the file
+ line by line instead of trying to read in an entire dense matrix at once,
+ which might require quite a bit of memory.
+
+ Args:
+ filename: string path to .csv file that contains the raw gene
+ barcode matrix MolsPerCell_Unfiltered.csv
+
+ Returns:
+ out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
+ barcodes as rows and genes as columns
+ out['barcodes']: numpy array of strings which are the nucleotide
+ sequences of the barcodes that correspond to the rows in
+ the out['matrix']
+ out['gene_names']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string names of genes in the genome, which
+ correspond to the columns in the out['matrix'].
+ out['gene_ids']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string Ensembl ID of genes in the genome, which
+ also correspond to the columns in the out['matrix'].
+
+ """
+
+ logger.info(f"BD Rhapsody MolsPerCell_Unfiltered.csv format")
+
+ load_fcn = gzip.open if filename.endswith('.gz') else open
+
+ with load_fcn(filename, 'rt') as f:
+
+ # Skip the comment '#' lines in header
+ for header in f:
+ if header[0] == '#':
+ continue
+ else:
+ break
+
+ # Read in first row with gene names
+ gene_names = header.split('\n')[0].split(',')[1:]
+
+ # Barcode names are first entry per row
+ barcodes = []
+
+ # Arrays used to construct a sparse matrix
+ row = []
+ col = []
+ data = []
+
+ # Read in rest of file row by row
+ for i, line in enumerate(f):
+ # Parse row into gene name and count data
+ parsed_line = line.split('\n')[0].split(',')
+ barcodes.append(parsed_line[0])
+ counts = np.array(parsed_line[1:], dtype=np.int)
+
+ # Create sparse version of data and add to arrays
+ nonzero_col_inds = np.nonzero(counts)[0]
+ row.extend([i] * nonzero_col_inds.size)
+ col.extend(nonzero_col_inds)
+ data.extend(counts[nonzero_col_inds])
+
+ count_matrix = sp.csc_matrix((data, (row, col)),
+ shape=(len(barcodes), len(gene_names)),
+ dtype=np.float)
+
+ return {'matrix': count_matrix,
+ 'gene_names': np.array(gene_names),
+ 'gene_ids': None,
+ 'genomes': None,
+ 'feature_types': None,
+ 'barcodes': np.array(barcodes)}
+
+
+def get_matrix_from_npz(filename: str) \
+ -> Dict[str, Union[sp.csr_matrix, np.ndarray]]:
+ """Load a count matrix from a sparse NPZ file, accompanied by barcode and
+ gene NPY files.
+ NOTE: This format is one output of the Optimus pipeline. It loads much
+ faster than a Loom file. The NPZ file requires two accompanying files:
+ 'col_index.npy' and 'row_index.npy', named exactly as shown, and in the
+ same directory as the NPZ file.
+ Args:
+ filename: string path to .h5ad file that contains the raw gene
+ barcode matrices
+ Returns:
+ out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
+ barcodes as rows and genes as columns
+ out['barcodes']: numpy array of strings which are the nucleotide
+ sequences of the barcodes that correspond to the rows in
+ the out['matrix']
+ out['gene_names']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string names of genes in the genome, which
+ correspond to the columns in the out['matrix'].
+ out['gene_ids']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string Ensembl ID of genes in the genome, which
+ also correspond to the columns in the out['matrix'].
+ out['feature_types']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string feature types of genes (or possibly
+ antibody capture reads), which also correspond to the columns
+ in the out['matrix'].
+ """
+ logger.info(f"Optimus sparse NPZ format")
+ try:
+ count_matrix = sp.load_npz(file=filename)
+ file_dir, _ = os.path.split(filename)
+ gene_ids = np.load(os.path.join(file_dir, 'col_index.npy'))
+ barcodes = np.load(os.path.join(file_dir, 'row_index.npy'))
+ except IOError as e:
+ logger.error('Loading an NPZ file requires two additional files in the '
+ f'same directory ({file_dir}): '
+ 'one called "col_index.npy" that contains genes, and one '
+ 'called "row_index.npy" that contains barcodes.')
+ logger.error(traceback.format_exc())
+ raise e
+ return {'matrix': count_matrix,
+ 'gene_names': gene_ids, # that's all we have access to, so we'll use it
+ 'gene_ids': gene_ids,
+ 'genomes': None,
+ 'feature_types': None,
+ 'barcodes': barcodes}
+
+
+def get_matrix_from_anndata(filename: str) \
+ -> Dict[str, Union[sp.csr_matrix, np.ndarray]]:
+ """Load a count matrix from an h5ad AnnData file.
+ The file needs to contain raw counts for all measured barcodes in the
+ `.X` attribute or a `.layer[{'counts', 'spliced'}]` attribute. This function
+ returns a dictionary that includes the count matrix, the gene names (which
+ correspond to columns of the count matrix), and the barcodes (which
+ correspond to rows of the count matrix).
+ This function works for any AnnData object meeting the above requirements,
+ as generated by alignment methods like `kallisto | bustools`.
+ Args:
+ filename: string path to .h5ad file that contains the raw gene
+ barcode matrices
+ Returns:
+ out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
+ barcodes as rows and genes as columns
+ out['barcodes']: numpy array of strings which are the nucleotide
+ sequences of the barcodes that correspond to the rows in
+ the out['matrix']
+ out['gene_names']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string names of genes in the genome, which
+ correspond to the columns in the out['matrix'].
+ out['gene_ids']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string Ensembl ID of genes in the genome, which
+ also correspond to the columns in the out['matrix'].
+ out['feature_types']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string feature types of genes (or possibly
+ antibody capture reads), which also correspond to the columns
+ in the out['matrix'].
+ """
+ logger.info(f"AnnData format")
+ try:
+ adata = anndata.read_h5ad(filename)
+ except anndata._io.utils.AnnDataReadError as e:
+ logger.error(f'A call to anndata.read_h5ad() with anndata {anndata.__version__} '
+ f'threw AnnDataReadError: ')
+ logger.error(traceback.format_exc())
+ raise e
+ return _dict_from_anndata(adata)
+
+
+def get_matrix_from_loom(filename: str) \
+ -> Dict[str, Union[sp.csr_matrix, np.ndarray]]:
+ """Load a count matrix from a loom file.
+ The file needs to contain raw counts for all measured barcodes in the
+ layer '', as in
+ https://broadinstitute.github.io/warp/docs/Pipelines/Optimus_Pipeline/Loom_schema/
+ Returns a dictionary that includes the count matrix, the gene names (which
+ correspond to columns of the count matrix), and the barcodes (which
+ correspond to rows of the count matrix).
+
+ Args:
+ filename: string path to .h5ad file that contains the raw gene
+ barcode matrices
+
+ Returns:
+ out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
+ barcodes as rows and genes as columns
+ out['barcodes']: numpy array of strings which are the nucleotide
+ sequences of the barcodes that correspond to the rows in
+ the out['matrix']
+ out['gene_names']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string names of genes in the genome, which
+ correspond to the columns in the out['matrix'].
+ out['gene_ids']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string Ensembl ID of genes in the genome, which
+ also correspond to the columns in the out['matrix'].
+ out['feature_types']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string feature types of genes (or possibly
+ antibody capture reads), which also correspond to the columns
+ in the out['matrix'].
+ """
+ logger.info(f"Loom format, expecting Optimus pipeline conventions")
+ try:
+ adata = anndata.read_loom(filename, sparse=True, X_name='')
+ except anndata._io.utils.AnnDataReadError as e:
+ logger.error(f'A call to anndata.read_loom() with anndata {anndata.__version__} '
+ f'threw AnnDataReadError: ')
+ logger.error(traceback.format_exc())
+ raise e
+ return _dict_from_anndata(adata)
+
+
+def _dict_from_anndata(adata: anndata.AnnData) -> Dict[str, Union[sp.csr_matrix, np.ndarray]]:
+ """Extract relevant information from AnnData and format it as a dict
+
+ Args:
+ adata: AnnData object
+
+ Returns:
+ out['matrix']: scipy.sparse.csr.csr_matrix of unique UMI counts, with
+ barcodes as rows and genes as columns
+ out['barcodes']: numpy array of strings which are the nucleotide
+ sequences of the barcodes that correspond to the rows in
+ the out['matrix']
+ out['gene_names']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string names of genes in the genome, which
+ correspond to the columns in the out['matrix'].
+ out['gene_ids']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string Ensembl ID of genes in the genome, which
+ also correspond to the columns in the out['matrix'].
+ out['feature_types']: List of numpy arrays, where the number of elements
+ in the list is the number of genomes in the dataset. Each numpy
+ array contains the string feature types of genes (or possibly
+ antibody capture reads), which also correspond to the columns
+ in the out['matrix']."""
+
+ if "counts" in adata.layers.keys():
+ # this is a common manual setting for users of scVI
+ # given the manual convention, we prefer this matrix to
+ # .X since it is less likely to represent something other
+ # than counts
+ logger.info("Found `.layers['counts']`. Using for count data.")
+ count_matrix = adata.layers["counts"]
+ elif "spliced" in adata.layers.keys() and adata.X is None:
+ # alignment using kallisto | bustools with intronic counts
+ # does not populate `.X` by default, but does populate
+ # `.layers['spliced'], .layers['unspliced']`.
+ # we use spliced counts for analysis
+ logger.info("Found `.layers['spliced']`. Using for count data.")
+ count_matrix = adata.layers["spliced"]
+ else:
+ logger.info("Using `.X` for count data.")
+ count_matrix = adata.X
+
+ # check that `count_matrix` contains a large number of barcodes,
+ # consistent with a raw single cell experiment
+ if count_matrix.shape[0] < consts.MINIMUM_BARCODES_H5AD:
+ # this experiment might be prefiltered
+ logger.warning(f"Only {count_matrix.shape[0]} barcodes were found.\n"
+ "This suggests the matrix was prefiltered.\n"
+ "CellBender requires a raw, unfiltered [Barcodes, Genes] matrix.")
+
+ # AnnData is [Cells, Genes], no need to transpose
+ # we typecast explicitly in the off chance `count_matrix` was dense.
+ count_matrix = sp.csr_matrix(count_matrix)
+ # feature names and ids are not consistently delineated in AnnData objects
+ # so we attempt to find relevant features using common values.
+ feature_names = np.array(adata.var_names, dtype=str)
+ barcodes = np.array(adata.obs_names, dtype=str)
+
+ # Make an attempt to find feature_IDs if they are present.
+ feature_ids = None
+ for key in ['gene_id', 'gene_ids', 'ensembl_ids']:
+ if key in adata.var.keys():
+ feature_ids = np.array(adata.var[key].values, dtype=str)
+
+ # Make an attempt to find feature_types if they are present.
+ feature_types = None
+ for key in ['feature_type', 'feature_types']:
+ if key in adata.var.keys():
+ feature_types = np.array(adata.var[key].values, dtype=str)
+
+ # Make an attempt to find genomes if they are present.
+ genomes = None
+ for key in ['genome', 'genomes']:
+ if key in adata.var.keys():
+ genomes = np.array(adata.var[key].values, dtype=str)
+
+ # Issue warnings if necessary, based on dimensions matching.
+ if count_matrix.shape[1] != feature_names.size:
+ logger.warning(f"Number of gene names ({feature_names.size}) "
+ f"does not match the number expected from the count "
+ f"matrix ({count_matrix.shape[1]}).")
+ if count_matrix.shape[0] != barcodes.size:
+ logger.warning(f"Number of barcodes ({barcodes.size}) "
+ f"does not match the number expected from the count "
+ f"matrix ({count_matrix.shape[0]}).")
+
+ return {'matrix': count_matrix,
+ 'gene_names': feature_names,
+ 'gene_ids': feature_ids,
+ 'genomes': genomes,
+ 'feature_types': feature_types,
+ 'barcodes': barcodes}
diff --git a/cellbender/remove_background/data/priors.py b/cellbender/remove_background/data/priors.py
new file mode 100644
index 00000000..39897691
--- /dev/null
+++ b/cellbender/remove_background/data/priors.py
@@ -0,0 +1,391 @@
+"""Functionality for estimating various priors from the data"""
+
+import numpy as np
+import torch
+from scipy.stats import gaussian_kde
+
+from cellbender.remove_background import consts
+
+from typing import Dict, Tuple, Union
+import logging
+
+
+logger = logging.getLogger('cellbender')
+
+
+def _threshold_otsu(umi_counts: np.ndarray, n_bins: int = 256) -> float:
+ """Return threshold value based on fast implementation of Otsu's method.
+
+ From skimage, with slight modifications:
+ https://github.com/scikit-image/scikit-image/blob/
+ a4e533ea2a1947f13b88219e5f2c5931ab092413/skimage/filters/thresholding.py#L312
+
+ Args:
+ umi_counts: Array of UMI counts
+ n_bins: Number of bins used to calculate histogram
+
+ Returns:
+ threshold: Upper threshold value. All droplets with UMI counts greater
+ than this value are assumed to contain cells.
+
+ References
+ ----------
+ .. [1] Wikipedia, https://en.wikipedia.org/wiki/Otsu's_Method
+ .. [2] https://scikit-image.org/docs/stable/auto_examples/applications/plot_thresholding.html
+
+ Notes
+ -----
+ The input image must be grayscale.
+ """
+
+ # create a UMI count histogram
+ counts, bin_centers = _create_histogram(umi_counts=umi_counts, n_bins=n_bins)
+
+ # class probabilities for all possible thresholds
+ weight1 = np.cumsum(counts)
+ weight2 = np.cumsum(counts[::-1])[::-1]
+
+ # class means for all possible thresholds
+ mean1 = np.cumsum(counts * bin_centers) / weight1
+ mean2 = (np.cumsum((counts * bin_centers)[::-1]) / weight2[::-1])[::-1]
+
+ # Clip ends to align class 1 and class 2 variables:
+ # The last value of ``weight1``/``mean1`` should pair with zero values in
+ # ``weight2``/``mean2``, which do not exist.
+ variance12 = weight1[:-1] * weight2[1:] * (mean1[:-1] - mean2[1:]) ** 2
+
+ idx = np.argmax(variance12)
+ threshold = bin_centers[idx]
+
+ return threshold
+
+
+def _create_histogram(umi_counts: np.ndarray, n_bins: int) -> Tuple[np.ndarray, np.ndarray]:
+ """Return a histogram.
+
+ Args:
+ umi_counts: Array of UMI counts
+ n_bins: Number of bins used to calculate histogram
+
+ Returns:
+ counts: Each element is the number of droplets falling in each UMI
+ count bin
+ bin_centers: Each element is the value corresponding to the center of
+ each UMI count bin
+ """
+ counts, bin_edges = np.histogram(umi_counts.reshape(-1), n_bins)
+ bin_centers = (bin_edges[:-1] + bin_edges[1:]) / 2
+ return counts.astype('float32', copy=False), bin_centers
+
+
+def _peak_density_given_cutoff(umi_counts: np.ndarray,
+ cutoff: float,
+ cell_count_low_limit: float) -> Tuple[float, float]:
+ """Run scipy.stats gaussian_kde on part of the UMI curve"""
+
+ # get the UMI count values we are including
+ noncell_counts = umi_counts[umi_counts <= cutoff]
+
+ # resample them: the magic of looking at a log log plot
+ n_putative_cells = (umi_counts > cell_count_low_limit).sum()
+ n_putative_empties = len(noncell_counts)
+ inds = np.logspace(np.log10(n_putative_cells),
+ np.log10(n_putative_cells + n_putative_empties),
+ num=1000,
+ base=10)
+ inds = [max(0, min(int(ind - n_putative_cells), len(noncell_counts) - 1)) for ind in inds]
+
+ noncell_counts = np.sort(noncell_counts)[::-1][inds]
+
+ # find the peak density: that is the empty count prior
+
+ # calculate range of data, rounding out to make sure we cover everything
+ log_noncell_counts = np.log(noncell_counts)
+ x = np.arange(
+ np.floor(log_noncell_counts.min()) - 0.01,
+ np.ceil(log_noncell_counts.max()) + 0.01,
+ 0.1
+ )
+
+ # fit a KDE to estimate density
+ k = gaussian_kde(log_noncell_counts)
+ density = k.evaluate(x)
+
+ # the density peak is almost surely the empty droplets
+ log_peak_ind = np.argmax(density)
+ log_peak = x[log_peak_ind]
+ empty_count_prior = np.exp(log_peak)
+
+ # try to go about 1 stdev up from the peak
+ peak_density = np.max(density)
+ one_std_density = 0.6 * peak_density
+ one_std_inds = np.where(density[log_peak_ind:] < one_std_density)[0]
+ if len(one_std_inds) > 0:
+ one_std_ind = one_std_inds[0]
+ else:
+ one_std_ind = len(density[log_peak_ind:]) - 1
+ empty_count_upper_limit = np.exp(x[log_peak_ind:][one_std_ind])
+
+ return empty_count_prior, empty_count_upper_limit
+
+
+def get_cell_count_given_expected_cells(umi_counts: np.ndarray,
+ expected_cells: int) -> Dict[str, float]:
+ """In the case where a prior is passed in as input, use it
+
+ Args:
+ umi_counts: Array of UMI counts per droplet, in no particular order
+ expected_cells: Input by user
+
+ Returns:
+ Dict with keys ['cell_counts']
+ """
+ order = np.argsort(umi_counts)[::-1]
+ cell_counts = np.exp(np.mean(np.log(umi_counts[order][:expected_cells]))).item()
+ return {'cell_counts': cell_counts}
+
+
+def get_empty_count_given_expected_cells_and_total_droplets(
+ umi_counts: np.ndarray,
+ expected_cells: int,
+ total_droplets: int,
+) -> Dict[str, float]:
+ """In the case where a prior is passed in as input, use it
+
+ Args:
+ umi_counts: Array of UMI counts per droplet, in no particular order
+ expected_cells: Input by user, or prior estimate
+ total_droplets: Input by user
+
+ Returns:
+ Dict with keys ['empty_counts', 'empty_count_upper_limit']
+ """
+
+ order = np.argsort(umi_counts)[::-1]
+ starting_point = max(expected_cells, total_droplets - 500)
+ empty_counts = np.median(umi_counts[order]
+ [int(starting_point):int(total_droplets)]).item()
+
+ # need to estimate here
+ cell_counts = np.exp(np.mean(np.log(umi_counts[order][:expected_cells]))).item()
+ middle = np.sqrt(cell_counts * empty_counts)
+ empty_count_upper_limit = min(middle, 1.5 * empty_counts)
+
+ return {'empty_counts': empty_counts,
+ 'empty_count_upper_limit': empty_count_upper_limit}
+
+
+def get_cell_count_empty_count(umi_counts: np.ndarray,
+ low_count_threshold: float = 15) -> Dict[str, float]:
+ """Obtain priors on cell counts and empty droplet counts from a UMI curve
+ using heuristics, and without applying any other prior information.
+
+ Heuristics:
+ 0. Ignore droplets with counts below low_count_threshold
+ 1. Use Otsu's method to threshold the log UMI count data (ignoring droplets
+ past 1/4 of the total droplets above low_count_threshold, as we go down
+ the UMI curve). This is used as a lower limit on cell counts.
+ It seems quite robust.
+ 2. Use the following iterative approach, until converged:
+ a. Establish an upper cutoff on possible empty droplets, using the
+ current estimate of empty counts and our cell count prior (the
+ estimate is 3/4 of the geometric mean of the two).
+ b. Use gaussian_kde from scipy.stats to create a smooth histogram of
+ the log UMI counts, for droplets with counts below the cutoff.
+ - A trick is used to resample the droplets before creating the
+ histogram, so that it looks more like a log-log plot
+ c. Identify the peak density of the histogram as the empty count
+ estimate.
+ - Convergence happens when our estimate of empty counts stops changing.
+
+ Args:
+ umi_counts: Array of UMI counts per droplet, in no particular order
+ low_count_threshold: Ignore droplets with counts below this value
+
+ Returns:
+ Dict with keys ['cell_counts', 'empty_counts']
+ """
+
+ logger.debug('Beginning priors.get_cell_count_empty_count()')
+ reverse_sorted_umi_counts = np.sort(umi_counts)[::-1]
+ umi_counts_for_otsu = reverse_sorted_umi_counts[:(umi_counts > low_count_threshold).sum() // 4]
+
+ log_cell_count_low_limit = _threshold_otsu(np.log(umi_counts_for_otsu))
+ cell_count_low_limit = np.exp(log_cell_count_low_limit)
+
+ logger.debug(f'cell_count_low_limit is {cell_count_low_limit}')
+ cell_count_prior = np.mean(umi_counts[umi_counts > cell_count_low_limit])
+
+ umi_counts_for_kde = reverse_sorted_umi_counts[reverse_sorted_umi_counts > low_count_threshold]
+
+ # initial conditions for the loop
+ # start low, but have a failsafe (especially for simulated data)
+ cutoff = max(0.1 * cell_count_low_limit, umi_counts_for_kde[-100])
+ empty_count_prior = -100
+ empty_count_upper_limit = None
+ delta = np.inf
+ a = 0
+
+ # iterate to convergence, at most 5 times
+ while delta > 10:
+ logger.debug(f'cutoff = {cutoff}')
+
+ # use gaussian_kde to find the peak in the histogram
+ new_empty_count_prior, empty_count_upper_limit = _peak_density_given_cutoff(
+ umi_counts=umi_counts_for_kde,
+ cutoff=cutoff,
+ cell_count_low_limit=cell_count_low_limit,
+ )
+ logger.debug(f'new_empty_count_prior = {new_empty_count_prior}')
+
+ # 3/4 of the geometric mean is our new upper cutoff
+ cutoff = 0.75 * np.sqrt(cell_count_prior * new_empty_count_prior)
+ delta = np.abs(new_empty_count_prior - empty_count_prior)
+ logger.debug(f'delta = {delta}')
+ empty_count_prior = new_empty_count_prior
+ a += 1
+ if a >= 5:
+ logger.debug('Heuristics for determining empty counts exceeded 5 '
+ 'iterations without converging')
+ break
+
+ # do a final estimation of cell counts:
+ # go to the halfway point and then take the median of the droplets above
+ count_crossover = np.sqrt(cell_count_prior * empty_count_prior)
+ cell_count_prior = np.median(umi_counts[umi_counts > count_crossover])
+
+ logger.debug(f'cell_count_prior is {cell_count_prior}')
+ logger.debug(f'empty_count_prior is {empty_count_prior}')
+ logger.debug('End of priors.get_cell_count_empty_count()')
+
+ return {'cell_counts': cell_count_prior,
+ 'empty_counts': empty_count_prior,
+ 'empty_count_upper_limit': empty_count_upper_limit}
+
+
+def get_expected_cells_and_total_droplets(umi_counts: np.ndarray,
+ cell_counts: float,
+ empty_counts: float,
+ empty_count_upper_limit: float,
+ max_empties: int = consts.MAX_EMPTIES_TO_INCLUDE) \
+ -> Dict[str, int]:
+ """Obtain priors on cell counts and empty droplet counts from a UMI curve
+ using heuristics, and without applying any other prior information.
+
+ NOTE: to be run using inputs from get_cell_count_empty_count()
+
+ Args:
+ umi_counts: Array of UMI counts per droplet, in no particular order
+ cell_counts: Prior from get_cell_count_empty_count()
+ empty_counts: Prior from get_cell_count_empty_count()
+ empty_count_upper_limit: Prior from get_cell_count_empty_count()
+ max_empties: Do not include more putative empty droplets than this
+
+ Returns:
+ Dict with keys ['expected_cells', 'total_droplets', 'transition_point']
+
+ Example:
+ >>> priors = get_cell_count_empty_count(umi_counts)
+ >>> priors.update(get_expected_cells_and_total_droplets(umi_counts, **priors))
+ """
+ # expected cells does well when you give it a very conservative estimate
+ expected_cells = (umi_counts >= cell_counts).sum()
+
+ # total droplets will be between empty_count_prior and its upper limit
+ total_droplets_count_value = np.sqrt(empty_counts * empty_count_upper_limit)
+ total_droplets = (umi_counts >= total_droplets_count_value).sum()
+
+ # find the transition point
+ count_crossover = np.sqrt(cell_counts * empty_counts)
+ transition_point = (umi_counts >= count_crossover).sum()
+
+ logger.debug(f'In get_expected_cells_and_total_droplets(), found transition '
+ f'point at droplet {transition_point}')
+
+ # ensure out heuristics don't go too far out datasets with many cells
+ total_droplets = min(total_droplets, transition_point + max_empties)
+
+ return {'expected_cells': expected_cells,
+ 'total_droplets': total_droplets,
+ 'transition_point': transition_point}
+
+
+def get_priors(umi_counts: np.ndarray,
+ low_count_threshold: float,
+ max_total_droplets: int = consts.MAX_TOTAL_DROPLETS_GUESSED) \
+ -> Dict[str, Union[int, float]]:
+ """Get all priors using get_cell_count_empty_count() and
+ get_expected_cells_and_total_droplets(), employing a failsafe if
+ total_droplets is improbably large.
+
+ Args:
+ umi_counts: Array of UMI counts per droplet, in no particular order
+ low_count_threshold: Ignore droplets with counts below this value
+ max_total_droplets: If the initial heuristics come up with a
+ total_droplets value greater than this, we re-run the heuristics
+ with higher low_count_threshold
+
+ Returns:
+ Dict with keys ['cell_counts', 'empty_counts',
+ 'empty_count_upper_limit', 'surely_empty_counts',
+ 'expected_cells', 'total_droplets', 'log_counts_crossover']
+ """
+
+ logger.debug("Computing priors from the UMI curve")
+ priors = get_cell_count_empty_count(
+ umi_counts=umi_counts,
+ low_count_threshold=low_count_threshold,
+ )
+ priors.update(get_expected_cells_and_total_droplets(umi_counts=umi_counts, **priors))
+ logger.debug(f'Automatically computed priors: {priors}')
+
+ a = 0
+ while priors['total_droplets'] > max_total_droplets:
+ logger.debug(f'Heuristics for estimating priors resulted in '
+ f'{priors["total_droplets"]} total_droplets, which is '
+ f'typically too large. Recomputing with '
+ f'low_count_threshold = {priors["empty_count_upper_limit"]:.0f}')
+ priors = get_cell_count_empty_count(
+ umi_counts=umi_counts,
+ low_count_threshold=priors['empty_count_upper_limit'],
+ )
+ priors.update(get_expected_cells_and_total_droplets(umi_counts=umi_counts, **priors))
+ logger.debug(f'Automatically computed priors: {priors}')
+ a += 1
+ if a > 5:
+ break
+
+ # compute a few last things
+ compute_crossover_surely_empty_and_stds(umi_counts=umi_counts, priors=priors)
+
+ return priors
+
+
+def compute_crossover_surely_empty_and_stds(umi_counts, priors):
+ """Given cell_counts and total_droplets, compute a few more quantities
+
+ Args:
+ umi_counts: Array of UMI counts per droplet, in no particular order
+ priors: Dict of priors
+
+ Returns:
+ None. Modifies priors dict in place.
+ """
+
+ assert 'total_droplets' in priors.keys(), \
+ 'Need total_droplets in priors to run compute_crossover_surely_empty_and_stds()'
+ assert 'cell_counts' in priors.keys(), \
+ 'Need cell_counts in priors to run compute_crossover_surely_empty_and_stds()'
+
+ # Compute a crossover point in log count space.
+ reverse_sorted_counts = np.sort(umi_counts)[::-1]
+ surely_empty_counts = reverse_sorted_counts[priors['total_droplets']]
+ log_counts_crossover = (np.log(surely_empty_counts) + np.log(priors['cell_counts'])) / 2
+ priors.update({'log_counts_crossover': log_counts_crossover,
+ 'surely_empty_counts': surely_empty_counts})
+
+ # Compute several other priors.
+ log_nonzero_umi_counts = np.log(umi_counts[umi_counts > 0])
+ d_std = np.std(log_nonzero_umi_counts[log_nonzero_umi_counts > log_counts_crossover]).item() / 5.
+ d_empty_std = 0.01 # this is basically turned off in favor of epsilon
+ priors.update({'d_std': d_std, 'd_empty_std': d_empty_std})
diff --git a/cellbender/remove_background/distributions/NegativeBinomialPoissonConvApprox.py b/cellbender/remove_background/distributions/NegativeBinomialPoissonConvApprox.py
index 1f08481c..1420df3c 100644
--- a/cellbender/remove_background/distributions/NegativeBinomialPoissonConvApprox.py
+++ b/cellbender/remove_background/distributions/NegativeBinomialPoissonConvApprox.py
@@ -118,7 +118,7 @@ def log_prob(self, value):
value=value)
# Use a poisson for small mu, where the above approximation is bad.
- empty_indices = (mu < 1e-5)
+ empty_indices = (mu < 1e-5 * lam)
poisson_log_prob = self._poisson_log_prob(lam=lam[empty_indices],
value=value[empty_indices])
diff --git a/cellbender/remove_background/downstream.py b/cellbender/remove_background/downstream.py
new file mode 100644
index 00000000..7f669379
--- /dev/null
+++ b/cellbender/remove_background/downstream.py
@@ -0,0 +1,479 @@
+"""Functions for downstream work with outputs of remove-background."""
+
+from cellbender.remove_background.data.io import load_data
+
+import tables
+import numpy as np
+import scipy.sparse as sp
+import anndata
+from typing import Dict, Optional
+
+
+def dict_from_h5(file: str) -> Dict[str, np.ndarray]:
+ """Read in everything from an h5 file and put into a dictionary.
+
+ Args:
+ file: The h5 file
+
+ Returns:
+ Dictionary containing all the information from the h5 file
+ """
+ d = {}
+ with tables.open_file(file) as f:
+ # read in everything
+ for array in f.walk_nodes("/", "Array"):
+ d[array.name] = array.read()
+ return d
+
+
+def anndata_from_h5(file: str,
+ analyzed_barcodes_only: bool = True) -> anndata.AnnData:
+ """Load an output h5 file into an AnnData object for downstream work.
+
+ Args:
+ file: The h5 file
+ analyzed_barcodes_only: False to load all barcodes, so that the size of
+ the AnnData object will match the size of the input raw count matrix.
+ True to load a limited set of barcodes: only those analyzed by the
+ algorithm. This allows relevant latent variables to be loaded
+ properly into adata.obs and adata.obsm, rather than adata.uns.
+
+ Returns:
+ anndata.AnnData: The anndata object, populated with inferred latent variables
+ and metadata.
+
+ """
+
+ d = dict_from_h5(file)
+ X = sp.csc_matrix((d.pop('data'), d.pop('indices'), d.pop('indptr')),
+ shape=d.pop('shape')).transpose().tocsr()
+
+ # check and see if we have barcode index annotations, and if the file is filtered
+ barcode_key = [k for k in d.keys() if (('barcode' in k) and ('ind' in k))]
+ if len(barcode_key) > 0:
+ max_barcode_ind = d[barcode_key[0]].max()
+ filtered_file = (max_barcode_ind >= X.shape[0])
+ else:
+ filtered_file = True
+
+ if analyzed_barcodes_only:
+ if filtered_file:
+ # filtered file being read, so we don't need to subset
+ print('Assuming we are loading a "filtered" file that contains only cells.')
+ pass
+ elif 'barcode_indices_for_latents' in d.keys():
+ X = X[d['barcode_indices_for_latents'], :]
+ d['barcodes'] = d['barcodes'][d['barcode_indices_for_latents']]
+ elif 'barcodes_analyzed_inds' in d.keys():
+ X = X[d['barcodes_analyzed_inds'], :]
+ d['barcodes'] = d['barcodes'][d['barcodes_analyzed_inds']]
+ else:
+ print('Warning: analyzed_barcodes_only=True, but the key '
+ '"barcodes_analyzed_inds" or "barcode_indices_for_latents" '
+ 'is missing from the h5 file. '
+ 'Will output all barcodes, and proceed as if '
+ 'analyzed_barcodes_only=False')
+
+ # Construct the anndata object.
+ adata = anndata.AnnData(X=X,
+ obs={'barcode': d.pop('barcodes').astype(str)},
+ var={'gene_name': (d.pop('gene_names') if 'gene_names' in d.keys()
+ else d.pop('name')).astype(str)},
+ dtype=X.dtype)
+ adata.obs.set_index('barcode', inplace=True)
+ adata.var.set_index('gene_name', inplace=True)
+
+ # For CellRanger v2 legacy format, "gene_ids" was called "genes"... rename this
+ if 'genes' in d.keys():
+ d['id'] = d.pop('genes')
+
+ # For purely aesthetic purposes, rename "id" to "gene_id"
+ if 'id' in d.keys():
+ d['gene_id'] = d.pop('id')
+
+ # If genomes are empty, try to guess them based on gene_id
+ if 'genome' in d.keys():
+ if np.array([s.decode() == '' for s in d['genome']]).all():
+ if '_' in d['gene_id'][0].decode():
+ print('Genome field blank, so attempting to guess genomes based on gene_id prefixes')
+ d['genome'] = np.array([s.decode().split('_')[0] for s in d['gene_id']], dtype=str)
+
+ # Add other information to the anndata object in the appropriate slot.
+ _fill_adata_slots_automatically(adata, d)
+
+ # Add a special additional field to .var if it exists.
+ if 'features_analyzed_inds' in adata.uns.keys():
+ adata.var['cellbender_analyzed'] = [True if (i in adata.uns['features_analyzed_inds'])
+ else False for i in range(adata.shape[1])]
+ elif 'features_analyzed_inds' in adata.var.keys():
+ adata.var['cellbender_analyzed'] = [True if (i in adata.var['features_analyzed_inds'].values)
+ else False for i in range(adata.shape[1])]
+
+ if analyzed_barcodes_only:
+ for col in adata.obs.columns[adata.obs.columns.str.startswith('barcodes_analyzed')
+ | adata.obs.columns.str.startswith('barcode_indices')]:
+ try:
+ del adata.obs[col]
+ except Exception:
+ pass
+ else:
+ # Add a special additional field to .obs if all barcodes are included.
+ if 'barcodes_analyzed_inds' in adata.uns.keys():
+ adata.obs['cellbender_analyzed'] = [True if (i in adata.uns['barcodes_analyzed_inds'])
+ else False for i in range(adata.shape[0])]
+ elif 'barcodes_analyzed_inds' in adata.obs.keys():
+ adata.obs['cellbender_analyzed'] = [True if (i in adata.obs['barcodes_analyzed_inds'].values)
+ else False for i in range(adata.shape[0])]
+
+ return adata
+
+
+def _fill_adata_slots_automatically(adata, d):
+ """Add other information to the adata object in the appropriate slot."""
+
+ # TODO: what about "features_analyzed_inds"? If not all features are analyzed, does this work?
+
+ for key, value in d.items():
+ try:
+ if value is None:
+ continue
+ value = np.asarray(value)
+ if len(value.shape) == 0:
+ adata.uns[key] = value
+ elif value.shape[0] == adata.shape[0]:
+ if (len(value.shape) < 2) or (value.shape[1] < 2):
+ adata.obs[key] = value
+ else:
+ adata.obsm[key] = value
+ elif value.shape[0] == adata.shape[1]:
+ if value.dtype.name.startswith('bytes'):
+ adata.var[key] = value.astype(str)
+ else:
+ adata.var[key] = value
+ else:
+ adata.uns[key] = value
+ except Exception:
+ print('Unable to load data into AnnData: ', key, value, type(value))
+
+
+def load_anndata_from_input(input_file: str) -> anndata.AnnData:
+ """Load an input file into an AnnData object (used in report generation).
+ Equivalent to something like scanpy.read(), but uses cellbender's io.
+
+ Args:
+ input_file: The raw data file
+
+ Returns:
+ adata.AnnData: The anndata object
+
+ """
+
+ # Load data as dict.
+ d = load_data(input_file=input_file)
+
+ # For purely aesthetic purposes, rename slots from the plural to singluar.
+ for key in ['gene_id', 'barcode', 'genome', 'feature_type', 'gene_name']:
+ if key + 's' in d.keys():
+ d[key] = d.pop(key + 's')
+
+ # Create anndata object from dict.
+ adata = anndata.AnnData(X=d.pop('matrix'),
+ obs={'barcode': d.pop('barcode').astype(str)},
+ var={'gene_name': d.pop('gene_name').astype(str)},
+ dtype=int)
+ adata.obs.set_index('barcode', inplace=True)
+ adata.var.set_index('gene_name', inplace=True)
+
+ # Add other information to the anndata object in the appropriate slot.
+ _fill_adata_slots_automatically(adata, d)
+
+ return adata
+
+
+def load_anndata_from_input_and_output(input_file: str,
+ output_file: str,
+ analyzed_barcodes_only: bool = True,
+ input_layer_key: str = 'cellranger',
+ retain_input_metadata: bool = False,
+ gene_expression_encoding_key: str = 'cellbender_embedding',
+ truth_file: Optional[str] = None) -> anndata.AnnData:
+ """Load remove-background output count matrix into an anndata object,
+ together with remove-background metadata and the raw input counts.
+
+ Args:
+ input_file: Raw h5 file (or other compatible remove-background input)
+ used as input for remove-background.
+ output_file: Output h5 file created by remove-background (can be
+ filtered or not).
+ analyzed_barcodes_only: Argument passed to anndata_from_h5().
+ False to load all barcodes, so that the size of
+ the AnnData object will match the size of the input raw count matrix.
+ True to load a limited set of barcodes: only those analyzed by the
+ algorithm. This allows relevant latent variables to be loaded
+ properly into adata.obs and adata.obsm, rather than adata.uns.
+ input_layer_key: Key of the anndata.layer that is created for the raw
+ input count matrix.
+ retain_input_metadata: In addition to loading the CellBender metadata,
+ which happens automatically, set this to True to retain all the
+ metadata from the raw input file as well.
+ gene_expression_encoding_key: The CellBender gene expression embedding
+ will be loaded into adata.obsm[gene_expression_encoding_key]
+ truth_file: File containing truth data if this is a simulation
+
+ Return:
+ anndata.AnnData: AnnData object with counts before and after remove-background,
+ as well as inferred latent variables from remove-background.
+
+ """
+
+ # Load input data.
+ adata_raw = load_anndata_from_input(input_file=input_file)
+
+ # Load remove-background output data.
+ adata_out = anndata_from_h5(output_file, analyzed_barcodes_only=analyzed_barcodes_only)
+
+ # Subset the raw dataset to the relevant barcodes.
+ adata_raw = adata_raw[adata_out.obs.index]
+
+ # TODO: keep the stuff from the raw file too: from obs and var and uns
+ # TODO: maybe use _fill_adata_slots_automatically()? or just copy stuff
+
+ # Put count matrices into 'layers' in anndata for clarity.
+ adata_out.layers[input_layer_key] = adata_raw.X.copy()
+ adata_out.layers['cellbender'] = adata_out.X.copy()
+
+ # Pre-compute a bit of metadata.
+ adata_out.var['n_' + input_layer_key] = \
+ np.array(adata_out.layers[input_layer_key].sum(axis=0), dtype=int).squeeze()
+ adata_out.var['n_cellbender'] = \
+ np.array(adata_out.layers['cellbender'].sum(axis=0), dtype=int).squeeze()
+ adata_out.obs['n_' + input_layer_key] = \
+ np.array(adata_out.layers[input_layer_key].sum(axis=1), dtype=int).squeeze()
+ adata_out.obs['n_cellbender'] = \
+ np.array(adata_out.layers['cellbender'].sum(axis=1), dtype=int).squeeze()
+
+ # Load truth data, if present.
+ if truth_file is not None:
+ adata_truth = anndata_from_h5(truth_file, analyzed_barcodes_only=False)
+ adata_truth = adata_truth[adata_out.obs.index]
+ adata_out.layers['truth'] = adata_truth.X.copy()
+ adata_out.var['n_truth'] = np.array(adata_out.layers['truth'].sum(axis=0), dtype=int).squeeze()
+ adata_out.obs['n_truth'] = np.array(adata_out.layers['truth'].sum(axis=1), dtype=int).squeeze()
+ for key in adata_truth.obs.keys():
+ if key.startswith('truth_'):
+ adata_out.obs[key] = adata_truth.obs[key].copy()
+ for key in adata_truth.uns.keys():
+ if key.startswith('truth_'):
+ adata_out.uns[key] = adata_truth.uns[key].copy()
+ for key in adata_truth.var.keys():
+ if key.startswith('truth_'):
+ adata_out.var[key] = adata_truth.var[key].copy()
+
+ # Rename the CellBender encoding of gene expression.
+ if analyzed_barcodes_only:
+ slot = adata_out.obsm
+ else:
+ slot = adata_out.uns
+ embedding_key = None
+ for key in ['gene_expression_encoding', 'latent_gene_encoding']:
+ if key in slot.keys():
+ embedding_key = key
+ break
+ if gene_expression_encoding_key != embedding_key:
+ slot[gene_expression_encoding_key] = slot[embedding_key].copy()
+ del slot[embedding_key]
+
+ return adata_out
+
+
+def _load_anndata_from_input_and_decontx(input_file: str,
+ output: str,
+ input_layer_key: str = 'cellranger',
+ truth_file: Optional[str] = None) -> anndata.AnnData:
+ """Load decontX output count matrix into an anndata object,
+ together with remove-background metadata and the raw input counts.
+
+ NOTE: this is used only for dev purposes and only in the report
+
+ Args:
+ input_file: Raw h5 file (or other compatible remove-background input)
+ used as input for remove-background.
+ output: Output h5 file, or a directory where decontX MTX and TSV files
+ are stored
+ input_layer_key: Key of the anndata.layer that is created for the raw
+ input count matrix.
+ truth_file: File containing truth data if this is a simulation
+
+ Return:
+ anndata.AnnData: AnnData object with counts before and after remove-background,
+ as well as inferred latent variables from remove-background.
+
+ """
+
+ # Load decontX output data.
+ print('UNSTABLE FEATURE: Trying to load decontX format MTX output')
+ adata_out = load_anndata_from_input(input_file=output)
+ adata_out.var_names_make_unique()
+
+ # Load input data.
+ adata_raw = load_anndata_from_input(input_file=input_file)
+ adata_raw.var_names_make_unique()
+
+ adata_raw = adata_raw[:, [g in adata_out.var.index for g in adata_raw.var.index]].copy()
+ adata_out.var['genome'] = adata_raw.var['genome'].copy()
+ adata_out.var['feature_type'] = adata_raw.var['feature_type'].copy()
+ adata_out.var['gene_id'] = adata_raw.var['gene_id'].copy()
+
+ # Subset the raw dataset to the relevant barcodes.
+ empty_logic = np.array([b not in adata_out.obs.index for b in adata_raw.obs.index])
+ empty_counts = np.array(adata_raw.X[empty_logic].sum(axis=1)).squeeze()
+ approx_ambient = np.array(adata_raw.X[empty_logic][empty_counts > 5].sum(axis=0)).squeeze()
+ approx_ambient = approx_ambient / (approx_ambient.sum() + 1e-10)
+ print(f'Estimated that there are about {np.median(empty_counts[empty_counts > 5])} counts in empties')
+ adata_raw = adata_raw[adata_out.obs.index].copy()
+ adata_out.uns['empty_droplet_size_lognormal_loc'] = np.log(np.median(empty_counts[empty_counts > 5]))
+
+ # Put count matrices into 'layers' in anndata for clarity.
+ adata_out.layers[input_layer_key] = adata_raw.X.copy()
+ adata_out.layers['decontx'] = adata_out.X.copy()
+
+ # Pre-compute a bit of metadata.
+ adata_out.var['n_' + input_layer_key] = np.array(adata_out.layers[input_layer_key].sum(axis=0)).squeeze()
+ adata_out.var['n_decontx'] = np.array(adata_out.layers['decontx'].sum(axis=0)).squeeze()
+ adata_out.obs['n_' + input_layer_key] = np.array(adata_out.layers[input_layer_key].sum(axis=1)).squeeze()
+ adata_out.obs['n_decontx'] = np.array(adata_out.layers['decontx'].sum(axis=1)).squeeze()
+ adata_out.obs['cell_probability'] = 1. # because decontx data contains only cells
+ adata_out.uns['target_false_positive_rate'] = 0.01 # TODO: placeholder
+ adata_out.uns['approximate_ambient_profile'] = approx_ambient
+ adata_out.var['ambient_expression'] = np.nan
+
+ # Load truth data, if present.
+ if truth_file is not None:
+ adata_truth = anndata_from_h5(truth_file, analyzed_barcodes_only=False)
+ adata_truth = adata_truth[adata_out.obs.index]
+
+ # TODO; a check
+ adata_truth = adata_truth[:, [g in adata_out.var.index for g in adata_truth.var.index]].copy()
+
+ adata_out.layers['truth'] = adata_truth.X.copy()
+ adata_out.var['n_truth'] = np.array(adata_out.layers['truth'].sum(axis=0)).squeeze()
+ adata_out.obs['n_truth'] = np.array(adata_out.layers['truth'].sum(axis=1)).squeeze()
+ for key in adata_truth.obs.keys():
+ if key.startswith('truth_'):
+ adata_out.obs[key] = adata_truth.obs[key].copy()
+ for key in adata_truth.uns.keys():
+ if key.startswith('truth_'):
+ adata_out.uns[key] = adata_truth.uns[key].copy()
+ for key in adata_truth.var.keys():
+ if key.startswith('truth_'):
+ adata_out.var[key] = adata_truth.var[key].copy()
+
+ return adata_out
+
+
+def load_anndata_from_input_and_outputs(input_file: str,
+ output_files: Dict[str, str],
+ analyzed_barcodes_only: bool = True,
+ input_layer_key: str = 'cellranger',
+ gene_expression_encoding_key: str = 'cellbender_embedding',
+ truth_file: Optional[str] = None) -> anndata.AnnData:
+ """Load remove-background output count matrices into an anndata object,
+ together with remove-background metadata and the raw input counts.
+
+ The use case would typically be cellbender runs with multiple output files
+ at different FPRs, which we want to compare.
+
+ Args:
+ input_file: Raw h5 file (or other compatible remove-background input)
+ used as input for remove-background.
+ output_files: Output h5 files created by remove-background (can be
+ filtered or not) or some other method. Dict whose keys are layer keys
+ and whose values are file names.
+ analyzed_barcodes_only: Argument passed to anndata_from_h5().
+ False to load all barcodes, so that the size of
+ the AnnData object will match the size of the input raw count matrix.
+ True to load a limited set of barcodes: only those analyzed by the
+ algorithm. This allows relevant latent variables to be loaded
+ properly into adata.obs and adata.obsm, rather than adata.uns.
+ input_layer_key: Key of the anndata.layer that is created for the raw
+ input count matrix.
+ gene_expression_encoding_key: The CellBender gene expression embedding
+ will be loaded into adata.obsm[gene_expression_encoding_key]
+ truth_file: File containing truth data if this is a simulation
+
+ Return:
+ anndata.AnnData: AnnData object with counts before and after remove-background,
+ as well as inferred latent variables from remove-background.
+
+ """
+
+ # Load input data.
+ adata_raw = load_anndata_from_input(input_file=input_file)
+ adata_raw.var_names_make_unique()
+
+ # Load remove-background output data.
+ assert type(output_files) == dict, 'output_files must be a dict whose keys are ' \
+ 'layer names and whose values are file paths.'
+ outs = {}
+ for key, output_file in output_files.items():
+ outs[key] = anndata_from_h5(output_file, analyzed_barcodes_only=analyzed_barcodes_only)
+ outs[key].var_names_make_unique()
+
+ # Subset all datasets to the relevant barcodes and features.
+ relevant_barcodes = set(adata_raw.obs_names)
+ relevant_features = set(adata_raw.var_names)
+ for key, ad in outs.items():
+ relevant_barcodes = relevant_barcodes.intersection(set(ad.obs_names))
+ relevant_features = relevant_features.intersection(set(ad.var_names))
+ if len(relevant_barcodes) < len(adata_raw):
+ print(f'Warning: subsetting to barcodes common to all datasets: there '
+ f'are {len(relevant_barcodes)}')
+ if len(relevant_features) < adata_raw.shape[1]:
+ print(f'Warning: subsetting to features common to all datasets: there '
+ f'are {len(relevant_features)}')
+ adata_raw = adata_raw[list(relevant_barcodes)].copy()
+ adata_raw = adata_raw[:, list(relevant_features)].copy()
+ for i, (key, ad) in enumerate(outs.items()):
+ outs[key] = ad[list(relevant_barcodes)].copy()
+ outs[key] = outs[key][:, list(relevant_features)].copy()
+ if i == 0:
+ print(f'Loading latent variables from one output file: {key}')
+ adata_out = outs[key].copy()
+
+ # Put count matrices into 'layers' in anndata for clarity.
+ adata_out.layers[input_layer_key] = adata_raw.X.copy()
+ for key, ad in outs.items():
+ adata_out.layers[key] = ad.X.copy()
+
+ # Load truth data, if present.
+ if truth_file is not None:
+ adata_truth = anndata_from_h5(truth_file, analyzed_barcodes_only=False)
+ adata_truth = adata_truth[adata_out.obs.index]
+ adata_out.layers['truth'] = adata_truth.X.copy()
+ adata_out.var['n_truth'] = np.array(adata_out.layers['truth'].sum(axis=0)).squeeze()
+ adata_out.obs['n_truth'] = np.array(adata_out.layers['truth'].sum(axis=1)).squeeze()
+ for key in adata_truth.obs.keys():
+ if key.startswith('truth_'):
+ adata_out.obs[key] = adata_truth.obs[key].copy()
+ for key in adata_truth.uns.keys():
+ if key.startswith('truth_'):
+ adata_out.uns[key] = adata_truth.uns[key].copy()
+ for key in adata_truth.var.keys():
+ if key.startswith('truth_'):
+ adata_out.var[key] = adata_truth.var[key].copy()
+
+ # Rename the CellBender encoding of gene expression.
+ if analyzed_barcodes_only:
+ slot = adata_out.obsm
+ else:
+ slot = adata_out.uns
+ embedding_key = None
+ for key in ['gene_expression_encoding', 'latent_gene_encoding']:
+ if key in slot.keys():
+ embedding_key = key
+ break
+ if gene_expression_encoding_key != embedding_key:
+ slot[gene_expression_encoding_key] = slot[embedding_key].copy()
+ del slot[embedding_key]
+
+ return adata_out
diff --git a/cellbender/remove_background/estimation.py b/cellbender/remove_background/estimation.py
new file mode 100644
index 00000000..24437248
--- /dev/null
+++ b/cellbender/remove_background/estimation.py
@@ -0,0 +1,902 @@
+"""Classes and methods for estimation of noise counts, given a posterior."""
+
+import scipy.sparse as sp
+import numpy as np
+import pandas as pd
+import torch
+from torch.distributions.categorical import Categorical
+
+from cellbender.remove_background.sparse_utils import log_prob_sparse_to_dense
+
+from abc import ABC, abstractmethod
+from functools import partial
+from itertools import repeat
+import multiprocessing as mp
+import concurrent.futures
+import time
+from datetime import datetime
+import logging
+from typing import Callable, Union, Dict, Generator, Tuple, List, Optional
+
+
+logger = logging.getLogger('cellbender')
+
+
+class EstimationMethod(ABC):
+ """Base class for estimation of noise counts, given a posterior."""
+
+ def __init__(self, index_converter: 'IndexConverter'):
+ """Instantiate the EstimationMethod.
+ Args:
+ index_converter: The IndexConverter that can be used to translate
+ back and forth between count matrix (n, g) indices, and the
+ flattened, generalized 'm' indices that index the posterior COO.
+ """
+ self.index_converter = index_converter
+ super(EstimationMethod, self).__init__()
+
+ @abstractmethod
+ def estimate_noise(self,
+ noise_log_prob_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ **kwargs) -> sp.csr_matrix:
+ """Given the full probabilistic posterior, compute noise counts.
+ Args:
+ noise_log_prob_coo: The noise log prob data structure: log prob
+ values in a (m, c) COO matrix
+ noise_offsets: Noise count offset values keyed by 'm'.
+ """
+ pass
+
+ def _estimation_array_to_csr(self,
+ data: np.ndarray,
+ m: np.ndarray,
+ noise_offsets: Optional[Dict[int, int]],
+ dtype=np.int64) -> sp.csr_matrix:
+ """Say you have point estimates for each count matrix element (data) and
+ you have the 'm'-indices for each value (m). This returns a CSR matrix
+ that has the shape of the count matrix, where duplicate entries have
+ been summed.
+
+ Args:
+ data: Point estimates for each nonzero entry of the count matrix, in
+ a flat format, indexed by 'm'.
+ m: Array of the same length as data, where each entry is an m-index.
+ noise_offsets: Noise count offset values keyed by 'm'.
+ dtype: Data type for sparse matrix. Int32 is too small for 'm' indices.
+
+ Results:
+ noise_csr: Noise point estimate, as a CSR sparse matrix.
+
+ """
+ return _estimation_array_to_csr(
+ index_converter=self.index_converter,
+ data=data,
+ m=m,
+ noise_offsets=noise_offsets,
+ dtype=dtype,
+ )
+ # row, col = self.index_converter.get_ng_indices(m_inds=m)
+ # if noise_offsets is not None:
+ # data = data + np.array([noise_offsets[i] for i in m])
+ # coo = sp.coo_matrix((data.astype(dtype), (row.astype(dtype), col.astype(dtype))),
+ # shape=self.index_converter.matrix_shape, dtype=dtype)
+ # coo.sum_duplicates()
+ # return coo.tocsr()
+
+
+class SingleSample(EstimationMethod):
+ """A single sample from the noise count posterior"""
+
+ @torch.no_grad()
+ def estimate_noise(self,
+ noise_log_prob_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ device: str = 'cpu',
+ **kwargs) -> sp.csr_matrix:
+ """Given the full probabilistic posterior, compute noise counts by
+ taking a single sample from each probability distribution.
+
+ Args:
+ noise_log_prob_coo: The noise log prob data structure: log prob
+ values in a (m, c) COO matrix
+ noise_offsets: Noise count offset values keyed by 'm'.
+ device: ['cpu', 'cuda'] - whether to perform the pytorch sampling
+ operation on CPU or GPU. It's pretty fast on CPU already.
+
+ Returns:
+ noise_count_csr: Estimated noise count matrix.
+ """
+ def _torch_sample(x):
+ return Categorical(logits=x, validate_args=False).sample()
+
+ result = apply_function_dense_chunks(noise_log_prob_coo=noise_log_prob_coo,
+ fun=_torch_sample,
+ device=device)
+ return self._estimation_array_to_csr(data=result['result'], m=result['m'],
+ noise_offsets=noise_offsets)
+
+
+class Mean(EstimationMethod):
+ """Posterior mean"""
+
+ def estimate_noise(self,
+ noise_log_prob_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ device: str = 'cpu',
+ **kwargs) -> sp.csr_matrix:
+ """Given the full probabilistic posterior, compute noise counts by
+ taking the mean of each probability distribution.
+
+ Args:
+ noise_log_prob_coo: The noise log prob data structure: log prob
+ values in a (m, c) COO matrix
+ noise_offsets: Noise count offset values keyed by 'm'.
+
+ Returns:
+ noise_count_csr: Estimated noise count matrix.
+ """
+ # c = torch.arange(noise_log_prob_coo.shape[1], dtype=float).to(device).t()
+
+ def _torch_mean(x):
+ c = torch.arange(x.shape[1], dtype=float).to(x.device)
+ return torch.matmul(x.exp(), c.t())
+
+ result = apply_function_dense_chunks(noise_log_prob_coo=noise_log_prob_coo,
+ fun=_torch_mean,
+ device=device)
+ return self._estimation_array_to_csr(data=result['result'], m=result['m'],
+ noise_offsets=noise_offsets,
+ dtype=np.float32)
+
+
+class MAP(EstimationMethod):
+ """The canonical maximum a posteriori"""
+
+ @staticmethod
+ def torch_argmax(x):
+ return x.argmax(dim=-1)
+
+ def estimate_noise(self,
+ noise_log_prob_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ device: str = 'cpu',
+ **kwargs) -> sp.csr_matrix:
+ """Given the full probabilistic posterior, compute noise counts by
+ taking the maximum a posteriori (MAP) of each probability distribution.
+
+ Args:
+ noise_log_prob_coo: The noise log prob data structure: log prob
+ values in a (m, c) COO matrix
+ noise_offsets: Noise count offset values keyed by 'm'.
+ device: ['cpu', 'cuda'] - whether to perform the pytorch argmax
+ operation on CPU or GPU. It's pretty fast on CPU already.
+
+ Returns:
+ noise_count_csr: Estimated noise count matrix.
+ """
+ result = apply_function_dense_chunks(noise_log_prob_coo=noise_log_prob_coo,
+ fun=self.torch_argmax,
+ device=device)
+ return self._estimation_array_to_csr(data=result['result'], m=result['m'],
+ noise_offsets=noise_offsets)
+
+
+class ThresholdCDF(EstimationMethod):
+ """Noise estimation via thresholding the noise count CDF"""
+
+ @staticmethod
+ def torch_cdf_fun(x: torch.Tensor, q: float):
+ return (x.exp().cumsum(dim=-1) <= q).sum(dim=-1)
+
+ def estimate_noise(self,
+ noise_log_prob_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ q: float = 0.5,
+ device: str = 'cpu',
+ **kwargs) -> sp.csr_matrix:
+ """Given the full probabilistic posterior, compute noise counts
+
+ Args:
+ noise_log_prob_coo: The noise log prob data structure: log prob
+ values in a (m, c) COO matrix
+ noise_offsets: Noise count offset values keyed by 'm'.
+ q: The CDF threshold value.
+
+ Returns:
+ noise_count_csr: Estimated noise count matrix.
+ """
+ result = apply_function_dense_chunks(noise_log_prob_coo=noise_log_prob_coo,
+ fun=self.torch_cdf_fun,
+ device=device,
+ q=q)
+ return self._estimation_array_to_csr(data=result['result'], m=result['m'],
+ noise_offsets=noise_offsets)
+
+
+def _estimation_array_to_csr(index_converter,
+ data: np.ndarray,
+ m: np.ndarray,
+ noise_offsets: Optional[Dict[int, int]],
+ dtype=np.int64) -> sp.csr_matrix:
+ """Say you have point estimates for each count matrix element (data) and
+ you have the 'm'-indices for each value (m). This returns a CSR matrix
+ that has the shape of the count matrix, where duplicate entries have
+ been summed.
+
+ Args:
+ data: Point estimates for each nonzero entry of the count matrix, in
+ a flat format, indexed by 'm'.
+ m: Array of the same length as data, where each entry is an m-index.
+ noise_offsets: Noise count offset values keyed by 'm'.
+ dtype: Data type for sparse matrix. Int32 is too small for 'm' indices.
+
+ Results:
+ noise_csr: Noise point estimate, as a CSR sparse matrix.
+
+ """
+ row, col = index_converter.get_ng_indices(m_inds=m)
+ if noise_offsets is not None:
+ data = data + np.array([noise_offsets.get(i, 0) for i in m])
+ coo = sp.coo_matrix((data.astype(dtype), (row.astype(dtype), col.astype(dtype))),
+ shape=index_converter.matrix_shape, dtype=dtype)
+ coo.sum_duplicates()
+ return coo.tocsr()
+
+
+def _mckp_chunk_estimate_noise(
+ noise_log_prob_coo: sp.coo_matrix,
+ index_and_logic: Tuple[int, np.ndarray],
+ noise_offsets: Dict[int, int],
+ noise_targets_per_gene: np.ndarray,
+ index_converter: 'IndexConverter',
+ n_chunks: int,
+ verbose: bool = False) -> sp.csr_matrix:
+ """Given the full probabilistic posterior, compute noise counts. This is
+ to be run for a given chunk of genes at a time.
+
+ Args:
+ noise_log_prob_coo: The noise log prob data structure: log prob
+ values in a (m, c) COO matrix. One chunk.
+ index_and_logic: (chunk_index, logical_coo_indexer) from the chunked
+ iterator, as you would get from enumerate()
+ noise_targets_per_gene: Integer noise count target for each gene
+ noise_offsets: Noise count offset values keyed by 'm'.
+ index_converter: IndexConverter to go from 'm' to (n, g)
+ n_chunks: Total chunks, for logging purposes only
+ verbose: True to print lots of intermediate information (for tests)
+
+ Returns:
+ noise_count_csr: Estimated noise count matrix.
+ """
+
+ i = index_and_logic[0]
+
+ if i == 0:
+ tt = time.time()
+
+ coo = _subset_coo(noise_log_prob_coo, index_and_logic[1])
+
+ assert noise_targets_per_gene.size == index_converter.total_n_genes, \
+ f'The number of noise count targets ({noise_targets_per_gene.size}) ' \
+ f'must match the number of genes ({index_converter.total_n_genes})'
+
+ # First we need to compute the MAP to find out which direction to go.
+ t = time.time()
+ map_dict = apply_function_dense_chunks(
+ noise_log_prob_coo=coo,
+ fun=MAP.torch_argmax,
+ device='cpu',
+ )
+ map_csr = _estimation_array_to_csr(
+ data=map_dict['result'],
+ m=map_dict['m'],
+ noise_offsets=noise_offsets,
+ index_converter=index_converter,
+ )
+ logger.debug(f'{timestamp()} Computed initial MAP estimate')
+ logger.debug(f'{timestamp()} Time for MAP calculation = {(time.time() - t):.2f} sec')
+ map_noise_counts_per_gene = np.array(map_csr.sum(axis=0)).squeeze()
+ additional_noise_counts_per_gene = (noise_targets_per_gene
+ - map_noise_counts_per_gene).astype(int)
+ set_positive_genes = set(np.where(additional_noise_counts_per_gene > 0)[0])
+ set_negative_genes = set(np.where(additional_noise_counts_per_gene < 0)[0]) # leave out exact matches
+ abs_additional_noise_counts_per_gene = np.abs(additional_noise_counts_per_gene)
+
+ # Determine which genes need to add and which need to subtract noise counts.
+ n, g = index_converter.get_ng_indices(m_inds=coo.row)
+ df = pd.DataFrame(data={'m': coo.row,
+ 'n': n,
+ 'g': g,
+ 'c': coo.col,
+ 'log_prob': coo.data})
+ logger.debug(f'{timestamp()} Computing step directions')
+ df['positive_step_gene'] = df['g'].apply(lambda gene: gene in set_positive_genes)
+ df['negative_step_gene'] = df['g'].apply(lambda gene: gene in set_negative_genes)
+ df['step_direction'] = (df['positive_step_gene'].astype(int)
+ - df['negative_step_gene'].astype(int)) # -1 or 1
+ logger.debug(f'{timestamp()} Step directions done')
+
+ if verbose:
+ pd.set_option('display.width', 120)
+ pd.set_option('display.max_columns', 20)
+ print(df, end='\n\n')
+
+ # Remove all 'm' entries corresponding to genes where target is met by MAP.
+ df = df[df['step_direction'] != 0]
+
+ # Now we mask out log probs (-np.inf) that represent steps in the wrong direction.
+ logger.debug(f'{timestamp()} Masking')
+ lookup_map_from_m = dict(zip(map_dict['m'], map_dict['result']))
+ df['map'] = df['m'].apply(lambda x: lookup_map_from_m[x])
+ df['mask'] = ((df['negative_step_gene'] & (df['c'] > df['map']))
+ | (df['positive_step_gene'] & (df['c'] < df['map']))) # keep MAP
+ df.loc[df['mask'], 'log_prob'] = -np.inf
+ logger.debug(f'{timestamp()} Masking done')
+
+ # And we remove those entries.
+ df = df[~df['mask']]
+ df = df[[c for c in df.columns if (c != 'mask')]]
+
+ # Sort
+ logger.debug(f'{timestamp()} Sorting')
+ df = df.sort_values(by=['m', 'c'])
+ logger.debug(f'{timestamp()} Sorting done')
+
+ if verbose:
+ print(df, end='\n\n')
+
+ # Do diff for positive and negative separately, without grouping (faster)
+ df_positive_steps = df[df['step_direction'] == 1].copy()
+ df_negative_steps = df[df['step_direction'] == -1].copy()
+
+ logger.debug(f'{timestamp()} Computing deltas')
+ if len(df_positive_steps > 0):
+ df_positive_steps.loc[:, 'delta'] = df_positive_steps['log_prob'].diff(periods=1).apply(np.abs)
+ df_positive_steps.loc[df_positive_steps['c'] == df_positive_steps['map'], 'delta'] = np.nan
+ if len(df_negative_steps > 0):
+ df_negative_steps.loc[:, 'delta'] = df_negative_steps['log_prob'].diff(periods=-1).apply(np.abs)
+ df_negative_steps.loc[df_negative_steps['c'] == df_negative_steps['map'], 'delta'] = np.nan
+ df = pd.concat([df_positive_steps, df_negative_steps], axis=0)
+ logger.debug(f'{timestamp()} Computing deltas done')
+
+ if verbose:
+ print(df, end='\n\n')
+
+ # if this is an empty dataframe, we are not doing anything here beyond MAP
+ if len(df) == 0:
+ return map_csr
+
+ # Remove irrelevant entries: those with infinite delta.
+ df = df[df['delta'].apply(np.isfinite)]
+
+ # if this is an empty dataframe, we are not doing anything here beyond MAP
+ if len(df) == 0:
+ return map_csr
+
+ # How many additional noise counts ("steps") we will need for each gene.
+ logger.debug(f'{timestamp()} Computing nsmallest')
+ df['topk'] = df['g'].apply(lambda gene: abs_additional_noise_counts_per_gene[gene])
+
+ if verbose:
+ print(df, end='\n\n')
+
+ # Now we want the smallest additional_noise_counts_per_gene deltas for each gene.
+ # https://stackoverflow.com/questions/55179493/
+ df_out = df[['m', 'g', 'delta', 'topk']].groupby('g', group_keys=False).apply(
+ lambda x: x.nsmallest(x['topk'].iat[0], columns='delta')
+ )
+ logger.debug(f'{timestamp()} Computing nsmallest done')
+
+ if verbose:
+ print(df_out, end='\n\n')
+
+ # if this is an empty dataframe, we are not doing anything here beyond MAP
+ if len(df_out) == 0:
+ return map_csr
+
+ # And the number by which to increment noise counts per entry 'm' is
+ # now the number of times that each m value appears in this dataframe.
+ logger.debug(f'{timestamp()} Summarizing steps')
+ vc = df_out['m'].value_counts()
+ vc_df = pd.DataFrame(data={'m': vc.index, 'steps': vc.values})
+ step_direction_lookup_from_m = dict(zip(df['m'], df['step_direction']))
+ vc_df['step_direction'] = vc_df['m'].apply(lambda x: step_direction_lookup_from_m[x])
+ vc_df['counts'] = vc_df['steps'] * vc_df['step_direction']
+ steps_csr = _estimation_array_to_csr(
+ data=vc_df['counts'],
+ m=vc_df['m'],
+ noise_offsets=None,
+ index_converter=index_converter,
+ )
+ logger.debug(f'{timestamp()} Summarizing steps done')
+
+ if verbose:
+ print(vc_df, end='\n\n')
+ print('MAP:')
+ print(map_csr.todense())
+
+ logger.info(f'Completed chunk ({i + 1} / {n_chunks})')
+ print(f'Completed chunk ({i + 1} / {n_chunks})') # because logging from a process does not work right
+ if i == 0:
+ logger.info(f' [single chunk took {(time.time() - tt):.2f} mins]')
+ print(f' [single chunk took {(time.time() - tt):.2f} mins]')
+
+ # The final output is those tabulated steps added to the MAP.
+ # The MAP already has the noise offsets, so they are not added to steps_csr.
+ return map_csr + steps_csr
+
+
+class MultipleChoiceKnapsack(EstimationMethod):
+ """Noise estimation via solving a constrained multiple choice knapsack problem"""
+
+ def estimate_noise(self,
+ noise_log_prob_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ noise_targets_per_gene: np.ndarray,
+ verbose: bool = False,
+ n_chunks: Optional[int] = None,
+ use_multiple_processes: bool = False,
+ **kwargs) -> sp.csr_matrix:
+ """Given the full probabilistic posterior, compute noise counts
+
+ Args:
+ noise_log_prob_coo: The noise log prob data structure: log prob
+ values in a (m, c) COO matrix
+ noise_targets_per_gene: Integer noise count target for each gene
+ noise_offsets: Noise count offset values keyed by 'm'.
+ verbose: True to print lots of intermediate information (for tests)
+ n_chunks: Target number of chunks over which to split estimation.
+ If None, targets about 5000 genes per chunk.
+ use_multiple_processes: True to use multiprocessing. Seems faster
+ without using it, not entirely clear why
+
+ Returns:
+ noise_count_csr: Estimated noise count matrix.
+ """
+ if n_chunks is None:
+ n_chunks = max(1, self.index_converter.total_n_genes // 5000)
+ logger.debug(f'Running MCKP estimator in {n_chunks} chunks')
+
+ t0 = time.time()
+
+ if use_multiple_processes:
+
+ logger.info('Dividing dataset into chunks of genes')
+ chunk_logic_list = list(
+ self._gene_chunk_iterator(
+ noise_log_prob_coo=noise_log_prob_coo,
+ n_chunks=n_chunks,
+ )
+ )
+
+ logger.info('Computing the output in asynchronous chunks in parallel...')
+
+ # with mp.get_context('spawn').Pool(processes=mp.cpu_count()) as pool:
+ # csr_matrices = pool.starmap(
+ # _mckp_chunk_estimate_noise,
+ # zip(
+ # repeat(noise_log_prob_coo),
+ # enumerate(chunk_logic_list),
+ # repeat(noise_offsets),
+ # repeat(noise_targets_per_gene),
+ # repeat(self.index_converter),
+ # repeat(n_chunks),
+ # repeat(False), # verbose
+ # ),
+ # )
+
+ futures = []
+ with concurrent.futures.ProcessPoolExecutor(
+ max_workers=mp.cpu_count(),
+ mp_context=mp.get_context('spawn')) as executor:
+ for i, logic in enumerate(chunk_logic_list):
+ kwargs = {
+ 'noise_log_prob_coo': noise_log_prob_coo,
+ 'index_and_logic': (i, logic),
+ 'noise_offsets': noise_offsets,
+ 'noise_targets_per_gene': noise_targets_per_gene,
+ 'index_converter': self.index_converter,
+ 'n_chunks': n_chunks,
+ 'verbose': False,
+ }
+ future = executor.submit(_mckp_chunk_estimate_noise, **kwargs)
+ futures.append(future)
+
+ done, not_done = concurrent.futures.wait(
+ futures,
+ return_when=concurrent.futures.ALL_COMPLETED,
+ )
+ csr_matrices = [f.result() for f in futures]
+
+ else:
+
+ t = time.time()
+
+ csr_matrices = []
+ for i, logic in enumerate(
+ self._gene_chunk_iterator(
+ noise_log_prob_coo=noise_log_prob_coo,
+ n_chunks=n_chunks,
+ )
+ ):
+ logger.info(f'Working on chunk ({i + 1}/{n_chunks})')
+ chunk_csr = self._chunk_estimate_noise(
+ noise_log_prob_coo=_subset_coo(noise_log_prob_coo, logic),
+ noise_offsets=noise_offsets,
+ noise_targets_per_gene=noise_targets_per_gene,
+ verbose=verbose,
+ )
+ csr_matrices.append(chunk_csr)
+ if i == 0:
+ logger.info(f' [{(time.time() - t) / 60:.2f} mins per chunk]')
+ logger.debug(f'{timestamp()} Estimator chunk {i}: shape is {chunk_csr.shape}')
+
+ logger.info(f'{timestamp()} Total MCKP estimation time = {(time.time() - t0):.2f} sec')
+ return sum(csr_matrices)
+
+ def _gene_chunk_iterator(self,
+ noise_log_prob_coo: sp.coo_matrix,
+ n_chunks: int) \
+ -> Generator[np.ndarray, None, None]:
+ """Yields chunks of the posterior that can be treated as independent,
+ from the standpoint of MCKP count estimation. That is, they contain all
+ matrix entries for any genes they include.
+
+ Args:
+ noise_log_prob_coo: Full noise log prob posterior COO
+ n_chunks: For testing, force this many chunks
+
+ Yields:
+ Logical array which indexes elements of coo posterior for the chunk
+ """
+
+ # TODO this generator is way too slow
+
+ # approximate number of entries in a chunk
+ # approx_chunk_entries = (noise_log_prob_coo.data.size - 1) // n_chunks
+
+ # get gene annotations
+ _, genes = self.index_converter.get_ng_indices(m_inds=noise_log_prob_coo.row)
+ genes_series = pd.Series(genes)
+
+ # things we need to keep track of for each chunk
+ # current_chunk_genes = []
+ # entry_logic = np.zeros(noise_log_prob_coo.data.size, dtype=bool)
+
+ # TODO eliminate for loop to speed this up
+ # take the list of genes from the coo, sort it, and divide it evenly
+ # somehow break ties for genes overlapping boundaries of divisions
+ sorted_genes = np.sort(genes)
+ gene_arrays = np.array_split(sorted_genes, n_chunks)
+ last_gene_set = {}
+ for gene_array in gene_arrays:
+ gene_set = set(gene_array)
+ gene_set = gene_set.difference(last_gene_set) # only the new stuff
+ # if there is a second chunk, make sure there is a gene unique to it
+ if (n_chunks > 1) and (len(gene_set) == len(set(genes))): # all genes in first set
+ # this mainly exists for tests
+ gene_set = gene_set - {gene_arrays[-1][-1]}
+ last_gene_set = gene_set
+ entry_logic = genes_series.isin(gene_set).values
+ if sum(entry_logic) > 0:
+ yield entry_logic
+
+ def _chunk_estimate_noise(self,
+ noise_log_prob_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ noise_targets_per_gene: np.ndarray,
+ verbose: bool = False) -> sp.csr_matrix:
+ """Given the full probabilistic posterior, compute noise counts. This is
+ to be run for a given chunk of genes at a time.
+
+ Args:
+ noise_log_prob_coo: The noise log prob data structure: log prob
+ values in a (m, c) COO matrix. One chunk.
+ noise_targets_per_gene: Integer noise count target for each gene
+ noise_offsets: Noise count offset values keyed by 'm'.
+ verbose: True to print lots of intermediate information (for tests)
+
+ Returns:
+ noise_count_csr: Estimated noise count matrix.
+ """
+
+ assert noise_targets_per_gene.size == self.index_converter.total_n_genes, \
+ f'The number of noise count targets ({noise_targets_per_gene.size}) ' \
+ f'must match the number of genes ({self.index_converter.total_n_genes})'
+
+ coo = noise_log_prob_coo.copy() # we will be modifying values
+
+ # First we need to compute the MAP to find out which direction to go.
+ t = time.time()
+ map_dict = apply_function_dense_chunks(noise_log_prob_coo=noise_log_prob_coo,
+ fun=MAP.torch_argmax,
+ device='cpu')
+ map_csr = self._estimation_array_to_csr(data=map_dict['result'],
+ m=map_dict['m'],
+ noise_offsets=noise_offsets)
+ logger.debug(f'{timestamp()} Computed initial MAP estimate')
+ logger.debug(f'{timestamp()} Time for MAP calculation = {(time.time() - t):.2f} sec')
+ map_noise_counts_per_gene = np.array(map_csr.sum(axis=0)).squeeze()
+ additional_noise_counts_per_gene = (noise_targets_per_gene
+ - map_noise_counts_per_gene).astype(int)
+ set_positive_genes = set(np.where(additional_noise_counts_per_gene > 0)[0])
+ set_negative_genes = set(np.where(additional_noise_counts_per_gene < 0)[0]) # leave out exact matches
+ abs_additional_noise_counts_per_gene = np.abs(additional_noise_counts_per_gene)
+
+ # Determine which genes need to add and which need to subtract noise counts.
+ n, g = self.index_converter.get_ng_indices(m_inds=coo.row)
+ df = pd.DataFrame(data={'m': coo.row,
+ 'n': n,
+ 'g': g,
+ 'c': coo.col,
+ 'log_prob': coo.data})
+ logger.debug(f'{timestamp()} Computing step directions')
+ df['positive_step_gene'] = df['g'].apply(lambda gene: gene in set_positive_genes)
+ df['negative_step_gene'] = df['g'].apply(lambda gene: gene in set_negative_genes)
+ df['step_direction'] = (df['positive_step_gene'].astype(int)
+ - df['negative_step_gene'].astype(int)) # -1 or 1
+ logger.debug(f'{timestamp()} Step directions done')
+
+ if verbose:
+ pd.set_option('display.width', 120)
+ pd.set_option('display.max_columns', 20)
+ print(df, end='\n\n')
+
+ # Remove all 'm' entries corresponding to genes where target is met by MAP.
+ df = df[df['step_direction'] != 0]
+
+ # Now we mask out log probs (-np.inf) that represent steps in the wrong direction.
+ logger.debug(f'{timestamp()} Masking')
+ lookup_map_from_m = dict(zip(map_dict['m'], map_dict['result']))
+ df['map'] = df['m'].apply(lambda x: lookup_map_from_m[x])
+ df['mask'] = ((df['negative_step_gene'] & (df['c'] > df['map']))
+ | (df['positive_step_gene'] & (df['c'] < df['map']))) # keep MAP
+ df.loc[df['mask'], 'log_prob'] = -np.inf
+ logger.debug(f'{timestamp()} Masking done')
+
+ # And we remove those entries.
+ df = df[~df['mask']]
+ df = df[[c for c in df.columns if (c != 'mask')]]
+
+ # Sort
+ logger.debug(f'{timestamp()} Sorting')
+ df = df.sort_values(by=['m', 'c'])
+ logger.debug(f'{timestamp()} Sorting done')
+
+ if verbose:
+ print(df, end='\n\n')
+
+ # Do diff for positive and negative separately, without grouping (faster)
+ df_positive_steps = df[df['step_direction'] == 1].copy()
+ df_negative_steps = df[df['step_direction'] == -1].copy()
+
+ logger.debug(f'{timestamp()} Computing deltas')
+ if len(df_positive_steps > 0):
+ df_positive_steps.loc[:, 'delta'] = df_positive_steps['log_prob'].diff(periods=1).apply(np.abs)
+ df_positive_steps.loc[df_positive_steps['c'] == df_positive_steps['map'], 'delta'] = np.nan
+ if len(df_negative_steps > 0):
+ df_negative_steps.loc[:, 'delta'] = df_negative_steps['log_prob'].diff(periods=-1).apply(np.abs)
+ df_negative_steps.loc[df_negative_steps['c'] == df_negative_steps['map'], 'delta'] = np.nan
+ df = pd.concat([df_positive_steps, df_negative_steps], axis=0)
+ logger.debug(f'{timestamp()} Computing deltas done')
+
+ if verbose:
+ print(df, end='\n\n')
+
+ # if this is an empty dataframe, we are not doing anything here beyond MAP
+ if len(df) == 0:
+ return map_csr
+
+ # Remove irrelevant entries: those with infinite delta.
+ df = df[df['delta'].apply(np.isfinite)]
+
+ # if this is an empty dataframe, we are not doing anything here beyond MAP
+ if len(df) == 0:
+ return map_csr
+
+ # How many additional noise counts ("steps") we will need for each gene.
+ logger.debug(f'{timestamp()} Computing nsmallest')
+ df['topk'] = df['g'].apply(lambda gene: abs_additional_noise_counts_per_gene[gene])
+
+ if verbose:
+ print(df, end='\n\n')
+
+ # Now we want the smallest additional_noise_counts_per_gene deltas for each gene.
+ # https://stackoverflow.com/questions/55179493/
+ df_out = df[['m', 'g', 'delta', 'topk']].groupby('g', group_keys=False).apply(
+ lambda x: x.nsmallest(x['topk'].iat[0], columns='delta')
+ )
+ logger.debug(f'{timestamp()} Computing nsmallest done')
+
+ if verbose:
+ print(df_out, end='\n\n')
+
+ # if this is an empty dataframe, we are not doing anything here beyond MAP
+ if len(df_out) == 0:
+ return map_csr
+
+ # And the number by which to increment noise counts per entry 'm' is
+ # now the number of times that each m value appears in this dataframe.
+ logger.debug(f'{timestamp()} Summarizing steps')
+ vc = df_out['m'].value_counts()
+ vc_df = pd.DataFrame(data={'m': vc.index, 'steps': vc.values})
+ step_direction_lookup_from_m = dict(zip(df['m'], df['step_direction']))
+ vc_df['step_direction'] = vc_df['m'].apply(lambda x: step_direction_lookup_from_m[x])
+ vc_df['counts'] = vc_df['steps'] * vc_df['step_direction']
+ steps_csr = self._estimation_array_to_csr(data=vc_df['counts'],
+ m=vc_df['m'],
+ noise_offsets=None)
+ logger.debug(f'{timestamp()} Summarizing steps done')
+
+ if verbose:
+ print(vc_df, end='\n\n')
+ print('MAP:')
+ print(map_csr.todense())
+
+ # The final output is those tabulated steps added to the MAP.
+ # The MAP already has the noise offsets, so they are not added to steps_csr.
+ return map_csr + steps_csr
+
+
+def chunked_iterator(coo: sp.coo_matrix,
+ max_dense_batch_size_GB: float = 1.) \
+ -> Generator[Tuple[sp.coo_matrix, np.ndarray], None, None]:
+ """Return an iterator which yields the full dataset in chunks.
+
+ NOTE: Idea is to prevent memory overflow. The use case is for worst-case
+ scenario algorithms that have to make things into dense matrix chunks in
+ order to do their compute.
+
+ Args:
+ coo: Sparse COO matrix with rows as generalized 'm'-indices and
+ columns as noise count values.
+ max_dense_batch_size_GB: Size of a batch on disk, in gigabytes.
+
+ Returns:
+ A generator that yields compact CSR sparse matrices until the whole dataset
+ has been yielded. "Compact" in the sense that if they are made dense, there
+ will be no all-zero rows.
+ Tuple[chunk csr, actual row values in the full matrix]
+
+ """
+ n_elements_in_batch = max_dense_batch_size_GB * 1e9 / 4 # torch float32 is 4 bytes
+ batch_size = max(1, int(np.floor(n_elements_in_batch / coo.shape[1])))
+
+ # COO rows are not necessarily contiguous or in order
+ unique_m_values = np.unique(coo.row)
+ n_chunks = max(1, len(unique_m_values) // batch_size)
+ row_m_value_chunks = np.array_split(unique_m_values, n_chunks)
+ coo_row_series = pd.Series(coo.row)
+
+ for row_m_values in row_m_value_chunks:
+ logic = coo_row_series.isin(set(row_m_values))
+ # Map these row values to a compact set of unique integers
+ unique_row_values, rows = np.unique(coo.row[logic], return_inverse=True)
+ unique_col_values, cols = np.unique(coo.col[logic], return_inverse=True)
+ chunk_coo = sp.coo_matrix(
+ (coo.data[logic], (rows, cols)),
+ shape=(len(unique_row_values), len(unique_col_values)),
+ )
+ yield (chunk_coo, unique_row_values, unique_col_values)
+
+
+def apply_function_dense_chunks(noise_log_prob_coo: sp.coo_matrix,
+ fun: Callable[[torch.Tensor], torch.Tensor],
+ device: str = 'cpu',
+ **kwargs) \
+ -> Dict[str, np.ndarray]:
+ """Uses chunked_iterator to densify chunked portions of a COO sparse
+ matrix and then applies a function to the dense chunks, keeping track
+ of the results per row.
+
+ NOTE: The function should produce one value per row of the dense matrix.
+ The COO should contain log probability in data.
+
+ Args:
+ noise_log_prob_coo: The posterior noise count log prob data structure,
+ indexed by 'm' as rows
+ fun: Pytorch function that operates on a dense tensor and produces
+ one value per row
+ device: ['cpu', 'cuda'] - whether to perform the pytorch sampling
+ operation on CPU or GPU. It's pretty fast on CPU already.
+ **kwargs: Passed to fun
+
+ Returns:
+ Dict containing
+ 'm': np.ndarray of indices
+ 'result': the values computed by the function
+
+ """
+ array_length = len(np.unique(noise_log_prob_coo.row))
+
+ m = np.zeros(array_length)
+ out = np.zeros(array_length)
+ a = 0
+
+ for coo, row, col in chunked_iterator(coo=noise_log_prob_coo):
+ dense_tensor = torch.tensor(log_prob_sparse_to_dense(coo)).to(device)
+ if torch.numel(dense_tensor) == 0:
+ # github issue 207
+ continue
+ s = fun(dense_tensor, **kwargs)
+ if s.ndim == 0:
+ # avoid "TypeError: len() of a 0-d tensor"
+ len_s = 1
+ else:
+ len_s = len(s)
+ m[a:(a + len_s)] = row
+ out[a:(a + len_s)] = s.detach().cpu().numpy()
+ a = a + len_s
+
+ return {'m': m.astype(int), 'result': out}
+
+
+def pandas_grouped_apply(coo: sp.coo_matrix,
+ fun: Callable[[pd.DataFrame], Union[int, float]],
+ extra_data: Optional[Dict[str, np.ndarray]] = None,
+ sort_first: bool = False,
+ parallel: bool = False) -> Dict[str, np.array]:
+ """Apply function on a sparse COO format (noise log probs) to compute output
+ noise counts using pandas groupby and apply operations on CPU.
+
+ TODO: consider numpy-groupies or np.ufunc.reduceat or similar
+ https://stackoverflow.com/questions/31790819/scipy-sparse-csr-matrix-how-to-get-top-ten-values-and-indices
+
+ Args:
+ coo: COO data structure: (m, c) with 'm'-indexing.
+ fun: Function to be applied with pandas .apply(): turns
+ all the values for a matrix entry (m, :) into one number.
+ extra_data: To include extra information other than 'm', 'c', and the
+ 'log_prob', you can pass in a dict here with array values the same
+ length as 'm' and in the same order.
+ sort_first: Sort the whole dataframe by ['m', 'c'] before applying
+ function (much faster than sorting inside the function call).
+ parallel: Use multiprocessing to run on all cores. This is 4x slower for
+ a dataset of 300 cells. Not clear if it's faster for larger data.
+
+ Returns:
+ output_csr: The aggregated sparse matrix, in row format
+ """
+ df = pd.DataFrame(data={'m': coo.row, 'c': coo.col, 'log_prob': coo.data})
+ if extra_data is not None:
+ for k, v in extra_data.items():
+ df[k] = v
+ if sort_first:
+ df = df.sort_values(by=['m', 'c'], ascending=[True, True])
+ if parallel:
+ m, result_per_m = _parallel_pandas_apply(df_grouped=df.groupby('m'), fun=fun)
+ else:
+ df = df.groupby('m').apply(fun).reset_index()
+ m = df['m'].values
+ result_per_m = df[0].values
+ return {'m': m, 'result': result_per_m}
+
+
+def _parallel_pandas_apply(df_grouped: pd.core.groupby.DataFrameGroupBy,
+ fun: Callable[[pd.DataFrame], Union[int, float]])\
+ -> Tuple[np.ndarray, np.ndarray]:
+ """Use multiprocessing to apply a function to a grouped dataframe in pandas.
+
+ Args:
+ df_grouped: Grouped dataframe, as in df.groupby('m')
+ fun: Function to be applied to dataframe, as in df.groupby('m').apply(fun)
+
+ Returns:
+ Tuple of (groupby_value, groupby_apply_output_for_each_value)
+
+ NOTE: see https://stackoverflow.com/questions/26187759/parallelize-apply-after-pandas-groupby
+ """
+ groupby_val, group_df_list = zip(*[(group_val, group_df)
+ for group_val, group_df in df_grouped])
+ with mp.Pool(mp.cpu_count()) as p:
+ output_list = p.map(fun, group_df_list)
+ return np.array(groupby_val), np.array(output_list)
+
+
+def _subset_coo(coo: sp.coo_matrix, logic: np.ndarray) -> sp.coo_matrix:
+ return sp.coo_matrix((coo.data[logic], (coo.row[logic], coo.col[logic])))
+
+
+def timestamp() -> str:
+ return datetime.now().strftime('%Y-%m-%d %H:%M:%S')
diff --git a/cellbender/remove_background/exceptions.py b/cellbender/remove_background/exceptions.py
index 4136b7d7..c8f321d3 100644
--- a/cellbender/remove_background/exceptions.py
+++ b/cellbender/remove_background/exceptions.py
@@ -1,4 +1,4 @@
-# Exceptions defined by CellBender
+"""Exceptions defined by CellBender"""
class NanException(ArithmeticError):
@@ -11,3 +11,14 @@ class NanException(ArithmeticError):
def __init__(self, param: str):
self.param = param
self.message = 'A wild NaN appeared! In param {' + self.param + '}'
+
+
+class ElboException(ValueError):
+ """Exception raised when training procedure is failing to meet expectations.
+
+ Attributes:
+ message: Message to write to log
+ """
+
+ def __init__(self, message: str):
+ self.message = message
diff --git a/cellbender/remove_background/infer.py b/cellbender/remove_background/infer.py
deleted file mode 100644
index a6a2dac9..00000000
--- a/cellbender/remove_background/infer.py
+++ /dev/null
@@ -1,781 +0,0 @@
-# Posterior inference.
-
-import logging
-from abc import ABC, abstractmethod
-from typing import Tuple, List, Dict, Optional
-
-import numpy as np
-import pyro
-import pyro.distributions as dist
-import scipy.sparse as sp
-import torch
-
-import cellbender.remove_background.consts as consts
-
-
-class Posterior(ABC):
- """Base class Posterior handles posterior count inference.
-
- Args:
- dataset_obj: Dataset object.
- vi_model: Trained RemoveBackgroundPyroModel.
- counts_dtype: Data type of posterior count matrix. Can be one of
- [np.uint32, np.float]
- float_threshold: For floating point count matrices, counts below
- this threshold will be set to zero, for the purposes of constructing
- a sparse matrix. Unused if counts_dtype is np.uint32
-
- Properties:
- mean: Posterior count mean, as a sparse matrix.
-
- """
-
- def __init__(self,
- dataset_obj: 'SingleCellRNACountsDataset', # Dataset
- vi_model: 'RemoveBackgroundPyroModel',
- counts_dtype: np.dtype = np.uint32,
- float_threshold: float = 0.5):
- self.dataset_obj = dataset_obj
- self.vi_model = vi_model
- self.use_cuda = vi_model.use_cuda
- self.analyzed_gene_inds = dataset_obj.analyzed_gene_inds
- self.count_matrix_shape = dataset_obj.data['matrix'].shape
- self.barcode_inds = np.arange(0, self.count_matrix_shape[0])
- self.dtype = counts_dtype
- self.float_threshold = float_threshold
- self._mean = None
- self._latents = None
- super(Posterior, self).__init__()
-
- @abstractmethod
- def _get_mean(self):
- """Obtain mean posterior counts and store in self._mean"""
- pass
-
- @property
- def mean(self) -> sp.csc_matrix:
- if self._mean is None:
- self._get_mean()
- return self._mean
-
- @property
- def latents(self) -> sp.csc_matrix:
- if self._latents is None:
- self._get_latents()
- return self._latents
-
- @property
- def variance(self):
- raise NotImplemented("Posterior count variance not implemented.")
-
- @torch.no_grad()
- def _get_latents(self):
- """Calculate the encoded latent variables."""
-
- data_loader = self.dataset_obj.get_dataloader(use_cuda=self.use_cuda,
- analyzed_bcs_only=True,
- batch_size=500,
- shuffle=False)
-
- z = np.zeros((len(data_loader), self.vi_model.encoder['z'].output_dim))
- d = np.zeros(len(data_loader))
- p = np.zeros(len(data_loader))
- epsilon = np.zeros(len(data_loader))
-
- for i, data in enumerate(data_loader):
-
- if 'chi_ambient' in pyro.get_param_store().keys():
- chi_ambient = pyro.param('chi_ambient').detach()
- else:
- chi_ambient = None
-
- enc = self.vi_model.encoder.forward(x=data, chi_ambient=chi_ambient)
- ind = i * data_loader.batch_size
- z[ind:(ind + data.shape[0]), :] = enc['z']['loc'].detach().cpu().numpy()
-
- phi_loc = pyro.param('phi_loc')
- phi_scale = pyro.param('phi_scale')
-
- d[ind:(ind + data.shape[0])] = \
- dist.LogNormal(loc=enc['d_loc'],
- scale=pyro.param('d_cell_scale')).mean.detach().cpu().numpy()
-
- p[ind:(ind + data.shape[0])] = enc['p_y'].sigmoid().detach().cpu().numpy()
-
- epsilon[ind:(ind + data.shape[0])] = dist.Gamma(enc['epsilon'] * self.vi_model.epsilon_prior,
- self.vi_model.epsilon_prior).mean.detach().cpu().numpy()
-
- self._latents = {'z': z, 'd': d, 'p': p,
- 'phi_loc_scale': [phi_loc.item(), phi_scale.item()],
- 'epsilon': epsilon}
-
- @torch.no_grad()
- def _param_map_estimates(self,
- data: torch.Tensor,
- chi_ambient: torch.Tensor) -> Dict[str, torch.Tensor]:
- """Calculate MAP estimates of mu, the mean of the true count matrix, and
- lambda, the rate parameter of the Poisson background counts.
-
- Args:
- data: Dense tensor minibatch of cell by gene count data.
- chi_ambient: Point estimate of inferred ambient gene expression.
-
- Returns:
- mu_map: Dense tensor of Negative Binomial means for true counts.
- lambda_map: Dense tensor of Poisson rate params for noise counts.
- alpha_map: Dense tensor of Dirichlet concentration params that
- inform the overdispersion of the Negative Binomial.
-
- """
-
- # Encode latents.
- enc = self.vi_model.encoder.forward(x=data,
- chi_ambient=chi_ambient)
- z_map = enc['z']['loc']
-
- chi_map = self.vi_model.decoder.forward(z_map)
- phi_loc = pyro.param('phi_loc')
- phi_scale = pyro.param('phi_scale')
- phi_conc = phi_loc.pow(2) / phi_scale.pow(2)
- phi_rate = phi_loc / phi_scale.pow(2)
- alpha_map = 1. / dist.Gamma(phi_conc, phi_rate).mean
-
- y = (enc['p_y'] > 0).float()
- d_empty = dist.LogNormal(loc=pyro.param('d_empty_loc'),
- scale=pyro.param('d_empty_scale')).mean
- d_cell = dist.LogNormal(loc=enc['d_loc'],
- scale=pyro.param('d_cell_scale')).mean
- epsilon = dist.Gamma(enc['epsilon'] * self.vi_model.epsilon_prior,
- self.vi_model.epsilon_prior).mean
-
- if self.vi_model.include_rho:
- rho = pyro.param("rho_alpha") / (pyro.param("rho_alpha")
- + pyro.param("rho_beta"))
- else:
- rho = None
-
- # Calculate MAP estimates of mu and lambda.
- mu_map = self.vi_model.calculate_mu(epsilon=epsilon,
- d_cell=d_cell,
- chi=chi_map,
- y=y,
- rho=rho)
- lambda_map = self.vi_model.calculate_lambda(epsilon=epsilon,
- chi_ambient=chi_ambient,
- d_empty=d_empty,
- y=y,
- d_cell=d_cell,
- rho=rho,
- chi_bar=self.vi_model.avg_gene_expression)
-
- return {'mu': mu_map, 'lam': lambda_map, 'alpha': alpha_map}
-
- def dense_to_sparse(self,
- chunk_dense_counts: np.ndarray) -> Tuple[List, List, List]:
- """Distill a batch of dense counts into sparse format.
- Barcode numbering is relative to the tensor passed in.
- """
-
- # TODO: speed up by keeping it a torch tensor as long as possible
-
- if chunk_dense_counts.dtype != np.int32:
-
- if self.dtype == np.uint32:
-
- # Turn the floating point count estimates into integers.
- decimal_values, _ = np.modf(chunk_dense_counts) # Stuff after decimal.
- roundoff_counts = np.random.binomial(1, p=decimal_values) # Bernoulli.
- chunk_dense_counts = np.floor(chunk_dense_counts).astype(dtype=int)
- chunk_dense_counts += roundoff_counts
-
- elif self.dtype == np.float32:
-
- # Truncate counts at a threshold value.
- chunk_dense_counts = (chunk_dense_counts *
- (chunk_dense_counts > self.float_threshold))
-
- else:
- raise NotImplementedError(f"Count matrix dtype {self.dtype} is not "
- f"supported. Choose from [np.uint32, "
- f"np.float32]")
-
- # Find all the nonzero counts in this dense matrix chunk.
- nonzero_barcode_inds_this_chunk, nonzero_genes_trimmed = \
- np.nonzero(chunk_dense_counts)
- nonzero_counts = \
- chunk_dense_counts[nonzero_barcode_inds_this_chunk,
- nonzero_genes_trimmed].flatten(order='C')
-
- # Get the original gene index from gene index in the trimmed dataset.
- nonzero_genes = self.analyzed_gene_inds[nonzero_genes_trimmed]
-
- return nonzero_barcode_inds_this_chunk, nonzero_genes, nonzero_counts
-
-
-class ImputedPosterior(Posterior):
- """Posterior count inference using imputation to infer cell mean (d * chi).
-
- Args:
- dataset_obj: Dataset object.
- vi_model: Trained RemoveBackgroundPyroModel.
- guide: Variational posterior pyro guide function, optional. Only
- specify if the required guide function is not vi_model.guide.
- encoder: Encoder that provides encodings of data.
- counts_dtype: Data type of posterior count matrix. Can be one of
- [np.uint32, np.float]
- float_threshold: For floating point count matrices, counts below
- this threshold will be set to zero, for the purposes of constructing
- a sparse matrix. Unused if counts_dtype is np.uint32
-
- Properties:
- mean: Posterior count mean, as a sparse matrix.
- encodings: Encoded latent variables, one per barcode in the dataset.
-
- """
-
- def __init__(self,
- dataset_obj: 'SingleCellRNACountsDataset', # Dataset
- vi_model: 'RemoveBackgroundPyroModel', # Trained variational inference model
- guide=None,
- encoder=None, #: Union[CompositeEncoder, None] = None,
- counts_dtype: np.dtype = np.uint32,
- float_threshold: float = 0.5):
- self.vi_model = vi_model
- self.use_cuda = vi_model.use_cuda
- self.guide = guide if guide is not None else vi_model.encoder
- self.encoder = encoder if encoder is not None else vi_model.encoder
- self._encodings = None
- self._mean = None
- super(ImputedPosterior, self).__init__(dataset_obj=dataset_obj,
- vi_model=vi_model,
- counts_dtype=counts_dtype,
- float_threshold=float_threshold)
-
- @torch.no_grad()
- def _get_mean(self):
- """Send dataset through a guide that returns mean posterior counts.
-
- Keep track of only what is necessary to distill a sparse count matrix.
-
- """
-
- data_loader = self.dataset_obj.get_dataloader(use_cuda=self.use_cuda,
- analyzed_bcs_only=False,
- batch_size=500,
- shuffle=False)
-
- barcodes = []
- genes = []
- counts = []
- ind = 0
-
- for data in data_loader:
-
- # Get return values from guide.
- dense_counts_torch = self._param_map_estimates(data=data,
- chi_ambient=pyro.param("chi_ambient"))
- dense_counts = dense_counts_torch.detach().cpu().numpy()
- bcs_i_chunk, genes_i, counts_i = self.dense_to_sparse(dense_counts)
-
- # Translate chunk barcode inds to overall inds.
- bcs_i = self.barcode_inds[bcs_i_chunk + ind]
-
- # Add sparse matrix values to lists.
- barcodes.append(bcs_i)
- genes.append(genes_i)
- counts.append(counts_i)
-
- # Increment barcode index counter.
- ind += data.shape[0] # Same as data_loader.batch_size
-
- # Convert the lists to numpy arrays.
- counts = np.array(np.concatenate(tuple(counts)), dtype=self.dtype)
- barcodes = np.array(np.concatenate(tuple(barcodes)), dtype=np.uint32)
- genes = np.array(np.concatenate(tuple(genes)), dtype=np.uint32)
-
- # Put the counts into a sparse csc_matrix.
- self._mean = sp.csc_matrix((counts, (barcodes, genes)),
- shape=self.count_matrix_shape)
-
-
-class ProbPosterior(Posterior):
- """Posterior count inference using a noise count probability distribution.
-
- Args:
- dataset_obj: Dataset object.
- vi_model: Trained model: RemoveBackgroundPyroModel
- fpr: Desired false positive rate for construction of the final regularized
- posterior on true counts. False positives are true counts that are
- (incorrectly) removed from the dataset.
- float_threshold: For floating point count matrices, counts below
- this threshold will be set to zero, for the purposes of constructing
- a sparse matrix. Unused if counts_dtype is np.uint32
-
- Properties:
- mean: Posterior count mean, as a sparse matrix.
- encodings: Encoded latent variables, one per barcode in the dataset.
-
- """
-
- def __init__(self,
- dataset_obj: 'SingleCellRNACountsDataset',
- vi_model: 'RemoveBackgroundPyroModel',
- fpr: float = 0.01,
- float_threshold: float = 0.5,
- batch_size: int = consts.PROP_POSTERIOR_BATCH_SIZE,
- cells_posterior_reg_calc: int = consts.CELLS_POSTERIOR_REG_CALC):
- self.vi_model = vi_model
- self.use_cuda = vi_model.use_cuda
- self.fpr = fpr
- self.lambda_multiplier = None
- self._encodings = None
- self._mean = None
- self.batch_size = batch_size
- self.cells_posterior_reg_calc = cells_posterior_reg_calc
- self.random = np.random.RandomState(seed=1234)
- super(ProbPosterior, self).__init__(dataset_obj=dataset_obj,
- vi_model=vi_model,
- counts_dtype=np.uint32,
- float_threshold=float_threshold)
-
- @torch.no_grad()
- def _get_mean(self):
- """Send dataset through a guide that returns mean posterior counts.
-
- Keep track of only what is necessary to distill a sparse count matrix.
-
- """
-
- # Get a dataset of ten cells.
- cell_inds = np.where(self.latents['p'] > 0.9)[0]
- lambda_mults = np.zeros(5)
-
- for i in range(lambda_mults.size):
-
- n_cells = min(self.cells_posterior_reg_calc, cell_inds.size)
- if n_cells == 0:
- raise ValueError('No cells found! Cannot compute expected FPR.')
- cell_ind_subset = self.random.choice(cell_inds, size=n_cells, replace=False)
- cell_data = (torch.tensor(np.array(self.dataset_obj.get_count_matrix()
- [cell_ind_subset, :].todense()).squeeze())
- .float().to(self.vi_model.device))
-
- # Get the latents mu, alpha, and lambda for those cells.
- chi_ambient = pyro.param("chi_ambient")
- map_est = self._param_map_estimates(data=cell_data, chi_ambient=chi_ambient)
-
- # Find the optimal lambda_multiplier value using those cells and target FPR.
- lambda_mult = self._lambda_binary_search_given_fpr(cell_data=cell_data,
- fpr=self.fpr,
- mu_est=map_est['mu'],
- lambda_est=map_est['lam'],
- alpha_est=map_est['alpha'])
- lambda_mults[i] = lambda_mult
-
- optimal_lambda_mult = np.mean(lambda_mults)
- self.lambda_multiplier = optimal_lambda_mult
- logging.info(f'Optimal posterior regularization factor = {optimal_lambda_mult:.2f}')
-
- # Compute posterior in mini-batches.
- analyzed_bcs_only = True
- data_loader = self.dataset_obj.get_dataloader(use_cuda=self.use_cuda,
- analyzed_bcs_only=analyzed_bcs_only,
- batch_size=self.batch_size,
- shuffle=False)
- barcodes = []
- genes = []
- counts = []
- ind = 0
-
- for data in data_loader:
-
- # Compute an estimate of the true counts.
- dense_counts = self._compute_true_counts(data=data,
- chi_ambient=chi_ambient,
- lambda_multiplier=optimal_lambda_mult,
- use_map=False,
- n_samples=9) # must be odd number
- bcs_i_chunk, genes_i, counts_i = self.dense_to_sparse(dense_counts)
-
- # Translate chunk barcode inds to overall inds.
- if analyzed_bcs_only:
- bcs_i = self.dataset_obj.analyzed_barcode_inds[bcs_i_chunk + ind]
- else:
- bcs_i = self.barcode_inds[bcs_i_chunk + ind]
-
- # Add sparse matrix values to lists.
- barcodes.append(bcs_i)
- genes.append(genes_i)
- counts.append(counts_i)
-
- # Increment barcode index counter.
- ind += data.shape[0] # Same as data_loader.batch_size
-
- # Convert the lists to numpy arrays.
- counts = np.array(np.concatenate(tuple(counts)), dtype=self.dtype)
- barcodes = np.array(np.concatenate(tuple(barcodes)), dtype=np.uint32)
- genes = np.array(np.concatenate(tuple(genes)), dtype=np.uint32) # uint16 is too small!
-
- # Put the counts into a sparse csc_matrix.
- self._mean = sp.csc_matrix((counts, (barcodes, genes)),
- shape=self.count_matrix_shape)
-
- @torch.no_grad()
- def _compute_true_counts(self,
- data: torch.Tensor,
- chi_ambient: torch.Tensor,
- lambda_multiplier: float,
- use_map: bool = True,
- n_samples: int = 1) -> np.ndarray:
- """Compute the true de-noised count matrix for this minibatch.
-
- Can use either a MAP estimate of lambda and mu, or can use a sampling
- approach.
-
- Args:
- data: Dense tensor minibatch of cell by gene count data.
- chi_ambient: Point estimate of inferred ambient gene expression.
- lambda_multiplier: Value by which lambda gets multiplied in order
- to compute regularized posterior true counts.
- use_map: True to use a MAP estimate of lambda and mu.
- n_samples: If not using a MAP estimate, this specifies the number
- of samples to use in calculating the posterior mean.
-
- Returns:
- dense_counts: Dense matrix of true de-noised counts.
-
- """
-
- if use_map:
-
- # Calculate MAP estimates of mu and lambda.
- est = self._param_map_estimates(data, chi_ambient)
- mu_map = est['mu']
- lambda_map = est['lam']
- alpha_map = est['alpha']
-
- # Compute the de-noised count matrix given the MAP estimates.
- dense_counts_torch = self._true_counts_from_params(data,
- mu_map,
- lambda_map * lambda_multiplier + 1e-30,
- alpha_map + 1e-30)
-
- dense_counts = dense_counts_torch.detach().cpu().numpy()
-
- else:
-
- assert n_samples > 0, f"Posterior mean estimate needs to be derived " \
- f"from at least one sample: here {n_samples} " \
- f"samples are called for."
-
- dense_counts_torch = torch.zeros((data.shape[0],
- data.shape[1],
- n_samples),
- dtype=torch.float32).to(data.device)
-
- for i in range(n_samples):
-
- # Sample from mu and lambda.
- mu_sample, lambda_sample, alpha_sample = \
- self._param_sample(data)
-
- # Compute the de-noised count matrix given the estimates.
- dense_counts_torch[..., i] = \
- self._true_counts_from_params(data,
- mu_sample,
- lambda_sample * lambda_multiplier + 1e-30,
- alpha_sample + 1e-30)
-
- # Take the median of the posterior true count distribution... torch cuda does not implement mode
- dense_counts = dense_counts_torch.median(dim=2, keepdim=False)[0]
- dense_counts = dense_counts.detach().cpu().numpy().astype(np.int32)
-
- return dense_counts
-
- @torch.no_grad()
- def _param_sample(self,
- data: torch.Tensor) -> Tuple[torch.Tensor,
- torch.Tensor,
- torch.Tensor]:
- """Calculate a single sample estimate of mu, the mean of the true count
- matrix, and lambda, the rate parameter of the Poisson background counts.
-
- Args:
- data: Dense tensor minibatch of cell by gene count data.
-
- Returns:
- mu_sample: Dense tensor sample of Negative Binomial mean for true
- counts.
- lambda_sample: Dense tensor sample of Poisson rate params for noise
- counts.
- alpha_sample: Dense tensor sample of Dirichlet concentration params
- that inform the overdispersion of the Negative Binomial.
-
- """
-
- # Use pyro poutine to trace the guide and sample parameter values.
- guide_trace = pyro.poutine.trace(self.vi_model.guide).get_trace(x=data)
- replayed_model = pyro.poutine.replay(self.vi_model.model, guide_trace)
-
- # Run the model using these sampled values.
- replayed_model_output = replayed_model(x=data)
-
- # The model returns mu, alpha, and lambda.
- mu_sample = replayed_model_output['mu']
- lambda_sample = replayed_model_output['lam']
- alpha_sample = replayed_model_output['alpha']
-
- return mu_sample, lambda_sample, alpha_sample
-
- @torch.no_grad()
- def _true_counts_from_params(self,
- data: torch.Tensor,
- mu_est: torch.Tensor,
- lambda_est: torch.Tensor,
- alpha_est: torch.Tensor) -> torch.Tensor:
- """Calculate a single sample estimate of mu, the mean of the true count
- matrix, and lambda, the rate parameter of the Poisson background counts.
-
- Args:
- data: Dense tensor minibatch of cell by gene count data.
- mu_est: Dense tensor of Negative Binomial means for true counts.
- lambda_est: Dense tensor of Poisson rate params for noise counts.
- alpha_est: Dense tensor of Dirichlet concentration params that
- inform the overdispersion of the Negative Binomial.
-
- Returns:
- dense_counts_torch: Dense matrix of true de-noised counts.
-
- """
-
- # Estimate a reasonable low-end to begin the Poisson summation.
- n = min(100., data.max().item()) # No need to exceed the max value
- poisson_values_low = (lambda_est.detach() - n / 2).int()
-
- poisson_values_low = torch.clamp(torch.min(poisson_values_low,
- (data - n + 1).int()), min=0).float()
-
- # Construct a big tensor of possible noise counts per cell per gene,
- # shape (batch_cells, n_genes, max_noise_counts)
- noise_count_tensor = torch.arange(start=0, end=n) \
- .expand([data.shape[0], data.shape[1], -1]) \
- .float().to(device=data.device)
- noise_count_tensor = noise_count_tensor + poisson_values_low.unsqueeze(-1)
-
- # Compute probabilities of each number of noise counts.
- # NOTE: some values will be outside the support (negative values for NB).
- # The resulting NaNs are ignored by torch.argmax().
- logits = (mu_est.log() - alpha_est.log()).unsqueeze(-1)
- log_prob_tensor = (dist.Poisson(lambda_est.unsqueeze(-1))
- .log_prob(noise_count_tensor)
- + dist.NegativeBinomial(total_count=alpha_est.unsqueeze(-1),
- logits=logits)
- .log_prob(data.unsqueeze(-1) - noise_count_tensor))
- log_prob_tensor = torch.where(noise_count_tensor <= data.unsqueeze(-1),
- log_prob_tensor,
- torch.ones_like(log_prob_tensor) * -np.inf)
-
- # Find the most probable number of noise counts per cell per gene.
- noise_count_map = torch.argmax(log_prob_tensor,
- dim=-1,
- keepdim=False).float()
-
- # Handle the cases where y = 0 (no cell): all counts are noise.
- noise_count_map = torch.where(mu_est == 0,
- data,
- noise_count_map)
-
- # Compute the number of true counts.
- dense_counts_torch = torch.clamp(data - noise_count_map, min=0.)
-
- return dense_counts_torch
-
- @torch.no_grad()
- def _calculate_expected_fpr_from_map(self,
- data: torch.Tensor,
- data_map: torch.Tensor) -> torch.Tensor:
- """Given inferred latent variables and observed total counts,
- generate a MAP estimate for noise counts. Use that MAP estimate to
- compute the expected false positive rate.
-
- Args:
- data: Dense tensor tiny minibatch of cell by gene count data.
- data_map: Dense tensor tiny minibatch of MAP output for that data.
-
- Returns:
- fpr: Expected false positive rate.
-
- """
-
- counts_per_cell = data.sum(dim=-1)
- map_counts_per_cell = data_map.sum(dim=-1)
- fraction_counts_removed_per_cell = (counts_per_cell - map_counts_per_cell) / counts_per_cell
- empty_droplet_mean_counts = dist.LogNormal(loc=pyro.param('d_empty_loc'),
- scale=pyro.param('d_empty_scale')).mean
- ambient_fraction = empty_droplet_mean_counts / counts_per_cell
- if self.vi_model.include_rho:
- swapping_fraction = dist.Beta(pyro.param('rho_alpha'), pyro.param('rho_beta')).mean
- else:
- swapping_fraction = 0.
-
- mean_cell_epsilon = (self.latents['epsilon'][self.latents['p'] > 0.5]).mean()
- target_fraction_counts_removed_per_cell = (ambient_fraction
- + swapping_fraction) / mean_cell_epsilon
- fpr_expectation = (fraction_counts_removed_per_cell - target_fraction_counts_removed_per_cell).mean()
-
- fpr_expectation = torch.clamp(fpr_expectation, min=0.)
-
- return fpr_expectation.mean()
-
- @torch.no_grad()
- def _calculate_expected_fpr_given_lambda_mult(self,
- data: torch.Tensor,
- lambda_mult: float,
- mu_est: torch.Tensor,
- alpha_est: torch.Tensor,
- lambda_est: torch.Tensor) -> torch.Tensor:
- """Given a float lambda_mult, calculate a MAP estimate of output counts,
- and use that estimate to calculate an expected false positive rate.
-
- Args:
- data: Dense tensor tiny minibatch of cell by gene count data.
- lambda_mult: Value of the lambda multiplier.
- mu_est: Dense tensor of Negative Binomial means for true counts.
- lambda_est: Dense tensor of Poisson rate params for noise counts.
- alpha_est: Dense tensor of Dirichlet concentration params that
- inform the overdispersion of the Negative Binomial.
-
- Returns:
- fpr: Expected false positive rate.
-
- """
-
- # Compute MAP estimate of true de-noised counts.
- data_map = self._true_counts_from_params(data=data,
- mu_est=mu_est,
- lambda_est=lambda_est * lambda_mult,
- alpha_est=alpha_est)
-
- # Compute expected false positive rate.
- expected_fpr = self._calculate_expected_fpr_from_map(data=data,
- data_map=data_map)
-
- return expected_fpr
-
- @torch.no_grad()
- def _lambda_binary_search_given_fpr(self,
- cell_data: torch.Tensor,
- fpr: float,
- mu_est: torch.Tensor,
- lambda_est: torch.Tensor,
- alpha_est: torch.Tensor,
- lam_mult_init: float = 1.,
- fpr_tolerance: Optional[float] = None,
- max_iterations: int = consts.POSTERIOR_REG_SEARCH_MAX_ITER) -> float:
- """Perform a binary search for the appropriate lambda-multiplier which will
- achieve a desired false positive rate.
-
- NOTE: It is assumed that
- expected_fpr(lam_mult_bracket[0]) < fpr < expected_fpr(lam_mult_bracket[1]).
- If this is not the case, the algorithm will produce an output close to one
- of the brackets, and FPR control will not be achieved.
-
- Args:
- cell_data: Data from a fraction of the total number of cells.
- fpr: Desired false positive rate.
- mu_est: Dense tensor of Negative Binomial means for true counts.
- lambda_est: Dense tensor of Poisson rate params for noise counts.
- alpha_est: Dense tensor of Dirichlet concentration params that
- inform the overdispersion of the Negative Binomial.
- lam_mult_init: Initial value of the lambda multiplier, hopefully
- close to the unknown target value.
- fpr_tolerance: Tolerated error in expected false positive rate. If
- input is None, defaults to 0.001 or fpr / 10, whichever is
- smaller.
- max_iterations: A cutoff to ensure termination. Even if a tolerable
- solution is not found, the algorithm will stop after this many
- iterations and return the best answer so far.
-
- Returns:
- lam_mult: Value of the lambda-multiplier.
-
- """
-
- assert (fpr > 0) and (fpr < 1), "Target FPR should be in the interval (0, 1)."
- if fpr_tolerance is None:
- fpr_tolerance = min(fpr / 10., 0.001)
-
- # Begin at initial value.
- lam_mult = lam_mult_init
-
- # Calculate an expected false positive rate for this lam_mult value.
- expected_fpr = self._calculate_expected_fpr_given_lambda_mult(
- data=cell_data,
- lambda_mult=lam_mult,
- mu_est=mu_est,
- lambda_est=lambda_est,
- alpha_est=alpha_est)
-
- # Find out what direction we need to move in.
- residual = fpr - expected_fpr
- initial_residual_sign = residual.sign() # plus or minus one
- if initial_residual_sign.item() == 0:
- # In the unlikely event that we hit the value exactly.
- return lam_mult
-
- # Travel in one direction until the direction of FPR error changes.
- lam_limit = lam_mult_init
- i = 0
- while ((lam_limit < consts.POSTERIOR_REG_MAX)
- and (lam_limit > consts.POSTERIOR_REG_MIN)
- and (residual.sign() == initial_residual_sign)
- and (i < max_iterations)):
-
- lam_limit = lam_limit * (initial_residual_sign * 2).exp().item()
-
- # Calculate an expected false positive rate for this lam_mult value.
- expected_fpr = self._calculate_expected_fpr_given_lambda_mult(
- data=cell_data,
- lambda_mult=lam_limit,
- mu_est=mu_est,
- lambda_est=lambda_est,
- alpha_est=alpha_est)
-
- residual = fpr - expected_fpr
- i = i + 1 # one dataset had this go into an infinite loop, taking lam_limit -> 0
-
- # Define the values that bracket the correct answer.
- lam_mult_bracket = np.sort(np.array([lam_mult_init, lam_limit]))
-
- # Binary search algorithm.
- for i in range(max_iterations):
-
- # Current test value for the lambda-multiplier.
- lam_mult = np.mean(lam_mult_bracket)
-
- # Calculate an expected false positive rate for this lam_mult value.
- expected_fpr = self._calculate_expected_fpr_given_lambda_mult(
- data=cell_data,
- lambda_mult=lam_mult,
- mu_est=mu_est,
- lambda_est=lambda_est,
- alpha_est=alpha_est)
-
- # Check on false positive rate and update our bracket values.
- if (expected_fpr < (fpr + fpr_tolerance)) and (expected_fpr > (fpr - fpr_tolerance)):
- break
- elif expected_fpr > fpr:
- lam_mult_bracket[1] = lam_mult
- elif expected_fpr < fpr:
- lam_mult_bracket[0] = lam_mult
-
- # If we stopped due to iteration limit, take the average lam_mult value.
- if i == max_iterations:
- lam_mult = np.mean(lam_mult_bracket)
-
- # Check to see if we have achieved the desired FPR control.
- if not ((expected_fpr < (fpr + fpr_tolerance)) and (expected_fpr > (fpr - fpr_tolerance))):
- print(f'FPR control not achieved in {max_iterations} attempts. '
- f'Output FPR is esitmated to be {expected_fpr.item():.4f}')
-
- return lam_mult
diff --git a/cellbender/remove_background/model.py b/cellbender/remove_background/model.py
index 30f5d448..205e886f 100644
--- a/cellbender/remove_background/model.py
+++ b/cellbender/remove_background/model.py
@@ -20,10 +20,66 @@
from cellbender.remove_background.distributions.NullDist import NullDist
from cellbender.remove_background.exceptions import NanException
-from typing import Optional, Union
+from typing import Optional, Union, Dict, Callable
from numbers import Number
+def calculate_lambda(model_type: str,
+ epsilon: torch.Tensor,
+ chi_ambient: torch.Tensor,
+ d_empty: torch.Tensor,
+ y: Union[torch.Tensor, None] = None,
+ d_cell: Union[torch.Tensor, None] = None,
+ rho: Union[torch.Tensor, None] = None,
+ chi_bar: Union[torch.Tensor, None] = None):
+ """Calculate noise rate based on the model."""
+
+ if model_type == "simple" or model_type == "ambient":
+ lam = epsilon.unsqueeze(-1) * d_empty.unsqueeze(-1) * chi_ambient
+
+ elif model_type == "swapping":
+ lam = (rho.unsqueeze(-1) * chi_bar * epsilon.unsqueeze(-1) *
+ (y.unsqueeze(-1) * d_cell.unsqueeze(-1) + d_empty.unsqueeze(-1)))
+
+ elif model_type == "full":
+ lam = (epsilon.unsqueeze(-1)
+ * ((1. - rho.unsqueeze(-1)) * chi_ambient * d_empty.unsqueeze(-1)
+ + rho.unsqueeze(-1) * chi_bar * (y.unsqueeze(-1) * d_cell.unsqueeze(-1)
+ + d_empty.unsqueeze(-1))))
+ else:
+ raise NotImplementedError(f"model_type was set to {model_type}, "
+ f"which is not implemented.")
+
+ return lam
+
+
+def calculate_mu(model_type: str,
+ epsilon: torch.Tensor,
+ d_cell: torch.Tensor,
+ chi: torch.Tensor,
+ y: Union[torch.Tensor, None] = None,
+ rho: Union[torch.Tensor, None] = None):
+ """Calculate mean expression based on the model."""
+
+ if model_type == 'simple':
+ mu = epsilon.unsqueeze(-1) * d_cell.unsqueeze(-1) * chi
+
+ elif model_type == 'ambient':
+ mu = (y.unsqueeze(-1) * epsilon.unsqueeze(-1)
+ * d_cell.unsqueeze(-1) * chi)
+
+ elif model_type == 'swapping' or model_type == 'full':
+ mu = ((1. - rho.unsqueeze(-1))
+ * y.unsqueeze(-1) * epsilon.unsqueeze(-1)
+ * d_cell.unsqueeze(-1) * chi)
+
+ else:
+ raise NotImplementedError(f"model_type was set to {model_type}, "
+ f"which is not implemented.")
+
+ return mu
+
+
class RemoveBackgroundPyroModel(nn.Module):
"""Class that contains the model and guide used for variational inference.
@@ -32,8 +88,10 @@ class RemoveBackgroundPyroModel(nn.Module):
'swapping', 'full'].
encoder: An instance of an encoder object. Can be a CompositeEncoder.
decoder: An instance of a decoder object.
- dataset_obj: Dataset object which contains relevant priors.
+ dataset_obj_priors: Dict which contains relevant priors.
use_cuda: Will use GPU if True.
+ analyzed_gene_names: Here only so that when we save a checkpoint, if we
+ ever want to look at it, we will know which genes are which.
phi_loc_prior: Mean of gamma distribution for global overdispersion.
phi_scale_prior: Scale of gamma distribution for global overdispersion.
rho_alpha_prior: Param of beta distribution for swapping fraction.
@@ -52,12 +110,18 @@ def __init__(self,
model_type: str,
encoder: Union[nn.Module, encoder_module.CompositeEncoder],
decoder: nn.Module,
- dataset_obj: 'SingleCellRNACountsDataset',
+ dataset_obj_priors: Dict[str, float],
+ n_analyzed_genes: int,
+ n_droplets: int,
+ analyzed_gene_names: np.ndarray,
+ empty_UMI_threshold: int,
+ log_counts_crossover: float,
use_cuda: bool,
phi_loc_prior: float = consts.PHI_LOC_PRIOR,
phi_scale_prior: float = consts.PHI_SCALE_PRIOR,
rho_alpha_prior: float = consts.RHO_ALPHA_PRIOR,
rho_beta_prior: float = consts.RHO_BETA_PRIOR,
+ epsilon_prior: float = consts.EPSILON_PRIOR,
use_exact_log_prob: bool = consts.USE_EXACT_LOG_PROB):
super(RemoveBackgroundPyroModel, self).__init__()
@@ -69,13 +133,19 @@ def __init__(self,
if (self.model_type == "full") or (self.model_type == "swapping"):
self.include_rho = True
- self.n_genes = dataset_obj.analyzed_gene_inds.size
+ self.n_genes = n_analyzed_genes
+ self.n_droplets = n_droplets
+ self.analyzed_gene_names = analyzed_gene_names
self.z_dim = decoder.input_dim
self.encoder = encoder
self.decoder = decoder
self.use_exact_log_prob = use_exact_log_prob
self.loss = {'train': {'epoch': [], 'elbo': []},
- 'test': {'epoch': [], 'elbo': []}}
+ 'test': {'epoch': [], 'elbo': []},
+ 'learning_rate': {'epoch': [], 'value': []}}
+ self.empty_UMI_threshold = empty_UMI_threshold
+ self.log_counts_crossover = log_counts_crossover
+ self.counts_crossover = np.exp(log_counts_crossover)
# Determine whether we are working on a GPU.
if use_cuda:
@@ -93,21 +163,20 @@ def __init__(self,
self.use_cuda = use_cuda
# Priors
- assert dataset_obj.priors['d_std'] > 0, \
- f"Issue with prior: d_std is {dataset_obj.priors['d_std']}, " \
+ assert dataset_obj_priors['d_std'] > 0, \
+ f"Issue with prior: d_std is {dataset_obj_priors['d_std']}, " \
f"but should be > 0."
- assert dataset_obj.priors['cell_counts'] > 0, \
+ assert dataset_obj_priors['cell_counts'] > 0, \
f"Issue with prior: cell_counts is " \
- f"{dataset_obj.priors['cell_counts']}, but should be > 0."
+ f"{dataset_obj_priors['cell_counts']}, but should be > 0."
- self.d_cell_loc_prior = torch.tensor(np.log1p(dataset_obj.priors['cell_counts']))\
+ self.d_cell_loc_prior = torch.tensor(np.log1p(dataset_obj_priors['cell_counts']))\
.float().to(self.device)
- self.d_cell_scale_prior = (torch.tensor(consts.D_STD_PRIOR).to(self.device))
+ self.d_cell_scale_prior = torch.tensor(dataset_obj_priors['d_std']).to(self.device)
self.z_loc_prior = torch.zeros(torch.Size([self.z_dim])).to(self.device)
- self.z_scale_prior = torch.ones(torch.Size([self.z_dim]))\
- .to(self.device)
- self.epsilon_prior = torch.tensor(consts.EPSILON_PRIOR).to(self.device)
+ self.z_scale_prior = torch.ones(torch.Size([self.z_dim])).to(self.device)
+ self.epsilon_prior = torch.tensor(epsilon_prior).to(self.device)
self.phi_loc_prior = (phi_loc_prior
* torch.ones(torch.Size([])).to(self.device))
@@ -120,100 +189,46 @@ def __init__(self,
if self.model_type != "simple":
- assert dataset_obj.priors['empty_counts'] > 0, \
+ assert dataset_obj_priors['empty_counts'] > 0, \
f"Issue with prior: empty_counts should be > 0, but is " \
- f"{dataset_obj.priors['empty_counts']}"
- chi_ambient_sum = dataset_obj.priors['chi_ambient'].sum()
+ f"{dataset_obj_priors['empty_counts']}"
+ chi_ambient_sum = dataset_obj_priors['chi_ambient'].sum()
assert np.isclose(a=chi_ambient_sum, b=[1.], atol=1e-5), \
f"Issue with prior: chi_ambient should sum to 1, but it sums " \
f"to {chi_ambient_sum}"
- chi_bar_sum = dataset_obj.priors['chi_bar'].sum()
+ chi_bar_sum = dataset_obj_priors['chi_bar'].sum()
assert np.isclose(a=chi_bar_sum, b=[1.], atol=1e-5), \
f"Issue with prior: chi_bar should sum to 1, but is {chi_bar_sum}"
- self.d_empty_loc_prior = (np.log1p(dataset_obj
- .priors['empty_counts'],
+ self.d_empty_loc_prior = (np.log1p(dataset_obj_priors['empty_counts'],
dtype=np.float32).item()
* torch.ones(torch.Size([]))
.to(self.device))
- self.d_empty_scale_prior = (np.array(dataset_obj.priors['d_std'],
- dtype=np.float32).item()
+ self.d_empty_scale_prior = (dataset_obj_priors['d_empty_std']
* torch.ones(torch.Size([])).to(self.device))
- self.p_logit_prior = (dataset_obj.priors['cell_logit']
+ self.p_logit_prior = (dataset_obj_priors['cell_logit']
* torch.ones(torch.Size([])).to(self.device))
- self.chi_ambient_init = dataset_obj.priors['chi_ambient']\
- .to(self.device)
- self.avg_gene_expression = dataset_obj.priors['chi_bar'] \
- .to(self.device)
+ self.chi_ambient_init = dataset_obj_priors['chi_ambient'].to(self.device)
+ self.avg_gene_expression = dataset_obj_priors['chi_bar'].to(self.device)
- self.empty_UMI_threshold = (torch.tensor(dataset_obj.empty_UMI_threshold)
+ self.empty_UMI_threshold = (torch.tensor(empty_UMI_threshold)
.float().to(self.device))
else:
self.avg_gene_expression = None
- self.rho_alpha_prior = (rho_alpha_prior
- * torch.ones(torch.Size([])).to(self.device))
- self.rho_beta_prior = (rho_beta_prior
- * torch.ones(torch.Size([])).to(self.device))
+ self.rho_alpha_prior = rho_alpha_prior * torch.ones(torch.Size([])).to(self.device)
+ self.rho_beta_prior = rho_beta_prior * torch.ones(torch.Size([])).to(self.device)
- def calculate_lambda(self,
- epsilon: torch.Tensor,
- chi_ambient: torch.Tensor,
- d_empty: torch.Tensor,
- y: Optional[torch.Tensor] = None,
- d_cell: Optional[torch.Tensor] = None,
- rho: Optional[torch.Tensor] = None,
- chi_bar: Optional[torch.Tensor] = None):
- """Calculate noise rate based on the model."""
-
- if self.model_type == "simple" or self.model_type == "ambient":
- lam = epsilon.unsqueeze(-1) * d_empty.unsqueeze(-1) * chi_ambient
-
- elif self.model_type == "swapping":
- lam = (rho.unsqueeze(-1) * chi_bar *
- (y.unsqueeze(-1) * d_cell.unsqueeze(-1) + d_empty.unsqueeze(-1)))
-
- elif self.model_type == "full":
- lam = (epsilon.unsqueeze(-1)
- * ((1. - rho.unsqueeze(-1)) * chi_ambient * d_empty.unsqueeze(-1)
- + rho.unsqueeze(-1) * chi_bar * (y.unsqueeze(-1) * d_cell.unsqueeze(-1)
- + d_empty.unsqueeze(-1))))
- else:
- raise ValueError(f"model_type was set to {self.model_type}, "
- f"which is not implemented.")
+ def _calculate_mu(self, **kwargs):
+ return calculate_mu(model_type=self.model_type, **kwargs)
- return lam
-
- def calculate_mu(self,
- epsilon: torch.Tensor,
- d_cell: torch.Tensor,
- chi: torch.Tensor,
- y: Optional[torch.Tensor] = None,
- rho: Optional[torch.Tensor] = None):
- """Calculate mean expression based on the model."""
-
- if self.model_type == 'simple':
- mu = epsilon.unsqueeze(-1) * d_cell.unsqueeze(-1) * chi
-
- elif self.model_type == 'ambient':
- mu = (y.unsqueeze(-1) * epsilon.unsqueeze(-1)
- * d_cell.unsqueeze(-1) * chi)
-
- elif self.model_type == 'swapping' or self.model_type == 'full':
- mu = ((1. - rho.unsqueeze(-1))
- * y.unsqueeze(-1) * epsilon.unsqueeze(-1)
- * d_cell.unsqueeze(-1) * chi)
-
- else:
- raise NotImplementedError(f"model_type was set to {self.model_type}, "
- f"which is not implemented.")
-
- return mu
+ def _calculate_lambda(self, **kwargs):
+ return calculate_lambda(model_type=self.model_type, **kwargs)
def model(self, x: torch.Tensor):
"""Data likelihood model.
@@ -224,7 +239,7 @@ def model(self, x: torch.Tensor):
"""
# Register the decoder with pyro.
- pyro.module("decoder", self.decoder)
+ pyro.module("decoder", self.decoder, update_module_params=True)
# Register the hyperparameter for ambient gene expression.
if self.include_empties:
@@ -235,13 +250,16 @@ def model(self, x: torch.Tensor):
else:
chi_ambient = None
- # Sample phi from Gamma prior.
- phi = pyro.sample("phi",
- dist.Gamma(self.phi_conc_prior,
- self.phi_rate_prior))
+ POISSON_APPROX = False
+
+ if not POISSON_APPROX:
+ # Sample phi from Gamma prior.
+ phi = pyro.sample("phi",
+ dist.Gamma(self.phi_conc_prior,
+ self.phi_rate_prior))
# Happens in parallel for each data point (cell barcode) independently:
- with pyro.plate("data", x.size(0),
+ with pyro.plate("data", x.shape[0],
use_cuda=self.use_cuda, device=self.device):
# Sample z from prior.
@@ -251,7 +269,7 @@ def model(self, x: torch.Tensor):
.expand_by([x.size(0)]).to_event(1))
# Decode the latent code z to get fractional gene expression, chi.
- chi = self.decoder.forward(z)
+ chi = pyro.deterministic("chi", self.decoder(z))
# Sample d_cell based on priors.
d_cell = pyro.sample("d_cell",
@@ -283,53 +301,62 @@ def model(self, x: torch.Tensor):
.expand_by([x.size(0)]))
# Sample y, the presence of a real cell, based on p_logit_prior.
- y = pyro.sample("y",
- dist.Bernoulli(logits=self.p_logit_prior - 100.) # TODO
- .expand_by([x.size(0)]))
+ p_logit_prior = get_p_logit_prior(
+ log_counts=x.sum(dim=-1).log(),
+ log_cell_prior_counts=self.d_cell_loc_prior,
+ surely_empty_counts=self.empty_UMI_threshold,
+ naive_p_logit_prior=self.p_logit_prior,
+ )
+ y = pyro.sample("y", dist.Bernoulli(logits=p_logit_prior))
else:
d_empty = None
y = None
# Calculate the mean gene expression counts (for each barcode).
- mu_cell = self.calculate_mu(epsilon=epsilon,
- d_cell=d_cell,
- chi=chi,
- y=y,
- rho=rho)
+ mu_cell = self._calculate_mu(epsilon=epsilon,
+ d_cell=d_cell,
+ chi=chi,
+ y=y,
+ rho=rho)
if self.include_empties:
# Calculate the background rate parameter (for each barcode).
- lam = self.calculate_lambda(epsilon=epsilon,
- chi_ambient=chi_ambient,
- d_empty=d_empty,
- y=y,
- d_cell=d_cell,
- rho=rho,
- chi_bar=self.avg_gene_expression)
+ lam = self._calculate_lambda(epsilon=epsilon,
+ chi_ambient=chi_ambient,
+ d_empty=d_empty,
+ y=y,
+ d_cell=d_cell,
+ rho=rho,
+ chi_bar=self.avg_gene_expression)
else:
lam = torch.zeros([self.n_genes]).to(self.device)
- alpha = phi.reciprocal()
-
- if self.use_exact_log_prob:
-
- # Sample gene expression from our Negative Binomial Poisson
- # Convolution distribution, and compare with observed data.
- c = pyro.sample("obs", NBPC(mu=mu_cell + consts.NBPC_MU_EPS_SAFEGAURD,
- alpha=alpha + consts.NBPC_ALPHA_EPS_SAFEGAURD,
- lam=lam + consts.NBPC_LAM_EPS_SAFEGAURD,
- max_poisson=consts.NBPC_EXACT_N_TERMS).to_event(1),
+ if POISSON_APPROX:
+ # Data distributed as the sum of two Poissons.
+ c = pyro.sample("obs",
+ dist.Poisson(rate=mu_cell + lam + consts.POISSON_EPS_SAFEGAURD).to_event(1),
obs=x.reshape(-1, self.n_genes))
+ alpha = None
else:
+ alpha = phi.reciprocal()
- # Use a negative binomial approximation as the observation model.
- c = pyro.sample("obs", NBPCapprox(mu=mu_cell + consts.NBPC_MU_EPS_SAFEGAURD,
- alpha=alpha + consts.NBPC_ALPHA_EPS_SAFEGAURD,
- lam=lam + consts.NBPC_LAM_EPS_SAFEGAURD).to_event(1),
- obs=x.reshape(-1, self.n_genes))
+ if not consts.USE_EXACT_LOG_PROB:
+
+ # Use a negative binomial approximation as the observation model.
+ c = pyro.sample("obs", NBPCapprox(mu=mu_cell + consts.NBPC_MU_EPS_SAFEGAURD,
+ alpha=alpha + consts.NBPC_ALPHA_EPS_SAFEGAURD,
+ lam=lam + consts.NBPC_LAM_EPS_SAFEGAURD).to_event(1),
+ obs=x.reshape(-1, self.n_genes))
+
+ else:
+ c = pyro.sample("obs", NBPC(mu=mu_cell + consts.NBPC_MU_EPS_SAFEGAURD,
+ alpha=alpha + consts.NBPC_ALPHA_EPS_SAFEGAURD,
+ lam=lam + consts.NBPC_LAM_EPS_SAFEGAURD,
+ max_poisson=100).to_event(1),
+ obs=x.reshape(-1, self.n_genes))
if self.include_empties:
@@ -341,10 +368,9 @@ def model(self, x: torch.Tensor):
# Additionally use the surely empty droplets for regularization,
# since we know these droplets by their UMI counts.
counts = x.sum(dim=-1, keepdim=False)
- surely_empty_mask = ((counts < self.empty_UMI_threshold)
- .bool().to(self.device))
+ surely_cell_mask = (counts >= self.d_cell_loc_prior.exp()).bool().to(self.device)
- with poutine.mask(mask=surely_empty_mask):
+ with poutine.mask(mask=y.detach().bool().logical_not()): # surely_empty_mask):
with poutine.scale(scale=consts.REG_SCALE_AMBIENT_EXPRESSION):
@@ -353,45 +379,57 @@ def model(self, x: torch.Tensor):
else:
r = None
- # Semi-supervision of ambient expression.
- lam = self.calculate_lambda(epsilon=epsilon.detach(),
- chi_ambient=chi_ambient,
- d_empty=d_empty,
- y=torch.zeros_like(d_empty),
- d_cell=d_cell.detach(),
- rho=r,
- chi_bar=self.avg_gene_expression)
+ # Semi-supervision of ambient expression using all empties.
+ lam = self._calculate_lambda(epsilon=torch.tensor(1.).to(d_empty.device), # epsilon.detach(),
+ chi_ambient=chi_ambient,
+ d_empty=d_empty,
+ y=torch.zeros_like(d_empty),
+ d_cell=d_cell.detach(),
+ rho=r,
+ chi_bar=self.avg_gene_expression)
pyro.sample("obs_empty",
- dist.Poisson(rate=lam + 1e-10).to_event(1),
+ dist.Poisson(rate=lam + consts.POISSON_EPS_SAFEGAURD).to_event(1),
obs=x.reshape(-1, self.n_genes))
- # Semi-supervision of cell probabilities.
- with poutine.scale(scale=consts.REG_SCALE_EMPTY_PROB):
-
- p_logit_posterior = pyro.sample("p_passback",
- NullDist(torch.zeros(1)
- .to(self.device))
- .expand_by([x.size(0)]))
-
- pyro.sample("obs_empty_y",
- dist.Normal(loc=p_logit_posterior,
- scale=1.),
- obs=torch.ones_like(y) * -5.)
-
- # Additionally use some high-count droplets for cell prob regularization.
- surely_cell_mask = (torch.where(counts >= self.d_cell_loc_prior.exp(),
- torch.ones_like(counts),
- torch.zeros_like(counts))
- .bool().to(self.device))
-
- with poutine.mask(mask=surely_cell_mask):
- with poutine.scale(scale=consts.REG_SCALE_CELL_PROB):
- pyro.sample("obs_cell_y",
- dist.Normal(loc=p_logit_posterior,
- scale=1.),
- obs=torch.ones_like(y) * 5.)
-
- return {'mu': mu_cell, 'lam': lam, 'alpha': alpha, 'counts': c}
+ # Grab our posterior for the logit cell probability (this is a workaround).
+ p_logit_posterior = pyro.sample("p_passback",
+ NullDist(torch.zeros(1).to(self.device))
+ .expand_by([x.size(0)]))
+
+ # Softer semi-supervision to encourage cell probabilities to do the right thing.
+ probably_empty_mask = (counts < self.counts_crossover).bool().to(self.device)
+ probably_cell_mask = (counts >= self.counts_crossover).bool().to(self.device)
+
+ with poutine.mask(mask=probably_empty_mask):
+ with poutine.scale(scale=consts.REG_SCALE_SOFT_SUPERVISION):
+ pyro.sample("obs_probably_empty_y",
+ dist.Normal(loc=-1 * torch.ones_like(y) * consts.REG_LOGIT_MEAN,
+ scale=consts.REG_LOGIT_SOFT_SCALE),
+ obs=p_logit_posterior)
+
+ with poutine.mask(mask=probably_cell_mask):
+ with poutine.scale(scale=consts.REG_SCALE_SOFT_SUPERVISION):
+ pyro.sample("obs_probably_cell_y",
+ dist.Normal(loc=torch.ones_like(y) * consts.REG_LOGIT_MEAN,
+ scale=consts.REG_LOGIT_SOFT_SCALE),
+ obs=p_logit_posterior)
+
+ # Regularization of epsilon.mean()
+ if surely_cell_mask.sum() >= 2:
+ epsilon_median = epsilon[probably_cell_mask].median()
+ # with poutine.scale(scale=probably_cell_mask.sum() / 10.):
+ pyro.sample("epsilon_mean",
+ dist.Normal(loc=epsilon_median, scale=0.01),
+ obs=torch.ones_like(epsilon_median))
+
+ epsilon_median_empty = epsilon[probably_empty_mask].median()
+ # with poutine.scale(scale=probably_cell_mask.sum() / 10.):
+ pyro.sample("epsilon_empty_mean",
+ dist.Normal(loc=epsilon_median_empty, scale=0.01),
+ obs=torch.ones_like(epsilon_median_empty))
+
+ return {'chi_ambient': chi_ambient, 'z': z,
+ 'mu': mu_cell, 'lam': lam, 'alpha': alpha, 'counts': c}
@config_enumerate(default='parallel')
def guide(self, x: torch.Tensor):
@@ -411,7 +449,7 @@ def guide(self, x: torch.Tensor):
# Register the encoder(s) with pyro.
for name, module in self.encoder.items():
- pyro.module("encoder_" + name, module)
+ pyro.module("encoder_" + name, module, update_module_params=True)
# Initialize variational parameters for d_cell.
d_cell_scale = pyro.param("d_cell_scale",
@@ -466,7 +504,7 @@ def guide(self, x: torch.Tensor):
pyro.sample("phi", dist.Gamma(phi_conc, phi_rate))
# Happens in parallel for each data point (cell barcode) independently:
- with pyro.plate("data", x.size(0),
+ with pyro.plate("data", x.shape[0],
use_cuda=self.use_cuda, device=self.device):
# Sample swapping fraction rho.
@@ -483,10 +521,14 @@ def guide(self, x: torch.Tensor):
scale=d_empty_scale)
.expand_by([x.size(0)]))
- enc = self.encoder.forward(x=x, chi_ambient=chi_ambient)
+ enc = self.encoder(x=x,
+ chi_ambient=chi_ambient.detach(),
+ cell_prior_log=self.d_cell_loc_prior)
else:
- enc = self.encoder.forward(x=x, chi_ambient=None)
+ enc = self.encoder(x=x,
+ chi_ambient=None,
+ cell_prior_log=self.d_cell_loc_prior)
# Code specific to models with empty droplets.
if self.include_empties:
@@ -500,63 +542,52 @@ def guide(self, x: torch.Tensor):
# Sample the Bernoulli y from encoded p(y).
y = pyro.sample("y", dist.Bernoulli(logits=enc['p_y']))
- # Gate d_cell_loc so empty droplets do not give big gradients.
- prob = enc['p_y'].sigmoid() # Logits to probability
- d_cell_loc_gated = (prob * enc['d_loc'] + (1 - prob)
- * self.d_cell_loc_prior) # NOTE: necessary to pass on sim6
-
- # Sample d based on the encoding.
- d_cell = pyro.sample("d_cell", dist.LogNormal(loc=d_cell_loc_gated,
- scale=d_cell_scale))
+ # Get cell probabilities for gating.
+ prob = enc['p_y'].sigmoid().detach() # Logits to probability
# Mask out empty droplets.
- with poutine.mask(mask=y.bool()):
+ with poutine.mask(mask=y.bool().detach()):
# Sample latent code z for the barcodes containing cells.
- pyro.sample("z",
- dist.Normal(loc=enc['z']['loc'],
- scale=enc['z']['scale'])
- .to_event(1))
+ z = pyro.sample("z", dist.Normal(loc=enc['z']['loc'], scale=enc['z']['scale'])
+ .to_event(1))
- # Gate epsilon and sample.
- epsilon_gated = (prob * enc['epsilon'] + (1 - prob) * 1.)
+ # Gate d based and sample.
+ d_cell_loc_gated = (prob * enc['d_loc'] + (1 - prob)
+ * self.d_cell_loc_prior) # NOTE: necessary to pass on sim6
+ d_cell = pyro.sample("d_cell", dist.LogNormal(loc=d_cell_loc_gated,
+ scale=d_cell_scale))
- epsilon = pyro.sample("epsilon",
- dist.Gamma(epsilon_gated * self.epsilon_prior,
- self.epsilon_prior))
+ # Gate epsilon and sample.
+ epsilon_gated = (prob * enc['epsilon'] + (1 - prob) * 1.)
+ epsilon = pyro.sample("epsilon",
+ dist.Gamma(concentration=epsilon_gated * self.epsilon_prior,
+ rate=self.epsilon_prior))
else:
# Sample d based on the encoding.
- pyro.sample("d_cell", dist.LogNormal(loc=enc['d_loc'],
- scale=d_cell_scale))
+ pyro.sample("d_cell", dist.LogNormal(loc=enc['d_loc'], scale=d_cell_scale))
# Sample latent code z for each cell.
- pyro.sample("z",
- dist.Normal(loc=enc['z']['loc'],
- scale=enc['z']['scale'])
+ pyro.sample("z", dist.Normal(loc=enc['z']['loc'], scale=enc['z']['scale'])
.to_event(1))
-def get_ambient_expression() -> Optional[np.ndarray]:
- """Get ambient RNA expression for 'empty' droplets.
-
- Return:
- chi_ambient: The ambient gene expression profile, as a normalized
- vector that sums to one.
-
- Note:
- Inference must have been performed on a model with a 'chi_ambient'
- hyperparameter prior to making this call.
-
- """
-
- chi_ambient = None
-
- if 'chi_ambient' in pyro.get_param_store().keys():
- chi_ambient = to_ndarray(pyro.param('chi_ambient')).squeeze()
-
- return chi_ambient
+def get_p_logit_prior(log_counts: torch.Tensor,
+ log_cell_prior_counts: float,
+ surely_empty_counts: torch.Tensor,
+ naive_p_logit_prior: float) -> torch.Tensor:
+ """Compute the logit cell probability prior per droplet based on counts"""
+ ones = torch.ones_like(log_counts)
+ p_logit_prior = ones * naive_p_logit_prior
+ p_logit_prior = torch.where(log_counts <= (surely_empty_counts.log() + log_cell_prior_counts) / 2,
+ ones * -100.,
+ p_logit_prior)
+ p_logit_prior = torch.where(log_counts >= log_cell_prior_counts,
+ ones * consts.REG_LOGIT_MEAN,
+ p_logit_prior)
+ return p_logit_prior
def get_rho() -> Optional[np.ndarray]:
@@ -582,6 +613,15 @@ def get_rho() -> Optional[np.ndarray]:
return rho
+def get_param_store_key(key: str) -> Union[np.ndarray, None]:
+ val = None
+
+ if key in pyro.get_param_store().keys():
+ val = to_ndarray(pyro.param(key)).squeeze()
+
+ return val
+
+
def to_ndarray(x: Union[Number, np.ndarray, torch.Tensor]) -> np.ndarray:
"""Convert a numeric value or array to a numpy array on cpu."""
diff --git a/cellbender/remove_background/posterior.py b/cellbender/remove_background/posterior.py
new file mode 100644
index 00000000..37587026
--- /dev/null
+++ b/cellbender/remove_background/posterior.py
@@ -0,0 +1,1708 @@
+"""Posterior generation and regularization."""
+
+import pyro
+import pyro.distributions as dist
+import torch
+import numpy as np
+import scipy.sparse as sp
+import pandas as pd
+
+import cellbender.remove_background.consts as consts
+from cellbender.remove_background.model import calculate_mu, calculate_lambda
+from cellbender.monitor import get_hardware_usage
+from cellbender.remove_background.data.dataprep import DataLoader
+from cellbender.remove_background.data.dataset import get_dataset_obj
+from cellbender.remove_background.estimation import EstimationMethod, \
+ MultipleChoiceKnapsack, Mean, MAP, apply_function_dense_chunks
+from cellbender.remove_background.sparse_utils import dense_to_sparse_op_torch, \
+ log_prob_sparse_to_dense, csr_set_rows_to_zero
+from cellbender.remove_background.checkpoint import load_from_checkpoint, \
+ unpack_tarball, make_tarball, load_checkpoint
+from cellbender.remove_background.data.io import \
+ write_posterior_coo_to_h5, load_posterior_from_h5
+
+from typing import Tuple, List, Dict, Optional, Union, Callable
+from abc import ABC, abstractmethod
+import logging
+import argparse
+import tempfile
+import shutil
+import time
+import os
+
+
+logger = logging.getLogger('cellbender')
+
+
+def load_or_compute_posterior_and_save(dataset_obj: 'SingleCellRNACountsDataset',
+ inferred_model: 'RemoveBackgroundPyroModel',
+ args: argparse.Namespace) -> 'Posterior':
+ """After inference, compute the full posterior noise count log probability
+ distribution. Save it and make it part of the checkpoint file.
+
+ NOTE: Loads posterior from checkpoint file if available.
+ NOTE: Saves posterior as args.output_file + '_posterior.npz' and adds this
+ file to the checkpoint tarball as well.
+
+ Args:
+ dataset_obj: Input data in the form of a SingleCellRNACountsDataset
+ object.
+ inferred_model: Model after inference is complete.
+ args: Input command line parsed arguments.
+
+ Returns:
+ posterior: Posterior object with noise count log prob computed, as well
+ as regularization if called for.
+
+ """
+
+ assert os.path.exists(args.input_checkpoint_tarball), \
+ f'Checkpoint file {args.input_checkpoint_tarball} does not exist, ' \
+ f'presumably because saving of the checkpoint file has been manually ' \
+ f'interrupted. load_or_compute_posterior_and_save() will not work ' \
+ f'properly without an existing checkpoint file. Please re-run and ' \
+ f'allow a checkpoint file to be saved.'
+
+ def _do_posterior_regularization(posterior: Posterior):
+
+ # Optional posterior regularization.
+ if args.posterior_regularization is not None:
+ if args.posterior_regularization == 'PRq':
+ posterior.regularize_posterior(
+ regularization=PRq,
+ alpha=args.prq_alpha,
+ device='cuda',
+ )
+ elif args.posterior_regularization == 'PRmu':
+ posterior.regularize_posterior(
+ regularization=PRmu,
+ raw_count_matrix=dataset_obj.data['matrix'],
+ fpr=args.fpr[0],
+ per_gene=False,
+ device='cuda',
+ )
+ elif args.posterior_regularization == 'PRmu_gene':
+ posterior.regularize_posterior(
+ regularization=PRmu,
+ raw_count_matrix=dataset_obj.data['matrix'],
+ fpr=args.fpr[0],
+ per_gene=True,
+ device='cuda',
+ )
+ else:
+ raise ValueError(f'Got a posterior regularization input of '
+ f'"{args.posterior_regularization}", which is not '
+ f'allowed. Use ["PRq", "PRmu", "PRmu_gene"]')
+
+ else:
+ # Delete a pre-existing posterior regularization in case an old one was saved.
+ posterior.clear_regularized_posterior()
+
+ posterior = Posterior(
+ dataset_obj=dataset_obj,
+ vi_model=inferred_model,
+ posterior_batch_size=args.posterior_batch_size,
+ debug=args.debug,
+ )
+ try:
+ ckpt_posterior = load_from_checkpoint(tarball_name=args.input_checkpoint_tarball,
+ filebase=args.checkpoint_filename,
+ to_load=['posterior'])
+ except ValueError:
+ # input checkpoint tarball was not a match for this workflow
+ # but we still may have saved a new tarball
+ ckpt_posterior = load_from_checkpoint(tarball_name=consts.CHECKPOINT_FILE_NAME,
+ filebase=args.checkpoint_filename,
+ to_load=['posterior'])
+ if os.path.exists(ckpt_posterior.get('posterior_file', 'does_not_exist')):
+ # Load posterior if it was saved in the checkpoint.
+ posterior.load(file=ckpt_posterior['posterior_file'])
+ _do_posterior_regularization(posterior)
+ else:
+
+ # Compute posterior.
+ logger.info('Posterior not currently included in checkpoint.')
+ posterior.cell_noise_count_posterior_coo()
+ _do_posterior_regularization(posterior)
+
+ # Save posterior and add it to checkpoint tarball.
+ posterior_file = args.output_file[:-3] + '_posterior.h5'
+ saved = posterior.save(file=posterior_file)
+ success = False
+ if saved:
+ with tempfile.TemporaryDirectory() as tmp_dir:
+ unpacked = unpack_tarball(tarball_name=args.input_checkpoint_tarball,
+ directory=tmp_dir)
+ if unpacked:
+ shutil.copy(posterior_file, os.path.join(tmp_dir, 'posterior.h5'))
+ all_ckpt_files = [os.path.join(tmp_dir, f) for f in os.listdir(tmp_dir)
+ if os.path.isfile(os.path.join(tmp_dir, f))]
+ success = make_tarball(files=all_ckpt_files,
+ tarball_name=args.input_checkpoint_tarball)
+ if success:
+ logger.info('Added posterior object to checkpoint file.')
+ else:
+ logger.warning('Failed to add posterior object to checkpoint file.')
+
+ return posterior
+
+
+class Posterior:
+ """Posterior handles posteriors on latent variables and denoised counts.
+
+ Args:
+ dataset_obj: Dataset object.
+ vi_model: Trained RemoveBackgroundPyroModel.
+ posterior_batch_size: Number of barcodes in a minibatch, used to
+ calculate posterior probabilities (memory hungry).
+ counts_dtype: Data type of posterior count matrix. Can be one of
+ [np.uint32, np.float]
+ float_threshold: For floating point count matrices, counts below
+ this threshold will be set to zero, for the purposes of constructing
+ a sparse matrix. Unused if counts_dtype is np.uint32
+ debug: True to print debugging messages (involves extra compute)
+
+ Properties:
+ full_noise_count_posterior_csr: The posterior noise log probability
+ distribution, as a sparse matrix.
+ latents_map: MAP estimate of latent variables
+
+ Examples:
+
+ posterior = Posterior()
+
+ """
+
+ def __init__(self,
+ dataset_obj: Optional['SingleCellRNACountsDataset'], # Dataset
+ vi_model: Optional['RemoveBackgroundPyroModel'],
+ posterior_batch_size: int = 128,
+ counts_dtype: np.dtype = np.uint32,
+ float_threshold: Optional[float] = 0.5,
+ debug: bool = False):
+ self.dataset_obj = dataset_obj
+ self.vi_model = vi_model
+ if vi_model is not None:
+ self.vi_model.eval()
+ self.vi_model.encoder['z'].eval()
+ self.vi_model.encoder['other'].eval()
+ self.vi_model.decoder.eval()
+ self.use_cuda = (torch.cuda.is_available() if vi_model is None
+ else vi_model.use_cuda)
+ self.device = 'cuda' if self.use_cuda else 'cpu'
+ self.analyzed_gene_inds = (None if (dataset_obj is None)
+ else dataset_obj.analyzed_gene_inds)
+ self.count_matrix_shape = (None if (dataset_obj is None)
+ else dataset_obj.data['matrix'].shape)
+ self.barcode_inds = (None if (dataset_obj is None)
+ else np.arange(0, self.count_matrix_shape[0]))
+ self.dtype = counts_dtype
+ self.debug = debug
+ self.float_threshold = float_threshold
+ self.posterior_batch_size = posterior_batch_size
+ self._noise_count_posterior_coo = None
+ self._noise_count_posterior_kwargs = None
+ self._noise_count_posterior_coo_offsets = None
+ self._noise_count_regularized_posterior_coo = None
+ self._noise_count_regularized_posterior_kwargs = None
+ self._latents = None
+ if dataset_obj is not None:
+ self.index_converter = IndexConverter(
+ total_n_cells=dataset_obj.data['matrix'].shape[0],
+ total_n_genes=dataset_obj.data['matrix'].shape[1],
+ )
+
+ def save(self, file: str) -> bool:
+ """Save the full posterior in HDF5 format."""
+
+ if self._noise_count_posterior_coo is None:
+ self.cell_noise_count_posterior_coo()
+
+ n, g = self.index_converter.get_ng_indices(self._noise_count_posterior_coo.row)
+
+ d = {'posterior_coo': self._noise_count_posterior_coo,
+ 'noise_count_offsets': self._noise_count_posterior_coo_offsets,
+ 'posterior_kwargs': self._noise_count_posterior_kwargs,
+ 'regularized_posterior_coo': self._noise_count_regularized_posterior_coo,
+ 'regularized_posterior_kwargs': self._noise_count_regularized_posterior_kwargs,
+ 'latents': self.latents_map,
+ 'feature_inds': g,
+ 'barcode_inds': n}
+
+ try:
+ logger.info(f'Writing full posterior to {file}')
+ return write_posterior_coo_to_h5(output_file=file, **d)
+ except MemoryError:
+ logger.warning('Attempting to save the posterior as an h5 file '
+ 'resulted in an out-of-memory error. Please report '
+ 'this as a github issue.')
+ return False
+
+ def load(self, file: str) -> bool:
+ """Load a saved posterior in compressed array .npz format."""
+
+ d = load_posterior_from_h5(filename=file)
+ self._noise_count_posterior_coo = d['coo']
+ self._noise_count_posterior_coo_offsets = d['noise_count_offsets']
+ self._noise_count_posterior_kwargs = d['kwargs']
+ self._noise_count_regularized_posterior_coo = d['regularized_coo']
+ self._noise_count_regularized_posterior_kwargs = d['kwargs_regularized']
+ self._latents = d['latents']
+ logger.info(f'Loaded pre-computed posterior from {file}')
+ return True
+
+ def compute_denoised_counts(self,
+ estimator_constructor: EstimationMethod,
+ **kwargs) -> sp.csc_matrix:
+ """Probably the most important method: computation of the clean output count matrix.
+
+ Args:
+ estimator_constructor: A noise count estimator class derived from
+ the EstimationMethod base class, and implementing the
+ .estimate_noise() method, which creates a point estimate of
+ noise. Pass in the constructor, not an object.
+ **kwargs: Keyword arguments for estimator_constructor().estimate_noise()
+
+ Returns:
+ denoised_counts: Denoised output CSC sparse matrix (CSC for saving)
+
+ """
+
+ # Only compute using defaults if the cache is empty.
+ if self._noise_count_regularized_posterior_coo is not None:
+ # Priority is taken by a regularized posterior, since presumably
+ # the user computed it for a reason.
+ logger.debug('Using regularized posterior to compute denoised counts')
+ logger.debug(self._noise_count_regularized_posterior_kwargs)
+ posterior_coo = self._noise_count_regularized_posterior_coo
+ else:
+ # Use exact posterior if a regularized version is not computed.
+ posterior_coo = (self._noise_count_posterior_coo
+ if (self._noise_count_posterior_coo is not None)
+ else self.cell_noise_count_posterior_coo())
+
+ # Instantiate Estimator object.
+ estimator = estimator_constructor(index_converter=self.index_converter)
+
+ # Compute point estimate of noise in cells.
+ noise_csr = estimator.estimate_noise(
+ noise_log_prob_coo=posterior_coo,
+ noise_offsets=self._noise_count_posterior_coo_offsets,
+ **kwargs,
+ )
+
+ # Subtract cell noise from observed cell counts.
+ count_matrix = self.dataset_obj.data['matrix'] # all barcodes
+ cell_inds = self.dataset_obj.analyzed_barcode_inds[self.latents_map['p']
+ > consts.CELL_PROB_CUTOFF]
+ empty_inds = set(range(count_matrix.shape[0])) - set(cell_inds)
+ cell_counts = csr_set_rows_to_zero(csr=count_matrix, row_inds=empty_inds)
+ denoised_counts = cell_counts - noise_csr
+
+ return denoised_counts.tocsc()
+
+ def regularize_posterior(self,
+ regularization: 'PosteriorRegularization',
+ **kwargs) -> sp.coo_matrix:
+ """Do posterior regularization. This modifies self._noise_count_regularized_posterior_coo
+ in place, and returns it.
+
+ Args:
+ regularization: A particular PosteriorRegularization ['PRmu', 'PRq']
+ **kwargs: Arguments passed to the PosteriorRegularization's
+ .regularize() method
+
+ Returns:
+ Returns the regularized posterior, which is also stored in
+ self._noise_count_regularized_posterior_coo
+
+ """
+
+ # Check if this posterior regularization has already been computed.
+ currently_cached = False if self._noise_count_regularized_posterior_kwargs is None else True
+ if currently_cached:
+ # Check if it's the right thing.
+ for k, v in self._noise_count_regularized_posterior_kwargs.items():
+ if k == 'method':
+ if v != regularization.name():
+ currently_cached = False
+ break
+ elif k not in kwargs.keys():
+ currently_cached = False
+ break
+ elif kwargs[k] != v:
+ currently_cached = False
+ break
+ if currently_cached:
+ # What's been requested is what's cached.
+ logger.debug('Regularized posterior is already cached')
+ return self._noise_count_regularized_posterior_coo
+
+ # Compute the regularized posterior.
+ self._noise_count_regularized_posterior_coo = regularization.regularize(
+ noise_count_posterior_coo=self._noise_count_posterior_coo,
+ noise_offsets=self._noise_count_posterior_coo_offsets,
+ index_converter=self.index_converter,
+ **kwargs,
+ )
+ kwargs.update({'method': regularization.name()})
+ kwargs.pop('raw_count_matrix', None) # do not store a copy here
+ self._noise_count_regularized_posterior_kwargs = kwargs
+ logger.debug('Updated posterior after performing regularization')
+ return self._noise_count_regularized_posterior_coo
+
+ def clear_regularized_posterior(self):
+ """Remove the saved regularized posterior (so that compute_denoised_counts()
+ will not default to using it).
+ """
+ self._noise_count_regularized_posterior_coo = None
+ self._noise_count_regularized_posterior_kwargs = None
+
+ def cell_noise_count_posterior_coo(self, **kwargs) -> sp.coo_matrix:
+ """Compute the full-blown posterior on noise counts for all cells,
+ and store it in COO sparse format on CPU, and cache in
+ self._noise_count_posterior_csr
+
+ NOTE: This is the main entrypoint for this class.
+
+ Args:
+ **kwargs: Passed to _get_cell_noise_count_posterior_coo()
+
+ Returns:
+ self._noise_count_posterior_coo: This sparse COO object contains all
+ the information about the posterior noise count distribution,
+ but it is a bit complicated. The data per entry (m, c) are
+ stored in COO format. The rows "m" represent a combined
+ cell-and-gene index, with a one-to-one mapping from m to
+ (n, g). The columns "c" represent noise count values. Values
+ are the log probabilities of a noise count value. A smaller
+ matrix can be constructed by increasing the threshold
+ smallest_log_probability.
+ """
+
+ if ((self._noise_count_posterior_coo is None)
+ or (kwargs != self._noise_count_posterior_kwargs)):
+ logger.debug('Running _get_cell_noise_count_posterior_coo() to compute posterior')
+ self._get_cell_noise_count_posterior_coo(**kwargs)
+ self._noise_count_posterior_kwargs = kwargs
+
+ return self._noise_count_posterior_coo
+
+ @property
+ def latents_map(self) -> Dict[str, np.ndarray]:
+ if self._latents is None:
+ self._get_latents_map()
+ return self._latents
+
+ @torch.no_grad()
+ def _get_cell_noise_count_posterior_coo(
+ self,
+ n_samples: int = 20,
+ y_map: bool = True,
+ n_counts_max: int = 20,
+ smallest_log_probability: float = -10.) -> sp.coo_matrix: # TODO: default -7 ?
+ """Compute the full-blown posterior on noise counts for all cells,
+ and store log probability in COO sparse format on CPU.
+
+ Args:
+ n_samples: Number of samples to use to compute the posterior log
+ probability distribution. Samples have high variance, so it is
+ important to use at least 20. However, they are expensive.
+ y_map: Use the MAP value for y (cell / no cell) when sampling, to
+ avoid samples with a cell and samples without a cell.
+ n_counts_max: Maximum number of noise counts.
+ smallest_log_probability: Do not store log prob values smaller than
+ this -- they get set to zero (saves space)
+
+ Returns:
+ noise_count_posterior_coo: This sparse CSR object contains all
+ the information about the posterior noise count distribution,
+ but it is a bit complicated. The data per entry (m, c) are
+ stored in COO format. The rows "m" represent a combined
+ cell-and-gene index, and there is a one-to-one mapping from m to
+ (n, g). The columns "c" represent noise count values. Values
+ are the log probabilities of a noise count value. A smaller
+ matrix can be constructed by increasing the threshold
+ smallest_log_probability.
+
+ """
+
+ logger.debug('Computing full posterior noise counts')
+
+ # Compute posterior in mini-batches.
+ torch.cuda.empty_cache()
+
+ # Dataloader for cells only.
+ analyzed_bcs_only = True
+ count_matrix = self.dataset_obj.get_count_matrix() # analyzed barcodes
+ cell_logic = (self.latents_map['p'] > consts.CELL_PROB_CUTOFF)
+
+ # Raise an error if there are no cells found.
+ if cell_logic.sum() == 0:
+ logger.error(f'ERROR: Found zero droplets with posterior cell '
+ f'probability > {consts.CELL_PROB_CUTOFF}. Please '
+ f'check the log for estimated priors on expected cells, '
+ f'total droplets included, UMI counts per cell, and '
+ f'UMI counts in empty droplets, and see whether these '
+ f'values make sense. Consider using additional input '
+ f'arguments like --expected-cells, '
+ f'--total-droplets-included, --force-cell-umi-prior, '
+ f'and --force-empty-umi-prior, to make these values '
+ f'accurate for your dataset.')
+ raise RuntimeError('Zero cells found!')
+
+ dataloader_index_to_analyzed_bc_index = np.where(cell_logic)[0]
+ cell_data_loader = DataLoader(
+ count_matrix[cell_logic],
+ empty_drop_dataset=None,
+ batch_size=self.posterior_batch_size,
+ fraction_empties=0.,
+ shuffle=False,
+ use_cuda=self.use_cuda,
+ )
+
+ bcs = [] # barcode index
+ genes = [] # gene index
+ c = [] # noise count value
+ c_offset = [] # noise count offsets from zero
+ log_probs = []
+ ind = 0
+ n_minibatches = len(cell_data_loader)
+
+ logger.info('Computing posterior noise count probabilities in mini-batches.')
+
+ for i, data in enumerate(cell_data_loader):
+
+ if i == 0:
+ t = time.time()
+ elif i == 1:
+ logger.info(f' [{(time.time() - t) / 60:.2f} mins per chunk]')
+ logger.info(f'Working on chunk ({i + 1}/{n_minibatches})')
+
+ if self.debug:
+ logger.debug(f'Posterior minibatch starting with droplet {ind}')
+ logger.debug('\n' + get_hardware_usage(use_cuda=self.use_cuda))
+
+ # Compute noise count probabilities.
+ noise_log_pdf_NGC, noise_count_offset_NG = self.noise_log_pdf(
+ data=data,
+ n_samples=n_samples,
+ y_map=y_map,
+ n_counts_max=n_counts_max,
+ )
+
+ # Compute a tensor to indicate sparsity.
+ # First we want data = 0 to be all zeros
+ # We also want anything below the threshold to be a zero
+ tensor_for_nonzeros = noise_log_pdf_NGC.clone().exp() # probability
+ tensor_for_nonzeros.data[data == 0, :] = 0. # remove data = 0
+ tensor_for_nonzeros.data[noise_log_pdf_NGC < smallest_log_probability] = 0.
+
+ # Convert to sparse format using "m" indices.
+ bcs_i_chunk, genes_i_analyzed, c_i, log_prob_i = dense_to_sparse_op_torch(
+ noise_log_pdf_NGC,
+ tensor_for_nonzeros=tensor_for_nonzeros,
+ )
+
+ # Get the original gene index from gene index in the trimmed dataset.
+ genes_i = self.analyzed_gene_inds[genes_i_analyzed]
+
+ # Barcode index in the dataloader.
+ bcs_i = bcs_i_chunk + ind
+
+ # Obtain the real barcode index since we only use cells.
+ bcs_i = dataloader_index_to_analyzed_bc_index[bcs_i]
+
+ # Translate chunk barcode inds to overall inds.
+ if analyzed_bcs_only:
+ bcs_i = self.dataset_obj.analyzed_barcode_inds[bcs_i]
+ else:
+ bcs_i = self.barcode_inds[bcs_i]
+
+ # Add sparse matrix values to lists.
+ try:
+ bcs.extend(bcs_i.tolist())
+ genes.extend(genes_i.tolist())
+ c.extend(c_i.tolist())
+ log_probs.extend(log_prob_i.tolist())
+ c_offset.extend(noise_count_offset_NG[bcs_i_chunk, genes_i_analyzed]
+ .detach().cpu().numpy())
+ except TypeError as e:
+ # edge case of a single value
+ bcs.append(bcs_i)
+ genes.append(genes_i)
+ c.append(c_i)
+ log_probs.append(log_prob_i)
+ c_offset.append(noise_count_offset_NG[bcs_i_chunk, genes_i_analyzed]
+ .detach().cpu().numpy())
+
+ # Increment barcode index counter.
+ ind += data.shape[0] # Same as data_loader.batch_size
+
+ # Convert the lists to numpy arrays.
+ log_probs = np.array(log_probs, dtype=float)
+ c = np.array(c, dtype=np.uint32)
+ barcodes = np.array(bcs, dtype=np.uint64) # uint32 is too small!
+ genes = np.array(genes, dtype=np.uint64) # use same as above for IndexConverter
+ noise_count_offsets = np.array(c_offset, dtype=np.uint32)
+
+ # Translate (barcode, gene) inds to 'm' format index.
+ m = self.index_converter.get_m_indices(cell_inds=barcodes, gene_inds=genes)
+
+ # Put the counts into a sparse csr_matrix.
+ self._noise_count_posterior_coo = sp.coo_matrix(
+ (log_probs, (m, c)),
+ shape=[np.prod(self.count_matrix_shape), n_counts_max],
+ )
+ noise_offset_dict = dict(zip(m, noise_count_offsets))
+ nonzero_noise_offset_dict = {k: v for k, v in noise_offset_dict.items() if (v > 0)}
+ self._noise_count_posterior_coo_offsets = nonzero_noise_offset_dict
+ return self._noise_count_posterior_coo
+
+ @torch.no_grad()
+ def sample(self, data, lambda_multiplier=1., y_map: bool = False) -> torch.Tensor:
+ """Draw a single posterior sample for the count matrix conditioned on data
+
+ Args:
+ data: Count matrix (slice: some droplets, all genes)
+ lambda_multiplier: BasePosterior regularization multiplier
+ y_map: True to enforce the use of the MAP estimate of y, cell or
+ no cell. Useful in the case where many samples are collected,
+ since typically in those cases it is confusing to have samples
+ where a droplet is both cell-containing and empty.
+
+ Returns:
+ denoised_output_count_matrix: Single sample of the denoised output
+ count matrix, sampling all stochastic latent variables in the model.
+
+ """
+
+ # Sample all the latent variables in the model and get mu, lambda, alpha.
+ mu_sample, lambda_sample, alpha_sample = self.sample_mu_lambda_alpha(data, y_map=y_map)
+
+ # Compute the big tensor of log probabilities of possible c_{ng}^{noise} values.
+ log_prob_noise_counts_NGC, poisson_values_low_NG = self._log_prob_noise_count_tensor(
+ data=data,
+ mu_est=mu_sample + 1e-30,
+ lambda_est=lambda_sample * lambda_multiplier + 1e-30,
+ alpha_est=alpha_sample + 1e-30,
+ debug=self.debug,
+ )
+
+ # Use those probabilities to draw a sample of c_{ng}^{noise}
+ noise_count_increment_NG = dist.Categorical(logits=log_prob_noise_counts_NGC).sample()
+ noise_counts_NG = noise_count_increment_NG + poisson_values_low_NG
+
+ # Subtract from data to get the denoised output counts.
+ denoised_output_count_matrix = data - noise_counts_NG
+
+ return denoised_output_count_matrix
+
+ @torch.no_grad()
+ def map_denoised_counts_from_sampled_latents(self,
+ data,
+ n_samples: int,
+ lambda_multiplier: float = 1.,
+ y_map: bool = False) -> torch.Tensor:
+ """Draw posterior samples for all stochastic latent variables in the model
+ and use those values to compute a MAP estimate of the denoised count
+ matrix conditioned on data.
+
+ Args:
+ data: Count matrix (slice: some droplets, all genes)
+ lambda_multiplier: BasePosterior regularization multiplier
+ y_map: True to enforce the use of the MAP estimate of y, cell or
+ no cell. Useful in the case where many samples are collected,
+ since typically in those cases it is confusing to have samples
+ where a droplet is both cell-containing and empty.
+
+ Returns:
+ denoised_output_count_matrix: MAP estimate of the denoised output
+ count matrix, sampling all stochastic latent variables in the model.
+
+ """
+
+ noise_log_pdf, offset_noise_counts = self.noise_log_pdf(
+ data=data,
+ n_samples=n_samples,
+ lambda_multiplier=lambda_multiplier,
+ y_map=y_map,
+ )
+
+ noise_counts = torch.argmax(noise_log_pdf, dim=-1) + offset_noise_counts
+ denoised_output_count_matrix = torch.clamp(data - noise_counts, min=0.)
+
+ return denoised_output_count_matrix
+
+ @torch.no_grad()
+ def noise_log_pdf(self,
+ data,
+ n_samples: int = 1,
+ lambda_multiplier=1.,
+ y_map: bool = True,
+ n_counts_max: int = 50) -> Tuple[torch.Tensor, torch.Tensor]:
+ """Compute the posterior noise-count probability density function
+ using n_samples samples. This is a big matrix [n, g, c] where the last
+ dimension c is of variable size depending on the computation in
+ _log_prob_noise_count_tensor(), but is limited there to be no more than
+ 100. The c dimension represents an index to the number of noise counts
+ in [n, g]: specifically, the noise count once poisson_values_low_NG is added
+
+ Args:
+ data: Count matrix (slice: some droplets, all genes)
+ n_samples: Number of samples (of all stochastic latent variables in
+ the model) used to generate the CDF
+ lambda_multiplier: BasePosterior regularization multiplier
+ y_map: True to enforce the use of the MAP estimate of y, cell or
+ no cell. Useful in the case where many samples are collected,
+ since typically in those cases it is confusing to have samples
+ where a droplet is both cell-containing and empty.
+ n_counts_max: Size of count axis (need not start at zero noise
+ counts, but should be enough to cover the meat of the posterior)
+
+ Returns:
+ noise_log_pdf_NGC: Consensus noise count log_pdf (big tensor) from the samples.
+ noise_count_offset_NG: The offset for the noise count axis [n, g].
+
+ """
+
+ noise_log_pdf_NGC = None
+ noise_count_offset_NG = None
+
+ for s in range(1, n_samples + 1):
+
+ # Sample all the latent variables in the model and get mu, lambda, alpha.
+ mu_sample, lambda_sample, alpha_sample = self.sample_mu_lambda_alpha(data, y_map=y_map)
+
+ # Compute the big tensor of log probabilities of possible c_{ng}^{noise} values.
+ log_prob_noise_counts_NGC, noise_count_offset_NG = self._log_prob_noise_count_tensor(
+ data=data,
+ mu_est=mu_sample + 1e-30,
+ lambda_est=lambda_sample * lambda_multiplier + 1e-30,
+ alpha_est=alpha_sample + 1e-30,
+ n_counts_max=n_counts_max,
+ debug=self.debug,
+ )
+
+ # Normalize the PDFs (not necessarily normalized over the count range).
+ log_prob_noise_counts_NGC = (log_prob_noise_counts_NGC
+ - torch.logsumexp(log_prob_noise_counts_NGC,
+ dim=-1, keepdim=True))
+
+ # Add the probability from this sample to our running total.
+ # Update rule is
+ # log_prob_total_n = LAE [ log(1 - 1/n) + log_prob_total_{n-1}, log(1/n) + log_prob_sample ]
+ if s == 1:
+ noise_log_pdf_NGC = log_prob_noise_counts_NGC
+ else:
+ # This is a (normalized) running sum over samples in log-probability space.
+ noise_log_pdf_NGC = torch.logaddexp(
+ noise_log_pdf_NGC + torch.log(torch.tensor(1. - 1. / s).to(device=data.device)),
+ log_prob_noise_counts_NGC + torch.log(torch.tensor(1. / s).to(device=data.device)),
+ )
+
+ return noise_log_pdf_NGC, noise_count_offset_NG
+
+ @torch.no_grad()
+ def sample_mu_lambda_alpha(self,
+ data: torch.Tensor,
+ y_map: bool) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+ """Calculate a single sample estimate of mu, the mean of the true count
+ matrix, and lambda, the rate parameter of the Poisson background counts.
+
+ Args:
+ data: Dense tensor minibatch of cell by gene count data.
+ y_map: True to enforce the use of a MAP estimate of y rather than
+ sampling y. This prevents some samples from having a cell and
+ some not, which can lead to strange summary statistics over
+ many samples.
+
+ Returns:
+ mu_sample: Dense tensor sample of Negative Binomial mean for true
+ counts.
+ lambda_sample: Dense tensor sample of Poisson rate params for noise
+ counts.
+ alpha_sample: Dense tensor sample of Dirichlet concentration params
+ that inform the overdispersion of the Negative Binomial.
+
+ """
+
+ logger.debug('Replaying model with guide to sample mu, alpha, lambda')
+
+ # Use pyro poutine to trace the guide and sample parameter values.
+ guide_trace = pyro.poutine.trace(self.vi_model.guide).get_trace(x=data)
+
+ # If using MAP for y (so that you never get samples of cell and no cell),
+ # then intervene and replace a sampled y with the MAP
+ if y_map:
+ guide_trace.nodes['y']['value'] = (
+ guide_trace.nodes['p_passback']['value'] > 0
+ ).clone().detach()
+
+ replayed_model = pyro.poutine.replay(self.vi_model.model, guide_trace)
+
+ # Run the model using these sampled values.
+ replayed_model_output = replayed_model(x=data)
+
+ # The model returns mu, alpha, and lambda.
+ mu_sample = replayed_model_output['mu']
+ lambda_sample = replayed_model_output['lam']
+ alpha_sample = replayed_model_output['alpha']
+
+ return mu_sample, lambda_sample, alpha_sample
+
+ @staticmethod
+ @torch.no_grad()
+ def _log_prob_noise_count_tensor(data: torch.Tensor,
+ mu_est: torch.Tensor,
+ lambda_est: torch.Tensor,
+ alpha_est: Optional[torch.Tensor],
+ n_counts_max: int = 100,
+ debug: bool = False) -> Tuple[torch.Tensor, torch.Tensor]:
+ """Compute the log prob of noise counts [n, g, c] given mu, lambda, alpha, and the data.
+
+ NOTE: this is un-normalized log probability
+
+ Args:
+ data: Dense tensor minibatch of cell by gene count data.
+ mu_est: Dense tensor of Negative Binomial means for true counts.
+ lambda_est: Dense tensor of Poisson rate params for noise counts.
+ alpha_est: Dense tensor of Dirichlet concentration params that
+ inform the overdispersion of the Negative Binomial. None will
+ use an all-Poisson model
+ n_counts_max: Size of noise count dimension c
+ debug: True will go slow and check for NaNs and zero-probability entries
+
+ Returns:
+ log_prob_tensor: Probability of each noise count value.
+ poisson_values_low: The starting point for noise counts for each
+ cell and gene, because they can be different.
+
+ """
+
+ # Estimate a reasonable low-end to begin the Poisson summation.
+ n = min(n_counts_max, data.max().item()) # No need to exceed the max value
+ poisson_values_low = (lambda_est.detach() - n / 2).int()
+
+ poisson_values_low = torch.clamp(torch.min(poisson_values_low,
+ (data - n + 1).int()), min=0).float()
+
+ # Construct a big tensor of possible noise counts per cell per gene,
+ # shape (batch_cells, n_genes, max_noise_counts)
+ noise_count_tensor = torch.arange(start=0, end=n) \
+ .expand([data.shape[0], data.shape[1], -1]) \
+ .float().to(device=data.device)
+ noise_count_tensor = noise_count_tensor + poisson_values_low.unsqueeze(-1)
+
+ # Compute probabilities of each number of noise counts.
+ # NOTE: some values will be outside the support (negative values for NB).
+ # This results in NaNs.
+ if alpha_est is None:
+ # Poisson only model
+ log_prob_tensor = (dist.Poisson(lambda_est.unsqueeze(-1), validate_args=False)
+ .log_prob(noise_count_tensor)
+ + dist.Poisson(mu_est.unsqueeze(-1), validate_args=False)
+ .log_prob(data.unsqueeze(-1) - noise_count_tensor))
+ logger.debug('Using all poisson model (since alpha is not supplied to posterior)')
+ else:
+ logits = (mu_est.log() - alpha_est.log()).unsqueeze(-1)
+ log_prob_tensor = (dist.Poisson(lambda_est.unsqueeze(-1), validate_args=False)
+ .log_prob(noise_count_tensor)
+ + dist.NegativeBinomial(total_count=alpha_est.unsqueeze(-1),
+ logits=logits,
+ validate_args=False)
+ .log_prob(data.unsqueeze(-1) - noise_count_tensor))
+
+ # Set log_prob to -inf if noise > data.
+ neg_inf_tensor = torch.ones_like(log_prob_tensor) * -np.inf
+ log_prob_tensor = torch.where((noise_count_tensor <= data.unsqueeze(-1)),
+ log_prob_tensor,
+ neg_inf_tensor)
+
+ logger.debug(f'Prob computation with tensor of shape {log_prob_tensor.shape}')
+
+ if debug:
+ assert not torch.isnan(log_prob_tensor).any(), \
+ 'log_prob_tensor contains a NaN'
+ if torch.isinf(log_prob_tensor).all(dim=-1).any():
+ print(torch.where(torch.isinf(log_prob_tensor).all(dim=-1)))
+ raise AssertionError('There is at least one log_prob_tensor[n, g, :] '
+ 'that has all-zero probability')
+
+ return log_prob_tensor, poisson_values_low
+
+ @torch.no_grad()
+ def _get_latents_map(self):
+ """Calculate the encoded latent variables."""
+
+ logger.debug('Computing latent variables')
+
+ if self.vi_model is None:
+ self._latents = {'z': None, 'd': None, 'p': None, 'phi_loc_scale': None, 'epsilon': None}
+ return None
+
+ data_loader = self.dataset_obj.get_dataloader(use_cuda=self.use_cuda,
+ analyzed_bcs_only=True,
+ batch_size=500,
+ shuffle=False)
+
+ n_analyzed = data_loader.dataset.shape[0]
+
+ z = np.zeros((n_analyzed, self.vi_model.encoder['z'].output_dim))
+ d = np.zeros(n_analyzed)
+ p = np.zeros(n_analyzed)
+ epsilon = np.zeros(n_analyzed)
+
+ phi_loc = pyro.param('phi_loc')
+ phi_scale = pyro.param('phi_scale')
+ if 'chi_ambient' in pyro.get_param_store().keys():
+ chi_ambient = pyro.param('chi_ambient').detach()
+ else:
+ chi_ambient = None
+
+ start = 0
+ for i, data in enumerate(data_loader):
+
+ end = start + data.shape[0]
+
+ enc = self.vi_model.encoder(x=data,
+ chi_ambient=chi_ambient,
+ cell_prior_log=self.vi_model.d_cell_loc_prior)
+ z[start:end, :] = enc['z']['loc'].detach().cpu().numpy()
+
+ d[start:end] = \
+ dist.LogNormal(loc=enc['d_loc'],
+ scale=pyro.param('d_cell_scale')).mean.detach().cpu().numpy()
+
+ p[start:end] = enc['p_y'].sigmoid().detach().cpu().numpy()
+
+ epsilon[start:end] = \
+ dist.Gamma(enc['epsilon'] * self.vi_model.epsilon_prior,
+ self.vi_model.epsilon_prior).mean.detach().cpu().numpy()
+
+ start = end
+
+ self._latents = {'z': z,
+ 'd': d,
+ 'p': p,
+ 'phi_loc_scale': [phi_loc.item(), phi_scale.item()],
+ 'epsilon': epsilon}
+
+ @torch.no_grad()
+ def _get_mu_alpha_lambda_map(self,
+ data: torch.Tensor,
+ chi_ambient: torch.Tensor) -> Dict[str, torch.Tensor]:
+ """Calculate MAP estimates of mu, the mean of the true count matrix, and
+ lambda, the rate parameter of the Poisson background counts.
+
+ Args:
+ data: Dense tensor minibatch of cell by gene count data.
+ chi_ambient: Point estimate of inferred ambient gene expression.
+
+ Returns:
+ mu_map: Dense tensor of Negative Binomial means for true counts.
+ lambda_map: Dense tensor of Poisson rate params for noise counts.
+ alpha_map: Dense tensor of Dirichlet concentration params that
+ inform the overdispersion of the Negative Binomial.
+
+ """
+
+ logger.debug('Computing MAP esitmate of mu, lambda, alpha')
+
+ # Encode latents.
+ enc = self.vi_model.encoder(x=data,
+ chi_ambient=chi_ambient,
+ cell_prior_log=self.vi_model.d_cell_loc_prior)
+ z_map = enc['z']['loc']
+
+ chi_map = self.vi_model.decoder(z_map)
+ phi_loc = pyro.param('phi_loc')
+ phi_scale = pyro.param('phi_scale')
+ phi_conc = phi_loc.pow(2) / phi_scale.pow(2)
+ phi_rate = phi_loc / phi_scale.pow(2)
+ alpha_map = 1. / dist.Gamma(phi_conc, phi_rate).mean
+
+ y = (enc['p_y'] > 0).float()
+ d_empty = dist.LogNormal(loc=pyro.param('d_empty_loc'),
+ scale=pyro.param('d_empty_scale')).mean
+ d_cell = dist.LogNormal(loc=enc['d_loc'],
+ scale=pyro.param('d_cell_scale')).mean
+ epsilon = dist.Gamma(enc['epsilon'] * self.vi_model.epsilon_prior,
+ self.vi_model.epsilon_prior).mean
+
+ if self.vi_model.include_rho:
+ rho = pyro.param("rho_alpha") / (pyro.param("rho_alpha")
+ + pyro.param("rho_beta"))
+ else:
+ rho = None
+
+ # Calculate MAP estimates of mu and lambda.
+ mu_map = calculate_mu(
+ epsilon=epsilon,
+ d_cell=d_cell,
+ chi=chi_map,
+ y=y,
+ rho=rho,
+ )
+ lambda_map = calculate_lambda(
+ epsilon=epsilon,
+ chi_ambient=chi_ambient,
+ d_empty=d_empty,
+ y=y,
+ d_cell=d_cell,
+ rho=rho,
+ chi_bar=self.vi_model.avg_gene_expression,
+ )
+
+ return {'mu': mu_map, 'lam': lambda_map, 'alpha': alpha_map}
+
+
+class PosteriorRegularization(ABC):
+
+ def __init__(self):
+ super(PosteriorRegularization, self).__init__()
+
+ @staticmethod
+ @abstractmethod
+ def name():
+ """Short name of this regularization method"""
+ pass
+
+ @staticmethod
+ @abstractmethod
+ def regularize(noise_count_posterior_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ **kwargs) -> sp.coo_matrix:
+ """Perform posterior regularization"""
+ pass
+
+
+class PRq(PosteriorRegularization):
+ """Approximate noise CDF quantile targeting:
+
+ E_reg[noise_counts] >= E[noise_counts] + \alpha * Std[noise_counts]
+
+ """
+
+ @staticmethod
+ def name():
+ return 'PRq'
+
+ @staticmethod
+ def _log_mean_plus_alpha_std(log_prob: torch.Tensor, alpha: float):
+ c = torch.arange(log_prob.shape[1]).float().to(log_prob.device).unsqueeze(0)
+ prob = log_prob.exp()
+ mean = (c * prob).sum(dim=-1)
+ std = (((c - mean.unsqueeze(-1)).pow(2) * prob).sum(dim=-1)).sqrt()
+ return (mean + alpha * std).log()
+
+ @staticmethod
+ def _compute_log_target_dict(noise_count_posterior_coo: sp.coo_matrix,
+ alpha: float) -> Dict[int, float]:
+ """Given the noise count posterior, return log(mean + alpha * std)
+ for each 'm' index
+
+ NOTE: noise_log_pdf_BC should be normalized
+
+ Args:
+ noise_count_posterior_coo: The noise count posterior data structure
+ alpha: The tunable parameter of mean-targeting posterior
+ regularization. The output distribution has a mean which is
+ input_mean + alpha * input_std (if possible)
+
+ Returns:
+ log_mean_plus_alpha_std: Dict keyed by 'm', where values are
+ log(mean + alpha * std)
+
+ """
+ result = apply_function_dense_chunks(noise_log_prob_coo=noise_count_posterior_coo,
+ fun=PRq._log_mean_plus_alpha_std,
+ alpha=alpha)
+ return dict(zip(result['m'], result['result']))
+
+ @staticmethod
+ def _get_alpha_log_constraint_violation_given_beta(
+ beta_B: torch.Tensor,
+ log_pdf_noise_counts_BC: torch.Tensor,
+ noise_count_BC: torch.Tensor,
+ log_mu_plus_alpha_sigma_B: torch.Tensor) -> torch.Tensor:
+ r"""Returns log constraint violation for the regularized posterior of p(x), which
+ here is p(\omega) = p(x) e^{\beta x}, and we want
+ E[\omega] = E[x] + \alpha * Std[x] = log_mu_plus_alpha_sigma_B.exp()
+
+ NOTE: Binary search to find the root of this function can yield a value for beta_B.
+
+ Args:
+ beta_B: The parameter of the regularized posterior, with batch dimension
+ log_pdf_noise_counts_BC: The probability density of noise counts, with batch
+ and count dimensions
+ noise_count_BC: Noise counts, with batch and count dimensions
+ log_mu_plus_alpha_sigma_B: The constraint value to be satisfied, with batch dimension
+
+ Returns:
+ The amount by which the desired equality with log_mu_plus_alpha_sigma_B is violated,
+ with batch dimension
+
+ """
+
+ log_numerator_B = torch.logsumexp(
+ noise_count_BC.log() + log_pdf_noise_counts_BC + beta_B.unsqueeze(-1) * noise_count_BC,
+ dim=-1,
+ )
+ log_denominator_B = torch.logsumexp(
+ log_pdf_noise_counts_BC + beta_B.unsqueeze(-1) * noise_count_BC,
+ dim=-1,
+ )
+ return log_numerator_B - log_denominator_B - log_mu_plus_alpha_sigma_B
+
+ @staticmethod
+ def _chunked_compute_regularized_posterior(
+ noise_count_posterior_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ log_constraint_violation_fcn: Callable[[torch.Tensor, torch.Tensor,
+ torch.Tensor, torch.Tensor], torch.Tensor],
+ log_target_M: torch.Tensor,
+ target_tolerance: float = 0.001,
+ device: str = 'cpu',
+ n_chunks: Optional[int] = None,
+ ) -> sp.coo_matrix:
+ """Go through posterior in chunks and compute regularized posterior,
+ using the defined targets"""
+
+ # Compute using dense chunks, chunked on m-index.
+ if n_chunks is None:
+ dense_size_gb = (len(np.unique(noise_count_posterior_coo.row))
+ * (noise_count_posterior_coo.shape[1]
+ + np.array(list(noise_offsets.values())).max())) * 4 / 1e9 # GB
+ n_chunks = max(1, int(dense_size_gb // 1)) # approx 1 GB each
+
+ # Make the sparse matrix compact in the sense that it should use contiguous row values.
+ unique_rows, densifiable_coo_rows = np.unique(noise_count_posterior_coo.row, return_inverse=True)
+ densifiable_csr = sp.csr_matrix((noise_count_posterior_coo.data,
+ (densifiable_coo_rows, noise_count_posterior_coo.col)),
+ shape=[len(unique_rows), noise_count_posterior_coo.shape[1]])
+ chunk_size = int(np.ceil(densifiable_csr.shape[0] / n_chunks))
+
+ m = []
+ c = []
+ log_prob_reg = []
+
+ for i in range(n_chunks):
+ # B index here represents a batch: the re-defined m-index
+ log_pdf_noise_counts_BC = torch.tensor(
+ log_prob_sparse_to_dense(densifiable_csr[(i * chunk_size):((i + 1) * chunk_size)])
+ ).to(device)
+ noise_count_BC = (torch.arange(log_pdf_noise_counts_BC.shape[1])
+ .to(log_pdf_noise_counts_BC.device)
+ .unsqueeze(0)
+ .expand(log_pdf_noise_counts_BC.shape))
+ m_indices_for_chunk = unique_rows[(i * chunk_size):((i + 1) * chunk_size)]
+ noise_count_BC = noise_count_BC + (torch.tensor([noise_offsets.get(m, 0)
+ for m in m_indices_for_chunk],
+ dtype=torch.float)
+ .unsqueeze(-1)
+ .to(device))
+
+ # Parallel binary search for beta for each entry of count matrix
+ beta_B = torch_binary_search(
+ evaluate_outcome_given_value=lambda x:
+ log_constraint_violation_fcn(
+ beta_B=x,
+ log_pdf_noise_counts_BC=log_pdf_noise_counts_BC,
+ noise_count_BC=noise_count_BC,
+ log_mu_plus_alpha_sigma_B=log_target_M[(i * chunk_size):((i + 1) * chunk_size)],
+ ),
+ target_outcome=torch.zeros(noise_count_BC.shape[0]).to(device),
+ init_range=(torch.tensor([-100., 100.])
+ .to(device)
+ .unsqueeze(0)
+ .expand((noise_count_BC.shape[0],) + (2,))),
+ target_tolerance=target_tolerance,
+ max_iterations=100,
+ )
+
+ # Generate regularized posteriors.
+ log_pdf_reg_BC = log_pdf_noise_counts_BC + beta_B.unsqueeze(-1) * noise_count_BC
+ log_pdf_reg_BC = log_pdf_reg_BC - torch.logsumexp(log_pdf_reg_BC, -1, keepdims=True)
+
+ # Store sparse COO values in lists.
+ tensor_for_nonzeros = log_pdf_reg_BC.clone().exp() # probability
+ m_i, c_i, log_prob_reg_i = dense_to_sparse_op_torch(
+ log_pdf_reg_BC,
+ tensor_for_nonzeros=tensor_for_nonzeros,
+ )
+ m_i = np.array([m_indices_for_chunk[j] for j in m_i]) # chunk m to actual m
+
+ # Add sparse matrix values to lists.
+ try:
+ m.extend(m_i.tolist())
+ c.extend(c_i.tolist())
+ log_prob_reg.extend(log_prob_reg_i.tolist())
+ except TypeError as e:
+ # edge case of a single value
+ m.append(m_i)
+ c.append(c_i)
+ log_prob_reg.append(log_prob_reg_i)
+
+ reg_noise_count_posterior_coo = sp.coo_matrix((log_prob_reg, (m, c)),
+ shape=noise_count_posterior_coo.shape)
+ return reg_noise_count_posterior_coo
+
+ @staticmethod
+ @torch.no_grad()
+ def regularize(noise_count_posterior_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ alpha: float,
+ device: str = 'cuda',
+ target_tolerance: float = 0.001,
+ n_chunks: Optional[int] = None,
+ **kwargs) -> sp.coo_matrix:
+ """Perform posterior regularization using approximate quantile-targeting.
+
+ Args:
+ noise_count_posterior_coo: Noise count posterior log prob COO
+ noise_offsets: Offset noise counts per 'm' index
+ alpha: The tunable parameter of quantile-targeting posterior
+ regularization. The output distribution has a mean which is
+ input_mean + alpha * input_std (if possible)
+ device: Where to perform tensor operations: ['cuda', 'cpu']
+ target_tolerance: Tolerance when searching using binary search
+ n_chunks: For testing only - the number of chunks used to
+ compute the result when iterating over the posterior
+
+ Results:
+ reg_noise_count_posterior_coo: The regularized noise count
+ posterior data structure
+
+ """
+ logger.info(f'Regularizing noise count posterior using approximate quantile-targeting with alpha={alpha}')
+
+ # Compute the expectation for the mean post-regularization.
+ log_target_dict = PRq._compute_log_target_dict(
+ noise_count_posterior_coo=noise_count_posterior_coo,
+ alpha=alpha,
+ )
+ log_target_M = torch.tensor(list(log_target_dict.values())).to(device)
+
+ reg_noise_count_posterior_coo = PRq._chunked_compute_regularized_posterior(
+ noise_count_posterior_coo=noise_count_posterior_coo,
+ noise_offsets=noise_offsets,
+ log_target_M=log_target_M,
+ log_constraint_violation_fcn=PRq._get_alpha_log_constraint_violation_given_beta,
+ device=device,
+ target_tolerance=target_tolerance,
+ n_chunks=n_chunks,
+ )
+
+ return reg_noise_count_posterior_coo
+
+
+class PRmu(PosteriorRegularization):
+ r"""Approximate noise mean targeting:
+
+ Overall (default):
+ E_reg[\sum_{n} \sum_{g} noise_counts_{ng}] =
+ E[\sum_{n} \sum_{g} noise_counts_{ng}] + nFPR * \sum_{n} \sum_{g} raw_counts_{ng}
+
+ Per-gene:
+ E_reg[\sum_{n} noise_counts_{ng}] =
+ E[\sum_{n} noise_counts_{ng}] + nFPR * \sum_{n} raw_counts_{ng}
+
+ """
+
+ @staticmethod
+ def name():
+ return 'PRmu'
+
+ @staticmethod
+ def _binary_search_for_posterior_regularization_factor(
+ noise_count_posterior_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ index_converter: 'IndexConverter',
+ target_removal: torch.Tensor,
+ shape: int,
+ target_tolerance: float = 100,
+ max_iterations: int = 20,
+ device: str = 'cpu',
+ ) -> torch.Tensor:
+ """Go through posterior and compute regularization factor(s),
+ using the defined targets"""
+
+ def summarize_map_noise_counts(x: torch.Tensor,
+ per_gene: bool) -> torch.Tensor:
+ """Given a (subset of the) noise posterior, compute the MAP estimate
+ and summarize it either as the overall sum or per-gene.
+ """
+
+ # Regularize posterior.
+ regularized_noise_posterior_coo = PRmu._chunked_compute_regularized_posterior(
+ noise_count_posterior_coo=noise_count_posterior_coo,
+ noise_offsets=noise_offsets,
+ index_converter=index_converter,
+ beta=x,
+ device=device,
+ )
+
+ # Compute MAP.
+ estimator = MAP(index_converter=index_converter)
+ map_noise_csr = estimator.estimate_noise(
+ noise_log_prob_coo=regularized_noise_posterior_coo,
+ noise_offsets=noise_offsets,
+ device=device,
+ )
+
+ # Summarize removal.
+ if per_gene:
+ noise_counts = np.array(map_noise_csr.sum(axis=0)).squeeze()
+ else:
+ noise_counts = map_noise_csr.sum()
+
+ return torch.tensor(noise_counts).to(device)
+
+ # Perform binary search for beta.
+ per_gene = False
+ if target_removal.dim() > 0:
+ if len(target_removal) > 1:
+ per_gene = True
+
+ beta = torch_binary_search(
+ evaluate_outcome_given_value=lambda x:
+ summarize_map_noise_counts(x=x, per_gene=per_gene),
+ target_outcome=target_removal,
+ init_range=(torch.tensor([-100., 200.])
+ .to(device)
+ .unsqueeze(0)
+ .expand((shape,) + (2,))),
+ target_tolerance=target_tolerance,
+ max_iterations=max_iterations,
+ debug=True,
+ )
+ return beta
+
+ @staticmethod
+ def _chunked_compute_regularized_posterior(
+ noise_count_posterior_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ index_converter: 'IndexConverter',
+ beta: torch.Tensor,
+ device: str = 'cpu',
+ n_chunks: Optional[int] = None,
+ ) -> sp.coo_matrix:
+ """Go through posterior in chunks and compute regularized posterior,
+ using the defined targets"""
+
+ # Compute using dense chunks, chunked on m-index.
+ if n_chunks is None:
+ dense_size_gb = (len(np.unique(noise_count_posterior_coo.row))
+ * (noise_count_posterior_coo.shape[1]
+ + np.array(list(noise_offsets.values())).max())) * 4 / 1e9 # GB
+ n_chunks = max(1, int(dense_size_gb // 1)) # approx 1 GB each
+
+ # Make the sparse matrix compact in the sense that it should use contiguous row values.
+ unique_rows, densifiable_coo_rows = np.unique(noise_count_posterior_coo.row, return_inverse=True)
+ densifiable_csr = sp.csr_matrix((noise_count_posterior_coo.data,
+ (densifiable_coo_rows, noise_count_posterior_coo.col)),
+ shape=[len(unique_rows), noise_count_posterior_coo.shape[1]])
+ chunk_size = int(np.ceil(densifiable_csr.shape[0] / n_chunks))
+
+ m = []
+ c = []
+ log_prob_reg = []
+
+ for i in range(n_chunks):
+ # B index here represents a batch: the re-defined m-index
+ log_pdf_noise_counts_BC = torch.tensor(
+ log_prob_sparse_to_dense(densifiable_csr[(i * chunk_size):((i + 1) * chunk_size)])
+ ).to(device)
+ noise_count_BC = (torch.arange(log_pdf_noise_counts_BC.shape[1])
+ .to(log_pdf_noise_counts_BC.device)
+ .unsqueeze(0)
+ .expand(log_pdf_noise_counts_BC.shape))
+ m_indices_for_chunk = unique_rows[(i * chunk_size):((i + 1) * chunk_size)]
+ noise_count_BC = noise_count_BC + (torch.tensor([noise_offsets.get(m, 0)
+ for m in m_indices_for_chunk],
+ dtype=torch.float)
+ .unsqueeze(-1)
+ .to(device))
+
+ # Get beta for this chunk.
+ if len(beta) == 1:
+ # posterior regularization factor is a single scalar
+ beta_B = beta
+ else:
+ # per-gene mode
+ n, g = index_converter.get_ng_indices(m_inds=m_indices_for_chunk)
+ beta_B = torch.tensor([beta[gene] for gene in g])
+
+ # Generate regularized posteriors.
+ log_pdf_reg_BC = log_pdf_noise_counts_BC + beta_B.unsqueeze(-1) * noise_count_BC
+ log_pdf_reg_BC = log_pdf_reg_BC - torch.logsumexp(log_pdf_reg_BC, -1, keepdims=True)
+
+ # Store sparse COO values in lists.
+ tensor_for_nonzeros = log_pdf_reg_BC.clone().exp() # probability
+ # tensor_for_nonzeros.data[data == 0, :] = 0. # remove data = 0
+ m_i, c_i, log_prob_reg_i = dense_to_sparse_op_torch(
+ log_pdf_reg_BC,
+ tensor_for_nonzeros=tensor_for_nonzeros,
+ )
+ m_i = np.array([m_indices_for_chunk[j] for j in m_i]) # chunk m to actual m
+
+ # Add sparse matrix values to lists.
+ try:
+ m.extend(m_i.tolist())
+ c.extend(c_i.tolist())
+ log_prob_reg.extend(log_prob_reg_i.tolist())
+ except TypeError as e:
+ # edge case of a single value
+ m.append(m_i)
+ c.append(c_i)
+ log_prob_reg.append(log_prob_reg_i)
+
+ reg_noise_count_posterior_coo = sp.coo_matrix((log_prob_reg, (m, c)),
+ shape=noise_count_posterior_coo.shape)
+ return reg_noise_count_posterior_coo
+
+ @staticmethod
+ def _subset_posterior_by_cells(noise_count_posterior_coo: sp.coo_matrix,
+ index_converter: 'IndexConverter',
+ n_cells: int) -> sp.coo_matrix:
+ """Return a random slice of the full posterior with a specified number
+ of cells.
+
+ NOTE: Assumes that all the entries in noise_count_posterior_coo are for
+ cell-containing droplets, and not empty droplets.
+
+ Args:
+ noise_count_posterior_coo: The noise count posterior data structure
+ n_cells: The number of cells in the output subset
+
+ Returns:
+ subset_coo: Posterior for a random subset of cells, in COO format
+ """
+
+ # Choose cells that will be included.
+ m = noise_count_posterior_coo.row
+ n, g = index_converter.get_ng_indices(m_inds=m)
+ unique_cell_inds = np.unique(n)
+ if n_cells > len(unique_cell_inds):
+ logger.debug(f'Limiting n_cells during PRmu regularizer binary search to {unique_cell_inds}')
+ n_cells = len(unique_cell_inds)
+ chosen_n_values = set(np.random.choice(unique_cell_inds, size=n_cells, replace=False))
+ element_logic = [val in chosen_n_values for val in n]
+
+ # Subset the posterior.
+ data_subset = noise_count_posterior_coo.data[element_logic]
+ row_subset = noise_count_posterior_coo.row[element_logic]
+ col_subset = noise_count_posterior_coo.col[element_logic]
+ return sp.coo_matrix((data_subset, (row_subset, col_subset)),
+ shape=noise_count_posterior_coo.shape)
+
+ @staticmethod
+ @torch.no_grad()
+ def regularize(noise_count_posterior_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ index_converter: 'IndexConverter',
+ raw_count_matrix: sp.csr_matrix,
+ fpr: float,
+ per_gene: bool = False,
+ device: str = 'cuda',
+ target_tolerance: float = 0.5,
+ n_cells: int = 1000,
+ n_chunks: Optional[int] = None,
+ **kwargs) -> sp.coo_matrix:
+ """Perform posterior regularization using mean-targeting.
+
+ Args:
+ noise_count_posterior_coo: Noise count posterior log prob COO
+ noise_offsets: Offset noise counts per 'm' index
+ index_converter: IndexConverter object from 'm' to (n, g) and back
+ raw_count_matrix: The raw count matrix
+ fpr: The tunable parameter of mean-targeting posterior
+ regularization. The output, summed over cells, has a removed
+ gene count distribution similar to what would be expected from
+ the noise model, plus this nominal false positive rate.
+ per_gene: True to find one posterior regularization factor for each
+ gene, False to find one overall scalar (behavior of v0.2.0)
+ device: Where to perform tensor operations: ['cuda', 'cpu']
+ target_tolerance: Tolerance when searching using binary search.
+ In units of counts, so this really should not be less than 0.5
+ n_cells: To save time, use only this many cells to estimate removal
+ n_chunks: For testing only - the number of chunks used to
+ compute the result when iterating over the posterior
+
+ Results:
+ reg_noise_count_posterior_coo: The regularized noise count
+ posterior data structure
+
+ """
+
+ logger.info('Regularizing noise count posterior using mean-targeting')
+
+ # Use a subset of the data to find regularization factors, to reduce time.
+ logger.debug(f'Subsetting posterior to {n_cells} cells for this computation')
+ posterior_subset_coo = PRmu._subset_posterior_by_cells(
+ noise_count_posterior_coo=noise_count_posterior_coo,
+ index_converter=index_converter,
+ n_cells=n_cells,
+ )
+
+ # Compute target removal for MAP estimate using regularized posterior.
+ n, g = index_converter.get_ng_indices(m_inds=posterior_subset_coo.row)
+ included_cells = set(np.unique(n))
+ excluded_barcode_inds = set(range(raw_count_matrix.shape[0])) - included_cells
+ raw_count_csr_for_cells = csr_set_rows_to_zero(csr=raw_count_matrix,
+ row_inds=excluded_barcode_inds)
+ # print(raw_count_csr_for_cells)
+ logger.debug('Computing target removal')
+ target_fun = compute_mean_target_removal_as_function(
+ noise_count_posterior_coo=posterior_subset_coo,
+ noise_offsets=noise_offsets,
+ index_converter=index_converter,
+ raw_count_csr_for_cells=raw_count_csr_for_cells,
+ n_cells=len(included_cells),
+ device=device,
+ per_gene=per_gene,
+ )
+ target_removal = target_fun(fpr) * len(included_cells)
+ logger.debug(f'Target removal is {target_removal}')
+
+ # Find the posterior regularization factor(s).
+ if per_gene:
+ logger.debug('Computing optimal posterior regularization factors for each gene')
+ shape = index_converter.total_n_genes
+ else:
+ logger.debug('Computing optimal posterior regularization factor')
+ shape = 1
+ beta = PRmu._binary_search_for_posterior_regularization_factor(
+ noise_count_posterior_coo=posterior_subset_coo,
+ noise_offsets=noise_offsets,
+ index_converter=index_converter,
+ target_removal=target_removal,
+ device=device,
+ target_tolerance=target_tolerance,
+ shape=shape,
+ )
+ logger.debug(f'Optimal posterior regularization factor\n{beta}')
+
+ # Compute the posterior using the regularization factor(s).
+ logger.debug('Computing full regularized posterior')
+ regularized_noise_posterior_coo = PRmu._chunked_compute_regularized_posterior(
+ noise_count_posterior_coo=noise_count_posterior_coo,
+ noise_offsets=noise_offsets,
+ index_converter=index_converter,
+ beta=beta,
+ device=device,
+ )
+
+ return regularized_noise_posterior_coo
+
+
+class IndexConverter:
+
+ def __init__(self, total_n_cells: int, total_n_genes: int):
+ """Convert between (n, g) indices and flattened 'm' indices
+
+ Args:
+ total_n_cells: Total rows in the full sparse matrix
+ total_n_genes: Total columns in the full sparse matrix
+
+ """
+ self.total_n_cells = total_n_cells
+ self.total_n_genes = total_n_genes
+ self.matrix_shape = (total_n_cells, total_n_genes)
+
+ def __repr__(self):
+ return (f'IndexConverter with'
+ f'\n\ttotal_n_cells: {self.total_n_cells}'
+ f'\n\ttotal_n_genes: {self.total_n_genes}'
+ f'\n\tmatrix_shape: {self.matrix_shape}')
+
+ def get_m_indices(self, cell_inds: np.ndarray, gene_inds: np.ndarray) -> np.ndarray:
+ """Given arrays of cell indices and gene indices, suitable for a sparse matrix,
+ convert them to 'm' index values.
+ """
+ if not ((cell_inds >= 0) & (cell_inds < self.total_n_cells)).all():
+ raise ValueError(f'Requested cell_inds out of range: '
+ f'{cell_inds[(cell_inds < 0) | (cell_inds >= self.total_n_cells)]}')
+ if not ((gene_inds >= 0) & (gene_inds < self.total_n_genes)).all():
+ raise ValueError(f'Requested gene_inds out of range: '
+ f'{gene_inds[(gene_inds < 0) | (gene_inds >= self.total_n_genes)]}')
+ return cell_inds * self.total_n_genes + gene_inds
+
+ def get_ng_indices(self, m_inds: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+ """Given a list of 'm' index values, return two arrays: cell index values
+ and gene index values, suitable for a sparse matrix.
+ """
+ if not ((m_inds >= 0) & (m_inds < self.total_n_cells * self.total_n_genes)).all():
+ raise ValueError(f'Requested m_inds out of range: '
+ f'{m_inds[(m_inds < 0) | (m_inds >= self.total_n_cells * self.total_n_genes)]}')
+ return np.divmod(m_inds, self.total_n_genes)
+
+
+def compute_mean_target_removal_as_function(noise_count_posterior_coo: sp.coo_matrix,
+ noise_offsets: Dict[int, int],
+ index_converter: IndexConverter,
+ raw_count_csr_for_cells: sp.csr_matrix,
+ n_cells: int,
+ device: str,
+ per_gene: bool) -> Callable[[float], torch.Tensor]:
+ """Given the noise count posterior, return a function that computes target
+ removal (either overall or per-gene) as a function of FPR.
+
+ NOTE: computes the value "per cell", i.e. dividing
+ by the number of cells, so that total removal can be computed by
+ multiplying this by the number of cells in question.
+
+ Args:
+ noise_count_posterior_coo: Noise count posterior log prob COO
+ noise_offsets: Offset noise counts per 'm' index
+ index_converter: IndexConverter object from 'm' to (n, g) and back
+ raw_count_csr_for_cells: The input count matrix for only the cells
+ included in the posterior
+ n_cells: Number of cells included in the posterior, same number as in
+ raw_count_csr_for_cells
+ device: 'cpu' or 'cuda'
+ per_gene: True to come up with one target per gene
+
+ Returns:
+ target_removal_scaled_per_cell: Noise count removal target
+
+ """
+
+ # TODO: s1.h5 with FPR 0.99 only removes 50% of signal
+
+ # Compute the expected noise using mean summarization.
+ estimator = Mean(index_converter=index_converter)
+ mean_noise_csr = estimator.estimate_noise(
+ noise_log_prob_coo=noise_count_posterior_coo,
+ noise_offsets=noise_offsets,
+ device=device,
+ )
+ logger.debug(f'Total counts in raw matrix for cells = {raw_count_csr_for_cells.sum()}')
+ logger.debug(f'Total noise counts from mean noise estimator = {mean_noise_csr.sum()}')
+
+ # Compute the target removal.
+ approx_signal_csr = raw_count_csr_for_cells - mean_noise_csr
+ logger.debug(f'Approximate signal has total counts = {approx_signal_csr.sum()}')
+ logger.debug(f'Number of cells = {n_cells}')
+
+ def _target_fun(fpr: float) -> torch.Tensor:
+ """The function which gets returned"""
+ if per_gene:
+ target = np.array(mean_noise_csr.sum(axis=0)).squeeze()
+ target = target + fpr * np.array(approx_signal_csr.sum(axis=0)).squeeze()
+ else:
+ target = mean_noise_csr.sum()
+ target = target + fpr * approx_signal_csr.sum()
+
+ # Return target scaled to be per-cell.
+ return torch.tensor(target / n_cells).to(device)
+
+ return _target_fun
+
+
+@torch.no_grad()
+def torch_binary_search(
+ evaluate_outcome_given_value: Callable[[torch.Tensor], torch.Tensor],
+ target_outcome: torch.Tensor,
+ init_range: torch.Tensor,
+ target_tolerance: Optional[float] = 0.001,
+ max_iterations: int = consts.POSTERIOR_REG_SEARCH_MAX_ITER,
+ debug: bool = False,
+) -> torch.Tensor:
+ """Perform a binary search, given a target and an evaluation function.
+
+ NOTE: evaluate_outcome_given_value(value) should increase monotonically
+ with the input value. It is assumed that
+ consts.POSTERIOR_REG_MIN < output_value < consts.POSTERIOR_REG_MAX.
+ If this is not the case, the algorithm will produce an output close to one
+ of those endpoints, and target_tolerance will not be achieved.
+ Moreover, output_value must be positive (due to how we search for limits).
+
+ Args:
+ evaluate_outcome_given_value: Function that takes a value as its
+ input and produces the outcome, which is the target we are
+ trying to control. Should increase monotonically with value.
+ target_outcome: Desired outcome value from evaluate_outcome_given_value(value).
+ init_range: Search range, for each value.
+ target_tolerance: Tolerated error in the target value.
+ max_iterations: A cutoff to ensure termination. Even if a tolerable
+ solution is not found, the algorithm will stop after this many
+ iterations and return the best answer so far.
+ debug: Print debugging messages.
+
+ Returns:
+ value: Result of binary search. Same shape as init_value.
+
+ """
+
+ logger.debug('Binary search commencing')
+
+ assert (target_tolerance > 0), 'target_tolerance should be > 0.'
+ assert len(init_range.shape) > 1, 'init_range must be at least two-dimensional ' \
+ '(last dimension contains lower and upper bounds)'
+ assert init_range.shape[-1] == 2, 'Last dimension of init_range should be 2: low and high'
+
+ value_bracket = init_range.clone()
+
+ # Binary search algorithm.
+ for i in range(max_iterations):
+
+ logger.debug(f'Binary search limits [batch_dim=0, :]: '
+ f'{value_bracket.reshape(-1, value_bracket.shape[-1])[0, :]}')
+
+ # Current test value.
+ value = value_bracket.mean(dim=-1)
+
+ # Calculate an expected false positive rate for this lam_mult value.
+ outcome = evaluate_outcome_given_value(value)
+ residual = target_outcome - outcome
+
+ # Check on residual and update our bracket values.
+ stop_condition = (residual.abs() < target_tolerance).all()
+ if stop_condition:
+ break
+ else:
+ value_bracket[..., 0] = torch.where(outcome < target_outcome - target_tolerance,
+ value,
+ value_bracket[..., 0])
+ value_bracket[..., 1] = torch.where(outcome > target_outcome + target_tolerance,
+ value,
+ value_bracket[..., 1])
+
+ # If we stopped due to iteration limit, take the average value.
+ if i == max_iterations:
+ value = value_bracket.mean(dim=-1)
+ logger.warning(f'Binary search target not achieved in {max_iterations} attempts. '
+ f'Output is estimated to be {outcome.mean().item():.4f}')
+
+ # Warn if we railed out at the limits of the search
+ if debug:
+ if (value - target_tolerance <= init_range[..., 0]).sum() > 0:
+ logger.debug(f'{(value - target_tolerance <= init_range[..., 0]).sum()} '
+ f'entries in the binary search hit the lower limit')
+ logger.debug(value[value - target_tolerance <= init_range[..., 0]])
+ if (value + target_tolerance >= init_range[..., 1]).sum() > 0:
+ logger.debug(f'{(value + target_tolerance >= init_range[..., 1]).sum()} '
+ f'entries in the binary search hit the upper limit')
+ logger.debug(value[value + target_tolerance >= init_range[..., 1]])
+
+ return value
+
+
+def restore_from_checkpoint(tarball_name: str, input_file: str) \
+ -> Tuple['SingleCellRNACountsDataset', 'RemoveBackgroundPyroModel', Posterior]:
+ """Convenience function not used by the codebase"""
+
+ d = load_checkpoint(filebase=None, tarball_name=tarball_name)
+ d.update(load_from_checkpoint(filebase=None, tarball_name=tarball_name, to_load=['posterior']))
+ d['args'].input_file = input_file
+
+ dataset_obj = get_dataset_obj(args=d['args'])
+
+ posterior = Posterior(
+ dataset_obj=dataset_obj,
+ vi_model=d['model'],
+ )
+ posterior.load(file=d['posterior_file'])
+ return dataset_obj, d['model'], posterior
diff --git a/cellbender/remove_background/report.ipynb b/cellbender/remove_background/report.ipynb
new file mode 100644
index 00000000..b091c1b8
--- /dev/null
+++ b/cellbender/remove_background/report.ipynb
@@ -0,0 +1,133 @@
+{
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "# CellBender `remove-background` report"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "This output report from `cellbender remove-background` contains a summary of the run, including counts remaining, counts removed, further analyses, and any warnings or suggestions if the run seems to be abnormal.\n",
+ "\n",
+ "This HTML report is created from a jupyter notebook at \n",
+ "\n",
+ "`cellbender/cellbender/remove-background/report.ipynb`\n",
+ "\n",
+ "within the CellBender codebase. Feel free to run the notebook yourself and make any changes you see fit, or use it as a starting point for further analyses."
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "*The commentary in this report is generated using automated heuristics and best guesses based on hundreds of real datasets. If any of the automated commentary in this report seems incorrect for your dataset, please submit a question or an issue at our github repository https://github.com/broadinstitute/CellBender*"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "Cellarium Lab .. Methods Group .. Data Sciences Platform .. Broad Institute"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "-----------"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "from cellbender.remove_background.report import generate_summary_plots"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "## Input and output files"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "(Modify this section if you run this notebook yourself.)"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "# Allows us to get the filenames without needing to rewrite this notebook\n",
+ "\n",
+ "import os\n",
+ "input_file = os.environ['INPUT_FILE']\n",
+ "output_file = os.environ['OUTPUT_FILE']\n",
+ "\n",
+ "# For the case of a simulated dataset where we have the ground truth\n",
+ "try:\n",
+ " truth_file = os.environ['TRUTH_FILE']\n",
+ "except KeyError:\n",
+ " truth_file = None\n",
+ "\n",
+ "print(f'Input file: {input_file}')\n",
+ "print(f'Output file: {output_file}')\n",
+ "if truth_file is not None:\n",
+ " print(f'Simulated truth file: {truth_file}')"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "# Report"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "metadata": {
+ "scrolled": false
+ },
+ "outputs": [],
+ "source": [
+ "generate_summary_plots(input_file=input_file, \n",
+ " output_file=output_file,\n",
+ " truth_file=truth_file)"
+ ]
+ }
+ ],
+ "metadata": {
+ "kernelspec": {
+ "display_name": "Python 3",
+ "language": "python",
+ "name": "python3"
+ },
+ "language_info": {
+ "codemirror_mode": {
+ "name": "ipython",
+ "version": 3
+ },
+ "file_extension": ".py",
+ "mimetype": "text/x-python",
+ "name": "python",
+ "nbconvert_exporter": "python",
+ "pygments_lexer": "ipython3",
+ "version": "3.7.9"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
diff --git a/cellbender/remove_background/report.py b/cellbender/remove_background/report.py
new file mode 100644
index 00000000..f1a34cb9
--- /dev/null
+++ b/cellbender/remove_background/report.py
@@ -0,0 +1,1888 @@
+"""Functions for creation of an HTML report that plots and explains output."""
+
+from cellbender.remove_background.downstream import \
+ load_anndata_from_input, \
+ load_anndata_from_input_and_output, \
+ _load_anndata_from_input_and_decontx
+from cellbender.base_cli import get_version
+from cellbender.remove_background import consts
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+import scipy.sparse as sp
+import scipy.stats
+from IPython.display import display, Markdown, HTML
+
+import subprocess
+import datetime
+import os
+import logging
+from typing import Dict, Optional
+
+
+logger = logging.getLogger('cellbender')
+warnings = []
+TIMEOUT = 1200 # twenty minutes should always be way more than enough
+
+# counteract an error when I run locally
+# https://stackoverflow.com/questions/53014306/error-15-initializing-libiomp5-dylib-but-found-libiomp5-dylib-already-initial
+os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'
+
+run_notebook_str = lambda file: \
+ f'jupyter nbconvert ' \
+ f'--ExecutePreprocessor.timeout={TIMEOUT} ' \
+ f'--to notebook ' \
+ f'--allow-errors ' \
+ f'--execute {file}'
+to_html_str = lambda file, output: \
+ f'jupyter nbconvert ' \
+ f'--to html ' \
+ f'--TemplateExporter.exclude_input=True ' \
+ f'{file}'
+
+
+def _run_notebook(file):
+ subprocess.run(f'cp {file} tmp.report.ipynb', shell=True)
+ subprocess.run(run_notebook_str(file='tmp.report.ipynb'), shell=True)
+ subprocess.run(f'rm tmp.report.ipynb', shell=True)
+ return 'tmp.report.nbconvert.ipynb'
+
+
+def _to_html(file, output) -> str:
+ subprocess.run(to_html_str(file=file, output=output), shell=True)
+ subprocess.run(f'mv {file.replace(".ipynb", ".html")} {output}', shell=True)
+ subprocess.run(f'rm {file}', shell=True)
+ return output
+
+
+def _postprocess_html(file: str, title: str):
+ with open(file, mode='r') as f:
+ html = f.read()
+ html = html.replace('<title>tmp.report.nbconvert</title>',
+ f'<title>{title}</title>')
+ with open(file, mode='w') as f:
+ f.write(html)
+
+
+def run_notebook_make_html(file, output) -> str:
+ """Run Jupyter notebook to populate report and then convert to HTML.
+
+ Args:
+ file: Notebook file
+ output: Output file. Should end in ".html"
+
+ Returns:
+ output: Output file
+
+ """
+ assert output.endswith('.html'), 'Output HTML filename should end with .html'
+ html_file = _to_html(file=_run_notebook(file), output=output)
+ _postprocess_html(
+ file=html_file,
+ title=('CellBender: ' + os.path.basename(output).replace('_report.html', '')),
+ )
+ return html_file
+
+
+def generate_summary_plots(input_file: str,
+ output_file: str,
+ truth_file: Optional[Dict] = None,
+ dev_mode: bool = consts.EXTENDED_REPORT):
+ """Read in cellbender's output file and generate summary plots.
+
+ Args:
+ input_file: Raw CellRanger
+
+ """
+
+ global warnings
+ warnings = []
+
+ display(Markdown(f'### CellBender version {get_version()}'))
+ display(Markdown(str(datetime.datetime.now()).split('.')[0]))
+ display(Markdown(f'# {os.path.basename(output_file)}'))
+
+ # load datasets, before and after CellBender
+ input_layer_key = 'raw'
+ if os.path.isdir(output_file):
+ adata = _load_anndata_from_input_and_decontx(input_file=input_file,
+ output_mtx_directory=output_file,
+ input_layer_key=input_layer_key,
+ truth_file=truth_file)
+ out_key = 'decontx'
+ else:
+ adata = load_anndata_from_input_and_output(input_file=input_file,
+ output_file=output_file,
+ analyzed_barcodes_only=True,
+ input_layer_key=input_layer_key,
+ truth_file=truth_file)
+ out_key = 'cellbender'
+
+ # need to make any duplicate var indices unique (for pandas manipulations)
+ adata.var_names_make_unique()
+
+ display(Markdown('## Loaded dataset'))
+ print(adata)
+
+ # bit of pre-compute
+ cells = (adata.obs['cell_probability'] > consts.CELL_PROB_CUTOFF)
+ adata.var['n_removed'] = adata.var[f'n_{input_layer_key}'] - adata.var[f'n_{out_key}']
+ adata.var['fraction_removed'] = adata.var['n_removed'] / (adata.var[f'n_{input_layer_key}'] + 1e-5)
+ adata.var['fraction_remaining'] = adata.var[f'n_{out_key}'] / (adata.var[f'n_{input_layer_key}'] + 1e-5)
+ adata.var[f'n_{input_layer_key}_cells'] = np.array(adata.layers[input_layer_key][cells].sum(axis=0)).squeeze()
+ adata.var[f'n_{out_key}_cells'] = np.array(adata.layers[out_key][cells].sum(axis=0)).squeeze()
+ adata.var['n_removed_cells'] = (adata.var[f'n_{input_layer_key}_cells']
+ - adata.var[f'n_{out_key}_cells'])
+ adata.var['fraction_removed_cells'] = (adata.var['n_removed_cells']
+ / (adata.var[f'n_{input_layer_key}_cells'] + 1e-5))
+ adata.var['fraction_remaining_cells'] = (adata.var[f'n_{out_key}_cells']
+ / (adata.var[f'n_{input_layer_key}_cells'] + 1e-5))
+
+ # this inline command is necessary after cellbender imports
+ plt.rcParams.update({'font.size': 12})
+
+ # input UMI curve
+ raw_full_adata = plot_input_umi_curve(input_file)
+
+ # prove that remove-background is only subtracting counts, never adding
+ if out_key == 'cellbender':
+ assert (adata.layers[input_layer_key] < adata.layers[out_key]).sum() == 0, \
+ "There is an entry in the output greater than the input"
+ else:
+ if (adata.layers[input_layer_key] < adata.layers[out_key]).sum() == 0:
+ display(Markdown('WARNING: There is an entry in the output greater than the input'))
+
+ display(Markdown('## Examine how many counts were removed in total'))
+ try:
+ assess_overall_count_removal(adata, raw_full_adata=raw_full_adata, out_key=out_key)
+ except ValueError:
+ display(Markdown('Skipping assessment over overall count removal. Presumably '
+ 'this is due to including the whole dataset in '
+ '--total-droplets-included.'))
+
+ # plot learning curve
+ if out_key == 'cellbender':
+ try:
+ assess_learning_curve(adata)
+ except Exception:
+ pass
+ else:
+ display(Markdown('Skipping learning curve assessment.'))
+
+ # look at per-gene count removal
+ assess_count_removal_per_gene(adata, raw_full_adata=raw_full_adata, extended=dev_mode)
+
+ if dev_mode:
+ display(Markdown('## Histograms of counts per cell for several genes'))
+ plot_gene_removal_histograms(adata, out_layer_key=out_key)
+ display(Markdown('Typically we see that some of the low-count cells have '
+ 'their counts removed, since they were background noise.'))
+
+ # plot UMI curve and cell probabilities
+ if out_key == 'cellbender':
+ display(Markdown('## Cell probabilities'))
+ display(Markdown('The inferred posterior probability '
+ 'that each droplet is non-empty.'))
+ display(Markdown('*<span style="color:gray">We sometimes write "non-empty" '
+ 'instead of "cell" because dead cells and other cellular '
+ 'debris can still lead to a "non-empty" droplet, which will '
+ 'have a high posterior cell probability. But these '
+ 'kinds of low-quality droplets should be removed during '
+ 'cell QC to retain only high-quality cells for downstream '
+ 'analyses.</span>*'))
+ plot_counts_and_probs_per_cell(adata)
+ else:
+ display(Markdown('Skipping cell probability assessment.'))
+
+ # concordance of data before and after
+ display(Markdown('## Concordance of data before and after `remove-background`'))
+ plot_validation_plots(adata, output_layer_key=out_key, extended=dev_mode)
+
+ # PCA of gene expression
+ if out_key == 'cellbender':
+ display(Markdown('## PCA of encoded gene expression'))
+ plot_gene_expression_pca(adata, extended=dev_mode)
+ else:
+ display(Markdown('Skipping gene expression embedding assessment.'))
+
+ # "mixed species" plots
+ mixed_species_plots(adata, input_layer_key=input_layer_key, output_layer_key=out_key)
+
+ if dev_mode and (truth_file is not None):
+
+ # accuracy plots ==========================================
+
+ display(Markdown('# Comparison with truth data'))
+
+ display(Markdown('## Removal per gene'))
+
+ display(Markdown('Counts per gene are summed over cell-containing droplets'))
+ plt.figure(figsize=(10, 4))
+ plt.subplot(1, 2, 1)
+ plt.plot(adata.var['n_truth'].values, adata.var[f'n_{out_key}_cells'].values, '.', color='k')
+ plt.plot([0, adata.var['n_truth'].max()], [0, adata.var['n_truth'].max()],
+ color='lightgray', alpha=0.5)
+ plt.xlabel('True counts per gene')
+ plt.ylabel(f'{out_key} counts per gene')
+ plt.subplot(1, 2, 2)
+ logic = (adata.var['n_truth'].values > 0)
+ plt.plot(adata.var['n_truth'].values[logic],
+ ((adata.var[f'n_{out_key}_cells'].values[logic] - adata.var['n_truth'].values[logic])
+ / adata.var['n_truth'].values[logic]),
+ '.', color='k')
+ plt.plot([0, adata.var['n_truth'].max()], [0, 0],
+ color='lightgray', alpha=0.5)
+ plt.xlabel('True counts per gene')
+ plt.ylabel(f'{out_key}: residual count ratio per gene\n({out_key} - truth) / truth')
+ plt.tight_layout()
+ plt.show()
+
+ adata.var['fraction_remaining_cells_truth'] = (adata.var[f'n_truth']
+ / (adata.var[f'n_{input_layer_key}_cells'] + 1e-5))
+
+ plt.figure(figsize=(10, 4))
+ plt.subplot(1, 2, 1)
+ plt.semilogx(adata.var[f'n_{input_layer_key}_cells'],
+ adata.var['fraction_remaining_cells'], 'k.', ms=1)
+ plt.ylim([-0.05, 1.05])
+ plt.xlabel('Number of counts in raw data')
+ plt.ylabel('Fraction of counts remaining')
+ plt.title('Genes: removal of counts\nfrom (inferred) cell-containing droplets')
+ plt.subplot(1, 2, 2)
+ plt.semilogx(adata.var[f'n_{input_layer_key}_cells'],
+ adata.var['fraction_remaining_cells_truth'], 'k.', ms=1)
+ plt.ylim([-0.05, 1.05])
+ plt.xlabel('Number of counts in raw data')
+ plt.ylabel('Truth: fraction of counts remaining')
+ plt.title('Genes: truth')
+ plt.tight_layout()
+ plt.show()
+
+ plt.figure(figsize=(10, 4))
+ plt.subplot(1, 2, 1)
+ plt.semilogx(adata.var[f'n_{input_layer_key}_cells'],
+ adata.var['fraction_remaining_cells'] - adata.var['fraction_remaining_cells_truth'],
+ 'k.', ms=1)
+ plt.ylim([-1.05, 1.05])
+ plt.xlabel('Number of counts in raw data')
+ plt.ylabel('Residual fraction of counts remaining')
+ plt.title('Genes: residual')
+ plt.subplot(1, 2, 2)
+ plt.semilogx(adata.var[f'n_{input_layer_key}_cells'],
+ adata.var[f'n_{input_layer_key}_cells'] - adata.var['n_truth'],
+ 'k.', ms=1, label='raw')
+ plt.semilogx(adata.var[f'n_{input_layer_key}_cells'],
+ adata.var[f'n_{out_key}_cells'] - adata.var['n_truth'],
+ 'r.', ms=1, label=f'{out_key}')
+ plt.legend()
+ plt.xlabel('Number of counts in raw data')
+ plt.ylabel('Residual counts remaining')
+ plt.title('Genes: residual')
+ plt.tight_layout()
+ plt.show()
+
+ display(Markdown('## Revisiting histograms of counts per cell'))
+ display(Markdown('Now showing the truth in addition to the `remove-background` '
+ 'output.'))
+ plot_gene_removal_histograms(adata, plot_truth=True, out_layer_key=out_key)
+
+ # plot z, and comparisons of learned to true gene expression
+ if out_key == 'cellbender':
+ cluster_and_compare_expression_to_truth(adata=adata)
+ else:
+ display(Markdown('Skipping gene expression embedding assessment.'))
+
+ # gene expression as images: visualize changes
+ display(Markdown('## Visualization just for fun'))
+ display(Markdown('This is a strange but somewhat fun way to visualize what '
+ 'is going on with the data for each cell. We look at one '
+ 'cell at a time, and visualize gene expression as an '
+ 'image, where pixels are ordered by their true expression '
+ 'in the ambient RNA, where upper left is most expressed '
+ 'in ambient. We plot the output gene expression in '
+ 'blue/yellow, and then we look at three residuals (in red):'
+ '\n\n1. (raw - truth): what was supposed to be removed'
+ '\n2. (raw - posterior): what was actually removed'
+ '\n3. (truth - posterior): the residual, '
+ 'where red means too little was removed and blue means too '
+ 'much was removed.'))
+ try:
+ show_gene_expression_before_and_after(adata=adata, num=5)
+ except:
+ display(Markdown('WARNING: skipped showing gene expression as images, due to an error'))
+
+ if dev_mode:
+
+ # inference of latents
+ if out_key == 'cellbender':
+ compare_latents(adata)
+ else:
+ display(Markdown('Skipping gene expression embedding assessment.'))
+
+ if truth_file is not None:
+
+ # ROC curves
+ display(Markdown('## Quantification of performance'))
+ display(Markdown('Here we take a look at removal of noise counts from cell '
+ 'containing droplets only.'))
+ true_fpr = cell_roc_count_roc(
+ output_csr=adata.layers[out_key],
+ input_csr=adata.layers[input_layer_key],
+ truth_csr=adata.layers['truth'],
+ cell_calls=(adata.obs['cell_probability'] > 0.5),
+ truth_cell_labels=adata.obs['truth_cell_label'],
+ )
+ if type(adata.uns['target_false_positive_rate'][0]) == np.float64:
+ if true_fpr > adata.uns['target_false_positive_rate'][0]:
+ warnings.append('FPR exceeds target FPR.')
+ display(Markdown(f'WARNING: FPR of {true_fpr:.4f} exceeds target FPR of '
+ f'{adata.uns["target_false_positive_rate"]}. Keep '
+ f'in mind however that the target FPR is meant to '
+ f'target false positives over and above some '
+ f'basal level (dataset dependent), so the '
+ f'measured FPR should exceed the target by some '
+ f'amount.'))
+
+ display(Markdown('# Summary of warnings:'))
+ if len(warnings) == 0:
+ display(Markdown('None.'))
+ else:
+ for warning in warnings:
+ display(Markdown(warning))
+
+
+def plot_input_umi_curve(inputfile):
+ adata = load_anndata_from_input(inputfile)
+ plt.loglog(sorted(np.array(adata.X.sum(axis=1)).squeeze(), reverse=True))
+ plt.xlabel('Ranked Barcode ID')
+ plt.ylabel('UMI counts')
+ plt.title(f'UMI curve\nRaw input data: {os.path.basename(inputfile)}')
+ plt.show()
+ return adata
+
+
+def assess_overall_count_removal(adata, raw_full_adata, input_layer_key='raw', out_key='cellbender'):
+ global warnings
+ cells = (adata.obs['cell_probability'] > 0.5)
+ initial_counts = adata.layers[input_layer_key][cells].sum()
+ removed_counts = initial_counts - adata.layers[out_key][cells].sum()
+ removed_percentage = removed_counts / initial_counts * 100
+ print(f'removed {removed_counts:.0f} counts from non-empty droplets')
+ print(f'removed {removed_percentage:.2f}% of the counts in non-empty droplets')
+
+ from scipy.stats import norm
+
+ log_counts = np.log10(np.array(adata.layers[input_layer_key].sum(axis=1)).squeeze())
+ empty_log_counts = np.array(raw_full_adata.X[
+ [bc not in adata.obs_names
+ for bc in raw_full_adata.obs_names]
+ ].sum(axis=1)).squeeze()
+ empty_log_counts = np.log10(empty_log_counts[empty_log_counts > consts.LOW_UMI_CUTOFF])
+ bins = np.linspace(empty_log_counts.min(), log_counts.max(), 100)
+ # binwidth = bins[1] - bins[0]
+
+ # def plot_normal_fit(x, loc, scale, n, label):
+ # plt.plot(x, n * binwidth * norm.pdf(x=x, loc=loc, scale=scale), label=label)
+
+ plt.hist(empty_log_counts.tolist() + log_counts[~cells].tolist(),
+ histtype='step', label='empty droplets', bins=bins)
+ plt.hist(log_counts[cells], histtype='step', label='non-empty droplets', bins=bins)
+ xx = np.linspace(plt.gca().get_xlim()[0], plt.gca().get_xlim()[-1], 100)
+ # if 'cell_size' in adata.obs.keys():
+ # plt.hist(np.log10(adata.obs['cell_size'][cells]),
+ # histtype='step', label='inferred cell sizes', bins=bins)
+ # plot_normal_fit(x=xx,
+ # loc=np.log10(adata.obs['cell_size'][cells]).mean(),
+ # scale=np.log10(adata.obs['cell_size'][cells]).std(),
+ # n=cells.sum(),
+ # label='inferred cell sizes')
+ # if (('empty_droplet_size_lognormal_loc' in adata.uns.keys())
+ # and ('empty_droplet_size_lognormal_scale' in adata.uns.keys())):
+ # plot_normal_fit(x=xx,
+ # loc=np.log10(np.exp(adata.uns['empty_droplet_size_lognormal_loc'])),
+ # scale=np.log10(np.exp(adata.uns['empty_droplet_size_lognormal_scale'])),
+ # n=(~cells).sum() + len(empty_log_counts),
+ # label='inferred empty sizes')
+ plt.ylabel('Number of droplets')
+ plt.legend(loc='center left', bbox_to_anchor=(1, 0.5))
+ plt.ylim(bottom=1)
+ plt.yscale('log')
+ # x-axis log10 to regular number
+ plt.xticks(plt.gca().get_xticks(),
+ [f'{n:.0f}' for n in np.power(10, plt.gca().get_xticks())], rotation=90)
+ plt.xlim(left=empty_log_counts.min())
+ plt.xlabel('UMI counts per droplet')
+ plt.show()
+
+ estimated_ambient_per_droplet = np.exp(adata.uns['empty_droplet_size_lognormal_loc']).item()
+ expected_fraction_removed_from_cells = estimated_ambient_per_droplet * cells.sum() / initial_counts
+
+ fpr = adata.uns['target_false_positive_rate'].item() # this is an np.ndarray with one element
+ cohort_mode = False
+ if type(fpr) != float:
+ cohort_mode = True
+
+ print('Rough estimate of expectations based on nothing but the plot above:')
+ print(f'roughly {estimated_ambient_per_droplet * cells.sum():.0f} noise counts '
+ f'should be in non-empty droplets')
+ print(f'that is approximately {expected_fraction_removed_from_cells * 100:.2f}% of '
+ f'the counts in non-empty droplets')
+
+ if not cohort_mode:
+ expected_percentage = (expected_fraction_removed_from_cells + fpr) * 100
+ print(f'with a false positive rate [FPR] of {fpr * 100}%, we would expect to remove about '
+ f'{expected_percentage:.2f}% of the counts in non-empty droplets')
+ else:
+ expected_percentage = expected_fraction_removed_from_cells * 100
+ print(f'ran in cohort mode, so no false positive rate [FPR] target was set, '
+ f'but we would still expect to remove about '
+ f'{expected_percentage:.2f}% of the counts in non-empty droplets')
+
+ display(Markdown('\n'))
+
+ if np.abs(expected_percentage - removed_percentage) <= 0.5:
+ display(Markdown('It looks like the algorithm did a great job meeting that expectation.'))
+ elif np.abs(expected_percentage - removed_percentage) <= 1:
+ display(Markdown('It looks like the algorithm did a decent job meeting that expectation.'))
+ elif np.abs(expected_percentage - removed_percentage) <= 5:
+ if removed_percentage < expected_percentage:
+ display(Markdown('The algorithm removed a bit less than naive expectations '
+ 'would indicate, but this is likely okay. If removal '
+ 'seems insufficient, the FPR can be increased.'))
+ else:
+ display(Markdown('The algorithm removed a bit more than naive expectations '
+ 'would indicate, but this is likely okay. Spot-check '
+ 'removal of a few top-removed genes as a QC measure. '
+ 'If less removal is desired, decrease the FPR.'))
+ elif removed_percentage - expected_percentage > 5:
+ display(Markdown('The algorithm seems to have removed more overall counts '
+ 'than would be naively expected.'))
+ warnings.append('Algorithm removed more counts overall than naive expectations.', )
+ elif expected_percentage - removed_percentage > 5:
+ display(Markdown('The algorithm seems to have removed fewer overall counts '
+ 'than would be naively expected.'))
+ warnings.append('Algorithm removed fewer counts overall than naive expectations.', )
+
+
+def assess_learning_curve(adata,
+ spike_size: float = 0.5,
+ deviation_size: float = 0.25,
+ monotonicity_cutoff: float = 0.1):
+ global warnings
+ display(Markdown('## Assessing convergence of the algorithm'))
+ plot_learning_curve(adata)
+ if 'learning_curve_train_elbo' not in adata.uns.keys():
+ return
+ display(Markdown(
+ '*<span style="color:gray">The learning curve tells us about the progress of the algorithm in '
+ 'inferring all the latent variables in our model. We want to see '
+ 'the ELBO increasing as training epochs increase. Generally it is '
+ 'desirable for the ELBO to converge at some high plateau, and be fairly '
+ 'stable.</span>*'))
+ display(Markdown(
+ '*<span style="color:gray">What to watch out for:</span>*'
+ ))
+ display(Markdown(
+ '*<span style="color:gray">1. large downward spikes in the ELBO (of value more than a few hundred)</span>*\n'
+ '*<span style="color:gray">2. the test ELBO can be smaller than the train ELBO, but generally we '
+ 'want to see both curves increasing and reaching a stable plateau. We '
+ 'do not want the test ELBO to dip way back down at the end.</span>*\n'
+ '*<span style="color:gray">3. lack of convergence, where it looks like the ELBO would change '
+ 'quite a bit if training went on for more epochs.</span>*'
+ ))
+
+ if adata.uns['learning_curve_train_epoch'][-1] < 50:
+ display(Markdown('Short run. Will not analyze the learning curve.'))
+ warnings.append(f'Short run of only {adata.uns["learning_curve_train_epoch"][-1]} epochs')
+ return
+
+ train_elbo_min_max = np.percentile(adata.uns['learning_curve_train_elbo'], q=[5, 95])
+ train_elbo_range = train_elbo_min_max.max() - train_elbo_min_max.min()
+
+ # look only from epoch 45 onward for spikes in train ELBO
+ large_spikes_in_train = np.any((adata.uns['learning_curve_train_elbo'][46:]
+ - adata.uns['learning_curve_train_elbo'][45:-1])
+ < -train_elbo_range * spike_size)
+
+ second_half_train_elbo = (adata.uns['learning_curve_train_elbo']
+ [(len(adata.uns['learning_curve_train_elbo']) // 2):])
+ large_deviation_in_train = np.any(second_half_train_elbo
+ < np.median(second_half_train_elbo)
+ - train_elbo_range * deviation_size)
+
+ half = len(adata.uns['learning_curve_train_elbo']) // 2
+ threequarter = len(adata.uns['learning_curve_train_elbo']) * 3 // 4
+ typical_end_variation = np.std(adata.uns['learning_curve_train_elbo'][half:threequarter])
+ low_end_in_train = (adata.uns['learning_curve_train_elbo'][-1]
+ < adata.uns['learning_curve_train_elbo'].max() - 5 * typical_end_variation)
+
+ # look only from epoch 45 onward for spikes in train ELBO
+ non_monotonicity = ((adata.uns['learning_curve_train_elbo'][46:]
+ - adata.uns['learning_curve_train_elbo'][45:-1])
+ < -3 * typical_end_variation).sum() / len(adata.uns['learning_curve_train_elbo'])
+ non_monotonic = (non_monotonicity > monotonicity_cutoff)
+
+ def windowed_cumsum(x, n=20):
+ return np.array([np.cumsum(x[i:(i + n)])[-1] for i in range(len(x) - n)])
+
+ windowsize = 20
+ tracking_trace = windowed_cumsum(adata.uns['learning_curve_train_elbo'][1:]
+ - adata.uns['learning_curve_train_elbo'][:-1],
+ n=windowsize)
+ big_dip = -1 * (adata.uns['learning_curve_train_elbo'][-1]
+ - adata.uns['learning_curve_train_elbo'][5]) / 10
+ backtracking = (tracking_trace.min() < big_dip)
+ backtracking_ind = np.argmin(tracking_trace) + windowsize
+
+ halftest = len(adata.uns['learning_curve_test_elbo']) // 2
+ threequartertest = len(adata.uns['learning_curve_test_elbo']) * 3 // 4
+ typical_end_variation_test = np.std(adata.uns['learning_curve_test_elbo'][halftest:threequartertest])
+ runaway_test = (adata.uns['learning_curve_test_elbo'][-1]
+ < adata.uns['learning_curve_test_elbo'].max() - 4 * typical_end_variation_test)
+
+ non_convergence = (np.mean([adata.uns['learning_curve_train_elbo'][-1]
+ - adata.uns['learning_curve_train_elbo'][-2],
+ adata.uns['learning_curve_train_elbo'][-2]
+ - adata.uns['learning_curve_train_elbo'][-3]])
+ > 2 * typical_end_variation)
+
+ display(Markdown('**Automated assessment** --------'))
+ if large_spikes_in_train:
+ warnings.append('Large spikes in training ELBO.')
+ display(Markdown('- *WARNING*: Large spikes detected in the training ELBO.'))
+ if large_deviation_in_train:
+ warnings.append('Large deviation in training ELBO from max value late in learning.')
+ display(Markdown('- *WARNING*: The training ELBO deviates quite a bit from '
+ 'the max value during the second half of training.'))
+ if low_end_in_train:
+ warnings.append('Large deviation in training ELBO from max value at end.')
+ display(Markdown('- The training ELBO deviates quite a bit from '
+ 'the max value at the last epoch.'))
+ if non_monotonic:
+ warnings.append('Non-monotonic training ELBO.')
+ display(Markdown('- We typically expect to see the training ELBO increase almost '
+ 'monotonically. This curve seems to have a lot more downward '
+ 'motion than we like to see.'))
+ if backtracking:
+ warnings.append('Back-tracking in training ELBO.')
+ display(Markdown('- We typically expect to see the training ELBO increase almost '
+ 'monotonically. This curve seems to have a concerted '
+ f'period of motion in the wrong direction near epoch {backtracking_ind}. '
+ f'If this is early in training, this is probably okay.'))
+ if runaway_test:
+ warnings.append('Final test ELBO is much lower than the max test ELBO.')
+ display(Markdown('- We hope to see the test ELBO follow the training ELBO, '
+ 'increasing almost monotonically (though there will be '
+ 'deviations, and that is expected). There may be a large '
+ 'gap, and that is okay. However, this curve '
+ 'ends with a low test ELBO compared to the max test ELBO '
+ 'value during training. The output could be suboptimal.'))
+ if non_convergence:
+ warnings.append('Non-convergence of training ELBO.')
+ display(Markdown('- We typically expect to see the training ELBO come to a '
+ 'stable plateau value near the end of training. Here '
+ 'the training ELBO is still moving quite a bit.'))
+
+ display(Markdown('**Summary**:'))
+ if large_spikes_in_train or large_deviation_in_train:
+ display(Markdown('This is unusual behavior, and a reduced --learning-rate '
+ 'is indicated. Re-run with half the current learning '
+ 'rate and compare the results.'))
+ elif low_end_in_train or non_monotonic or runaway_test:
+ display(Markdown('This is slightly unusual behavior, and a reduced '
+ '--learning-rate might be indicated. Consider re-running '
+ 'with half the current learning rate to compare the results.'))
+ elif non_convergence:
+ display(Markdown('This is slightly unusual behavior, and more training '
+ '--epochs might be indicated. Consider re-running '
+ 'for more epochs to compare the results.'))
+ else:
+ display(Markdown('This learning curve looks normal.'))
+
+
+def plot_learning_curve(adata):
+
+ if 'learning_curve_train_elbo' not in adata.uns.keys():
+ print('No learning curve recorded!')
+ return
+
+ def _mkplot():
+ plt.plot(adata.uns['learning_curve_train_epoch'],
+ adata.uns['learning_curve_train_elbo'], label='train')
+ try:
+ plt.plot(adata.uns['learning_curve_test_epoch'],
+ adata.uns['learning_curve_test_elbo'], '.:', label='test')
+ plt.legend()
+ except Exception:
+ pass
+ plt.title('Learning curve')
+ plt.ylabel('ELBO')
+ plt.xlabel('Epoch')
+
+ if len(adata.uns['learning_curve_train_elbo']) > 20:
+
+ # two panels: zoom on the right-hand side
+ plt.figure(figsize=(10, 4))
+ plt.subplot(1, 2, 1)
+ _mkplot()
+ plt.subplot(1, 2, 2)
+ _mkplot()
+ plt.title('Learning curve (zoomed in)')
+ low = np.percentile(adata.uns['learning_curve_train_elbo'], q=10)
+ if (len(adata.uns['learning_curve_train_elbo']) > 0) \
+ and (len(adata.uns['learning_curve_test_elbo']) > 0):
+ high = max(adata.uns['learning_curve_train_elbo'].max(),
+ adata.uns['learning_curve_test_elbo'].max())
+ else:
+ high = adata.uns['learning_curve_train_elbo'].max()
+ plt.ylim([low, high + (high - low) / 10])
+ plt.tight_layout()
+ plt.show()
+
+ else:
+ _mkplot()
+ plt.show()
+
+
+def assess_count_removal_per_gene(adata,
+ raw_full_adata,
+ input_layer_key='raw',
+ r_squared_cutoff=0.5,
+ extended=True):
+
+ global warnings
+ display(Markdown('## Examine count removal per gene'))
+
+ # how well does it correlate with our expectation about the ambient RNA profile?
+ cells = (adata.obs['cell_probability'] > 0.5)
+ counts = np.array(raw_full_adata.X.sum(axis=1)).squeeze()
+ clims = [adata.obs[f'n_{input_layer_key}'][~cells].mean() / 2,
+ np.percentile(adata.obs[f'n_{input_layer_key}'][cells].values, q=2)]
+ # if all are called "cells" then clims[0] will be a nan
+ if np.isnan(clims[0]):
+ clims[0] = counts.min()
+ if 'approximate_ambient_profile' in adata.uns.keys():
+ approximate_ambient_profile = adata.uns['approximate_ambient_profile']
+ else:
+ empty_count_matrix = raw_full_adata[(counts > clims[0]) & (counts < clims[1])].X
+ if empty_count_matrix.shape[0] > 100:
+ approximate_ambient_profile = np.array(raw_full_adata[(counts > clims[0])
+ & (counts < clims[1])].X.mean(axis=0)).squeeze()
+ else:
+ # a very rare edge case I've seen once
+ display(Markdown('Having some trouble finding the empty droplets via heuristics. '
+ 'The "approximate background estimated from empty droplets" may be inaccurate.'))
+ approximate_ambient_profile = np.array(raw_full_adata[counts < clims[1]].X.mean(axis=0)).squeeze()
+ approximate_ambient_profile = approximate_ambient_profile / approximate_ambient_profile.sum()
+ y = adata.var['n_removed'] / adata.var['n_removed'].sum()
+ maxval = (approximate_ambient_profile / approximate_ambient_profile.sum()).max()
+
+ def _plot_identity(maxval):
+ plt.plot([0, maxval], [0, maxval], 'lightgray')
+
+ if extended:
+
+ plt.figure(figsize=(12, 4))
+ plt.subplot(1, 2, 1)
+ plt.plot(approximate_ambient_profile, adata.var['ambient_expression'], '.', ms=2)
+ _plot_identity(maxval)
+ plt.xlabel('Approximate background per gene\nestimated from empty droplets')
+ plt.ylabel('Inferred ambient profile')
+ plt.title('Genes: inferred ambient')
+
+ plt.subplot(1, 2, 2)
+ plt.plot(approximate_ambient_profile, y, '.', ms=2)
+ _plot_identity(maxval)
+ plt.xlabel('Approximate background per gene\nestimated from empty droplets')
+ plt.ylabel('Removal per gene')
+ plt.title('Genes: removal')
+ plt.tight_layout()
+ plt.show()
+
+ else:
+
+ plt.plot(approximate_ambient_profile, y, '.', ms=2)
+ _plot_identity(maxval)
+ plt.xlabel('Approximate background per gene\nestimated from empty droplets')
+ plt.ylabel('Removal per gene')
+ plt.title('Genes: removal')
+ plt.tight_layout()
+ plt.show()
+
+ cutoff = 1e-6
+ logic = np.logical_not((approximate_ambient_profile < cutoff) | (y < cutoff))
+ r_squared_result = scipy.stats.pearsonr(np.log(approximate_ambient_profile[logic]),
+ np.log(y[logic]))
+ if hasattr(r_squared_result, 'statistic'):
+ # scipy version 1.9.0+
+ r_squared = r_squared_result.statistic
+ else:
+ r_squared = r_squared_result[0]
+ display(Markdown(f'Pearson correlation coefficient for the above is {r_squared:.4f}'))
+ if r_squared > r_squared_cutoff:
+ display(Markdown('This meets expectations.'))
+ else:
+ warnings.append('Per-gene removal does not closely match a naive estimate '
+ 'of ambient RNA from empty droplets. Does it look like '
+ 'CellBender correctly identified the empty droplets?')
+ display(Markdown('WARNING: This deviates from expectations, and may '
+ 'indicate that the run did not go well'))
+
+ percentile = 90
+ genecount_lowlim = int(np.percentile(adata.var[f'n_{input_layer_key}'], q=percentile))
+ display(Markdown('### Table of top genes removed\n\nRanked by fraction removed, '
+ f'and excluding genes with fewer than {genecount_lowlim} '
+ f'total raw counts ({percentile}th percentile)'))
+ df = adata.var[adata.var['cellbender_analyzed']] # exclude omitted features
+ df = df[[c for c in df.columns if (c != 'features_analyzed_inds')]]
+ display(HTML(df[df[f'n_{input_layer_key}'] > genecount_lowlim]
+ .sort_values(by='fraction_removed', ascending=False).head(10).to_html()))
+
+ for g in adata.var[(adata.var[f'n_{input_layer_key}_cells'] > genecount_lowlim)
+ & (adata.var['fraction_removed'] > 0.8)].index:
+ warnings.append(f'Expression of gene {g} decreases quite a bit')
+ display(Markdown(f'**WARNING**: The expression of the highly-expressed '
+ f'gene {g} decreases quite markedly after CellBender. '
+ f'Check to ensure this makes sense!'))
+
+ if extended:
+
+ plt.figure(figsize=(12, 4))
+ plt.subplot(1, 2, 1)
+ plt.semilogx(adata.var[f'n_{input_layer_key}'],
+ adata.var['fraction_remaining'], 'k.', ms=1)
+ plt.ylim([-0.05, 1.05])
+ plt.xlabel('Number of counts in raw data')
+ plt.ylabel('Fraction of counts remaining')
+ plt.title('Genes: removal of counts\nfrom the entire dataset')
+ plt.subplot(1, 2, 2)
+ plt.semilogx(adata.var[f'n_{input_layer_key}_cells'],
+ adata.var['fraction_remaining_cells'], 'k.', ms=1)
+ plt.ylim([-0.05, 1.05])
+ plt.xlabel('Number of counts in raw data')
+ plt.ylabel('Fraction of counts remaining')
+ plt.title('Genes: removal of counts\nfrom (inferred) cell-containing droplets')
+ plt.show()
+
+
+def plot_counts_and_probs_per_cell(adata, input_layer_key='raw'):
+
+ limit_to_features_analyzed = True
+
+ if limit_to_features_analyzed:
+ var_logic = adata.var['cellbender_analyzed']
+ else:
+ var_logic = ...
+
+ in_counts = np.array(adata.layers[input_layer_key][:, var_logic].sum(axis=1)).squeeze()
+ # cellbender_counts = np.array(adata.layers['cellbender'][:, var_logic].sum(axis=1)).squeeze()
+ order = np.argsort(in_counts)[::-1]
+ # plt.semilogy(cellbender_counts[order], '.:', ms=3, color='lightgray', alpha=0.5, label='cellbender')
+ plt.semilogy(in_counts[order], 'k-', lw=1, label=input_layer_key)
+ plt.xlabel('Sorted barcode ID')
+ plt.ylabel('Unique UMI counts' + ('\n(for features analyzed by CellBender)'
+ if limit_to_features_analyzed else ''))
+ plt.legend(loc='lower left', title='UMI counts')
+ plt.gca().twinx()
+ plt.plot(adata.obs['cell_probability'][order].values, '.', ms=2, alpha=0.2, color='red')
+ plt.ylabel('Inferred cell probability', color='red')
+ plt.yticks([0, 0.25, 0.5, 0.75, 1.0], color='red')
+ plt.ylim([-0.05, 1.05])
+ plt.show()
+
+
+def plot_validation_plots(adata, input_layer_key='raw',
+ output_layer_key='cellbender',
+ extended=True):
+
+ display(Markdown('*<span style="color:gray">The intent is to change the input '
+ 'data as little as possible while achieving noise removal. '
+ 'These plots show general summary statistics about similarity '
+ 'of the input and output data. We expect to see the data '
+ 'lying close to a straight line (gray). There may be '
+ 'outlier genes/features, which are often those highest-'
+ 'expressed in the ambient RNA.</span>*'))
+ display(Markdown('The plots here show data '
+ 'for inferred cell-containing droplets, and exclude the '
+ 'empty droplets.'))
+
+ cells = (adata.obs['cell_probability'] > 0.5)
+
+ # counts per barcode
+ plt.figure(figsize=(9, 4))
+ plt.subplot(1, 2, 1)
+ plt.loglog(adata.obs[f'n_{input_layer_key}'][cells].values,
+ adata.obs[f'n_{output_layer_key}'][cells].values,
+ '.', ms=3, alpha=0.8, rasterized=True)
+ minmax = [adata.obs[f'n_{input_layer_key}'][cells].min() / 10,
+ adata.obs[f'n_{input_layer_key}'][cells].max() * 10]
+ plt.loglog(minmax, minmax, lw=1, color='gray', alpha=0.5)
+ plt.xlabel('Input counts per barcode')
+ plt.ylabel(f'{output_layer_key} counts per barcode')
+ plt.title('Droplet count concordance\nin inferred cell-containing droplets')
+ plt.axis('equal')
+
+ # counts per gene
+ plt.subplot(1, 2, 2)
+ plt.loglog(adata.var[f'n_{input_layer_key}_cells'],
+ adata.var[f'n_{output_layer_key}_cells'],
+ '.', ms=3, alpha=0.8, rasterized=True)
+ minmax = [adata.var[f'n_{input_layer_key}_cells'].min() / 10,
+ adata.var[f'n_{input_layer_key}_cells'].max() * 10]
+ plt.loglog(minmax, minmax, lw=1, color='gray', alpha=0.5)
+ plt.xlabel('Input counts per gene')
+ plt.ylabel(f'{output_layer_key} counts per gene')
+ plt.title('Gene count concordance\nin inferred cell-containing droplets')
+ plt.axis('equal')
+ plt.tight_layout()
+ plt.show()
+
+ cells = (adata.obs['cell_probability'] >= 0.5)
+
+ if extended:
+
+ # Fano factor per barcode
+ plt.figure(figsize=(9, 4))
+ plt.subplot(1, 2, 1)
+ plt.loglog(fano(adata.layers[input_layer_key][cells], axis=1),
+ fano(adata.layers[output_layer_key][cells], axis=1), '.', ms=1)
+ plt.loglog([1e0, 1e3], [1e0, 1e3], lw=1, color='gray', alpha=0.5)
+ plt.xlabel('Input Fano factor per droplet')
+ plt.ylabel(f'{output_layer_key} Fano factor per droplet')
+ plt.title('Droplet count variance\nin inferred cell-containing droplets')
+ plt.axis('equal')
+
+ # Fano factor per gene
+ plt.subplot(1, 2, 2)
+ plt.loglog(fano(adata.layers[input_layer_key][cells], axis=0),
+ fano(adata.layers[output_layer_key][cells], axis=0), '.', ms=1)
+ plt.loglog([1e0, 1e3], [1e0, 1e3], lw=1, color='gray', alpha=0.5)
+ plt.xlabel('Input Fano factor per gene')
+ plt.ylabel(f'{output_layer_key} Fano factor per gene')
+ plt.title('Gene count variance\nin inferred cell-containing droplets')
+ plt.axis('equal')
+ plt.tight_layout()
+ plt.show()
+
+ # histogram of per-cell cosine distances
+ # cells_in_data = latents['barcodes_analyzed_inds'][latents['cell_probability'] >= 0.5]
+ cosine_dist = []
+ for bc in np.random.permutation(np.where(cells)[0])[:300]:
+ cosine_dist.append(cosine(np.array(adata.layers[input_layer_key][bc, :].todense()).squeeze(),
+ np.array(adata.layers[output_layer_key][bc, :].todense()).squeeze()))
+ cosine_dist = np.array(cosine_dist)
+ plt.hist(cosine_dist, bins=100)
+ plt.xlabel('Cosine distance between cell before and after')
+ plt.ylabel('Number of cells')
+ plt.title('Per-cell expression changes')
+ plt.show()
+
+ print(f'cosine_dist.mean = {cosine_dist.mean():.4f}')
+
+ display(Markdown('We want this cosine distance to be as small as it can be '
+ '(though it cannot be zero, since we are removing counts). '
+ 'There is no specific threshold value above which we are '
+ 'concerned, but typically we see values below 0.05.'))
+
+
+def plot_gene_removal_histograms(adata, input_layer_key='raw', plot_truth=False, out_layer_key='cellbender'):
+ order_gene = np.argsort(np.array(adata.var[f'n_{input_layer_key}']))[::-1]
+ bins = np.arange(50) - 0.5
+
+ plt.figure(figsize=(14, 6))
+ for i, g_ind in enumerate([0, 5, 10, 20, 100, 1000]):
+ if g_ind >= len(order_gene):
+ break
+ plt.subplot(2, 3, i + 1)
+ plt.hist(np.array(adata.layers[input_layer_key][:, order_gene[g_ind]].todense()).squeeze(),
+ bins=bins, log=True, label=input_layer_key, histtype='step')
+ plt.hist(np.array(adata.layers[out_layer_key][:, order_gene[g_ind]].todense()).squeeze(),
+ bins=bins, log=True, label=out_layer_key, alpha=0.75, histtype='step')
+ if plot_truth:
+ plt.hist(np.array(adata.layers['truth'][:, order_gene[g_ind]].todense()).squeeze(),
+ bins=bins, log=True, label='truth', alpha=0.5, histtype='step')
+ plt.xlabel('counts per cell', fontsize=12)
+ plt.title(f'{adata.var_names[order_gene[g_ind]]}: rank {g_ind} in ambient', fontsize=12)
+ plt.ylabel('number of cells', fontsize=12)
+ plt.legend()
+ plt.tight_layout()
+ plt.show()
+
+
+def show_gene_expression_before_and_after(adata,
+ input_layer_key: str = 'raw',
+ num: int = 10):
+ """Display gene expression as an image.
+ Show what was removed.
+ Show what should have been removed.
+ Show residual.
+ """
+
+ inds = np.where(adata.obs['cell_probability'] > 0.5)[0][:num]
+
+ # ambient sort order
+ order = np.argsort(adata.var['truth_ambient_expression'])[::-1]
+
+ for i in inds:
+
+ raw = np.array(adata.layers[input_layer_key][i, :].todense(), dtype=float).squeeze()
+
+ # size of images
+ nz_inds = np.where(raw[order] > 0)[0]
+ nrows = int(np.floor(np.sqrt(nz_inds.size)).item())
+ imsize = (nrows, int(np.ceil(nz_inds.size / max(1, nrows)).item()))
+
+ raw = np.resize(raw[order][nz_inds], imsize[0] * imsize[1]).reshape(imsize)
+ overflow = (imsize[0] * imsize[1]) - nz_inds.size + 1
+ raw[-1, -overflow:] = 0.
+ post = np.array(adata.layers['cellbender'][i, :].todense(), dtype=float).squeeze()
+ post = np.resize(post[order][nz_inds], imsize[0] * imsize[1]).reshape(imsize)
+ post[-1, -overflow:] = 0.
+
+ true_ambient = np.resize(adata.var['truth_ambient_expression'][order][nz_inds],
+ imsize[0] * imsize[1]).reshape(imsize)
+ true_ambient[-1, -overflow:] = 0.
+
+ true = np.array(adata.layers['truth'][i, :].todense(), dtype=float).squeeze()
+ true = np.resize(true[order][nz_inds], imsize[0] * imsize[1]).reshape(imsize)
+ true[-1, -overflow:] = 0.
+ lim = max(np.log1p(post).max(), np.log1p(true).max(), np.log1p(raw).max())
+
+ plt.figure(figsize=(14, 3))
+
+ # plt.subplot(1, 4, 1)
+ # plt.imshow(np.log1p(raw), vmin=1, vmax=lim)
+ # plt.title(f'log raw [{np.expm1(lim):.1f}]')
+ # plt.xticks([])
+ # plt.yticks([])
+
+ plt.subplot(1, 4, 1)
+ plt.imshow(np.log1p(post), vmin=1, vmax=lim)
+ plt.title(f'log posterior [{np.expm1(lim):.1f}]')
+ plt.xticks([])
+ plt.yticks([])
+
+ plt.subplot(1, 4, 2)
+ dat = (raw - true)
+ minmax = max(-1 * dat.min(), dat.max())
+ plt.imshow(dat, cmap='seismic', vmin=-1 * minmax, vmax=minmax)
+ plt.title(f'raw - truth [{minmax:.1f}]')
+ plt.xticks([])
+ plt.yticks([])
+
+ plt.subplot(1, 4, 3)
+ dat = (raw - post)
+ # minmax = max(-1*dat.min(), dat.max())
+ cmap = plt.get_cmap('seismic', 2 * minmax + 1)
+ plt.imshow(dat, cmap=cmap, vmin=-1 * minmax, vmax=minmax)
+ plt.title(f'raw - posterior [{minmax}]')
+ plt.xticks([])
+ plt.yticks([])
+
+ plt.subplot(1, 4, 4)
+ dat = (true - post)
+ thisminmax = max(-1 * dat.min(), dat.max())
+ minmax = max(thisminmax, minmax)
+ plt.imshow(-dat, cmap='seismic', vmin=-1 * minmax, vmax=minmax)
+ if minmax == dat.max():
+ plt.title(f'truth - posterior [- {minmax:.1f}]')
+ else:
+ plt.title(f'truth - posterior [{minmax:.1f}]')
+ plt.xticks([])
+ plt.yticks([])
+
+ plt.show()
+
+ ambient_counts = raw - true
+ removed_counts = raw - post
+ removed_ambient = np.sum(np.minimum(removed_counts[ambient_counts > 0], ambient_counts[ambient_counts > 0]))
+
+ print(f'Fraction of ambient counts removed: {removed_ambient / np.sum(ambient_counts):.2f}')
+ print(f'Ambient counts removed: {int(removed_ambient)}')
+ print(f'Remaining ambient counts: {- int(np.sum(dat[dat < 0]))}')
+ print(f'Erroneously subtracted real counts: {int(np.sum(dat[dat > 0]))}')
+
+
+def plot_gene_expression_pca(adata, key='cellbender_embedding',
+ input_layer_key='raw', extended=True):
+
+ cells = (adata.obs['cell_probability'] > 0.5)
+ adata.obsm['X_pca'] = pca_2d(adata.obsm[key]).detach().numpy()
+
+ # plot z PCA colored by latent size
+ sizeorder = np.argsort(adata.obs['cell_size'][cells])
+ s = plt.scatter(x=adata.obsm['X_pca'][:, 0][cells][sizeorder],
+ y=adata.obsm['X_pca'][:, 1][cells][sizeorder],
+ c=np.log10(adata.obs['cell_size'][cells][sizeorder]),
+ s=2,
+ cmap='brg',
+ alpha=0.5)
+ plt.title('Gene expression embedding\ncolored by log10 inferred cell size $d_n$')
+ plt.xlabel('PCA 0')
+ plt.ylabel('PCA 1')
+ plt.xticks([])
+ plt.yticks([])
+ plt.gcf().colorbar(s, pad=0.1, label='log10 cell size')
+ plt.show()
+
+ display(Markdown('*<span style="color:gray">We are not looking for anything '
+ 'specific in the PCA plot '
+ 'of the gene expression embedding, but often we see clusters '
+ 'that correspond to different cell types. If you see only '
+ 'a single large blob, then the dataset might contain only '
+ 'one cell type, or perhaps there are few counts per '
+ 'droplet.</span>*'))
+
+ if extended:
+
+ # plot z PCA colored by a few features
+ def _pca_color(g: int, layer):
+ outcounts = np.array(adata.layers['cellbender'][:, adata.var_names == g].todense()).squeeze()
+ rawcounts = np.array(adata.layers[input_layer_key][:, adata.var_names == g].todense()).squeeze()
+ # cmax = 2 * (rawcounts - outcounts)[rawcounts > 0].mean()
+ cmax = np.percentile(rawcounts[rawcounts > 0], q=80)
+ if layer == 'cellbender':
+ counts = outcounts
+ else:
+ counts = rawcounts
+ order = np.argsort(counts[cells])
+
+ s = plt.scatter(x=adata.obsm['X_pca'][:, 0][cells][order],
+ y=adata.obsm['X_pca'][:, 1][cells][order],
+ c=counts[cells][order],
+ s=10,
+ vmin=0,
+ vmax=cmax, # min(20, max(1, cmax)),
+ cmap='Oranges',
+ alpha=0.25)
+ plt.title(f'{g}: {layer}')
+ plt.xlabel('PCA 0')
+ plt.ylabel('PCA 1')
+ plt.gcf().colorbar(s, pad=0.05, label='Counts (truncated)')
+
+ percentile = 90
+ genecount_lowlim = int(np.percentile(adata.var[f'n_{input_layer_key}'], q=percentile))
+ if 'feature_type' in adata.var.keys():
+ feature_logic = (adata.var['feature_type'] == 'Gene Expression')
+ else:
+ feature_logic = True
+ features = (adata.var[feature_logic
+ & (adata.var['n_cellbender'] > genecount_lowlim)]
+ .sort_values(by='fraction_removed', ascending=False)
+ .groupby('genome')
+ .head(2)
+ .index
+ .values
+ .tolist())
+ if 'feature_type' in adata.var.keys():
+ if (adata.var['feature_type'] != 'Gene Expression').sum() > 0:
+ features.extend(adata.var[(adata.var['feature_type'] != 'Gene Expression')
+ & (adata.var['n_cellbender'] > genecount_lowlim)]
+ .sort_values(by='fraction_removed', ascending=False)
+ .groupby('feature_type')
+ .head(2)
+ .index
+ .values
+ .tolist())
+
+ display(Markdown('### Visualization of a few features'))
+ display(Markdown('Focusing on a few top features which were removed the most.'))
+
+ for g in features:
+ plt.figure(figsize=(11, 4))
+ plt.subplot(1, 2, 1)
+ _pca_color(g, layer=input_layer_key)
+ plt.subplot(1, 2, 2)
+ _pca_color(g, layer='cellbender')
+ plt.tight_layout()
+ plt.show()
+
+ display(Markdown('*<span style="color:gray">We typically see selective '
+ 'removal of some genes from '
+ 'particular cell types. The genes above have been picked '
+ 'randomly based on fraction_removed, and so might not be '
+ 'the best genes for visualization. These sorts of plots '
+ 'can be an interesting way to visualize what '
+ '`remove-background` does.</span>*'))
+
+
+def cluster_cells(adata, embedding_key='cellbender_embedding', n=2):
+ """Run PCA (if not done) and use spectral clustering to label cells.
+ Returns:
+ cluster: np.ndarray of cluster labels as ints
+ """
+ from sklearn.cluster import SpectralClustering
+
+ cells = (adata.obs['cell_probability'] > 0.5)
+
+ if 'X_pca' not in adata.obsm.keys():
+ from sklearn.decomposition import PCA
+ z = adata.obsm[embedding_key]
+ adata.obsm['X_pca'] = PCA(n_components=20).fit_transform(z)
+
+ if 'truth_cell_probability' in adata.obs.keys():
+ # if truth exists, use it
+ n_cell_labels = np.array([k.startswith('truth_gene_expression_cell_label_')
+ for k in adata.var.keys()]).sum()
+ else:
+ n_cell_labels = n
+ spec = (SpectralClustering(n_clusters=n_cell_labels, random_state=0)
+ .fit_predict(adata.obsm[embedding_key][cells, :]))
+ cluster = np.zeros(adata.shape[0])
+ cluster[cells] = spec + 1 # offset by 1 so that empty is 0
+ return cluster.astype(int)
+
+
+def cluster_and_compare_expression_to_truth(adata, embedding_key='cellbender_embedding'):
+
+ cluster = cluster_cells(adata, embedding_key=embedding_key)
+ adata.obs['cluster'] = cluster
+ cells = (adata.obs['cell_probability'] > 0.5)
+
+ display(Markdown('## Gene expression embedding, $z_n$'))
+ display(Markdown('Here we find cluster labels *de novo* using spectral clustering, '
+ 'in order to compare learned profiles with the truth. There '
+ 'is inherently an issue of non-identifiabiity, where we cannot '
+ 'know the order of the true cluster labels'))
+
+ # plot z PCA colored by spectral cluster
+ plt.figure(figsize=(4, 4))
+ for k in np.unique(adata.obs['cluster'][cells]):
+ plt.plot(adata.obsm['X_pca'][adata.obs['cluster'] == k, 0],
+ adata.obsm['X_pca'][adata.obs['cluster'] == k, 1],
+ '.', ls='', ms=5, alpha=0.25)
+ plt.title('Gene expression embedding,\ncolored by spectral clustering')
+ plt.xlabel('PC 0')
+ plt.ylabel('PC 1')
+ plt.show()
+
+ display(Markdown('## Agreement of per-celltype gene expression with truth'))
+ display(Markdown('We can now compare the gene expression profile of cells '
+ 'from the same cluster (found *de novo*) before and after '
+ 'CellBender. We expect to see the ambient profile match, '
+ 'and we expect that each *de novo* cluster will match the '
+ 'gene expression of one of the "truth" clusters, although '
+ 'the label order may be swapped. Here are three ways to '
+ 'visualize this same data.'))
+
+ # load truth data into a matrix
+ true_chi = np.zeros((np.unique(cluster).size, adata.shape[1]))
+ true_chi[0, :] = adata.var['truth_ambient_expression']
+ for i in range(1, np.unique(cluster).size):
+ true_chi[i, :] = adata.var[f'truth_gene_expression_cell_label_{i}']
+
+ # figure out the learned expression profiles chi
+ learned_chi = np.zeros((np.unique(cluster).size, true_chi.shape[1]))
+ learned_chi[0, :] = adata.var['ambient_expression']
+ for k in adata.obs['cluster'].unique():
+ if k == 0:
+ continue
+ # get chi from mean cell expression in that cluster
+ summmed_expression = np.array(adata.layers['cellbender'][adata.obs['cluster'] == k, :]
+ .sum(axis=0)).squeeze()
+ learned_chi[k, :] = summmed_expression / summmed_expression.sum()
+
+ # compare learned expression to the underlying true expression
+ def _gridplot(adata, upper_lim=1e-1, scale='linear'):
+ plt.figure(figsize=(10, 9))
+ n_clusters = adata.obs['cluster'].nunique()
+ for k in adata.obs['cluster'].unique():
+ for j in adata.obs['cluster'].unique():
+ plt.subplot(n_clusters,
+ n_clusters,
+ n_clusters * k + j + 1)
+ plt.plot([1e-7, 1e-1], [1e-7, 1e-1], color='black', lw=0.2)
+ plt.plot(true_chi[j, :], learned_chi[k, :], '.', ls='', ms=1, alpha=0.5)
+ plt.ylim([1e-6, upper_lim])
+ plt.xlim([1e-6, upper_lim])
+ plt.xscale(scale)
+ plt.yscale(scale)
+ if j == 0:
+ plt.ylabel(f'Learned {k if k > 0 else "ambient"}')
+ else:
+ plt.yticks([])
+ if k == n_clusters - 1:
+ plt.xlabel(f'True {j if j > 0 else "ambient"}')
+ else:
+ plt.xticks([])
+ plt.show()
+
+ _gridplot(adata, scale='log')
+ _gridplot(adata, upper_lim=1e-2,)
+
+ # same in the format of cosine distance
+ n_clusters = adata.obs['cluster'].nunique()
+ dist = np.zeros((n_clusters, n_clusters))
+ for k in adata.obs['cluster'].unique():
+ for j in adata.obs['cluster'].unique():
+ u, v = true_chi[j, :], learned_chi[k, :]
+ # cosine distance
+ dist[k, j] = cosine(u, v)
+
+ fig = plt.figure(figsize=(5, 5))
+ if dist.max() > 0.9:
+ plt.imshow(dist, vmin=0, vmax=dist[dist < 0.9].max() + 0.1,
+ cmap=plt.cm.gray_r)
+ else:
+ plt.imshow(dist, vmin=0,
+ cmap=plt.cm.gray_r)
+
+ for i in range(dist.shape[0]):
+ for j in range(dist.shape[1]):
+ plt.text(i - 0.3, j + 0.05, f'{dist[j, i]:.3f}', fontsize=14, color='gray')
+
+ plt.xticks(range(n_clusters), ['BKG'] + [i + 1 for i in range(n_clusters - 1)])
+ # plt.gca().xaxis.set_ticks_position('top')
+ # plt.gca().xaxis.set_label_position('top')
+ plt.yticks(range(n_clusters), ['BKG'] + [i + 1 for i in range(n_clusters - 1)])
+ plt.ylim(plt.gca().get_xlim()[::-1])
+ plt.xlabel('True')
+ plt.ylabel('Inferred')
+ plt.title('Cosine distance between\ntrue and inferred expression')
+
+ from mpl_toolkits.axes_grid1 import make_axes_locatable
+
+ divider = make_axes_locatable(plt.gca())
+ cax = divider.append_axes("right", size="5%", pad=0.05)
+ plt.colorbar(cax=cax, format='%.2f')
+
+ plt.tight_layout()
+ plt.show()
+
+ return learned_chi
+
+
+def cosine(u: np.ndarray, v: np.ndarray):
+ return 1 - np.dot(u, v) / (np.sqrt(np.dot(u, u)) * np.sqrt(np.dot(v, v)) + 1e-10)
+
+
+def fano(count_matrix_csr, axis=0):
+ """Calculate Fano factor for count matrix, either per gene or per cell,
+ as specified by axis.
+ """
+
+ mean = np.array(count_matrix_csr.mean(axis=axis)).squeeze()
+ square_mean = np.array(count_matrix_csr.power(2).mean(axis=axis)).squeeze()
+ var = square_mean - np.power(mean, 2)
+ return var / (mean + 1e-10)
+
+
+def plot_counts_per_gene(adata):
+
+ order_gene = np.argsort(adata.var['n_cellbender'])[::-1]
+ plt.loglog(adata.var['n_cellbender'][order_gene])
+ plt.loglog(adata.var['n_cellbender'][order_gene], ':', alpha=0.5)
+ plt.xlabel('Sorted gene ID')
+ plt.ylabel('Unique UMI counts')
+ plt.show()
+
+
+def compare_latents(adata, input_layer_key='raw'):
+ """Compare inferred latents to truth"""
+
+ truth_exists = ('truth_cell_probability' in adata.obs.keys())
+
+ display(Markdown('## Inference of latent variables (for experts)'))
+
+ if not truth_exists:
+ display(Markdown('This section deals with particular details of the '
+ 'latent variables in the `remove-background` model. '
+ 'This analysis is provided for interested experts, but '
+ 'is not necessary for judging the success of a run.'))
+
+ display(Markdown('### Ambient gene expression profile'))
+ plt.figure(figsize=(12, 3))
+ plt.plot(adata.var['ambient_expression'].values)
+ plt.ylabel('Fractional expression')
+ plt.xlabel('Gene')
+ plt.title('Inferred ambient profile')
+ plt.show()
+
+ if truth_exists:
+ plt.figure(figsize=(12, 3))
+ plt.plot(adata.var['ambient_expression'].values
+ - adata.var['truth_ambient_expression'].values)
+ plt.ylabel('Residual: inferred - true')
+ plt.xlabel('Gene')
+ plt.title('Inferred ambient profile')
+ plt.show()
+
+ display(Markdown('### Swapping fraction, rho'))
+
+ rho = adata.uns.get('swapping_fraction_dist_params', None)
+ if rho is None:
+ display(Markdown('Swapping fraction not inferred. `--model` may have '
+ 'been "ambient" instead of "swapping" or "full"'))
+ else:
+ from scipy.stats import beta
+ nbins = 100
+ xx = np.linspace(0, 1, nbins)
+ if truth_exists:
+ plt.hist(adata.obs['truth_swapping_fraction'].values,
+ bins=xx, label='Truth', alpha=0.5)
+ plt.plot(xx, (adata.shape[0] / nbins
+ * beta.pdf(xx, consts.RHO_ALPHA_PRIOR, consts.RHO_BETA_PRIOR)),
+ label='Prior', color='k')
+ plt.plot(xx, (adata.shape[0] / nbins
+ * beta.pdf(xx, rho[0], rho[1])), label='Inferred', color='r')
+ plt.title('Inferred distribution for swapping fraction')
+ plt.legend(loc='center left', bbox_to_anchor=(1, 0.5))
+ plt.xlabel('Swapping fraction rho')
+ plt.ylabel('Number of droplets')
+ plt.show()
+
+ if truth_exists:
+
+ display(Markdown('### Droplet efficiency and cell size latents'))
+ display(Markdown('Including empty droplets'))
+ keys = ['droplet_efficiency', 'cell_size']
+
+ plt.figure(figsize=(5 * len(keys), 4))
+ for i, key in enumerate(keys):
+ plt.subplot(1, len(keys), i + 1)
+ otherkey = 'droplet_efficiency' if (key == 'cell_size') else 'cell_size'
+ plt.scatter(adata.obs[f'truth_{key}'], adata.obs[key],
+ c=adata.obs[otherkey], s=5, cmap='coolwarm', alpha=0.5)
+ cbar = plt.colorbar()
+ cbar.set_label(otherkey.capitalize().replace('_', ' '))
+ xx = [adata.obs[f'truth_{key}'].min(), adata.obs[f'truth_{key}'].max()]
+ plt.plot(xx, xx, color='lightgray')
+ plt.title(key.capitalize().replace('_', ' '))
+ plt.xlabel('truth')
+ plt.ylabel('cellbender')
+ plt.axis('equal')
+ plt.tight_layout()
+ plt.show()
+
+ display(Markdown('Cells only'))
+ cells = (adata.obs['cell_probability'] > 0.5)
+ plt.figure(figsize=(5 * len(keys), 4))
+ for i, key in enumerate(keys):
+ plt.subplot(1, len(keys), i + 1)
+ for c in sorted(adata.obs['truth_cell_label'][cells].unique()):
+ ctype = (adata.obs['truth_cell_label'] == c)
+ plt.plot(adata.obs[f'truth_{key}'][cells & ctype],
+ adata.obs[key][cells & ctype], '.',
+ label=f'{c} ({ctype.sum()} cells)', alpha=0.5)
+ xx = [adata.obs[f'truth_{key}'][cells].min(),
+ adata.obs[f'truth_{key}'][cells].max()]
+ plt.plot(xx, xx, color='lightgray')
+ plt.title(key.capitalize().replace('_', ' ') + ': cells')
+ plt.xlabel('truth')
+ plt.ylabel('cellbender')
+ plt.axis('equal')
+ plt.legend(loc='center left', bbox_to_anchor=(1, 0.5), title='Cell type')
+ plt.tight_layout()
+ plt.show()
+
+ display(Markdown(r'### Droplet efficiency $\epsilon$'))
+ if truth_exists:
+ display(Markdown('Comparison with the truth.'))
+
+ order = np.argsort(adata.obs[f'n_{input_layer_key}'].values)[::-1]
+ plt.figure(figsize=(6, 4))
+ plt.semilogy(adata.obs[f'n_{input_layer_key}'].values[order], 'k', lw=2)
+ plt.ylabel('UMI counts')
+ plt.xlabel('Droplet')
+
+ plt.gca().twinx()
+ if truth_exists:
+ plt.plot(adata.obs['truth_droplet_efficiency'].values[order],
+ label='truth', alpha=0.5)
+ plt.plot(adata.obs['droplet_efficiency'].values[order],
+ 'r', label='inferred', alpha=0.5)
+ plt.ylabel(r'$\epsilon$', fontsize=18, color='r')
+ plt.legend()
+ plt.show()
+
+ display(Markdown('Comparison with the prior.'))
+ from numpy.random import gamma
+ rh = consts.EPSILON_PRIOR
+ delta = 0.01
+ xx = np.arange(0, 2.1, step=delta)
+
+ if truth_exists:
+ cells = (adata.obs['truth_cell_probability'] == 1)
+ yy, _ = np.histogram(
+ gamma(adata.uns['truth_epsilon_param'],
+ scale=1. / adata.uns['truth_epsilon_param'], size=[50000]),
+ bins=xx,
+ )
+ plt.step(xx[:-1], yy / np.sum(yy * delta), label='truth', color='r')
+ else:
+ cells = (adata.obs['cell_probability'] > 0.5)
+
+ yy, _ = np.histogram(gamma(rh, scale=1. / rh, size=[50000]), bins=xx)
+ plt.step(xx[:-1], yy / np.sum(yy * delta), label='prior', color='k')
+
+ yy, _ = np.histogram(adata.obs['droplet_efficiency'][cells], bins=xx)
+ plt.step(xx[:-1], yy / np.sum(yy * delta), alpha=0.5, label='posterior: cells')
+
+ yy, _ = np.histogram(adata.obs['droplet_efficiency'][~cells], bins=xx)
+ plt.step(xx[:-1], yy / np.sum(yy * delta), alpha=0.5, label='posterior: empties')
+
+ plt.ylabel('Probability density')
+ plt.xlabel(r'$\epsilon$', fontsize=18)
+ plt.legend(loc='center left', bbox_to_anchor=(1, 0.5))
+ plt.show()
+
+ display(Markdown('### Cell size $d$'))
+ if truth_exists:
+ display(Markdown('Comparison with the truth.'))
+ cell_order = np.argsort(adata.obs[f'n_{input_layer_key}'].values[cells])[::-1]
+ plt.figure(figsize=(6, 4))
+ plt.semilogy(adata.obs[f'n_{input_layer_key}'].values[cells][cell_order], 'k', lw=2)
+ plt.ylabel('UMI counts')
+ plt.xlabel('Droplet')
+
+ plt.gca().twinx()
+ if truth_exists:
+ plt.semilogy(adata.obs['truth_cell_size'].values[cells][cell_order],
+ label='truth', alpha=0.5)
+ plt.semilogy(adata.obs['cell_size'].values[cells][cell_order],
+ 'r', label='inferred', alpha=0.5)
+ plt.ylabel('$d$', fontsize=18, color='r')
+ plt.legend()
+ plt.show()
+
+ display(Markdown(r'### The combination $d \epsilon$'))
+ display(Markdown('It is easier for the model to nail this down, since there '
+ 'is less degeneracy.'))
+
+ plt.figure(figsize=(6, 4))
+ plt.semilogy(adata.obs[f'n_{input_layer_key}'].values[cells][cell_order], 'k', lw=2)
+ plt.ylabel('UMI counts')
+ plt.xlabel('Droplet')
+
+ plt.gca().twinx()
+ if truth_exists:
+ plt.semilogy((adata.obs['truth_cell_size'].values[cells][cell_order]
+ * adata.obs['truth_droplet_efficiency'].values[cells][cell_order]),
+ label='truth', alpha=0.5)
+ plt.semilogy((adata.obs['cell_size'].values[cells][cell_order]
+ * adata.obs['droplet_efficiency'].values[cells][cell_order]),
+ 'r', label='inferred', alpha=0.5)
+ plt.ylabel(r'$d \epsilon$', fontsize=18, color='r')
+ plt.legend()
+ plt.show()
+
+ display(Markdown(r'### Joint distributions of $d$ and $\epsilon$'))
+ if truth_exists:
+ display(Markdown('In true cells'))
+ else:
+ display(Markdown('In inferred cells'))
+ if truth_exists:
+ plt.figure(figsize=(9, 4))
+ plt.subplot(1, 2, 1)
+ for c in sorted(adata.obs['truth_cell_label'][cells].unique()):
+ ctype = (adata.obs['truth_cell_label'] == c)
+ plt.semilogy(adata.obs['truth_droplet_efficiency'][ctype],
+ adata.obs['truth_cell_size'][ctype],
+ '.', alpha=0.5)
+ plt.ylabel('$d$', fontsize=18)
+ plt.xlabel(r'$\epsilon$', fontsize=18)
+ plt.title('Truth')
+ plt.subplot(1, 2, 2)
+ for c in sorted(adata.obs['truth_cell_label'][cells].unique()):
+ ctype = (adata.obs['truth_cell_label'] == c)
+ plt.semilogy(adata.obs['droplet_efficiency'][ctype],
+ adata.obs['cell_size'][ctype],
+ '.', label=f'{c} ({ctype.sum()} cells)', alpha=0.5)
+ plt.legend(loc='center left', bbox_to_anchor=(1, 0.5), title='Cell type')
+ else:
+ plt.figure(figsize=(4, 4))
+ plt.semilogy(adata.obs['droplet_efficiency'][cells],
+ adata.obs['cell_size'][cells],
+ 'k.', ms=1, alpha=0.5)
+ plt.ylabel('$d$', fontsize=18)
+ plt.xlabel(r'$\epsilon$', fontsize=18)
+ plt.title('Inferred')
+ plt.tight_layout()
+ plt.show()
+
+
+def mixed_species_plots(adata, input_layer_key='raw', output_layer_key='cellbender',
+ ngene_cutoff=5000, count_cutoff=500):
+ """Make mixed species plots if it seems appropriate"""
+
+ # find a list of genes with expression exclusive to a cell type
+ if 'genome' not in adata.var.keys():
+ return
+
+ if ('feature_type' in adata.var.keys()) and ('Gene Expression' in adata.var['feature_type'].values):
+ genomes = adata.var[adata.var['feature_type'] == 'Gene Expression']['genome'].unique()
+ else:
+ genomes = adata.var['genome'].unique()
+ cells = (adata.obs['cell_probability'] > 0.5)
+
+ if len(genomes) != 2:
+ return
+
+ display(Markdown('## "Mixed species" analysis'))
+ display(Markdown('There are multiple genomes in the dataset, so we can '
+ 'analyze these genes assuming they were from a "mixed '
+ 'species" experiment, where we know that certain genes '
+ 'come only from certain celltypes.'))
+
+ for genome in genomes:
+ if 'feature_type' in adata.var.keys():
+ if 'Gene Expression' in adata.var["feature_type"].values:
+ var_subset = adata.var[(adata.var["genome"] == genome)
+ & (adata.var["feature_type"] == "Gene Expression")]
+ else:
+ var_subset = adata.var[(adata.var["genome"] == genome)]
+ else:
+ var_subset = adata.var[(adata.var["genome"] == genome)]
+ print(f'Genome "{genome}" has {len(var_subset)} genes: '
+ f'{", ".join(var_subset.index.values[:3])} ...')
+
+ for i, genome1 in enumerate(genomes):
+ for genome2 in genomes[(i + 1):]:
+ plt.figure(figsize=(5, 5))
+ for k in [input_layer_key, output_layer_key]:
+ x = np.array(adata.layers[k]
+ [:, (adata.var['genome'] == genome1)].sum(axis=1)).squeeze()
+ y = np.array(adata.layers[k]
+ [:, (adata.var['genome'] == genome2)].sum(axis=1)).squeeze()
+ plt.plot(x + 1,
+ y + 1,
+ '.', ms=2, label=k, alpha=0.3,
+ color='k' if (k == input_layer_key) else 'r')
+ x = x[cells]
+ y = y[cells]
+ contam_x = (y / (y + x + 1e-10))[x > (10 * y)]
+ contam_y = (x / (x + y + 1e-10))[y > (10 * x)]
+ contam = np.concatenate((contam_x, contam_y), axis=0)
+ display(Markdown(f'Mean "contamination" per droplet in {k} data: '
+ f'{contam.mean() * 100:.2f} %'))
+ display(Markdown(f'Median "contamination" per droplet in {k} data: '
+ f'{np.median(contam) * 100:.2f} %'))
+ plt.xscale('log')
+ plt.yscale('log')
+ plt.ylim(bottom=0.55)
+ plt.xlim(left=0.55)
+ plt.xlabel(f'{genome1} counts (+1)')
+ plt.ylabel(f'{genome2} counts (+1)')
+ plt.title(f'Counts per droplet: {output_layer_key}')
+ lgnd = plt.legend(loc='center left', bbox_to_anchor=(1, 0.5))
+ _set_legend_dot_size(lgnd, size=10)
+ plt.show()
+
+
+def _set_legend_dot_size(lgnd, size: int):
+ """Set dot size in a matplotlib legend.
+ Different versions of matplotlib seem to need different things.
+ Just give up and move on if it doesn't work.
+ I know this is a bit ridiculous but so is matplotlib.
+ """
+ for h in lgnd.legendHandles:
+ worked = False
+ try:
+ h._legmarker.set_markersize(size)
+ worked = True
+ except AttributeError:
+ pass
+ if worked:
+ continue
+ try:
+ h.set_sizes([size])
+ worked = True
+ except AttributeError:
+ pass
+ if worked:
+ continue
+ try:
+ h._sizes = [size]
+ worked = True
+ except AttributeError:
+ pass
+ if worked:
+ continue
+ try:
+ h.set_markersize(size)
+ worked = True
+ except AttributeError:
+ pass
+
+
+def cell_roc_count_roc(output_csr: sp.csr_matrix,
+ input_csr: sp.csr_matrix,
+ truth_csr: sp.csr_matrix,
+ cell_calls: np.ndarray,
+ truth_cell_labels: np.ndarray = None) -> float: # 0 = empty, nonzero = cell
+ """Plot a ROC curve (point) for cell identification, and plot
+ a second for noise count identification.
+
+ Returns:
+ fpr: Acutal false positive rate for counts in cells
+
+ Note:
+ True positive = a noise count that has been removed
+ True negative = a real count that has not been removed
+ False positive = a real count that has been removed
+ False negative = a noise count that has not been removed
+
+ | noise | real |
+ ------------------------------------
+ removed | TP | FP |
+ -------------|----------o----------|
+ not removed | FN | TN |
+ ------------------------------------
+
+ For the cell identification case:
+
+ True positive = a cell called a cell
+ True negative = an empty called empty
+ False positive = an empty called a cell
+ False negative = a cell called empty
+
+ | cell | empty |
+ --------------------------------------
+ called a cell | TP | FP |
+ ---------------|----------o----------|
+ called empty | FN | TN |
+ --------------------------------------
+
+ """
+
+ assert cell_calls.size == truth_csr.shape[0], "Cell calls must be number of barcodes."
+
+ # find the real cells as things with counts in truth data
+ cell_truth = (np.array(truth_csr.sum(axis=1)).squeeze() > 0)
+ called_cell = (cell_calls != 0)
+
+ display(Markdown('### Identification of non-empty droplets'))
+ display(Markdown('TPR = true positive rate\n\nFPR = false positive rate'))
+
+ # cell breakdown
+ tpr = np.logical_and(cell_truth, called_cell).sum() / cell_truth.sum()
+ fpr = np.logical_and(np.logical_not(cell_truth), called_cell).sum() / called_cell.sum()
+ print(f'\nCell identification TPR = {tpr:.3f}')
+ print(f'Cell identification FPR = {fpr:.3f}')
+
+ fig = plt.figure(figsize=(5, 5))
+ plt.plot(fpr, tpr, 'o')
+ plt.ylabel('fraction of true cells called')
+ plt.xlabel('fraction of called cells that are really empty')
+ plt.ylim([-0.05, 1.05])
+ plt.xlim([-0.05, 1.05])
+ plt.title('Cell identification')
+ plt.show()
+
+ # counts
+ # just work with cells
+ cell_inds = np.where(cell_truth)[0]
+ real_counts = truth_csr[cell_inds]
+ removed_counts = input_csr[cell_inds] - output_csr[cell_inds]
+ noise_counts = input_csr[cell_inds] - truth_csr[cell_inds]
+
+ # count breakdown
+ tp = (np.array(noise_counts
+ .minimum(removed_counts)
+ .sum(axis=1)).squeeze())
+ erroneously_removed_counts = removed_counts - noise_counts
+ erroneously_removed_counts[erroneously_removed_counts < 0] = 0
+ fp = np.array(erroneously_removed_counts.sum(axis=1)).squeeze()
+ noise_remaining = noise_counts - removed_counts
+ noise_remaining[noise_remaining < 0] = 0
+ fn = np.array(noise_remaining.sum(axis=1)).squeeze()
+ tn = np.array(real_counts.minimum(output_csr[cell_inds]).sum(axis=1)).squeeze()
+
+ display(Markdown('Take a look at the statistics of count removal, with the '
+ 'following definitions:\n\n- TP = true positive: a noise '
+ 'count (correctly) removed\n- FP = false positive: a real '
+ 'count (incorrectly) removed\n- TN = true negative: a real '
+ 'count (correctly) not removed\n- FN = false negative: a '
+ 'noise count (incorrectly) not removed'))
+
+ print(f'max ambient counts in any matrix entry is {(input_csr - truth_csr).max()}')
+
+ plt.figure(figsize=(8, 1))
+ plt.hist(tp, bins=100)
+ plt.title('TP: noise counts removed per cell')
+ plt.show()
+
+ plt.figure(figsize=(8, 1))
+ plt.hist(fp, bins=100)
+ plt.title('FP: real counts removed per cell')
+ plt.show()
+
+ plt.figure(figsize=(8, 1))
+ plt.hist(tn, bins=100)
+ plt.title('TN: real counts not removed per cell')
+ plt.show()
+
+ plt.figure(figsize=(8, 1))
+ plt.hist(fn, bins=100)
+ plt.title('FN: noise counts not removed per cell')
+ plt.show()
+
+ plot_roc_points(tp, fp, tn, fn, point_colors=np.array(truth_cell_labels)[cell_truth],
+ title='Background removal from cells', color='black', show=False)
+ plt.show()
+
+ display(Markdown('### Summary statistics for noise removal'))
+
+ print(f'Cell background removal TPR = {(tp / (tp + fn)).mean():.3f}'
+ f' +/- {(tp / (tp + fn)).std():.3f}')
+ print(f'Cell background removal FPR = {1 - (tn / (tn + fp)).mean():.3f}'
+ f' +/- {(tn / (tn + fp)).std():.3f}')
+
+ display(Markdown('And finally one more way to quanitify the amount of noise remaining:'))
+
+ contam_cell_raw = (np.array(noise_counts.sum(axis=1)).squeeze()
+ / np.array(real_counts.sum(axis=1)).squeeze())
+ contam_cell_after = (np.array(noise_remaining.sum(axis=1)).squeeze()
+ / np.array(real_counts.sum(axis=1)).squeeze())
+ print(f'Per cell contamination fraction raw = {contam_cell_raw.mean():.3f}'
+ f' +/- {contam_cell_raw.std():.3f}')
+ print(f'Per cell contamination fraction after = {contam_cell_after.mean():.3f}'
+ f' +/- {contam_cell_after.std():.3f}')
+
+ return 1 - (tn / (tn + fp)).mean() # FPR
+
+
+def plot_roc_points(tp: np.ndarray,
+ fp: np.ndarray,
+ tn: np.ndarray,
+ fn: np.ndarray,
+ point_colors: np.ndarray,
+ title: str = 'roc',
+ show: bool = True,
+ **kwargs):
+ """Plot points (and one summary point with error bars) on a ROC curve,
+ where the x-axis is FPR and the y-axis is TPR.
+
+ Args:
+ tp: True positives, multiple measurements of the same value.
+ fp: False positives
+ tn: True negatives
+ fn: False negatives
+ point_colors: Color for each point
+ title: Plot title
+ show: True to show plot
+ kwargs: Passed to the plotter for the summary point with error bars
+
+ """
+
+ fig = plt.figure(figsize=(5, 5))
+ sensitivity = tp / (tp + fn)
+ specificity = tn / (tn + fp)
+ sens_mean = sensitivity.mean()
+ sens_stdv = sensitivity.std()
+ spec_mean = specificity.mean()
+ spec_stdv = specificity.std()
+ # individual points
+ if type(point_colors) == str:
+ point_colors = np.array([point_colors] * len(sensitivity))
+ else:
+ assert len(point_colors) == len(sensitivity), 'point_colors is wrong length'
+ unique_colors = np.unique(point_colors)
+ for c in unique_colors:
+ plt.plot((1 - specificity)[point_colors == c],
+ sensitivity[point_colors == c],
+ '.', ms=1, alpha=0.5, label=c)
+ lgnd = plt.legend(loc='center left', bbox_to_anchor=(1, 0.5))
+ _set_legend_dot_size(lgnd, size=15)
+ # midpoint
+ plt.plot(1 - spec_mean, sens_mean, 'o', **kwargs)
+ # errorbars
+ plt.plot([1 - (spec_mean - spec_stdv),
+ 1 - (spec_mean + spec_stdv)],
+ [sens_mean, sens_mean], **kwargs)
+ plt.plot([1 - spec_mean, 1 - spec_mean],
+ [sens_mean - sens_stdv,
+ sens_mean + sens_stdv], **kwargs)
+ plt.ylabel('sensitivity = TP / (TP + FN)\ntrue positive rate')
+ plt.xlabel('1 - specificity = 1 - TN / (TN + FP)\nfalse positive rate')
+ plt.ylim([-0.05, 1.05])
+ plt.xlim([-0.05, 1.05])
+ plt.title(title)
+
+ if show:
+ plt.show()
+ else:
+ return fig
+
+
+def pca_2d(mat: np.ndarray) -> torch.Tensor:
+ """Perform PCA using pytorch and return top 2 PCs
+
+ Args:
+ mat: matrix where rows are observations and columns are features
+
+ Returns:
+ out: matrix where rows are observations and columns are top 2 PCs
+ """
+
+ A = torch.as_tensor(mat).float()
+ U, S, V = torch.pca_lowrank(A)
+ return torch.matmul(A, V[:, :2])
+
+
+def plot_summary(loss: Dict[str, Dict[str, np.ndarray]],
+ umi_counts: np.ndarray,
+ p: np.ndarray,
+ z: np.ndarray):
+ """Output summary plot with three panels: training, cells, latent z."""
+
+ fig = plt.figure(figsize=(6, 18))
+
+ # Plot the train error.
+ plt.subplot(3, 1, 1)
+ try:
+ plt.plot(loss['train']['elbo'], '.--', label='Train')
+
+ # Plot the test error, if there was held-out test data.
+ if 'test' in loss.keys():
+ if len(loss['test']['epoch']) > 0:
+ plt.plot(loss['test']['epoch'],
+ loss['test']['elbo'], 'o:', label='Test')
+ plt.legend()
+
+ ylim_low = max(loss['train']['elbo'][0], loss['train']['elbo'][-1] - 2000)
+ try:
+ ylim_high = max(max(loss['train']['elbo']), max(loss['test']['elbo']))
+ except ValueError:
+ ylim_high = max(loss['train']['elbo'])
+ ylim_high = ylim_high + (ylim_high - ylim_low) / 20
+ plt.gca().set_ylim([ylim_low, ylim_high])
+ except:
+ logger.warning('Error plotting the learning curve. Skipping.')
+ pass
+
+ plt.xlabel('Epoch')
+ plt.ylabel('ELBO')
+ plt.title('Progress of the training procedure')
+
+ # Plot the barcodes used, along with the inferred
+ # cell probabilities.
+ plt.subplot(3, 1, 2)
+ count_order = np.argsort(umi_counts)[::-1]
+ plt.semilogy(umi_counts[count_order], color='black')
+ plt.ylabel('UMI counts')
+ plt.xlabel('Barcode index, sorted by UMI count')
+ if p is not None: # The case of a simple model.
+ plt.gca().twinx()
+ plt.plot(p[count_order], '.:', color='red', alpha=0.3, rasterized=True)
+ plt.ylabel('Cell probability', color='red')
+ plt.ylim([-0.05, 1.05])
+ plt.title('Determination of which barcodes contain cells')
+ else:
+ plt.title('The subset of barcodes used for training')
+
+ plt.subplot(3, 1, 3)
+ if p is None:
+ p = np.ones(z.shape[0])
+
+ # Do PCA on the latent encoding z.
+ z_pca = pca_2d(z[p >= consts.CELL_PROB_CUTOFF])
+
+ # Plot the latent encoding via PCA.
+ plt.plot(z_pca[:, 0], z_pca[:, 1],
+ '.', ms=3, color='black', alpha=0.3, rasterized=True)
+ plt.ylabel('PC 1')
+ plt.xlabel('PC 0')
+ plt.title('PCA of latent encoding of gene expression in cells')
+
+ return fig
diff --git a/cellbender/remove_background/run.py b/cellbender/remove_background/run.py
new file mode 100644
index 00000000..7b777e4e
--- /dev/null
+++ b/cellbender/remove_background/run.py
@@ -0,0 +1,779 @@
+"""Single run of remove-background, given input arguments."""
+
+from cellbender.remove_background.model import RemoveBackgroundPyroModel
+from cellbender.remove_background.data.dataset import get_dataset_obj, \
+ SingleCellRNACountsDataset
+from cellbender.remove_background.vae.decoder import Decoder
+from cellbender.remove_background.vae.encoder \
+ import EncodeZ, CompositeEncoder, EncodeNonZLatents
+from cellbender.remove_background.data.dataprep import \
+ prep_sparse_data_for_training as prep_data_for_training
+from cellbender.remove_background.checkpoint import attempt_load_checkpoint, \
+ save_checkpoint
+import cellbender.remove_background.consts as consts
+from cellbender.remove_background.data.dataprep import DataLoader
+from cellbender.remove_background.train import run_training
+from cellbender.remove_background.posterior import Posterior, PRq, PRmu, \
+ compute_mean_target_removal_as_function, load_or_compute_posterior_and_save
+from cellbender.remove_background.estimation import MultipleChoiceKnapsack, \
+ MAP, Mean, SingleSample, ThresholdCDF
+from cellbender.remove_background.exceptions import ElboException
+from cellbender.remove_background.sparse_utils import csr_set_rows_to_zero
+from cellbender.remove_background.data.io import write_matrix_to_cellranger_h5
+from cellbender.remove_background.report import run_notebook_make_html, plot_summary
+
+import pyro
+from pyro.infer import SVI, JitTraceEnum_ELBO, JitTrace_ELBO, \
+ TraceEnum_ELBO, Trace_ELBO
+from pyro.optim import ClippedAdam
+import numpy as np
+import scipy.sparse as sp
+import pandas as pd
+import torch
+
+import os
+import sys
+import logging
+import argparse
+import traceback
+import psutil
+from datetime import datetime
+from typing import Tuple, Optional, Dict, Union
+
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt # This needs to be after matplotlib.use('Agg')
+
+
+logger = logging.getLogger('cellbender')
+
+
+def run_remove_background(args: argparse.Namespace) -> Posterior:
+ """The full script for the command line tool to remove background RNA.
+
+ Args:
+ args: Inputs from the command line, already parsed using argparse.
+
+ Note: Returns nothing, but writes output to a file(s) specified from
+ command line.
+
+ """
+
+ # Handle initial random state.
+ pyro.util.set_rng_seed(consts.RANDOM_SEED)
+ if torch.cuda.is_available():
+ torch.cuda.manual_seed_all(consts.RANDOM_SEED)
+
+ # Load dataset, run inference, and write the output to a file.
+
+ # Log the start time.
+ logger.info(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
+ logger.info("Running remove-background")
+
+ # Run pytorch multithreaded if running on CPU: but this makes little difference in runtime.
+ if not args.use_cuda:
+ if args.n_threads is not None:
+ n_jobs = args.n_threads
+ else:
+ n_jobs = psutil.cpu_count(logical=True)
+ torch.set_num_threads(n_jobs)
+ logger.debug(f'Set pytorch to use {n_jobs} threads')
+
+ # Load data from file and choose barcodes and genes to analyze.
+ try:
+ dataset_obj = get_dataset_obj(args=args)
+
+ except OSError:
+ logger.error(f"OSError: Unable to open file {args.input_file}.")
+ logger.error(traceback.format_exc())
+ sys.exit(1)
+
+ # Instantiate latent variable model and run full inference procedure.
+ if args.model == 'naive':
+ inferred_model = None
+ else:
+ inferred_model, _, _, _ = run_inference(dataset_obj=dataset_obj, args=args)
+ inferred_model.eval()
+
+ try:
+
+ file_dir, file_base = os.path.split(args.output_file)
+ file_name = os.path.splitext(os.path.basename(file_base))[0]
+
+ # Create the posterior and save it.
+ posterior = load_or_compute_posterior_and_save(
+ dataset_obj=dataset_obj,
+ inferred_model=inferred_model,
+ args=args,
+ )
+ logger.info(datetime.now().strftime('%Y-%m-%d %H:%M:%S\n'))
+
+ # Save output plots.
+ save_output_plots(
+ file_dir=file_dir,
+ file_name=file_name,
+ dataset_obj=dataset_obj,
+ inferred_model=inferred_model,
+ p=posterior.latents_map['p'],
+ z=posterior.latents_map['z'],
+ )
+
+ # Save cell barcodes in a CSV file.
+ analyzed_barcode_logic = (posterior.latents_map['p'] > consts.CELL_PROB_CUTOFF)
+ cell_barcodes = (dataset_obj.data['barcodes']
+ [dataset_obj.analyzed_barcode_inds[analyzed_barcode_logic]])
+ bc_file_name = os.path.join(file_dir, file_name + "_cell_barcodes.csv")
+ write_cell_barcodes_csv(bc_file_name=bc_file_name, cell_barcodes=cell_barcodes)
+
+ # Compute estimates of denoised count matrix for each FPR and save them.
+ compute_output_denoised_counts_reports_metrics(
+ posterior=posterior,
+ args=args,
+ file_dir=file_dir,
+ file_name=file_name,
+ )
+
+ logger.info("Completed remove-background.")
+ logger.info(datetime.now().strftime('%Y-%m-%d %H:%M:%S\n'))
+
+ return posterior
+
+ # The exception allows user to end inference prematurely with CTRL-C.
+ except KeyboardInterrupt:
+
+ # If partial output has been saved, delete it.
+ full_file = args.output_file
+
+ # Name of the filtered (cells only) file.
+ file_dir, file_base = os.path.split(full_file)
+ file_name = os.path.splitext(os.path.basename(file_base))[0]
+ filtered_file = os.path.join(file_dir, file_name + "_filtered.h5")
+
+ if os.path.exists(full_file):
+ os.remove(full_file)
+
+ if os.path.exists(filtered_file):
+ os.remove(filtered_file)
+
+ logger.info("Keyboard interrupt. Terminated without saving.\n")
+
+
+def save_output_plots(file_dir: str,
+ file_name: str,
+ dataset_obj: SingleCellRNACountsDataset,
+ inferred_model: RemoveBackgroundPyroModel,
+ p: np.ndarray,
+ z: np.ndarray) -> bool:
+ """Save the UMI histogram and the three-panel output summary PDF"""
+
+ try:
+ # File naming.
+ summary_fig_name = os.path.join(file_dir, file_name + ".pdf")
+
+ # Three-panel output summary plot.
+ counts = np.array(dataset_obj.get_count_matrix().sum(axis=1)).squeeze()
+ fig = plot_summary(loss=inferred_model.loss, umi_counts=counts, p=p, z=z)
+ fig.savefig(summary_fig_name, bbox_inches='tight', format='pdf')
+ logger.info(f"Saved summary plots as {summary_fig_name}")
+ return True
+
+ except Exception:
+ logger.warning("Unable to save all plots.")
+ logger.warning(traceback.format_exc())
+ return False
+
+
+def compute_output_denoised_counts_reports_metrics(posterior: Posterior,
+ args: argparse.Namespace,
+ file_dir: str,
+ file_name: str) -> bool:
+ """Handle the estimation of the output denoised count matrix, given a
+ posterior. Compute the estimate for all specified FPR values. Save each
+ as we go. Create reports and save, and aggregate metrics and save as we go.
+
+ Args:
+ posterior: Posterior object
+ args: Argparser namespace with all parsed input args
+ file_dir:
+ file_name:
+
+ Returns:
+ True iff all files write correctly
+
+ """
+
+ # Ensure that the posterior distribution has been computed.
+ # TODO: when we make the properties non-private, then this should become obsolete
+ posterior.cell_noise_count_posterior_coo()
+
+ # Choose output count matrix estimation method.
+ noise_target_fun = None
+ if args.estimator == 'map':
+ estimator = MAP
+ elif args.estimator == 'mean':
+ estimator = Mean
+ elif args.estimator == 'sample':
+ estimator = SingleSample
+ elif args.estimator == 'cdf':
+ estimator = ThresholdCDF
+ elif args.estimator == 'mckp':
+ estimator = MultipleChoiceKnapsack
+
+ # Prep specific for MCKP: target estimation.
+ logger.info('Computing target noise counts per gene for MCKP estimator')
+ count_matrix = posterior.dataset_obj.data['matrix'] # all barcodes
+ cell_inds = posterior.dataset_obj.analyzed_barcode_inds[posterior.latents_map['p']
+ > consts.CELL_PROB_CUTOFF]
+ empty_inds = set(range(count_matrix.shape[0])) - set(cell_inds)
+ cell_counts = csr_set_rows_to_zero(csr=count_matrix, row_inds=empty_inds)
+
+ noise_target_fun_per_cell = compute_mean_target_removal_as_function(
+ noise_count_posterior_coo=posterior._noise_count_posterior_coo,
+ noise_offsets=posterior._noise_count_posterior_coo_offsets,
+ index_converter=posterior.index_converter,
+ raw_count_csr_for_cells=cell_counts,
+ n_cells=len(cell_inds),
+ device='cuda' if args.use_cuda else 'cpu', # TODO check this
+ per_gene=True,
+ )
+ noise_target_fun = lambda x: noise_target_fun_per_cell(x) * len(cell_inds)
+ else:
+ raise ValueError(f'Input --estimator must be one of '
+ f'["map", "mean", "sample", "cdf", "mckp"]')
+
+ # Save denoised count matrix outputs (for each FPR if applicable).
+ success = True
+ for fpr in args.fpr:
+
+ logger.debug(f'Working on FPR {fpr}')
+
+ # Regularize posterior again at this FPR, if needed.
+ if args.posterior_regularization in ['PRmu', 'PRmu_gene']:
+ # TODO: currently this re-calculates the first FPR which is not necessary
+ posterior.regularize_posterior(
+ regularization=PRmu,
+ raw_count_matrix=posterior.dataset_obj.data['matrix'],
+ fpr=fpr,
+ per_gene=True if (args.posterior_regularization == 'PRmu_gene') else False,
+ device='cuda',
+ )
+ else:
+ # Other posterior regularizations were already performed in
+ # posterior.load_or_compute_posterior_and_save()
+ pass
+
+ # If using MCKP, recompute noise targets for this FPR.
+ if noise_target_fun is not None:
+ noise_targets = noise_target_fun(fpr).detach().cpu().numpy()
+ logger.debug(f'Computed noise targets for FPR {fpr}:\n{noise_targets}')
+ logger.info(f'Using MCKP noise targets computed for FPR {fpr}')
+ else:
+ noise_targets = None
+
+ # Compute denoised counts.
+ logger.info(f'Computing denoised counts using {args.estimator} estimator')
+ denoised_counts = posterior.compute_denoised_counts(
+ estimator_constructor=estimator,
+ noise_targets_per_gene=noise_targets,
+ q=args.cdf_threshold_q,
+ alpha=args.prq_alpha,
+ device='cuda' if args.use_cuda else 'cpu',
+ use_multiple_processes=args.use_multiprocessing_estimation,
+ )
+
+ # Restore eliminated features in cells.
+ logger.debug('Restoring eliminated features in cells')
+ denoised_counts = posterior.dataset_obj.restore_eliminated_features_in_cells(
+ denoised_counts,
+ posterior.latents_map['p'],
+ )
+
+ # TODO: correct cell probabilities so that any zero-count droplet becomes "empty"
+
+ # Save denoised count matrix.
+ name_suffix = (f'_FPR_{fpr}' if len(args.fpr) > 1 else '')
+ fpr_output_filename = os.path.join(file_dir, file_name + name_suffix + '.h5')
+ filtered_output_file = os.path.join(file_dir, file_name + name_suffix + '_filtered.h5')
+
+ def _writer_helper(file, **kwargs) -> bool:
+ return write_denoised_count_matrix(
+ file=file,
+ denoised_count_matrix=denoised_counts,
+ posterior_regularization=args.posterior_regularization,
+ posterior_regularization_kwargs=posterior._noise_count_regularized_posterior_kwargs,
+ estimator=args.estimator,
+ estimator_kwargs=None if (args.cdf_threshold_q is None) else {'q': args.cdf_threshold_q},
+ latents=posterior.latents_map,
+ dataset_obj=posterior.dataset_obj,
+ learning_curve=posterior.vi_model.loss,
+ fpr=fpr,
+ **kwargs,
+ )
+
+ # Full count matrix.
+ write_succeeded = _writer_helper(file=fpr_output_filename)
+ success = success and write_succeeded
+
+ # Count matrix filtered to cells only.
+ analyzed_barcode_logic = (posterior.latents_map['p'] > consts.CELL_PROB_CUTOFF)
+ write_succeeded = _writer_helper(
+ file=filtered_output_file,
+ analyzed_barcode_logic=analyzed_barcode_logic,
+ barcode_inds=posterior.dataset_obj.analyzed_barcode_inds[analyzed_barcode_logic],
+ )
+ success = success and write_succeeded
+
+ # Compile and save metrics.
+ try:
+ df = collect_output_metrics(
+ dataset_obj=posterior.dataset_obj,
+ inferred_count_matrix=denoised_counts,
+ fpr=fpr,
+ cell_logic=(posterior.latents_map['p'] >= consts.CELL_PROB_CUTOFF),
+ loss=posterior.vi_model.loss,
+ )
+ metrics_file_name = os.path.join(file_dir, file_name + name_suffix + '_metrics.csv')
+ df.to_csv(metrics_file_name, index=True, header=False, float_format='%.3f')
+ logger.info(f'Saved output metrics as {metrics_file_name}')
+ except Exception:
+ logger.warning("Unable to collect output metrics.")
+ logger.warning(traceback.format_exc())
+
+ # Create report.
+ try:
+ os.environ['INPUT_FILE'] = os.path.abspath(os.path.join(os.getcwd(), args.input_file))
+ os.environ['OUTPUT_FILE'] = os.path.abspath(os.path.join(os.getcwd(), fpr_output_filename))
+ if args.truth_file is not None:
+ os.environ['TRUTH_FILE'] = os.path.abspath(os.path.join(os.getcwd(), args.truth_file))
+ html_report_file = os.path.join(file_dir, file_name + name_suffix + '_report.html')
+ run_notebook_make_html(
+ file=os.path.abspath(os.path.join(os.path.dirname(__file__), 'report.ipynb')),
+ output=html_report_file,
+ )
+ logger.info(f'Succeeded in writing report to {html_report_file}')
+
+ except Exception:
+ logger.warning("Unable to create report.")
+ logger.warning(traceback.format_exc())
+
+ return success
+
+
+def write_denoised_count_matrix(file: str,
+ denoised_count_matrix: sp.csr_matrix,
+ posterior_regularization: Optional[str],
+ posterior_regularization_kwargs: Optional[Dict[str, float]],
+ estimator: str,
+ estimator_kwargs: Optional[Dict[str, float]],
+ latents: Dict[str, np.ndarray],
+ dataset_obj: SingleCellRNACountsDataset,
+ learning_curve: Dict[str, np.ndarray], # inferred_model.loss
+ fpr: float,
+ analyzed_barcode_logic: np.ndarray = ...,
+ barcode_inds: np.ndarray = ...) -> bool:
+ """Helper function for writing output h5 files"""
+
+ z = latents['z'][analyzed_barcode_logic, :]
+ d = latents['d'][analyzed_barcode_logic]
+ p = latents['p'][analyzed_barcode_logic]
+ epsilon = latents['epsilon'][analyzed_barcode_logic]
+
+ # Inferred ambient gene expression vector.
+ ambient_expression_trimmed = pyro.param('chi_ambient').detach().cpu().numpy()
+
+ # Convert the indices from trimmed gene set to original gene indices.
+ ambient_expression = np.zeros(dataset_obj.data['matrix'].shape[1])
+ ambient_expression[dataset_obj.analyzed_gene_inds] = ambient_expression_trimmed
+
+ # Some summary statistics:
+ # Fraction of counts in each droplet that were removed.
+ raw_count_matrix = dataset_obj.data['matrix'][dataset_obj.analyzed_barcode_inds, :] # need all genes
+ raw_counts_droplet = np.array(raw_count_matrix.sum(axis=1)).squeeze()
+ out_counts_droplet = np.array(denoised_count_matrix[dataset_obj.analyzed_barcode_inds, :]
+ .sum(axis=1)).squeeze()
+ background_fraction = ((raw_counts_droplet - out_counts_droplet) /
+ (raw_counts_droplet + 0.001))[analyzed_barcode_logic]
+
+ # Handle the optional rho parameters.
+ rho = None
+ if (('rho_alpha' in pyro.get_param_store().keys())
+ and ('rho_beta' in pyro.get_param_store().keys())):
+ rho = np.array([pyro.param('rho_alpha').detach().cpu().numpy().item(),
+ pyro.param('rho_beta').detach().cpu().numpy().item()])
+
+ # Determine metadata fields.
+ # Wrap in lists to avoid scanpy loading bug
+ # which may already be fixed by https://github.com/scverse/scanpy/pull/2344
+ metadata = {'learning_curve': learning_curve,
+ 'barcodes_analyzed': dataset_obj.data['barcodes'][dataset_obj.analyzed_barcode_inds],
+ 'barcodes_analyzed_inds': dataset_obj.analyzed_barcode_inds,
+ 'features_analyzed_inds': dataset_obj.analyzed_gene_inds,
+ 'fraction_data_used_for_testing': 1. - consts.TRAINING_FRACTION,
+ 'target_false_positive_rate': fpr}
+ for k in ['posterior_regularization', 'posterior_regularization_kwargs',
+ 'estimator', 'estimator_kwargs']:
+ val = locals().get(k) # give me the input variable with this name
+ if val is not None:
+ if type(val) != dict:
+ if type(val) != list:
+ val = [val] # wrap in a list, unless it's a dict
+ metadata.update({k: val})
+
+ # Write h5.
+ write_succeeded = write_matrix_to_cellranger_h5(
+ cellranger_version=3, # always write v3 format output
+ output_file=file,
+ gene_names=dataset_obj.data['gene_names'],
+ gene_ids=dataset_obj.data['gene_ids'],
+ feature_types=dataset_obj.data['feature_types'],
+ genomes=dataset_obj.data['genomes'],
+ barcodes=dataset_obj.data['barcodes'][barcode_inds],
+ count_matrix=denoised_count_matrix[barcode_inds, :],
+ local_latents={'barcode_indices_for_latents': dataset_obj.analyzed_barcode_inds,
+ 'gene_expression_encoding': z,
+ 'cell_size': d,
+ 'cell_probability': p,
+ 'droplet_efficiency': epsilon,
+ 'background_fraction': background_fraction},
+ global_latents={'ambient_expression': ambient_expression,
+ 'empty_droplet_size_lognormal_loc': pyro.param('d_empty_loc').item(),
+ 'empty_droplet_size_lognormal_scale': pyro.param('d_empty_scale').item(),
+ 'cell_size_lognormal_std': pyro.param('d_cell_scale').item(),
+ 'swapping_fraction_dist_params': rho},
+ metadata=metadata,
+ )
+ return write_succeeded
+
+
+def collect_output_metrics(dataset_obj: SingleCellRNACountsDataset,
+ inferred_count_matrix: sp.csr_matrix,
+ fpr: Union[float, str],
+ cell_logic,
+ loss) -> pd.DataFrame:
+ """Create a table with a few output metrics. The idea is for these to
+ potentially be used by people creating automated pipelines."""
+
+ # Compute some metrics
+ input_count_matrix = dataset_obj.data['matrix'][dataset_obj.analyzed_barcode_inds, :]
+ total_raw_counts = dataset_obj.data['matrix'].sum()
+ total_output_counts = inferred_count_matrix.sum()
+ total_counts_removed = total_raw_counts - total_output_counts
+ fraction_counts_removed = total_counts_removed / total_raw_counts
+ total_raw_counts_in_nonempty_droplets = input_count_matrix[cell_logic].sum()
+ total_counts_removed_from_nonempty_droplets = \
+ total_raw_counts_in_nonempty_droplets - inferred_count_matrix.sum()
+ fraction_counts_removed_from_nonempty_droplets = \
+ total_counts_removed_from_nonempty_droplets / total_raw_counts_in_nonempty_droplets
+ average_counts_removed_per_nonempty_droplet = \
+ total_counts_removed_from_nonempty_droplets / cell_logic.sum()
+ expected_cells = dataset_obj.priors['expected_cells']
+ found_cells = cell_logic.sum()
+ average_counts_per_cell = inferred_count_matrix.sum() / found_cells
+ ratio_of_found_cells_to_expected_cells = \
+ None if (expected_cells is None) else (found_cells / expected_cells)
+ found_empties = len(dataset_obj.analyzed_barcode_inds) - found_cells
+ fraction_of_analyzed_droplets_that_are_nonempty = \
+ found_cells / len(dataset_obj.analyzed_barcode_inds)
+ if len(loss['train']['elbo']) > 20:
+ # compare mean ELBO increase over last 3 steps to the typical end(ish) fluctuations
+ convergence_indicator = (np.mean(np.abs([(loss['train']['elbo'][i]
+ - loss['train']['elbo'][i - 1])
+ for i in range(-3, -1)]))
+ / np.std(loss['train']['elbo'][-20:]))
+ else:
+ convergence_indicator = 'not enough training epochs to compute (requires more than 20)'
+ if len(loss['train']['elbo']) > 0:
+ overall_change_in_train_elbo = loss['train']['elbo'][-1] - loss['train']['elbo'][0]
+ else:
+ overall_change_in_train_elbo = 0 # zero epoch initialization
+
+ all_metrics_dict = \
+ {'total_raw_counts': total_raw_counts,
+ 'total_output_counts': total_output_counts,
+ 'total_counts_removed': total_counts_removed,
+ 'fraction_counts_removed': fraction_counts_removed,
+ 'total_raw_counts_in_cells':
+ total_raw_counts_in_nonempty_droplets,
+ 'total_counts_removed_from_cells':
+ total_counts_removed_from_nonempty_droplets,
+ 'fraction_counts_removed_from_cells':
+ fraction_counts_removed_from_nonempty_droplets,
+ 'average_counts_removed_per_cell':
+ average_counts_removed_per_nonempty_droplet,
+ 'target_fpr': fpr,
+ 'expected_cells': expected_cells,
+ 'found_cells': found_cells,
+ 'output_average_counts_per_cell': average_counts_per_cell,
+ 'ratio_of_found_cells_to_expected_cells':
+ ratio_of_found_cells_to_expected_cells,
+ 'found_empties': found_empties,
+ 'fraction_of_analyzed_droplets_that_are_nonempty':
+ fraction_of_analyzed_droplets_that_are_nonempty,
+ 'convergence_indicator': convergence_indicator,
+ 'overall_change_in_train_elbo': overall_change_in_train_elbo}
+
+ return pd.DataFrame(data=all_metrics_dict,
+ index=['metric']).transpose()
+
+
+def write_cell_barcodes_csv(bc_file_name: str, cell_barcodes: np.ndarray):
+ """Write the cell barcode CSV file.
+
+ Args:
+ bc_file_name: Output CSV file
+ cell_barcodes: Array of the cell barcode names
+
+ """
+
+ # Save barcodes determined to contain cells as _cell_barcodes.csv
+ try:
+ barcode_names = np.array([str(cell_barcodes[i], encoding='UTF-8')
+ for i in range(cell_barcodes.size)])
+ except UnicodeDecodeError:
+ # necessary if barcodes are ints
+ barcode_names = cell_barcodes
+ except TypeError:
+ # necessary if barcodes are already decoded
+ barcode_names = cell_barcodes
+ np.savetxt(bc_file_name, barcode_names, delimiter=',', fmt='%s')
+ logger.info(f"Saved cell barcodes in {bc_file_name}")
+
+
+def get_optimizer(n_batches: int,
+ batch_size: int,
+ epochs: int,
+ learning_rate: float,
+ constant_learning_rate: bool,
+ total_epochs_for_testing_only: Optional[int] = None) \
+ -> Union[pyro.optim.PyroOptim, pyro.optim.lr_scheduler.PyroLRScheduler]:
+ """Get optimizer or learning rate scheduler (if using one)"""
+
+ # Set up the optimizer.
+ optimizer = pyro.optim.clipped_adam.ClippedAdam # just ClippedAdam does not work
+ optimizer_args = {'lr': learning_rate, 'clip_norm': 10.}
+
+ # Set up a learning rate scheduler.
+ if total_epochs_for_testing_only is not None:
+ total_steps = n_batches * total_epochs_for_testing_only
+ else:
+ total_steps = n_batches * epochs
+ scheduler_args = {'optimizer': optimizer,
+ 'max_lr': learning_rate * 10,
+ 'total_steps': total_steps,
+ 'optim_args': optimizer_args}
+ scheduler = pyro.optim.OneCycleLR(scheduler_args)
+
+ # Constant learning rate overrides the above and uses no scheduler.
+ if constant_learning_rate:
+ logger.info('Using ClippedAdam --constant-learning-rate rather than '
+ 'the OneCycleLR schedule. This is not usually recommended.')
+ scheduler = ClippedAdam(optimizer_args)
+
+ return scheduler
+
+
+def run_inference(dataset_obj: SingleCellRNACountsDataset,
+ args: argparse.Namespace,
+ output_checkpoint_tarball: str = consts.CHECKPOINT_FILE_NAME,
+ total_epochs_for_testing_only: Optional[int] = None)\
+ -> Tuple[RemoveBackgroundPyroModel, pyro.optim.PyroOptim, DataLoader, DataLoader]:
+ """Run a full inference procedure, training a latent variable model.
+
+ Args:
+ dataset_obj: Input data in the form of a SingleCellRNACountsDataset
+ object.
+ args: Input command line parsed arguments.
+ output_checkpoint_tarball: Intended checkpoint tarball filepath.
+ total_epochs_for_testing_only: Hack for testing code using LR scheduler
+
+ Returns:
+ model: cellbender.model.RemoveBackgroundPyroModel that has had
+ inference run.
+
+ """
+
+ # Get the checkpoint file base name with hash, which we stored in args.
+ checkpoint_filename = args.checkpoint_filename
+
+ # Configure pyro options (skip validations to improve speed).
+ pyro.enable_validation(False)
+ pyro.distributions.enable_validation(False)
+
+ # Set random seed, updating global state of python, numpy, and torch RNGs.
+ pyro.clear_param_store()
+ pyro.set_rng_seed(consts.RANDOM_SEED)
+ if args.use_cuda:
+ torch.cuda.manual_seed_all(consts.RANDOM_SEED)
+
+ # Attempt to load from a previously-saved checkpoint.
+ ckpt = attempt_load_checkpoint(filebase=checkpoint_filename,
+ tarball_name=args.input_checkpoint_tarball,
+ force_device='cuda:0' if args.use_cuda else 'cpu')
+ ckpt_loaded = ckpt['loaded'] # True if a checkpoint was loaded successfully
+
+ if ckpt_loaded:
+
+ model = ckpt['model']
+ scheduler = ckpt['optim']
+ train_loader = ckpt['train_loader']
+ test_loader = ckpt['test_loader']
+ if hasattr(ckpt['args'], 'num_failed_attempts'):
+ # update this from the checkpoint file, if present
+ args.num_failed_attempts = ckpt['args'].num_failed_attempts
+ for obj in [model, scheduler, train_loader, test_loader, args]:
+ assert obj is not None, \
+ f'Expected checkpoint to contain model, scheduler, train_loader, ' \
+ f'test_loader, and args; but some are None:\n{ckpt}'
+ logger.info('Checkpoint loaded successfully.')
+
+ else:
+
+ logger.info('No checkpoint loaded.')
+
+ # Get the trimmed count matrix (transformed if called for).
+ count_matrix = dataset_obj.get_count_matrix()
+
+ # Set up the variational autoencoder:
+
+ # Encoder.
+ encoder_z = EncodeZ(input_dim=count_matrix.shape[1],
+ hidden_dims=args.z_hidden_dims,
+ output_dim=args.z_dim,
+ use_batch_norm=False,
+ use_layer_norm=False,
+ input_transform='normalize')
+
+ encoder_other = EncodeNonZLatents(n_genes=count_matrix.shape[1],
+ z_dim=args.z_dim,
+ log_count_crossover=dataset_obj.priors['log_counts_crossover'],
+ prior_log_cell_counts=np.log1p(dataset_obj.priors['cell_counts']),
+ empty_log_count_threshold=np.log1p(dataset_obj.empty_UMI_threshold),
+ prior_logit_cell_prob=dataset_obj.priors['cell_logit'],
+ input_transform='log_normalize')
+
+ encoder = CompositeEncoder({'z': encoder_z,
+ 'other': encoder_other})
+
+ # Decoder.
+ decoder = Decoder(input_dim=args.z_dim,
+ hidden_dims=args.z_hidden_dims[::-1],
+ use_batch_norm=True,
+ use_layer_norm=False,
+ output_dim=count_matrix.shape[1])
+
+ # Set up the pyro model for variational inference.
+ model = RemoveBackgroundPyroModel(model_type=args.model,
+ encoder=encoder,
+ decoder=decoder,
+ dataset_obj_priors=dataset_obj.priors,
+ n_analyzed_genes=dataset_obj.analyzed_gene_inds.size,
+ n_droplets=dataset_obj.analyzed_barcode_inds.size,
+ analyzed_gene_names=dataset_obj.data['gene_names'][dataset_obj.analyzed_gene_inds],
+ empty_UMI_threshold=dataset_obj.empty_UMI_threshold,
+ log_counts_crossover=dataset_obj.priors['log_counts_crossover'],
+ use_cuda=args.use_cuda)
+
+ # Load the dataset into DataLoaders.
+ frac = args.training_fraction # Fraction of barcodes to use for training
+ batch_size = int(min(consts.MAX_BATCH_SIZE, frac * dataset_obj.analyzed_barcode_inds.size / 2))
+
+ # Set up dataloaders.
+ train_loader, test_loader = \
+ prep_data_for_training(dataset=count_matrix,
+ empty_drop_dataset=dataset_obj.get_count_matrix_empties(),
+ batch_size=batch_size,
+ training_fraction=frac,
+ fraction_empties=args.fraction_empties,
+ shuffle=True,
+ use_cuda=args.use_cuda)
+
+ # Set up optimizer (optionally wrapped in a learning rate scheduler).
+ scheduler = get_optimizer(
+ n_batches=len(train_loader),
+ batch_size=train_loader.batch_size,
+ epochs=args.epochs,
+ learning_rate=args.learning_rate,
+ constant_learning_rate=args.constant_learning_rate,
+ total_epochs_for_testing_only=total_epochs_for_testing_only,
+ )
+
+ # Determine the loss function.
+ if args.use_jit:
+
+ # Call guide() once as a warm-up.
+ # model.guide(torch.zeros([10, dataset_obj.analyzed_gene_inds.size]).to(model.device))
+
+ if args.model == "simple":
+ loss_function = JitTrace_ELBO()
+ else:
+ loss_function = JitTraceEnum_ELBO(max_plate_nesting=1,
+ strict_enumeration_warning=False)
+ else:
+
+ if args.model == "simple":
+ loss_function = Trace_ELBO()
+ else:
+ loss_function = TraceEnum_ELBO(max_plate_nesting=1)
+
+ # Set up the inference process.
+ svi = SVI(model.model, model.guide, scheduler, loss=loss_function)
+
+ # Run training.
+ if args.epochs == 0:
+ logger.info("Zero epochs specified... will only initialize the model.")
+ model.guide(train_loader.__next__())
+ train_loader.reset_ptr()
+
+ # Even though it's not much of a checkpoint, we still need one for subsequent steps.
+ save_checkpoint(filebase=checkpoint_filename,
+ tarball_name=output_checkpoint_tarball,
+ args=args,
+ model_obj=model,
+ scheduler=svi.optim,
+ train_loader=train_loader,
+ test_loader=test_loader)
+
+ else:
+ logger.info("Running inference...")
+ try:
+ run_training(model=model,
+ args=args,
+ svi=svi,
+ train_loader=train_loader,
+ test_loader=test_loader,
+ epochs=args.epochs,
+ test_freq=5,
+ output_filename=checkpoint_filename,
+ ckpt_tarball_name=output_checkpoint_tarball,
+ checkpoint_freq=args.checkpoint_min,
+ epoch_elbo_fail_fraction=args.epoch_elbo_fail_fraction,
+ final_elbo_fail_fraction=args.final_elbo_fail_fraction)
+
+ except ElboException:
+
+ logger.warning(traceback.format_exc())
+
+ # Keep track of number of failed attempts.
+ if not hasattr(args, 'num_failed_attempts'):
+ args.num_failed_attempts = 1
+ else:
+ args.num_failed_attempts = args.num_failed_attempts + 1
+ logger.debug(f'Training failed, and the number of failed attempts '
+ f'on record is {args.num_failed_attempts}')
+
+ # Retry training with reduced learning rate, if indicated by user.
+ logger.debug(f'Number of times to retry training is {args.num_training_tries}')
+ if args.num_failed_attempts < args.num_training_tries:
+ args.learning_rate = args.learning_rate * args.learning_rate_retry_mult
+ logger.info(f'Restarting training: attempt {args.num_failed_attempts + 1}, '
+ f'learning_rate = {args.learning_rate}')
+ run_remove_background(args) # start from scratch
+ sys.exit(0)
+ else:
+ logger.info(f'No more attempts are specified by --num-training-tries. '
+ f'Therefore the workflow will abort here.')
+ sys.exit(1)
+
+ logger.info("Inference procedure complete.")
+
+ return model, scheduler, train_loader, test_loader
diff --git a/cellbender/remove_background/sparse_utils.py b/cellbender/remove_background/sparse_utils.py
new file mode 100644
index 00000000..4a0f26fa
--- /dev/null
+++ b/cellbender/remove_background/sparse_utils.py
@@ -0,0 +1,101 @@
+"""Utility functions for working with scipy sparse matrices"""
+
+import torch
+import numpy as np
+import scipy.sparse as sp
+
+from typing import Optional, Tuple, Iterable
+
+
+@torch.no_grad()
+def dense_to_sparse_op_torch(t: torch.Tensor,
+ tensor_for_nonzeros: Optional[torch.Tensor] = None) \
+ -> Tuple[np.ndarray, ...]:
+ """Converts dense matrix to sparse COO format tuple of numpy arrays (*indices, data)
+
+ Args:
+ t: The Tensor
+ tensor_for_nonzeros: If this is not None, then this tensor will be used
+ to determine nonzero indices, while the values will come from t.
+
+ Returns:
+ Tuple of
+ (nonzero_inds_dim0, nonzero_inds_dim1, ...) + (nonzero_values, )
+
+ """
+
+ if tensor_for_nonzeros is None:
+ tensor_for_nonzeros = t
+
+ nonzero_inds_tuple = torch.nonzero(tensor_for_nonzeros, as_tuple=True)
+ nonzero_values = t[nonzero_inds_tuple].flatten()
+
+ return tuple([ten.cpu().numpy() for ten in (nonzero_inds_tuple + (nonzero_values,))])
+
+
+def log_prob_sparse_to_dense(coo: sp.coo_matrix) -> np.ndarray:
+ """Densify a sparse log prob COO data structure. Same as coo_matrix.todense()
+ except it fills missing entries with -np.inf instead of 0, since 0 is a
+ very meaningful quantity for log prob.
+ """
+ return todense_fill(coo=coo, fill_value=-np.inf)
+
+
+def todense_fill(coo: sp.coo_matrix, fill_value: float) -> np.ndarray:
+ """Densify a sparse COO matrix. Same as coo_matrix.todense()
+ except it fills missing entries with fill_value instead of 0.
+ """
+ dummy_value = np.nan if not np.isnan(fill_value) else np.inf
+ dummy_check = np.isnan if np.isnan(dummy_value) else np.isinf
+ coo = coo.copy().astype(float)
+ coo.data[coo.data == 0] = dummy_value
+ out = np.array(coo.todense()).squeeze()
+ out[out == 0] = fill_value
+ out[dummy_check(out)] = 0
+ return out
+
+
+def csr_set_rows_to_zero(csr: sp.csr_matrix,
+ row_inds: Iterable[int]) -> sp.csr_matrix:
+ """Set all nonzero elements in rows "row_inds" to zero.
+ Happens in-place, although output is returned as well.
+
+ https://stackoverflow.com/questions/12129948/scipy-sparse-set-row-to-zeros
+ """
+
+ if not isinstance(csr, sp.csr_matrix):
+ try:
+ csr = csr.tocsr()
+ except Exception:
+ raise ValueError('Matrix given must be of CSR format.')
+ for row in row_inds:
+ csr.data[csr.indptr[row]:csr.indptr[row + 1]] = 0
+ csr.eliminate_zeros()
+ return csr
+
+
+def overwrite_matrix_with_columns_from_another(mat1: sp.csc_matrix,
+ mat2: sp.csc_matrix,
+ column_inds: np.ndarray) -> sp.csc_matrix:
+ """Given two sparse matrices of the same shape, replace columns that are not
+ in `column_inds` in `mat1` with the entries from `mat2`.
+ """
+ column_inds = set(column_inds)
+
+ mat1 = mat1.copy().tocsr()
+ mat2 = mat2.copy().tocsr() # failure to copy could overwrite actual count data
+
+ # Zero out values in mat2 that are in the specified columns.
+ inds = np.where([i in column_inds for i in mat2.indices])[0]
+ mat2.data[inds] = 0
+ mat2.eliminate_zeros()
+
+ # Zero out values in mat1 that are not in the specified columns.
+ inds = np.where([i not in column_inds for i in mat1.indices])[0]
+ mat1.data[inds] = 0
+ mat1.eliminate_zeros()
+
+ # Put in the new values by addition.
+ output = mat1 + mat2
+
+ return output.tocsc()
diff --git a/cellbender/remove_background/tests/benchmarking/README.md b/cellbender/remove_background/tests/benchmarking/README.md
new file mode 100644
index 00000000..e52c1423
--- /dev/null
+++ b/cellbender/remove_background/tests/benchmarking/README.md
@@ -0,0 +1,62 @@
+# Benchmarking CellBender versions and changes
+
+## Running benchmarking tests
+
+There is a benchmarking WDL in /cellbender/remove_background/tests/benchmarking called
+`benchmark.wdl`. This can be run using a local installation of cromshell
+(`conda install -c bioconda cromshell`) by running the python script in
+`run_benchmark.py` as in
+
+```console
+python run_benchmark.py \
+ --git "v0.2.1" \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/pbmc8k_raw_gene_bc_matrices.h5" \
+ --sample pbmc8k
+```
+
+The `--git` input can be a full commit sha or a tag or a branch. Anything that
+can be pip installed via `pip install -y git+https://github.com/broadinstitute/CellBender.git@<SOMETHING>`
+
+(The caveat here is that older versions of CellBender cannot be installed from
+github in this way, so this is a test that will work moving forward from
+commit `cb2d209d5aedffe71e28947bc5b7859600aef64d`)
+
+### Datasets from the paper
+
+To re-run the datasets that were analyzed in the CellBender paper, plus
+several other interesting benchmarks, try running
+
+```console
+./run_usual_benchmarks.sh b9d2953c76c538d13549290bd986af4e6a1780d5
+```
+
+You can check their status via `cromshell` using
+
+```console
+cromshell list -u | tail -n 5
+```
+
+## Validation of outputs
+
+Validation is comprised of a Terra notebook in the following workspace:
+https://app.terra.bio/#workspaces/broad-firecloud-dsde-methods/sfleming_dev
+
+To prepare a data table for the Terra workspace (once all jobs
+have completed):
+
+```console
+python run_benchmark_result_tabulation.py \
+ --workflows b9d2953c76c538d13549290bd986af4e6a1780d5 \
+ --output samples.tsv
+```
+
+where, for example, `b9d2953c76c538d13549290bd986af4e6a1780d5` is
+the git hash used to run `run_usual_benchmarks.sh`. (The script will `grep`
+the local cromshell database for all runs that match. More than one search
+term can be used.)
+
+Then manually upload the `samples.tsv` to Terra as a data table.
+
+Open the validation notebook and enter the git commit
+`b9d2953c76c538d13549290bd986af4e6a1780d5` at the top. Then
+run the entire notebook to produce plots.
diff --git a/cellbender/remove_background/tests/benchmarking/benchmark.wdl b/cellbender/remove_background/tests/benchmarking/benchmark.wdl
new file mode 100644
index 00000000..6a0781b8
--- /dev/null
+++ b/cellbender/remove_background/tests/benchmarking/benchmark.wdl
@@ -0,0 +1,29 @@
+version 1.0
+
+import "cellbender_remove_background.wdl" as cellbender
+
+## Copyright Broad Institute, 2023
+##
+## LICENSING :
+## This script is released under the WDL source code license (BSD-3)
+## (see LICENSE in https://github.com/openwdl/wdl).
+
+
+workflow run_cellbender_benchmark {
+ input {
+ String? git_hash
+ }
+
+ call cellbender.run_cellbender_remove_background_gpu as cb {
+ input:
+ dev_git_hash__=git_hash
+
+ }
+
+ output {
+ File log = cb.log
+ File summary_pdf = cb.pdf
+ Array[File] html_report_array = cb.report_array
+ Array[File] h5_array = cb.h5_array
+ }
+}
diff --git a/cellbender/remove_background/tests/benchmarking/benchmark_inputs.json b/cellbender/remove_background/tests/benchmarking/benchmark_inputs.json
new file mode 100644
index 00000000..dea4a2e5
--- /dev/null
+++ b/cellbender/remove_background/tests/benchmarking/benchmark_inputs.json
@@ -0,0 +1,19 @@
+{
+ "run_cellbender_benchmark.git_hash": "GIT_HASH",
+ "run_cellbender_benchmark.cb.input_file_unfiltered": "INPUT_GSURL",
+ "run_cellbender_benchmark.cb.sample_name": "SAMPLE_NAME",
+ "run_cellbender_benchmark.cb.truth_file": "TRUTH_GSURL",
+ "run_cellbender_benchmark.cb.exclude_feature_types": "Peaks",
+ "run_cellbender_benchmark.cb.expected_cells": null,
+ "run_cellbender_benchmark.cb.total_droplets_included": null,
+ "run_cellbender_benchmark.cb.low_count_threshold": null,
+ "run_cellbender_benchmark.cb.fpr": "FPR",
+ "run_cellbender_benchmark.cb.debug": false,
+ "run_cellbender_benchmark.cb.checkpoint_mins": 200,
+ "run_cellbender_benchmark.cb.checkpoint_file": null,
+ "run_cellbender_benchmark.cb.hardware_preemptible_tries": 0,
+ "run_cellbender_benchmark.cb.hardware_cpu_count": 8,
+ "run_cellbender_benchmark.cb.hardware_memory_GB": 64,
+ "run_cellbender_benchmark.cb.estimator_multiple_cpu": null,
+ "run_cellbender_benchmark.cb.docker_image": "us.gcr.io/broad-dsde-methods/cellbender:0.2.1"
+}
\ No newline at end of file
diff --git a/cellbender/remove_background/tests/benchmarking/cuda_check_inputs.json b/cellbender/remove_background/tests/benchmarking/cuda_check_inputs.json
new file mode 100644
index 00000000..b9a3c46b
--- /dev/null
+++ b/cellbender/remove_background/tests/benchmarking/cuda_check_inputs.json
@@ -0,0 +1,4 @@
+{
+ "check_pytorch_cuda_status.run_check_pytorch_cuda_status.docker_image": "us.gcr.io/broad-dsde-methods/cellbender:20230427"
+}
+
diff --git a/cellbender/remove_background/tests/benchmarking/docker_image_check_cuda_status.wdl b/cellbender/remove_background/tests/benchmarking/docker_image_check_cuda_status.wdl
new file mode 100644
index 00000000..8763bfe9
--- /dev/null
+++ b/cellbender/remove_background/tests/benchmarking/docker_image_check_cuda_status.wdl
@@ -0,0 +1,38 @@
+version 1.0
+
+## Copyright Broad Institute, 2023
+##
+## LICENSING :
+## This script is released under the WDL source code license (BSD-3)
+## (see LICENSE in https://github.com/openwdl/wdl).
+
+task run_check_pytorch_cuda_status {
+ input {
+ String docker_image
+ Int? hardware_boot_disk_size_GB = 20
+ String? hardware_zones = "us-east1-d us-east1-c us-central1-a us-central1-c us-west1-b"
+ String? hardware_gpu_type = "nvidia-tesla-t4"
+ String? nvidia_driver_version = "470.82.01" # need >=465.19.01 for CUDA 11.3
+ }
+ command {
+ set -e
+ python <<CODE
+ import torch
+ assert torch.cuda.is_available()
+ CODE
+ }
+ runtime {
+ docker: "${docker_image}"
+ bootDiskSizeGb: hardware_boot_disk_size_GB
+ zones: "${hardware_zones}"
+ gpuCount: 1
+ gpuType: "${hardware_gpu_type}"
+ nvidiaDriverVersion: "${nvidia_driver_version}"
+ maxRetries: 0
+ }
+}
+
+
+workflow check_pytorch_cuda_status {
+ call run_check_pytorch_cuda_status
+}
diff --git a/cellbender/remove_background/tests/benchmarking/run_benchmark.py b/cellbender/remove_background/tests/benchmarking/run_benchmark.py
new file mode 100644
index 00000000..efb8520a
--- /dev/null
+++ b/cellbender/remove_background/tests/benchmarking/run_benchmark.py
@@ -0,0 +1,225 @@
+"""Run a github commit of CellBender on test data for benchmarking purposes
+Example:
+ $ conda activate cromshell
+ $ python run_benchmark.py \
+ --git 2927346f4c513a217ac8ad076e494dd1adbf70e1 \
+ --input "gs://path" \
+ --sample "name" \
+ --truth "gs://path_to_truth"
+"""
+
+import tempfile
+import json
+import os
+import sys
+import subprocess
+import shutil
+import random
+import argparse
+from typing import Tuple, Dict, Optional, List
+
+
+def get_parser() -> argparse.ArgumentParser:
+ parser = argparse.ArgumentParser(
+ description="Benchmark a specific commit of CellBender on a dataset "
+ "using cromshell.",
+ )
+
+ parser.add_argument('-g', '--git',
+ type=str,
+ required=False,
+ default=None,
+ dest='git_hash',
+ help='Specific github commit from the CellBender repo: '
+ 'can be a tag, a branch, or a commit sha. Not '
+ 'including the flag will run using the base '
+ 'docker image specified in benchmark_inputs.json')
+ parser.add_argument('-i', '--input',
+ type=str,
+ required=True,
+ dest='input_file',
+ help='Input path (must be a gsURL, i.e. '
+ '"gs://bucket/path/file.h5")')
+ parser.add_argument('-s', '--sample',
+ type=str,
+ required=True,
+ dest='sample',
+ help='Sample name: used to create output file names')
+ parser.add_argument('-t', '--truth',
+ type=str,
+ required=False,
+ default=None,
+ dest='truth_file',
+ help='Truth data path for simulated data (must be a '
+ 'gsURL, i.e. "gs://bucket/path/file.h5")')
+ parser.add_argument('-f', '--fpr',
+ type=str,
+ required=False,
+ default="0.01",
+ dest='fpr',
+ help='FPR for CellBender')
+ parser.add_argument('-q', '--quiet',
+ action='store_true',
+ default=False,
+ dest='quiet',
+ help='Include this flag to avoid printing to stdout')
+
+ return parser
+
+
+def update_input_json(template_file: str,
+ substitutions: Dict[str, str],
+ tmpdir: tempfile.TemporaryDirectory,
+ verbose: bool = True) -> Tuple[str, str]:
+ """Create a new input json for a Cromwell run based on a template"""
+
+ # read template into dict
+ with open(template_file, mode='r') as f:
+ template = json.load(f)
+
+ # perform substitutions
+ for key, value in template.items():
+ replacement = substitutions.get(value, '__absent')
+ template[key] = replacement if (replacement != '__absent') else value
+
+ # write filled-in input json
+ json_inputs_file = os.path.join(tmpdir, 'inputs.json')
+ with open(json_inputs_file, mode='w') as f:
+ json.dump(template, f, indent=2)
+
+ # write options json that disables call-caching
+ json_options_file = os.path.join(tmpdir, 'options.json')
+ options_dict = {
+ "write_to_cache": False,
+ "read_from_cache": False,
+ }
+ with open(json_options_file, mode='w') as f:
+ json.dump(options_dict, f, indent=2)
+
+ if verbose:
+ print('Contents of inputs.json:')
+ subprocess.run(['cat', json_inputs_file])
+ print('', end='\n')
+
+ if verbose:
+ print('Contents of options.json:')
+ subprocess.run(['cat', json_options_file])
+ print('', end='\n')
+
+ return json_inputs_file, json_options_file
+
+
+def cromshell_submit(wdl: str,
+ inputs: str,
+ options: str,
+ dependencies: List[str],
+ tmpdir: tempfile.TemporaryDirectory,
+ alias: Optional[str] = None,
+ verbose: bool = True) -> Tuple[str, str]:
+ """Submit job via cromshell and return the workflow-id and alias
+
+ NOTE: the whole dependency zipping thing is way more difficult than it has
+ to be and took me hours. it is super fragile. but at least now I think I
+ can put this benchmark.wdl anywhere I want to.
+ """
+
+ # zip up dependencies
+ if verbose:
+ print('Zipping dependencies')
+ dependencies_zip = os.path.join(tmpdir, 'dependencies.zip')
+ dependencies_dir = tmpdir
+ for file in dependencies:
+ shutil.copy(file, dependencies_dir)
+ subprocess.run(['zip', '-j', dependencies_zip,
+ ' '.join([os.path.join(dependencies_dir, os.path.basename(f))
+ for f in dependencies])])
+
+ # move WDL to tmpdir
+ tmp_wdl = os.path.join(tmpdir, os.path.basename(wdl))
+ shutil.copy(wdl, tmp_wdl)
+
+ submit_cmd = ['cromshell', 'submit',
+ tmp_wdl,
+ inputs,
+ options,
+ dependencies_zip]
+
+ # submit job
+ if verbose:
+ print(f'Submitting WDL {tmp_wdl}')
+ current_path = os.getcwd()
+ os.chdir(tmpdir)
+ out = subprocess.run(submit_cmd)
+ os.chdir(current_path)
+ out.check_returncode() # error if this failed
+
+ # get workflow-id
+ out = subprocess.run(['cromshell', 'status'], capture_output=True)
+ d = json.loads(out.stdout)
+ workflow_id = d.get('id')
+
+ # alias job
+ if alias is not None:
+ # solve the issue where an alias cannot be made twice
+ hash = random.getrandbits(128)
+ alias = alias + '__runhash_' + str(hash)[:10]
+ subprocess.run(['cromshell', 'alias', '-1', alias])
+
+ return workflow_id, alias
+
+
+if __name__ == '__main__':
+
+ # handle input arguments
+ parser = get_parser()
+ args = parser.parse_args(sys.argv[1:])
+ # args = validate_args(args)
+
+ with tempfile.TemporaryDirectory() as tmpdir:
+
+ # determine substitutions to be made in input json
+ substitutions = {
+ 'GIT_HASH': args.git_hash,
+ 'INPUT_GSURL': args.input_file,
+ 'SAMPLE_NAME': args.sample,
+ 'TRUTH_GSURL': args.truth_file,
+ 'FPR': args.fpr,
+ }
+
+ # create the updated input json
+ inputs_json, options_json = update_input_json(
+ template_file='benchmark_inputs.json',
+ substitutions=substitutions,
+ tmpdir=tmpdir,
+ verbose=not args.quiet,
+ )
+
+ # get the path to the cellbender WDL
+ this_dir = os.path.dirname(os.path.abspath(__file__))
+ cellbender_wdl_path = os.path.abspath(os.path.join(
+ this_dir, '..', '..', '..', '..', 'wdl',
+ 'cellbender_remove_background.wdl',
+ ))
+
+ # run cromshell submit
+ if args.git_hash is not None:
+ alias = '_'.join(['cellbender', args.sample, args.git_hash])
+ else:
+ alias = '_'.join(['cellbender', args.sample])
+ workflow_id, alias = cromshell_submit(
+ wdl=os.path.join(this_dir, 'benchmark.wdl'),
+ inputs=inputs_json,
+ options=options_json,
+ dependencies=[cellbender_wdl_path],
+ alias=alias,
+ tmpdir=tmpdir,
+ )
+
+ # show workflow-id alias
+ if not args.quiet:
+ print('Sucessfully submitted job:')
+ if alias is not None:
+ print(alias)
+ print(workflow_id)
+
+ sys.exit(0)
diff --git a/cellbender/remove_background/tests/benchmarking/run_benchmark_result_tabulation.py b/cellbender/remove_background/tests/benchmarking/run_benchmark_result_tabulation.py
new file mode 100644
index 00000000..ffc5e5a8
--- /dev/null
+++ b/cellbender/remove_background/tests/benchmarking/run_benchmark_result_tabulation.py
@@ -0,0 +1,146 @@
+"""Tabulate outputs from CellBender benchmarking as a samples.tsv for Terra
+Example:
+ $ conda activate cromshell
+ $ python run_benchmark_result_tabulation.py \
+ --workflows 2927346f4c513a217ac8ad076e494dd1adbf70e1 \
+ --output "samples.tsv"
+"""
+
+import pandas as pd
+import os
+import re
+import sys
+import subprocess
+import argparse
+from io import StringIO
+from typing import Tuple, Dict, Optional, List, Union
+
+
+def get_parser() -> argparse.ArgumentParser:
+ parser = argparse.ArgumentParser(
+ description="Create a Terra-compatible data table TSV from cromshell "
+ "benchmarking outputs.",
+ )
+ parser.add_argument('-w', '--workflows',
+ type=str,
+ required=True,
+ dest='grep',
+ help='grep to choose workflows from the cromshell database')
+ parser.add_argument('-n', '--note',
+ type=str,
+ required=True,
+ dest='note',
+ help='Annotation describing the git commit')
+ parser.add_argument('-o', '--output',
+ type=str,
+ required=True,
+ dest='output_file',
+ help='Output TSV file')
+
+ return parser
+
+
+def find_cromshell_workflow_ids_dates(grep: str) -> Tuple[List[str], List[str]]:
+ """Find workflows in the cromshell database matching some string"""
+
+ date_col = 0
+ workflow_id_col = 2
+
+ # use cromshell list to get the database, then comb through it
+ out = subprocess.run(['cromshell', 'list'], capture_output=True)
+ s = out.stdout.decode()
+ _RE_COMBINE_WHITESPACE = re.compile(r"(?a: +)")
+ s = _RE_COMBINE_WHITESPACE.sub(" ", s).strip()
+ greps = grep.split(' ') # separately match all search elements (space delimited)
+ selected_rows = [r for r in s.split('\n') if all([g in r for g in greps])]
+
+ for r in selected_rows:
+ if ('Succeeded' not in r):
+ print(f'WARNING: Skipping\n{r}\nbecause the run has not completed')
+ selected_rows = [r for r in selected_rows if ('Succeeded' in r)]
+
+ workflow_ids = [r.split(' ')[workflow_id_col] for r in selected_rows]
+ dates = [r.split(' ')[date_col] for r in selected_rows]
+
+ return workflow_ids, dates
+
+
+def get_cromshell_output_h5(workflow: str, grep: str = '_out.h5') -> Union[str, List[str]]:
+ """Use cromshell list-outputs to get the relevant file gsURL"""
+
+ output = grep_from_command(['cromshell', 'list-outputs', workflow], grep=grep)
+ out = output[:-1].decode().split('\n')
+ if len(out) > 1:
+ return out
+ else:
+ return out[0]
+
+
+def sample_name_from_h5(h5: str) -> str:
+ """Get sample name by parsing h5 file name"""
+ return os.path.basename(h5).replace('_out.h5', '')
+
+
+def grep_from_command(command: List[str], grep: str) -> bytes:
+ """Pipe a command to grep and give the output"""
+ ps = subprocess.Popen(command, stdout=subprocess.PIPE)
+ output = subprocess.check_output(['grep', grep], stdin=ps.stdout)
+ ps.wait()
+ return output
+
+
+def metadata_from_workflow_id(workflow: str) -> Tuple[str, str, Optional[str]]:
+ """Get other metadata for a workflow id: git hash, input file, truth file"""
+
+ # git hash
+ output = grep_from_command(['cromshell', 'metadata', workflow],
+ grep='"git_hash":')
+ git_hash = output[17:-3].decode()
+
+ # input file
+ output = grep_from_command(['cromshell', 'metadata', workflow],
+ grep='run_cellbender_benchmark.cb.input_file_unfiltered')
+ input_file = output[58:-3].decode()
+
+ # truth file
+ output = grep_from_command(['cromshell', 'metadata', workflow],
+ grep='run_cellbender_benchmark.cb.truth_file')
+ if 'null' not in output.decode():
+ truth_file = output[47:-3].decode()
+ else:
+ truth_file = None
+
+ return git_hash, input_file, truth_file
+
+
+if __name__ == '__main__':
+
+ # handle input arguments
+ parser = get_parser()
+ args = parser.parse_args(sys.argv[1:])
+ # args = validate_args(args)
+
+ workflow_ids, dates = find_cromshell_workflow_ids_dates(args.grep)
+ output_h5s = [get_cromshell_output_h5(id) for id in workflow_ids]
+ samples = [sample_name_from_h5(h5) for h5 in output_h5s]
+ list_of_tuples = [metadata_from_workflow_id(id) for id in workflow_ids]
+ if len(list_of_tuples) > 0:
+ git_hashes, input_h5s, truth_h5s = zip(*list_of_tuples)
+
+ run_ids = [sample + '_' + git for sample, git in zip(samples, git_hashes)]
+
+ df = pd.DataFrame(data={'entity:sample_id': run_ids,
+ 'git_commit': git_hashes,
+ 'sample': samples,
+ 'output_h5': output_h5s,
+ 'input_h5': input_h5s,
+ 'truth_h5': truth_h5s,
+ 'cromwell_workflow_id': workflow_ids,
+ 'date_time': dates,
+ 'note': args.note})
+
+ df.to_csv(args.output_file, sep='\t', index=False)
+ else:
+ print('No submissions selected')
+
+ sys.exit(0)
diff --git a/cellbender/remove_background/tests/benchmarking/run_usual_benchmarks.sh b/cellbender/remove_background/tests/benchmarking/run_usual_benchmarks.sh
new file mode 100755
index 00000000..98c50f7a
--- /dev/null
+++ b/cellbender/remove_background/tests/benchmarking/run_usual_benchmarks.sh
@@ -0,0 +1,74 @@
+#!/bin/bash
+
+# Example run:
+# $ ./run_usual_benchmarks.sh "v0.2.1"
+
+GIT_HASH=$1
+
+# from the paper ==========================
+
+# 10x Genomics pbmc8k
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/pbmc8k_raw_gene_bc_matrices.h5" \
+ --sample "pbmc8k"
+
+# 10x Genomics hgmm12k
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/hgmm_12k_raw_gene_bc_matrices.h5" \
+ --sample "hgmm12k"
+
+# 10x Genomics pbmc5k CITE-seq
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/5k_pbmc_protein_v3_nextgem_raw_feature_bc_matrix.h5" \
+ --sample "pbmc5k" \
+ --fpr "0.1"
+
+# Broad PCL rat6k
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/PCL_rat_A_LA6_raw_feature_bc_matrix.h5" \
+ --sample "rat6k"
+
+# Simulation s1
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/s1.h5" \
+ --truth "gs://broad-dsde-methods-sfleming/cellbender_test/s1_truth.h5" \
+ --sample "s1"
+
+# Simulation s4
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/s4.h5" \
+ --truth "gs://broad-dsde-methods-sfleming/cellbender_test/s4_truth.h5" \
+ --sample "s4" \
+ --fpr "0.2"
+
+# Simulation s7, hgmm
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/s7.h5" \
+ --truth "gs://broad-dsde-methods-sfleming/cellbender_test/s7_truth.h5" \
+ --sample "s7"
+
+# additional datasets ========================
+
+# Broad PCL human PV dataset, hard time calling high-count cells in v0.2.0, wobbly learning curve
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/PCL_human_PV_1817_ls.h5" \
+ --sample "pv20k"
+
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/PCL_human_PV_1799_BPV.h5" \
+ --sample "pv15k"
+
+# public data from Caroline Porter, high UMI count cells called empty
+python run_benchmark.py \
+ --git $GIT_HASH \
+ --input "gs://broad-dsde-methods-sfleming/cellbender_test/cporter_20230331_N18_Epi_A.h5" \
+ --sample "epi2k"
diff --git a/cellbender/remove_background/tests/conftest.py b/cellbender/remove_background/tests/conftest.py
new file mode 100644
index 00000000..8bafa2da
--- /dev/null
+++ b/cellbender/remove_background/tests/conftest.py
@@ -0,0 +1,202 @@
+"""Test utility functions and session-scoped fixtures."""
+
+import pytest
+import scipy.sparse as sp
+import numpy as np
+import torch
+
+from cellbender.remove_background.data.extras.simulate import \
+ generate_sample_dirichlet_dataset
+from cellbender.remove_background.data.io import write_matrix_to_cellranger_h5
+
+import shutil
+
+
+USE_CUDA = torch.cuda.is_available()
+
+
+def sparse_matrix_equal(mat1: sp.csc_matrix,
+ mat2: sp.csc_matrix):
+ """Fast assertion that sparse matrices are equal"""
+ return (mat1 != mat2).sum() == 0
+
+
+def tensors_equal(a: torch.Tensor,
+ b: torch.Tensor):
+ """Assertion that torch tensors are equal for each element"""
+ return (a == b).all()
+
+
+def string_ndarray_equality(a: np.ndarray, b: np.ndarray) -> bool:
+ return (a.astype(str) == b.astype(str)).all()
+
+
+class SavedFileH5:
+ def __init__(self, name, keys, version, shape, analyzed_shape=None,
+ local_keys=None, global_keys=None, meta_keys=None, barcodes_analyzed=None):
+ self.name = name
+ self.version = version
+ self.keys = keys
+ self.shape = shape
+ self.analyzed_shape = analyzed_shape
+ self.local_keys = local_keys
+ self.global_keys = global_keys
+ self.meta_keys = meta_keys
+ self.barcodes_analyzed = barcodes_analyzed
+
+ def __repr__(self):
+ return f'Saved h5 file ({self.shape}) at {self.name} ' \
+ f'(CellRanger v{self.version} with keys {self.keys})'
+
+
+@pytest.fixture(scope='session')
+def simulated_dataset():
+ """Generate a small simulated dataset Dict once and make it visible to all tests"""
+ d = generate_sample_dirichlet_dataset(
+ n_droplets=100,
+ n_genes=1000,
+ chi_artificial_similarity=0.25,
+ cells_of_each_type=[10, 10, 10],
+ cell_mean_umi=[5000, 5000, 5000],
+ cell_lognormal_sigma=0.01,
+ empty_mean_umi=200,
+ empty_lognormal_sigma=0.01,
+ model_type='full',
+ dirichlet_alpha=0.05,
+ epsilon_param=20,
+ rho_alpha=4,
+ rho_beta=96,
+ random_seed=0,
+ )
+ d['genomes'] = np.array(['simulated'] * d['gene_names'].size)
+ d['gene_ids'] = np.array(['ENSEMBLSIM000' + s for s in d['gene_names']])
+ d['feature_types'] = np.array(['Gene Expression'] * (d['gene_names'].size - 50)
+ + ['Antibody Capture'] * 40
+ + ['CRISPR Guide Capture'] * 5
+ + ['Custom'] * 5) # a mix of types
+ d['matrix'] = (d['counts_true'] + d['counts_bkg']).tocsc()
+ return d
+
+
+@pytest.fixture(scope='session')
+def h5_v3_file(tmpdir_factory, simulated_dataset) -> 'SavedFileH5':
+ """Save an h5 file once and make it visible to all tests."""
+ tmp_dir = tmpdir_factory.mktemp('data')
+ filename = tmp_dir.join('tmp_v3.h5')
+ cellranger_version = 3
+ d = simulated_dataset
+ write_matrix_to_cellranger_h5(
+ output_file=filename,
+ gene_names=d['gene_names'],
+ gene_ids=d['gene_ids'],
+ feature_types=d['feature_types'],
+ genomes=d['genomes'],
+ barcodes=d['barcodes'],
+ count_matrix=d['matrix'],
+ cellranger_version=cellranger_version,
+ )
+ yield SavedFileH5(name=filename,
+ keys=['gene_names', 'barcodes', 'genomes', 'gene_ids', 'feature_types'],
+ version=cellranger_version,
+ shape=d['matrix'].shape)
+ shutil.rmtree(str(tmp_dir))
+
+
+@pytest.fixture(scope='session')
+def h5_v2_file(tmpdir_factory, simulated_dataset) -> 'SavedFileH5':
+ """Save an h5 file once and make it visible to all tests."""
+ tmp_dir = tmpdir_factory.mktemp('data')
+ filename = tmp_dir.join('tmp_v2.h5')
+ cellranger_version = 2
+ d = simulated_dataset
+ write_matrix_to_cellranger_h5(
+ output_file=filename,
+ gene_names=d['gene_names'],
+ gene_ids=d['gene_ids'],
+ # feature_types=d['feature_types'],
+ # genomes=d['genomes'],
+ barcodes=d['barcodes'],
+ count_matrix=d['matrix'],
+ cellranger_version=cellranger_version,
+ )
+ yield SavedFileH5(name=filename,
+ keys=['gene_names', 'barcodes', 'gene_ids'],
+ version=cellranger_version,
+ shape=d['matrix'].shape)
+ shutil.rmtree(str(tmp_dir))
+
+
+@pytest.fixture(scope='session')
+def h5_v2_file_missing_ids(tmpdir_factory, simulated_dataset) -> 'SavedFileH5':
+ """Save an h5 file once and make it visible to all tests."""
+ tmp_dir = tmpdir_factory.mktemp('data')
+ filename = tmp_dir.join('tmp_v2.h5')
+ cellranger_version = 2
+ d = simulated_dataset
+ write_matrix_to_cellranger_h5(
+ output_file=filename,
+ gene_names=d['gene_names'],
+ # gene_ids=d['gene_ids'],
+ # feature_types=d['feature_types'],
+ # genomes=d['genomes'],
+ barcodes=d['barcodes'],
+ count_matrix=d['matrix'],
+ cellranger_version=cellranger_version,
+ )
+ yield SavedFileH5(name=filename,
+ keys=['gene_names', 'barcodes'],
+ version=cellranger_version,
+ shape=d['matrix'].shape)
+ shutil.rmtree(str(tmp_dir))
+
+
+@pytest.fixture(scope='session', params=[2, 3])
+def h5_file(request, h5_v2_file, h5_v3_file):
+ if request.param == 2:
+ return h5_v2_file
+ elif request.param == 3:
+ return h5_v3_file
+ else:
+ raise ValueError(f'Test error: requested v{request.param} h5 file')
+
+
+@pytest.fixture(scope='session')
+def h5_v3_file_post_inference(tmpdir_factory, simulated_dataset) -> 'SavedFileH5':
+ """Save an h5 file once and make it visible to all tests."""
+ tmp_dir = tmpdir_factory.mktemp('data')
+ filename = tmp_dir.join('tmp_v3_inferred.h5')
+ cellranger_version = 3
+ d = simulated_dataset
+
+ barcodes_analyzed = 50
+
+ write_matrix_to_cellranger_h5(
+ output_file=filename,
+ gene_names=d['gene_names'],
+ gene_ids=d['gene_ids'],
+ feature_types=d['feature_types'],
+ genomes=d['genomes'],
+ barcodes=d['barcodes'],
+ count_matrix=d['matrix'],
+ cellranger_version=cellranger_version,
+ local_latents={'barcode_indices_for_latents': np.arange(0, barcodes_analyzed),
+ 'gene_expression_encoding': np.random.rand(barcodes_analyzed, 10),
+ 'cell_probability': np.random.rand(barcodes_analyzed),
+ 'd': np.random.rand(barcodes_analyzed) + 5.},
+ global_latents={'global_var1': np.array([1, 2, 3])},
+ metadata={'metadata1': np.array(0.9),
+ 'metadata2': {'key1': 'val1', 'key2': {'a': 'val2', 'b': 'val3'}},
+ 'metadata3': None,
+ 'metadata4': 0.5,
+ 'metadata5': 'text'},
+ )
+ yield SavedFileH5(name=filename,
+ keys=['gene_names', 'barcodes', 'genomes', 'gene_ids', 'feature_types'],
+ version=cellranger_version,
+ shape=d['matrix'].shape,
+ analyzed_shape=(barcodes_analyzed, d['matrix'].shape[1]),
+ local_keys=['d', 'cell_probability', 'gene_expression_encoding'],
+ global_keys=['global_var1'],
+ meta_keys=['metadata1'],
+ barcodes_analyzed=barcodes_analyzed)
+ shutil.rmtree(str(tmp_dir))
diff --git a/cellbender/remove_background/tests/mckp_memory_profiling.py b/cellbender/remove_background/tests/mckp_memory_profiling.py
new file mode 100644
index 00000000..fe6e831d
--- /dev/null
+++ b/cellbender/remove_background/tests/mckp_memory_profiling.py
@@ -0,0 +1,146 @@
+"""Script to enable memory usage profiling via memory-profiler"""
+
+from cellbender.remove_background.estimation import MultipleChoiceKnapsack
+from cellbender.remove_background.sparse_utils import csr_set_rows_to_zero
+from cellbender.remove_background.checkpoint import load_from_checkpoint
+from cellbender.remove_background.posterior import Posterior, \
+ compute_mean_target_removal_as_function
+from cellbender.remove_background.data.dataset import SingleCellRNACountsDataset
+from cellbender.remove_background import consts
+
+import scipy.sparse as sp
+import numpy as np
+from memory_profiler import profile
+
+import argparse
+import sys
+
+
+def get_parser() -> argparse.ArgumentParser:
+ parser_ = argparse.ArgumentParser(
+ description="Run memory profiling on output count matrix generation. "
+ "NOTE that you have to decorate "
+ "MultipleChoiceKnapsack.estimate_noise() with memory_profiler's "
+ "@profile() decorator manually.",
+ )
+ parser_.add_argument('-f', '--checkpoint-file',
+ type=str,
+ required=True,
+ dest='input_checkpoint_tarball',
+ help='Saved CellBender checkpoint file ckpt.tar.gz')
+ parser_.add_argument('-i', '--input',
+ type=str,
+ required=True,
+ dest='input_file',
+ help='Input data file')
+ return parser_
+
+
+def compute_noise_counts(posterior,
+ fpr: float,
+ estimator_constructor: 'EstimationMethod',
+ **kwargs) -> sp.csr_matrix:
+ """Probably the most important method: computation of the clean output count matrix.
+
+ Args:
+ estimator_constructor: A noise count estimator class derived from
+ the EstimationMethod base class, and implementing the
+ .estimate_noise() method, which creates a point estimate of
+ noise. Pass in the constructor, not an object.
+ **kwargs: Keyword arguments for estimator_constructor().estimate_noise()
+
+ Returns:
+ denoised_counts: Denoised output CSC sparse matrix (CSC for saving)
+
+ """
+
+ # Only compute using defaults if the cache is empty.
+ if posterior._noise_count_regularized_posterior_coo is not None:
+ # Priority is taken by a regularized posterior, since presumably
+ # the user computed it for a reason.
+ posterior_coo = posterior._noise_count_regularized_posterior_coo
+ else:
+ # Use exact posterior if a regularized version is not computed.
+ posterior_coo = (posterior._noise_count_posterior_coo
+ if (posterior._noise_count_posterior_coo is not None)
+ else posterior.cell_noise_count_posterior_coo())
+
+ # Instantiate Estimator object.
+ estimator = estimator_constructor(index_converter=posterior.index_converter)
+
+ # Compute point estimate of noise in cells.
+ noise_targets = get_noise_targets(posterior=posterior, fpr=fpr)
+ noise_csr = estimator.estimate_noise(
+ estimator=estimator,
+ noise_log_prob_coo=posterior_coo,
+ noise_offsets=posterior._noise_count_posterior_coo_offsets,
+ noise_targets_per_gene=noise_targets,
+ **kwargs,
+ )
+
+ return noise_csr
+
+
+def get_noise_targets(posterior, fpr=0.01):
+ count_matrix = posterior.dataset_obj.data['matrix'] # all barcodes
+ cell_inds = posterior.dataset_obj.analyzed_barcode_inds[posterior.latents_map['p']
+ > consts.CELL_PROB_CUTOFF]
+ non_cell_row_logic = np.array([i not in cell_inds
+ for i in range(count_matrix.shape[0])])
+ cell_counts = csr_set_rows_to_zero(csr=count_matrix, row_logic=non_cell_row_logic)
+
+ noise_target_fun_per_cell = compute_mean_target_removal_as_function(
+ noise_count_posterior_coo=posterior._noise_count_posterior_coo,
+ noise_offsets=posterior._noise_count_posterior_coo_offsets,
+ index_converter=posterior.index_converter,
+ raw_count_csr_for_cells=cell_counts,
+ n_cells=len(cell_inds),
+ device='cpu',
+ per_gene=True,
+ )
+ noise_target_fun = lambda x: noise_target_fun_per_cell(x) * len(cell_inds)
+ noise_targets = noise_target_fun(fpr).detach().cpu().numpy()
+ return noise_targets
+
+
+if __name__ == "__main__":
+
+ # handle input arguments
+ parser = get_parser()
+ args = parser.parse_args(sys.argv[1:])
+
+ # load checkpoint
+ ckpt = load_from_checkpoint(tarball_name=args.input_checkpoint_tarball,
+ filebase=None,
+ to_load=['model', 'posterior', 'args'],
+ force_device='cpu')
+
+ # load dataset
+ dataset_obj = \
+ SingleCellRNACountsDataset(input_file=args.input_file,
+ expected_cell_count=ckpt['args'].expected_cell_count,
+ total_droplet_barcodes=ckpt['args'].total_droplets,
+ fraction_empties=ckpt['args'].fraction_empties,
+ model_name=ckpt['args'].model,
+ gene_blacklist=ckpt['args'].blacklisted_genes,
+ exclude_features=ckpt['args'].exclude_features,
+ low_count_threshold=ckpt['args'].low_count_threshold,
+ ambient_counts_in_cells_low_limit=ckpt['args'].ambient_counts_in_cells_low_limit,
+ fpr=ckpt['args'].fpr)
+
+ # load posterior
+ posterior = Posterior(
+ dataset_obj=dataset_obj,
+ vi_model=ckpt['model'],
+ posterior_batch_size=ckpt['args'].posterior_batch_size,
+ debug=False,
+ )
+ posterior.load(file=ckpt['posterior_file'])
+
+ # run output count matrix generation
+ compute_noise_counts(posterior=posterior,
+ fpr=0.01,
+ estimator_constructor=MultipleChoiceKnapsack,
+ approx_gb=0.1)
+
+ sys.exit(0)
diff --git a/cellbender/remove_background/tests/miniwdl_check_wdl.sh b/cellbender/remove_background/tests/miniwdl_check_wdl.sh
new file mode 100755
index 00000000..2da9ad83
--- /dev/null
+++ b/cellbender/remove_background/tests/miniwdl_check_wdl.sh
@@ -0,0 +1,13 @@
+#!/bin/bash
+
+set -euxo pipefail
+
+# runs from the root directory of the repo
+
+# find all WDL files
+# WDL_FILES=$(find . -type f -name "*.wdl")
+WDL_FILES=("./wdl/cellbender_remove_background.wdl")
+
+for WDL in ${WDL_FILES}; do
+ miniwdl check ${WDL}
+done
\ No newline at end of file
diff --git a/cellbender/remove_background/tests/test.py b/cellbender/remove_background/tests/test.py
deleted file mode 100644
index 5f72ec8e..00000000
--- a/cellbender/remove_background/tests/test.py
+++ /dev/null
@@ -1,193 +0,0 @@
-import unittest
-import os
-import warnings
-
-import cellbender
-import cellbender.remove_background.model
-from cellbender.remove_background.train import run_inference
-from cellbender.remove_background.data.extras.simulate import \
- simulate_dataset_with_ambient_rna
-from cellbender.remove_background.data.dataset import \
- SingleCellRNACountsDataset, write_matrix_to_cellranger_h5, \
- get_matrix_from_cellranger_h5, get_matrix_from_anndata
-import numpy as np
-import anndata
-import sys
-
-
-class TestConsole(unittest.TestCase):
-
- def test_data_simulation_and_write_and_read(self):
- """Run basic tests of data simulation and read and write functionality.
-
- Test the ability to read/write data to/from an HDF5 file using pytables.
- This necessitates creating a small simulated dataset, writing it to
- a temporary file, reading the file back in, and deleting the
- temporary file.
-
- """
-
- try:
-
- # This is here to suppress the numpy warning triggered by
- # scipy.sparse.
- warnings.simplefilter("ignore")
-
- # Generate a simulated dataset with ambient RNA.
- n_cells = 100
- csr_barcode_gene_synthetic, _, chi, _ = \
- simulate_dataset_with_ambient_rna(n_cells=n_cells,
- n_empty=3 * n_cells,
- clusters=1, n_genes=1000,
- d_cell=2000, d_empty=100,
- ambient_different=False)
-
- # Generate some names for genes and barcodes.
- gene_names = np.array([f'g_{i}' for i in
- range(csr_barcode_gene_synthetic.shape[1])])
- barcode_names = \
- np.array([f'bc_{i}' for i in
- range(csr_barcode_gene_synthetic.shape[0])])
-
- # Save the data to a temporary cellranger h5 file.
- temp_file_name = 'testfile.h5'
- write_matrix_to_cellranger_h5(cellranger_version=3,
- output_file=temp_file_name,
- loss=None,
- gene_names=gene_names,
- barcodes=barcode_names,
- inferred_count_matrix=
- csr_barcode_gene_synthetic.tocsc(),
- ambient_expression=chi[0, :])
-
- # Save the data to a temporary AnnData file.
- temp_h5ad_name = 'testfile.h5ad'
- temp_adata = anndata.AnnData(
- X=csr_barcode_gene_synthetic,
- )
- temp_adata.var_names = gene_names
- temp_adata.obs_names = barcode_names
- temp_adata.write_h5ad(temp_h5ad_name)
-
- # Read the data back in.
- reconstructed = get_matrix_from_cellranger_h5(temp_file_name)
- new_matrix = reconstructed['matrix']
-
- # Check that the data matches.
- assert (csr_barcode_gene_synthetic.sum(axis=1) ==
- new_matrix.sum(axis=1)).all(), \
- "Saved and re-opened data is not accurate."
- assert (csr_barcode_gene_synthetic.sum(axis=0) ==
- new_matrix.sum(axis=0)).all(), \
- "Saved and re-opened data is not accurate."
-
- # Remove the temporary file.
- os.remove(temp_file_name)
-
- # Check that anndata matches.
- reconstructed_h5ad = get_matrix_from_anndata(temp_h5ad_name)
- matrix_h5ad = reconstructed_h5ad['matrix']
- assert (csr_barcode_gene_synthetic.sum(axis=1) ==
- matrix_h5ad.sum(axis=1)).all(), \
- "Saved and re-opened AnnData is not accurate."
- assert (csr_barcode_gene_synthetic.sum(axis=0) ==
- matrix_h5ad.sum(axis=0)).all(), \
- "Saved and re-opened AnnData is not accurate."
-
- os.remove(temp_h5ad_name)
-
- return 1
-
- except TestConsole.failureException:
-
- return 0
-
- def test_inference(self):
- """Run a basic tests doing inference on a synthetic dataset.
-
- Runs the inference procedure on CPU.
-
- """
-
- try:
-
- n_cells = 100
-
- # Generate a simulated dataset with ambient RNA.
- csr_barcode_gene_synthetic, _, _, _ = \
- simulate_dataset_with_ambient_rna(n_cells=n_cells,
- n_empty=3 * n_cells,
- clusters=1, n_genes=1000,
- d_cell=2000, d_empty=100,
- ambient_different=False)
-
- # Fake some parsed command line inputs.
- args = ObjectWithAttributes()
- args.use_cuda = False
- args.z_hidden_dims = [100]
- args.d_hidden_dims = [10, 2]
- args.p_hidden_dims = [100, 10]
- args.z_dim = 10
- args.learning_rate = 0.001
- args.epochs = 10
- args.model = "full"
- args.fraction_empties = 0.5
- args.use_jit = True
- args.training_fraction = 0.9
-
- args.expected_cell_count = n_cells
-
- # Wrap simulated count matrix in a Dataset object.
- dataset_obj = SingleCellRNACountsDataset()
- dataset_obj.data = \
- {'matrix': csr_barcode_gene_synthetic,
- 'gene_names':
- np.array([f'g{n}' for n in
- range(csr_barcode_gene_synthetic.shape[1])]),
- 'barcodes':
- np.array([f'bc{n}' for n in
- range(csr_barcode_gene_synthetic.shape[0])])}
- dataset_obj.priors['n_cells'] = n_cells
- dataset_obj._trim_dataset_for_analysis()
- dataset_obj._estimate_priors()
-
- # Run inference on this simulated dataset.
- inferred_model = run_inference(dataset_obj, args)
-
- # Get encodings from the trained model.
- z, d, p = cellbender.remove_background.model.\
- get_encodings(inferred_model, dataset_obj)
-
- # Make the background-subtracted dataset.
- inferred_count_matrix = cellbender.remove_background.model.\
- generate_maximum_a_posteriori_count_matrix(z, d, p,
- inferred_model,
- dataset_obj)
-
- # Get the inferred background RNA expression from the model.
- ambient_expression = cellbender.remove_background.model.\
- get_ambient_expression_from_pyro_param_store()
-
- return 1
-
- except TestConsole.failureException:
-
- return 0
-
-
-class ObjectWithAttributes(object):
- """Exists only to populate the args data structure with attributes."""
- pass
-
-
-def main():
- """Run unit tests."""
-
- tester = TestConsole()
-
- passed_tests = 0
-
- passed_tests += tester.test_data_simulation_and_write_and_read()
- passed_tests += tester.test_inference()
-
- sys.stdout.write(f'Passed {passed_tests} of 2 tests.\n\n')
diff --git a/cellbender/remove_background/tests/test_checkpoint.py b/cellbender/remove_background/tests/test_checkpoint.py
new file mode 100644
index 00000000..59ccb9ec
--- /dev/null
+++ b/cellbender/remove_background/tests/test_checkpoint.py
@@ -0,0 +1,626 @@
+"""Tests for checkpointing functions."""
+
+import pytest
+import random
+import numpy as np
+import torch
+from torch.distributions import constraints
+import pyro
+import pyro.optim as optim
+
+import cellbender
+from cellbender.remove_background.checkpoint import make_tarball, unpack_tarball, \
+ create_workflow_hashcode, save_checkpoint, load_checkpoint, \
+ save_random_state, load_random_state
+from cellbender.remove_background.vae.encoder import EncodeZ, EncodeNonZLatents, \
+ CompositeEncoder
+from cellbender.remove_background.vae.decoder import Decoder
+from cellbender.remove_background.data.extras.simulate import \
+ generate_sample_dirichlet_dataset, get_dataset_dict_as_anndata
+from cellbender.remove_background.data.dataprep import prep_sparse_data_for_training
+from cellbender.remove_background.model import RemoveBackgroundPyroModel
+from cellbender.remove_background.data.dataset import SingleCellRNACountsDataset
+from cellbender.remove_background.run import run_inference
+from .conftest import USE_CUDA
+
+import os
+import argparse
+import shutil
+import subprocess
+from typing import List
+
+
+class RandomState:
+ def __init__(self, use_cuda=False):
+ self.python = random.randint(0, 100000)
+ self.numpy = np.random.randint(0, 100000, size=1).item()
+ self.torch = torch.randint(low=0, high=100000, size=[1], device='cpu').item()
+ self.use_cuda = use_cuda
+ if self.use_cuda:
+ self.cuda = torch.randint(low=0, high=100000, size=[1], device='cuda').item()
+
+ def __repr__(self):
+ if self.use_cuda:
+ return f'python {self.python}; numpy {self.numpy}; torch {self.torch}; torch_cuda {self.cuda}'
+ else:
+ return f'python {self.python}; numpy {self.numpy}; torch {self.torch}'
+
+
+def test_create_workflow_hashcode():
+ """Ensure workflow hashcodes are behaving as expected"""
+
+ tmp_args1 = argparse.Namespace(epochs=100, expected_cells=1000, use_cuda=True)
+ tmp_args2 = argparse.Namespace(epochs=200, expected_cells=1000, use_cuda=True)
+ tmp_args3 = argparse.Namespace(epochs=100, expected_cells=500, use_cuda=True)
+ hashcode1 = create_workflow_hashcode(module_path=os.path.dirname(cellbender.__file__),
+ args=tmp_args1)
+ hashcode2 = create_workflow_hashcode(module_path=os.path.dirname(cellbender.__file__),
+ args=tmp_args2)
+ hashcode3 = create_workflow_hashcode(module_path=os.path.dirname(cellbender.__file__),
+ args=tmp_args3)
+
+ # make sure it changes for different arguments
+ assert hashcode1 != hashcode3
+
+ # but that the "epochs" argument has no effect
+ assert hashcode1 == hashcode2
+
+
+def create_random_state_blank_slate(seed, use_cuda=USE_CUDA):
+ """Establish a base random state
+ https://pytorch.org/docs/stable/notes/randomness.html
+ """
+ random.seed(seed)
+ np.random.seed(seed)
+ torch.manual_seed(seed)
+ pyro.util.set_rng_seed(seed)
+ if use_cuda:
+ torch.cuda.manual_seed_all(seed)
+
+
+def perturb_random_state(n, use_cuda=USE_CUDA):
+ """Perturb the base random state by drawing random numbers"""
+ for _ in range(n):
+ random.randint(0, 10)
+ np.random.randint(0, 10, 1)
+ torch.randn((1,), device='cpu')
+ if use_cuda:
+ torch.randn((1,), device='cuda')
+
+
+@pytest.fixture(scope='function', params=[0, 1, 1234], ids=lambda x: f'seed{x}')
+def random_state_blank_slate(request):
+ create_random_state_blank_slate(request.param)
+ return request.param
+
+
+@pytest.fixture(scope='function', params=[0, 1, 10])
+def perturbed_random_state_dict(request, random_state_blank_slate):
+ perturb_random_state(request.param)
+ return {'state': RandomState(),
+ 'seed': random_state_blank_slate,
+ 'n': request.param}
+
+
+@pytest.fixture(scope='function', params=[0], ids=lambda x: f'setseed{x}')
+def random_state_blank_slate0(request):
+ create_random_state_blank_slate(request.param)
+ return request.param
+
+
+@pytest.fixture(scope='function', params=[10], ids=lambda x: f'set{x}')
+def perturbed_random_state0_dict(request, random_state_blank_slate0):
+ perturb_random_state(request.param)
+ return {'state': RandomState(),
+ 'seed': random_state_blank_slate0,
+ 'n': request.param}
+
+
+def test_perturbedrandomstate_fixture_meets_expectations(perturbed_random_state0_dict,
+ perturbed_random_state_dict):
+ """Test the setup of these randomstate fixtures.
+
+ The state0 fixture is one set of params.
+ The state fixture is a combinatorial set of params, only one of which matches
+ the state0 setup.
+
+ We want to make sure that when we have fixtures set up the same way, then
+ randomness behaves the same (and different when set up differently).
+ """
+ prs = perturbed_random_state_dict['state']
+ params = (perturbed_random_state_dict['seed'], perturbed_random_state_dict['n'])
+ prs0 = perturbed_random_state0_dict['state']
+ params0 = (perturbed_random_state0_dict['seed'], perturbed_random_state0_dict['n'])
+
+ if params == params0:
+ # this is the only case in which we expect the two random states to be equal
+ assert str(prs0) == str(prs)
+ else:
+ assert str(prs0) != str(prs)
+
+
+def test_that_randomstate_plus_perturb_gives_perturbedrandomstate(
+ perturbed_random_state0_dict):
+ """Make sure perturbation is working as intended"""
+
+ # recreate and make sure we end up in the same place
+ # the "recipe" is in perturbed_random_state0_tuple[1]
+ create_random_state_blank_slate(perturbed_random_state0_dict['seed'])
+ perturb_random_state(perturbed_random_state0_dict['n'])
+ this_prs0 = RandomState()
+
+ # check equality
+ prs0 = perturbed_random_state0_dict['state']
+ assert str(prs0) == str(this_prs0)
+
+
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+def test_save_and_load_random_state(tmpdir_factory, perturbed_random_state_dict, cuda):
+ """Test whether random states are being preserved correctly.
+ perturbed_random_state_dict is important since it initializes the state."""
+
+ # save the random states
+ filebase = tmpdir_factory.mktemp('random_states').join('tmp_checkpoint')
+ save_random_state(filebase=filebase)
+
+ # see "what would have happened had we continued"
+ counterfactual = RandomState(use_cuda=cuda)
+ incorrect = RandomState(use_cuda=cuda) # a second draw
+
+ # load the random states and check random number generators
+ load_random_state(filebase=filebase)
+ actual = RandomState(use_cuda=cuda)
+
+ # check equality
+ assert str(counterfactual) == str(actual)
+ assert str(incorrect) != str(actual)
+
+
+def new_train_loader(data: torch.Tensor, batch_size: int, shuffle: bool = True):
+ return torch.utils.data.DataLoader(data, batch_size=batch_size, shuffle=shuffle)
+
+
+class PyroModel(torch.nn.Module):
+
+ def __init__(self, dim=8, hidden_layer=4, z_dim=2):
+ super(PyroModel, self).__init__()
+ create_random_state_blank_slate(0)
+ pyro.clear_param_store()
+ self.encoder = EncodeZ(input_dim=dim, hidden_dims=[hidden_layer], output_dim=z_dim)
+ self.decoder = Decoder(input_dim=z_dim, hidden_dims=[hidden_layer], output_dim=dim)
+ self.z_dim = z_dim
+ self.use_cuda = torch.cuda.is_available()
+ self.normal = pyro.distributions.Normal
+ self.loss = []
+ # self.to(device='cuda' if self.use_cuda else 'cpu') # CUDA not tested
+
+ def model(self, x: torch.FloatTensor):
+ pyro.module("decoder", self.decoder, update_module_params=True)
+ with pyro.plate('plate', size=x.shape[0]):
+ z = pyro.sample('z', self.normal(loc=0., scale=1.).expand_by([x.shape[0], self.z_dim]).to_event(1))
+ x_rec = self.decoder(z)
+ pyro.sample('obs', self.normal(loc=x_rec, scale=0.1).to_event(1), obs=x)
+
+ def guide(self, x: torch.FloatTensor):
+ pyro.module("encoder", self.encoder, update_module_params=True)
+ with pyro.plate('plate', size=x.shape[0]):
+ enc = self.encoder(x)
+ pyro.sample('z', self.normal(loc=enc['loc'], scale=enc['scale']).to_event(1))
+
+
+def train_pyro(n_epochs: int,
+ data_loader: torch.utils.data.DataLoader,
+ svi: pyro.infer.SVI):
+ """Run training"""
+ loss_per_epoch = []
+ for _ in range(n_epochs):
+ loss = 0
+ norm = 0
+ for data in data_loader:
+ loss += svi.step(data)
+ norm += data.size(0)
+ loss_per_epoch.append(loss / norm)
+ return loss_per_epoch
+
+
+def _check_all_close(tensors1, tensors2) -> bool:
+ """For two lists of tensors, check that they are all close"""
+ assert len(tensors1) == len(tensors2), \
+ 'Must pass in same number of tensors to check if they are equal'
+ equal = True
+ for t1, t2 in zip(tensors1, tensors2):
+ equal = equal and torch.allclose(t1, t2)
+ return equal
+
+
+def _get_params(module: torch.nn.Module) -> List[torch.Tensor]:
+ return [p.data.clone() for p in module.parameters()]
+
+
+@pytest.mark.parametrize('batch_size_n', [32, 128], ids=lambda n: f'batch{n}')
+def test_save_and_load_pyro_checkpoint(tmpdir_factory, batch_size_n):
+ """Check and see if restarting from a checkpoint picks up in the same place
+ we left off. Use a dataloader.
+ """
+
+ filedir = tmpdir_factory.mktemp('ckpt')
+ filebase = filedir.join('ckpt')
+ dim = 8
+ epochs = 3
+ epochs2 = 3
+ lr = 1e-2
+
+ # data and dataloader
+ dataset = torch.randn((128, dim))
+ train_loader = new_train_loader(data=dataset, batch_size=batch_size_n)
+
+ # create an ML model
+ initial_model = PyroModel(dim=dim)
+
+ # set up the inference process
+ scheduler = optim.ClippedAdam({'lr': lr, 'clip_norm': 10.})
+ svi = pyro.infer.SVI(initial_model.model, initial_model.guide, scheduler, loss=pyro.infer.Trace_ELBO())
+ w1 = _get_params(initial_model.encoder)
+
+ print('initial weight matrix =========================')
+ print('\n'.join([str(t) for t in w1]))
+
+ # train in two parts: part 1
+ initial_model.loss.extend(train_pyro(n_epochs=epochs, data_loader=train_loader, svi=svi))
+
+ print('no_ckpt trained round 1 (saved) ===============')
+ # print(initial_model.encoder.linears[0].weight.data)
+ print('\n'.join([str(t) for t in _get_params(initial_model.encoder)]))
+
+ # save
+ save_successful = save_checkpoint(filebase=str(filebase),
+ args=argparse.Namespace(),
+ model_obj=initial_model, # TODO
+ scheduler=scheduler,
+ train_loader=train_loader,
+ tarball_name=str(filebase) + '.tar.gz')
+ assert save_successful, 'Failed to save checkpoint during test_save_and_load_checkpoint'
+
+ # load from checkpoint
+ create_random_state_blank_slate(0)
+ pyro.clear_param_store()
+ ckpt = load_checkpoint(filebase=str(filebase),
+ tarball_name=str(filebase) + '.tar.gz',
+ force_device='cpu')
+ model_ckpt = ckpt['model']
+ scheduler_ckpt = ckpt['optim']
+ train_loader = ckpt['train_loader']
+ s = ckpt['loaded']
+ print('model_ckpt loaded =============================')
+ print('\n'.join([str(t) for t in _get_params(model_ckpt.encoder)]))
+
+ matches = _check_all_close(_get_params(model_ckpt.encoder),
+ _get_params(initial_model.encoder))
+ print(f'model_ckpt loaded matches data from (saved) trained round 1: {matches}')
+
+ # clean up before most assertions... hokey now due to lack of fixture usage here
+ shutil.rmtree(str(filedir))
+
+ # and continue training
+ assert s is True, 'Checkpoint loading failed during test_save_and_load_checkpoint'
+ svi_ckpt = pyro.infer.SVI(model_ckpt.model, model_ckpt.guide, scheduler_ckpt, loss=pyro.infer.Trace_ELBO())
+ rng_ckpt = RandomState()
+ guide_trace_ckpt = pyro.poutine.trace(svi_ckpt.guide).get_trace(x=dataset)
+ model_ckpt.loss.extend(train_pyro(n_epochs=epochs2, data_loader=train_loader, svi=svi_ckpt))
+
+ print('model_ckpt after round 2 ======================')
+ print('\n'.join([str(t) for t in _get_params(model_ckpt.encoder)]))
+
+ # one-shot training straight through
+ model_one_shot = PyroModel(dim=dim) # resets random state
+ scheduler = optim.ClippedAdam({'lr': lr, 'clip_norm': 10.})
+ train_loader = new_train_loader(data=dataset, batch_size=batch_size_n)
+ svi_one_shot = pyro.infer.SVI(model_one_shot.model, model_one_shot.guide, scheduler, loss=pyro.infer.Trace_ELBO())
+ model_one_shot.loss.extend(train_pyro(n_epochs=epochs, data_loader=train_loader, svi=svi_one_shot))
+ rng_one_shot = RandomState()
+ guide_trace_one_shot = pyro.poutine.trace(svi_one_shot.guide).get_trace(x=dataset)
+ model_one_shot.loss.extend(train_pyro(n_epochs=epochs2, data_loader=train_loader, svi=svi_one_shot))
+
+ assert str(rng_one_shot) == str(rng_ckpt), \
+ 'Random states of the checkpointed and non-checkpointed versions at ' \
+ 'the start of training round 2 do not match.'
+
+ print('model_one_shot ================================')
+ print('\n'.join([str(t) for t in _get_params(model_one_shot.encoder)]))
+
+ print(model_one_shot.loss)
+ print(model_ckpt.loss)
+ print([l1 == l2 for l1, l2 in zip(model_one_shot.loss, model_ckpt.loss)])
+
+ # training should be doing something to the initial weight matrix
+ assert (w1[0] != _get_params(model_one_shot.encoder)[0]).sum().item() > 0, \
+ 'Training is not changing the weight matrix in test_save_and_load_checkpoint'
+ assert (w1[0] != _get_params(model_ckpt.encoder)[0]).sum().item() > 0, \
+ 'Training is not changing the checkpointed weight matrix in test_save_and_load_checkpoint'
+
+ # see if we end up where we should
+ assert _check_all_close(_get_params(model_one_shot.encoder),
+ _get_params(model_ckpt.encoder)), \
+ 'Checkpointed restart does not agree with one-shot training'
+
+ # check guide traces
+ print('checking guide trace nodes for agreement:')
+ for name in guide_trace_one_shot.nodes:
+ if (name != '_RETURN') and ('value' in guide_trace_one_shot.nodes[name].keys()):
+ print(name)
+ disagreement = (guide_trace_one_shot.nodes[name]['value'].data
+ != guide_trace_ckpt.nodes[name]['value'].data).sum()
+ print(f'number of values that disagree: {disagreement}')
+ assert disagreement == 0, \
+ 'Guide traces disagree with and without checkpoint restart'
+
+
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+@pytest.mark.parametrize('scheduler',
+ [False, True],
+ ids=lambda b: 'OneCycleLR' if b else 'Adam')
+def test_save_and_load_cellbender_checkpoint(tmpdir_factory, cuda, scheduler):
+ """Check and see if restarting from a checkpoint picks up in the same place
+ we left off. Use our model and dataloader.
+ """
+
+ filedir = tmpdir_factory.mktemp('ckpt')
+
+ epochs = 5
+ epochs2 = 5
+
+ # data
+ n_genes = 2000
+ dataset = generate_sample_dirichlet_dataset(n_genes=n_genes, cells_of_each_type=[100],
+ n_droplets=2000, model_type='ambient',
+ cell_mean_umi=[5000])
+ adata = get_dataset_dict_as_anndata(dataset)
+ adata_file = os.path.join(filedir, 'sim.h5ad')
+ adata.write(adata_file)
+ dataset_obj = \
+ SingleCellRNACountsDataset(input_file=adata_file,
+ expected_cell_count=100,
+ total_droplet_barcodes=1000,
+ fraction_empties=0.1,
+ model_name='ambient',
+ gene_blacklist=[],
+ exclude_features=[],
+ low_count_threshold=15,
+ fpr=[0.01])
+
+ # set up the inference process
+ args = argparse.Namespace()
+ args.output_file = os.path.join(filedir, 'out.h5')
+ args.z_dim = 10
+ args.z_hidden_dims = [50]
+ args.model = 'ambient'
+ args.use_cuda = cuda
+ args.use_jit = False
+ args.learning_rate = 1e-3
+ args.training_fraction = 0.9
+ args.fraction_empties = 0.1
+ args.checkpoint_filename = 'test01234_'
+ args.checkpoint_min = 5
+ args.epoch_elbo_fail_fraction = None
+ args.final_elbo_fail_fraction = None
+ args.constant_learning_rate = not scheduler
+ args.debug = False
+ args.input_checkpoint_tarball = 'none'
+
+ create_random_state_blank_slate(0)
+ pyro.clear_param_store()
+ args.epochs = -1 # I am hacking my way around an error induced by saving a checkpoint for 0 epoch runs
+ initial_model, scheduler, _, _ = run_inference(dataset_obj=dataset_obj,
+ args=args)
+ w1 = _get_params(initial_model.encoder['z'])
+ print('encoder structure ============')
+ print(initial_model.encoder['z'])
+
+ print('initial weight matrix =========================')
+ print('\n'.join([str(t) for t in w1]))
+
+ # train in two parts: part 1
+ pyro.clear_param_store()
+ create_random_state_blank_slate(0)
+ args.epochs = epochs
+ initial_model, scheduler, train_loader, test_loader = \
+ run_inference(dataset_obj=dataset_obj, args=args,
+ output_checkpoint_tarball='none',
+ total_epochs_for_testing_only=epochs + epochs2)
+
+ print('no_ckpt trained round 1 (saved) ===============')
+ print('\n'.join([str(t) for t in _get_params(initial_model.encoder['z'])]))
+ # print(scheduler.get_state().keys())
+ # print(list(scheduler.get_state().values())[0].keys())
+ # print(list(list(scheduler.optim_objs.values())[0].optimizer.state_dict().values())[0].values())
+ # assert 0
+
+ # save
+ hashcode = create_workflow_hashcode(module_path=os.path.dirname(cellbender.__file__),
+ args=args)[:10]
+ checkpoint_filename = os.path.basename(args.output_file).split('.')[0] + '_' + hashcode
+ args.checkpoint_filename = checkpoint_filename
+ filebase = filedir.join(checkpoint_filename)
+ save_successful = save_checkpoint(filebase=str(filebase),
+ args=args,
+ model_obj=initial_model, # TODO
+ scheduler=scheduler,
+ tarball_name=str(filebase) + '.tar.gz',
+ train_loader=train_loader,
+ test_loader=test_loader)
+ assert save_successful, 'Failed to save checkpoint during test_save_and_load_checkpoint'
+ assert os.path.exists(str(filebase) + '.tar.gz'), 'Checkpoint should exist but does not'
+
+ # load from checkpoint (automatically) and run
+ pyro.clear_param_store()
+ create_random_state_blank_slate(0)
+ args.epochs = -1
+ args.input_checkpoint_tarball = str(filebase) + '.tar.gz'
+ model_ckpt, scheduler, _, _ = run_inference(dataset_obj=dataset_obj, args=args,
+ output_checkpoint_tarball='none')
+
+ print('model_ckpt loaded =============================')
+ print('\n'.join([str(t) for t in _get_params(model_ckpt.encoder['z'])]))
+
+ matches = _check_all_close(_get_params(model_ckpt.encoder['z']),
+ _get_params(initial_model.encoder['z']))
+ print(f'model_ckpt loaded matches data from (saved) trained round 1: {matches}')
+
+ # and continue training
+ create_random_state_blank_slate(0)
+ pyro.clear_param_store()
+ args.epochs = epochs + epochs2
+ model_ckpt, scheduler, _, _ = run_inference(dataset_obj=dataset_obj, args=args,
+ output_checkpoint_tarball='none')
+
+ print('model_ckpt after round 2 ======================')
+ print('\n'.join([str(t) for t in _get_params(model_ckpt.encoder['z'])]))
+
+ # clean up the temp directory to remove checkpoint before running the one-shot
+ shutil.rmtree(str(filedir))
+
+ # one-shot training straight through
+ pyro.clear_param_store()
+ create_random_state_blank_slate(0)
+ args.epochs = epochs + epochs2
+ args.input_checkpoint_tarball = 'none'
+ model_one_shot, scheduler, _, _ = run_inference(dataset_obj=dataset_obj, args=args,
+ output_checkpoint_tarball='none')
+
+ print('model_one_shot ================================')
+ print('\n'.join([str(t) for t in _get_params(model_one_shot.encoder['z'])]))
+
+ print('loss for model_one_shot:')
+ print(model_one_shot.loss)
+ print('loss for model_ckpt:')
+ print(model_ckpt.loss)
+ print([l1 == l2 for l1, l2 in zip(model_one_shot.loss, model_ckpt.loss)])
+
+ # training should be doing something to the initial weight matrix
+ assert (w1[0] != _get_params(model_one_shot.encoder['z'])[0]).sum().item() > 0, \
+ 'Training is not changing the weight matrix in test_save_and_load_checkpoint'
+ assert (w1[0] != _get_params(model_ckpt.encoder['z'])[0]).sum().item() > 0, \
+ 'Training is not changing the checkpointed weight matrix in test_save_and_load_checkpoint'
+
+ # see if we end up where we should
+ assert _check_all_close(_get_params(model_one_shot.encoder['z']),
+ _get_params(model_ckpt.encoder['z'])), \
+ 'Checkpointed restart does not agree with one-shot training'
+
+ # # check guide traces
+ # print('checking guide trace nodes for agreement:')
+ # for name in guide_trace_one_shot.nodes:
+ # if (name != '_RETURN') and ('value' in guide_trace_one_shot.nodes[name].keys()):
+ # print(name)
+ # disagreement = (guide_trace_one_shot.nodes[name]['value'].data
+ # != guide_trace_ckpt.nodes[name]['value'].data).sum()
+ # print(f'number of values that disagree: {disagreement}')
+ # assert disagreement == 0, \
+ # 'Guide traces disagree with and without checkpoint restart'
+
+
+@pytest.mark.parametrize(
+ "Optim, config",
+ [
+ (optim.ClippedAdam, {"lr": 0.01}),
+ (optim.ExponentialLR, {"optimizer": torch.optim.SGD,
+ "optim_args": {"lr": 0.01},
+ "gamma": 0.9}),
+ ],
+)
+def test_optimizer_checkpoint_restart(Optim, config, tmpdir_factory):
+ """This code has been copied from a version of a pyro test by Fritz Obermeyer"""
+
+ tempdir = tmpdir_factory.mktemp('ckpt')
+
+ def model():
+ x_scale = pyro.param("x_scale", torch.tensor(1.0),
+ constraint=constraints.positive)
+ z = pyro.sample("z", pyro.distributions.Normal(0, 1))
+ return pyro.sample("x", pyro.distributions.Normal(z, x_scale), obs=torch.tensor(0.1))
+
+ def guide():
+ z_loc = pyro.param("z_loc", torch.tensor(0.0))
+ z_scale = pyro.param(
+ "z_scale", torch.tensor(0.5), constraint=constraints.positive
+ )
+ pyro.sample("z", pyro.distributions.Normal(z_loc, z_scale))
+
+ store = pyro.get_param_store()
+
+ def get_snapshot(optimizer):
+ s = {k: v.data.clone() for k, v in store.items()}
+ if type(optimizer) == optim.lr_scheduler.PyroLRScheduler:
+ lr = list(optimizer.optim_objs.values())[0].get_last_lr()[0]
+ else:
+ lr = list(optimizer.optim_objs.values())[0].param_groups[0]['lr']
+ s.update({'lr': f'{lr:.4f}'})
+ return s
+
+ # Try without a checkpoint.
+ expected = []
+ store.clear()
+ pyro.set_rng_seed(20210811)
+ optimizer = Optim(config.copy())
+ svi = pyro.infer.SVI(model, guide, optimizer, pyro.infer.Trace_ELBO())
+ for _ in range(5 + 10):
+ svi.step()
+ expected.append(get_snapshot(optimizer))
+ if type(optimizer) == optim.lr_scheduler.PyroLRScheduler:
+ svi.optim.step()
+ del svi, optimizer
+
+ # Try with a checkpoint.
+ actual = []
+ store.clear()
+ pyro.set_rng_seed(20210811)
+ optimizer = Optim(config.copy())
+ svi = pyro.infer.SVI(model, guide, optimizer, pyro.infer.Trace_ELBO())
+ for _ in range(5):
+ svi.step()
+ actual.append(get_snapshot(optimizer))
+ if type(optimizer) == optim.lr_scheduler.PyroLRScheduler:
+ svi.optim.step()
+
+ # checkpoint
+ optim_filename = os.path.join(tempdir, "optimizer_state.pt")
+ param_filename = os.path.join(tempdir, "param_store.pt")
+ optimizer.save(optim_filename)
+ store.save(param_filename)
+ del optimizer, svi
+ store.clear()
+
+ # load from checkpoint
+ store.load(param_filename)
+ optimizer = Optim(config.copy())
+ optimizer.load(optim_filename)
+ svi = pyro.infer.SVI(model, guide, optimizer, pyro.infer.Trace_ELBO())
+ for _ in range(10):
+ svi.step()
+ actual.append(get_snapshot(optimizer))
+ if type(optimizer) == optim.lr_scheduler.PyroLRScheduler:
+ svi.optim.step()
+
+ # display learning rates and actual/expected values for z_loc
+ print(repr(optimizer))
+ print('epoch\t\tlr\t\tactual\t\t\t\texpected')
+ print('-' * 100)
+ for i, (ac, ex) in enumerate(zip(actual, expected)):
+ print('\t\t'.join([f'{x}' for x in [i, ac['lr'], ac['z_loc'], ex['z_loc']]]))
+ if i == 4:
+ print('-' * 100)
+
+ # ensure actual matches expected
+ for actual_t, expected_t in zip(actual, expected):
+ actual_t.pop('lr')
+ expected_t.pop('lr')
+ for actual_value, expected_value in zip(actual_t.values(), expected_t.values()):
+ assert torch.allclose(actual_value, expected_value)
diff --git a/cellbender/remove_background/tests/test_dataprep.py b/cellbender/remove_background/tests/test_dataprep.py
new file mode 100644
index 00000000..8fcb1562
--- /dev/null
+++ b/cellbender/remove_background/tests/test_dataprep.py
@@ -0,0 +1,101 @@
+"""Test functions in dataprep.py"""
+
+import pytest
+import scipy.sparse as sp
+import numpy as np
+import torch
+
+from cellbender.remove_background.data.dataprep import DataLoader
+from cellbender.remove_background.sparse_utils import dense_to_sparse_op_torch
+
+from .conftest import sparse_matrix_equal, simulated_dataset
+
+
+USE_CUDA = torch.cuda.is_available()
+
+
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+def test_dataloader_sorting(simulated_dataset, cuda):
+ """test dataset.py _overwrite_matrix_with_columns_from_another()"""
+
+ d = simulated_dataset
+ data_loader = DataLoader(
+ d['matrix'],
+ empty_drop_dataset=None,
+ batch_size=5,
+ fraction_empties=0.,
+ shuffle=False,
+ use_cuda=cuda,
+ )
+ sorted_data_loader = DataLoader(
+ d['matrix'],
+ empty_drop_dataset=None,
+ batch_size=5,
+ fraction_empties=0.,
+ shuffle=False,
+ sort_by=lambda x: -1 * np.array(x.max(axis=1).todense()).squeeze(),
+ use_cuda=cuda,
+ )
+
+ # try to shuffle and sort at the same time, and expect a failure
+ with pytest.raises(AssertionError):
+ sorted_data_loader2 = DataLoader(
+ d['matrix'],
+ empty_drop_dataset=None,
+ batch_size=5,
+ fraction_empties=0.,
+ shuffle=True,
+ sort_by=lambda x: -1 * np.array(x.max(axis=1).todense()).squeeze(),
+ use_cuda=cuda,
+ )
+
+ # this is copied from infer.BasePosterior._get_mean() which is not ideal
+ out = []
+ for loader in [data_loader, sorted_data_loader]:
+
+ barcodes = []
+ genes = []
+ counts = []
+ ind = 0
+
+ for data in loader:
+ dense_counts = data # just make it the same!
+
+ # Convert to sparse.
+ bcs_i_chunk, genes_i, counts_i = dense_to_sparse_op_torch(dense_counts)
+
+ # Barcode index in the dataloader.
+ bcs_i = bcs_i_chunk + ind
+
+ # Obtain the real barcode index after unsorting the dataloader.
+ bcs_i = loader.unsort_inds(bcs_i)
+
+ # Add sparse matrix values to lists.
+ barcodes.append(bcs_i)
+ genes.append(genes_i)
+ counts.append(counts_i)
+
+ # Increment barcode index counter.
+ ind += data.shape[0] # Same as data_loader.batch_size
+
+ # Convert the lists to numpy arrays.
+ counts = np.concatenate(counts).astype(np.uint32)
+ barcodes = np.concatenate(barcodes).astype(np.uint32)
+ genes = np.concatenate(genes).astype(np.uint32) # uint16 is too small!
+
+ print('counts')
+ print(counts)
+ print('barcodes')
+ print(barcodes)
+ print('genes')
+ print(genes)
+
+ # Put the counts into a sparse csc_matrix.
+ out.append(sp.csc_matrix((counts, (barcodes, genes)),
+ shape=d['matrix'].shape))
+
+ assert sparse_matrix_equal(out[0], out[1])
diff --git a/cellbender/remove_background/tests/test_dataset.py b/cellbender/remove_background/tests/test_dataset.py
new file mode 100644
index 00000000..08fdfef2
--- /dev/null
+++ b/cellbender/remove_background/tests/test_dataset.py
@@ -0,0 +1,40 @@
+"""Test functions in dataset.py"""
+
+import pytest
+import scipy.sparse as sp
+import numpy as np
+
+from .conftest import sparse_matrix_equal
+
+
+@pytest.mark.skip
+def test_heuristic_priors():
+ pass
+
+
+@pytest.mark.skip
+def test_feature_type_exclusion():
+ # TODO there seems to be an error
+ # TODO see https://github.com/broadinstitute/CellBender/issues/121#issuecomment-1443995082
+ pass
+
+
+@pytest.mark.skip
+def test_barcode_trimming():
+ pass
+
+
+@pytest.mark.skip
+def test_feature_trimming():
+ pass
+
+
+@pytest.mark.skip
+def test_restore_eliminated_features_in_cells():
+ pass
+
+
+@pytest.mark.skip
+def test_remove_zero_count_cells():
+ """Functionality yet to be written too"""
+ pass
diff --git a/cellbender/remove_background/tests/test_downstream.py b/cellbender/remove_background/tests/test_downstream.py
new file mode 100644
index 00000000..bf578823
--- /dev/null
+++ b/cellbender/remove_background/tests/test_downstream.py
@@ -0,0 +1,119 @@
+"""Test functions in downstream.py and a few more."""
+
+import pytest
+
+from cellbender.remove_background.downstream import anndata_from_h5, \
+ load_anndata_from_input_and_output
+from cellbender.remove_background.tests.conftest import SavedFileH5, \
+ h5_file, h5_v3_file_post_inference
+
+
+def convert(s):
+ """The h5 gets saved with different key names than the h5_file class"""
+ if s == 'gene_names':
+ return 'gene_name'
+ elif s == 'barcodes':
+ return 'barcode'
+ elif s == 'gene_ids':
+ return 'gene_id'
+ elif s == 'genomes':
+ return 'genome'
+ elif s == 'feature_types':
+ return 'feature_type'
+ return s
+
+
+def test_anndata_from_h5(h5_file: SavedFileH5):
+
+ # load AnnData and check its shape
+ adata = anndata_from_h5(file=h5_file.name)
+ assert h5_file.shape == adata.shape, 'Shape of loaded adata is not correct'
+
+ # ensure everything that was supposed to be saved is loaded properly
+ indices = [adata.obs.index.name, adata.var.index.name]
+ all_columns = list(adata.obs.columns) + list(adata.var.columns) + indices
+ for key in h5_file.keys:
+ key = convert(key)
+ assert key in all_columns, f'Saved information "{key}" is missing from loaded adata'
+
+
+@pytest.mark.parametrize('analyzed_bcs_only', [True, False])
+def test_anndata_from_inferred_h5(h5_v3_file_post_inference: SavedFileH5, analyzed_bcs_only):
+ """Make sure the extra stuff we save gets loaded by downstream loader.
+ TODO: tidy this up
+ """
+
+ adata = anndata_from_h5(file=h5_v3_file_post_inference.name, analyzed_barcodes_only=analyzed_bcs_only)
+ print(adata)
+ print(adata.obs.head())
+ print(adata.var.head())
+
+ # check the shape of the loaded AnnData
+ if analyzed_bcs_only:
+ expected_shape = (h5_v3_file_post_inference.barcodes_analyzed, h5_v3_file_post_inference.shape[1])
+ else:
+ expected_shape = h5_v3_file_post_inference.shape # all barcodes
+ assert adata.shape == expected_shape, 'Shape of loaded adata is not correct'
+
+ # ensure everything that was supposed to be saved in .obs and .obsm and .var is loaded properly
+ indices = [adata.obs.index.name, adata.var.index.name]
+ all_columns = list(adata.obs.columns) + list(adata.obsm.keys()) + list(adata.var.columns) + indices
+
+ relevant_keys = h5_v3_file_post_inference.keys
+ if analyzed_bcs_only:
+ relevant_keys = relevant_keys + h5_v3_file_post_inference.local_keys
+
+ for key in relevant_keys:
+ print(key)
+ key = convert(key)
+ assert key in all_columns, f'Saved information "{key}" is missing from loaded adata'
+
+ # ensure other things are also loading
+ extra_columns = list(adata.uns.keys())
+ relevant_keys = h5_v3_file_post_inference.global_keys + h5_v3_file_post_inference.meta_keys
+ if not analyzed_bcs_only:
+ relevant_keys = relevant_keys + h5_v3_file_post_inference.local_keys
+
+ for key in relevant_keys:
+ print(key)
+ assert key in extra_columns, \
+ f'Saved .uns information "{key}" is missing from adata: {adata.uns.keys()}'
+
+
+def test_load_anndata_from_input_and_output(h5_file, h5_v3_file_post_inference):
+ adata = load_anndata_from_input_and_output(input_file=h5_file.name,
+ output_file=h5_v3_file_post_inference.name,
+ gene_expression_encoding_key='gene_expression_encoding',
+ analyzed_barcodes_only=False)
+ print(adata)
+ assert h5_file.shape == adata.shape, \
+ 'Shape of loaded adata is not correct when loading all barcodes'
+ for key in (h5_v3_file_post_inference.local_keys
+ + h5_v3_file_post_inference.global_keys
+ + h5_v3_file_post_inference.meta_keys):
+ assert key in adata.uns.keys(), f'Key {key} missing from adata.uns'
+
+ adata2 = load_anndata_from_input_and_output(input_file=h5_file.name,
+ output_file=h5_v3_file_post_inference.name,
+ gene_expression_encoding_key='gene_expression_encoding',
+ analyzed_barcodes_only=True)
+ print(adata2)
+ assert h5_v3_file_post_inference.analyzed_shape == adata2.shape, \
+ 'Shape of loaded adata is not correct when loading only analyzed barcodes'
+ for key in h5_v3_file_post_inference.local_keys:
+ if key == 'gene_expression_encoding':
+ assert key in adata2.obsm.keys(), f'Key {key} missing from adata.obsm'
+ else:
+ assert key in adata2.obs.keys(), f'Key {key} missing from adata.obs'
+ for key in h5_v3_file_post_inference.global_keys + h5_v3_file_post_inference.meta_keys:
+ assert key in adata2.uns.keys(), f'Key {key} missing from adata.uns'
+
+
+@pytest.mark.skip(reason='TODO')
+def test_scanpy_loading():
+ pass
+
+
+@pytest.mark.skip(reason='TODO')
+def test_seurat_loading():
+ pass
diff --git a/cellbender/remove_background/tests/test_estimation.py b/cellbender/remove_background/tests/test_estimation.py
new file mode 100644
index 00000000..528ab675
--- /dev/null
+++ b/cellbender/remove_background/tests/test_estimation.py
@@ -0,0 +1,358 @@
+"""Test functions in estimation.py"""
+import pandas as pd
+import pytest
+import scipy.sparse as sp
+import numpy as np
+import torch
+
+from cellbender.remove_background.estimation import Mean, MAP, \
+ SingleSample, ThresholdCDF, MultipleChoiceKnapsack, pandas_grouped_apply
+from cellbender.remove_background.posterior import IndexConverter, \
+ dense_to_sparse_op_torch, log_prob_sparse_to_dense
+
+from typing import Dict, Union
+
+
+@pytest.fixture(scope='module')
+def log_prob_coo_base() -> Dict[str, Union[sp.coo_matrix, np.ndarray, Dict[int, int]]]:
+ n = -np.inf
+ m = np.array(
+ [[0, n, n, n, n, n, n, n, n, n], # map 0, mean 0
+ [n, 0, n, n, n, n, n, n, n, n], # map 1, mean 1
+ [n, n, 0, n, n, n, n, n, n, n], # map 2, mean 2
+ [-6, -2, np.log(1. - 2 * np.exp(-2) - np.exp(-6)), -2, n, n, n, n, n, n],
+ [-2.5, -1.5, -0.5, -3, np.log(1. - np.exp([-2.5, -1.5, -0.5, -3]).sum()),
+ n, n, n, n, n],
+ [-0.74, -1, -2, -4, np.log(1. - np.exp([-0.74, -1, -2, -4]).sum()),
+ n, n, n, n, n],
+ [-1, -0.74, -2, -4, np.log(1. - np.exp([-0.74, -1, -2, -4]).sum()),
+ n, n, n, n, n],
+ [-2, -1, -0.74, -4, np.log(1. - np.exp([-0.74, -1, -2, -4]).sum()),
+ n, n, n, n, n],
+ ]
+ )
+ # make m sparse, i.e. zero probability entries are absent
+ rows, cols, vals = dense_to_sparse_op_torch(torch.tensor(m), tensor_for_nonzeros=torch.tensor(m).exp())
+ # make it a bit more difficult by having an empty row at the beginning
+ rows = rows + 1
+ shape = list(m.shape)
+ shape[0] = shape[0] + 1
+ offset_dict = dict(zip(range(1, 9), [0] * 7 + [1])) # noise count offsets (last is 1)
+
+ maps = np.argmax(m, axis=1)
+ maps = maps + np.array([offset_dict[m] for m in offset_dict.keys()])
+ cdf_logic = (torch.logcumsumexp(torch.tensor(m), dim=-1) > np.log(0.5))
+ cdfs = [np.where(a)[0][0] for a in cdf_logic]
+ cdfs = cdfs + np.array([offset_dict[m] for m in offset_dict.keys()])
+
+ return {'coo': sp.coo_matrix((vals, (rows, cols)), shape=shape),
+ 'offsets': offset_dict, # not all noise counts start at zero
+ 'maps': np.array([0] + maps.tolist()),
+ 'cdfs': np.array([0] + cdfs.tolist())}
+
+
+@pytest.fixture(scope='module', params=['exact', 'filtered', 'unsorted'])
+def log_prob_coo(request, log_prob_coo_base) \
+ -> Dict[str, Union[sp.coo_matrix, np.ndarray, Dict[int, int]]]:
+ """When used as an input argument, this offers up a series of dicts that
+ can be used for tests"""
+ if request.param == 'exact':
+ return log_prob_coo_base
+
+ elif request.param == 'filtered':
+ coo = log_prob_coo_base['coo']
+ logic = (coo.data >= -6)
+ new_coo = sp.coo_matrix((coo.data[logic], (coo.row[logic], coo.col[logic])),
+ shape=coo.shape)
+ out = {'coo': new_coo}
+ out.update({k: v for k, v in log_prob_coo_base.items() if (k != 'coo')})
+ return out
+
+ elif request.param == 'unsorted':
+ coo = log_prob_coo_base['coo']
+ order = np.random.permutation(np.arange(len(coo.data)))
+ new_coo = sp.coo_matrix((coo.data[order], (coo.row[order], coo.col[order])),
+ shape=coo.shape)
+ out = {'coo': new_coo}
+ out.update({k: v for k, v in log_prob_coo_base.items() if (k != 'coo')})
+ return out
+
+ else:
+ raise ValueError(f'Test writing error: requested "{request.param}" log_prob_coo')
+
+
+@pytest.fixture(scope='module', params=['exact', 'filtered', 'unsorted'])
+def mckp_log_prob_coo(request, log_prob_coo_base) \
+ -> Dict[str, Union[sp.coo_matrix, np.ndarray, Dict[int, int]]]:
+ """When used as an input argument, this offers up a series of dicts that
+ can be used for tests.
+
+ NOTE: separate for MCKP because we cannot include an empty 'm' because it
+ throws everything off (which gene is what, etc.)
+ """
+ def _fix(v):
+ if type(v) == dict:
+ return {(k - 1): val for k, val in v.items()}
+ elif type(v) == sp.coo_matrix:
+ return _eliminate_row_zero(v)
+ else:
+ return v
+
+ def _eliminate_row_zero(coo_: sp.coo_matrix) -> sp.coo_matrix:
+ row = coo_.row - 1
+ shape = list(coo_.shape)
+ shape[0] = shape[0] - 1
+ return sp.coo_matrix((coo_.data, (row, coo_.col)), shape=shape)
+
+ if request.param == 'exact':
+ out = log_prob_coo_base
+
+ elif request.param == 'filtered':
+ coo = log_prob_coo_base['coo']
+ logic = (coo.data >= -6)
+ new_coo = sp.coo_matrix((coo.data[logic], (coo.row[logic], coo.col[logic])),
+ shape=coo.shape)
+ out = {'coo': new_coo}
+ out.update({k: v for k, v in log_prob_coo_base.items() if (k != 'coo')})
+
+ elif request.param == 'unsorted':
+ coo = log_prob_coo_base['coo']
+ order = np.random.permutation(np.arange(len(coo.data)))
+ new_coo = sp.coo_matrix((coo.data[order], (coo.row[order], coo.col[order])),
+ shape=coo.shape)
+ out = {'coo': new_coo}
+ out.update({k: v for k, v in log_prob_coo_base.items() if (k != 'coo')})
+
+ else:
+ raise ValueError(f'Test writing error: requested "{request.param}" log_prob_coo')
+
+ return {k: _fix(v) for k, v in out.items()}
+
+
+def test_single_sample(log_prob_coo):
+ """Test the single sample estimator"""
+
+ # the input
+ print(log_prob_coo)
+ print('input log probs')
+ dense = log_prob_sparse_to_dense(log_prob_coo['coo'])
+ print(dense)
+
+ # with this shape converter, we get one row, where each value is one m
+ converter = IndexConverter(total_n_cells=1,
+ total_n_genes=log_prob_coo['coo'].shape[0])
+
+ # set up and estimate
+ estimator = SingleSample(index_converter=converter)
+ noise_csr = estimator.estimate_noise(noise_log_prob_coo=log_prob_coo['coo'],
+ noise_offsets=log_prob_coo['offsets'])
+
+ # output
+ print('dense noise count estimate, per m')
+ out_per_m = np.array(noise_csr.todense()).squeeze()
+ print(out_per_m)
+
+ # test
+ for i in log_prob_coo['offsets'].keys():
+ print(f'testing "m" value {i}')
+ allowed_vals = np.arange(dense.shape[1])[dense[i, :] > -np.inf] + log_prob_coo['offsets'][i]
+ print('allowed values')
+ print(allowed_vals)
+ print('sample')
+ print(out_per_m[i])
+ assert out_per_m[i] in allowed_vals, \
+ f'sample {out_per_m[i]} is not allowed for {dense[i, :]}'
+
+
+def test_mean(log_prob_coo):
+ """Test the mean estimator"""
+
+ def _add_offsets_to_truth(truth: np.ndarray, offset_dict: Dict[int, int]):
+ return truth + np.array([offset_dict.get(m, 0) for m in range(len(truth))])
+
+ offset_dict = log_prob_coo['offsets']
+
+ # the input
+ print(log_prob_coo)
+ print('input log probs')
+ dense = log_prob_sparse_to_dense(log_prob_coo['coo'])
+ print(dense)
+
+ # with this shape converter, we get one row, where each value is one m
+ converter = IndexConverter(total_n_cells=1,
+ total_n_genes=log_prob_coo['coo'].shape[0])
+
+ # set up and estimate
+ estimator = Mean(index_converter=converter)
+ noise_csr = estimator.estimate_noise(noise_log_prob_coo=log_prob_coo['coo'],
+ noise_offsets=offset_dict)
+
+ # output
+ print('dense noise count estimate, per m')
+ out_per_m = np.array(noise_csr.todense()).squeeze()
+ print(out_per_m)
+
+ # truth
+ brute_force = np.matmul(np.arange(dense.shape[1]), np.exp(dense).transpose())
+ brute_force = _add_offsets_to_truth(truth=brute_force, offset_dict=offset_dict)
+ print('truth')
+ print(brute_force)
+
+ # test
+ np.testing.assert_allclose(out_per_m, brute_force)
+
+
+def test_map(log_prob_coo):
+ """Test the MAP estimator"""
+
+ offset_dict = log_prob_coo['offsets']
+
+ # the input
+ print(log_prob_coo)
+ print('input log probs')
+ print(log_prob_sparse_to_dense(log_prob_coo['coo']))
+
+ # with this shape converter, we get one row, where each value is one m
+ converter = IndexConverter(total_n_cells=1,
+ total_n_genes=log_prob_coo['coo'].shape[0])
+
+ # set up and estimate
+ estimator = MAP(index_converter=converter)
+ noise_csr = estimator.estimate_noise(noise_log_prob_coo=log_prob_coo['coo'],
+ noise_offsets=offset_dict)
+
+ # output
+ print('dense noise count estimate, per m')
+ out_per_m = np.array(noise_csr.todense()).squeeze()
+ print(out_per_m)
+ print('truth')
+ print(log_prob_coo['maps'])
+
+ # test
+ np.testing.assert_array_equal(out_per_m, log_prob_coo['maps'])
+
+
+def test_cdf(log_prob_coo):
+ """Test the estimator based on CDF thresholding"""
+
+ offset_dict = log_prob_coo['offsets']
+
+ # the input
+ print(log_prob_coo)
+ print('input log probs')
+ print(log_prob_sparse_to_dense(log_prob_coo['coo']))
+
+ # with this shape converter, we get one row, where each value is one m
+ converter = IndexConverter(total_n_cells=1,
+ total_n_genes=log_prob_coo['coo'].shape[0])
+
+ # set up and estimate
+ estimator = ThresholdCDF(index_converter=converter)
+ noise_csr = estimator.estimate_noise(noise_log_prob_coo=log_prob_coo['coo'],
+ noise_offsets=offset_dict,
+ q=0.5)
+
+ # output
+ print('dense noise count estimate, per m')
+ out_per_m = np.array(noise_csr.todense()).squeeze()
+ print(out_per_m)
+ print('truth')
+ print(log_prob_coo['cdfs'])
+
+ # test
+ np.testing.assert_array_equal(out_per_m, log_prob_coo['cdfs'])
+
+
+@pytest.mark.parametrize('n_chunks, parallel_compute',
+ ([1, False],
+ [2, False],
+ [2, True]), ids=['1chunk', '2chunks_1cpu', '2chunks_parallel'])
+@pytest.mark.parametrize('n_cells, target, truth, truth_mat',
+ ([1, np.zeros(8), np.array([0, 1, 2, 0, 0, 0, 0, 1]), None],
+ [1, np.ones(8), np.array([0, 1, 2, 1, 1, 1, 1, 1]), None],
+ [1, np.ones(8) * 2, np.array([0, 1, 2, 2, 2, 2, 2, 2]), None],
+ [4, np.zeros(2), np.array([2, 2]), None],
+ [4, np.ones(2) * 4, np.array([4, 4]), np.array([[0, 1], [2, 2], [2, 0], [0, 1]])],
+ [4, np.ones(2) * 9, np.array([9, 9]), np.array([[0, 1], [2, 3], [4, 2], [3, 3]])]),
+ ids=['1_cell_target_0',
+ '1_cell_target_1',
+ '1_cell_target_2',
+ '4_cell_target_0',
+ '4_cell_target_4',
+ '4_cell_target_9'])
+def test_mckp(mckp_log_prob_coo, n_cells, target, truth, truth_mat, n_chunks, parallel_compute):
+ """Test the multiple choice knapsack problem estimator"""
+
+ offset_dict = mckp_log_prob_coo['offsets']
+
+ # the input
+ print('input log probs ===============================================')
+ print(log_prob_sparse_to_dense(mckp_log_prob_coo['coo']))
+
+ # set up and estimate# with this shape converter, we have 1 cell with 8 genes
+ converter = IndexConverter(total_n_cells=n_cells,
+ total_n_genes=mckp_log_prob_coo['coo'].shape[0] // n_cells)
+ estimator = MultipleChoiceKnapsack(index_converter=converter)
+ noise_csr = estimator.estimate_noise(
+ noise_log_prob_coo=mckp_log_prob_coo['coo'],
+ noise_offsets=offset_dict,
+ noise_targets_per_gene=target,
+ verbose=True,
+ n_chunks=n_chunks,
+ use_multiple_processes=parallel_compute,
+ )
+
+ assert noise_csr.shape == (converter.total_n_cells, converter.total_n_genes)
+
+ # output
+ print('dense noise count estimate')
+ out_mat = np.array(noise_csr.todense())
+ print(out_mat)
+ print('noise counts per gene')
+ out = out_mat.sum(axis=0)
+ print(out)
+ print('truth')
+ print(truth)
+
+ # test
+ if truth_mat is not None:
+ np.testing.assert_array_equal(out_mat, truth_mat)
+ np.testing.assert_array_equal(out, truth)
+
+
+def _firstval(df):
+ return df['log_prob'].iat[0]
+
+
+def _meanval(df):
+ return df['log_prob'].mean()
+
+
+@pytest.mark.parametrize('fun', (_firstval, _meanval), ids=['first_value', 'mean'])
+def test_parallel_pandas_grouped_apply(fun):
+ """Test that the parallel apply gives the same thing as non-parallel"""
+
+ df = pd.DataFrame(data={'m': [0, 0, 0, 1, 1, 1, 2, 2, 2],
+ 'c': [0, 1, 2] * 3,
+ 'log_prob': [1, 2, 3, 4, 5, 6, 7, 8, 9]})
+ print('input data')
+ print(df)
+
+ reg = pandas_grouped_apply(
+ coo=sp.coo_matrix((df['log_prob'], (df['m'], df['c'])), shape=[3, 3]),
+ fun=fun,
+ parallel=False,
+ )
+ print('normal application of groupby apply')
+ print(reg)
+
+ parallel = pandas_grouped_apply(
+ coo=sp.coo_matrix((df['log_prob'], (df['m'], df['c'])), shape=[3, 3]),
+ fun=fun,
+ parallel=True,
+ )
+ print('parallel application of groupby apply')
+ print(parallel)
+
+ np.testing.assert_array_equal(reg['m'], parallel['m'])
+ np.testing.assert_array_equal(reg['result'], parallel['result'])
diff --git a/cellbender/remove_background/tests/test_infer.py b/cellbender/remove_background/tests/test_infer.py
new file mode 100644
index 00000000..06feb679
--- /dev/null
+++ b/cellbender/remove_background/tests/test_infer.py
@@ -0,0 +1,301 @@
+# """Test functions in infer.py"""
+#
+# import pytest
+# import scipy.sparse as sp
+# import numpy as np
+# import torch
+#
+# from cellbender.remove_background.data.dataprep import DataLoader
+# from cellbender.remove_background.infer import BasePosterior, Posterior, \
+# binary_search, dense_to_sparse_op_torch, dense_to_sparse_op_numpy
+#
+# from .conftest import sparse_matrix_equal, simulated_dataset, tensors_equal
+#
+#
+# USE_CUDA = torch.cuda.is_available()
+#
+#
+# @pytest.mark.parametrize('cuda',
+# [False,
+# pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+# reason='requires CUDA'))],
+# ids=lambda b: 'cuda' if b else 'cpu')
+# def test_dense_to_sparse_op_numpy(simulated_dataset, cuda):
+# """test infer.py BasePosterior.dense_to_sparse_op_numpy()"""
+#
+# d = simulated_dataset
+# data_loader = DataLoader(
+# d['matrix'],
+# empty_drop_dataset=None,
+# batch_size=5,
+# fraction_empties=0.,
+# shuffle=False,
+# use_cuda=cuda,
+# )
+#
+# barcodes = []
+# genes = []
+# counts = []
+# ind = 0
+#
+# for data in data_loader:
+# dense_counts = data # just make it the same!
+#
+# # Convert to sparse.
+# bcs_i_chunk, genes_i, counts_i = \
+# dense_to_sparse_op_numpy(dense_counts.detach().cpu().numpy())
+#
+# # Barcode index in the dataloader.
+# bcs_i = bcs_i_chunk + ind
+#
+# # Obtain the real barcode index after unsorting the dataloader.
+# bcs_i = data_loader.unsort_inds(bcs_i)
+#
+# # Add sparse matrix values to lists.
+# barcodes.append(bcs_i)
+# genes.append(genes_i)
+# counts.append(counts_i)
+#
+# # Increment barcode index counter.
+# ind += data.shape[0] # Same as data_loader.batch_size
+#
+# # Convert the lists to numpy arrays.
+# counts = np.array(np.concatenate(tuple(counts)), dtype=np.uint32)
+# barcodes = np.array(np.concatenate(tuple(barcodes)), dtype=np.uint32)
+# genes = np.array(np.concatenate(tuple(genes)), dtype=np.uint32) # uint16 is too small!
+#
+# # Put the counts into a sparse csc_matrix.
+# out = sp.csc_matrix((counts, (barcodes, genes)),
+# shape=d['matrix'].shape)
+#
+# assert sparse_matrix_equal(out, d['matrix'])
+#
+#
+# @pytest.mark.parametrize('cuda',
+# [False,
+# pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+# reason='requires CUDA'))],
+# ids=lambda b: 'cuda' if b else 'cpu')
+# def test_dense_to_sparse_op_torch(simulated_dataset, cuda):
+# """test infer.py BasePosterior.dense_to_sparse_op_torch()"""
+#
+# d = simulated_dataset
+# data_loader = DataLoader(
+# d['matrix'],
+# empty_drop_dataset=None,
+# batch_size=5,
+# fraction_empties=0.,
+# shuffle=False,
+# use_cuda=cuda,
+# )
+#
+# barcodes = []
+# genes = []
+# counts = []
+# ind = 0
+#
+# for data in data_loader:
+# dense_counts = data # just make it the same!
+#
+# # Convert to sparse.
+# bcs_i_chunk, genes_i, counts_i = \
+# dense_to_sparse_op_torch(dense_counts)
+#
+# # Barcode index in the dataloader.
+# bcs_i = bcs_i_chunk + ind
+#
+# # Obtain the real barcode index after unsorting the dataloader.
+# bcs_i = data_loader.unsort_inds(bcs_i)
+#
+# # Add sparse matrix values to lists.
+# barcodes.append(bcs_i)
+# genes.append(genes_i)
+# counts.append(counts_i)
+#
+# # Increment barcode index counter.
+# ind += data.shape[0] # Same as data_loader.batch_size
+#
+# # Convert the lists to numpy arrays.
+# counts = np.concatenate(counts).astype(np.uint32)
+# barcodes = np.concatenate(barcodes).astype(np.uint32)
+# genes = np.concatenate(genes).astype(np.uint32) # uint16 is too small!
+#
+# # Put the counts into a sparse csc_matrix.
+# out = sp.csc_matrix((counts, (barcodes, genes)),
+# shape=d['matrix'].shape)
+#
+# assert sparse_matrix_equal(out, d['matrix'])
+#
+#
+# def test_binary_search():
+# """Test the general binary search function."""
+#
+# tol = 0.001
+#
+# def fun1(x):
+# return x - 1.
+#
+# out = binary_search(evaluate_outcome_given_value=fun1,
+# target_outcome=torch.tensor([0.]),
+# init_range=torch.tensor([[0., 10.]]),
+# target_tolerance=tol)
+# print('Single value binary search')
+# print('Target value = [1.]')
+# print(f'Output = {out}')
+# assert ((out - torch.tensor([1.])).abs() <= tol).all(), \
+# 'Single input binary search failed'
+#
+# def fun2(x):
+# x = x.clone()
+# x[0] = x[0] - 1.
+# x[1] = x[1] - 2.
+# return x
+#
+# out = binary_search(evaluate_outcome_given_value=fun2,
+# target_outcome=torch.tensor([0., 0.]),
+# init_range=torch.tensor([[-10., 5.], [0., 10.]]),
+# target_tolerance=tol)
+# print('Two-value binary search')
+# print('Target value = [1., 2.]')
+# print(f'Output = {out}')
+# assert ((out - torch.tensor([1., 2.])).abs() <= tol).all(), \
+# 'Two-argument input binary search failed'
+#
+# @pytest.mark.parametrize('cuda',
+# [False,
+# pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+# reason='requires CUDA'))],
+# ids=lambda b: 'cuda' if b else 'cpu')
+# @pytest.mark.parametrize('fun',
+# [Posterior._mckp_noise_given_log_prob_tensor,
+# Posterior._mckp_noise_given_log_prob_tensor_fast],
+# ids=['inchworm', 'sort'])
+# def test_mckp_noise_given_log_prob_tensor(cuda, fun):
+# """Test the bespoke inchworm algorithm for solving a discrete
+# convex constrained optimization problem"""
+#
+# device = 'cuda' if cuda else 'cpu'
+#
+# poisson_noise_means = torch.tensor([[1.9160, 0.5520],
+# [2.7160, 0.0840],
+# [3.9080, 2.5280]]).to(device)
+# log_p_ngc = (torch.distributions.Poisson(rate=poisson_noise_means.unsqueeze(-1))
+# .log_prob(torch.arange(10).to(device)))
+#
+# debug = True
+#
+# print('\nA couple normal test cases')
+# out = fun(
+# log_prob_noise_counts_NGC=log_p_ngc,
+# offset_noise_counts_NG=torch.zeros_like(poisson_noise_means),
+# data_NG=torch.ones_like(poisson_noise_means) * 10.,
+# target_G=torch.tensor([1., 3.]).to(device),
+# debug=debug,
+# )
+# assert tensors_equal(out[0], torch.tensor([[0., 0.], [0., 0.], [1., 3.]]).to(device))
+# assert tensors_equal(out[1], torch.tensor([0., 0.]).to(device))
+#
+# out = fun(
+# log_prob_noise_counts_NGC=log_p_ngc,
+# offset_noise_counts_NG=torch.zeros_like(poisson_noise_means),
+# data_NG=torch.ones_like(poisson_noise_means) * 10.,
+# target_G=torch.tensor([6., 6.]).to(device),
+# debug=debug,
+# )
+# assert tensors_equal(out[0], torch.tensor([[1., 1.], [2., 0.], [3., 5.]]).to(device))
+# assert tensors_equal(out[1], torch.tensor([0., 0.]).to(device))
+#
+# print('\nEdge case of zero removal')
+# out = fun(
+# log_prob_noise_counts_NGC=log_p_ngc,
+# offset_noise_counts_NG=torch.zeros_like(poisson_noise_means),
+# data_NG=torch.ones_like(poisson_noise_means) * 10.,
+# target_G=torch.tensor([0., 0.]).to(device),
+# debug=debug,
+# )
+# assert tensors_equal(out[0], torch.tensor([[0., 0.], [0., 0.], [0., 0.]]).to(device))
+# assert tensors_equal(out[1], torch.tensor([0., 0.]).to(device))
+#
+# print('\nEdge case of massive target')
+# out = fun(
+# log_prob_noise_counts_NGC=log_p_ngc,
+# offset_noise_counts_NG=torch.zeros_like(poisson_noise_means),
+# data_NG=torch.ones_like(poisson_noise_means) * 10.,
+# target_G=torch.tensor([100., 0.]).to(device),
+# debug=debug,
+# )
+# assert tensors_equal(out[0], torch.tensor([[9., 0.], [9., 0.], [9., 0.]]).to(device))
+# assert tensors_equal(out[1], torch.tensor([73., 0.]).to(device))
+#
+# print('\nNonzero offset noise counts')
+# out = fun(
+# log_prob_noise_counts_NGC=log_p_ngc,
+# offset_noise_counts_NG=torch.ones_like(poisson_noise_means),
+# data_NG=torch.ones_like(poisson_noise_means) * 10.,
+# target_G=torch.tensor([3., 3.]).to(device),
+# debug=debug,
+# )
+# assert tensors_equal(out[0], torch.tensor([[1., 1.], [1., 1.], [1., 1.]]).to(device))
+# assert tensors_equal(out[1], torch.tensor([0., 0.]).to(device))
+#
+# print('\nNonzero offset noise counts with zero target')
+# out = fun(
+# log_prob_noise_counts_NGC=log_p_ngc,
+# offset_noise_counts_NG=torch.tensor([[1., 0.], [0., 0.], [0., 0.]]).to(device),
+# data_NG=torch.ones_like(poisson_noise_means) * 10.,
+# target_G=torch.tensor([0., 0.]).to(device),
+# debug=debug,
+# )
+# assert tensors_equal(out[0], torch.tensor([[1., 0.], [0., 0.], [0., 0.]]).to(device))
+# assert tensors_equal(out[1], torch.tensor([-1., 0.]).to(device))
+#
+# print('\nNonzero offset noise counts')
+# out = fun(
+# log_prob_noise_counts_NGC=log_p_ngc,
+# offset_noise_counts_NG=torch.ones_like(poisson_noise_means),
+# data_NG=torch.ones_like(poisson_noise_means) * 10.,
+# target_G=torch.tensor([6., 6.]).to(device),
+# debug=debug,
+# )
+# assert tensors_equal(out[0], torch.tensor([[1., 1.], [2., 1.], [3., 4.]]).to(device))
+# assert tensors_equal(out[1], torch.tensor([0., 0.]).to(device))
+#
+# # # This cannot happen for a properly constructed log prob tensor:
+# # print('Argmax > data')
+# # out = fun(
+# # log_prob_noise_counts_NGC=log_p_ngc,
+# # offset_noise_counts_NG=torch.ones_like(poisson_noise_means),
+# # data_NG=torch.ones_like(poisson_noise_means) * 1.,
+# # target_G=torch.tensor([6., 6.]).to(device),
+# # debug=debug,
+# # )
+# # assert tensors_equal(out[0], torch.tensor([[1., 1.], [1., 1.], [1., 1.]]).to(device))
+# # assert tensors_equal(out[1], torch.tensor([4., 4.]).to(device))
+#
+# print('\nExtra genes that have no moves')
+# poisson_noise_means = torch.tensor([[1.9160, 0.5520, 0.1, 0.1],
+# [2.7160, 0.0840, 0.2, 0.3],
+# [3.9080, 2.5280, 0.3, 0.2]]).to(device)
+# data_NG = torch.tensor([[2, 0, 0, 0],
+# [0, 0, 1, 0],
+# [0, 0, 0, 0]]).float().to(device)
+# log_p_ngc = (torch.distributions.Poisson(rate=poisson_noise_means.unsqueeze(-1))
+# .log_prob(torch.arange(10).to(device)))
+# log_p_ngc = torch.where(torch.arange(10).to(device).unsqueeze(0).unsqueeze(0) > data_NG.unsqueeze(-1),
+# torch.ones_like(log_p_ngc) * -np.inf,
+# log_p_ngc)
+# target_G = torch.tensor([2., 2., 2., 2.]).to(device)
+# print('data')
+# print(data_NG)
+# print('log_prob')
+# print(log_p_ngc)
+# out = fun(
+# log_prob_noise_counts_NGC=log_p_ngc,
+# offset_noise_counts_NG=torch.zeros_like(poisson_noise_means),
+# data_NG=data_NG,
+# target_G=target_G,
+# debug=debug,
+# )
+# assert tensors_equal(out[0], data_NG)
+# assert tensors_equal(out[1], target_G - data_NG.sum(dim=0))
+#
diff --git a/cellbender/remove_background/tests/test_integration.py b/cellbender/remove_background/tests/test_integration.py
new file mode 100644
index 00000000..624c59df
--- /dev/null
+++ b/cellbender/remove_background/tests/test_integration.py
@@ -0,0 +1,52 @@
+"""Full run through on small simulated data"""
+
+from cellbender.remove_background.cli import CLI
+from cellbender.base_cli import get_populated_argparser
+from cellbender.remove_background.downstream import anndata_from_h5
+from cellbender.remove_background import consts
+import numpy as np
+import os
+import pytest
+
+from .conftest import USE_CUDA
+
+
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+def test_full_run(tmpdir_factory, h5_v3_file, cuda):
+ """Do a full run of the command line tool using a small simulated dataset"""
+
+ tmp_dir = tmpdir_factory.mktemp('data')
+ filename = tmp_dir.join('out.h5')
+
+ os.chdir(tmp_dir) # so checkpoint file goes to temp dir and gets removed
+
+ # handle input arguments using the argparser
+ input_args = ['cellbender', 'remove-background',
+ '--input', str(h5_v3_file.name),
+ '--output', str(filename),
+ '--epochs', '5']
+ if cuda:
+ input_args.append('--cuda')
+ args = get_populated_argparser().parse_args(input_args[1:])
+ args = CLI.validate_args(args=args)
+
+ # do a full run through
+ posterior = CLI.run(args=args)
+
+ # do some checks
+
+ # ensure the cell probabilities in the posterior object match the output file
+ p_for_analyzed_barcodes = posterior.latents_map['p']
+ adata = anndata_from_h5(str(filename), analyzed_barcodes_only=True)
+ file_p_for_analyzed_barcodes = adata.obs['cell_probability'].values
+ np.testing.assert_array_equal(p_for_analyzed_barcodes, file_p_for_analyzed_barcodes)
+
+ # ensure the cell barcodes are the same both ways
+ cell_barcodes = np.genfromtxt(str(filename)[:-3] + '_cell_barcodes.csv', dtype=str, delimiter='\n')
+ adata_cell_barcodes = adata.obs_names[adata.obs['cell_probability'] > consts.CELL_PROB_CUTOFF]
+ assert set(cell_barcodes) == set(adata_cell_barcodes), \
+ 'Cell barcodes in h5 are different from those in CSV file'
diff --git a/cellbender/remove_background/tests/test_io.py b/cellbender/remove_background/tests/test_io.py
new file mode 100644
index 00000000..06d23ce1
--- /dev/null
+++ b/cellbender/remove_background/tests/test_io.py
@@ -0,0 +1,223 @@
+"""Test input reading and output writing functionality."""
+
+import pytest
+import numpy as np
+from scipy.io import mmwrite
+import scipy.sparse as sp
+
+from cellbender.remove_background.data.io import \
+ load_data, get_matrix_from_cellranger_mtx, get_matrix_from_dropseq_dge, \
+ get_matrix_from_bd_rhapsody, get_matrix_from_anndata, \
+ detect_cellranger_version_h5, detect_cellranger_version_mtx, unravel_dict
+
+from cellbender.remove_background.tests.conftest import \
+ sparse_matrix_equal, string_ndarray_equality, h5_v2_file, \
+ h5_v2_file_missing_ids, h5_v3_file, h5_file
+
+from typing import List, Dict, Optional
+import gzip
+import shutil
+
+
+def assert_loaded_matches_saved(d: Dict[str, np.ndarray],
+ loaded: Dict[str, np.ndarray],
+ keys: List[str],
+ cellranger_version: Optional[int] = None):
+ """Check if a loaded file's data matches the data that was saved.
+
+ Args:
+ d: Dict of the data that was saved
+ loaded: Dict of the data that was loaded from file
+ keys: List of keys to the dicts that will be checked for equality
+ cellranger_version: In [2, 3]
+ """
+ if 'cellranger_version' in loaded.keys():
+ assert cellranger_version == loaded['cellranger_version']
+ assert sparse_matrix_equal(loaded['matrix'], d['matrix'])
+ for key in keys:
+ if d[key] is None:
+ continue
+ assert loaded[key] is not None, \
+ f'Loaded h5 key "{key}" was None, but data was saved: {d[key][:5]} ...'
+ assert string_ndarray_equality(d[key], loaded[key]), \
+ f'Loaded h5 key "{key}" did not match saved data'
+
+
+@pytest.mark.parametrize('filetype', ['h5_v2_file', 'h5_v2_file_missing_ids', 'h5_v3_file'])
+def test_simulate_save_load_h5(simulated_dataset,
+ filetype,
+ h5_v2_file,
+ h5_v2_file_missing_ids,
+ h5_v3_file):
+
+ # get information from fixture, since you cannot pass fixtures to parametrize
+ if filetype == 'h5_v2_file':
+ saved_h5 = h5_v2_file
+ elif filetype == 'h5_v2_file_missing_ids':
+ saved_h5 = h5_v2_file_missing_ids
+ elif filetype == 'h5_v3_file':
+ saved_h5 = h5_v3_file
+
+ # load data from file, using auto-loading, as it would be run
+ loaded = load_data(input_file=saved_h5.name)
+
+ # assert equality
+ assert_loaded_matches_saved(
+ d=simulated_dataset,
+ loaded=loaded,
+ keys=saved_h5.keys,
+ cellranger_version=saved_h5.version,
+ )
+
+
+def test_detect_cellranger_version_h5(h5_file):
+ v = detect_cellranger_version_h5(filename=h5_file.name)
+ true_version = h5_file.version
+ assert v == true_version
+
+
+def gzip_file(file):
+ """gzip a file"""
+ with open(file, 'rb') as f_in, gzip.open(file + '.gz', 'wb') as f_out:
+ f_out.writelines(f_in)
+
+
+def save_mtx(tmpdir_factory, simulated_dataset, version: int) -> str:
+ """Save data files in MTX format and return the directory path"""
+ dirname = tmpdir_factory.mktemp(f'mtx_v{version}')
+
+ # barcodes and sparse matrix... seems the MTX matrix is transposed
+ mmwrite(dirname.join('matrix.mtx'), simulated_dataset['matrix'].transpose())
+ np.savetxt(dirname.join('barcodes.tsv'), simulated_dataset['barcodes'], fmt='%s')
+
+ # features and gzipping if v3
+ features = np.concatenate((np.expand_dims(simulated_dataset['gene_ids'], axis=1),
+ np.expand_dims(simulated_dataset['gene_names'], axis=1),
+ np.expand_dims(simulated_dataset['feature_types'], axis=1)), axis=1)
+
+ if version == 3:
+ np.savetxt(dirname.join('features.tsv'), features, fmt='%s', delimiter='\t')
+ gzip_file(dirname.join('matrix.mtx'))
+ gzip_file(dirname.join('barcodes.tsv'))
+ gzip_file(dirname.join('features.tsv'))
+ elif version == 2:
+ np.savetxt(dirname.join('genes.tsv'), features[:, :2], fmt='%s', delimiter='\t')
+ else:
+ raise ValueError(f'Test problem: version is {version}, but [2, 3] allowed')
+
+ return dirname
+
+
+@pytest.fixture(scope='session', params=[2, 3])
+def mtx_directory(request, tmpdir_factory, simulated_dataset):
+ dirname = save_mtx(tmpdir_factory=tmpdir_factory,
+ simulated_dataset=simulated_dataset,
+ version=request.param)
+ yield dirname
+ shutil.rmtree(str(dirname))
+
+
+def test_detect_cellranger_version_mtx(mtx_directory):
+ v = detect_cellranger_version_mtx(filedir=mtx_directory)
+ true_version = 2 if ('_v2' in str(mtx_directory)) else 3
+ assert v == true_version
+
+
+def test_load_mtx(simulated_dataset, mtx_directory):
+
+ # use the correct loader function
+ loaded = get_matrix_from_cellranger_mtx(filedir=mtx_directory)
+ version = 3 if ('_v3' in str(mtx_directory)) else 2
+ assert_loaded_matches_saved(d=simulated_dataset,
+ loaded=loaded,
+ keys=(['gene_ids', 'gene_names', 'barcodes']
+ + (['feature_types'] if (version == 3) else [])),
+ cellranger_version=version)
+
+ # use auto-loading, as it would be run
+ loaded = load_data(mtx_directory)
+ assert_loaded_matches_saved(d=simulated_dataset,
+ loaded=loaded,
+ keys=(['gene_ids', 'gene_names', 'barcodes']
+ + (['feature_types'] if (version == 3) else [])),
+ cellranger_version=version)
+
+
+def save_dge(tmpdir_factory, simulated_dataset, do_gzip) -> str:
+ """Save data files in DGE format and return the file path"""
+ sep = '\t'
+ name = 'dge.txt'
+ if do_gzip:
+ name = name + '.gz'
+ tmp_dir = tmpdir_factory.mktemp('dge')
+ filename = tmp_dir.join(name)
+ load_fcn = gzip.open if do_gzip else open
+
+ def row_generator(mat: sp.csc_matrix) -> List[str]:
+ for i in range(mat.shape[1]):
+ yield np.array(mat[:, i].todense()).squeeze().astype(int).astype(str).tolist()
+
+ with load_fcn(filename, 'wb') as f:
+ f.write(b'# some kind of header!\n')
+ f.write(sep.join(['GENE'] + simulated_dataset['barcodes'].astype(str).tolist()).encode() + b'\n')
+ for g, vals in zip(simulated_dataset['gene_names'], row_generator(simulated_dataset['matrix'])):
+ f.write(sep.join([g] + vals).encode() + b'\n')
+
+ return filename, tmp_dir
+
+
+@pytest.fixture(scope='session', params=[True, False], ids=lambda x: 'gzipped' if x else 'not')
+def dge_file(request, tmpdir_factory, simulated_dataset):
+ filename, tmp_dir = save_dge(tmpdir_factory=tmpdir_factory,
+ simulated_dataset=simulated_dataset,
+ do_gzip=request.param)
+ yield filename
+ shutil.rmtree(str(tmp_dir))
+
+
+def test_load_dge(simulated_dataset, dge_file):
+
+ # use the correct loader function
+ loaded = get_matrix_from_dropseq_dge(str(dge_file))
+ assert_loaded_matches_saved(d=simulated_dataset,
+ loaded=loaded,
+ keys=['gene_names', 'barcodes'])
+
+ # use auto-loading, as it would be run
+ loaded = load_data(str(dge_file))
+ assert_loaded_matches_saved(d=simulated_dataset,
+ loaded=loaded,
+ keys=['gene_names', 'barcodes'])
+
+
+@pytest.mark.skip
+def test_load_bd():
+ pass
+
+
+@pytest.mark.skip
+def test_load_anndata():
+ pass
+
+
+@pytest.mark.skip
+def test_load_loom():
+ pass
+
+
+@pytest.mark.skip
+def test_write_matrix_to_cellranger_h5():
+ pass
+
+
+@pytest.mark.skip
+def test_write_denoised_count_matrix():
+ # from run.py, but should probably be refactored to io.py
+ pass
+
+
+def test_unravel_dict():
+ key, value = 'pref', {'a': 1, 'b': {'c': 2, 'd': {'e': 3, 'f': 4}}}
+ answer = {'pref_a': 1, 'pref_b_c': 2, 'pref_b_d_e': 3, 'pref_b_d_f': 4}
+ d = unravel_dict(key, value)
+ assert d == answer, 'unravel_dict failed to produce correct output'
diff --git a/cellbender/remove_background/tests/test_monitor.py b/cellbender/remove_background/tests/test_monitor.py
new file mode 100644
index 00000000..68735074
--- /dev/null
+++ b/cellbender/remove_background/tests/test_monitor.py
@@ -0,0 +1,21 @@
+"""Tests for monitoring function."""
+
+import pytest
+import torch
+
+from cellbender.monitor import get_hardware_usage
+
+USE_CUDA = torch.cuda.is_available()
+
+
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+def test_get_hardware_usage(cuda):
+ """Check and see if restarting from a checkpoint picks up in the same place
+ we left off. Use our model and dataloader.
+ """
+
+ print(get_hardware_usage(use_cuda=cuda))
diff --git a/cellbender/remove_background/tests/test_posterior.py b/cellbender/remove_background/tests/test_posterior.py
new file mode 100644
index 00000000..e5382b3b
--- /dev/null
+++ b/cellbender/remove_background/tests/test_posterior.py
@@ -0,0 +1,428 @@
+"""Test functions in posterior.py"""
+
+import pytest
+import scipy.sparse as sp
+import numpy as np
+import torch
+
+from cellbender.remove_background.data.dataprep import DataLoader
+from cellbender.remove_background.posterior import Posterior, torch_binary_search, \
+ PRmu, PRq, IndexConverter, compute_mean_target_removal_as_function
+from cellbender.remove_background.sparse_utils import dense_to_sparse_op_torch, \
+ log_prob_sparse_to_dense, todense_fill
+from cellbender.remove_background.estimation import Mean
+
+import warnings
+from typing import Dict, Union
+
+from .conftest import sparse_matrix_equal, simulated_dataset, tensors_equal
+
+
+USE_CUDA = torch.cuda.is_available()
+
+
+# NOTE: issues caught
+# - have a test that actually creates a posterior
+# - using uint32 for barcode index in a COO caused integer overflow
+
+
+@pytest.fixture(scope='module')
+def log_prob_coo_base() -> Dict[str, Union[sp.coo_matrix, np.ndarray, Dict[int, int]]]:
+ n = -np.inf
+ m = np.array(
+ [[0, n, n, n, n, n, n, n], # map 0, mean 0
+ [n, 0, n, n, n, n, n, n], # map 1, mean 1
+ [-0.3, -1.5, np.log(1. - np.exp(np.array([-0.3, -1.5])).sum())] + [n] * 5,
+ [-3, -1.21, -0.7, -2, -4, np.log(1. - np.exp(np.array([-3, -1.21, -0.7, -2, -4])).sum())] + [n] * 2,
+ ]
+ )
+ # make m sparse, i.e. zero probability entries are absent
+ rows, cols, vals = dense_to_sparse_op_torch(torch.tensor(m), tensor_for_nonzeros=torch.tensor(m).exp())
+ # make it a bit more difficult by having an empty row at the beginning
+ rows = rows + 1
+ shape = list(m.shape)
+ shape[0] = shape[0] + 1
+ offset_dict = dict(zip(range(1, len(m) + 1), [0] * len(m) + [1])) # noise count offsets (last is 1)
+ return {'coo': sp.coo_matrix((vals, (rows, cols)), shape=shape),
+ 'offsets': offset_dict}
+
+
+@pytest.fixture(scope='module', params=['sorted', 'unsorted'])
+def log_prob_coo(request, log_prob_coo_base) \
+ -> Dict[str, Union[sp.coo_matrix, np.ndarray, Dict[int, int]]]:
+ """When used as an input argument, this offers up a series of dicts that
+ can be used for tests"""
+ if request.param == 'sorted':
+ return log_prob_coo_base
+
+ elif request.param == 'unsorted':
+ coo = log_prob_coo_base['coo']
+ order = np.random.permutation(np.arange(len(coo.data)))
+ new_coo = sp.coo_matrix((coo.data[order], (coo.row[order], coo.col[order])),
+ shape=coo.shape)
+ out = {'coo': new_coo}
+ out.update({k: v for k, v in log_prob_coo_base.items() if (k != 'coo')})
+ return out
+
+ else:
+ raise ValueError(f'Test writing error: requested "{request.param}" log_prob_coo')
+
+
+@pytest.mark.parametrize('alpha', [0, 1, 2], ids=lambda a: f'alpha{a}')
+@pytest.mark.parametrize('n_chunks', [1, 2], ids=lambda n: f'{n}chunks')
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+def test_PRq(log_prob_coo, alpha, n_chunks, cuda):
+
+ target_tolerance = 0.001
+
+ print('input log_prob matrix, densified')
+ dense_input_log_prob = torch.tensor(log_prob_sparse_to_dense(log_prob_coo['coo'])).float()
+ print(dense_input_log_prob)
+ print('probability sums per row')
+ print(torch.logsumexp(dense_input_log_prob, dim=-1).exp())
+ print('(row 0 is a missing row)')
+
+ print('input mean noise counts per row')
+ counts = torch.arange(dense_input_log_prob.shape[-1]).float().unsqueeze(dim=0)
+ input_means = (dense_input_log_prob.exp() * counts).sum(dim=-1)
+ print(input_means)
+
+ print('input std per row')
+ input_std = (dense_input_log_prob.exp()
+ * (counts - input_means.unsqueeze(dim=-1)).pow(2)).sum(dim=-1).sqrt()
+ print(input_std)
+
+ print('\ntruth: expected means after regularization')
+ truth_means_after_regularization = input_means + alpha * input_std
+ print(truth_means_after_regularization)
+ print('and the log')
+ with warnings.catch_warnings():
+ warnings.filterwarnings("ignore", message="divide by zero encountered in log")
+ print(np.log(truth_means_after_regularization))
+
+ print('testing compute_log_target_dict()')
+ target_dict = PRq._compute_log_target_dict(noise_count_posterior_coo=log_prob_coo['coo'],
+ alpha=alpha)
+ print(target_dict)
+ for m in target_dict.keys():
+ with warnings.catch_warnings():
+ warnings.filterwarnings("ignore", message="divide by zero encountered in log")
+ np.testing.assert_almost_equal(target_dict[m],
+ np.log(truth_means_after_regularization[m]))
+
+ print('targets are correct\n\n')
+
+ print('means after regularization')
+ regularized_coo = PRq.regularize(
+ noise_count_posterior_coo=log_prob_coo['coo'],
+ noise_offsets=log_prob_coo['offsets'],
+ alpha=alpha,
+ device='cuda' if cuda else 'cpu',
+ target_tolerance=target_tolerance,
+ n_chunks=n_chunks,
+ )
+ print('regularized posterior:')
+ dense_regularized_log_prob = torch.tensor(log_prob_sparse_to_dense(regularized_coo)).float()
+ print(dense_regularized_log_prob)
+ means_after_regularization = (dense_regularized_log_prob.exp() * counts).sum(dim=-1)
+ print('means after regularization:')
+ print(means_after_regularization)
+
+ torch.testing.assert_close(
+ actual=truth_means_after_regularization,
+ expected=means_after_regularization,
+ rtol=target_tolerance,
+ atol=target_tolerance,
+ )
+
+
+@pytest.mark.parametrize('fpr', [0., 0.1, 1], ids=lambda a: f'fpr{a}')
+@pytest.mark.parametrize('n_chunks', [1, 2], ids=lambda n: f'{n}chunks')
+# @pytest.mark.parametrize('per_gene', [False, True], ids=lambda n: 'per_gene' if n else 'overall')
+@pytest.mark.parametrize('per_gene', [False], ids=lambda n: 'per_gene' if n else 'overall')
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+def test_PRmu(log_prob_coo, fpr, per_gene, n_chunks, cuda):
+
+ target_tolerance = 0.5
+
+ index_converter = IndexConverter(total_n_cells=log_prob_coo['coo'].shape[0], total_n_genes=1)
+ print(index_converter)
+
+ print('raw count matrix')
+ count_matrix = sp.csr_matrix(np.expand_dims(np.array([0, 0, 1, 2, 5]), axis=-1)) # reflecting filled in log_prob values
+ print(count_matrix)
+
+ print('input log_prob matrix, densified')
+ dense_input_log_prob = torch.tensor(log_prob_sparse_to_dense(log_prob_coo['coo'])).float()
+ print(dense_input_log_prob)
+ print('probability sums per row')
+ print(torch.logsumexp(dense_input_log_prob, dim=-1).exp())
+ print('(row 0 is a missing row)')
+
+ estimator = Mean(index_converter=index_converter)
+ mean_noise_csr = estimator.estimate_noise(
+ noise_log_prob_coo=log_prob_coo['coo'],
+ noise_offsets=log_prob_coo['offsets'],
+ device='cuda' if cuda else 'cpu',
+ )
+ print(f'Mean estimator removes {mean_noise_csr.sum()} counts total')
+
+ print('testing compute_target_removal()')
+ n_cells = 4 # hard coded from the log_prob_coo
+ target_fun = compute_mean_target_removal_as_function(
+ noise_count_posterior_coo=log_prob_coo['coo'],
+ noise_offsets=log_prob_coo['offsets'],
+ raw_count_csr_for_cells=count_matrix,
+ n_cells=n_cells,
+ index_converter=index_converter,
+ device='cuda' if cuda else 'cpu',
+ per_gene=per_gene,
+ )
+ targets = target_fun(fpr)
+ print(f'aiming to remove {targets} overall counts per cell')
+ print(f'so about {targets * n_cells} counts total')
+
+ print('means after regularization')
+ regularized_coo = PRmu.regularize(
+ noise_count_posterior_coo=log_prob_coo['coo'],
+ noise_offsets=log_prob_coo['offsets'],
+ index_converter=index_converter,
+ raw_count_matrix=count_matrix,
+ fpr=fpr,
+ per_gene=per_gene,
+ device='cuda' if cuda else 'cpu',
+ target_tolerance=target_tolerance,
+ n_chunks=n_chunks,
+ )
+ print('regularized posterior:')
+ dense_regularized_log_prob = torch.tensor(log_prob_sparse_to_dense(regularized_coo)).float()
+ print(dense_regularized_log_prob)
+
+ print('MAP noise:')
+ map_noise = torch.argmax(dense_regularized_log_prob, dim=-1)
+ print(map_noise)
+
+ if fpr == 0.:
+ torch.testing.assert_close(
+ actual=map_noise.sum().float(),
+ expected=torch.tensor(mean_noise_csr.sum()).float(),
+ rtol=1,
+ atol=1,
+ )
+ elif fpr == 1.:
+ torch.testing.assert_close(
+ actual=map_noise.sum().float(),
+ expected=torch.tensor(count_matrix.sum()).float(),
+ rtol=1,
+ atol=1,
+ )
+ else:
+ assert torch.tensor(mean_noise_csr.sum()).float() - 1 <= map_noise.sum().float(), \
+ 'Noise estimate is less than Mean estimator'
+ assert torch.tensor(count_matrix.sum()).float() >= map_noise.sum().float(), \
+ 'Noise estimate is greater than sum of counts... this should never happen'
+
+ # TODO: this test is very weak... it's just hard to test it exactly...
+ # TODO: passing should mean the code will run, but not that it's quantitatively accurate
+
+
+@pytest.mark.skip
+def test_create_posterior():
+ pass
+
+
+def test_index_converter():
+ index_converter = IndexConverter(total_n_cells=10, total_n_genes=5)
+ print(index_converter)
+
+ # check basic conversion
+ n = np.array([0, 1, 2, 3])
+ g = n.copy()
+ m = index_converter.get_m_indices(cell_inds=n, gene_inds=g)
+ print(f'm inds are {m}')
+ truth = 5 * n + g
+ print(f'expected {truth}')
+ np.testing.assert_equal(m, truth)
+
+ # back and forth
+ n_star, g_star = index_converter.get_ng_indices(m_inds=m)
+ np.testing.assert_equal(n, n_star)
+ np.testing.assert_equal(g, g_star)
+
+ # check on input validity checking
+ with pytest.raises(ValueError):
+ index_converter.get_m_indices(cell_inds=np.array([-1]), gene_inds=g)
+ with pytest.raises(ValueError):
+ index_converter.get_m_indices(cell_inds=np.array([10]), gene_inds=g)
+ with pytest.raises(ValueError):
+ index_converter.get_m_indices(cell_inds=n, gene_inds=np.array([-1]))
+ with pytest.raises(ValueError):
+ index_converter.get_m_indices(cell_inds=n, gene_inds=np.array([5]))
+ with pytest.raises(ValueError):
+ index_converter.get_ng_indices(m_inds=np.array([-1]))
+ with pytest.raises(ValueError):
+ index_converter.get_ng_indices(m_inds=np.array([10 * 5]))
+
+
+@pytest.mark.skip
+def test_estimation_array_to_csr():
+ Posterior._estimation_array_to_csr()
+ pass
+
+
+def test_torch_binary_search():
+ """Test the general binary search function."""
+
+ tol = 0.001
+
+ def fun1(x):
+ return x - 1.
+
+ out = torch_binary_search(
+ evaluate_outcome_given_value=fun1,
+ target_outcome=torch.tensor([0.]),
+ init_range=torch.tensor([[0., 10.]]),
+ target_tolerance=tol,
+ )
+ print('Single value binary search')
+ print('Target value = [1.]')
+ print(f'Output = {out}')
+ assert ((out - torch.tensor([1.])).abs() <= tol).all(), \
+ 'Single input binary search failed'
+
+ def fun2(x):
+ x = x.clone()
+ x[0] = x[0] - 1.
+ x[1] = x[1] - 2.
+ return x
+
+ out = torch_binary_search(
+ evaluate_outcome_given_value=fun2,
+ target_outcome=torch.tensor([0., 0.]),
+ init_range=torch.tensor([[-10., 5.], [0., 10.]]),
+ target_tolerance=tol,
+ )
+ print('Two-value binary search')
+ print('Target value = [1., 2.]')
+ print(f'Output = {out}')
+ assert ((out - torch.tensor([1., 2.])).abs() <= tol).all(), \
+ 'Two-argument input binary search failed'
+
+
+@pytest.mark.parametrize('fpr', [0., 0.1, 0.5, 0.75, 1], ids=lambda a: f'fpr{a}')
+@pytest.mark.parametrize('per_gene', [False], ids=lambda n: 'per_gene' if n else 'overall')
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+def test_compute_mean_target_removal_as_function(log_prob_coo, fpr, per_gene, cuda):
+ """The target removal computation, very important for the MCKP output"""
+
+ noise_count_posterior_coo = log_prob_coo['coo']
+ noise_offsets = log_prob_coo['offsets']
+ device = 'cuda' if cuda else 'cpu'
+
+ print('log prob posterior coo')
+ print(noise_count_posterior_coo)
+
+ index_converter = IndexConverter(total_n_cells=log_prob_coo['coo'].shape[0],
+ total_n_genes=1)
+ print(index_converter)
+
+ print('raw count matrix')
+ count_matrix = sp.csr_matrix(
+ np.expand_dims(np.array([0, 0, 1, 2, 5]), axis=-1)
+ ) # reflecting filled in log_prob values
+ print(count_matrix)
+
+ n_cells = log_prob_coo['coo'].shape[0] # hard coded from the log_prob_coo
+
+ target_fun = compute_mean_target_removal_as_function(
+ noise_count_posterior_coo=noise_count_posterior_coo,
+ noise_offsets=noise_offsets,
+ index_converter=index_converter,
+ raw_count_csr_for_cells=count_matrix,
+ n_cells=n_cells,
+ device=device,
+ per_gene=per_gene,
+ )
+
+ target = (target_fun(fpr) * n_cells).item()
+ print(f'\nwith fpr={fpr:.2f}, target is: {target:.1g}')
+
+ assert target >= 1, 'There is one noise count guaranteed from this test posterior'
+ if fpr == 1:
+ torch.testing.assert_close(target, float(count_matrix.sum()))
+
+ # assert False
+ # TODO: this has not been tested out
+
+
+@pytest.mark.parametrize('blank_noise_offsets', [False, True], ids=['', 'no_noise_offsets'])
+def test_save_and_load(tmpdir_factory, blank_noise_offsets):
+ """Test that a round trip through save and load gives the same thing"""
+
+ tmp_dir = tmpdir_factory.mktemp('posterior')
+ filename = tmp_dir.join('posterior.h5')
+
+ m = 1000
+ n = 20
+
+ posterior = Posterior(dataset_obj=None, vi_model=None) # blank
+
+ posterior_coo = sp.random(m, n, density=0.1, format='coo', dtype=float)
+ posterior_coo2 = sp.random(m, n, density=0.08, format='coo', dtype=float)
+ if blank_noise_offsets:
+ noise_offsets = {}
+ else:
+ noise_offsets = dict(zip(np.random.randint(low=0, high=(m - 1), size=10),
+ np.random.randint(low=1, high=5, size=10)))
+ kwargs = {'a': 'b', 'c': 1}
+ kwargs2 = {'a': 'method', 'c': 1}
+
+ # jam in fake values
+ posterior._noise_count_posterior_coo = posterior_coo
+ posterior._noise_count_posterior_coo_offsets = noise_offsets
+ posterior._noise_count_posterior_kwargs = kwargs
+ posterior._noise_count_regularized_posterior_coo = posterior_coo2
+ posterior._noise_count_regularized_posterior_kwargs = kwargs2
+ posterior._latents = {'p': np.random.randn(100), 'd': np.random.randn(100)}
+ posterior.index_converter = IndexConverter(total_n_cells=1000, total_n_genes=1000)
+
+ # save
+ posterior.save(file=str(filename))
+
+ # load
+ posterior2 = Posterior(dataset_obj=None, vi_model=None) # blank
+ posterior2.load(file=str(filename))
+
+ # check
+ for attr in ['_noise_count_posterior_coo', '_noise_count_posterior_coo_offsets',
+ '_noise_count_posterior_kwargs', '_noise_count_regularized_posterior_coo',
+ '_noise_count_regularized_posterior_kwargs', '_latents']:
+ val1 = getattr(posterior, attr)
+ val2 = getattr(posterior2, attr)
+ print(f'{attr} ===================')
+ print('saved:')
+ print(val1)
+ print('loaded')
+ print(val2)
+ err_msg = f'Posterior attribute {attr} not preserved after saving and loading'
+ if type(val1) == sp.coo_matrix:
+ assert sparse_matrix_equal(val1, val2), err_msg
+ elif type(val1) == np.ndarray:
+ np.testing.assert_equal(val1, val2)
+ elif (type(val1) == dict) and (val1 != {}) and (type(list(val1.values())[0]) == np.ndarray):
+ for k in val1.keys():
+ np.testing.assert_equal(val1[k], val2[k])
+ else:
+ assert val1 == val2, err_msg
diff --git a/cellbender/remove_background/tests/test_sparse_utils.py b/cellbender/remove_background/tests/test_sparse_utils.py
new file mode 100644
index 00000000..01230da8
--- /dev/null
+++ b/cellbender/remove_background/tests/test_sparse_utils.py
@@ -0,0 +1,203 @@
+import pytest
+import scipy.sparse as sp
+import numpy as np
+import torch
+
+from cellbender.remove_background.sparse_utils import todense_fill, \
+ csr_set_rows_to_zero, dense_to_sparse_op_torch, log_prob_sparse_to_dense, \
+ overwrite_matrix_with_columns_from_another
+from cellbender.remove_background.data.dataprep import DataLoader
+from .conftest import sparse_matrix_equal
+
+
+USE_CUDA = torch.cuda.is_available()
+
+
+@pytest.mark.parametrize('val', [0, 1, np.nan, np.inf, -np.inf])
+def test_todense_fill(val):
+ """Test densification of scipy sparse COO matrix with arbitrary fill value"""
+
+ mat = np.array(
+ [[0, 0, 0, 0],
+ [1, 0, 0, 0],
+ [0, 1, 2, 0],
+ [1, 2, 3, 4]],
+ dtype=float,
+ )
+ coo = sp.coo_matrix(mat)
+
+ print('original')
+ print(mat)
+ print('sparse version')
+ print(coo)
+
+ print(f'densified using {val}')
+ dense = todense_fill(coo=coo, fill_value=val)
+ print(dense)
+ brute_force = mat.copy()
+ brute_force[brute_force == 0] = val
+ np.testing.assert_array_equal(brute_force, dense)
+
+
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+def test_dense_to_sparse_op_torch(simulated_dataset, cuda):
+ """test infer.py BasePosterior.dense_to_sparse_op_torch()"""
+
+ d = simulated_dataset
+ data_loader = DataLoader(
+ d['matrix'],
+ empty_drop_dataset=None,
+ batch_size=5,
+ fraction_empties=0.,
+ shuffle=False,
+ use_cuda=cuda,
+ )
+
+ barcodes = []
+ genes = []
+ counts = []
+ ind = 0
+
+ for data in data_loader:
+ dense_counts = data # just make it the same!
+
+ # Convert to sparse.
+ bcs_i_chunk, genes_i, counts_i = \
+ dense_to_sparse_op_torch(dense_counts)
+
+ # Barcode index in the dataloader.
+ bcs_i = bcs_i_chunk + ind
+
+ # Obtain the real barcode index after unsorting the dataloader.
+ bcs_i = data_loader.unsort_inds(bcs_i)
+
+ # Add sparse matrix values to lists.
+ barcodes.append(bcs_i)
+ genes.append(genes_i)
+ counts.append(counts_i)
+
+ # Increment barcode index counter.
+ ind += data.shape[0] # Same as data_loader.batch_size
+
+ # Convert the lists to numpy arrays.
+ counts = np.concatenate(counts).astype(np.uint32)
+ barcodes = np.concatenate(barcodes).astype(np.uint32)
+ genes = np.concatenate(genes).astype(np.uint32) # uint16 is too small!
+
+ # Put the counts into a sparse csc_matrix.
+ out = sp.csc_matrix((counts, (barcodes, genes)),
+ shape=d['matrix'].shape)
+
+ assert sparse_matrix_equal(out, d['matrix'])
+
+
+def test_log_prob_sparse_to_dense():
+ """Test densification of log prob sparse posteriors filling with -np.inf"""
+
+ data = np.array([0, -1, -2, 0, -4, -2, 0])
+ row = np.array([0, 0, 2, 2, 3, 3, 4])
+ col = np.array([0, 1, 0, 1, 0, 1, 0])
+ coo = sp.coo_matrix((data, (row, col)), shape=[5, 2])
+ print(coo)
+ print('scipy .todense()')
+ print(coo.todense())
+ print('log prob densification')
+ mat = log_prob_sparse_to_dense(coo=coo)
+ print(mat)
+ print('nan densification')
+ mat_nan = todense_fill(coo=coo, fill_value=np.nan)
+ print(mat_nan)
+ print('8 densification')
+ mat_8 = todense_fill(coo=coo, fill_value=8.)
+ print(mat_8)
+
+ truth = lambda x: np.array(
+ [[ 0., -1.],
+ [ x, x ],
+ [-2., 0.],
+ [-4., -2.],
+ [ 0., x ]],
+ )
+ print('truth with -np.inf')
+ print(truth(-np.inf))
+
+ np.testing.assert_array_equal(mat, truth(-np.inf))
+ np.testing.assert_array_equal(mat_nan, truth(np.nan))
+ np.testing.assert_array_equal(mat_8, truth(8.))
+
+
+@pytest.mark.parametrize('mat1, mat2, col_inds',
+ [(sp.csc_matrix([[1, 2], [3, 4]]),
+ sp.csc_matrix([[0, 0], [0, 0]]),
+ [0]),
+ (sp.csc_matrix([[1, 2], [3, 4], [5, 6]]),
+ sp.csc_matrix([[0, 0], [0, 0], [0, 0]]),
+ [1]),
+ (sp.csc_matrix(np.random.poisson(lam=2., size=(10, 10))),
+ sp.csc_matrix(np.random.poisson(lam=2., size=(10, 10))),
+ [0, 1, 2, 3])])
+def test_overwrite_matrix_with_columns_from_another(mat1: sp.csc_matrix,
+ mat2: sp.csc_matrix,
+ col_inds: np.ndarray):
+ """test overwrite_matrix_with_columns_from_another()"""
+
+ out = overwrite_matrix_with_columns_from_another(mat1=mat1, mat2=mat2, column_inds=col_inds)
+ excluded_col_inds = [i for i in range(mat1.shape[1]) if i not in col_inds]
+
+ print('col_inds')
+ print(col_inds)
+ print('mat1')
+ print(mat1.todense())
+ print('mat2')
+ print(mat2.todense())
+ print('out')
+ print(out.todense())
+
+ print('assertion')
+ print(out[:, excluded_col_inds].todense())
+ print(mat2[:, excluded_col_inds].todense())
+
+ print('assertion')
+ print(out[:, col_inds].todense())
+ print(mat1[:, col_inds].todense())
+
+ # excluded columns should be replaced with new values
+ assert sparse_matrix_equal(out[:, excluded_col_inds], mat2[:, excluded_col_inds])
+
+ # included columns should be left alone
+ assert sparse_matrix_equal(out[:, col_inds], mat1[:, col_inds])
+
+
+@pytest.mark.parametrize('mat, row_inds',
+ [(sp.csc_matrix([[1, 2], [3, 4]]),
+ [0]),
+ (sp.csc_matrix([[1, 2], [3, 4], [5, 6]]),
+ [1]),
+ (sp.csc_matrix(np.random.poisson(lam=2., size=(10, 10))),
+ [0, 1, 2, 3])])
+def test_csr_set_rows_to_zero(mat: sp.csr_matrix, row_inds: np.ndarray):
+ """test csr_set_rows_to_zero()"""
+
+ out = csr_set_rows_to_zero(csr=mat, row_inds=row_inds)
+ other_row_inds = [i for i in range(mat.shape[0]) if i not in row_inds]
+
+ print('row_inds')
+ print(row_inds)
+ print('mat')
+ print(mat.todense())
+ print('out')
+ print(out.todense())
+
+ print('assertion')
+ print(out[other_row_inds, :].todense())
+ print(mat[other_row_inds, :].todense())
+
+ # other rows should be left alone
+ assert sparse_matrix_equal(out[other_row_inds, :], mat[other_row_inds, :])
+
+ # specified rows should be all zero
+ assert out[row_inds, :].sum() == 0
diff --git a/cellbender/remove_background/tests/test_train.py b/cellbender/remove_background/tests/test_train.py
new file mode 100644
index 00000000..66bf7290
--- /dev/null
+++ b/cellbender/remove_background/tests/test_train.py
@@ -0,0 +1,135 @@
+"""Test functionality in train.py"""
+
+import torch
+import pyro.distributions as dist
+import pyro.infer.trace_elbo
+import pytest
+import scipy.sparse as sp
+from pyro.infer.svi import SVI
+import unittest.mock
+
+from cellbender.remove_background.run import get_optimizer
+from cellbender.remove_background.data.dataprep import prep_sparse_data_for_training \
+ as prep_data_for_training
+from cellbender.remove_background.train import train_epoch, evaluate_epoch
+from .conftest import USE_CUDA
+
+
+@pytest.mark.parametrize('cuda',
+ [False,
+ pytest.param(True, marks=pytest.mark.skipif(not USE_CUDA,
+ reason='requires CUDA'))],
+ ids=lambda b: 'cuda' if b else 'cpu')
+@pytest.mark.parametrize('dropped_minibatch', [False, True], ids=['', 'dropped_minibatch'])
+def test_one_cycle_scheduler(dropped_minibatch, cuda):
+
+ # if there is a minibatch so small that it's below consts.SMALLEST_ALLOWED_BATCH
+ # then the minibatch gets skipped. make sure this works with the scheduler.
+
+ pyro.clear_param_store()
+ device = 'cuda' if cuda else 'cpu'
+
+ n_cells = 3580
+ n_empties = 50000
+ n_genes = 100
+ learning_rate = 1e-4
+ batch_size = 512
+ if dropped_minibatch:
+ n_cells += 2
+ epochs = 50
+
+ count_matrix = sp.random(n_cells, n_genes, density=0.01, format='csr')
+ empty_matrix = sp.random(n_empties, n_genes, density=0.0001, format='csr')
+
+ # Set up dataloaders.
+ train_loader, _ = \
+ prep_data_for_training(dataset=count_matrix,
+ empty_drop_dataset=empty_matrix,
+ batch_size=batch_size,
+ training_fraction=1.,
+ fraction_empties=0.3,
+ shuffle=True,
+ use_cuda=cuda)
+
+ print(f'epochs = {epochs}')
+ print(f'len(train_loader), i.e. number of minibatches = {len(train_loader)}')
+ print(f'train_loader.batch_size = {train_loader.batch_size}')
+ print(f'train_loader.cell_batch_size = {train_loader.cell_batch_size}')
+
+ # Set up optimizer.
+ scheduler = get_optimizer(
+ n_batches=len(train_loader),
+ batch_size=train_loader.batch_size,
+ epochs=epochs,
+ learning_rate=learning_rate,
+ constant_learning_rate=False,
+ total_epochs_for_testing_only=None,
+ )
+
+ # Set up SVI dummy.
+ def _model(x):
+ y_mean = pyro.param("y_mean", torch.zeros(1).to(device))
+ pyro.sample("y_obs", dist.Normal(loc=y_mean, scale=1), obs=x.sum())
+
+ def _guide(x):
+ pass
+
+ svi = SVI(model=_model,
+ guide=_guide,
+ optim=scheduler,
+ loss=pyro.infer.trace_elbo.Trace_ELBO())
+
+ # Looking for a RuntimeError raised by the scheduler trying to step too much
+ lr = []
+ for _ in range(epochs):
+ train_epoch(svi=svi, train_loader=train_loader)
+ lr.append(list(svi.optim.optim_objs.values())[0].get_last_lr()[0])
+
+ print('learning rates at each epoch:')
+ print('\n'.join([f'[{i + 1:03d}] {rate:.2e}' for i, rate in enumerate(lr)]))
+
+ # Ensure learning rates are numerically correct
+ # scheduler args include 'max_lr': learning_rate * 10
+ max_lr = 10 * learning_rate
+ initial_lr = max_lr / 25
+ final_lr = initial_lr / 1e4
+ # https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.OneCycleLR.html
+ assert lr[len(lr) // 2] > lr[0], 'Cosine LR schedule should have middle > start'
+ assert lr[len(lr) // 2] > lr[-1], 'Cosine LR schedule should have middle > end'
+ torch.testing.assert_close(lr[0], initial_lr + (max_lr - initial_lr) / epochs, rtol=1, atol=1e-7), \
+ 'Starting learning rate for scheduler seems off'
+ torch.testing.assert_close(lr[-1], final_lr, rtol=1, atol=1e-7), \
+ 'Final learning rate for scheduler seems off'
+ torch.testing.assert_close(max(lr), max_lr, rtol=1, atol=1e-7), \
+ 'Max learning rate in scheduler seems off'
+
+ # And one more step ought to raise the error
+ with pytest.raises(ValueError,
+ match=r"Tried to step .* times. The specified number "
+ r"of total steps is .*"):
+ svi.optim.step()
+
+
+@pytest.mark.skip
+def test_epoch_elbo_fail_restart():
+ """Trigger failure and ensure a new attempt is made"""
+ pass
+
+
+@pytest.mark.skip
+def test_final_elbo_fail_restart():
+ """Trigger failure and ensure a new attempt is made"""
+ pass
+
+
+@pytest.mark.skip
+def test_restart_matches_scratch():
+ """Ensure that an auto-restart gives same ELBO as a start-from-scratch"""
+ # https://stackoverflow.com/questions/50964786/mock-exception-raised-in-function-using-pytest
+ pass
+
+
+@pytest.mark.skip
+def test_num_training_attempts():
+ """Make sure the number of training attempts matches input arg"""
+ pass
diff --git a/cellbender/remove_background/train.py b/cellbender/remove_background/train.py
index 0f47f18d..8d778082 100644
--- a/cellbender/remove_background/train.py
+++ b/cellbender/remove_background/train.py
@@ -1,35 +1,40 @@
"""Helper functions for training."""
import pyro
-import pyro.distributions as dist
-from pyro.infer import SVI, JitTraceEnum_ELBO, JitTrace_ELBO, \
- TraceEnum_ELBO, Trace_ELBO
-from pyro.optim import ClippedAdam
from pyro.util import ignore_jit_warnings
+from pyro.infer import SVI
+import torch
from cellbender.remove_background.model import RemoveBackgroundPyroModel
-from cellbender.remove_background.vae.decoder import Decoder
-from cellbender.remove_background.vae.encoder \
- import EncodeZ, CompositeEncoder, EncodeNonZLatents
-from cellbender.remove_background.data.dataset import SingleCellRNACountsDataset
-from cellbender.remove_background.data.dataprep import \
- prep_sparse_data_for_training as prep_data_for_training
from cellbender.remove_background.data.dataprep import DataLoader
-from cellbender.remove_background.exceptions import NanException
+from cellbender.remove_background.exceptions import NanException, ElboException
import cellbender.remove_background.consts as consts
+from cellbender.remove_background.checkpoint import save_checkpoint
+from cellbender.monitor import get_hardware_usage
import numpy as np
-import torch
+import argparse
from typing import Tuple, List, Optional
import logging
import time
from datetime import datetime
+import sys
+
+
+logger = logging.getLogger('cellbender')
+
+
+def is_scheduler(optim):
+ """Currently is_scheduler is on a pyro dev branch"""
+ if hasattr(optim, 'is_scheduler'):
+ return optim.is_scheduler()
+ else:
+ return type(optim) == pyro.optim.lr_scheduler.PyroLRScheduler
def train_epoch(svi: SVI,
- train_loader: DataLoader,
- epoch: Optional[int] = None) -> float:
+ train_loader: DataLoader) -> float:
"""Train a single epoch.
Args:
@@ -53,12 +58,22 @@ def train_epoch(svi: SVI,
# Perform gradient descent step and accumulate loss.
epoch_loss += svi.step(x_cell_batch)
- svi.optim.step(epoch=epoch) # for LR scheduling
normalizer_train += x_cell_batch.size(0)
+ if is_scheduler(svi.optim):
+ svi.optim.step() # for LR scheduling
+
# Return epoch loss.
total_epoch_loss_train = epoch_loss / normalizer_train
+ if is_scheduler(svi.optim):
+ try:
+ logger.debug(f'Learning rate scheduler: LR = '
+ f'{list(svi.optim.optim_objs.values())[0].get_last_lr()[0]:.2e}')
+ except IndexError:
+ logger.debug('No values being optimized')
+ pass
+
return total_epoch_loss_train
@@ -98,232 +113,185 @@ def evaluate_epoch(svi: pyro.infer.SVI,
@ignore_jit_warnings()
def run_training(model: RemoveBackgroundPyroModel,
+ args: argparse.Namespace,
svi: pyro.infer.SVI,
train_loader: DataLoader,
test_loader: DataLoader,
epochs: int,
+ output_filename: str,
test_freq: int = 10,
final_elbo_fail_fraction: float = None,
- epoch_elbo_fail_fraction: float = None) -> Tuple[List[float],
- List[float], bool]:
+ epoch_elbo_fail_fraction: float = None,
+ ckpt_tarball_name: str = consts.CHECKPOINT_FILE_NAME,
+ checkpoint_freq: int = 10) -> Tuple[List[float], List[float]]:
"""Run an entire course of training, evaluating on a tests set periodically.
Args:
model: The model, here in order to store train and tests loss.
+ args: Parsed arguments, which get saved to checkpoints.
svi: The pyro object used for stochastic variational inference.
train_loader: Dataloader for training set.
test_loader: Dataloader for tests set.
epochs: Number of epochs to run training.
+ output_filename: User-specified output file, used to construct
+ checkpoint filenames.
test_freq: Test set loss is calculated every test_freq epochs of
training.
- final_elbo_fail_fraction: fail if final test ELBO >= best ELBO * (1+this value)
- epoch_elbo_fail_fraction: fail if current test ELBO >= previous ELBO * (1+this value)
+ final_elbo_fail_fraction: Fail if final test ELBO >=
+ best ELBO * (1 + this value)
+ epoch_elbo_fail_fraction: Fail if current test ELBO >=
+ previous ELBO * (1 + this value)
+ ckpt_tarball_name: Name of saved tarball for checkpoint.
+ checkpoint_freq: Checkpoint after this many minutes
Returns:
- Tuple(
- list of training ELBO for each epoch,
- list of test ELBO for each epoch for which testing is done,
- boolean: False indicates training failed
- )
- """
+ total_epoch_loss_train: The loss for this epoch of training, which
+ is -ELBO, normalized by the number of items in the training set.
- logging.info("Running inference...")
+ """
# Initialize train and tests ELBO with empty lists.
train_elbo = []
test_elbo = []
- succeeded = True
+ lr = []
+ epoch_checkpoint_freq = 1000 # a large number... it will be recalculated
+
# Run training loop. Use try to allow for keyboard interrupt.
try:
- for epoch in range(1, epochs + 1):
+
+ start_epoch = (1 if (model is None) or (len(model.loss['train']['epoch']) == 0)
+ else model.loss['train']['epoch'][-1] + 1)
+
+ for epoch in range(start_epoch, epochs + 1):
+
+ # In debug mode, log hardware load every epoch.
+ if args.debug:
+ # Don't spend time pinging usage stats if we will not use the log.
+ # TODO: use multiprocessing to sample these stats DURING training...
+ logger.debug('\n' + get_hardware_usage(use_cuda=model.use_cuda))
# Display duration of an epoch (use 2 to avoid initializations).
- if epoch == 2:
+ if epoch == start_epoch + 1:
t = time.time()
+ model.train()
total_epoch_loss_train = train_epoch(svi, train_loader)
train_elbo.append(-total_epoch_loss_train)
- model.loss['train']['epoch'].append(epoch)
- model.loss['train']['elbo'].append(-total_epoch_loss_train)
-
- if epoch == 2:
- logging.info("[epoch %03d] average training loss: %.4f (%.1f seconds per epoch)"
- % (epoch, total_epoch_loss_train, time.time() - t))
+ try:
+ last_learning_rate = list(svi.optim.optim_objs.values())[0].get_last_lr()[0]
+ except AttributeError:
+ # not a scheduler
+ last_learning_rate = args.learning_rate
+ lr.append(last_learning_rate)
+
+ if model is not None:
+ model.loss['train']['epoch'].append(epoch)
+ model.loss['train']['elbo'].append(-total_epoch_loss_train)
+ model.loss['learning_rate']['epoch'].append(epoch)
+ model.loss['learning_rate']['value'].append(last_learning_rate)
+
+ if epoch == start_epoch + 1:
+ time_per_epoch = time.time() - t
+ logger.info("[epoch %03d] average training loss: %.4f (%.1f seconds per epoch)"
+ % (epoch, total_epoch_loss_train, time_per_epoch))
+ epoch_checkpoint_freq = int(np.ceil(60 * checkpoint_freq / time_per_epoch))
+ if (epoch_checkpoint_freq > 0) and (epoch_checkpoint_freq < epochs):
+ logger.info(f"Will checkpoint every {epoch_checkpoint_freq} epochs")
+ elif epoch_checkpoint_freq >= epochs:
+ logger.info(f"Will not checkpoint due to projected run "
+ f"completion in under {checkpoint_freq} min")
else:
- logging.info("[epoch %03d] average training loss: %.4f"
- % (epoch, total_epoch_loss_train))
+ logger.info("[epoch %03d] average training loss: %.4f"
+ % (epoch, total_epoch_loss_train))
# If there is no test data (training_fraction == 1.), skip test.
- if len(test_loader) == 0:
- continue
-
- # Every test_freq epochs, evaluate tests loss.
- if epoch % test_freq == 0:
- total_epoch_loss_test = evaluate_epoch(svi, test_loader)
- test_elbo.append(-total_epoch_loss_test)
- model.loss['test']['epoch'].append(epoch)
- model.loss['test']['elbo'].append(-total_epoch_loss_test)
- logging.info("[epoch %03d] average test loss: %.4f"
- % (epoch, total_epoch_loss_test))
- if epoch_elbo_fail_fraction is not None and len(test_elbo) > 1 and \
- test_elbo[-1] < test_elbo[-2] and \
- (test_elbo[-2] - test_elbo[-1])/(test_elbo[-2] - train_elbo[0]) > epoch_elbo_fail_fraction:
- logging.info(
- "Training failed because this test loss (%.4f) exceeds previous test loss(%.4f) by >= %.2f%%, "
- "relative to initial train loss %.4f" ,
- test_elbo[-1], test_elbo[-2], 100*epoch_elbo_fail_fraction, train_elbo[0])
- succeeded = False
- break
-
- logging.info("Inference procedure complete.")
-
- if succeeded and final_elbo_fail_fraction is not None and len(test_elbo) > 1:
+ if (test_loader is not None) and (len(test_loader) > 0):
+
+ # Every test_freq epochs, evaluate tests loss.
+ if epoch % test_freq == 0:
+ model.eval()
+ total_epoch_loss_test = evaluate_epoch(svi, test_loader)
+ test_elbo.append(-total_epoch_loss_test)
+ model.loss['test']['epoch'].append(epoch)
+ model.loss['test']['elbo'].append(-total_epoch_loss_test)
+ logger.info("[epoch %03d] average test loss: %.4f"
+ % (epoch, total_epoch_loss_test))
+
+ # Check whether test ELBO has spiked beyond specified conditions.
+ if (epoch_elbo_fail_fraction is not None) and (len(test_elbo) > 2):
+ current_diff = max(0., test_elbo[-2] - test_elbo[-1])
+ overall_diff = np.abs(test_elbo[-2] - test_elbo[0])
+ fractional_spike = current_diff / overall_diff
+ if fractional_spike > epoch_elbo_fail_fraction:
+ raise ElboException(
+ f'Training failed because test loss moved {current_diff:.2f} '
+ f'in the wrong direction, and that is {fractional_spike:.2f} '
+ f'of the total test ELBO change, > '
+ f'specified epoch_elbo_fail_fraction {epoch_elbo_fail_fraction:.2f}'
+ )
+
+ # Checkpoint throughout and after final epoch.
+ if ((ckpt_tarball_name != 'none')
+ and (((checkpoint_freq > 0) and (epoch % epoch_checkpoint_freq == 0))
+ or (epoch == epochs))): # checkpoint at final epoch
+ save_checkpoint(filebase=output_filename,
+ tarball_name=ckpt_tarball_name,
+ args=args,
+ model_obj=model,
+ scheduler=svi.optim,
+ train_loader=train_loader,
+ test_loader=test_loader)
+
+ # Check on the final test ELBO to see if it meets criteria.
+ if final_elbo_fail_fraction is not None:
best_test_elbo = max(test_elbo)
- if test_elbo[-1] < best_test_elbo and \
- (best_test_elbo - test_elbo[-1])/(best_test_elbo - train_elbo[0]) > final_elbo_fail_fraction:
- logging.info(
- "Training failed because final test loss (%.4f) exceeds "
- "best test loss(%.4f) by >= %.2f%%, relative to initial train loss %.4f",
- test_elbo[-1], best_test_elbo, 100*final_elbo_fail_fraction, train_elbo[0])
- succeeded = False
+ if test_elbo[-1] < best_test_elbo:
+ final_best_diff = best_test_elbo - test_elbo[-1]
+ initial_best_diff = best_test_elbo - test_elbo[0]
+ if (final_best_diff / initial_best_diff) > final_elbo_fail_fraction:
+ raise ElboException(
+ f'Training failed because final test loss {test_elbo[-1]:.2f} '
+ f'is not sufficiently close to best test loss {best_test_elbo:.2f}, '
+ f'compared to the initial test loss {test_elbo[0]:.2f}. '
+ f'Fractional difference is {final_best_diff / initial_best_diff:.2f}, '
+ f'which is > specified final_elbo_fail_fraction {final_elbo_fail_fraction:.2f}'
+ )
# Exception allows program to continue after ending inference prematurely.
except KeyboardInterrupt:
-
- logging.info("Inference procedure stopped by keyboard interrupt.")
+ logger.info("Inference procedure stopped by keyboard interrupt... "
+ "will save a checkpoint.")
+ save_checkpoint(filebase=output_filename,
+ tarball_name=ckpt_tarball_name,
+ args=args,
+ model_obj=model,
+ scheduler=svi.optim,
+ train_loader=train_loader,
+ test_loader=test_loader)
+
+ except ElboException as e:
+ logger.info(e.message)
+ raise e # re-raise the exception to pass it back to caller function
# Exception allows program to produce output when terminated by a NaN.
except NanException as nan:
-
print(nan.message)
- logging.info(f"Inference procedure terminated early due to a NaN value in: {nan.param}\n\n"
- f"The suggested fix is to reduce the learning rate.\n\n")
- succeeded = False
-
- if succeeded:
- logging.info("Training succeeded")
- logging.info(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
+ logger.info(f"Inference procedure terminated early due to a NaN value in: {nan.param}\n\n"
+ f"The suggested fix is to reduce the learning rate by a factor of two.\n\n")
+ sys.exit(1)
- return train_elbo, test_elbo, succeeded
+ logger.info(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
+ # Check final ELBO meets conditions.
+ if (final_elbo_fail_fraction is not None) and (len(test_elbo) > 1):
+ best_test_elbo = max(test_elbo)
+ if -test_elbo[-1] >= -best_test_elbo * (1 + final_elbo_fail_fraction):
+ raise ElboException(f'Training failed because final test loss ({-test_elbo[-1]:.4f}) '
+ f'exceeds best test loss ({-best_test_elbo:.4f}) by >= '
+ f'{100 * final_elbo_fail_fraction:.1f}%')
-def run_inference(dataset_obj: SingleCellRNACountsDataset,
- args) -> RemoveBackgroundPyroModel:
- """Run a full inference procedure, training a latent variable model.
-
- Args:
- dataset_obj: Input data in the form of a SingleCellRNACountsDataset
- object.
- args: Input command line parsed arguments.
+ # Free up all the GPU memory we can once training is complete.
+ torch.cuda.empty_cache()
- Returns:
- model: cellbender.model.RemoveBackgroundPyroModel that has had
- inference run.
-
- """
-
- # Get the trimmed count matrix (transformed if called for).
- count_matrix = dataset_obj.get_count_matrix()
-
- # Configure pyro options (skip validations to improve speed).
- pyro.enable_validation(False)
- pyro.distributions.enable_validation(False)
-
- # Load the dataset into DataLoaders.
- frac = args.training_fraction # Fraction of barcodes to use for training
- batch_size = int(min(300, frac * dataset_obj.analyzed_barcode_inds.size / 2))
-
- for attempt in range(args.num_training_tries):
- # set seed for every attempt for reproducibility
- pyro.set_rng_seed(0)
- pyro.clear_param_store()
-
- # Set up the variational autoencoder:
-
- # Encoder.
- encoder_z = EncodeZ(input_dim=count_matrix.shape[1],
- hidden_dims=args.z_hidden_dims,
- output_dim=args.z_dim,
- input_transform='normalize')
-
- encoder_other = EncodeNonZLatents(n_genes=count_matrix.shape[1],
- z_dim=args.z_dim,
- hidden_dims=consts.ENC_HIDDEN_DIMS,
- log_count_crossover=dataset_obj.priors['log_counts_crossover'],
- prior_log_cell_counts=np.log1p(dataset_obj.priors['cell_counts']),
- input_transform='normalize')
-
- encoder = CompositeEncoder({'z': encoder_z,
- 'other': encoder_other})
-
- # Decoder.
- decoder = Decoder(input_dim=args.z_dim,
- hidden_dims=args.z_hidden_dims[::-1],
- output_dim=count_matrix.shape[1])
-
- # Set up the pyro model for variational inference.
- model = RemoveBackgroundPyroModel(model_type=args.model,
- encoder=encoder,
- decoder=decoder,
- dataset_obj=dataset_obj,
- use_cuda=args.use_cuda)
-
- train_loader, test_loader = \
- prep_data_for_training(dataset=count_matrix,
- empty_drop_dataset=
- dataset_obj.get_count_matrix_empties(),
- random_state=dataset_obj.random,
- batch_size=batch_size,
- training_fraction=frac,
- fraction_empties=args.fraction_empties,
- shuffle=True,
- use_cuda=args.use_cuda)
-
- # Set up the optimizer.
- optimizer = pyro.optim.clipped_adam.ClippedAdam
- optimizer_args = {'lr': args.learning_rate, 'clip_norm': 10.}
-
- # Set up a learning rate scheduler.
- minibatches_per_epoch = int(np.ceil(len(train_loader) / train_loader.batch_size).item())
- scheduler_args = {'optimizer': optimizer,
- 'max_lr': args.learning_rate * 10,
- 'steps_per_epoch': minibatches_per_epoch,
- 'epochs': args.epochs,
- 'optim_args': optimizer_args}
- scheduler = pyro.optim.OneCycleLR(scheduler_args)
-
- # Determine the loss function.
- if args.use_jit:
-
- # Call guide() once as a warm-up.
- model.guide(torch.zeros([10, dataset_obj.analyzed_gene_inds.size]).to(model.device))
-
- if args.model == "simple":
- loss_function = JitTrace_ELBO()
- else:
- loss_function = JitTraceEnum_ELBO(max_plate_nesting=1,
- strict_enumeration_warning=False)
- else:
-
- if args.model == "simple":
- loss_function = Trace_ELBO()
- else:
- loss_function = TraceEnum_ELBO(max_plate_nesting=1)
-
- # Set up the inference process.
- svi = SVI(model.model, model.guide, scheduler,
- loss=loss_function)
-
- # Run training.
- train_elbo, test_elbo, succeeded = run_training(model, svi, train_loader, test_loader,
- epochs=args.epochs, test_freq=5,
- final_elbo_fail_fraction=args.final_elbo_fail_fraction,
- epoch_elbo_fail_fraction=args.epoch_elbo_fail_fraction)
- if succeeded or attempt+1 >= args.num_training_tries:
- break
- else:
- args.learning_rate = args.learning_rate * args.learning_rate_retry_mult
- logging.info("Learning failed. Retrying with learning-rate %.8f" % args.learning_rate)
-
- return model
+ return train_elbo, test_elbo
diff --git a/cellbender/remove_background/vae/base.py b/cellbender/remove_background/vae/base.py
new file mode 100644
index 00000000..2e8b93c8
--- /dev/null
+++ b/cellbender/remove_background/vae/base.py
@@ -0,0 +1,164 @@
+"""Base neural network architectures, for convenience"""
+
+import torch
+from typing import Optional, List
+
+
+class Exp(torch.nn.Module):
+ """Exponential activation function as a torch module"""
+
+ def __init__(self, eps: float = 1e-5):
+ """Exponential activation function with numerical stabilization, useful
+ for outputs that must be > 0
+
+ NOTE: output = torch.exp(input) + eps
+
+ Parameters
+ ----------
+ eps: Numerical stability additive constant.
+ """
+ super().__init__()
+ self.eps = eps
+
+ def forward(self, x):
+ return torch.exp(x) + self.eps
+
+ def __repr__(self):
+ return f"torch.exp() + {self.eps}"
+
+
+class FullyConnectedLayer(torch.nn.Module):
+ """Neural network unit made of a fully connected linear layer, but
+ customizable including shapes, activations, batch norm, layer norm, and
+ dropout.
+
+ Parameters
+ ----------
+ input_dim: Number of features for input
+ output_dim: Number of features for output
+ activation: Activation function to be applied to each hidden layer
+ (default :py:class:`torch.nn.ReLU`)
+ use_batch_norm: True to apply batch normalization using
+ :py:class:`torch.nn.BatchNorm1d` with ``momentum=0.01``, ``eps=0.001``
+ (default False)
+ use_layer_norm: True to apply layer normalization (after optional batch
+ normalization) using :py:class:`torch.nn.LayerNorm` with
+ ``elementwise_affine=False`` (default False)
+ dropout_rate: Dropout rate to use in :py:class:`torch.nn.Dropout` before
+ linear layer
+
+ """
+
+ def __init__(
+ self,
+ input_dim: int,
+ output_dim: int,
+ activation: torch.nn.Module = torch.nn.ReLU,
+ use_batch_norm: bool = False,
+ use_layer_norm: bool = False,
+ dropout_rate: Optional[float] = None,
+ ):
+ super().__init__()
+ self.input_dim = input_dim
+ self.output_dim = output_dim
+
+ # set up layers as a list of Linear modules with appropriate extras
+ modules = []
+ if dropout_rate is not None:
+ modules.append(torch.nn.Dropout(p=dropout_rate))
+ modules.append(torch.nn.Linear(in_features=input_dim, out_features=output_dim))
+ if use_batch_norm:
+ modules.append(
+ torch.nn.BatchNorm1d(num_features=output_dim, momentum=0.01, eps=0.001)
+ )
+ if use_layer_norm:
+ modules.append(
+ torch.nn.LayerNorm(
+ normalized_shape=output_dim, elementwise_affine=False
+ )
+ )
+ if activation is not None:
+ modules.append(activation)
+
+ # concatenate Linear layers using Sequential
+ self.layer = torch.nn.Sequential(*modules)
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ return self.layer(x)
+
+
+class FullyConnectedNetwork(torch.nn.Module):
+ """Neural network made of fully connected linear layers,
+ :py:class:`FullyConnectedLayer`. Architecture is customizable including
+ shapes, activations, batch norm, layer norm, and dropout.
+
+ Parameters
+ ----------
+ input_dim: Number of features for input
+ hidden_dims: List of hidden layer sizes, can be empty list []
+ output_dim: Number of features for output
+ hidden_activation: Activation function to be applied to each hidden layer
+ (default :py:class:`torch.nn.ReLU`)
+ output_activation: Activation function to be applied to output (default None)
+ use_batch_norm: True to apply batch normalization using
+ :py:class:`torch.nn.BatchNorm1d` with ``momentum=0.01``, ``eps=0.001``
+ (default False)
+ use_layer_norm: True to apply layer normalization (after optional batch
+ normalization) using :py:class:`torch.nn.LayerNorm` with
+ ``elementwise_affine=False`` (default False)
+ norm_output: True to apply normalization to output layer before output
+ activation (default False)
+ dropout_rate: Dropout rate to use in :py:class:`torch.nn.Dropout` for each
+ hidden layer (applied before each layer)
+ dropout_input: True to apply dropout before first layer (default False)
+ """
+
+ def __init__(
+ self,
+ input_dim: int,
+ hidden_dims: List[int],
+ output_dim: int,
+ hidden_activation: torch.nn.Module = torch.nn.ReLU(),
+ output_activation: Optional[torch.nn.Module] = None,
+ use_batch_norm: bool = False,
+ use_layer_norm: bool = False,
+ norm_output: bool = False,
+ dropout_rate: Optional[float] = None,
+ dropout_input: bool = False,
+ ):
+ super().__init__()
+
+ if use_layer_norm and use_batch_norm:
+ raise UserWarning(
+ "You are trying to use both batch norm and layer "
+ "norm. That's probably too much norm."
+ )
+
+ # set up layers as a list of Linear modules with appropriate extras
+ dim_ins_and_outs = zip([input_dim] + hidden_dims, hidden_dims + [output_dim])
+ n_layers = 1 + len(hidden_dims)
+ layers = [
+ FullyConnectedLayer(
+ input_dim=i,
+ output_dim=j,
+ activation=hidden_activation
+ if (layer < n_layers - 1)
+ else output_activation,
+ use_batch_norm=use_batch_norm
+ if ((layer < n_layers - 1) or norm_output)
+ else False,
+ use_layer_norm=use_layer_norm
+ if ((layer < n_layers - 1) or norm_output)
+ else False,
+ dropout_rate=None
+ if ((layer == 0) and not dropout_input)
+ else dropout_rate,
+ )
+ for layer, (i, j) in enumerate(dim_ins_and_outs)
+ ]
+
+ # concatenate Linear layers using Sequential
+ self.network = torch.nn.Sequential(*layers)
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ return self.network(x)
diff --git a/cellbender/remove_background/vae/decoder.py b/cellbender/remove_background/vae/decoder.py
index 56b0a58a..8d56eef2 100644
--- a/cellbender/remove_background/vae/decoder.py
+++ b/cellbender/remove_background/vae/decoder.py
@@ -1,9 +1,8 @@
import torch
-import torch.nn as nn
-from typing import List
+from cellbender.remove_background.vae.base import FullyConnectedNetwork
-class Decoder(nn.Module):
+class Decoder(FullyConnectedNetwork):
"""Decoder module transforms latent representation into gene expression.
The number of input units is the dimension of the latent space and the
@@ -20,6 +19,8 @@ class Decoder(nn.Module):
expression will be embedded.
hidden_dims: Size of each of the hidden layers.
output_dim: Number of genes. The size of the output of this decoder.
+ use_layernorm: True to use LayerNorm after each hidden layer is
+ computed, before the activation
log_output: Whether or not the output is in log space.
Attributes:
@@ -36,37 +37,10 @@ class Decoder(nn.Module):
"""
- def __init__(self,
- input_dim: int,
- hidden_dims: List[int],
- output_dim: int,
- log_output: bool = False):
- super(Decoder, self).__init__()
+ def __init__(self, input_dim: int, **kwargs):
+ super().__init__(input_dim=input_dim, **kwargs)
self.input_dim = input_dim
- self.log_output = log_output
-
- # Set up the linear transformations used in fully-connected layers.
- self.linears = nn.ModuleList([nn.Linear(input_dim, hidden_dims[0])])
- for i in range(1, len(hidden_dims)): # Second hidden layer onward
- self.linears.append(nn.Linear(hidden_dims[i-1], hidden_dims[i]))
- self.outlinear = nn.Linear(hidden_dims[-1], output_dim)
-
- # Set up the non-linear activations.
- self.softplus = nn.Softplus()
- if log_output:
- self.softmax = nn.LogSoftmax(dim=-1)
- else:
- self.softmax = nn.Softmax(dim=-1)
+ self.softmax = torch.nn.Softmax(dim=-1)
def forward(self, z: torch.Tensor) -> torch.Tensor:
- # Define the forward computation to go from latent z to gene expression.
-
- # Compute the hidden layers.
- hidden = self.softplus(self.linears[0](z))
- for i in range(1, len(self.linears)): # Second hidden layer onward
- hidden = self.softplus(self.linears[i](hidden))
-
- # Compute the output, which is on a simplex.
- gene_exp = self.softmax(self.outlinear(hidden))
-
- return gene_exp
+ return self.softmax(self.network(z))
diff --git a/cellbender/remove_background/vae/encoder.py b/cellbender/remove_background/vae/encoder.py
index 6862cdf2..909ada9b 100644
--- a/cellbender/remove_background/vae/encoder.py
+++ b/cellbender/remove_background/vae/encoder.py
@@ -1,5 +1,9 @@
import torch
import torch.nn as nn
+import pyro
+import numpy as np
+from cellbender.remove_background.vae.base import FullyConnectedNetwork
+from cellbender.remove_background import consts
from typing import Dict, List, Optional
@@ -23,6 +27,9 @@ def __init__(self, module_dict):
super(CompositeEncoder, self).__init__(module_dict)
self.module_dict = module_dict
+ def __call__(self, **kwargs):
+ return self.forward(**kwargs)
+
def forward(self, **kwargs) \
-> Dict[str, torch.Tensor]:
@@ -47,7 +54,7 @@ def forward(self, **kwargs) \
return out
-class EncodeZ(nn.Module):
+class EncodeZ(FullyConnectedNetwork):
"""Encoder module transforms gene expression into latent representation.
The number of input units is the total number of genes and the number of
@@ -85,35 +92,34 @@ class EncodeZ(nn.Module):
"""
- def __init__(self, input_dim: int, hidden_dims: List[int], output_dim: int,
- input_transform: str = None):
- super(EncodeZ, self).__init__()
+ def __init__(self,
+ input_dim: int,
+ hidden_dims: List[int],
+ output_dim: int,
+ input_transform: str = None,
+ **kwargs):
+ assert len(hidden_dims) > 0, 'EncodeZ needs to have at least one hidden layer'
+ super(EncodeZ, self).__init__(input_dim=input_dim,
+ hidden_dims=hidden_dims[:-1],
+ output_dim=hidden_dims[-1],
+ hidden_activation=nn.Softplus(),
+ output_activation=nn.Softplus(),
+ norm_output=True,
+ **kwargs)
+ self.transform = input_transform
self.input_dim = input_dim
self.output_dim = output_dim
- self.transform = input_transform
-
- # Set up the linear transformations used in fully-connected layers.
- self.linears = nn.ModuleList([nn.Linear(input_dim, hidden_dims[0])])
- for i in range(1, len(hidden_dims)): # Second hidden layer onward
- self.linears.append(nn.Linear(hidden_dims[i-1], hidden_dims[i]))
self.loc_out = nn.Linear(hidden_dims[-1], output_dim)
self.sig_out = nn.Linear(hidden_dims[-1], output_dim)
- # Set up the non-linear activations.
- self.softplus = nn.Softplus()
-
- def forward(self, x: torch.Tensor, **kwargs) -> Dict[str, torch.Tensor]:
- # Define the forward computation to go from gene expression to latent
- # representation.
+ def forward(self, x: torch.Tensor, **kwargs) -> torch.Tensor:
# Transform input.
x = x.reshape(-1, self.input_dim)
- x = transform_input(x, self.transform)
+ x_ = transform_input(x, self.transform)
- # Compute the hidden layers.
- hidden = self.softplus(self.linears[0](x))
- for i in range(1, len(self.linears)): # Second hidden layer onward
- hidden = self.softplus(self.linears[i](hidden))
+ # Obtain last hidden layer.
+ hidden = self.network(x_)
# Compute the outputs: loc is any real number, scale must be positive.
loc = self.loc_out(hidden)
@@ -122,6 +128,10 @@ def forward(self, x: torch.Tensor, **kwargs) -> Dict[str, torch.Tensor]:
return {'loc': loc.squeeze(), 'scale': scale.squeeze()}
+def _poisson_log_prob(lam, value):
+ return (lam.log() * value) - lam - (value + 1).lgamma()
+
+
class EncodeNonZLatents(nn.Module):
"""Encoder module that transforms data into all latents except z.
@@ -175,66 +185,59 @@ class EncodeNonZLatents(nn.Module):
def __init__(self,
n_genes: int,
z_dim: int,
- hidden_dims: List[int],
log_count_crossover: float, # prior on log counts of smallest cell
- prior_log_cell_counts: int, # prior on counts per cell
+ prior_log_cell_counts: float, # prior on counts per cell
+ empty_log_count_threshold: float,
+ prior_logit_cell_prob: float,
input_transform: Optional[str] = None):
super(EncodeNonZLatents, self).__init__()
self.n_genes = n_genes
self.z_dim = z_dim
self.transform = input_transform
- self.output_dim = 3
+ self.output_dim = 1
# Values related to logit cell probability
- self.INITIAL_WEIGHT_FOR_LOG_COUNTS = 2.
- self.P_OUTPUT_SCALE = 1.
+ self.INITIAL_WEIGHT_FOR_LOG_COUNTS = 1.
+ self.P_OUTPUT_SCALE = 5.
self.log_count_crossover = log_count_crossover
+ self.empty_log_count_threshold = empty_log_count_threshold
self.prior_log_cell_counts = prior_log_cell_counts
+ self.prior_logit_cell_prob = prior_logit_cell_prob
# Values related to epsilon
- self.EPS_OUTPUT_SCALE = 0.05 # slows down learning for epsilon
- self.EPS_OUTPUT_MEAN = 1.
-
- # Set up the linear transformations used in fully-connected layers.
-
- self.linears = nn.ModuleList([nn.Linear(3 + self.n_genes,
- hidden_dims[0])])
- for i in range(1, len(hidden_dims)): # Second hidden layer onward
- self.linears.append(nn.Linear(hidden_dims[i-1], hidden_dims[i]))
- self.output = nn.Linear(hidden_dims[-1], self.output_dim)
-
- # Adjust initialization conditions to start with a reasonable output.
- self._weight_init()
+ self.EPS_OUTPUT_SCALE = 0.2 # slows down learning for epsilon
+ self.EPS_OUTPUT_MEAN = 0.6931 # log(e - 1) # softplus(0.6931) = 1.
+
+ # Set up network for inference of p
+ additional_features_p = 4
+ self.layer1 = nn.Linear(additional_features_p + self.n_genes, 512)
+ self.batchnorm1 = nn.BatchNorm1d(num_features=512)
+ self.layer2 = nn.Linear(additional_features_p + 512, 512)
+ self.batchnorm2 = nn.BatchNorm1d(num_features=512)
+ self.layer3 = nn.Linear(additional_features_p + 512, 1)
+
+ # Set up network for inference of epsilon
+ additional_features_eps = 4
+ self.eps_network = nn.Sequential(
+ nn.Linear(additional_features_eps + self.n_genes, 512),
+ nn.BatchNorm1d(num_features=512),
+ nn.Softplus(),
+ nn.Linear(512, 512),
+ nn.BatchNorm1d(num_features=512),
+ nn.Softplus(),
+ nn.Linear(512, 1),
+ )
# Set up the non-linear activations.
- self.nonlin = nn.Softplus()
self.softplus = nn.Softplus()
+ self.dropout50 = nn.Dropout1d(p=0.5)
# Set up the initial biases.
self.offset = None
# Set up the initial scaling for values of x.
self.x_scaling = None
-
- # Set up initial values for overlap normalization.
- self.overlap_mean = None
- self.overlap_std = None
-
- def _weight_init(self):
- """Initialize neural network weights"""
-
- # Initialize p to be a sigmoid function of UMI counts.
- for linear in self.linears:
- with torch.no_grad():
- linear.weight[0][0] = 1.
- with torch.no_grad():
- self.output.weight[0][0] = self.INITIAL_WEIGHT_FOR_LOG_COUNTS
- # Prevent a negative weight from starting something inverted.
- self.output.weight[1][0] = torch.abs(self.output.weight[1][0])
- self.output.weight[2][0] = torch.abs(self.output.weight[2][0])
-
- def _poisson_log_prob(self, lam, value):
- return (lam.log() * value) - lam - (value + 1).lgamma()
+ self.batchnorm0 = nn.BatchNorm1d(num_features=self.n_genes)
def forward(self,
x: torch.Tensor,
@@ -256,44 +259,55 @@ def forward(self,
# Calculate the log of the number of nonzero genes.
log_nnz = (x > 0).sum(dim=-1, keepdim=True).float().log1p()
- # Calculate a similarity between expression and ambient.
+ # Calculate probability that log counts are consistent with d_empty.
if chi_ambient is not None:
- overlap = self._poisson_log_prob(lam=counts * chi_ambient.detach().unsqueeze(0),
- value=x).sum(dim=-1, keepdim=True)
- if self.overlap_mean is None:
- self.overlap_mean = (overlap.max() + overlap.min()) / 2
- self.overlap_std = overlap.max() - overlap.min()
- overlap = (overlap - self.overlap_mean) / self.overlap_std * 5
+ # Gaussian log probability
+ overlap = -0.5 * (
+ torch.clamp(log_sum - pyro.param("d_empty_loc").detach(), min=0.)
+ / 0.1
+ ).pow(2)
else:
overlap = torch.zeros_like(counts)
+ # Calculate a dot product between expression and ambient, for epsilon.
+ if chi_ambient is not None:
+ x_ambient = pyro.param("d_empty_loc").exp().detach() * chi_ambient.detach().unsqueeze(0)
+ x_ambient_norm = x_ambient / torch.linalg.vector_norm(x_ambient, ord=2, dim=-1, keepdim=True)
+ eps_overlap = (x_ambient_norm * x).sum(dim=-1, keepdim=True)
+ else:
+ eps_overlap = torch.zeros_like(counts)
+
# Apply transformation to data.
x = transform_input(x, self.transform)
- # Calculate a scale factor (first time through) to control the input variance.
- if self.x_scaling is None:
- n_std_est = 10
- num = int(self.n_genes * 0.4)
- std_estimates = torch.zeros([n_std_est])
- for i in range(n_std_est):
- idx = torch.randperm(x.nelement())
- std_estimates[i] = x.view(-1)[idx][:num].std().item()
- robust_std = std_estimates.median().item()
- self.x_scaling = (1. / robust_std) / 100. # Get values on a level field
-
# Form a new input by concatenation.
# Compute the hidden layers and the output.
- x_in = torch.cat((log_sum,
- log_nnz,
- overlap,
- x * self.x_scaling),
- dim=-1)
-
- hidden = self.nonlin(self.linears[0](x_in))
- for i in range(1, len(self.linears)): # Second hidden layer onward
- hidden = self.nonlin(self.linears[i](hidden))
-
- out = self.output(hidden).squeeze(-1)
+ x_in = self.dropout50(self.batchnorm0(x))
+ p_extra_features = torch.cat(
+ (log_sum,
+ log_nnz,
+ overlap,
+ torch.linalg.vector_norm(z.detach(), ord=2, dim=-1, keepdim=True)),
+ dim=-1,
+ )
+ eps_extra_features = torch.cat(
+ (log_sum,
+ log_nnz,
+ eps_overlap,
+ torch.linalg.vector_norm(z.detach(), ord=2, dim=-1, keepdim=True)),
+ dim=-1,
+ )
+
+ def add_extra_features(y, features):
+ return torch.cat((features, y), dim=-1)
+
+ # Do the forward pass for p
+ x_ = self.softplus(self.batchnorm1(self.layer1(add_extra_features(x_in, p_extra_features))))
+ x_ = self.softplus(self.batchnorm2(self.layer2(add_extra_features(x_, p_extra_features))))
+ p_out = self.layer3(add_extra_features(x_, p_extra_features)).squeeze()
+
+ # Do the forward pass for epsilon
+ eps_out = self.eps_network(add_extra_features(x_in, eps_extra_features)).squeeze()
if self.offset is None:
@@ -301,39 +315,53 @@ def forward(self,
# Heuristic for initialization of logit_cell_probability.
cells = (log_sum > self.log_count_crossover).squeeze()
- if (cells.sum() > 0) and ((~cells).sum() > 0):
- cell_median = out[cells, 0].median().item()
- empty_median = out[~cells, 0].median().item()
- self.offset['logit_p'] = empty_median + (cell_median - empty_median) * 9. / 10
- else:
- self.offset['logit_p'] = 0.
-
- # Heuristic for initialization of d.
- self.offset['d'] = out[cells, 1].median().item()
+ assert cells.sum() > 4, "Fewer than 4 cells passed to encoder minibatch"
+ self.offset['logit_p'] = p_out.mean().item()
# Heuristic for initialization of epsilon.
- self.offset['epsilon'] = out[cells, 2].mean().item()
+ self.offset['epsilon'] = eps_out[cells].mean().item()
+
+ p_y_logit = (
+ (p_out - self.offset['logit_p'])
+ + ((log_sum.squeeze() - self.log_count_crossover).abs().pow(0.5)
+ * torch.sign(log_sum.squeeze() - self.log_count_crossover)
+ * 10.)
+ )
+
+ # Enforce high cell prob for known cells
+ beta = 50. # like a temperature for the sigmoid's sharpness
+ alpha = (beta * (log_sum - self.prior_log_cell_counts)).sigmoid().squeeze()
+ p_y_logit = (1. - alpha) * p_y_logit + alpha * consts.REG_LOGIT_MEAN
- p_y_logit = ((out[:, 0] - self.offset['logit_p'])
- * self.P_OUTPUT_SCALE).squeeze()
- epsilon = self.softplus((out[:, 2] - self.offset['epsilon']).squeeze()
- * self.EPS_OUTPUT_SCALE + self.EPS_OUTPUT_MEAN)
+ # Enforce low cell prob for known empties
+ alpha_empty = (beta * (log_sum - self.empty_log_count_threshold)).sigmoid().squeeze()
+ p_y_logit = alpha_empty * p_y_logit + (1. - alpha_empty) * (-1 * consts.REG_LOGIT_MEAN)
+
+ # Constrain epsilon in (0.5, 2.5) with eps_out 0 mapping to epsilon 1
+ # 1.0986122886681098 = log(3)
+ epsilon = 2. * (eps_out * self.EPS_OUTPUT_SCALE - 1.0986122886681098).sigmoid() + 0.5
+
+ d_empty = pyro.param("d_empty_loc").exp().detach()
+
+ d_loc = self.softplus(
+ (self.softplus(counts.squeeze() / (epsilon + 1e-2) - d_empty) + 1e-10).log()
+ - self.log_count_crossover
+ ) + self.log_count_crossover
+ # d_loc = (d_loc + d_loc_est) / 2
return {'p_y': p_y_logit,
- 'd_loc': self.softplus(out[:, 1] - self.offset['d']
- + self.softplus(log_sum.squeeze()
- - self.log_count_crossover)
- + self.log_count_crossover).squeeze(),
+ 'd_loc': d_loc,
'epsilon': epsilon}
-def transform_input(x: torch.Tensor, transform: str) -> torch.Tensor:
+def transform_input(x: torch.Tensor, transform: str, eps: float = 1e-5) -> torch.Tensor:
"""Transform input to encoder.
Args:
x: Input torch.Tensor
transform: Specifies which transformation to perform. Must be one of
- ['log', 'normalize'].
+ ['log', 'normalize', 'normalize_log', 'log_normalize'].
+ eps: Preclude nan values in case of an input x with zero counts for a cell
Returns:
Transformed input as a torch.Tensor of the same type and shape as x.
@@ -348,7 +376,17 @@ def transform_input(x: torch.Tensor, transform: str) -> torch.Tensor:
return x
elif transform == 'normalize':
- x = x / x.sum(dim=-1, keepdim=True)
+ x = x / (x.sum(dim=-1, keepdim=True) + eps)
+ return x
+
+ elif transform == 'normalize_log':
+ x = x.log1p()
+ x = x / (x.sum(dim=-1, keepdim=True) + eps)
+ return x
+
+ elif transform == 'log_normalize':
+ x = x / (x.sum(dim=-1, keepdim=True) + eps)
+ x = x.log1p()
return x
else:
diff --git a/docker/Dockerfile b/docker/Dockerfile
index 9b35e0f2..1a37bc40 100644
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -1,40 +1,35 @@
# Start from nvidia-docker image with drivers pre-installed to use a GPU
-FROM nvcr.io/nvidia/cuda:9.2-base-ubuntu16.04
+FROM nvcr.io/nvidia/cuda:11.7.1-base-ubuntu18.04
-LABEL maintainer="Stephen Fleming <sfleming@broadinstitute.org>"
+# Copy the local cellbender repo
+ADD . /software/cellbender
-# Install curl and sudo and git and miniconda and pytorch, cudatoolkit, pytables, and cellbender
-RUN apt-get update && apt-get install -y --no-install-recommends \
- curl \
- ca-certificates \
- sudo \
- && sudo apt-get install -y --no-install-recommends \
- git \
- bzip2 \
+LABEL maintainer="Stephen Fleming <sfleming@broadinstitute.org>"
+ENV DOCKER=true \
+ CONDA_AUTO_UPDATE_CONDA=false \
+ CONDA_DIR="/opt/conda" \
+ GCLOUD_DIR="/opt/gcloud" \
+ GOOGLE_CLOUD_CLI_VERSION="397.0.0" \
+ GIT_SHA=$GIT_SHA
+ENV PATH="$CONDA_DIR/bin:$GCLOUD_DIR/google-cloud-sdk/bin:$PATH"
+
+RUN apt-get update && apt-get install -y --no-install-recommends curl ca-certificates sudo \
+ && apt-get clean \
&& sudo rm -rf /var/lib/apt/lists/* \
- && curl -so ~/miniconda.sh https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
- && chmod +x ~/miniconda.sh \
- && ~/miniconda.sh -b -p /home/user/miniconda \
- && rm ~/miniconda.sh
-
-ENV PATH=/home/user/miniconda/bin:$PATH
-ENV CONDA_AUTO_UPDATE_CONDA=false
-
-RUN conda install -y "pytorch>=1.9.0" torchvision cudatoolkit -c pytorch \
- && conda install -y -c anaconda pytables \
- && conda clean -ya
-
-ENV DOCKER='true'
-
-RUN git clone https://github.com/broadinstitute/CellBender.git cellbender \
- && yes | pip install -e cellbender \
+# get miniconda
+ && curl -so $HOME/miniconda.sh https://repo.anaconda.com/miniconda/Miniconda3-py37_23.1.0-1-Linux-x86_64.sh \
+ && chmod +x $HOME/miniconda.sh \
+ && $HOME/miniconda.sh -b -p $CONDA_DIR \
+ && rm $HOME/miniconda.sh \
+# get gsutil
+ && mkdir -p $GCLOUD_DIR \
+ && curl -so $HOME/google-cloud-cli.tar.gz https://dl.google.com/dl/cloudsdk/channels/rapid/downloads/google-cloud-cli-${GOOGLE_CLOUD_CLI_VERSION}-linux-x86_64.tar.gz \
+ && tar -xzf $HOME/google-cloud-cli.tar.gz -C $GCLOUD_DIR \
+ && .$GCLOUD_DIR/google-cloud-sdk/install.sh --usage-reporting false \
+ && rm $HOME/google-cloud-cli.tar.gz \
+# get compiled crcmod for gsutil
+ && conda install -y -c conda-forge crcmod \
+# install cellbender and its dependencies
+ && yes | pip install -e /software/cellbender/ \
+ && conda clean -yaf \
&& sudo rm -rf ~/.cache/pip
-
-# google cloud SDK and gsutil
-RUN curl -sSL https://sdk.cloud.google.com | bash
-RUN conda install -y -c conda-forge google-api-python-client oauth2client google-auth \
- google-cloud-sdk google-auth-oauthlib \
- && conda clean -ya
-
-# Add cellbender command to PATH
-ENV PATH="~/miniconda/bin:${PATH}"
diff --git a/docker/DockerfileGit b/docker/DockerfileGit
new file mode 100644
index 00000000..05bbde40
--- /dev/null
+++ b/docker/DockerfileGit
@@ -0,0 +1,35 @@
+# Start from nvidia-docker image with drivers pre-installed to use a GPU
+FROM nvcr.io/nvidia/cuda:11.7.1-base-ubuntu18.04
+
+# Specify a certain commit of CellBender via branch, tag, or sha
+ARG GIT_SHA
+
+LABEL maintainer="Stephen Fleming <sfleming@broadinstitute.org>"
+ENV DOCKER=true \
+ CONDA_AUTO_UPDATE_CONDA=false \
+ CONDA_DIR="/opt/conda" \
+ GCLOUD_DIR="/opt/gcloud" \
+ GOOGLE_CLOUD_CLI_VERSION="397.0.0" \
+ GIT_SHA=$GIT_SHA
+ENV PATH="$CONDA_DIR/bin:$GCLOUD_DIR/google-cloud-sdk/bin:$PATH"
+
+RUN apt-get update && apt-get install -y --no-install-recommends curl ca-certificates sudo git \
+ && apt-get clean \
+ && sudo rm -rf /var/lib/apt/lists/* \
+# get miniconda
+ && curl -so $HOME/miniconda.sh https://repo.anaconda.com/miniconda/Miniconda3-py37_23.1.0-1-Linux-x86_64.sh \
+ && chmod +x $HOME/miniconda.sh \
+ && $HOME/miniconda.sh -b -p $CONDA_DIR \
+ && rm $HOME/miniconda.sh \
+# get gsutil
+ && mkdir -p $GCLOUD_DIR \
+ && curl -so $HOME/google-cloud-cli.tar.gz https://dl.google.com/dl/cloudsdk/channels/rapid/downloads/google-cloud-cli-${GOOGLE_CLOUD_CLI_VERSION}-linux-x86_64.tar.gz \
+ && tar -xzf $HOME/google-cloud-cli.tar.gz -C $GCLOUD_DIR \
+ && .$GCLOUD_DIR/google-cloud-sdk/install.sh --usage-reporting false \
+ && rm $HOME/google-cloud-cli.tar.gz \
+# get compiled crcmod for gsutil
+ && conda install -y -c conda-forge crcmod \
+# install cellbender and its dependencies
+ && yes | pip install --no-cache-dir -U git+https://github.com/broadinstitute/CellBender.git@$GIT_SHA \
+ && conda clean -yaf \
+ && sudo rm -rf ~/.cache/pip
diff --git a/docs/source/_static/remove_background/PCL_rat_A_LA6_learning_curve.png b/docs/source/_static/remove_background/PCL_rat_A_LA6_learning_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..4f66e009c04c33fce8ccc0860f64019e9f7f515c
GIT binary patch
literal 23433
zcmb5V2T)Vdw>}zr?@g30MLHPi9Z@L?VxbEt2qG;IO6a{w@4YA>D$<c<@1mNQV$=
z=!D+k9e(%z@4R>C&6{}{7-l#*`|Pv!+TU91`__(lsr`(C^e!m~1ftMTS9=8l;kJQ5
zIJCrsz<*c{h(7^;WL%yYxV(0>a&dp}{1Noxy$i(7(Z$Z@1Bcs3XQ+*%16bsd$U`BH
zPcAMHsH~`{{r`MH#L?MWR4{1K02qV>qHYKUfyn4?e{nz=S$9Anqhk#<l{c`g?KyXE
zy_Kt*y$MW9Rj-vJ@#9ec2e=|HvkFTE^h)lPR2j9px__FhnI7LLt8vvgDe(%#lM<*b
z5@HSjRciEdEVb-zp7hkysrKSJV_FiMa$|cuxG^RZGWAFe=as}wla4HC`0b~fyX6y;
zvKsIysNg4Nq@kgSXMCno{+~<Iq=8RvFLdcr3vmG-;g~>k;0rW1PW1R%o#|8WtQf=H
zWHts6En-zV-`-Ku#E0g|8Weas4U71zlb<u>kyJlX`}Wwi{1GWZ+sIbBp*%I@$b`9D
zruISmlh1Fwyl%L-C^THlH_!%!JJhsv2%$m5wS$$6mEH#{t--A?0v_k-Q;}9R(+IbW
zDJrYuxO5)<J`7?j*2=9my@Wv1AF;1*XXW-KeNd{33!M4UD!(|Zb*5Mve2=|l6$18}
zLy#iuUYs)Im-W{xfvdW#^z5vlh%E!Kv4hpn*b;M;SXG2oimg?MmE;5hS?p}KRGbCo
z_4a<ldg5n!dRj5&JtsQ2wbXoO8m@luP1e8_qE(P$9wB>ht#wVjhBr$ggr|+mCIxPV
z?(9LvOO++DcjL&&LD$5VIG8u}TVrtVriss>u7y}4T&mz&)`hZ@ob-<}80XCg8(Q!K
z_!3+vK(1rOL&Tds4CDw(49E<KfsZO@dddgg4;X6y*g`l-PLAXG&((qA7wX2&oB^VU
z^(ZWx!Mj5dMYTI|;>F1|7`$OV)+U?+)y}L5cYwcxl?8s*Y`sH_bN?5MAL%I#_ANAK
z;fl`P;dIt5d4#1LH@w&x$pm9?iq-514LQht%`?xJeB2XuFl)5pc@UtHd~8PG?0>4+
z(;YhGQyOD^=vs$}__OQ;KNaZd4vvle)pIpZWJg3Rz}iUh0C(j0$m{Ei%$iBm8=_*$
zalNbWz$%jY2A<6Cuz({4=dsuKcJL1g#H!T5jU9*-9JzppXg^r8)4agMC0s^-?$g?C
zHq=&NH%w_9(|Seq%e1)P<Gz0cvc6KYNAZI`nfjmhGTK;Bjyp2>Up~u`?%Y)?z9lZ7
zzO0)fwiGI&)MQq+8xsCV@K?}VqFVJ0O3sQryua*aswLVp*d8Uk3EUIB`Nig#d!Z+v
zg?{$S+_SA4Kkk=9*3*LXGY&T8naXnYR}N&CnpKOveXOmFH|iX0o376`H*`bKr<bm?
zp^<}jln@+jO*#c3C9XqRQc24^8I6tN?aMmE16x37l2C|1<Ts~}#_5s_yb4+0I6#N6
zQ>MVES<vqg=J~APt&Z(%Z%8}Jp*G7(U6%!MXn(O3|5W|Ger<-k1Fmvu`T`TgzArpH
z_Z3wuID5(ZTej}WD{7$%^nJ98c`?-Ct?*ewz=%MPNQ}YHW0VV;9voSADC;2PS`CrL
zd_im(qgBBdiO|vwg%o2v@>1umq-LQdzko3RD*prjiC-SB2D!HeV*f7#kp>o4q@`aq
zWh`m=Gx-nr4-IaW!LFKKVV|<G?_>!o?cPT}1RICzGruE0jDhuS?&`qF(eJ=s<hy^y
zSz%@5k6p981X1R3Wr!^rb2>5NPb*no53(=}=x&H#Qz0yQmu+<WAF#Pmh!)UJm+cYk
zr)J83-podq)kF?()3`u$8zIFe8`oc+e`QXL!Ez+0c-VVQ0hINyWW<fY%9RAV=R#DL
zxQWDs==m3epC^G(2kp~XL@u_!hQ5v##@t)HH2c%a?U8BMDLqO#ksW)EPWA%rj?J(t
zb75ssU=K&p$a7jOdHLPb@0v3G8)wbb@!y}gLbfr#C>;V2GQ#FuwER|*_>OmT_mK4@
z!y>Wa>RQ0Y{+!QYP=J#@r4%5#XIgmFb>WDjB6mSDJgu_h;u_4}FrT^yc-Khzh=1!+
z^h*)AG4q<<+ZHaaXjRwpf0;)AjgrG(oNiHaWbwiO!BL;U6V*nyZ4*5`z3NC#t{1((
z4XbLTG1C*-uQ!0lsvYauyOsx{o8yv`yAbK%{CDIuHmV3a*Xn`4CO;?{<D1_av;N+p
z{$Hd1Z*(4pO9Pm^ulkkRcA~Vf{3Th?*fUK{(zyoL4$N!gv}%)E7%%9Xsr*Q0_4MZY
zieYk$lb*D~VNSNjZb~(blJ(m=miQmH82=&#N{NAEz@R|Rp7D0tgvF}@zOL7_#0j-k
z<o<O>VD;W~O~aG>{GXJf1_^kQ2vw8HSau-=4mmHnLv8GAl@hj=RqT+W@4a-HujIZv
zy{siSLAs1cY3E;|m%d)~0`cn2eJQEb-@k<j;790#v$e^J`^EZYdelO~89-?HX^eE~
z&A~OyFQ&TwtRYGfOdQIhanI^QaVzgX3;NBG(DFa!Nr8(XrPXb2M{8!*{ncTQ3n9rG
zW@HND>#27r6l-BBzUk~fBJRkB_XQK5EgRK!eQhTwv>G-j$NzxjZ}+}TXKMnEC0If_
z&b}Z%DXn)TK~kCsbFR3wb4t=l)2YKkhg<Zp;|P=amm0Jam8s0BaXE5>vxmP%>dk!0
z9E(5Hy$cqpq)J=3oMpSAC1Lb8jBV%H6`+pcY%`BLJw5%zl;0N~8_P#ctk!skOSh22
zK+Syck-;AGp$w%b3!!mJM|X(W8u@I7jo88);uP9_K5@(>iNVfYqZRM2%{gD4BG!ZY
ztuWg?y1ZcgonDxKj03niXiA>hk5f>R34)V&>ND%NNAoR8T9G&hbcYIG1cY9%YlZyw
z6QRtHlC}yvH@|~Pozrfbhbio_g$Ad<*6T$pQS%|+>Ykh4JJvBO;)V<(az`7aiTq!x
z$`g=m8|F@C#)e%+*Zhpgx?2Ckos#_CKbB*H+pk@;0pGyma0NtuhUQegR!fj16jPN(
z_x~Qe-2axSt`Qz|(No1^K^!rIdrGWQ{*UaO0lx=8C^wo6as@#ef;O%SM<t{8?gLee
z;57+CPd062plDz2wb?3FQ^YZW)*Y~`yf0JUVR6`*S8D5U*(XVBhZ`yjluz>L7t}JL
z8NVVP(}R^NDin9{9{9$@bSAEYp+$6$rqdsW%VHI&PO|#2+ERA6rZG$WN)9`)YfWuN
z4F}zar*Drh>DYDs&1KOpisEi~p5gCEaEJ4f(ef0>Fv(Rei#aSud2<=IXzIb1A{+qm
z-1Q}D^M8Fu-lg2#&*T1k6-w%M46Y@md<x=yD#ZN6*VOY;0oyi6O^!a`i-R4Hy7b?D
zQWQ)tdetD>Z2f4#g*fHIHGV=5ibP`}rE5lS`2DsJ?2JVebj$}EGFDT|5CSYQoC0Rm
z^Tn~r{yJ}<ZcjC%*S;*b;7|8{q=Z@shbgeaFR;UpVAG1u`^q`xxIXC`y3@fVM8Gl{
zit4H5=Nl39W9$)a&DB4@OlOtbuk#(eq0#YQ9J=XWnbl~lJep;?Vbm8<iW>EK%}mFH
zkC#jK*LQ&2R7*a`COZcGcYOPkV-KCtK4UPhfr-3zmYawRyT-|QpO8BCv9@g*mJwfT
z?}G01%n<Vye`zRr-r}YeK4AU0oV-;j;Mn?#pdog-eE(2fs$-)}0QYaFg#&lRh5D<3
z(P#5(3i<n>!(OvpHf(GYEn{n|pwDF2U!Ts>!^fhblz5+StiN8Umqk8y^`L!~cQPzd
z6#1x<|Mhpsv`tZbX6C)^Z{d#&0dP8g=41T6sz^EY{ofVyPh)q#)U<VGHd*A(w+i;M
z`p~sm>~h@MZwu|R6WSk@co`T=%^7vln$=xG=QTW2vU1mU==-4A=#beL!|=GEUF#PM
z73QLZm8^@N`~9=RS)-qM6$3$?#&LV^#`l-df_?X3L}eG=lhcZ)C{L!q!p<|gx~=xe
zI=fmg8G7eRft@O|v5VirFrkTyPj$=983+Gdqpx0w4!zb#8|mkqn>;R{wCIjn2&U6U
zAI~_>V{D_?6pV1ndwqexEj)AWB4KbS(&RyaKhQMZvyiGXzm&I2VtSsbukuuC_Me}p
zaFU3+`PvgZ*<(>5GqX<g3fSPXgO`#jiTYVMM>q7f9%Oe}V?h)m)gdXgR2Ur5Y3vVj
z^3HseaX`7HqH4k@Qxfy>d9j^$Vk?|r0M0(dZ*4ZfgZGTaC1!C%P$vH^ON|@#b6Ebj
z!sFoF9D=!vt%lgBd`=%FyxwkNF@`)%GF?5r?<&4r(HyIR=HN!QI?x^5-|onx65gRM
zc($^&ISy73Gy7X}y(7R6Cn$hDKrS09s;9XNlBdifrpFt?ZG|F3xLnO>gHZL9t1VkT
zJ380~wu0)Q*Ec(?ncjTfCpyF-O@}Q#k1<~%h!WHDnbvlKFXC_lEfS=6jO#tgy}4sw
z0Fo`XD$ndhjQh)#Nt_tD)}m6a?a@e^T%5Q&CrrM>XV0EGq^tZ*rm2SE`t&$I?(sb(
zm$l+{GRNqiYEtnzY}a;vj9wk`!uRm`8adJa*)V^dUON_#cInbICG;~7L;RK7oa==6
z$A0?n+S&cNH<#88ZrlAd&scP6PFpT^8ne81>Myq){g&FvM3-kATMK{u2ygMZ*hdFZ
zIWI==;bOa(%pmO~d~Xm~G2csAU>k>qhDOuzv~58_0gzf=#}}722xQa>(vE%*AtoHh
zqr$=S*DB`Cn0kPoho{3+;Ws#mh%FFimJ&PwQUqO-MiIskq8?T;AnSiF3^d^{<IVZk
zz{B8@=r(m!nwrW=%Ix-3`5NnKmzL>t&0BPn11`SSR|weh7%uaSV;L(7qagDHlj7s?
z{5EU}_mJs(Zlplx-{EcT^wo<*Nu@t4*No9@$X$z&#;6kH&jjAZ{ELxcsa#-G{R%>|
zKbtN;sBaxLublo4QICC|K)1+YEoQ0HmMf7_{EI>@*5CBmKkrzIVN$^n#OxL7M98Q$
zU$<pE;N3|5LAA<J|HY(5Yz1VYeXikg6uVN+gsE3FlPN5~I>#e4A%PaiJC?emE!u-g
zxOsT?dhf?V#tp42K6H^z*E*VQ|1tBsay9i>VaS0klAwGq?2!s5CfzZtT0=Ps_qI%*
zWF)4M1!hCb@Y1GE?^c@NG_7Y{NN4Se!d+o1d-&!oFY)D<v!HNGuoqb!ZQ5t6W`8>W
z7G%9V#}I-I-{F}G<ly0flGxr;cvH%3yuR|V-h;ld&PUP<+G!A35d+O$rlARiG7HZ<
zQOD~uurET;D(zjH76}!#j{n2RrSzyxgA@pdTnz&axeRW$NGA_V+g~)Uo0~$yrYB4H
zJDZ3C-tPDe`e&YJ2wWLFPA}4y-@{R_1%HRXvH15x_tuL*+<bg^kcAK?y@Gk?0b%)5
zbPys<JSb5|v3}Z?#eLQ(%lCX86m?H327!3Kip(zNFvt+Xunb5t{w?w$b-E5UBj#HT
zM0?)REL6cDWvs2B?L7mdm|#4#b5%g3Kp{p3qxVK*=wX!$#$@Rct9pkZ3WsbS+aPRt
z?C)CD)Mf>nW6)e0Jv;t^@wi^;qE8J`|8g?SzwJBin2nt1!b?+_!cLLKtWV?|JU!+>
z&gaJNdpRo)%b&;G8SrT?wvA@Rx%_tQ%-_pk<8ox59?f8AXxO;DJ6oS<>`)(~k;q%|
zae&cVk^SuK%nsF5-)-x%GwV@%m6Sxke{gVL*ChUGG0S4T+glTllXlVDu2=ATIk$S#
ze0c1EuE|}?>jrs6le&Pn#-z+X%5vAPZDMXIgtwVkqtE9#vjVoe!o-qQ7=NwLea@}W
zQpLv~NBx@wdYl6eub%T|S=Vw*B1$`LpL~B7%Xai2=TX`@qN&{STkYQt_-oD*x~7|Z
zHLM39sN7A~V1R{mkp5odLtrP6DaKgTWU8bBvEQ=hTOPNv#y-bl73I=*mF%^IsW0Pw
zdzQR>Z_aUj_Dn^>!yiR$q_~xOZY_kduNIfJERO0Z@dyf%0l9di+ZL+jKCME4d3Xew
zuO-x%L}BJTwR6l$Cg~|`_^Ir7o`((wQr}A_+ZWjnG#nY8-Dum8uO%U?-x{A;4)aQG
zJwPpGbl5S6T`(TJ`^0fjqSWD8v&t7FyLXtB9MyF5+AW1Z=Dj&lPK%|Q1kbcr^c1Z(
zD+DUM{^Y1hTQN>|#elc`+bd5!iFG-FLVb5%k7<W2s1B|UE;+lR$)6$Ak0At^7O(2W
zdkA`{$zni}(2s??z>|y#|4SPaQL6TS>E7bd*AC^`h?DY<dlS`ON8X^>YQDOvBRmAL
zJR%`h78fk8Wo6WOAuVH1kRtJgj*fD<{ZLK&-gL+*ag!bm&CgivY}pF$vwfq3t;vd@
z&Q7)I8hb4p8-6Oq)4cC$i9aS^y=-58a?h~79^~&(l6WRwmKNJO5APQa@mgk|R!C%?
znz^+U?WH2+Mv4luCfmEk1>6>w|F$VN|KiV5cfi%B#4KeE4z->`waQf`)ft%^d{t>P
z+<?)WC+#?4kDc$uDCV&`4jT5X=?OCT^~5dRU8b<jCul>~<L2ezh)7V~(QC0lO~(@9
zIqGRD<@d{k1;?fqS*IB~VrKg_*QXW~+efJT_2)^GU@F_E5h$vy&djgX@JwyH_I2t>
zlGUIA#r=`i%POf@hoSq;ZJ(u09>BZL5~q}vLaTSW+{}zcV3wpkjvn!ui{z~`fdQ{%
z?K8W6bd)WP#QZ$8$<=JtwrM<H!R>jrYZD&+rfwnhUftn{Rv!XU090a8cPgxl$D^fC
zy<1aN>m!9^>FMd0t5U5~`2)ed*Tf+J2CN($XB`P$I}uTSN@UU5Ou~1r*)HIw=H}sy
z8A^5d#yx9~llO?43*|_@k<doR#M+O|De0SPNvs~MTCcMw{CO2y_+!cGx$&RqpvI3Q
zvRz({M%Czdq>O%=-Qi|&leW^{r(ToicqZ5k{k#+2zym9}ikHkJiGEaMgdHJ{5X*Tt
zmijfew1Y}1i_uv>SQXp0H3tjSCvZ@Rvdm-;Rmy`Sjp?5m%*9oFR7C=(M|*ORzy!0d
z8L@HcbGY(fq#k!lH?9ALAWI%r>9l)@bY!wFN}?m4)-|sdHWk)l9<0%rUf7<$E<ZP3
zdnC=$8QK&(^}30Dd|8q@b#0qS%l%!k{8bi)Q>4#Wd@$FaEalbW<>khvrIpTEC}Nq8
zB$Q?EnwF5&;~~N~>}5GQCU&WkOyt|!T)h1K2Wn~?ww>WLLMC<P<MOO>ZkGGYeT}S*
zXRGf-U-v{YzZO@x3x^IV)b>bJR#lx8XN?5^!>&ruJ2$nWpVcs>j1>6@G|v>3J3~2R
zy(^>6HuoP;1Y<9v0<uS?3tU95#s?QWM!woPfTPCuKGHj&x|tteIIpj@915lp4Dcn+
z926C>LFh^DVN{XEah??RwL?KCLjf<Dy-(Z|L_Jb1puBBzLmef8oYt~^jxl^w1n;K=
zO+lH%B_C!fK}CwVqS?*1D?|CEVcL(~3Psb2+dycySi$4R2!b|P)uUNM8$avQe8BoG
zl2v941RW~}__-#Ed4n>kwpM(2H~2eknS7@knxpvvQdE0OaM?V3#~gGRm)o;JjR*KM
z`YH>Iuxo2mcLElZbIwC~-2N|l(oCK#>sQo7T5asRS08X$F1&UvkcXP<BykcM3CeL1
zg66~LFX)rFhf7Dzuw`c%T^f(;4-w$HS!?fd2p&P#$Jv9bHW$Z4aw6BCX&#)WK|=sB
zEYqJ(S^Ql&_3PgyVn_ZHbLQXKDt$4+r}GUgWXUw0LsvGnV8Ml#o;5GtZy#qP2Korx
zb{!kVEgrNoGY;t%zFu=%EVD7LroDzr1^3rI*uy(L^TIAYOmhKWy2#j>;fm%`e0?~L
z@V1&D&C5mMg~gmGQia^`7Zbmvy3f+!aZxu16d~0Y3deWfU1S|&t|q*KO!yrp&BU4T
z<uE(AJHPG^9l#I%6bykrM}=oWGyi@nnohpapK=o0LWEpfaPs3>DG}|?gPvWRJ#i-x
zGdN3U!e>A7LlL@)Kr_9No0nqSc&N$oNDs%9a+w=&4SiAq{qB7Q{+#N9ox=^N3{{p)
zFy*=iC`-QJfNf$JK?bltTljclOZ@OQzLPB@kHa{vbWb!pY++#6!+Cq?ROC@o{%`Wu
z5G>tdR3%lfgAcc@v)+OEp3ibjeFTo-T%Y%mj=_)MIMD#KD2nInZ)#Id{Fpmrwk;)(
z9DUPIY8!9(A=5WOoo^glKcDvszCRUhuB|o9Z;nmI`?qud>y1t`wDM_?QpFCarlxiM
z$nexgw9ABANS9G4X!n{TcYfDr=fF#;nyHz>Cnf~d<Jn5j{O<jEVxw=~MNVB$p-C-`
zcT+wQUF&r`YCCn9r;4HeNuiyv(8sfG+!m9XAmugnP?y0xTTfb_=^_XcyAuWAv#Gx0
zt!R+fv|wW_>7v-1If252mECA_G!wNblPId<i<@Rtq{tO%7d`$guXA!PXVXF}<n?oc
zp$&tH8!h|T9&n#iiU_qya@Ek@u#WBrVH-i$S&j=>+lBGBrczIG{f0RHbbuU#Hi9NQ
ztKY4Z2`bzF0M!KCYoW&{lg7B(hC33-jkhh!VNKsl6XS%vtJ-X>k!)_FaJvX_n|m79
zTo~seQ-A8+J-z*;32d#PCAB&#E0c4wT2|z3m5bjR8^auN>HJ32*Y}eY<yxBFi?4v@
zkx~Dr%6v~wY+J+Rbo;Z-&YsYwGbOv|G(4k5Cn2>e&acsf+rU4sEAio_H`}HUF(2Dx
ze5Z7)dv@026LK+t?tZc$gSTkB&D`0|=W*}Rb82IK|GPns4h;CUZ<2}r_;@#-%%j$C
z{KpL;{T<lh-R2Z?KU{J)%pmsV6C2FJbAz#%6O^G2vZl{YRA5H<DvmAv@@VU8mJ@^t
zNE`Zu1@c`rarsGn{~Ue^S0A8%P>5JeYx;D*qt!2^MyUy>jrKB_SeiXn#I~cA(Jy-5
z`tWmVJ5F<~`4wD$?e!xfzwuALEoC^Lz~moJQUlQ`53FUMOWW(md0{>NmW<FV-1l9N
z?)|-*`tfL7Rned>fWY=3Ei-YTVK$p(x=O-=$bAtm!W@P895he#Gh)xBU;US~=-}@C
zk`b4W09pA$)rZmH*RBNZD~Q=;x{@m_%{rBLBvhBYd`CX>REh(1jZZ0i`R023{evEg
zwU^?}X_&gJ##pO6ojalgZNkG;>|xffa4(63VG2q`zZ3ZdvlIZLVlgF{ZJAAja*@~<
zXFQ-oyUgc6#w9Q2K-_a^JKrnC0c<{EJwKwMtACr4cT$)Cd>6^a-zIAR3vHGkWec1A
z13|B&=8@?`s$nXqe_y|Hv@s2}Rkn791mcHVeD&in6PaljR+5QP+<yH)gzSBK58CWP
z*fQGcH}2Tee2zMv4H;5zzZ2h|w#<bx)`M;iu#X~B#*u3oQ(tt6sx2U?<l1-t#RF;3
zz52g^4WiAjSRC*9GXFm7zA!fQaF}LS++z&9>sdO{e$i&{-xgKWr+{&NrFg2ad43tV
zoqsOwG5oaD<cxK_l)M!O-yp@Ccu7Y(A~L4;*Wr^!)w28FoN57*6rJ*8=uX{X&}XyV
z-ve`fKD(kHfaJTfmVQJgpJB-4VV1>(I~?;`$rNCOt(?Dij8HYLH#NitpOv*`F)ieo
zy-_;ShxyLzPGH10{Y0H{LfM~jyle-HHo^S{*m(VX1vOXwBIm7>j$r=t4lN8!0v)?L
z=q9L$|Bm5svo#Jqey#j`D@Dn}uP12==Y&3oSA-5eK8kH|OKHB_G9D1?Qu*&E4q8-^
z;l0<)U@LOIEC_DxyzhmtneFB2UT0U=-r=Z@lyaX@Ue-Exmfti^?kHF6+UJhX{irix
zZth+2Ji8h>e?GIckdD2xk-;eKjUiDX91~`hNIoBQMU%@c|1wi*g<@HnuFuRcFAt70
zq2$_S&4?H<FK%R8Qmcxyu_!U+z?5F+V|L+{W{*wge>~ZKIZ9E#-S~$Ei^9Xz3#5qz
z6+;Tf7m|_fk;=iiV<KG_dogU!2QK&}sd%4fYd$T8JxrvN_Ejd@kz10ct%>#D;y<l=
z3VmWvrtsi2T81>xm==7xV)<a26n@<PYR7Ze3zb>R);d(T*Xe8G^PC+hgLFlH0E(f@
z{|f~%QmEyz@v{haIQ*=yukWk-F6y>o%yJpg009C(nAy$wgpffwsC9nG9|tc@+KEd?
z;fP+e`H&WLb1{Dd>Kz#9*gbi+eVq|(&fnxyQro}2(7N9s^QJ0c3V^pun-N-Gk<)kp
z)JDw1mxx@@3s6el`}HIDM&1zAy+{V!hl?Knz=O`h+U4UygM6vyd&bQD$oda1j@|6Z
zPcTt>!y<*@>W7ay-x#>UwkmoI>L3dabDTMz8~kkESQQj(i6WGoxjl^Cx3|t|DNW8@
z#OLC8W!kpvBTylDpRJ?@t8!WmW;{_<#jQno3Pv&tzj6<6SbiSHJfwUb3)z#SP9<h)
zCa}MFNxu(v|5xQQDH|ji6tgrZv7LT&!~ztDr_NJW<Vz{JhnW6+FFI$3iADMmp&xx^
zg=see^5S<%B`Z8_IhmB?Sp6r;6cgsN<zM2Ihd)xfks(-x2^Rva+m%Nd?4s}q>bhO3
z>53t@tx73CXE%tXsH3va^c>g@hnLZ)YpPt>g@<i3aDCTZoZaV8_xiL)>7mQoOQ6!_
z%k_TP!{%l8WYYn74cqbGoP2y?r>8JVrAx@GSN}dzU~y;p{SFlMYFr7Px6q|LKWxc8
zFbjBV)>awYbQnXle;LKj>^skwb@O#8n7M=bc$RX<(~$-`ZpY_kOM{77L#EcoEKFNm
z(C(h_gB(Kt(ayf&{-~UZS~em~Rgg-KgEqm=-XFf2H>06305h2MUdS7e%FIhlr<(rx
z%m9!t!V_kGK3x^c9Llv0bD+~9zw3_X6V$*D-F!fKz}zxgNV#2Zky7SMlKI^66Nypu
z*=PruQkP1xB&54FZ^-Fo_29iVz3Q<G-$nRk`MQ@X(PgloUwvS!cXbj_m|-$3&Vf3&
zUZ(ywyc4+&M0I9sl7A*_Re2?fr~N%|Lc?VL@DWg|EQgrIya5Su4bBwOx%=$I&ig2F
zdYoLf@U=_bOpZxa&+DL1<=hI3TxS*hp4QHtoCFTlUT96`D0n%a%(+94r>yleTlLFK
z<uA6XK!Dr(Cu++^m^OPlO?^z=b}HfJxs6H^ES<cbRbv07s6}Q;MRPX?CnAG5{{E5x
zYTfmBVQuWbmNWai>X|13iWp+ro&%Y89_tIDhJOm<M#J0|95LiPs9lAQ{PUW~UVF{0
z&+V-WuSAh;2@%1#A5E^>g&0c*mMRXY7Rpif`R^Xm#82ufGK@R<O*Wd_vAb;ii&z;*
z)!&+jbWqH?PP{{8%Q0bC<lJ|bH_KYB`Vx3R&(p*Ljy5Oo;?mPKOiWC=(6zTEv-66k
z>erW@{rAM7X8VB#O#9boX;<=1t`zhRQ^`fn?qN1Mp7)fU@2xr2oeV{vHT#)pdsQ!9
zFn0DFj7~Tq{34}W4^QNG<YEQuf}W++@PcJ>Zro9v^3FSL3^?W0&p(qm>{Paf-eeEC
z6Dhza2C86a1ZUQxAFr#mc(`c2?%LcfzUNREpe5Rz1dvX*%;(9xpAO*6sivO%#d;-N
zii$b!J3{v+TCW-WhZHZO=C6-yZ%h2~5@To^0TX2S2@?(k*v$v)MY$t2_QqLIySaux
z>j+BmWjjaRFZ*L!ujSfjQ{Dfp`ks2F9XrCC4n>H(N}9votq;1nT;I5DYFF?takOMt
z#Q@aLnOz?*<L`l`A1(IvC)TN>jXnYsPY+Cac+;UG(;q)Q45eb<u|j4sZ-3azei`0G
zkSKjDobQ)f)8yUtKFwxDD!%)JRaF+WBmZ5@YF>>jY;UoPoJG1kUh7}=la+J=&4tWo
z=<v-Z5&l0%B)%(L2A8EL^=AudGv)|;hCZZ>@8lbJ7?~7y_~hQH?-iwJu}Z9D->l0k
zKW(`mRoAP3Ab-Z%ZY}Dq$Ywl~d0A<<BL^)Pz@wpw$oVt65UHfn=qt7sHElc}dns`#
z#vE9P@3YsGJd+xoD{B3z_IV9s3bt%Wzud(_(F*b@c6?US4U|vysRxJkwM1gjzsNxJ
z>Q?q<Ue5xecc+U}!DzI|=X13dH_%O^n@7J$%VA4zB8NA9RC198QQF>0XLnkPdSm66
ziZ4{6zB3`7>@{=;p+mW)9(GvJoBS?An4oQN>=%Feh9E9u&o}lN=RO*R{C8Y5Hn@l_
zPc*JMaXCt+_F44@D<(X5-0T1Xcv%*-N}!O6=$e3>&=-Hl1=(!C-;9TqhH$+kmnXYK
z%1q!tbJr+w{_b_~Pt=nMc_ZqgP>lrD)d}^+dUVmJ;P(pe3_7+bWsPssrri|Z7dKzs
z*hZs)aKCqU==S<J8+bn;du(!%1XQHK3(=q;1(_c+e?O&as%3;8!5_mx=#gI$R+6Nr
zQV?D&57rOP4VHt;)OH{q`cwJ0;^qhB2h<|$LZA+xPalfnoGDA?31d0SDPUp=QoPYZ
z)(y2UpOkmb%-2kl1=>~=<^?<5oTo_4d=ZHG^X24VUNAI1XIc2Ix7`gTjbLT95NTj$
z{*ztU=wKZO{sLtFL+IyyN1C5RnAuH`m1Ik(kUq#!7NbasxgdkEn!c3BxkE&cLnY(5
zeDwcCFyD+}=fA~Cl?%JSjW*1NZcQ?yGnlL^oSuoF<jSNS|BNW$?j8mRh3tp_=2Y(H
z02b+%aWou)6w-0T?9zq)ZH`WP!Wd5|P7f>f?vcSD?6gl|9_K8z7ylmrZ;p31#z!~}
zsvYFF$-fAGd?ig~C6CvF1xfe?O5v$Vtp1)tCW|6IKwBn0T*e+yE>wSg$7QTYLys?|
zWCa;djJu-1cXOVpeq@pvbQGpi5CZ-ma&_m}(?o!xx7b;TNqv6ErQO4gEO2m3vq{g=
z^Kebq;%h=e!eoQ1p#4nETOa+f{j=c21g^=3>2(h2GtN(5px?QFX6Qc^*U%><tQEOB
z<AN(|osP(=dY@~fPxz?JArs^=TlvW|1x^o$9b#muLXeNA4DT+_Q7JC!6Ma-`RO6!c
znZS{nNHBe)W%Th(c}$as&On!6p*Z7^HF>01u2tS%3R%w=zz6Ueb>CH8H^-y@*iKY>
z9IQ_Fi@XQ47njZPk_FMMD~jrEJezFuWd>A*qJ=<$bqNYN73-(f8!&H;|2U>Jy##jl
zrDJVewtivY{wp$Iv-Dx8l=JRr*WE*;wMwR%_d3O$<3(1({vQmYY1L*MJTrM@=!Y|s
zb75&^o!M%9My9s%m8V->ELJfhHns?eQWu^LkAg?IVB55gGD}cFUc9f|)XQPf1cwf}
z?T+(lIz@-vzA_C`nFm^TCS{@l2mlDZ;0KH*MgqVzLZB$m&43SOTKL%QFD0ksPp$q$
zI4`Ib{1mH_Zfd3YZ?5f2hhfhDU#?9dq9E8(dY7ShcFg``+qArcr@hFnx*&iWLRNCz
zFe?RG5Pb?Iicq|r1#{N9Sg_lRs<Xj;Ve<71m8Z_$^klqTFA&=>C@6J?`T7$GSzk7|
zr4MU^oXNNGrBZ!@B3Ck|`rt|dUf5NVmWGkTPw2tNm(t~w4{kZAui<m?YU8m7=3bI4
zr{J|3TheMNtCE#kM`Y0pO<ebEH;z#~`}}7!mIQEGFyN*{n;6wsveoMpitHc)9-P(H
zW#^UQ=6qvG8k<*HP<I^nv3QIJ`pjs=8YJ;MgbZ2#0#?#7lC<(XY>U7A@^!B<8PO&F
zVq&kbz{+l7s`&ej&oNu0DSq$7q7cjY-Ry-s0<^}rk4bb0U1Wu9SAQv`9@V$!p>}2&
z)6cg9+F)>GFwrV!t2D#}ZR}J_IUj5z#1|qzK0lSryj*-JBJ+*UPkKceNa2S?D%Ahn
zj)Z9f$&Ksd!$I>x^O_&BVjp|e#C&K2$ge&l1%`3&%@>X}F0|YQ@ZnmRpY3<0b>Uzf
z>fl8RE#?cohk!3gzc+v8aIK+4=%QJ6b<U;B2vxyeml``-8Nvmc)_wbnM*CcmuEj3C
z2L`o@^oP7$I=fr;?~yVs1`QDlh9w$bPM%fxb8*pD^*4eoqeGW#$}LzAWv8{guE`CK
zc+63%ObV1i<A#k$g(b7P3DF>P=2JSDs}`d%sV3}d<Nlq}w#gB-P7`W`Z>g~vzpWo;
zcKUA1Q{Z9O9rG#+ywaOfUs~Xx*S?eOUj{==TwQV>Flwu*eHcfOh^;>@vkg*eBlGpk
zB9jt57@>RJ{7BU_(y=W6{oxCSq9JwJKqk7~ueLiv`?9Rvn+LCDme5F2|3Zz3`WRVc
z{X{m@5gmQERJ*I;2<W*(i^B3Fhkg$o36>c9MM6bH<#kqq;A0nN=TDQstTor@Klm4J
zH<F5HW16zx`Sz<GS8o-Bhoc7<8ZEi-@le-`?i}<Wro+J58Xt7KRC|CZWci3X6@PW!
zqHpQj2q^W+ny^Zh2(7fri<<mLfI)vxJb$v#t#<rQs0M$cu+(*P;03q}AMJ$}!OSux
zBV&6Tj-suO+_CM<rd=yKSlmUda}wFa%HGkw@a0Yso_f@4ePV#tvhA`ix4L_%z(y+B
zF38m0wX3>o(Zu0?B*iete8+{?<!<HX?9Tm#1b27W)U_7d%2*9D$v87w^pw^!F|J|?
zYl_Y8ns2X4PI+Hw1Hwggox>F)nz#ur$yP^x3P(^vk%MI`-jlJeoA(CC8L%56xBVc#
zFOn4szXf_OC$g(+MJ+XTe8~F^xBzP;DnMV9ki5l%9_hy*;!EVowo-V)jF;M^Zy#q2
zbXu$_CiEy?1h|{sB+Pz$xT9bpR)*~T>fT5n8b)Ml{YQ&4GU;ifM11H-ns8fmrGpR5
z=I?HMk5+(R+SEFH3VvyX6m30r*l2X{=}e{y)Y>Gsd6y?GDlaiuI`IYo{Jp-#=-K_m
z>jFNT0CK=IMVF3UDJ!teOzE-;%eknz^Fv$)6$kqohLrtYJ|(d{=-rY;y1MID=mV1A
zI{Piq?~y1qt{1xhWd<Kt)nI)&F4fKP-Q@oJ&YM5{L8s@t<%f86hmRc%`-@wrm+m5E
zt{>TmXOAS-J*P{E46l|~dLByUw|P~vCNaG^YK*ckeU*!97I(V##O3UHY)w)4*ptip
z^5EXB%sX3NN?`(v?`8tzj1%`cPp}+|=%gua3$hC?4&UvVh}da3k{NyfA04Pgy6zB9
zl-g8DYPNr7c%W)Sa7~8PlZ|qHF7KVvZdNc>ut?yhO$We;DO+a`>~GF$&6j58fCc#S
zGf&(Le*YL(Z1!5(uHRgr7mo%M1TnB_>$&>4o!~bM&sTU%rUiQd346N%sSa^>D2f>H
zBK=4EF*KBssD`7d*KRMw*hqIahA#GH{+_5wA11t0sZ?U?)(H$k9@pFjNb{9zjYt`Q
zk3cF`d*#oDhKH*eB3;ssr8|q@L)SR0gzcv%?u8TijgL!G=k+>&kImG(+TDC9rOJMo
zu$J@{WRMXCNbI*Zz|8su+>AT<9*&m`-Q8}^_;vZEU*^sUeRY5EO70tW?g@^vA5peu
zIWNMhxv$-jKrwtYOiN>PEPxVV^b)M2xFWD*{WG6>^?Aw)l~WbNb8p|Ur5*&9_WALB
zCTW!)kg=`GzU#kfj@?rv)I26nutjGrtxjMnILp%X%M=3<%fV0UaDcoz<0Z9~BH4^y
z=)LDG0Qs?dEJtiw`JO0MNhz_qQ<~6q(EJc*Tl%}7*9#xs!IVJsrytu4Zxt?_k_lC3
zm7Hi2ne3NiUQ<~b#E#Pl0iEid!P)iLDIW-2<H$km=fL6HnFj(CSPWH4Tn>FtSgTBi
z+s_CJeXIC8g&fkPs+`XZAG|sE!039oG!3LEi`GTr_r2s&V>h}!KKuCOdmgaSe7fY!
z7w$pUVCptx`o1_NRp#4KS8S)pxsG~HCti_m+p+9`S&hgE+H@h9#`%Nptv@E-YK#*s
zqhV3Y&6%So3w>JY@y*%KkP2yF%Hc;6KohC<ijCOBPzaSnh(ky7+qvoBqPE$g0LMhe
z%U*?D(ZvEhx!XXfp&h;i7g=}BeK7JI$fPWxP-hRP>CflYIx7)ndRSD`k-XFJSMnVr
z*Q=O~LDQ;i>lV|lv#_I=Kovg10z^rYqL4;e?)E2+o+YixsS+ri$njltw?4*M{p9)e
zyhm6Zec4Zm!j-)zjUZu-2g`K_*&)?hFP2NpSo@4yWxvEhvyOHhJ$%NejfUQ}{PZ8-
zLo8?Y6pe^0M9g2gDBn*uQd7J8_L$yroQP>3cR}Ly*MYVRm!HX*WOe5uh6uwdR(}cO
z$WzDyVX3DRFLTMF$K|)j3?Qw@!W8t=!?|B22;<BiGhbh78;w1dS%t(CUxhDbt@2@+
zW|=)0ElH<KI3@X3mbTsuy%Eo;6H?rMO!AhavVu(gFg0<3>5NnhRRYXflEmFbT26x+
zvmlfzXOF!j!usj48?PuzJ@=SjXl6}#HX{Lz2^pZuf1I&;nVh1f0=a+gup;SA2r$>b
zq{l@h;+cG962kaIm%4WKz)o$k3^oAD5X<(gUebEj>b!3+yuXfm`g#m#4=}d14zVWj
z^lK$*Zk7}b10?PGrGI#aE&RC*nWXhsNdYY5#lnD+7?<#aTG>eQity35_2Qa`KeW2p
zO6r_^*=u81;V<}(tf#uYQyc}%0X|*No*&m-eZmDPUoc}zl}ASt&uh8oh3yZ4T-*NL
zEx5XSSaaILIxFP__~oC4%KS<JJ#V`BcR((Z0I=XPw$AlfY>ve0-haeO=1IeR=o2-g
zEMJR-<yO`CrIYo;!84vFF^SBiPU)O$PX>MxCwXt=O9D8T#7pW~b;jR=`P|eWEXyQ^
zMZN%gCqZ}sXr<u(-rg&CyiN-QK&O=huwkw~xpHO_zG=(8SjjQ(Z6Y0QIsMBX6x9oK
zxp8s551@tZJzo<eM&8Vu=C*vgqTKPg@oF)ykJet-Xu8?db+MiM(kZP}Li=blNZtN+
z>VQ#(%Pa)!U(tShM-N~CgHr^ce3)usO-Vypq*`QCfmzRbzfwZpn+mEi+I~@hy1k2C
zauo{Otr->pP_Hv^?DhL$IWZUu3uIC8-D32Vtzji+ZVyr^_%+qW2%Lc4=KA&?)-IgE
z9Y73NU|7%&c?q&o*nAOIlL}T5nko+pQ!?%TyOW;vumBoNDlpQ-PwapXG$iD4=Mjjg
zsdZ-FOoTKI1>rch=BPjn#X9ZFhA@LNo9f!wgEF!(n`5GQp8}3L8u}!Bkh;#u5+jYx
z-mTm=Y)tDRsuxTKyhyWbtUI7nWQhCJZLFazd2O>IdRtz-87JP48~~}mzBL{zQOBM;
z9JviqzGCc3OHCZ`p<}v*pxS6wxP-;|BUP0r8B%7|WPfRwnY@JCfW~t!E<00l_A9mq
zmRDt<)5yY1=>xO*^3Pf{RN1a^xaRSM-@f-feTiaK6>j6=QoX<t^QCpUV_-c#W=vKO
zgx|G!{%rj+S(9jfE(&#B!}||*DdRJK^aul;>O2L?<GM%vo(HxEV&I%Q?5&$nJfGvd
z8_@6=Aj?A-a#fH~vQ>jd&C$1>4=~(^pNdyvpWFqMh^OWG235D;np{?DPo)}9-~L&q
z7e6eJQx)cJ8>|{pEq@f#j}T&(#$5y15w^G@8J_$yAM7ig;`~P|#C4&41OYpt1gRgC
ziciV(L~U(*SAW90h0tYuiM((%)n?x+gq=KAEY5Hpl9N8SmF2~YyYg9YKN7F=EUEw|
z(G=O+qN^|)W$98P{<EALJ{7C;aX3~^UFXYE!u_hdWhPw0&$D&Y)3b3e$zb$rWZYp&
z*Q~HdSt5UWMMdrl^J^EMhxaPpj$ro=rz$rryMBzfRl%J#(s#CPyNG-_(m19N2;WO7
zIt2tjb+oYd@MvUxk07ft$*hIdJ^y)Jy8F6yt6Z%HTD+0HzQ8ypzb{#MQTRSO{(9++
zJTe&A^~@N{q)fJ1%tm1-T|9B|(8bwTc%($Zl@u@uY@qqkI&Hkqch9<d>BCobz9I##
zsQWhOzwCwP(uR1R9RtMb2=NA8`q**i+J-uln`;;~&D67BVVPvP?9+$8B-iU5tNK1M
z46Q%5-2n@6Q4}c(w?PBQKC8=7rA?R@+I;`Vs;YK8YVw9kifUe-bZUOf33m3~Tk+Bt
zkBJQxq(^f7fc|;G%1@nx<si7N)i@iB#DK_!07>Labjju1Wy|)T`x7@7Ryhm^bLXh?
zeXDxD0OZRD0u3i$5xHOIdLL+PV6HWQ($0$u)b=<70oe<&wIC^J3@Y{$NoD|0PmPS-
zuJjM_&G+C&6ykfjVSKu98Rt-!RSj2`bw)a?5&t#>ocYP>-k{JB!YX}=`7LtS+{!V-
zuqEAIvWjT%_t_BVdTiKyS<wBm`E6`Q2_WFQD`42igMe;zU4bn1f!Af4Ju)Xke<@D+
zAv9u)Ya3Z!zrK>a(I>F8dOWy=2c9vvLL#2NO)M%)_xS)E^r{(0IQ&Y;v9lV=Fm!-<
zbt8-<4x9a@rfc)mfX!3scejGglu90d*6iJfI^>lUgyQdA7+#vC)%t>vy}!P{%EF;n
zz)i+fxMr#c|7FEJ$Q=l;sq*~g*%GJtpAS;>&a=ne-n2{JCLS)k4;XLFd*X*e)~jM>
zdyw_10iyoC$pX!xZ)^$3lW-kzi}X@IS^1(B%*Fq*Nwd}sB9J9X+q&kUzwB#7<eizp
zEo}@bgd{Tb_<)1iMtM%koHRgpD4Il=#_DllV3Qc2rcx#`<%onjBtUmOFtdyplU|t<
zNz>ewVUX20nKVvamrN=q%Q#|UZfOH(zDcV-#5=vzp(cIH;(5NA{QO;ga-;UW0-1m(
zcMW|Q2h<-`T_CpX!Hs>0X%LyeA>0To_bWzgE4e|y95ed?Gy57@pDtmt_$iMMD~B<O
zhdM}N@fS!XXBWLVAz@I5e5eB*ga{ImE&-z`;7NR9=_lc)U=sFqp^VHQio}8bc%LB&
z-VB})Fj0g#K8OP*rprHL$04XnAy{E?BqipBd)lPb4Aji0n>_=keR(li3QfjIWPBt0
z|B84{b<vN^nZ?-3l8+vn6JyCCW(B0Rv7xvFSx|!Hm_aLjf;o2eZ?AaRv>ihx1ZML9
zobU$@=2*6Fxhnvx9nu&#Ry|fLqJy{pi)Db|nAm(!h@h;oQw=j)<eH^JF;O>yu*#T1
z3}#AARE3Lg6kMRC$OPQXveO^iaCDXc6pE>N{DU7lmTq^&W@S$D)34|plfNZqK|`#L
zi8zWB&3X|zcp^V}2Ck6xGn>#7*?(_Uib9n1kgu;j{xFC^c$8z=6pkCjK9c}M;9u$^
ztSfx*PXW_dmC%{;(`1Wu<&~6o2q6kbhQ=50t|RIH;lprl$?*-t?#Y#tYexJ)PSbM-
ziDyUlTDiwRU9-5k;$P;bpdVpoV*%P`OgR8rD%NC%|4*u+k+Lk-;l&J|m;T}1u{Uiy
zKa#YQ<yUfd^GGqXiOBi|;AeZLc7C&c>!ntC_*UyxX;?OtV8ORR!v2UN5QEdv?~br-
zmrv7#GGrPWqhB|PvsuewG>fxx-&-X=3t+{yl*9azG=Ib{SKDE%{}@RQQZ%lNIP&NM
zk~BS+0iKb-?SZ66Z9ZqJ6b3jGAK6gm@kyFJl<)Nkn&}^Wzg?yIh?8FxD#`9)>iR%X
zq_EE2SU4A}s>7a!!yG3Penq9iNbgu%^<*02qN?pR;KmCFayVv-6xuCt8KxiJKA2en
zG##1~tVKnY)JWmD_4FXQl}68(<cTp3%xByPU=W<!;gq@}Y3c)Sb9;X$r~8=LDE;a8
z4Wi@)SJgr41q94N7!xPyIsbk-e%|1NBpFN~Tp#`!jRU?AHf2W_``?h6Sr?@@69Z^Z
z)-h2>1Z%mpeN+(mCB+;-;&`9&v?J^`JUjW=JyhwgKXk_ye8FHUHp}ejkLp9#?@QU}
zFHi(;5@fyoS+<FH`ww3xQ0x+gzX(oFS0=-cvkftjgAqCG9VgI0+W+&h9395-dh~u|
z=dW}8wzouDZ{F=X4IQzc=#(&sp%7HX0)?qmHFfkbqzu!es!^;FL(Sq5(RCzr$1fRq
zMuDYr`*YcXGy7M0W&VXnF8ZI^9{@8HZL0yyaA$((vwGwZ<&<%5JD_)Dk@NUbGqtpm
zJ_qBg31+nBZ|0kKKkxzZ1KBHN@^!2i@l?{pX`|FX8#=W`8gAQP#UkC3a;S=~#4_rJ
z@Zcag3~k$hu|IEl8gQWxlmf`wocz$6>@^<<pB>xW+kO>M{pn&%-o`2d{zjR`ltMEo
z2G*)GIv=jC_JhrVYk%BRMGSkiKS;ilpmNKPGHDikv<-JZ{Ic_zhB!3*WSu5LUF~N=
zgsb%>oqu2g2s*y(9Fn5W7|{^A7C(Q8v5(4A?zKS&g{x(3EbxBV_)>ze8Z+!O^Q7?s
zmvGLvDq^x=)y|(h25JQ>Mpwemr73Dk(N#S6G&Hba`QjIVKsK^CHpSYG$ob7<=<|1j
znpk%65;%flIscQ*VG+0MrBSe!`od^ev2=r(7wzIdLLOTCqjfry3H;hI?Zo>zm3jA(
zMg7`p$B`Qg-+vD|{l!Q#mJV9H_9f||+Pbq%8$o<T7I?Ffb<y%&4J=oWGcdY^ugbaj
zpE&po$at{p3DIe7Y1bL-2HcGQiXxmpudsq=^mwlP)T5hrsBhun>YHaBzt?e**vk9c
zP4&3JdPGE`Vh5_my|_V_{nozwx2>7L?U_Ft5ReTBU^w#c@sN}l0&OwFL_Zh|xb^<=
z5tqsyX}`O86+ze?p^o=bVfIaWo1u<)>(UT}5@^Mx0p_Wmh8lflV5p9_q^4%n5VE-s
zx-<J$rCj~@k0E#+KKR0?@Tx%I!Zq5+VHj71B8o2l3dmG`p0DZxN@k>-yh`IgHQ9on
zO1C%YP+zh_J``TryghEU{YxtW4#?{$Q@o9)K<@Fjq$&>+k1#XrM3=6ot?vc^H-L%V
zkkK~5&X0@(UWO-~o?n)6a=lU*>v%C4pp<{=|N3kJ87sY~JV+%(tM$)Nq8*pBXo?e<
z=C7A=T}nF~cbv&LzFot)D%4aU!O1o`027n)6CyY}pZ;xK6eLz>>zPEx&6UhA8z}}%
zoW|y>2{C2km$SlbAUDjnwJQFUs;(xWOU?2=Ag-f~e*_pf=if18q*=(t$0jx9hl&%I
z_|sv-V1nCq<GXbvetv&Im(2cXSv?};##Qljlu<2B?>{4>8c3Zd%I{<OjQIG}8Xs}J
z3JF9Lj{<8G&-l4_=qq<_lg>cz=S<q|k#Kbe(!l#bY7IQ@hl(|B+{twhklND*{ZnIp
zkNV=|i2xKq%_;5P`Eat3$%?9|L%q*X+6aj+NG>7j5qcZos`jxIw$^vcWRHR~3<zx>
z;{NAZKjv_-qhXUd(?l<+g=iz`@t?&(MVjm?*{Z_TBS#weSlk#qL>T-Jxo;2UJ<U$y
zdO+<A>a;P>0(!?D-bhSVCYkTQzYUik;(;+q@jFz?zpDe5lL0!2GW9_*cvl+aRr3j`
z&u?xO=C==IBlDCW;O7MZ=Us*Se)0S3?;4j!b@)UAZcZBnJpJVpt@7IfqMD`o>0=YF
zwA?t+m|ayo4FKwCq5oL90{h$ltFRdf<o*cf_Y1j<Zfd9}f?Dw=ru<N<<*MLa{RaNO
zrHs5jKhk~7t?nTZ=~Ha1CG9<th~If5DS&`9g~*Ypn#CSRrT_W7=A{b+Gj+1VZ6)*l
zY_apFDV-GD3jA!j*OcL%SpWwXaH~gNNfe~s=3e7Y*+BUInOxJPX%mZ!=YOrNWX^%b
zUNd;<+`YrcHG<3PLzV@`K~x)Msgp<6&p-b^8scFWd=Zk<+~Ql;b)t%XgPHBdR>=9E
zAhrlGCVMT(-v533pPhAkWfbxu#TU8=-NE{ZeF^Xk?T(Rn>B!!-GObwgNX3-0V3Xbg
zVU=wTdi635&?Z!`pBkJJZp~5x2Z$p+Ml)uoPeB)2fS7zDfligfStD@i*j&XXq(@T}
z6#}`%4x8?^ho{6!%m1V<yKg!B16S~tZVRjtuHV;pO<T#Gp`b}!mRrBbzgs2O`Coa5
z#6Koyq>qu!2?JOTrw(*V-Alzj1acPip)AkOuvH+DClT{Lh15!|{QZwFTfu)q#^~3?
zsIf+Ye=IB2?LJp?Ov>nk=pT(w6mzMaAy&%i9}tG(<c7qu;%?!mNGvl^eERYT*q1ya
zG_ve(DvE(a%oT#Siy8^8;^sij{3*ctdU%dmkSzH4N&A$iM?b2<2S-X(f&C->^>MPG
zasf;MUWng8>8_w1oE*nq6{qt5tCH)EYHEA42_0#VCZHls0TmDoCG<xXDI!SkDugBw
zjAH0SAyT9&XaFfvq$5pg1VqpWfk2Rs6nPkW5u}52g1?zHzx8I#nqk2|tmNEt?>*=H
z_W8cOcky|+1Lc4#+4d#$!(Kgv>jmuvmhkEuSiPsw^@6;0{N?6b*Rme&Y`jV)1zIR>
zx$kVmeAr*c{DO>OjHVauzo!S*lH@`j4=vcfm!$P}n>$^|%ljq;M33*~a5l9))sh3M
zFn;AO2*BGP3#`(Gm!+o4Y)Dlj12x$Vn?#JpgP_>sG$?btJ}4{K2i4;LAk+cQVDXS4
zahMCGfkeGN-V}zUvTIn%IP&;6@mbLr-&7DtQX?Bu%Gk*D!vu@)4MSL!mgnc~{xz<b
z#+Hf`t%M2Lsvy|dT?AXIb$_|IYcr%u)NIN7k_2s*(LK#-{d2-KZhOhsnJ=w!mT=E#
z=Q0*Pm=A!Rrad@aN9V@U1U0-;wRv&(%34anul&@mCa$Al7C1AUvo25Qgn;>E^SswB
zZVhD7Bd0{GTtWDwiw#<CVE4<i|8CPj7Ck=Sqg~5Kt=u8oF|P4s(Oga_UAi^J=TtlB
z3|ow-&FYf`nk(|S{*=bl0$o$=T-}@`d^ME2(Y8$kImSf^D|&s_H0I1tZ)Q3Hiv+Bx
z`-nd|o%(xtJhc--9m+~~WYf0m-&4;GusU63(L2#6<u$}&KV6-m;%MXAlh2cqro=8x
z2((j0=^23l(pz?c1RkH>eYq*pPxqzwODgDMpZq7y*OeR|Ms>YHwc_CdMb-@yircv;
z?>dw>dVgzq(fXsf6)1g$iY$#g_`2ENifYh|{3zYc%C3RQqU*TUUw!8jtGw5!Hc&WK
z$8c<Xo0*yK&8D885B%s;k)NGXMsyuA-I%?C`H}LqnW`L4{=PK3JL_)L|Bid@&{lfO
znEMIfdXPlV%S1<TB#w$R#0j!IeU|U`DUSc{cwJ(deIu#J{1Rbv!<O@yB;@>6seJ#g
z1pNn=y8D3r<3fh|a2q$3vCWEnxK4G2uoS%(;sKvz?5QtV+XW40<#7Stro6?fS15Da
z$h5fW7p>VZH1j53l+xx=__ed=*rP<{Omq{#o!qQ;?-vEM9{qdwDygQFUh)Ad;r->&
zWTP~Boklpz>$1E>%$qP@gkKZxJVXIXLC>3Q7$G!fy|twg4gHY_P7CHOVl7&Tl(T2W
zAD&**&Rv!KR&?+l=rRYg;EBBXa}tj_HNi4BGPbtnE48U9QqD1R8z?eIEhw2Tl0ZAJ
z97B;M__rvw0JO_Ly~)y3Uo(5!K_Nz%Jie#zI78LctiBk=*P@K>uaq;=8JmiL-5Piv
zt<|Fl+CxRgZ}pgOK8$1_Mf;M>3Te@+vTw-dL!_V6Dnx4H&_=*avEn;M-rDWF7!w^^
zd?^Xu0+3){YXT)hy#yKRq5v<Ma%><4)hI*C_B?8}Y?(UA41+E3M2eFPI~RlEx5L2_
zn<(wnG@v&I>?}Yo(t}l*7q>@!RH1F9e84DPltY7}<bA)tpp5tWsT+UR$lKeyjZ4oa
z|Cc?)@xZ@UYm5^72{*$}YPhsgcmvuduu~<X6%sNm-<~~GuCB?V36lrl@iypDfUNn-
zB8D?6VX>?+j-qj1#cwC2rs@N~1wQo<6&^bzbmmk18}N$kLJxm(adXRfkHeK7ENBM=
zD79f(?g04;_t5UQn9GqdV{$6D0bADYexHt=Z#%a$0zs<rw9G)io+nP6XTSe_2KnV@
zBmbyP0l-)l*QsYfHb3DW=j48JC1f@5Wa(s|xZ@&dB8>ZwW+`o`ZVlf7BAEov>u?MZ
zy7I444gX*Dn;tgU?Vm%A<YY^S=_s*`spZ<q@8c*Uh3dHpBuyHt=Bedg5qamtcHEPv
z;{8mtbLpac!DQ}#ELINJ%Q+?I!9~splnQCi{QGZb)0;Qz*FG>^GcXVV>@$kCFg%<F
zL3n90BNwN<0>7?)cpq>tfsj-L276YIBFj%0>GiMgQm^f}F%X#?y*)LKw==v3pbH>~
z1y_)z{^|bz*DJa&jBa9qRH9RGt~4UA+Q{6JA<-%ygv8EALf_=Tyg%*~?H(5W&w@ZW
z`Bd>sQaa0FM6G#6>D%XV^VbH_1%)d#*==Y5n$mNA@S8F?l<Eqi!Z}e<3VBZ7z@Q#H
z2ifzfvQKaHjullXHf#6iM16!Gufqsguo-B+0)`<WUGW*9>(Q|kz3=J(vN1gbLe#$D
zeztL@8qk=o_PJ;}XO<kDJo>^iXgIcvG;t3DH26)oXLfB-!W>M;sLB%H^v}F9w`rJo
zq)OzB!5ki@aj;B%PxC3;2?DZI8Y)cmA()GIyr;9P>#C(<9!}q#4ajQtXf2cQc?@BI
zCT$}AJn{?B1b_w2AJPqq&m@%xypTJ;sy|IHl}11|K?8uJf3r*tJTOX9%N@Bm%sjyS
z)~=?V=@AA0jd5k`Bv8Pg?9En)$`xKUylZ^!l4L}+vB5bxlgqX`w{am_A^?CMXVEhe
zR7g^Aum9ZHYa_Ny^8*&80M}pFK+@1Y5&~z~NZG6F*Q9?$dprYOlfUbr{HN;w6t1lO
zV#ud){B=VwM;4v2EqJ=4T}R3z(|{U-Sj}v0XCDxpvnoNTjEa;sc;<~EUmux4?ryr_
zN5MWxw0grJ`G`v6Y7+=2+Nqf#>X_^F8a<m|S<6R@a6;Otz^sa$i&^pb@(POm(>dxb
zkA|E+%s0}0D3%Y@KoXDlygfn34dI8!0ak%9%+|qZuN|>fit|#al0<A_a4dyMu5~yg
zHpe>=ID4z}_{TWBJ7O2{ZiH}Fwu&zFq)Y1`zTmJI^-|_e2XZoaTkdmBN4x-W_J!kw
z;7}4`W$YJyD+f}8Aw_fRd35e<8WTllK&jcDyrI1y0;Ky+s%w8BLs{nCfoNYL>yq^8
zonb-}lpsO913^=}vDmSwcq6WwHFOwE$A1lL3xi>7Pi@J6xS8wj2Urc?b8;aTK~Ub8
z0tZ0<)gbIYf(_bxs2^jquAO@X6zzUJ_?7fPYAZ{vJ@r?pXRlBZ6at4s7N@jey321#
z_9fpQBf1HyG8*YBhZe?`|AGfiPW2l8lC3h^kvgdn|BQuDG+@8;s-UI$`1sD;`z#c6
zu>OggOrgG;<me_QCTgU~d%=Hv78!XTtjwmm`3Hsi&M9grC?}Wv<%`*)R;OxjqubWE
zFCI+V0nQVAg$?;vV{h=ese#=~wOqaBIRN0~XDJdmgH#-P_qEWW;?ke$zYbQWv)Gwk
zP$eqXzSa*&!R0BF{bbErsWeJ?#U!nI?yt)E4$egjW()l@SWeLk1SKd;pdSlqVb_30
z_zH;HK*en^`98QF*a7e<`$;&4?4HlQ5ouk8NK8m*DXpF8e614`851)D$g|#SmHIFf
z8wSP;2_)9NS94y;&OR4%CFbTPuYnLCOB*|A=#C)baK$cTi*dn-LAD(><TK!G6$A%V
zy(Bpvy>gLP`u-O~V=r~z&Bt|*H~DWkBpXxD<XZ0%FG!1xKNGk5zDx{kAG<91V~gkg
z^m6$QbH*WKP(Y>v(rv`~a_~I8L%@-)&#x!OO@~c26Gzd+?c&IIwLS6BgBbxnrTgM)
znejlYj=K}Dn)2&#H5790+_|@i?o;>f-J_7n*V@>`&g)RW%srd_n5Uyvd1&yZ%e?L?
zk&=m1u^mr)%cw_(DplVq@W865F+@zW68UWT^XF{NW`AWcE)|I2fAjiYq1wRjSglX9
zg3ji5Mv$l7qXrFWd8jleTC&quPtA6f+0;aPmg7?h1VZ%k22tG=fTxE8p+_46z+NCn
zH}*9k;Enr!;s+5-_d=SiZ?+&a+`smy*<5lgA)c6J90+zR<#F<VI6jtqtRdXxT@yl<
zqgDjWlgLn4Geti=K)MXT;(oGj(@Pl>{d&y3geF{D6WSENh*_n^8;Gs!jSDpIhiLv-
z*$*}j_c@xTZ8@UxxjrznI-Hy_=(~Nu_tkH*T+A_fvi$$ZV3$n%x-&CX@T*v^ixj)-
zTw^43!2zZfDN8G3xtM01jyn&f@ctsUyZYm))M$@x7(tjP1nan9jqL;u!V<cv^`0>m
zoNNKjgs|bP1JGIHUM-v~{sJI&gdv_Jq!jM!`Lx_f#%QsW-h6Xi!tuSQXgh`X-MG>a
z12QPk<&%3}1s(nzlCZ9%31j3GeH(HZ0J#E#Ma9Gb6^W&t=WJ(kxMp_bbMw@<fHZ3#
zw6NBu3(l+V`V?0SZ+gFPw59mA%+-?TAp#B$vCV1L8hKu_VJImcI3N6U`ReM;mH}XB
zkzD2roxqn<BOt3&MqPtWR6;&>833ln3<RHh`u@Ugs5Dw*6>M>!i+1kK(x70P@qoor
zyMXz;YVWV9XH^3zV9C_Ng096iGq@xB$yt41Uv>42FAX^}RE@{}?^qoE;78X535i@l
zPX%zu!;fJdPCt2|zVqYq1X&KKwUY2P2_dOxjdK5rtRE5jO(aYTxvN?xLtU_bxg0bm
z92SuuX&yVc6oQjzBJB1@hLciubz>C<QuX2BFL(3+M45th8+_n4Vzk<w#YQd4BoKy*
z;<Hm?JW-3z9C88;d`;H9;`3INYY5iWi!&7}hezrW{=%(vTXBeQ?Y)c7v_c8mnL^e}
z1*O=pGz<oefSxIocMw*EVUiF|e2wU2U@QQY$#YxI#_M!<zcm#eb-qw3D=Tih+Wh2A
z++(M&ut<VcO{{mRZ69!QX-l!x7S?6A38$RY^q4zOEq7bp4k?dN@RU7(vt{tIq@^rn
z)UK`rJ1HMTqz#<>G7%q}Xze9z>CNtrTwrAvqNGTU$DOWHh2vuA0ZRv`^ufQnsy3;I
z-~JIGKuv#C(qm9IYWjKbgQ+eht&^UEw3*=KpGm23vqwpOy}B~NRyWj#)Ip!9270aL
z4g3oG%_EEQ3r7fNA8IFUi$%R;D{bLF!5r^*8-A}Tj-UFh26Pl$iTM2bd)sjtmas~X
z6Y#Y?<`*jqFuNGQ0q7S!uW_k8N>&CG!*b0ZWwKj)xuvD@0w<uQWZ)?Q1PE7w!9ZDx
zze)QdlZ}l{XTyU<y=&K^{CsnTHcu`8<d$z9(s`ZYlGN4xd7=Zz;VW$NZ*V3K=YZE*
zndbeqfnrx-QSM+Qcw@U@Cq+t1D(>l1=Fu8Yp5@niS_Nf9q@?_cIHLQVu$0<sR}t<T
z+NYv{^Vmo%k({ITg!TShuh6ItdsJLeFxIo8Sp6*>n5*1z3%)WQQ8KD~?@^|PJqg?*
z`-lO>R!w`WZ*yt+3`yh}y!j|ST2^kal$)J>gx+7d%c}C$6L12RHTnIy)}?_lU)EAC
zF_QMN&*lhyrsRmWW_tYNKxbW0ysl(3+z1=Y#Cg$K48a1GoeXGV`c3B)<<F>gVJm8h
z6>m^w!MP{;lL}blPN;su=D3bXpDJOs_~VQaLcU*7L1lE-ZDI4w>9`D+wqDfTQ=7pa
zya916SBaIkWdBB)J!JS#<IWgh?Wq@&BM85j@zf8R%J{FbHqWHd1!tbZX54A(tGdg}
zvUP@1`vgBWXDC4E_jmsF587510^tC$#Y3rA>>{?XxSANh`rA01sw=_;VR6HK-#l4`
zCvD^e?9!DgcZ8o?b?0dV!V(WczMbIr(mE@3vjVY&$1TfNMIii4lQ-WZwxV#{sUtS9
z#=>=d0cDTCID~u$!7e$I=u`Z6AL=Q4(c|rmg{j~-N_Drii@{BguZ1ujFvFSWPO5Eb
zZbtVz8zOu1EJlyaGn^1PVWHGRkW!37*fnsz&8zH4T5f3$n=~_$u9|2`Ysn0Iy7sA-
zAaN;ug?@tiB=s_;A2UW+8t)KCKP_BW#L^Z6MXz;67MT}iM!$M`pmksO0kSRfI{#97
z|6thZ2@l=olIlr%B>GZ{Ha@7L4oZX?4<)HV(b=T;P|q-f*Y;ahBVjUsW_4CHVlJ4+
z89`T=LqjBP(BxX^H}mYMjBaz@A^OJOF(>^zbI0u_f0HrM@WI-Vy4FBzVqhG5@#6uG
z@i2k}=Ny!9QT=<mEp!Qj)h)C#YPUwb>n5Z`IJ)Fp=%4OfA%20PdSRtq-bBf}@{Ji|
zH=J;&vMlWVk#-;8p!%+rAew4D>zYgE6#=)wGtkOn3~nPzkH18)?iQ#Xem7UlV4ABj
z={|xox7e~B7*6p{73$Bs&*4^L8Zy!{JP##6)h1y2FujLf4OS&JeP8v{aJCY960ARF
zFq+71igO`0S>WiU$t+sb!64kMO;Mpsv9?c<Qd(piha8L(INGd#^ufOJJWy3o9ST79
zcp-bbQ9ZQjsQRAzN$6NuCbnu$v4a&n<3ArDKNhfhLlR3KbJHjYlI^4%)C+ITe)_;a
z)4zU%O59_NP`BBcQSl4S$8I}LH+4u_JgXW*F)4@J6TQy%5(%ardM{EVh7ZX%i86Fx
zB<Z5Zr?CrDpWPFbtTd2qfgC<&x69W)kI-Y{(@`URk`_xRjAN7djt(GGJ}r$s_1};w
z|6Vfb6Zer|G^S6cITybD>_ZxqNZr=Gq(Z4tZiH#A)73+QA`~L$pfFDt1U<DS)L3>e
zR@>T(*hh$CGH#xMRhhvWndTURxYX>4GVcmlCk((Eezh=#Wv@|!kBUIk#^`o1%~ZB)
zsdJMwqdJXt6-`jbDg3B7jRlmAdfxnp9dikdA5}j_E60AtYB+2iEizf`Z9TnN(Xu^F
z-cm)HCU5LhWqyk-v1c1Uvm6**Lyp|`KtI?aqd3=uq&0~$@(SR&ZW6%T-><sKe`cD@
z`<K%22QkWCj;!WAj%lEn`rj;l|FO&e;XA3v^>PFavpu_f4-BDQ^q3Zypfb549A8?4
z{cARY)o#4|+b=Up6qE!3`jV>Yd?z>>0Z61mW}vAF>e_`GbrIj#VS(!7SJv^k!)pBu
zPr}_o+$?V7Aws@Ee-92zb4S!u*6&JrQV1SbDp$3P*b2w(o}QTiyma8j31D-paLxYt
zHSq*S6-bi{!Wx3`JBz%Gl<7T*KC#nF2*CD-o-l;0P>n(l!`Lp`H;VPL&rv^x^w5Q{
zPwxX^v_6>2M=%$Rdblm)#drGdK7hlbOr`=9jWArx?sT;QMl2oG|6!g8*|?PV>l7Mv
z{6pZuPf#n1`a=_p2Yrf`-_mVagHABV{~a40`B8A7l>>Blbf!Ycc8J+hQyahJrjb8u
z*T`Qn=H?_zPy=;_c6sbJq@|-;_4gzGv^k6c<SU5+)|3KJmp5%J@z@%|suK%E@eg<4
z=;WW!#0x1PMkknezh7{2{{xKd@eqsS_dy&Xe^pJlgtoa8p_aH2JxVd%0P`60qHX>c
zO_@wpCbqx7`>xrY$*@R0&U_^+VO@>44C5x<XOiDw`!k^4G(W=Bsa%1f6fhQXpj)AH
zpm|3%1QAmvTRiC(`^K_>hENP`Lr9U^=Y>gJl<7&6?W(&3Ie?cL(-bFpJh)FP1~{hX
zfT4{jrTBj{JEgR?fGnb&(uAQbvH#6I`rnsY`Y*O0Z0}YR@wD^678`O+=el;8rqiSU
E0Am;H3IG5A
literal 0
HcmV?d00001
diff --git a/docs/source/_static/remove_background/bad_learning_curves.png b/docs/source/_static/remove_background/bad_learning_curves.png
new file mode 100644
index 0000000000000000000000000000000000000000..356a8d1c6b45f2017c885017be80d8f424a60c7a
GIT binary patch
literal 214054
zcmcHh1zQ|lu!ajqaS0Y20t{}!-F4955?q734iX?}kl^m_?iL`pyF+jo+;!k|-u><G
z>_2e&x?rGvb+5Im>Z!XX^oxQNDiQ(Gt5>g3!P4T&uU^6PzIyfM0s#(Klan{~75G53
zm)3H6^$HpD-_Pr*RiCF<uRgs3i;JkbryZ@rYhr8Q!u^ekBqvAuD6C|?Ad3x&()Ep4
z1@-9}&11)ED$ig?=;lgBa8&>KX#A70LA-ZulIF&LAzJUocg)ny)G)DyX5e;r1Mx&*
zIeF3WB+coxdg2@VC-VQ692CC@=l}W3q{c_g4Vy4X{J&+_unQ*rKY!v!Q$`_S{NKft
z6U7<x|E^Ff=n<y=zbi0IdH+A!#duUqQe6Dq4=M3a?CdDf(b3R#G&Hn<jL|h~2o<kn
zKz|fz>3Vv)Q`gc^7(@VHw5J>6>WQbud_0nfhiIhZDJu&X0Tne+g^7A$ajBq(I%{O}
zpmq=XXlG{*FV#=PZ~H%u78g(c9w?dXffY4Pc%L0YV=5}zW54mlqp972LSw<s^6#pP
z(bI#7%}#rM<)}`AH8mjvG)GI#SdiznfS2b-^w+niV$Yne`xCD-y)Hr*Z$<)<aVLC*
zeHWG%j4rlGm6{#%%Q_BYv9WTyYD{HWEG|HGR`Vp85f39}U1PyfB*xu8&KEko38@rQ
z!sp7r7A~7-huaNIWOMsG;dzxj7(84Z@EeH-F*FJ3(KeWeh75qH4$S1_rt&oc!7bv5
zrbC5~9eQhluTkK?C_QeyGa&tb_$uM<?$PvnIVflT_^H|H+hPsn(UCO<4o*Z;(zkjE
zb#?WDOd%nprsizvX!38<55C`4@>rrYK7tf6=ASzV+C6VY>LlOFYJT#4_NuC@`u+3o
zUz6isvWd*bFVEi4+YA9nvh_Q-_wP5VPL;0n?gnG~@g2?eJ00E<a+U>~lPR*>%%Pwy
z)^^SvPkw7W=q!Ea_J-D3PV9ZZCkQv(Fy!{w<gU<bja+WDE4@Ak`^>rQUeOW>_`JDs
z-G%(u8RLk6_HDbXfm;%vm{`GFfgrHkbxHicMjIV1u*gkOhEY$ilHGfH3^%$ZMK?k%
zd|NV4Cy6KvnFC!a+A0l^Luqmm(9wHh_r_=DxXs+YB!>`u`+$QJSc}V4L|Sb&YZ*>*
z`927haLhQdccs&pRHaZsMPzJEz%xs?)z#j03>6vqlil;v;1Yf?8tf;TSdb{3wEOMq
zt$__WHw`*1D8$Uh;ywLvxEA^2b|}`}RqeP^wzZaOktT*jdz;?j&A{`}))5M!EPYdJ
zYcc4(Q%!w?Uur6R8ABww`B)l?qnQIIXSLR(`kdGIe=V;380$<+D>udQ2c!$UUQ?ij
zGBhdzgXh;5dgOj1c?HO8zNomKZsG?Looox>xIit<KWjIOnp(#DRoVN~jm+Jmq93rZ
za)Ug0+zo^}@C4u#{WG=>EBkZB5$)LR*1`+rK3;tK`&uEBFIXos()|T?tJLaxjP2`g
zp{ZSz$z{-KAtWYw&ybMt(ft-?`)#6%EbX%C_aJSdR^sXAravi_V)y=}{_pA9Dy>BM
ze?uou2Y4HKpTYAVfIFu*j$u*~P2OGpd$_a8$s?kWUKRVg>A9zn=4Lg<d7YDdXaV#g
zetZ^(#MTyWS*t>a$32lQ(2VgbISmNUYQ9`z)zd{ELiMSgH#!=_QrAQ0;`iVnt=eII
zO-)c>Lvr+@9cy{HQKLRV!yG4Pr8h)8V`U6VEg=1={#P*kC!zm_awX#bJi>@zU|5J_
zOPP6+r*6J3B&`$-pLpK;Rc^m{wMb4r#BKyKWg%(CC|57%h0az)c*CP_UKOh}x!p2_
z;m{XQtvueYH$hIB+S<Z$Jg=c8wnLpSM1MM9j}8x47In6(NZyZE9Gd&F=Hu1A>v}{&
zK14n~J~ZlO+?J}9LA)-9WE!lt&f^&ZrE`$01(v2(NS0P@L6Kf=5FLA`Q@;X-wOp6~
zd!<}6`G9+q40v>sPx{YKMH-bE1XMc%8$EZBbwQ%bYl-Cso5@x;y8sW9sDWQ}QjtXd
zr<2AZ;UU5tq@*)NTJ(biL>e$ywf8co&edhlXew7Uy#`Ygj=k9Zh3RCTg;Ue^V7fs@
z(f+ri<yKtC?QuY7XD3}wCbOL$HiIexDk1xqNA7x)MRSGBE}<8q-_aCOYK-BHC2fHO
zvsPKwi$8qdDr5=<{KRE08YjyT@Q$=uZrGU0mkkMx49c(P;H|UE+@Jdq6u9-*?=83c
z*{2+1qZtz2>t(xv48~MXAzmu&<5(z%Rmqn5NIK-U`m*wPq1GJTk%upL=xh9Bwz0X%
zc*4~g1()4=Une|z_wByYN{b6J4xOraeTkY<qmMxBA)CZlAP#*(DU6&%&?`iqkgMnt
zk)?W_c(H>@KBZ{gZlxs<t3Qg29rP}}wGEkG7P+2zhkvEfj+CLi?6cJ_gSV1bu?G2i
z=qkju)R=krmTohU)pR&Zul?bS&Zs5Zb))+|5uYn|?@t`5S<6|JAdMLBDz|alrsn2c
zv+^a&&U8LEv{tv%ZlM<v(dwMui7X%Q8LNrk-vlV`Gl&G85x8u1q>g#7Q5q<;8}1ia
z?(TJk^JU_d8gzKR8L@NQ8wk0cECrToROE>EB7K{!$)9v$=*X3d+W6fC<90pb4#Q=R
zG?P#HTV5h)*~$qF&|iNIuo=}+?Aou@!y?m(*ow1_jm>|$Ep_BVZw7uzF1sE5rW6uV
zE!Td3eaHb^3P+yF_;{%vis%ROjbl)H<A+YV`M2xsLZcgLg>GkLlw_p%YKsf%ZnMKB
z)iaO#E2${rexOZXij^s6%9P&DTF!j;W*rl1&Nx|bM5k6vSM8T`_wdNJh@}Dx*701O
zc5=Hu;VGnXBO=Sns$Ld{Nk**A>9=BS`5yDSo-9Q61V?>}`iY#vWmPbiZhhk#I@aiN
z$k~SA8QCq-_IFEVcH3OL@x+aL&rz}-I=cbg%<)x>96q&D?;(>yQ@t*RyM1#%dn^5{
zceLC{s+hswl?RT+G|3V2d60vx^0*%1G3m91hz25MYu8(TyK*(=PQBkiR8rqO++oD+
z)oF=3+B$9IV)K3~DnvgsSQX_La>1MwVV<#aeSgtaz4>)ux{ghXgM8>uQl4zV|4dS$
zAv8?L5z)af_z{(fZs@Oms?=(c<O+TNddloR5n7Yw`+N_7O~~P&UXSW?w*f&RV5jL)
z_+>CVku6qbvdD_f3=MR#n$K0GwY%0k=B;<w-e;+^TK-6Sa~gF+%=@>w;U<tBmDgpz
z$X{#%F3d-(4nZs!RZLtp&Ql)FJ#@07ziW*|2*wVd&g;ga4Nr06`bnYuy_PMRn;Xs#
ztLXyN3Z2%VUHMd_X;NRj5&cqi#yb+B$dp;H&yX7;cAF(^zbO5=Dfdb%4F;{J3sz6)
zj0FlI=erk@^V!OA^0Au+-#y|dZc`b5xCZ{ROx7K#@JA>;%&NuDdecW8ape!H`oLS5
z0&t^$4yGOE%(H`xr(igiQ$-jJ){8NlSZdI46cT<s$nEJoGvA3m=RJdRnt*SI*TcM|
z&lQH*XtJ?mviir52*{}N&h5&+xjyZlO-?TzgBH-4vQT~BN7VNhdm4Tf+Afx*aEmFk
zrS>OJ9^tTy;djev>$5niTy`RXNIMZsY-bx3UXRD(BhLraFe`^~4rWC~2JJc;3H*){
zAL+3_0aR<fKTCz%G;5p@;e5CeqFE-Je2#e~eb0G(beolh%Dwg~blP)nn%{kc&Xq<D
z?9;%^BOn!_EV)u!m*awpAHw4dw(bzOSk9O2W<5IH&oKjnwkkWH!${g8xytt!N&FtH
zRPq_&#I3Fe^SR4~FIKZiPQEEW+dOnp6MP>}blYCmAa{EcWFskps4#f1qL-6)cZ_hq
z063!^=qw(OHFpZ9jX#ra3p5z}=V3os$Y3I?OR{rWF;hV5<r!Axvn2_06=}6GZId;c
z-eq(zRi@-eMN$^~0!-1&lUpA$PjnK<VxC()acs03m6O*T<|%mH_CxN4#<?CokMs$0
zRz2@5IM|Qv{J<df&dXFxU~_t4N5MPGerSAm+2?zS>XXK)Cgi(vsaTPV?4y@e0q&EY
zfJ`zYvsmxjyFC9M;Mq{hdR3v@nqXktTyT3b{dG1!w3E3v3ti~Xk8>>595Jt&HM>6F
zfY5dCLSHGj$fn(~2w-wI{u8pneRkSeq~FK{3~$4#ci_*>q<BR)xhE~8nL0m*5|}G%
zj(i`lxdI-vVrITHNgJBf*7t-bv;B3OL{~AF<61~q<#WqXNauyFJC>(EPFs8NLfK$G
zp)Xvf8#xl+ZWZ{hCj$0y2Ck@6Dia&uaRyuYZVHhl{A|@J>by9=ZGBGRWMPVpxFVdE
z+Ss_T3Gu-+yz-_u|Jvkig*Lt+)aE}5B6k8SEUTR;0>&{n#@V!KOs!O%1<du6^1$bM
zl*`^>to-CFCHQkbxaDxZvWNL+Vvzu^?;3wBxaFatG#{tIb}{?`mB3^Gw2EJAR=h{q
z6O0b{EwRqEsxc?acAe$VAMvbOS;2`lB_lB%(-ie=K|0OWYm8($Twx%{O1qPkH|Q%N
zp{7=QnUJA5;c|}@fo^{!Nnm<9@e;drku8e;yY}x+Y2SyVVyU-&RvXWrVty#m=oB1^
zqsu4pI$f)U_eo^Vg;cJEgoKDkksvZ#v_o#YPY>sc;TK&FXGB&#uYb+Ys&UA0*{&4%
zr;TUy!f9GZtb{CDy~S~iozIa#uFp_s)XtJiW_8EFe&^=q2CM7nkea<95%ef~gLHj3
zFV%dpGxHfN<rQVrukUNP8ia~~hzS0mxzytF6UkA#`wfSss#I?nj#7&%?KgIoCWoz$
zzzZuTcr)ivBbHA#-)X0hvmN(I?|8}OOM{M9Q)@HX=g(i>v)>%A2e)~stUU@7e4YB<
z40#bubL7s^gLxKxE#%CAZ~qeUsu?kM{NnTH5A+)4Q|oOUmNN(m2?-^BjD8Ho0YHFJ
zuC}ER6kJm;`(VbU!DP-+HWNwAcYeF>TRJaD^!b(RlF`}Z<Rpz!PE$!HMq-<w@ALRV
zhZDTzY*98rT_g{&19XP=sUF6RBX_V^3x9Je^y`xW1v;1K)qaJ)2ijsyG5lM#5>>xa
zSrWhyE5CMlA>^8<Hl{Aqs?{cC2Zz<yM<DIW<~m)@l@A<m|KPM)CU11Iy(SKw{PVj9
zW;*c|8nLjn^of;$AOId!^-{G|`#lc5x@ZQ!hyCa}X~PoiY<p1U+841qkyhnnZy2u1
zI-|#g{ml`+i=aXhw-U!xXC#*J{G=&#$iof;a5yVY1~0Sr0KXjl&0ErvOFhfFfnRd)
zx2OGn0r1rpmuA<xcx+~6BDs^@-OB4#337k$tm~4y9QDRiIFM~^-%r$-BFA0bND2$T
zay?$k?^q`F<wc68S5v6!%=eop)lhtDHvNf|F)S9XU2oMxsYpV^=R-s|XZIU&>MCJx
zACe--kv!0_v_y1azqwkGKI<`2?_@qxq(>Y@BGlyvd#rL>_bJnR<ntHvueLpIdFRJ+
zFDQGQK3l4>@lbC?qn(G~RiYN4F`0wV<c98X*++#q;N~&$8@<|SM+WEotXTWKi{*S^
zaB!>T$kF|y-ILp_zChmN?QfCUFA;WkAsb12z8#+R;?+4tTD2c#5}EQlqJTJVV_!+1
z-{%R`<weM8E!=2J`iY+oQ)2tIN^TI1R!yF#Vh6qN>0$tn(_YDvMCs<cRTshMZDVi8
zfV5InLQenIudm~;&ihGyvm}N;hTeo!%?tjnX1SOUk!sY~@wCS9CAE5?ee}60OXhZa
z_V*Olx-W11d$Qqo`21{dKaeeuK!8HZiH+=Y|9&nZq3A>fhr?>dztCJ{rNapqa&y#I
zy+u87-F{p3e7io`;$juMEG}MMC!%mruhsR3$zdu1Y;kzFqr2X}eyS?D;-VpVnk~sF
z6aG&4t=Y5R!1q@P4rHD^(&%mEdV0_eSLtsE<P^DAr1m3C)uB`eyUruU=*|TiF%_Wh
zgBa)hbQ3zT+f@<^uxaQmtjSM71>8bEU#9fpGYf9f_kq<f#yaCNYUR8zoX0vRGJUoD
zl}CSkH@Ve!@WOZ;p7$pG_3PV7z}F1`t#-VH!;{0}0&9$SzhU+dIESGN;mJ=C3wX+v
z#5i-=F6H?s%Etg=$2-qn#HIB%CwK~>`x%3jCU6Vzu*(&v%ICA9Cqb*HoiBRtd8bqI
zTFs8Nff*D2=2-012Zkj-mCo-HHVD%7n7iE3(8rgbivM(S#8V4w&}}tr=M5?bPr+^;
zPp`o(0+5>->VpFIiDJG)dCG$!HWpBH-tsl<cJ>AMm(y-em2{syKCe{-A}?bVb8;I?
zmCp@!yRYmlc`2j<YrJ`E0j(w_CB3^m7+{g31K-Y-r$Qkn3}B5h4ls=(1K91Z%J4y$
z8i?|82hRKyEWW)dN?f|{Ienkx77#+CQp2t!{L+2zv``Oed{}K&wP@Ov)%>GR@HU=a
z^$XYEzMA#Cje3KwH`kCJ&+97ZMb?0z!SZFdiCQVQ*-yl*CJ5Sf-z@B>TMW10tk)ZK
zVS@I5h2>ItXASPXAFd=GZ;nw2d<&$1bxrq(ncfn#YQ++^BVth(^+od^B@%uc{4tGB
z9uX<@{QFl|K=pBDXO+vLnKPs_)n_e!2~H%DDZTnGu~*Vgxk9Jee}#!zXnjXEu?ljk
zha(q9_y9N*pd_gpte4ibi$8IQ>t5ylX=9N*EH)49ZMSO>xYO|@7UInC2f5?No#8~o
z?T~2bbpoSy5$S;%ZG@lK{l(@_=8iA;euI=+V>SHf=;%l5xXWMSB`uVVvbyy7+=>V<
zS#nXOh?0bSN#M{4BZm`Tdh9x3VLq@2{I*VmjqwteIBE?hxo3x(=Y@R~Voo9;%I$mh
zeeTN9_fZmw+eR^`)GB8)ANhTA>|30buc5>&m&QH$sdOAW+2?6qf2&q0H(Ws}#{{wL
zYwMf|Dw#zYxCiWS9W$#{ol|T+Q2^c|tNe+a!2H$h!&Nj~_iqtWqsdLmJCA3ttNlr~
zNfaHGo1^962N$Tp^EfifZl~8hYm4i#KieL2QFz~--xx480Vb__F+S7x{C&p^l~~C0
zE5ZFaR3M_6<4TzlCBUFoOzRHQ*T}L-<z!A|e?Sh$=ldumml3YZ=e9WytpD0*i#P{b
z{v`6%gZpM$(J2D08H8R82yT@=r^~JAyCcaqXU`iw!IQb>=q~(ls00Z5`uYL4l#i|l
zG5OHj3cybGd9PSe)^(|_3Qk9(=y=?H5*ag#AT+*7<2hqcm40~+*`TO0*m&c`eqkKH
zP~|cFF}}QvLaVORf-nyAT|zLLx<iF7Fi1oT6^I><cSqStgY(`62Nr)S)H(FMGmM$E
zwiztbZD42+;WZn<3xDoD<Hfi-pmDMJ6^sLWxb9bb=^Gr%qaNDCM0V_>d_CTd&$57D
z<*6PcGS+0W)Y{vzvo9%b_krpFaACkLa#QbPos;v1LAvBy+fm=EGbtow)cKY3OuUa3
zB~+&KZ^xs_DQ`(=m6c<cGG7D_E68%n4*qrp5Lof?$-BbCkGie<Rw5#zPQMfM%4>$4
zW|*yac&9Ww`A*H0X{pG^GPVZnHm>_{EGPn#v+8McOOxgEO55VLU>dtcOv;$xSI*Hz
zr@e56TF9+!M5&b?$7)iw@pWbYH};f_cF*GZpC1`Z5Tioad0(#Zr#W12ZY*1XFbd?Z
z-&o$@dXr*0oS3{a2WWA<K2-B<DV@&`!`sfr{ShkDfYT>`%yqtfz$OobVY7#9=3}3E
z?7Y4<SgT!PGwbPWxg99i?8y$bKi<wd?iBjIJWCcvMn+Bo1L@I{fM;&rZjIgjWdK9}
zS}zonlAT&1O>WkeLAOObmhr3W^(8s?!nQNuR&j7H9osCGB@Z6R#53g=985P>b{&d{
zFtekf@vt|zx~7Z^KCWy(L2ggf_QKMpYsTU<I}O{l-DXSFI5aBsICu)vj1?3Vsx$K~
zrm@Q@IwsES<$(}rwq|!U)%?s>)j8bk=|||p<@020A>3)DPP4Kv+DMw=QyRb5T@{tx
z+QR8tWTz`$o7>}wfP5^K{0!uBZ(L(tig<Ay&}Y?82h9aspmmcwih%#&tFu$P|1S6k
zbi5tPQBvwOndG4FQ{2(Wx|-D$pwjjt8Sx-DY85+hE38_g3YMQQb}|n*#%{SzWYkc&
zT=#v!PuCwiCK2!~I*>`EE%$ilAJ}(xERS(wetu54f0-|c@NCySH+ndfaU8go+>PCG
z-x<P*b*l5;Bj(WE8Hz{IBV~2|cH}S;Pd|CwFiX<kNXVsd)?r=SM*a9ZPBxLL*n*#*
zpPkR`ggr^-`DA}=7U($3+x43px&0|<#C#<Og8OT_#!qFF-zpyTRSu&MJ~HV1I`Rg`
z!ByrnevACqb<q?hzFr(0TE8XalcOe313RSNY3?}k?}f4=>H;yrt?$DDbuOXz)xHu8
zCbP$+%N~^}INe!jg+3GoEW1aGwUUVr7u&S6OtA;sf|^UJ4y$c0U#E+dD6V~gi~;-W
z*XSr~i8ubWfLr){hCzYJ@N##KmC_xEeAq}?-B4PA%8<E=*PLk!>wGCbu$-;P=W*JV
z^DpMKUMcFN4qHY`@Yw0la8ope>gk+#Z#A%tWeQd9Mvv3jpc;2~_sE~D|3Sg~Ed1hS
z2>9QK1!;obDIK~jrqH-|r^B^y9^=c)e6z>USu6F!<%-bIW!Lx@*;IxOkIP-?xR!-%
zFPF=b>a|aY*L}EP?ypC!)#izIsVMQ~sXpuF2Ho{J&kuOt99x=fH14C##R_yf9mHw@
zH752{c3D0_X{t5vfvGn>rlB;owbg98KpyUNm}d#dP{0IZVOScD%M~p*l27;f<>bd8
zk4|8CN#4d1$$Gq$+I4`>o}Xcl9Pg2^Cn_lzfzvwxj!Z}Ac?1cYq%TaXfnmcV*Qyn|
z)G+14Y6i$nmc*>9zx@I*<s9}W$XtkA#)UQ?VYllXJWhQXT)=i&`*Q4PB_*hbhK?iJ
zActmS=Fa*#tnIHktfAL;?(S6RrD&^i?y(CV+7>38Vd&xM*}*?|u;N-jt9K|bP#R|y
zcG4DYPH_%?Y?V!{fWn^ck&CR@lsv%`y@T_^kM#P?b_W{X%AR*tz6uYX9zxfEd=5xP
z=@%6risR<i(#Rpz+qYvtK&N7&ioH9Q{!=o7AWLod<DLGY`rY3E^v_1yQkDZmla>1V
zb*B@9`+JEf3ix8{cxvc>Gks_HVl3!5l)^0*s>o#G7%JB8cM|olRHT{vkoU$**CE#@
zHg#@Mo|Wcy-U+rwd5MNkdOi+=U)XbO+glz|aQdW=!Fy>6+&)lj{|q4w6>0Sx^Xvto
z*TadGzin+47xLm5@FBz<-j9((3@AvRh6$HPQ4xeQFQ!EiFR;ppNej?h`T1d$$m`L=
z%Fd$%6R8Hs9&zj`SJC<x*vt7=Kf+IUEnlmipldpcFP@n)b7i#g^qu)NY^F`=KXJzk
zANTM%t=+>4eDC(FOGe4sk4(662X*381UDb93F#<4=Of|Lg+Mth3DRXP>*MHjd0p?h
zDpQ#}6GC=JzZMY?^+`*bGxaZwm=D&>HGA#|r3n%*;{(FA9#!zSgOy?|BByQXUUmGM
z>ITjfhe|eKKK9vWCv%7Mr}w$j(&c0-^`amM?5_~~>?vzpQh-=DnXCBTnC`E64$eof
z^QlBJDw$Ed!KaU;zE{;H^f5#+i>-Dqua7++v%V>9&1-&xSPk7CuH65$6m$^!^lGoJ
zN4<TyzO>^;!3*|)`bB_ZD%Co3hMP6s@6h+Uq@!;r%P&CgqeNUr@6K>P8V1$ktaw~e
z94tF5|8UWwMS6xF)XqZz3s7zo#qYaW>=N>D{a{MY$_=gE8=pO^a$NU1efB<D?<Bq4
z9qp|$?3tk<nM$o5ebFW?_n@L!4JYOrGa>E(oR1<UhP=JexMCHo@Cv1g9^D9M3Y)H~
z0qKzx0oj*~6{(l|J+pU#?@E<^m^g>s<>?h%`Nu#q@}$Shm^&Y)wyUCppJ7rFte|C9
z+avCP=n!WUQ0zbR=sTJ$Aa%_3Etsnv!qsqFIupb0!pwBr(d^XDu*`6bv9TEgA5W*j
z$s2W5BCC+z)3lx8ys_B}jdwj>ro;=CM!{!AN#}J|?UYsI#24kN_+X>Rd(d0_D3ISR
zDt7JkE*NBoKUK^8EG_<}#~AI4n6C_>8nmvDvYn?t62H(BHndOMm!UKOI|4L<ndTa^
z(bFb_oCqVCZ<<FpQc@CI%{R#s$^^5Oc8(x&vfRETf>4AdB~KFs4g9X%f#~h1nh)#d
znLa65=Ba790+1_J0aa1~@37JrSYnCk@aJryXKu<aX-P_%IN{bCV%0~Z$FX#{aRQU8
ziq3}9Fa0$#J2H5Do^t*>Lu^4Eu!rJP8|^{NHSfmK4QX+0tiC9z=BfPIYbi-+XluHI
zM<R`1C>9kSaj5jzvCzvCRyM{)33Wj@8g^!#x6sp>n2Ugb$8LBDJ(1h<4!!y(5Wk#&
z)vxIEO?~=o&ok|8iJ!Z0l{OE?xljzIekIALFh0!aNEC{-{bNa}zJ4U3)mg^XQo<H<
zd)7{h8P4evB?p&OIMM4Z<GJL_tEJ4(hW?Z0@l(6cu-kfQn@6K8?sX%DpoQ<S{n(V4
zp6Bk%&2pGV-A)|N!T1#R93%e8*6A%Bk`IUo{!?I@7qq|ygUfae?SEMS@wh>7fLu^;
zGhrb~phEjnJ&XIm2X?Eq6_Z~q`W;VHq(efG2UVOR+}V=*A~Juu=xn9DqSkj(J7)4>
z8A6_SAJE?z2T-V$;xS{Q9gq6GYl<PzR>GZNk?@zDfa427Ar4F0$K!BT)KPxixj9}&
zKt@)n9J(3&(#p!7!#-ewhtZ62G*cXFi8oidF8p30mBT9IFO?iN?Pg%YXo~Uhjg4lL
zJsIcAJ@r}%xy0^ajzet0*&NzNv~s^<nyS;l8Lyf=wkGa-WY6EdJdG{IExE+d1DBe*
zD!IC|w%pv@hfe4u2K7F5&w_Be)3{nh9AKDB$(%N0HmpT^L>;c7IM-0v<7>--kH*(r
zBl({4Td(_?w&C12{~$saWOK<_`9*CKQRv@Az!`2-qk{m;I&&tio<K{ewcgwNzBz*3
zK8sknU+=L!?oxn*UqC<}4bjvt5jA1SuxFD5op!=kA2J%9?I%a8+fqj$Lt74-ppD>A
zv9yeuhCyy;CnU}X-v(`lrX`AuXKd__E(9~$sj<SUqzS&?U+k!e`H)6k{@xl(Uxs>N
z+i}@iJL?&f2%`J}Lb4e?c(fCZ2qk0+^s#O3;VKmtf(T3Q^%f5m4$sro_TcaE)raLO
z{m#dv&3(9x>ezeC5Bc|+AjvmrlvvHXgAGGx4UUrX$-M8GJajO1as$)b1(;rN`BrZ0
z4-u`pKPWNSWKFTfw~FB&d8ygx$DS~0CO|zhN%#;7mx!psX=*1Z4{VpNf?M1oA|sW>
zbl~{izth;rH)zD?(Q^n<f;B8x$N*YktIwy{|6M4HdCW;$^W_z|W${QC2;76cqe?)l
zQ~AS1WCpRALg3q0yH2)bnnX-m@KpRw0~2_@)}V{Ml8-SaHjlo?7aHSyTrXA%d?16^
zfKw;P#8gyoh>g;6fLq*b*CnOTkY~pL_CgKZvIk>2h;@z-cng+u1Y{4iLMAX*zG3e+
z=s*zBvjRAQ#~&)9LJ(NP-)>wlkN=*o0@#E`z^j^>_qABLyN%6>@<1V#SH4~sXoE)K
zg&H6$oQ<}CbS+v703={!pM?%DEn_4-58``k8=tpa=$7fcYCF`jxfag=7Fmty#l*Wl
zzm;vl)1phx<th|gCAX(ZeDa?YzcPzKDF?EiS>{nB&_30^>G5=G^tAMJwe%l({=%UV
zR?J-~_BO&3w1p};T+PZaBGh*hpO2iDOM<LbRjsoh^9a5SS{N)+V}8SF(WZ^@bHCop
zjMd4Q_(SWsR_4KGt=al903KbPD<G!W_2SAxE|cZ$V$EV{_my&zkS(fmGgz<2nYJ&I
zIKV^brEIsg)`Jbm0iyg=lq&ukhOM3W5KfytEE%6dr=ZX*H%9yVR~iqoS{55S!+GCL
zsZnw`%HCZU(Q0)pJY3708*pft0_3<qZvnr!6)SP(p?G&I*PpAqZ%SqsnO3#vSIOZg
zbh|MiBMoiU#4qNEA+pq`G|ntYsQ#_34BE#pws8j~yz=9|)}CsL4(&KDDsQA%JXTUd
zBoxQ|-UtT$EqOD&w0?YqRSBO=uW@vHE@5Bw%u(D!n0u_4DuC}|>_SdKeuaN{ktO33
zlQU`7((kioFE07VN6&nZKXi;%r1y%-*LQQyFN_%ka?KnBCYB}^t*=-Zw*14%A#rs#
z12gsB=Pmx}#mA#y@no!6XMLl8s{ClgzaTc^C%7dPQ0YxMp_XN+bOegDtTrho0}pqP
z$qHT4sqDAASNm+gs6iOTSuqg5H=$-K`3Wm^`?fkp;ZGt_^t@vcc@>}A`L7cC0uU6*
z5falr={n18E@w?P2&8b>`jd&lQGmuKIQqmTBujEm9**HAO4R7DDa7V&7n+nz?R7gH
zR2Z#p6om|jcSkkbJtIWs(oGJ>Qh2b8`@?%%>N|}O8*EC>Dg4YkU!KJZnRF-Hzj86?
zRDBv8FT%>Tkp}0|ReLj|GOa!y;3CQT5*&98A8z_D)|z!q4avNuuwV)^>sRQsaAL=b
z1?5JkuCwO1RhyhEtCoFBQo+H%4-^3sPlGixx9Wo-+bssgs4J4xlhP63ZSItxCpyva
zNPYe}ta92$d$AjTLKO>T*FxxR?S*i1vFLl?&_&9ZIE;nF$jcZQXD8$)eDqd1c<FdI
zTzhdWK(Ep;6B2d_g{5m?oexlYcWj2FZ3mYAiGFdr&rb2W9yVQLbyvHfoDl$~U}pKJ
zC+zltbhncJ=?TU_PBl5=#9H3wOLZ&zUEWD(ROh;ileeqW_3riu0r9+8J->Gl2xQw~
z@GYRjtM(<mvyn5M*K2oFTm2&~CZYikN~~qO*d8n=_+mbmA=3VE9Wh4FW$AXhChG`^
zvHa20yrU68wYM@BiuMvq#CjEFdzXof%Ve-ymaN%e9XPb2+wRFR6vvoqpf62v(c|}a
z7G1S`d?$&;815h6BE)6W-QaqRRUV3w80?yaGnm)r0OOTO_^cOu-}@8y!Fz8LYU!LZ
z@_Thv*CwC;Ly>MFnvU+&Z69Z`fwJR2U_zrpH$VkVN-pU$i^wl5s3aE=aBuRN=;s7H
zyUxn>-u||5Tit<9MPIVpCL@|X9vg-*73c@=;<Mv*F19&a@(!1dy!+E4i;=<?m(5kK
z_3|c}lVPFFq#sgIj}Go@NJ3SZ%wbPlmG?dE!p|n<T!3~lkj@9BeL7INZr{-Q6oIsM
zTjVaSG%Bfh%U{%dV^0rQORyegjUpd<#uTa-SEnW_)JhvBPf;-P2WZ(&Dl|xr*rn_g
z9^a6dEt8PT_{vRrhk{#@abum?aH$RmjACquK7$ikk*sWN#_HbT<lCA{AS$G=EDWxt
ziO114wd_g+Uf|eY-62{px4iF>3ZH>z*@}VxORYepfj($8Uy|=zsCm9gqts1kKBoUW
z45!oRZx)#(o0Z1%4Pke^SYDpQA*ao33y{+Z13;?OW3m<$Lvk0TqvInM^EG}VK}WnO
zQXy9DDuD{y)t2Jy#1#Slv%5t$ohEz1MZYmBc>?31D+~9FJH+R;+GUfh%RiyUAt7P@
zwZ;nASD$?&uq^7JwsoTzMC|$UhjS$?O!e$}T>=KbgU$4Vsp{R%sh$dngJ=q4Dd<!`
z3wjy|{|0>b#__VlQ3NE2N49mdBi-4{vK{?ovy(=y*sir^<J>0*dwpJZ?LkyTNkJHk
z)qJH!rKbmu(sF{b$K{?_U>MHX=?29DkcLU&WD_)50)k+SyHA`SmU1LyVTWsLYnBW2
zU`x$9cNR;K@dAy~?@{7-dJT24$U&DXsqr<*%IX4loNIscjc&{KLr&{2JsJ2;n}5Lk
zURML#jzo@s=3JURwt@yN8eZxJdLKDD-q^nXd$5uHrnQD2jyH5_B+XBu4Qf!x<LR_p
zHusxhSya3dE!NrizkxW|3tFu+{5M0Y1b|!OCN)#CZ4vA01#XAFG+5~V0*+a7=c(bM
z#%RY>eiS2+Q>OSkv*@%?;}JVKJNDusXuBF&UVVPvpSiD}s7)=O!s%CDUr!)ozNYAe
zIRXJv;2dk-s+jndgXgx1OjV85UA`KyYWqo1HGUx$#SDS@PI?tVp9hA?1E-zHcBObZ
zIo$@k6;r&}Q9;5ZYolP|$=RRMkpNj&^8Guufqw0-TP~Mn&1r#RX10&rb>|$7`ie7g
z@XXN3FI`kWtn$gfO5}bqbur{;@k5CknQj0ab=K)BFi9EoBd9nSmo{1eV*PoqFM=>6
zOb+r^l51OS;_{PDoBi58yGnr^_?G;kGJyYs<w{X>%j8mr6SaDo^j+1%P&|DSpB)Lg
z@syxktPbf%Ms1U;q+!pG%z9-$m0GzU`H>6w|1w{)TTH3KI=p(3(s*q6T|XUqTs+}3
z82XV2K%n}G!8{QJ?d3lD4PHDq&oH-|Ir{m91b|p;O>b}M8_VS48uU2nFWw?wAAN8+
z>|UC}d?Dif0m!>}>D0wm)}sW~ev4StiqOg(ht*bJ(&+*@`8jhLxzLz3_d;Dhj^1r>
zto_~DhInD3Vf=Qhf_UHk-Qysstdse)bQN3sP~$$}u$Ye}>;T}JR;?{AhN&Ll_K<;F
z3Sy39WTtRp@~lc2Eyh5)*|Jj{xVka6ogLmYm@eQJa5QMuX1TF6BLPUKY{&f14m*J3
zBvUzgr7(p9tiAdT(!HFdJU{|ZMX*%cji%>Wg*C@G2Z9b#JegutV=6TBY^11znZUS9
zhX%^y<D1j}?aM)#3!H^-iD;NYhTiE}jZLG9ZFI5P)>}yWkUS(Lrd+Fb_uE=qbIQ|J
zlwjXVvy<8f8Cz7lm9|i=S~DYP9BoK=q-bp8x@4`HJnF~Q>s2?ULRahTb45Yoe|FCJ
zK5S>9!KO8(QFplk0mZ9`6?wYEJrwiqA+>T=`6H7~h+sRUc%&B|{mrb^#cv&da^(P%
z-@$LsrqdQ5QoGH^GmQV78RGHTrXeZslibMmRj~8#I5AJmDMV-K*xI<Ro#^sSR;y4m
z=+(U28%v*t2T|gYi6Nkpvgo{TxoB<4k>r?ue-M!@=-uLZo$33cCN=<YUeY$F)8^x9
z!nwYkK6LadK~wIHRvKz@t1EQsMIruyMzt?+DEm=7wdSQTr?bBz*B+x<Rwg$`Japed
z?T$cNF_NQH<8YiqYdqp>qy69Cue}6(_$4FpB3s?C;7l;LlqTNV-d~mmJ?`my`Q2ac
z83Q6av^I@D`9guk<#0}8`H`IlVT(Gv9~e+G)zQYFdifwU!l3t^PdWalcI5bTC7#!o
z&{$MrUK9$U2ZT=8P3eQJVmc?zY1^g3(FX%Ff+Xn7!Sw!mB{Kl@<vznUB|bOUy&#i1
zzX%I_b{NkjU2SR4gzmTj`0_vU#`OAhU6EMvo<vqmbL<+BxQVRhE7YbV@!1Iep_`ZJ
z1gx5)%zIXqg6*@{^3cBn_n-DouL-%TOb#%w4wmN)I(7j<;jmSA=5(tQKMKgFl)lO`
zc`f$%Ss{Vq@MOMIXm>Qz?z3;>iNKcvXNzk{rw^=Z_m<<qqf)OujK#Qr^S0g!O8^}n
zl_2k#xhD;Ee;pm5A4Tk?5l&WGqau8t`~V~KiJKiX`=~DE30!j;evMM*uZR6-&)b9Y
z3fhZ%(H<jGydgm$fFT2tIss_M%Udy-`YH3WS3$Cu-$W$yAXU>u?L5fNX01sfdPXr7
zWdAZITulFdYXI)fU=K)l=$o8IoIEkLO_h_oWzgj$=SP9*CRDBq7m}Dc(+`t`!d#qQ
zbt@tHMt7vENq6A)C63$nzZ<bSdq)Z9)m}k!f&O1-G5w1|3XS7x<apU_X4JqXvjp1Z
z>@{T&2Y>1Pc+<NEdpy6)34|JUMgY?<omkL_W~YyU;PahvTD!f7K4@*ffuNZALt-FO
zAL^g2e#Nt%?Lp0IN6H49WjJ$lb8c?#BEqPNUmxO$KN1m?u(EJ05s=la*?eu-1W!0U
z@%wfVh)TZ+R{pxPIBiacM@WJbDtUdFYhJnb@IY{TveMNjpIT))tiC;b7)BHsS=Csp
zCXz(0nDKXVdO>Mi=<$e6wNi;fqQA)FXTqK33bz+P*l0GJp}L+fzSYpc6_=8r0MHP|
zueO+-kI#H<u$yJX-Fa`$MZ3X@^AAxoD&OWSh&y1<kfViqWB@ewk2O}{*jDY4M4e?x
z%=R?7Cpf-OTCi80TnY;C3cB<1DlwWJT34--4EiA8a>0JFGh8~xF*t;Zir8CgK0dSO
z>M$13B^H#oUmHiKMj+sM^(E7EdUurWk)_aBLG5EMkOEz3b}Dr39~6{r@r$KyWGhfi
z4a=sGEE0Xf{O#c66c8Uz31m3JAN~C3e%YCt2CmPSvuOv5x`pQ{{GwVL^^#lhHZ{tW
zDL5TI4cuU_vDSr=8Taeo>dBEuAS5#{{oxQ5d;x;y$yP;Xj<)ME9w!5Ho3~b0^w+Cn
zvGlmlPfl!0b@!xiNQKm00+nSG_#8Ifz4)PQPuKJM|6qIa+?)^gLc$zt+kSI1#VSe-
z*4-HbbftWf#;i^nQUqQqg>t=+)6U8Y7e{_TbSw?pk3PKs*x4D9gcJ#ipeX6jzGUd=
zvq?t}wi-H}9>}D=tVr+8M^xv(@C=Iji(SnRH*rL5YqL6Rx`l*?ch}(1&zvn_CXK`o
zgHeNVFfc-Y{rcd7iWGs+|Ds*bs#!Us^>9$wTp}2YLNQaR@B5EZ_t?>Pe_Rtlc{mdz
z;<pz@mH249?VdodfffLV^5t-?8p1V}#*-Ul7rRlKcb+8hm6JiQ4&l`0w4x%upkvcf
zZZosL@21x48PG2`FXBt@_G1<QvH+9mXG&jOQBl#^^@Q)gh3h+o?zdyb6Q$y20fPsA
zRL#(%6rvRJn6kn9>U6d1T=exvE)j;(8^BDBer*I~30x6oRj@y~1kx6eI4en1QwFz`
zYLs`Mv_sTlP6B`W0BALV&PCJaQ2elFhgYB=_;&GUoo>G~xFuKcMU?UYuG8I*WVJO{
z^?=n<Q_I7^=XT8I?*88JGGUZPv+9H3{kBRxcY*8_qiWFy0*-Irb9!a7vH-GWIg<~9
z2C)BDz^o8dV>6lazC^wFjpBH>z4J)6KgQm8de4wr=~P_aqe3v+18CKc)F)&}t)%2}
z0l<<Chh~Z^8Nl<W7p8zJk)y>HN2>w{pz8zi^tf%Z@vL9d<-k+rDvcF_V7H0jw@ST*
zDZKM}Fz3A!-IX>UV#v*j+Bk~f<C7EijokB=j&97!>f^0ijBRXRgqkA1#aCYL?axK6
zDMSDr5)cqDnPo^}GM+&UFg8YA-GRW(aIq}V0?%o#Onn?51iqa)U@N*%H|6Hr>yqe1
z)_fW^@{X~U%L&M#iedt=iZcw$A{`Gw5aQf=_{)+taG-jHs*4LHAh0A=mw3tO&9{|s
zHC)Lk0?^9g9HXTVc#4?Ci#vtgCWu_3O%<F16ml4!>VfTF`m>HpWT$lN+(TuiP@8;7
zbz7W8sy8WTzO=_|+uyF~rU^K~@wuM-01{E21^E=<7TH8bm1AX(lgneb^&G>V#@V)*
zJbGdT;O!NRiGhJLnS=t5Xr;-aq`v}`PI@iiPwJayJqKDn23#DgFhDFKIwM=Noq89D
z<UyX=p5EfR*5M*ci8SoI?!_k9yDp!zU84}(LJ80!SIHF1H>J+s{$MImfw&qakZ_bm
z+E%Fgo)zWJH!1B;PXw@U7}B=u@?|`X=QG-YXp%R9Q8&lZdtaL3+x|rM{$wq#j_w6s
zy1wl`m(2qf(CA*jx9|?oZ7{aIlh{_{x;iSehmiF0ml4A}Ns-B%-m|j}F>J;J=ea_Y
zWbeyst`f4d>6PY%r#&<8DvTwAZ@m{glhwvj&d@0d(|AN`YHGD%lhudoJIMHY69F@Q
zZB{FxAZ#1U`QRY~iDAFab{c@DNs1(MO`bP14BHdjmg>5sD5P=#LK<uf@lqalo+(l)
z!Qi!+0v++We?YI}z+<Qv+YNtk#x7AWH*)qtJdvp?c0ORG(~Q;L>hlf-SpUD{LN9*+
za?t*Aj`0V;B0M0qkof^cR6(q3l!nH}0Ysr!V=hkg-{0I{<k&}dOIq>DF#o>5p);85
zme4#ubPRBkcgLprE~48CfM%yT&c#!-Ynl39=Ya&arjHFaXScWgtB~6d#KiCW)0m@*
zolaJ0^s|Xb-tkMb#x}cUcG$bP^cLm{_LlRxoj-NH%pAsS(7!(fGR6L+g-;0cE6r-h
zDNUJLKEsJzm7bMAwNh2Y{FVZNu7|G9jN9zo*xLTX@$vCY6`S%;YDJQsrO%GG5FEy^
z%ROd*PX8&=DI-fy@9ybY^1XTno|U6cWh^a{4uK&8jqq#hyVFvuo(h45Hjh}#cj^2d
zZ+gQ%s`%Y{q7oFr#Rr-m)nM^4#~9fdM;P`5m3qTg-PTL<5Yqcl35}la&fO2o8^f9M
ztq-P3CZR@~57-)8R1psQXJk)jJ?NUXW)YrvxfFYi{5qb6x)7IimI9f0x-$V!A^=wS
z9kw&9`kSpno5M0oWBf;Y9_Sl`KscZ4evzmD2|&}q+=!P>A>}C?lO5g==c5JMuBR9f
z5Sz1v6AKWO0+85iaWd5X;#lh9hXDW!k@%)r#7iBX%ziHw$DoC2+!wAS;(1?PUDwTb
zyw(LI+DrZGFRJpz&o`ldnmU}E4K&*)Xz1u}E{yVFI_N+h(C3egI$54~nVsJjtW-!y
zTn}at(|KJ)%a>e~eY!uGvP?0>-&pXz=W_z)Yk96c(b37&>D+PQT5DvCb^^I6`4DhR
zP*R@N|N8;d1(o%SGC2ZOX^?4c-02_qg3mrZv2585T;|$@8C+gZFeGXvyTlNh)j$dH
zPoF-$s$i<1BE2j8fx@_<J4HygVR*OEYRBUEN4nPMhk;8jP6%M$OUp{Pe-RFUhkspC
zO%p+;E6h$C<3WgbCydjUNTxgjZa6rBK>#;~OKag-o}#CHrPUW*fhQg1%?N(9$%hUE
zAHS#ji)2O>?%LP~f(XLS@*rB>hC5v6;|JR?!Dtj*KGO$wZj<4lLd6UqFg8>m74Z4-
z21qZza`lW376fPlk!P=aHA!=s!F&3788DLc!9D%nG(dUM;bp8GqStYS)+j(8t&5+>
z2CB@WH#Y1?4=oG)g0Tvh&G6Na)yiXOGPpMD*i=%dlk$P*VTM06N3=}S&NIxgXcjq2
zmZKp)B<_5&0)BrUgie~U2@%DA(6@vK$_g>oefU@g4|en<6Z7Z?LwE<lQ-L>nY}KIE
zP0Fy+zruxR#Ak<LIB=IY&{aafSC%Z-u2=iON;}oZ=YIBG5CH4qcmX@%(xWpZu>0U*
z1qlA<ki-<3sd!~1p_jJ5rZKks7Qi_<KeU_Q+Iq)SM1qd%70AIFG7{3S>MYerIXXeA
z1wkmU+no9%-+^1E9xsAPsfu0e$_|0lJ^(H{?T)C4$Ql>OXBNEK0ZLLdLVxw03|2lo
z9>zLTzUE~DPuWa2lmplMA#EJm5gKjKc!l{hGw<+N7wK@mT*b+DtwzF|8Zc8Ufyi{4
zc428;Q>9vod8L4PfIZ&MQy&a@(3^tu@`5UK>Guu*B14tcF0n%Q{8~ew5F{jXA0z1f
zm?Z*(yE}P2RYTvpdqXPtXMx4I%3_*oB#j5_n9q4){reyUD8A{-Onv@Swpgo*h)nF!
zchQ5szO{ERv{gGU*c<#7p3o*095u+?`J_yF@a(kzc?Ee}@pA+NeKkne4IMq+p4a;9
z!z|lp%FRaI{IZ{*g9nS)?C5ZiE_f&bwAv*Gmn(IfXV<>{1Vqa+qCXUXghzI}I#qlg
zfHSOOjZm*&11K7h4cS8DAQS>?coy|jn}!00u7!sZe486335|6|Cl&mW!eMm)DeR}*
zs{xJ5#GE+U=+54OTL3;>3EWcCS*qc?D5g8q22|t#cxUI+1wF@=t8$)AhGvZI&yS;O
z0AT8sw|9v+wB+okaKPrpntmf*L*NYz#re_C);5YZ%*uekM57YjCu!t+1f(%TGkqQq
zKYsjpciKrWX%`ksvjv8>uGX3>KD9VfmNAx}T`ksP!z0VT=M4=RIg0QgkDd%j-~EQ0
zsKc4uGl!{E!vdaXr#eXaCpB{Y7fA5{@^IzP2ldCh$8?#(7RsDf&R^R9u-W2j#1XYr
zHUVhv2aI+k0OFU>)J!5U5)I;rOU1*l0da%yQn_rD#YCh(YDC$t;%$S5#5rx2iZ~nO
z4p;{cw%-LGp&=lnOWB_xmuc4{b-r|{bG#FTS*w;1d<E(eSr53wZQN|`6I9{(Mkkrf
zY!t|4Xb(0D4iS@yz?2tYyPm4E&m2!C{gLUTKKuag8H_#Sh&`oJCH(xOeLjZVPk&rA
zaHMuavVqn|_cvF1otv^_MY{r~>d=0drjsnRm9ZTwz&}83{PztpW&$#p5v=Ksv;H36
zRUK6zRmSnT;KoD8a-5on0Ml9#c~8TOB-J=IR9WmrTz>g)?9pVz5^w6R>g(Tn0HdPP
zrQkHP%Rc-?1gQKBz}|iKnF6W(tOCG(IzBZ^N1(n|QphB11{_V`ARgxd6dQk?Wr{I0
z%j+-(2q0DYU(@CahrzZnOTx-~)ErmWTFP35!67vCGN%UA-!}?kykdXty0Ta~VJdk7
znUl#>KnfO*OUBoU2WDQ?&&vKW2iQk#6{xNmQ|<eQJFs|sPp6Yu8^uQ|OpJCCc>=D!
z)mFC+$TmVGsgNHoSr<?NESwG0EOy%J&H(H=j-u~uvZi<(0MXYefsv)Wy&e9{<C0}(
zTP7`uX1rH}MLz(6ref!F=v){$e1^LEM_@^&n7H_0Vn&w&Eq<bOOr_G%zf2FnCx)&8
zh0K&WBO{{5+nWHWqpP`GTPw=z)nI<HqDZJZO4+5z_D+rP53=(`7;r@7404nA9;e>B
z_r)k$^s=(tygX!ywN43yQ$hRknAlfu8f{m_?d=h|dwNtOa0h>lrc*1X1UolFPotK(
z?7D}R2<olocRj?!4|XCLD`O)$_!OztJ9-CP(`2k3Q|oy`_tm)V|Gr9K&;<t9=}NQJ
zue?L8rYb>jq8KiwDHEUD(Vq)_jUte4aXldyc)m0wQ4q{ke{Bgs2gq{ioTxm+PDQ+}
z;^Mm|>=U62LF!C4<6~C{u1Lj*bgyTbr<;flMzw_2tp2xRKDnQea9<n_!PQbq?SYSx
zK4vY=l$fsC_<vb|1Z&yhf5IA&2otxXqB_7-GBahJDBe2UX&2{vQt(88H%RQ;={p@o
z1jaN_R1y+0vKxWsOzs_!K3>-~rcyw4c72g>W8Gd?_M@HWzddM5YviC#z|aH-*jex%
zXV$^-Xr;#fcWlU%xFv8W#H?KS>+(Yls_YFqHg;p|@BjZh)YO39rn}=G$wf)(YkT+B
zCjS2pqf@B5Oip$yU|Lw58-Q5YPJ3SAC(TKG`0slHOtlA*fIA}eF75?#4gW1BoJ+2D
zFlQuJ+X=Mj(keiy{<kPCQEKI~2_FCAgVdwe&kt(aC=k!>djTCs(s0s0)6mL1m`<&P
zoR^ok^c3^oHyPxZ)K&qd>9``?ZYM-x*yo~KRPqL<$(u2$c3u%<cT>%)eBJ%)b}f~O
zx00`HrfNRT?SlS0u8gS{kk_rRzy(%;^K~x?mci2;fqr9Le*lO|saBoQMHsoIm4b?j
z>i=A*PT1Z39SOe|0#MPt(&A0%;mB9$3~$yfOg$yE(2OShSjQmj>6D;VGwStK+@c$R
zZvWlkjI3--g^T4JbnDbjhHqqSZ65o`#Ab}4ySp3HwZRo30znE$eqrF?C(pVe{`U}S
zw6wK0HUT?0VPRnrIs<AkQ#)U3-$ipRpdCP*+a0kV%ixLFppXI(-+xb^24^ZJFSSIN
z+KA9wu|A@M(t_BQIxvrG?^>ft&@*CXWp!qVKAt&B|0AQQruaK_`2Q}#s`XWCV>>j~
z{kFa0jOic(U@)BK^S^U0Xx(v8-4mGV(~<_4jT!*+_`mlR=LvWR@-A6eOTg^X09X>t
zqoe1rNHU0%85!_R0??kNg@vM*<>e)yszvm_1IU4|Ofa-DMWrug=Xj|p;|6!&Q!OM{
zssy7{;a~%X;=i7#%*+UGF|#<3X_wPyK2>>Dkbmir*7d)A<ciKT&!8|Okpht1y0<dB
zfb7+O?X^)(&9kzofB_H<7qRfTGk@e{<)>dV|Me^JxbLpw6o{@&-szD3@>6(z?*Cb>
zZHDYj{{PVQl~HvxO}7x--5r7j*Wj*!Ai>?;-Q5mOfFQvk0fNK9g1bv_cXxNY!}H#c
zzpQnpXS%v%@2cJP8?@r{+z{uCB<@weq+j#O<NsT$gv9k1*(wz&PH(+Vh1>rb2qnAe
z?E4q>_o>sWIKVhgrU+I5lm;~9Bh0YgzaIt$^?@4<|HA)2Hv<U^P+1_mYuQ$X11~Q&
z65O+oZpyx`ElY*nHUB%6pu}G&?u8Z)xKQe2_3y6l{SRBQ37YSs#((gSt9A=C|L1#e
zuREEdo&S5n7oE19AJF`IO<a4O|I_e@>gQ{92f2GLR2;N)<H?OGkOj0?PI?Lc_iF@+
zup*k8Jo%JV7&utnA=j{A2tykiBpn?hO&u+>o0acIGF4MRha_KM?BXw7Fxa&=$R|F&
z=COk2-B2mNkgt5mf+%dyUwQe>4X}b)i*R~uY=jch(nkj1+FF^}kc54nt3iPj+?2yN
z+45oR^ttQh_Xj7d-$8vizn`Ew?PsKG4uM);UN*8x%Kpmtg_$fsn7;Cx!BE*R=_(a0
zf)I;{SF!Bf6;Jv9xs6SkRQ2lmAdNGe)BRJ-W9WbC{HFFP+a3p)0Kyps1&7-8c8CKF
z4c#@t2KOc9S_&A5oj44zTh}LV@2rT7iUTO2&s9Ei%*XS8u~MI(fm)j?nNAtmLG4e}
zSQcO27d0}?UL?<ONs0Bjo#AR5fX~PGYQ)4Khy+?b^F1nsB4Op;cZbgHxl&JB(-IyS
z8(0075%!y!k#Vw|9mWiuZ9*9o7bz;^$ol+J*)?2B7m`H~3NYRDb~>8j;Bv#U^2t!i
zt;5_%crrgE`_}e>k>>wx_dwZXsbtGG7cmV&SqY;yuuXGSk8=(~$@H2vwgkGFE>s?*
z769P3F;O6WZ?lJLyY?3f$mEdJE44zpva=uU1pSZkwHvHsX+_63lDjb~Dwye%$c4wU
z_~QfO?~MTvE{)d&_VM<_0=+6G3iPf`Q0?CBZQ1~i$jZv<+=~anOTRup7&W#3=UIQx
zZWWEAlj6cM*w{4gjz$opq9VJSVXz}cL-W2DIsX!rmeCsH#YjoD{2EO0or1G`MoEsb
z8)%uu2Yw6#0uD!(&i}JYGXe7eLLf>D5(ioUQ?zK42`!|)x8%e_Gh@z0IvLz{N_Agw
zk{NVxVU}0EQ$<i!7+m!qOck*?tm<a+_z}J*;FS#O`+Jd*aZ&)zp_C^Eqgts>lARrv
zz>vxT%$f&{>TqxE!62OjV_pTuy-~&L<v%+g%gRj@a}coFM1lEw46GuWWEvXShm~c_
zYiR0sCtK_DdQ8vx6Q5^qC}bJv3;alKrsN#JR8T)ZztHtIgd-6iK}E?P6?BR$Aw@6=
z%l|u5&cnrFVQVJBY1dgWS^t{9%w=bEq&ez8yQCW(StPu@xg3L8FP-z%%zXAn4!ujv
zsnM+hs|}|h=(`Li5do8iqPNEj=NlcOnqYS(^)k)8$}FvVU9S;fd=KW52-Ru{7?GGR
zH6RNL3es$MqN+ce4{9h;Wz#+aJMT0om8zG!KIy(6A5$3(4Z(;4sylgHGY6Z}@K3Xr
zoS-P<T#kYXSpN1e@gha=`~-nu67Bq+S}QH~!T*I^nf_%--(M7Wiz-6}ITGu>8A9-y
zf=6x+U&1Nse1JY&r3|cw_}``>{b<gIrr-wg9}TPD-kSW5QeK0!eJ>ru{&xV8f7xI<
z{t7_<ICwWvzq5Swz8;BWvvj7Dxw+wH)<7cyE9upHU(|d9=+5#xJOV=-{Yt%Nk1c;g
z^253EjN_Apb2E_+sGH`ZBHFT;>s@S(3SH6lZV@i=Ka;;V^lD6Z*Np!AO^H9hNKWYV
zyj}XJW!p61zV<}t?XB}Z!;v!o=X-{nM|j&N5@Q;#>=kvx0ZULmZbts=bwJ6<NDLfo
zbOM6Nt^Ro9l}D8M_BROP;}bTM$7|EdOF-BLAn3u5t)efF2;t>*wJ|5?%N`vNA#)^P
z{`KZFntX#dNW(CN|2-r)ynQI1nDq)x0v0;=yjZqu>3?omD@yA_77)~KAg4LH^VL7>
zTZix9d$v5XR^&>D15SN>aA2Sqd)%S~b^IWSR;fr7Dr8QZ(<WRCY*CO9AzhR5&ZwD>
zXU8(kRpfb$lNM|KN$h;WO~I{dXpr{T(3mb$Ly|<q>#PGXw&!^^VAv2YBm$+sH`+gq
z&yA3gDGAkMwp0TT7~|(6#NX*?{BIM|!=xpOy?>S!q;pyNxU@}*(AJ);CJd7(`wkJC
zGU#$gpZfjpt*l38Cb^|O2B!c-;S6YaKZB^6M;_+LwYmPGf9(^F<etsvg*dFi7}4ri
zZ#quZ*&Q$R3O%O4i`A0)IRLU{jf-2~_wH0rx78a*r@=v5^KAA0^@$A7i0G9_WkNFs
zO8Rg~xs=@Zc3fWLkHj?=pVOr3-A2;b(gYcdNTvbW)1;@L?ET120FxAeItt(5xXNTR
zHURe<T3OB8sCxg-Zgr#>eE82ZW()bm!9m`AP0<A@T8}PSjH9A`Zg5@lm|!DBCgQ^m
z6A)x`9Z=dH*htukSHb;<8#r}zpJqf0<(T)<Ovi0PnQ2YCf=czjH$R>v-oFO2TX{7_
zj$^nV7eI>D&;tZG?7-Gtd*Tu)gR?lg#~BHp<Wjpa$)dq9q&+s9G2(emnzNu7W1}Gb
zR&OE{S8p8pnBHiEz9^t2EeOc|OctVW_NV_+0j?H}Uq>e?I-Ao9vn`z!^7ogNju8T|
zQocka^pvKSHaVb-Ahq}AxDK=;xFTHPF?Pgd@lP~bsm#;X2whw~KS&v%A)tr@kP}wc
ze`i+axyhHH?e&q!)Pq%i5)1~byv4^vNw~Vk_+x(uf#9c0^oZIkwUX?gSU!FFMvac^
z>d_6i(#-KMoy+%zPx$xrNy~L7?K!?5?t7#WYtH)b4A<my!tm9JU#1mc+(iJPNq<Ec
zDDf1mBDz;B3&~5eq4{6_=f2qGtHAmpu$o|NYFk@OG}u0kT#+EtuNw1HSU?v=FF?~e
zJ`AZ5&$p8SQ-SppR_%a*P*Tw|h%yY(sB;`ism}UGvOn?6Qmxih;irEG0XLt35r&A?
zXt}6Vs^AcVVtJqUUCIB|*wmE&ZRbl>T`Y?2m%lJyV_=s&PL<wLCcaz(kq78JzBvOF
zsIYH7{j7LW^G>a|oh1gQsY61<QAxW$X9BVuIxY=HDo3|HX7GQV)7|+0Gn@U`lv&cO
zsnO{pcF`A1yDD&gy@dq~N13=HyDs(o!G)E<_K`qOX0!qh!KdMZ6aIsyI9&za@BAM2
z5E@Re&%6aeW8VQW1@DwyH>XZu#3qq=lba1UCFO8sKMnuJ@xN!#wZ5EXdHl;&rus{y
zW>ZsJD`9F14M+e1^#)>rlwY_M3K0dPfNncAH#he|liIAhx@L>(pOSHcc4Fw(@A_VL
zV?9;&|0`r#55AG?Ps36)XBfTto&cHu$~+NpJQcUDkIO8=-40}{uKM;|riD4gfIg2f
zKtpE%Y+=m=$cxEV)hi0UV1xHYf`H~Tpz22gsE8u(XiCZgN5ePB8=sA??17E_G&Ya_
zRN8*eXZ7bfj%PhDUREY1%YlRwK{WM;A)A>Dc*H;*nsdKRdJBxu4x~>ZR4`ZH4Hqe7
z^$n`fDwkO*0|cbBaDTUSywl^f>~f16(%9G-omvU@7ee~GTrBKs;(CCp$`)+)pI=VH
zN3maN`)xgqckQTAsRvjwZ5&w$FwX~AEZoBJiD$4m<zjuQ)>8RyZt0SiXBX|jRLc^P
zFZ8>CCQ^LD{_gRDj&)DQxj|7iw1rUmVYO-`y=4OH9cYq^A3HF==NwPej)UfhoW28w
zb&Lx;N+1#30pW#E$^wfFpNY!p_w41Kyki?G<*e-qH<ej0oSN}po49nyT^p$|DTn!3
zXi@=H9EmV00N*4oUw1^HVPId2%|OMT-76c&rwe@*?6F$>MPFk99@pjnuhd0ufsgJg
zfzqhH9pGw#j}L6ELs@+A+-&uxUQZSrHG7--(Ya3<nJmHMcKOkIraa`X*<CUsj)W~Y
zAOIQ~wnW4#W42r^_X;4X_m>*bP|0`^V;@=S&x&Cfl`3or7<C)raTo|urKljMH+!S<
zzV$}U-}jR}^4&F@E{8WXG!VM&jHH7tmu>@ij(t>HX-x{|HJ1!$N`6?*l;9}#jTA%>
z)L#=rfIV81CDuwqXf}KQxttTfhWy1-e2>rwJ>=KxEFhs~f}bEq%dT#SuOAtYYr4A@
zad9j3t4f98XCJ;?8e*EU|MC8`0myN6%cqxg*9DDy1n~QtcR(N;=!WpYgJWI{+s6}B
zI&P-k)xT!Ax;G-%uD8rP)B+$kY{_Wcu+}URR1%@!fLC77(%tEwuMchqm?0foKz7b%
zb{T(2@fB#*V)OiGXt(+#2gviicTozxxF!=i%K_{$1i-{e7Qmt?m1}bWG7_C<lhi;<
z9$L|Omde!i!3+|UPTg4TQ!)rt6K;uePRWV5U|VizVwZPjCn-^#9WJ$zq1In@x2&hR
zsQG8Ei;}9!a@>r#6%%yeXWdD!(}(I#6ES^+4b1fRo(eX9XE6}E-e151#w`JXDnAxL
zh60b}K2LV5uUMss8xi-q<6z1HC;=4Q(3Z=Wo;3ej;|!(Hn^Z~tfhWYkFw*qVFA!Cg
zM+S7Jeg$XFRqFLub&4nOE85$xG{Lk<$LIM)pf-9=j9WTWm@xvWLfB|RTC?h6>$?QN
zcM6Ls@s1JU^UZ03<jl6@|6CT`5eQlGe5e!D<o3AZ$fErd-A*fd4-)j(1LO9qP|(){
z8UbDq`*&>N@4$A9nH8>P7WqBDe8bkdzPNk(LUgQa8-7ES5Wby0e52$&ah6`wNh@xv
zR;G&Zba&F9_<L9th0-h6*u<>-LpF63PH(D7f1b+T80?koXs&>lGZoOOZok?Rb~NYk
zQL}!Ub!cld82_t$=FLv1c)cyp^%=!8fZJQ^P3my!aQnQocX!Qh-2mNr9i$BS@Z>1a
zBsu#h!g_A?s%mz{du2ZNFe!;YW)ad-8|x2G_gD8QnXRw*f3s@xWF^r!D|O*_5Pudx
z^G+P$5R`YGPI(xn$DGCYtm!(Ni}Na4%=e`WQEG?$c&ad5dq@>kAVFe4%51&DPMXd6
zww47dwv>>x`d3>6$GBB~xm`!=dCvIpl_Ts8>_K|2roKn^)_UgSar-eOK@{HTYycCz
zC$@A}tz5fIyS%)hT+QZ3{o&dypW`&&VM!@yQ$iwzoN9zsOp^j(!DscR;%zU}2Mb>7
z&O9M-{O;=Sw(}YP7_Y<WsL$A$m%k!XRLRD&Q;1Ghb9-?`X*bE-ayT9?i~ZCG_x7x;
zq>us7dCFWxbKJ|*$Jg6?EY0h=cEbbOoLUS%e({jFz(5f|J*>>EG=kI@JyqYGAf-|>
zs9pFbz;ow^#{LxN`9bVrW_)#kCWO7yAr_tQbC%G7>(rfe^<GKz_uFGIJU$0lam=a2
z5+HBxG76fXFrY=zp`E;q-;Rw2yp}w7%`Gfr)g1kwI76#E@s0i3i)i8V0}kR4dube7
zs~_jeBgwI(IaDUKaQAgur{i!vPZGm~1(|xs_V?+0TTBrykr1nF=9pDgle>)?sJOXN
zfoTLOsXi@qMSZ6atZs{(j9N9&B_(3$J?`<N`Ag})aZA0XM_RnDlswDOfl5xyNSWgB
z=*Y9y$JrQDDu%dsNgN$m;YlYWYrNW$vi^~TXBC2SkU8BaCH1{fQA|THzGLk#)P;U)
zW!0&`jcOzO&zy*VLV_D*thfl!aSgA1Y~xiqjpW?vYY;yvYwlp^-tMVG(IDi@iq)-z
z#m|y;)r=I}k7t3ulPVh1I(c^s*`U2xP(}M5HIbEsI~|0a-w}<8sh(dnB~Wv@`~DR~
zYoc?fgZBl4A%+{gf_I=BVqE(4P@?m0{B~>@z}3{v&`tk3m5J>$Z?1oU;E&&aCU=Jy
zuFr-w)IaG5j`op_7XL&|9fncHbXxe0bbh}T&|Mfyq^zwzYplG4gIj<R|9qfmSN}M=
z;^C5a^>nj#r)3bi+DU}Gtg^H?&hG4T`ST5{;hX{*nJlEc-7jrv?3UZv9_m)lodGka
z1u3NBuaC$x4JS7@1%nM<rULCI@cAFl<LtKPEo{e61qa4fqiLF<u=*4`(wt1#c$h1?
z{n8rKb8_O#$`}DS4nW<5@cG?|rQ!*qLFZLMhd6+Bh@!v5G90-y-uWYSnw|i}3Upl?
zbj3|&r{fM*VCU?3IW2q_-ZLhE5(%^#kqpx~*)%u1NZF6oj3i(4Ld@QV&8Ksn>HccJ
zxwIl!89KCi{AL19&&?H)&ldQrvxkU`-0}Rd=Wb&VOb)-<KZpkR>HOwaz?7t1)X5Br
zvddN`8X;J3wz|C$mOCW$9|Xtcmm{yAFjErhCyZi_RWMjdl>gQfFfa|ptmWZ%<6_0N
zZ5@lY25Kha)E1Oy*1M1l%qaNV7hRMb9OE5>R|*I8Kg;;$q_!_a^dc2aInX<4cN;H5
z)9^edP9R_ux(9Er*Lj?}8ZZZYX9**pM3*o!WH`+E>yo&(t#u5P6LK@6095zE`3p@m
z2}9chtPh}9BI(5o*lgY(LP8=UuXRH=H5@`BLI|=j1@J;J0f%|5U4+KAQ9c^m`rj73
z-QRY)Nn+;A9Rt4yd0gN0e6ttvwuSF83El0^?%0`+ZFYN8^?Rt|y;F!P8jsmqu1=fF
zgY7bmI-wg@{P~?xNj7h6QDC8#qk?}?Ab9Q=RG#N9CrxmYkg@0m8?0O~On>IOfPhM6
zyt*3u$p$sfqJOMG!~NeT`oXyvCeCvfhCD>JH_o+w=-l=Z`E6Jwvf6-oIh|oh%pO$1
zDVhm;pE0pO1kPz%t-2W;7+v_0o+w1x!49$Iou|iC1zO0=IFmTMioY79{_jTxX&YKF
zSXkDgGrI_xJwfJBB|mlNeFs#~#PJqFrNqm{cL`EP;bC)E=}q+9t%hxEZn2H7Hk<`6
zr)2R7N&eX{rPvL2FZxjzTf62JcE(P^!a<WU&ZM{6iJD<Wm5_kY?kF~s$YJTELO|gD
zzHV~2T;^;D22oxCA!uBcqj8z4nI0&_Pu4#ns!#SJBH_d-KP2E1bsZ@-Vb5T3I%1k&
zQBhQn`@u!yFzP&QGyZ$+GOl)=9}Dghdw%t3@2)W$ZmcE3#g3tf6i3I!g>7CL8<vJA
zF)Mc$k<R!`7V!x#2UlRWTgu!G8{YOS;V4N2s<1QOVf*3~tx4xl<2IT%MC6^o3~m2i
zSb4h6#mLm#_J?08DLQ5fIRl*`a5@rWVV~Dx{$A~Nnb&P(;4qvY+Ei|&7%T_ltocS`
zP)+KbG?<TCKVQx+kq@)`2durMvAMun+gP8^%o&&rPX4}MdAZvvZ&<tK#mXRu4<-k#
z(pFB!#V%0O72Ad?C$c)rz?;WMMT}0rcm?ZLVQHHXur%%BCBjo%;2>l{JN;Y-|4UC#
z9Lra=xD#v0=ATF`qOID>OjuYrGTsZyc_WQf+(p4>AOj9wFi)0Na}oYRZ(RcNx8WK=
zJc*W{=EL9o`E4F)?e|2xPJ;=bEq43`UcvQa&^ZpH{labYC`EE1%${)z+4qyAqZzzv
zis9N{lVW1wzBpACqjsBwIH`LhQ-0vDJQ78Rg3OTZ%C{-<6oL{mOzh!~h99AMO!-Tt
zq-54#9VMlo3MQa&?hDntcpC0A@pXna{Tfnna5B5lVz2&bxn$OeYTh3*B}i{+<-pzK
z;%a;=$!ys1h^b+HR>wh7NeR8I9)dY;my|cAw4!N92Q#>5kPK<n%h?#Md5I?DN<vfX
z$2DFGdB8ptozAmXP7EWz{dNn?%cgY?WA0h=n|`Qt_|~$^-qu=WYj_;y;M&_d6tlCw
z*TatHb5}I07gQQ^+cxHF(#g|A;@gHInlN(6_FpGMt$8xvdx~Rb+SevajMw`PD5ie?
zw_qe?J-t{pn_0u1hPBN;A_oB&K^kD^;K`I$lx(7Nxn{%>ql55W&Ed9e=MCaC+Lz+}
zV7P3ZA@=!fA<fH{)j}obzKSOFMmQ6n<f@V0NGKkWj7u(xqkpuVd1#&B3MN;86e_3j
zw25=&d$|1mtk0)_MDUllOn>iwB-u5^&RfNSlt`a}%tF?vU8(^S@ZCKkoxk_z-G5=y
z6of{~8VTtP+K!iZ7DAb!`>E;riuG^Qc$40wV|xi)g=mexZbb_^4iHKVzUk8G!3CE|
zA;%(i^a~UB9+jfy%Bqu>5{~2t=Hfg#?M!1o8?ROcxfsFcLy6f6;-S%gjM#|Ri0j89
zJCiVygo0MyMuOQ8RTpsGfgX`@s2E*!O6J5;vU`<MMq{-uDZP00Iiuj>B>)<Q;QHbn
zNG!Cpl!Gp#5PSOivS(jDJ{K0*OFjT^Cd5ly&1<!OJ~8k#CH(ly-eT~nlImtqg~e6q
zX8Mi71<?L{Jk#*(Gn2;tk-Qm!LUiAJx9Xx8J%Hd4Gfj<W5R4{#k6zh!i8$WE?FkmU
zy1G0EvMF+!Pl$L1uF9na(o&&Yg*j{c7<?KPJ%Fx4=Wz3T#FGffQzZ@dc_I2&5)Ww)
zmpc^L`610NAv_D^=sE*p!c$6*tXZdf%HVi6J(Cjc=QS+DU+l6ajhF`Yo%Pe`W-GQz
zH|2-|j%RFXgy4N)U%iLMBD{Rn2d9=U)5@gp87?m7hz&&tZq-SW4MS#TFZvCk@8|mw
zNo(d}BAm)|r`dLXh41Ih=$bAoCRnO}OU8*wiYAnTy9^MX34g)4KYFV#=sMjd_z{d&
z_u)m<$n7Gt98a=+K^dZ1{6Zn&x!*kb{A$69M|Ba@&?X#u32PA@`#b;8GBL~3Wd)LP
zB{6zHgky{>H4upgP1Lg^oonGE2rK<3VdU?2VZ7!mo9c#KOe)MSFI)>W-Cqj)*piZd
zeo&Z6m`m^!jLkRAXKg#~_21y!zC?1w15y5$zN6qJ13J%U7m*P`W(-xl;PR0h|54it
zq`*RH#Bm}Z0uG36!h<bFyGB2+#00cHJ4ZAV3wc6lHP@iykj91f{Ng@~$LIZE)g57U
zbW6hHyUz;nEW-xlqoX@+ynC50sK&-yi^r*n?AV=<keYWdcXTW)RC!6jN8A6byCfX`
z6!9ivQ+6(~o*k;~4RQFYR;BmC*^_Y3u`9;3mJ(mXC!KMLVT&S~^sxoJ0e{-!Xxd=y
z9+kH;9BO998_FRNokBkLu?d^6)s2<xBuuDscwkg(Hq~Sghd({`B%{^OR=we4(M6HW
z%aPnNALxCrVA&Z<CzwhFx;rhC7wZzQBNVwwdWBQXE}DMQjxlAwS(9Zy87oGw94)h5
zG+u48f#6vwjx4$SpB`a5=pJl|j#wbosbHm{bDL!Q@X{adO^21S)wv!%79Xw-M$eFo
zhK)(eOX)V^T=n;E{J`O*tFW|HW2q&G(2-1%la$SZ@<A%O7`be{-we9KATrMzhjHpN
zPKK8Mg0K}&Mx|bg14#Dw)RXX6`*(XIE64<rgO*FUfl^f9`d_!N>UFk6ELIVbVEL#P
zT*sNEwmHiZo*VxYrCcpT#oygf%=@PQ>jfZ`NrVJ+O!?Nn*;ptyq}N=sPjgP2tY-2)
zUwN`4gkYhQ>}Vw{SG&OB@uf2{W2q1f4h{kW=nh~5K>F>Qib^#VLo^)6`;iJ2wnW)c
z8b%efDGA2s#_vnp2poLNC78>u+|a#9mR>i#+PmHT*MwO2jQ(@A2`8u_0J;;?S4y|E
z%bNchHRfpHG34-@(#pP+YSqG9$ne-H*B3QIV!l?NCroIkp$qB)a6_57gAh)ZNxiTc
zS?js^#+t#=O(bE4ggM?@GSEYO{1MnT1?rLT8FEBjOG8TYuLaRaE)?8U9~6si`f;zh
zNy<8!1SK_dRQm3C8JGwy;z7L&RB&ztkbiu@xUu?q|H}92N}SezoH>)H{o>}Xl2^k1
zwI#NGC@4Aa`|E>y^PYpF52X>uPid>r+=Ab6=5kcf<BFa9a~CZvwU}C8VKEM>DT?i!
zw+4UnS8)}MP?fk-I9uyuP>z_3c~-&T#gW8_Q@d#<sdPQ*tDS8wWWp@xgSWOk7bziN
z5t>T2Sl*VR0#^Lpi}q-Mn#^z^jqok*xW|8pp^7eQ@^lFUW8;1%tG$cGp*=)+^?5vi
z)ou})Silv1jqwc@i}q{BM1hxMQ>FlD)6`VbYL2&*s&s7EWZ+A`dbY&d2GiN3%I7?@
z6*mFviX|<9R|IUK*D<SyjPrey_zUlhfVXzq?>we5sY^>4$IliLrXS_Q>*WX-kv@{M
z1`4w<G3^Lzk&8*1Q3sF`F1GEo^_q)orqKp=@3^`y#93tf!2*ioJ(0WjWz3Z;8-c?o
zht2`->At*3i0`>iqq-2<@56HhYGQ%ApM~GAjDEV{A%7S5J1o_9!#=HJ`S24Fq}Lcy
z#n}J&ZC=<99dfR9`px7|Q=aO!7}FwBohRCL-xm<u0c%fu5VkO`vB*3^(#vfmYQ5)M
zMSZK*?%lW)LLOwHJ(nVU^i|BC5?5{O0=FD&d}HaZG`@13y^rvl(^G%mQzd^F7R?J*
zT(j-n?W~ejl%`#K=(x;4Am=>>T=fE!8Oz7&cb~(<RrQ#}kM?8BP?qEa6HSk_rn2Tq
zN5%3N&PG$@c$BU*3Jg#*vY3_{R!N4gr<)rpb}Ayiz~_tMdg7gUVO?(0xy-Cd`{wlA
zzb}?JHCLq@uB9?DsO!<M5pqOTrMZMRU+r_s3}Hfgjx4Kk^z2AXF~8&Ego4PCj_qM2
zf4Ji#e=rtjl8HkN%m1U>Q=9$~dyc2NvKSFHqU*_wgD_WI+};Yx-ph#TP>+3_`vjTB
z5wRrdiWXT^GZ5zBf`W&^4;>aJLs@sLW31(qD)-;@WHrXM@A2p4(X?r^5e%cyYa`ui
zbmea3G_`gq2tfH4L5|qCpO2I`fS8}3&u1%b>B&<l{}VI_%)7kt!a3f|%;h&V5e}yY
zgqZXro-bGovwEHz3S6IyHy#vbR9jA%R4v%T!wZo1j`!fD1aUJmm8;?Y9K{U;22R8s
z9c7yLz*=@`95%R^n3($w7KW2ki<$djF-kp5zVMLU{p;rrWJ1(HuS%U8k1IJ`;&zVY
z)x6V|&dl^OFesbt|7exTWyXHmx~HE5sI6VoM^e7&Hb{f-J&mhGudd=osOK+!K27N-
zbPbkiwZaNNS7GGOdsnMYaVk->068@oWDQB&O!UU;qwuOm=fByR{;XEa%e>Jly+;dt
zX}y0BY5RoBS~=q7BN2x658VEVkVAPl322oOY;|wJc0VSG8v{I{PW{&ELA8wc(!$q*
zx>1PqzVPxK{*YJqau6o{e*6x`A6B0>;63E6BLi?h?l-{}oOa=uP*L#OK!9Ob?|G#3
zk=dw>h+QzY%PJvfAVGTDe%1c?Wvt_^NUX5;Zp=Vt5RWRnunIr3I<Q-I)LW*MW8^M$
z;LsWqa<_`a+u(6xkv5-xH%(S-P79P}6GgTHnvrwr$K9vTpT?~ebbYU+yhxcVdNr01
z(XlN>{R%>s=nfO-`I5dR-*2wNWjH(C?rZjBsY)kBW-2nS#G&xw6j#(;{Sa9p7nydA
z|JRR;{ZiF-(_5Gl8HX36*x?4AVx{ZpI<h6Dw+U3n{*Fxo0*NTr(zqm<6`9}g)svNo
zDF1+ZiY~3bvBWuM&WcWInt$N+f-TB7l96XE{EgDpiX_!geQ9Qr$W90U<1cfgDvFkL
z>3I@3k(`j4RP<1e9zCmJN>Yqg6q$9t|E~3JbuYFijlaH`P{^03O|KeLN~QnC-A*0p
zwE0eG*{MHkK+qYVCJ8}2G0~9Cs|}3H;vd9vs>#$pZj$hk5twBF<xd~p6c4~ZI}BpJ
zwyuyKznBfDxZMj)oLx^>?}JwnVNuinLTssL^+GxOd~4vi-(<^b8$YEM(rSWk2V|Z7
z$$Z{cTbD^wCQKW)HX_KvMcLO+8gwiQz9jD5WkrH$#Gi5`Z#`P6qW9x$&jQ`$vb^Go
zJied!VFo-cZ53Bm+T2h+Wu|JVZ#ErO+&%ETC+1e5!mjF9VMh!9?eruMR<xbXqQlUa
zsR%<3ZA1BtDJKIrl*<YhinQlUX~GyshnM2DP=()?pEHL^!@l{Wf50Ml64?3{_`MWW
z+k}^R_zkh{5Z?Cc4DQ!0n0?{-tpZ7E{d>IodHZlMoe@Yg)+zkMM|hmZZ;9)gw-ljO
z(4dXT1Y_wA9#cw%TEj_eL<+p)K(C%<?%B@zhr2EUOeAUxE{6v4^rcq!rp<6egrJ9G
z9sYK6Xi;PyJ|rW<iQgv?cu)O;D|K53^zz5VR~H(B+2?g|r3(j*ek^6O1y2Pw?=ML(
zl+yh}v$+|*m+R|5sbyF7Aq^;etcFJ5Y=++qBsCsv=%d1Pt1X^H#E+UvW&g#v^lnu<
zD}<#U*Ax1igZf90cpSfB^J7#-@a5w|cLM%FWM3`eC<xJ-L4@<@dx%(Bqh(G1?CCDv
z+{I&>u5WUv!qNb)ctkJs)}j?Xeim_LPjMjX(GPu2Bi4G7^6gD)*>b2S#3OV9ODV6A
zr1)HmuWDyXs%Fup0BHDV*#qR02=ckdh_{2N$k$BP86*zZ`}J3WabsA9Ho})J1#^CW
zIs0TSW73g=0$!D{P#g_;)pBzi`&;c&y+mtnNBBk@7H*G1PcibDtO)IGR|=X;uS|VX
z5+okt-bbUKtL?jCm+!$jdG&>`0Ew_wsl#YaJSaGnqmit!9zn+4VO(uIgsqrT*L$o*
zT>hs#sI259yZ_*NdL!z5cGHw#CND4V*~>&+JeuFX)Q|WUcVlu|mQ{kI?AF3(dUmFJ
zKJGU&-~gO<f$7-57CYH~E>6WzguEuyxdz4VG|q>Q7hfZ^f54UJ5o)`UUiW>=%^LZ~
ztN>B|Eb|sq-i$HkkiA*onq{7MiL8imI)X+KR@9N{^fe+wpN9hSZOIjzSJEwwVrSj@
zN(ve8azUQ>(TD^QO?z>GD*0CxG+n`XgjpYxZknr|7n$?er(((Gyw7S~>@RrbX!gi~
zW_@CG-FqEb50J@zrFmS*HAFBD7)r4ggM>-r!A_fMtg(D(&a3M9Sbc7EW0=rvRhD<m
z0vyy?VT{l9c-l>IGb6|2q+kRxaJu00fL=?Vxx=@1GZX;l5Jk73t0^%FP$r=aVFV#v
zd2~5t9?}rZe6{~2<<aW-dhU}B)W3)`s05#yC(OC%{bZ@H*`gD*I@P&JD@l5-{Z7Y9
z`<jX3&iRsPcad*?bPN`7=&@Ji@DoRVe_iURT1=N7BJv%o+Abz6M2_vlnM!P0zD}f%
zyx&WGYNU3QEI`UR8;B^g_O<2Gi|8d6Nkm@$dAEHri-RMal@bgQ9iZ<2-*J~!1oMIc
z2zT~YvOkH;AESO6PwD<{?zyH0CGE$ET!(9K4{)PT0qr(1PsiQsWi(6m<&uL=T+gxP
z(UG9NJTQXO2eAE~;TEF)kO54eJ`qCVBGmv_tQ=eGAc23?V*kL|d432*z<N+jvEp^^
zfq!jT*M!8Q>J|&uIqC12U8edAkAx~}sB{OFV!CjWbi?#+DL?ZY$<JOff3^<Y9m|OZ
zY8!m$FMR9bdy2za{kQSF`<6SGBI?2o@og@M`=pZ{{Rs*V3rWlT7t3ZDxky5cW4RYX
zpYjz2Y^X=$LBX87)Iz-jdffw-1S0CbA+=4<?!CQYA)StfC0&TmRY2a{j}G^pBqh|!
z2<@t@rO8kH?SYBxKXhiZkxSF9dEV}n`WOSXv(U=xK?WLZA90NxiGMY^g6bW48lLek
zQ||J9jfdT|tlbRbV?l3_k%ST5s=Hb_Ofwy@$sfxH7$Km-t&VW8`5;-yZiwuY0&h0~
z<K7SzjL45!U{^&b=rN#%Vvqk3lA?aL-clZ<^SmSETg#h@6!6JaUM8mFtO7qiPq_`r
zMf%gvs{1R!fsyj<B>W8Ju5pW^jY5}iODnHOB>h#!in7CRemFcYd9t(TmqWyiE1Bcv
zMu?>+md*^iBUgL2^4D0LB3Q2W6Z@rV<+)bysH|IfcyT>-97|*o6@XeM#bYCAV2t-8
zAHGX}?^Eqd3YD5=BF<|jlPceGwk-yYu|A&}{HtRxv_8l6<DPm!C$r12?)kAQ%0WoC
z<J#{N3(xmz$^tc={$LAo<kTqQv9=ZoixEm%q2qf$>Fwac0J{--xbU630rwvq0+qtt
zsw#9?hg%ac&%mK6DNjO(-;>LGb5kT1ZM|)`48hN+2@@%k;aNL6W@af@*Q1}!Phhaq
z`MMFbX1UB)IU6s({eP-y$gs96KO+<K#X`es_3uqTSgP4E;p+ObQBv_&Y{>yDdTyLB
zgG=rvBFBFjZ&5`ngVaXIqP$*oF0NaxGb7ubKB2&GrcJoT_S$+PebUGw`Xu5q)7g5Y
zlC8CX5edmodE5SBbRC12zbAu<;Y<h8;u$|}_Z9Z?dn~#u%C|#R{H_SsmP*rb3M1lx
zbMMS1M47%6M~Auk_;u|Y#>%f-tB({tX&bEZdjD9q<k3UVY9IOr<Yg^V>bAU%knw<o
z{+Li^E_h~|QNiy+SbY0d_w4(bJb=ggq&49qY4)KgI3OS?)@<qOFKd531jjc@W$isN
zQ7w`NYe5F;4-I2vv9k0=Ee#$mmry_)v83iSCoSnXVX5raY0nPn#L!!U-&I?zKJTs5
zMYgH1(<-voe7?~?W+?7ka~QG&k3X&D5DpX2y#AP8P*%0M6J)rAc2(=2Hd1c1=){Lk
z`+DE5=waw%|Jy>p5UHhYXAt@)N~84`c4h$vjKS}aD|`+d2N!v)1Spn?{BgQmj{G{1
zMxr#&#;2_E@`zUl^KbW@!z$xf+d^`bv*xKA9;pW{DSHh8ZSX|5shbhUn3|(BBcxan
z^z<aszu!SY3v5A@-Ew}M3hA;KBp2=>pF^}P*e59_0XzhcmvlI|oKN<R7gD-!V4}s$
zub40`Ri138Pu*m|l`ys~&y&Md*@kPGiJ75)REnJPfDzi&zWa;u+Qopa!qMWgKYe;0
z%RwE1Tb))2k&&8JhL9=#kMuP)?nc`KTVh2k8$C5<ITw3)`ks)KU4G3L?_i3*?|u}L
zwjKg>^xK$(<7dc{<{u~}=)^Fpdlwy+ot7aLgZ`wdA=qo&Hr18NU%+|%r4S(jlnld%
zs<-9Pg}PIP;pQB1KgLCDir2Yo^vIOf2{J2YSjg{Je^h0@!QqF;L{PXVpT!yn(`z@g
z6MQQaiy}d(V~s}2^+#xe7QH4aZ{uQDnN0GgN8)I&EKqHhTkeV=hk>Pd>rD0bWi(N8
zNM@x5mgyZOH<4@AQSof;TgQdQ;7%4&xX+q_;>ebvtiCnrb2emc7`%oiC!o)A<%m<i
zSJ{8nATMoJPU3^auw<&jUR~3)8mXafy0y25EUj7v)2R++kYE!(*FGvGo`lpTTgSEF
z)F!MMg~Gd&mVT<Gq{@|*8bVi6{Um@WQ1I=1y3n*qSGp@t3zNczaB<?)7n4Y_!lYM1
zL>bfZ<~mO%`eweZwaSt2@E8%%=GfAC$kRp*)p?j`-Epu*K`pm8F#VwV{)W1h0}_lc
zsq3cB{@Tjd{}k2nG?{6}$Hx0XWeV1)C&5u(Au@1qJ)@+FsuH&DtDW8XdL2wfrH66<
zg!Kp^#;A?&7zs%<Z-FsTv@Q>&05_5NlL4Ow#1R4j(CWt*g{1Dg$)ic>o`b=D85L`R
zLuv-OPILCJP1x9=^$RR??z2gM`d&&;$<bSM>habjx+vDP7`yuB=bGVnWOltb)u8XZ
z)~N9&wLw%(&9~;-MBt(aYdLlc0MoZcl}AG(!q-c3o}qJmAL;FzPuzR@dW;ui2wLPj
zT*d{X?}VSxvHD}XM+hu(F00>7OMxm4O7t_+&|jaRRX)0T-wmiL!gBPG{Ehj08r@R7
zCv%vkVEC<HaW?w5INsuFcX1=Oz6_|{e_yaP3tw7hbmRrx9A3z_R}3C#oU<JP(V^}-
z7kb7i`-2~(avPDX3T?2#roj*S!p_Y)04OL#8{kqUtKm@Wac8={JB>QM%E3S;=ZlH@
zc5Bprn35?k(^uAJP5B<1Q0q`(dez)Hc}+tjG}j#&2AGRQjm)q<9iC_g<U*rHe9%kO
zG$hN6I>q2s%oW5?abvi5d=XXs{FmG&SYy*B)-t)uR~oWyRXdPv5V-k*n{;Z*n4C~D
zR1+D5d@1ozy_mgr!mw*+>3|R)`I8YV0U_%tZpKJOfxsP4wX3|CMBiC|Zn9MUFh8)-
zkmyj?zR?s1gB$A;k-W{BX1g90ouHZCT@9Q#4(y6(GnH(!q#4g~Omyl#)lY*`Z>
zCHHhxH0?)ns`9#W9nfeXdWQ7<2tFTB^&M7U$-q#GGo~eAMi%F;#s;>$Ji8K?TXsxb
z!Y0q~uB)o{K*Jd-fd~Scg`ku>UvtJ!RdXAP;6HcwjL}#O2P~XzW~}WPZUwP#r*N&`
z@LQOOv6}?}AtrO<x|W-^U>wY=7;ffK(|-Vg#j{aYZ5N-;VX>JPLE2^T`lzw&H{N<b
zDtOmr%G&jN03HFkYfAAA3J@~42Jb0WG$Ep$t;Jj9c)?z=w4if?Vcb*q&K|4s=sl0C
zCGpuPUuyBG0AMOVCqW*4z9n@b>(Ppf0Y9cgmU1`p!%wf}9VHKz2yM6N&v}X}M)+RL
zF|sNOFPiwHjBqoRm!D}Vp=;|RJ#hEL7QmTsu1|gfoe*3O{Qu;{XuAb=DPIA^>pE{I
zgM)+m#Tg>?Mr?Qv2uzGb`&8)cE|8OJh~i_T9`16-rf%{Ise^!#^oWEbgh>5)$ylJ;
z`8vQ*zO~?t;{KVNjJZ4eSzC3QpN)|l-_4!iw^}T^M9?@#TCb%`A&u{dHgt}w%zRaQ
zCQ?||OrPlz0<7U#wf5`1s~jn#6arhGI&wTxcP=^-;-?JqX0Ggq`BS#sEk`Rsp@rm#
zhB*=-?6nq+lGu*3<4`O2)NV|`bIQrYB1mfc;ty%T{+8Qa_zlzn1Es~%%5@|gK{Oww
z;kJgX1_h2TOS8fPyBnMg*>h&-3_Zp0!+XOGJ#FDjGM0JNR{}*l)#YW0LJ<Zr_H+L*
z#emRu<J-31?9+n9f5*<J8bs2h4sr1Ge++fMZ=S|Exaedhl+1nkr2tQvknI>OSYAlc
zo)0M<b>yjfRvpEM$%hb3KIZFv=zr}lWv@W`lI=Qj@3iGOSLq)H^!|5Rveq}R05iL6
zz@2x#uf2QoZFWTM8iaz@oE2@4)~2Sl?d!v*sHA?|81lk4pM;(^U5I%cA#lMkDn(y=
z)im_NHfX1>-vMLLE&5kn{KP^Lp<2@5&h0t;$dG=ISAQk!k*l+?tI~Ec8#)UNVF43x
zNV%3P`-qCeaq5vNh!j?Izt}D-_Xr)Ti#)8iZozzb%r7mb-ptWA*vyUZ(7QYUu;*;Y
zyeXT`DV^`XTQPRLzZi<`G^(vp_-rU{ENkt~(PN#xtQ<^!1=av54KS-`?zqMp{y5{N
z#JEIJS_PYiT#<C89+z}&;{dEE_RZ)ANmPZXvRfkn`Z1d3uXj#*uQCn#3M^LB_T>d{
z`$!mr>9mw)dICs4{po)^#$>LoIqpwlvQZ4BMmc3D>C4V1Y&0uZdr}OY>6%`smROH_
zpmKVcmYA1^)%<D~@7LMWS!|!^<!%C+890*RPC_>lx5|qDQdHv7c^<Fvwddb{AeO=D
zx8-BB@LuRkX@?GWXO=8Z=*oc9?kL}S(6&JT6#Z7B6}?QHkUw+uVBowGQM*tRb8cfF
zFL~cP{k2cj64VLQ)%Wz{y8<g|+<trkuD!|hNe|ryyE&h8gn`X~?{bLY#+NdH-WiaJ
z4=ayfh$!nTu+pB(|K&8l&;b(=%6}zD_pp?AWUsjTi(J&AzTqcgkeQFSjY(6R4sv(h
zxVz?AO(nXmk?y(LCA$}joWX<|p8Vc2jQBDrMNpF<lMGeO$ixgD+l;4>5Y@F{h7USP
zIHi3hLhCqc79B)!QRF%T5PCm%XJ-1n2W#9X4si2^4pU-$O(G49E$+Egyrk#=-@YNt
zD0h`U?rwBfO*l`2xE8Py#O~XAg#S?#fWE68nJTVk%LI%r9gXV?Ss@N7>c?KM>&vme
zg9&CONxUN#3SebB&KuE3h)JWJ{(Q&J!5%vjbC>cA{4T7pdm7!h8IGs!v*^>MS#unr
z{x;{zoX?z=IsavQTkz?7a(0W`E4-$>R7(px(NDRJcAwynn^6Sx_cLJy+ObfjIE0;y
zURria%3akZ;jgxQ?}6y<nlpogf$FmblO8P)FA0E5OoduPj@1w~--#Ba-;pUqbhvS~
zzIydT@rKham1Gym|HLIr+T>AJ#)*oYY&2B;J5u>4G;b9ypl3`bWHR@2viO)bs-6Zp
z*($`|m|0b(TQ_}s?Jlo}sJ{<CG#QM6o>NyEcZAM$7m-p6dnvq}|A3^Z8CA(K!T`H2
z$g0nF-+RvHQ+Zsu6}pL(doznr9=)U83;bpMn<z0mWIfUw1ir1Gk3ul{b5}xoVm9=8
zLo2K8J6Rl`71ur9GX@|s6mNgwl<M!n3JtN^p3I44{ujx0vG0U~k&iu>`;NM?o;%wu
z?#zYH#_os=O%0cYH1||`*zup>$RHkB@@u@-5c#RGI7{>(mwwc~SeI`;%TwP&LM}GD
z(mn56<)=UMK`3v2&&><Mo#t6aZ*`;JqWs{r7R;8$&v%;~e<jjsitaeIY^2Fb4f%v<
ztvt8gGoLpBLTl)n9y>G94E$~Ot>3!;zeJp7mG|nH)cutA@dj9)>iW<l=1^N@9=|n&
z7Ho`E8#xOK0orhF2D#>AeGlL<Ch(yQ(_Lt5AK3vjZ^5lUnQuJgbG5)40s9{`pwq+l
zcL|)bi@1q<@HfiPrF|ezgGw)fYt8%MeA~b}{!A@j@KhA{3#g$O_nr~QaK=TdH2JF+
z%qWY9zFB@4*&~f77&e1E&p8v7%vw<imO)o42{+8)!=~z8R*0X2Rj+g%>1|zT<sOLm
z-1YY|=een*PSQ~lErfTAPcfDVNPnPV9k?wV92Yfb-6b{!6)7GsiIl;IF4JWOXnWa|
z`t2Ns>g=}N;gq`;9vwe9y$!ju$>EJ5CiYdUb_{vr)VL|Qm&5nk7o#V=rK1h!v%|eU
zxqdA?V6&NnEn42G9})V@#0PI40ul_XTfx*{c%b|sucR5POg6UAqFB`%Low0X0Kq26
zgt;c@P2o;jE(XvbQncYFJ?rQ^D}?U5YE+c-F&k{yJ17pfpGtlBEL`PNmtf)Wjo64b
z64Z!G>pkeKv4HBRNKkNew+x_wDVDSW=pB1S8R@A;Safiyb8-Np^(O0&b_>~IkQ-}`
z4_V9~Jsk-w=m{a?d30#qD}l@-g9!dBzx|SoN;1#GpO*jBV7_U77)gDBDa6?G4vuXO
zF%35466Ki)RYo^T4<L(4stj`)fHwD;VbPFnnCoCa!QHRHI=!ic${%BLU$QOYF#^ei
z?bP&Z*Jx{5WwLNwDe2opSPGmj{yrqscOUJKP6>EmBxk3XPj+I8QEoFEamO>6{D_~(
z?ulD&iMq?BCcFI+6a@3cecVD=<Iv()IDgJcG1cVAxrvJ!*!P}Pg)^>rVvjSgLBWB4
z+qF%k{vtJ#eEb6bH?7-0$tWumkJJ_GBHgi}nt)Fdx=lw^Y4(@52-)C-uIqL=DrIC6
z3xQR((MbK1JS;+hehrDe<d*c_6<!X}FDlA^$VFv&gVfy_0t*+rs{sHgyK#&$(aOi2
zNN0HYPv+G!d`7~Il5RpU7>WtLE`6i8p`dVeEI>MqRpEVMyrzY}B5uW$Sj=27p(h#!
z+&NqQ7^a7GD~l=XvugL*r%MEQQeH6{AIVs0awPG>$eR#`W9rA+(NfZp^01EszqP<=
zf2H^(dXRwG*NHDom@wkU-clxH8ZZlTE%OP}X<YPwRMgyiCJgTZuWLUbPY-pM*KgIy
z(!M~aEO^upebi7TSqPg!@0oS6z~FP<rfZ?l2kDFs=M^AfXs8&1h!BPs8axc-Z*f#0
z<o8yG=Nd+?PQxke<=3*~$>H0L*6_Eyd2HQU(Ko)lzfNBxHcBiij1Fq@=3j@~&eKUx
z!LMlFfY0d_DpsjI@B!9MF}G&jZ_u2}@OoWoglOmQK~5TOY;tiRcRvBPaj%&LwUPW6
z?9=BvZ`J9=yy(?lh_h;%kKtR1gW~tRNTtV`f1mIPkJ;0f7uFNK{FL>s=NG*h?ZddW
z{*>uqB^96(*!E={p>}PQ3okY)dPj>9k(hR_!xPyHI!x0KI#PQ}fo4KQO24lm95e-*
zsJZ__Mf{~q%^XN<_kX<r0PzVn?ENL?Jx?{X`w>20(MtOjI{6YxZw?t7djHhBQT-Kb
za$G+MA6)@2oZLIm5GBazp7(j;Y(MIAY0Vif*0vwRy8hcIH0y={SK%_G*Pf#&LWd`a
zCNSxj2aFDVC<=Cr@PTc2CQGfOk-pn7SrW>WMjEP~;?2WArhmj=1p-^Xk|>D=W2d4H
z_Xb>=YGi+iVe2+(xtLLL1~{_k{&k#BKI~4-v{avqQu+Yb0<@DJwo=dk(20Eso8i%&
z4WucWw7KE>TXj&LL~ltrVosGG%&3HeqMV1@6Q+e>$(GkL*2%#_`&+^pa?1FcsWdJ!
zH(r+uL5`bL%Bh?LH8pDA5D8WqcJnh{{N_?s1<(S;sw+rE(2Qgabbq3z&YJkSvt2S|
zC{68)NhA=sr+z4s6S6nRx54M)K~By7(F6YCv7KTv73BQtA03^XPr=I#ZGt6^VOnkI
z`vsc)8x@A-0ofBlmvC@7EcBAVT$J8zSCrKMQe6RW!s+4V_{;q}ARlZKqhCbv>Su?i
zDRr<&h*3hk>u7zF0}9s63MA2M_|bOnM8D8kO_;fQG{8<#OcpbR_~Uh}P=e{<o0=hw
z4Oz!sZNts%ocULnS!F)3e}gj;cc+K`Q@;Kuie^B?9#EljbF2&?^43H6#867U|0%Nx
zp)k;}M_vGK_TclyV>jR=&k?hlAx2JjUyq7s;ek>mKDDx*w0kY(BzWmySDiwjdjjpd
zibXF$T3z|B+us}9Jc|!Sxg`P&oI;F&7J_k6XuUmo%DVWj&1yZ;u;lcUE-E%2f2z`c
z*_(-u$;5YUPs>!wFi;7Ngx{anM?=f|+SYqz@lq1v$%CZbjYz(fW}p*cz(=U`1tn^4
z|Cs5fBS^r-3+6R3{*HU}$It6+2kYV?rC@-*qI04~Dir1?@ZTu8qK`)35GB9m#D&79
zVc4>Hvq!;TJmAC%-%?P3{Av05i2%jO&`!(iKlF^A)O>t|E_-9Yqor#pIH>&;Sa`@M
zeACH6LcOgAcJO`7?7WJ>k@=VP@rLQ%P&4UNAp87el0auuj?fNAP>?b!jUsHQ3R!KL
zDrwMBM!UGsz2J1?cl2NrLy+uQlFI_KFMh?j>426&hrksszRGh$hxrjVp2C-o#be6j
zk#c)_>gg6CPG)N2&m#)q6UKoThZkiKY}tA(&_C2>=;pdvh5U#~(`JZ%+<U{}+UCod
zXJ5#oOT2Zq`K8{7>>pc|W4`xSa2Cww8LB8CpyGTsd>AuXu2<*u1}{GEBQ{5xQ5tI7
z!Y9JQY{_Ursh=^0mQJ8(#_)LP3NUA9#-u4BaefuCZN#CNy`@Frb{RipOGEh})eHAj
zwc}(gR{rcnxpx$zys{oFiiv^y&-;7U3~cMurn10OZOue11Q??_4>t#_7#S1+T%Y^T
z_!@M>=UAJG%nahdTbY5U4j`J2qWWhYMIPGMiOz*Rx$XotbbAUO7w*p?^Ff}Jy;bKf
zjpf(p&1AM*bRu%Xm2tU^At;jAS5?JvXn7@E^?tH@-i!UrW-PJBD1y=YMRBkBJ^!;g
zNCfiP)Z#Fo#YJFzgif@4B;>R>#1N_ANF04dWONO3?k8UEu!78nxZ(3?>I-{JQ!DA0
z2~yJYT;e4ZY6k{nUDqQbB%Ld*Iq@&VXyPLS_GCa93+1KSs@_tf9zQhgJXQHd#l|ia
z6P;VGTLH~b(k*3VVnxrn4=QtWe`=Ze%5Lz6zzQ7E^2YVWYV6C|cyeg8;ESeS5sljo
z0BIx`725-^@N`nvtW&(Z)2e2*vK76U^1nME)OzFHeV!Ns;>F2+E+4t`=^`y9PLdmD
zf|A0gOvwW#@j<*FFp`o=xiD{yR)k$4BP~K9a_UY;<sJF65N+QKA&M%pbE9sziE<
z4P7Fb%uQypMZRRVL2HdxrNzpC;6>!=7G^OmXeyY_L|1RwT>oM;4MXa=LFU6m^?v9i
zhUlJ$sgiFu6EO`PNfRPL6FP^8S&dN;MQ`l5-^NmOG^r-4Xh!|h9WFY>`Px=eA|o#^
zKy5e0Nus6ITy58{psJ>po#=wOLJM$qD-bKw-4wnt%#fn!ezCjzg{AN=lTOnjlt9Z~
zQ<!!NlqDkcAbh9k(Pt!`t%M8K@iwRmwP>SkHzIke-J8gP!QfS-PlVrzaYA!RL7Aid
z|Bt3`4y*i&_Rc)HsS{3~Fq3WDnrs`B?M|9J)nsF`ZM&u>Otx*m=XdXY|L>{iY47ja
zAFZ{<`1KI<7xTtZs?r5<xh7@A>{=?<IEn|D+57|h)_RUJUh#Rkc9VDJCR_QLvs&YS
zkYLB2Op*L6Xe4+Vk>p<NKpIiZtpdh=W9spKzsTcI7hxDjz$Fm%gaN3Oy(1PtC>s|7
zwT4A;Ars7HDJh>NM3<Q<{EsjIC`wwOxw>i)5|K*N4+KwzPS7bSKf^G+eMcBO*7x*1
zVuUDiI;Yltt>1RcIXOTY)E2)9O@b?1%uHYybJ8bl?Kn-@WLBe^mSa=M6cNumUD^Fz
zMrq%*nWfpi$fuZH;qRF~V0cdCvY%#ACXpx4;Z7Ts7dzl8uW9Y@-?5nQsyBsf7w9tu
zY&FI|mP`IHe0_XT3qVwd=UR1wg-j+z?9BN`L%=_~lPJCXS_1taaJ@nJ1VOR4bius<
z{ggMsFenh(Z>&}?4hWRvISZCY&hKU^mOuDPqH{zZ!D9WkL<qf)tl1aOP7pql7nrxR
zY1=*HHuBOfQA-%B`Jh|N)eKkn8$xJWl^-{M?zCkQ5SXPnO->lEZrS|JR@GzujC_R!
zs)m;U6C3iGRGBDh-z<Ypeq5lNj27!9a}nc9Kj1F>C+6{YJ9(G8(FHK0=b(w5($d1O
zK#7sLdQcK>7$C7Hi)gN9|79K>V6!huwyw!`p)o4FV+=n0`q+$x!vJYGO7e}-nuMA0
z`1S*$`$VXr*CRi@tz_icOa#yAlJ>TvbUYcsuj{w~gL5Gma@|C2<&!+X{2@!Fm-F@Y
zR9#PyqfD2B9odAezd<1qPH`JB*t!SX(##NMbZ#QbDC>0bh-+eoLO<tt{xC6>B$R+4
zV}qP;0ojhXFTRx1y`89fy43WE&;3kP&kjJXkQw8=Gc0CvfI%=mKQTT%{EwsDxuP_p
zd+7cbryJoPshG_JGwq@b&i6oyix=(o`j&p^aeDKMlHwxrFRXQ`Qo)Dh54TPq=8%%L
zYpsYWo<oU7g8bPddV@gB-33?`V2rq$j#MBrkg37~@rSrW-IW<b$s!hjQdt-~rB5Bn
z$TRidc?Bex5${K2n*?@YZH+!|p<9Z8>MKkl2^p=qo_Z=#wr8}z?FQtUr*}n8-uuqW
zECF$S4Gpy1tAuv@QwJEeSV1#SpbqrZ-vxE+VGsigH5oiccy+a3P|8nkM7qvVv-*d(
z<4PZ$<XPtB;ZdKm&;4%B>Uon92|6L(BjE4WLyeC9f8r)Rz7Uew>ykreG+9KIRjUuA
zX{410@?cgj8k7ipP~#;$6PtgBP~A)~-%bY7!Eb1oX!efE+QyGunis<pw@Xa~0%tvy
z*5A+;CkVq&rb)t=n18DXyW(&f&*KWS-fR0|1(W@gDyIvz-4klZ<LN}A{#vbKCY=S{
zpSx9ZR+^Hkx#<}9f%ftI(Eao=i9op;H8D7SWNE(nchO+@vS9)1<zX#=)Zlrs`_ZMq
zY_qRO*In$d+v{sM(dE!dbwNuZoTi@hXTnnj-}a4F7sv7D^jV{Un`t3%7%G1D_Bu8w
z0TK)JLUb_+$n|1_a6`Dc&>Uf&x`ZP4XJ1|s<(fVLG<R*-a=pn+kFJpkDYpWmIDIyP
zEojL$Pq;0d?LgX%8J4z(EkmMZy`J|YF+E;j;m@su;zn8(CW#<IlmzqI<DQ^DIvwr%
zm<~chS-P|S-6g>CdpwHf3gjEfVdL`_PjQ4_|Fu-G7(+`(B8MIok{K5>A-Ix%XEo`o
z;15w-<XBJgMjTH&)j8B_ZXap4*L9T{SZt5Fd>J2hQ#*u%lS3L<Qh~&bka~@dU3DAn
zgfchy;ZODgLU0pSD{rviW?3rFgCW6$b(l*H3)yE+#{~-)<EF;{9#m`M|72Eq^q-uf
z_=D(0g_7DQtioEFZif_@_%>oX^2`puQWbH``T|_xWtT;)e!|ZX%I2<|kReEl0e*zx
zCLyp;Wh49~dkum|t#R}axiErOM`Zt7kCwCPgGZQoVsOvICXSg~j<TAu5wca+-QYsF
zMsyadaX=u=z2@+`-`w5%M@eWfvwnK$312S|vTZFlHy6rwv9jNX)Hm>?x*cz$LMsxI
z>sr6&TK5cz&ZLWsY&mOMVYZpK3Jr%VD6GJR3_X$5(<9z+P<CZE@H_j~@&0_G)o6i6
z@^+I08TE;DUPvZGK`oyq-Ms)Co&9g2%Jj3zRD%dPl8KRtd=4suKgGZ{NH8jk@VEh!
z-bE?gyKlp^muw4Ss`_>}!=>hMgjR8iWbPkMd2_P~L9M!LO?NI<PetUhFvqsz-h*=b
zp}z`#R~TktAf~|hak=^Zgj-8Bx<i(a#Mc8Aowz)REYes1aP0zPYps!oA2n&&Ozdw0
zFBMF8rx^gEql&0-z^%jx>TaNea6>gO(!yGlzlz#*RkQv5XI=`zt;B?fR&0-sZd#LA
z{{XH2x)5W2#N1>efY9Gyz8gxd>2~~h?@^O9G9<<J4Qj}NwnLrflT#oCKRRP#S#8^H
zf{c^lCWwgQ_V`f#HJTZ>PZ1D+RU(anS-{vUnJs&YJEZ)t4dJRSO-w;`K0Ncked9*@
z;IK5<HL#6={3JupslEJ1W6+-yTfTwN&<Qx)3CQta^MkGXL;;w%h^HHZs3NRT3z%uc
z41qxG#ozYVv-a^WU3&?Ks;_WqQdy`G2(*Y?S2^!w(iS7j8_;J+SyiAM(Q^I~q4scc
z9Ajc<hw_KUHi`)H+PCLrgH{uGe4c|W7j!{`YZe$-Se^Cur~M4$My<=|49D%P=8Zl`
zgYmRsz(e=Po)A8RPN;2lXVc%Ss+L!j()FJ>^`oKX{?|BmOEp2LSqC{DB-BdTzOpY;
zkKxUQ<K_ayRn1l{$idv-e7n1ewf9e$4f(>gYV*}<)dFEsjrzMY4+^BwY;FXj<^E|Z
zhuu&Qc#Ov;l0^gsEvTan(*%V2I)!jqaf6Cy@t@Z@d!ep65NoS?0#Eb3D>VO*9om#D
zX#J%u$=@!39JR0vNPG_YGX<4F=DR0zCl<G!D!AS7jU+*Ylf6YhEdej&*HH*_fT(Ed
z2A7?mwYXl*hG_;a0aOSQt<B1V)>Gu;2wBtojJ;mD;kCl<ZJ(agDqm_J_tqbSW<u%+
zQa*L-Oo*P|^Mr%gZ-u?`!_1T`0RM+v4^M$<Eg2<N`Wl+P%F_vbw3-W@bR6=KVE`v`
zcN!mnK#Wp)_Z6#m0QqUgyTzR$#Mc!yN7)Ur@tBs_kUrR_KV#-tu2r8p1Fl8}Ho&1q
znBgDn0VG%Ueuo2V6KJSgON!B%8Yv~X6lMNv7(qB1sV{`FeQW&@T`yxjsP2TVltaQ^
zQj%c};~~6x^L-0c+DarqkXr|8`*Zj~TUViI3CeA11goM~h1>45cQAZkPiZ<UrLzcd
zdcGSIc$+RUW(7K}Fa<2j?%M6gVF-@oywx~Uo}$wPkQm))y;zu^IXtA%{4;7bhmXp?
zJYY;MuWl2$&CIrNS@6nAgqR1<nP*bAP9J<3R|`{;ul;`T_lIS#$?`L17%830ubBc-
zcOst4&VmaqU@?1|z@LKzWAr2OW0IKSCA53s++>TJiqW8m{n8+fpNPG7kW)wBJH)Gw
zj=s<Wz}Wn$>Z==)ndmE@^q=?{D$#kM$>~nWy#e?dJ-|BrH9gbm3$MlnU7umfX^&;b
zjd$|5Wr70*Sz%F6ql`>NcR`VF&TGqJ#;mSNBnfd2VgI!DHx%u<cI(7=+Wuy+3S){p
zihLS^RAM9mO6_)#+|YJ4sa=>uB0pzS{tRnpV>_;I9S<ogHI1pLDi!Bm@_)WV;Sikh
zag7MKj~PJX^?hU$k74B<v9pY@Z3SK5-8)pK;FF0utGwk0cK9Lt2EyPP(&F=C9`Q0j
ziQ60z2Th|~O^xhA4?)TZx1!#G6f>l&do`>QOk{+c9*;c$mpu<X^CkH3h_Gk`U#dOK
z$@^Vksl^Mmf|@NQ;0Ep9+*0hDlGY3s-2CF-EAq<U6~Yn{JlJ4BufW4<v?FASKHIcl
z^c_*xFl1Q;i{%vYL{0#SykxpfR`IR+Ayq;W1mF#IcP&GNrPjb2Ihh`u<GnOv-ca^3
zpxA)yKpx>=Z%mX%q}|^kHQzu{ow5MO?4@;z_9|0PU!WcJ7zBu*C!9GsJ~1`q`(eWL
z8S|ca1LyVurB|%t=wW-?Qbr$-56lW3*bto%S|u0~IBVR6!e9?CKsV8X85qHpR*}IB
zL1;YP*i?pe9pho+0!VNpO#dqQ{wO85KV#vZkMb(mzFKbL&t5w^nHRGYj1}4Ly}5XE
zf;))$AthgW@hhPV+m2;(D@So?8ozp9<fpt9L}C~DFQu{7Vq&A(Tq3VCn0J0gOOX4M
z^ty27MePd&-YMoSb3yEw-d0x|?8I%+TIzD74D^)znxUxzuVsb9woKJmr&0id$iNbL
zUW-pJ_W)SuNq`{Q|HNX=ecM%haqDrPZp24Z*n`WHx2K;=;;@SM+^|4{G{lz*C*>vp
z<EBNf?X=L>jBVAYJFq8oqH2P=0|(mc=7sr37nG`}$p?DA$uQbB%Ic>!^dP4m<>I`p
zumdtls$TXzLLhkZeB&NV$VQ90!E3?Z7fYljR1}4D)q)#vWone|$JbnCK21*eL=j<B
z^OW;<6i;W8vnwFuCWfEXq=5#hhuKpv<H2iX;9XbUd^rUGUI}Yn_bQpQ&qzO2Hv1)%
zBlMwHq1Qzm$LJk{TBO_L&>zjz710w2wR(4AxRW6DzWSv+3AhB&gALJhO9HgA3&!O!
z2g-6w$6~i_q{e@xRhpHNgeqm>U{b`2lj+qUv+BhgEl!rs$vdG^BPo+N&?A0fcwt-2
zR*A$)<6%vyA{B8B)-gn=+@_ts4w?D(6EzDKYp13j%W2M*89+E?yT_;T3TtyteOM)F
z6)P!h-a96OSrNo<q>U6A?$&E~wf&4@H%2EJlGK`DwVKU2LZ8=L2~&pvlrY$+n_@JD
zbU3rOy^iAwVBr<2=&c9l+q4^smi&h8B8>b&LL9Srp7w*OX>U7v!dLJh0u`-iP%ZqH
z*2qwE&tW+1cmf)Xnzbe<Ac9FF0tvuFZiF#cBo4ON6D%<7plb>Iz(uk@>mYxb{M1<<
zh4vzc^~rAVE%is(ptN?T5N9UQN58GClTJQHMnKeh2_ZgImZ2a53oZ6gav3NNJ34&w
z84!%nEsn15*4Ox?zr^^g)~yfiu-v{)7gPFGKoF0&oZ~ofK^Yptn<0Dyab++d+z9=&
z*&SQR1YiLH6xDO@MB(zAah%n0Y4AKJ(e4uCu|`RLbGtcBFO20F{|HQ+J{W(?CYI!c
zn3~qwEn5NwEC6N_4j#HHn7Zb+Q#=R66?_N8yL$&a<xJfJA1SXG8C;gzS2Q`|`zZ{M
z&nMMN+CmYKFpJ&P|I9l$WQC+uzr?S|jMz>6shFE6?RB0N4n)sVICcGL$-aiy6f%8d
zP)>*@g6pO6cZ7UP%k>PGN?hyc8B3hT3hn32iCkWN9qR9Y=bg90y_j4GI${LGIq%4_
zTa48$`^f4R4^M3-;Psh9x)$xMxZ;o1h_i1(z0j6a7eSko^JW`Znr2ics<OBSI<*)@
zcZ!e(AK!pn+=(-tQxT(_+!KDb&aWC4pLQxa5S1ar210j{_}wR|(tdPRl0#bY#IAE|
zAAK`Q=40lHdKexm?81#u?*^=2DsOBc9(3}g3}z?5O9f&MVA_}3IT<U7e8O!!n=EFG
ze6$;8A*OZ64LY&cngnppIxRTo#Vo!kXJ(H|SI@qgj>7YiJH9ME{38j5I>3*HD%#O;
zO<zIYJ9HOjeOd(Shq$|DT&&oBng6ZV;vN9~!ljgdRIt%kdv2&#aXbI%&QWhS#I_YJ
zUPpIz3ogvObj&BU;3%XWx^=XA2S+@>)c98}ef!+v1!>BpHFIyW^P08Snbv~2x+k}d
zjcr1PPw2;n32Woa;YcDU9SfGnnj8K?r9QZ^T$7xh8}6%Swd7O~Y)OYu=m+GJrGaj#
z0E<aAH9|NbYa)d`Qk}am#-#Z4K3E3ZdF~WTD_}#~m(LZsDnY2<ZF{lBR~*WY8sgi9
zZs#M5#S^bJNpgr_R<cklh|E-{l5gT)^m!i@VA7MrdW7QA5Zf9AW~LrJ3FWkB(<4SO
zxaf4&!qEj=#(vE<k}8kxS!_pP@lmJmgOO+Gqta!&mMA<$Qdq<SnhoW~SJm^FONc3|
zfX0Y2ODjC`*Ju5eO{BDbU}N{eYAqptInJ`Wb@@m#?nVqYx($%}7Q$jo;W1B+l>d+$
z12LbG^27Xwk3Ff@gZgZe|MIRh&Klu;j#85p@e&xK348ZMRC4pNQTxJ(B#8>A21Vxn
zG5rN+Yqh;QIK3HUGwOeM&uehek12$rdJIUDoCE{0H!4kSe|(!cj?P=u6n0q<X$TiX
ziVVi(1qOd!8b)Z#J^etpaSIww|Fh__a?+A(MHoQP@Y{4I^Nu+#4HeiqJhM7;#<9vc
zZ0M?RkJKAJMI%x19$~wq8}H$Vp#P1(DKi-At;eRt3|>{7LY~d>YUz})-C%Y0SZ+(_
zDf3`e^#F&+0`!}BGSu`gx+eQ0aKn!(;yu`JBW=1Br|R2uv(Bp(ep5@HE%S6MD#WIl
zyH9_6K5xG|%$M+lAnrj4d_zEe+1sG0h06-P`#+H4QiXkY2xOFBtVl?{ikcy~&vZB*
z(!TTYV#DltUtV5*v9pf3mRqu%`JZ$<ZnX))-aqLxQ^UJ`J3v!7b9`<f1X>^tQRw`m
zQz4C{k)$^<MF@j()Pry;QSLFJD;V@JY_^@C`#en^My99>-~?(@A3#m*!!hC1BVRWu
zk!ks&niy@6G^ckB+8iLU_j#_i%V2K0w_R;vw2sh~9eg;pGZ}G6r*YDwB!MOKa`U0f
z?^B$>%ptV7|H#Gu-lov2P=9B&EIF}2@6WBNS_E$A!#rk&Acwj)2kU@iN{#BLO?zPo
z)J#P;UB4nVq#GRv*)u%)(xI#V5vK(&ykuvUl`#(hLXz2wr2cE!WT0~W&+;d3&fj5@
zYueeYC4_?`>>Xoj@VNOGvjy43G*tJI_PI>M8B@oc<hIxbDM$sA5}RDIq&JxtHu!H&
zgd(V_f@#a(g>^pN6q5=TUmji;6Ce8LYh^sf2bB4PCPYR>8^XHr6KXI4+irL$y9Xfi
z1hBP_WD_}Qgl^!$m>#nOhYe76E?wHrn>$I4XD2aZ-YpEq;yrxP%{Q8r$77*Brq{iM
zrQFu6Uw!<dAd62#;Z-CqjB21)fcV>lsROlhT#=pHG*iWvTyArjc6C$K&MPf6?shw|
zEQ7dn<?+IxI|Io%gEf~6X;GM;1d$Y*?^C~E6Zh9Ni#s09A3g$dXhmQDWnh}TM3O@u
zhpq5#o>`*U{)eG}KQbUi4e#&5K6qi`)u9vT)Sg_gh%w<wPI@1Q5o<;7Zvo(SPu9Z>
z2@1fAhY&{XCZ4oRQt&DZ!p^p-?k$}b=uSDVOM!3l=%D1|;OCM&0U`G&a+wjx4pq=g
zhG@w5O3rZTzkn#3V`rEYi5DDXNB~A+Vp1M03S_eP%)dXuurCjk3<98&#p;sr*Woz-
zP{^`e3K}kqTIbrDl~XHw`@z_(6%}o5<dVtoWfxc+f_yR~RRfZ@k0d1l<o=n2_;_J^
z7q)t83P#TT{Bt4G2~nDpeKTetekKQenWxVOogLq-^u?N|fv`866I(N4&PXDexP;h*
z78)Iea<Z5meHdN2@%a;9XbU@%waaq!kda<f+U12Ob|!8IP#Q^!YvJq5Rc0Gp%HQe9
zNPMUKrl0QK=^^aSiS2OJN`uXg89%HnH@jZ)h%9p}X2X~zdUh^8<rl7pF~-UAiS&b5
z8M3?mM#V=&;2etYqVkA0?}-k-ms2u|HPtVE)5PO>Jy>%YN*?#3zCR%iYp;r!qJzVE
zGCukeO1Bv7q1USb3mEbWM$Q5tg!JZOzQo2l{K=vfrVu5v>?<I&<Zw$b4qKu{tV^u!
zKxH#=`F$Sn>m60B#EVnisveL&EyBxpw`xA$rZSdEI-m9~R^9t>?g+JprYJ~>0Y*Xz
z79WH!D8gMw6Dp#xV(u&4I}7!=A<Y9l(+GzmQ7A7xuWhFB&gWy~re1)%Hi4&KyLkH|
zsL%F^WweM*Y<rd|p(GGq4Yc0R=h|ui^?+%5Rog}J_<25?2ul@dyt7P#uuNEcY0Si8
zWWzuDg@hfs-S2D9Fpe(OiIpEZ;TZWz?#sKlC#rx!QA?O@I!8^{Fb$nd!fjVza)x0e
zJw5KF>KBB^WV#vDjUWRzsFdoB1I=2IF&!x9V>e}s3fKWmRB}eL-&J!ztpz^NCFRwr
zvc0dZeQ#zRx0M3k*|L4^d3N{rd(NGd6YATF5ge-VKkjaF?ik+w$T4j%#=YT<Mj}fq
z%7)J!i7P0eeNoI3HPhnu|9BZbZ0XeNc4u%{?jWe{yh1S8sDYl{!m8l|Jzm9q;5@AP
zkot5RQr*i>bi)e|0T-u26%j(5Rca!GPF((gI4BtiH*~V1=Hm=Q{GGU5#SlqTV5k(&
zR@lvl&@fHd$Z=9Xk+VaG_8RH&YOOOKmY|~tF!3W5mb?7&C7ymxeSlZ~<CW_KKgwdv
zTsf4N+Bo9IBD?PW|FZxON2hWWjEahAEz)nd*-xd`cbi(@h<wvygCmhgqA=dOcp3GM
z>U}=z|6-gT5j~jNriL^c$77Mk?s6l3#}@TSBt&OTHEuu8xa@q817>lskD1Q&O_TT-
zUVG?7WDBDH83;RY)dj|U1yQ@Q0wLE<^cr1<S0Ip#><`_}?@<{~dnNcuh8nmajot((
z??f7>@mKj>s<50F*b*;eZ@wNJ6?!DpS=304HqiS+09rxi7cxTh8v0K@aK<>#k&`tM
zWG(kA@zfIp?x-@&dG}W?fm<Z4&vE!=lgR-uIpl@%u%skHw=N8*T_7b<zDJwQgWl6`
zJh4$RDd$0LC)!Tm;q0P?{O*;r-5;BTpzF)1oSX%t&gE1^#s<wrR5Jq^GXK41zGr|;
z)=f$g4N;#E<VSQ_H!UkK^{v1gP?m8)@w^244JX!I^jyJ>_8UzjDD!b27}vKa<O2Y~
zjmn>A;YSYWUb>@*Bm8dM4d%d@D<~Rax)x;Bdu*WnPi8(MXeGqWi2MyP(Nnm~t;n6{
zB8rsS?#E>%<3oh+G-w;&1OndqJ1+ZBmumg^Vu=N!wzo|@1xU;+Eq7x1E(TV4Tv$BT
zkIqw54d3}tP~(g@-rpcys^IC9qNYX+{zozmdc)YTXgoAv&aS86ef!4F`^$xt!W~^h
zLz4)$gnS0?TgRCYCGF6^$*1Sn>up*0QUbtGSh(xgGX}@JwammGQ)hqd?`&C4jxPPI
zZr%ZXd<eV<m$NuzQc6bEpF*5}iJ|_KlEGd?z#;y5I>H}cMATyI%&9RUJt)`yC68!G
z8#d`cK`~C!WOoK0*SZ${8OVT~;Lh&uHG4ESvvvQ2do0}V(~@TmUX=b2hg#C*PW!YU
zetU8N2g9TKLl;t~D<VCi^BiSsTF?%ac*$HqB+#ZU{g59uq?fyyIWMW#h#8pt%*phF
zFA&P+Yn_7ub&8TguaJl4W>!B(4LgX9ntC&(l|YMw4PF&<fueVN<&FN<)Dcz?<LVm8
z6;0M1Nv$f#Z`4wHc}2o6Aws`0U|5Zr*GAE>1|3BynAE62LuBoqM%v>QFCJV4oR{7=
zcR7mSiSu(Jjw$}~_|Oga5>*Wla5!?eLga$;2g;;T1vJZBy%+??^TVqy^i&{Gs6KlL
zNwG16)SB?1Cbk>9^f|8;K+75Ar^&sEi>d^%?^sw33l(Z~ilCN#hdY%c<$kN*0x#>~
z)%>}2n?YA~`zYTOMR`1c)(?9i-GCxZLa=_GjHlM;*d(C3a8VzI=AV=@XJ-8NSfllf
zWMn*pdEh)DhVk}wjXQ|ct34?x(P_oLd8;_XG3fYVE|OTAnnw`3G3RW$!^`DpXR2Kg
zS${4!gL?9!Vfj?7>MI8ia@nZu(Z>&xjcTK&XlMT=HWneAo=H>UOg9%+U@DJ=2V2au
z39$0#Ch;sL_IEPsJ$={l`_8>eCtK*|U@LG;@C48)A7^YEkzBDXWX~QPpDc6=OgGM}
z?FXw>8c<1mO#=#bp(ql>T!pTyBDmd<T9H?$v47rdV*mZpF9-X*C_Rff)~%k#=q)%B
zsTuY0G>V`9ecxC(#c0-63w1`}*z;~=l5P6Lzes##cK#^rH=&c45?|wd^NUjpWW}Pf
zk%J<PU$%`_>BS{9Hb!?9CC`@ulReD%?<u14zfd~&nPXC9(<b+70Hlv;u9qy@x0%N5
zxFRcefWoH-(QNNCuN6Fm9oew|dm+^uTVVA4@i5QSF}LF5GjPMTwOJm|?dv!1Pa@pM
zxKmwfEu=X3r-k4O>NzpL1RxAal^7BGqc#x*KwVLUzrJ8}p%hdxdt$}z<Syd%d0cWV
z;jQu=s>y3oPW&s#bBx&vk!g~+f%-%CGrkS+UnB!_XQe<BU9ZS*l)nPg*`Hk#-6CeK
zp%XmVL0KJ8f`X`-epcnCbf%oU@E#&-@0wqHW8p>{oz~y`p|kP^R5d2L5oS0J(IhZg
z_eBI=L*iqh!y-l*J~8tVk{9NzC43d9`ZM-8!y?wyZq7W!l$uS;MM$22l(LHja5q77
zKwW!}L(bd@u61|Om}Fw|KV#*=10d`G8p3guo8K`i_F28rY@YnsU9!bBoSLMiT^z$$
zUz(!bm)uFJX~Ox`V`|1JEG#AF<{H?&K2d!?Q@p-=+?qD;_I?945m!r9tjBzcKG?po
z)hgx>VZqf<<C!nG2&!K;wqOWOt_N;TulbSC@pEfR><#35H_Yp@Y%<^^B~f18yx@_2
zURd`)yeunsNx(Z8Xh&2s)^(=L<ixVwFElz^m7k#45IIOFwBi%yXQ#(u%#rMV>3Zn%
zC$6+9aM37?n4!xN_P4>EHCFjMFTuAS+4%66ZSIA0;*3jW_EgkB6|Qkn06(4rNaqM&
zsFyUgGV_FEg0X%BqNm<f6-o9Oy4SvB1;)4_;_wnc-Y=0dhokfM1quqw7KRyf>3t)6
zY>W(7hx)~dd$q&&o6dP&H|eX1KczkUerJ3&l`roSqs%qFS<JTRu=D;@BU5t0`yu|9
z9FL!Ak_i^karaz>UfbUhI!zp{OoPJ(`mWu3Zd-HD{Roa@EsdP<LepGB1unUO`jgDI
zvGCVoVgtu^q00uag1q)U!J_NmHuZm~H_XQdU?wh&BBM*!JCeT3bO^VzRZj6*No0WW
zu^CX^O;9u2?KUqY5bdHb5b-vY*o1%FtFcSDHT()nDeAp-%e?bpF<sA{CiN0V7d>v6
zWf+GP)tH-6tuV!f$x$P{Vym!=-j{`G<D)fLj~OWXZd5Y;Y{w7Wp)y#fiWd=Oc%OrG
zfg*((`e?=7gAz%jKta)}<q)SKnXUAz()aENF?>aLsh*O9UiVn)cY>I4l(|4fy-@C0
zJV7TcnS%+x<B>9S{+lhCQ#!}s9%n5U9^{9r2`R<xL%s?gUEn|r53@~NjMgQy5zpaL
zv0~8;06BrPa0vZ0G#MU!aZ+~h_*a=SM-rZ9JWm1YRb>J(XhD<eLQUay(N?PVNDIWi
zb4wgNUY&OoijT+r{I=AtsT(^_LCFaCxK0JpR@j5{jy<<FG{~8FwpSgfere$#Vs*=4
zc`h+yi31Q!<!i2?Az0Z+*+zFQ2$gkQykQOj>`S7Al>$9o9@od(A);NLp5><O!7syL
zCT5r_(@x4@sF{oUJ$`C-3-Xc}yK#gYI1i)FWXR@%d<lwQ?0RD$<7X`obl{p*s5?jV
zzwNc*lf4vs!}R}lNGF#$3lIw>rVKH2&JOaMq;y(+mRnl%=Q7%}IqpTX)5fG>0w*^b
z?H-I7J5_w}*p-t6V<yQsLUGe&qV#5Biv-}!S@d{q80JoNx-8{!e+G02)G=KOnZY5e
zzJH0hYPF0ziwNbvs1iPxq}8G>)NYP-@c9pvIU7I<eE@u5uh_2;Fr@YknMW8HB28xq
z<r6&h_n|6dbV6S&4ExBFFcJS`An|-fn9NV{LP!0MBWl=uJ~%nt9m#V|XNEL6;&CPd
z9v@%-0~j)LA|7a=mrl|QtREC6dx+j%3Y$8Un0&l-<T7?q0d@Qia_ElbLf6=tGVB8H
zaF|94ZROGXL#0%OHK=p-T}JWZ*BMdpSzBSn<7zCKliZ_HeaB9uJPT^Y%^prrV_0&l
z!}5PS_4<^EZhV)4K?eWIbWS_iRfi|OYz|D9VUw`USi`p(2Zw7F_;3kuwmmT-gv=tI
z9732J03poP^&Wc<P&>{n?n^3!SWxukF$@mb`>&wRx5|kRGpF&=wHZZF@T@vY!4=%5
z4KzL$h9Egpd|8rU5{e0BzID(rZH0tYRnalH=mqgT$jozzIH>$WnOwAzHY6$Q)+VbN
z+<*NN^uAD2;H**PmC5Op5GnSdj~vpm3!PSl2OMxuB~I!MwZU$gzQU}>NBq4aCj2Xa
zD|xwD)R7;$pcY`^ffJerxI|u_KmM%cDaHPYk3kJ}*+Ek}9d3B?gJpNa9`zzb<TOa1
zfzqrAfh?bPbTgh+Y8h?pf4<y7t#%&6976-WQ3DXd4y)Kh@54J2bbyDletM=Ixb3VO
z)1^At6?%u0dVu$G$7kQzQUw<MTa%H<I&G-pEV^Sa1m5_<AUKBn(?hPh!NnBtt)dBW
zmgu{RAMsn`nb4%tXcY2vX5!Y)&06kMZvQ6I)10#rz1Y=iGruFp6HPYVM({7887iGa
zXrd$BJP1ePb;t#$BPBvw1*mb@0i+rXhqqaT&&^vy6&giAyY}BG(DgL8q^rZG4BDa2
zqT&8~T>SY9emOXcJIm|+9qET)9ppRe-tyVB+c`;N1E!+L|L-yGsh5LWY(-Nf#QCO5
z&JvLeT={a#$&r=uY9q^5(Y&7T`u7E}Z~swVUU)%2U;U_2XZlnyGRRbU$Yy{@MO0X3
zTWa;CcnatWD5yk9X=YP?mKvb{&H<85rD~O_84Og)fPSAi+mndUjy`f`K8_M?s^070
z>0PEI-=l{mO+ZQe@|YcMOrddYtir6rCwg(Lo{6~=4mr`JlD?+<m3dB0p?7k5C6`U$
zED?2*dq$MFe0Cl7y?G9){(Wfa%%XYeA~h`>{C4tRWsELanw;RF8M05S{HOI#hsg1e
z2C*)W0FuK((Dfagb!-xM9#oU<hu`f>-+xE!GL;?qw7dIy#1cDR#{;RUn3Z<;w{>gt
z_bgUa1n&ynVYk0N_%09!N`95vIT!q{MO9HGa?no>W2)d`kWcgh67KZKHP9x;$%EWG
zB0gc5T%0R#k6N0fwqRo*mjT;u6MBs@<=k`ZM8GSE2K+DUS*sR{4Xr9#YmsF~2?e_4
zWc}`w{WE%JHNWqJ4y3F0$i4>_!QKPa6<Y}S&@F<%$Qe{*G<UBS4&kyk6|lhyqBhB(
zx`Be$anf5Ps_y#QIk1G6OwbcywwOXRYO%F`!>aE`c53b-sKf;if`c%>R|m72u1PGX
zILDcb+R{U}9m7f-)bxM{gQBGafjKTJr#n)?u1D$Tcv^z^S&TcxTbL0(zf67osu=Op
zW=0|ZBY;lt!u(N@&bMR*N%^$14iIiZW0Zw_ILa_;jm}qVk6owK+yRTRhXtiPF5IAk
zrjn9+ti$i@^|PV=ry#A#pEZ?lcmoZ={wtc$A5$g9P3nO-<|CSZ$a?D7GkVyxL;yTB
za6Bh`Q+YS%p#-4pp3|}dJgbgX6O1N%%h5Z!#m<&=E~_f&cPkYk4VTIbudy>a11Fe>
zkCnmUnKD2S5t&40@Ic!z`qj8Wz3<pAaPJHTp9EEqpFs53EP`$>#PH!M5T9(i^^^A%
zo|cYBZnQhqh;0ft!B`BDl<#j=Q-(Lqm3OM!eN|pTuQ*lX#rZM@mSN=h5<^6<<%3@E
zpuy_Sf8;2JO`Bzi@()jUN80iN1i_OZ*K+mBkSnGoE=8qNx~e953}rTls4kul8gh~h
zuh(hJb9l_jP1E6+)G*Uv<+O9=SR9nm2!^)pIeRnp)9m*-zc_+HYIS^9x?5_cS4%zC
z#2tbZ#GyIOevrRbVtsNC&?S~fX<Au3PKyOb1|4bL+n+kV<J7H6#gB4{_8hUaS1=Rb
zf2Q<bI9uAC4EwF2sv?8SZAKF#NO|iUk8M93=a3ay*y8l44%;1YgTyQF<r9MxKRR(d
z@t(M711M!Q2*+!_(WAdcH2sLDWJ%K>Ii(!>Uu0!GzAyH5VXZ;7LH3MRsaSPC?J`MP
zz>%dK{eXq=h&I4$=m+fny<}J(0=4u}$JQZ>z!^kogfIBcRB^8ZJo|gg_IW@a3O@^l
z#RYmdm`MBKoG1K$S%C1jXv7&;@4OOe7>fSs9LQ$9^}xm>T5?WRJv-`2C0yzWIMVg%
zb*-EJkKPMo_S8Q_FL_(gLEu?En2%T*veoN3K_cCC5}ZSrT?wb}xD&D7NI3h#OdB}q
zG!=(|@hm<|k?2K647W)cxtgbFZ7+$Z^{jfsdAp&R^~91_QZ?)zGy3(JhoSrdkEi3%
z5v+x$MYkgDE|C)SWw|@U(?wxvoS^Hvm#vpgkInG9<+HF38hH4oj4u34+ObGw6o>qi
z8QG@(QX!P`P8hseX~hu)(kxgyQ^vy6F!q4t|1QZ+>@U+*GKfeHOr_Is`CD3LJ$hF>
z>jdL_pMUe+NWH)Bqf0je&F43i3=|V$m3EvaWa&J`)#K^0Oic$P5af27t^tU~usIz`
z6Q6I<HGV>Kb*+z4MA#8!@!1|pnNynsdo-Yg56r8dMzvxwv0G56SxfOR!iTBN8i(c0
z#`HKF0!S3JqihRKBhZ=)^2<<-#~zc=q*;dlwc<ScGv*`(ZFKnIra1Iq!+DQ8{=%DN
z6vVkMSE8Q?rbiB(P^&yuFK?z9zoGcGIA&Z^{l*x&`ct{M3TEJj3*VzG|L0j~z{$>x
zEqc&SQ9a7@WEBGN<sJQIUPM}sB$U<9UfM^NkY%3IarTBQvM2unF#|`mk7{Jj5vkQv
zrkW4sLn<K5psk>!Ea>8^1+&hVz+&vfBEG!Tv(UO@^5Y23Y-R{J|7UxMhc(<RSN4+p
zf7B7bCdYt{ps{i-9uE3yOGKS1At@kY84Ue6YpzCN)NZ0q8shbxhTR~vWBaCtp{4m^
zy6cT;E_N!nIl8_yu#CiqXe^Ttru)0NsC<SDQ$D>R(Zsz?`e-SBm)wnD2;oW{8I}3N
z4<4%Tz+Kxe^qUJq@bX{od`uegl3wTWfY=sq>uZi%6J@?vNhg@!<|642$To6(UrI~)
zo1&U;V3jsVQ#R1t4%o!HEoD+`*hubI0-#x}Wu<i~1}sD(CdGY9Pz`;2+4Ka3hz*Fk
z1T9JS_dmz=oBq%1Lo$?J`2^&PhrmRXS|?(LeP9YG=(h#eI=?dP2*vaCmW${nxEbT`
zW=c^j(!oSeOxFf}azCoJuV^MHtVCW554^9qAu(BD$~^SdwMwK$M1Dk&MpwF6!Z$o<
z)H;FNV8tti+mZm!AuTK1%F;>&Vk@hl;r?|E-tkhulqpqDGSZ@^0?zP6s;lH0LELN$
zaY~+l;lVMq5MbKq3c6VB<R?N_;q36J!OMtSoc(tjibO=XvO15RHo)dwEEBi)Bbv3=
zP8J4Tqk)oqBAhxE!>w|SS*+Ajh!m4thR(>nF~qoOv8Q{l+n7B$kVqEp+<B4W+<EC9
zA1PrJGh|}?o^%=NZ07*-Lw&04&sk6Oz-zoiumPC;f|1t60oB2$J!5O*RcPj`J0i;A
zbPNy0jf+6oo%fTXmZLlQY1apQA*9b@N(aAo@N=veGx?BZ9Po+*`~B@B(ESwDa{)6i
z+?!<BH1|-#|H3J)rV(fTiVFR|>@%pviNH{oFL=aUdq&YA$o*CU$RTwk^E0WB+-xE+
ztsQ3UR5!dE#HkL^reNF-)LfTUA`5qSr#)a(NK>jr4|8bB%9%$7Lyyt(U@{P0NpF95
z#+Ou{og63QuOhO15RSl~^ziK*UL==yAzg%jgzjAsdZ}Bo<C<uf$E=Rhd{zi{`bfI$
zc&3f{*Vmh4b{tP#8R++>)B|tKAjg<}THcLbUV!d>*SnP?CqDK1v!?oSY-GnE3W@df
zYd55_3ceBrfJ$@u{ETEu=?b+^tR0Yrj%7K}6SsvUqeC2MBXG=IZ3{-I&6y>Ge*bn_
z(%!rH86i<04v@y|MLp%S#Ep?ST;=WV@T?G}3+V-xI1R9b%cp%OSNxSQ(L7Rk)1)@|
z*S-dT)4}XXUNBspS6GY|jkk^piMp`+2&VZkn*8`hg1@ebtF?jFax&g~@d!)Y{OUV_
z=MNa#3xfX`8a852R`sBLaoY%FziU%D`?XTCQ--0mz1Gh$!w#$O->I~Fii7Wl(`v2q
zxDM(?<`lacZjRj{^?^yH|LqX}X~tGMzwQ_haa`RR{ebcWwR9vhdmHSuGRW$J_Ji^p
zC@l#D{tI7rczx(fDiJiiHOA@d)IwN<rk|GdDY|G*^{6|PxjmG&L>{tsE^xualbOYs
zzl~Y@lGN*K{IYfFj7?lmsi6EppLA$KyauyHxohV$Gs88POkTHfpwd`K8tW(3^C%Tg
z&P-k3l@F8xmXUDS*pOKRc(;oiOtq3ft+`v54j-ruYfL@YZtz1wV#y5vp_m2Pw2XHm
z3n{eIMSlKm>c@vJPn}(2WP{LIPweamf3)GY3wXcPSMWQs+l!hj;~eX4PeaN`NWGMp
z_Px=_hMbxI!pnYKt|l>5j0iDEIz;+TB?5vNT;8Td3139}|8p~Xjbe~$gD)K(p&OD`
zV7)kYDKd8;a>9(9V6gx9uMGmz{qA!Q;3<pU&&O(xu74kv#&n?bShToFe0@2EuUTzl
z`>wsni?a<<*JbU*bVHf~&q8*L8W(#GE5K>-asgkJ^j#L_3Ohda{0O{Q`v^Wdhnf)i
z&gx0vs|Bz-@rR;wF__5^KK1Aqd}D9e#>o5W0M>SaOR3J#P`o;tBU9)?SKZH?aUaKn
zKJW_&_Si9nSVZ8W=c6z}Be*Z?-0+zuAsx4Dlpfgz2-4c9WPb1F-dJ+iAYpn7Q#LZ{
z-AtTzeQi7zwTdB2&;{kcQ_<~ryEGhs))}2&z1(1KV)Rn~Pxkiq5m9E#B$^Zq3=Eq*
z==oH=)b%ol%D)T^I?SgBUKO{OxT6Mt*)xPV6HxBbhiP;rNovFWFb~3ryJ*khCSBUy
zYL?pNCKMAbeh#K<ISpI=S2PUAdSX`>mYG<Rhwl%k=+fJM<WXbzf_Cq2$!sK2Eg+he
z!adHDAHj;NZvYbVb!KWJaXdX4lAJ!3ZlN$Sc7S0Z?e<qh;1~R~;pNYyIPbRgn{Kql
zU4!@R9iZX=e9??@b(@k1eIgQ<#w)~!2ci2ce0<PzbCigU-k^(?H{SP$&VnWVg{VD9
zlC{rGjfLm4lVO924TT>-vPD5$K27wNNQ)S(rm(i7bjbdC=)YgjOgNrPx(3zRgfzqN
z7i?(683mTk|54=wXQF3yKvG|im%Z49u`>@XeCXjow;?f1-CEVe5hpy>?G4~F?mczL
zTTV9X*s8Na6szU|)IUpGIeZ@w?b-)6)WwnSF_2zc6aE_Jznq?<IEdeIu5ra^5F;X^
zQXp`xs=@Ev;*P=575gW(O=hfdRx%#D%P86O_hvc(Y*2eqp~WCaIc878U;WNknE$ZV
za@QUFWufkM?XvU6l($4KY+WP{UP?0;zY8($oQUP&#iZ|l?==9m73#at9oa+`uhSUW
zT4;l+Os-ezdH_GlirWSZX(}{1g_qGC)32CGJ*c;l48LwxeT82A8qAf2TkRT#5BkT5
zKv1a+oSJTTtIV2E%Of|1UQbGpnS3l17T;h@>!rOHvE-zPuo~>oKfBaYiDYI|;J4pQ
zvAmi=dg8HXqSx}C|CX(>PuA2BjoCl-AKSp;|3qi8#x6V^4APtYibMe=ha-yfBSd5$
zw$6`46Fz#C>RFFj00d#snOr8OJDbPH!P6%q*4A{8{d;APJnE^4JJ-ECSGnyg@Np{&
zh8H8Hax`4##G<E`0hrqs-ypiMg(yS}ASjs7!MH2v?o$n!Zl^(~zuta<{F+}sKf$((
zg=(@{=g9>`NMnN-3&>1e4<N}8B&T9{Qt^m2v`siB2C{D$9WdsjZ$UCP2s<d@y9)_k
z7>epWQZ&J+aYTQ6JwJpwns7OSg~WWRgxsFk<(1-w1T^<Hvp4J|o!$88vr9I_rX9Wx
z*6S*68aku}Ay$zX%u{S2a@ennNKEyNajSFqT$V7!RNM6!@#w5YHa}>XC~2d9$<Vvi
z@^|eZ@nAqxJdJi5iEu~Q7m2vnSYGwJ@;C;x<v{ufpDCVojz0Oo-Z)rh#mOK3^fBhY
zgI}upK-jfjMC534w;^2(cwyx>nPUVnq9wHQQv4p3BMJ0-!VlhXn#;wT9vD^ly0&%@
zF%*eC5aShpur9JT7EH&<b*$8LXqO8L6UiuZA<pWjx=Df!RTLAjCn|r*CGmx~uv}YS
zs1q~tTFE>2f&RY!huq%cNAdSV1)A`tB@ix2czamaP(Mo6Z=&qQZlaMlVuuWPES*`f
zQ}^XJ?hAk4fyNh%PK3M1;%tlL1kP`P?j~PfeGJ1{9ep|F{?8>7Gu6KAdKXJuQ=mjO
zfjIS28<6<u(L-O;{t0yS4}WR8BqPQ*;>S|6FZ9pr?*xs3KzBQc*Zd1$zSZeRL6i6K
z0egZUQV#~YuGous1ScLJuK%|9ZY<uW?qvzN#W$8}92(MYJeB-J#m_M;?&*`#-n!Nz
zp~iO+7m&z-Xo@%$74>YaV%#vAp&BLMye6=Yo)R+qa<-ztgB~+4$@2m~lzHgs8^j|X
zk^t*ne6=MBznIK{Ta?LvA(u~sHGkJXaGNvjvwyr!FMC|$8_vzp$XSx9z~7;?YY7;n
z0u)=r!HzbJOaBO^{c07{j{IIM$!&%%dY?GMs|yiGa0U6Hw+!DAW7>9@IX;QnS<LYL
zCd8Xu6400p=0eXzIphox-gZ-(Z`=Vhol9vap=Q!%rg*@iQ`2k~ofT=Nr3TVjTN0^0
zy5S7lwmY?Sz>|_x7QwfK_<fFVTODF#in8TP$no~h?A)_&@Z|cUC!*Q{Q)2(>chAw^
zCl%)A>2)Hb6(ugAI9C^sO%xvH)f+jD{!7X=?T1c~OJ-P9_|#3;S{{X>4s^je@GP9z
zJv$oT=;<E>MIt&GJ5%yT|EiLFG<d?}j$vlx=}tp{1RKO+GJtx51^eM1Cpmcz0r{5x
zip23a3gX(8b~L+X<;Lyzvsec+@=9OnCqLyo0n-v{SI-3W<M`cfRUbN_rU>8#8EKf^
z%hd2j{(lzWEpjHa(-Yl7?o-FgJAlQOybzfUk(1ks1o-Vm^xu@Ax{{GWa_)=Kvw*s1
z;kK#LL5t(YMqa8NVlbTz9@UpjY+HkuKhtPW?5sIu7cYF1=V*imfo&(&wDpb%p;wFS
zy;T|z@pT6UtY_L?e}S7%K^*Zz83KQy+ljfoo*^}%^WN$bhJvZ@v;u5yGjlJ95vB*Y
z!)MuJrmbJhdG}(-J@?Q4CmD~57#s1UuyBgiSQ}yT2N^!rSf3A>f)3B=@32GUpiBEY
zJz6i`Vsu6d5>Zdl5l1*d)-XXJAXqo1Pwo9=Yb#4DIXUi=49>E%ai(U3Im6MmYpWi!
zD=6K#r%v8|hRH=-B#>72w$bvb!qn8?%Xt{0wS9%MLsmU~1=Mzcre6k(n5WRjZi$RJ
zzX`8BU<f||)<s0bmBqs1dC_7s0&sQchiS>S2hQZE^UKDg@!gYW`*_D2YD{dfHRKhB
zIM^fu3==WU`$>hG;heEJrC|_1j8R2J83%UDPC<i4$}K8uA1|&r0e;QE>eEVbsZ_)c
zTJ&}EWiiuKQU-nh(n6OP0H5}?i&{G(;<#?9@bbsBxV)c(E#14NDVBV<qQoDzcFC7H
z+Z9<W7hCqZ>smA^Cx`-AtAg(P7G9>s=h4^VbM*1#{5gAF17@5$XFUKl!gb8-8L5}|
z(qf1wAr&v|koi|s;m(hB#gSZ0abV@Qnfp#87O(4?g}tta0OYw90@wYbBZpw`%Wp_H
z79kkXp#4!ie#bZ5;{+~*hzwZF5g)@)jShm8Y53vF?m}mU{vTe9Q{@&e@bjHavmepc
zY-DB=K}R-7EPW9r3$^c+0pThQ$bA9<aHMM=H^#HCD=PL(ohi(R0dK=Xfe<~;=^n{x
zmJY%uzQinC2eOfPkpyREe9r1Y{6KbDs0`U}(aCnVGH=i2`iXwN$_1e2K_BK6rPIRy
z^Bo`Ne{v^7pA@6V<2E|=kz_FAf#&_13yQ-P6+MID29Juu+yMB*4{60~1AO^zqF3g%
zbk!+VB$+47argdG0l4~BfQBVy%Lv?M?J}k`lK0ehZrkV4L1HDCVcXa{;5=d^S%}hC
z;1*1DFOWba<*Iqgd{Y7%87rwq6d1rIA2W{eXPJNq=kcAQS<tL6$@!@dP36}#P~dG~
z3hYA?V(y;UlLbCV#V0iZr}wsqZc0EjEqGsf-Uis}`%eM)%k2r7f}c^g=f5@3x@?9X
zJc;3P(`P1!E<R}FjhFW_B0r7_+#Gfvj$W!_bR*BL$zQl6w|b!Sj(SmxW5~n`3h-3o
zuJ)}T0VLyb7u%f*8wl>Q49sZ4Obr}huqj~&tpe6OVU+9{gFmBSCMVdlLhMYwQ3-<W
z;5c@}M%O1R&E{35kyV#EcqIqA;-`!b)>l`dqd@rx<@h9Etk|bB&(}9m;_9=&;c_u$
z@dh`q1R#;%xxss4IHc|GtYGFYh=Tbkz}XGaU`;z{CB+P$V*%p6GV^hf)EcAoo%JrK
z8qh`{6_I+g%}f$dWttT<<^MSqmBzsbv|&!hSJ0iFxB?KqPtARWoCOcSKP~OEkhKx*
zGg-?yWXiFDAFwO#9bkP;2%I!$CXC+Fx?{GFJq&9e4o*51xbewDis0XjIMxm4oxf{i
zi@+wktdZ^)D|LV;MpR5RKJrfs=X8pOBWj7@K9&X>*%=|ye+_Yga}b&+#`f-D=*J-c
z0@AwrmPS>=VJ4?nw*k6z{P=ph=zmku?Zx%M6%!?}S*YS_PDnj5l$6Kh0c`|9MvxfR
z)_jBK%>G}aH0TTk3FK9(jI~MVcCd1AH&&_S_AU}F*Bhee!N@lGs88NcVKvP;RD^4!
zC3aeJOZ_jX=ONA}Y%xvgWc5OBNTW~67OR+lyn!*jHD6)m$6;q#bIT^0Qk=K8yhE80
zLa+$MA)%pu0i2u2pJx-<h`-3o;G80mGXX=NqG-i_x$sTmun|V(k$8UkSNm}I7K@pI
zzJj(9l-x-*P!(T0H+vQ~UNDOjAR7bXJaPqtj~v3>A2|-)hONO7>`&>r{{dVUUjO?s
za^eab3Ab{@--cqLbKwKhd|h?!gvI9Krv2-(PKn$md4#pOim~4zkqx_!hJw(4OX&YP
zieQgw75{myMJusgU5!e5uJU?%J;pNCktohx&FH%v0(aMqm7FQLpIi}XYwP&C-;N4)
z?$&<JsJqN4elwYSX^xW(!Q(Z}#P$o7Etead+D2e^f3%W1?8sc|5>0Xm)2V|^*}(Z{
zMJ$!dlCIciR~c+j8;w%rtFB{so2cSeW4SsbE+%&)?8u93{pUL9+`3qCIv}MP=U;8J
z6|}OG3z%OZj_~t}aV9TZL1omGfGiHa*fC8p%BUA61-7y4JOXG|W_lVc5_6~i#s9O#
zMEF^oW8bw~Fos;9%yqMr8nGSWW`QcX*%Z7ZA0=-nK3TCMct0kjC@lU04X1;$jVWi&
zt}<}XWiARqkld)=)>U22nZ@kW^kuIVH}v7L_ICNQ7ED1UZ;RcONYgTFM}gM)>;F>a
zmcB9v_gNR_VP|DClF@`9S%dKn5cd-p`*ke@9=3z+OT(UcSX74TuY4M=tFhk!AqS*!
z2(y2!ga0Gsw>`^VIe;Nrr+f9O#4kDu7_~U*|0%rqStPwKa?_CO$!yh>I;(|1=R!?<
z;*aDv1==LWa;;0=-82NWt$C9{Fm(bZ&kV)PAHTqTRPngOFP@f8zuZLf{vS=(6dnn<
zEH}1o+qSvcIN8|TIN8{m*tWIV*tTukwtZ*MIrn*9=9}s7@2=|VD)(=w9HOww&P>FJ
zh6=+5Iwb9_W%zbPmHrtT1M#*@O2Q=yVkuGsSDX9+Z&jm#3UeU>*y3*FIY;*Zd#|%a
zO^rmnknlly{8G!1E$rnPsY^Dt*?h%V0wIG<@dY_{mu3stY5Ox4XRtXoBm1YH?iy@t
zN%BwaZC&BOc|pi9I5b$u!=32!dWAwd0%2nD)=PO|7QrgiOg)|N6js-()aGvX@_D3G
z%zSgH!E<Mc;hbm2uY@xPP@Hj%!`o{DQ%k-*qitq*vhV0BNW2oJM$Ks25i37XDVf@e
z#2XgEiU3WcA|G>R8)6#yTSvjOTkq$2T)Zq;=f;%S#uKWWp5M#@s#YNQP>($am@b0<
zA*>QWt=8FZnx=RPG=~#G(p>J^F(Mfxo4fCTQ9wX`7CkiV1;kJ-c1LL4stPZEp`2KZ
zz8U7x6bTVGMdDVep`1rCf-{Jm_;=h$EM-;xSJT$Dz_6nbOkD-8m&vw(sm14Xw9Qtx
z$rijuu*c%#^yT!7Q;xEmedEp24Z{DllsfgA0t2$~-DIsLf1?rz3i#5L*WjtRKX;dX
zrCY|fpZlk_SFXu$gAsoSz3m3=^^C+->pw+pi!R_q65HRa{ywGL{;4TX`mZX0V|P)s
z!|6A#t1nnVdB52Bm}&(fZ>6Wu^+sn84tsWkDRHl5VJQ4r9MXMyNGvvx<T8kMRFCAl
zkwM$k;Gp-!%`o7Pl{F>0aRcwxhL&6j5pL8YCOzq&>dTPGb9O?`9O%IFkBV=7)4Zlh
zASs>zihiJrDYml$8&voA-zG8?9WngH+{LJ9Gupz&neUHQs;h0~c?E0SNGVGsA2ert
z4o~^<|NiQeqQV7L*yrYDd|3{m0nNPaenI%YkAh_jm+F8}lMT^w<7V<9TR6o1+A0EP
z)pmUwt-R{6N>p2!?t<j=@M_5a*q!fi7cL*_Y&3-I{Jg&ev(tm6k6(J(ZMVrC5dVsZ
zss@`uy<JJMlA4AVldxcu8jI{N--I+}O()^)$)y{-eq>ch)@TmjRXO|cd41shJm3u%
zhjQg7hFNoCkFY6d4|_p~p{^h|aH@GC9ag&|u1hs*)Q5~EIO}?VoiAPP(~|5T9X#iY
z_S~H#sgy9`VD40r4mm!lm3mQP5Omz{;-SqU>@>S*YSct0o9eizVy*6rU^bzIoz*}V
z5uXep$ci(X|K`tjL>6|u<QR<_Pbnfj?yuxa>Xwz@6pkvw>N%53N+p39s9H@Z>CGam
zuA+EXnz;@5y@2(W{R97M7;L2@hmUPqz%puAKqQ>$Tq!WTa+E_qD7Fr~6Um35mV%K=
z8d;?{q*DOv+y4(wa%B-_mhp3zCSBH`g8V4<05Ad)LLwt5yc`($7B>vR*}^hpmYtky
z+gF0|(T-~AMi3atBX`>FTMIS={X1umS##z|3{Fi+@*}C*ZBbEezIrndWgWp_h6>Yv
zf{cG{QxyoIYHkz}HhDFo>@xsoX2pByX%@qgaR%&3;H~66fiWZP$4uxa>!;U$Kr%$(
zq{CehQ}2%Y5)8T?wew=EG~8^hU_P%OYq}fCEDT=iRlUp($oG@^LB66(ETAIuS|I65
z)QNz~Z{$_~6hrMd-B2{MtiTQv=0Plv3B%(9A}~`v!@>!_L;)h_EMA??owhxCBPN0i
zaZQyc%~kx75mghAtybs5+{KbEPkR!Ky+(qvq0d`P^u@-x&$>Nlg618nt2_fy`ioH=
z0Xl>lLsm=$`N1;JvfQ{x`N}U3Kj;tu@Pr^0`wG80+D0lQ=AEX(wda5h*_N+qvAp1*
zvIy;_cajyp*jcMMNP#g8dw%$BRQo2&{3*9(Z7boyV&-yXNQfG=sV213wmuoa$4B%v
z{t%zT*G?<=^}N`5q-{q083uG63O0H_2;A*tff*|y-^t4`BX$h6)f8k!!KdmG4c$m7
zhG3+}6g5#cFvUj_WPP}ARyt9ghixfKh{Q|!E%`*G<lW()`Q!H_G>Jkz)(>@(DIFxJ
zky!kE#o+NDv|L_5`n(fi?CSKS7axBT6U6=Hh@=<`zD<wC0Toal1myQZ2<*0lp|{!0
zL}asA&Vj6~>2yYT-h~%tvS7dmGnI(eZ=%AmJ`A+$?LDdhvXy)im)v+BP50GRkr;**
zBqOWx6zjal9^?Q!)g7{35;-lL`#UD`|0q5p>CpQ?aaTL2ZI1J?y7pfgDXk5V8&bPy
z6%R}x^lwe$WJ)rNw(+r+M*ZrffJbs~_W?hs1fMntwtMr52hAYqQ-I0|*+jA4iAhp^
z;&$+|RY2>BR%p|Es-8XDf2)vrkjU<eFX-xPkC=YLkK+2*>Ftj1ra-WHW775!m$?z8
zxk#u?$x4dy<ZMLe$6wpLjyCh#c~DeR$(8SONUTqv=gG?NH&K1t=K^T?0ihKV2;9nU
zs8vT^@bb7usEOnIaHNrktZw?==s=Aq`mviKGb35~;cWKqPq<9B<Yj#lokfxeu?R%*
zr<x!5ovuUp{IBD->`pJQ;L@bOS0zAVHg%Y-JNF{*fg|u8$Jgt=yfu?8s8lNp#I~X{
z9ilqNUf2ZH^=(&0Z~QxWc}7<B(xd83^z0^duUko^ggZ>LY0pX!?ndX%k-MzM^>c#*
znN#r)&&~A1+_kT^LT}OOv~vP%(oQ%42S@37zxEtu$o|dg{cex7-pQjoXJusd4)h^B
zE$=xaZmGzw#YcwCbublR6XZ<&_oCYosY?7D$I%{~)T^{nsP7b!q9nnDJD%~4CYF*B
z9$LpdI->)5HJTdRXDOEq9`+>Osb5q&o7bKR9A|dk85xyIjA;#|<F~%_U9i6#WlqU)
zT^rNX`sMfwvfvYayWJPKnK2Fh!j_7kNdVf#wh7+zu?cnA7Xs<VWLyn{JF=NB%Kcw&
zGFQELp2<S~Wpx4k`Q^u=9H3GM)Mf~cS+~M}d{U1jwR!Bj{PpB7YPwEvi_-?OXe#o3
z;c-3q4r}vVIxgwwQ1SFR0d(@$A3*D(cIN>Q?JeSHunESlS2_{9zM%gq0vsTAm*jxs
z4=B?G{ZJ_<QpYwWR<Fc^?y^%k;gT$ziQ4Mn2FS<r8}lhF7!MPfFl}vrJ?QYg39o$X
zGB=@-4yQfho&|8IZ#H$ntUpexc4-HPL(OfquowF8k`LI92Z<go1lya-0#<qBtp#`o
zh9G#kTD@_z1&({q-@4$|W+{JBtO0dMoyTvvIF#&QP63<uf2BilJ$>}V+K_zRx?-0T
z;Li~cc=&NY2_qM4^D16FP{^@$AFl1*&>FxC7d?T?ThJUqym;cCTNjaG=w&NOUU5q<
zZCTi%3IE3UeRDlbAV@=*o6EmWXv5mB^IZo2foh26222&`?V!5_k{L#dP=9JyqQHLI
zfZfuRQV+q+6F#U{5ezC^rbRGo3$v4nTFL+j94QA47sj3FGcuWvo5`#0f^5X#JZ9u^
z!s3jXin&v~JWLVIOot5j@+=cMoJa)MgUgCbwokKp%Hw{ymT{}DApSrWMP;3$*!P;7
zKg@%mM5^G^P9@WgEOhVYzmCa0+z~$7by_v0DwZX@<=OYtO#RIfnKcF8#Z?UAFiHz#
zH6cH<_ekV2kz_NG1R3GXRB(V^hse8Cvi9)^Ub1S6W7jC(?KdB`rqAFE*^_1;;%xTz
zYDvgVVwMNoDKk+a4_fWM5xqlt4EXp#mlUSvEk#8@C+w5ZI$4`OEQzHrDVeQ*GoAhY
z9McZg@kqRVa!!*N>L1IGFj}4gLeqT5<-SoEplvXZ4k<oo0RDXh#PlSPjMF`~MZ4gH
zd_E=zO4jE13GW0HlOPsr6O0C!g1qL<@;@BuZ!NUxC)BPEowHTgbJqdXqF)cWMyq&Y
z;0`iUD^`ASdNFw;NUG0W`>zH=l8}Oo&Q1ly$Qs(w4PrpGgmLtm^{DWpc$uI+Ct7R+
z>gd@2sh9tQ1#(t7k)nUrD~s&Xb8o8}V<9Fr85-)50_BIZ=D!Ardhb1dHDaG+l3==F
z#WPFvS47mZ2=NgzZ20UKOGdk2v?*|YOXu0<j^%GOorR%Ck1){ekt0C;AwIiMfME-$
z-A*8R6l1~O!&(dQ1YI-31!2hznW~%|ir$rEZ}##RN<eJL9?pDAIWn@;BBmWWA6X(x
zE>8Py_<Wh=7>k4D71ui?fneYEzMWJj6j5)`QHj>}L(Ab$ewbEa{M22E+WvMlMV~7-
zD%`?J8Vk494D7I|{vtKaCvpBM@WZIFC6hq_^(!_)zrXC#V64A;C9=D<pq|j5>F@UC
zuehyeF(OqA_&$V)9gYq(u}9vhnKsXk0p2TEzo-h-T5a~NC0T8a2=`l>RN}0t3&&d&
z_b9uaxFEhd?68q3#{ujl?Edji){turJi8hWE68?V81HX-{_><%ycq_qxF;dut58yu
z!sD3;rk4bM%;EW@6!@d<e8<M+!z$`M=j8#uLkWJTc&*RFvmg@wHY*>%=^xj<A^`sF
zcFMmlNKB|1#KRG$;q7d~q`{-rX`s*|7WUtdddYG%226$ece&>fBvQ`HiM7~wG-pAf
z=?k7M^)mr(EqVb0Apd(No*!pDE2gLGhDB>+d<0h5`?*E9D7JJiAqP7L!E~Zv{BNj(
z1ethopZnF`w;D2%9E=tBA|llPmPVvQnEC=(J~-r=!8Y(LWL*$=J^ae`MJOo+=3J$u
z=5rDNmsLA#C}K<R4Ih3Qd~#8kMTOsjh>))$IY)25-W}%Khk66^VyvO|D>=zTRFChS
zH<71lRvsI2uQtGm@8QgCHgYZpZ`L3!3*G;Hp4y~8$nc|=epo=_AgDwOG%X4<7TZEf
zXBoP80mlh=H-Ox}^2H`qOJuksEddvP%2zsR;u_lQvIsyKa;66T{AlEH+YJ=;a>|GC
ziGelh<~26Lugpn{wlXh{<{m7PY*y(htcR{gcMrHs=1<VuNk?k+bg&2PV5amLXg`5H
zZ+_C;14*1jlFlh-A#QVippr(0!AeT66#Ok}3#{p;NfD_FXV{0IEoX+^APu*F-a=7L
z2$Y4!_nE~V<#u<f_YX~gshiX$<QOPbh)8;)t#I&hOL2pCkP3=n0;FpOS<cp9HMo(w
zD}TpL59|#&X1;Xk0np}VlWQGvnRU*2OlK0h9#OLZG`TE2^q;Pz>m<gO9`murOZ6T=
zk}`z-Jc6H{aH0F=6tUj<jhTW07eyW-D9&mcdm(DzQm)$F(g25&J|Z-}uq`8V$e)r1
z4evf^_(0vrHU#OrQeeaic}m;UTUh9F;{cLyPCdBf0SPwn^kwzz7Zfc~28<vln5YLx
zKj235d79{dgT7>nx|B`ZTRlZzng%5<?*^E>SH0zF<(6`u3f<>7$p;jm+6$!WGZGXi
zip?Ew4-4akhlF{LA$%#0{||@+7N<iPqI&lV`6>!WYVdaTh;%dz-t&)H2@(2BH<aWq
z&&-Yh%=e+Kkj4)fT2@#UwZ6{tJF%^&2WCDdf8P5CLeK;m0v{85ggm<dV$mu(fa$Mo
zprEsP+}+fiM~mOz2zsM>diJ1?q#+5GBG}^b;(erbvioS76$OowY#ejd$xBZLJvXD9
z*gbpS(JB>FtIeC4J@C`ZSww64Y*+zi2<b94LWRLrmD7oob2Thd11diMJmnw)%M`1|
zRvywNg;Q*)Lio@kHr5fsTr1Qb9MHp<!)aNQ5V~2&wpEaBKOE>Ql^}`><hA|2y0`K!
zJwTG12#`JX)XA<t<4Y?1+n?ExRm{OVYz6sY9`QP%NX&8SJElac1Xt@E>ltuFy)gUJ
zxWv*PV6Jz1R$_l>-Wlk<A$!?<G#_QM&3hF{)bS7CEJ48<O@bA+y%b$g{IPHq+jf(c
ziSCY!)JWuV$t#|S3K-XLYIepiwJxs8lS7RqB=R4Bw#OG7=&5xhmvqUktbm-*6?}f@
z&E$kJ4F6VT+kG%Y0*&vi(d=Lgd8wBjrItd3f{63AuIV2oWm#HOuY1U1Zp2@ilNB9|
zNc5f_4UvF);#plj5_NiLyI&%yC~|Kg!3XRT#FVn4?ge;IwCU3`6*B>4_HE!w+tUYZ
zJ5L9@SD|8<e`_cgLU>?o?f(W}OrSjH_Jq+gVjO6p=Wo(To=ag2%R+}3rp^a@$at^=
zZF#gep*Oz$MkT9eW&+W*#TanD7?zJ3$ZyAs<B<wDI4OSFXy471_x+#TJCY7wa~Fjr
z^x(U1X87;c))Fjbha_K)=4b2cNsVh6Y=nHX{~<}B==Dcb`Lq2xGaG{4Wf|zCeJjkK
zvrZD8KaB+#TTAmF@$0`(nQW+D%gDM(Sa5TIm6dNDMe(+N%nm3?3?Z5JmxiL{U`_Pt
zp-V1FPnyQTcevoTiI35Rc+z*X@a&L#zx%5)mYtdriSWw;|3dMW78gnE{F39t;O5R2
zs~d}HzV;JVR9*XCVc;$=tgL!Guv1sJ$HKz~IpC~o4&!LDNy)S7j}*PwgX%W>H|Xrh
zOc*;K%!xUanwNqhUa%5eEkXV#qTBg|@vmQwP`DIf--h4>s?ek)2$A9RXE&A{soj;j
z&D9Qfvy&1GQo#A&_Yuz-t#F849>ep%jmTs>B}cm6ajKx0ZrzUT5gMu{euQ@imX+iZ
z?^*Nd#miZAo&Gxy!IN~Pi7ZC#o+bgRUUOm*NLVdYMbRhzWY!>#?w1+9wi?u{sx)zu
zP-3*RQN|MSh{nnK@(S>BzM1Dy9_rgS9Z^7cCbNMnD2p0{h7*kbtrAOfw^HP6Vqg>k
z;^KkIgo)wr;r+KKG^v@aPHKm3P*&F;DNl1tc};hX^&>%YLuVF)95m9)+MBX^TCXb?
zMTIpHyz8X`psen4QQ1pzE;24}7Az<$hkA~Dw|9GZa5v|T%9dfWymY!gLj0Cj&6v7R
z+E9N>4Zc^0fr@7b<{a8&D7t~f5BF5yDk2c@Pt9#T*{+>97hD~2itm#5R(I+3m(;TT
zVjF{V{%Mb9ZgXAV*EUg*zfh6izJfu0xXaV-SeSg-mGdWa`_I8cNRTr#J)GbK;CC~k
zf;GDR#<&yz@4NaRvtHT%Yr!aZl*Wtw+}2E%enI{q=li0>qooF?fF36MsAw%WeGhC&
zM+kvBDgN6Y_l9%>W=0AbytbR3JY1lz`6iCjQk`AZnexBH8_@b-SJjO^D-+O!xB!2A
znrN`?@+RW-+CR19<AS)~;wqe8A0WTVd-b~a?7blEle<N*j8ZGscuivM4kcRt@cA%^
zGe7$D(bq7qP_ve35ap7du-EfF#heR!Z&b{V8<B-FuvHzUZ0u`IIb6tFQCtG~Rte5)
zo!;x@)#7>c&U5DQv7oXqaMHHnO5<m}lZqk9EG3B{G~LKfsEf>V(?i{LREcZeMj0aV
zqAKlr{rw01syMk+-PD$SY(Y~oF$P^p(e`pFqEqX~utG%tV!?x(p*qh+JsIhPoFkjV
zMmr33)md)u;(NAjyB?vZ0ZixDE}=F5>hzPKba=J#MLBzOJFYjnNx#w12;@1nEiSP#
znqm1`)tr>dced!I%ibls)bIvJD2EWzN+rXZdlFLy?j|U)-+ojv*M+%QB#nQsh{^}s
zLjEev9TpA!&98y8jS6+YaSM9`Rilweom4jrR|F%(l!-*gUs9P)tH<|p04;8%t+v);
zRi1RK!d!4NvNH7UHkIe7*%JqA<w)o>%ZR|e?oUROj&UIErR*ug(O}_#kjIOOQD8T6
zb;m>Zmx>W|=XG4H-ToQX>HPelceXQkR@D2)@G<LQnDyg}5Idaj%sfX@Ty)FNo4W&F
zDrf-l>f{Ra*5j!J56l-bvaa!Byx5S`L~DEIoa5#!3Zuj7;sx*Z^T2}<Vmg6n-4p(m
zDCKi)Zw}$r%~SybNF?>{wsY%P{#VOUDZDr?I}|s{X-$I&Zu6v;PQzNdnI`-th+W9K
zbl=+jowX$Z5-`m)3)DR%)xOHVU$qVWD%x$2v@aiZ?2SVPyDq;=`)|88ZWjTB$G6IB
zq$BE~+VfYH{=XLB1duz7mN(aH<cUH5Geqqc<@r9qZ<+@Y2+91tDdeQ1YYb&P=xb45
zokPM<_j#H`WOHEZ+7Luv>vLIoVRl?VgxXvQ59K5Kn6`Lf7Zj)ahsegu5p7E3z>Jg`
zx&7t1SQ)`oxxA{p;5<}#fLD&^vMo@)@7(z9PK_H6JGYp!@=^VhXmZmXxVU_w%{U>m
z#h06pOqRZBxOjqp_gMMZaP54Kc+crAyZvv~5;>g=d`;K<j;0yeC%Wf|(D(RUC;jn7
z)sK>(X>L=KFgBWwTp2T@iEMK~N-mt@q+Q;(b(_P7cd(uqyPMW=5!LHS*Yl=tCi%}z
z#yOjUZKTewcPD>mS|8(2?|lDU+q@V$ds8r!|F<z6TH?oZe~ORp8k8r>to4KXG^ex7
zXakuQ8Fx@!zGCY1WPq`IHs6sON35^QA1cGbj6@Uc{@bhlCi}iSL6(|=5rLT!4v>bh
z*`SrN<n4y>+A`YbMzEedX26G5RNlf^9?Y)P>@o~s;gps-Vao`w=C)g*a3x<!g8E02
z`p5n)wX0<6?$5&DR$B6Bj0tas#(VhNWu-Yeop;<0O+pKyNcPRsgt$4g$3Ss73W4dH
zI?rE+4lT%?4}|Nfc0!<RLA2J}@PDl`({~gXNLxfM3qVrVCy-kI`u%bYuufe>J(eNg
zHCO&@ZcZlKFlh@vyH06>v!EKxqNCc;w@VkHS_88OdQY4f*OLzdNLDLzOSOmaI(ghQ
z5ao{RWf6Ms8~nOqKd@5-JydAu#h|Fxa$ULhI`*=*$w?y-&U;Bor{>n=jXpBr)BC=6
zGFqk!%o8R!DOY612OKgY?UGqbhv9fP73uCeOhls7XT<{^x|X?*9l4ddg9Uyz`#zwX
z4Dc+s*`wcM@dWxityiXUdP$i?s)!nuqnS=riohv`n%Y`;W69v@VCfu_qne~h6<nzs
z8LtZs{sjtpH$E1Cp=z*0f3*xzB$W1hkcJ}!Jbk&xy1(7sbux}+tBDD2Gd5L~Cxs^_
z$yNGk(7=gBx}}4({OMKH?#70T((U|l`^@qm?}181>E?H!mcW0^e0tjJcTYuV-&S`-
zO`4+ZE6Vm+@g#I+FJmT7q#k~)bk)|~ix&NKV-M|uVSfpz%Hp@(BDC$vf14NJ(9Gjl
z+X**ByH?#+h>{egTZjPA+SzC$$LsRQaAJdX1LfL{0P<cTkR&9YI`{~)BeSQsfn;bs
zl(D#b0B)0@(_`qm?~3Gv$(4qIR<xbGMd+d%BEr!5-9MkiBBz3s?a}CNjdd_J(qRR8
zqb|cYH~hsq-DVRw|NIdZ3@2=m`gXvb$mOaMN6kco-__|W_~Be4B?d4)AJ~19(cW2#
z%**L^0NPNCsI3!o{)vnTo+|4hM+Dt95T!klocGM|m4X<Frf}djZLo>`%;|XZtE%M(
z9!7dDrt$Esc#07V_eu(jh-WrSHOyO9G*<oGLbS57d0RC-vszeO<URZK*0k8^tRQSc
zLp_}Rofw(eGsf}vA=_upo8Wh8pocs;rM>q@CejdFj>GTCyY(e+-4s$oo!{uSdt`2h
zmUf7$ZPz2&wC7BAfa0@Zx#-^eb2s`_^r<q!_qVw|H;)&+Y*QNQ+gK=cO`fgF1GSN3
z10_hTb0P-nY4`NfGP~~Ri#7Xlpi4o9wn39tTy|QR#1UgaA4(_6`;74rP%)9nH6egv
z)ryZn5coZTs;?06)eIg$+i15DV6)Z)6tTfN9**|_1DhxBbw!3#cdDiBH6WLf^;t!}
zr%Ust#4^(#PM7!dir<>Oog3B9_f7{OQ%`?~CIKMd7QGF%25NM6z(Z-c^Rpy=^v`Pi
z-X*Jc(}E?`?}QQzeV*q+tV1o{9juG*W2(&q)3O(*syk_ohEyj;7iVYhmrV7j>UWIS
zFjKZwCO0A*53Nq$(RU)kr@IW%a)k~kjge(vK<dGTx^s6$_`8HhQjiueh|xPB6uk3S
zY$A^A*UL1#5)4roDlzGd#RdE=QK@6k8B*^`QAvR#bU-Gy@F2<9b9B8*M&1l1?`(J>
zHQ!&Yjt!Y5mxoKNqF=u(`pR!$q$3D4jQpu`3)*|4mN@ZyAIT8%VE6+<*Y~BzC*-;(
zA_nhZZK-%wmx6P??%;q6Jj4%7&}S!o4pM@m=BHY(Ob!;1DsP`qb9>klcbxr7@79>D
z#hEdc?Inm0x3G2l7$O--^BhGG$R7=+L-+y1Y`z16GP89DsP~UUW3%5OVpdx9m^o~o
z)oihqa_`u0hsk2LPfBtm#l*lsdx(jQFkZULASVAv^Ivy^JwG;eZWssYdq4c-zXciw
zJIkF3l2K#B8{O>QUH3*Xk@A!t^_-@20NrIjo(h_uzw!GJ`3jjAqG9Pdkz~8=GKEHk
zUEbh}4j*^R5G8Rn_5dkYXN}j)0}1GpoU9a0lp3bg;H1YgW~)zC6WMIY9`^1qZ~(F>
zG>OTwH88DjYl%=if@@ln5b1&fOo*veOhp@+2PSS=eFn5^D`b~!p&Q0uzTuJhz3dYj
zn-=f~UIX(3OI&I<^qgG{v~Kg_;{EH_(X{%~_pq8BD0SI`B^$Fe^BkNdLGZ6ub7Xog
zIB8b!#WH_gCxq#rf+BkI_(*rVS0W>PzQZYDu}o3c!I`me<ol#F-h$;;LVh8p`#~`5
z)$o5LROYTHlMDVhzA24@X7#8@Po4GMmt*TeHVt5%-w;0vpx~K`iI~BA+e)U5oXBSi
zuz=XvG6s>jAc@xGrZN38c*vH~vmxeLjW#`s9Qx_#IW?-=d4a8VQ3)8s6OUl>i^s)^
ziFG!KB>ZDL;GeEmojF3J!%{SK_tk2|1l(DOu@fO7!Ndd9^pU43Xb(ZETFva5w#I4#
zpeQ-AhTeE*O*9{=b3${i@=)8)T`;9hKqNE+iY<}Xh9sOP{McrD>*@mqF_hxe&<myV
zY6k-QFe>r{^9=*S=&3FKX#EmY_H0wc6vtfEPl00KOD-zNw|lcW*8pq~S1(6W-<ZyT
z;Zhb#7#fA6{ge|#SLf`#KyZO9ZZo=(Q!8R3;{+|8;Vj=$2gw?cDw^)SHr?}95Yyb!
zO&LIz*p&RK&M}Uk6RM4Q6y4&D2<w!s7oC*Ub9A?pZdEC=S@=CPCZ_Ll!}Gca>Z+hT
zAJz5a$u%x96+l<noWJJb$xqnKZ<DGQ8*`p{NujCwb6~zy#jxg)-*5B&<p@};T~Wpu
z7|pl0V?h5+R@I-X+f`FSF<<SUykGflmn%me6XX;CuNKSnj^~4}rX2(4KkiS)TvL5K
z?l<-7&^+JEofuRK2Y0?Aog&Aw12f5ad&vJ4BM!o6T^=~2Z?zO90-P53yGVOSDY|)7
z#8Bt0nUCksgbef1Wm(<PYFyp`C_#T>x>!`*>D6>e`sq?!Vec}b`f;<CyX88SYJ^XM
zE6`e)d<NXn=NCWzW*<N6TdUJB0q}$p;nqh(@E?MKTD6285+w7lfwB3sBouPNi9BG%
zr70rJQp9d1aEbNrB3(=*1-gLm-5n`Mdp+t+qFiHHTV(Xr&_9PZ)a69Y$`LGTT;|)}
z`3)2o_C;LYuj}7lMx5DhD!R%orfyN)awDM_L1{a`lG5x~e+~QA8+I^V>7>VSc7-SL
zjfrY{#xFD2PX@?W`XQ=2(|z9W=#7v4fL^P+D`E2WXOcOc*+-}ssu7s08k@qK;-5$G
za;e5mQ8goG8klj{&(I$lYNK`ISw*W^vbcS&scGg#C&wcul(s@z6q#F_gImK2eR$v>
zL6`pI8dQ^$>cWX4h*1cw-(LUzJv9+GiY9G_y414ga2Hp#>6Mv^3eoW{KJPCtVj>p_
zLy!Ns246T`MomK@qKMp3=FQFUROA7j${8&2ibe2-(eBd%=i7Gms76ZxyaL14*d7sq
ze;>rlNVAg%<ds7jj$O!9)y_qQRCQ5vGL#UNx%gf8U8iwWp70;~-SzKbURj(e^EWW?
zScfG>H}&k|TBgn3xTpp6?S)>79B$v9^v`}=&!&rR>bDyw)H-tp;9hpQU}DOC57tb!
zuQx>H#85MqvS2VeSEEmj2S~T!v>|fjwoHUr@YKBT-Hv?Hp3_<N0QRNnhkx|8sKmEE
zSP=X;*1lrDT_*bUqXOD$5YpYJP-S=f`aZR}UA`Q>bWuU_@~t~R`R-*jZSv+K^o}lc
zeoVIU54%5Qx$eeE5nj;Syd~k;^#g-muV3-(wt*SZ{Xf<9+>5VpMI|Xx6eKbfg7g=j
z{iP#~iA_^@&B&D>sYs}5(_^Oo-dTYPAyfS4ku4C@Eo*xE?HCZ+dyDDywoCK;)`}{a
zs<npOy`zP7JC~V-l+c=Nir_&E-a*!AYfW5NyfCIPfO>8w7`t-eqX74AmE)i1PO`d3
zOd$xx;_=JlzJ|EiVpM`zu7#N0&E}W2xL_gy8MI<q#J2Ve^?ZT0q5dOz;(CzSZ6_d-
zGG%H6Ac1nTv@pP}cD4&+xbj{8{b|;VK06*-L0HVAoq+GhnLFp=s;+4)IY@#^f!kPa
z9y=KuY}My6Y5Rve&E}c|`les-fQs_<LeL&F8C9|%SAlAzZjYFLPT*Io32Y+~Fou4S
zNL0d6yJ?%?D+(1nyXNhbD#XLYxbtm*#mgf%-Q4dHj~^YJy%qG&UsqQ+AS4$dOhxku
zgMZ;5u!6JVV?*kjd9j!JQ%L8z-ja}DpPLWCGlHn26C0h@;y`WnYcIyr_CWAlaB&_d
z_M+_GYRH>FGmZXOV`h7xc^^7awM}tj9|1xY1=i=b2zeR@o+C|sz0r--mYb^Wjzbgq
zDjw$~FLIFaW*cW^h)$_^V5Pd5_~2RaR|@W&b8SHX6N%$mJWR8h9%p0Xe179f#L6BT
z3uCFd!hD2TVg5py=mL8`H~o7Do{k`X*^_~h40%;{_n+3DtAw_DhKoV;*PON#dC}jI
z2nvi_V{f0ep%yYu;9Vd278bFqSUxhA<UpyAC>Q1R(wRm{%ii9ztNPHFJ~u<ulX222
zm29Lb`Z0f$hHSB(`r#m+ZMQbiLvr-=RF9~@%A37buqSS|&+Ve&t2>{%`QT8y_2EiZ
zMz%l3_uhE@qvOhO;qTAp<NOH221PqoR#6RJpKdU+<DeZ^5#|+QYBlI^h}dJt<)39#
zq2J(s-{$ZGS)LDTq6MU@K`gg4C>Bzf;(!PejZAN$kXW|R1#)Lq3I9AiuSO!^7zlil
z2{a&0^uvf}lu)Rt-3gy06OB~tK@l5FckF-uXc8_laO`G3l7x<fLC&H<Hm;Mj-dGpJ
z5ETwGjNhrIW6n{OpH}>s{*Xfk#|a=0868990p2si=*oP(Zd_M+rbF>SC;@yQa`H^3
zKuk_iqz8FByxYm|2YNuUQzl2ZQ~I23s)tFH^JC1J>TVm-91ych2S&X+YaXaK-*>6I
znCM7Azdm>3No;V~NoYWybXB^U$~jOP$Z+{;cdduR$9np%E>1Ya;vljR)i@rTy<a9-
z3HR}_A6t;~>eat3>u-iPw2f6aR>Rgx*dnO~n`3uuStLUMH_4J}Uf{5-#x*9Eu;<Cy
zhQ-lcF)e#oh9B9q?M)ZdE(A8>GMNZb6+Z0sFYdQgYvAF;#H(hamYD>m4>lUHDWPSD
z2iJbOP>tdlKTd0GT_K3<cwA*@m5ptd9sVGCj-@jTuwfn@Qr!_Q?lhYFLgF18dsi@}
zC$AnEYkXr#hmwkUIyOzqxS_h2Bap>BXt-m2Df-h3p%&|+oMJPN;W>b&-zX5h8CFr}
zmhR=Mn0p7m^WcGU{pM~^%N0+aH?b$e{HJ%v-f8k1^wDySF@MHgKE;pN%-K2~UlAWc
za*^(Ijxr?`m8kOZaW-$PSd`o*Rbs_vT&)wVX)Bpr4*Vd42tXn`in0vI{`VuXk<^*Z
zj^ug|FX+ERU@SEsCQi72k$NiZMD0p<yTXl(Y~LQq0l9mAO6V(@lHlybxpQHZz~W)W
z*2^Jk1qMdTZcxs-3_!N5uG?I{dQG!qP}V_kW;9HNb(NDUF=P7?-}56G2qeFH2>pK5
zP>v1mIFP$8m29$x$>spvjxVKARIltlkU?Wq=iK4A(!u|e<<k2aQLXzK!DN~--{@ut
z4SH2D02#<|8&;4)S~#o<T?<b_2^K5QI6VwX`I#H+XC^#%??cspYdzJ%B_dEumUvQ1
zJ>4;Z+-BB=$~p)hu}G}Zj%)PJElo3*Ycg@#2+DKwNJ!TotW<}ZRlRD`&sW}FVk;#)
zN*>XN8(K(ww48`+C}LX1WnG;BL1Ij1EnsV9|Mpxn?f2LWV$#3LOv8Z6R8HGFR~Alu
z%zHJ$y{ueIZ-8r&Y}i1p!A=A7q!TB!eN@||#m|wZ&-?t7AO~yVcD9tBnTLr9r|$Qw
z?MGj-+vYR3&xxg%5qM|7xnPbb@8}QbOoy{|EZ%sb2_QDjYi@@2^?G+;Z_{}`07JTO
z?X~NCz+9}*pL+AEEgwHWr`K_sp-5$?sao69Xi4;bd%8JmX=w?KiaNW|nyIN79?J;z
z4wA5MXw1HsY+NHFzE|<gK&EtaLwi5V4$$gM$nk<LuW8~4FekD+Xv-tqKy0wbsOQlS
zGW}#I=X=8G+E~vf_Tn#?+h(hJefz<tis9i%aizl`Ne)VR+9o<2CQG^4FX*Q_->p{z
zBhmv;V{;n%U`q;oQU?Eacs}i>nP&%$UG(k_L^#|{aET>Hn@7EXf}F3_L1l2-^adl~
zDQS6N%gM<NPfkku_<$rccAy_IYRzBpL<%8ymsNejSuIuWC3q?nLI!eQwclM&tl1VE
z!k6%<4cb1<Kzg-Tp^9|j(zaP1$D-q;OilK#%pgERYPn%XNQHLWGT>P{li{vg7}kkN
zYRMI}u_2JnS_T{d2(y%BDe;9gdVZYZS|SH<cm+%OYBWw*;r&qDJl^@~xA}G=@gXSn
zJqKhi!0sBJ`z*k2X5WbQu=M8}^7134J_3&wXX2XP;Zn@JXLZpZNEjBE!v0k^zoIj{
zq51J&{Ety}O0k4NR~^kLZDZ-~u~b}Pps0l<R)=PHf-#En)DoL|TKBU&dHs2#luVo$
z3kJUKXj<$3dNJy5!Nm^kg*JshcPO;abU@nMDk8Tihm~WF?qGyOnJ84&-t3|k!n%Qn
zJ$JLi@@J(nrRk<9M+a0la3o^>lCKJ46ciNAUnN38)Zu@s%ESx}QIwVAucrf%b}QSi
zE^B&wgk!(jybffRE7ZW>LBJmran@jf6TSY|=~DCM>lt6yxrW@y-%&-F_2uT~YBqgM
zKUP6O5Ndi+ZEfu?ycav)FFqi=aJK8FhqiBM-M)L9y4RI?&08e&yZYD?wqgvlH*&{+
zOp8Z{jtAY)9r|$!U5j(r!j*NC(UA0T3=ep9q~v}@(hA3<8^`h@x^B3&p_S;58UBsR
znp|B=qSYM}bU2aPd_>Hzoo>)_j>W71w?&lk55DHfPU?TWF1a0bJUxGv{S7zOf75Ep
zJ=ol6bNF~zD#&leMa&?;!Lg8*k?DW5S*!8b;N`zNXM5fQ4uor64;8jk#z#$7YE?aD
zUoS|&bP@XG%Og0SE+Uqi8ZS{h{hv|8Cp%bNlfC;0Gt2rj0Nxa&85g!<)Bf$#I!6QB
zGC=h)-!C3zzy?JUw;#J(^uo^6)kUv6gax-ympMmSoa)?ZVD~H!0)c!6j9F2NL4NOQ
zQ|5aBN2^ymd+rz-fCSe5rT+5`ZnXIE8x=_j$O2qJDL*;~D!6o7p0zxt*lT_@E>c%3
zM6864mm21me2vSCwn+l%;nCntwku=R>L)7bAcfQnRp9+~gX55)rv-uFmOjC02u&!n
zQx=k`&&7J!dtHA(X4mr*rnQhFR=Cfo(KEc_hTnG-6!W_L+J!zMV|A0lLB4AV2M$`?
z)!t-kU^*&z2n53>F=+axRds?~HIE&v+%{za76h*pNX$~m>ME|@r7iQU%3N=8THUtU
z5R4I&5^7v}I%wO(;@*wp(Zqm?0>6J6oB8cG3c4>9Jw1P+j~4}v4i|>Y<~_tnqMV46
zxtJ7dC61&oy2sja;Jx|09%YKR$kWkDNp}CFfalPyExA7FS9nMshx$GI@9a=hhf`cH
ztL%V8Z_ItbI-5sRi<ii8xl_eAa&9QaMd`+#)?-A@i!5)rZ>*+HsK4{K)|uhrPY0gk
zji3i|x~H9v8ImqObbp!g0lO0FBwbZ|WD=bkv8xUuA|WrlN-Ar8Il)-@N}a=1z9WnA
zpltuZ(B0iqc&)KUg9nGt-P+dfrl)WJ72jDoPu7$%*!b_amXj;rFJ`lebcBu%ujCmg
zg6|_E===O))r1=Q+U$s(xzUnML^CLddm^yrv;g<5!LQ4AXw%N$^EfEe-XLDp57>e<
ze+Oz>b=F(mdQ%$nqUY;HkpPrMQW~=DMBDb#_Nyp{O53+(?G#EsRq*gfs8R@IhEvJY
zbH|%8MRgGGU&UhUj_w*1Il$zyzX2Yt8^KE&mbRjeqx8V%6}!WADT@bY&He8UIKDp*
zAh6Z^mg0dOOP^jol=83!zh%p>aZ;m1O6{uc1Ybj16k~r<p4Dgb%Y`kO3tpSb@jF?$
zX}<dMj<!r8;;<#+JTE(3N#OUxEv|)#*rwvC2(3-C|1(L@$W@YeIxH7{I?hgE!k#aD
zM14KEX1AMcw+WfyzWsr=9J4>E+t&NonaSo#NIORN{2)dFbu*=|*CX`xqTsdHxK`tO
z<->nJebp-P8BSNZ4>DJ|rSkIfk}7D->vJ7Qv>ny7*{h-;6hxKdg?jjiCf_08wxX`N
z63ryp5py^|pyzR7_siDy^3lT@(^l78!1n#L*54ZScd(H++B})Id-9=QC@QiOS@jOY
z@Y3gJ<mYm|JJ`EECk^)|%e1mAXif#3*Go|bG)|NTnMdFBwj)Jv(;8flHI%B50W_F(
zz6g*9&D0znrMvF)yc~$9z%tnBg=qdd5<+VGtUJmKGSBvbgn`*83ok4wGCIA-p{A$r
zx3X*ZM@sPxO{+6|q&6HwM@M&p3y5b+kwA$<-EiDJtRuSYE0GfL!-fJTMj;*3vTAjR
z@V_~r0GCQ4@&F_cF&uaWpugTdN`|oK+>3iEn;;<vgOBV~7+zz~;W9KtIDOAAJ5cqF
z^FDwIj1y#spJ;-~2sWvVV(pRWh#(wvy|etS*<OZo1)(6d)v>TwA;#}ip?(2V1~-QV
zpZVtYcl1~R6ZXU|$^w=Axp3{tF^MwugO4~(E1C=%C-^?2R6AQ9(EIqFslmGv^Zwml
z919@&@9E!{{&a6Ma)H#-@c!2N?VQO<Tul^w91}vM!kiTq56x+HyxCnI5_rSl;s&bK
zlm~|?9TmyG7r%mozE56&`|sV5Dn!%7Ik6XqZ`b3CfN~j^5kSe1Lnp#Yr|JL9;<6)T
zY`+%uI(uBZIVSq@^Yn2;+TDwOS~$0PL^+!qkpb)Y+I(|6x^TTdE{Ji4<Rv2$gK8m+
zp*nY5F!*MZtvI4p*YfoZoo@#|1cJ9OD;IT^r21{6=V>i*0{Z+5y0MH17D8}J_}pM*
zL$49{qRs_|L`)2sTmx$z_Kc_LcGmD8(mC1SvipF{xgMq5w_UEzTN=`2Fw<bz(WY81
z#{|aJW<l|ejg945Xf5AKdw*L99zH3lTnIsGOfnvbHcpL0i{WF+j+C}~zUy{vB($l@
zT=4GYY-7Nlt}$9uTDR2>>^@X?Zy=rn8Y#+SbCD+TQIbnuM{thgaMZ-9rLK+P!7?lA
zUYVIk2Ko?G<e_&W5e96_`d$3pLfFOuL~?=}gmZtUE&6+;baix91#^eA3~lXdXGHJD
zJpsceqb9+TlVM)(d~F8nTr(rE-)QEOC`%folWEqo8O|TAo=xm97O$#=iF1P(cp@H&
zUkmT4y-a=7n$}Vn(*2H>@coEkm-Sb`n}ncCBt$CKh#sHzGGSmeSUG}xZGA^&*y@lx
zJ#lo)BzhDov%}j>PT77)a=|)9gQn=a;B>*4lvTE)A|31-yjAMDU2V8J0F9jyjyK{H
z;$vg6V{XGoC&y(M6WFNu?jf|4^rAUqN~FH89~PS>CMDevQa3Um{D#4y_4w&<Ozc?g
zL}0lAG8~pLIqbc2S9bj>W*W1s+MgN<V%Co;3MpB<B?xT#g+fyTPxy;=%YHfU&qJyF
z?yQTIx}M5wky)jS^Ygc7X=!P(9%=O2+Ug-3R*S+a9Fb!a5-u*77&_chFFy&U=(l20
zglJ$M1i=)Q&wBAPkoH=x=zAl_F9q&2Fn<Uycx(`t3q#TE(Tb-TKI>j6SQ^Tiu9a;@
zMpo}_1cO|?mgEvVg9mWH7HjJJv%Y|4%6>=5_zw6k=o76wcS>+?1^&nw^0c9UV-UNH
zK1Ihu<9-yg%t>m(UCo;Rz)0k4`q+kcbe>(Rqjxo+FT}X3=}y&%%cp!ah<0@`Sk=iJ
z@rZCTx7zXZdBTcs`!*yV4WJj087Yw*i#x~WKr%{etP=nhVyO#V|MdtWKs8crJpl)&
zY`>zJ=kOU(Qew1INbaSl6Ow-mxnC@9xY^&Yh@2d&6fIa}jy!;#&awZy4YGY!Zerq>
zg{}X9$)%c_%k4hR`_@LAYhLSL*uBB1v&Z()QJKV_s=MuuQxZBl&zemDcnGp>qy<gx
zG?H-P?Z^!%+pbYhr&?kvx)xL}Kf)}J9}w_zLAlGstE3=d0R{gdz4rI3;~SpV@5DA!
z?wXJU-DI>KuCfI*I(Nl<KHdi+JWU*3GRGL(PKs@tq*gE7?q{TK9y089&^(|5<nAA!
z0yZXrxA3CCZ3%oJtE5!I!CLeVkySW}k_d9%|7!uVrk5fgusvM{^!Qze0;UmygSU4k
zn1@QpR#3$y@Q}r)40!l1@N@P71`&a|+{B@*@Gb>;(E*EuxRHabe-?uHuW1!DoPHk-
zU*K~FA%)i2{9Hlrj91r}MVkuR!i4FlfN5<*?1_(3yv59!WktC)RpB6bsPO4T50?l&
z09++s6gB*@wSRh^G}I{8Q8I-a#cP+2C{mszZ*Nr-%GRqPV(9$VFSt68_t8YpI2&11
zH|x3Hlvj-yQl?Ejx<lnNF%mWa*vF=4l<v7NzfAD4HiP+5ObHy?jbR{o70|5Jl<X*e
zr7>tX?jFY#yuXg)u<_X~ITsAk7gVZM>Vz#ea_V$C(a`J4OcxiG?7d4wRGWQxoNsot
zY3C+%9NuD#y*xxPwj<xbpdXX>NjZs&S^W|SyQ3h;cf3AumB%`N3*bOS`o*iY79bsS
z)&(AP^gz^eT`w9M8REX<n!7*FXWT?Y(y^kw8DPx82(F5;r&17?b69920Zp=B-&2eQ
z8F-qjPNjo<v$oYr(3-br$MkWL$QiS?5iBHh1CE%8-+QsujxlFGe(WMdw}y#@7Xq_4
zl$UEjJF7gWsHEiIoBzSB=!NXM?u_MfwGuc)=xHr%qhi~<18Kcp4YCWv!0e<O$b^0V
z;BI14V?B8!#2B0l*ea1Ykke?Zf{Z937q#Mt{0f)KTR`7SJH~(S_oZ<GXCsCg0$ZX#
zqu}k0EIWz`{4N`i7z;Opf;58*G5n!`OW+dxjchxRc$isZd%?@H;`=Tc1Apq2jGbE3
zGvpUgGgeITMBMyoAq|nz>oIu&>|w@%l_VP4z1qQHi_)PeFj)*S8o3Ecy=E&Lz}B7G
z=G9f5iv||}>kVGw^*5SQ!4EX6P@Tw@PRu=_WQv&TO`ta2w6I^pVF)4t+3>_Y^5F={
zr7yOydaL#PC8DHoH`uY(LMnC+_TOLrbK$4k%>Y(hTy}IR1VpHUEP#B3nsGfQ5DICR
z=<AB;Pi0wO{6Itva2T0P+Ux@64*TvQ0K*b6fZ$$QRz_6CnCuY?)9B;%!R@3d#jPZ)
zt<#a}Ds0yO;i~<z^~^2zt_O+;Fkm#0e>b^Tr?KVxM4%Ty&BFqMrDIP*FdA=2VLBD4
z<{A66fzRCMw=Wxro#r)oiNE+`cQS%1ZvhW5h9#`#_2n;Larh1XXVA6<==&v#ro8B2
zdXpd@fJ=xeDEIVvG-&QWd*uE4TJJT^_V?Sd_jfzy%eF6{esO{?Ij_?Zs((OCuh!*q
zy@jeHOjWK=3`qds+y2gaIG!FxwDDFzX!CR2jP0`hO#c0~XRF(YH{1CL3H<s<=mVzf
z`GDzW>F6_@;wUHP-@a<=iGG-A`^$w0fQ}6PT;uZL@!B1OO4t>A%8Mob<JrgUFCuyp
z$Bl4_Rt=ViHkkoy-+pO#JI>1=8ra;{mZ%hXlY|zwT{8QpfdmvA<Gx?=!!!_Nt@#TR
zdR}T8U?m%X=U;fY%9+g04Cd6~u~ZEHsj*;vYoIhNPl=mMBNid!ukn_U>FZP|$0oj>
z5I;x;%tYnWe#?xB4oe$tL40yuhAb*0($Eqt;t>U+XY{4y4&VVQ|00{>$Y+55Na+$u
z7K_KVI@*)#sG&uTFOcl%hi0{0XmE2=n6y8z)MVbEk4P0_A!Kvhwb%231zv16jq~*T
ze0jhm<C0cCC=x*M0>1-jR!JCV6L44bfR3A<5f`5b7#^0+uKB5{8CAT9*%@y>l`*D3
z-R3iq1dOxZ2_ZVHQ&Nc+o4Ruo_*4ETfNML1KKyx(=qq&KhsUF`M&r5Fa;Gt$-uv2?
z!<-x!370EBKOs^$5+Q(e?D?Y)$#^jG#Qe3Aq;K47krlkLeUy#e-~Zv1{oRIH*M178
z%jz66Oh2B*56otPXtfiF)P6yZ&jT`J?Mcfiq2mD>bwnzo`ciA!41R4e$wfjz=s&1X
z|4~DY-fxe)rRF&V_q<+jU7Z^^$Llr5RD|y5ESUmN<%ASga)*k<%1Vj0A|spc@8&{7
zL|^Yz`*Nv%CD~rMdqXi-H`2*PX-R1-doI8%UZO2mqS4W@n{)lI@EEm<UI;9`d7A(`
z?k>;cJVFXfO@-<3Q5o6E4Dp<6rGO|?;mePAMPvjG*}OMRwc<n;1P^BvyE%uqhKN{o
z)rc;idZeHOhH?OAWoA|wTuG6a1zl4?Jz<x(GaPtbKu`keO*aA?3f{~^9c?%N#VCnq
zms<!#{4e4^H8Rb-Q?VRRH3FVuBJ!+@*dSc>#Rbex2wIL3O|tvxluUrfub9RRLl$Qu
zwVY&(!1eoIiK?7p&?>C=neF0&Cz-LE%M3DA0z42K!dmqRUygnB{mXp{yC>WAsl1+E
zj|ShA9+}$Kc}c`#{*?g?8s&DD>t@ew65Y)}u^e?3b;#I(q=<;%@4=|Pq7<Egpd&_6
zFCFV9jBGAn!OzFp2#uE+T6}KB$klW$jS4mT76TZw$;_}ZrZ#+)5;-_+%%z>s?!A~V
z94vh=7B%mB*c{tsKL7w+$F~om`F!>Vi`C;fn|D(z3NbKAto_;cS?v=f!1UXJyQ<KB
z`3kdKSGdcR!LS*3=Nn;UE|Et!97VA?(l=3P0~?{0h99&>xUM2Lyx;n+LP~e2A$Va(
z0Fmw~j2-wyRT?1Q3D*@xJv=0j9-6k?@sbu6Qdzu#KqF(zX>*ct%%{<TtgG`NV&XXm
ziqOx9bteUhANi4^>#gJJ=9{U9CG6+JC>=F5wpVc0^rmf4$K;M!$hJ${D}qz-WRxeO
zL53W=n?QOs@!@+I%yWA)r**doifN{LEM}{qk6_{c$h0<{COEov<V$8d_&K*KFkddJ
z9Z)3nN0iM^Z!jlB*Crwcif1NU&YHkLVI?66+XjrzV|esO-S3PAIYE=bkSY9sL*Hg)
zdK=1b6HsvN`%oKWF%^;l6A>$ZP#Q{v0G|B2&eK88YbJ0pB)ntU@Y>z%v69^;J8LNk
z8#?jt`#9Oww&O{6tPX4O{IBJ)f>{DII7HTTVy0G8w++F6ci!JkxO)@-3d|(cv0YC+
zK=pop){O9}PmJXeh{*E+cS>^ARf1}_N98S!c^0@#6`*<@;(pzT^PVStI0upYX=z;R
z&5=_w7jA?ZaickFQ-_9B{!2RE2$ik3mU^6_0#c_%klZAF1HTtpkAJ?QEVrQol{>6j
zt*|Vtapr2y0vZ%xWW1tL!*5qQ;=^eX2!F#Mx4F>kPhi2*8e~y(+bmC1fH2x*+!)PV
zb1Yt^XElqNs9a-`Y3H1dcCa&F>P27&H0xiuqn_Idg}4-Njn@n8|Izdn3{f@C*YvXF
z0@B?mDc#+TASvA)(%m2}-7O{E(%mWDE!`o_dwqWI{}b%JdneAEIWr@l$+A_L65ikq
zP5O7n@Sr_~=wzdmKLv$D0@GoT8h+8!L;$92oyTCzGXcK^&~+#YyGm^;#*s_*UCDts
zgL_!3P{a&sl+@VkMP^W?^*<HXH9ojMXl6(RjPtjdn>v11Z-(J5uMJMbBZNs8^f$j%
zs#pfSC2JUVUN!Z$Je%0CdRJUjz98RF>Od*%XeGfM$G*=X-_wUV4qQR_GJI?@e7R%L
zc`Jceh7RbZP6s=2M+lbP<D|LZxNXf+drWvHGazA|zUJhvBxSuZzY(bV7e9XW=D&)k
zG~_@%uT7}2*EV~{l#vvQt4`|Q{fZ?PhAXMun^=zbg_SSzt>bn-CX&W*?e#<AT(jsL
zDqMhhFa@M<99OX?aaQ`VK5hmLfC+4(UY6P2k`%n=KeBggSdksl#UHu9GY6s3aV}QA
zGQvqwz^rVbt(jQ}BumI<ppE|cE2fy~K5+h?69vnJ_J|mZLb96`!-=<g1j+fpfe^`x
z5kHPKGmbKWd?2RaMGjy-pzbH!I)bR>Cc+q8g#6UrVXgZ<F}4A_TaBj!;h!DLJO6z1
zbw6)+zXV844xMVtYqMPw^Q;>)Qy*TThg~e+tfZ!B2W79&85xlltj^aFhG$@`!rh*Y
zb7*v=Lr=($A4fjo{Hwy!gNkT9hx0a7c!|YbFimJ3v>68!(3mu2*MKeq{kc9Pxf>zI
zj<&;tA;a^Z;eDnC(H3%7Z*`B7%Ta9Lz~XD3Cme;j^<E>|d0~YXW4!jsuuQgFHfsb$
zk$1n4b~hb49%L6)@9-=aPl~PF3YP2pc)P!?!+kf83w`R#5n{wf%K#4^>87S*qt+3<
zo>NdSM<g!cH=C#<C_(+W#9_sA(__WIzW!8g<_151S|RQf1)m}>H<({q`jT9J+C|g*
zyyF#F<8=JdyLtyt_INE-NDx|MVz_4&?DO`D!K$ue5&>Z~J7i3{@Oh*tm+L_8CV)Tg
z|0|N_i67{tAxggjeZLH)5(vS7_qh7OGTD~w>e}k(<AgX_#UyG~k1AJGwAFb1RjS_!
zgLPhOyu2&`I`)IH`4wvO_OfL->vSn2z-ay^_ZxZV3ogF7AosQQ^yIB<JF-euq?>$L
z*?2)y61;uK{I&J6E*j{hY-_pdf={s?c-b53^BWrAftL6@@(Bq}wGs9=Kr|R3^nB>(
zOK3f8REL)9CV>AHr4vmA=OfSBH>r+~=ta;7rxX=9;b!0b^vVbXrlP0J$ERSF-8#|R
zOHo3}MAE)u6)yVfpB1XHyI($+z1$Zzhx%16)mG=4>;5FY{H{v_bMWDumae8;t-q&E
zO$q)9t2Ql^)q2(+So#tKUn7qH+Aw+L%Bp7hjU}L+{-2HkMPIIj-G)09t>OoYE%Rzr
zIG<h!9>>+GbCH?&<*4xOr!TQ_z6z5O+bN`a9#=jw1?8-EUs8LSgEPVZ@F#u(N+)4<
zlP}WkqT)f)lcAQU&~CkDTte{$@*(jt<IKrjLBrf2v+8fD$_;q(Vm_RMUFwO&4ecpp
zMhH}iIqtt6CRq!~huC2l*>Wxe4SMzRzH?IFRN6)C-Tj@sFp?v0?=S<F@Q3UD9j}ci
zU-LNU+p^f|lVkls4h`K9d^R&Op7&r0iGIy?Pi)t-rBTxBEWWd@A1S($e245<vnWJ{
zrcc4jaVADrU}&UyzR497X(Tw+$<5rg50t||hcOb@Wt`;I6J>I$R{W|q5b@EYaoeH9
z@^bqOmovY79koV)do(aB=cFQ#s~Nsiixc$iv|F*kO6ZG{9<0bGGV=-l+m<jO-Na-Q
zf6xJj;z&F{e~iinUCoY?NwZVQKmkNkz&La71^Y#3f5$f#b|HTI*LN6A?I<HjfFA32
zK+Fe3a^p`1f5WCmTt(tcEgJd}Ngks>yfSZ^vFSMWP%wO2oG4V1OHV#A_SZrl-)Opz
zf)mq~_WZzyA>-a{iJHnY7;0ffIT*k{=|tlo&I>sW7yo30DW2!W%q9WUIN#B#1|uUP
zkcb34Q3k->Q^I73VZGCg(&>qb!DIF8<*)a~Z7&mi{K_MGJR^sFiEZEbzut6^Zh#^!
z??(gTzxUVrz?OyL6=O$N=<U+*v?dvRCPeG~`=~_5ZA5N15$dwO0KK&A(K!cfbnNMa
z&1x~rZR`gXbHG5brI&z(2~lBW8DxtH*;^cV%vJOv9;!MAZ57DVREBaYjDjl~A=ZDw
zWVr}!$06_7$mk*E7<tD9QSlt0$F?dlc}D+u@%;oUDxhy})vx#3#ICHpW_Zu_W_>V9
zYAu2%!kjZC4CFqWJYiaL6Zub9I3jKu02xh9myq1mvgx|~9kvf9t74tZk7blUGWwu%
zz;81lCAVo#ltrqicDEI%*b8yhBZ-do`d;nWi9d)hs{sgk2uj!alG<H>p78zq_W~pH
z6PW@a=JtC`_6qa!Ml^<AHYO&LKTIN{@W?m@<1H?9>-ynG%U1Qj)*Px0tUh55M3YW>
zeT|<<(86%MU1g9T{WiFDWClpjPLkhD+0}(BRwsW`k*kvidJMQgTH$yp$61Xi#@d)^
z<aYSgS0Yvm5rsK>J-Y>ZTgLgJiNmb54J{;Ng=s0JyCx`6+d#Q<R97wr`xc>RVVlyy
zQ>Frn@S{d9PiiIU>+Nc1GOUv=zM4&h_&UxS#5cd0$!^pu8@a5YtcGdDNDfev?L-Zp
zG9bn?<i(T0tiI694f4R>cT-?*Eef<gcrguiopE<}y`>ovRP*r{VX@UXn!TadO>#IM
zV!z(yOqY5Qx?7Tqu)?ypevi}B@?fTf4TUNZAx<yGPRH|K*F>eZmh?gcz$HYqM@CkD
zH^m`8>w7Af{kn{8Q&c^>FS$7m4zEmX-Sz0N1k9EyKKj_zFB7oAl2ag$XZ~<Aou;rw
zMx=aOGrhay_Qe{{zu~q*IQ3kQ9#Lh4c~`$|C>#b*5Eouz=Ijy1I1zCM-!u1y_Gq_V
zOT1t8#g<XQc|K`LTVpi}3;6@0vU+kTni7bfA{-Sp#tJfV*UoFTx?h}9R*`EWYj!g|
zcru)DQ63mSC*rNLyEEF0L821Zjw2uqXj9wwe|o_LL^-~Ibp%K0=A%%bKX>5jUsG!O
zy$j3}QyeMCLhT}*ojw9z9G~=YPgov4fwiHEmh=OrwR@`~V}d(Km4@)gtm3tdKO%<u
z!=00-ZusKZcJQu84ms{;KqwOqcK-D}AXlbXaH&Ol6Dv+7rmbYu3V+hlb(iERGF_1)
z>8^4uWj<erGdc%LNXpiuhT?S~=T#o2ren+q4x{-Gqoq~!LaZD-s?F@N{kj3MwH`Jd
zr_li|-qT-i#EJuBymcHs&cq}HoPv0AK+XRC4ai?YtPywwD4pkpFj-U1c+it^m65JC
z)&~(}!p=pm$St^^>$fNA@478jSj@i=owb?~Y!kgpqz#Un4ur`Hj+;rgpBcUIdzKPR
z(U{1FeHa*oUUD5ccgQK~-TUk@kVY2%EZnXX4{rSNc5TYcNp37!zZNVp^Jy!px}eJ_
zL=7i+%ND^ITl*{af7J&>R1!QNrlgpK#KY0KP)0cO)oAlJR?O9i3aNHS(z*K|XFd+H
zQdv3*S^QC7oM+{&HIbc&$BD?R_IsPcBR45asSdMI!4|qCJ#>$6q~=E<4q&o1)nxFh
zGfK5mmer_dDd)lJ8a-Z=wG1t2+AbKgK{=GoFAAy5ztEjT&sZwoKm10-gp0(M%%xU~
z^<a625ol>K=*EEiif7eNa?n5>mq6>%T2(uZDl#*5^fA|xw|s;wB?X8M)%pK|%?%sM
zJ?9(N@^@x~)}tm!`2;j$?XEmp76TXtr4zTM7~6;no|ogZn4DLv;?K_k6so_lMMEQI
zKI~ks{JGYc5h%Urz(-EHZ8F{A%%)kM_>asjP&{8U+25zVmdkNQg@g$nd=%`Ck+RQ<
z&3XOGHj*KlX5C_EV5QFB>-)l>7w^XXL!9*dy7Z7&@39$je5cV<6;dnDl`^|SnnTjZ
z^GnA4x$#$!4yJC<@!@e^>+$^q??Qu8L0+%2fO4#D(B+?(>MxgaB#CC4lgi1LvIBzC
zB)w0DM~q1}aOTR#l$!aW^)`%@=3|>F4uHluE3KIGCkmkt0+`%HEzjOmbf^TEK^t{L
zD)yq^=4le$u=P<Sd?L{oS0`LCz~I{!8ytKFn;Jq`R~3PUE%HUe3^!1yOAEk)@zL`8
z!86bHU4^9(qMEp!I!it*pxqE01(h6dU(jhEy2fAt%$6}_rwJCC(gXn-<o#POL%25O
zfgIQ;R6i(r46zQQgin;0cS=ls-2_|_qXZw{v*GM)_rh^&k*d^l(GcfnU&x#WD5^kN
zY5Q34D1&mWoJX1FPz{d1^ZuI&xj*X3gzjwIKh3#(jd(!_mjG*@iquznL7$8JV!y2)
zit4Ho(iBN;9q}wRyQ}l&m$TRNsDHz_jB}b0xtJ;vF{MS0AifIjr)-%|inh&Kpk(}G
zI9@~p02y~@c40*P+rW^YWYxXWoL&>>foHbF%DJjBqoFQ`KK_&>@oeDyAX!P)b$=w?
zIkFMVd`kH^ab9s+6U}PDo}Gyz+kbTdgtbRA_p6v8HS<MXQ8mI&d_fC{7xqjQSzZ<i
zB^H?_%@EEfPgJzGIpWthjOX8k^Nk_5^d|UDp9aRr-`{??@c-8|l2D%6(zvw~nLs0{
zicmtPcQ{>IVv%{V582F{gfM&&xM0v+LOHzPedlF7or^8ZE<;yo)lX?}JGc)jyS!lh
zYWUsq>P*vggb&^7Qrlv$72!FpBrahqS|>1#vU%e3c((Ersfzy$=S{m65b!ezcrnR>
zqde#3{E%i=GXOl)+jid?mvDv9I8=*0k|>vJGKRzvPa5hZYOLZ@hYOS8X?*~6xyj^w
z-riY6ONsq3fw47qN%}8$Ad;KM#iy&|r9(?o4A62za9!F^ey(eWGy>e1-|?G>KqKtb
zbI>|VAV;9SxHSWu#{sK^lDfk+{Y8}r<~llQpn8j`wBGd^qf{#@!5ybiGQ8h14DXSS
z53J`}7j6|w3uytFjHcCEnC5$kHjauv`RJmycG#ZID8j9_i8Dq4n?bf8a^Fy)a@Bl?
zASf_GmJ+SglAH8z2O#2=m8#^B{#=kE^=3;0R?J{EItRX(i2+?f0)+JlJWr>|d@neA
zs3EGzf16fn&Pu8F0s>!Rl(;=@)qg-_d*4L3Q6|*oupq`FA)M_EJyvJOJ;Oyw-frY)
z+#1W-yN61Fr~J|O3z5-@GX;>E0aJ*-OW_nXSPDUiJE*4=+;(SS)vU+qCO)TVzV@N>
z*cWMebc2+)gXkkaeBu5y6Nu16n!swr!bVT<$%eCBLjX~v=HmKux5?9|zg1^tdjd>a
zGD?~hN)NZ{=aqf->x;v3GiZfSaWe53rli8j8I__{LOqozRvst-9?<Avu*y<09o@Sv
zqpP$D4Uqpp%^Uh23-@}{&0yu{lq{D^WjHj3=o6axf)MhborcXNy6NA!t;-kZ(V(~#
zPLTTE?6v!|&MWLR21xve=(o{Lto>MK(9|yKMd!y>3d#@LqdBVpek<fda>f)LKp2MO
zailmcSM5`L#n~@5P*>z<7Bb<Yf60D#@L6JNg#daQ?&DMW$Li!@Zx5KV7@_u7WT_)D
z{3-Mee)O+{gHt5`fCnPJh!=gjXoV6Av5bm3x9wn9XP{R~*ja)&B8DTs@=^+yAbodP
zdy>P`=T#Q2Fc2^NS2+2Hl$ar;BrbP{!ba3ustW@nJ*!FKqQg}_sGJNC<j4J(s%NLa
z*=CXw`qeQklG6tYv)IU;G`p8+0mf6qf^IfKTL@~!;Ps-1tujbnt<Fwfl_oP%%!>cD
zikp~6l)AQ`w=bfigf(BKY@k3|sb3H1aD~Dm1o9^|*Uv`-I<x`m#hRvk_+z)v+rlnb
zpYl&MSs795oQ%kBxau-|C|>OD7x<yOxBVeWhzxC5vJ<NlPe0RR%<1XN;p2Lpiq4sU
z*LL|=jy+siK$p0L8L|_>+o;4P&|D?R11x&C$zHbt>&;kg`27&{gs_?b3yh$Dma;5`
zxD@{>hUjD0BO5$BaYZ&%j`n7SAe`R{Er6R0nh6bBw>9b?{+Xd@=+b(v8404{o((1t
zji|?PpP@btBUqSvkx?t1xrw0QS=<80-jcamB)u1GSM{S)=5el%JxUdyFEg&mvmgQW
zOwY_+%3>eIA7;zG5tTXSbz|c>aGLwdvfJsh8VXl2(wSx2aH%HGS>mnbScg-5GNzy#
z@<?~L)FGUIDDDabS0jr^2O`YHO&8Hyh#h<#-%h4zWhyJNzTByplv6psGCH58VRwJN
z0C7mihhw*XsCJg*?eLTE?QA!mix2?nE|ezpd^*u;=gFx`tE+l_HUo7k@s;<0--=2S
z{<A;H%(Jo}`?Unbvj>*hAQ8witqTDF*786v$}~*Dr`x2kMd5<D%((=Mp3NlIZb=2T
zOF=w4m=5c+Y9{?5Xz~rApWx8#HboHLb0YuIT7YGR2W+9S$8+>rSdRLp5n;$ub#N#U
z>uS$H-yDaUJwrDvSokvrjE;WjS>=8LR}}JjfPv%h_5ay04FhFO`9yz#^Cc%+e8K73
zy7X`=JhiczMd&hAZM){7P9#N7F#)&s0_S#Xv}yc&1N{AiQ-snn#_q1BA0|rWxv72d
zUGONh4S{&7>GS#c*GE(;7lX9wz$G*kUcK^y-_rj2+=5Z1#|SjVypCHp_1ztK_&*}(
z)7|~oVBtv$fDNd@&?0dxDrG$j@;Tc{T@>%&F)wq*!yYzj*I$3_z38CAGIo&ZZFu8x
zFp`W-#0+YAfQMubOudKqLoFhM=C#2h>SVN0$F6SF8JknD7)`@4`K28HcxT@KaCM2n
zFv5NE8*_b;F~kI!%tc6ecFpB^qicS_uzt47cV;H9R@4xV1J;ed37SvNihKk!r>3>E
z89}M7354!v%*2$QjnrU+>wPXXv>`RDgF^j=8P1yYoroJ+_P03Mw-E#^RtOB7=2)Hx
z&Gb}WU2f5=Z?mbzRQgXcT9sVjn*%|+!~=%lwXWtJb2-!Y3y&62g2ZW{2*(&77&dYj
zIy|zzi(wSeLPV`8#ydBmLGpm?8?DBj3kyw=WRE3RUfUAZomMTC$ebv(2}dt*Bo6)Z
zJ1q0w`#r{h#!LIWG7;FmM|L|k+I$5*K}()~%YL^MMG4Pxm#!+Y0UBZ$bm!NxN1BNg
zx6A5c5Y6-YSH>GXK`9$Se={iSI}O9}jt_|EN2ysE{+-^(nIr)ordJ`)fw`dWP-Fj9
zfr13nn212AH5P+UCfv5{9@+m)_`?|hJw96Pk=(Dpq23hLD#x)AkvsRiAELn{UdnEF
zw=*SoCAIW7Nq+doknxLl2s1a_{h%VN#0^B&ZxY&Z1P2Xpd!bc(Q5##)wj#|wA94=m
z5%G-}3L5^p;s2i&K-Dfs!5$A$dxz|Yd_X2!?BA>a=`g+jz^hnv*3$FUT*dupnAa9K
z6o!;x3h>IE(paRLQ->LuPGj#xR8EQQ^KRK<buKTD1BnU&e}id)`=*yL4NL+i=+DRe
zjL8UinF81+#L0|9B8ggCS_z(BjYO?>76?pf$x(FHbW;V^2(lyTpj*57uO0YS-;7Xh
zyw#+FFoJH65}Eeo;N?eSx>r{h4L3-9SZ`F_S$R}MGDF3eP>T=30(YS0zrWq48DP}k
z<=GHFeBz-QN@5A_U`x~TuQe{W7|rv1KHAXn^s6=sicehFF;lN}^$vsuqakk-nbE<H
z8T`X{PXDEO)_SPtpKmY~4P}4H%r)}iDx8=D2VF&nQ!(zyNp1KBU^4}>KOkASJX}uS
zu`=+{gGLgmAr#x>1E$v`x#qQ;IT2xcD+mb`6y(1FYCT3QKY`#7i#3sf<{P?abq!%G
z5A-~}AG}D-3@7lLKwzh#Z%Q%tvV=uZIhfao&tg`eYuXB`kYeFk1OH^bmoWz`vb~eA
znS{WP5VT82Okrrgp*-^mZKi!|JbuOfWOG+to+-3Qor*&<WqH3grCmmrZVQpWBj?#~
z2SQ{SLHnYnS~LI~I{ZWFTEROB6qg0`#d*!!SnOV~hy(~|^R885qnxTiqp>(u5t?E6
zD)U&OB-B_8a(-y%ghD}E)+Z`^rG&aWJ0SbRNR@X{*68j5DKrA>T#hCHCp&e7CSqT@
zg;lDh12i-+jBF&I#I2ob&oo}6dy^3oOTd~x|Lm~2z@Pv~Wrl^C4S^cml)E{_JZ&L<
z{gm!Zk0M6uu%N-IUjHC05b;5GNY3$MWl{WM*^IQ@3<Ioa-E<xf`?*51p#Vk)Z6UvY
zB1EQCMLPDY!QUD)`El;L!DPVE>;I#N#Np6=2ABga{Y_05!9>V0l^!r76@JH#fA4-v
zy3ei4ts%yTUm>6i9aif;c_Eq+QrFa#YZGg>{|ub;R|s$1s^v+OF8q)ljDqhPZ>thR
zst3Qg@D85mKg1gP84#wVufRZhO|^$9T$6t(xB6kl2TMw;l*Vm;CDR1CQY1o5cd#V*
z66ac2Xa}JejmV`jb0eqLqI(-XeE}VTE|0w5gz(lNqMau`%M$qyg%FqFzwRuTYHN)q
zFqJ}lp#IMl7ET`DGYFzjnw4cH%iTd}b1m-4`&&ml_*tMbUCp|mFcR1bB{p|t^b-NY
zhaa0wBtOLG$Hc<Z@m%NXUs1wwy$F?Hdy!X{(dH9b>Gal8Jrs(-I?3~*QRzmYf>`rJ
zgh&qHHm}MoVg+JRbc6f`2@%g;AV<^0{zah!jLTl_BZo7&yUZaT1E2j%Q}p5rH9mj-
zT+Q{FT&`Rq`4>`UIJ=EwS{$*dgjpNxMvDU`>o(<uQ0HP4D<!yJ8|mrB&6fs2HYfmg
zPhlv$VXI+Mw&q$I7EQg0f~F@JEQ2<dlNwB|sya1`(|5kH`i(g3=;0nSgT3ZS6U&vf
zY_R1e>HdZBtJnDhFg$P4_b-UEQUJ(KEDqrziE?D>!=11jmv9tfQsIMzIsE$Kmxy#T
z=4BY4o0}<)!v0Xj9ym5)Z`b?sYG5F0X}Du8zsu~Xbu7|v<(TEU0&)U|U=4;b?f??S
z3A1EAddG*L52R4>Jt(pm|BBPNyzW4wJl6d7%#9(M`^IO|XO|2zNorB~9Qc(zbd-a3
zEy;W>plPH9UMQ2~B2`R?zY?tgmt(~345F@N-?-f!yvvmqpglhtPg1s)6U{<lbQL~&
zBC=<ps9^oG+jrwW5d4l-a9sl)7J*>-xeb?sPs;CkZS8(r7!_TU0LGzYBxKI(gJXSV
zh<V^L<KQ6N7$AY5-lu5xLj_jYu@J<FT37lXNC)GkKz|g(9Vw)yV{q9V`Xd_tsl?e%
z>T_|fU2e;8<n95+!L;SUMcZF%xW83!YC>>Ye{xDI>&srI$|CdT-}Sh5w?k~Edpo}r
z2TY>PZk;lZK$fa;H!eyXME3(8ubDegb{Kz2JJ1b9HM!l**-W%Rfg*mA$-ZW{mbMGT
zl84I9hf>yu4ZPjTfV1^JBi3z+DXS?jS@dQX*xArkN+?9)0-~&%;(Yu-nIn{<x+vvT
zQ=&>o!?_qI@8f)aWo#WF+7bWPRlx8(-I_WX>t=PfXrShwdYHhKx?E{;6`2tORkin{
zYN7+u)~%%osy^IV$C;e@{O(iOci-$=e|vKIULN7_J3cL~DRQ?OL0eU{Kmv1x(;WZN
zPKO}Ahtq@U_lO8EWfLIZP`l=D|A0VnCLO&X42C*<rT_XhzY3^m06)B32uAF$m^DLK
z>PIi-6_#bYGw-#lP)-u2k4id7GlkS`H)Xjf`g{<Ml1pLa$Q3%nz*I>&4#-8L1`DPa
z`W;%{H_ML(A2on21R}6uWjZxA)dawlnme<2o?H<%pOAjXPZ;w(nv3{m;>C|nb=1g?
z`|Hqf+PO^)!hc)I>E@MrG$V-_)4|k)O<e)%%^ZAx_67{)EMd{PL7!3Q#yc>;&YtRm
zftwwBdFZP4KU}pW%M69r>R<%xu;wGrf0TWCS@f=k?tJ~LKKtvX8d3cuC52qirr%wa
za{Kt?{^&s<s9&g+msH=U4F}rZUup9XDSNUSp>mCdHo<dxX#9x!Yr0XYNWlRjTfEWf
z{SL^B42i~jtr6oo3D$!1WD#quYxFF{-Rv{Dvu;=!etQfbwmN%WKbCk@d+hH~AZbaW
z{A2lOe-sWx-+Fy`AbULlgHw};XO5xC21i>mcSSmIgd24FuYX45S$Zx8&Dp^a`1;If
zIb<&)@1oqzq;Tz61ZCL!$BYlYu4!q<cKUt^_F|em|K+#{+{^_j!?K8@2Db0Z>xoLc
zwK-SFgm&#WA$ucq7)DvBF#hMSY!krQm3s>fd`?oUgZaVu&-3B+8P&k*;E57^#IuTk
zE&CwW*Gv)H{CYf)Noh7nFlIJv;wRY(*-)e5#w-w!Efo`1uQQijmtLwf{}LbAqKbX0
zbtR+u!Mnn?g8~;w&`4Civu6cgu&=^AT8{kq`GjviO3tis$D$nsjHUI>$MgDTW)j)<
zGVUw&ZNJACKzJXWujFvLHCWw+AF#oxA8fkM)YVv!BeEchg=C$+?0>Q~o{2*8NTfZT
z9@B96)P4HHW+~W^XP<%YS7fNm(!f95U@o{E=9!X{l8M$8Ou+~Zs(1Is7!zH-vtmeN
zBe_KT+SD)>QMZ*u$}k`W2swMRuuHB9n!Pi_um9{C-`%_G4Y}&b2mt%H5b#15)1ag7
zKE~D2Qsw8(1pzAoS(aF|KJ;?uWyWeectCW5xgZDL-ufuU4yIp$6o*0@0Rs_52qBIV
zyqi#LYd@sxyDR^(uW>3R`voq$)tLN~anEGYJBpovB1(S;Ontr>ccN0qys}E{eClT#
z9`%@{Tl&kN1VV91EfdbB)KO}rD7xrLlSyIrji(m%oo(<svFyPccD2ENmCW5u%tK;*
z-?Xy2IT2&H!`E?qYiT5YWU8!0$0eAmQ@u7&h9f+uqij_E#aO*1h?=(&R@=N^539HR
zUn+*sGd^`TjSa$9c?XP#9k~yXwo*xHU#|U7t#&XY@Yjb*d4aK-6CAqS0X_c_85dTN
zBScSUrEaiWeR-Dje!GH^myRL2Tf>_Y6X4aFk^sB$GOm!l`rVzu@{b&k5x%}zs&Wvx
zfWg?oRKodgz;TTTFS%7CzL-;!9qtS+E&JMx@r)C@`0)%~T0Po{6cSq3Ol~H}a|1%-
zAXEhTDwtr;bJd(j-Tf#C=7*@l@4^SKx(i>T*x8g8E;7I7eBIEy!)v|M&esW8ATJoM
z>cO3NtZ&RBE$*%s86_pP?Lp683!vFJtS=ac#@_JsrYC80#SwJ7oIMjRRrh=6eg?g}
zcrt_`U#_4Z6{4E38Bgtu$1B<r5E`>ta{jaN$s;x3&w;r)5{$ebWxX+}Xj&~^+#p%Q
z8TaggYR$eTCDqem<fB@<0xG8V%ZEnd(m@X7ua76Ivua!&PW>p}9-nb~fQ86M6Z@Nb
zPjDvGS)Cn5k_6e$16KSGuNdC$%_JfMtn}$G;B)SENE$o?lwJ+irBU_MYBtMED?&M1
z%q^pWLabC<g4tJ0(V>U`S_>sL8SS`DNjpigN}@*?4)3VI9JbKUEBcLFmsPsI&r7&U
zHz_Ozg#rx-dlq6$FB2OWbgc%ntx^N0>@i6ibr0?MkXQvQE=x7F<3ih<mHr+tPb(mH
zUeiUzXb;UWhddp<AurM>u1!!TRa7&4?OI^$?W^G>1uh!j_;%xf{@Qki^DjK}3yUG*
z6Ft4kUGoAc6q-{@xNHGh1hvP2m?IS7qaf@Qj6dsR!P|c(MuVv?ca5w6C8nj3%$3+)
zQac~{`f-?_O6kM5$CB#W_ZwbTAWo9<;|bvSxTCGWa`*^s<JMCXb5Gl9a2=C5m8fAt
zu^;xs=K6IpS<g*5+zI~@N^2j&1OAXaQDHP=xU`(Z_91#qe!%!9-%e(=XSUz6`#W50
zB~iNwWWqbOouqNvngMNpx$<@v27FwZ!A*#oyrQeIp9^76dcZ);RLNPvq4ggQ<p%^b
za*;@~5f*E-?wcmz8?)#?G%-3IahwN5$3<YjA-oG0`q_ZZpKs8on+U>*ko9}|p(cUl
zESYk5Cw^b=eTvuT$@7%$1wu42HhH3o^RBm!7W8}yF)l`aeC-g-Uoat}#ApOs{su<5
z(zK+<23kYo7u)em33<u^ilT@-_R0W8!)&NF+rIo8+p}$n?`0g#q<(#Qx|Y;_WS41_
z;lnr^o9&Nnv%{4)U0n}JG<Can!;!zb`tDmy;4V-TTGY$5(=pf2AEKC(p`cD%a0f-U
z0_8`~AeN&nL$&k89IUL3!*A{tyB%}LiQz=9BS4FsGB&mwpm@=ZgN-O0m`evCa%2%B
zhP?mZ!X?bFr_jV;JHE;<EIc%k>(#{wI<Xu#V5b@e(a)&1i}%C6iBw}ya;|*{hcTgJ
zCbaC70rsM1R}X9kf9I1znvLu@k!Wc1$;P7<Kt~tzXGryrnul0<e5u1`J-}mk{TM3Q
z4;bz+CDw%eYM6_~aye}ipP1BN>P3ndtz+zY_Y?oPVadJ7m5GGuRM@o^`RV7fJHgh#
z-87o>Jghc%gj=?Z|DJsa{*}m0%n|UR@by4MOoWpodk}fw&j*31-r<d=U3t=ULqzwR
z;23wnw!?gF*s8Lu=1KPtGy0on1p%#)8P03-B^|o+)hpFHwJA9wzRnGW&2&+(KcdW(
zp_Wh3Plm2TH8`pkV(D*nTgiLysYwd7$q>;A+JCH!IAbnB+vP{)M@_w|!j8I)1#vgb
zlN4|G?u9pG7(zo@6cVfYS+ZTrQNxTkt;G2Uhr%*=Kt8j%^Ls$gmICMA1on*S{Y2ME
z=Bw0iB$+gSOX2F)hXcXv9#@#GyS<s2?*~q~jqUc@q@&f!sae69CP*z1O!{)weYLoR
z@3m$<j(c3t!|>DC=+itc1+V|}wX&S{#<25;!eDL&=83McVM=7l>MosP=00=3TN5y5
z(Jx+>3L7l%D$Qzq!9JfYrRicW;y}{%MhF8vgSCCm4}5{5WbC3F?X|LaluR3jzE-cF
zD%u-hj>4X!4{ke&)Xt&o!D_~v%eV1ePgt+dBJcYuor}C~<-O<J&;uR1-+g{Qr}uJr
z7?VdV8U~#-K8h;4{9I^vW6+}H$=RP7=0)#swE>nMsPSK}e=K((73o;?x1t9pnX%r5
zvaxW{f|0Kh5Y`;jtrt>#i0%ECa8_6C;}Sfc+r=(^zD%M;!*QaL_;G>Sco|pS4pGL^
z;sfJP;J6bP-?D(N7(V_#Vlv+>MO^P);rab2aK3KSaT)X_(`VwAB}0*-?Q{~NNigec
z?wVdLN<Q@PwVL_K?)h{wC%;l4e8PWXS-@ct?f}7{gjaKAfVg<`ENSNA*Jst%l&PZy
zcgC>xy^7S#lb)V7ars|=V<9J({sL(E)<*F7tFHJP&7L%`d*5QoojImF2C&Ed$s*h@
zmX{KpLgG~VK%Qgm-_Sn(qbi5!kV<G8lQ{D_<|aYS13V@q2ka}>2BEy<#Y%)%BdA|D
zt$e_uP;7IsaU<z@KR@KcOEF@hSy#%uAY97_3wja-nt+J(=pK;CJy|793G*HQQio`A
zupxt#IP;1)f{TTCW0HmO!mL)NS4;ERX1Q!n-zNi;8H`YC{KRpe$jMQlDb!wnjfA4P
zuX&^jP-B;p@X6LD+;8ra?<En#Eb)O2O*aK&plhI)9ETyq*?E4hfz>iFf=y*m4Ytr<
z)b5a&aGI_=S8kU@cgG$HI-EVWw9tbOWP`unE<Bt}1XK&(cU*D4Hmz<n3?leVPRI@X
zpvJ#{uzsTOp^op{7}X0Z+Zw=zZ~l9g|IK#<nAm8O1Y)1`ED<M$wVA0Au*(8bc7x0`
zP_^47@+(L$(<bQp>J9c|JJiurXUMenu(mb-;s5~*|Bv#*uEra6cHFfnU;%k|P0K_j
z=V&S4C*(9R(`r3;?826bi^!*K;%AyP4`V5Kk`>K}JO7B7IHY2RZDuq=CZ^$@Ci_$#
zL#4aDI=0Jp9h|MO$<s})Q>F}mj_a?EJHJa)IlAOE|6${QloE4Ux{tRUyjGwXhH_j|
zm$_Aln;>XsND}cft$Mr@JQ6jha*pv6RLcF1gn@@m8ngbL^AQCm18Q<*_d!WmVx~}$
z)F=`NvVOz&`sJ!n*!ks{2i2L?<2}>vkKK#?C;E)btvt_E7(kljPlossN~JK@m4saP
zJ?$OtFjR<iwQ4w3z^QMJDGc?mL(NteQpM@O877y%@KFfehasu_@RBc%;(wkb23gsq
z)7U_nhSC(81dCoco_PzdP=?UShPfsa-d{ZC7@Cs~CXNk66fcB}r_a;aJfL*9EuF4s
zb}k>}kiKs=0;g+T7QO4y%rpn<GsVAX$A{=>$e)^nfxqDY$2$c~c<;jqOdEl9z+DE8
z<8NjUkhccjv|<BgE}Jc<-+kfbtFuSKUqX9YB{XE#0##U%-i)lVSfh)ED}^L-;L4uy
z>y^LZGM|WO733@&9Vh^kHMA!Y!JnxUroLjuaNS>yNzp3LFD||CbeiR>qHZ-1)2bwH
zbt}k$z<qj~+nKELLDU-`Sr{p?|LO_`O}Z0HZ_d31``(C!gAd~(7rLiEGbjC5qXLi2
z#Ze)m7Zm%QjErvywyqAt>{jpqP}HBQ3^9!FLCUiefLiP4g^VSZC#N9ejBpqm_R(+9
z=7<d335hhgrWBd?T~X)z_^LsJZuKm5CP59o6_tR{I~KuKZ=~%k*5yFnMq>NzTV8m*
z^}M$(?4Mg&>^#tJy1c}fX!w5a=LT=*K~Nf^m=VD{4r-htX#l@=H`{p6dGLd@_xB_7
z%uJZ6i0oMCmuk$oN9HFnMO6u0POa;Y5^e@9S`nEZ^aW^6l0oAqijj4BwGTHLOAYKY
zXMIKwaN43U$;IXuodw4@d)u{IS{7HO{KDzmj$Oxz9{P9y^>R)4ua*>J<rt6d^gjK>
z-*Jh^Nm?o8C4FosUt3OcvVTgOUi0DrTdAEGbnfj+@}-f@%&$z)!gPh4%B=g>l^nl-
zj1mLi)VD`8ymDo4Kd!_;e_h&i$Lp(Sy+be8WKIKp59B^Jvy55^xL#gC-Y>DEe@w``
zc|LYNs!vYz&F8F!Q#KG!p<-ErC%T0oXa3k00T>pci2{D*oC$G`=bAyJO@cR^$2-D`
zp8UQ~dy#Qx0;3y*|8unO%QU2F3j4s3j3657#axYf(JIF1F{F3m*~>0;{EA?qHXjV%
z`x8ZP_rknZpAi_R99<5qYmKa(G`(baVABPIJ{!6dQcUwl-Iu%HX8cdn9~0+Ya$6V?
z)ldIu5Tb2+ag7HUA}RM0i2v%0=V*UliztElJ?)?m75VydHrcrS5k^yC7N!R6?fcD#
zv0`+}-7DcVUyq!Q;Aud)7+8C6RzfP!ZT^;(L?)d53vFc9w4lQlH_&w^bRn>AnusnK
z5t-C9mK`hQ+eJ24?ydyFG~3?_#({Mq-`O>{elt`hNqRdCPRhA?1RHOXKW(?(2u#RP
zwIJZDU50f00=^n2Pk#t-6foBBq`zX%{l38m@vzwb-ovYpY3ufM&l~Q9n)<L^J#T?j
z%DuPuS-(38aq>k;On&UpH)}zJXVjl(<|xZsoP%Mu3EuNZ!{IQdf6D}m_A-pLgQsLg
z=-|TyQ)Fruu|%0RBUEG-XbFr*9=<Mvb>)u1N>tpTs|g}};$vlmRxcrzl!Gsikz5D=
zcvqfl{GZ4IQ_SU5B)HYes;E!u=)=g;HRR1tFTm4;a;yjoG8zA;R_HwmpRezR6H(b&
z7jdlJ?zzJd<|j-dW2Z-^m&t``{NLYC>UUmUW`OA0cD7Ks``*Ft^{;U)@k^=Y<<6SG
z9HbS2tIiHub%w=DmVoo5jNh8<w`Ia6`Z2f4#<#~tGE&W#@niG++R`9UGPTTZw`w?&
z9W&<l&t51jVk|ERRSHMV2kRe~lw64|w>J#%ySq2&=#(Rpdp!;p225I+;JT|r(LRTv
zfzvJ(QCKdyK5Y5e3;UeoviL-^q^!a3o9geA>hN>mzd1{QeP8FHvj^0}!r$@uQ#p&8
zbX)h!ScU>pp3H)OVkYtR)z2Zhb~l8a@2H{ePZ-Xz4%pL24V>9Ut9dj(qJ@8JSAC|7
z_g_uqKKWoDu%xsa-_O0TAQQSu=>mC*n&G_AST?D_9vwsCMoSdE-V=M@mTF6$^&Ak+
zYV)-+C{+t3axQ&1z4=xFm_bfXmQyoHZTZG>e5zuD*zp9Jt?=8Ci!WyS0Xzv1nWdxV
z8WD1jXh>a?Q@nX|-*dNBPyRCOqtEz`9UJ8iI^qqqKr1P+xQKpyZ=#i8V^p|Cz;bhK
z)y4Ev$1^{#K!ahXo}sX>_4}WiJhB-=%eI0y{46_(zRXN28=l|(2EzTrEu0L@vH@nA
zi=UPeQxqgcT0lR|6$TVWuicC2-t+}S^*o?L*}&1`71i+a+y*_z1Z!ss)PC<BPX9uD
z&s8)%!?6SRZdJ4x6#@D0qAuhe?QRHHZ_WLk%33nAf?_%-NB5I|sH@PbtEIR`bH}+$
zDsh9idWGY&G2Yk8ognB`9rZsu%hVo=+RNd!17tW>Sq?-BTemL#^)(e0fMrtM{f`0d
zEbFOdKzu=ItrN%iahtX1)Qkb2u-&4{6;Vkw^y}KtWmZ^z_>^n&{+C>l;YtmSS`&PW
z-=LM7FdKla!{@%z6qu{)+&N1sVq)?u)+5a7S=zy+YHo3dwt7$YK6V9bx1q_tvI^JE
z43mbN1hJledL;`+iKf$ZSfO%0q7hBUkm=1zL>5soxFnGMkQ?SPZuB2|z$82sNlN({
zY+xvx?xhxUl(|YZds^`acDXEFN@%L|ZYYxq4D;^c8SMOeR1xHU)m{KOErw0A>C*oy
z1?A`ahS$*soR2m0bS!Tnj{9+T$A=t>MB~sa@gRN8Tm4`_3@6j|I)W4If<;PHvU3gw
z24`PB;JhnxP3f&+e52^QvRy3t60GlAmUkspWw<uR;JyFnp*r@eMkQiFctRf~D{1SS
zQjiMJz>y*rpQH*u3{kF?b>~Y*QBVm5Ni5(D?w0UWy`u5@xv+_w6ZOq{Yql&uxnzXP
z$^FJS+#7<OUqu4;VJ4GE$o_lXeUAT>0aWze{nmnaVeo`Mbzbf-uX8BqmqM17hVPH&
z`jsP?LRno%2x3TjijxG1JY9r-Zxd^>Gr~&Xe@~|XAkm_2F>LP102|wMG(3E1v!$gV
z4fS1xTT?R`yvF1tFr3Z}gJtAat=eS5PCz2~a|9Db)0U$xd05=Z%8X%diq_<&Mv*Rg
zq8y9hxeU6Oot%2_B5gX59&9wg&MBBP0rx+rVRs3|P>UWy)Qn8h!Os%jFVCbZ-%uUT
z;d=PYo=hwfs0X4a3f`0ni07*($TjaZQy>sXuWY0{(v1muyd%qgy?3u7Gv|^RIOif6
zsGu+2-mupSmeN~~*2S)1S$e2pD5U;XQ_q5}-(HIc`nmj2yhwKVl5{%d42;wN`~zFM
z<bM<e2g;FIX+L<Tex7ctz#JTD6iF1X2f1Xw&NI;$bnM%>R<jy&_)S<9T-l;O8Puzb
z*z+UsP;x@Ig2iq{elDVt3lckh;v|b@AZGtLeAxVK4G}0^+ds6~P3-yRivwWmKw$t|
z55-j@^qnt|iimkZsj{^JV<a?!+}h}C2SN@*W`^_5a4It-ec2=GCk*z%y2)GhODY?s
z`M=hL#r8iopR#7ZG^6~CV7rXAmaE=3q33-j&_|F{c?yw>bX?p1Le*@$R=_ja_j|$D
za@=`rT&ah$>P*Vk%Y^eFaFxp2wx2rgjwI~lMpXL!W~ml!d={|u<EEiK-0eeZQbH7;
z2DWm54<KwMG%-}A*if~p?<_r5MoJbNVF)E#As^xtcAeg${fEE5eIJNnc_wf7GE`*^
z)r({!s4Fh&VwD%hnLcPzZaDqSeY75me<zk*%l!u!0v?oSgFYd-j^ZR5#q4n!lr13L
zJ+^eydNVb1`0%UJaN9tFUVTiuH}*_3Sl3j6TkAY*l*w39kpf`2x@o{e($$n~JZK!J
z&}v*R+rp-NcS%hbrgXL8_*ybAmBv;rgfOEh)ZrGu00PEr+%CzAaaXWgC!)_rM)mVJ
zgx@mm{Qfr8zSq5Ij!5KUX&Y+(|Fi%h=OXeG{;66p3&1YDRmED>BLh*yw78O#3<xV#
zKy?74)R9A59#+qpEdR@R!~OnG^R2<j^FXqQYPcI&!pO=Sj`ANpfqAW?D=ogbX{JMZ
zMpbpwO-Y7AuXmS@W5aTNjlU7<;d-9i9Xa9uR%Q<nikj~@$~N2lp=ztCgSjP!?m&wb
z6Q4NwA$++io&H57f}bW%ENPru8Jp*hLG$H7F7lv<)+Dj|oJ%5NOxnC|+GE*v7;QR1
zF)Z^RqBeQQ?h*`9Q$P&t85LE`N(|q8d<$4oN9J+ENQ`fQ?0k}H+or$~PLf#fusu?n
z&b?>hSVTJDEuRQYJ=ZM!V*%INTK~R%v;$7)9yKa~4@xF5Q3Ynd#lGL0@NuUWl|;=1
zoJgdLpU#Sz+3cjpi82%DEU6zb*p0AFlj|zQC$G|_&qcg-t28}_UTxy}_IHw}ltj>F
z0TX`qfzi37W_;1)Q3s;^1}S{T1`?pz<_pS0uEr@S3KD!bkZRW(G8w$&09RB+VwQ`+
zX+Iz?L<Sg^Brp{p97hheMkv#!WTW26L^6@-8y_v+tX<N&H}GSeKd<Pa{Sn<ZAnE7X
zXgwicFancZjx&g!y#k(l=N=dT;yu@vSrV8jK;5FaFmQnAGG~fsNRW70-qsBgi7I?^
zTsCR8@j1@FE96S-IVuTC%tW5%149OVof*h?u#dYOU0<TX$?ytnhhG2%J)rf;Uq(Y^
zs=(J+i8W0A0()GgBvv7WcJLtzYlXaMcaUmoc=-n#Ydg2vOpL2OA|q3zIWndV6Jld?
z>bZetlYxsumdPnk#%e3EIVVY&id$Tmh=iNd*n&{p-i#`SdCG(Q2wM%Ow^1hS^_8k;
zyOUX5f+LT~mv#pvBvvfPwS1!IyVIcW>30SJpTMI?s*h_h$$P{2`@uT)gZr?|<einU
zLv2i)eC7m*_HXp{0D9tkhK^eHaj5XE-gtccv~7Y|3tKb~R=w@L;TYKE&48c~n^|+f
z`6?V+IPQvYVk;-90&m$X@?TV{4|5KN`5>;<Qz*`rDO*?*Ph{R3E%)_%5<!@4*b$ku
zPp2MSYB(+>D*0R@BoGmK8~+^oSs(}ohecqNZ#QArWyMfrGn&dT+-fV!AV4wuld8C9
z&k}6M^7_Jmb@nRfUxH@A>x7SsvZ72VHtMbIKYiohLy<>!1pQakOU8aX4ncPa*+;ML
zvE@H7pLMQ?n)94}R&El4e_!Fha7I|VT@FW7X%%9Mpq%1|^=081L<+@Epn0TyeN!nc
zNI7L0!b|y803Pf>6SKrI5SaYt`9kEqU$ydt<}|P1eg5Ym+^n3{YcvAC^m+P3wWhcY
zKHczgV{-UV$kx{6r7K#-^uFz!xjItBKm*Tgx;7OKKf3WlNyZzTG^eIEEga{8#(|&t
z?#V9)y}ts8;51+v1EpoKj58HsIu5@-DJfHEt0nqf=YHQLZm4o1#)`=N@^=M9`}_A_
z++IQ#E=*KA#8zG-E{gQUCeqMXGLI>5`9a)+fYlFx{w3+q!nR7-V!JmQ>gKKLza3zO
zJMOsT6G<_t5+$S+wU7Sizw`oC5sf>uqxWY`OE~$1Dzh1Pnh4PhZ&%C7$B%BBDOUQ;
zSZd$#%$I7DW2xyz=A3iW+oxxFCwp>8;;C0>6$=>!@2-NqcM<V<MIc^<*jSL+mLrKY
z-l_r<e1fVuaKDTcFdqO!7OEST!1EWFtqLXKxRV!9Iu*R<Vc3O`D+qWKoKIJY7!|ct
zv>20qUl2qvRo01wy_ChAzHBV{*5nejt1c}6Sf|!Ji9Gk8)<dR03Ch&tE=u{`@=NQY
z<>Y?J+G~0i{|}oI8)laK=7?6ah4vfU$0yY}qCV!0)!19FGQ)4*<(10m$(da?J792-
zT&PUDPuM-;G^!Sn3!YWOsgEVw^gn%K`qCs2F_j?{x~T}<KFs2$TOb8WFVp28lzRK(
z*qjV%2sS03%tZpxs^~+nd2S&v$E1NRJ3OxoSf2YBfEB7{n=RV#BhascPp7;fb^D@e
z7Xp%iK<f3MECZ9n^;lzRH?p3X5i0>x*CF#`K09B3NJg@JxBUWXps8kCUSMT?J&B4$
zjXP#wF6D;$H`GTWHG)fTy?uitz97xcpO^~_W8XY=1YAQvJX%(ytrS<Dj7<Hz9PDGl
z^h)&aBB$GSMCzDVbyxanMpHArleb@!q5r{$I|i~v@JAj$Q6gGVqY%`o>Kem8MbOji
z7${O^wP9}Uhy_;Y41V;xJ)f(YTS8)Z{$Zyf#9;LDu$J3NU6Bb980>hMx;p4?+*ee`
zzx;yvwKV;|h9U9^O7u2ID0yrqWIT6T);#kud`TnN$y4Tk=rX>!GDSu)Fv@n%pzDPm
zC6md|h`UC3N#~f?ad*Crp`9?;`dg&5iYYMYoY(0n8O)W-(5271<uYZcd??=C?59L_
zX!wfGTc?pvM$RYc1y2W0PqubTn7Rcejy@Md#F%cE&xjXlfFFLnnA_)MR)6B3$_^iR
zvR&%f(0eY%*Gxilj8(SSZ+@g52>vLTWws5p)!_e210`uiAy)$v(vPHAsfh={VuwCr
z7Fh7!t{(q((@tPTb1N^ws3cBpp_8jzia=1heyhX~)f1p0hT`M%aShohG=_#YPQ@wC
zKKHPD??3YJ`<RHf!A3BM^4*}#t}$O36en^JPPlBOLulVjf=DGbc2cu)hJ&J__Dpbz
z<jw<mu8}kY(4<Q}jO+4(zmRE4o{x!h)y}j(w@w*GItv}{t>m40WC`JJm-haHM1a@G
z>$A-&bRvkqZ>dkPF*84fhHK(S*e|O>^BPY`SZ((9Mf<d!MJN9^YS~7SrW}6?C8()i
zD?{=7<fOm6`tJ+YByaC#G(K5EvZW_irA7p`XTx3*{+ov%K6=J$AenD6?((|+#Oma%
z{3%--Y!P7%sIVAVwIW+p)7bO!a4;Vx!+VcfWIrmBT4%U$L?x``rPu}?C<eYQfwY#H
z5jkn!)*&W@H!t?0YMI|}8vex{#{aSNn!+9lv9YdL;#j4ZFR_6NPV~)gE*O=TlOno0
zR3WEqq>0-Vwu0Vb?N~fcgv+17jPc00lc~WY>Ov!$va8U6SvE{YW+Pf=<k{rR{UL3M
z>(DLAUX1kxq%hlkvNE4kHlWv{W)TyqFg(mNriTMr?zpAzxXC<FIOF1wqejVh&PF3H
z(jt-WEv`}Q3$NvF%wR&h%@uS4aJm2iauBda2zXY6Vqzas$r%~Mm;-YRq*~bJnEkCU
z52nIIdMT-+G~1o$+Wa)*vaJ+(EWOD3L;fo?J(dai_p|q1R*q-f^N5bQzEGp0F4l3A
zskhvZXKvDSQzkS8!_rcZHN4M`L4hk8*Z^Y|b!~<JowwNJB$anr4I`NbmJ|8C9ifrl
z(L8UKz}O9kBuu7&3~Rlyj)B_AP7qmUicc03u%ZIL_`iZT{&%Qi@<)(ear9&&5u$Zj
zp%?LUw|u*YAVBPe#m2Jd*Hpkp(t2Gc9i|P*JChgU-><EJfX?CwS@yJpanlccJ>2>6
zEdFgQYVMtfZ~-~g-V)|47TRq1_*T}1o+4=nzJs4RanpSrZ>*S)F;*Hy`UvDq@SnU9
zf@i-VIsFf>_&a*D!I>WXRc+(cpCh5;K@M%dIwLD45q7vXW+7C(@XWiz)(^~NV>T;p
z;2w>9YqrHe>D^9HSLH9@@^UV(F{aC+uBCQP>gJh=)KS_l9UnA&f8qk`a>`5M)%h-m
z$bMk*Oc&uC5}2i4P2A1(Mp_v7xq7n0YDg+zPU`dDPH^2JRk<?Z34I6mHg7dnz*}h*
z1>Z@R?xkxRR(b%O5)CO8DVrzW7Gtc4&?^R+CTvmho{zXB<Y-Vy)Pbg}=H%>~uNB}+
zU-H-OfG&hKZ|_tPL9Vp)HTR~A<VJ}}fnEyk4+AX~BvQ47>sO8m@|{<Db0aS;<PS$H
zhFMl8&Zr$v=X}Y1g$!-+tx!FvLV=JJl=nQ3o;f?Z9%8i4luj&P^Haxxv2S9hc^iN*
z_~*HlgP#^8lWOne0}+1~81@~TGnT{taFJf_gg=|_eRgviUQR{e)hbPW4{O3@1kKyN
ztt#v&>d-Hyn1Lg+!>3q1_WyYL28K%CFWPL|wr#t~HQ7zJT_=08Jz10OlbbZzwlz(z
zd*=7Q_kMzN-Us_xd+oi~qMax&-49Ptu8;Nzs0XuhtRi&uy9UFqK7YPB_d|!v*$ig?
z%PL6f_iOz&<ydqWu*n`-N*{A37AEp0+$?!5A=*p)a*Mj#yDKYMwQ7CJu)qumId$4?
zF-sfJ1nF}irdX^qY{p^bAZrWE-BzX(%Nr67ij{ks5iY10pS_$i9XFl(Nv;3orIW^;
zf2Isebcy;0qbu?%Vi%6hLwA!mJiQ~>IaS3z9n{jV{+<Flv2swCKsx1LNHX==nIWr>
zEgo~qEcu3ph*mIYm4oh6BjU+aUIz(cH0Jh-U(px`eyJ=BeYyXSbp#!ES}BW_b^1TJ
zx`ZLTo?rKpcu~YqUvCo?hPmW-Abptu6Pd(<Jx62`PQp-}3n;T!5wBac<nc`cH;9OH
z^|t@gT~$Oi{>>|1W-{jKc-Y#^)niVliwZF|^S|;w6M>zJEy!)!{$bN?OZ{$#ZJ>GJ
z%OR~3S)RLNt#)KM4<!1r@%ps{CS*oXIfq=rNho;NI14k>)77uUNgiyC88_PlSACD@
zeR_)+XA$-C_j)~-Xh8!$Gv`XgO>5tP$D$YQ3q1vBkw9`tpha42ecE;Y0!JvO-i-`@
zMP34L#{1##e%jAl>!~{gszp6h;_#3<O8HsAfol%NZJ1ByUgSB`u9ME~HZsb^x{=l7
zj43?iK5(whq848H<Cv+V9BhQEmKlEmSCNQ_S|l<<o#*;7M)4U)`QKS<pe$1kR_hJG
zYvF;08}}L9il2%kxY;akm8_SqFEmNW4%M5D#hE}pZO8OtCe2B)hZgkK#la&0x!9fX
z`RQN#?|Y0teK^G4^s!z8yI`wLGa(B6v{SQ!NH==HbmarAnwHukz4`O<nZIIszLlTj
z*Cap62D%&sWHMfbp2!osC#1cV_GbfRtOp0{7tl5eeCS^Gw>2S>l_an<%EWGb$dwsF
za@id=@**2(EEOYNZ$%3&XK+&6{M4*ElXkSoP*v+ZDkfE7<6<_^A_N9k7fiJ-2~{GK
zLBuESf6qy4AszCz)=^RCY66IFZNHcWgolX)LWJyvWeTKiKQ*nv7GLVCI7;%N9aVag
z>Tm`~zz7(k;BUWQKzwa^Qk!3i4u_@>1@t$XaFK+2?&OUn{P4{0v^dFw4O6sV|3Y~C
zi<b7s84M0?$!=?!2Kq!Y=YI=}nP=e6A5tL1*&NS25t5n=$^X+?M>YYv?c=y~Gl)s1
z5gRz*w8(9e#H)ZtM$%B<ycId{r#&+_Pb2hrC41GLH~m-Z@lw<lmX0Z{N)|$nAk8@u
zsJ7|z3h}1ljDn)wYWk&<@N}#9;Yx*G&g-5@Ls&i}KBI+Nhbm*hBK*l-{GngpnW^Wz
z#&7WKL?NY2TAf_h1!?t}&tNeaWXPwcw;c24&YE81i$|^NqH1mb?Q7~E3hVLkr`y}u
z`s&Vj4<Rz@Et00)#dMsIRW;Fr)sDA;IqpE_D3@{eH3CqS`WEb43s0^#q2NN!O6{cs
z%hY|!TQ30~#lm`_4Z!`YE6I8%`n6-6g=i6Z#T=Y`GdGbon;WY!1w>;HG(qv>ONOr+
z!A@}a6dTHonfv68m}$Uw6s9hE?Bo$2b|2MGumIWaeSDdF{(><XV@@WLewikU+F}y8
z{cm+8htzX=Rd>a!*oYB-ehWfVdQ2bJm;0;m;j$9hRm&Vz6`?FA4D+s+zV!f=$b?Qx
zTq9}fE3T+~8X0+3@8ZNw%9e59vUk3m>fUF`Xawo=%S`77kB%)H&ADI(MlMG>PF~Mc
zr(+<?zq)X?B`FmDoK_1HWLLA0Jr074{0~TYL{UhKk+X`&h|LsPKjV*O)cAeey31e-
zcr+?>w(6T4Ueukf#rfNoYK0;l2ve>_F$FtDC{83`<Bo^4wuXn|5JNZ%sH(#*`7hV4
zn6G{uKxl^Kp(Vwp^_?3%D5UMg-dbBgE~G5H+dlPxHC@OUxNDP$QV0(C%Z>X7+PR_(
za@1~(jh~uinBQn+nx|B;5y*3I8ztXSFK`B$P%q8#mxus5V1mns(8d1#qaiuXT%rE~
zSL}*9X~O9I02mPY7j6DMHj)q1Ted=fUeqSt4CDPfpvNT;nd}l-SSCVB!%dAny{NuT
zu{C@x4Q+{s(it@J1bnGpzkLB!;Ul6`pSFUPwFn}XBh3@whiou!H}!FLmYps}dt<af
zwGk}c9$9QzgzEv#0z>d&8^}Lkf7aA9FI$q<DyN~M;8VytF+rnl_)FmH8e0*v+Z(83
zbEm$^$w)Yk@AGBqSRg`sxO*J)%pWPk5Zz@scf4ZiFFD9FINoc>el=t-=tMqrRgJef
zc>TgG`7)`G5%ekhATdIC@?zz<nV^6vQDDv11f-3t-bW_UpQl=aqB5IXy|g&bCqrvN
z<AVp@gK<vQ6|v*pY~{DN#mx!_L?G_mZtKKNr^CPI6N3(u#E1+svFWT)QhQHYyJmEW
z2xYZ{+5<~59p28aZl?LLSzS~fh(Uzb*`)vMa_}Rz``q9Xr7jldZru#NzY?D((JwTa
z*VaJ$`UW%g+l2YpYQNsG_WP`Y)TcJ^;g!wHM~fSWXF~0E?xl1Sl@9C{q2VX2%!e*g
z8?&lCbwg<2oY&GanZ3I8Ik@q43-rU3UhQr+T2m^l?{0O2-4x1zm(AjWH#`eHbpy6P
zgI8CfI5!Z0SHV~rDD6UtcNzU1ZzN);`XDQ0)<A!v5F5`Ks>^6YR22t)@!)B?2Cm7+
zx}0eZ@FrJMW&Mj65-BU8k_{g=&mJTok2AqRFKH^d<!p@s=y^@3fXzX}z~*svk+p$T
z@A0f*n!d<p#wX@Up7y!T0H-99FJl;WEZH)_FpSr5feCA=U1(_xh2=M`V$-i?8B5v5
zS-4HX^)h`N$I<aOXZ($!s<ukd4>B^1%8m0*H6Jbr5Knes;>k!v#8Aiz_(HdlHXY{?
z(vb|N6uI-|;w3K}cntmTq=qGO5<A`8_9ug*=d@H-usaPBSnw{<`z7{UdJ*6#Dn(km
ziP3jrwLbHTXLb4yIS}Ejzm2ps-Od13kQCA>ED)fO>EZe*q3b<mC=fW!GzwDGo`X8%
zinY=!z|2)rQVhFM0(B)}83+ZQ7}r&rckf1Me_U>ZG5(%})Q;<aA#=x;(Wm}xJ5V7E
z)%1Z!RNV(oPycc!q&WqyQ^eQ5*@gTu?RlakOAhy|-s-#2$v~HJ$VpUWs^&mb0DUqT
z^%7(aHgD=_^$+5$u^GQ+-72VAn9&$Nnhj9WoQMei@92gy{CwQiSFxAYjO}}MxKr?R
zzOJ%)np?A(^_pp-xVb}<nHv@hd3FB|yBIfFfw4*7okY3PjfTlGnVO{kciC*YK_<$x
z<Ml1Kv##zOX)oFfW&cMPZoz~S)j^1rcB-5TUQm#wJ8Gx1;ccC>;X83vz<&^m1ESJV
zkl>v`irh6t)6x63y9En?@>8nlWK?P1q5b|>N8%m_8J%)|Jckwwi}L8;BlBq8GU-r5
zHW-T=>B#1<SLYaUK0@8Z`=d-`a}AHC*AB@aKo_fIDR7Os9-${X1Rh0tKbY(f9;SJ)
z5>6Pa0$Z&6S7*$ZStZr^vH||E?4P>w$`=BwkB=c_yX*1C4MehN-ikpS*D!p`B6x$s
zJ<isoAK;e1QGvT*jVD@s>6^j~uVq-ZVI0SL*6RhQe$0PlQ3rcpfE9KvptOU(?_ZxJ
zN8xm0@aN-&az2dAh1z-&cWYqZE!^GQUcKklvma*WPGfM3pKcbk5-*6eZa*?aeeEWk
z3H?b(9ART>rHctw`6*Lxqg&zdc<<iog7U%|lsHS#fc|)48Vq=J*S1TxDdYN&?2y}R
zH%AEkoHiB1FoDNt9OCicekbZqClzcln9O`c9;=*}J>+XDTPo$Gy&WB_BbI(&3UFQv
zb*4lT*EE3^B;)I0*UhLp6$)VX9!cEd6bQY3k1MydXrc8{EqAt1T;kt4C>x=q-U<&h
z;a;-29tgo7MiN1<X%gFh=V+!^_;J7zXItlFs2R5Gnfj=|S32kpba*zY#8EUh)TYPl
ze7ugfSmcld*XcxsLrm<>DsJOYi?~nz7cwFO&CYk0!+ro;50qDMw?%x{B|(|aSG6rf
ziLs!q6}9+c4`Q^TE#9!26A=Z-KTEX;`7{sy*rZ;*3;9)9aA7U*w=J}}D3wTgI6+2!
zU@~s_HDc%d>#!0dAW^E+E`j2Ks#(Lj|3*W@;x+>8%8IEb5QCvvnZQ@gIQJ&Ox+P#)
zG(3yhIAhd_+1V)95&2o*!RLd)?Ah!+b*fk`C$y1M((n&cjXcPeGBsZdi49OaMmc}m
z{Z(11W)K|y?wtVJw!$+QO#Vj&4kE5q@1%S#N5)jfQs_$8aB_R){%WgU@uV_bsGfdh
z58N;m|03=(J|5hNb-rHa=b>4%rlk9@q`U7AMiwr?rdf23A7{l)GC*`xQ2%n!Wf7}X
zqaS_#*o87rwRWpy;y#&A(@ys(Ip6a8M1?&h=*BjP&-=iULsmJ$Q(!H<FgMSr4-(Zq
z0&!K7c8H#;r35>OL#QMb$HGpOuWFOnaWM9k%YRuQOBt_-#T*ZaY%_VxguG#OH~*~J
zUe-+%prv-~U(MYeyrz8G;3fB^70>E`nV#q!IoAhGia)i{!Pievymdwwf?UW{|2IID
zq#?EAWo#Hex-A%k8sBP`Zc4V58k{(`&Z_PTdh!l!QiQx$^TTBWw8^313yeF^7`yS=
z^+hJisN6MA-T&myR>GL#6V%k^D9)x4HF?DIrf3+@DRQYU7(t>RCAJzRV&feOn!>D;
z`>M};1cDj&xWYh943TaNPZrH2;VNR@LL-{0$OmT{fYCcJNO{e*^Cy4QM;ls9t$1{U
zGk>62np2x@Ei7a%<yfTBD-`-aP@f=|dS6bV-p&_XL#Gb51E!ERK%zJ>a*Es>A^r21
z`U+PeMjf!egK9c?9%w0dWTh&h$!$SLAVGxxOTdAznv<*RH<2<?Gb!XzMu%=kM=XCr
zsY0S9H~yIHKyOWM`R3WLd#0xx#f2tWqr3&>tmZ<4gHyVnGWd{qQQmjNya{ST4`(&t
z`4$vv0z9+d`OGHv0A(GUH<%oXjb#1g3OACG#-aeL-v246Xro9Ey`PtXdQ!w>-_2JD
ztp@FdAn2yq*NZD{SmdtE?JKdp#W;%B!m*<anLU--IyIIVgdy{az<ge8qv)T5NJ27=
z@cykW6CV$+T6nRmwYai!qDkxW;2eY+K#G@X%nAhtHz*^)y8R*MHsgh>Fx&fN*wFhS
z#r1>p9Kv^;FFMSdwm0CNs3wQ>)m7S6kKA)SP>g^4m0pJ@R}J3vZ~vS;qcraQzXl;h
z)PGM1WVSFhbfqBU?;XV3tRP|C&S>gz8@am`*f$}<QY&ZylNW-ZX_+z3I+FR&DagF+
zY1x~WS50@AmOvg3sIhqV&Dz-`P)64_GdX&_B@pz8vUnGfX`|O2Keg7u9~$@Tf&lqq
zd?#2Ln1*IJ8z>=1T0?pEm6%$ITV<Jl2@<aNkx=I{t==6wzGXWsRVW;V?0<~#P`MT4
zMX<-WKs+~%jTPS#6f|kB*CNzRX-Qrc)TK+!VhvAV!^J(j;j(vn+QDA1)5%dH?73b#
zyv{7FyvLxFhF`t*l<bnRy54fcS}Pe1b(~?LFpklNUW9gA8YB;Vc{WoGl#QS2FlF7f
zq>Cd)W6HNKm73!pEyBrb+KEt9?G@X>AZB{(e=kF<Rmtw=z}N1M@+v<p`cK*UB+l_z
zM%<&jg*Zr4ZW8EWcm+gBG-|r%iIG1hw0;-uPKN9l8=fV;d316K2S1S=e2`v}9-kF8
z*sGyx?Eua|Lnu&Rf`N!2qM!=bk6F?PlkBGq!eb=gUY*e24JSA&mWIeD;&8Pk>`L5<
zW8q;I=$uHg30IbolFADnCcf#o>9vD=4amNOrbb<(p0l{*N*jvb{zK>cTWdc2bYnMW
zu2rz~TPr9<$mivdc`=F(O~EA=hy923kdn*^wpocn(V#g)*gGAjmPLRs8uvM)DYyhr
z@nhbmX8U>aEoWEy!?jS)A=C_XfMp{Kk7^Tj#a_!L?QI$(%r69Pe`I2EEhAizr;SCq
z9mt?)Z0z?ZRal~Uz-KEopUz@LW0_nm&?3coUOK$p1?09{hAaaT2sjKrHudBYtJ^?i
zceMAr*FD>;GtiCW<Sf+bwZwZE1y_I^Li9QyvvbBTpCej51f_gkbSzQ7c8=4~INA*l
z9iG;%c_}LhlU8qd;EkUaX=?Jb!aqA(pd<_+e*eZsMsA*){hz<;=#jGgP&a*aaU-Nd
zAhso6a;aQl(KTdtc1|?Ms5zC_k{DKzsB(Ixc&-p^5v9o4Ohsq#kyzBRw}pdaBvOS;
z$$QVO*kwC{JE-F^4{b(qcs<<ERg+orI|*vZqy2j4-?0m;Iz#c?d7NSo5YEi1_>VpP
z>L(^E7hB?@Mum=ALqBrKcwa1P9RdfMTxm^?b8{_xWxOXFDwjk%l=I9cU)aNQ;!=tn
zv4zwNN|E`*;Gog@o^vhQ1PQJwnpfLeoj#-QS5<G)NT`j?6tWW+xg#H$i1(V!`#+`t
zm`&Op9TInDNc{8v^#XWL&ZrD(f;xQDoE^U;*9(eV{uX4@`(^s9PhN%%6nX5-WO}uF
zGzyzegqdR7^-%Yn39f&M7sO;4J}cWqLyy0bOotY~70dFZC%qwLR^t7Art#dhdijyP
zy6Wu(z?R4kZ#|OXa{3_5%^M|)LgCX2L=UlZ57}u*&Oebc1YNfELo9hH4$KHqG=Z9&
zmG(XuR{yoI^PGZy4j0zRW-~rebI`XGGwS|B)t%<~79|jHFs}e<v2j#{E;^{{<|%A>
zqp|UdT5OQA(HG<W+OQ+{P9~=|f5g}I{$cXnw556>YF*s;F)S|J>+$Q4qE_NPu+xl>
z#ugaf!1Gvlmj4xCcg!`wemam7A-kJ+?(3udQ9MCpiOkh=R$lWwtHmwM2pA}s;ADx_
zE#%y^J5CT;`6{Tf4;kxIp?Lw2*^+9U<;5-%!L&Qfo4stQl*|H9rdCcwx~FnMTK^4y
zx`!%9uC`JcISTcF6Avx&6kGiR$E6XLnu_>s3psISUyK&fA<)uBB=F&=H?ZypUrvY}
z&LOhhD)?gmW<p@lzQ<)EHDsI58F{orA_LF9TtB#cCsaf!&>~~U_4fk^R{#I?IAlJ7
zlZD7bksgSvInn6MtxKhkecG2OC%{4VY@8HU)=&K&APG(Ng-~+LnYh+5lI(#uLx`fn
zw!#UOTzKRm=fAa}Ne-YTG>w5!FCk+YrF90V$4)DWN^iW+h>aWa9mz^KQ{QA3Tfnmg
zFk7SO%QeFXoRN%t@jh;XbMS_Om?%tTgLiop6-`nGkOEemvHaq-_b~SvO1kvbDAFZy
zngc_&XgU!jC0M%g;<1g1Twk@h6`|ftdy(oUV7in-HatGJZPs>ig`5?@!}M&)VFQi>
zMj0no9>!+{E`HiZ+Tsa4jn}SIwGp=&e+}CB`dFX-TJ9CU=dHvD)rFig?8yg3Ok*CU
zdaLe0LDr}z7&KAaZ&Q3_1Ip<rB|kv`5e8cw*_#-FVErvKm-t;y3RI=-CqTOEK+Yzv
zvX!-HLJdxg2&s_*QJuo&>zwI|^Lri=P!nNH`NF>|d5I!l#x?xQs&n<=O0B8MG}s~%
zf5^G7Q&3*O_J3gP(_eymp3&xDCn((hjG2Z>ke8H|Q}POS&G?<2ovUdWLW-Xd3Ueqy
zk|e-tAR>?>qZ!&d+DKbSLlY1Zm^7^?vfE&|-rw!i&x-Zk9;sTu_BDJ#^B)rvu3iFX
zXGc{3h4!mvNVPOI)bbm{R}C?XEJ)V3GfD$NT4cLpb5qy58w~);(41!AjLj*ySS*L;
zIdoMScN8khqL@WE2Q@nCi)f@DF{w$KD_kW6Vm-8J77f#2$j*&Kj7nr`A5^^Nqh$Rk
z@^fJ`?AQMKr-uv-r_s;udcvt-IfivMbNl%+a_KWEr7@V}B;}e3m?B~?f0lN8j6mNR
z*SC|^yp{SkgmflgND!+b4{?F6NcU>)<hBTl#Y2LLF3lP<t&u9OX+&zY{|}Vc7A8C`
zb5+k=!=k?+;nPI)|Jqnpd6}jLL2j@hQ6O@9mNl5{=-(j5x@HTN-G<`;-|h#<Yl9Z{
z<smQ4`!<Dxq`v2TZvY<q#6V35{uG`?gkk>tigt4rKdzHfwM}GPwk&&|B{;PUi~@vH
z@RK#08gHtRBfV?xCq=Y)Zh}fbJ=f^CF+qFfTBdCj2><yu<wUXbxPQg!4=(yLn_wbJ
z3HH380T@;R+-YQ%!4I<e_LNl-Xc;yl^u2CgCnn`@Sv~8w+5=}u1F){0{7HWjU%q)k
zYn1K0oICX|_uYCZ{u(QE<A}dtb=vdc72Fw!ZeFB|Lltu_6>t6?@LI}Tv2S$<y>hN1
z3C9DM4mGhH^?7a1?3uy67um)dv5dWb#VejlHbOm!^c?iGkjdiLU1<wn$Xp1-(7l)q
zfd>a57)L0*C{89bCYte>9%7;1HEEF39ExRg5#N<0n_gQeg(hI1g;B#<pCT3i;9Pmv
zi>%zq*!8gHCaz}^B+7x*=@p7QXm4s8ouTfcpcp1Ry*5?C)0m(AjZX9zr{O{!ykJRl
zdT?*_TZ&%XV~s8vG!L2n{&5VlYE;vN^Lk}~)W4wfzqd;(IYpF>oBs^b;w9Tu%OBx#
zNcPQ&{TqyPJVahHRM-FkfJBv>#2WACAMxXyD>q!~rWPVJDx>4@QX`Ci-XZ1{Wo4tg
zAB&tf=h~0`t}C7MD~&R5R0)vV-KHp?;|T8kClwLc2omFYkZ|q_GN2#ctE#CQY3qA;
zSXfxR{r3C!@1T!6%I4<gqJ~mv#N!!|d81lgR$gaqaY+djitP3)$hqfnN8$5^I+a8;
zSYOnS6J&<;eL7zG?RlK`_GX4eT;GrJ<F$`eBE;)0<1~{lp4;;-QL!a6cyKtw>}roH
zlN?*M$mw(F*ESFF@ZW4SMTLldSC@gE#~{#XP`adtH^3OVRzErP2alv~{fWvok$8?M
zQVzxL9e)bDAZ>78*`y96y8}TXEcr@AhTb301SwUNFu}7>kmV_#WyxN95;!N?NNNY7
zDo5Ql)9r3%eohptu_$VAef&3%pCDGii*-08%97I+IYt5x=!x80xJ7)j0U|QLE&h!1
zT5l*M!<N6g|3*_>mUV*kHxr*Jiq7c|vY!u+d=T$e^xs4THMGu|Xt+GyIA3LF3;F_w
zJ0GD#&|u0~yI7BHSYjrp9)k&wwp23L8AEC}^zj85ZF^^kM@Uq?EvUfb%s`>?(~c_x
zXj=_wuMBE%pm_@?J7xjX+JbHRNnmCPPnHls3r@aIeEzNDwNV1gigxcgVdCcY*7t;|
zCCzk~w2nn4L?c4W3u^bc4|;wZ9%vc)c4$S$M>}or1`5gtirep3pVDkt10Kr;q1r}%
z$a`QGxXSMz(=(gFf>$p>=BIy{$R--e#yJe8w?ZF4MR%)$fOJ^?&)Z<wn;yU?DiyWV
z$99kRBEyEqa^BN^0m*2~vQcDpU1!WRsGm^pT2hYJr;u}to>FuWk<KXUn9i~pq}Ene
z`-sr`(A?5YqRldS1T6RgXGgQOrvE)ALRkzWu>fbA_ykM1@=ZKw&z{v)!=S=ExPt@}
z5_%%Y=4?#SjU1r<$HIHMj7%6sV_A7W_}q%8pg*_@C9Kc`@a2M_3^*~iV%|>7q09pf
zs&9#iTZu{fsfJ6Y@6=DUTga1uCfIgz3#6wSy?4@OiB}&h;Zk{@XI?sh8(yniLU(&C
z{p>fSWz+~yH*%1H_e<g%ZqM#~SP0(#lx;~~NI@QrZeJ|niL^fmFBBXjg*@Pgo+jNq
z6EO{3y*j#sY{eE8KZKr^5_SP^h;ZFcPO5F6m$?qx&r%>HeZ1y05HLaM%+%4uXoC#W
z(Yk1i!seX2F)q*#dR~?9Hmaotqdlk0gNC>+!8WbDWYc3b#TOv{2HGmjrE`#bUwM8Y
z!Q+-6Yi;*kOa)5l;ORjV>;B%xCTaq0tx94_Wewfv_U8vtE{~vN&-2k087-&X5|jV)
z#N&Pa(7IFKcN1%~WdR~qlwA8W+iD%5OAE(e9fXDBmWF39{ST|UJ9H)4p=r*fiI$m(
z;c{$I!*VuD`ky*r&u)z6<m}Va=g#>9kl%4`vMhY)Ch>fkr-8Y)^mTa|-=*5}>dqBI
zOn2^9b)8qu-_8r#c1=788T27<!acZN20`-qyXAG)qgN>o#cn1MV#~--izGljaWQF)
z<EZrU_}Ag4CX+K9DDe>LKT0YSBQQ^c0w5NT&XdTFU|1NG9k~7|?kYh<BWxg0*Mtfo
z-%|Z1iaNsfvx3CRtZx>%czdRC8+BhKKi6vILUzfS@UA13_+NXu+h!ZO*W-SW^A&`p
zQ94DL9qd3Sie|#X9eTI!P7>@ySZTa{$=-9^2MRpD*UHJ|f|Pa%trS=*Y2EQU#hM|S
z&^THuADk!dq!?!&^}k)fu&dj0871!4Izk29Kw1k6tif`o;(P(2h+hm0^>0OoW79cv
z#}6QTQ3)Jr`;B&VWaI~T@_3Mv$y(<`(h$EziQkN2VG8u>90n2O^ii0NiLVyK0r~Wi
zIhWxzN}(>=j%r2xu2?!2%|7p=F~}gw$^#(SZV%Dh*oyaPpv7f-uH??xe!d6MTbr*q
zAyKs)&C24wJLMqjQq;hiJkWCPAfe5c1Z@U*9_dwE@3pq+ga<4a-e`nm149!B51Zwm
zUOe2`me(T$(@^W>1i)!uSQ1EK2ZzoDl_@F+28}?r>WR_c)t8OLs_f3|?T%6GUk@--
z^XMcM6gUkWd1DY_y^F7kuo;uQWu}rkUwpgo?k245ZO?c?04<xA2^r+sD+1RwH3F~%
z&U=9zOUk=E#gktGr`eDuL6tIXLl5jJc*623eF*Q1JW4hu#HOYc?<(R9hd1sTj6LOq
zGn5GAu$Hm!O_{l{^yHX{_Go+a;LKojwQlXD4Z{RO>ndEu-DSj{IdRF|RQYC+XdJC&
zR=y3Vg#)sVq_s9y`1a((rTwjEj2EKXQTbSM(G8Wjpw}`iju+(g1`QQTF7$TudB@DF
z#91Zm4cMo#-|R#dDL@5Tfb|bRBi7rmx>C_TD65P&GF+^ag7`ih913pg(wRo}`8m*G
zLxV*j6_@pUtxqUi+)icTU;s%l-3=>i>#3gH*0Ui~@GHzr>3ryGDM~(%BlttAP>kOJ
zv6$XC+_AS%m1KR3h|h7%zu~zkTX%-%1JGlBTH)qP<9ko5CSHx8IvR%|zZxHgwpyr3
zG>(d#o%gudu#|xU`n7(%>BOu|$uIG@Z*Q=fyXz3=t6KJQAGA?pQUG!xt%-;EMmNzD
zc6%MTh!+@g#%zepB<{DhVeu)c<ZG&{+urd{8OERTPcfe&P?MZ<s|XazJkN#rI2A!C
zpuOHL6g!4KkQIFc5BwYzl?V$e8Il_E*#MdE0VnEu5#Fu`+izdaZzqgD5Gv`%Hw6t1
zHN#~|e4s{u{d(z<7FqoIZKC^=5!D@5VDJZ8zV8{m(Ea1_(~MPMdcf7a*H*O0`O-sG
zj>Jg92nseSaiy%J9ax>cNJQ!U{O%V5KwGX+#&=lEm4m6mGRloKtuZZc^pb-MIR{KD
z46^m<=hl(rsvK_*ef+Hzy^-SlPv=W}teTv_-3v1EeJK}Zjy?M%G1~KrwZLu=kE33A
zz*3?i2ZAUPudnwr`_H(>Lxx@(!PT{$995D(yet@fcZYFz-B)}%rCX6KN17?#_Kp+G
z7O69G1N%%1J3G@wRpgbzl}^{8_dG|*xtWRi{Go!yc52q^eeNx}xF1mz-##z$V7(0%
z^ZstHmnjsp=uR<Q_3+w%+Yk9xMV=Uzd=YVx>~AXR2|r%OOpn&f8jf6oA!pJQOqGb%
z3H%Dl1M|4$pxdJ`KjE<LHZ9j^?xCA3JNNl0bnneH6?Y}<#U*gu{rUOY$n6vmAk6js
z;v_P;F#^pquV#3)Gqd#Q@d6;Nr3?86$b>mRW0j;g@RyIb^9-b@HR3|Y^wMf<bb>Ms
zRY0QT21fYRW`%?_2V*bD=)Q>y&gd)H1^&PZ^K})Biy|#rgzj)ghg$C+u9%fkjv=<P
zeTS0*MQPaO_-w=P!d3*k&5ZSo6eHAv-rB5~kbQplLLZr(YrbRG0Z2_EP*FSw_?kgk
zi=}8YMk;ZT3T8)>Bc1fCz4GfH`Cz%#w??VIbv5RL42mg^%7*RBTZF7Fmm!3kz^;lC
zLdf7k$X=k}^NOQ1tF?a=7J^~YDoM%s!Wh?cb1m#IRBCu1;3n*lvUUf7B!J!?B~e@Q
z+;798y`Od?cOQHOS7WPY;XaNNK4aUqd|-`1e$9Nu$F#>a^YgZ`nj-<n>)RtW_?(4p
z)x8$tZ_>lVRIc%=Iwx;}uoG$%IM9tz7hJr?OmR}`x#_KX%VG<BPKI`FlGFM8wsqt)
zl%~fvnkED=7?)pWR8&J%8RI3Ksl>l1Mw%!TBAABYL6seKnznsm3$TX-U-7F-N6do~
zM}uO=xP5-3qvu~w=H2uvQ_Pw(b<oi=3ii9|fc4>D1zEuX7LS*+1RM~0-%43cX9u9i
z#>c~Fu1AvwBO4xi9_tUMcwHtO+IAPUjfO8)o7VE@8AKRdm)A)fnSe!f#3bvO#?QaB
zoo~j_eLqn+UuU@YVP`HzR==R$$4ZxBNu{e!B+HpB6hKmB@Wda?+gSmdG{V5aO$QOm
zk=`acD|L7U{rc-F8^()mS{)X3^wuPhjfE-JKf!*Q^C~KZp2ySvP}E`9Dd*T5SMuCb
z{7xESS;)`VE>TWfvTF?J47d1d@+F?|F~C-Y=DQ#(6il(e$2KD^+fVgj<RXy6946--
zM({63Bh~qaK*tYiQrA`$^<zYcY;Z~Rjyxv4++G%#Na>*xdJR=vghAq;gFqS@ch9$1
z^a*wDde<|)73O9oB{V2;B%~Kaj+pk{2*NHboaGu_Ncr>|($7HDz=v~VFEoLajcA>|
zz<`X<ew>1Yw>{jqk15`OMvVtMJB!^HZo7K-)y4*qx8exn9$zf|1kIu{dc0~@EGRCR
zqNg2e;&c_I>q9Q33M-KityWyu4+uJb4e*K57KIjFs;K1tVP)<r@}j>}D0U^qXH3Q|
z(hK3M=ON<EqH#A9+o2{N5@Xp=uh?Fa?hl?l@F6G2t>kG1M?)PjqBCc!9-a9kaiAlO
z*8s8sS#Lj|Fpij5eB|#XL-2lo;7!cPng84JdiBcsRZDw64c9nq(DD4m$I$=s@4Dl0
z+h2o#Ma`!S!M-u#a3jBOh=HGx+xz9B@49X0%ZR2q7eBW#!?zIM18)XNi#j_IB)TZE
zsNtukl%AH_dN<GSqd(^HRBsRTW!2R8gRRrrGcze*I*={PR^f)TVhV*3uM~t@TamZ>
zDRUBvSSX4legO|yroOzQG~L>Nwzf3aoW@zhscMR;lVw6c>9u*77$k>U8FO?4PYQcI
z4c+j{G<W;n?HvK8P^mTu)(AN(7kNRLMVQ)?S+U%VINyBkb-Uea$_A~F{6r=)-;X^s
zU?1>cC1i@Q3yit5WZ&(#^n<@RB?1sKmV8*xJaCmJBnO|mW_2`>S1Tl5w&4CG*qUbh
z)|f<SGMvz8P5%O>-QWE-Twn&>DDgiq|KdY_C++ccR_sBld2i+gW0%+UK&`M}gZZq4
zV&+EAp+%}veY6^|b>HLZeQBRPKm%)yW*1#)^c}GM{kXR`U;Dr#mL%Nw0J-ATIEdjB
zf22y#)bI~}ew%Zp38qE4VT7m|8Q$XBP~R*{Q@)zMz3-~o#Fj#D)$CV<;>tK<#_?=0
z@v~CMRI(@$&Mq?Sb(JOHZ{!8<LAWq(kZde8B`A!^kW3zy7uE-L!9Kv2=AAtKnOE6W
zQ4tQWdw&JEGjwX|2{5kq{cNBQAYaXk(s$~CcItWO$qzWvecOQ*4I3stZ*=X<ecN{C
z+c;#jvBd(aMr~)LZ2@R7wYow)x&p6zujjjb;rI;q2U|5f*r<~b6)6-`PaWUL+%c1P
zq+^!jbBY}&dxl1B^19%i`i^7_sroPFkl*)e;xr8-x3)T4EoNlhOT@BW`G@{M;C;}@
zMenn`tjJ>Oj!(BX!H0Q370b-L;=LD+VA{G#D!A(HberTkG)3u%@Duk5B?WWN@mYW8
z*Gb6?$KUVqew0H_Naoh7tAyuc#yKcQnOyT`llRcQ4fkjW=p%4KYk8IIhAXML-4iU;
z$m39tb=9q7z8d3XvoWLoo58i~Elmk(2g}q?j5(ZdO1vl$2CW~Q+0Jj4bv)DK=>#<m
zjnJ(5vWL?!NFbS!_lHm5=gWDm>~wn#38}H3QM@%@T=&~sAhTTb)7~tz?31~EcVS#S
ze11zWTYlg*R&Twh`r~L#H5&seSOf~CD(bB8RdP)34tEp;1~CuRfH+qqw^Z>rsgng{
z>&uIhyw-dG>-Qtq%!|FIUtfl0@L8eQqb8ZBgWmqSR0yEOpRE|54ewo`ErdD^+<bKv
z^nOkIl$TTnoy;L&MponS<AQ>1)^okTUM#BtfsP$s+kscDcafrRB!Jh~rejiAa3v+B
zw-4jbJMHfKzw6!lXG77L0WbVMK0af=eu?==TjV76cM2dPB0jx?e2jai9Z1IqZhVF4
z7@5PR(|6rb44mO)2R1#wW8lh?l}I30Ce|Fc)rXXYr06lEsO=!g>kxwMBH3aKCAEv$
zxWGS7hT-!B7!4e(<jfsr7ly*zS28{kO>EhaF+(2aI_ZW9SD8E`JrEEmAd=v^fpA3W
z!05e72ZZL99NHBoXC1*9w@E{9EpY)dL^fzhK@}~D#iPAa!_v1*Bn>H<5adR(qQ=>T
z>=CjrwgJQ!PORJ=^wGbkzBpF<HIDFB$nH1CtgzAIR+m{(*5KDiK@e%hbW?zKmxTmF
zK$bqq8Y{wd@6V}*_feMBO_wa`VOr3ogPk8Jpb-Gx&pG8>x*EJ)mV-Ye6tLIV80#)>
z9!@pNcgjVd=|L>f`Z3UA)9M8}&$H%?mxyFT2)f7$C5_tygMWr{M^0CAhLsEs&0Zdj
z38X;UNY>8D`W_YAzi>fi-tHjhsgs!Py+sh~Rgk|0(^Lsz2J~Oi4!lh!IWrlkv7^+R
zs9J%l>0u(dsJ;sE*g^)0wu8^Qjzimq;3-W`x((VSnc~6NH%7?-@O&3tq_zUEKxTUO
z^!3Ki$iIfELEpDWLPJzmMrFKU+FJWcAWRZ?9;3<?;&yoW?rT`L3|p`~+Qb*``8RRC
zd}`;DM93LzQB#wSh2`p|`>e_~RqLzCp+yEcHMa+qK?Hl`e63gm<8Sw-&=%_R3-hAZ
zb;r<ki|?_fJ~yqyzZFAeBjen>940oji<pWrF9ihBJnaOs^6Jto6Aal{7KfXZbg@cU
zO9pXgXWdQ%w<~Q>MhOIzAX#$eO5N0Li?A_AI$WxLK^BO{gi<Sc|A}S1*@(2dJpv-s
z6r>>nqz-u++Ul~ya#Mlmrm(R-!#3@b1yRjXf(VTfGC?i$6n*F*oHJx1E^$nEFnGD%
za2LZUkh>D%<92&T+1DQejp!?E%JV(&{yVcOGI4cXJA~Vjy93z1yFk*3I|VB%E`Ntl
zMQxRot}ZkRsc?jfK>&9zU_0PxiMMaH9K82_8Zf*i+(q7usl9Tbrm5LGMja6;)dgWJ
zb#B>UP61qsO^ezKyKhRk9kp$-A}c>Ix?>z=fCAigNE-&KFs^?;x$1KP=u?xI9&U;X
zq{rAR)wCewF%>tpG03ZhWTuWff#8LR@g^Und^8GH+4>Jej1UVBDp?97Cp627crOjf
zLL2;lIPS&g#7cC(trleTy}ei4)%WY3hZOL3D>5^wyL$l&-{F=8sOQTyr#l>}c{vl(
zlaH+qgX1aX0&saXVo!NL^`GB8&X*}UIZ+U}qm{AB1RLxn1hz!=FzvGJB2B>NT{kn6
zyj&pOOxT&dZo<fyW^G#en8u61awOb_dlNJE4sMx|08*i%eBLe!i%ebWQ5Fgy-oAb;
zPqdT+!{IH!xvDW*)O|e(VUjgJTCgs6!0FP9_{|jgnxs~o4P*Fi&)wXpgii#IiB9>j
zEjC;iW*se}P{&6%<aZ>J`9Ii=N5`cQ1F2jf@|}0vAMy+7o!Huar3#=k<r0OGpBuYc
zt2T@(rlPh|W_%SRSHuHGMd+3a0~3?oMjdsjxCT>k@g`^c4b=JQfB2={U|Kn$LFebw
z#bw*R-Q!m5+s3`w(sW+eP1k5(B<<&u?3Qq+`K_6OOYq4WcFlzJVlxWxLh;nxm@*>4
z?&x^`^ijF7f0+ev^y0Js>dgdk5Vt2HbTvmobKJyhTyx1u08<3s&mmwB;MzP%l>NmJ
z10T}Pbu#HO3H8yiPt{8D6Fj!Hmt#GFaZzY;j4^1_#!65(NU^W{-$7lGnK!%FUV0dp
zlG7Ii-u?a|xzg%D$kKJeQBqnyXx1XW%moz+-u*VG@HET{0Fpi~cJ(1dzIW^No9Q_A
zKGU$cZ+UG4OgGwH=olFG>OOb>Iy{Zf1jby{Fx7OJA$yDt)7!vnZ8Y0M#B2Mi?IThe
zB8Y#p<6Di-ZKm?~eJ*)AL5YPa<Vgny(ulQ+j~BS9g*SJcDsTAfqKBAbt4$p|CrDjW
z37CuFkY@(5;z^SUc}rORq+Re71#Wqe^G!f9Pb!LcI)fqRr8!^vkdpdw{HW<fY;rYF
zr3yBl>9ZFMH|`C@@|6l7pX`!GFECIKVr_tsf4!+i%$6j`_D9z-M|X^b$hh1Ce30@2
z;=T)BNi_*$rVmil^}fWaCO5(EJHvy{LlI_kv1hbtnnBH7s^-tX8EK5#W*{x;ynJqp
zJM-=a3&cKS*DIpGS0LAOz^C8a)nmf9&)@*>&v#VcT-)pQF5Jij;?dM-%m&BJP8QNZ
zrr6Fi6vK9B@S~{$)V`U7zW06p?vEE7CgX-=(O%lXzrO*J-`y`Zu&ATGqFD?(-K7jB
zX8VcH%gkn0eTf(!#jGn+xjir=%JxSrq}%OXUn&HQ>Ah%}VnvaGSxEc=7-aGJH-1xG
zKnhbGa>mzO7rInGakPw>J_o^2(mv-P?ZBMmF+?Au>iQjKQ{gMUZn(BuJeAs(e<>O;
z_Q5Q?HhZmOr=_(O$GGFwkwjM~6O>83`Fr;m5%GB^jy6;A<vzXe_Qt1Gvt0k#P|9%}
ziWC$X&;WdJYSrifV_nV(&u(U?4eRx5o@Si<L+dP}pDY4h7-+0$i~eRtu&x_N+Q6ZW
zDh<9Z5+)hqk#6ShZUQ8q4qw5=LkybXeqHwM*2+Sbl1&ANXkrSA>LU)o`I3!+jMx~-
zkdEpO|6eZvs*t`X*Ff1~b$J7l9QHU0V={)oIzbbK6lNYlm|y7{=ANCse0mo)q_8-0
z*O%`0U#+ACO^$~=o!_ZZHg0Z^Hd<T|MEWc#K%j_=#OIPtR7Wl8nZ0TMc|Fp~yX>>g
zDSO$X?L-wM5tASI2t@)oCA&`@7Ol&Q6XaD4E$s_+0}rm!kaNv)F2+f|Mnk)8Bzc|v
z`1=p+vY=DB8=0*%o4%)M@1Z6L!$_YNj3wvIcQIQ7pKjlH0J{>P+re!-=HOHrav5*j
z`gLI1xV}(<yoHxatzU^ZAQ>@FHl8SWbYI+8Z>AM@{HpfSFta6`E?JDzpuP&|WS<=i
zcscB1n_8I!94;s_F)5Fhl}Z<1`5#_nz3~v3^2k;#t<8?t*vTeDDWFkC!vL^muIOy&
zNs^||vyveep-BAh(UxYW#mWSG8-4!4%RO~AhhJO<eh-T}4QBl2mHPtDbE))CACG+r
zn_XVTpp>e`qh+JWb92qZ;7|8awq3pHV45L`hm4E>$;)-Yx0hy{C+}lrP%K1%hi7q2
z@>?AF{x3y6@g3zauMVKIgBzp?uU8cSFDHT^_zP6efKtrG!G&d|F&C>}TP=gj`3w9#
zfW|6ua)I(N1RLf5hZc^9w<H9*p80*F@G~xH8nTV$q;aBc1al<*+ajOF`SE8#F6j7c
zi~=`DQGQh&k!_E;>6$wl`jIP&1$Zp;+iwF<ShQgvocrmak|<yk&Ea?J)@S-Hv~NCU
zzihcMG$ZC&JR->t#bbJ`c$9%p(3B!iXtMc`)W8E-IF)d4bUUq+QqU_HCBM3h&-S3r
z-;>iOjmiH@zlo{bN5Bo@y5Ikj3!_=G=H%slPoCqrSncW3|G|ldv{yGH1X(^6LC9*>
zaO^xpb6-UeZiCk&(kqB0nE>xNRL?FlHO$HU>&zb{aQ<|Gt-)4T_=up(xdNr3{k+Gr
z1t;zCl`C@hRyOgZzL0Mo+9lOXMH-oSEAiJaPVy5PUsu;T!fA@T-2yVIlAAc?^94qk
z@8PWC2Db63JcFb$o<3L?Zez!Y`(M)v!j%8LWnlG|mwGii9{jlctXKkWO-(+>7Hn8D
z0g}QY!C+XrGOo4+>d6~DD5QRqJ&#)x1;(1X;MJS)7d0+uvabUssm`6hLg$Xya2OBk
zPs#Cov5JJDU9aBS`Ne1RkQ18VJgVYi*0889m`AHDDnlJNesHHt2sUA&N%tRFCsmyU
zY~<HUO@~^LyuB<!u4v9dQ4%!8dhnS~ZfZn0ZSvvnC;R0m$m+1WNbO6s4l<C>u6Af}
zVa?Vbnoui_OND1epte*-Wph+bQQ?I|l_kblwH3(7Et<yU)e*$HRXswhZT5$)m|_Sh
zVT!aGp__TC$}1zHr1?3Q&4v0JLn&8>sW4v~@A{FMndGn%wxs8%**}_Ud{AJmSSJMk
zd*}t^PKYF*0z?8Ewe%l%j}k;b@Y*2EI;d+XJjqTXqJ{k)Z>%gWEr5>YQ|n1#>*V{V
z0OP^Io174aJB?SZV1kb)`U=5Ml%&=E>2;{}<R94AuY-cH3^zT;f5kDN4EO~JAB_YF
zr^Ux|qpUJ@-kk``is&oIi55*xft^bgf`|_{mi5Mg^!%`2W+|em+%bj<3m?^M1|6y@
z@*_e2kBm>z-M(OA6`Ki=f_;aA_4jRW&y3&(9V&+*@8^Sff|Rq^uYv8ms}To{Vm5*q
z@#*LUN9BlILF8*J05)J#EvAtCn_+mwVa<s<TIc>6n>*PGl173An-pAY?Pfz%eEs9^
z8RLzFo?`+Z+|u7>xP);8E1t~+34K0DpFV@RBz2lT8fX!Y5@F!d>E2dH;npy}PY?&O
z6mwPQZ!&vfOvi5Uaquo^vkCl7vE%cJGN8~Kx%_=<B1{KJ9qA{zl&ENRZa;o}5qfli
zhd{^1oWp!!b|<O3E1Xu8kmQ119EV@Y<V^?|K%IgIT`jb4B#FQIec@!kr6gbBgrkby
zvBi@J0<WDqcSjs&UXvn8cT;*)0**L^<CVTHnq<SYA#{SA590TLr0-DxkIQWC4U8x}
zUA~_bxmtvCm|eKTtWuz6Gs{$F5#dCkMHi1=lpJV1xo{Q*>%X!ch6mQPtCc&Q!_WZ&
z3WNHiTD(D(chrT}2?=q8A$Z7&GYTPt41-8g0-E-OWgw2)#H=;;*sD<q`QN1iVFJpB
zEjDFMKkvaT<EDOL2lk7J5y%JxxxB737W|Nq3QpKXJx)=rMYd;ur^3y2+A4oOttBuu
ztlegM9pP_8WMAr8dO+@e6DCmEWToHB9j8U=ADoe=7hwfFoSRU-NT%PsjUGEMkKp?g
zZxW;E4k0TbsvRQ^TBIe;4S&Pd1r7g09wLKRB&yo+t<}cEVKy?E8JQw{es&O`l6f)7
zSr&aWm4To!r$t_0sh?gl(b=~{#}qK|L`=@p@|I9z*@)nR1XiF#mR>ig(|#c&?+*hf
zu2hE48c(HE_!s4vsodi+zD(y2f9Kr_l24)8$0Xz_tgnw!+1c{mSL(j|HNoe3M<qH-
z`hF9+d1T=j@MxH-`L&;dv3>vE>kl6rYu_IV$F57yjr$fqHm55C$Jj2XZKLj|xe4$4
zsWy}TAzI$epz7KV&l=-w_R!g{(Bht8jS%d@y>~*GL`yuJqyVY#ktRlck&i!cib~=5
zf5qzg$X6nr#L;<IFubXocYlWOD1smGtuc{Bp)30qVG|x7?MMF{MN^;RS+@sA5V80~
z`%G;oY&8xb{g@ng-n`{9bo_7Uq7W<oN;y}p!D>Xc95Lr4<1?bg(k3-QM`{N32Q*4m
zWP=Cn$csGPcyp8ZozsP=Vd_S=X8s;;661XSenL!>%n8iWC6#sWnHX()?)Y(9;fV4;
z=gwib4w4_N8!b4kk81b)@b-7_h6$naNGx2<m+G8|5L{A{m6inSQsFIRLEAO6W5kgj
zqCGk=)w{a!I3tKakL|Gm){c$MDDG&rhM0h?;upLr=oGKJbbK1u1_8#pC^jLRAtfS~
zg&6-(yV#FDu3{){2?!eH%rLL5Ktk5O7u(2`MtTI-#>Sh~nSgk!WcN|dnGa-mWm+wx
zK>j;dH(w&WwsmCP^xJ*XPrS?H%NROxZfg908K|i54N%0Q6s;H>GkzU_*z-n)q{3fv
zaq8ls936ijiw%$qU2<%}_x)s-5`7YmE5vP`=!1KGsT*#jk3i?#i^RVL!ugQvU7*=J
z!fDZxq0GoOdo0NWQo@;8KMi!sTi=A=+@1>L{(O`~gBpv83!#GlD+f8#$OAdW_3{sK
zilZ5-svS!zNd`Ia6`JG{R>B8(kq|iAH%@+o_m(XuiT#UD*XcOGWp-98O>>;^;oQm-
zL`Vt>V_0m|qWBpq6A_xAsn%sP+WDz@3>aBHV|HQ0WhRFH99V|JL8$7cRy`>yJ*M_R
zHS!qF+_!Y+SB$qlkW6Idn$BeECdICTUHM*m+MwOfsaDyyq?L9M%iX1Xk3qx^p2^>7
z17j-;41~`+t|BBu6hKmwiOA{X2W13NKO)1X)g&A}c|L9qQdWB0Xyb{6LtaK?phCsX
zEi5S*84)M6Yloh?yGc|$2O}6=RxK3R0l)zxSc~Tv?3Knt4<bHiFimdm8=r2uFUk^s
z+g&mRED+LJ^m&$8$D^pKvMOv5C<YxOH|Gk%Q^EMZQS=Q{PKwCbH-y1Ux%d9I7IYm3
z*Z3|`zhW8@5QGDfx|<b6n2#Wp&;wcQFB%u&_B-bnc?So1$)V(AE||p}!`Le4zN;dL
zk|*76H|2@f@^(3&*$ShN9aC~KF*79zYjPA*C68+MS!7fb#SQl%kNWL(UZWOH5O!6)
zuQgi4NF0*32M!8*QL{>4d0i@MqDNohJL4@hMqJ2gj9kV9m?WT-XM^meym^V(>V0CX
zat!FcNtAQMWqmIOZ<~hzQn=~U8mkV!MdN#aT4NK0su}DUA!^Jwxcc${rcHJl$Dg*!
zk9OC>p}pO%byG96O#*v~8EL3~cpp7`XLW>+w?Dq{KPAw%k{1#<>tfYpX9-^v&s?NI
z?&`>vo28M%L%6Evp`l@}cKlw&14~>wUN+e3#-F$6NsZ#q%;dXKwD@=RPZ&W#jY8+~
zpwq+d*@4W!ZHhY}5NG1ZpwS$@54||D$BOW_dI~lvWpAkFmcHH5g}luO|8xw(q-g>!
zlIv-Zf$QF1^Jiu%BG5K8H4~I;5lxH}!YpW9Q{Dmhe=9uRDI&4B!eRU=a}yi1$sFAd
z541*1EiDN3xT47)o+GWwmRJdFCIFS=C#h(UuuD~45ycJ6Xw0dbiN*Mcf_WTK>@nTU
za*P$bDFsX;dZI}Dk$vCg$yO%)hr#?f!ot4NVQR-`nW6o3XI;R&9=Op0{%J+Xs}z$E
zNFAV7rKyV#EvEXnMZjUq1?L+BOc5b9MDgSXMWHyxAu?5xkSrsjj0VbQ>>QW2Z@jr7
zF;Vpd<N>Ea#tL>;0d^^))=A4?rfSMr@}3J^k*cT2AlG_e2G3-FC}2C(*0|llVUmNL
z){=Jr(|{Wj$L2<B@*?dBB=Qj08@6*k1FR50d3PuN`U<H^!{p?bJH^n3B}BmHj*0w1
z*O+oS$RYES?MD+k)}dCDqP>M-o+6UxI&;N+B)Yohf)8bYFSCl=AZ+9>m(V*BvZf6c
zd?f)!ssd4>atTm;h%=Qtp@J*=-eYy<{A2)$lEy`Wrq=PBpCC~ReOwD#lN>qUxj8f-
z(PEDSu1y#PcouY`2+1VL%2Tm8iWc|`eqxkvXM0H}*2qH5o_+`v5PIMxg)k?qX{x(g
zp5aqp3WLqxMlDww*(*YtoHT}uRaB=HH~2!O7BuWA{&f)zPPVWUu~4y08sJq?E-H1<
zvn^3449|(Nlio+9l{kR#f0+6Tt~k12TRejg?(VL^ox$BLxJz*N!Ciw(f&>p9+}+*X
zA-IG<g5BYJ>%Mh=L7(nZT~)jGt}02iZg|^&0xKCoBCK3(2(dHkqLG2W4Nu}}>fD?d
z{AZQ0O3EcgZ$uWb3|j96JiOmscu{yWn&w{G=661|3n@ztprhU(OTs4o)&ojkX}wA7
zv$F2-Co(VV9HRyjq92i2zF|~H+*`>4s=B;otLnw$^+f!3P^Z!8CMS%bRV2&X!(eW{
zyr?3?CWsuU*0lgj4b@|dJzYg^_;-heCQ0?dojmO|Vgz{FPy_>u%kQh?XHhluO6J}R
zO?K8%gwmgE1{FLfbYA_}%hzYc%U>_Z-D%gpq)*9RUm6DKUhMviK;S|11;JBCe@)%k
z<4(=k6Z^pihr6wJhI;?6;UV9?kE}UA+|!pxd&Qmg!539<tFnG?EI0v$F8cmGT+N6}
z(1`;-6SFvljl$e1L8eOcGq&*))KmBL_Fd()K47^2FkCJfgr^C*tR+=B7<>!atqCW8
zs3%Xe5JRQ-cUkiddm{Ygl3%b{7+i~>v53KqVk5oGPfgrpTU{9}<SPXe5=A#>h^xtM
zdoA-3PQtW}XY(6r(1!xNt`A(NuuKeBr+4a3jNx4ua^wf00T(*nY9lsOEq?>co{Zq*
z=*b!Krj&zW`gM@^cbLSup@u!sp*mkaQ@sC5e5LgaVQ+!IfEC%H9pj2A7T9uHvN--i
zPjfbmkte6^@kh0LvfvpLm_h_8WU*y#D}cZ^su>ilA0OL(NR(dw5VtrhVt+>cl$mmS
zJot$;3avI_>Nn_v3C}ItGVh1K-KF-=kcmh@rKphLjX03S_uoHO{?bifh1()IuUq<C
z{e^SLT>hY-t-az9J$NWv?Jjn;LXdUP*r@lbKE%Q4n^RtDXtZcmf7}VMD!5SA|1C@{
zH!W(D1{J7?m2Z>~>rwc6^TT|_w-NWRJj^R~l~jQyDpBF%oeOj_iS<qRN34jKjrZ5A
zgGGe!^NlXw8VMW@Z8SiR#_M>L?QaO3?s&zytw~5ngO<vr)X~msX0ZxQ%D)Xm7)P}!
z1;2ugg4(XM8w*i3H4;YhyZgINS;bCr9;@tp4J^w#X>y9khtn=(9>cGCP~a(Ot@~W=
zB!Qd#Uy&t%=Fm+X8GDW2>Nx<Ci@=0Gs$uLA9HB}8;UF?amX$kenyYI_``Nv9)PBi*
z`<?JsG_g10T0t|k#%KQJ{2^#nm5#)j1kt{(XJqpi7SOwoaDM%L$)Ncp0W4mLiAXC)
zP-H?o**C?~Y8lk0mx|rLr4Ah=x;{`NdUk~=zP8v%0Lv7CvSiohW^F&$Re#jPiRebO
z8i3U$4)xq9LS0JB+-(|;PywnQ$6j{^r)h8{o1{9o6k>Q)XT?lc(ML%AT3_oJ!`kyn
zV?eWSK0B+fHGB&Q`~fnc96;V<{KQi}`E!8X9WJE5Tc?-VPiJaeQg$)<u&LiNK3>LB
zgWNN}F!(`5bKfs$GrXN=3ZyS_O!-T=^sH{*U3hwqg*WA<^je)y3Ma+n8>?OK2N|YX
z$3J{=Dn%`5kT7OuLtc<d4p251J8Y_Upv|#qCTIKM^vBD06m83=h;$3ac0;q}(wp%T
zVS!sww>VR51Q(-}F^9v=|Mr`e^KYE{o!Fg*w?OhJq((`DU>swpgDt!U2WUL)`=&vi
zL}cC|k_j6?$x-AiT)i-L6b0=n&N{xy<U;Srcc!)|Yl)g?a$46v)R#Xo#<|s-g?o90
z2E%^OasYhcPOGg=q&owMKm5t1|Gn=dp4WmYe^gu`BJm!3@j3sbFL-#{_Aa>SD-2CH
zqHeSWI0{^s@zt`97D#6{5(XhgnlN&flso1_^Q#uNO>L&@>e$;$)8*g-w23CldZf6-
ziAh)$Ls1Q0OKBm7Xbx?29nmWLH?6Yt&VDXCU@gNt!_5O)=+6<xtI!gB1pxw~f5278
zIfI7LPNBoONJI6#bmJ}wJrqBKr_k?d1%t>KugupcEWzvDa^JwNohM{kqYaIO{j|X6
z=R1X@UzaDv>c@x_nCcYzW>@p{)S&p401a-nO=1Z*vB5{a&jYlOKSx>w0@>ku`9H0N
zX<JU1rX}ZW;q{J(&3+n{NB1$x)@-4_MBu`oy~$ZdGhL1*1jL|-r(6fUb>Ur+{TO;_
zX5P`};6Mr1n&50-5%E+)gDWbpc$>zg$v-r;BOJfI!f8rJvp>dLK$ALc+F1ZRQPA^<
zEtBh30Ue?xJul%nmQ1@N9ZPX&pC-7WjR(S(KEXGX6Gzmb6>?$7-`sKt%SY4@o@OQf
zW4)wU`dh6<0yb(w-)r=c^zrUTIGf_`^w??73HVF^q_M$~`KCz*{T^o_FU79?{myqH
z4Dgdih22akVLi2>?hy~{@hf^p5juk7IzhQ1?{B13^yG(wiBZ`WzdIA+)3ON5{f0Q@
zr8$Nu7`sk!R!1b3?bLrQ2X80tuz`;mhGoCQT1Z&?iI%SKAiI5E40Ldff3vp{BMF(t
zdlg6*_OrPcN^LkC9vXgMF_@c4NB4W>fJuP#7z`#(?Iw7n3JMK7Cbw-%bF5y4`-@no
zza#JnC`<vwkR4(c`<<sT2hJRm)Z}eE1&ohyKBxMCY81zh^ofb01$WWS3HxO0Q)0Dw
zHIHdkoK0kvLjggnp^;a*FhBO78s`Gog>WD_%W-p{&+KoU3(8v_6xb5nU-N2p@*0kY
zOqoKG16FE~{C2Gfo9!sAN8s#SQ!>;!1){tF8O6X7$f4=tAWjW{;EdYPk4ArdTG602
z7C8!6W_3k_lZw!<#~NGsD5+n?`N;;AhFxMSP!OtT#*{l9rX)f{Yv>*PfQakO>>z_{
zyaf?iP<q*G%}!fdIvpG%ahBN_Kq=!P?ir}%h_KVZ<%v7Wa?1Jb``5ms4US`djQcKj
z?yaHKhgL1!kJruU$^{ktaCLb^*aikf*<d`+G?Uq4Ma79WKAmE}nw<W1&rLMsdiC>A
zkAJXM>yUl8?Z23Ds4<*KDtfGIG69wTfy{JAv_<6e@eTK{O(9aywA<Y#9)4V|sgD9t
z|9x(xYz7R7sjF`h=Y%1_sAE<IG<awc*Rnhbb=yClk*{sKe9VOO0f=^Z)rXF(#+1o=
zUwPx8<(3Z(G#CH_pPGf{)sw1Jk?>$szPD$P@YUml6n}d=l9H}kSO3*<5^ZNi*G|$z
z22}Itsxpa<)zxxYUe<6wzBceD6vY@gF~8XGpJ0}iDU=k~)`$7>PpIrWT4o#4yaxYK
zO$&j|P)}hMh<pM235qyS=eAb{6~?OgO@eOSCpMzj0M{m6<vm8IIvRh3HbDEbXYo^%
zma$GNo<VHJLvd1RXmXsbE7q~BmVxs1?5H*L_B6Q+v0>0$>$KW6?F4s;fuR;353B3}
zB!Kf)@8Lsl_iYwWS~{h3o^+%RiA`DCuhru^ylDE?Tt&;<xvjo6NTRf07-_Efjz1%o
zZXn~Uz3ig*apTG$DGX7piRN(lM*x@WW{aa5Mw4RzUI}5uaBb0QW071#Av+axPC9n4
zWzi_ZJIm@mV@wc2uDS&rh-PFXgKS*j@cE?P{Xxoo(3o|toRI9>uiy<2X&S-h(C?F(
zxdQnLA&My6!Z>w<GU#>bc%gibm~r##|C)V16VA+)O4R1!87FG}y&lR}&n>_2rH#$h
z`4Ulyi`!$0VeYAH=O8}816MUs?`5MLNgfrbo^-U+tWBAfYJ$re*29^O*)aU<TyK^H
z-H)Uh20k_!#Hp&4QWfxBx`#h9bgV0w4s?JRGaN=Fd#~-*BX@4qtxHC>#AxDZIjzG!
z?H-r?5qU2UDE4P>&&;=UvV%*RBy4*Ac^~k#m@W@LtYc}PAjUDaYKv)5-58{$7+I)`
zov~BAqpk)<<4z+f=QNlZ$=k!4!Sm~n;Z|R=F3$_%35Ii<=!uxfzin+fgN?Qj6Aoy|
z)|x&KHgbb0X(lw&$Edm!<)0{;6+T3z@Rn3tVa6OPZD+U5B@$F#7N2iFDO>VToOWqm
zeQY`|8ArN~c54SRZYB~^kdU3uJ6E4rZPYVEbc5=9I}q3m!M!Mu6p16k7NyMlAPsiK
zs|oS`m7NTr#p~Qyh*;q$@k>F+nmF~So;|+y$zL8z){~#yN)c?i<pET`r8u70%gc}q
zxBPYdo|lGfcp+1XvtQ?&mh1YBh0<Wn-@GR4KcBRA+5o}|nk`%cy(`ezE?5Vhx?56s
zogx-&VsSW3bOr(kH!<Ki{B^BAhjAVv0a8%F%%hNp#4Un<YiyhI_fw94$KLbaZy+0C
zfS);+V-N9oVT{87VVJp+t}eZ!4yt^T(lZ~53Wd2khF$h*?F4agR^~|xa`XEcP&^8Z
z6?O2nXxE^zth2S)<3O`_^^sVMsVy{5K`^rN!t83$Cv5@<O>!fg(0@oORR5QqbCc(w
znD6?B%2fYwsol-LEp-5bb8!_$19GZhN9Ekcgzrj-zDk(}1ve9GlK^%x-95!^<V6Fq
zAJ%N*>%<k){L*0B%-@ZK21w&OW<B-p5{?w4kD8*1xD2G%`PHks$?TLC6dR&69@Fj)
z)j$Pj_|QWMmqZQcEYC2sNK{%tAO2kQp1l@Cj&&lM#r1)@x;wlu*)AC+ZHvCXb>nDv
zJ`5PpNk8`-6brC__UD|gI!9`{))VAKm;a;4R~-cPN9*mwSn+ry_qNjLSM>37`={Jx
z2C;rqb4-yq+`U)pKa|*u&TkEVs{5!n_CvU&=R=&_Xj<I+VCx6x_TsXoy%(FIDtlW!
zMW9OzWb%?C*hB%&V(f<+5}t&6$)78^62xA!$WM*bj=aK8v78<sw}}s5O#%jIC?Gv=
z5CZogm;Sx{<=%VA6<=nD=3F^TE!7_rh6LEcEQ@UJ0esm1chyO5;BC~j`y=F6<=I_Z
z&HD1bGDu>=bCDak`WNLns09^VdI|bHW95Deq*)tLtlza?o*sm7Wkb(>*pHNq%9?(y
zIa(FtA%zp5tt@(aO1yj4a}(w%XR}`e%wm`V!_zhh+KOWXRnJ`>jPeA79K)zblsGoQ
z1P(Dd$Xe>hgL*uymhm=dpyXdo`@QRUroCa@Cm8X_xe*mqY?I++TS8>WF>U#kL||Hw
zB9mE>rAfD1)CM)?e>JjQ@PGH>+A$Za^N@Cz+Rg4qQ9%enBbb%ehweGNrrwo4tfdn?
znuCuo^sQGK$T#JYTL*<EPn5bmdT&dpZc{px@{-GgJ`@R+SE1dfaKHb}2d}t2s9~kt
zi0-D9@uY%}$n*W5sIR(k+RMIKEW12K>*rRYyZRMrJIF7KhO$%_mNtoNIsj7|%+rR7
ziiu*#qmpAi7Bet>Uky3~MIZ)-p^Li-!~fG{;nV%d79PdnM2MOWg2vs$g<K(w?U>z&
zo<_vKYlx8n%w&Cbi{?9Y(?)5ug+s7^qC`}P*RGgdqTiq}W&hMA?+M05k%$E$IPFi)
zeB;fH&~8h7(_#FymclJ(`NObx>1}i>b7tuz@tReUmV1gd3eiK4;ZG*)7v??1yc&hc
z$AtUTR$$A$u7`n3OZl(N<cv5r$6PR+-}X_Z`X7`LA2*7G93*}()*!a#`kG>lx5sM~
zAqt#wM4k<|O$Smru1}AW0$%W@d&T2~L^b8B5E1y@`1NkIz`2MPvL0rgyDKiP%6S*L
zn)|Vd+hzAI*>S?RnW_O>nh*Wngzq37KN7+g;umpyQ<YWSNjAE3aNq@>mH-(EVK$NS
zW~rRWpHD;mLs(V1F{hN!hYrQ+_k%9a&TiV5(e{6arajw-nC{K8{WxjDbQwuzhE&@{
z0bP&Asw$t43z^O~I+D~ac8&`w>r}>%w|U{dHEvB4`M3St|LDHp(@#I)Z0;Kfm=woh
zsIG1wXk62X>Uy$2nY0XcB;UQ%q4*xx&20{WjA14VJ#5DOai>R6-17s%%W|$pC;zJj
znCg3M?9wxe$6@B;5fi*8y;|G(ebD*&B+1RHr|nup(oRss!A3l6>-8h0;Ohs7RzsFQ
z^;;S#T%67Heh=+ER(fyO`B&*z{P%>+JX`XR^jv6jMZJvC<n_5#wj~gGVSoJ_Z@Pps
zA>ML-5fbi({R=DF(rJ`fuVjIL6=4+~uTrL|aTUHIBsn+ioI(eDYvYeA`L-qyl_@CF
zgx>SnngtV&Z+a2_8&RH=c@Fd5(0Gi?&L=9u5H(#(Xya{e^ktSyV{l%sFjjTYt9izD
zr7ha-_jJnN<@+1g;f2n+|42RD>bxab42M$yQX(~qoTX)6=wEgZg%3M&G8gq9sNFeF
zDLEOr@uq%ZspWp*b2t`zf0cJWQCI4u^;!c6C%eg#{hK~6q9K^Rk6RPq($)GY>8OW9
zN4Y;)UPY09!Ci5})p#vbo!EKf(fn~M0Od6=C=wTy5(<wqTsn#!12UmV8qdtB5l{#m
zSFvmoa2qN$8}F)j6wJfcJ6g;@j#a-NWb*rUipbNeNd<%>C~=UGpCX9iq5%|=Oa%Jg
z<-1)YP_bJgdP%#BCH5?6X6j>8t}m#L^ewr(ya2Lk8#Ty!QFw>isPsP%>irTvGfHAk
zx-IE%`?pR09oO(wk`#Nv?FnP^UI$PJ1JUDGWxO#l+YZhobdTp@&&)*wi4xPhXG7p1
z#jslA`^{O3KkpIG9Ib@yv#55gai=GMjxpO*9`kCbVXUWqRIdsJY3yJTACWdwkr7!>
z1NbpV4~Nmcv%7cy7!ib076IK0N34w)LoLp#k2bp(wp$aKsV}>k+u5^aK=!q2H+;*#
zN)f9ecK$JS!xMg04o#B>UTqyHrn_i*RS2(!DTQrGBUXXjLm9i$V&WVdnwonorR3SE
zXL{h=yHx{D8jCIe5d(JyAATK;W-z`%1gMm*dFhYuVsD1BN>orGU)O*4Igjs1e~3P7
z)DzpPRZacu49JeYbptpk5l~H<Rmw;v4L9+UoXUmj+$BEfqj3?9QfyReWb#iQq1{ZI
z6mKCpW<r^aB5Nx}0Qf=nAV|8S1~ixMR)(t9lSr?`(o&#`q8o=_(RM|Bk%g}+jQB+B
zi&QrlMlkDgv3JwrC<mRt<w<yWF9L8KmI7co>!^%}j4ur*dKyq3+_*t#%fBDP#ey&a
zW@X7z7`Bdr9j2%uC6ZX!VItndR+0=gGZ}Q%As)Mv?j+=Wb-%<xXkADzj|S`=KiNAy
z*o!El<W_+YaGj8Yau#9`ad_w3VnXmeef(0)ece1M7EheDEiOnYuY>S(canXTFDx5u
z!sTu(e0OmuCG!XMjo|F`9h5|2j5O_kY<Y@C|EnMW!$n`G6a#go<AInpCtIXYd2E)D
z#4J<wCpcRBz6<cWKT5*qUIH$S-3U69@HcZLt!m(bIhX!DHjBhY8t}rtC8*uM$iGA}
zQ7!$0E)Fq^m3qa(f?8i~5g#t~Xlmh+y}Qiqh?a@c=&@$}?+Zdum~{ynZS<73xN2~9
z4nS(bw6o=&WwRPhFJyC0EmzT|xf!Vau`C-gcH`x*pX6+&6gylo4VtkI<9pPfspb)n
zQ(GB$t@!)*@$f7XTH9q9mRSkV){B=xWOYngBPb{pi`+DR87y*$bmCzR={&)G2({6H
zGpZZKfLq{+^f{)INu;d5^09d`-fc}w|K8l+_P9hLN~;}SsY41@&`c)R{KkALW^?ON
z_K?0aDKF?5#%UrJ&WxnXC5rB4%oV#n?0*;0`Z$XGv<|Y<{?2GaxHL67*lj4>FfD4#
zppQc1g%DSZF5Qwy;0!1J2|EQq-O4P6%62QG`mx}vH<|vT8lVu4f+7^5g=8s<2uyHR
zz?<G{Ll3(E3|kPA9sOIRv4Ul}f*TYO+bln;9yQwfJTBiW5h%HUzNzz}s0(g)QZ!`i
z6MQeDIMe#Dpy7W}wiqMs^l`qOjvQaQR6?0H^PEpF@`CQP&Ov=GijNCy5STBT7{ybY
zpuEX$7Lr1FD_10SYJf|PQ3MeNdF5|KH)xRp((m^8u>`wcyVq|X@pPy}H^Gf?g2A+;
zCq5O0tqKJV!sXSyUemIk^q<u#)h0pgbNowIWhQ=zcGP2d3{^SvuDw^^=Bo!#q(B%L
zT9H<nweE}awl+d`y;#t28W9X9vMS^mg#YLFbg=tZZYk9<?{c*&K+JL{_9gS99jCV9
zWFJn0n;$WLQ}2W;?aF8Bom%}J2gD^jI4qAz#T9!DY!HD}Er$Nzz%7CzE`u#2B|QXa
zwTLfV$i9c-nWOtRJ0CL&TD*ykh2jJ+`(H~%U@Tx)8CW-3O$Aa67=<$w<@<Gbl!tlq
zU-8uzgQYNT@^Wy^1{)oy*Pia`Se<9}EqG(5$)ilcdaCUgxQ@_+BBnxM8^w((J~)kK
z6P~o$@D`wW#25V$ry@jcUGr15D8ww1Df!w~)4U*!WotVZae0rMx*2(qkaNY*TtOn&
zrDC;@JnyN(3PmB|iborGebK1pAEs#!5l&s@fS_5JazvS~X(sLnJI34xT2;;fO<w)a
zr&78&_0SjSP8T&jSz}hROH*h2wRU@629ObqK}I1c&B8ovcv87le^74J2-y8mA-rWP
zQ)CJkY|>|0*L*yctR4~oy0MHkd6i2jFlvXgRO{EH!AU(M@No;!^3zF^cSp?H&-kZ1
zdZKX)hYjU+euz##S5mqd$$P)q!1X)x5NDdpgE$EW{K>Ni>CU?hHo8E#M!X)*5pCw1
zU@cs8O7igxARHtSIZl)ZQt+vp*s=JeZDpJCoAH}t=6(&#TtBL>o`dgjlYuihnjmCz
zjiqcVVb@zcfM2Gj8trXh^Io*p4c(;dzlaJbK8uJ>i6E~kgep}MPthnizzvDQ(#7$4
zy1X5tB|G@bE1BO0sm|f3>ysK*7|m29iU~I3NiUi@-rHw&iCiYk0=Ok=L=f-=iW_f$
zsds{b+Zy1R@rQng>WI69H?YZB(9h_ANgb$BPm%U7?+#EJql)@m|IGAL32|tVFKy%6
z8M(D$^b?kav|Pizq^`v_5?%<sU3FyjCXC$WSJm=g2ysjPD)YhLW6xy^{r=?sh@|&o
z;cQ?Xa=V8P6x}4rQK*X><NCQpL88{LlenW{4C9U?Awe64`i96~l-bTh(Zn5GP4vm1
zm{q9P#vs58q8;Dosav-N+af(vFvf1Ge*C84@)P>v9IfmF&%SFVJdcln(I9tmtK5I3
z#Rzmr&CU#^-@ShBAC!Nu?xTOqI1i(gGD5?PoWqYV*dR?_c)`G+e{|%>6JcMp0llaA
z{~VPEv^lG|nwXxXj?^uh#{g2?5CIFwqoMyn@2~-5mR#R5voNZ3TQr`@%JQzribvQ>
z_#gBsEJn*`PSUqqM(pC`&nV&puvJA|tCp~L!DcxGq}^n5DV9kzYaORdyBUE{on2u%
zeaL@6<OyNQ_aj769d`;RTLO(Z?KzlNn5yai2!GOy52~RH@UB9QT%v*muYVgl2fQY}
zyqR(aRAWIbjf$+2(oEci@#+-R{<DlcN|%PDPw0m4_uRf}BA-my%xK&8ty<*YG1k|&
zZ%&jb!fQbV4c~Wu;xOS|pxAtWqNuNOwMO#tC2t#Aiy!f!_{GYU%Y*S+1~yO26BHnV
zWvs}S98vOY#KA?!o;?lI4#I@VA&PimmsI*(^TkGI2emSay}%8hOF)<+P0*;$D_aTI
zM2)q!+Cz8ka#%oMx|@y)!q#~HM$PQ<#Wz)PAd{NWBgb^7msL8YMT-9a#Z)sCjd@}T
zbMihZ<zBdw!<Hu={0LGet&6^T>U<adX_dJ%|70|ovD=vy79$Av3Q^jH%mN^{h%4>)
zTtKhi-B)c4+n!*hfKvv|9a+fE4ND0p%ib=+w6qzct&{=64Z%l)u&U#%!;D3PnI1nJ
zk+><&5QPcP^s)6F9PW-KYxt+WnsDnt`@$t5h_aO%XA*|IJ)#vx8iC21HEWusT+5&L
zdH%ib&!{+tU$l{lF@750%H}VG1HqIfGD2i)G1juv42u16H>0&04+!rSv@sW5YX8d^
zh7})ABE|WGOS~TwDe@-Sm>@YYpPB0fe9w_xti-G&SU1pAlmi9JuHBOOtZDw`<F)8t
zJwObWx-`al4r$mI&R6w3n@<gPM|>S^XT;Rycg#7H!^b!S^?VvBpPW~#0AVmdY>d}h
zgp7-4HxU8tP~#3y{Crh7T)82^?QOp6d0d>4a;ce9<$ja}Xj9|$<_%?g#ghe;u)_Y8
zS%N56(-hM>+1hcE<9x)4=pl!?>iE+rB+kY*EWq1K&P+}PR6`y;6ML5Nkl8OjQU!lv
zh9hQAtA0<`|GaU;pGClbHz%z@j=3(xgQ7^b?ROci9e`3`eFf$=CHS)UUf|x>OQiyX
za^S?%9}DZ&uo7aatl*Q?CU;_4q%hb1$)?$09qBREE%^fdCgjd<-M_7p32eJDLWx-M
z&;QA$rL36ho_6JdD{%9w*z!h8_6kDfItbm75B+?hW3JT;5sWcZsr7ztZgRi}Y_N{8
zM^a#UhxN6SeeJ}HD8(%lBuQ8gL~iwFXxzUNlbs0T4iv9?v@*HT36<Vi^(!FpJL?!(
z#R-Axyru7B+eBZd>q2bUB{uQu@469u++Wgo{EPg-TS~Y_cNXQql5V%x0MFD;ne6lD
zpR(syb>7h*Lf>up-)OzuugC7MdQrv>B5?vi_zqeJUo?qFT9A-0zO#ptMln!PR~i7n
z32-{kjfj(MI5X`hQ~VY+*7z8g&Tobyg=zXP$80xGA~V)11k{?hS?4d0t38Du`H;Bm
zDwFodqWz-fDLd<#Y2IxUG7Lq^22FYwU0QWNzQ;6e_TLvf6`;o}V0ubXp=IPI9Y>ef
zpNOX&%k}6B+6}B68iD0~a0L-S>Nlye75fc56%Ypz!hb^=r`mhkczP*^$4cPIVF(J$
z*MVz)&GEO^R&sw5(ITY*-(2Y8*q+%yI{#|((BCcElj0o`urJ0#b3dAgd@UQn|GJff
znvdA&&h~qaYvydK$ALLAmzO9cz(VsE|2gL;?JOPnCGHPT*Kq|9g8wGGH<E^Ni%&Ct
z0>osMpwb^>OW;4hRrg}Lu^$Als_CH5jhMx7+@7ZE2I*@4$-<nOv`Wm!=6*Y+oPJDM
zABS2u-<KdtyMj`53uqj|%$#M%gOD)zq#LXXRYS(W))`yRd3U8+${XNBEt&KHY*7U?
zbB8X2KGLtdHzHoc$KAyI5YO8Hu}evv@?tTzZS?7CyyTYG992y9S|4KqJboz!L9lwd
zs`l9+@7I6$Dw~~1%JvT$p4!0!Z>&1!;p&Yj5ZD&#ZUr%YNG-JLYihsKlRP0t41T6z
zCQPKIM*CM$S6ap6+=d&at8SO|VwHAUlYED(2m6yY#CWa2rX#vBm<chsLZ*1ZPuD&(
z-#r_n%+GNVAGYi9(Yrgy1eF)>ZY<1<T&!TaAf5bMKS#IJ+EteKyNv~c>-EW&NwScK
zLl6lbhv<$A`TC;;RKf&l>c-hA?#v#9;poGW+n<%|`^OBM6gNT6_<{nw6O)I{CJVui
zna-;MxDzx@y35%iVFR+l1D}Jy7}a}My<6-dr~u(N7CxsZ5*&|r%qofYU!}SHCISP+
z)_;s&ShV~B%s$D}E_OmVu%QQZK%}yQ!n)vD!H%|0K?tqe7*|M>Biw>a|1u@B=q8F(
zKG^$igUiFp(sOG|00cox_+lt(|3JpD&6ljrR!mB;tgVFyV99tNtBj|BS-Nzs5E=sK
zbnk0SC9uuYx(!cqAuI$wG+PY-ml_!)Lsl&+4h)rmSswwKcX!8kNNj}pdeQ-JdKV~@
z>PwSR_57%p;+?=dis9sg0=WB92i#PFHhy2v>IyrtQsB1p@mzWYK4jDwCdxDuU(vfZ
z^$-K(D?3-}ONn#RaowL>lYexBeIMgjH~^II8U)}#l$El$7b_Q)B)w@+!jF3B6N*4l
ze6*gOKSD2VpJse;_k3KIt&+dB&`Gvj=&(x$LN+ErPH9LsM33l+xZP??%7!UourK9U
z0okK{Jh)3jcXrch?RoP8UTVen%eW9RNOXGpMJ43?Wace#ocE2^0(H}yuHFWj1J~4z
z!Py-vqo+_=&-AAWekR6RH1nb2SJ0I9?=h^X^H)I#q2MDo6$|k-_<~6@azJxz+FLDN
z_%&BT>w`bNO`1ScJzu6~qy;8)an?_By^b_na_;&`c@B}|R^o+*d+F@BGULPFY$b&6
zp&Q2}g%pjmIxxou#v_m7e$y`$Io$-6Mj@35I*CPRMQ&hQkKY=Gskgw{N#GvzwI<Pi
zni?F@)_u?^VnfmRXd-V_Wv&60{jA2Pel8?fBqh0vyjP8@aNAN|Yy4!0-z%LtI}_EG
z0LL+xtQQzhaXr1wUJ1mL0`Z6fs1_!gb*d^asS6t5`F}t)*#|;_I{c4SYx(x~TJ+Ua
zx;{|NY)o56%ceg+J~cf+5PG<;@dFxc2=$(c9z;7%H_m(*q@kOvFer|nj}X%T=v
zj%Ys#Xl{a(+yHXQ3X%Q;$pC|Cr75MMH00s_LfK5?Oa#5tH~m<>@AYPUja6?|F%z3q
z2j^M8l`rTV+c)II(4{>(no6;ajEgmTv&a$YKVR=crjtpMZ_F9m64H+mwozW8-XXIy
z2!`Xas*FT`rL>U4taEe}`J_+9YES=x(ylv9nd>yv^B=74@wTktf2Ncq@t*#xb@b9#
zWW#>i;q-t_t#AL@X#ca+hptpD2q}aLNBsz!9*Z^fM3z)7l$T!+L-cuq5;8@vl9FO}
z_L54~glNg2;4*-;gcNL2hn_Vv>VOb0h6Rd6{b+{E7_ZsoEbo$6!<yDm+WYtK_MH%N
z@9F8ae=PIP6jQ00vMFP}_c7em(%-+!!s0j(*+E#L0E#%2fwIB(>L2#-GMrz^?G0XK
zy0YLnauuiK2C%$pz*-o8CoM5{cnjIHW*8VC(jj+Y{bUkVq@&^CI(P7vE|LcF{gY1>
zr-<5*{;gEY8I(V^n*q2QPu-}7rdb{z7ZTr4d`<~RQnev7wmTlN-n)^aOv9pc(M6gw
zYDVX<svg%${`v=UtLtgzM)-u6t5Tcx%qeRo^8W$&PW+HbmZK(A48cGxtrYGkt77V_
zh{nPpjbbL!ic(sPP!G1?|9Ot<#Qve`YC{O!Y!V~0=3hZ%c;H!DNjIsq2Eo9=bBmzp
zJ2L()_TQJ!hG85dlJK8%%tk1|EA}7+K%|R(S3)R7U2z7B*==KYvv6`?lb_=G0W9ZC
zZyp2gu=R2RSTOC<!Q^_aBcpAf<4SHR1kw2Y5HMfOktyj!`-az5zELNQ)QH;&5|L=r
zFnq`Ld?Nf6%`?HIJgm?@RZzS%U1P36-q-!`wknAJ*itR9{pOe_a@wiTwBgr`Md#_8
zhy_DUkpl(IlVoUOdsW9g-ermS*;WAtx!`+}G@>%aVtLacu0cJ^_~xkx_uTI+gQ=%n
zt`ugy!IU4>cj3`88csQW(|<Fz*>DAk*qh5R2kzt<Z{@fmx$<Y^FTRH#!^xkIipjuI
z>kz^n`Uj2@BWbw{8Xk+=J&CkfvxVEJkiA_<)Wjaacc?&BrGPa^#2vCTz&cE)AO2|Y
zC~dU(UOOtsoXAF;fsdh9k)r%<rgxqBa?525*kkf!vR4~9g-21}X7g%CC}dj|$lGX*
zZ;$<F6a*S%qY+Htgp)Acr2XSPSFxF=RHO~+QD2kLVkdCCrLokoO_L*&uyBK?qDY&;
z7<-IFKBGnK(xs=$TI-pg!&hC2mnys9_?<>U4k4!d^-(jHSQIRrw=3uYoz?wO5xt2f
zXp#N9aWq1NeY+E?HZs=27bkE#&J`o>_ZRrlf!Ed`qc<n9#QS?vD#9o^OO2PP&L@S;
z#iyN?yV4$RzyxwN_XtA+IklC<gdTf@F|~v;i+E0MyqCxda?K?)7>wQgb9$W7PT)F&
zgl-li##Gr8^Rn9xtrF>SsA8;&3&+riJ@3_9M~$5@wv;s*Q7a`f`Q*Tv4=etzYx|aZ
zksOXvU`u!*SZC~P2X&}6P+35t7kG>IhPxcquGeGZH%iBGHKaoI(X`eqy-vKAMgn-~
zO)(=&E>>zAKJGXux>-;Bq~bS&8TIEzvqWU_YnvS@l18LUfL7jd5Gv%wo%vFM6)bKQ
zU0l`EjIx;;Ul4y?V*!>gBfgCQ%jYH|gE+r?l4$(`yd2uIXD5V$UYz~+(37~16HmBs
z4`_GZHd`@iH(6V*r&OnUTB<m;r<I3av3|IqKW;sti?xl%<bsIC&N|M)eeTv@TOZ-g
zPiK_RgSBs9qDVt$b>aD)Zo+WLf-Yw-_*M@hwX~ccXa-x%U&sfSLg3UiO;pdQOf3Q%
z7QPy3E|_*Cy`MH0IZ|eUFPA_*l{eUi9ziXggN-TLht0hxZ&lbi<?a(T{8|<#obfX{
zUz5Mrm=}Wn^7e<Bbv}$)M1lUg7HqWOVz&dtGkyP+wmWX}*eDyTY3gQcADnU(97ZK5
zRZ5P|V~+}$Pr)=Ok^KZt-YITkcKXkcLu3y-svaM{MV&8b>#akJ&)m4QlZ046#e^*t
z`Q5A!Q9eKS67AF`us^zD$gUz&tmKk|lgF3y#c@R%>#<oY!;ZWBaUd>}&uC32^hhi$
z?<!i{`C_ia#Q^W!_?&5BPN#jUMu+_|%2RqnPz1m9YgN^zpI}lyV1Kc|+cP?+UO+1x
z6q7zbCiyz4m)3>VdQvUghteokOC!#x_kpI!9UITrN<1}PhjqN)PlixUj>gVUD`(d4
zyPhL6oAMxOHULKo`+KX5^^+`P*g%|jaxQhBjp(5N5=FxH<4@lEe4zd_hVCJKA4kB<
za`cpD(}s<xq>kC6v6v#@4@eeqwWw`_`2DE4$ml_Cd|bd(s5Wuc+`O<j5D_j;vDhi;
z$}mh+*FAgcDgo!{YB;$;0f{%pY-o*gG^!lkT%+I!R#bs7Umseonl!KCBh`&1P!rMf
zr&D47VXcHFiUlJyb)<5i4gcyLv67FvlW3)s88o!M?l06qti0VaXz&1;hGp{T?eE?w
zai@x~EoY$bo!TGX$o?#RGX&FP3*{h1mXQ6L7gB=wE0g$2G)Fyal2dLVl3FhX7a1I<
zkQFB}Qt!=>ve5nOsLN)G{3~aa?I4wJA`}VEtUvRoIy8m5eas=&oUjkod9l`A?(j?L
z#hivlZTW0LssUgHPYU0*|K2`M)ZIv5s{IHK<985^HW_v#Y@skl$Z2qHyTj=Pj<^`e
zox(~Vi1kB^n~#yKQmZTO<j*$w?bw6K5($3VbnfVvV*cxwAl=xrbc+Tw<<?5`)YqKz
zdU7B~o9hy743}Y%#uAHqY_=H$Qmf)^0vN0b7H6e2d2ghpU|-d><A1c<53y*^_L>1;
zH1#4330kf>E3n`WH9_jQJ=9biNHI8OQ)XhMYi{}f{VB2G&FIdMyUL#Lu=IegHq4JS
zj}N3?J0K;%y_hgMDTOp1@kex;!Yv|PY|5DExnC4-XRcT3Umx}AyTh3|uUkAXZ&uPd
z=$VxSCpe?PK&$7a2fjMgQ|=?lpgQV!nxwy>j2H-l+k?QW6Fp>2B$D<0A=yn(M~5^V
z(^I*{k0J{06}qaqNd43!DoQ8Iv$hCpK%dv`Z@ds*fw?y5)Yb@fVmFRz!QeJ@{({fD
z4K=R-pJ!sa?=%Iw3wf#+<w;Hd-mgkN3JfH8lMso2u|fj@wa;DjRfgHjmv@Vkqp5ra
zBvRsZ&=z)InT8+yty<5y*mG2RP?}IH#pvzUuIb(+n3b4K@(3Frk8DEG;cd~1+n99?
zvcV)W24WzL8WIPHLd=-GqlYj}!k!<o;hy@8ysrA0{fuzOnU5IFniFdlMU%&!J&mYR
z38XKMm`ju&w#$#&`3m^5_ff;~P0LTozf_2be<{brgKjUC0Xazk>8nCW-np?rWaxzO
zrcCHLHBl_$OvuUG-3n_*KqCQ1-nVmICQhvVbO86P+;@8!WRG2s^CcoD@{pRBt9J5=
zkaKrDB+d<4PwIrt7_VJqdxvMnSjzMC*Su`^t=rgcoxRa|JLaT|$9ft(9~;80NGuqJ
zA8<=rWF)at$M1(OMHiYK-JLLDu8)%PK@8vt_B=p63g@JRgO-+o{awO1lI2Esk089d
z5}om9S&|)vTA-@md-w!1@FAnhOkJ5xocaqAi+m<e!g}?eWORgvUezS)_-_kUmKEv$
zei;Unio2CY%QUC=H`2>`u3x|SEt{!{V=h!%+*=~o*A}zGtA+M_)>sL=j3+=10Z9f^
zhuWyg{!vTWg}Cm12`<Hs`WnKZfGapZ{+Ek6&@p07g0R+PnUfuuQBGGiGY0jJ^Jz4F
z5*tK?!uz}+M|dGCy`njlPX^tX>G{l1j|-e&?Iy>{c8q&kN^Zat-+U%740;ohotU2b
z0W0ASGNUwVb_fo*`8lE#`+<I&zNVkexU#QGk!~r?ijr+Jz6vBr=ubgwYNU15-;sOw
zwG`EfrLaZ>uBeH4V4+}Y6D4^(%;o*&Q#3YlgK=jwzstL{5k)UOw#djcm^GJ2F)*wU
z<;4qwq`T(B+gYgCY>=f;OC&({rdm><>~WZKZ<68(^^z+pTB}=0KkS0=?FVBnGA9yS
zGx!nBsYO$@S|2qu<wiKw>{RZO?XTyJA-W4KTCf@c1b!Z9*)tc?$)J_buwI69n37b)
z;;p&i{rvpK6Q=*G1z0BEe#1MG6fEATE)aLO?ybM!f?(#*?u7q-ceM~Y0sdTq7*PlI
z@(wqS&U?HgbP9QP;g3DkD(-Vr!_+w;Qvl$?hnsKg9+_ZmKn!ZgXZK3_`+%Sj0*A=_
z%y8$mcyb95I2%Pm8Ie-U;cBF^)VrPNmLx`Zu`wekt_5&HQW_1vlwMyIGoGu4gRq=(
zj8dA3#^+GhP@6hQMFta0es)v<W=8U^^AMK4f6{3^hm+hI6n<QG8Q&`0IMG{lpC=nL
z1!B<b^4?$A2KMH2?m+Kt{guV&n~N;jLe-mU531ZYvb+KCx7CMFI9KNjnd!W!h#36?
zV<`uSdS>H$L=*xVz<>)G_BCZsLm8zb1-O2bW;6p<b;p0qOO{prs0<9U^}`RpnnF>T
zGndu~V8FpwH4GrR60CR<ia-2pm*qnpxl-ppwNL}k1(x8lrw#SLgy0NXFJ~Av*dS$k
zl8_UMSKTV<E^l&CtcU*hRd>|`885-IHCa9kp40-v-S{4jrs_{fU@`7YUuqFte2R9+
z;;>P#1TX+DD;mxJX$q&-`0f7jQ6kZ0F@nRysEO#V9?Q9yV~am0V{bq>=eWnVK<;t&
z?1UJQh5(Bl?ICWL4YqMOk9`o3v}SRjmJ3$8Afqf%L&(qqi&>V%(Pz8y!e;$V`tyq6
z`y#HrGaT3*0>`K&Wh~=PqW)^9yR-psxkO{$MfzWW3Vt_%tYKrX{9fZfxnQ<p$QBQm
zi#h*|x`GpjWwO?&vEEppE577|dWmEyT0hxXYQYc=YJn+bOv#`p+KDzKphjqi1F}QL
zhNvBK$#QysTv+B?906iK9gTR@fEWlCP3!XbyKO31n;&)?eLE(i3_fwgiz>Go#oIoc
z!X1%dZZwKHo*7y~(Ept5pGZ`7U39$(nV%fwaBjH&&OG!FTUM*-`-tLpw%`1LK{HK|
zmRwa6w?gJNd>IJ=PZIkfsvyx2w(WZ7q0iz5>a-F>B8g7jstKS<6t+6@z(1FXwwNo-
zNSEzquVE}T>K(XhZT=-_;?&pjtJk`2CgXXLDT7SA5Nn-Hj@vTjJoVAk+a4A8z+4e$
z8E)&WYdJBhJn(1#$3*x2;mgk!Gh5bt?<aajdqLoWJJ77JRl{=mMn=?+AJc&o<V9P2
zE>%^eVN?gNJ~bg|1#t_JT=#sM`|))0im>puh|YSYfi4-;fFx{;2<b)IvfF_X)K!B{
z3V-Ld{-c$@1f4I0Qeyf4Mt4TX3^TCzb5yR1r1Mt%+6X3tkP68Mqg9NA$r`wZSy$x}
zlei;eAWKbp@s->YI%TY7WF?zLLlgxCe!x--!1q`si${s|<6lR9q>pt9(XR!r-CVG*
zSDH5EixvOGRZ=5TwlC+S!Q1W4xr+Sqh3n|72jXmrm3jXQ1mq$QXBR60W3Y3=v!_su
zL}+FY(N@pbT^?0Ts){2aYVBXM4id3BY;P3~X!x{f$LEyHWr|e_Q$Ydj0;ekhr*1si
z08ie9L*9PApG4HB!<NgRVnwSmX68u6Qu>!v^5+;8f=_#len1Uy&p7fW(*aJdrS`Be
z2~V0B757{%0KlrtU7s!&zFdx8(`3Ec;D2g)b~i;sL{c+{VX)%51^wcEiWPAqn6EAS
za3Vqd{@TSHsby6Uy%)-TF|AU}y9&C;0;X@JQityAvv3qj2y4uYfv#^fBMSK=Jzt@0
z$I-8k=DVe;okU-;JF%TSC&JfYbeRNhBBu?K-4b5zrz)&KzrD%FEw$=gJq6m0bF+dq
z3D1#+b`tF#OO3nHC%ISKJh5gAtFV4oCZ_EMiDK2HHygeZ=<)^<h76LThq?ZD@6oo`
z5p`!4Oi{8#{g7$-iitoRlKxd3L=i(!DQyq%i)lwC8?9EgNeT{SE^eR_8Jj1WlX>_k
z43Rws@#ouJn<DhXk4Ddzz45|bc||a@1-vBijra%8mXKbk^b?f2gemcbQfUlh6&X4(
zCEC{PuWws<hUPQQUpyIdlGI};XmEPot|x)SayCtqeQW|pMW4nc)U=U6`3DfBNIb>Y
zkCjI}g|hqC(E1-OfBUz3fh#h*fitp&-3Tq8Br7LS)!rZmZLeZf6ZHsxEZxhHVpFB6
ztqfgs;+lbAe?Z4HoDL&-r6L9Zdb%u-_dc`E@GfXC2;Zf(Ae<FywHn%31?y|VWre(W
z@G_aHSEswg+|ST++^o%Il)fWUZapQ~-3YlAI0H-{fz<x2xS8!{a)tBU4T89Ov`(*|
z;XsqZH0qs}tU;Jhzwwwe#p3D0QsU~;CbCgL9#}#CkO4HC+TpU?)tF-6J4NeFc_e1>
zVK88cj!O3D!dIF9{&h;sFpV!FZEjC2dR!$VcO4C=t#0?~TRO;+hkeuNgekN{y+i#@
zQicH=g^$qnZ-rIpZ@ALA$M4*L5|*gpZT_YGM2lv6UH#0cD675r7=gyeBIaa;Xtcd~
zP|fb)$BPxO3Sn0BN<ME4<o~dkaBoyXjZhOZ&QH>G#lmGSwR*~yJH*F_^N>uD54XAF
zrh2a9UpBISkL6b)sM(N8#Q-BNl0K|Fs-=q&v9`&r-~$~Jc<-1|XhUvK(K<AZJ``lI
zpa}E)+Nt=?6-Oo#1QhLXx-VnabDwUAa(F^aqXB<0f;p!93JRG-0z>E~ShdgRi}n5m
zHq!xmm^=6bjQy4x`!YYtDX<q;>qck#*@YH?u#>e(7WleRwYK6h>r=x;@6MmV%exTr
z{V6)0#RIyjt$DduUZ3;1m}hOX+z1+ftK;X-PdJ(CX3{PoqA(|hQJ4OdIn{y^kL6uB
z-jx4kC2QM2Auu02z)>YZ98nrmAv@uO#XK!Hvz2z@fNrY@*(6Pgw&Z<t^R+uv+6lO;
ziXzz64KM27Mf|}|s8$w}w8fKxH7ke5^Q?E*nH7*T+4mgikAaG@X%Q$Sin2}oMh_lH
z&85DaloZ6%*x-^ZJSS4Q>}`^a*Ste5u6)|UirX#sO5B|dT3Z-SbrlapqOOF`V3es7
zXo4NUHp$(9R-Cz~w*m^a{lEkjw9lU`tLR<a?8nFWGBYXnH|MX$UPNX?NuA4Ct!X9W
zRCEN?LSK>DC)kIh!v4#r`ne^t9`Ro^ADze`gDiuhTc0sZm9s7{Ar0HBgij_a__+*Q
z&uNL3mjq&NY}|vs<G0cv*V2Y2awJMVsB-l52|%V~UT9Aq-6y)SGi=)R&N#NYv+@xp
z(m~Z^QcV52&PXJEK?I*e6k-EE8^E5OSU3FXPxX{BOxRFCQ3I9pQc-G!Q61$ew6iYD
zNd=aSE|+g{riW9&^bk@OqlD^eJsp2*5c9kB4Uy$Tsy|kM1z1OR8(^pK%Eu8cl3IsA
zU_o~vEIVZ|pZh)iER|}^a)15NB|{wr2`)xwT9E!rKy$2mec&^m9k9%aF~<RzibPo`
zTm}aD1BL1QD!{gd#j`|KSZLX`e;y0VKBMosOVZ_$6j1>$-*$cJMU;|d)I1@Nj|1dT
z_7=F)37{6YGuzxq?a8st%{QWIa5)d|5mT>~yd`~JFQ7daJ~Wk6ZhQT%PDGi6LCdHQ
zLR9GvLb<!s?q8z0`<x8v_N3m`wI)Q-gKl`ZT~b=P^$X<)d*O&$0w4aep={by{2&4u
zly(tGNRf}FD6n=g?DggfWy?xrW=gV0u66$+lHo2F^PY!Ca>1{(l@e;*=LMYlltoCT
z6Ne}OLTA|12qHeUIF4N^)B0EFeha@)1wzJX45p@2^{~r6Q>uu{YZZ`eM$90;!H;Ia
z)j88R<h&w=SpKUFh@%7BoOf*FE=}auoeV-|MUvANgTpL52Qqs-TQ1TcH;l;v!;m}U
za`g<Lkqh~pbW<Ga+v`QcD~W+xLSW3J_EX&O)s%c8!nqPq&IOG-G9rI4O!I#)?z<jq
zt|Kx^OMQr~nUS)@^tSvj%UTCts5&pQaDrdKknQEk)`a9G2;Xf;9C+V`zpO>^R8G(?
zHR?CCQgO$5DWYha=@e|15{jo!#_pYpQf&O{)nX~VE^!rvY?YtGa$r`2G48-zBT5i2
z-83y6cs~F4*WrXCtoaRkU>T)%FQV<%Q+5<S)sx7XZyqN{Dxz@c#j&w0;i<+rOs1J6
z-Ybg<lPQ5m5i1rs<tChnuABMM#(C$<2XN+8N-iUc|NJX)hjYP};=tGN#TA#01lfrb
zMAlh}qzw@KhaAr9hyd91X(s8wU;YC>hljhzdrdXOHJNNgtD~T?vEg8h=|gyJhEJhL
zW!;13K~8dQZq#F*vUQa$7UJ67!%9x5aYIGHi-_O7PZ47I-><im53(A>ni@9)?m{Zk
zwikH~i91@arJ(1-cMErF2|5@(Y#;^+pTm&^4Pi_w=Ps_a_7@9w`xNgRyga;YuxFAO
z@^4ipY%kh3tvG1N<oE05&Bdyd%TJ1_-{>d#86f~A*9gYp8HI^EV=k6LASKBr(JOB5
zMOjdwd=bpqQy+(W5g1QW`TMD?K1yK3OD&Coav*U~bs~7nvYeduus?9WyY4BR0_$z#
zG;*0tBJf|&lF+-WeJ>Zkqio=1rx|IJ77y23!<IVp0C^SGH5HL{UV!Qs6GJ!N;^I%B
z@mHv#U0D@)#^~^U5lja7q>L?;@fh?$OQTZYs4`Pd`wYn1-sD<&c@|W;`XLkM%f?^D
zdU|56>&oSdOd-oc!@<c)<fOgE-+!gszB(Fe#D$0dq9<RZ_kGEjIqe_$VyIh?+q*sL
zTHJO~c9W}Q;)P%qW#iJSAdabLY2_4!5$f{jaWW~h8qjLNqT=q#hXsX-zw)Jen~p+P
zEgQU-3+f1HL7cv*5RapMZlo9SSoA+Ao`}1&PxYw5N-t5w2^qm4OrnY!k>E5Mq{J_F
zx*l{)zP_3Ke$6~7kun}p9HL3T6rJpX-2F#VwjTZgNB71*9*ZQT$hdyKgkJYcM%<6~
zso!Tuaa3aSsI**Hp8B%{OaJ6bITWZRd=DILXD<HX_{h55cLsvIgb{z7dSp&60yXO2
z1KE3J8X;PqJ$HXcrX~woV=!UF;Y;PX%S$91_$0G!mr)Ncdr<LN=S17vyr<@}6^)?Y
z68}m_ArYNYt4`Gd$buqX7h;29#ldC5!FgQR;weR%zv_Z1W7py02d=gdip>IIeeLvp
zE&B{v3>v}d<@R-WqtPZ|U=vwF4K;$FCwhwOGsW|db8b)K&t~TG(LPDi*9Yx%b(Wto
z6ny_gyjZ1qgRS`NOvuk3QQD*8VDyE)&b-O}SI!-R{(nyVdlsvZ`-7eTOQPjTlOK6D
zT6=#;%|;7;tSi=&BS1Wi(;7?Bm?>@z3CMyZu4dYK8k;nyKtX#vgk?cxE0{%H+<M_4
z=0pg`Lb9K9!d*D?M+mj(mAz0K?@Jn)l=~Kox9hMRes;8G2E6^T)gJ|S^`qmqRn!sb
zZ_KE^ejlp<rd_(R%V{5FJsFaU8=k)PlZk_3r-jNAqnPs=f)q)lWmwvo)-cI+TdPna
zxiDrN?fsg$%gMjB4JMA`S&$q54qo8imo@w_7S8=Bahx%e&B{EY6fCJVL$H`!(+Dep
zO^zF85~|biKQw)Vf8Aa8Y}nYgZQHgQ+eXvawrw{_W81dV*v1JP>z+RE{rv~$bH01;
zHEU+p8Y2m7R{HLWQ|nzzjE7%2R*n-|AnSqY!ZFBK%Bp+rHKT^Xa@G+^6r{gxz`D2j
z+stvIShtA5UA<YM;mY-!*^!E<_owf4G8%mT^iB~cV4%>r(Pe!YL$kwDf;{AX+8~5v
zFyqF7TELf%8ue_b%FeGkK>hbWhxP9dkdt3)&J#vG(3}>IE2^#hkV$MbEHbYU(VRq9
z_h6?gxPTu1*7=BpsP1*@;MELGxzwIbOmPJ2h3;eyf~w>a$JtbTyCc*9*1Mw-+Dey=
zj4%RZ27k3jwpmLE)@UUOl<r&?L=mYFh@<b`N@+=*5XV7d!9S#YZCwB}(1Lj_jKvQ~
z+4)^6*5ZkJeIhVSwp(iIW=9tf5?+T!ig3I3q&J1q{(~Me-~Qb{wus3SOJb_E&}0R{
zmId8E=zBl8=tYxEZ{{1#kHQP4y5S|oDBzVisxGdx@9^mR3rR~ek>SDjto)PxyIb%j
zy&2^F$qr-Ml`RatRCda|D~#Qi`u}nSx-1QvC>ix&dWl;e^<b)>=O^HUcn%XJn2`9K
zn+KU%&~7RsjJsi^UC}Cm-h>rJEW{Cm*2$747o8zI{HWi)sb%In546HA;`=E26Fa!4
zhwT;(otgH=_S<?_C~qrVh==c%;LO;ElybL30B2|3L0G+*Xe9v@qp6f&rFv{sj%dkv
zG*ZI{d2$f^STlj4d2kMf_fXu&`rxv0d{0Ka>CQFn5MthV<lPFDZU47?FfrF~;51xA
zz4Z^oG<5Q!c{~wo16W0Nw+53GL1f}B`{84WH7x-?f;Aa8{`wE{2*T?HL)Dxps2csT
zq)s8fJ>)zZKFHUlxPxkHzCU5jzkUnkmom2D)<5|{v0aAew8E2fdYZ5QS7RF6oPTf{
z2q*`)(q_L^admIE>tk-V)Oai%icrC3G1JXBk)lV7U`lX*`|ga5yLBoXhng(25*6ur
zIfA1r+qn7s3Yh#cdqk3BeHJolVgANSXMKsGk(O4n_lbwaGF-;fE>joRe{J-1%Xuky
z28uE-C72Z_XRWfe^R!~}(Isz2spFm3I00s1*_J`8-k=sLk+<Hf^zQ)6EB%MqjdScC
z>yx@5DuGngZ)rl(@o{AFh#L5apRCLi!y#xRY0aUlUF1+WjBv?<{OM!1JE8MFL2-{F
zX<N?ok&5TxGC{feTeY|;NzGPIl9~L4(jY+Yh&4vM%yp}q-or23>)Uq1HxtyUX{G+?
zH(PImMVzIA;eCv>3VIP<cU01eI{J3kg-bX@%hLUZiZ*)%6uF?I67JjK(FvykI*=I<
zacSu=NJMSZ+}!UivRN>hiXcLtL%i?zj&4RP<DB@*YCpu~MAA0G95<V=FXYrH`%xdI
zCa6Ojp1+-@I+9qe);F?24aF3((}4=Lxij2($sk!$X#cn#LG`&;ddq~jzt%}%FN1ut
z&zs~`w+3Cna?ptY>M-$xuDN5fO^+~j_S%{;tw>x&J!C^ciurE<w~}@B&5jOecl(@B
z2+ePHEB*-1x46(PLnP_>|6R4);|=%f5h_AIC@H)H$hIY3^N#_&&0bH=w8PG__fH@X
z<a2v0VFxc&PfPaqgxvR+t-5)~)KJ+s#mX2@9tSr(rQ_o^W<Fa+H30WYk5N;(!)@Uj
z)>Xa8DT0C@q2pseUc{q}iv{KS&7}Y4*p*r~jIQQh;kF01swgLf_kz5T*?-s-W;aq1
z(ej%}H*m@V5f~MJ3Co&fS!ze4aD|w<i=meZqTqsxP||o-)T@-~N|OS$B2u}ARQwxO
z`O@mi+bhR9hci@N1`+b?&Z4+I?&mMyz(%A@2QP0h3dBR(5=E*<N@!*x{+?4QZUy~r
z)-N8r-|<VX4gSRmyWx0<It2wPnXEyku>|s;E|N_yZsYR?-P~B6icPGwGojir&$l0V
z3H$gud9e<pBJr%Aj`vf9eJ0`NXPA_;=58-|GN$M(NsXAoJ;tCWE52yGUeC3}0Nmut
zZWjk$RxilNj&?tD!3I^m=RsZda+_1LlWUGf)gQCR?HQ@_PaX2J2b0mSQ1J^}nACS`
z1vL7NEz0n{e$A*-cwNuOQW~V5{T!wZJ^w)|J^CiHDz<`pr9#O3*7S^Ly<4}j@5cjr
z&=G1v_3kYCj_e{8cq41Gs?RwNIz{2g0(YEKVXzt1RD}jF6x5kNM_m-OFO4FwK>@;;
z-|DDvr4@Ofv%R&J4aA-IDNgLE-5rK<T;K%#jy7p_Qm6kZ$iJcf_YzlRJw8vM_StJ{
zn-i7^nMl%*XYHmO&Q~g;n^T8YuEs9>!8y+~nu}A6X1zJ%MxIwO07%GrBj*Xjf7N+%
zCX1dUq_Rd?DnPnGaELc~JTRpWrp_b_9|5)d*=&r5Dqda_S&ykl7QJpS#7}o`I$9S{
z9qN!rVf3{F;&cw}a%;GuS!)QV_2qj?9;9VglP4OYW4{~ZW6zyH;z(kg`}z$}$O%OC
z;&!r#g-i0PjMT7S663{rY*+AS+>5w<d1%$E8e*<}^5ooATAXynJ1-sC<`Obm8cm3W
z8J;47X1%`m*ymlo-8_(#N{oBC!3eAMZ75tdQ@t6>`~S&XQ^3GkvIUB1hx9;%INYI@
z4=|MjEM-hRm{GtEB>L#mHfU+w>Gny1raky=tBAHY6+N9Ge^u+kEjR%BV~XP*B@)MJ
z&EwR8SCrvfle-`2pe~#=>fv#5A2t=Qr8yNb(|DYYV_m!pLn!9(?5u)t#oPO5t;N`e
ztCILog37$z{>Nf-xnk?V)ugQF{{Y_sA`!z6V2u=OCLk^yzq4hm<YyE_Q$xAP>I__V
zkNdr8*eUc)gm~)buyt}~`<F02&`TwhzY%VsgbbY0>^N#4lc{&_UNA$03s*N!Mqi&_
zI)@wX?Cn6!jT)oE#-A~xL}$A(P)B?FKotq&^5r<L(t$57g7-2d6JH$K%W^ZLI-1J(
z@F-eCLZRi8(g8e@7|OM7<ZlZFrf=5{ryZJ2HS&@qU<cVZKVLEBi=oLtE<^8V%=aw~
z_IfOu_*^z<7q?Uq<=pQ7c8@dOD3nAV3@MjoRv=~?j;k_8N7I*1Qe%UTTWVp+yz>*x
zm;X@en=zekv$1A;QvUZzW=cvLS_rag8YR>UOc`zV_h|Y9GP<D?7}WenVFLDeAOVMn
zk7b9|MDrVG969=<ohcDbWR4)1SP-3Xs;YtzlT+U<vlNx1iC1$#rg2a~OHZaUYPTkm
z(Cu^*VyRm4Pw%-OPF&V7HXACCf7jf9eNGAEm5AG{7!l*Q6bv>6LW4Cz?GOI1G(r52
zJ&sz)Y4b7W#N`vEmUhiwc@p7khVud#M9jcP&_D`oMJyuVvTJxjMnhh#djIfvpU;b%
zfyjouBi0vkfC5-F6bxe-gW4N@mwL_AXU45C^0F<gYDa)U)YWU5C+A)~kFNy>MNgFF
zc>Lg}&4&YS3CRQD`30i<0+BhG32hbrxZ|c?+|_}xoabSgM+>W!b{(XKQE%<{4OY#5
zs<)B)r-diI1GP+FQkW1~DP_UfI?U;>0DdP;sP4F<vQFc53VinHIoeXP<FUR@dzb`t
zomfT@;%^W#K{C*Oud)<r=_RIqSHt-ywtObf{n6^&UZs#h^uj3}uo@7cmO8%NdIusZ
zucz+CR$(x?(Cma^{fq}$BKC!|f^wvSr_-avWx$L?ivKD(clh}`bT({60olAspMf&0
z{+#f-Eij_<TgY#s3Rr7W6Y)4K3K>*sa^@I!1jU=*{@<VjjeZV#4X*s76IW|MEcFti
z{)kFX01bPA87xFZOV23KVVN$s6-h2_c4YVN1c>=P7L1}S2)aV$fey&WE81B{BLLf|
zI?o@E0r|qwY{4o7i5ZdT>5z|yU|NoEh|+!0VS@<V^}!nUxY8~az9R+y>s+z^_hB7$
z)B>jEoab<Wb4xHX^Ges^E^m&+)Qr@PD+-n3Cxc8xBfl|)Tn2|;c`~2JD25MSKdL{S
z-VsThbW%i)5C|%U_}5qEBv6<Dg0XzM{@|7uehJdZsunN{dX>s4=OGFhXQPGNctJuH
z^>K<SmwYTyfM<rOv}9y*j5LM(VAijSdf@WU@Wt?%y14uv<v1*)CAP@&s~Y}k_WBDC
zI0s$L5I#=y`cVrLscnFq9`v<Ga2_a$%b|#)0W^>0Z9#3BUn*mEL~<gz50x|AcOD!l
z(g2bQ`tdacfDe~En*X*Z>ixBr+x+C3R~!7~x<Tl0R(P-%VYf}gsI?sx_8(yI#W%Z*
z^w1I#{z4{*PFj*hdx8bSar;!2McjcZcqmc-FQyEI(Ew$tG4n5I*S+TAj{o2Kqj$eh
zUnRJc$X$xZ*$n)}5_KvS0oqS;1pFr)m4ZPWQ&tKujZjrhfXcYFASZ*C`yNrbRJk4!
zFx1k@U9721^5+j8v<UkH%MRWQNC*=Mqz^khx}Xx6u^rzlXL=g!k02@Dv@T0E8iUGm
zslK^ARWqNviI>$TV$;5rQJ0|V&Wc+_bAcchG@{NX2u<Lwf}HLM3_VS*JdvSQBoX$+
z%euHsCc%H46t`Iaf5Jcz<>L2n)~<re!ocUO3`XMlAY?u3MH?~or6$u^8YJ6rpqY2{
zSS|*8fR8X-<$<^4{g*Q4Rv#rex+AgVhJ8j#-6HSv@70-izJqaD9qV>iLk?98I)`NT
zSCE)R?;oS7ng_?rFUTE1+bjx}XowP+iI<mjd}7Ht7M9?|G|UM0c7L)5_u$7L2N68Z
zg1ypE(7cvO=XL;BY5)%#Z-%SRDx+hEp+_C$^9}Pb;w;_*cm~ky%tP@Yg;1uVLeeC}
zXvzWP7(tmC2djq0aqk}l7)CPI#FPHWM1vR1c5XS$4Yx-(`T{#LaBAf%)C}~v0CB3L
z;`G#_OD#Mdu|&U5S@EQsKZTMhxmKYYHvRGXWk&Y&P$b_<!o>Q5D(6o64*si$UBr|F
z=1N9#8~KS<oMM<o!tfDSo1nl5=ngg%8$N^!j4ybaUsmT=n4)tFC<B!W*2E^5SJD}=
z3XoWP-P9S%vP_oxg+WEmu3={(KAf2Sc46Xub1;4A{8KEtD(5Ua2(fr(@W!^{>M9kw
zAJmmL#+0wk_YJFEp~q~l$pwp8AvDO)A;ja>04`1-hyc0}`2PXOa5m<lkc>u%hFXl&
z1s=5j^8!c${30^ddK+4}fH4%5zeS-%vYGfPcSM??r`CZCbVb1mQBX#69o78W6%xyu
zh>ekPD?A}YD;+BR-CBT)G#NFT?2pPe0tB9P<HrUD<;_T^$FoiR_xq_hUS#A5Syb_>
zi-EluzP*e<z-ZR~$tq})Lc0Z9!^Do%7@e>*V}SjZP!88T7XUbZ-!5n5bq*>McBn|?
zjG%^A<+OoOKtTU2%ExcJ5iBrF)Ln=;(oTc514%OM;{>HBl3u>k5Jt&L6p|sJ+VEX)
zvvf>e)Q3JJ9WdhWXNS@AIpp90uHq(u(%BRJxrnp8T3yc;BgJ=rqckB{LHTV_7Yg-J
z;T41?)y?5iV%6(T8RuUsg=U3vx)D#@@@8E!ADtvzJ9(S%y;ZgP%+a`2*%ZT$=mX|3
zhDXhL>x^TqVuG0ch0H2KV$g1zr|s0okavLB?_mN3zuY>IT>AO_0j~b8C0vU%hP1qx
zDRwr~Lp+eUzoOx&U{sEKq1Vc>=_cHid6fl6h#qjhyJzghzgXWwy|7ai;Uq7$!ncO9
zg%ZPDh++aQ?N%PVe)sQ7hO?#Jh*JS28tgT&i*r6uFVh6;#zLSewC*V+9AUzm*yrR{
z!{@ThaSzYEYH0J)1BCzm{K=S|@EZ_#8TgRdr3*0cUz6xSTTEIDr;yB*NWNY^A2b!S
zmDg(r4w4QU<Gp{cngFZ%URu6#yn{A5cs$m-muLQyrPs_XwI61u3IxW}_&JM&>Lku1
z7l<>4MTq-|{%A`*37gU7YGHYx)Wf34?%5=$69qC4qUOrp>bT)CYO~K<UohSfXQb|l
zR#R~s(I;wTEIN!uORf~4;i!y*(usOVWv@J;MMh#!{FZjrW&G^?X6z~hyZ!uhanp~(
z<fhJXkeN%Rt!aBr>*aO5wTiEo)rK<J>H`}WBw^aZUks|Bp5Ln+cxQw37E9hq-)4D0
z?_=f5lC1sjI?gOn=)8P0;UkLHJTyYK{;aXhz_F4zY-j3!S?dMeU*h{(Y%ciOc17dD
z&e;%>T8dL{`*dKHOEkossYm^tis>`+7&-{rawQ|}%GRHm*Mja6Nm473Py#c=!1U~M
ztOz8!A^<s1!PqJZT_1J2_NrED?N$R>20|V!H?ge)eb)7#@N3z%_pgNuo7<Fu8mFv^
z0?VeYhgcuWAS7F^kQC~9)xyyMvpo8=3f)iCVfsAC6!19BYo?ZO9<<^GD{kRuWWu}A
zn{$uk<XUol>MP%ZDv}~@SKUQ8TlKF%z=6>ykAeu75S7Q%Nh;OtXu-n-u6=D|z&hTN
z;`20PbRXK2hi)2Zr{0PMq{de1)^jxx&DpMsY}`X21J#Npb}uJ><RFFAdgH6P+}K6X
z^?u|iD`oP;+_z0J^dJ{8-xxH2C0dktH0dm6@G|rOF{vV6BoO?&jrlwOq=`}mV0Hfz
zw}j0hdIj>L9cTnlJ3|Hx;Q*1wN^MgSKh_#u17=3T%1O9n<Y=7d<;rjp?N`TIjlnoa
z$d+9aGA!E>A8W=OfO_@w0Fm1OvwVcU?6*84cp={&e84BQp!Y|R$Vb(G{7ZXbq#WQw
zs#8`lJh82$XFXQ|y7!Z8wdQ3aWwVnudk}>Zi)cO;#nnYs>Bsr}>a_e{yL{~odKqfS
zjf7YY4Qv^I@Xv^_Jt7XWL867ZtvQ)!q_Bp?!e*E_1D41-5>lYQ^2R2W=#8HipB(t+
z!HRI)+{i)g-Ade~GDLfr{O5%{fc>xK3xOi?Md~RC8(^n$(9;gc);O9S)BpPGJvz64
zeQixShFW3MMgO8)R}W!g!Wnb!`(qYfz}eW4a>H2rCzi0|6qMLt@4E%mgi<q@OA#@1
zCx*}U%Ll!(KQJZ^R0tT^AzYqrSUXx3wQ9+V8gm+f8{!+jA(;&RV;ozseb<Q+i(RGB
z^9N(gRdpZ}y6Mr3K(E?j2qhnJsvaygCcc%*&4ZL%pXF+O6Su1*IWLjt3-V`2>o*)l
zI!p}my~6lv+&^#5YB|anC{4x#{Q6;z+!M;QOUC&;cssD$VF8Q>ugmSv&&MjK$Cp9J
zn>8I~G=z(rS}SXjoQ4n;<P8^>qny*jms7XNP}lGX2DeH~?zBa7MPhHP-sKRCp!aTs
zWW!q`g)sX&UE1SdohG$RW$p^mL8HL#$5!KzjYjuBpeyi~?X7?dKhOxTzgYztna&@j
zg-#G@>^w!+GV2t$R8uX6ltk2Mi4X)f9kJqk9xr()ITv2E0Fi`lBJ(w5=N+f1+aDCh
zv(7+2psYj*L=rl%9C496VZ_MjINpXiGaWII9QuTTvay1c4>OW@Fy<LWum?kvIE3z!
z^^?nu!G`pVn3=c&6&PH@U1N|I{wp-{{hwVPf2OTpN)Bj(AjW4yi9~7oR%8Wq%NzS0
z$7}*VBLa`(nB^D|zf050f+4Q7kQHHxWx6AT4CzHYqDNRu6yJRadEO#VT&$BkUbQ6#
z2oLpry?3KV!>J{3222$WRx}`PUR4;p8b*`%TgOjMx(q@4eL7>T^MPAk!IMqfOSmPW
z!QG=e9X0T8g92+2>+uQY?)?fw9~+<2o|P$8go7whSZ0(SSdPXV__4NC<<@xz(7{GH
zN?-@)u<ysI#}oL(uC|s6W{L^Axp^-$;T5rEsx{A+XO`j~HoH}wkL0f>P`BXa4XRhs
z*T4;V|FBJk9(JLhkTcoNw>>T)8Rl_wZVm=+$OREeH9}peO$bjiYBlovTvPOOTm`1y
zBg?pjdKx<jZAK(3<T8sWXB>~bwrgv#@tJR?eMtu+0RmS81mNh`a5$Wvn=Hj6XTCXN
zH`|j?BIWNZ&o7~fnA`Kw!DprB2AE6GNISzbE{^=;QGmH>b4u3`DR9yns0i!hxyr|S
zoJWsg)RWYyIyd%DlhTg%!j~%~TD^~qLm@f$J?RyAR+6pf@gV9*LE`3|1|e3NNzOev
zBncSE(@5pCRpjvDz%?@onG-MwpfSVSh+)>PR={m}_(}@d&2p-m$b}HP1rZ@@7vhi@
z(~_c_1w9*gYM8fEZ=>~1P9Bia|7&l){ny@rD7HQq;@vU=$DNEswmpN=OA)&P=;D=d
zMSQmFcEp@a_{1Qac%c&C6{vsVtpQ951OISMUi>KOA-E>6H_j_KA(aNJ$mXs8a>1)H
zSn%JM11g3w#STN=`c|W{v~o@~q%kvU$aLEhKii4;f_R9FA}U@Lg)5~_Dtzn=k|q&Y
zxpQQ~NMYJs?LjfqOZLYy-Pqki5>xGl-`~HZ|H@5O&rLTET_hzd!nqlCvD;FMtxP|z
zkGndP4d01pXMsk-ADbD&{-g70eL~ZvV=#)E_ZPj#kjX3_Z7gDpv4mHBN>)MCZDCmM
z&tqv<DLBx?Qf~jyc>g~8_2VFfoQDSTH=>W3K1`nP8?}7&=4{!bed->}kn#arUF|;i
z`Gu{@TH%0oeIH)bzDqF^AwT5Jb|Gb#fl7SsRtH2n19tWRHLXg{s$7~@RHp^f3t@1A
z@px;f={LW>t*|*j_f|U@Shl)&(eyx6w~a11{qB{U77G^@XnwmF=rt$sokLNWp7yv}
zvqVOes&hG0)$=@3_wDG?-B`a_h-Ml%xMju*4=lJV2y@uY7y9zlTJkVqC_VEyscx<N
z>hq(v`2vH|U#vG9N7eIvSblfHhC6-P*x}sex8hb_+xQ(pD#r|g!Qz#gY=%4x%eISS
zP=p2+Q$woxYVdZ(g<Mk(flgQQR~2ZC6rgB@u{1i2@r;n+l)n0oZGo86Pd$*Z2RAB7
z?#Yj81#0BZxWyCj(BHm_9y-i$3e9H~;US}ohy^2xs=%Hku0g-jozkSz^oxyuLE=Bk
z?4;;_l&&y%2D@@x4Q$t%wwSwkh6qxT<0pJcF(NYhs^GgAP=8nR6I)Y<X=FvKyD-}b
zm{nu?Or@}%8gyY_RR@fIy&W*Rgpoo^xlju6Wt_)QgCPLGmA>Iu4AFW;igZw71*mf-
zI{DCvUUxx<@H$n0kJ@r(%tmi?B_X}6SdcoVd_PuN6$BE<338mmG}>IW7vn(BP0T(!
zROkFmU*HVF2JLx<6iEkIE7Dgts_v7<0Whz3@O7cdx^=nuTS8Uy?$5_EEu)RK6!5`&
z3~8ER_vL&mnElZ1<%l8pKdDN$8JNJHnhG~YyiIMvX@$d9B`F;?n-%xO4Xk?)c_@Vb
zlABQ2)jNa&dJ7SfPHv<Y{7%p6<g=_M3y!Y3V8;_|FJY!7eD28AvuQj&U*Hq9MeETh
z^$Q!81A{3{5H@$nd}-JKI5?Ozw}dnDlwT{H%oPF|N~4vL&Aj_G0zW|R-I#25qCZk+
zVRBK!=FwP$_*oujmJT@ys|utmv<i^`69lL3!X2<dr_S(HQ%}JsbT_R6%ePMp{L)n2
zSP`{Ue3IlYk|wLZRhbxEz85p<(#N?#gItIh-0SCC?-!oX%i8Q}(dLggz^cc~QC)>)
zCx{ZEP#lxb9vrk}SW9Yw(1()W`JzDNj1_s%Xub!&lEB+#!FBuX>qY%b6dkc#mzslo
z0fB9Zo>qrsUY0SPIy!P{l@+DKs(4V-=1^PSJ2UfdDXL@`$n1E!Sg68W&sf)88i@;!
zqMunLSk+-4kJrL=RQ=whLC1?6m*SG@)FTryVLF(niEY(^uGqIoGMH;^N!XM&4-h#m
zJ8_utP@F1aWyXcSU-DL9OC#5`!Q%=3L90O=y8i?u&?S%eh|}A^vW<#9d%V{&$1q-C
z`v*3VD~~CqT(xLHw$<s0xLs)@=^EWjWi3t85f}^CSm(EW>QX0Ju`M$E=gH{SoJ?MI
z)g6IDq7akblLp)!kAYf#Bv^BHG_k|2%egL5fU=xZniDIc7iI{(PKW1PV!KI>g!-W%
zEBnZJF0+o(LIkIg(to!jiCNkCq}F`dqq>VQOvWqE5pH<SjW&wv$dp>qorLYjYfd(e
zf*^wRP&8PR2qmsKt`ARqyt=6_vQ)N@9oggeM6sdldLHJ82#9<z(%)m)!kF)(#9rut
z*$=!bNf5lh;+M)b66)j#x*8aYZSvd-e*`=}d^VfwAx5*`fnD_gk$aBGEF=e-mlT=B
z#>6hBDD`*w!*RdwcrxIAm<sG%FN4qdnn)rj=5AHKFKNHDRC1>BYM(-GqcxF=FFm>@
zDVNRrx6)?@5a{lq`168Mz;`@skR}wHNGnD+ph7H!5OW;DMF^%`SglBq=i&c#EV8V6
zCtKfA5d`^rTRW)E7>sr)q5s^{%6WWAqbah#)OV1-bmar3hmkbEb#J^Bp(lhxICzf{
za>csgvCDiC#q--F6B#n}qYU<Z=Meexa&hm-y>9FNRp)gWFe?0^3q(5-i{0nC@Y?7s
z>6bii6P~F2;qmbu`H2<@?5Xm??&Y6md=k~W)76L=E_R^}M)Lj9K-<LBg8k7}yO11S
z{$CL5Q)^`UbReRu;G$G9E?sP>`D!yMz}%t26oVN|j&&`l1YS1-zU|FOTqWd2R!RT-
zugU4?AJ@n2|AMZoPMDrm%@Q#}77r^|h=$W0wPR`pG`{s-8vTfHD31jrhauYH&L;Bt
z??h_;??mP}+>%V7<^n~}yq%q>5Fg?m4d@D3nraUn3SMjCXU#7U7{8YZ<M)SPUg(gz
zeh-TzaH4h>PE|duT>>_c?cCY@!F5ERA%@+T)o<_QKI-`}!j3>!4{(X|5e2xrRjNh%
z8UP@s@a!!C?k$f3(e~3qq>3hB%^;XEDTtgU#$s^|ZSj|LYxhB?g%sLb`UFuVrh)PJ
zp1j6-Fz$A>ohUbIpc<%X$s}5MUOojVUwo^{&aK8~9j&7dV35)fV*E5<&b0<z93_d*
zckF4-Aw1A09cCX!b&$Cr8w*p5Ai`1L&eDg*30Az07zRRv!~n(-C)Hs-I^6A6E#JcH
zZBSRrTQL6HZMHP2O=4&$_Y7j`WD4#3wwBOob0nc7N-QxLLQfdB-|K?NKZnSlK%d#!
z782xTZ-Wh#xV@eH^=J)$^e2a`EhUzZS@W`+?Rg8gUUyNpb&Meb`f-_FIaPDLMtY(a
z#&HW;?@oJvf^ESCkZI!8E@~)E#&Ki6V!n8Gw&?Ob(cp-U3E=y^C+<xg#nnpGS5KB&
zV8~(nDQb3dyl*82GK#lRU&;=2+nP59?Dn8lOFNfq@o|B><{@ua+KE>(MaZ3)lhCb&
z`cmmA`;}tcJ175m8NjEX0>;oFT2hxcU5*Of{-D=sL)J8K#prnoCW3#q3EzqXX5b<{
zZ)Tp`uyYK<fn9gp&*VAG{M4@Zfn)@Z%Do6EnxFHd_*nMh?KH!?4ddXCXTky}bW`-a
zwULRXVb4t?p-1VFQ5z3f1qTBtIUcTHH3Ys*hAH)7cEXUKho^Y>o$1aj9j*bWZiBIq
z)bepTC3&`RhenC~@bbHS5U)1)J0b@+(fK+Nuq4p;l+Dsk?GPmEjSVRO*;EPreuIgp
zX+N<54ofWFc6w|vVAZ5nScCqPj_?0tOxeR;e{}252TXVwx<}y4%-mDmr3{!m2|n}3
zgo#yUTkulw&M$l3#qy;cOtvx(mpXPssK-!2B5Ym04QPRc^<r2^kqwpPpo~9FC>+^s
z;D}q~VW65JTlpG@fwy65hmgZP?<>e9xDcN4I}#6cC|jANco~zUPX%*4=q5Ct?2Yzr
z##0)q<p(r7$GIo+f59UAvq{rK%%}2Rx&o2+0ZibpyV_^xt@s<Fabvqhjor6Byv$y-
zj@iSG$Zw8TT7IEwk%+ZK=6hgEz%XJdtI^!XMo>>HnWgWo80|EOOD#|&IoS-R67-5C
z@kv}K*E_tRwm*5;!_ipT@i|!Mt=dG)Kzsd)VYzA<vC~2>JcN^vmKg%`9E(=MaLG_G
zbO;BMf>s`eJ%l)I=g3dmmDPj9%GT?iJlG5bl~<_7*Mdp+PHOQ-)T0{{)6Y*m+t{iO
zfE(9|ULNj6_+%T=06^`tZwWpJs2gvUSZ0B@0i<e0nEUvD({#ph9#Ao<@7~TDf*Mlu
z7FEkptbS^^!S8j??mIjPY|^+gpBcdNSQYf}6I|P-7ppQ2*PsEO*P_Gs70mQ=N7#|K
zKDZEWx>XEz9{O*4C*xZ8S{=Xdy{bb#KjzX=UIhTrDbLqyt)Y~d7+G?S{Zo_S+<*pQ
zjriJ?fL4nKi|onFQKt{82nI++5zXo^`aay>NuE|naEk#Aw?t)^UW>o6b1y<omV~?l
zE{>5;$r%md+^Xk|vzrky>5z65EN7ZRZ?<{~%uEZeNElMROScb%ydRi3r>2K4FYhaL
zV=$sfAJHEZfh{j_+t-9D@D3wwApY6=UW{~qpXBsXWM@V$U;B*mkGwDa=2HTsfYeTf
zUkvkCLo3(_k)!d`t%C!w1j`32lHm1vN!@QSu*}BJ8rST3P1VJH9!iI$<(f>FtI#=7
z01CUeN@7CIP{`(5HLs+0ctQeQ)hKYHG7RGSYH;lHQ8sKCRY&ET{F$;7jC9f?QAO-}
zjr>HOG$%-LmgIatmIv%y!c4}bn+DDl15)ZT1kxSK)CVED@zPU%{u%W<;Ch+O9nAa-
zogjgtSd6AzC5<kcFe4)g0~SahpK3Zgw09f&gFGt)eBD{PBIvEQ{k3b7m1`qb0j}Zp
zxpPz-MLIr(By}29!Y(U6azPxNq!-Cg&A1_iW{~<B=ikP&K+XeIeAv`|)^^lp@=(|L
zMYBR0kXbvSNT%&JCsn&iX?e^3`t3P}dn+?GId9k?&}3{f-Y_fGo2?4T9J@QDXMTi<
z%&bU&RMzdhT(mbKl9Ve1lxz%QpaN>0wIAk1j2;&kINgaw9hU#WnZWJQweNd0>iyUI
zjejpokr^!t`Bak5E#$?d7IEGwY?toL=Q=qKPZHLbrYEBLh$ug54`wqt4A|L`$g<qA
z>Zwkepuc<f@ez2l79SrtR}H?&<uGuWz{JEv2Ga=x*H#Eat3&@Fk`Vts0t&3u%>bP4
z+DEoyJ4X9juzMF|kjV1~s8txKdRk%mM8Q@*nPrE5rc}88=*Aog_pOVIfGi+ZdZ=$D
zv2o5b8N__sLJL!0=+~cSF7~tgp%X3<b#lrVYt8VL=}5+5BK>&7#37ipk&_stUhnRt
zSQb0|C^Qw)0pTYWvR$lNf|2+uT|Q&);E9vZ)8vanRYxxtuyABh7oCjK{q4zid4_*y
zkCO18fw7c?C3v?MNMT63@`+b2SWYOE$@O}4qsm3AQ7W8_i#M19cJ;GFRT+(VpiA+@
zKV-wB&jUQJ51#XhgutgaV!T+uzp;3i(y$|Yk9Wzx-)wV`>D)j3`hCfX>@dm$iYpRA
zkKoS=#6^@yz%gnSlTeHGLKt5P73*EPHT%)coU;>}(_5(m9NuI!CTFBtRJ}w9;2aw^
zio$4rmfLsl_XOes;_1lY8e|93YvQ&CZoyNhz0B2bC8gwhLf1N;B9ki%%j*79V?!ze
z;gBYfE)+Kq1^sAU1+xTjrw;=2)}i4oiGN9Yy4a!C53FTl`=ncuGPuo(=9uS{(ptFK
zVf-ChExm#;GRZD0)Xi{+wv}U`YO;3@e7$z%q2f58NFrh30<uNZFBwosA|b+lBVogH
zb{@r^@h&LJC4Q2TY2NI^{YkGA=I0T2M^P}X)S}ZQoktnb40i9AS~7MYSTLj27gU+M
z*?$yk`$3UvLRcLF3!E^E95R_wV&fqJjpu(KgRE6jQ$x%1ebnZz($MQL7PyGOjhXd*
zY_zhrHm3#od6HtXe7@ZFl;_`f-h7eq{^SH5CMxgH$`wUk*u6DaGQi`Tc;-d{0Tl!G
z&e#EyU*84B=Vh3`A36Q3Pf*uJf<+(fJ1M5-CAV;)sWI!b0rHn|8T947XKyJ$T%Dqg
z9yOvs7{Bz8!SY;nYc43;E?kAWxt?YH&#H3{XVO|m@MlOI*DIyVmX(T!jE-JC+0C00
z8S*KJ-kHa+7K~#*M@FjgV1>4|)7^~ef9xUhJ1p=`6B|$g*M0gkx2E2kik1Ok1HUnW
zcT+zp_k2YJMq`R7eek0%6FzbL527LULN)C3>9po9pTA%Lr$4lbW-CSHnU)}3Z01*d
zh`^|T)fF`JS=Jg)q?XPq*Z5wGt`RJI7gGRoE~K-1rskJ3w;BqKBTUFeZ{X;TV_zXB
zXnhSUOj;dDU$fkOor%Uo(?)A`53&m_2_wi$cZt%hlwt#d;Z_4gnGP7Y#&sxp)FprB
zN()+HnI8nHZX!j=UW7kl6~RcI(34(yij9~Qc2iFtB)zu?nzh6k8Wj~Ma{j7bRIN~Z
zxWHm=m)WRtIuoV6e2M2@iBM}THabYR$uG-m{=Nh1*=Ap*jhgLF`!(n5t+t^4Qg`kq
z2vkvC5}vbh9+^*kXVG&Ql8^)lRT3CvXzsrzeMZ6F*DYq+-L2JXpSBj}rp9YYn-soJ
zN8R+*ZWfk3Yq>ASzS~cp`c<n2!(n%&#T6n*ES6@Rsnh)q5KBlIJggTtv+4>YCvB_h
zx!pBQO9+@T<AS8z@crfex_o#y<1D3PZ`h<Oi6q0|ey$o|<`$H6#bS{%j8$Z`jP@v8
z!zU&_DzH3{x3;kLzq=BhxQEQI9%2%oo+d=+8@@}*9{Fr`=NN`CHmgm<y?N<^rO#tC
zt}yFv$h0hp>o7~+oN*=k$c&v0I>qV-9inveK?Y~n`W?~XmzVD*?tG?AO<j2}KZUh!
z>nPevL8Tjd(V$T(@z}t;<s3U(o0XQSv6n>L9#AQ3YEE8lb*?qJ;`zTW8hY%eKi${I
z`&MaQyYKiu)!T1ylAi_NXMOzS#-~J$CU9TtAA4t6<A1m^FiRJWO_g4~+ik4ux2tsB
zl4__ZZtEI@v6hB8D<{L+&vP3L_?-}&S|iD%3%<vBL%y#+*GS-KP!X?W$Q~7!a5crf
znT5;?#oG=V%@j%mHLehy+V|$g0yXt_@IEUAiHb8H!k#2JeTB&Bbrot8w^7paVf}U}
zs!6OIq4+m`BQOnp9zSobDYOcnOQ_A9GC7PK=}`<Q_(B9Vh9L7o27yKxYd5uoCp2I6
z?su3kETLeA_dso`)kqMqAgyIbT*_SDPd^UK<18`a&LDlY6RM@k+;u$v#X!qoqph4>
zZt~aaiLQpii#(%Tna1@>nLCq{CYN&fm7t?>My70bP)?W&7Fgg<48=N@;mUw@)U5I~
z!r0=-=uV4&`SVpjTFQCweJ{Ln^Bd{|!M6;zqabkuuN??tzQwJUr3Rw#xmcgW6w#?P
zT4%k~!}g%d%Nt`X=De|+t$Hi+r%&J5+=cBU&qbJmNj*02Qyy3LH2H~Qva0*33ST=0
zd#6iR_4sHxDy%vy6O;#$+$l35iAc;)oX8ngxRDd*S+Z+YgT<g<w6{HTX-DK|Q+U~v
zu0O+giw?wU9mKdgq6!}`oYlbJKU3{vX5<PR>T`Wv?b(|ptl^{xQ2H?O5F@J<=aVs%
zh#Md^%O<{aW$+aScr!**fv<H-68YLgU8lCN(RKxXooFj2>Yc0-$|haW+iJryXavwT
z!xKhNn=814kcqF8o?Irm2s(dSMKz*I-&oCTp7fM){d5}jn$hHX@RUEP*w8MOH;&I-
z8y;wx<mM(V8-B=3rB~r%6>oX7l*-lneX8-s#z@lsF~rF*DDdXiq;ir!cZ&f%gURQ1
zbr_13H+pd1P79EBcpH;~fe@57(SE%4zt7%z53{Z5h}r-q3Wq$6xbbp-N9Zv6<X25X
zjrWS(IQUV)>DYawAsL6OAS=s?fq`MS(~{e|?G?@7fj9eqUI3!6%QBw}-5ram43?>^
z<wlOfp6<8RE^D>)?K!swV|vUB^X{kJeKI&Xt@`4yia*pMlhcT7`8Z9gwcc8J-h8dI
z<02XyB_(!CO@+IN@Z~_rv_tm#cwGliz*AkWDQcOdbTkkdPM={q<i;@ieN&oH$VMa5
zS40S~!NA=}pT!dxRt8q$X-X~wFdMTEKR|LrU0r4i%V_}owz=Bf({xi^+1zU)@IP&g
zmWKN~!!QEd8no|ve9`cGk-p5R-HeKaSwN&s0~!J>aQBiVL^03^#AE~fsN`F@9gM5(
z)}aapYjQ_xVl==vef5<p@t&#>Q&Dmn=SgI?5j4*JMRy04glhR<Q&oD~?ECe&Ik}hp
z1tyzAHWw{dB%KB3p2{1ea8{a(t(JSx%QPU7(SLDr<>%fbROU6EE7&lR?HhW><%sLh
zTZa8Iac9B@`0SM`F?P8;jWn_K8FiTX>zrApc<o4%btt(v)b?Kn6D|6Q)(pZ;8THj+
z@>+cYo&^?1J{5F6Bi<ug+6%FFH%(Hgzh+Wqbk*0~cHPUif1x#}=}pum63nkJhZdZ~
zh52sG=-j@pOV`Y(^1a<|R;Fl8_MY`7F8HkMU{o@ar46ExG~Yx>Hw;Gg+QH19`Ny6k
z?q5mH1pV&jwfUbuLGyh5i5LJVtVthwKl_zo7kuAuWA(e5n`yU^n$?UF<k93Jql?^r
zSsu;cYFt-SnPh~&m98&B|5k%1$RafQJjIthq-3S1%It>92G*0bCJ3a(FiVyi>QIRI
z{fMR~rj4-%*Jiylzj)nE$7D+Xxs~af2L2EN<ec&*w(m82D4r~{p=E&>q*QCy<lhKn
z;m2g}8eCKfCtKdL;oA!}mvf|&gfq0y2_GdI>bZ4Jo1`TAcim<?H_N-DDJA&w<9p-n
zGcSBfVxc96F-2$nh#HGgo|Y~Y@1)@%EaOt?H1C#c{flR>-GR&)1lSL*%%}>RtL`U%
zX$6_1K&Xz#!|%Lk59kv8zib3*+!sp~s&<+p_Y1|^c7ukudPaWzp2}k@oBSye8aiY8
z83cK`*_>>d&Tz1HlEz|w&VKSRdH^N#ZWu&Bpij#SAe?Iny-K2xn_aCxWjZW-dEPfa
zcv&aXtk<6S#orJzd4Iy}Dh+!w+k@&Zxg&o68J3fSNXVor#iwDbkl$)hyDfK^Y8Ze4
z?ueQ<`H9D2dl*32*gsZu>)G`&V`4yES-hM)jadwW21i%K=I_I(ImPcPYRPO;^qVzF
z*Y((Oyk{FO#aWdzFZ*iP#(uxqy`}L-^<|T!^qp;4)A@Jef90Wv^uG(S|5UCgV2j<W
z({ffmftKAA5I$+?vXiYUR2R0|qG|kwm$B$5JWYS5PL%jLb(Q7j_jk#wMU#Vtk}EcU
zn`IKFArE=p&1twX0br8eQL3y*-6DUd=0i<IjZ5XtVYjsqhvUezx^(2&(w4cY5+}o^
zvNnwFDny5KvdTqB`uEy<UXF`+ob|j7?pFT}0O~6b85Ah`qcmTMi2!gpEv(4%gAP>_
zF&1_G7B*}<s_W5uD)Y;!LRSf$PkXp?x3TS<Z&mY=kU*oeu@s!6+82{otnxFa=lFUv
zStu6;P4c$2!ma{0c(cWboNI>Q&#=XV9@%nj;(2Sbo1pY2OwZj<6K@S35jDC6&_+S$
zXGb^nOM&$Gy~gyI(i#eJajutR!)^BpHo)O;vD6&5{pq%%?-q!&Mr-Hb-<v}ta7xjb
zqg~by{J7U9J~uu)-29$0QB~i6fSP_+1jACI<E9TIG!fHAo9#9ncX7nw>d;K2Y@SDp
zR88R1XKI`sGipqf@{`;xKR@B2us$6$tNPZtXSL%H`dx-f%3(;@?XRqEo9mLo+EuBA
zX63}x%A-@TH$|2!zqjt%i{%+0#PAcxdE2AO4y6h6FVgXB{;XK<D~1k|$!?-_lDwkB
z_MYYEbGVdtql3WHrY{$tyPc)x$@?n6t<m<0@Z{vYf)*G|@5)bzbdu}7&9d^5y~BU_
zF66m@J`#(!a>DZ0Q()8M03mL!-Sq-fM^{%yiO<i($k@-Xb(V^vEW#jal-RE)M)NNg
z0>kctEhQWKAXfo_)%gY~&|_p!<h?QOm^~t~7@nM(y?FAV)o(#XGZN)Wyx}YO<^e$c
zi}YQ>VuWC))O1k9@6CedNP&c$ZGs_Mw(aNbDVj4G4`zS)Ra*eh_!cVpo7boO+I@d-
z_3C64!7{rm_NMC|wT2w5g2nH~Jc#_`p+evTFHFp>8&R1c%nVXb)vqbfboe2_QMI+R
zYlgQ0%FVZn?PR_A^UTAi=E{+FI%~^q95f9@OmPhNhO{hp3(djQvly*QUrh(NG|+88
z)Y56~DH7m=*6c*;b!(C}fR=cP=P^mnH66d85NfZqs?N|{ZOjX2Rc?iN*X>_*g14gR
z4A!QzXGWTZMUJ2_|9rU0j1`+bGPH?Q;Z|J=%RH!+7It`UN|6lwm6h;m4xJj_;Hvad
z`>Q^(R20sB0f(!NuQ>reo0QE>e=w(BF|F9$dg5^GB)4B!S7NMA{y1*>!?pk}3u7|m
z>~?6&mdi2hL~7U8e(?@{POYt^$B;99eXQcTX>yYNhCN1cYj*RnGjl6il&Y#IBxuhb
zs<^6qn{rq+oGto6gddD#h-(@)T%qqutr`*C{9q>6++EWHk9@2MQ1{Dr<+=}hjUdnp
zR%s`gozl&DQRpkKx3Z&5<dPp{ZDm^R?;~#w*-(~wke<RzmMW}_9nrsTWJnn}M1j;>
z?}F1*==`a0g`p8M2KQ97bzAoHU5yNRh=!1J$}<N(79n<Kp||bJ@2a==;^pt0_HpWp
z!NKy1!B4-qkWR;TU=l(+|0}g%Yj|aC`*}z!H3@QSXRJ7B#z^Foa|``_NIJpTYI<C=
z_eGAubec75J#i|@nO_sA)u2DQ+hk(Hy5ygw5~H*GOnwWJ&|xa(*2pGf9F(%yKyxZ|
zPc_8@4Y(Ofzx;epHK{ri|Ab(KBds)XKgl9fb01o3Bnjr}nG$vw)UNLGp{9bWmRwhF
zyqr}@S5rLmX8XsKEfRsfO@&u9AeQLg6^70Q>j$AAo1YHO=Z;tJJohvo1RYQb&ZNks
z$Io|fiz?zfj4?@3ri0K`R)zgGJUU1i|Knxcj^2J?LBU|%DL}3iGbwLZvYxC#5`GHX
zN{~w0i+`#@e@3C)paYYFCf)W#e>#Jo=@$oJ3OmvDJ98`5rTi|_hcSlkByJ8do8k^n
zwB^o@5;N-28`rE;(ELOjMR^5Wvj?VW!zlCd;6<vLnPG!z1r?P=?js8ZGz3-nnhQv=
z8<T}DgY1meQWs3Hk+|z&h)Y!l(zvfnd^6f`wd^ym1h;NEZ6%<c^+tN?kvBqOe;^e?
z_I5z6ZPpTMr5rAz+6$g0A&mD9ET!Zsbf=*M(Ll!?<QE>w;#KR*`k#q5pUqLD=B2^r
zB?C)uEz2&|{_n2tL<|-RVct5=Hm{v#J<mURY28e4RP;<S$*7LGm(v^U<s5URsRTkp
z8%OF~_Vx#1py%zGkhgEbpI_Z3i9PGQ=$r}H_d8fkepSwg$qNPAqe8zDB~vNwPfRk6
z9-ZUrwE_%V?Keh77<9)b$Bjp(GRx2Fd%t?WwjArl9<;SH_1pmuRP2b4nMNx$1T(YN
z*49>6f%%7?rS*_!58k%Gbfxh+>T#LAUx@GQjC|v(N8hVxXCTPY0UxeqRXxF2f6@;A
zd~Z-4sXMf#Q!ydPn0GVl1-5JCX^M8UvMGSc_D%-|Lxq~-g3PF9SN!^==y^#-e)F?<
z0+KE3ji1o*e}0-bS4?<Z4)xE3#2F10H1%w)k@H0gG7!%I8z<m%&<S0rMvUFJx*U8G
zJK;pwJ?z~taI2a1x+ilfOoiTvsC)a{A=FT?HM4Ao>7WA#ax<d^uGpDP<0nOpc55+$
zfE)4xOxPQ0!63L7^i!QSyo@Ri>&p=2fu%-j!>=wFZGeuO2@NZq+tS_nNU$F3FOcgX
zB?5PWJi4CXWsP_on;+t6fx=FRrs6QYFUX0@O1wvod7=5Z=%Qub=Ll_7hOuipu{c-R
z*lTE+<1o~+m1=8O5azD|mkI;E^Xc<Xz2R%hQ5w#6aKoQLhcW~17q#A1SzCJRbHAXr
zRJ2eT`p4?OK1u8@M0H*A_}Kw%wT+cNBk1HP`YYl^OJNd*A!0baYa&Ip7$u_RYr$sL
z*ic+#@#8;&LY$AC$Xl5h6(9Rm4V*ode^R-@u;F5C)W>*;qLebYNWO0B{%J5dHMb3!
zqrqn1be?3GDdr>edo~vIe%jKj^I++H131R>kt2P+ik!T8JS?x^o2UAcCt9Wa`NPS&
z(3$65#FC3}v2wqd|1tIAyEgz?LY~_pjwh1S(&A>pB`DX&5r^qW$a5Ab!or$W!R=NO
z4lQZgZQ$mwz=!nZ-RLVe8!=!u3fX^wNZbiEx{AR0_qAKx;1#7l*OAPc@m*2LF_N?$
z(Z-<g;PzxbY<i^oZ)JHVnWC3VP$G7RId3H4_5pb^eq`?AaSNo1_Z2yv;_Taq(Nd-+
zq#h${vbaQ<Fc7BAFp8d{Lw{qAC19C2-;lZOltF5P{`<=^h@V>#$&K=W$Z%sOZ9&p3
zeSdXu{m~l=kFV|PC02gMJ8vM3L#rj%<aYA_Hn)KbWfszKsC*-SO(#&58&J|xT_`Y-
zb|*P{d|m$3&=@4f(n4fsaXa*0?3Lh)&EftoV;>DXo59qK<o>)#Mclrwd^Qu>-Nl6C
z>^OPJ(uz@*eFU*lBghB@C`liLtf=?5l=%a)NZq-=*9pP$6v{0^{JC;>T8}W&(exUK
zz$3S{z^Pu@%bf4E%A<y!2VDMK*_PCkIj77ywT1uM@?X5#u-5Xq+y5b;Q@3WDshbp-
zR&3Oqp~xa{SogOMAQzvbAGB@|ZFXIG0nX>n(FtKcq*t_ID52{8<(sq^UX(g@(NP_#
zT|Q`G?8<J`!eJ7c{bt}~jH}f@BFIRN3@>jo@8PejzXBsw?4h@%OIxGe7*@UEUO+>e
z|2&jCN)Wwb=$-56|Ca6Z%9d|@noy|7G^b2Gg*Tag-x*F+`nx(cI-PzmNsb9CzhbfP
zGnMJ)$IgY;9W5uhzHpqk6S?c^%8;0VG=-UHq98YnR$B6^4OJ&)hkkHrY`>9s?Y88p
z;}bU{V^Uc*CR<X}q3ZmkhUq5&&ZD;Ty5t%Fu#xAR-^eSac=0yY-H6mT=2zPg3!{sw
zY)CR$9yK7th}Wec8(yh&{ny6|J9Qw7mI*Tlcsgow^m_Sales9;D4yg^jz(seAQ?Z|
zcstm)>kBR%A<e}8uBxwImMgU56>zUQj`H|*Q7>*Z9IL~<=9vRq#1gE0x#}Bqxu0kx
zF2yCUHxc=O4y+U>l+G3`I^#Y!WjAH^glPcX=u+9uW{;G`V9$cvg%0P>=F%AQ{1HgB
zyXVt^<XH_Nak~}SWBhlv`(^Mr$n%C(ekv5|2X4-*(f#)2BgNjMcXlS6NZ_8F#Qf+}
zbzPS831++}9z{fF8x+}d<fWiUdF(s|*`n}c<*E&v-<M`gTsd4ZaZWMehHGKwLyhl%
zht*PK8W=)$^-{7@Hdkp}wXk3JME6nN9I$IN2ADBS0bFm4F8kjNDR(m;RAT1kRFWDJ
zHi;mR08twD3mn)Ox-1(P2OB<m@(G(Bn7&uk8{D%?7tEu2A!`~6E?aEkD*}-vkAha%
zqlz}C7p!3YeM6TnDLhL;S6c6PflxWUQ<(zMdxPObPtA<!Bc;_F&wnk2+?RH`7`a+E
zUC<1>PNENw%%YW+t%;4d$l^Xa%|<d3O3|co$uV+KXEvu>$fwnHZ`d8E+zJpWIC@9T
z$dbwIB=n;WmdLRK#5W?l-qEJwB$d4~(rL2}yih$4KvQh}52wr(QxN&Wj4VA&6E6WD
zKC893de;WS-=7F0Wcjajrh4M;Z#&Uh(L!9S|L>K+hky_>7A-_JsKv^Ede**}W~jE?
zZc22bo=UM^l$HbNe}z|^+i*IpN~@ZHC6~$%U!WxC{0yuZK~8-DeCKFM&fivWD+nNf
zI`HK&f0`FjM2qBDz_yUb7P+cJP1$6zY+B&p=ogE#2Q|elaJlbElQ6IW$*1c9Jn=Uq
z9oI<=eW#wexjmYcO1eY~R0}K++9&*XPkqeNv#~gPMD?jdcB)aMPQkA{IKy2y39NRF
zeDRJbr*J%@{vS=(7+h)BY~P7(8z)XCwkEc1+jcUsZBK056Ki7I#z`jjJ@ZxF|EW5u
z-Tn0LwR-j1;8WY8x}v!P_Kg^3h28mnZUqAHxu{!*;WM@*!zYwf!;@u<xT7hgA8iV0
zYsym3sriP>m;F(pOjn<qt;tmz@t2R}C0GeOUn;(j6KC(c@eamTaB0b^Cc{;%6q$^A
z+;eA5?i+=uF*K4#JzFid^i^fl<bh|yRx$WVdt8I-4b5zQ7w!B5f-xZaAaPLiGSA*`
zi+eaYB`5X}>6f`XvEqvOpTUd|RI)n3-BmGSmCI)I99i%gwVexrL@76NDpM}xs*0pj
zpLikSQ28!r9eKv{1yCmaINFb9BxPrhMJ1PZBreKYI8MuUI0h*d&Ht=6@W&0s;Ieud
zt+5~x@k!d*)ws-DJiSRO$dt5ua=A=rZyfkF?hest|8*MB(JySVWtHCZRlZUu4*pLo
zEjP+UNB*_=OK!x@fwIv+Jll!xx?+XU<-bd>xyoGjv8WIltxG$mRV89G-?WGZ1hxU2
zJmI?aIUjsKDRII2#M9Cfr<wM%5Q!pwg5gtH%7$bFADK?9fwEK0v~yeMp45ntPzja)
z&n+I|Q-WQ-uXp%O-^nRaawP%(lhK!P1^@Xg#G<YI!#H01lVNOqM_b=ZS?%=W{l-bg
z18h)}U+zyCzS`x%nE!rArPbq&m{@M7Rf=-+_HuefeAG+4?Xjs(6SN@1Wv;w~C!`D*
zO)`}f7mg1U?r~<de$3EU<v<_+)5o6eiA@OngJcF#gI>zws%f78<n_{um**28IM|e&
zp_DXSTu^bp%d`R6M%&=%D^TUWqO?Keerb(tq%nmsDP=yLfeaD2<a3cRscv4`g@bD=
zwBy1ltI0KsT=0X5Nr`_fH!VwP{0PR?HE#nNt<J9U5c+HLhs>g|^47H5c|0(s!BLo~
z$XQ;FM0Ju_V(6-<b$S@82#2*_f{Wnl&f8t14K;44JO<Zu3r@f0OP|i<ZHN4Q^NDE#
z-w{uihv>}7ho5df2-P04OT&80KP~SG)*0B5VX?C|O{YUUE`P(5b_U|FS!fykV=hCo
zy-Zj#xmgCCaO-`N{0uAXhv~@Pl+!@=5Zk^Pr(N;)Jv{S`3Fyj-@MZVU${=*6G#?ew
z?Mduy;?+%DvACu%RV%$@Qtox&#P|WFy*EU(Vxe=ie!<Fn{<iqGfF+}6lK%Ml075P!
zwbPds!VAZR(cjM>NB{r`&2G=zy&bo~9}QP~!_a3jyP59%wTaO;&8kJkh`10w1QJSs
z(WU>`;FiR}`r%@J1<h-F{V}^-rpWg`wX4(-k#|lc>u6{G3s_E$S4mHn`a7m&x$n))
zG#iX#Rz-?+?e{L0lyk^-@Efl>GMyJzk4l)!oFr4e$_)%eRGmhh8z*bYivCTaXsIva
z|28>$+j;nH&zbBWOZ6EG>b9Ad*x7yWbasA`V>r7v_SDqqaF1ef*Y&>A>ia*ap?&z<
zcRXMvB_(<OjimD2uPC#$u{%nXAf_Pa@0oP|$C8t`ZH>IjW#edTj=8PqDTtXtDtz5p
zeF8t<B1J(S>nC4Sw(8SpMq9<T>^R7)nUozB{P(i6<gR?)(2{V{%S@ZmU+6Wbhu*S}
z(3BV@`wuoD#^%_hw<zE8!y~CkoY$A=bY$?oOk7ItB8Fr~Fe!`5o%5HQ9Br|DXv1&9
z+uTv(w8B?%QEzS<;H8n^zcCMzQC9d)zwt#X<B;2ua9*~q{4hOe%BXQ0lj5e%6xla!
z{Xufd+0HwI-x_NmD+mIk0Y%LMW1K_=u|^et<{ura93YuxN+)FZZ2DGSHdelmR6Shf
ztTAMPo=S+{wNGuQT$*|4$w|*(zzj+Mf}$ZsTxqUaQ!*#O>Q{azVo83rc!*z97dd<(
zHf~)R&UdW!lZcD-eb#6y!m?T_lI^R^SZk9%(cj2?<QlYy&TdWZTvHZvJR)X#>W!aM
zJ&wsvN(tt(<+;vmD=0?f{}h*|)xPx--SJ<I!+5fTHiP8$z|h`X)bG*77eKyha*4ds
z3e@j_Y;vwr>8CkS_`#5cmk%%1^F}~R6Iz~J2lzWum!R8IXFFS}8+j;o`EUUV)<0el
zm4<9Vz)ec1ApFYRs#b>0i?pwfX*wBd1+Uc^Xm%#gIAU@eOtsulS2`u7o<=AEk$Edk
z6Y371iWQqNLlrXpEofUY&TQl4$vMsmMoOJ2o0eIdR*6+c-jeC^k@Hb^RQ^W#@1sBg
zeJRCX50`Iq_41Vx$7Ke+3H+&c5U;Z*D*vNN7eYY^NL^}mTV?PO_3<X13EBYeewgMR
zE!PSb<a>pw>v_b(t*EeFYt3CeY+w0G?|V11?>(c`_dUYY_j;kQ?|n%doLs%|{yR=Z
zn0Pl|f_)cHmG`@3SWw>JYkeh?_lb|l?-2ES*ZsVJ*#?|EVS!^zXRG78rxq3@?&!>8
zw3__|UprI?ul%n*)h_`1?(ucWK|8QRF7BtN>3^JS7yBwQ-mQa;k+!;FT0MkrV{D70
z3(Ts5n6X##@2mFgDT#*y^y>Giyo?0H>@HDW7CM+dl{%SxY`Gw|LjJf|&QB*~!@q$*
zZ$-;2#wfBCPb}5b=*6elOm{e^hOrRnx3Y%fwg_3y`R=Popc82LYu9<$Zbnvy%sHpG
z=OcEfuNzUMjX>gtts~aB0}WY-vI0%ezAI*@uQPc@kMA|*GIgu2p+5drt;ieuH4ig`
z5K(l;J0faNaJ11*Rc^AOM-Kje=EC$e*1))#fJ?XE%hxpKI9OEU^V`FEKG!ti=x`%U
z-<!lh-4|JH>Uteecf_({K=X}lj$L6-|AwafP7oLQT5Q8EKv1!Z(zHTr*b@EwF#XP<
zDm<5362Fm_yr4WV$>GsX5is7+*g{MdTD_a+3V9s(0D_L*<+Bm#%GBINx8FP84!1S_
z0WN0kAm;5&<zalC$^=04@PSJvFGsp4a2Hl+Z3HUosum#jA~ac|%fuYtRht)nc=p-5
zS$nM8+kz$)YruQSL7PM%(jW^sp<FsNRNNaXX6WY*C;CS2Zmd%hzQ6oRHMYbW&P*wo
zx~v#OGuY|jaDmjd)38PKYb$CNHHms!-j(PdD}a&%4^!as^Y3<L>idm^;C@s7C!2|Q
zynD#>HB2aZ_v^Y1PTKHC-}{kK;sPZvU3AxVXEY|Tce0MyJEE$#_t6oU)Y!=Kc^luy
zOWNq*f0rFDCZ9c5Byqb1^bKe9d#LKZ9dX9`&eJ!!18EWV%bo`;<EXt}ivLlzUb_m^
zK{Ja8-B2Rp>5#tV-M-$1aj(UEY3oh&cDhSQCdekslws5x$~shXGsMiNimnLA9bWAm
zWchNhFA0`5b`-ISECARYol^zf6+ix-g>HRboy^!5*YaPg9#+pvuxeKrR^`oPH=qem
zdN8qRiiI^b(K<VCOWB40L!76rsi{JV%d)|?sI7dW_a-F_4}#079ENLUk}_l@Vv3O)
zf0wf=;*MTRKir4+Do}Zl=g{V@?BjQpF_qCb<HVjDnD8bK)UX>vSZwitD;vxDMpVw3
ziJO9NC}+@|#;FdmKB*7Ojv~rTw5{kHcO2T|w^y^ljGHo`Wf{q>2hd120AtB(<21W(
z#CA2@<{eZ1p(sZ!2smHofyt})ty`uy?C=qF<4GX#sBfrzn{bAxv92au7v~aL=EUW3
zbGYZ;W-M?U`bwVNS()hiP7u0oC*I!?9w*Ur_!2d({ZV8dP0?9muWPT&!cGSgG?#R(
zcQ-(P`m}>AK_nf`+7M7++)<Vi5(j(Q<_UIskqS27yb@S$C43_yk4{vccB*<pW`l^Y
zymrx2TE-o;Y<>q!kdgu5LCIoB^F85hneme;YBV!Hm)gB6(5olP#dY{Xc`jud<iEQw
zT=fqtK$L+O2q6C82@(=`yp3lquK)S<Iql)K4-DWvmbc$(%zU_5;SM^A$z+XxeQS3*
zz+rC;Q3KYM*7n8~27#TtG|nrx0yl?)I$Xx5r?{;<B9lSF$G-Q!(Y~)u`2|dZ0rkXb
ztnzHuifaU}+Z?QG;a$*gf-4bTW07k(a?k{O_DovkYsM#tCx}N;xMJ@sCRG(S@d?SN
zM#p3H7W98gonu;gU|P8_vx$lDHwtHL>EogG%plE%e9K-dE13_6lGNg*m47|;Fuo4t
zWx2bnwi>PTj{`I@8I0ZTO+pw1qpD5PmKAxO3Am+1NU%hwQCnNUg!+CUqia7<RZrCT
z@~DrHuiNTJ$C49V`8bTbV5aLG*$2{OU$QxSvG8Q8cEozKqE6QS&``*2F;tO=!CodZ
z6wxBrnb46!XwOtn83GOAWC_#Or_ivZA1!rR?MJFXi;$0K(qC)WnmN#$v@bND{@lmn
z4@fb~S<`trJ*qquuthC!MfNjJq0?hwWpkW;n_a)N&&79M;a=0yFL9xrMQ|G;(~A%@
zw;yVF>)Nz$2@Vf+$$%e_P1bbkI_Ty;au$Nqw#6YBhnEeW$J=`_u@?C34qy{*f~2+%
z4#VozaZuH9;`9s7m-r<=@FwclJk2)`0i>`Xs`^I(gYC`9QCA23GvA)5AaF$7`L|?&
z$TFAa&WOpR$ITUf>48fwn9ZyIPl*M^O+0-uVeZmV^_(1lM>-#X#n9&Q|Fr<P(Kv9(
zVQGe|sQ5DbR3@cQ;agbuVC)q6<$~kWHy7EKR_BsZYt)fi7QPuh_K1Bea*g)JIF$Y6
z+?KtQ|6@9zP(7Bx6Y6|lCNJ0e^7}y5Nj|ich_{tcT+EE31bQM<zzt3SILA*Aj)`1G
z()SyceWgz41Sa?K-uLaXeacEo3N|xS!Xl`otW1TaJ7Y?xZEk%Nfr{Go)mNZwUVhdi
zD_E(zJg}P9l*&iZr!m_FS`nU#H@-hgBKc*J;=pkd?Zqm0C)^TxFEwCbwPgVI{?*}`
zd%gUwD<*+3Vu?~eGDiTR$(m}tUMr*!%8h}viyJ=EAKCv2*4*athl~Wq785dbwVSE2
z`>ZD1XtXXuW>f<Rir&*(7?^r|UD8mYAP4~d_*F$-=}5(uDpRL&t*cu8_i);6DNNMo
zXVrDUQ!mA6XhMGtFhp~q-Tsb~?lZA{QGng{cc{|2o-D?lGhubR-{dXY)5eGW%V$G&
z<k*j^LqZszXUi+2-y4bJZ0OHZ4tK&#NmBlrU~-@>zCN8h+2cK>x`fN7{X@c04Jt+J
zom}ydY#Fwv?Ey=*WwuJo8ePitMIn)!cN(`}C0FZzHX1gs-(NVqr{w8*`^Rcq*{|Rx
zWF;rWCVASdbpZ}6+oAb`aO@^*8XzWrC|_Y(T#mt&RvlER41YCT|Ddxz(#G<3Px{X3
z*wUH=gmN&BWrmPx&}{3Ayn@!r;Gs8DBzV&*YJ%}pqdIFR`FM5-2T?lFBc@HhE|Bx6
zahk20G`Lv9+*0&7YkJOUzXZuTz{g&acYmALd)pY<1XUJb$hsfnT4ebeWW>ulqc?+I
zPzS&!{;v}|!2khKL%yd?m-p*LiT2NRn~uE^`~zMOf`a;zj`5eiU%yb=E>3@?+o@$I
z;N&L&H`-QS1hE#l9e01-xtP0i<k_de11OV{a|dhrePE$P_>um?Gj>b5$+@M^rqRnZ
zH#eiiBIcZ@vW5~;R@+;~$CsHUC0~m0_J1Y>$L_}<tDUXx<ZF(dDVcL#w)w<Z5yK6w
zbJASXtORYzRM%i}CkTnD(aA3;Kqi6XeFL7M@-!v!wO`>Q=q#lGEUJp9ZH>C>qH|xe
z(^EsLJA`=o)k;9+pY7L3UsGBwZ(zS~#q^87XyNfs+^dUYjJx?#=?NR{-`rw&@@_mz
z$ra=*?cR^GDt`qW9O{l*YFhp{0#DSRc#FqtLL8Nb-BMO6`0lRt!}Pen)m6QPkgTOo
z$8Dv`9~ErftzRkA$X?`wc6pp)xwP)Xl!GRF<C7BLlET)sfq<ub*hyhJt)DiJ60^?Y
z50w+DQ+TkCMfrx%WNN%PnG0QcJD{(3HvjAthLx-}m`x)&8OOvxYm!jr)zjGAPa?$d
z8R8A?QAGTNk|%!qZwVdcJX%zvzXikU6EG>J;sRktW)CzUb$|7TgF?Et;xxYWqO+gc
z$%Y765FW>mw|S?qFhZbDLhhb^adRN8`qmmo`L}BY0dB!hmrbd86!qW0ZA}6u7d)ag
zO}+Q=p>FTa+v}#^;K_OvftBd8QB3{?r=O;Ul~s@+j==vq;kmqGD<A|zY<P+0C!Y5V
zC(F~vslc{gN8`rL@^9F)JZ{&)!cLmY=I45o5iJ^jRVNKs<1yMKMo43&;vbcIWL}p+
z=u5#lXuu*Z{S_jI24kV99l75w-tukkM`tOdXP0#sj$k*ga5+)M35G3Au9gX?ON{8r
zaEdPCxv75?Km@IEyZL4i$Ak6fvpt~#aGtpuXYZTpTxR0wad_L*;aA?)pL6~e`M&8t
z5XhXloH>)F*j(<LWkC;b=T@f?2{m50b(b#7tL9$RN>$VgDq1~tdCQ6zbr-AqN86Rh
zgJ1=@J^BgTf;Bc~4tSQ|${HUW&CD6=7c2n*(&=W^rEt}T<l^+JR-trHR=G&x7~Cx=
zHJ0lY8&dykE7_Ai@D!*QV8pw154`iSAI2B8BFEfW(6P?^k$V?h4P3vQfF9-CdZbma
z?}v5t#NJ%8GYinKOb33AB5ZG~55ImqG>00?GLD;HXH*xVJ41dDt>s6(y)&Y|k7AQx
zxv%dMpfCJzraH*`qB#tSLEnDpT!x|ZSTxk8^p772a_VUJMyHz}JYCwDYpeUAY_=Ue
z<OxU{dQZ&!H|GF_nmZ7MXXYD`&t;wOy$SJX&122J&Ezrz0u3$C`x|H9C%ZdvYji<D
zK+0K}m|@8~k>6cXH1cacTWw^;N}Ip$8Qx=O>S0Ci<i3}uyPJM)adY6Q_o*juM|Kv{
z=MP=h`oZW@jK)Vr8Lctjs-mtZwm%eT9u|1sA`|PQ=H&dV_v7Y?cKn2_-|tnv&7j*D
z3||_5Rc|SZLIyz%|0U+htaQw$A&6x{2TjPL$t)$s7sL!KG?;+?^%D;5TU(2_b54g}
z>(p*68pg-JrN3xo?-H}m-IAOmGB?@UGIUO`i*g2R)KxhX<brd(iG((v&qzqbXc8{$
ziX2?Gipz|)wAS`9&Ni}B6CInE0s@4_2}6@|BIi8oR~UG51u`PNsBwq&@;J0B2`9C?
zG*T8U{z4#Ss!-S<<xLQQ0*rGS!zfEi695R<K0cE3qn}z`dG<_7GtH32OBy+S?@#x%
zVpB)+0oG*{xSKkVjigj5&3Cpj%}Cuk4tj}n*m|RSw}aQAiQGj$l7nNu^=<~-)xiqB
z?VPc%)hVenl^H74V5K+It7i13c9i<CSNzMs|bFSdow$_1xOoZL&FvXPToIye#x
zUTW$!B?VT?#Q9|WukAbm6F%jG<|+oy@`Kg)m(0-pr&4@Ah<AA@Nq3%0Dha1!zxq*N
zHRM2VTQEQChOn{F$NICJ_7+x^31^6{GvG_pLd^X4S3ktaL99Y^(cHD2nH#@27lFIK
zdxv-HWcTZNRhM6HZ$TpQClJ|3SAQ1sjkQD)&A(1uf?8bPd%%g#WB@1_x274E#E(Yd
zL2rHBv2=27i^58ew?2@v?9NmL=$X+`)ZY>6E?x#u?2epV7f75-?W$7{ea*jgw^ikV
z13#lFF#A4#b&j0B>w91Y;<os}d|utPA8&?v^jmmStiypsw=&wlMTDY=(`i)_(wKh5
zh51GO%^-|%Ra3!3N<deXuNbpZNjRq3RkDoN;=}9tU2&|7jW1n<`J<5xVRCxNLzLL;
zV^=(8qNaUBkD%_fEm+v?YWk~{ci`so(+6SkJ#wrAohfU=C*ZS9Hl=&ca@eJl?<&&?
zJ@0JAghK;O{{_3y=F=anM^H#Pq^1*e@(37cd~nzb(?$|v<k0KvCq#=#a+0ND{@!39
zK@>Cf+pARXf9kT(tWjn|^#gXG2P@)LTvy;ga68Gc9aCyv@_ho|!qf-S0;|Q4EKuPi
zcr2?d!p;xGiDrM4I7LApF8VgpaPH@MgXjF_yBG`^<RRR{soBd&_MP1T7G28pHyzSR
zqSqd0Vr$fMrvO7J;8~%Tdm{g(#4F>!jA%bf)5#efzd5g}n@2?|y5RBHdWamfQN{>m
z$mZk~>}xWH%<sNPF3*}ECh}YK`c6SH%#6Fy=fJop84?mws}*Wzch$@W`b%VDRO~)B
zxsq~f`xz}YH8Ca(cW*Cut=f{%U}$6><Lnim{k3p+M&v4N%z~g=Pqs|p7F#NiIdvpX
zGrqNA<hpJ4sRm$;D^H@*bHo|!Et>97vqv>!Tin<iyLk`s`Ivt~?1E6^2rqu{?XZ~7
zw2f2$0hXW|*zJ{Snk26hFiqYtxTK=zgR62YLodbM>i0|Ks0v;0mb#G&_vueE?-Pj3
zT$s_55$BE%jMyo<8P=jLnBNH&jA~}VrW6D(Q+(IvX-UxFRI6`o4lEmBO`w1uRM%u6
z)Jkl3wMLcd6*=*YK0P0etn~}Kkr4#2Tr^e7cfyFJ<rbr5$Ao<9sqN)GIsM5Z#BqD`
zcE7KuPe}n|d($p(Mb}MM9c|RZ=Lud-J2_~z={aLQG9>Fj%l$9!Ul5>>N@?6W!t-@A
z1~3=kwU`9*O@X^`LaEEpS{+qL9TjO>N$Jfo^u@)p)z`HZTZUrm^#i1{BAQW_=d8b!
zHrp+>FZpbV0=y6d^z;<mI{L{1C<XJp{*sl+F`<#-MOz4U@%u@Rqye&YA>VCP(CgTy
z9N}m$MZ@|t+W&#kAQyB}S0oe`R#qtmDYL2bC7coHkFHCEdPAnecU>*5;Yr567+=v6
zQGfq$!^4BCAc1}j!<1=pnHz4xR+op0y87`upV>spGmo82R>MI~ew+Q8u$~?RNTsQl
zR#jVTZlHzBS}IRLQ3G6Z3{sJ(?1N66Vyj=b&dpYmQU@u15ntyN9=j^+kV(7LEOcAd
znwg5Yu1>MFu@YMpP`y8)%e5#hSsg}LKLiQj`$`QFq#`|BW$tqWty3*|_X7N|C%J2$
zP)r2??!8YMDM7I19R~_ukN=QYgQY`10+f)F(iUu4_dMTtmGeoc$xU}h@(25sYHL`|
z?-;=3CR<kqq=}1wX9E#gs|k0~#=$+Yx1$)tOpC*SGQGv8DdW@mF`Rkaw7I1BSM#~D
za<n7kJ6|Fvv8y+ORO{PMR!%-L#U-ocQd0yG5ZnP#t_O9j0esevmiOQ`1h-8ZnrPD6
z0w6q7<`oZW;eR;3GB*S<?8=!>Z9{={decX~p6X<MUzhJfyq~-i_#Iz6*%6_L7&F+O
zz)9=}JIJ)a!&_ldVWj&8{={Df3x>}X)&iCa>6VCL3Tw3?MdIBjeURgR)Y7Zka}#Z!
ze+8hebrRW!Gix>)lJdWAh{}5lKi#7J>pNhkNkGjz$)@k!)E|BzpJ|5;N9xAaUc)Hy
zK_E1->0Du`?`^cf(7}r#3#<91W~p5E_5EXWm2h-)*Q0@Un?HlkBk-hfI9f_dj{n_P
zUx;*}On&d9#b&L^=SfNUB+6sQ?X%Bu9Idgb5$bJ^yN%fdL@SHeZiFxC$7^$$)yKnt
z_96~kDw-u+wTqbD32g<8_wD^r5I6=vdJ|GBqL|JPOK*U^2E6ZkkP=&KD6AOb2r~ai
z^5-mXyrKIdk^dJ_qFPKCi}8kmOVEOR06XX`<qL>4<CU85Sr=ivhZ>aBlJ1n?C`sQ`
zpr`wl^~$-6>dAx-V~kqU&u9XoGs)*O6xpI%eDI$K$bp6v379ubHma=$L5?^JqWEnP
zP5y#DRqNF3tiMZ0JRY%|uAVL@X2T8E2Ryz^yz9L2a-%IN?sB4%PS>fJCYQNZ!0DE8
zMVlZ`7QqgKq%4!`uGAQ*xk=&OF3}{ouH4khste<E&Q(L;K!Sl^0q7e|QDbyZb4>5F
z+Lh=z7e7sZ%bg(NFG?(*I3*s=l`kPyTh?n#vNcP123R1{jKczaMp+LB7|Dlu<R{-H
z9k2I2=*B8zj563-J$6{mMPWu|2XgJI(0tDLFk^7u4{G!X%I(CLJHIawvA-<!ak3uQ
zcJT8=f<6M!De<6c6nq;<C?q?XBeYqmb<dlc9^ZSvV?0%!JDXSiuHP9#mG7AZ^gZE?
z!DEktgva6v&;h+jgUomH^X@L!!=E7j0=z0F=$)^&=U7|4PH(j5X<O#@W-ol_D>L9D
z`s;0NMiLI-*k34C*1eLpi&5l#^QU|1tltbOJCXkz^FXVbi#gc^65A{)^J>r6e{;=K
zr*E7;iZc<lR(2V3erL#HD@#L3JAW59;g(h(H!5&CJM3od>>0Vn9Br-Qhe|pIF$Dw{
z-Mk5WQWlOjpt-~?)5b+a_Z{^}p2r-%Os?&x$k?ZBO0`irdff>BXVN$3IPF{6;uOAW
zN9gtslKi|Mc`FpG9<ESY;{w(?Zs$C*wAT2)4kJu4H2&K;+A$eDMtnbP@ZkF4oqj7g
z)EdlSOXFZZk&v>`SYrXkZUZ={*b%MUj9O9+oYM!YI!k}Dr4$n^b~F1mqPQz`Koz+|
zH?S*=BwNJMxMyj4W2}Ts9^4`S`{eFj=*b)eBlO$i4?FW�`evtVr{IhQm`L+50&2
zVigiag|Es>_uSNIHg%;rd^{4;*NqfLiacM-!ArTq`^@%-?bPfks|}=y$#wZpJPKg2
z_Sfgk7$x<2xVN|1p?>ysKxsXYN~Q59i}3IXrj^?xH5lu#%ald9|BE{bmk)@*MnQ=l
z$;pioD-wOs#O|Lz4#RiY2gUD11DSsa)TGaZM4~(T@Kx(Ig@qlPUw0oMrO|ghmj}L%
zNeE8-J#zTFPsiBg*bg=mpZ`P02OodA&f_)o^Je?Y_`2IG*zb7;4)IC}phdQ{626^~
z#8*!nBA<jy&Z4EUyt^66(FWpoI|=+wY^J4m))V`?tV3M;MlcW3oL_R=OgPkw%<^-I
zyw%$=>E`F2ASob=ZhxUWVx7T&^>+lWX^Y*A*BPn;UBQf#jPpkYxzacM0w4a#Nq8Hy
zt8Q|h8dU5byeT8`usYBB)3uyOw%xSjWWAr&M#?c`R&&Fduj#1Ax%Ko_ijEs7Qi3m)
z;)wL53K=kp3WjU{oc0NW=34d!dAxba*OUz7&)?tC_KJPU9Qb~ggeThK>*%zn(U)m2
zv$%~?uJr7(dRhb=Hb<`i?^In>(}9E3J0d}}=!syLb1o5?w9#P}2(NIAa;b@_?UGh9
zGOZi&Gj8uT20Dt+`S=^vlP774^i4CcH9?;Wg_FkZRj74KAsTITJn1<}eo>-`^b4-q
zTmWxx&bR0@%FqJ8KQh&Y+tR=O$Xc}J#RXov3WDtB*gSGND^>ni$mJnG_cfTavkR`8
zvhes=;@H-z_PSP>Am}^H<g!8w@OgJi>w76RnYMabdC!mE>Yw7=iK@c~e#YFpe%`0|
z-E@Tq3<`eUdPv+XnufwNb|&y%w+C0%f~;VWKeP@!7#&VRW1*GIGUFL&_l{qNPfpYF
zx>61)c%8Il(pL5HkER|;Trh{{(3J}x2yEe<>&FY3%ac`@QZfhQcls_}jd44Z2{$b+
z2^URrRtG=?ONz-^NSYsUkum%#n1u&ff*u7`g{F7H7=EAdI8LDYe1MNZPj5pIkiR7l
z5I>P51?g)&ZrRNb@t;Ck1Ef@r!6(*tOk?&`oGDBR*xYYzQPtz}RTD+<=nTI0sip?N
zH#|f<RpLBuVj=vNh|{nDBXwJ0!S-R;@(RcUrN5Mga_tE0MeR@O7yk>NMFYcD-{+!z
z%;}Y`p&eDE`|{afy{V?bx}=LF)f%topnMiRvHN19BYGpd?3C-P;l{Ky@d(k(vEgDx
zRug?BMjyNy=E`LMSi~%4Z}mHPU)ffsM9XL|X!qykaKuqokgrYd{-L&Tp(;Z+L3Ns8
z{X&$2b4IMJLBZhX+)HN5$&^)OkH@!vqzY*AKoT-7^cyq}KjDiD=YQ!vc-+W^o~4t)
z%uXp5?fsit#}&W*ot*3AvdXT60VmqxJo5G$^mI+|mzOhP5*ew{T;Zf`wE1qbgi`_=
z$A#DsP=jLPV#DkD!|xAHp7-*I_9G>?6oN!H-Q!P+Kt?^)MG#1aNtByG-vABkCpYe(
zxKZIUZ|($t(B(l&1Hmv)(A*7+E4d{B7)9UF$njVcycafZ0k}#TQ?+fZxOHj;VR#MN
zmGRm}MxF6m6u8yvYS6^jgH}RP>S*ZUaM&T&iAu`RH+cb19qms4JMB=u$76jC4T<_O
zv%18<D6!f39G{ZTWv2Mv&R|A;0sC2`j?!`Jq>z~CFIiu#t388xbE<6Q%phPTF~Ng3
zZ|VNMN%M&aCE%H_JPI@fCa)|^DrnvxzMiq3)!g_&!uU}vPG+tdQk2Wlp0;LlvCh^v
z5k&gE7+i_3QfTO<lz4zlaA_u}p(FSs;ooJrOHUpmFVvPJnV|$@eS%ctCL}TF?4_FY
zG1@>rSCp~G0ho$4-T6_`E5xb3b^Af3zCiH$+k3dx^u@=NBcSQnb=zfqL<A}>hF(jR
zEwtPgCNb~yUnlq3><fK_C`5@>T5l?i>LPBo|LXYbU6qVZ5_j;|!>QrNZG8-m_8x?3
zdbu<j%rvC2gJ4YY!`>UU6iyAMC<kA(&-pES*0dPHLSY3%4E-n*L!@e-yZEa&_KC%f
zM@5KiTX-()LFPHWcpNH)dM_%2M>9E|1)#HIuN(EU+O~ulM<@*-k>OUw@^U6%ug?XR
zPHQgkF90~3-91I~k>IQ21xhkUl6TTwl`b!>l|l@z>rX$mPyBu{XlVCC(N7B7*Ihh5
zl=5?0BXie6N8yq8J=)zrWHa|fQbQ&uf?I8Ur^4~X33PQGMi$S$JA+lb3QES{aa#aj
zL=%4U&uw(k%GSYy_UT}W91qCEi&BRQQeF}IvS;*J<o`tfukT$(6+Uq~g@LlNu_-Dl
z3f80#n*maxNRrSn-^M`fLj<AsUwf$U%p3WjCGWc>r#RZ)vba04?pgGPV#1PHAdosv
z27~q&&z&63BN|G&kIGMKrp@2tVo7(Cjf*8oE9K&oOEyqhpI;1CZbpusB|MuQE^RnM
zJ-tM@Ij4xnvA^T{JX1co?q)gBd7oJ`8TaUPnURjefqc$?po6vjaBOAPp`E>u@zI#_
zf=H?^0Q&YAz~#zQdfH&_+<Y$@kTv&Z3$tERzbIRIrQN+_dVj4m@^<_Fx0Sp_4NdzN
zE0HrG_EJiW(BoS#{I${RjVmQa>CJJgwd`6U6T@#am8f=GZGp9MKVJENQ9(fvHRjix
z!0F^JkMF9cF3<fgB$q}EV7T7=?nWFHt>G$8P{2BhN#O?Ug8{C^m*9KWeUM&9rsv=K
zq2v&zG#kBZbtDlImM$RQjYg8=dL`+gEijPHJ$uaCe`K9rn@MUu>~$UDi_U!E0c&@J
zHsZXGrN1A?kDJdu`=UWN@VERA4-gJ5mE%8NZ%sxs_KuCnWpgR2k9>Z-S=rf{jLnio
zMTIFUD-XPT-L)Iu8Y>M`bDfl1A`CBa(QUhsQaIZ3qBXTZ(QQUy!WU}TL$(FX63>P`
zMl5vx_0&un(6yt}=v=9%$Ptk~C)nOj8K~@O(X2ezWamDCf|Qlo%Umk?LqBf1efnpi
zb!&b0%<_PdwDTAi&_V@|hb{|lcZsG;ew@<^S~Y;^87<!k2ju@`&jgsTt36C6wpL*k
zx6*#UukL42OR2Q3zELjzeTK<XVn}}C)-7y%<PuWp&ebnsx`vurqly1|oE{mQF{;0#
zI)Qh91Un?#>obT!i3nQI0ab%dQNS78G(DF1q<^Uyt}P$+a5A;v{b!O9i`~$;+OG$X
z%~q=<vLcg4O7)>omGGqe^xzy<+w7UwM^8Q5ACY^hQ=SavR%CBU!xp|MsFG4fMJ!?b
z;^>r=3^jzUk(W}4KglLBET~v($c240sW@316eg*BOw*=`2Itp;KTiL51?Y;x<VqOY
z;48d;0N+nX1<eoA7{b2{6;|2R%%(aYrU#D;paZ5_R`e$0{qIfC(4agZb9Oh49(u(;
zZy)XHDYjjdKzl*fK+UnJn4E;bIn}r<c3~V8Uca?>OYJ9qIy)b%wdv6o)oB$EX6^ME
z>J@v8;o&b;L<)@ML_yUStEWChsG{l(ozXu@F$>m47hR{5nu#37_-WkkLtBsA&Ghsl
z5)4^EHlKRCXeW$;yE#92BK12y*niq(S~6hX;${$%$W2G*cgd5Bo8azY1Ib%Y>;PR^
z!Vj(tAoT`4x#(f6fiu%Bi*+_@j!co@uCcAqgcct4zy6eiOiIJC)GAU_)y_A)xvh85
zEK;Q04YPrdBTIcDG$z!FYU`X*>KigKf581O0Rf`)D6(<MwEoUhMR&w(k=s?1Xtmaw
z8ofCyskDRjf<rk`H*CYu4N~Td82hf>%Y6?TPgC$orW_X!5uf|wwozbO$K3$h2`Py4
z08M2@K*#>iv+}kfen(0)p1HC(q)Whwid-UY;NX@^*aCo7(At_-mj8mSAONxqd^#;s
zU0J-j2bo$<Uj`fA{9S2#ycwGKNLle9z}lF|=6Rli*LhxkNFJBwSvOa}XGW&+qgF3}
zXogOIgKtIgI=UAo@-)JqM=dN4k;)<<ogo~sX=aw#B1#ufi}kMgseZT05!TADWa6{h
z5XD{7!=r}WEAtsG`KW0O<d0M8Bi7Znkic26j|e|}+{^RBzsBVxr8;^TB>$(aPx*tY
zS-SQ_(riXnf^|R1+;Hc(Z9g@%SWBW~XC^dF|L%9Dw5}l=JnOO}#HbTrO!Va!)7g`<
z_@3FIJ8a#QkU7YrQ1#LN53*n;7SnOzFSdz2AKmF>cAw19*JnvavE6vBaOlSmL&{Co
zkICLMwUVSxmSQlm`Q5aRMVw<zE~hBXt?mJoNpauruIw@jgumyiFJ7^<if_)Ny4mk!
zEwSBq<c}C6MBrLUAms>nko0$iOC(Fdw~s5_9k^6I>ry3R{4-1d>n8y%cIC3QG&`?Q
z3GPMY?7WU}KL05b6cINUZ~HbnS=*k(xcTAry>A~bjsJ^J@O3W!`COyVTfJ3>q<+o4
zK9x#6)%jOLk&IY@k#=`eQ5I+i$7yswQ6k#*yP);*xVtxw=cd0utQ!L=QGiXE%7eAe
zQmN#X@7x>QK?`_!)-<tkbOj5VR?gbM0PIO(yL1$1=k6&!%3GKq)o_3~-o4EAe|c+d
z`*=R2;t%ro=Y*PwC6ePE)R{$!&f<lDqo4E1Js>?eSMuu8gY`8|%sT&;h6ypaOnhOi
z-a2Evn$N1C2Un`!2tt^JX9k}sht}V-?5r^TK2URP!+6<q(9^_{B3`?w-b9h{=TZq>
zVZ;3&krhdj<O*q#l{nMG|NmNm4A=kzwQ-lKCuTm9qV|J0gqhJ&6t<p2%JNN)00{J`
zg>k9dM81f($F*ANFYmq-XoIytql~!8_}nm%+6DZ`?TR{TVt!nLDz`P<<Q7gqtN&pB
z#uL3i_vfWN7u_2vSC{W+BkT}hDw4eT1+?Guzu?@W2wPD7TkNW+tT<ZEw%_mCd6{-H
zxVZDqVE5se;eVsQ!(!Mk6a1jb?|Xyy<S!(XcXapk@WEVS5HhW(N-+O2pI~8hsC@pq
z+=BKbIBMa%`D(rr>zSl&HPgB7F=R5tkD)0IO$(X*E+1s|<S+F-VcQ-IR!Wi_@J*OC
zVHW+j;qP6r+SPWMH2TSNH=o?hCm(b5GzOVg^+vP0;ya%_Kl(bh`Ac5%t3N)|K1j9M
zf@n}xq7u|d>+OCQm67e&-HEI9WkhmBxF?!6ijAYl&2RTHn`ZmswuVJYBb3J0=cYNl
zt(e(QD+T-*SIT0_bXzHj#>3S~e^}e`BA|g|%f!m#L7EeNpG8*eXrz&U+IR791pB~p
zXr`BtE5VpuFc1zHiBd*e`U&U1)G}d`$C1W;_fv1nq9C^tp1OC+fCD7PdF}W`Ar-GX
zj>0ff7%9av(}0#9Jm)KW2<s=(vuO1{>=XwEi5;2@3jJ<jh?=;W&`?^8R#mwn2hOj{
zPY%@9u8_Bf;bx-{+$LP0M;{GFCKRfe9CWb=N^AS8pcEv5c9mO-S%`c95y98CWKxrm
z@%Yngj|XYFKN|=Se~AOz@?QLJ=NjWt{DCVVXXyP5NT#WFB#;e%GK#RIalsFZaErVb
z=y?_8dgI}!_O`j8TuyxtLVfR-e9w#66{oF-733RzG-p+kLc^(Ap6<3DD4i;8QaUQj
zk#VaJ0%7kyc{IT*_hzQ0pn;AxnA#UsvI`CO7>i4B_EeVx=RvIBcew2<t;XZ2JHwSh
z7}Xm)v$%?j_SO_}qb)Ac%r&RB5AFxly;5T&q^UXZR_iQdygN`K`e-W6>*fT=3ub~*
zOQS>~c<a8o0V=ub;I<=yldAD4b&&9oZ@r|+4S=H5##2~Q*w7@`@l2AHD^bfOa>Xn<
z69%SV#xw`@Lw|lbf_<Xr#f1l__s;-6+y{x%=626VeXdYc=T1dTs*t-IE>boyZc%b{
zVKRw!XZ?98<?|Mg?r5;UdpkV|wE4c7a@9z(s1*KnEcMPtS!+kD?8D4444U^oO##a-
zD6gc#qtpScPATWued4A(YL79f{n#+NsnA|CGixWt6NVv$jR4CfoR~!Z<+`aUn=jC#
zu%Xvri>u!I&O%#N3FQccIg^HW-u^m-vRzv#XW#n<aUIe?FAwRgLZ484Kxw50sc1Ga
z#G6@8{Ha0o!*#~mk<w<yU@X#bDH~POi&6HP?zhv~>w*`Jx-Pd1#B_E-^!`1%tqct5
z=m&p;?z<htdrd->$A8iQtf9lzIsLBcgSsBmP^-Bflsjk#sU>d>6u!}5W+bfSZ{~x)
z<&G<Ad6=-SN5cpxDMJl+X|y6xlDMd)(h-_(S~XDgKX5b`cQR8#;;K*?xX6V2*+*J3
z%&14d?sj6tm*9%V2CYhbWaLN)&i1rW<FoP?!AFeDL;LjT^fuUMGm~vcKGGGqUPG4g
z#C#kH)J|~2LJqrEjaFvBp_bIZV=~r(fvf&wp_l<W@K~uvZJ9UIBI&1e;R;dAOa|~=
zdC<Xd(`gVli%{X9#N7uaF0|T`bkx0{SjaLsB;f|Sb7^-+16=kar!Z5lud*gx_5?F#
zx-9OuuCojSj}erur*DO+_@t)s!HP_n#_N(i{-dz|e^=s2%J{Jixlrd|T1wzFoPdVd
z0f=~}@Vy>auCj~*I#pb-q`{aL@<9);J0$JbR@9gmzrnsU^}$QCg_QS0sWlEJ=>PY|
zZ4j!;=b=b!5sWr~FliLx$ttd+Vu-OAN3$3>J9EKSN=CH=RzSQG5rV>^`0qH2PXpZ|
zeE)LRM?4k#4Yr^GUv$C!ZT)aPyd1>sqf{72rGFO6Qjx%V*G-dG9XBr?{wG^c0->*F
zgzXI%MufTj!cnNB4`JaRG=m3z_i{_ckw7|dC_0oCTfPgS3}@JCku!ghWCmF|aZglV
z!9m{cX?#Ke-C)?saRz(LgTQP;&-9c3;h!tjXQ4+K(zuk=jc2!|y_Jx+w)W<R-A=EF
zLj8Oz)uHUmi3N1O6xHX$5?D?|B~vy9eXi2aNps9!GCq>rA(k&x5w@S$8F-;q<835)
zI&7Ac#7GbwM#uyCpG77V76j;rL8u2xfi5LItEYVFK{=q`LuX*h|B@?Ts0*A|1ce=0
z4Bk_l1*P@)w<g!OB#LQ)JlM;mAeglEJP}1{XhsEaY&_`I{XVD*`0Yg4xa8j`R|jbb
zc9OzqJ%VV$XfPCKNr8nA&2hH=s8MB5IhN-SWtRQZXtj7dlIt*5({L%KYp1KHOMZXa
zt<Asbe^sMZQvX~blO`DS=SG(IOw7+klK49=lfa??ei*FtAX$M?lJt9U?ivqS;Lj`Z
zA~b}k86+}xTHZXhs3~NE+K=6b`4(N@M0v@g=_TS2P#TgIf$C>Tx_*>JegF|8-xmSM
z-K4j@_14?Zb3byF^e@==_L~9L`t*>YX3roltyjI)!I&y@OhZ|FxYBs((BWBjbd6Yu
zf@NTQNjl3n<3~?8XCkV`TgaSt62}=F2;KgGsLf0R0sJmm7+hmdpVCh*<C}1fE*3~U
z6gW>qv>=os$X%!#KS?}x;zafUuw-z*4MmjOl`XPsKB|y?gsg6g5CD|)JO#K|?M|fS
zi?*Y&fwaos_T^Ue$F&b2A=hFCBA9*0F9=kz{$p?k%g4|H7D$D$^xY^eNa`F`q_F1u
z<X-Emtk4QnN7@x7qX+~}b~&PIU4uYLF1TQ&pr@!gNA1_0b5az9(bsoy$J0Q0o)Q?+
z5ClSuwNuBLt*29Bkn>$#z4ybpexH69Tl-llzS7QW2t_)5|0=V|^q4<dA#QH!qF3O0
z@{0*8sB>%JYGW(XNl+QgzGTY#Ld2Q*;z2-X3`<awJ@6(Fo5|yMswmC79TVfP&6KX<
zh9xc28VWz{kA<=GMcQNhbeVtp*)Oca4ki?Y_ZM6o-}B&isI!+2Pe>A^A0B61@@<E5
zQ0%p7_bW}R7a#<H#30r7cb1f2ic=qSFMH2l$L~+cEj{<kY}_nh3l0a)2tYtDpHz@x
zA&_|Qvf>?>9F}R_7F}5Eo@<G{@^>un)$2+PF3B+daL!WSZ+q)GUgwv4w)7xxRc{^4
zb|Io)`G7j63jnh^PoLhWBejr;T0jU_f}n!)M&88w8lNP58X;CUhP{&wof5s!|95N*
z_tVQp14FRS{h+N6;*Z18_Q!Rbe^>B9K{%0ce*~&tue=g&p7%M>^L|c3NnZ)JYAvSk
z^TENwX=dEf^(0Bz_XC+2eB_!_0xEDJIERWtGGRm?Mk@PW&8%k>a3+n|^~O%N9#xFE
z@n~{8@Q1563q^Pzeften*x@T*Us{Jg4{@;kWtV<hh~K;PY~?rCnR0zSr@9(R9GX|z
zfb(42Zp|}7W!YY_8fx4P3;uSJMgJgo!-wMJ&ue_rko(ZI;wm-r`6SoT`23$@(O9{z
z7beNBjD7dYpd9QZ4A*^L;N^AtwGFR90!EJAkKhC8EcVS$zDyaj{Ms9D^}rG`W#our
z&>96bf_hgKv!EU<xZ@;beRO9dR^0)HGB|=4JJtB?UdqA=IKFnw6uHZj)lEygi$wde
zO<rSAE$Hm(6TY@O`l|Nt+cYl1%q0{R`W=Ny=eU5#`I{JSuOi?byhvJl*Ok%;<gb+>
z;D^`P*l5_0aU0vEekvrNcyc=zPE5)ks_u};Y9sDMYj41ZW!-!c`dtOb|E9!B(qo8Q
ztGY(Bn8aeH{qrc$G3Y1NvMMRP3XoiQ?2Nn^J1CU@C)zg`PN7?bz;V(3|E3TB1O9;o
zjuAcQQ{fqPoIU+C1-hA8gs|@B@4aZE6Bb$YWDob&-cW@k1iApE+A2!K6y|$eUYrRk
zYp);gngLsVA0G}lg1s2P-Y)@Av6El#nQ$Qn4jxaaw+|>ef<Xt51?LJ6NAw%m7#ng`
zLrD(IRkZP}L{6S+w0O=PB3=Yqi1)aL^M?=rr=&2=`ZGL6(hv97eFLHhw3Zo67}+jY
zsR7i69GYkElOh$o7YctplC=*z$y3V;KWPYx^YLudv<ZNl(DU`Ij5*+w`}v5`KeDvc
zT+|F#Egb0#`wDDkki-fb9IUqcT>2dZ%*4I5tp_dzaRr($=IBl^_JB4MzgVCWPaIdo
z4jOFbNt6EE+xYQBXqk`@fAZC;364e$1fADIUOciH?iJ@i1f+%}^zXNTgB`o5i;@Z{
zA~02c%$&+Yn7Qfm-MmvCmaQ|93gT+a&<RJieDR0xQxs$M-^T{*<r?^--_f)0;*vNv
z_O`%%Lv^$?L3(Fz$GIpn=_iF7QOyzjM1U}cB?ZV!B)LvbPTmYeGoDLzt`8csk<!tj
zb7W(-w76M%bT|T3g{3+VH8+r2jethT%@leP*Y2t#1)<!pobuf;{X&>_-uumXB-T}?
zEm9OnmQJBRFY|ot!-N|-ZXucg1Lh?mhkMgj%mcsaJcn=T0R8axx!^{3A58L{)idNp
zl?b2paHY8+2=!UR?dq-*(475W^M1P!sQY6shQ0gY-ZC(tju9}YD0;}Xj{gy`b!*W6
zu)I-k^5&kBo__E{3}jS*EgnlhGL~2bvcOV>b4>c~ONyIkh+!Ta0jP2nm;RZu<ny}_
z&Qt04Vj+eucNdC0oO2$>eo7Sy@p^UoODr}tE|89uBv^!r3^~+@Bm@vq$pKjh=^zxE
z6%}eiwSU5W0QFE%VSl@C7>utLYv24Dq{%~0%mp4OA8Re&-&EmYWOg7R%eGt~S{+_I
z!8OF9SZiwJ0zigi6&AzcJ7B6H!PJ;y;30n#hatT!1n4=_uJe6s7be}!;9dC-y+MRw
zh@~2M=W#4+9x?24nsHAuWj?zxX$jk0Vhr9W;DZyW$HFCnBny8HVG-3rX8AyL)n}Ok
zu_|kE+!+D*UX>UQ?JjxB#aY;|2_NP^E3<M=$72#IM?fi}%e@s<ru3zub2XJ|yNhVf
zV*;wE06<h!&d)RoUcXn@523M0OFs_%uV%O|n{}u1;b_+`*aL}T<MBnl!xC*M+s)?a
z#jS8{lMlR%s6qs21m*73a@@i<0Wb>%aEF;$Vs8hw`B0~MWzn`?h?g68$3oJ)A~51=
zPgXuc0!CwLB|JoG(!=HkJjhFxzxRFY%51;V$uah+IaAB=Ho$j1iFN^+f7AiF{c7;$
ziGz0VQ8yphThw<aK^N3AlU7xxYLMJkSe7o)?#YBNR_>dfj`Z|Uw0bfyQhcDx>6I2m
zY3>lVpp_j)NH91KD@U!rKz4k@-VHEvhiP~#k8On+Ho^#H(L8s18aM?wHw$F1$iY5i
z2a*l^#~e_pk99Bkt7#4Wx}EXF5VO$n6<OHc8bXCz5|{izt;zH`(GxRu&Nk-o;#+?Z
zpd{ePqy>)RlYD$Nrr(NrDE-%rYJ^p&p)*7dIeY?CYncG3Hq0BimWYngcpg4nc^)j#
z_LN1st2AQUSFdS>&wwmoSM5%I8xv$iD%l13Sh!UQNd3XU9L_|_I6!c%?8SOE=p{Cu
zfVtr=VbXKNem+EX#k>FQIUa1txCK3|2QJUI2H8{D#`o!S3U>Ll;dB5=4$U?z2ZQS`
z3x!G$eGX(-5({01(!?3!rUi{fd(A$O*VW)zLX>fsiGF9kb#+)Of7j+Ct=`g!zY1S}
z$Xh9LPG+uQf{ZRb9JEA9uuV+J77mz%C5?<va+Tt9Cw4zLayxpiu7+eF5+sS_;;@K<
z7KybT?Hh1Ww6y;TDfF?HJQqZ4+^?E}?VB9eFGV1gVQs?U$QO*K9_aX=H1#a^0K>N8
zC3u{Obu%n3u^>{whS10QjQ>o)ecrL@vsBZKXO9^cfi47O$8fvXbdrB;f2}oVW!8tG
zc40x~0Aa*X1Eyh?J7%F+(>DV;Lm5b}EOE*Zl%$MP!b)iEHRALtLJA5BJ%J7npg^^+
z6Gf~28kiJ|VgJj?-sju3LP8Iv#X6|<@HCk;x*4hM6AvQUzl*}+Ybqr$=<~e%d6_zZ
z(@JO!eehqA7AWlRYfaKq@;#B=o9c*jSu7-PoT`#qCO^#t#h(X9Ni^&rsc?D8;0fD`
z*3XH#qbQf6%7WnQ8NE2{^L1kC^b}bT#KTdXDKm?UB0+CR-iKi4BnqE_i&%djRkljr
zrjkqu+C94XZ!i;{7Eg@GfH#2cdsa^Pxw;@mCMn%;%q9k1%>FK+2B^#xQ>{x+1N+(`
zp)a<y2i0$sS9~<D8iQ0Y`~pv?tWlMpL))j?@zf!__1DI6G`aq>y7$Fld_f`^Qu$`_
zEZvZK&r%Q@1&&{xz8e%;@{@wlcz%v(m%%>ZWR$P$XE<05IWK&;Wy>W~nmOjJlO7F@
zC;cvKKknPM#mhMI<n8jZwJN7qKRkhI;%L(aElLz^`&oA))y}G6aEca~BE0@T=pF}_
zJ6i5w`LBhj_ut~V;j61l^SKFxc3jc(*sE#V9?<Chs|GPpFpa3*(+2nDf3fqt8;p!*
z2prFRoNf)KdXJPvxmixD{08qzWP2|JJNYVS?OzF5G5tIlUQWvC*@Pa|2iZ+d*qtj~
zofo6Zg+AccC+7Op6><HFe7L^#6OHbLtQb?6py$6Z)M%$zW@K&}_hRmt1c)pERY;Eh
zdX)U>|Iu^}?0GcII<{@wwrw=F)v&Q`qm7+3wrw`H?Z$1?@NC}iIzJ$rXJ-fZ%sqfS
zrts7t7wZRM@HTdGVmfbvvgHSF&hB?zrn+fJnR*UE5rd)!(ATY}g?QLDcGh(PMrSJD
z4+mCL-EVMT2(f(II+1F*eiRPu_OYF@XHQ&1&5U|0{5Hq+bw)z?>lpEEU1RrB#}&al
z>8ehIzHhNg`v&>mlp&+BYy;}Dr!IF4LC{MVEG-i`w_GJe$O-KJoS=q4lbmPF+ep!f
zWcEaLReLUabY}-#$HY5Pt>{h2R&++VYCSAU(k4Cqn&4_s0fy8ZCLc(nL|8)gdLB4^
zpKeDDa(k}>-l2=66GpTgdvFV!U{3Nph=3h=5%wD`rVGc6Pv&pluUGXud)q!&f-yeI
z(@<&5-UGuN>wd-nyY+?St-iXDcey77w0&IDoM#`|6|Ynu_Mmv8;90M)&_txT!MA(Q
zzMN(Hoh`P}0pbt^!;#(7wn7L<3PnJ9Feo16NkS2vEy<CCB%RXy5RX^V#Gm}g3yu81
zzVye6{u^k4Zpg1#NStamEv5z6%<Ll%O%y7iFqp^ALgEO4^-XB4U3b}*0dSj5Nl$~C
z-d^*8G5zoL5Sm}zP!Gi{vzmy*2GB}uLCUm&iJBnWxk?rc2jA^d@+%M&OsxO@5_DMn
zjW{Z#qTxgvR%hr!xhNxXodgqgwH86&ISxWec3vu4#J_l=>ghlAuATU|ZTnsK<)_Mb
zlOyUE*4v53@D~<zF&JF1(~73SX=e9R(P@fK!pCh7VW!>-Hi5@im!|{bkI?g)Uc-;V
z6+=Hn`831%_9@%i-tcpWHq_o1LP)-dqIp!>r>cgZqsFDba7qr5^>tPWHLAU9kuJY#
z(Kn`<UQR%;Lf?>CY(GSB9VbJI!G&?cl3Jy%83*>U9^O{VSRtwup_T!)u>Lzyoq-3%
z0N8EQA`^DmHenyIRLVXfRIxE*QQ3~-TJK68v()10+HqMfa%x^=gXlx6IMAOI_dBN-
z=K9*1MQL?Y^38#)Q1R#93QgO4&w=cH)@<NV{GJss0v9$>&}maNBFS5U=C(V^k$=1$
zTS=^_<Hfx~tlE%Sf$9cnFi?9l40uC)3%T}xB;EA6_In;pp%zAn7z`r*xGPQP=xLHq
zRC8JSp+!p0h*-^NUnZa(wqH2V*z|a4_w*-t&eOsSbF)9vKgjd7`G7EH6miwiehI8<
z-L0I4ggc{wAC}a-j2&M>0yiPVna}d8#z`z!*><>gt+Bt8!F2O8*=~UtWQfLL;%^qJ
zK11NGcmI3qX~MJ+P9J<9rFjiptnZc{KDn9SvW5PR{ya}evAG}Cbp0w?PC;*gpOmh>
z4@$D)2l}%Mp?{eUKujcYnZzqF7JQ1<cToK<f^Smn<~&N91ELOe(uW7aA@bAEd8|+)
z-{3*U<)p)bBuES|J1z%GL__0;e-V%U`W@4b?-^x%)?odCoX=}Pqvcd^)tI%TEn#SF
zW_GlBW#~uuS2n6+O49T6`TF&^{Wb#X<L?Lx*AtNgEOZ+|-@KH*A5#^uWIfZOh>fXO
zalxvP8fSfptR8Ve$?gQ3ZyN^L>0Z$!i=jinpTOgPRMDF<zA7-n0X_vfKbP!dC$eDV
zcnbAiQUvh`IPxp=Yd?t;7WRE+v3z)1<uQh_u4sjSZH2J*MZu3RQ{OVbU0a}^#qx-2
zuR<?vz%04pL_YBcSW6|t7O`E@OxKZ8hb3poY85;ZOWsTolXP9ljx`$%3Gz)?VvXS4
z5=mcrUj5RMq~1*B)FNwm5e(%Ph&nyCDuA-Nrv`q|kR&}{Lcu*3YgVTvyN*5An~~_L
z7uwa>-&c|>Yl_m*VKrw`X^pC_l|Vl?2ldwHyG5;5<_hKG2TigvVFa*<UWUAc%;mry
zx)Pc}tPhVQ@Id2bt$Z}V@&qQ6VOWUr!lVlg)8itjr$pR;@6wTJZQ%7y;M3hspETF&
z^CLNhk$KuKhzjZn^8XvcJpvNecs`xv8qK&<bJ>pT?ngrg*|Z8&ME-<LKI~q9KmM=l
z;5SfYi4a7zT;!ViT&PO$iikP1{YwdC&%*)v&s!`8uELWij4J_uW*jot+Myd5z~vxN
z%G?>F_Si>3eq`MzWYkhFcm6o{yNpgftotW|zP+Kb@-I7e*O)b0TYwfnhWaGiNpIKy
zo(0NUSdRXuT>*s;5CfxPYApGk1WJ1YfD=AB>XN)??iWvpV|9g)>^p1zFh3LtGtzZm
z<m1KtI+1N3xpoI&x8y-`m~<VHFXyoP-IRly2_URyLbZAD=hGow2hhaENQ_fGohsm{
zXzSon%jOc7zuC?b@JU*DzNf%-J=s9AK?+mF7&cX4aY!+E+Mb(G=qNrN&wmlj_$>Qg
z1I5!jjRwgfP8aF@m<r9%B2@+AA5mw+v>fqPnah8vE}%#M2BO7t9$kE)Ne|L+5XrYg
z<S9Zl-BXY$I#FZYGf~3hM88=C8BvG%$G(r``nAX7&Gn+@+n!SvF#)(`tlVJo_`f!1
zrGD8~bhgoco@=bSL~LJgR+4!+v;U+4IF$5XcZM)rdMiuc+@%@JWRtyf6$Iw~ntK`(
zPHpLRncyIR#TG`E6$Ziz<bJfhHbL>dk>Z{|aGGA`dczPz{2j3}yGmEkj=fb@r|Hxb
zwGl>uK{IAi%?J1xC|LXE1NNN`7&K%YO3<~CD&fD=U%9!ya|ee+u7!s@HJC7(-;`;p
z-Ku?HYIk=MVJlB9Of&%Ao|hL0@oxpqB3`~rOzIroChYhu<u`|d<l|><`Qr)44<N~#
zXwFlbH&WCj8O__=270J8s$&tYq<_|2Kn?;S>55fdK;uiU!^tlnA7-vEf=2DZn2msy
z+ls?O13k^MXr4%W))#L<=sdpnU1*cUX>;f9^JDGL;1y^-D4$3vw&pvZHFr!k6ShOw
z85B>>^mM0GuW<gy%QDb_I`9j=faVYD2Fz-+3pT{BNWvB|RLH3s5|Te-N)7oG6G4nW
z7@Z)=S9Q(@25T>eEE?feG7orul0Ltx@cZAdyQ{bUyN6eYLL$DZmj0HBxYJ{aH*zv^
z`*uWvp&XJ6`zs3TX1Z(7{|P`6v$Iq*`_FoL`2b^N?F8NTm+a4MFhY&u&KwPMA2RxD
zwSUt;a#t}lF&TnSRayyxV%C3$(HE^=ik3U@gc5A_CT`Jc?!%xR3L8e=KVMORX6EU5
z)U%?~9$Qd7HiEQT(%hQQUU2%U!SChZ2Z<pWHf7$oC@ZRPVRrPbRXDihLBwwRmyKxh
z)@rL`eohg`wJea(_p9bNg0IXqIz4Xt3HtYJo%gRJ;(Nbl4_o8yc9Ta{G~8OVKd4uN
zvP+kk8RfuV_QR<E*gJ~*;hoX5Ry2Pzcy2|qu=eOeiK2S5zs`11>2|*6!HUoy8Eg<?
z@O09_7TaNo7i-FFA<qL^GSP$F?9Z+4K^_{L<oX$MjEm_axwJ&8*M6o6uupfe?iq#l
zlMGP6-HZ`53_nbt2v*<{um@ugCVKl2lZT=d76CwpSU1_BlX?;Lf@RR8;61?#ZH9fZ
zGelP$>>|cotd2-gVEO9bn+dKg7C6NmNVz#EPYQTNFQk}zIhP41L3reyu~9X4VNZ0P
zI1AlD2;qIY5F2<#{8PveOhm&sqiQ?*l0k^l3}DO~sCf!voL)Eq0k|+fA0>J5k(3oL
zZfqL2%`leYo5rHoohI%&WYNOczIDI*G8<&qv)lp;Jx4mn{zs$LSexN9nknV}p+XFb
zj064^cOqw#(SCQ99v>~F>g1_Z@ak2y^O42i!;}E>eh>itfnJJ!stDZmD|fJ?FZV;~
zRubYt6+T*zquRGr@L^moyIS`><bFy|@$@)^?JyGd;CdRWBg(TMs6Z*<E*?l00|@N?
zxH<f!JX>dM8hBKg*POTM?bL-mGv--80+@0y#`o)~%S24zDmDLmy(wr_QTf!Q)8qa+
z-0Qz?!JNw4W|xy9LAO<-^MRROKKK+VieXN%m#+pnmjVkpR1|t=9(h1qofQw7Mp`lS
zseUIZW=7(A>AUSBnrv7}{Qq76Cp#y|-Gnj~T<`$>k)JJSF7g)iE$STN8aQlfRZrV)
z3*V;Gw_gg|UP(E;#J!K*xd|em9ASh4npz#jf%i{bSUw-4>tn!0N&ZvNah~5q1O@MP
zU-coOnQqx{u>j}P#u+#Q5`kq2Z3gLmA7aj)KyQUjBD|zfN<$ht6!a+t^~Fy96jwDn
zOTT*(UKcik<LRCb6UiWEib2chH>B2o;<KhYRc1$}OA_?85uF_mB!*l1kd6{(*+y(4
zK{~MmOIM?M=a0tf%XZob-Ffs!ZuYa}ISP^`3-eyvIoV)XNF!M2e+5Xsh>?8;YN?=U
zwI=sV-KLgilf5u_lhzHWJFBE2we8*cknHseMI#we8nQem4aKU9DiG07qBh^tFh$Sb
zSMVy|-T%N#RgQy1=mb+^TuyEaMU#%(h0e)vivdL3V)xB0Xj?Gc1t>F?wlo<dC3N;2
z64nEyZ@!w#{ic|D>(lML)sILzIEkRasyQ2N4qEKA85)^DK?<Nq`9Cpm``l9)Fs>$<
zqnpGu4xsP~vZYiL?nH>w{F=Rj@3(9??dCBED)RfKDDdMCbzi9u{#=;`{CPTfa{jLA
zAd|k?W~ja``Zi4#e44EQ1Mak;w5?H&R$l9TaL(EEliapcWee%eX3nm?%`Ur~anGIT
z&9pur8vqs1ung963-jLZO0|3aLx^s&`P_cx`!g0_SKxJf>n~VhBpq-Rba2krA0Ho5
zLNxbTlaL*RB|IDyFe?h4t;A2c8JqgEaa}mh?-v_5MMp>dqs7Vj@r1q!Jo;JLqK}<8
zt*BfOB&3)O81qdOw&n12S<lB?>cWU2+Kr_PoFdu7Hi~0L&jEC`yoh3J@6x!j^zVJ?
z8UxFnI%_0CdJ{#8ke2b01;dx+Ap6ZYrwYvX!r`*A$5RX0`$g*xCNf>(7ZnK547Us&
z=NpsQU8M8Bim6WY5;c#%4A(Cf`%_oDG@^$iU9EgE)zy_7Q>&4eHo%nF&Z`pvJ4}?2
zDmN_!1;@Xu4Z1U6sVE%AUv6EXvc*;y6LrXh%*NYkou{hyAaoF~WOkmfm2bqaH+SxN
zuU-FE!5SutI58&Euup%IaMe5-{M%l_bF>##uno7OH!hwFa7JvpY^@8HAk1))Vq&wb
zrn8W|oED0rq2{Or6g-4naLxdH!7Zjd^FNv}>JI{kwqA`Wh?X*{!b_qc8<rA{VP~u)
z?=wFcXjApaoSq60yBVauDNS28c3GP9xogqHUas<n4jSYy=&-%VY%_lvm*X0=N#;8V
zG<PafRRdvy8Ij<b5<u|1c1gli6cQ9Q6n<oYGN;n<*^(*^@8@R&nafuu{M(|=2M;(-
z3J%>lFJP1rO`o~)acs7Di-yRA116n|4qPb)w6kIZHJZj)AYwb&9J%HBwt1&|4!qcd
zF(**=pCEV2ie?}k1J_*jJ$sgZdNgH<Gpm^$l~l>2e?f)1qx4~lu(<7Oe@0c^)-#83
zF@pnuX%!|-8ED4J>CEkA4@o!%x^X@5xe9XDl1!PQoS1S#0ImQ^HTmkM7*hWCI(WT?
z0TuZCzW$P=Ql7Rm8#EM=gJp=C>8>_P1imh+24PMmN}DmhbE{B-Oa>+cC;ah502q)3
zOPi+i0+$sGceKM=erfuV$?!Y6kMllB-e4_;D`>YE!(B{&H-xrW>Bkb(tf5xar7%CG
zkMXg>aUl(zM60ugA)ub;!~Fy!31B&)cMCz853At{6R|0G_-T|2e-hJdMmMr5Zl~kF
zK8c~y#GV5)lma>?);Cb9C$Rh5z4m>1X?ZdAyblgWt=$!nmpv>v$gz`k2j!j5T&xVO
z>SBmZ;BpvUErS?`fCCU6U95jr&8kxQ-k-L09m5uzSjnRKN>G>@bhQDyI}NeQ4hW5i
zM%LqrZxIh<G{5$gBTKTFmuA+B%X~E!27+ugQq~C$l|u@_dN2M#dvP9xp(|Q}BdDE6
z&<oGbiK;b8&5Y~jX?jWRc1PJSdQdmOKvksjYkz0K7Y3ZP7|0XlFp>vDT*@hv{`P~$
zy54E<N>z~(@*WXtD|a!D$L$XQO#<yk7^f7)d2DLcfb+tK#`BeF#!38|!0DEkcs>+b
zNwvSBr6@sCxc^?tW$a=x0UMHoclkby9=Xfy=rGC&bmzFJLnT)w3n8j*?+(@sUOp?9
zYdRmU^A7uMn_yWlOhB-K?y^fSkOK6T=$IXk_Zn6bKUj|LqMGu;u~`Cy=6R|~LW}sm
zx!@U<T?v{t;9bTtobn&iDGX8~T_ci4-WbMY*v!(nLr1}|`zc_Z!U|QI8}bS1tD9c}
z@UiEYcSj26so49}GCuYycm~4*-g_lI-4%J?pRS1BzT>1F3>nDFCw+}`<nwrMLzpDY
z0+o!re<@z>0{&4r9h@uLyt0dQ=~+TbJ&xz=Lrc5%KUqkjO^uk$77$;P_EAQjpidqR
znEskSkJjgQr0DS$g>CuHZR(h#*L$86OU9Iyv2#M$qW=fQ>Gng+VN*IH3xMdfA&cME
zMnIFSaS9NX#dzz8<hBfPuoxn2u2#^V!#KTyt&w#|>JZXVB;~j$w?DzLm00vvYz)=O
zzDNi((2Vgv!>Cn=!0puG2ZKXV*XSlVyLFdrbaJ^9QkY=fopu@Nw^S@CVu_Uni)c4v
zyN%HBA%V*_3{I`@`9XrsBZ5LR0H8{9YoPf6y860ceU$3DkVHhaoK=9Aw~=DEC5-a<
z;H1`W)K;%T+0&%rejkGPrP=r2LWrk@i^c`IZU*fSlj$*6?YOts<|D|+wSzNj04I_R
zoze4!t^QW5rf4?~civSS)5yAiL~QnV9O$KS%!2Q|VUXad13|kUeU3@{ubKfvD!y~K
zkMSltT{E|)^qt!~(T;P#0Zm2qTC$sa_2;u14j5A9nYM}mUWVD|94h&kZ2kTNB&AFP
zeqzp7am|LGMYF8<ZQB{rbR(Z~3W0#+;|I3Qu&=g#W2c~fCq<QneR;eFpIOKfj+*%%
ztPV7ILiNjB`*Ag4<^rYXGkeDOa_7DvDP~<i6G8YiK)r5ES>@^<p;;s}KpAWx0^@ce
z$hB9~Zp`LRk;LMj`t!-AOd^bnmnZ=Kf*4>`#&%st_gUNi0(;3{-p8_h>YV-B-}(J0
z+I85?%;9R%IfMCrqvb4+?bpc5n|;7`qE<iSm^h;Iq@ahxV&^;<m;QDXH(%M<rGWSJ
zp(`H92sAGU0U#zt0f6@h<(WUwud5BWYVCR7EK~4?Qnnd0Gb6ZLh~_62*hEj8g$L?0
za@N1jJYz9y)>IEg%hNOhMjMWwf|E5g%XbkvNE(TrCWgm~+C>SQEhxArU9^TS%2D_m
zTd9cM9=1*zfG|9WVH8_3>ftNrm-X{gNhk9QO4}2>C<p}&`ygaVG$XNu;FXj;QBPcH
z_&=b3nbd2ue03e4o4)}F$my2=UIfmIku<6BJ(!KU`k>KmDB%(e3KIo)w$w?Spd?&P
zK7!&br!{W~!!BroQSr~#fK_<Vx!D5|4X%H?uwycf_;=1rRJ1+4A98At1GeuFSWY1+
z2zA@_Xr{~Bd^n#t1<zp4L;Kqb9D}|mDMZ0_+;$ukI2>FIl=X}Z8J*+Jo^cUA`8**`
z($}D2oE<MS6rR`Jjx*2WBDUAR3n3~Nkl8e+mD~sY5}=j6Wj{rLvJHYbl`KdM<tsTF
zY|um=oY~x+C@TtTSivy%yzns&3)_U(@?#WWrIiJV#V_Ndwl}NS0m!{G&^k$-{>iKW
zZ(X_VhwSe2IE2F}^IT~p80==dQdn3Bw&#RXBo#OC4_iOd=eoUg)Se^4_AAhG-aB|a
zm8Ghq16ny8Jx58|=6d#Jv)z}sI+xr^p~)0{;5=6(vts`^`Cpsob{x-zCM(})d=&*d
zA&j$_ZZNrHOUSrtnWAnb`Q^3MOwe@NGEra2FHH&^gmm5XXrK5hxXZBL<ty7XAfgZw
z_(R_X_((pbMg-UaL%TUFQ^(W3|IX%3H{%q!_^$~5uebZ*a4-T9!5sAO{wrSJdb_gC
z+24KEC_c$Tp<26s`aDh3jsE4fkQ~)^WOVe3*uDxQvFdd=As>d~aV=50Oj&AjK!#$;
zV?a2zfc7hZ!Dv45pNT5lA{L#!UhsKi36nPAx`zeo+3$F8vK?-2&D!11;55;(v3E5Z
zu<DGnxKU&kuD#cD+b>t`K9KM4cMB*yNLu>eATQkcw|x%~eY+e8k%+2>Y6ISAlC$HP
zHV7+fMz%b5;yg5~#oR5SQke{*M@Fc+-X1ozo4;b}>gwGBj0w`Xv6Z4Nh!0XLIrCKf
zt2+I~aZeui;hpIm%3(;Jt$cMqZ-a_wJY-mVJ?8ixp`(aB4Y&fA4CRUFB$0BFGT$vN
zE{gONxF`72q3_!c`kLBCyf~vF0a!%v;*<OgoJ|;uLNC^&8Zs6v7*xd&;A$vt0SgUr
z7bjP#OCCBC34wo?!B{8wixW~bGgwl6G3km|Z~m2OKF8AcWG&7bKD6_D#U*Sh7<=@>
z{yXp2e8Q?~D7gvIv8f}J=#zO;Dj@!S^|!&Ql0wBw2~_1Be}<dsCtENXB2TG6b@+fE
zNQMmmL)*7#G*>PIL;A*ok?h(YkJf?BXjoON0$f`$5UTz8hl0Oz$Ey2vUasqAKu$ZB
z3ZSD=!z3kld5Jt~+kn0JJ5g4-K1f<jdf~u=hOP<s+TQm}qQfp=SoifPiaIeTr}|NH
zv%w3UsmBx6|8Zl-q~Ghr<o)#@`0KBkuDF-1n^Fy*J6UAL*?%WTCj)jev1lWMKge|Z
zVAtgYKqptgO3HSd)_667e0R~ue1yT$;kI*SgOmzis;VUOBEGsnvY5TLvN)PD<vWAy
z_~|}6*`hG}e+dtnGW3g=#o(dPIk=8vK2{1%H>k)<sP7#C^jK#9zX0_bS4hK>5xeDp
zqa#hyM6GLta#=EY=BiSjWg-G^*8!#q1djUz-gt4P`EpK+Jj3i@X~&$Ho$h-`q~@pl
zlO_cV51PU}uXTY?cO<qS{dR}-zOueYQXCbvm<)&UKL}u@NmPs=0l3v9*FgAu5Sm_5
zo35ZfxdTTdit~eHs4hZDLJa1tlr*Jy0!KnNc-U^9xH}0G^17=Nu3rWx;o~*|hu>1+
z1WT<%VdT%zMEz^l&d@=?KX@$09vI6M)en6)MMOjYGmEzaXkt&0Fi6Opjj0@NWMrW`
z6#u&v;)9R`p@B`W6NGjMYsS^Rs4?&B1q7V}k`tYlo6+lF<R+8&e@gm2#(0*KxnR?t
z=EDTw!9;JFO7x0wwC}qZxa)=++CfW3i{~fd<~R{v=#=>pMn{U6_D^#$S*q8t?o{&=
zaD(T16?YO5w#S$wtV)ESi$uX?pYkm%q~Amecp5bGhB^gqX?6<itS`rx=K13h2fThu
z*zoiFmy^zK<=%QZHj5rz;rTm+P5&;XRds8B!6?tY@kvq<{};Yr>76%qoXq^hj!^Ty
z@9WvL%gx`~Nio?Ha9&x<)mSpNOJM!-#_hqm<>^ny14^DoQpIkp`R^7eF;-#<rG^ba
z6e^YgPds-gdmZjYUB3e(i9mZ?Du?44uTai(lI4Z16aNfUHnJW4Ds7`~T-16#MU-b9
zo4KL7d4p~WL!*t#ogSAXCmSti#LW6lWg92d%DJ(sl9JvZUusXR#T2*pK6jANTh3$$
z1sB#liWia?Z*THGk6!n-!Xd`Rl|DtIDZ9mJ^*S!?x3-U~2mi5~sKcMF(7lN-g~)sU
zyLa#<p7(>-=|Q2}mF9X3%&`G^#e{5BwuLYU8cu4FC}I@<%gygnOl#{$k4mg#SgCts
zD|_3-@KgXzBubp|5^~u+VE{n`yLHFgHtl5M@p>@F=!1``{mGO;yDopj{r>*GjL#J&
z$*FkbWNh}&wPM80=;Pz5$KGhF4_)o!KK93YxU@q?3`V7^TrID#MyF?-LEm)SBbe{E
zGC~bR2lJ9umTDf%R3Xb9XG`kXg&HU)&YU<vW)PCd*Alz!A7CjKFtp?(7?XkD^D7^J
z%F6d!#Rvbk4N8~AgaTOfUZ6|ta>HYHW9OZVDppu``HTF#-`^>26@B%^gYL@I;!dx>
zmal(}K7KcB=?s2z>eScn_U5dptV~MIjLyCwz(^p7YFfPN&PHA#kkawurBY4@sciP?
zt0^`|fP;f`+T9hk+v>nf({GSKCb`^hl>YFwIrB6(H+SD)A}0qQhJ;-0aKm5U&`rw9
zS*k@vl_+O-%7P(-6Svnx0oVcAj@`s77OKg2L}}Hqep8fw8|spwJ!0aE5qbOvlOw+v
zmBbp#5ZF3~moq0~s`(I#7@W#&us-7W)7~9j`C_xB7U;VAE?kA(POa<%9`)v@y5>UM
z_HuIT$j{;-FHk%*X2CJ$$mH`0Bj|^(CvDL7W>Thu0fR7<DZ|Hee&+NU&~C1T%HHNb
zXv?)Ghyci(sx{S^$^$xZG-BD^yM~r^fFg%Ei%MrlLII}wK{zjh);&RF`b0`KlMT*U
zCH+cZNDmhgqbLkYG)7f#$Di+?4jU#qfSL9E7$@LVt?AxCm_e?iqiqP+RPFr{2W}_y
z#f3Y5wyf;)sKwFCy@(YD-FdA}(f!dyxkSy@nS_G_6CE<47|ZO-*HS|Mp1_UE-+!{X
zm}lqLqstgZc|pJ@u#;~Gs4+b)Fa03~DssyztAB6$mm%^awX?L7(2_p5@XVQJ&G@U=
z%n-8VE`Y0Mi;D16<<Fr<L-N7$22sSQRz6hE=!A=rhIZSdw4*F0BAZ%P#d@e~?jcFt
z@m3tzv_Au-W8=28UeyYOVGVdCl>6jg*>iITk3jiX3=I)@OjHOpMN!Kd@`%Sw>;e;x
z_#Vw~15i9UAaa6_93|{7%~X(7B8+Q-rz%ZEEFA_W3L<3>g9;Tvh*^FcAOx{*xX(!V
zRoA*4#u0%EEZl=CYAkal0!pYWDl+rU`~+?L(ex}eO&A$xXU{#CYn@s~NzYd=*Cq#6
zd3#$Ll`Qt_i_`W^L^*gdYY2v??elRIBXk!}Z_?LO!Ts7dU+`OVBMTu?Xt6+k6m>%f
z5qs_i)C1+g8{ZWOZN@|#jcNaUHTbeO9~Vv{qi(H_kNTy^hn3FTjezT+4?i!chU9(6
zC^4n2N_?dKgzGCXbO=GGp>JS)o-FZudn@ky89vkJE|Ao0z7#rfyM6h8z$WdFj4O9U
z`gRaPyX&@a&qeJx;l`Bf_noqO^_;#~Ae?EEvw7Qu6*m!ocF^QQb?qu>rekeElm5^p
zK?SmKUmy#oKmwcTH(zwf^ttjYffSj7dUhiNY}`p#z~uFIOK?|`2;Dh5otq0@GX>^6
zE?`zZ=iB;WO`<2|*=7(8mj;KKl?HwKfYb2g5c~+hF-Fb2T-S~Od-e=gFa6F5pC%h;
z$oS8#Wmo^s^n#pZyCBUk%=6_fXnyQ`d%1j?_UH+D6nX;Ld0ze)geLZ#{pVHUe)ltm
zT&1#N)b-6=n(N$)pApIP&_$OepL-&eaqv5>TIWRs$&I(stg!IA=cXB(O6Df1Z(XA%
zbHXn|t@SS(C*6P{ZEOYF&J^qpXWU(^gL}KtTt;fQ!C1;H;q-HpN5Sv@2#3|~E`rgT
zWWW&u7kj2ybAU1<@QjBhz7-W}!W9$gY)K7@L}pn{O#3=q7{%-PT=e$b-CxgxiIa6Q
zl%0Ha)S2+pdn2FYyQtezFe_xq@P{IFG~k!>MUrsE_7jgdOs2~a2);Uj`mPw^Us`HU
zWM*k+r>cmL2Qx`L)CMJTVwq}2xx?py_(*;_N(tH7uU|yNB1P0io6OL(&PTJyzyKcV
zO2rNmTrDwW{jVH3px{xpB`r{Tt&Z1!{h?mpnLd=A`8xdKQuX;v1|T*AKHj_&6sxQ3
zkNW8_u!q+}A_*BKni}FN3Ll<a4&yO@ST4*B<&o3#lFwNuY*-&zk|)WG&Yb#1A3pG3
zo#M${YBTETX~j_I`Bto4LT55p<+n^Bmq0otyy)nvaO1`2&u#fc89G06bco+TnV#gq
zA#pLq#uU%6>d}M*!nEKw)Yp!L&|A*ur{q#@P}r~W@MwgkwUSRN)X}Jt;E6Vw2p6L+
z+~?+SGvb6rOCbxbt*o|ZW{ty19Z?*-)0gyZPE?p2+XmxVCqSy;h9rKM2C~Rcc?Kzo
z?Heb1>I+qZWYK4;%w0eRC_L0?8XgDV@}_4PcXJ$tJ3|<GiB#@Duogiq*9vCbr0fBI
zsx(RAY^-m$%*Id|X*MS-<`jWJ0f}UiT0lf3O%;n;UAs2kIR06zEYWjlhveRNRci&&
zb&RE^!QV*XbvY*EWW)q0LnKX2S13NhX>FtN%*;U;X)+#vMH>>Fc`c2v(b1-rn3AF*
z{{D7_J>><WJ6BJBw_XC)*4DE|DNE}<c@mdCVyEw2Ym40<Cu{k`tPGZ@H+Qdt_g^KX
z09WbOCewh|=LV6}^rr|xK>Y@2LDF-+I0TcamyKM$64@=+>(59b!I}xSx%~)i{fYdz
zM-KAEPRj3F$FK4-wVa#D4s@G;XDIX1!L7vIHA>C6^{@N4We$k>6~mRn9v?^OH%EWv
z`bKXfHdMS(L$kk>T{f$9D{g6q12#$;6%DkH%{B<@ZRpM&JCma!optl+>gwJx!$69|
zgbz4%8sy+>Girto>^1i698Th-o%flw&J2dobDK$PWWBRDCkF|4S;Ao{U74k35rCx(
z8tWqC*iY-SohK}KppNHrX}PDW?JvZbQ6hfELS0=Q!RkZYg<74(wqNHt_HLJ=z8$To
z=s(_`O6~<&GDLlS*q#3FoJo9^FJ(GAS?)=#*NE2zACfS$MpLt<rp6^j4T)OpUkLnr
zk!{Ki&=u+Fk#?P?Ox_MM8ZFO1{t#A5a9UVcSlfLai0rC)beNi&qM>IV4W0kOlbj6o
zW|MKnVfHYlKv4DQ+S$8ULnu|e#?nr6Uv;_TWRBrANh3G>;`94ni#QmDyGq44Fb_B1
zS%4`w6S>I&EQMSoACbW`ScVff709Bx`p&uo^~kA0qff+sUH7mNJ^%J3*>Vu6lopYI
zWJ*sY)=E?xg-7`UcLxo$qOzr{0Q9OFF3?T#{__42zGbqaC2Ret?r&5zG#x?gAKC6=
zCC*zS5wEFRTGb`?k%oZ0Hg2iUTLj`>R90Iu?w<LAacFKrOlFZM!rKM;)r@j6^{bN<
z`ctgxZsVv39xt)!_Thq)7`w<W$+0pRd|SU&g7R38J)r;lr5fmo2m*zu8Ue1fS=12)
zP=YFN)w`FtyVX{`zxt#*f8S25^opqI@n0#!jmnj|AZW<(jpW^T0~vO|Jq`7l|I^-A
z4o=Q9V&!y}`oI6?Z-1t9`SBqwEgBz`<`o(bMKW#p^}M-U>==cwv0IUF-dmF}2Moc)
zGSc=xBjT%Q=&Ui$wqO3WpS}8n&7>PT85;{4;0LZ)Db-}>PU^ngWa1GgpLfpF0KvWW
zb%-+MFN<J5&{YDBtj~{l6!40(WbfZ*0h{o|G*Z-tk!{^)kBvbtJQeMfL!8}@Cm+tW
z-WTPHtjy;`z<>q*N{ZVP87reMv^VzMoc|`RHm@?^0)D(M!}|AU>1|f}8rbsYjxUib
z?cxHDmzBw%w?c(kO_d^)p{?l-0Oe#FGFsm9Mk5&9=hP{>J_~j!k+TCaylO`=^T8dG
zYeU2lbEoY5#}{)GY;`Tf)m&tf>nG$P8?8Zsj~2D_^6^_1r4O`XR%{WQn&8AtX-et;
zxJ<M<g~EJjUU6rPU#yN2d-k<rOnTtWmn_A&D+3(YEsf8nmAYWSNRQr~kO@~&C>|%G
zQ5mJoKA7iEvEa@s+`n>V7@w0XqsGefZta4rxBs}z%`HiVg^4ZYt2`Q)KN>3v!4lq|
zRXtz*H(du2$HsV`4NaQ0>#X<0y52|dpT=kW!(q{>3!9qMKQ`YR*t**OND3NzjS_nF
z*$L@NaR~=N)V$57-2gWa5tIe*G2mii_(O8eS87Lz{Xd}nn05<?vG2==9LVU$TMj-P
zGQ+61cVeQUxCc}*Y2C?dGSotW$}v*Z3ePGJBiVb8DQK>z2lK;n>-ma{uI$E7wZC<)
zQViD2hp{0xbABr1LXqR-rW$!LXWEL6+;TTQ<Tnj81E`cWCIuv)O;3Bl+qZn1E{|Jn
z&lQ!VFn-8PQH7C7MN-JKxk(bmiYRxKc4Zy5sD}42-+wa(k^~O2zUkcsmVzijs4Ih$
z*}lU_fU_7m&CvEZD_x(63JQK!p?8~eb(3M33t`3(;Nqe9mLBjVCRWbax5Um#{lciH
z_vr%DglJbnvUSJ%K+tcMJi)leiuOcc6dPcduCM<q>fJ2>(zmy_8L4jFfs*~sVs@-h
zTv~dLY13l26+D&IDh4^YD4TNAnD^1O(PGr`4ppVsibQYZkITL32@afbaZYEne0`~I
zt+O9RV$ktLW$N|SNh2y;@7_RMF-q{*@xq^}){n}}#AD$30Bh5~^n7bxQ8Cbq`DnlP
z&t?yJ$0dviO3gAINvO#S&rA(`&&^KzT9QAilWLB2ds#{KyYgZ)sLVe$p0XrvkvNn<
z8I4C@uoC3T7y~OLu*(%c4)t?Z2BG8ECCj&c31G~?hB^XuyRCtJ%xnHf=5vKSX3`Y1
z_+%EG&0V{z06<}$Hp%K0VPs`%9@YG`4cm+|o<rg32RU|xj)9o|+TPgU&E00EiTE!a
zRA?=bZ@SzqNh*KaZT=E)^4Y6^q&BSSS^mElU}|!5V2V2c(Xr<i3TSk6lQ`6dl`5Tz
zw^=ALo+})FTA$)3XQnRQXc<?6Ch|W-#Yn&ld(`ZX*?E27PSn6m8A;pc-+9Km((fTC
zrybvF!B_hd{5entj4x8(%t;gPNelmDMJSA1YqZ?E>eSk`ZQX>LV67C1Nbcg8Nc1%b
z$-r_d$!9-<XEZs8A2v&HK)omLxE80W(PeRDSNQ7)ZRMW3o-XAJbL~aUa0qFo6wcu-
z<SGByTMBqj8L~opdNHW<%`ZV?C$+yVz6N>9N#54)o>C{MeYQi%P%}DNe>Y88S2)8v
z{Vp{XauhyFa-1sWWC^~%mQ$tZmfo-3X()&K%&#}v!v-14YR+?M<@Nz6B@6zAYzHva
zv-AGo;k0qtI(Xntju-X&8qu;5m#dR-ZA}Ly484|b>eu8@2(my~TA5pOt9TM7MQI5M
zJv%(q*^n9m`bfekv;VC;2RC;?cQqkYNa*)P_2N=13JF;w0Dz}$8b;ha_lnunfG(m8
zCI2PCoSbcSxR*3Wfm{U+{B+lPW}$4|LVfRGd;1t*4ERsa6GcM+W)gv-M66~V4C^kC
z{jRz<{P)*D{WsnrRJDU=@Ah(nKS78pwyomj8>Lo%VZG2vv8KW~o&`h*@ZKze_iE#F
zm!w|UHi|#%9?2j8riwUBq>eM&CAGU7v2v28&7A0*rmBJ)+dKg2PIKAc7jIKH7HL@;
z*?=nA4_hf$3Mosyv_q0hog#}(Q(YPgcO4_EY$19xD{cG<c<|hkkWq$^rb3AWkwe+V
zSxlwNH--X3#q9O6rCz?a4Q)C7xT5^2&DF_L!)i`SjN){*lOjnDbSH2;0RIjO@sn|r
z7S?XZbhNJxO1<Onr2y~kw!ez%4MGcBZ|g0huqPf@rI|YFhUPm03AX(RAF~2_%^SW6
z+~xe`6t`JVxEaAFIaUm|IRi&=p`P3OEUU`x&bAk(fJcn622D$_`>5lkPK}Lr^8dbj
z6oV(0{0^(ai+UhcyiX4=HD?3VRznj_<ZXHLTL{B{+eU2-x**JfpMwB%c*zi~77!%&
zj5?VJKx-2@1vo(Ah<THGs#WPmbK^)=a>FQ-6jE>NDfr2FVoNwp6F3PWr>`Syr^;Dj
z01y4|yWOSy(k8&KsNXf~Ei;~D$){8Co7!cR-I7O@dWGqFW0&ahgwHmHfH*rnX);$)
zg0R9U;J66;INLC8zkN)MRz^#w1ZFzwB!Whp2Gn^J&?JJy8!j*IcVkD9YqH`qhHc7L
zZ?L9a-k_UhLfj_3cuq2hbiCw%cS5ib_Wee`{VI@J%<tif=O6IG>)M4LW<Iaw{543C
z@w6QBXK<|_aYZj~wt2A4_{RXx+xni6+#${<FM>WVzmOS67VY5GqgQL`s2#>k5%NtJ
z6EplmU2jvI#`1?)*O@ZM;pYdW{O<$8nR!d%j<jw2v|DFnNJ%s7u4!fGLo<tKX^FkR
zBHpwgkScazeA%0fcPu7ry3~;9c~WK=^J-w@)&)`5K=J+yjp7i1)g_pM#1H&(Hmaym
z?$qx{6*-2$6q(6V^z6nG%@C<fJS)mS9xi}}6C1NB*lGYq@nU}!QlXL3!ViQ}qAb$z
zm@(iC4=_2cSMY@t{pA&l_w+b7dA%9GH%<<8DLp=*=AI5+*t{QDCypJ!Pc}P2KS<0{
zs0{vu!N{MScLK%_(w6iT(dW<2BiR`DTf#RpP1P	(N%jMdSvM5)yI}46DOL=$LU-
z%I{G&%8LS2tRo!qXu~*WjtMyMLcq+gHqHFcw1eGU@7@hA_TCR$%Y=h?y*2}k)s2^~
zsc^INNyp>#oPYiSJGGnrOS;3LKSx=Nf%5T33oYH@WDiV@KdkpSR;1r~!6637RFG2R
z1BsYX7s$A7xV5*}ShOJfV(+2yjs71JZhlnNLRe1HIvdu>byNQKVa~H-wSXjr6_f9m
zAvPK`b!cH#68L94zuM7JQB@};<S`3rl3mY=lhrx~{T^Pqh|BK;5K*8gBH)wLEX2#J
zFVvAkH$ZhXj8DV;?0?HF+bQJ<WAteKNIKrLFd`%SjpsQ;9<BN`Vg3aFb8~OJ5aQ>G
z49(h`@iGy}p>@EK0*zpV==c?Kqg|<Afp)Svjn_dYF*c0-2Zq$Wkc^V@9X(0qtVcCM
z+JNG(;}U8zITA}X&ftmMP*pVqSg?3iY;oSDASI0uvw<b|HP@mCl12Z6fhDBK$jH-B
z_MrlHEtc6eSWhM&7_wCmlT(!iANTvceDY#1!o;i;?BB0^u`qAX_`aa(FToEpf?ovZ
zD2IX9rUZfFi!xp8;4UQ?F2-oy5V1a}5!nq9mi`b;hI{l94$96(yuMc+;AzA3{>M~y
z);*8_4{G*-=d)~86^a|!<H1NUXz9Fwj!KC(3`t81UbY5}dNWwH4oz8fO{>$-m*;jX
zNg^oQm=IDIALfqh0mT~?8L5<-UV%DCYnAf>t{_t~MxX>4;Vr`PBo5OI=L%<pND;A+
zJ7)voCesj{-Gr08&nPupBPS&xv53n$D~mzCV5b!Gqv|?+&SJEE#1QUp9~A^QbCCJ{
zgaA~*%U<_I0x;x`o2@B}e_n!^iTAD&3Z65d1%1ZT_&2mSH7h5w487w1jHdFwh!san
zsR4dV?|v(ZP-$!nE+3EYFn5sjrz^h&ecMfpn83e&th-+d1CVAu2bqn~BBqHcfUJwr
zsw+|Jodr}&j}t-C)Ix-<^S7~c=hmL)2~nCU-PsD<j}gn@W36?DuWXnke1>Hp@OJh7
zSfz(mf|No}>j;N0TQPXx0P4blb4n4ddXAeT*JKqFPvObbLG0v-vi^tFSiAdIJPkY9
zc;XhVQ%(SKQ;?OUXNsngDkZ3_OG`k4b-A64vc%#V1Y!_&s>7&S4ENx7icJnP7KbcO
z45Xj`o{R}?k2le=ZSU7o!Ef-X+`liIUJ`KwmtsT+7%%}<Uny*@ks6Y*dWGd2D5DS4
zF+!xJh5WkQS0hB`iV$sjzqzP_#-Xx=5n8}YN*80iNEk@LzsI~4rt4zereDpHltuIw
ztAuJ7T1FSDY5;)>H!zMUDj|`B0G_ib#ZnBcg(IlIlKha3c9$wubnAENuSLNV!Y^os
zH%Yw$Q6PU!=ZBPg$8}QfLY?g*>`kK(|E&pfyOdnUey3{)hv|%@oF5s<hfkk-=4y@7
zU}w||b2gfB;T#O<8g%*d{^54~NS(DJ0O^sFjI2<tk7FwJtHi||Za^m12W-CvvmQQl
z{KBzV_h~>?^~jeGjHG-13&1T8yo#tu%U~z6y@1>wJuLWXK&xKDtO@J!-mcNkQMbh;
z5)T+5kI|8Y<rll3NRf~gDuo42Lulz}#3`rxBjn9Q1cAe|_zygB1+kuImlhX8gmrqD
zoj+Wz>ZyljR?yu|oyP=$Lz!N*eLKy~+Hgp$G9<R`c}3%Qw+q25P%d^j#bL<@L8ppU
zZUe=tZZOrLe5lf{-Z^`i!m<M~M^rJJN=qrgR1kIgr{<2kQHDPW56at#+_Wjm%FK<5
znxm`P?wp48{<!^YwL?6cl^}cTvNuo6cDwwmn^APt_a_v3u_6rSim2>#P<Ar9)`Yx-
z#P_ku!P1&~%CmV{Qr=DXM71y3Fb?|W1~&DTBD7l({k*@){H+SPI5b3~`8VNYMe=*b
zk=;%e#stH~r%z*~q><GbfO8p45Z>kzNfJ8>veV&a`-&`d@T9-xNITVdp8#_#XPM3D
zt1?VR7JkJRxY-SPkTZ0?9riq4v)XeeaA4{IhO0MSuDi9LM(}xDkq%WfAl$azy?%eE
zZM)v~VYQhr<^o?@T#Xd|IGXXiykAM+3v{uZ8L_SHcGvIk^Tc9<prB`8&eZVA7G%@o
zNsfQ{`Lz{vR>~-L_qP^Fc@2#rs-fix?~d@*3<aPFqkM-v>PIS-j8CrU)y|Ge9rGO^
z%U2Aexea_iCFSBt_0?^7dDk@3eJlliLubc)hfHt69E|e82X~cwLO$=`C>~KzQRh~E
z(q>cQI2JGJmu0+xra@Z{3qh>(g5z}LEM5?}ebXW);tM8|m2p6Fl}l97+W<MS!EP29
zUo$6}XvBJRQB&Q*k$9Xum0uD2+sFDUfW(W6>ZXg}N>w)_^n#Hf@XFG~gpK#+@40z7
zp^1TowLX*O^8W9Ek(wLZuoZt88fo{3meKGBqhYL`LpNj$R(DKSkDs2Q739<A)b-=R
z$0IocE*s2%fWFvF<}GhqKFMBic53b!(MwFx55lY;Xo`Wi5USZux!(vUL-g{m^@yjT
z1o22og8kbp-%)gXhq5f>B^EKY%Q=-4M+)#<`xEC%eu>x?mg9M!DanlEsx<!Ty`B4K
zXUm)W&m5l%dnLb~6y|JpSj|L?hVAfJBL7MbkwPZ~hGTV{ONHaAQbIz4T6T;DnfmM(
zg6jzWD6oIZCoVMmdsIo7aD+}9pMHImtO-WYkL&cq`2jZ@(t;Awz~yEGHUT()r#+i8
zJA}f9<pd#}xtUW6z#U%ToGPevm>2`h`q+6|#My$??w&NUWU+jK*=XFTo?}m(keAn(
zjJ)7}{kO8}>V7Mus=f`!?kHgAa--tTv^cgFGfx25bT&`@MC|43+Tflo6Lvaqors!5
zPRa<lYZB6zEKJhmWASvT;c#zj56Ke;LNNvZc#${A`~FL<y-h;mj<db#IN*R1S9B}V
z)9x|g9pTQ+Q3L{9B_bQ%?Kq)l&cqQ8&+~9AZq@LMl`{Je!s?XF?_840bW;@HZMKp~
z)5IPu!$WcavABM@o4g@XnpYGGlkLDrviotbjP%Inq4Y+yIA4tt#5acf#$&t~&oFVr
zqtW*j)dWtR>uJM{P44mQgS`m%XN)YfHOXfe1^PSuR(s5-bA67D%52Mof6G}~7H_o5
zKq2FW;<8(=2Tdd5bCVWF?AmEOJR};^EdCtf8mt7}e#%HHJVpriUNVER^TsZ?cNE>}
zq|z^0#z;)==x2Q*;2ALj<tz;p6apMAZ>yT^_$E^5UF1g%JUL8}FE<G-%R}wQh?eKv
zFEWf3j%Hc^=%K(^EAW)t!Vd*~{&Hpqp&U#|0Cy;rZ1iweQOa_!ExhaJ?nh{Njo0*0
zC^%q7kNs}suhN2|lfJJH$ryaL46613;{E{eiA8|T`GX<If}xI6=~47GU1$Ms=SB%W
zl=sJmIW{&ni9GP+{^d$M(o@Zj%TY->MsTu!MNGlH$O=PEA5ZO$ku#?(4mODLj=o<d
z>Sv%7c;irR^xiJ*=&f0&4a9^iTeB!)iZTx8Udn!Ir*aqN*%u7Yt+048HHt~x9sC`<
zjR^;8i?=OPP3(S}E!ZM&;qMM_1xL#9_Ap^0);oCl1#mLgkc^$o`6IJCK)9zt0-Ri6
zP1hed3ZE`d-~Qm*X6F}Vf<ElL0EZ6&n|TVJrk2i_MSbFf5q!s?pb+Rw9m>IQr2Zk<
z6sHaQR+FbX6V(7sASeouxy=2(URzxJUbk@;nq>Y$1~_&~B8?=y_}-pGQ1*dXrRr;I
znZ^Jx-bU`lEAif2)*C*JPLXP&dM5`x&GIgUT6C{9+D&{wj**c)P?D(^dbT~o$f7&{
z%5(0n(w!79f|SSucYmh=y#hkUr%pz9wmhSgN0va8{4bXMJ?Ok1$t@%kc(@rnT@!q?
zt=_b)?FmLsV;U}IpBWlr-+0`0wZ`jm;Lc{E4oAk1USH?CSgi-g=XM-fP!^7ZR2B%b
zt!byKn0X!~CK_n5-_T{+E^Nm!N|^yBj8U}k%W192inH9ug-WRg3+gmsfVf2v%|JrC
zrx0MP`IB}X8vkf9*Qxm7RVs}-KWkM*J><TVuI^IgUNz6d(+i0hU;4;mcT;ue;t$}7
zQQh3hW&?Li9*s4z#GWelGM(uU?}RpMFwlMSScO3`AnCd71}i&-M<`Rmce7F3=YF(s
zx<z)t$)KPN04``8Q3L=aNZ>i&jIVqAv~Sm}X36e)1yUVzxbYp*8jo=0h2L=!1pd%i
zD&@uNId<RMR<C}ZX8}6@Ph^yp`u!d^Lx2OK8v-@F9FsZx*lCm95N6a0B2v0^?w8wf
zpYB{f)i*HB&<nEnX_#3>g$3qdG~xW_?{c+X1ar?5!5>r$U4+5RYBXW7@V;K5&FFzY
zP-Dp&T*b^($X8g>@-h+<Q5Pz>UMA^f_Hd-pvOkRUW>Gwy$2A4s=G<YnLW_ST>Qeqo
z*HpH)<qeG@Ix{>K>mz1MOrs)Y9t`thhhnAu#Qhk?$qv5k_`gtf{|VgvF?evM+lmmy
z=L2t`$5dHU`%y@Kf)w;UHzGPPRPdh3x90<d3TYz7<c>n-3Ij+?NP|IT#t~f6P*qO`
zMNVif4X|H#-yXkh5PA%Ck?3*R4+MgOU*yr#RXSb!J(6)!LMjueHfY-9{wtt?%T|R@
zGxEFC*z`WPpH;_17XE{+Ab67PK2yJKbv&72e>HKvv(sbXeRfnv!#tV{bn+B>V4K{f
z&$;8So3($uzo;5%c?>yMKD_Ar%D#S|oJ9=!!9P)7UpMgnawBy1`K#t+Q+~VcZldl*
zBia~ZbKDOZ)xFx;>hO%wf;Kc>-fmV5%`)@DdaW5ET;zr)y)|6)l`@wV1^oMRsL2%6
z^v(`JH`2%Zr;CeS5HOk3Ay$Z+Nol?MQikkeejj}!1>6Cb6;HDE1&}2R=A`0|U7u^v
zP%h?hmv!X)d)%$1hzPQ}iYg2ZuC<-!>T|WCzw>(pY?%F3U;=+0{9j9(P?Isl5)iap
z8+jR#n7h&hKpM;OEX64oFZ>C-foSH6b1E=gbxoK;r92XX{EDT-7#=GqE9P*uClm{4
z53_HJDP_;grK8z(t{}~3_eRgonXAdh1L4CAB{cGW5wdXT#8HEfgVdjSdFK6i{;N>t
z-BGm<Pc$tfLFQx7KWzqWR~@#OA}7!H=c;OIF;$bXW<8cFqM}1Dm0!OW(A_p>MIYzF
z6!JQqt0WNiQIe58Qyp1&-BebN6is7GbgV#*!Xmlqj&`?dUA0)*{z6SotwPfhWZ!(6
z;zZw_U)Qv-)@=eWn35+^K##vAH9oYvg8y`=Q4Z<a?L`FmnJKOj4{{dy?9*dP%!_*`
zqDtpMvkxC4U$Rj+e=h6K)Ch67V)#tp-^}8lt7@J|i*VnJY_adUQ}j>b1r|P=im+YK
zcb6M#*@>5Vk(7F0VBr8y&q4jP*PDD%91d%hN%;Bac}vwfp_dt}Tt6(prcAR=AVWxB
zW{`zmVJ6Z7;AB%()V6whQ!x);w;!)0{M5+b|0Pg=+iG4ReB3QAjgU8Z@)b%C|F(H;
zc1juv^ERr4xxiu4g6GNk8-o|{inWgaQrWuR{}El#)=Ed~4OgjlJFN}*@+81`t_$Y%
z=pr2SX|d2h^szs#jJoX&@Ihc}a)^#O54ERvz%Ft@Z9wueTAociSs0u*Z<|pI06uGu
zdx8TL@aU9hfBq<MsuDQmcnOMyRhnQp2CZo$nS_$?b{Y7YpVGL3^JrfV_GeP1cbEiV
zNBCV%ig@`Iy)}*=1a9;L*E7g|>pm%REBS6Ytv!7^aJmh{7F%fIdaSI7o-*<ayU*e(
zkbcxx;Q2qM-ZG%7u4@~mI~Al;5K##U>24&Yq(QnHlr9OSyBnmt8|m(DkZ$QdbL0J-
z_gg=<ZuVMh)`)Aw9P{4(w)rKUMlF`p(zV+JtNT6qF9&(M>7nCyV`r82#xh3!B|GoX
zm|aJwmvnw7EXFey03+JwFtBs}_{u*qy=M?H)l{dyjkEDWgyO}+SLejPiWzw*>A1s?
zgKjlNd~9_3ZQ#L)zuVEPqPnVX9o<F41CoJB?+=qW$xN^50}^26TtfaFiU7Rx3>hh|
zD8es3i>QRk*iB66dZ@xPZ(h6kPgY?u?`wM;3~A8__3ydpd|KlxV&(lmb;QN=+^|OU
z_0^QLS%3%PKfI%1H^_QH7t!)~e^k5I@_@P}mFP#4dgrg{Tvg??e9t$K%12g(etvbU
zP4@Fmk4ZB9cR^0#gt@~9O)jUkOOGp``^$=9UeujDT4d&-i@+}hks_+Ay_7%`eVr=o
zga6p<Z7bYsU$846ZvgvncOiP$f+|+I!0_v`%m|tPGu{BswgFcfdOYc=t(sU}ozWk?
z>rEei-%o!fWN`6hw#iPa)E=1+h`>rGyDATxd$n&Bi)kUIu~X)SMV8lz<O<0sdClhu
zm_AGvo1&#Vy^zx)^Y(-zK#6FQsZ})Gcn&P@-~2{+*17hDN{*~q{%PzOmoS<YhIEJZ
zdoqfXQ(w9&O?n1;Itogf<_zV~G{ugV+(6`ms%MZ!?5O#c{;=Va^zII?rL{D@eZ`tn
zUR~kmmiAXt+52_chP#70g<orpr(0ALQ1diD4GRiC%j^KdC~{?4dmlQhr<+(c;`pKd
z-U2511uYlZj|_pKKL3R0E`fJKL$7<fCc>La?e*SMDV0UK7Y+~n<$V9m8e@pG&s`fZ
z=1Etvs6J`2D3zVJ3Ncq@*&l0;;1bQ#pU<LMGsj(W-xX1%SJN6l8e3|6gWB$5Udkap
zA+*+#%B;9fikT~q``V0-9vHMmc$Mkn@wq_rB|n2r*7RJ^Sd)>TOP$5RI}($S$UAo+
z{K$LL(=74@o(`}1#Omw9>R)>drS|&ZMC(o^l^U1o>5bn5JL5LP3?6~aO)fMOQ0jP2
z-|ermGP7OD0rMd%%5QOPCL&6E<)lH&7PxaS)wLjNGu|c?wrl0nBvBbq-<|c#{O)Pq
ziDS)S7{YHIGlh&z9O-$6Qlr#*otf)<2eAHtTahjm8Rju)IS_3Rv2vSae5TCUSHyRB
zGlj*hO(j1*N=rP;%{BQw$1^!oZGdOU<s=ZtWa^DOaxV?_DU*=5;w!8m7au$ttgW-t
z6KN}aNvl9*w_BsS+l16!5e+VfjVeMX;)TKFz2brs5^!-8rJ`7DZS{g(HJ>!I+xJEO
zx*Bg(>b6B66&;<Co(?VJ54*ENfX{lvG7xX4`%#W1eA}<IAC7b_0E_z-5bZg*XX2d=
z1wYA2HfkdeZ7Pv%{oI)GSPFJwKbM!`3&tEUsV*n7qHP834_QbF3*&U3z0Ak%<#q}t
z0vZeEZHkcfMgi7?S_6-#K70nIx!24cG)y_1MJo<FjR&G>>Zqn$Z7%MwH*I^02L!Ex
z^2fcKu(gkq4DZ|4TlcA;#R)N@6xS9zi5MUI%(Lpbn)l77RD3<cpn9XbR$l@18R#`c
zuQQQ5zOPPof4i{?v{}G6ET*wfHl5Kbj&)^~M)oK%#ngaQ_wH=mj@QaHUzeN;ac1yU
z5-?ehPk7Ooe`}u6^6R?s)U0+CFHuB0Ts(ota(i?RsAm$FPz@V*;g=cWsDp)g2s;T3
z_b`%yJBZC2{kZAY%(vN@ETSE~cgLhcKjN8;45q~+oTvfC0^U7lC9ht?v8TM5X+iv~
zvRR(Nd|w(n*u%3}`kEa%o08-<8qKw_`I+0zQD-CVHAzUS<J@<-n-DLg5B$A!D_;RG
z|LWkUkw^0-+2+1;jrWqFq~t_qicG4#%(uCk>@Sd4o-~feDSi7MNoXY|Ug*7%;PUfo
ztVgY~ZVX$L_Kqw8j_sH!o0im-dN$pC4UE@>QFr63@otz-&w*-+>|q&wHIkQ~G)jQ!
z`jJe`cn|RaCP}Q^JI&Am8|H`y?!b=*+Q=4<n9h{W7>T>Z9#Q!)-&0P%A+lg0sa5XF
zt5w+tRGsO97%3#_?!C7tHs?`y7+lSq-CA8XkQE7)C(OY2l}hWq;kz}o-iCNJIqY;^
zXx*&VU5$2Tj?m2-s1TI)DWmTtkFLU_HR-=SbKA?*Qru&#aaHFybTMEBK1G2Ze$jl?
zmw(ICh2K+j6fU4vl%W{95$M>GQGmIa4Jj1OvJu^){S|8|WPScI497_ECuJV}HnnzJ
z&F+RTDc+Bvd=lKNy@<irdFNGTJk;$sZqpQ4-dKK`e$|y1GdvFfwmm~OvVYc@CYwX#
zaX%!?0x!Lt6WAQDg+T^HFw2e>I$mnHKdZ}|lM@w0D2c9w#SL~d<C=h<E4P12klb$=
z3aFqb8q4HLzGbwF&=vn(4e@aOaX&RMs+{x##{Y#uh=v=jj;5zPi+FE%Y%l=SGQ^9F
z2dBzhF1;vUCDH6tOxqZ=5z*Qo+;QmMz9v*VdA1PQGXGsHq;NKgfo>iFU)i86#3Ppa
z`SvnhUVkW13=Sj!xpubhnvK0|T90spK&Btvd1)#+-NjM{24*D-st7utJ7gmx?Vy=8
zy!fy7bUs80$@KAVM)2GWi2}4)*^5)-&ii>Wz^Ub3$MMq6F?$HfvDD2Hm*)~(Q$Buq
z!+J^zc<b5VZuCZ{p43?I+s&nDPvUh0b_s8QF)~dVqJo}J=X<UtnxL4|VyLJFzA_Nl
zezlp*N}j*<3|O2n5LN>SL;q%j>gD2rCzZsEZ7Bw{rmzNdT{ok++Ot0s8sdj&4N3cD
zg-c3-gNUXt`Glg&^xfWH#IOmoea=TTt?>$r3?=ytUXx*!!n)HhIYtt;-nnwvYwZ-4
zB}|htD~Bi^-cmu5h&KAUoo=G@o2J@)SA*DyPFk#n|LB;dtCKNo4toENb*pa)0+2w$
zY0x7+nr;CA1w3C50*nMGG$+l7b1Z^5_rkyMBaY*6O0UoyB2JYhtd2|zs;!~+Yt?T+
z<%ZN{7)%D|UVg`-Rar=w+%ZGSph&ZG0^~>eBm9r85ZqoyV#Kr`GP?%xA-$A>ffUIc
z?661qV7}(OB{_>o7kZN<W67>Al7e?euorCX%<8b-iQ5^HcWM(<Z|wjh%l~o#5@EV)
z->#-yXBP8L03jpsK06c;OKv1VSYnWVDC}ybC31mdjF=X*>UX<+gOR^oY_Wwr7tvMv
zC8(s*0$KJp<EBR&a;u8h=i%VF*UFNYZJvx@T%(aCP6Y>dyH5^(7ZS{=NbUMR@O;ui
zKt~dq*)T8S`);XovAtA#Jyt=!0&=ki$|;dD*Z&C1Yh$o$V%3xVMsqmVg)nmvGl<}n
zmW=s7-kamSA%c-3x86#^?vZq~MG6FrJ%)qa#LP2OT7wc{qO+sYc?HfFaz7BkALm1D
zDsN){yk$oAd=fp0qn60o@N3BZs~ik`3VuCVGsrRsRgT6|)S8qfv%jEAM_gq;aOJnL
zCT?Ia5PNh}3F&Sq_)&VQ!<wHcxIIlr;NzUNnqPsyZ~4x9Opl5j4xF<pE{Ss+&x~v;
zN*32FDxMNl%-aFxFAx&O@!9Qy;+_5Eo$&*qp{gsG*^lyV{7ZlhXqHWyn9Eukv4xxs
zbX*_GK8}g@??u^+bNI-AE+@oo_9WS>+;a4@&!Vccp8Nx&|4O8G<b}Df$K=})r+2Wf
zlb*EDKVWE!5R;x;dCwWKTw+1Qsj-K_bFEufwl{n267|bm9&qIfaak^JOzww0KiG2$
zs`L3~iOT<&Vj*uBTTOCRfC)wE{PGoOnq6LEgDUqchv)IVGFf$r3<su7E6vAR;`z@b
z+ELjfI9-h{>zvxm^WlTY|3n=wya0nUfea3IoTXjXviWa8=YqNqkgka~J6Y~=E{3G~
zsy~`0$KNN(hkfrAvOAtiaKjQLpUqplArN|hP?^%xZ_1a*TzQLng|~=V(M-7@lwXzK
z#CTEc<ve$NEcM8n=Ez7C1!$-ld&;~-_5+#BSkuD-1w>Lsz+D!SwZB`$$gIbYa1K&U
z5S+;t)~s-cqmDE%+hHsy|2Q0k6s@3-D{1%`C)$T%$<zv5WPzU2F#%t6Q1YbbnM5{A
zfa4{ZFv5-a2YnC0%oVBYPmj>9*dt-%qg{+tZN*s~L=L*=e=AN7><~pUY!5X;a<;6z
zR8{79duFBZq2W}T19C)KDvaELoEZ&t+Gcm({^@<5>?Om)QAM*5`-p6MsMGMyLO2IH
z7o&D-0dcDbwRRE=CV;b*AX)f*smu#?8T<n&&Ja@24NmCAj^@5s<};em+C8l%*Q)p<
zbOiK84wNB*E8U^3xo=?fk$1xg_?I&oY=v6+MchV)_QI)=U)~yux&jFlm@|EZaTyd5
z;&w9eX8481DQ8ql`wFoW{2Uu7^acdir;n@|hHo&NY}!$5U%A~9T=BZ0+|1SJ-Vj0e
zLixf-#H62bpbwci2_AeXEx-g(Gi-T9NWji(6c`@ixT3&j*`CO?L|e_ntX|7D;3CIR
zQdZF)AQ&e)9VYs*P$@)RM8(x=tcatA$Q0mLu19|~b<c_fcfLuG*=lnC^78SF!5s?e
z28z1ABr;-C*!r3;4Lt#cgTvni9fD=`Ki`+DC%5FI6kvDaPQqW#IzbsN>(Yd(BAb-6
z;Yx~<lG~_{PV4wWF~jL5ge7%vkDDa($8n|^O9=wdFB1wjnBm9y8Dlv5CEAzYVS)*L
z6G}H1;8F6XyjY|9jFJ)l_@YlRH$R9>)0J@HVn7C2vW>%993|_FRL(gBgDru#4?ei)
zwm_Z?M9Iv41g$nH%=2`c7YZ8Bm*@H<eNgq1R|*;lF|ZD)3%dQ+W5LQ0J3*zdhlC}6
zvZQWB&m7{b5V1aAz8J{o;djGroMvF(b$S@NkWIt0^4uTvPMXA2dk_xc`xhbsN|RF+
zB&A8MKbdPV;N*}m=ZBI=br$UoZy&qtUw=LArySsFfVy66ytp6IW%f!RZ#OXHDy^t|
z5e&WP{>fcNvg6}RzsW5H!Nlgt(J)0^(N$AV2rSEZ-X}T6s!o!N8(#qz_$*nziR@<X
zJEJKElppkJiT@wTz}hM0zq;1v1<J{0x%XYnBY-iI30rnRpIAmV^NSY*WSJnZNiVG=
zbuVZ%_gxQjFlU0{l*~_5Tg(&1{d4wRDr)ko-JcU$id}M6t)gu6YDzMK;;w3e22i(5
zX*)A#`!eq5;DFaB5-IlLDV5e?3Yg$%#e>|vclkem^W_r$i84PV0n#JSMihc6v5IKy
z1vmfAM6V?xjz~3wj%Z(9_;>yvzt&qf!!MLr(#k1}wVy{z=_o$P(>t>-bbpQ2i)6ub
zB7pRCfYZss?Qh>$m2jY}^d5%Ag3?IW!1ts%RJD?HF-;?f3F<JH5YOBa-!U`xOALSp
zCZ0&!5}s3?l~Z>|4d6H2E{q=G8bv(GUDtz~LnI*`>lP_{*`GXVwKav%tmqHNE-0{i
zHzJ34*?(RTS1k{hA?fn>5e&Hb0dWVX$4Z$&-xChEwOu!;Jsx_`bwq4KsxwsHaw_zx
zwgwAe=PuB<_hG)CZ+flso>7T#hJ0)UutqzzJAI{|7IyUI;P_#ah9&MYj}uaPM6v}L
z)+;P?Px(4T=h=$@^bZc>p8e^Rli9k~CHA$IJ=(aFNdj`^H_()Ruac+i3~)nqnQIXt
z!c0ebXKrdTATI0Ugvg8p`siKrN6$jvdG(%uZqF`=+}Vzd3@UiFl}Vc>#D<1J0-j>p
z3fcEs7@&Ih)rvXB;%ym&X;!V`oz4ZL6iLdwWf`#UM8+Q<;4;c!)Y??#O5AZq<`5)@
zi5BOZag|xphw!*J^ErM!Vo5VBBgZ-axv3Mw5=9ysFY$|5^5=Tu6~hfYRX2_zs>QUn
z>bGFr!cn!51rT^NXShjC9;SFnnNiDFXoN!Ae2o@#KPreMA{rqibQ$!g=V#70a>4U3
zVtV{+7jiU`h!@v+-&B>MD(}rlKUhs$aT7l;5wV_C@C*w{L8kX5`J?C(>TS0seNhO3
zRh(rv`ZTM3f%l_0jQdoYevn*@fY<UTk{}TTTb^diHt+bm6bqbCuBkE1NOMij9~tk2
zS<K%o<i9SUvf2$yYGNnWPwBE}ROCYk^l7><2HJe@knTCYLqBc}3mx#0njXs4{#rnI
ztCnB#4HlBrlK;Y8d(V&%OVFRnYy?5p$F8p^EYc2~+X^Nb?!KI7+vx%dgH>Kd4-Wgy
zGAxTS^n;o)EX{lqmpU`i8{Mf`e%r6)b)P6)m}?WCh=!l$siU!pyPYY2omt@QC)jqd
zRYV8-eD)#`_n9in*`7y?n=L<fZGHdhcrCMEV(LvWMxjsebLz1B>dt(@rE|O8Dvhn}
zGL;5f<An0efoHXTjUrd|+FG^_KD`ln#EH)5oAuj)@?=IgbUE8wfx{0_lBs3ZXbC92
z(Z3giAj)bRj)~|E*4-60{%uX`S#-P0LFC9y>gXn{kam;*$PF9xAUa1A4L9CnOrkGP
zZLAO%0Qk4=&_o6&W>RPyf$&~zrznp--H~i+vlw=g_fB^d;mSK_vEkz<+=j|UuY^~P
z*Be?veIL6;p8wt7@^dWWH$HDUh);0tq`z9e!HsCXFJx80QA-^iaSrxGx~351b=mxt
zRZn3;m`2(FM-Qe58{3VU%IKM}0;_KeN<>y`;S4M>_BP9#%ipDkHO>d>%Zs2-oE)HQ
zUrK*vASXJ;&KLD0CHzp$bhf<$LyX#s*@PF42N<N97Uz7e@nXmF3ig$Lw#62`YJ?8T
zO1!LN6&z=kz!pDQ4R+gb>i38_*pz-!^IOAmQ|-@XmJpH_w*Nb;q(}uXsGM$i1|4-F
zeJx_~UO26{_>7X;z~(I=xL@561A5Amp`8rR!k5pOyf+(wC*Y#g<+nufJ0BD-vB&hg
zkg4&BDxWxsIG``8iFBgmnRLEyOUU_fS@&xZzK1Dky%>ah27_gRTjLkX0o9^pTm-Uz
z{)&30i;V)d1GdC8k1h>`8no<Q!)<t@3J*#}8KI%Y`?TXRMxc1yxf8HE`T1umw0zNO
zNl^Sl0s_{4H&*;TP@h06Xe<g@i0aXXmlhX~@fQDEeN?cMa>HY34+JcGjiV94!|@P8
zgRRee9@B`#M6@wUWk{HQk3{?hFeG1NGd`if%6yoF3)BhYoIW6oI-kD`7#WgHKhj36
zut(f{q2bTQR-caO!&szGiqsH0%%4rX_4cg}MsoL{7*GWS$o$d2CF<#eCF&uw`A<T4
z(Zcx>U|0j&6hv)|^n9^2uAd2V#mpNNI@4sYccolk)sA$y>r?8~Gioe-KdR$v<u_r7
zLV&LLKxX^qV-?_YhgTPMW~YCdBw``N{EPq=NI=H@ND*D5?l+RRyojJFn|`F<EjO%j
zleZKcJf2hn%g3a}XNGGa0O@jO?wVjvpF=bGg?|zn$#}%c{l`f4hP6y>#C$Zch@Ov1
z>!!AN1MpEyHvuR9CXKWHMjUW1j9yygYXoKG!r43fXoyb!@$xo5K~8yHOe7go@3bN;
zdY>C?xPQe$e@A=PQ~=FmjL2b84pWy$c(3o}u7=#TcN@;+$UBjt2FG?smK15}Yj;L$
z2|}AUqPPdkq3C;46gVg%uIlY9Hml$9+S039^vleq=?g!f!2kG(nbw>Ry=Z4AFDK_|
zHx!S+A;>?>2tQS#?!VaJBp^Q*YhA@#pshvP6+)6IM>Napg!`%1x2VrfC8M~j=FAF<
z^_Ie$*I1|pzWZ490%=>s412B$EutN^HyL^>?qjVLL^3^5a$(TtOpdWU$$=lhnRSIV
zBE4`qW@^q*Bt)*iGR!6=JhiePV_>FX&v!=B5A{|;jC1YFFE{SU{YPq|$yN_iAIEDU
zIYen6y_q0u1a?R?!Fb~b?p5!iQ)YcVf|pL(Adz|`7P?S3sE-vMAHHP!IF!s~v+UOL
zyL_WxPD?o?%w*<LU94*24^9xD8_dJ;2pOY(&&r3Rt;^fvh122HUU*r-7z6yl5-zH&
zz1r)XL`Ksy{MX1Vr=yjq*d%YX`T;wt=$Su2iEdc(D^mcWj{(n#X)b8pY39jDPV<jB
zk-b!kfyIyjlTmkvNdqCn#xU}>b;Z#?;}rI=tmh9g)TkEMNa?{x{GT*Q7>cL_Sd=Jg
z&uqSyTod^u0jVH(<tqJ}mafW3Ps?$3!@@gs>N>_G;rJ`J1#f+AE&fsC>4zj48q3BL
z1Q!HDL**AQ=rsbqmoRkPlDT>|G<)D_eSLtp-*aT26^nzZs;WZTrqJMYqQAM=K~;#O
z3stX>q><rfUNavV8sM~GO<Ye>yA3Q9j#|kIO5s<c(Ai~ti;DS}IP)1f&b^g|oI|0O
z2&|vr%H-P&`~Yhg8KT=H@V&A0m4P8uLt#PPg?2t6>_w;ao*eGgw!HWc<{N4y7GFnt
zv0T9L?(22Izzpy$g$a8p8qxtOA@eI*Ef=X{q<B1hlS<`yPT#$%WrwG2kL`Y!Nwx|?
zd7WdAcZffiJus&=T{Jr`<2>dq*l+hM8afbI3y<LFB<i`HsCP{#9wI5IxA=$lVX2M8
z`{(*in$j*v(JONBfk*ie56A+nC%C5hur<<~@jFrk@Dt|LyQm}_6`)6`((7rV7E$b#
z55489x-qo61jmw)ES&wH*p5h9V(w-C7%9`24kQhS^U1F(ZHOnczfxEG-j;;SELXfO
z%f}{380{rZ7S;^4`yGuvJ<D;ybPHUoQQ}W+PPV{PjHgVQ;Cl-JA!XB+kQ*Mq?~N6)
z=v`4hg}8_#_t9zA?c0}Ey^IAHw(B5>i)=8sdv*<LY@F*qzv0K$`*U5I#-U;HB2}$m
zHbo2t4kov_$R~YITt>$G6o2?&e@^k`myjnNM$HJ(Pu6OWmbH%gp1KL!;j617Ky7H0
z7Hb>+s!KGof0Ju6hRp<}gx}rr)sLVPq-z9(gud&aX}o^B;5c9OnvPcB@^CSnPltSn
zd);1&>vACuCPn$!b~0b-VtBnP#DpzTA>#MZ#bq~HxAx-&TSJY*<J|+kqMYXrvmtql
z$74woZqrg$%y1y0z@jgbUkf6>_x~%6+d918CGnD1WFHZLel_+;ItcO4;Q1BoxiQ+q
z+L~gvvOrOHmAHb0gor^4kdsH_>kx#pO%W9pwHBHEF`a|OwbnlGUY=Hq!LrGM*WK;$
z-qGa3xbbu<@vF-yCECDID6NsK#U)DWu+d|k>nxMLfkAp>Bh1gGTfW=o@$S&}rppC?
z0_GgCb5+zA@A!3Bm#lsE4X12BOiZ_v9d{36Fqdn;C{E_g<&c;rK?EHDm;D!rYI-RW
zlV&{v6xGnzWI`%4gr89Pgg!`c4!KRTK7XXVMu-agOXV3#M}*`MQre|K%)lZ_Me)mC
z%%EKK%uqmANa~1T%*IH1`|+>`cc9lbY*<tTjB13Ox`GcTNUUgaS#kQh9qPgF{^(wN
zy_v}6dKCRuW!0bWF5NBacEjV5?eEQn#cA#C2&2^!w}#_R_V#oMAGcO0IFi{rIh60A
zmgM+ucG&&LlABiRihvL2U1JOP9p3&-71yB?8kdBVPRwFKBv=W}$XP}7igF~w0|9Z`
z*%9YZzvF2ETk3fw9f)p&2f7Vh$gLExixJ7M3oz5NmO}aMc2v8wh*lnthr<+A?BALY
z$}yj=5Pt}6fXBwJI$1^hXnh~d_Io|V!fyox*GFq)Q)5~}iP$he@}|{gBQuSRu%6J%
zhpR0<ux;n4tH!GNI*!`B#&X^dl|IfV;l}xHcge-(eDZXTr+dR%=i*z%%OMPs&x}~q
z)QBe=^$wrUhPdS6Iq?{@ymNB8K3uALEtijd8i?a4XrE4%`T_SEB{NA_Gy1XJUw8io
zh7#YC)AkjWgx}-+GOT+`bZ2PRyw+AT$hJlHf-p2?23%Sm)2Al-sA+f|ID0su`le6C
z5^VgBrtgEwNjw+iH>;5;q}Yw_k>qB7CIke`?WQFDG*S5V`ylWEU-!J|jQHeEFFLE~
zB!}o^x1b~~w-&!ZrKs!bWSx@N%1R#lm25&#y`S^lN^!&SY3luc?$Ni(;R<KFCEI?w
z+VhL~lWRQfwDNqyc!Hg7jnt(kXVFGJYSzg^E=ge=1}djF1Ab%2V=T&Tx$k^c<<YQB
zJwa;D0c1^lAP$O&Yx@aS_=8A?5_O@3#1=t-;m8UjJWaF)4@+p{VGOFEaqJPL(#aW5
zc_*5!X_$nA-O$X2;e6qG2$K|=VJ%j0ER`rJO}F;*6b|MC!9hNVq0?p_f({N!9kOp0
z8t~S;gDZTJmYYvRxzBEw$u6IrZ)0O<sBM^|Yi^0sWcX4Tox|#j)k+nav=-9(F*<Nv
z%S-MIp-V&h2Z@kyKP<gyY&1ut66yM;=e$0x61s5=5kS@^D54iKV&ZIRc_`d6xRb#Y
zjdiw|yo+k^+Q(`G3cSA=tME;7D+q01{B#-W|48~L&7#ib)>(%%P1O4n0iYxk=q4g^
z%*r}TmK+f=W{F*Gu^bZ5Y&5rFTGDc%)-b#|ts7lj^))ggEjQ3E&=*kpAAEa^sQm}I
zid)zas<`uPxpIWoLI%LAhkw_FfFI9i**@@#j<k1pLIE}{7BMGK2}Gp+L5Vq-T{RSK
z2S{Xvhc4G28>=#~BJ!amKOA;<;+Np+dRCGt+~mIAKJ@w{>6H|ZSoF9P7C#HdPZOt$
z;(Hdh$i*D7rR|D+dvWAHeb5zGX3;-oIai6SqV2|t5fi2JP*k%9O3oiz>taUR*1RDj
z&b!h<PPy4qxM)&d9yzP5^qsv!U0VKu$2SYIAxRcdTT<}q;25!~=?m<O=Ir$QZ+d$<
ztkE7{)BwCgO$?FuI@Qkk`HmA?z}DLS^OqfiEO81)S+BPXHZ0k{`EUn`zjNyb%GDID
zgHZcf3P=p03cT-t+)7JJ*BObLQhMu#LBIsJv$Jzi*q)K$_PN55?%`rv@8a9^5t(bL
zw6*Q@;bNS}t*iNB8SM4_U+W<bg_BKnD}jJ^<lNsZqnTouAhF^sW)5y|s<@Lc)g1-Y
zcKEhEn@RKOlY*SQR>-1@<9a(L-*3FPD!M&sFd#YXGOkI8D>Fu@v4ctza3rOq##0Q4
z(}q`A;j3;OFe#?{EV3Cxa(T&x_>tP1kL%J)pE<Zc&$A>QhV&d69`@_n35}e*=Jz15
z5iUDhAjSaaVBSBiSV%$fZf%RNJDAO_VUWq<bITHjq}s(kZO5d_un#-P-xgf?&0%+Z
zcgoOuU^{n-d${QHE9C(%#r=lBet1y8RubjS+c$DrxQCA%rXALR2mBeoj->elHxQzy
zQc44G)BX>d8-h%UcxE6qbDuxY@qBao8EAYO6_9>#-1U3<<b)<wr~L*!H$UIEcrT9K
zB!Z{i?r)ZaWAEI{kAoOQAJAIrFSX!=(vL~RL^+D);Oj$lh5Y%=&gXy{U5)qP{SKtR
zuB>kQ{Wn+yG*EGXS2P44Y@{?wQwb$SuEDD;7R@?Y0BiwdHvW@6B0c))(O<O;$b(@J
zrV@&VUNmkAn>sP)A7=TQ%pC);ac_3dvmy3VsUW_4G&rBOn0WsmltN0Zw<qF|(mI!m
zPENIj%7UhQ&ngK_^5WQz2sIq9qaDc=k;dugi}UkIu%rvWRQXqdmX8;rbNf+niv^7D
zTXdBVs>?6agX3U1jL0jS^@{4J7TyCbNdeKB)O%<1SVb}~z7>Nog^l4y8xr@EwKw~8
zd0uk3K?kXU3{EPmBj;M98ht}IM~HT!wx^(<w0iI?n9;GpG5N2%n`)=fMqe5QzG7-;
z1UW*akF_B4&Txv9-}*pM9HXX*RJ51upt&i}i2+s-vj$)wYB<0V8T!(jES;<7Y|^@$
zAi~&*ClRyXXrBH$%pTURdrNn4NHUHot?${`BF5K63=L;z(pTt1n^P|V9m*AboMsx5
zm~fz-b!*RVF6uyh`%OD8OricXf)TtNhjcO^LqE2NNAaQi=1WhrA((EX#)Kf2P(x}H
z*mCbhY6twt)Ibu%y|6!WHe1*2?@g>RA~fz$!0@ATfobY5l$;M}K7*ci0&euPUMy9l
zK8-IZ&*81q%w@0hvXZnNYO>P}#6LKBv^FSzL}(sjW^DBC$~$*sLHyCAv3~kBFCOg2
z`2|8*G<d&3O&t}N_LmG3PDaiHFXHPq<jNw^ta3tZ`Pj@@bCKhI6>PQ37gUpz#DAg=
z9tl`Gjh!Dp7-w$&e!ps)s|s7^12M$9fDIE>m|_Us?#uW)zU*1Wmp}k=fNlS2VNm;s
z3K_A$EiQhoLkfW%ctGmWUUFwMH`Xh&QU-!RME$cSedfELhyGfP1QBCNaDmE?7(G8U
zhd$Ka;KpUouhg%Q0$0(37~qoewHnQg*Q%PsHwM4H2x}oAoXdzgqV&6fF1Vhs<7Cm)
zwck5Y5fEJ}+n@q2b6|x*`arXzrj4xsYUh~)NgDzj?*ma#1WA~pfASrEIMr+@5n{98
z@p?DZ%hqIw+lX!m_yZ$^qOJD|+pZ3S1TIkZpK>JFE6xf&y;G~e6bwK3nUHG!wyow6
z)M1QggVsA`uB`(sR1bsP(=SsNY4795@0cH%n4J3)sn8A8d6dvM@j}dIQS#ZyWubu1
zJtcd$#_vT)Thvc-*G)tbe+9%2*ee?fp<C2zUnm88#pn|k>^qC-LAAJk!+o&1#W@pq
zvKt2jqGtD}T#ZrnIZh^kzxV;1380Y={L2j}kch_I$7Uwsv=`2+c-A4@WODKNkj+7u
zAEzPm#=+JP>P6Aq&+^@z06gY?BABabvH99~$YKNtNz_jT>jK|_K*=1Yyvy<ShH0p1
z1cm8Goe9#F6o3vOc|#CHmz-o7_Y}~4c3qIH8#kk_H9QC~DugTrp~>+}VPoH;UQ>8W
zFR+58Quwl?29|0O{mm|nG0LtZWLAjigX>@b7T#60V`Wi7@HH7+4yfgRrw&XS@RHWT
zEHM$})aT;i!^k0h&dCG)9l#&Mae^d@la0{gWfW`KUH@CMVQcnaE4USy08iNnwirsk
zlWg2dzld>2K@$^<+nYQaB3eu&`T@FtY8;vMwiiTXe;?>X(935;f)h}_F?;?^7L#@l
z>Y1RZ8&3zjj(Gn!LV1*3&I_<-%T6BGmqKoSdmUo+EU-HX%bLF{{QWj{-?Q`$hWIK$
zin8FPhyude?-RGWGRVq%inc7hl5lz`T#aL4PfSdC-PS8HEBBX1kgpV(5oXlk-E$c0
zYylERf{3@@8$f=x1H_Lc(wji>QeqG<#NauG$iJqnO_`cK+K`g$)T~@V0?|C?gpQ&X
z!VvKsY`e=Z8zIOtVhae(G}U2J?@Td*#qYP4u0`o89!|q6D!Wk^7CGR78M0BarlA0#
zsGk7oy^hcE)D_`Mq(yN_24it8a~6~B87{Q~vpvCz++f6}7GclI+NI?tJn=1$zSi&&
zO?r6HfA+Jk!9JOrlCNh{1ngy!b~zt#|2mkOkMZMB7yJq3XCI*z;E%rE1NaAWopiE5
zw{sIO$l(Hdu75^dLDsnZ4MTERcXg~X90|JW$e7CQQlzK1Zh;R%7JBs<KvMK38G;>n
zZ$Op_7!{;z?~6zm2P-&>s5BvtH_edA66PHziVfQT<pOjWVa+wdrCX5w<}M|m#PQ9l
zqqrnQvFJfFhF7T5gZL73fUiq(DoAFW`%Xi|Fb!grf@|$3PiDpBJNkGFm4Nzp7Tah~
zMs-izlo$@Rf;b&6pKn+ialtV3Q#XL#ApcEML`PN#6c1!<$U%@<R3#|$b7-D%4at3T
zcMi#N0QmpvOO=L+W^h2Z8N|)}I=sx@!h0MYmdgob#A@=?Z)~Pn9W5OwU+G97W8mNk
zMa-q%eRkuCh>RQsr8x34h#g8lSj_ou^v6=XWPYxYlN7v=l;lhFd6b*f-x%V^dTNCI
zC3%}U*psRBGGfN@J7hQzMU8wI9jo0H%(iS^@+TP7gAgNgL8JH29F*iM(}8xsonNLL
z%L~PwN@Oy5mg4pRf3}}8e|Q2~T8~k;YAe=kIM|YOJdz~9)3nDX^|(XM%Fdx0fI=Ow
zWY(j`#&GWYzt;5<UiqGDpGxrllOghydTGpF2TKY8kV5@*Z*%<xu~nv6(73GGj92*2
zcE*1V>b1!s>4}DV_3xNp*5kfaUb!JE?rvU#jr(?3SG)hW-)5y1n)`Nu`75yX_P6d%
zQ)NLR1L_WPE+u%kiU_O;=_{dE8X=j(*8e5|gv0C$&VSgjOH8-@2l8~lwys<q@$M22
zw-(Lf%?wof)<mlo^wGBbrTfSRNEWgPEBVwdgT|E~$)CpahGkI0fp_ljj8n_0rc?qq
zHtO<MSQW5+U{i~mjRK+0$dBl9m<SK?BM<_R3n%_#_$U__f!8LLCQYRM)7pByFJcgA
zI=xW;diWjB`u3m(pxP}FP|S#*e7EiOoO;0Y!=+L!iHz&k=27K0P|Cyn?hg%L8EgUB
zyK5|7kF%qH9sptC?qSKX`}0Y?lfRh1R+stWLmp<eLS8{AkNO<~=X1G}wr9Y9;u=Kw
zo|x}`XBZ1j6wmmqc9BsETS>=uI7QIq!<MSTkf4mEOhdOnIsuAO%n7Jy0V=g*6<9rc
zqL}|aZu|f_!qmGxSmYKLbUr?uu)Z#b6JpCy>;08-L!;$(?tzU0A0Cwybaw|`dpKwI
z9^$VYW|sCERBtZ#AZ6EqVq&hCYf=$ae&L5k8Wfz!mRX(iV%x=aegFKGKtxu@vhly^
z`uB<R9Q1x*J@k>}sQHrA9ULYY$7CS+b%PTLXP08};S|7mQ`r+%b)Jhg)Zz7-mV5Ng
z!G_J_Q*>BFOkX1tLqh}LI$jRuCM~&->)bNiwGP93_u2DzlStS`d;gn$rrq7$70j#k
zFE8QIz4|!$q@X|lj88m75GwtDJMjwvWU;r><IPXNNrcduZ=QdhyK@6Z$JGHWW}4G5
z>DA&9Y0!;ZR}evPH|YcF5SIr=BPoN_c<Wz-{(Yz|NV@Cofs@Ym$SOYTZe`5$wkSKh
zDg4Wqpt5QL;Zl>YiRi<M^27N+OX+Su>o0db;`ZV(XYS)0H5?U|4EgDkHlD4Qy?Gr6
zfB%zc)8m{H(|s<7;fI@AU{ef`B|p7}*pGjo4zF#hTLym<sJ2`Rx_h`+e|V?>rF0NK
z_8;$f#LQBgWjf3n#8$a%F4Wn*cbKaKzRDs9yBu%pg|1n7+!7^Wh+QF5-fXalbj!N`
zSrwC0P#E9jRg#yJR}I{=s(pWh$>k;J7s#oNG})YXjwk*;=zqekx3r-Wy2Zr}{eqQd
zr}=WZB7AdmgA775l2Ug$M2Mfp$42SiqSE&EzNZUnUp2qJuG>m!Pqb<dBr0ye5c0<m
z(Nn0^C>-DMT_&TyCq(%!r>%n6zHY)s#1e5ffumt_z3~;W$H_&a)oQ56ox;a7W?GKk
zhw-gb5m$@X`22c$Lj6qk`U|x-*s1cWN{dX)XNu3KIwq6_)H%8cz)rxYizXpDOMU!}
zgvw4@GKks&<zUpsZ$w<X{TuaQe!9$d@=Zg{W_^{}d$Xy57^etag4@f38V{}VD>sd=
ztMGC-@cprLp}XU@wzea5Twr%6X7#{8rL*nDQLNEpqsOB=3kwS+ZSVZ1LltY)>o;?V
zHSNeHY&*5n18*|lX2fATGsR&ss?q%<tTSbRwEkpFqzct=It6F1K?1g*KkfWR-mmf5
zq1ye>{IFgTCK9SEONJZ3!9d>{Ei#VG%pfTRx>u-7Pt35@jKk+Cz1RjWyL~>kLhaD8
zv1e4&8#UiFZEW6OkZ}j(DddndGc!{l%<`(RBs@~Gv2p!VFd$ST*1^%n4vD;CQ!@bu
z>tLN;^gJC(M3=DS8XVcNT_uy{4P53vh8l`@#Bg_5cC!|fuf7O#!lM(+9QH>Y;v*|o
zrZ&Dk+xcntOQzHRW9*mp9O<56KH}}lp<hetcPaO$g{l=JMe2@%XyuF|1)Et(Lr+Q%
zsPt}@dF7X1=5e2BLm!CTfS^jn+ZPftHI^4md3kyG-1<n0A|~IaVzu2(E-#Z@r&RHI
zOmUpg4}&AdKHVAZe0ryK#|uIT#{OrAwfBy-n*OlK%Ex~rKGK)LCi6XB53e}5fU-zv
zS9j3wJ|DA_daNOz4GeZx+T~CW;HzhQ1#jAQzJ%@jk0ef7helZw@q{B<8UO$JXn^9u
zDOX{_N1EF(Q%B@1cvEnEHvahgIqHj*#|kI1WBI6SvhEgU^>PVuK|#-v@EoMEpYS7@
zuO&|u3j;D94mKtlPWLxtxR1<@wq(kZX7+3l9Xq{IUVr9y#-@daNzXnU8hB)llz~x|
z)2X;_bSqMIG;%UBjBLE<y}Z0jnTW#q{_J_vLAB+4PATpQf-vTiM@vW^MJcz-0!IR~
zQOGNv(;yk%4Sg$~Vej9976ZNVtw+<^kIu2q=eq=KG&7o8S>{b(i|cbJaA8Ff&t@+M
zpmi04F9CLN$>*&U$(WN?Vxd!M0<Z<XNyW|gw|lGpWD^r6R(suK9uqeXvyxIWW3A9)
zc}}U#giaT;y9?K0gtDJLeGw1ky{HW#a!<;+Gy0Jun_3)0r?$Od*S$CKUT>npD72Qd
z&M`Y}Ty~EP(lkGsZFiC5w{Ut18t`q6NWh4P(R@bk=u@PhNggLCz1h^ibR6#|b8{fE
zM7>nE-Vvb5^!Y44r-u_7@_1QSId%7&S=y;BIPk8<%u39F2n!CaJ?(ZI9~u8~1js!B
z9EU+N1(gm5Lz>CptLNADAQ^H=31oMN9!Ui)dEr2c7Bru}&yZXB&26T)f@Z2!vdntY
ziLk+6)(8b?u(E~Jvv)$6Z(i@Czv$F~{22<%F2Wny!8jtHL`0Vw!4+o*20+DPY<n1a
z$>T)R+1G*N@Q~XARFl<nb}yKn6|zu4)6QOLPZJJBE+hB!hntQ+i%*?|+zcf*jb#ts
z^WT+gXRL`Tywgw3;cP#L$%HOcL`c2EV05lm^zR+uHcwdJwC4e1QwQJGj_rzTYy{XT
z{huR3o)37hO*b@}d2~lE<X5A-mHe-RA&?H*qKPp{*7MNZxPD1Odb*cY-4SQETZvkv
z{+r0Z8=K2SX|SlfNk}-WYsv&T-a&37SFRJO_sdIL&JxE1Uu}Xx(-9V2)VjZJeb!HH
z<{<Y#z5it!UcB6$>IVLgwS3iqD%|TR_A^`HSL%PH=94rQHooz48?s(W>(3dmsDFJL
zF(G7md|J<rlnBbjh_-^%t}5^X2r2SMna|73;=k`9ge)GEC%!(HsOpy5`*g@8oP>ln
z@<&k`pxTc;JPjKw)0|=B9qo|8J_=-|yJ@t@B||i^?Z@h0K%U0>`DP>O?|3j|*ft@M
z{^Qv@n<dXMp}X5deupZf(U8x48~L<KHbcWiXu`wCw~ZF}IUQLtkl*<WZOj`!h10{T
zfG-_v-hhzWU22psF@l?d6@=X1Yk*-vdGe?bhrDLs(+*jmxcbbNYZeu*b#Rff_GM!O
zZ~yOm<hC?I7V)Kn2lIj1EO@>l6-mi_j04-;%eb$K@>%^oxZo8hkjed|7NKn-L7PMm
z{s46azvObHDII71tJ-q$&3v-k=fDPiajIvd({`J0c<8G7++-j2w4JX;#q@i^GV>`+
zYS{}jLpUWb7d=`CjSQ9N>r&kxRM|kR$Z$9zTRQOrEf&)20@WB9;7VJxzCX7)cXSr1
zKY07Yz8At;Si?yFIIPt@GO<DR)IIK(HC2XaFrm^nRVPD3sxOo5^UN;o2z2UYpL)~7
zB8XUWZMAE2KVQScm%l5RTaTer&-qDTFP=ISm%gd0O2}rJG*ztri&i0LdurZl`odzK
zb+s7TX1t1qZnUIQ_=i1cQ;zRH^7p*&D>o<cxI1w^4Q9)5SIx9<>UBA4P<eD}dlb2^
zWn22%j>_SWPFe`O)OoHs+Y5EerYo&pH(Pu(yED2%4g*&1?Q9TwlIxCZ)aAD~7kg7p
z{JkhSk?M)azJ$-0UrefKpI_|fBwFT=Vbr)>9@ZaJt5?x6$dXB?()Sx?L8xhh6_6j!
z!s_8bHiQLT?@t%u6Uua8f1yJb#2tEkp@C20%59YGms*o)^rrh`QwVosi(0j*(gyE1
z%S4Hl_R)>YQFDRA?B3ei!9T7!M=mn+dM*4;Oq)TgWNS%##lrRGeX_mc)+?I|KOiEY
zp!Q|`OpFiaSe-#djY^w#QC8y-{=oz`mDFap$?|Wfi-*0_%dVN;cdiFL{R-`8ex~#F
zMK@<h+6L7*lBmia8poMVO?F#Drej$qRa>SLwPl>d_F<dux5?G!OB#D-N0P586qnR4
z`4g+C#I))Alv|DRy9MG2N9vH5=RxxU>JXljN~M5zUFu3F=i`L*sZ@oS_F384c~afY
z&DzI&?)LEe+nrsr;!%kfp1e8chESXP(?FX1g_IS{?V0tEN%ZOEv}U*4sAW<NvC1!<
zvP<WMg@%%7)QT=cpz0f|)sn_Xl|ngpcYz|Y)9bBNv+>PACCk%m*S_A_Tx)jM))umN
zVmF2+BH@Q|nOvWqBxS9r8mw?z5nOfAj~qUC1vDJUQtB|&uCC1*g(^L|-%n&1d?a8x
z>x|XjUVd~})*eX8uG<PyIlW9Q@OXSsN@-N;e?iFRSQx_L9EPW{iHBpnJCWyn9<6e=
zoTmVe^I^I->hqY)`Sf#cZH@BCj2tW1M(qB<^Vu^=7`eBfT;#rWr#==Eqz)e&W+`v7
zf10VrxT^87;L`f2^>PJ64W0Vcu@OY0b+GZCz{r>H|2VjCq$7Y#UT(qC&9H}kZztYc
zR#u15kQH#k!Kb|#y8cI=yQ_6p`yM$~<1h{P`4QcA`yGSWy}q~fmxBET#r2ZRMni>v
zxwk_2h8w7;{rapx(Tmmk?cYD`4~|MkRX`EhJc&sAe!7~TqqjR<3f1Bp#dCf2-wbBm
zsuga3#B3!o%lIH-7TYEcvs78!IxgA`H_0c5y7khMG;ms)6u-RI%SwjGmeLZyck<T6
zVj*u3ns{xW1l_rmrDgF4i`g~J3*p+_+`g%F9XVTuXgFB+D&v^!0*^~BxATqIgwsVQ
zwR-zJPWx@SX88@zilF(BDGL>shig3J&BED)y9K+O_M^wNs3?S`7j0wHwDk1FHlgj|
z0Y8598M+~uE##Ej7=?}x+2j&Y7xxcM31RZ$JYU9&(Jb@N(H;9Irh{t{0=(5{_LMv8
zJ4^kDucgSI_>aHIcIoN_1@Fb-Nw(YVg?wtxlG{zLe5UBqIk5ai?`tm(=JHgPW0XOy
z_)?`Y!4Fu+1=5KTAFB5%DwIA8&D2@!DHo|$blt9t1JI&aVz6%?J`KYry4#Z+abiZ$
z5*qiUU?6hXd-|a}|FhNiA&5?bI^OFa9#)}xc9Jrkyg^Wuk?{o=_l0wRP*hch!{c)H
zN0p-+@9(Hp?)NvQ!_~U?&UBh(f{P6f3To|VoR1;($pMs(z4dQ_u0?c>1+VjT`vG(q
z3$a*?LY)JhQrS}zW-cysibd)*LoQQ74Wuv5D~6L;$IbHQ3yfr`6-|H6=Sveg?Ahiv
z-~Ihqq)}Zsw{AF?a^Ob7x&N8fVy<$iCH1g$lKw#S50jhyO8_YNU4g#MSP@NfG*K#C
z$A^2;q@{QttY-67wA@tznN5U{eA|$P#N~O|t~nN6mfiZ^zMJU8)5`6N5AHkCZvK+3
z<$3}?n8>FM|DD>*w7$sLXfUDZ_OL+~61aDor;gRg=#Vt~R&6@^=~obk(xLHTx)|xG
zY=Lt0^5uie(Ox<tnc{1#kr4`?OJf9+Lfgjo1)9xH15Y@hvEVj!&FlVmE>q^X?2#H|
z3TE``5^e2I>1uLIbLEmTcM2yeSxR#4PclT*D)j^LCeS%|UoX0z4l1QMkN*B%|F>^B
z&t_eU#C}9T@}pU<#e9v)syDn^m0@trApBPFrQl$nL0*oD@&2sQUOOi5NYvv!*ej}*
z$Ln@uI)8DH!0T53!E&iE(W+ho76IeOPquLzVv%?Mw35dYcOWixVyjng^qd(wGJXE1
z<tGcjp(cB@!9%pv*{o6OTXXd#=B!sV1=`EJoaYk?O{Gj`zdn@vJ8~$3G7-~{IS(pi
z?q}hBo9{jh^^BAW`R^7o4^M25(!9wVGaq8IP^%hK80OBTN&PAawnF3ZT%V0g9wV{<
z`Je1OXTo_}Sh4oygAuY&xFj>@J%UI*n5F<Lh>CjV!<+1^FLrLW{2BY>?qvG=e;Vo{
zj7V6^<>1I+BU&ZDv^3W>5G!@L#8-zuFkRT{#pcjlrDYMalk|v1&0@})^^8lBE`D|>
z>J+}U@}^&M&4O*ggZ*Nv;%CJ_MVbVX;>Yx?O<I~8Gu6Mbe>IB7KGOuvcW?m&gksr{
zN@_vm3;~5utxI-Bvh}flva)!bG1x}z8C(kX9f(-OL*|_r{=C>BofL4JHv&L|hx4CU
zzY-lmhCBy_g}6YucxXZR^qMqYowEOf#ZV3y0ubxq-{r~aKUKw#gVJ7|^H{1}z_ua5
zWUn#&!ZezoWe+kX|E>;BNYAc2W0oG>t9VswwL&&>?F#I@|DiaklOOh5C6l|sT#c`^
zgj}i>=Le{ksKApQbdbkF)?b^?YW3_0zms{A;N+V)IO;bm^HmAW*2!${JXr>uPaRq#
zG=m6EADMoeM)bwU+dG7R6OW~^pZwFFpdMlq3_&2bHM0!OT9nl`^Y!Qx0_h+WtJaG>
ziZqFU;C=Gu@F^&H>w+(IN*J63@187xFF)0@h{`^C%Q(U9UA}h&#>77+Uh64-{Udr<
zr}Q6kr?VU}QN>+X`e(U0L5QKT?MwCrA8-Dbr<UG#S*Y(Rk~Wk>05Q}Waq3sE*R=vh
zok22U?jZjMT%P#r!J5gE{l)(*5_U0&o|W&^8MXhM`DAEdxe#}US*XAzztVd`y&RA!
zqAoJ~Ur%FS{u-N^l=p180dFH8Itj-M@Lp%vQYutPuwwe}6=rYncIp=cq{h2l0PlF#
zZf+gGo5tdx<ZAoBG7Vqi|JJFS?R)j#<ls96%cyIG7ptfAKZA&>0H-WgejKI<*m^@n
z6&VOTh9Wnl_Pg7iE-O>ZqKU=XdG+`IbT8y*s1E27aVM%L7RlxNZyZlQ@FxrXM{tm@
zX+!A_bXecN^pEBzxpLV4VUw=vm5MO2-h8?j@_~W#MBzQ(dOKvRzNvftVP-w|TQ!X#
zU7PraCI9hZt`!oe*XwKj9!6<+m>j9`)yjXr_6;Zgb-3o^o9rrrU%EQ?VEd2iy9go3
z(sU+3&xPowDV0$2-)fb0GfMujmu{hLPkXYlp3DHhmmD&^P@`5i>;{k?dB%adnHCM^
z`=&i<q7a_@XUH^61va*2&_<3@LT)1Osy(k;El(Ra8u_P1Lm_%>eWqN}2pBvLOv(@m
z2Gv16U)`Aem*}6(_W+^PQiCYOWa@uv2&&^niMUFTCK!s`lMu{p917HB{2x^Sehttp
zSLc1Ay<9{v0mkh4NStn8lAp+ie?O1K@jU*n7XqbEhKF4=_-bKE+F4A4wiYPD6gaWw
zNX82vj9u753rue7`{)8#B{3&S-VsK1{$qfk`QV=9QIoHFw*Sa~;pKLa9bb`J_ExHg
z7Cb!s?)a}z(GX&4?QAGpO~>k`ulEiRH-|Kj#ezV%34ZyzkJHQ6K7ws&PA)d=8Q`RF
zu?oZCnBk*K1chkT@PJF_OtH|ApLL;@X=Y;Yj$DPGqrgJu#G0SofEgV80tSush{U-b
zM<gl}$0#7ncU$NbrYFOx+Ko;|uWj;q=jPkwaQ#+_OXVvrml_&Motv|Uf)ZLx&Yvi#
z=OE{EZ(~|(aYJv2`+y}F|Gl7lYRORQpQsMc5Z=N%rzQ!tjAyQB_)kPqM!oi8(~pkB
zvW@8<!Gft=pP9Fk9Y-B*Nh|CMj&8Ead?tuIiBY|$>%SANqNU`%6k<wev5<CF)=f74
zm?J%Yjh6)WUuFHhU9j*$rcU2#@_0;izSySiuI@joF;^c{uMGGZHa=M|A7RPW6gs^*
zZPA-Zye9U~W>zPBb;FulDPChEDNez+_dk);kw?Py*M9UnUcPoqGa<Krz~NQDwkh{W
zR!3Nkw6yZWS>#l)>iVMHZrpnm1k(q{oym$=1Po&Q>8<gDtH>1M056`W@ueBk5mNu>
z|M)K3XtVqGExn4CR<mW(#tS;tYWotfsSr4=0F>nXd;G9Fva_OTX{3_S`%f=mp#9Ur
z%OmYn7)|#VT+XZ6vG#L2zjor5??gLAI$d1DzFl7fgq`zsS}c^X=*3{UW1WSK&F91q
zTc&_8e*E}}TUL1U@Z<o1tT2=gnVzUo(rmMv{}6{!Qi74e4v1sS@A;_LI2M8Egpxs4
z_ReuT=JG-HOR-?Z4US|0*wv;r?27y_hzQRKlMme+C`=4^TeOMT%){1+uLv@A`$wnl
z)*`AT+<3F8y1M5wpW1<ZIN%BxCE2J?S<W_V%JSV;J`+NAaj-sABPDY^ZvOES8=aI}
zn8*2|(59iXlBLRI_LIeYwNm?;{l=d^#=Q~LH`|f2js2P)*FPVQyZQP8)>t3Ue(jwd
zI4$?(R#xc}E}2jKw!ZRgkvLVh7?oxi`1EC@q_}AO{8yy&YuX-PV@MbH_x}l#b94bC
zv^X}8<{YW`j1f|VNW6z-5lIOXxrd`zEoTXa=DYGCu1k$zVh-C9%z^TERv&VSI7r>j
z_^D~AmF8c?g{4(=hfIp#X=M#1W1d*rsZY5xxz*>$B<A>&xhaC0bv4eHhed5Zh+_A1
zmbI$xivh1F7y168wD)PXNQ<^AirUIy_vf54L;6IXGw}GWz|z&-@LTwyL(4(>AGm)i
zOdBp4F^<dS@98k_q=nN$SvMC*S{5fs8tZLn28TsT*Y43fm@msbUUmB@RiED8Z)gj(
z=@bp$p6}19)#?Xg?99l6NI`*3qE744_c$f3?yd1K3LzHjk<q@{T6u=%zdv%C?yExx
z8TpZ`8V^Z49)6;rpym&3pWV8t)i@?2pc5+Y@gA9lyyW0j@Y_*f)aeUKM*BxP!$TtV
zG1B1>iwXECx1cvS4O)11^sB6<1$ZA*s-kSvyJEhbWXD!opL_}a|JeEps3^NOTKx<R
z5DNsPkuK>H5T(13mX?-=p;Q!Uq@}w{T85U+p@yCTqz4A1bKsuA@BjaG*S)hApfb*!
zIq!MTv!A{9^PHzbn!wa&b&R%a=3fYAuV0rVWFO9wiF%BClgDo6r@qJ*P529xHyUhW
zAbI-HN#BJhrX;xbjD1zF-n|U8St|j++x7IjK7!)bb+ZH<A7xEWewKMYOU&gi6(Ah8
z5dCEB=9zhnu`zZ&$sag(d??CMh;gy$=c3Q7vK*na@9l<(mV(`~u_rTWxD=ZZmzodk
zlx#7r_OUaJ_LepjEVhX!JN7VoHi)JxXB_hUE-qagg<?wh!y4T1&tkr?)_J+VBf%y3
zyG_v>YlSpnL|*%s9WFS4_e^7>Tn%|d6DBIg$jG>Z^N+?ohPgxQj+NB)27FQVxPh<l
zM)QW?)%hENz6yKi@OqDm27{Fm{X(bB8mET!&_r$(z(lMlm7B%fED2*BaTXuGyyF9S
zV$D?kvZY_rxD;`PT)(^Jcee$Gdx}{AYnpQ>+O2D*|DPz(e>8yazTj?Og{!AAI0ho7
zhDOQKG9NjW#5U~;P-_-N<AtJc-Kjm&7OV&rSy?|p-$=A<v9))ODN&KbWsnJAUE&f%
zQ0x2sP4>dmD#(9O$GUxskQQ%``0=XcZ77Z~W+8lk#D70;3e55q8F4Xn6#UMOe@H~n
z>k%p6YjPFg9xJJ()+G&^!caG@k&e)PaC4%HFXV^5QuuDzIxSn~u6=%VV5j-6D3G53
z?c;S5U`8-|SB%3oYNf~NiRu0e<o91bkEV2cH$qw4VmCcjzt__%D<eB*fAd1JSf_dd
z9oXH~^=IlgQo!g{%rs)f9+B~duWm2gw_Y)nb9mwR_34-~0mk?Ty@Hom`NPETHw1MH
zxsR}PHR|f80m$2~Z$=QG9n&i7af259zGdfR#~mUvw+#o>aJF1AeYSnm`N>u^qfVo7
zts~!xm8U_<!$+^JlH&7SajHh<rAQy}(T_2%on)4O5eP+8o?r6Yb1wtcWd-qST#8@+
z=f<!Q9^6c;MrQN2`s0P3m;&OW{yA65KmCY6ceF35Kp0hO5@I0a@_)beb*i@$Ctyb%
z#*D9HC&eKc3sdILSJ<@^NA`&AELr{osDH((l@A7~0T@ej2|_{Qkw&=L{3kl|-tcOx
zoO%0+!fWT!%U5k9si7NZVIAJr8UKs6_@xGKi?@)G-1Nr1|92^Ola+-D>`*38q*7zk
zm{E$qDKQwphrj>8N4Lcfe)0G50qRc-;2O)c%7mjSiT|0%)Wv^(^&{72dOit$ZP^D`
z1ppwOpG|JFDZ3TO_&V7fOVm%hV(xl3xnpUe%eUodV;yw4>%uq;hFO-C_w&tk&6e`B
zhnj8Wi}zhSgqvpRn?%zV7foqt{SVgWA3Fb2^&TgXnQvn-gks$DsjF2ATW^F{(&e~b
zhP=NpI69w6<+0slI+j;^4XqjsKo(!VG|uQUt8Pr4N7S92T@M;m`rGGuAbB?9v4z6!
zrGWX3)P`o*Hd_Q3QDZ;FMskcOCacr}rs?1SDo2FErU&}Q)-00p-=_UmBCOS|u>^NC
zBcEHvli%<*&mxocy%A>eb;6Q?6dtwUC+v~4h8QKvShGTt&M?8VKdtL$b!+77LHh&}
zP_jh6OU{#2xqUvsk+%qNCMw3Nk*aEHF-bEob4r(CP67CvA&4_TWlogodkDF{181e(
zq5B*?^xQU;ltPSN`L5eT@3!)#pW8V?(%b=9Om^k1J(hTao}db=Y^Pndt2NNtHZq;5
z&tJuda`xEkHaZ8|{XB=Tg5wo=Fxq|k=54Gp`^WO3q%Iksz~+Gd^5w?KVbeu1eF{W2
zN2^??N+ID?X8|T4<$byz9vmMYrmd2-J}W_q3uv(>o4)StJH3&SQSfSorvskCkwmh}
znIX=$fVU5QI%X|r$0kPqo>E(}{q2Olc-(*}c0PZ&{fhYEy3wT;9NX`L`aV)moMcAY
zx&2GYjWx9=f!8``cR*BWeax)aLNht@9g)Atu`|@%mrJUX+_UJ_?b`$}3n65Ja*3Lv
zfVS;)W$_MVbK_Gh{?=CLE(#h+tq)S~ar;YVjP%Wy8^?~Mx}lsT+lr%V@$sRRKHk+f
z%o)Fa8-MxKVu%sP@)kds2#6E?*TS*{Y%fO_H}j)Rw+3_`E4i&AEY`c+?}FwTA9GKw
z+Z?I)*sWg&q~Smzd1)Ubt6sy-POnQ_gjC1A?sD}e>*Wm|K+s4Md;cY7&WDthS-q0g
zGnJIfTmA#XPl9%{NxFa8t6q&1?PIw<%(8U@pnk028MQ!}ZRoYQBSr0SQKG>*%)@}a
z1zrL~f8A;UWTO)6I)14n*zp~;&%{<>;hZI$X~_;zt)&E{l+}f`k%lT>C82$1Kph*$
zO%oClGI?Sk#oBTD$N^sD&v^&!>NU0@nLZFefav38M0qDdR;|GM@ntH}%>H{f{^VqR
zQgC5!qzatgX~t_;1Vl-zu5T@Z$BR*!KB)dgzBMdqEYdodCg%4TSN($PtBT-<I>2w4
z359v(<eT>=6u$1fN5W+u0(D*637+j69Ua{n7x-2XqQ9nlDeZkodzANO;arO&T`i|O
z`jTgu(A8{HHR3M<YKwKtryF<ZuJf!X+SgkozVIkVdK~HtpC7g5eS)$#oxNHgvmczx
z4A+Di2m@9$Ev@-faO<mw6|%f`ti+&D$)fL%ohN30+@%v;BSluC2w)JaPGz2K-kJdn
zu#rY8^B*yw%W8mHt#n)&GnmK>K9deyz$<64>DP5_?ZV@^u$y4I*N;{Hou@@Z*hrgp
zKKZ>u=B`gt=?7ofdGuhNeG?NGE56Xia|MrGRu_@ViiZy$E^>KKMa1RSAR(`75Rk+a
zp95`|tuX~4ydT@8j_VsQ=F2E~jN{xR$uig0hF9OdehoTzw(ByP9fwLgAgjkPu&@j-
zD<nFv+tyf332m)bPupvChLJU33i=C68#BPMz#WU8TZC&<G2^S;WY7oT7H<QMpOnUh
zwi;Im&LmpPS;$7Ki`jn8ee!mFz8QHRWNn(I1`T__{+=fj)#0+N`xU#VnIv$Uda}l1
z*HBuc#|qj~@Z#6qs$80k8+Qo{5Ms8y7ohBO%W9fi#YEq;ziO)5`|CMmT>oSwwPZXy
z&VbH|uR?IVM8D7t>1$Z5!#!>vCXmx%`|y((9x-Ru6dDBJL6dN$QEiK)9Cq`8;Ze(x
z(fuE-R;*SS_mkY9&xIvhYr8m=uKO^--vKmM9(#=zF;z5<vgXc_P9U(grNGSl1{?5a
z3~ifwMCt^1HXJS*y$>u0*QR3QR*Eets*4z^!^i-A!}`fb&4R+|hrf$GT7!pljLuID
zON6B%EfEDV{y_3-%gV~bN7$}UQ3?p4(4EQN)hVfGNtz`m+7%Ws)r)hCTP-?DNo3v|
z?f6cM%w<%?^USsKUF;I3tP4-j0+v$TpXdZ<pz*E=mr;@{-gaev82jLEu`ymQP&GLK
za(Kx?!s8fSUY^ji#T>ed+g)v=?|Qt1OxW?QKR85rtx_Q)@!6)UV`AVospK-0bK^~P
zis}n}@@BjRy2Q6G_w08QwH^g-*v<OG32N0OY@6*=H8_@);0`UM^jl(`SxxEQQD!*f
z+^+Rv*8|T6DTRdS%DVpPve%l=RC8ywYorg0GV#X-?Epuve;LvCzpGjLn-;Ah8xq|n
zCu6kqY3Uyh1w5eNsQzZZV&WuvO7MihqnGz5nF6SMVAK4`&oZ*2#z18hN2WHT3wtW+
z6IYrmTe`+ZrSj$tC7?i+pF<DoP9OeDDhAx{jbqa=K;E<0BH<gDSz^P00HLvYU~2^s
zR!0(K%3$jRR;cA+sHUp66#h|k`0U{(Man(f6ciRafYYp4Z|F1{ef2SD(r10jHz=l{
zHu^%CV;SaetHoUu`nk1k#;(;r6ml?WF`6l*WWV<LP}E&%76@{{ppqwc?dIwZbBVH%
zLs{d+fspqpT8~w)Qrdc~NCg6U?SdMWY4F@1dmu%wVTR`&;~x*&$$v}WwJcDfKi3RQ
zzrJ-)f955;yEk&@%PRxEXeE(j@l<b-*B~&fMs+BR*10K#lDK4$a0fI7_m%=J0}L31
z{3mu*wbm=5MbYtRXLofP%!@gPhn?TlY}U<W<B+>IpP~kyBlPRE5BI%9QoVT8OX}9~
z6}-+Iwoc|Tr_2pTB(qrd$YtVmws^?ekZpXtm$$pjrk45cp4Ipp8qb=cK#Xo@yKbFn
zw=u%KmlZWqm<!q8D5j?u&foLLi%Cji>2#=1b3z`@S_5km69b!SuX%rT{3P~M!#nyD
z8}<+A<-eyfm)4i}Y`}j1Hr7AwTz6Jp9}{$(ds$_JoT>9(=m>>^IDQy}^72V7ssAl&
zPiG-13H33k7@J<6aah<}e56GB!|Bz(T;D-(sDJ=;eLyG%&eB^*XFr@(CO?wvvD-La
zG-+tPv)FCGb9@bdrm6S2^V)Fs$xdNuz34EuLa`2k*5+y&Za75df+C@OkVy3O^^Kz7
zh5Miy=-tWQ+KJD3IBK#u(Z_RGhGTjPzy?YrfkaYX_~7%bw|LeOF#U)^oC`j~bw@X8
zpVaXJxc8U<Gw<4pf``@jhr`VS3XiKvyB(tM=QTD$JyRR%^e5XF4(MwFVEg?^wa7kL
z;3N8iX``=<M7Tf_-x8_O?cTaL!r5OR(E>5+@VtSMUWz%EePnj4R*Dbd95q#ApBv@l
zshYj3<EV#6!u<`#&5ST`EJ2Jo!h{c3Q#{X0kPW9y_iw#x_20+ZYxXvrV3w=js00^6
zxbAfcP=Sc7Vj}>VA2B=eSm4!m<UUwcL``si`D>-|%=<xUsFrgkHTjf}N#OC0&f8jF
z^=?EYbFaf97cSzorjBQMbj0nr?Qqt&i5hI%UrCFA#IoUv1H}Kr{H=O$BAthBJ|iP?
zeQ9aQ+S4b+&9`(DvEM}#Uf^BlS|0<}70ycoBk=`_WAXH~(y=G_QZ_EuQR%&ks|Egn
z1lyVe2h_p`$QQ-ak?gI8+b;~nSGc01w6dC-c(kHKa6Kx>cY|3P+|hGZAJ<;1t*ot^
zJzYA4-#4EcKKLDw(y&96O-p|Cj+KKd3-tYdwP5{5tHR`K7!*{+7drD;x+&}Kqo<`p
zIk|Ye+HNT%3o5jJO;tu_RP*NGCVMPvLl+|pzW8flHr=4;Fd=Tx2Of~ani4F!T!clt
zOrjeB+38=asIijPl@ASbH-qe(Fx4XGt2Exi7Gb=06KTLfaSuIRs9R%4!f6@+@_vR2
zf6WTBlY{l#G25xtJeu$#7aYK87Bkv6E)~;OI!~$s9teavI>DnkKX{#+J(=f*gRCrW
zd%La-Hs8t%{uNefgVtmWijVX+W6l)&ME#&pw!SK}YZwX$ZGYFvY#LA9^f6mjhGVHR
z+G&T%-LXQaoho-UuRz@dK_&yWb&Z<HVJf+B-TLQ8lA}iuA>UYJrGQp#B3}$_Kwe2<
zBwTc0{A7R9YC1<j{WR%2(Qx|{Lsj&_P^P4Z*h&xq^Sy~$N-T_%j<FbcT1Wi;lAdra
z1^xb!X7<Tiui+;pKdF|tJMj<Mwg*^;F~;){?+UL>tt_F>*i=hdizw1^^xVy@^SQ=|
zxzFJ?Ph~}MU*4!c9NBJ7?7#3OyKd?UbCCCpVFI@$iuGC<GdLr>!LD|numOOxa%;iZ
zfTgO&P|fBot%W{Gc(tPK#Na84ABXJ<9T=4fkz-Pj;SG_6=7YIQJt>Vu3P|&1>zn3S
zXxJ8%^+9PoY?ICRajO1I_PqqNXZd0b%1Z}*qu*}2*{1vYLdD@Lme7`L?ld(uGGu<f
z&9q9J>*JzyVYCcpI$!CuqYhUvm_@Gpc2&&;f<oSDT%i4Q%Qx+j(hHSplS1pb*YL8D
z7Su#pW`e1?)oaBta-UTy_V3)r3_mfnSj8Y(LF%wT#CYM52&|baR|s%l$6o3asMxgG
z!AYr_3yD+GJKqKE|FrZ>YNlDv79Wg_I5T-aCYcbBt4SHk?SZE75WBy7<W^#ue15M7
z0aXA+aHYShvn}Wph9}BQ3TK)iXdU70Qs3@s>x30;KEUx;y|>V-WBu^LzpR&h_VR5>
zrDKAf6(mam&LEeO+ga~<pg+pZs&aRH(P|{jIGgWQfqoQsGEd&?Vk(RJi8Ajx^S<xt
zA`K8!-qmXbWbmDAxhl(E!m2<=+W{yR28Wt>s{LLnhUNc^AGq|&1hBgTtV6I?jshxF
zp2f_!5j-KM-CsROojHsUjev5g;-|G;Mvmyb4VOuSRqc4)$A`_L=PFSBGp{#`ChYfb
z^fDnjo*$=QrqOMroz)HdIpUW$z{+scHPp<DD&&`>5)fQ?9QPQC4Qdp9k6J1AX290g
z^65*0plt$hIdN}UtG<%12DwI>0b}Bj_V9RO%Vh>kc_3a?PW;y)H_Fo<5m3_^u@kQa
z*Umm9AS4`f_XVCtEj^zvA}4=dd-HokMN-ZK0W-qHz@RP(fu0LEWMDH++u5>)v)V=0
zd98<D%S%i7MKaReXxfXeUiQ50{iNY45x0#AhvorZEJEtEzEHCaRh0Z!f4lr7FA=5*
z{yFu>u{NEWFr>Gu2QC8YGTzHxpH^*)Kefc$0NEd%*=igg=dwU8D2b?8Sp|jhpS|Bl
zkI$~34CT03-}vWhm6d*(qQ$G{d3=VR3gB1`K;CHkcvhJA?=7#FGE?0z+O8@uFW>7Z
zEM*c9IR23AhSjz-EpPi-{=}q^{i}86@Eb@7UvgBV8l4oP^<%<HGL~roQr3>4heuR-
zHYSXaW)@gu$u<WKXo~8Z`1C^<DTi@f=>k&JYBcW}q~mzv?&z-ws$ScRg*GLa6?Toi
zJS!^-qSlYS)cd`gLjZeo`>NUsDMIB~!YCdjSZgv#oR?3OcL;W-*(_%cW1Z$O<Nxbl
zIKyU&5B7)!t(83$FF_2Y`CpFkB;a3AlNOeFDVgm*_?KV1T-kX}|Jv9t|28U6S#rb5
zXfmwPpixF-MEO;7ZkTrdof~<Vmd?wk@PEV7I5x}ytL0KTy$)H6mEX=?wkcigB6=(@
zx3;!_^#}^=o%9^<^_?Ty*RpcAdOG<pU*x_CP&(<dZ|=k7z}an=(#8E1>`P8;G7MEe
zOH~p{rD~;I`zyc2_DDeSUQ=<zV?WaVRSM6K4E(O$y#AK%ev+4?F%HztTA!fx(~9~_
zOFi;J-LRVfc{Nw4a0FvTguLXtK73cx$yy&<{Y!S8T;`$rIu8nmDTq3PurKrBjcIf>
zjpOtL=cSJ9{c`a3D7B5o9l`vdzl0fOAprk~0Ps|0!y}K}kBABuDOXFLA{cl5UtjDM
z<J*@kC6ji@fPT;b+IhZ2+pZ0XlgT&hr3O<-DSG?-&C)@63TappQ;lwHHb^>A2fFqM
z_EU?=Fu;er5<(jmfxu|>fDw~y%ug#uZz6#m*~_9bmTq3G>v0vxx})^%9(OLdr6LHp
z%q-X=8$7Z@_IP)Agv8G^8y>%#HM~RD*>+nS*o|@&lM9qz<@KPCCB7iXKfp>qG*LI+
zm7<;ko-?1xn&0omo9a-qSpUSyksm(y+vX3L9ca^a?5$S=Ne_{!u(*74+@POaPi8!>
z*hQ462zr3MC5=3wD6upHmZI!F{}tIn;;+Ps89oRNI^$N8FcdC)d3sfre_8cd*I8(e
zezB5kKzR$<u(W$DdpHhgyrF;y8nfx@^ON!0sFAvGx5VBu+iV{BS^Lu!W;|_gmuL5}
z;A+15f+M&UX}bNsZ651+1U#@}47EN5BS^jSD5wSG)Q^@~gNh8F(NN{1I8OiJeEg@x
zKhBHJqF3;X;z{sXKfX}fN{&LICb$Q4EFogKPaVlYlBXg(et;ILgzQ^o^V;MA3Cs#1
z0vHn8$@2EPl*j~4G>6?!tGeP>^^?g)lhriSjM~&Kh9rF*Bv=|XH1e9@@pnJ7@BZb(
zQSj%rc6-^E!E%%WD-3@)=~Nk-_TBw8U`4f}81goQIx#kX)$i+~=(<riMNoc)nHG~y
z@aNVlJ5kSr@dWNQ8CHY(H|t}?YC1XzI#qVX0FP!22wK~OPK%y(wdzzm=R4LcLOxnq
z$rk8T+T@p+bXxT>=Dsr-pwkCxV*-zL?BRD%fTYyZ)%6X{kA!5z=|kfE@kKYaM0N~;
zX*W+|Vf5R%yCc3mP3wBNg4thla`jIMC_ULocjE1D@Nd|=<^Dxb<7p{1b5raCvv+HD
zPYjM@f|kKKxtp2m8^_16Xhy+{`rS(4_$=0~98|e>^WE`|8N|Bhal`3AozBm9W#?X<
zWJ}ZRhGnX`DT0#L<rVN3%yO91LoKlBj}>WWUMM6k7ZadvCfAv?{ej-XS}#6*PQ+f(
z4@Ap^YX)wPO7=LH>8nd%1&tPG!}%%a!$?|_T*CYa<4Fl2Q-1%D&(|51IX$^v?6yZ%
z@gIILIw3eeX;1JQ_)X?WO<Tbj`u5Y}tYvJTiN<UV#>PZ3Y(RL0^cZusf#0KCGK6%=
zC9{(H!II5XR_p2hc+>f@5*`V^GRRT(UG08-7Hasm`)PY&zbiAGi8QRsnZjhUr?+?H
zWWxpN8p`$5a#&4iy<`0%O9ohugtqF{uHD3ow2i4OQFssY#{COof%W3r<r08)FwPVd
zAL2*O%Y7>~Vx|p&YJVJSC?u7b=PNDvgKmO|z58*<C@Z)r*!?8B%Cc4n{vH-KvB$i#
zi*Hx)CMK(pU<V16h-goBzqofG@H()yw9wJ_XvLo3-K2MTE8K9z4__p&{?{vghsk5W
zvB|e{;rPia$_|4h3!G9UGM%`uP3X*~Q#p7F=~&D4T({9Qu;yaqSiF&BM)Hilmx4ii
zoSTC~CH?J&+ii}_iObcCy%JZB_G{PG1YIqU-b9t}GoyzK!;*fV8CE}y=wdU--K#RL
z^w|I&%F*Yv60iMnRe(>R{_Kn?j!hp5nFdE}D6f}SG^``)Vp#g!c8hD?P3|o9urutD
zKV|<ubbHiT%1AtT%q|MX9TQ2aJS{1+6R+M4Q03+0PcV^4Jfni?(W9)JYn)ymtzhQY
zADYqeMab8<*j}^0mpb1scy_}#_fRXmGvu>+JYKnM%9v+z^5!At*ZA3&nOm+4%l$iv
zIoOU^rL@H+LK-QRKk^1op}@Adv=aBZg%9afA#!@A)}Ml@x+k!w>>{xE7LS}${q<K|
zjl&dSc!u`aRiPZiYhcq0rWm(@<H7pXjhl0kXp%M55F#c;4-?0FIZ~WE*sVQ{jq_8W
zEUCAY%-yNC8E2u>8bU!YJFglLp*V-jpZN~nCG}ooHM7;JwLn|j%AM0CTiLjiZ20;o
zv9;E(@Q&!J*Ue7z?KT$=5wj(KS>y7yLKf8mQ#kMqCEu_n@&~z2UddOooY7JBEndE^
zv%L`*FvuyPzy%&IH<R1{lNk<)qru-Q1@x$PYZ#g9%doZa%Q)5aJfmOtuIUM`Eu(aV
z*_PGGi@&43^xdQ>a(dlyCjB`r%%_$BiA%n5lAPZaJzf8UTSw6${U52YLI-Q#vGh9M
z*th%IDW_`Bl^?E|am$V)+%6P^A<p3^@TWQ1s443bIDwX4_6^bVJc-~ZsD3`=u(f?-
zdYMIPajIsqsxd=4%42q9UkU&I7}&y%Bv?xCyz*yEHR|C$F8d2fW6P-#F0~##{p>kd
z6w=lJwwXnHgj&#J370&@Eah{*6UK?!%>jcKkmf7Jx=mtR^``>{!bm(hz)I$fC<-gU
z6nqqMz-~EXtIOj#Ccov=S=Ku}KNY6dThV^=;nOt22wBm-oat*xk1kht07t9_P%28M
zFTrDgXJgt!$r<V|!?euxr<qm1kb<e{R87qKD;0LorC8m1jL&j>E|0~)3`uxPHCeGi
zUG>@Q?sEV1Qdw+Oqk;^hrBmX4T&7GFe>d-;3<;Otq(e}`Jc%WyH*N*hEhK*hMWRP#
zf7v*tG)sEg)ZL$Va%Xa+4~&VCC&@x(zjn;3HGbXZi|e~m+#1He!=UdC*Es2Xjc<A6
z(79GrW?>1abpz#E<90{yL&PBjklD`QBnZ@UqTSBDy=XN>mFGmGjJYWVCHhbX8ex6W
zvp)?6K1&3cN=VhoAv!Zf*fTmc8juMLj39YuYtmz{|MLlQfXY^8$Woz^$_E}54!0dF
z(h@HUN*h@r>>LZ71ez__jv;?DO~YDiYhHfe@;eqySzUak-94}vo6M|f3Dux~5K@fW
z0Zv&jRpnFDUs6)m==D!1Slp}G-!Qy5(W<hWRCPN$W{24P78e$DFh0g?(l_^o-~w{4
z?5r9{^{6N%QMkSvB0G5@^rIAcKJe!k_So8`9j}PamfUtjX2V2M(_tbPdMz(s0!;_N
z02K8!6T-67Rp^RGOxfjnZhMHuiZ2_Mm6TTloM;apGDbIe`}IZ3cT-Hyn-lZ79i3Ic
z7Wu;joi{3{Z80K%^g=P|d##zBxNhq8Cm#BN1}LLN+FV$17p&~ka!ehlJU~yot&RXE
z#Z;AbbZTlTUFg$7sc?hKQmw-!x2C$OIv{Zkoov`_9S|1lT%6#XGh>r^6^|DEk>_UW
zg&vn<n`<ub)ftCMKtUaq6?T-~OZeUL;r+^XZctY=RBF1eLI^<|JJeS$4>E-Z^hrND
zPe=K=M(Wx)HFI#R9#kX5jJrP!B@cD~TRUaV<WjO<@W)aTz~B#m0jrrt-`rj;rVP)=
z=JGwuzWhbJQk5kV1Yg?Q6Am$l2BIql4W7&bjL>s9*4zUKL`I(Cbf_~Lr4S$XB-W(c
zL3Zjq+B6BEeC&&~#WDeP-%RVT+9V7%KmMJW#g*vTG3AfS@8aT;6cVjp%mK8x025LI
zWrc#q%~$ZqdFAoAIJpltDyldvM%a{wds2Od=UV(?;PvtT_b5tyq&fzTWdGv@AT6n^
zR7bckC=jgsjXGQ%DOJu;GH(V1GecDQ--@!3GMi53;g~1Rf|oA4G=sOmd*nbW<ujr#
zDe0diyOb101MFz>sijj_>!?;3=~Ih<2BEZ;Z%N|1@jN>_#|l+x30bNdj|Tn*CB=lS
zK1pH?@*0~-PQXaTC!AsqtwzeT0cn(9wLZ4ycPKh2d^nqX_Q?z3JBvY1QMybLUz_&8
z`UC;gjG5)Gr$+p+{cH7v_<@tW#Hk;Ik_+pwH|?p8;vDR$MeQL^xAG-4V;d6$)U*{6
zAvN}3iR6PD^DljuQUDzL%)TK)x=uTH+1;sD0_Ho{tHuqZnz+T2`Rr3};*)34r?|gr
z4S!lP-x^-HWe?Ob@Hl<^*W4EVPXFoGpy1$4k-W1{ue|@ve|22(K?1}15>;-YgLw{H
z?U8XUPue4j)Kyzub2zMUZ>q2qGW!w~{g`-kt)Zr>3hio-=?J8#PmUJC8;JjBwQ4R6
zOyjzDj2TiI#W<Y5=uhU7%E1Nsap@Gs?wv<hOENbQ-zpsz{CB2(b!t`bK4BkgzItu6
z$Ya+kd>xtj^~pD%KP{d;2C4e9vkx!F48DDUbUJK!L%J$ULb-A*t>%H(94kGWqjh+!
z$K-L`P+2PUDS2b8x|%|7+!9kYOEqVa4be%_dgyakv{0T#ucpqAcV)|c$C9=>H3w5;
z=U~uC3=V}?0N}Iq&<^qO@q#~1q%hAfq->|lB?{)NG9!p&qa8ks6EFr%l(R^IG_~2j
zAdFoAFejIfW|0+U^z~ZXX-(i0;Gf^eN(VNb+m0#1@mV$Wz&aRWJB^f=J5=qkPdM2H
zH7iXRj6I3(<`EN4y^CDQVU<SVnCz{ag~_;waZ0jQ1#aHQ<x9x%0JF7~R4vc(r+l`m
z3kxJ!IR%y@Iao{8&d!dNaQ9ZzVPiR;X)iYu9dTk#I;(DV0Y9lXGtzs$HavE<zf`N|
zeU@4nzx2v^&LgACctM%-ST~INLE<GA=Ncx-DWq(A38+sE%KPemH@kV3ZwH8OM*0J>
zw+B(LEi#h(_LX$Y6)}3N`2n97)*IS_^G6Htd^LLQ*kZm}h+1j=PI)YgZpmHtMsf6X
z)i^jesW-K&oNTM`w6}c^KGQw1#eB@mSWtOjp?yR&A@R@F?lA!i=8KuGX7Q=d%IBul
z;P?vT9oB(<cFlw37Xo3?T5XWa#q(iecME)mVr}MnOvYvO{<K%AZeI=$BNM96-ge|W
zl@;VCpJNM~I3%5pT{574aD3)tZ#>iZYLNJvfyX{UD1~pMDv7#1Ssn8xyI*{Txks^S
zZwN70N1ggugC`%tXWEv;7wG{xZ-3&_ogMe?y#DxfKMi5oT~)i95$tkyo5oc0<i-m5
zLN$7N|FCJc!tl_|VqR9P7<T{K-!!?eyPtU&s&^S8Y%Hg9LV1@L-!1s#F@+hcKK}hR
z5idI@o3YSxOeC?N@zcP>+@ZPX&FK5G0T5r@n)Y>abzwb7H}Dc>;#aw^akA<|)yV{%
z|JWVI9dr$_Gkmj!*HhZ?KRjrpqo<9XCA<@}Hv;N(YKv`vH^s;(-A$ON2=f#4$<<vz
z1INlBDAG}1UpUMz*Dhre+PYkmZA6ztD*}Ymn*NMy&iwmrwu)PAIBpLUg@U;r%MP9&
zNy~!SQanS;=x})*fi5GS<?dTX2q-fvj{;Wv-dXDH^&8j=!L2QY8Xt2K5E3xbW{bsi
z)`~5}&P2Mr5ONW0J0-bw1h2Mzx^)8LS&X6b8(YI_dqe@eP|I`Ok52QB^r5<uN_rB5
z`R7f!h~c!6Z)#>8jF;mp`*gVq*Jrf7PaVxEQ1p$N#vfqZ;6ANXa(m|18@n?ZClg~(
z2yR!@V0uGi>P2R?Eg}P0@CshoZU2@4b_7tvW+x1BH$>!>uBYqQukWtFM*Q*C8`f50
z%q%Rd+Y2;Mvo+zvTm>Q)`U`l32(b(kzz|{I-8fQ1kn3^)fZ>Qff<ID2LevV!i?(0&
zY$80PDV*a<^ha}*K(XLg3ZFM<Q!hZ_>QozNgVR`ZB(bIz`=NrWll{%&$=SZ4o~7Qx
z7ahe^$=geD$o++|rzPg6?;|ELpricM#rLze%a3cU+(s4>H?a8ZFWV+&^N4<|sCP{G
za3&M$LG8I(TQlp^Nd|TLbsg-Z&~5*x1(mrj<Ex}TGi9LS#2K^Cw+i|$FgDo(`V=zv
z5vP9r!fpP`og02jhgDXSuS^ul7WA9ubDmaNI@uk2Q`8?ep1U5-MyEs;n1qSwn3EY2
zJyT6^5)4cK{9)_m*1gL?!T+sz4j8L)$jJMl3=<XAu0gog4G#}vwLehw2ro$0P*$d_
z`Ph%GjRB1*^Ie-QQaoZsPB)x{w@1|km%p2g7VGlS($ZTWY)tkrGYn!Yl#E)#iVuD|
zEDXS<D10u1XEMIXn}(CUgR$MS^HYk))vOrM5MmJQ80&kteU=y66B^$pFLKrQOw^bc
ztH+Inj%M1n5={tQ4pa6&!z5|qkBEuMt{5OGe0@*gO`;>8^?W2Pb~k^bUTr_Cs3f11
z2k0+3yb)D^B7l8I9{*zDsu@r3aFKoE(4La0hNENoIq1|31&8=WM-vqsd<UT$9E}?1
zys$ss>q_c<wB32KJZ%qOn+t<D?3e-s>Ed#-h&&%A72etW$3xB;LQv)(YdceN_u@F8
zuxOI-n&)NJN3)^?^c>Vl+iTg)_klrEZW4zrv@LLSEQz55f%|7^<q1Qk(G<BTudy~0
z+r4ojyG=y{pT>u;1DmW0r#AN2;rS-??9Iu5X}NPJ*r#d*oFxSt9hc)np)4Ae`HeLu
zd-Rvvg5IEJ8v{3mFBp!!MzO(%()HL&#QXFDjlgk_(wN5i<9moSi{rd=Bb_Cl+LMNX
zNzSx@;ZM*tiT8fDsPh;Y_#a&Nif8(#k&;1Io!2Y|&-UZ1yV=fz@hiY<L}xhtEc;7N
zW%q@fel2GLeZSMnqQ}5br-6^xh+ZtpeD)t}>4_-}(t}C4A}L}>6YXKK5@ik!xAKWU
z-NgI#XWY;?W|dqzaue#O;z1r|YKr6O$wD%H^-zN&`OjrfF!cmdAy-^b7$}h}=&}vU
z^6nyBmqx2JWyePP8gk54x}eEnb`q(}9j%Cl2|r>of*T4W^Do``Z{m^$^!4=_Ur-zi
zOiWWz3z9!EiGp9m4k_T^NWB|O>pi}(a1qo)k`|@NwcmHb!I2*RqL^#K9UPyWOtyhW
zYuFTXmo24~t99bwfX2NXjfBlR*4hts1IwZfnS4S@&AQ99o;TLemkC1lZVgCDQgL#U
z0Q;;P=q4{-lP*ebW@?(Z9w|SJct6Kd=_LalR_~YZ{k4(m#`U*#yR5Kf9UUz#4Hxqx
z;&@^SNFC{FmTEr|DW<PS`muF>-s;EDa27jasD)h2j<cqMjI*56VQ1sUPeFkU$^Ex*
zOUvA=^weQxe`aWfCa9LiSoMMeKSniND;Uq~Xc;hmmb_7>W#T}?@jxWS;1<i1w>+ns
zB+|zXu4uvE_H@m&ojB*`=d_L?<+L0eTBDCQLhQGrdVo92BHhowK!d6E_+C_I7q@=w
zy(iQ(aFR9nr%!XMZA@+sJ$+r`rp}d+cFNe<R}tR@rlwol+a=&;fq6$9GUL3@WyT%9
zFTT^3AFg7FHS9Hhm=7i{w*496S6$*W{R60SKH*r}RgnuX-lSyx!8{cc(>2-i5ASJ2
z$bm*8RfwarFrr;i&dgR+`S`AsloY9`5AL-TqR*gF`a*E6AteXL(4B)K=l4!dt>tFD
ze~{6PM5n*r;I7W5n$-}=9P8KEks;mJS?e)lx~pTEU#a|3H&Yst$r?eiV=)zGInh4N
zZR@K8GfzweE1dxnNhD%JbbfMb?{(T=lc!hPBP`IVL&I37O6==$R>%)oBm%yUTZW>d
zy7+Qsk<iIt_QpJeoy>3%<KqTA#;ERY#()6+u(`RnN5?0zs6#n*O)v#e@Kydj%Njcc
zFX+~E?nODX2FpOPbC6Lt@=QNgj|rZXOv+(1@yGLo(Ei!X=i~-5FE6kAgoGXcsP%E{
zyDq_Cv-t2&e#P%#vR@0tG(gvUjRJK^ARE~TleV(~+X$q44h2`C=N>IMDCW*ntt~GY
zy^f>xT3R}w_-p2bho?p9xP~YTE&;Rp4<N>mn>IMvvs3U33Q`t{y*@g5Z(I30`uq3m
z_)7TuK`c#Q#(|%^d`B~|=NT9-ANV5t?~wI-jU!8zz(w8po~6s}7M0Zt+L%obDhagZ
z)3eYpGh4{gcZ^e)J5PFij!c(mSEA4&%T4`p3sXs+gly%kOfsZ0HGSh_C9F$xEm@FN
zsF9_qon4rwRhix7m^Nb=^x_^1znq%O_c8l-<^-^dhi(lSFC;U(ba`S3wsnmPap$xD
z>uHy3?0*wX+lkb4#U8Eozef>{?~si8t;nIsnC>D%xgOK8ZxYgDi8+yB@dVblp-CS5
z@jUj#<!s@Kxm>1CRJVuMh_vJ)^3jV&%=<Xx6fAy|9sl-?xqi?2AtI$P_%6hG>!Z2b
zb3SfF3JVjm)<kN?_W~0VdqcgLB*ac1O3D8d3P<Z&Drd{y-?SgSMZn+LC^35HHnxqq
zO1?2I^Guy*sYfuO)sL?J{{FGoDjTGvr0QPYm@{?ZPeE~og?Gtq5Rwgaz3dNE4E|j}
zH~;@vgOh(y#`$SV3OR_6PyD-I72LovX+7E3thLV;1NOJkFD#j0&pZP54PyPfYvZM<
z&~Ik)Dcpe(Rrf9ErRw=S$~881E$8Mp!$8?@*I4UIaZ}Kz7%bEOXOc<(E(y7?<e0Eg
zF3g9^3C?q$;0U7)LTh6?`%q`RDz`}c-^7DI<rU5U_tIf6--8D+0;H?YE}#AHw+v3u
z%je$-{`*6H`Tc*|?fv^=Js`a{6<^k`BM+~z7?|scAp*;oD$nZvFVwh2+nqJgEbb=}
zS?JN{>#ODEb6YE43tC!k-M(FmzN=SH>bGk_d!TiHZY#V1UXW=Xipn;H$>uI}(!Ad_
z|0gLaNhQ~g%>8FwqIc?(jEwi^eZ+cA55k*2yu7Ye=x@~m*tM5j85!)qjn$vAb+%n`
z{g4{YZuR-|W1`Y1dL?Q>K@z@4k5t&6-%ocz<yoKj=C7^s|MV3t0e$pjWjSXM!piBd
zJh-i9h{EpOGpETOvE=taJ<FiZ{1%-Q&=bRy>6#4vbGTpDy=u@v5-b<TM#aiXu(`Q;
z%X;u-kmbj6XA(sjTy}A4M%vpNOgQP1{Rs?Dh?M7Ccj6^Tg~C=>878KM(j1q13T*1?
zYGb<^k#~pU*ygSJ2(of>TSj`<t6i=fzZggo`C`4jEv=yNr6%r4PYiQ7odU`10U~!s
zw!JJ<>i(lg*FZ1?i}&-1kx>oiMK2TM*C*`y;u|r{(n@l#W9Hso$GJgZX`hwiidww=
zPLS-?I`ZokU(pI!XlQV2ZN<uafo0AaJ$<-GsbaFA=2OFeFRiXou4MwxW`V}$<Cr2)
z2W2#tn`#Vvk^zCC8W%#9nztZ_5TY$bNKLcYGhH{}9qdcw!}mPc=m>n2`<_`?L87g-
zAJ1idR1CNvIUs#R6&!ldy{7M?7oMd2@2~JhMa}<S{p=q~@)q<8A>vy0;1}YMq7hwd
zMk5NJ7%v#Bq`$i1b+BgfZn8YPd&OtBpYP34>nqZGq9LOpxb8op87i6nM4ZkHDP+eR
z4Ktytw8AuxoR<4KQ1GTX(>9qldZfC&QjZDnu>W}19iFF}pE0DHB8a|LbiQN(zz8WZ
zl;y}BlJ2k|>FzPZzfb(&hyAQb-|?A~&15-oT&W(9__Lp~1eMm~_tcQ7L0}6Hp?1MY
z{$uWQM8$j7Xn?n|e&zbnNrqXWraS0bF<M<sxjE%>R%w+jD4Ejq{_4s=iiD*3%~#f6
zjrS9H-n+Q+KD|Lqbh0<1{ib4BOFW4`pGp3s&pFDS!*yeKbvQe?8$FOB6b`a{(BYpo
zpx#|YSd*3E7Jjv*-#C3^*y(1WHBW?+qD)I}s*gX&cIMA&odQBa#32yD1Rnb^4Q8!`
z89FZJyIYMyw?PlI_H*x3yw1BgLPF}L26!O9Z2o+^G*Mw9Vs|;KJF3iW%RnzO>&_J_
zo`OuZPQ~xMnAb<g?z->8<zk<O_oZHt4Wx>OgSI6FA5Yp61Kj~i2A$~_?s|vCeLOs*
zQ|O4K4Ti7m_Sej2z-Yq1du{sTHJNPA)G{?XY>&g|%##O|UmQ%BeE&0b-5mmnhQE(}
z7Ir(1rP?CD3mp0A2I3Glx)OuU5tnX_5dYtB^;HIH_8He%Xa37-$p)fM>n<K1oK#=0
z@*JpvcL~`Bm-~~g-9gBm$9;db@pMy&AQw3QtXri*2C)X122qy+*9(lo=z?SX`TSQM
z9f<!>@L4&eYI$ufU!d;iI}<~9vD3BO)Jivch2ej$U4y$;h$L$S9aQcSy>Jhps=j*l
z<5X;aqCz7dW3M6`8&gMV)46{&IPn1;%7;{+_VDm1bX1(&<JjNc9!l|9Cbk{em@GI8
zCFYVWRI3Sdoog1$ly2DD>1Aqs*lS7Ql?HsPqE;}GL%oCbBhc2M+P30&)gUZzXsXiI
zgkLhm7tKSEt7S-5s*$HGCXCFMdy<J>bmxHVz1u(h_U%48GhB2mw)9O29%M&m{|$Vu
zK*!zvNO&?gvf2Vh&dR{h^(VPhvp4}-qF+}Xo8;KLQ70p-sOb>9g4Eg)K`(@hYL-!O
z7=Jc81uVf++qm!Fv#PcR#e}6q>b9P+3^tyhSnmsTn<NXljb*SIR3;pOMcZ;;SBsCY
z^%z(&m(6QxJ-Aear3UoF57C7RkS$>iNzX%{i9b5%sVbi1W4&5OxN_6#SaD+Y+m?UO
zz&);>8u4PRK??uwmP5BrlaC)29XWZ5PHnRz>Y#U_!!qB}@{13<^?31!o}NBlY1O+S
z>L}PW)GS!D(3SNy#&+np#FRyaC|8@_d1?O1E#k5<R};&I`vl!%#ky5eCwtv$qIp_P
zDZEC%#9hwmckfY%6rm3L6L>US+`Z4THjzNL(lBIV@K=qCgnnykYMP%pc)5-<TI;{^
zWTd2{!wMMm2dzen(yJh8F$XZwi|^hgqXM)k!#?nu{w&3Pu$E`0lNrpiiewPGExh<7
zD9w|d2|#$Uo(!8s*&ku@M=Em^60Pj)HvJo~Gs0A188*YB3@*}^_NWD>IG~dkxT>3~
zS9VM52REfFBppG=@*HU3lHY8z8kjq+bWu+NLhEw8_u*Pope>4eog`{*2LTKVb*}5Y
zE*C$o$E#|PKDyR}KR%)?B8W7L-pH(vtf<(I<S3Ze2ss@jp;{wg(ia{w6ra4pfMdU=
z*#Ifz=o7iM%|#<suvJpImhGEn-{fOHt0rllbI_4!pAVxv+<rBsIXuo|wZ}m4+G(+#
z(H1$E<a4XG%hi#s!H1@$;5YN^mP{0VVI$fVJ)Q5G?G-NiU8F&J#hr0|7>9Yk63{@_
zg}xa0hEsL}3+9&U7;UxcK`ROOAhuir&i8t`%t+ewRns{J-eS7?!^lGragY3k=_=85
zgD3uLiw=PJHKfqly>S}3cIL5)$<n1A9r^x2Z7@XZYay>i>69nm29RILA<`Jk^j0H}
zI%IK?>YAIPK=Q7mYW4+va`nSR#;vAOW33WB4|rLkp=PY2hH+K|L#khjmk6^~iTX5H
z8(JD&wurOiU5yvNe*IGSh$_$P4K(U0R@bQB)w0TZ;pusrhh7x=mZ(m>#1(n(u`WHq
ztCb|l6z-q<oo@sryC@U6A<z7epN>1u!eE9xjY>)U=^}SobB);vBqx_aEdtcmq?<(A
zOt<lHM_81CIKNoGZtM?+^<CQTR?~%MSzJ%8GsBuAd7PM_jAX?4H%W!_7P@1@6EXf{
zU&wGkMA)O_=|x7hBD%(h;V_X?b?@+OBR0IJfr-v-@ich%WO^cwx(*z@z5#DwY}Tzc
z(@2vK+=@G%J{uDoSI<lIcS*ILF}@THXaG)V&&Q#^-Cj_?N(8)qfmFALp)vI-gWsPo
zWUJANpe4(<k8W!l>erHk74V%=>&~TN6s2Lake%ZmH_3BzTQYJ&<MnIdHZawEkj2!s
zK0`!yK12M*;_<hUn+TK0l7FGQZuJi%>1Y`6SL2bPa}<5p>yL~ND1kE^7eV$d$QUeH
z4czSX$mx=%bJcJ;!r!&KE4ydF>EG(a#I355Zo@f86)&vX9rZEw(@nwbc~;-J4Q&ji
zik&KlgVSLb*Db5UXEeG@9Khnl;!S&&ZqN`?{Y}xMdn&SNYr1AO$AmeC$NNTqG!PgQ
zNcp2ydKc7k6}Xt1e7B~uNm5jOmWWKOO)u>E1{q_F*4(-H7>FHye`Y!y{?n%~NjnxP
z!5*5%-jrlfnas_$#3SUok$3Y(vBc)XgYF_FJE6h+rD-}*%-ecWU#F<6K5NpE<Fh}9
z_?fi@lkLWpS=X~v^hZA`5?4km*2WZS5(^&4<2m<yXcUQ>R(}}oEW3^)wpMN9Jr{x2
zW~%yMAk#eq!3-ok4W56K9&GN_OX#Va3XNlEA`_~nS>7QS#lw?OUh0v=evDw9k@rgl
z8X*X>KL-b5F9#s_@!IkJPre-wxSj9EK>>w&?N$%pcx_QCV1{s9JWs1_rRC+B+C8-$
zMUIbX#FJhGf-;(BWaHfAR2-XbPKj=%jpa@)IMzX2Qj-q!pwz7PbByyj+m@>0wOQ(k
zdH!xN3e|3Ei&y*DbRlOJZg_gQZn4J_?<vZ{B9AbzG!GTko!{CjJ>V8*MJ$qAU^x$2
z0wyBzC>Msqt=U{Vx2>`Myt6TU>kmJE+$=cjA@Qn~WNw432L=Wn&DyV15;S#$k`QAW
z^p5YWj@ELqb+pax#jE*u+RxMn3ZEX(@c(4eseETviNGi%7C8L|7NhrXFMJ9E&ytJ0
zRR($Cg%+j3(H9jub*`a0`0bP3xAp5tO2bbkGQ;!7Lj(BJ$Mni1%Oz!0&Q7<QzGIG8
zRz=f9PxjHld$=Q=52-q@X{&(_TL*vko*R7K{iw-0zxf+Dp<{K;Yy!JZx{uvv8#kaZ
zoRQpC+HpZd`Ay$La^X=lOf<E1_-zMpx&%C+kxWx<53)k@;&=*GEvFP+97v~%tPWK%
z4|8!jY;;AuV5eM9usz#Vs<c7k>O-A*iXzyo75J=?|0vxf`K)F!U>uG&rqCa;Ph&A~
zb_?$3w>_3<PjEB?FZ5NhP8l!M!l#iEv3A!JNd;}F(p{w$^a4VDF{QL_Pg$3?v=7||
z#~1lqImv2ApJZ0X10=3KInZ?T<kyY65s|S$AS2D&dTCeMc*1n%<@Ns0p9dSWzH`V8
zbw-enUCqzG2k?P<zWUJnkQftjfHG8S6!l$v4I5ntQ+>|^+oDZFJk3b0;03ntntF69
zy~#a@#A2A{L7VK>&?EQ4a3gpc6Oo`P!<5_V2P=(2>gQ>5RXh?9NxCuni8NzT)79sd
zHj~4HV?cKj2S9!vJ^vAJoo{Y6BGVt81zQHbeRW$Tpj*{3qF{Tpmrr@2jhL=ByZgID
zzozeA-1ETBVD^Th#?Fsd&%tVkqMv>{*9iR5CRH7`;76*v<O)hh{`|Vjp4a}{clHMr
z%r#zH`%SrNV9?g*ak@o;67?eGtFTa#Y&%#2PoaD#cMa49^|tRj$2(*_aZemOj=lXB
zbyLUQ8b#i|vLR&6OubmAe`7Z*xz5+nwP`%jzM=SX4|H9d(m*6$V9x4hnAh6@|KSxz
zE)<5VPe|_dqpEw5+h&Fc1op1cwfhdgueWN6tEfcW5_OigRj+d;YwZBuUH={<v<?;B
z6KDJqk|HL)ulN)lp`Smi*80*~BJ7*y>$a8`j)yXb>Lq%o&n<{4T^-+xM|I>pA+*Od
zUqhYzZh!ZTmNpLlW2VY*yhw$8sMyo3>H`Mqvi|(MYTJDYuJs3Tw)Ot^qRb+Nrl8XS
ziyhsJdd+mDrO6KKi$THD@8u>d2dedlv4`)skhk`BJAY9S{$A{kVzLxt>MQ?_0Rqee
za`Jmx^z?gApAY^qFlICC=zf$tx3YwH@!@6WAp^lqS{f7TK=k6wM7V`X5>TA{&PM?x
zKZ^9a`2#zrdmrsP-CPKlN>UD$+lH0sqXe5puefa0!QKgZdg4@C!S9UQ%`{I|SoBD9
z%OXMjLEwEOJd(xZS&;7eDUQ4cuv@+{7vv~+r95%=$tFQ)9DDMeT2@Y}6VUuBRkCLC
ziM$H7dV?>UNE0KX=1d9<029Ce;QnfOuU5&_o=12<58K}?Aoxj?b1=nUx}+TtOV}@3
zkQrn@#SskL32Y!KB+t_<hIS8L*SP(uLe7<N9AvOWjnq23>!tkKNo8*T@d9{leD4V<
zR;kD_S}5w1=V_(`{G38#UfoxEuLke;+rMeEvXM+*JhdE+OY~Z#xGX52KYwXvzXTiP
zUV^;(CyA6Sl;k%AUSJV-yqIOG7aCd>t4uH%LGu+(7avc&3a_9iUmwR8Wzz=T<(?S#
z?V1c_bz_czv3s|E1OA;TYl+*(Y7Cd-n7#1kjeObJTBjAcN%FWNkYUgF^BG7roT4%$
zf(eO^H>ACe=5cX}If-B|Ft}hzhiA*h{hA%`{L>g28TnpVggnNN>1kNgaSXGefv_X8
zW5HSx0!biAi@PdEL9xIhPJF#Uol3}`1CS3~Ty<RRoSZ@_<mBWit$W7;{QPNF?_FOr
zoOgtJLc_D+z#a2v+JC=51vAqX;ej-BR?j-o17x%10Ya-l{eiZ&c0gEIdH<DV>Za%4
z?uB+o(d8G-e}c-G{t!8bxaPPj$1wfi8j=A$2_$|(&EwX&Eclg`#Le*ky<Ffj#e&xS
zfN$JvXnBV41e9sJAGmy52f_>Dypp~7Hxp2U5(Dat2d<OJW;e`?zm>t8&VxWfO(7#)
zN9fbs9F4)=;k9oODe2|xq+|*d38bVO^NSmw_bRzpgT70c*fw6Adv2V#@iiW!m`96@
zAj-(cjL=ALv!KGnM2LEM443Uso7`u)dF!;nzMp+RP%IC```t7$l9rcMRZ~)k=e%KJ
z2)Rwbtnn&Uid+TO-_`dy$1G09X|kJ5U_{aMd-u&MT38r)XmA``;o3JJg9aV_Z&BX+
zJIo7%yXgiV8R=`>E;g>Nj{eEXYox5XNsj#s<(Aq4*W$uBSwrJZilWpGC><e^{^>7z
z9f%9`5@DfE4&7|i@$uTRvk{z<<zG_0c|D<3PYk>s6gvto?QuGiES(S~{NCIwv7#rA
zk2m9g=%yqn-E@AS`(i8?-yiR(&-8c{XHtxs*0-YsLC)SilY93jXK~gl>8P2kgA%fs
z)qbeVbI8=V&sav5oH=Iykb&+n>*z*{%D9~x`nJq=O1b=?pM_@M3>Hn^zb&~W&v+Xo
z+<_u8ZXn$PUDb20*$<|rpR_LzpoNsvjc>@OiXcoQH@y$xMv<Fj(5Pqn0wvF74mSy;
zBJK9t<>ZI^W3q%U5=&VlvzX-`D;oedkn^?H%UB7c9G$d!Ub~q93#l;m<S+{Fy@-P6
z!^x91_NgQ`o>^}8FSgs#sGIz2%k@VYiHc%G1UY~3V*J)X>ixq=sm^8He>IQ6tLE)h
z^S-0O=Y!e|FKif8sS=M)){k7!I?n4e#F&#qsoLh3pdZVF;5e3!K4PxXJ#vr)Fsaku
zA^;Uf+?pp&iOKpybWMz_Jht(sgW)8?cUGurILt>G<P*5N`CY&8KVRi0Pc&1A;xq6i
z^FfMy=gM7O?~4yEXWA_q0pxTCSi44%cH2h(<I4eO?rHyArsfFD0`e_BwTj3|kn^v8
z@u>6&My9>wFoTp7Ttlj^R~xTy;<tphiF+M%)4QLn<%aS`hOj6pDOF7zv2qcY60fcO
zjB%7>EwL>rH$j^U6Loh)+oXZ|&}XpUqmW_By**BY4>N2Q4fo=2wnabPo+P$arE=@G
z<e#GU=6(P2gKvtMV#D@AXW_?zyHBVc0NGaM6DWT>r&PX|<CnUHnz8Z0-b;foZ%B~|
zV1-qaEPCZh;q!QF)HYvfMT#J6N)yOeeV_8a-JF4OY%V#V9O&t#c+llV@Q4^Q=&vIC
zMbCJK)9d^z+>%+ZZ0thgqoX(0lxAjTa6=srld7otcz$KSDq)|_36rXp9b5J|7rkf^
ztvh_X=LE}Ni&sz(U5I!z`}v=o$~qrb13VGWJjqO3$X3JQeFxNI-eSYj0bhutfs^aN
z=Hj0|EqW(MXHM>4YLStn$}i7vnm9T><m9Xn60cV<4Sw_>hCqIp70~9aBL4%^TdZ_z
zv4FuQqyUP7kX~jKLuy)D0qWyIXIMOL|0Gv#L~KH`z)%{MFP)FjbhX|8sqMVuss7(S
zUin5tNkVp+MabS1A+m+CjuEoQF|y7{va&L>NA^4+*~B5+vG?BRaO^$rH~Rg4_x-s4
zy>EZiqj1jod_M1Sy<e~AbzPFDCpdQ!KM-+=#RY+aJL9t?oHkjBw;!>wb<9Go=to}|
z$Z~>JFC>iNovVtf^;BmFo!LPm$OTLR6_veTwSQgSYha>J(m?1wg5+1W&~}F{za`7%
z|HifI3L_+J?4uK{t~QHd(IO^^th)U?$#0wc#mkpFTcXiqWTlSSm~vL5ZdMbUy{0Lo
z@pNpJsO*{R!~09cYIk)c!>UG9UmQHmhk<`C<KX?KcHBETyz1Rp<?aT&1xcO{;R!Kj
z8^!F6+Uf(KVJSD+S_s+%DA$w2JTr#5xoGH=_|6t5kGZUK2X!t2P?qHivT`phMtM@O
zg@yKJRpePpkt`+$LWbt4altMt0v01>L?~gND>QP4zn9r9%?9%G3^EcOpNI(<Yx~#a
zr@y^%J<BL;J$yEVPP(oSn~e0H6Ig9Fa!^n|hN46>RtU(qQb9!orOdBfkH~E|Qf0TA
z*_k<2<eh!4q+8L^GPeUQ@5qh`-ElBGQ=!luk;=%3=9C#wxdECjZON1?$pkuw*~^1T
z^=nk7{3iU(=M1TUXJ==~vXi-QeVJM%`443@z!vXxw#}dI7q1diD-8AvD`dOd!GD!B
zp2ce9e5U^p@bMiLn1(fXE{|7H$TT-+zR8cngfneAE)OPij=!_9Yng5KhL}XtMg;T=
z+#IJ7u)Ej1PtI-kRm_&w^+EpIBIC8q-IC^}r6N=zw|x6HHQ=nK?mDci>_AfukN(`^
zDSg89vWj}lH4}U0CWax78x!H#OM<``FwM76E6LaCONh;>WMDLIe2q&zI#$Hu!l_64
zn&%Ny5Ikzs>M#h4DF;`>OED6A{d%bbO9fY^N(3`wBGLUjm!zffv;(T&f5<hnQR~ft
z&_oMTb;R7y46e4n9zx2NIRfO_pM!_&t->P0>w2-grR1bO&;=A3wi!g|%VQ-PWBFPO
zqDc0M73*QSv7j<LVjm;{nI%)=LbWFzA*D7kKwpcupMu5!CJ8X|*b+<RZ`d)%F*15>
zgL9%_wAxaOe^6+03thx06a%VG=LVtKI@sP1RC{_&xT8~BTR4wm=ooo<QO=5NXn`JD
z8K0k{HFgYte$Y5zK4%wa{jwVRl>1}`omGcK+Vb$;mZ#t|CJ`l!koh}N|K)K3lXR8p
z??hz>6IBar4lNn423E|bwqu){%?Rg@&dwBqeNv*kloA>rYn-rklMg-B8fQ|S#+AiJ
zYU1bHV`|%nBpqHRh&l+OiVP8d493cry%i4~v6C1J7iA}NCbC$vtSn=8@pS1uLnN*P
z{P%AUZ{NiQF9ML7fa`4+=7rFB0!i*qa?V_k;3d<ls+I7m>l&wCU+4Sbx=Wo>!~VAL
zjSk*F|0F5T{PAwY+gd^>Wc!n{at<4VL1{LrfE$oNDcrjuq%6C?k9Qx}{p1L_^>J-Q
z#I=t@o%{(9bS(O^Bg#wM8g8iRLWMp|UfRKq1Z(n`JxF!5jP*BPLA$ickIkUx{G_R2
zZ^ll{7a&reX38PT%-6+n-NYqM#k@m9j_N;NHv4FQ7(^w7I9$R&dL52k^xt<C8Wmtw
zO;rZbUqRr<VX$Xo0ZyF4qdBC;sA!y;k$0M+o#i~t>e)%@mo+QjCwEPppZx)_ZdHt9
z>v+IpzpQ6RMJKR|9jrz!JFj-dnGTY4BqCO<$18_^U_A+bNsb97GpFp#`dv#I7%zY_
z*>^rQsK*Y-Kcr<21k>2<Y$^jK!x=2RV6hz+0?!vn^sF$TY?J__GCa5jK|t-k%(0y$
zr%JaO4gr4N65rjm?E%Ql?yVYR^Z4acvP~NAjsTF2IF)6R@Mn%iV;z*1FaW5znex&o
zX*92-L5x=WSIzr^_K#@Ub?d)DIwb1-h%oi84>|Apld#wIQwepR8L|_(&gc;|=j6#@
zY?9?tKflBLlXQjXPyXKw?<Cr&>I9Q<k(YG$be7s03U4VX#s-Dqe}C2JZEsh3dUEfD
zlyr{a$9F{30xtJJBlG-hVX^HkRcB^W#b7rQP>i}B4PAKZ+;T7C`Tayy7^wVUx$VMv
z?e65Nlj82GZm5*QVceh5dOyU>Zx2tP<ujSC<(+J6<N`+Z*nRn?;bx^Odih$GGL?Vi
zlUkHmp30nh5W(Rt&F{)G=-GL_03}91V?7d8Co#=1r`(}zf->Su>gn#W*=o{vSPnO|
zSoz}<K6IN(h!)sOZIIGCt>4&HjWi9|QSQLw>fNM<%@Rt?m%u7ckC+bKN^EMc<myzY
zQ&pt7tqv8D&itvmVj?ZQh21S}zPi~;YZ|HT`$c6vV%55Bj>U}KDlK(#xoi_d$!Loj
z%X+L@6!^jHAK1tb-H0BJl*Y+Ffge(!6J$j3RMwU2-Uo@Q849X0gm)nV^G6%4?0dG&
z!8SjNkl=~r(U%Y5uk%B3IXH#_h4v-ey6B{PXH>j<y5$uG)b4GVf9(LW1S<$oy?Zhg
za|UJifsw%0e^ibO(;Z(NfKTe2veSNd^|&|yoF58+!gvNDaQ?+A02~l$SNJuHSl4$&
zR>iKczXT%I-qj$vxwX}Mb_?=A1uIN-aBwg${KO5n^b(hGn^HmOD?PoWc`P;sm@x}w
z{AH!@p~k3wu<<$TJ`xtzHXkborDUYq$%ly)@w>@+%raMii>{?r1Cw|>GSV$MO}~Bu
zj1jqN1P<P9v*cuuM`AaIfWwi;^X;+b00Jn%;<A664cF2{C5WhLr&hLbBms8?Qyo5*
zkpLT>b!5@V>(16d<R=XLG*!i}SGyMjp(*Roa`;yluo%WuM3_bL8zwpdwe+4}3KS%G
zF=h<}{fE^^Zvwu3Sj_?E73+fsN+PQs{REjCT5Lb!%GpaS8LELbGTfOg!NbSrwC(=Q
z++BDsjv=@i3+fb+SysGXr)_T>&^Xo!Dz?jKfmBzPi!C-Y72i}=mKsS-GkFhESMm=m
z;($-)<Fegd?zEb}(rPryf(Y+S_!h87aUCPHikVM0K*tej9j~hyX!De2@3dn2ym#pu
z!5$BOPfeg5J{|_r(mQ2ao7%XCgKIr~GA(tqb&AKUT+P9A<)ws=CknO$w_uB-nFew^
z6Tz(34TwMJ$Uq2rJsqg?Zb|w?TW>?)8yiHu(ZD$fxG((~sFYz(OX{}#p8W$e(1fYS
zp0Om=Ew^~=jVAjGubmyC)l5T8VBZw`5;6S*_J^AL67Y~z8eeBIfy1`pv-Yph5Ao#=
zukq6qTn^gEu!~(O&TK$-QRirv9|i;-YSXL&l$SScq@#UJHGSlDb6A=No=d$7$Z-;P
zv>jMIRKyKl1LMFF*r|*YU#khErdez&VRX-R#E#pzFzetXKBU(hP__OimYG_`C%n9p
z5-0Tpy4OUg#uRaIz*HpKT%-eh1lCebp5n38-Eq}M-z9-L4b1+=UJ*(}L(p-JI+XWo
zz4&w~{PZ$0#80)i!9SZ700URUwmSi~K@r3Phyb;-<zb_=I&sP$B{slWWRXI`ydJvo
zXd#mZu5GZm8g@`kDKvMVrLlK(w#-g9YynoI=fT`_*VWkyB);_m^_#3DELHK<iF%$U
zk0rQam_gP#&>%x38#}c!l{{SEFD~1;-}I6gvVQ_9<0=;;Qfao^hM-^xC<upHi9B{X
z5Vd}sLBVIm>AVuwxN!&&$6m+h&!1y9!}eM@7C?NgzJ1FS)#BV2bAQ_Q;O7ozZ{6(F
zC-{DH(DjYdaD8;jV|;W;DE8Zvw+0(SGZ5GIPDMvd?_*-E$2AbQb+0doEqKEpouceL
z5+)&euHs75A(|T2tE(O@{L#}hhgUcI5MBx|Ui@z8yLyX)R~+Hvk}Ms(a`fTm57QaO
zHprR?5IjTDdSs-$y27l}(hcf;2|p|*`vkzQ($|FV1HbkM{1g@^?0lu~Q%<sN7~>+q
z12i>F`_*j8L=FBqa&tdt3zkHh5aE&RkRXt6`B0GgMqjpuG>GOq5fHJ@bGfF^lxpm?
zwvbnrM=A&!gK0dAif)43ct3!YXY$8t#_24NmJ!wDzfFNf)AuNS?#&Kll#iJNSr#ot
zcgk4(0VeUqsP#3;e1m^iDGV!kgma{vr|Iz1as(G_bvzYp_133{G$%)!sO3G~s!s>o
z$j)q|o5CE6GO^L2tiY6qbj17Xz)w>O)%8vA034~d-!(@Gog45DW0Dv5#Jv^pYogRP
z<-p4)x0%j-yEjLJhE<DO`qTrbJyN;(g>JE(z%Pq2Z@?4U*xS8`s;^dK!KdOK%}&P;
z3z!-z5|h|VwjqWz5C1GFY7IijNv8BnqUSs^5?ydvHQ%A0KB$p{9p;tvE<8}}d%qzN
z^}sv}6jM>ls||^*-#y%)n4W+;&46z6*&WI?y6*PaVdCJqHcox{*?|a`gT48nSRk$(
z5Szd3I4}Vv?#6{J#><HjK@>jf1j2e7(ROVi5N^sr_%kY0Rb^$@4`n*x(F#66rxgo_
z_2N4t@uX|xdajW-<U96}J_L=R5=mWO^K~jA2bHBT>|pV$JX>D#{wF-DgwedvdAZ8%
zl?se?wXW_<MbxtsSYhuLw;LyvF@fe#JOR2lh4eY|$Ct-T2-I?3OSO{Z*&OXDQwH*T
zkk{5maV<c1sP6r4?jv4d%ijZiOym-RjnXqcaPTIVy>4c{&5IFq@}Ut-UEwnwyaPxP
zjw8E{NSU!>$V6j->89~E%M93*A52LJJ04eP#aH@(UWcm-D84V9*@>_^b8z-X6H$11
zWx+gj;8=0(BZr@6U6epj^>0^nX64EYAQQPEwvd}MkfbZpd;6B!cJmCALiXNmm+9+A
zKRWjYEFth?;&GQHDtdZ`@6J7k%78BWWjU0cZW;-30L6n)jMFSdKZS+t=+*NMt}7@g
zG&MCn^Ja6^R1h#38WGAIita~f`LGFa6)dvq7TyQxcC?|^W8(hl^2%9uM$3cMR0BkJ
zJ8P^VZv2M0pFB76yI8Q<S*jBn@Gz9LE*A@zYBA~OwkteKVT!W6*GTA#w83n=7t3q;
zsVIy~rAH;%+OK6@Eu-~QVojnTjmU6Du#0w3Ub~jjZe9y{5s*1vqWEHi($RPNAAlOL
z%iNelP&DUJBtkRGm>L5-t(d!S`uD<TP@nyX)<La-NOTk^DvwQ}4uQ1J2O*Z(?O+XM
z#_i-hYVOZdjn@}bf~SEVF*$;$sq>wp3rJusvI?XUA~6Sf>O~}AV%@eKFj!wrs4~m?
zWL<qAyFKzl-<TXIz*Z8>fdvH`Q<ICnEyOHj4ho?h=+BC;UR7AWd*@*b^^|NPdtXjX
znnMCb)vO6cGg47;x_^4|2UZbE?t~Sz8n3(#P*5Htu4u-_k2xBM-{ZBHFQ1&O**LMk
zPE9q`xyyKTd_3hE`}mnfME!!4_?2!I>)M7sdHJk&Rv)?ez-yDc5;qslzi#;Abt0GR
z$9*64>uX;L$@QCES+UxQW8et)nAR!*KB)M<c%u?|&^JRi<ah@)Xt?U>AED3?8IzDt
zUO(SttKOl1a9KC&L;9>%V(UHArq*)Z`B7PURU@r_<<RE~pE^Q}-kub<ZLlmm;|ol9
z1#ID$mQ_qR{?<9Mwy<~-7pJ3CrXtCE$KQL3`7Q8tLx4Qh!2u+y|HxOnD<IlB!)!o2
z&ng7s_X|Al>eMrnW&_vAm1{s}50WY01;@5?y&Z_RPT=2v;X$jRS`UQ_5RCcSNI!%B
z|34%M@XT|@@b4eW(a3*&g1<l-GxXnA@s~I>x($ZJe@7NG+b)R32wC7G4c44q>VTB~
zCiu@!1Pq$S7zZjBw8VqgWNre9Vm6C5{FGFb@?x)gCMOp^I%0=YbgNv_L?c7Gx~^rZ
z-6Yt>ctx;i)YgvY->YPyrH!q}u{pbcfR4lup&wh5m|7!PlioJ#+v@^&<Wnw@WboNu
zN4^YRtdsO#o9Rv~aUUoWjd{z6E65YCpp!w*G_koWe)0Jmki((&>^ju}k$E(rCpBT0
zNZt7*vpi^jq@h7tOc%V3_B;)s4%Esrb#|09wzaXh|Lwkw`IbfXt7Jj$@ZweaKY|{~
zFJ6zH-<)?w!0HPG3)zO@-?qUjUjO6UTSKW{go!ppVdL(2mpwe9wmO6cE1kI_Ss_7H
zt~*JN?;pABEPD_lrK8-_K#aL}kb@|IPw3Wb`#neE-s-Ayb;2j#qF=rlm2WasXas5+
zG2Ql}dEw?JY&LL1GMIYosZA$9sXr_@=f}s#o8eoJoOS-quiK#-Q7$r;7w#=$0vcPx
z^hFok*SH*K9Q_xZ^T%j0*#4PfZ(|8DHG{qip|K7~d`id`q`*wHy;xwl=e%up^b=zO
z4nuz)y#z1$domf>Ysut_SKPinL-?lL;A-{$4n?lkAw^}S@TLV3bneE%?y3*man)nC
zTyRk6cWo_ohGJr^HAR2ULIAl4Zkf-^o}M0#^=-c>(4n!M9Vd~HN%3q|(rfL{d!z3A
zIQU}RWZLV4XG2JgcRoPhMg#nUmV@IuUs&kIa#T8ca(WpbetL<2*1U;<p248SRrRSk
z*vQtV{V@dqy_x}0M4h)h1&8hQQWVFG1agyrlHd2Cgimd4$L$$iYM<S7Igg^S6qh3}
z(nf;;u~UP3Rl2pFY&CDocY@yukv=)kJLctA8-d>4l`c=QU6P|=SYR`C<#TVY(U!i&
zSOHFkLcGf9!jK8j@5XQvQ3-+U)q7Vc;1j+O_0vy{)b0TdQRf%ZBzu0d+otNLz`#*A
z?)pxnB$i99$!#6E&|PWDB`om%j=~E{zG3~WJ0QGhd)o9E_=-MY>B<>-_3fCHl)dr(
z39sn}{rgM6eD^{3h$(r-eFsH9_w=U-uB~l1#5gSd0AL}K7qlRH1He%{jRw$n?iTnV
zw|%>)r}|gQ9hbfwmHq4p3X0;jBo{rh5fm2}S4vZnbU57R8-HdXTj;|2cg@iP;$e4x
zo4X&-$Uvq^Q@5$qVVIB%-{Z%GeAbM5ps@F~e3gc;sj*rs^keHYy0w9y|0y05gP(c=
zgpVxL9@?6idbmIXf55;%1~3KRF9CHwAFH-8qVu&2#2L`kLN51AIuejM?7^xyg$HNt
zq?uvoXmcT_hlj`e-XDbI-0#0GbpO!D;Za(|%v=3L>L%ziX6&2x=aP~0U6arrb+7>N
z1EwHz(Vzdr+Q;(am5SB-mUO2n@*}@6JD~o4M!srKPoKRQCr6i-b#)r-t>&9Ra-J6%
z9r1w)tUaf`z`J<B#p3_>>15*n9c2nSU~arKIy(cm9WBz)Gt4KJt9zz|5xG@X1~?9a
zFglQekCV)Ihe~jIa*B@M>7=q84#3{j^whfMT0(HDWh^lr`E?7XQ$aH2gX}v7+Z?SR
zv+P)8ZO`|$?RCjAySad0z>DWx{-(g`PW2}AVg+rvA=7>9=1mDHW8(y;C+1H5naVRk
zq4vSl{z@|CDB(@H<S_d*EHf~e$RlOD-B*tvd(-Pk<MHZ(n@e;)e{%yee3-Q3<EO|c
zA&XRD#ZqnS8$0T78F_iN0W_MdsoSWwwFh~`o1n{F!LzNt!&{#FBsDD~YlL%9^Il=T
zUwr)fVGr8=30Td}zKrgjwcO6+KgY?NT3*g%Wu8ggf{90}tG?#8wwPfgovU6?dpq&%
z+d*cQM9}G(nZ6~U3}l<Ur0Zj4qTr(wsd~A(qOjw0F_C3KUc_BOW!p*eGh+R-*5i`3
zfsS}aoC4dKRU4?|^Z3qT46N!vpP5cA#tbMrPs`c?@z?9%ADP$dkoF})iG}{f*m$`S
zFr8U0HmRZ93Yl)6^9A;o-#G`bE;UiwEvMaf>nURz{S2TicwF27;=m7G5&o2|{v4%C
zkki<56*tLYsUNe`YpNRfYah`4>wrhh%pFuliDXbif0Rf7)uQhsI(&CRdMnLeYNT0{
z;VtE*zg^I|qEf*C*|)`Lf~jI2(sLsCVeJygaq$JI?<rKRZEfP4^Q@iUC7~%bmSg2t
z5%N(V!AXGJamSG#wtU|kpX!F|p#E%A$lB&Qudq}1Z1dnRccK219SYD4+mjUHc75me
z4!`UC6J0m-Rr-u{gSJfNOtqXy{aDawmXZQ?N_7*qcNshUvv$;ZQ;C)TjZqsT1ET@9
z+~NJucLo>HUM|$VazW#Ui@p49JdpkOO$Bf5?ELZ1Sq!rWOhed$t$U~y_G^rkOgX7~
zE*M>6r(3ECmN!5@^!3KmZ2B?LV6mXdq$J@Hnb3SW<t4r8V~{%1Jad+d5E9XepxIdH
z&%+pWm?J99xI57cGoFB&)*o#}=&g`Klb=hZE8vN-TQ?1tfZi5YlFoVr;8TcFA_80)
zP@58F{q>}Y@E|meLZj17l1%?FPdbp4H8r>5F(u(M4b&kH7rdTy_e;`vc)6O}MAYs_
zwtY$VP+Ljaj-=&aPzK}FGtP0SBrZN)e#NiKeW(7Of}}Q|*To21d3Zj;IHoT?>V3`+
zPh)s3HTBf9xJhEU%O)0W0bh$x#d)vT%v*0@0#3)vd%GG|eYV_yT~~jjm>RaH&u!6r
z4eT+009WwUg@xE-?h2v07(S>X!B-aeuD2tWPmzZ*m|Exy3x`G#4LgnRaIpl~8Ah~G
ztUhJX`3Lm$^|PXC+1g}~wuC@HnAA1<sga+rs^1GQB89l+T>FrB>8LwkQ>*i8df9v9
zmSozEIYQ@$p%3Y3X;t~zwJsu@+!#4hS7F-k;1S8kaI5cbTMmGrd%A9abI)}kUz`bz
zDPfll%^#<MAL;YT4Dc;{OL<(PQt)n_HX?PT3~G`x@Dc#)p$-0Op;i?;*tZ2>l{LOt
zW%bYg`Yx^q*S>hGM>M~5{Ct}zU8TIit>mUUeEsjZF4JP_lRB8Nylu5z%JRpm#b$T>
z|28p?Fq7+Mp#6Sy5&9rMF>9zV#e&N^`-)IM`yMT{RO-A-G`QB!=Jt_*LBDw68-C{h
zhw<jcTb{q||IUzU?2;AZtWa5_z8EB++v2~1>AFj6a3rnXdk<VT_*uHEjHJSgd~v-`
F{{^2d5Lf^J
literal 0
HcmV?d00001
diff --git a/docs/source/_static/remove_background/simulated_s6_learning_curve.png b/docs/source/_static/remove_background/simulated_s6_learning_curve.png
new file mode 100644
index 0000000000000000000000000000000000000000..017868cac6df9e8d1cd8dbb80e71597224e1b9fe
GIT binary patch
literal 30054
zcmZs?bySpZ{4F|kmvom%N-9zVq9ZLREsY?fbl1=zp@axXmq>Sm)X?4CLymOk5a;E4
ze&^hK*SdcU46|P5sZZ{`pRjkT@<jM__#hC7NKpZz4g#UKf<S1GaIt|;7*23M1HZ(b
z-{?AP*jYHcnL2&~y)|{Vx3+V({_>H<^^>F17du-K{%8Cmd@P@xo$Z|@1O#mU?*sgH
zj!=PtG{`XUC3yA<ADloS0`i9!8Yn4+1_W{*SA@uFx~J?dx}|7NUR@pyi%8DHOt;ET
z*%Zh~Nhv5v8(L{1hy=ruKC3*lRC)G`=53kxbD@@*uQO%`L=;$CjDH09_)_<p22+-q
zp{{POGuQ*x-b<y^jjjjaizRWV4@1(heWxKZGeRZq>{xZBYR<^b5<m6-{&@Q~(B$DW
z?#cbnq(u*Jvga6y0^%OtXf|ZYmi_mbrQ0)H?f;(Ni^Oi_K|t1h$nrn$dp`RW%10XQ
zF=$AV9ea#p6Bt*JT3OGVp-FH`|B`b=TPc4t)27Vw5glfss2U`|AAh_j&NN&o8QmcJ
ziAzzsF?aTj;=GAXpgzyrkatf5LLuLK?!y%RhQ@yD(I&f!8y32Kv#v!p`;@Jdt-=t(
zL>7&kz7`(3z{^Gw+GCVK5nAZg@q`Wey~mS``%$zOoA4;|hAwIqVxEIc%7Uyjl5vwp
zgVW^0iPN0~sG<ki)Y;gqAnPh*v#e}RbPgnP-+OY|ezLQDWvkGJ`0St5TU@7z^j7-5
z!btaRgx)7sHy*gL9Z#0u=yPSyQP~7m6WK;paq2#oEcfN`#XCh~0@a{t_(LI&duUF=
zks~E9=?+DIU=U*NW2yV4Um2XsnSjsB>0X8*cT5nb!@R>MKFr@0z6m|a^)G0~zM*!<
z7Nf2k)<_pm8x}%$r@o}CWBiBP2zvZ%l_?yj`Do&58ct@yHIkL-w1Yrh>0mVY@%q~}
zE!E-(k|k`-+U~Y|RWFDb5!IbXxc1Xz33y|jVl6~@Q)ud3Rr}G?>{Z@-AV7$`4Tjb`
zGPio<!sAk5j6dkA=}uH8^s7upC^^r_`@~^~lg1jTXJq|2(sRduxUuFQ=MP4!5h}-H
zD0LhT)?3ea`V8)-G07QjR!qAkr#rg3$2P}cHA3CtQlO1?qwF#TytfKJwNw#JiRv0x
zw>)EfRgH->xt}v+XB+CHJ}+HSeQ69L!=sFD3GEThpy9$<>JaVn_Aa@q7s{y_9g(3k
zJ9oE(HH&VR-VS@0Za6J{MrfF+)K51H<nN>9$06UZiXmFM62{~vvV<&AH>GF}a1SF?
z38u7g*E+0*{>&~j)>7kP^g6>aXWYq0k1;UHh%9X$fsOCc7BHnS`MRVd>BGFAQEfU6
zWlEI9?KPQ(b5oL8inp{nm3tU1AziwpU!62JlSPMPm(WKI4lhrKN<5o-pT0~}`2jKo
z(m`Vm?IA^&BbZ9t>44?NT1x)UXe_$bm>=NUh;50!TKDEVLib*plb4_TLkWx4QZ{M6
z!btxinqFJ)q)os*W@oyM4e7;7C~%Kvb#wxCr;)xym&u_Wz(J@W2uhE;?j9_R_&*Da
z2ORBbYdrkT%N!vK9dP#>&?jkB36XAb<f0?Oiwy7Fli@`Kwx?LQ&>OZ2nmRqmoo)Eh
zZFt-{=(X2*s&BZ97lu1#9mS>iq3hEqV)~Q+@06cYl0~OaRrnGd;@dMG7Kcg&51xGM
zh?HWo3eT51P)10Q&4${9mtc9~G;T&*xS3!bMu$p8co&^f@3Oyk!$2;DF%lj(g-YQu
zV5_3HMF>T?M|u#{4aP*(QqEs8JzrGy4i`ebF-JROshM7Tp8ooOZ<%2gQAv7D>)sFS
z)`EPovhnT9B%APhWk?RX8%9evv2#zWsIZ+UBTAnI7o}50;7(FUd&y@~Esfok@I1{}
z*tdJihSwf^DDh(+4*b`;vvvEWlxs$>t2MDvc1U^D(j9sb`hm$oZ+Exu<g_2P|00}3
z`i#qN{lP<AMn&uHoB-Di>7X%$L-JpeMQ6@@@v@*Z+Z1*4j#V!JkBaI}$^$|0Mgek1
z<Q3u_J4KC}_97MRq0yL0anVr9uTh;hap40@oBRG7_8p1jt9np_@JQ<1B~P|y4B&vz
zlg=@DA7W^}tGrIQ#{UhI|Gg<#98E!({fLb<S65e0YM@if*EXLl(4y^1ASPnYr~7SF
zroIi#8rzE0ix7B1x8CuSlg?}I$#0=O`<o@N^?~^LKMzO*{r`kwV2py$o~LXTDv)mw
zbNQzm@nJo3WWDE&RAD`|MjK{t)-}k)AG6i7RVYFDkLu+rLK`{RuGkzQ5R31jchc%g
z&$(@!fvJ1UlSrXUY06)zn*b4tNdfZ?&+7@BL)tMW;wj=W3zIvk`)i0C%y8}b@U$&F
zAFk#nd2>k%Y@e)zpem3lD9r{WL2Sa6k;l%ZZ_x~1VvaEO2z7aFI`!eORS+lO7Li69
z!Ob&Dg`QCNP)5(f!SLO+;@bH=i-V;m3aafk6Cylf;wR$bi3te_n<d}W2bKOyOWhEN
zCupZ!c|yYPMT8Ljwpa~$jgEUNbXt&QnkHu3WyX@_4di3EkcTwVn8bi&BuhR@Vkb5_
z+CNm#1P)>#_U`#WZARGH*||GkOY#!Kbh6}>_6aD93Ny+-}i2>Z*Eji(vOInm$8Y
zr9Q)5F$2GO8^JDn!P=lzVQNs}g}KH%Jy};%T4gtj3bE&b`0?~+yXA4j$oO~aw!^)=
z+}#_ZyF+~I#gwqik85b*Lc4uj>k%6;u0nfet*9v_$YyA0KJu)#%c!5GvYUn+rvp^!
zhvm`;eSFA&m@j0}40U_<W5?-pBy#6HS)$OUxY+=g*+3pK<93>@I45UFB%`DyXaPjV
z%ElH@V%Xq#dAPD!5((@MEFjKOeysAFnEJcE(0zwcg~>aN`c7-{AH%e_k_dZTjt%~{
zfIz{nck263O8=xxuw$*qYnF+ELyW?roxhVn6>4=|p%mpZwF&>;aoQEDAh*lG`S!GZ
z2DYkOv}&@7SVJtWHiUMi3;iE3VrEl6^el8Xe5eqKS^xJA*e8nk2AZ+wzjoV%2@hk)
z412k_K5Y&0oqwV>6MQTeTp^H~5c9WVMCeJlT@Aw%iHY8fAt64s{x(*``8SP*j$kws
zKV>=RPn{)v+WSLYTcoROuX?z+IQV8{6+K+a=E1Bf7~&k7bwp}jO%bsjxT^b9-AlV{
z%}dUqcZ^wAs~5R1Mq3h$7YP@P)I3JZjYeoRa0keu-?II*2*0WTJmI&$ntwaI+9-De
zqg*=v+X;^(KY#pp1Ol-JZd$}6A<1oQY^*QL)T|t1(aopo0FJPaD+BuRajUHS#pk%#
zh8}@@ldT~YY?nw=WHfpWGrx)T0R-FmTTa`?la@aLmp!DvU!|^7#h<iC68_A3?cwxB
z;-F}b;AQQly;9%esp_}&f^x9?O`*+hFHhSRsi#}0dG`uB@CB@A_Gq8@FW#T!QBdt>
zEPS%}c$3ClvS@tfoi{G~<!kZ}7MvlL%)EilEMr?N7FOXPea;K$+MNOMMqHVX9lzDM
z6V0?FnHi$jW8Og`ND<>_;zSDuuN(9Pc*aLM2j`$j35v=@x0S9Gm?uG2euiPU|Lqhm
zf-x+vu`5L4b#B=o-X(blVN1?5%0xk1l(T)iKij6yr3a`}(_&?xDt$9IRuu}A>=tZ<
zF+5njg&Hw6(RUUOp`ULgg%yT*R^zf8@E)xsm4g)(znxCYxWx!}QGl?}{x*{T!|}pn
zVGDyJ;i?ErGnKE?<vV`5WLq}jvs<$-&=pSBahYEn$n0v~o*Hpy163s&(s7=}!Oj&3
z$EwYXjtPf9Sej=Y6Gpl52)F3^-SneIo=T_q`Zffn(tdST!{y|c3QAnl&`nMYY_;bq
zjr}?wHSw?dbc#bX9K(Y{Q@Z>hk;WhNC9x(@Kun=-C431IES6Y%MT#LIS1I-M`*KtL
zFJN=0BX<<-e*|*mC#F0vV-Go2N)T%1#EZ`C2$Bqo!ZjTt`4X#HM>F*_5-{H~Q)7OP
z4O@Qt$%wrW0-6Q!q(PkHtyBs{;GBj<8B9!&G&Rh9xdR)UKUWN|JjYpw80W&G!K-)#
z#03Su+tKKKJPU-|*6E?lJx}FGB|JQx`gs=2seX-#i>5T>5IyeLVv*`dvkT8Zi@(0&
zbi;=4i2{R%;^w#X(vMYJ;B0)#`m1BeJSUG}VN=~b2sUZpHU)JhJW{)=hJUp7qsLpo
zKLs@(#k!$yq80k>gbVQ>?9iLJhV@7x$jIu6CEGig{h=85TR@CXs@i0z%@MVn;L1{V
ztugMRIY<8bK|>pW`-h3&d?!<?hLHy`VT4Dx|EQh5f2{W@Kd1PU*!yA?s7dZ;(f$54
z1s`q>O8x*y-r;FdS1?YDzQKuuxfevRQgx>@pg2T4#kg;2_hVZCV~|8(C=rRpm>;oo
zR8)qYBSnKSexW+9w7#$3eX$bqx2YjuztPoO;*Iresh6Q9`{KTWN71dPd!%opIzs6V
z;x#n8oR$yFuX{>P-$vR_fCaG`B5RllufK;D33r8C{?~9UkSrh@mTB12j$6>CL5edJ
z-_GFeapEm(j?DHgxdSNbf%dM15ODWnjA4+5pE`mA!EdIr7{!>^LbTNAArgp9WVzdI
zeMZG^(DRkrNI_L~$iX<PX;EZS`hU{!GhAEPSf%xW^(%8+V{0im%Ie%<X=KPis*!it
zD_+S{n+$N4qPf(!j?Og6#Oq|M&YEZYp32|wIkGZN^Dby6Mbgs3qJzY~KG@k(3x?rp
zTm62O5ufTH;-xKK49;xY^ia9!W4W*B@AOta&bOlf?F=DjrX52-dwKZSrP=BVY8?mD
ztbwq>CTei<I9bDrAV9yO<`Je^(zCGM7*3~Q@WnGLu--UWW83{<&JsbEwNp7r`z>Fp
zHAnhKY_EHG_%IOt%p?vMD0*(L>M!A}vcwXQkczFBOd{R6q^r)P+s-2`X3sDr-jLI7
zuUFjjb<wUPB8($GX(~ziqRh#2|2(!IE#p6%G9B|<FF25dgiLV}xL^xaB_CX7{^hvD
z>|^)rcyVozt=a$H$fzXnCe{-Dp~>4r*~q)QY}pRU)QhAS$-D&IV3qw97r&E(SW6I)
zJ1kOvwY8_)pdkwN)u2F#6=m+pQnnUrl7Iki5pKG;IuoM>IUs&t-8{T1lWJ&&^8X%!
zc~i-qt(Tr`*PQs3fV#rG_tD)^l<!Ejl}=X#y$<E2&*`YVfzQ>3)ZOKpf%n-YpHXAN
z@84|K=X?FTvz3X$pYTBhRD34K>%E$qGcr@qp_H)x4K_Z$@Y-dc<enad{iX{e+q#wD
z&6_%<>dz0WK(?xhM(Z0y&fB=#MWmA66}}Bo2qiRiR+37@`qUwhJ<cr}E~BzHtZ4X=
zc3LtF{Nc^%0nOX{$ZBNQ6eWl^TQ3bq`Y!q{>V!^ugaq3yY^9k%&2pk#P3HC){R`Wo
zXksC!O_e1dl#P+c2BneL(HkXYWrv1Em2~OUdMD%~Dyps+HR;~$cq1iCOCFcO_?scg
zgTpC(yQAI8;p0{uX8Ws+coCbqXP)P~TeZUln)<(G%Ph|#*W(@M@24CsOEIZ(AF35{
zKhoU5ag-Y(p4{`j=H|zbEsxn=Tir)9&pka)P$!91?@NE8k^tX`Hzs}w31WnMo=xiL
z#4S5?GY+JIlPP(%?Uy`|lb#WkLlWP=^Z&*5+ZxX$r>5?ntG0TMt6ls`+mu$8twZja
zA<si$KiR{_wJ~2TPctHvl)3ovLybtxy;po+5?CC_l$G`{Y!Eq$L+I>}D&aQ8U|wvH
zMM<@Kf%vl;Xic_)qGEozp@G2wquW9&G~HWHP7a$mB=F|OE0RH69r;|sZQsC5Qds!s
z&CQyTFQl5Cf0!p_NN+ROuBd7Djq~?fU-;m%S7=%oSwxZxZbsceXAe$`g*}o+8np;n
zGD;ju>HOklKmMh9DDgl5IAE$ySo8EGtl~p6?|0c$vrEr$S?LeHeks2lA-g8kWB4x&
zKei`G)Un6lxUKKJ4XIl>T{<mfPn8ACFktSN>-pkE2f5ne?ZvW)@0}M=w3Cwk{P|Pw
zY;3XMQ?CFwvevF>Wm5!oJm4nzp^`IM%QsPdr~J()N{aaU%jLeNKTztU?k^bJ#r8sW
z<EE+y=@2<t+V~q3&b}BF6c)|8aJGaEYHl!lw7ANiXWxM5g8lxTUTf~oh0`Y7Tmtg7
z!c$Fjk({mWB9Iwv#+fB*VbyWp9SR2Vx_%OBwo`}i`va%ZDfUBu;L|XlsY)qk7vRhX
z@-ph$rny8e5<x*dq?-eDgWHR<6u;^p!gbeWYymfXNIYM|NxS7%(U|xP0qZ9=XXf#R
z>h}ec|0eRuC0%#(E)p6WrCz;$t*D~%r>H1=wi22?k^)z{y1uB*=bP~iT1<Qclu?-0
z{N63=pABqZ$S{v%Kz{_1Ozo~+*cEehY~@|2GCW$(XxrPI%!d7=An$s<X5II8<6A}?
z)0HD{h@&zWS25!Ia{5rqrX6zj!0i#^_WGX^AwJ-X-R(&?_+P}}xR@_(zFxSVnkzVA
zMp^O-xJI%{3kT`M?{6j^sM9WUM$7w13<zAj3^bYKt03slv!@55p1S%z+tbP)GR+PC
z?H<{UwVuoe#Y-Wt5tGI#eNiYUao9UyZwUx1+w&YsD@r}HvYB_O<MHha1g%=wvYCQw
zS!Vavcg+ITh6lRKSwiUz)wlWuR7k)ZKTQ>V=;-Vm1p<*=6qEOaQ0<~CeW6Zij&eLd
z@c9vH`Tj017H?Ae_HOWbimxMe4&lD8**!4imG5%S@#S#vq;5o|X#Jawfe(!Y89X3t
z_p6Yqf&;U*z$?pfR%t`u3^@>CnLR4H;0uIgGj0OcM}T(eD{t2+V*Y7}$2lrh3*|c#
z-CC&ZCQ|1_7;>Vl+awvX#K<=5@#hCjfMy{4h&$!%68)j9Vl^}i1qz&%bHX7M@|Dq5
zCr#QFe_YdS=VSR0M%3HFK>lA2gwExJ_gvg=4(Dsj<v<A!<Yyj(V6S1{oO^@kKU)ny
z_HbA5<`eBOO@}`}3q0KrykJu+qfv*I%_PA4E)vLz_P6I<8xV2n?%NE6BGNU}3i`MX
zO&QY=-4o-4dz}X307OxP=_)gwPDmQL$puLn;^4e#`>#O6l}ZEEE65{AM>}kKUgA5v
z%J(JGRP<db7Zz0a_M&-*Yvv5h=Rs$%aqg6U1^p+*a9$iVLNil~?{kxJ=i9IM?<4Kq
zxHeV)XYSLqY6e*4_a1(}Qa8WzW5W+c%EK#}!=e~vD|h*j5Ufh!r{C*BO^iNNB-bmu
zp7aTH36?L9iiG}Q;zk@p;}6@yM}fRP7?^Iox@a7=w&sF!>AGX;3Z+7`Um8*MI$L86
zV0Z=8f3$0!j@?!M;dk>=-}((z7#E^TMH$f%qv>e_Cj@R4d+*SYBzptK0@C@Gt+_y#
zbayrA`iT9t`);8P2ofTGH4=l=SKt2)6i8biI)Um|9@4|hjZ7X45x06Lj9#SDzZvN^
z75&}>dE~qOZ3QSic7zu74f$SI(_C?EEL@C6?e=qH@1m2j!#YIDG@czrv!+N+5byp=
zoJpycl5@fOLo11BPj1!1@3Lvh@zFm?08R^uPOqzlZP^9ie{=4M=Zz>Xv|-tFokbLN
zyVFDd@)iUN<$?3|B<S#o+vIC!F6_lj55JaOjc@NsLtBTiCeh5_<=rmTWwa65-7xgw
z7s_;Jd^f!rtZ9xzGwp-SR2KbRYW6>iyed0%fcd<`6${v{H)$#HE!-a-*`T0zKhXaa
zxP-9RKPWvYE`cIfCYR|=x}~@VE71hIeq}$Ex?a+`^Ix@}8xEOS;|*_W7NIbpJa1;H
zx$DwGodG2*Qy0yhxG&DKwRi__i&R`f+4cugpX+oEh5_D{Zk(Zf7pkn3-@-oUyFckr
z!Atj)_YI~DEeQ?iem$hHCBp+-6Ce(X?)GnFDhDG7ABFq7=^U_UWrQai=?`W$K9bAr
za8eq^RNYF1r}JP7Bhmyw?#IVuUUja<T1NW^PcArlCVG3)9QYIm*Qv=pk0x_o2VsgQ
zi2fnZJ$t*lMDoIMuP#jsP4K0cxxyr9Qf-Raq`%b*h_c=j&RO6`>;=tHcm$h2$tGp@
z6ER{uP+dn-r&|^D5XEFf4bYSt~>++01>lo9jr1I~-J;BnMLmA3w4vt;QAwPGxI
z??YdQ{S_XIYNbVcBhONUL%~pk_G|uAVD7?c5;FJcDW>rZ+;VZJ%(A2@O@y$-gSo<W
zZb^}_&M$S@Q}rCv?ckp7c+943)BaV6?2q!!G<ITpSTo&=x`nb8#(9{Pfh%ktIXN%J
zQchxSU3su!|I!P8*Vt^h8%@#M=J3VZ+IEJ^z2)Vqn%@jca9cw6h@lQFYX|vVEWd$d
zaL2R0nwT10@uL)gZ{N?M9;5eidThV5kb!D)at)nh{<RI=QfEB<mOCV0qm8o>*EH7b
zf4$F91{FG`!8uq2QF5GL+nS7gE%P>VA#qQyG?|Z2Lp{P|FZFk^n`%r0K&r9nWa1;=
z6YHpd<i<8;kSMVnqAf_CNNAu32Poxg2HhfDZKeLe?RbK+{X<Gpue{a8w8vnfk*-pG
zKPkiG_4pai0s!*4a&AWhx-x^ig|Py&_5en$4NjM^m9JKX<MyvoC%4F1RdM1YwpA6s
z(#Q83R?}^jJ~_8gQDQUm<T$*k&U&cp|CO2iZV&J<?Hc6<FD+e8>ZZuPa-Jzy^PS@5
z{z+|txoMj8aGregah>0)tbu^vQ_6W}YObs|=63ODrfKF}RP+gn^To%R$_`t*YxsFj
z08u2>4Th3qK0bD-Pq)uoeOy^n!D|UvQ=F4Yk28!2nt&wp3Iz|-6reAFzpotYSmtP^
zoRE55LDfLCnDsDg#&G(DE^-ho;EkUV^yRg1Y4?p=do2md;9!4y?}WJ)spVslcXfB0
zJL|pJt+`yfF|5^+UPvhF`wwpcJl{0z7jd9vyf#6<S6ET&IqO@1q00WlH^brwJre;|
zh}6972-14y&~<P6WIsCn=>1tUSlMLhcJf|yQLXwwV^$J;M}<1RTgYG5Zs|Vc-BWe~
zm?I`i{uj1p7~R2*<z6I;?=#IQVuhg&bWT+5Wwzx~ca<wz-3&tl#eJozR~36J;A$Oo
z;B4w`lzh`JJ@q53el=Jt?eMmU#Zgo3pPqt{NB`8+Hp>NSkK2A50$#`ER1ZLRA|-BY
z%B~iUw_8n7Nl_Q=$*Sy9e<Wy{EvyHxvN?$qO6X$NlA{)WxT~=&jse70QU}n*k&!*>
zXz`K$-5PZ3iS<!%a)H1`AX{YM+}6|BZ9e{RQ%m#`AOCB#h{=_Rs2~CU*WQb^Oy)w2
z>}CzqwfV`Yf${iRBgF%MJtu)G@DKW>fDh@@yYI1dQF4PzUHfL1*iMVxFLN`E#2anh
zH-2tA<=kR4*A;rA_!@Y$B2Hl8#VF4>m$PP)=7R&PfwQu!nC%OxJ?+cw{cXfuY~s1r
zGz>Fjl2B~8v0)QR<776t7d<_r<Ly$iP_j^KX*>9`>0(PX{z_M-7Q02M8n$z>Dfma3
z&(|Hsh~uQ}z7y<F%>)h4Y_k#jY24>FA9<Lp=zZmbBAYrG-1t?AIvT#3uSa4?rq0l}
z(DiqD*K3Klfb5K9ew8=b_?Xo98&YE4d#W>Age$KAxd%<K#_i*zL!6I`ryaf$Z{2MQ
zn!$B?=c!8gy0K0;V(ko7k67oIpRAsYY8592N@_ot<&x{thwAwBs4bt5XovgRhkb9(
z%YEPB-kfrv5}Hb+CaweG#=KDOV4F-GHQV|{=_=X^w>J2RDQfyT^BS{(yq0ZVAz|Nl
zD{LiDvP=Og+GDP5+d$WBx5n)Up$m@DQun1TDs9}s{<c;B_4LVPg5$D;vvmCtnz;=U
z<r5q%t?^nl5siKsd`%Kgezl;KFPdG|)~U}scD2R52*4^}h+ZRmAg{4*7bB9;<ZiQ_
zB3Apbzn-^#Ed}lLZOv6l+!voxzQw^?=39l~4uu75e&)qg-Y#dTjvOmF$^{YQg8J}R
zgdha|qeO^shp8V(8&62~d+t(J>HKXrev>hpy}DYf@&^s1KF7{y&nh!hp}DU+H3u!-
zp1HqDB?iJq{akKHEqI(o7fbWQ*W`HWgViGc-kY;kR!-OQuTdV@5~di5wNOoxAbkAb
zyvQFdEfOSW>*=!#!#B6Casm5hCI|SEb(Z5HqIKZ=uq2PHA;HUas+T6Sm4bT*lYv<Q
zuMq3J0QGi_k$>&^R^y2_iButP>8U$<sTq+a#tLL-WaNtTI?<A#Do(wx-&kg;!>a%M
zo!<T?0R3A1!KqserZ+a!0*eGpSDBoMGPBjzg|}a2n^VR*c-|8aNIm`PrnA>+v-`|E
z_u}B?1*Rl!k>jUo4bQGhk*pu%*H5%I`j?zq5fVGK`_1{!yGF(6#B-kXTEUYFa%9W!
z^RV<uVl*^|o}0naF2_Q$N-6`}!pmjtrlZTC`MEk&T)ma8mY<~ew(HBzUDSr4_Gel4
zwf)2_$&NCo3_2o$b61QAkQBJ7^Ml}8Ofc+m(<F7#7nZ5UDyR<k#)>OFBu_8<SwmU)
zi(dz52%zFS*}cf;yBeRx3<JRMQvP{EO3=NRirbQ|EsjLgK}^r;go^f`PvqOq_Ik^Y
zr5Ek}RvoDUEa&JuUQ-=neF!jgsRCkeDR4LRRK`=Oa!!+<IQ!;xpb6!>e%DWARJ$f0
zTjcm*y805Y-%|y3x5H1YiERuqYMv|`i`bp6_nQn_rQH~?{H@%h6t$J4e_5ZJvzG_|
zys~z7aeAF4uZ|&<9c$^xU%T?sR_Npf^X5Ui-R6s!Ysc9<Td&Fu!Dv?11#E2RaEqNx
zFJb&Xce^vAySpo@VCOIZXIB~ex9lHp5!uD+S`%bQ`!xle$L{Oni%}{t-<}uiZj>aR
z^ibC1(i4LOzr4hen-HS$>Z+p%2xKq-*T`!uVQ*wU97r+F$=0$E9n{RzzX@rBRkto9
z+PSj0!?NLgIh$hE{}oI3;CUc?%WZ=2=Z?~wku*JE#!fX118bE;I9JG<y&wt3SnWCw
z?^&iArh!s?U9_~OO$wtsQ|eNI<#y)f>cA<gy&jWoL?jULO|`agn<ODxgC6tZT_qo<
z+GF+sh}mHBx!RuAMaRVotL)9kR|5CGx|3WFv1HDZtHmM8gSaR}&Bqx0lxk+HTfL=P
zgxmUd0=xGt;;z)DK?tME>x){kI*;Fzw@DrAy)pI3G??n8o(OcPW6SxtjX>Zs)57Mv
z5H{ILmJ902!Pqu}!%LZWL5nxt`|k>9k-SZDn8yMrB-DShc_3C<A9n{G{fpO6l&;`a
zbAM~5o%}t-ZR2bMR)9Vi1Lkds*h!tq+ZxwA83DXk1pq$ekCkahg#NAslGp{b5hge6
z{PxVy-}s1UJAG6<GWiHQ91vPqs)w8?(VQH5Sr|*ihRDpnvUTu?a3pg-{K;BJ#5d;<
z6mJnO2MNV1CH`(JF`lE;@$9322vZo@gWYP)XQ@A`^S1Q$5&he6FCBX<Nfe;<vTh;R
zcW9Xy?x%W&*}9~>cW2AUu<%YKRaWO8pz2_KvFs&6K+SL7`;+5ljrl&qYP!UF@oyY5
z-8U^a&GXyosk7(B{@?>BCgH*ZvYIMdscpG^N-t{7!o?Npj{(5jx#`C@C&MRmR;6f%
zo5^<QKa!J`-;Lm@yqWD|7<jsp{50gD%Pu)u0hls)MQY;i02gMS67*2CNS}|Ic;{7j
zLK(IWn{p(vWQix1cu?~2wwnR6AP10<;D=IE@0ky+KFf*v*H<U)vi`c*gHie>;x<N#
zYa)lAq2Jj|iF-rs&L}*W>i+@Cj)2*Sz|O6gye|xci5dH-jlEh=x5hs-xIXnpop0nM
z+kTv@`r>u5=vKFAt6i-3MnM7BZOMZ(S=7d|c8e4sOzQ4hcpm`j&h~}}R)fT=aVu|)
zU#wmPNxU@>*^YIy+-S4lY+;;}Hjz(jS)@nzR`b1&8&_#icgtKC#V|_%oko_;Z5#h+
za`OVef`7h|JDbhQDD{r?=D*IzKw~0gB)vRDis34h5<eW`VmG^2!Pd}pWY~~;P1NX2
zJaC`Pn9;CjrQdw!q@U*7V{4ZVt#!JX+ncact4mS?pHW;a!35faiP~4}AJ8_3%|xj7
z*?w?PP#JWXX>5EPU_~u<XUZXROyEUreT#5Y8&6_2-^$dg@K<RmwGqm9w-}wD_<#O>
zHRH&V{z<YDBzeK__ZktFPP7KRkT)@sCk?%~mv@mO&q3N+fVRL#W5x00wuXJ-XvlM+
z%l-f>uyE5D)1=$Wea%t#xayXJe4Hp{g+(@LKO@%C2=>c;^U=l4`?Th&-X+iJ62k@O
z@wWiDq1DuUd#-=JH;)7F7ib9wc-*!iywmN;?=3eQ0-Bnd&pzBHDJ#~0|LI8mfS_~x
zCU@J|6>pT6);O%RP#vu|?imO-5-cwXh4st1aGoik^^aH7nY&ZliZ!`SKF^8T)iyKH
z11N#bORj>ifUv{m)wh0wx^Py2+qr8gytJJs15)a~6x3w0PfUxr2+Kk9;)V)p`8QCV
z$KJwQ@7DXB{ovk@(4=|qTsX;n0aM_}J%boZU~*%W(Ta(AzIgLo$X?9yT;^<dHU@{u
z-LId&p=fFicDCKAEYO4r4<@2d09>TW^KA9+PKo5UUh(X8lepCNhPJIlrGOcX^Ei20
zTSHf`wZgC_<V9gwxI+{9Y~p)(Y>dS5#nx*8vJ(Q@a!T9RK%@YGa*;-P22gY+-D=+T
zmRhKUZ5NHt)iPz#VoZBn?m42tM!pW!C~F{;1DiaK0s)PkGSPI6dD+kuZ~}Nq_K}CG
zsS@R8-H`=*`gTomsDqYtpWAC8>lr~MB_-WzOG-sWMM4I#9{_YL{Hc%Puk+R`0|N%t
z+)B$s{an>l-M`M<wzUnDI2|p%PERF8@Wb#3ES#0l0e(+ibF;u<AUoID>^ijM`7eTV
zCNFTg@uS{R(*@CN-z*(djlFp6*6oC*3g;KeB<p0mq&_pzcyIvXAY&#Q5O{K0V4K2&
zlelY$)zBQcTs)m-4DbPy5Z~Bk&z<6dr03u3>csxK?)62<+$jK#S!pr&5ilW492^nH
zbp<Bat`MmCYLkBd?+*X*@o{KW6tbYOWHrgE1Pu}2a>H?g9Oe`f0z_+xiHXG)uA%tx
z$(%?VC_u$_ohh2TX_J9u{#cZqnfPdd7*?GHUF_u_K-`tZHA_r5==Nt;{ZY8I3f)Mi
z<<U3&p{Yb#&33^2&xE@UhYFfHf^aHdG$j1VBfUFr?J771&hKiY!|3MYzS5J^op&Oj
z61_1onw-Jl9tdg-p~F2H8NmcjilBcnHS@x(i_l^q;r-5$6LP5dM6U^UVKH0r$*0&@
z<3o!;d~VBP>-h3MqD$l&;Yi<gs$_Ci!e@(Go-0_BbB<bzCx?+!lc^m1$-=?J77uC|
zVw?$ju4D5RyLE4Znw|u}Dp!A!8SY2{JabBCej?zd*I9j2OOjuE-a#h^NCbvxIu#D8
zN5`oG3MXRjzYpGP0SXW)R|bAihMW3hfS$sK7aeh%n0V+>;YG8CtA;_h5vqg-Ox=h3
z4ud)$CZBvEUBK4jr`#%D*w}3IeahCuD&=&>y>PkB17dp>kRV*g&enD@ANhlZ^3`;m
zy|i9!I?41tbaLMsba!?_Mz}}U4}j^Xiji#mY=64WhM6YHEjSvXo+F9OfQsZw0c`M6
z;YXlMn&c09qw<y*fWulnGfwUM+f?ryia$Bap!u@8jzt$@i&lajjE;&B^4<}56}9KI
zj}Y3_D!t9}yT>6!BL-bu#d7|~-aP`?JGCyF+IyMz<Ja9t{ReJIbU~(Cj-~PHC6hk!
zCdT3R*rIbJ(Hg+4O}A8hYccYA1XNFVhvyOiNB@A#T}<s3N?hC#;5e<6@t>o<t6Cbo
z`5Wcm>-+V#&}7%{8P+R6zCt8POg+NOT_Hze7hflsA{$vy?-Pgrhl&~#1LZI8=MOLX
zOZ+HW2wp$KG7a<PI}|^p5hN31bT@{mM+hlPg*0KWV=VeX5Ckj)347COV}v}++{Ft0
zo0)1aM%jgaE~{*}14IajaGRk$#Oh*AH28b43o8&e<XaZ-yV8u`>AhzM8Mz&3eq^UJ
zdeNkvdwUL%{sst&M6J7{(Sfor+f3zbtmC-_l{ybx&w$iA+ymz&#A3=NJTwSh0DT(6
z?vg?c=bN8=xDfWG9%Si)p#c!ipvCw%cO@9GOpj-Ik=U@{MiRUHG}e2tPLstYgQ>m2
zjs6H6JV#uhbE$VVP`f)#GgJn1^TaNvO~!NIUw}cr>C&;a+JDF&ts@TNi8=Vh{#{}q
zm}UIZ)>wIF)zVdfjkBuc2ZsRSfXl%@(hu=*$+$A=L8JjhXi$knH4f;Atks-4?gdm!
zOL9Vwld_Tdva1j}Juhey!3ip<nyYN7ka|$^Xr!^R2{lubNAJ8b&kik;a$FDfwLE|V
zNJ5LZ84*;b+3!ZZ;oHbX8usRyahj5nI%ekvzg+g?9}J9oE=JA?F+RYcJDr@PB0>Sc
z-mU~DtFHOJh|E$8;riS4mD%_!K5_fJi0Oth=k;i;`dbx<Em`y@HucABl9t>pDiG0c
z3N^@T1;|j|njZ37C|NYM*E#3+9vItC0XBJvc}-}_;Xzi?U~@s#{Nl4&-|5D>hokG8
zFFZfKr~0G5FARWx@JTZm`r%?Kv|>B0_pLf-HqY}9YGt7VXqM^sot`~3PHOBWBKUQL
zG)CR5qc2N_n;BSg{2isb*FC6d%+fH@Fle4zYJRsijo<5AU2{Pd)&w#H^xG6l)F~4(
zFF4cfj5*QHa3>iie%0!J6Ef?2XVsl;zgI_buf@ef>r+Y5wszz8Hd|BU3m1nCw_qoY
z6Pj^a`U@|Vuck%#2?PYxLC=3W2uo3x))R8(>n7Ei*$!sLkvwI{=cV&mhT{*AD*
zu}7uZ4lLt?oo<(|hCBi8MEITrTE<x!>KLC42sGGU?H**(ybXLen87{HYKv`Kn%kI{
zNf4@ldnjxlBb1X7*?#O#k?!rFiBo&^()#H}+^6EU;HLVUmA&eL=2P>BJ}dM`MnPC7
z=h*@L=4l8TKtIVrXr?*I{8Co8yWgVOIRA7RlFQc13RWKr7uqsP4oJlmLjR}g5`qpV
zpQ);rT|(mUa1`s;+U_nk#Q(bPDB*CBNbTtA`WYKbF>;Y_Uyt{wb%WN~v3;xQA<rr3
zEE~5fJ&$@q5CE37GW54Hb3s!@H#)gimSMV#3Qe&JugBNLF<5n4!X5UJGB$hjr&h=!
zSH~kMSJsviU{T8eD1ZFEGGXS>WQZ<7rS=l{fj6nd=Fq`vZ3eU98EGVIonxZ3EO_{Q
ziB(aI9nMGhXL0I-s{hH>)TS8V$z2IVMjJ?z^eMwAqM)Y9WI&|UyCw!C!Et~qS`*j0
znq@b(&IHf`0L<6ebc|tO?E4psSt9xN-dnSOY!$mJER<Zo^#7<oSQ!7UIQ2c+&1#qi
z&mUk!h4wJc$UBgi2X5rJmsBc4Ha#NiiBVkpQSeIt8-u%=7Y#HZGnU-E6u?>pB{<TI
zRsMCZ-29XnJ|XtP5XQfN>*?Mai|t+9wbZC@*In^)5W#6GaYD(X6|6n=CQUGgXTIHo
z;_1u`s{cG;&`?SWHHae0*&1!Hz0!F)dawUO&kU8Fm{onDQ?2lnAGb^TQNyrR!_0{m
zAV)DNy=Y}yMX!huN(6K(3}oUYmoE-=b}1}hToe9oz?W1zl+K_4ezK!o@}X5q7KQ>O
z)xS83H*!AQSsJTkWPtfSB2&Nkd$eR#(jA1lIms2M<!#DVk!tqv9*=+RN_TVfG<3^r
z7s;EAz7?b~=0$6U?=Q|)p$Gv?$M|Dzt{0%W%ZVJ1<mM=T9^U_M((zlLrS<)JQFF?~
zMxGE-sTvoUvS`;rRO)SJd3Wwag&o#AC<QcJJbaJQw))(h%27RD_s*m9FiV|@?RDw^
zLwZ{xhIbC<31&TeXRGCSi-nr~wPON|tgdet-%do@QjBz|3DeaZ$9pKCjs392YEQUV
zHZwxfn0wOvCibT`HRlM)<)egeZ^cn@xqSF9Ax~g)JEGA)b1Z8`vdX&7Z8WZTUcQ<9
z_>dtdhiTf?(m%Y?vmiqf?-@YX94Gp0jb!(&SY-{g&Y{2CpXeA-G{Z0rvPDB(;@XA|
zwn;wL4hK#PR3iC3A6=(o*h+%wON8gMiU6%uu_^}lm(U(SH)O&|ins9Xtp=*gDXe_r
zX7L#FoqsO%-{L=i@-4e)q;(sjDzRP7r=yc(%dp`E9OEo?lvfu0tbho}?4bns+Y$WS
zJ~+PWtEkv>{#W)QQ|0v%GGZ|5-D@VE+^uJ}Tzy6No+$0vHMr3WDW()t1WzrvmzzB5
zXdIx<LnSoEq>ND*B<TNJvvbvz@D?zPz$GB+vt_i=R4!Y&6{;y}O!_ROcNg0gXBj(E
z%#+NA5K)sHw0+8Tw%bhN){-F#&<_Kco|5Fs3QO2KBn{BYkyQ%NV4Um+CX|p;5K{8<
z;C{@9k-9PIYNcY_>@a6=5x5EjWC)6#dCtRb)=(oTzJD^EwxwRQ!Nz#rLqTuubyB5;
zZj?|G(J3FZH7l*SI0i-BOWtGrlj<HW`egjyltgzcKk&t&9RgL9luKvmX3<aeh~KjI
zyhjMmWGibK`%%RXd2KbfV4<>GUa_ij3Ea2F#FCq{9N}u$cmtO-V^qgkX1MXYgJ0uw
zW|s`034<$p4h7z_F4vGZLVPf)Az$?7OLomzc<_B|(eR#lvur7V@p4a=6wSsMp56Y7
zbqBbH8H=I_^@v<x)G}x2lY^2a<#gicKc2_Mf2nQ#u#|10tS^KP4}^ZyW01~^NFt<W
zj)m$BYjE7*9_MxYW;A%=H~a5rtf{B-o`=*5d!vIel4Yv$iA^x}tG)^zk<gt2YGz4o
zP3_jP)DK$oKrmV(lvtrGi?e)m9W#gD*yJcMnyJck*ZU46*C#5`bg!X%C^*nGaz6Lu
z6b*wWp2&zdP<ySy*iE3oV{{4XC&KVJ@0?*NHq5mJb9R;q+i<tKSo~p4ON_k&8vxl_
z@uDHLH&6myun1;w0^Afa%K*1NHE>A$E+EQ;2#9GIc_EJPr)WA)s<WX&xRGWwJ8yyv
z*k)MF_YExFFM(y~T?kd@icIJ%Ih1YQJk0%5xM_Q{fkbC;Gy`K7xv|}+`Ao007FP;j
zDJTDQ^u?QeLpDot)FX0a?FniDEpMAJH8Oo65s*W)1qRO03ZW}4mlC`42C6S-RG;*o
z;S{HxhDu!hdg0voP>5NbQ|o(dHfX9IEc@42%AM@}Vow2cC02r);Bjiaah6X3)!53W
zow$m*0%`8K>PyJKL>gJ!UHb%s?u=(ddcb`eEI1T^<PL$}+W<A@NW2+m?<uo)rBw{H
zSWi`eb;~y|yh<m2&o`IWN|H~jvODenF4qwM>L{IvT}ljhN*W(4WEb%a(3$i2om)_s
za~6iCr;SCpY43dkX#H0y#7~G6Y%k5T^|$~RNjRus+}vOuE2zP7{kB=+grjYfKEW~N
zvqo4c@bC?w2f3J7X;3IBr%p>9p&x+S4+S+VKFFAWs{bQ<Kyv*rCB_G5ONje2dGfUR
zBKatMm&_RD553C|v-|@9AK>f7WUHCf51Q4CjGOC!_{&qqL^5fuvSaW6QG9ti@U4pP
zq4}8v$LtXM0jtjkJ)oGv_(~29L^zI01?LYa?6(F^aw>+2@M?{|7GVEmT^WV|xvLZ)
zSONMEy$k;?xbGAAH|pdPN1)4@t2!-W13D=QTdz-)fZBoEsXozFof9N?7CsHqtrkxr
z9(@G>Wb?~LYAOjav}+o4Vs(U%)LEoT@GNs4CJcyWNc`d&Dre(NbvT<mt<_2xr(-nK
zPkaAZ!hq3PH~Cb}9^e&2p%miLvr}{ovDX0MmYnnPVeqemX)H8|KI&iSBS49{k4HAB
z#gUD!G|sB$2Q;S0X0=aV=@5RpsqFf~EMPFkxvCZkAi4PLJ)&MxseaFMSjAa~b!pLW
z3TSzc5uL9|ql48HZn32s#?P@iNvtp3fnzB^OD5TLxa&==zX)h8I{pK}6sJ3qRD9la
zx^Y&DTS)gN5JB}I>wD<FKVURHVJtnsk$OxgLF8XHlT)1cV5Z!)bt$JsA4_N!10#h@
zQ(ja#Iq4`)a(IgmG?DTs?Ruh#NKd`I6a%`zfpDRlGWNhD?-+Y*j|z>k`1TZnU62@*
zX5g|1D!x8C;;mXap#zloJ5Z@B;y5+tTK8jr9Jk0XH4^!9O8TGIF#Nu$H!MC1L}-ub
z7t7j<D1+5|S^-kz^liaXVhyQ5_Bl-FL;o$B4&d}Wska?&4{chtapcDL2%Ng3$w?65
zJ`h5`PdlMwew`Cz<?6sreEL?)&gIKqs>fNXjxk~v3D*lmSkyQe^Mi6vPm_F}eENNx
z=@2)4hjBNXYHtrVxOhW8r3p+?-y|GRYQJ)2;$9yWs69Ne(<towNuQLO_uLW-{IlD9
z`eMZ8@u@VqXZonHjK<t6dXSlQ*;2w#<9{la`73G)M`uJMly;yAv`{kF9~Fq*I%sWI
zWB9Wo<bvXvY58HL4q_yD?tc6UNq`Asd10*j_RJ@$Us~SAw4ijbU?KoI(7h0Ac%Hu-
z?Ho6|q#=W;`u2A-94yB0v7x<g88BgZU1qoLeRFph7w_@*0I!vo-B}KaO`1r~c`hPZ
zrbHFvXk5dQC?NTRr?wE|@7Go(9VfGUKjN^V3uu|6^6AekwPPmLfR)ZyoG6$^?0M9t
z$}X?xSrokW%LcfA?W(lXwqQ?4_p)`1c5LfKRQn8l6c~4@lJE}Zuy`;!<LQYry4Am@
zR5RZkWbE~+)Pqm9Mub9;MHa}IcuiKnNSblArEp5{TeAH>1Ew)n`TZZ&!FVBTXqnYX
zvJp(4BW(`zdAJiOGZXjz-P!FQG{!kw5TWQhe0;PEJYcn{s_(@Q2kRnLImD`ZyD-@h
zg#OE=frn-WgLab!x;Fwan>FcX8RA9^7e6wc*6Z5BZjp^$f1jAWGnsq#o81=D>83lY
zH1E=;Zj?^@uVKUviba^#=Q1~y6mdFl+QpefzVK^zkl(`nqAp2NlE^=jv-My~*7@mo
zsFXCWoDl&KmK3-102kVuNqnzh|Ksqi7&^i&T2x<1uDjR`A5e>uu`dvU^FBKqk7k_J
zV=U95nAcMF46pqZ;uCZq9`gO3h`po+70CfAA>33Zu%63ipTRDU%rLuOgG;`{nW^rn
z+^YfC<4n)8i1AJN{8}yg8;GcQ)id|WEL|EpHw`CvfKboWVa%>xL`|x-7fVqW#(?pz
zJJO#cDV$aWORs3+Wk|qXv&_Mlu~}-u*s9lf&UAC;XPSU}xE84mXWKbG<5r9wG>n56
za0O}!w2$-y3YnSXFN-3?yLz}ymUus{{%FLa1X6$3Xa-b$EptLM&uNRh53D-K+~N8s
zK9oj2jWGlx#7@!iW;8>L&$DEzxg<esTm}4f`16_UeFt$6_A?V-?~n{AC{Hg=2>HUy
zPU^Yj<!VWmlV1!@t3Y$ny`IbsIh8baZBcnTSw=f(=2+21O$76C91wXds(_4xUz-VQ
zguXLS<o9wRNjZnnW7?(Tc)(U-6DK*t;8h{XS$~0GsaOBI0$+P?2v8QLJ*;Q`7Q@Sj
zv*RKlpgnCC-|LS+Nc=E{o5vo3+9mO!5)<Kl168mAcH%tUa>Fj;sYPz9{0kjq%{iF9
z#&;v}jD;pghNMZYR-nv(AJ$TVfdXh=GQ7!YO8-U-B7wBC#oeU(QJ9Zkm~UQ5!@N!4
z>fi0%TiX6_Chhmu!`8zNlN*dU(~E2SuKr@y-rmdYbtY{#lW6wYeVV__Xbs%LfzU6f
ztS*8^v)F%v!AM2mJEPw885MOf&+)S<(8MhAj`iW*6`kqq5Fc<o=E?ca$buGLqcpUw
zG=9dY^G?G`=7~+_TN%GBn5T$F=KN=R8Xasm4v!yyK{ej*89PLu<yl`59B!l+<eop0
zcX_$b6S!wQf6?VfAo-+ceh3TD-=J>0+@gG1c$cj&jl%y9`Cywx5_8EkG2)ZVH!doj
z5e*T9`O!Dmf?HP{CXo|y1;EkY*kKP`FJSk@dEp?>l2<DQXA;bq|7S5WyFD3B)-+8`
z%!(e!3l$hmige&6(3X@Mk9y{Kya()!ZJi?S=wkRq3=M5tf)O53`iS1d3;RLIe*syN
zKfPd{i<F0Fw=D1qB*6AU7d1})uGv<A3%j%~GK3E#p5D3*<~Lox^-o>a#u-IQVHz(P
zx<w)HCToU>h(0<sSo@ULeiN4Tvpswxid(!V-RW9AUgEJ#e~_~L?IV2;dXWcKurU!>
z6?}cW^K*t8+t4J?!j9LtYPzIRpxT>X1_?rqvD78fnCSxdjYM!B&wJ3;dSRb1`{C8L
zmGSP%R)d#9_tDLK@PrJLYuX;ntr(ygsBf&ufD0Y&9-9eC<Bl~JTm389UqoM?GS$_b
z(n`K=ooLaL*}VTpCHZpaSf_9}YM*f02e*NGHg0Bfkz`95tsGpFsw#aAMy23=zZ*W^
zq>CTyKC3wbyI)YBqcYx!twkL!$`qV9aNe~xH)4#bzTP2T*LdQo9a_lw6me_(dXtz&
zu=tv;3$Cy8+DlFBVDv9WA71U$!cqf1VVY(T5E$5Ez!;dyCX4ifCzV||x#f$yCep?Z
z+~;!&mBn5_SuY}<p!1-qZvWdoci&+*Tqc8HUjB&e+L6XSK5qwfRe$av!@d>7w|l;p
z55+<>5Lg~S><}N}AiUe3NevQr8dssr8X~EzP(Z|0n21aop;u}edugz7c#|mjkF5@Y
z*HAkA;YV|k=M={B1ZJL(O3T*kWZ}{Klq7Hq?AKfdgA^FPpTtcW*P3;Zb^kY@uj>z1
zQ0E-ens*^f4_^G>*D^85lUMH)*liz|ofGxEt=*#SoV7n|I422+4NFT~afGGJdwT^=
z9Fx3){48dMfAFoj8fl<@K5ywmhxsA~xIhD<nGnYwTdIx{f=e^y->mQq`D6LkhF!+V
zmK82YtB*83Tt<(GH8f}Lb}M8EUDW=gQU!*`HT^M^Z>@l`3X@?k8lBAGwUpM|5UN~R
z4ZflEOh+SNwsRN!<u;Pk%&=)-GVBoCTIK7m&evC1>hL@2j9x;fAn;bI3yYHf^4h3J
zvDRg-79ZQ#;Qopi3FmCV=vW$7HV%j#-z^*qBUGQ4CNXNDC_(k9qBTJCiTq$Gp4R!4
zK0wnTGV8^{?B}2Qn>8<z><1$@94v1wt(GmYA265>-_uXdoO-autfE)*-{VO~dyHDA
zOwN4wCZICD<maUoEs{D-gjUA2QL|S1PiAQ($u-<lZWkqP`jV!l(pq&JsI&zht?&e{
z4$Q%PFnVE6u;fXdX+Fd9cl!0uwPXF?2A!YgL9Nk-O8I3*Wi1Wm+tzWcEjtRt4<{E7
zcREE`Y~}uEhySIKsygeIF%nWQ&fZ}G18g(%%a4pI?t61GgCjzLSf&+K^osru2c$Za
zrLJ*iXlevS_)nyFU-=f>egYi2JII?Uh*9D>KNKzb4;}X?QQ%4!W&?YB>N|W;eoUm^
zDzXD%huOkXjhy^UY`W8Xpx+eo_QRN+ni%<}2hao=UX${%k-^1SXvNU^M7t=k3O?E_
z27A{N?lYLZ==;u^8PDjDG*P<g&udh5pV)oVXG<bB-5jgF_pPRN8&cJqhS0DpLyL3s
zGy#NwqcJWxPOp6}Pt0q5h?y5~6_@9X`p8GURmLxFrnE>7E~I~@%|A(q3<pY4b0>#i
zw{Dhbbd;FidsGm3EJIdqC7}zp;T}SuGs|(H^#JZMcEp)Uf&Zqe1zRl@-Xe<AXS-j-
z$=WM21wk0W`M`yQLCZeT|2IA1-6w#tkHM3hcSTk=PjiwwEffu3_V4%m%)(eVMpN0P
zkwz$mp@OF57w;!St0A7GaHa41p4POLxb3yxOcCkXTRZLeF?6viGS6_Zr<cm>)`cI+
zdpfoA=W?#J4y);z$KLvLRoUHoxi5W#0rh0OFE+zDNgS2nq*b77S!fE$jf{!tA$}S2
zKXd~kDGoNA2%&WBs$Vt(4{SpC?p7jsT}_-GkKvW;U`4L{Y#M?f9p9QGPW|aCk*e1J
zVJ`Ig;?Xt2g>WDpXgB`{4~9VeNc=vUgbO_}gm_|nHJWxO>cRPo76zw&`3osh00QoE
zunP%S)fc|k@FozRfG&B9X8b+Hl7#&%4GyTEu(Rgr_@V6b!$X5dplg%yh2XLSxuuHQ
zA+!H{xX`fs-aOuo7K9kx4UG<yIgs+p2QkLBk&^OrfjLg>+1?6yBiLzcbL&xu%uI_>
zoci9In3~qE1h<TWi~5>}>%*-j4wPulIF5e#@Gmc4n5q0irle|dhMuB7`Co;dbyQSe
zzyD!qlo%1|8iS5ex?@nJr8^uH=?3W%7#b7=l@KXulvZL81*8$Fp$4V9Tlnqqd!OgI
zcm38~cP;*a#X4u7ea?<gy<eLkFdp78#<~vl!%wK0g~pX}Y8){i5qnkBC1Z)QWQAmS
zr4Lc5FD>=aRvq=|?ii$KZ+@%G6Fqism;bHU$Ehdk({3~lJ$Lr4(Is9fO|#sF@p=Zc
z3r)kP;!VLMQpUc@K;WJx{Puu`vJ{kd)b6f^_J*$bdk^UwaSu<Lm@)CeWDqw<xCo~4
zr-j=cx2|TBtwC6+liOg=d0<qrDvt<7FL#Qx=lPktizI=JhQ=NglUe%Z_7xKoE!Qr@
zsqNLVapq46#eeGe<_`|BL%CL`h8Aj<{UM9U<l4$lhlfx53_B`r>(7^dHtk_!izf?P
z^*69$<dS;~bB{&wfj}8_{DIoBI@B#zC3JL^A{%)(y3@U1L`QkfK6}Ro>Z^s=QJzDg
zdJd#VF4Zs6ZQRIxn*v<nwS>+a;y&I}yK}UU#xM3jBw;5i;?=MRId49OpryJMcKaSe
zU8dr0#TR7UI5IAp;9(^<&85be7a^_k+yeyZk-ZenL%OUKb0=oEJ9hWqgA`slU2-by
z<4$Mt_4u-PJIVS+cNGzrR>?MGi5~|*I5kLo8ve*EmfU{yaNWVXOE7Ba0H!*yy?@J~
zyxF^Pv4mzQFfQ=j6a-@R#7y$5$r^jKtpoR|L}gz#?Yr3k;gHPca1owpmH$7}CZZ;b
z83{ke<(m4~Ft4HmWq1#+J{sr#vC^fs;&E%Txi_fAmZ4FYw%m{IM1#^sA54Z0ld~b>
z8Jt}B7~U@o01$)=sLor^`*+#{qMc05$6DR<(caKwG1r@0NrjU>8b{*&qh}`5za_8|
zArErzzCxDehJ`oW1pddbAIVd%Kl9uxF-w}QHhsIh_4-je`n92BVwsyiWE4yD=G$BD
z#|qkIcadS(ar6Z22R@VZMq2`T^OYkS*9bY@pW3n}*VQ{j7ZT}_m$zzAjX<={XV(`{
z4FuYFlK}>n%#<H4S}y{5E;p0*1vk!K`sx+s!4^8F5L#NGQ~YRKjma^N8;=~Kckgd3
znW=izgs6YiOTHAfchE)4il66;v6;8stQe1)K!b{xjH7P}wxUs=_>P0UTVn=WO0)7?
z(^oWe__HxY@GwQR#!y*v0VbXpn7KWLF!wuK;#(nx4`f|S;0rJ0^~+hOWC#?yLfI;H
zSJ!q@C<sQLLa?eOp$?5*t3CJbw9gS0{~&Q!!vk8L$3FGuuI-uk`YF~{W97yly|{Pr
zV*x-BF!2S>;mTj5|4dOkirwtdP5ei`x`u63!5TPX+QhQd8cEIw-p~XHidtQWcryye
z!9OExShr9pv#3DE!+Z9d%LYNXm>I}CtEuYPw2_Q1KDG6DueKsa&FmaphhG{<V9-oj
z)f~cJm`S$${6ap!Wo$S6V3McB;ioiMR?-l2{A;M08^xUVf~6Nd_ORO|=@0rjs2GqN
zc3<<GJNhge2&C-J<SLt*+NAf3cxOs)kCx^HR5OC**DP^ex!_(RjcL)kwgp*^CW%|B
zudzbk%5m_jj9V^?uy(p#nta^a+o4Zf!wE;Z6tcvJR*U%MdTYlf<4b!atb0FmNsXv7
zT;j&(iX56z83bocoy508(mrt8r<Q-|qiLjiBXxE~ltPmTogAifa`Rs1QG|C-U3dfh
zV6~hEWW-i{){N3ra|WrCNr^}}_(q%!1so<l*2whDi-7v?o46M8ls3dRWs_Pf%L(bP
z`R94!g3K_bt2QDig2KK7viqUwI=5`$w+AefqNcK#(8(*39g!TV_>S$As1>V=s!`ka
zj)T=Sp9$Kz4T-0)VQtkK#h@qQOsBGk1c#%#K~$pJ<(rJFACR|Kc;w<V%yl)`tKO!3
zXrj?PQ*@Il>a&Ma&#Zo}Sl%;fbsiu(*dGz8^oO2#CYWC_F2b|xdB~81whZAwE3R4Y
z#eQhQCxW?`!xy3=%W2R7!b-i+AFCm10SOEW``F~i-S?zYOcgu2^hiSvUGQwVdQM8+
z=)~>(VvQV3xraT^hKWlKih%t7^ZSfEE;;o?(O;@6DxuZN-pwZ*HD@eCV;QrDF3*mx
zrhASKF1nn4(F{L0ndg7Z!#^x4Qown{z2>)X3AytjxG5OljQs4%5tFxtO3M?3EarGj
z-kXg&(fsutg3mEf^y!#cm6Wts_EORA)qbzn{A}y;7VyG7RK+y>SXSy+4Tq^*)O>#{
zt&4n}F2qnFUr+649v~c<evEPar0`BI%nBsO*Gqm-X(>zxeLgWcS9t0l$uhZcxkM^N
zr2Ty3PK~sghUwBP;-A&&`rq^>hSS-P6s=NhqkDppBueTc?N?(ur^ZBP{}h2Ccptqx
zmJ%twhOXz%%6cr_?VT0Xd+(Ij6|a8K!B6-}WPmX8GLX$J)2|NGOs?*=XPP~Le3PgQ
zj%Pv8y}Ftt$&+uA2d1As({p31aBN;@5^;j#Gr@BkTQ~7zTW_g-CSVTBudO&J36yP&
zxuuTpl;<W`9P1Gl$Qhpc1G_tX!(^cNVez`XIyJLXoCKa#93U|frG~=+d&Z&!?^yZz
z&fo^##uF?fZ;0fDr(;4uHQ4Hak1T0epakaSLD|REtkD|nyZJAI9$IQqMzH=~ah6We
zae4qY{9>n~%VApeoMY>?O&?olhffbGK7`lC+m0h`g~Z!R2*#%5(6PE!Xy1WuUYRfY
z)+CXmcNe$yHQo;r{BB%wf7Fp<?^iu0)PE&OZ2T4LNF(MbqY$i=u5y>1!7zS|AuhSw
z&BsT?4qqYk4`ZRMcYDx#HL`1XlqY=J+7M%iQCyN}T{sXpPosF*LKWI=S#wE7#)EP1
z4Mca<C{G*LQXX8vwH_5PU&AtA#j@ZZCW*;u>|z(6tTJdQ<giokdr<<hXz4j1rV<z1
zO4V1bxdGbz?R{g0xb=Q=P#4)jR#sKH#Y9)dXF6Z^U~?=_#Aq<FE?f0bKqPC8Vy%uw
zIb5M~Guys7H_1!S{kk5|=p;}`om4DFx@U!DXq(`j8nj{6*wvY+@-a}S@L@O5o$KN%
zC^ES$k`{|=f7oo+!(z{kJ%R(9YDISZ!*f=j0F|B8ZrgO~fdBJ@qvFTI>9oID-&dF+
z_BiXIlnXDZ=cuJLG_={$xBMrl_v8xQ;?c*6tyP6eGga0D5%?7qV21%Xn@m*Lb8=9w
zUb0vyXK~Y4JHyLy@#x@cA^$U(SQ(i%$=z`KAkMwQ9NY7*oAOD)#Y5gKJ-i$ek}2#_
zV`uMyH>ai~T8ocbB{OVvzK7w-zY5+<ByDN~Db+z+V#nMmqilg_v&=W#eH1jej!Cs;
zyj~S|C~QXD_S1Ma!?$1BJfw2zaofwp*+M}D&knSn4)-vl9aHIw9uRU$=S7^w=iQNe
zf31CRi-STkC;!<*Y`CxYcN=v1l{9HmMHr)x`-!{Jw)9*<+hpuot&XZOXkU^xRX^jY
z*49fP6S~9Uvl(sp{**rbh=d|UKNg$FZX}`@pM~9bY3ti$40gwSc}<f9BvrvgAmUjg
z@vQ?xlhMq4P1d5S>KpWO&Y-seNj@ftc!&d$*PO~=(s}3QOMXzU74{=$B@63vPClg-
zF>Ncn0r?b`ptJrG=hXn^FDmxQ9CPb(w{_n!w!HEVy?&qdOmJ}^_i(%LEzbI!Vbu9*
zVodd;VGl4|v6h0NJMY>p<#*qw*!Ndc7<81?Pbx3A7brQNfAy}F?;U4~E0U0_)Nw)p
zJEJ2GE;pjj1zSA+AnT+vZ10x3Or7)Kz>Ck5h$SDxiyJz;JRZ=&F;CP65<Oe3;4-h<
zai#=Pg|4$NyjyR9zO>H~3T{tRLuaBPGkca^n1XQqOOKvSEDv6jA-mboYl<uM$#JTb
zbjD$27?}vEKH)tMq%53qwuQS#hJ9w%{6zT6pQt)ZRV1pNu(i2go)Kdp%M*^BOV=K2
zqF<PbnikF1yAHl#g6V@-u_(}yxs3?t+}WQ@vs;gWkn`c^)(9<yv|0SJLY>|m+`QP6
zqHyA6BiT^}!g~%ranklFLZAHAA04Vi?LS0~Sm==Izip$qc%N}<MQ*Z&q6q8lB;r0%
zdItsXIYL!ypyx(>N+MWS3-0fOIm1%Icz|6;Yn4yy9GBy!PziTE|2^Q9H$@p;8M+xP
zSUu6VwK5J%vA@npLl}kX%x<b}3#escY?$1bPfMYp*=W$PoTdomRLZ{drViv-Mic=K
z$_&(_;>DZtT^A<5+gse=k~W&LH(jO7HGEh27O4ff`ZmU5p{nU;g@}DP(9f*(P(O;O
zT+m0Uj<5qq!U(rg=;%3zWAFA-esB@_S~#YMQf<10VYSN4wWxVOK@1F+`x@GbqGfNZ
z!`=&VK0|yl|J=H6zrh0{3}oPoZWphSQ98Wm9HW1_UN~@JRzEj0eCn0pH{5Qg=}H6m
zHSCf-s|He@$($gZ5~|NYc<-6;R>?Gdy`#XhZ>;789zRP&OW{>Y?`ib3brkfTfF$gt
z=A#!*t~iPn9*=;Fjb>$=wJw`v_9P+jisuK|#nGy~W+@GmwVX$tkEI6W)3QO}n$}8R
z!Bt73Rf=vbA|u7#R+yBs5-c8gp!81YIOu5q-9lLsxS@AWeb4#1T_xvHtX$g}$(y%#
zRcrWO(=b~R<y-s(3=Si9q%(ym`Y7G@KWo1pB@_WG(O?!9ulrC~k!%j_AQwKR_PS0H
z#8)s8iKFYBV6hQ!o84!lA7YTa{Ny!=XodlQ&<c3+%*RP}RFm_)X4gT-oy<+{a$Ksj
z6+#A(bM*NI4ZECP8M&xf<|)v-^sax2p!7D&mvJ|tvy-UD<?-M-53B%32k#celJ-St
z$%0~^GuS6Jv;oiI+BzBNn9js56Kz4T3l}KuX9NX<>=2+E9}dPI%_z1MB{etN{k^Y(
zYlway<=}A%)9g2o>KqVmrFEU4q8pHreRn?m3#8e2<wDuA!W!t7H#u?4MP{qxeA>7a
zr)SO=SAV40DS_-E4^{_UAlQ&)VTA-b-nZ`v<1U#!^7gd>I{`%)L1|@mn47{I8PmH=
zjQM)8w_tUs&f7QwB1=t*<t-SK@{NgIlN&JGBW8~>)EO7ReSI`VY|s^1+*Cn4SDus6
zI)*}D@xx9|JGIQ5+xo6Y?3D?U3iO62ak+{sgt2oG>f91FznN-}FZ@6R4S+_D-j@}^
zWIy6piVCY1y>yb>A{C-e*H`<C3R?)_bp6A<^$O)F(msdk6vE!LlH7s|;T>HYBaz;0
zXm^m89iUJlY{vT-GLC<$x84b4Q@}#-yLERYP}_*l?1F#Im>7g%XVj)Tg#T$2jG-Uo
zbf#e>t#OHtRDT=2)<eHb8G`C5&6S>V(8Z&p{u>^UMO!0#?qwlsQJoO0^u<q2A>ZXK
zk?I0Tys)ZJ{{i+fL69dW7~pSCI1W853p=u)PR5hbtP#w|aNyoz-lRtQkTroN1xRc;
zg`oEEMhFJSsz2d4W76L&A*af7oaiXD71p!ZKQ#)!p*>x%Lam%w2XUjFK36R*^KGDn
zvOJ4I<pj~i8?xx8Sq#;i+J~w&LX3nYMnEDy5IN#x2<8$^vIIKLkJ1-c|A#Dp6gN)r
zHCp9{D+Be4z5y3Zkl=#oODMZlEqsBu|8G^`O%G6p7V_8o_%EPYyI#fA#G6O{VTmT(
z0)Y*IFyE1>0c)!}Kh=v~g1ihKFSifWWX4_w%mU$mYVwwh1fTqOCzl&m6>|E4n~FCi
z3(W-Ec?(|<95cfjWvhMjZC)7NVeQB7$B$za>0)+ZnaBlD08bQUnS_kViMy24gZZ%(
z2Dafq{Co^qd>=#|aL07`L*WHz`WFxSSJT6-1U27gFP`d#xWvJ`<KWK79)UIa`b#qN
zf9X%_UgxCjy)UkpY7R7K{L0nk_;F29xkxjrAcgLqFYpMb7cHixh8Cip;kSD0evzI?
zoga-+v4jGoSv6|%5)N_1yhGqxfXetK(JbXEQb~@cx!O+vSfKR%@(ps3e$ioD$8Pu7
zF5^0}sOkH%L%I3tgi(b`x44XKxpl}jqDboj(Bs2wYp*#4n75_=A*YSCu`-sti^uC!
z(xf-6b-6|e7MYNu`1HtW>SSm18`uvrE<9NR=iN^W`sh1jc0;Nkbc<!KBxwyLr0CZl
zcSixFP8ndT_z&}~fnY}<F`yR+X6CDKgGCan+HL+!IZ8C(EwCIwantDsi6*bqnQ)4f
z)z9s@=-mAjw)Vd3rDsbK%2LNqrh=T`{Ky6bk_1_K@UX?iftqY!m)bZPbp}cE{4=Z8
zmpH$-o8R_O>XQ(X1my4!D(p^TxQ7+(z`7wR1&Dw}^1OUyR{L3wxAxSE|Aahe=ISM}
zHCmrntx@6^kaT!5<<}1nM9c+3zwbA`j8UmHJm!KLZ#3L3MwzX91idITt&nL2Z9$yC
zEqEp)w*#*@$i1E5st>}yL1q8Lk9+ue&DRX9-)I?KG9~cj8k6sQ2p70(m9~LEJPhp*
z?Ffk;OO2VPOV&g7T<L%G+HuNnoea0za`gwd&_GeKr1)=b6>aW}suYdnep9vualSbC
z=EHjwqYI)|iSUIWq<T5L{ZX8;U7VYLS0VBqpWMSZZjtr?ipE6SZyXg~tD9RHEi_DL
z)AXqR+<zW*9;ffC(-M1OtW6FS_xx4R1|F<sA&{Pnx~-?!kAni~Nm5(LGHj;^WfKkH
zunk=$0rZva^uL=v4%qjfckM`ymL1wEcbhw#IsbFdv%b^O$&cp7Kwpd^FPq1Av19dQ
zrA+>=VJMMr5_&7P4`GvlO82t}c9_e54nhTj1mDqYA#}<%VTFz5Udu>Eu8vjuO+QOP
z7_YL;eDd{sjK<0()nB|d9kg%&-C$t5mz<mf`BZe`<-3RkSs@sH&<X5v`pB~y-QszN
zE`hE$p_KcLtj=>@JE1qTs-wxZKDE)Il`ZCL%G9z)<$WI+Q?#cq#=&#JkEG|)Kqrq0
zu*L8c!)rK=fYdB_y;q?81Iy<3OR_N&!-Rm_vTyfssV8xDq=HDD|CINzj3DwfC4OtI
zg3sl-*J_DhSSE%B(0;pG4t&F~OG0lT$<)U8sLy^#Y_)1===0~B#9td#9v}F$Kej*L
z5z~1yt7f&IumHmLv}=Y0raVUpC4(kceORXHR~Rb)*AU7n{nFdaJmKSM!X_o=nyKJ=
zVK?AmrmL?}WRQDPGudGC(|&t;KFs~5_uZ(`5OZ`~*K&UxTmt6Ky5)kcK;vWYRxA&`
zjUFiz<OjLVoPQpWipR{k;KCNpbU6Q<ndL5^jwyZqw(~8KWR*48{bXIk^nSicG^o=T
zt9EW$Os@;ATjjRI|Jfk#r(0%fsblRQ)&`Kh;0>_>QiS!NX*0zzW*gen?Ya3jZl7_=
zw6o_sxmsL(p-Eh%_HHd3srKF{EwB)9NsWPX-5a3o&@nO^o7A`44hs*zuc>+KR(r)A
zJ|6uNB>*0f_Fw0YTg_B4a;2Vesk%egFj;ds?f68n5-B@*a{Z^M$1|PvbmCV0Jek(w
zny$`Hc?7+X6&_{O`Ps-fdL=hE`02rTO`tHfL8!OTs6fmh_pMT%Kh|d1h;r76a3;-7
zk3F(*fm#nj2fWgs5BBOz@$9rJ?aSQV-IvBb<PxJPsG%?zYzb4t8%4$Dys`3Lw3EJV
zk%<xqul!cTz`VCd=o3@KD#*$_hjmH2obRuQ=zRgmHmzO*c29H-jEfAE{KE*|_q-^i
zeaG8gWY9kE4rMm$-YF?In4O(XNgI8hSoD(&<NZFwvZeNuy%fL^0-jVoiugRZKrpI-
zF0ABon4)$2y@95_oZG53YK)mg+6}f@*d+F*-s1ftuX%2L-X=pHZ65z1-bb`A2RV4E
zXR{lsBeWj?ED6DEbp(ECzE0VTySDgxv$YDYmAb(SJozb--l+t$K>E7CHsoZ|26l}T
z%?~=a!tM~d%4=6n8K@4wRdB^iZCFz=FI>>)8GgraR-aT~U%yZuX@4*4B8U^jJ}0nl
zQ8Ef!QDWKcH?Mz!49IXa&WbJljZ53k5P$+GJ<@#p9En!YE|8XirSC`Hef2EqSN<Cf
zi}e+7w_gbGj?X~p6XVu^?Hl}~?<nVeRv-yXsYhXpU$M&`uRr@_7`aQEOsXUpUeC&@
zXPA%S0sGl|cZi2=Qh^fJuvu4kw^vfO{f9WX*TP|6fUm@*TGpK4j`ziZN}t+A&wpi4
zQQmCWDsj=Qeyv(_AEV45KcD4^vI0vd54&^o@namEE}TkMMNjW40r6$4g`QVtUc>r5
z84@p^JtK5(n7<Yi6O%3*D48kg1-o@C5-d17`=geI=C-Y2MB<slfe@gBodZg$S;LkS
zS~oX0&vi^H0BhAs$iB0`4qO+ry*-$PQGHeWxsq?bS5}o(i8<kQfp^I=EzN!Pm-?DF
zkNWhBSq6X!j&n3dvVhJ`W@|gq5tBW7V^ahu5)RFu>f|N_o-7xR*SjI(JyEaj*zO!m
zxa@ZER&4!>71S{?i2<s>lfC7hOsTZWN};TPo%;e!dpbJ0y2}GON@b%#i<v%8SIg?2
z9*o-_W5cQH&vv?Bi8?XL_#f;|*y^u$gibyL?1D#N&f>N<Ov}2qZ05r@d*)SlrrYGj
z;lHcnew#IX=qT#&$<kSW=KX+JyP-RNh(j{)xsHY-kR#%laD+reP$*OZyhcOIK#{@0
zK~Q(^(ZS+FedWrR7+T(y#&%=d%x}tc=Fi$@t-!+qY2yd)bxgr2b+HCFpO=1Mi&#m$
zeQ9T+MX;IXMX=n~uXpCs2g3%Mflq^m&x<&{qz}Y|i9v}?*!2!eyL^l|a1`E>rw!_n
z4&VAe(Pl+b@=KEq9`Wk{LbtQKOYp#am~t-7K(OFBNTxU<gn-8Qc=Ne1VCE@WxJi^8
z&Q<zpQ8vARl_Ugi{d!UVu*dDPBX4h1T48Gn3CKpAZ9{=V8@OO^Z>O1(d;`u|4gdPu
z=gnx&EG!TW+Mb^s=jG+G!*1NXX&q-*Y+vmC>*S(*uH%pInd_*r$btVMq3yMov+hj5
zIIondLIK6Tp$1q-Zaj-3)Mt@;tJ|~fRUqDmT{&dWIouAU&d1>62L0sx$nszcdF6kJ
z6#WvvcXwYG7Eby4717nzl`i97s~KGY`2`qy_EVNdCJ*)lA1iPJT~Ds=6;M6;wLA_6
zs?G3JV#vQO@B3euklZ`s`E@z8*_LglLA8d4F^5PyqWcNhMe)7CVEO1iulnoPFY)C(
z%#f(_=XFu=$I_4dOMxeQ0FnC(;I{H1MQGs06G{LEVw)4YsC4?*PU-_2+K<1r@$mq@
z&t}e$Dn>0qqK{pC6EBZHZ0ByNK%t&gz6@9n7Tg^)K|`W`drw52Cyh4YEK=U3@SBY2
z^|~1gQ1o{H#ryZ~H;UeX$@O4VMNLgA$FKMxKW|dM22G@QNO-6c$DUgmM_%`aJ!;@O
zJPdFFZLVU_bv^UtANbr6oo?r^r?*7{_Dqld5CfP~ktA^b<mUx)m#Km2PC0)y^>>o{
zbL85dEm1x#O8)T3HQ{Q?FK=Wx_8BQCrn*&8p8@oW00)3aKOq(RK!zrv1@fn`wT4Ym
zk25%cIUjQ)egN`bk)$0&Yd%-I`&>h^X7vM*EWRT(GsCYNaGsPNh?Z6*sSFKl?2Wm)
z#eJf1$qv0tIgbEAXE|H+wjnzYgP+#NhE~-g6JTD-&%dV30crUk-dYKpEiD@JAO#YU
zPm72oP6pb*0#44c0wq1IC~uv#m7B>?-cphC$qp}?AwSZx$3#->e+#p~Lb}Wg+PN%&
zBfj2dUD4{~ajo!y(F%;`=Y7YgLXT*@MFzop5|b!iskfeb*szx3d}d=~2-I5HMnu=n
zSYoPXPh7OtBvosy!5~P=glP<Mr0A4Ip(bIlZbY66VkOgG6GyRzCYJ;rBwk6L&vM8c
z(aGY2ZAPm!qnNkN3N`-7N@f^B<UvXEv{`dh<8yZqm=I}?J|b@Y4!4%}8y9Lnyf3-p
z*Rgiif&Epeb%txqsd&)N4hQ=lCSNCE#b+(vWQQ^kWL#q28fUu{K-qZdD}flKAJr*d
z(Ppt5uuhyoenwc|#+K^B&Ep?dczHv``+>xpgGDIA%8kwkQ}8Qu02W-r)_F_$7#=K^
zBj)L7&MU(+{8L!*5<58bK($8nACnkzx*~ZOp%_XR$`o=t``5BObPEM9_({JjvL)-<
zv^WXq*;q=PvHX}*h>3}>&R)Q0pZr;rv`j0~Z=8Yk_N8vrWv`K=Oswb60G)JtuM7{!
z>1|@sMgoR?L0GPye<9Qa&ki8Oi{n252_Ah_-ph~HXhvNRjg{f7KdJLL_cnXyDl|HW
zmbdJ^1!KcU&Z9g5IO6F`g#>c15z+u&{|L^#nq{hFi|(7*NgowKT~y`1?uD6Fk*F$3
zJ8qI<J(p$}z=;o>Fpm(1ede7c_(IlQGux}K3_^_wO7B9=vI8aL)d}XGCVb!ZZCTyk
zgB>mnexwTo$T_weL1}T+cHr}Lfb0dnS`BflU=O{rvzr7gXhoQeGU|<i{L_r!@<KhP
zeBAc`d6G~HASii~{sVs?{#{4hn;}6{IQ0FJU?Vo_b&a?fLEWgX#m@mE#0|oy4E5jX
zb+tRJkL6?fo5hR6Cj!tu=M~}=9~nW#!hiX{M=$Gu;$;4g&wb2;Q>6^btj34pyP_-)
zJu3(e*1C#$fWgxi7H(fE;PZu9VIZgF;@=O{03PV?m^qXF?=rKMnmCg4>##*eWI1F6
zKkX<(>`N6}|0v)huYV^g=hM2=esn-xM=qiKNd35F(}Dkzn<R?V&HX#79Iy0XgY3Yt
zzoMcv(>no+zo<Ay)+bh~$5b$t0_>LCINuN7bGKKl%5_$eHDF8O-Me@6Lg@0!eH<~t
zrjz<163GC6NN5?juRkmSP!HB%N5tvml9JFz+rPi}^(`ddW1X!TwLor6HXxLhp|rFS
z-Sfq(WfOc*w-^=rzxmoKiX_TCbb+E?^8rCHvGX?bt}Hb-%m<}{LS@9W@EtOTz#r2I
zQ9T8|pfth!IjQsnA)6rkfz9w#^v0<ayZF+QCuxh?DgJazYaehoieDTMk|b$F4FG#W
zGhZ;OIwI*d^iv{Pr>XRxy&=cR`VV&{zwY+S^#eqOFJL+_NP2Qv25b`z6>8b8V`}!+
zC(_<He~PvnD%hU#nVI$eYz)-z92^{zzKfZTnCc?_LM7o58FZk<#J8-9hjDKB{!LWZ
zpma%2ao-7raN#LmbW3Q<!$R%PgSOODIOm>$OEomDa_D1?s&Rw}XQwM^Jq3zch5(V2
zyCB_)51-Rnl3;!KiD=;j#FQ}_4Dycz>L}xMojSS~+Zg-z%FEq9H*N5LHaD=T6Ki&;
z(`^Y&>kB;&4sW1qt*vl69_e@TJ?Hq`6qq&DHwFH!72s8Z19A9oHl<m&Mb5g4i*GjW
zitqN^@daNJpOlQux<5-MudtAYnYm}|12Q@;&JkO@9))w0<BpnFBxJwQ?-utr)3&F>
z+tXc<>YFBAR7}VMCgi(#`2F^Dsd)IkL>`BmA7kx_TfsEK1MX}B<Jy8_46w>cRn`a;
zF#Y*@fnezYHVOxVq`m0iN&gigXEa6mZ52z5jZ2fmsDSgQ<y&A^Ew-2mI^_o*ik*vP
z(D9>5XEB;v>TeAzECApkI0uZj?<<9vtkN$#j1z89@4;9M<^uCLYU{I;WorQyfrnE*
zhWgJPIM@O<a#<s|hwXcJUV8|0)h~FfAbQuEb+J*w<9Om+gg~igdy)T1`V+GEH}doF
zq%1LIb|F8o0p~(VSg{|w1xoghESmhbf2axKS6a6b7ieaxR#-IE186|Pr<XeM$>v6P
zlaz|Db)&ZT2%|<A2}hfZB4>aYlROcPjV161%*T*AZ4-Q@SBfwN{q-&eX`gLqhF9Gm
z#<v+(A%BgQd;48Rb<=*lhcm%)0GaZSZx(xv|6%t_et>h5O8awqNI*?T7d5negNy6^
ziiwoZ#@BV{8m$HWJDof{IAXK?gcQpc!`@Ig{F6Vf<IMFt!J!ScI!-7s-c=g94@jTj
z5Y1quKiM)}{-5a3ptJ3Jii(QqA`aI+d#}U9#nW-d{1`G3f2&t)#&y)(Cpj(cUO<4f
zp}80BsF?=^9uBebN-kIy$am?HojpCJ<VmSJ24)Knf!-<7EAcv@ARVEAo#5!~YW9fQ
zOzo6S?>Y-iKJ`-kN5Oo(CSV{Yxu>n)|LgiV>@5xA=3RR;!=1-M$h+&YxxJZEb$t%|
z8&hu{b-z5W2`YB-+v_fC+-nn9nDQnzjsPWBk$R~>^*Z`;clD2_kd;iUoHUUPCoUoW
z_jRuqTuvefh6Zqi^7SS>^4?wzNwqTuBJgBtH&1G_i5+i+Pz%1?QtIYrx*ztIjY{15
z8o&|F>lvb1%e(;ALb20QDm^~>B(TAT_9fsluCR-j-<GR?$t8}BeFRb=*;GHz7gUQF
zaXlgTc|=``3kwa8r1(D$GfI7_0qXDWKj3H#{4FkPQJ3AReOfZ?lt78d)c9a@R|R2J
zUZTg83pEKzF$5n4MCtVN=WBD30JLQ0tkjV%bhq%v)@W>$o<{sadHx+yu7O=bw&}|=
za^yt!&Vs?HrUyn=9>^OM{k7T_00PP0(I<wi<V2QltdHxV{r&szC4>a<TG&7Vnaerx
zBO;u#)=53~Qke4GHK_3|7&x{i_vMZ)A`iXz2r(yGfjo)nl!kI&O>TnikYqa`P@QV9
z=f8cMxL~vAXUyimzxXb;yJd`sea}wQFaDlx!a84!S#JEYu4Sj(8@#LeVOYUcU(1zW
z?%H&yi~M6MMsK3YK+<1+Cl?si8aNQR5+4vG;!zD7Jc(-<q^Hq8h6F+M@LV7$73wW`
zCBqV)8R6Ljguvma9$mD&72W_FJ)!h$KNXbq%GdWPga9n_$JaRjHNwb<9$+{aNyX6g
zO2)<vrw8lTZ{Pk1v`(+J?>A^^xjOR8wAP$mkR74LL$E>v!&IsE+ObUZw*P<%M0Zk{
zLK5TQr!o_T1Tvv=49J*}Vb~6z>-DXdg&l|XSO~I*qu>7z(GjS<6lqU6+8p8<mh@U3
zi3brFfC5;c6$MnMD@;skI?rYI?X*;RCq1JnkipJm9101Hx_J>q6L{6F+&>^t!V3RO
zF;&)U@&A79i4iDIhyZclAx7r%*yFfuO_T}6<lJ!(9yM|44mVZD0eJ<LQADc!O%UN(
z9xgEir>_)U1F<`esB|IsnC|<i%dNAt1Cf$WT9qX)lil$9$)-;jYzpnDOZBb`kMCZA
zzWV)rs%&+??sxr~V2BKQ)Z!Fp>-=^Pp|5ZcNGLj!C1!)UH=#<eCZ}R}fe?-SZS|yb
z`jA3e{T5^RueYB#>ubk3GQgDhZR;^>ICFIINwr2nW1J@7j;p`SymE&-9O{PG0y#(8
z-?G3RXc5(c?`wgxqP9<1E0D$iW(TC)QUURkcX?kl%$^F>X|c#ZRnd<B*p-FFT!BHd
zDDknR_DA&jdP795S$l-RedD(&mM%BWRRyyQSR!EY@QAFy>(ZzAwUC==6)aQ5&O@s=
zY^#M}HpcVt`#>34x~e%c6L>C_6>=KgKr3u520v@oeLG`wVQE?<TJVj6jrPCA@SC0S
zrh#p<z!gBXM$CFG*%ivLA#q@ghg64e@oX7F?_W!{0TI7oFYOEv!EOhqrO+^%`6oKe
z7rWM!(HriH%aOFs->f_jc6|~aTmdHR@D-BwDT2wx8go!vHEe=XF<i5%uL{%>ZLden
zDLM|kiPrFa3XJ7~QR^0xb7imo&Clqv^P*;O_uBI6FXkd_l@;|{RsyWk5e^;0a)uRc
zGYYP*^k;5q!4*Y%Ot&Pdp8SsPggOd(AaCL`;1_H9zdIb1+mk)%_%Pj(1<vXTEmrq`
z=j2*7c)G;q$M&bq#i<fR_HMv<<KellKeLu)PT#v2`1?Oosy}r*R85KJIp7XwzHFvu
zz^4Qec94e!AiTf7;Z9SdEM0v48RGQpf4R<^P2_Lb5-d<se*>Yo+~<Ekm(3LqrILh_
z<^gz$8<vP^+T^^2aziy3%n;S72>niz9AnZ|7~Po&)8PXNL7QSNUrXpKPTWyvhKTb4
z&l(pAK($zuosaGeg8Et@t^_1B7$9_%fm`)5x|3UWJ{MurP$;dBD0jwa4c<m|D#1Q>
zdk9j+c1pp1w3V1`B6;7gyZnB_w8n-ZKGmwKZf&OzJU%9xj5m)j7o(0fPu|$YJ&zy$
zBAy&cIRQ^Vmmh&LdRU&+l~$WH?`c<fNbi{b1uP^PZH?VoI5+EcWn&`-X$6K&ya398
zIepnykIe!|4muOf!QQURdWjkTe2F_7aux6Q>(eTcwkBrMuYS#mcv*PYRs$W)1)YMK
z0<)_r4*COsijD;ze*8XxeS9(Cv>jnSxI||$5r`LVUXwSzI{MrKooB*`B?*a;J-r4w
z^lg?K8j`=~n<_^dFfo~3*kFkkX76|K7ubp!t;O>XVaQrH&dN62mRz94&M%zv^#-tA
z2PqHrro^(<{Pj)vjG$BaAAh+TT*HpfMU93y&eRdVLRitoU4g)%VyB`XQ6IK<ik$Vj
zkT%c_yokgH_0dFLDRs8C1gnj;4wGs~(@2q@jcE?jr@S>2{@gIb7!`q9l35@0pDY+?
zC8V|`DA#Fgdr0(%P}qwJi1{y+{hJp~*81@uzOLmdGZ5aNX^HHbxCw2zS`VFY)%)S^
zdpQnX!GwGm+Beqj6JDsYiClfVUJ<<1f9f_Mdh}J~pS#TaH@BaxCAJ;heIGfmd-2EN
zed81~xjSSrG&T-i(iF32!AgScQCRzP_VrGu@_+6D{~<co^17H}gnbMbyC0KIO-?;U
z<cS?rB=7g?m5n@Ru{i8U5)LP(kKq4NzVR_($6Fck7{4Dv$B1m+=Pq{>R7Ql%f*!X8
z_Sw}P{)nP^lQu#GdK&gmPvHZi;#QCuP}IHSVH#k4kJE+7&zfnbE>!UKXAA!22o`hZ
zNS0;e*<;zuZ`P(sq?UdyjzR0>3J;RMYvIpkv7ITQX?y=Pyq=YvhS)%ISGlr7ddno4
zq`FSM?GxU|w<Wx9YSRr8U3ay)U8{tB%syT<Me&qvY~BvFFI)ROt+pk;SqqHlMOk~M
zwd~U|J}~8jw`z}egsBtp8O8)o4nzLVGoYjWaJ@Frk~?e14xK3fVR2P2d-oKwD?_m#
zggsa|SMT*?qtM<s?zh<AkytRXDqOS>Akp0BW?LV(T$9*X`Detlpl!$OvdBXN(GK!#
z|BNY@+)Yp%cixcr_aQ{pnE0}2aJoy7x)c-s_lm&y@AJm~o{|Nupa1jl#$pCkCph_V
zaB^Q86gsF(SP=ozqPHgE8cL-Ib@iw*Qq*S(EcRtX;5;)<aQ<T$kQzLYt)S>N$F<D~
zijaC0L1Qb}<(#KF2-)pDbx>Kbc*4Bv03GMSoiC<`xOI}Eg)0gNM=Fvi&v9h$m?CT{
zP&TohRj?~RuV!n2+7(HsaQ(Nqt@BS1zBJKSd@6#?SQYXR)EV<r>d#reur2AW2-MzQ
zO7^85q79PgA?N5ti_r}-Pr?(cQR$Y;^)PhyJ#ZTua2qtB`d*z}U%vU?&o?#DD()CV
zuWKFWOJPK2w#@y8NZKLJc!4_@YP6|;Zi;0<_HbtHNF#e-6%KR+hm1(oc(_!^G5WoH
zkT&RE=)by$<v^-u9CnxR^)8MHzkRcKx`7tR-qGt13b|@!DzoW7M*n*UBS83$$<hg2
zM)R-BR?>934=nl*mupfE57tN~j1DiU*MPTau2S^Kbi!($4ty~E7I1W(Bb01sE;}AF
zy_M37?@XNxhyA*mY>alrdTAZEgj&XSHe~J4LthR2WRnNog#W9XXZbHBcm%gnnn)*!
z=OBlWG)AO1$X8pCu;N(|-43-RSY$wY`*+EGqg;7P-wCOOEDFyqg1v7<egd3?@RfdO
z(o4Ho`ARYo8oYv;*PjKzh?#!ER0;ux_utX=KTi4tu`jS&{cju3f3~py90>Y%djluk
gf#RlruDoCiH~7FaOQ^~R&OF7tudInESF{TIe?Oh$W&i*H
literal 0
HcmV?d00001
diff --git a/docs/source/_static/theme_overrides.css b/docs/source/_static/theme_overrides.css
deleted file mode 100644
index cc6ce10d..00000000
--- a/docs/source/_static/theme_overrides.css
+++ /dev/null
@@ -1,15 +0,0 @@
-# https://rackerlabs.github.io/docs-rackspace/tools/rtd-tables.html
-
-/* override table width restrictions */
-@media screen and (min-width: 767px) {
-
- .wy-table-responsive table td {
- /* !important prevents the common CSS stylesheets from overriding
- this as on RTD they are loaded after this stylesheet */
- white-space: normal !important;
- }
-
- .wy-table-responsive {
- overflow: visible !important;
- }
-}
\ No newline at end of file
diff --git a/docs/source/changelog/index.rst b/docs/source/changelog/index.rst
new file mode 100644
index 00000000..534b1dd3
--- /dev/null
+++ b/docs/source/changelog/index.rst
@@ -0,0 +1,85 @@
+.. _changelog:
+
+Changelog
+=========
+
+CellBender has been publicly available since 2019, and there have been some
+significant changes over the years. Major changes and updates are denoted by
+a change in the (first or) second digit of the version number. For example,
+major changes were made between version 0.1.0 and 0.2.0. Small changes,
+edge case bug fixes, speedups, and small new features might bump up the last
+digit of the version number. For example, the difference between 0.2.1 and 0.2.0
+represents this kind of small change.
+
+Version 0.1.0
+-------------
+
+This was the initial release. The output count matrix was constructed via
+imputation, so that there were no explicit guarantees that CellBender would
+only subtract counts and never add.
+
+This version has been deprecated, and we do not recommend using it any longer.
+
+- Imputes the "denoised" count matrix using a variational autoencoder
+
+Version 0.2.0
+-------------
+
+A significant overhaul of the model and the output generation procedure were
+undertaken to explicitly guarantee that CellBender only subtracts counts and
+never adds. The output is not constructed by imputation or smoothing, and
+CellBender intentionally tries to modify the raw data as little as possible in
+order to achieve denoising. A nominal false positive rate is approximately
+controlled at the level of the entire dataset, to prevent removal of too much
+signal.
+
+- Uses a variational autoencoder as a prior
+
+- Computes the "denoised" count matrix using a MAP estimate and posterior regularization
+
+ - CellBender never adds counts
+
+Version 0.3.0
+-------------
+
+Further significant changes were made to the way the output count matrix gets
+computed. We compute a budget of noise counts for each gene individually, and
+partition those noise counts in a way that optimizes the posterior probability
+of our model. This involves solving an auxiliary optimization problem.
+
+The input arguments ``--expected-cells`` and ``--total-droplets-included``, which
+have always been required for the tool to work reliably, are now much more
+optional. Extensive testing has shown that these arguments can be left out the
+vast majority of the time, making things easier on users.
+
+A great amount of effort has been put into refinements which reduce the failure
+rate on difficult samples.
+
+Output data loading functions in python have been distributed with the CellBender
+codebase in order to facilitate downstream data loading and use in
+`scanpy <https://scanpy.readthedocs.io/>`_.
+
+In addition, the code produces checkpoints at timed intervals, and these
+checkpoints can be used to deterministically pick up where the code left off.
+This is extremely advantageous in cloud compute settings, and was built with
+`Terra <https://app.terra.bio>`_ in mind. The checkpointing functionality,
+together with the re-written WDL workflow, allow the workflow to be run
+seamlessly on preemptible GPU instances, which are a fraction of the cost of
+non-preemptible machines. We also hope these checkpoints make it easier to run
+a workflow using Google Colab on a GPU for free.
+
+- Produces checkpoint files, and the WDL can run seamlessly on preemptible GPUs
+
+- Computes the "denoised" count matrix by solving an auxiliary optimization problem
+
+ - We demonstrate that this approach is superior to v0.2.0 in the publication
+
+- Several tweaks to the inference procedure lead to speedups and better performance
+ on benchmarks
+
+- The tool produces an output report in HTML format with plots, analysis,
+ commentary, warnings, and recommendations
+
+- A ``metrics.csv`` output file is produced for those interested in running
+ hundreds of samples in automated pipelines. This file can be parsed to look for
+ indications that a sample may need to be re-run.
diff --git a/docs/source/citation/index.rst b/docs/source/citation/index.rst
index d55e2e3e..65b8beb1 100644
--- a/docs/source/citation/index.rst
+++ b/docs/source/citation/index.rst
@@ -3,10 +3,13 @@
Citation
========
-If you use CellBender in your research (and we hope you will!), please consider citing our papers:
+If you use CellBender in your research (and we hope you will!), please consider
+citing our paper:
-`remove-background <https://www.biorxiv.org/content/10.1101/791699v1>`_:
+Stephen J Fleming, Mark D Chaffin, Alessandro Arduini, Amer-Denis Akkad, Eric Banks,
+John C Marioni, Anthony A Philippakis, Patrick T Ellinor, and Mehrtash Babadi.
+Unsupervised removal of systematic background noise from droplet-based single-cell
+experiments using CellBender. *Nature Methods* (in press), 2023.
-Stephen J Fleming, John C Marioni, and Mehrtash Babadi. CellBender remove-background: a deep
-generative model for unsupervised removal of background noise from scRNA-seq datasets.
-bioRxiv 791699; doi: https://doi.org/10.1101/791699
+Preprint `available at bioRxiv <https://www.biorxiv.org/content/10.1101/791699v2>`_.
+https://doi.org/10.1101/791699
diff --git a/docs/source/conf.py b/docs/source/conf.py
index f41e45d4..64a32149 100644
--- a/docs/source/conf.py
+++ b/docs/source/conf.py
@@ -12,6 +12,9 @@
#
import os
import sys
+from cellbender.base_cli import get_version
+
+
dir_, _ = os.path.split(__file__)
root_dir = os.path.abspath(os.path.join(dir_, '..', '..'))
sys.path.insert(0, root_dir)
@@ -23,8 +26,8 @@
author = 'Stephen Fleming, Mehrtash Babadi'
# The full version, including alpha/beta/rc tags
-version = ''
-release = ''
+version = get_version()
+release = get_version()
# -- General configuration ---------------------------------------------------
diff --git a/docs/source/contributing/index.rst b/docs/source/contributing/index.rst
index 2986dbd4..7d89ef94 100644
--- a/docs/source/contributing/index.rst
+++ b/docs/source/contributing/index.rst
@@ -3,10 +3,12 @@
Contributing
============
-We aspire to make CellBender an easy-to-use, robust, and accurate software package for the bioinformatics community.
-While we test and improve CellBender together with our research collaborators, your feedback is
-invaluable to us and allow us to steer CellBender in the direction that you find most useful in your research. If you
+We aspire to make CellBender an easy-to-use, robust, and accurate software package
+for the bioinformatics community. While we test and improve CellBender together
+with our research collaborators, your feedback is invaluable to us and allows us
+to steer CellBender in the direction that you find most useful in your research. If you
have an interesting idea or suggestion, please do not hesitate to reach out to us.
-If you encounter a bug, please file a detailed github `issue <https://github.com/broadinstitute/CellBender/issues>`_
+If you encounter a bug, please file a detailed github
+`issue <https://github.com/broadinstitute/CellBender/issues>`_
and we will get back to you as soon as possible.
diff --git a/docs/source/getting_started/index.rst b/docs/source/getting_started/index.rst
deleted file mode 100644
index a9bf9b9e..00000000
--- a/docs/source/getting_started/index.rst
+++ /dev/null
@@ -1,11 +0,0 @@
-.. _getting started:
-
-Getting Started
-===============
-
-This section contains quick start tutorials for different CellBender modules.
-
-.. toctree::
- :maxdepth: 1
-
- remove_background/index
\ No newline at end of file
diff --git a/docs/source/getting_started/remove_background/index.rst b/docs/source/getting_started/remove_background/index.rst
deleted file mode 100644
index 90de53f4..00000000
--- a/docs/source/getting_started/remove_background/index.rst
+++ /dev/null
@@ -1,138 +0,0 @@
-.. _remove background tutorial:
-
-remove-background
-=================
-
-In this tutorial, we will run ``remove-background`` on a small dataset derived from the 10x Genomics
-``pbmc4k`` scRNA-seq dataset (v2 Chemistry, CellRanger 2.1.0).
-
-As a first step, we download the full dataset and generate a smaller `trimmed` copy by selecting 500 barcodes
-with high UMI count (likely non-empty) and an additional 50,000 barcodes with small UMI count (likely empty). Note
-that the trimming step is performed in order to allow us go through this tutorial in a matter of minutes on a
-typical CPU. Processing the full untrimmed dataset requires a CUDA-enabled GPU (e.g. NVIDIA Testla K80)
-and takes about 30 minutes to finish.
-
-We have created a python script to download and trim the dataset. Navigate to ``examples/remove_background/``
-under your CellBender installation root directory and run the following command in the console:
-
-.. code-block:: console
-
- $ python generate_tiny_10x_pbmc.py
-
-After successful completion of the script, you should have a new directory named ``tiny_raw_gene_bc_matrices``
-containing ``GRCh38/matrix.mtx``, ``GRCh38/genes.tsv``, and ``GRCh38/barcodes.tsv``.
-
-Run remove-background
----------------------
-
-We proceed to run ``remove-background`` on the trimmed dataset using the following command:
-
-.. code-block:: console
-
- $ cellbender remove-background \
- --input ./tiny_raw_gene_bc_matrices/GRCh38 \
- --output ./tiny_10x_pbmc.h5 \
- --expected-cells 500 \
- --total-droplets-included 5000
-
-Again, here we leave out the ``--cuda`` flag solely for the purposes of being able to run this
-command on a CPU. But a GPU is highly recommended for real datasets.
-
-The computation will finish within a minute or two (after ~ 150 epochs). The tool outputs the following files:
-
-* ``tiny_10x_pbmc.h5``: An HDF5 file containing a detailed output of the inference procedure, including the
- normalized abundance of ambient transcripts, contamination fraction of each droplet, a low-dimensional
- embedding of the background-corrected gene expression, and the background-corrected counts matrix (in CSC sparse
- format). Please refer to the full documentation for a detailed description of these and other fields.
-
-* ``tiny_10x_pbmc_filtered.h5``: Same as above, though, only including droplets with a posterior cell probability
- exceeding 0.5.
-
-* ``tiny_10x_pbmc_cell_barcodes.csv``: The list of barcodes with a posterior cell probability exceeding 0.5.
-
-* ``tiny_10x_pbmc.pdf``: A PDF summary of the results showing (1) the evolution of the loss function during training,
- (2) a ranked-ordered total UMI plot along with posterior cell probabilities, and (3) a two-dimensional PCA
- scatter plot of the latent embedding of the expressions in cell-containing droplets. Notice the rapid drop in
- the cell probability after UMI rank ~ 500.
-
-Finally, try running the tool with ``--expected-cells 100`` and ``--expected-cells 1000``. You should find that
-the output remains virtually the same.
-
-Use output count matrix in downstream analyses
-----------------------------------------------
-
-The count matrix that ``remove-background`` generates can be easily loaded and used for downstream analyses in
-`scanpy <https://scanpy.readthedocs.io/>`_ and `Seurat <https://satijalab.org/seurat/>`_.
-
-To load the filtered count matrix (containing only cells) into scanpy:
-
-.. code-block:: python
-
- # import scanpy
- import scanpy as sc
-
- # load the data
- adata = sc.read_10x_h5('tiny_10x_pbmc_filtered.h5', genome='background_removed')
-
-To load the filtered count matrix (containing only cells) into Seurat:
-
-.. code-block:: rd
-
- # load Seurat (version 3) library
- library(Seurat)
-
- # load data from the filtered h5 file
- data.file <- 'tiny_10x_pbmc_filtered.h5'
- data.data <- Read10X_h5(filename = data.file, use.names = TRUE)
-
- # create Seurat object
- obj <- CreateSeuratObject(counts = data.data)
-
-Use latent gene expression in downstream analyses
--------------------------------------------------
-
-To load the latent representation of gene expression `z` computed by ``remove-background`` using a python script:
-
-.. code-block:: python
-
- import tables
- import numpy as np
-
- z = []
- with tables.open_file('tiny_10x_pbmc_filtered.h5') as f:
- print(f) # display the structure of the h5 file
- z = f.root.background_removed.latent_gene_encoding.read() # read latents
-
-At this point, the variable ``z`` contains the latent encoding of gene expression, where rows are cells and
-columns are dimensions of the latent variable. This data can be saved in CSV format with the following command:
-
-.. code-block:: python
-
- np.savetxt('tiny_10x_pbmc_latent_gene_expression.csv', z, delimiter=',')
-
-This latent representation of gene expression can be loaded into a Seurat object ``obj`` by doing the following:
-
-.. code-block:: rd
-
- # load the latent representation from cellbender
- latent <- read.csv('tiny_10x_pbmc_latent_gene_expression.csv', header = FALSE)
- latent <- t(data.matrix(latent))
- rownames(x = latent) <- paste0("CB", 1:20)
- colnames(x = latent) <- colnames(data.data)
-
- # store latent as a new dimensionality reduction called 'cellbender'
- obj[["cellbender"]] <- CreateDimReducObject(embeddings = t(latent),
- key = "CB_",
- assay = DefaultAssay(obj))
-
-Or the variable ``z`` (from above) can be used directly in a scanpy ``anndata`` object. The code snippet below
-demonstrates loading the latent ``z`` and using it to do Louvain clustering:
-
-.. code-block:: python
-
- # load the latent representation into a new slot called 'X_cellbender'
- adata.obsm['X_cellbender'] = z
-
- # perform louvain clustering using the cellbender latents and cosine distance
- sc.pp.neighbors(adata, use_rep='X_cellbender', metric='cosine')
- sc.pp.louvain(adata)
diff --git a/docs/source/help_and_reference/index.rst b/docs/source/help_and_reference/index.rst
deleted file mode 100644
index 1f52c737..00000000
--- a/docs/source/help_and_reference/index.rst
+++ /dev/null
@@ -1,10 +0,0 @@
-.. _help and reference:
-
-Help & Reference
-================
-
-.. toctree::
- :maxdepth: 1
-
- remove_background/index
- license/index
diff --git a/docs/source/help_and_reference/remove_background/index.rst b/docs/source/help_and_reference/remove_background/index.rst
deleted file mode 100644
index 6c0b585f..00000000
--- a/docs/source/help_and_reference/remove_background/index.rst
+++ /dev/null
@@ -1,81 +0,0 @@
-.. _remove background reference:
-
-remove-background
-=================
-
-Command Line Options
---------------------
-
-.. argparse::
- :module: cellbender.base_cli
- :func: get_populated_argparser
- :prog: cellbender
- :path: remove-background
-
-WDL Workflow Options
---------------------
-
-A `WDL <https://github.com/openwdl/wdl>`_ script is available as `cellbender_remove_background.wdl
-<https://github.com/broadinstitute/CellBender/tree/master/wdl/cellbender_remove_background.wdl>`_,
-and it can be used to run ``cellbender remove-background`` from
-`Terra <https://app.terra.bio>`_ or from your own
-`Cromwell <https://cromwell.readthedocs.io/en/stable/>`_ instance. The WDL is designed to
-make use of Tesla K80 GPUs on Google Cloud architecture.
-
-In addition to the above command line options, the workflow has its own set of
-input options, which are described in detail
-`here <https://github.com/broadinstitute/CellBender/tree/master/wdl>`_.
-
-.. _remove background reference troubleshooting:
-
-Troubleshooting
----------------
-
-* The learning curve in the output PDF has large downward spikes, or looks super wobbly.
-
- * This could indicate instabilities during training that should be addressed. The solution
- is typically to reduce the ``--learning-rate`` by a factor of two.
-
-* The following warning is emitted in the log file: "Warning: few empty droplets identified.
- Low UMI cutoff may be too high. Check the UMI decay curve, and decrease the
- ``--low-count-threshold`` parameter if necessary."
-
- * This warning indicates that no "surely empty" droplets were identified in the analysis.
- This means that the "empty droplet plateau" could not be identified. The most likely
- explanation is that the level of background RNA is extremely low, and that the value
- of ``--low-count-threshold`` exceeds this level. This would result in the empty
- droplet plateau being excluded from the analysis, which is not advisable. This can be
- corrected by decreasing ``--low-count-threshold`` from its default of 15 to a value like 5.
-
-
-* There are too many cells called.
-
- * Are there? ``remove-background`` equates "cell probability" with "the probability that
- a given droplet is not empty." These non-empty droplets might not all contain healthy
- cells with high counts. Nevertheless, the posterior probability that they are not empty
- is greater than 0.5. The recommended procedure
- would be to filter cells based on other criteria downstream. Certainly filter for percent
- mitochondrial reads. Potentially filter for number of genes expressed as well, if
- this does not lead to complete loss of a low-expressing cell type.
- * Experiment with increasing ``--total-droplets-included``.
- * Experiment with increasing or decreasing ``--empty-drop-training-fraction``.
- * As a last resort, try decreasing ``--expected-cells`` by quite a bit.
-
-
-* There are too few cells called.
-
- * Try estimating ``--expected-cells`` from the UMI curve rather than a priori, and
- increase the number if necessary.
- * Experiment with increasing or decreasing ``--total-droplets-included``.
-
-
-* The PCA plot of latent gene expression shows no clusters or structure.
-
- * Has training converged? Training should proceed for at least 150 epochs. Check to
- make sure that the ELBO has nearly reached a plateau, indicating that training is
- complete. Try increasing ``--epochs`` to 300 and see if the plot changes.
- * This is not necessarily a bad thing, although it indicates that cells in the experiment
- had a continuum of expression, or that there was only one cell type. If this is
- known to be false, some sort of QC failure with the experiment would be suspected.
- Perform a downstream clustering analysis with and without ``cellbender remove-background``
- and compare the two.
diff --git a/docs/source/index.rst b/docs/source/index.rst
index ce71b411..5e23cc4e 100644
--- a/docs/source/index.rst
+++ b/docs/source/index.rst
@@ -4,11 +4,23 @@ Welcome to CellBender's documentation!
.. image:: /_static/design/logo_250_185.png
:height: 185 px
-*CellBender* is a software package for eliminating technical artifacts from high-throughput single-cell RNA sequencing (scRNA-seq) data.
+*CellBender* is a software package for eliminating technical artifacts from high-throughput
+single-cell RNA sequencing (scRNA-seq) data.
For a brief survey of the package and its current capabilities, you can read
-the :ref:`Introduction <introduction>`. For more comprehensive instructions
-and quick-start tutorials, see the :ref:`Getting Started <getting started>` section.
+the :ref:`Introduction <introduction>`.
+
+For an explanation of when and how and why to use CellBender in your analysis
+pipeline, along with example code, how to choose parameter settings, and
+tips on quality control, read the :ref:`Usage <usage>` section.
+
+For a step-by-step tutorial on a simplified example dataset, see the
+:ref:`Quick start tutorial <quick start tutorial>` section.
+
+For details on input arguments and the output file format,
+see the :ref:`Reference <reference>` section.
+
+For troubleshooting and an FAQ, see the :ref:`Troubleshooting <troubleshooting>` section.
.. toctree::
:maxdepth: 1
@@ -16,7 +28,10 @@ and quick-start tutorials, see the :ref:`Getting Started <getting started>` sect
introduction/index
installation/index
usage/index
- getting_started/index
- help_and_reference/index
+ tutorial/index
+ reference/index
+ troubleshooting/index
contributing/index
citation/index
+ reference/license/index
+ changelog/index
diff --git a/docs/source/installation/index.rst b/docs/source/installation/index.rst
index 86beb63a..4defef3a 100644
--- a/docs/source/installation/index.rst
+++ b/docs/source/installation/index.rst
@@ -3,36 +3,80 @@
Installation
============
-Manual Installation
--------------------
+Via pip
+-------
-The recommended installation is as follows. Create a conda environment and activate it:
+Python packages can be conveniently installed from the Python Package Index (PyPI)
+using `pip install <https://pip.pypa.io/en/stable/cli/pip_install/>`_.
+CellBender is `available on PyPI <https://pypi.org/project/cellbender/>`_
+and can be installed via
.. code-block:: console
- $ conda create -n CellBender python=3.7
- $ conda activate CellBender
+ $ pip install cellbender
+
+If your machine has a GPU with appropriate drivers installed, it should be
+automatically detected, and the appropriate version of PyTorch with CUDA support
+should automatically be downloaded as a CellBender dependency.
+
+We recommend installing CellBender in its own
+`conda environment <https://docs.conda.io/projects/conda/en/latest/user-guide/concepts/environments.html>`_.
+This allows for easier installation and prevents conflicts with any other python
+packages you may have installed.
+
+.. code-block:: console
+
+ $ conda create -n cellbender python=3.7
+ $ conda activate cellbender
+ (cellbender) $ pip install cellbender
+
+
+Installation from source
+------------------------
+
+Create a conda environment and activate it:
+
+.. code-block:: console
+
+ $ conda create -n cellbender python=3.7
+ $ conda activate cellbender
Install the `pytables <https://www.pytables.org>`_ module:
.. code-block:: console
- (CellBender) $ conda install -c anaconda pytables
+ (cellbender) $ conda install -c anaconda pytables
-Install `pytorch <https://pytorch.org>`_ (shown below for CPU; if you have a CUDA-ready GPU, please skip
-this part and follow `these instructions <https://pytorch.org/get-started/locally/>`_ instead):
+Install `pytorch <https://pytorch.org>`_ via
+`these instructions <https://pytorch.org/get-started/locally/>`_:
.. code-block:: console
- (CellBender) $ conda install pytorch torchvision -c pytorch
+ (cellbender) $ pip install torch
+
+and ensure that your installation is appropriate for your hardware (i.e. that
+the relevant CUDA drivers get installed and that ``torch.cuda.is_available()``
+returns ``True`` if you have a GPU available.
-Clone this repository and install CellBender:
+Clone this repository and install CellBender (in editable ``-e`` mode):
.. code-block:: console
- (CellBender) $ git clone https://github.com/broadinstitute/CellBender.git
- (CellBender) $ pip install -e CellBender
+ (cellbender) $ git clone https://github.com/broadinstitute/CellBender.git
+ (cellbender) $ pip install -e CellBender
+
+Install a specific commit directly from GitHub
+----------------------------------------------
+
+This can be achieved via
+
+.. code-block:: console
+
+ (cellbender) $ pip install --no-cache-dir -U git+https://github.com/broadinstitute/CellBender.git@<SHA>
+
+where ``<SHA>`` must be replaced by any reference to a particular git commit,
+such as a tag, a branch name, or a commit sha.
Docker Image
------------
@@ -43,12 +87,16 @@ A GPU-enabled docker image is available from the Google Container Registry (GCR)
Older versions are available at the same location, for example as
-``us.gcr.io/broad-dsde-methods/cellbender:0.1.0``
+``us.gcr.io/broad-dsde-methods/cellbender:0.2.0``
Terra Workflow
--------------
-For `Terra <https://app.terra.bio>`_ users, the following workflow is publicly available:
+For `Terra <https://app.terra.bio>`_ users (or any other users of WDL workflows),
+the following WDL workflow is publicly available:
* `cellbender/remove-background <https://portal.firecloud.org/#methods/cellbender/remove-background/>`_
+
+Some documentation for the WDL is available at the above link, and some is visible
+`on github <https://github.com/broadinstitute/CellBender/tree/master/wdl>`_.
diff --git a/docs/source/introduction/index.rst b/docs/source/introduction/index.rst
index aef82c46..7bcf04a9 100644
--- a/docs/source/introduction/index.rst
+++ b/docs/source/introduction/index.rst
@@ -1,29 +1,30 @@
.. _introduction:
What is CellBender?
-=====================
+===================
-CellBender is a software package for eliminating technical artifacts from high-throughput single-cell RNA sequencing
-(scRNA-seq) data.
+CellBender is a software package for eliminating technical artifacts from high-throughput single-cell omics data,
+including scRNA-seq, snRNA-seq, and CITE-seq.
Scope and Purpose
-----------------
-Despite the recent progress in improving, optimizing and standardizing scRNA-seq protocols, the complexity of
-scRNA-seq experiments leaves room for systematic biases and background noise in the raw observations. These nuisances
+Despite the recent progress in improving, optimizing and standardizing droplet-based single-cell omics protocols
+like scRNA-seq, the complexity of these experiments leaves room for systematic biases and background noise in
+the raw observations. These nuisances
can be traced back to undesirable enzymatic processes that produce spurious library fragments, contamination by
exogeneous or endogenous ambient transcripts, impurity of barcode beads, and barcode swapping during amplification
and/or sequencing.
The main purpose of CellBender is to take raw gene-by-cell count matrices and molecule-level information produced
-by 3rd party pipelines (e.g. CellRanger, Alevin), to model and remove systematic biases and background noise, and
-to produce improved estimates of gene expression.
+by 3rd party pipelines (e.g. CellRanger, Alevin, DropSeq, etc.), to model and remove systematic biases and
+background noise, and to produce improved estimates of gene expression.
As such, CellBender relies on an external tool for primary processing of the raw data obtained from the
sequencer (e.g. BCL or FASTQ files). These basic processing steps lie outside of the scope of CellBender
and include (pseudo-)alignment and annotation of reads, barcode error correction, and generation of raw gene-by-cell
count matrices. Upcoming modules of CellBender will further utilize molecule-level information (e.g. observed reads
-per molecule, transcript equivalent classes, etc.).
+per molecule, transcript equivalence classes, etc.).
Modules
-------
@@ -31,6 +32,5 @@ Modules
The current version of CellBender contains the following modules. More modules will be added in the future:
* ``remove-background``: This module removes counts due to ambient RNA molecules and random barcode swapping from
- (raw) UMI-based scRNA-seq gene-by-cell count matrices. At the moment, only the count matrices produced by the
- CellRanger `count` pipeline is supported. Support for additional tools and protocols will be added in the future.
+ (raw) UMI-based scRNA-seq gene-by-cell count matrices. Several file formats for count matrices are supported.
A quick-start tutorial can be found :ref:`here <remove background tutorial>`.
diff --git a/docs/source/reference/index.rst b/docs/source/reference/index.rst
new file mode 100644
index 00000000..ead0ce72
--- /dev/null
+++ b/docs/source/reference/index.rst
@@ -0,0 +1,188 @@
+.. _reference:
+
+Reference
+=========
+
+.. _remove background reference:
+
+remove-background
+-----------------
+
+Command Line Options
+~~~~~~~~~~~~~~~~~~~~
+
+.. argparse::
+ :module: cellbender.base_cli
+ :func: get_populated_argparser
+ :prog: cellbender
+ :path: remove-background
+
+WDL Workflow Options
+~~~~~~~~~~~~~~~~~~~~
+
+A `WDL <https://github.com/openwdl/wdl>`_ script is available as `cellbender_remove_background.wdl
+<https://github.com/broadinstitute/CellBender/tree/master/wdl/cellbender_remove_background.wdl>`_,
+and it can be used to run ``cellbender remove-background`` from
+`Terra <https://app.terra.bio>`_ or from your own
+`Cromwell <https://cromwell.readthedocs.io/en/stable/>`_ instance. The WDL is designed to
+make use of an Nvidia Tesla T4 GPU on Google Cloud architecture.
+
+As of v0.3.0, the WDL uses preemptible instances that are a fraction of the cost,
+and uses automatic restarting from checkpoints so that work is not lost.
+
+In addition to the above command line options, the workflow has its own set of
+input options, which are described in detail
+`here <https://github.com/broadinstitute/CellBender/tree/master/wdl>`_.
+
+.. _h5-file-format:
+
+Output h5 file format
+~~~~~~~~~~~~~~~~~~~~~
+
+An h5 output file (this one is from the :ref:`tutorial <remove background tutorial>`)
+can be examined in detail using PyTables in python:
+
+.. code-block:: python
+
+ # import PyTables
+ import tables
+
+ # open the file and take a look at its contents
+ with tables.open_file('tiny_output.h5', 'r') as f:
+ print(f)
+
+.. code-block:: console
+
+ /home/jupyter/tiny_output.h5 (File) 'CellBender remove-background output'
+ Last modif.: 'Wed May 4 19:46:16 2022'
+ Object Tree:
+ / (RootGroup) 'CellBender remove-background output'
+ /droplet_latents (Group) 'Latent variables per droplet'
+ /droplet_latents/background_fraction (CArray(1000,), shuffle, zlib(1)) ''
+ /droplet_latents/barcode_indices_for_latents (CArray(1000,), shuffle, zlib(1)) ''
+ /droplet_latents/cell_probability (CArray(1000,), shuffle, zlib(1)) ''
+ /droplet_latents/cell_size (CArray(1000,), shuffle, zlib(1)) ''
+ /droplet_latents/droplet_efficiency (CArray(1000,), shuffle, zlib(1)) ''
+ /droplet_latents/gene_expression_encoding (CArray(1000, 100), shuffle, zlib(1)) ''
+ /global_latents (Group) 'Global latent variables'
+ /global_latents/ambient_expression (Array(100,)) ''
+ /global_latents/cell_size_lognormal_std (Array()) ''
+ /global_latents/empty_droplet_size_lognormal_loc (Array()) ''
+ /global_latents/empty_droplet_size_lognormal_scale (Array()) ''
+ /global_latents/posterior_regularization_lambda (Array()) ''
+ /global_latents/swapping_fraction_dist_params (Array(2,)) ''
+ /global_latents/target_false_positive_rate (Array()) ''
+ /matrix (Group) 'Counts after background correction'
+ /matrix/barcodes (CArray(50500,), zlib(1)) ''
+ /matrix/data (CArray(44477,), shuffle, zlib(1)) ''
+ /matrix/indices (CArray(44477,), shuffle, zlib(1)) ''
+ /matrix/indptr (CArray(50501,), shuffle, zlib(1)) ''
+ /matrix/shape (CArray(2,), shuffle, zlib(1)) ''
+ /metadata (Group) 'Metadata'
+ /metadata/barcodes_analyzed (Array(1000,)) ''
+ /metadata/barcodes_analyzed_inds (Array(1000,)) ''
+ /metadata/features_analyzed_inds (Array(100,)) ''
+ /metadata/fraction_data_used_for_testing (Array()) ''
+ /metadata/test_elbo (Array(30,)) ''
+ /metadata/test_epoch (Array(30,)) ''
+ /metadata/train_elbo (Array(150,)) ''
+ /metadata/train_epoch (Array(150,)) ''
+ /matrix/features (Group) 'Genes and other features measured'
+ /matrix/features/feature_type (CArray(100,), shuffle, zlib(1)) ''
+ /matrix/features/genome (CArray(100,), shuffle, zlib(1)) ''
+ /matrix/features/id (CArray(100,), shuffle, zlib(1)) ''
+ /matrix/features/name (CArray(100,), shuffle, zlib(1)) ''
+
+Any of these bits of data can be accessed directly with PyTables, or by using
+python functions from ``cellbender.remove_background.downstream`` such as
+``anndata_from_h5()`` that load the metadata as well as the count matrix into
+`AnnData format <https://anndata.readthedocs.io/en/latest/>`_
+(compatible with analysis in `scanpy <https://scanpy.readthedocs.io/en/stable/>`_).
+
+The names of variables in the h5 file are meant to explain their contents.
+All data in the ``/matrix`` group is formatted as `CellRanger v3 h5 data
+<https://support.10xgenomics.com/single-cell-gene-expression/software/pipelines/latest/advanced/h5_matrices>`_.
+
+The other root groups in the h5 file ``[droplet_latents, global_latents, metadata]``
+contain detailed information inferred by CellBender as part of its latent
+variable model of the data generative process (details in our paper).
+The ``/droplet_latents`` are variables that have a value inferred for each
+droplet, such as ``cell_probability`` and ``cell_size``. The ``/global_latents`` are
+variables that have a fixed value for the whole experiment, such as
+``ambient_expression``, which is the inferred profile of ambient RNA in the
+sample (normalized so it sums to one). Finally, ``/metadata`` includes other things
+such as the learning curve (``train_elbo`` and ``train_epoch``) and which
+features were analyzed during the run (``features_analyzed_inds``, as
+integer indices that index ``/matrix/features``).
+
+
+.. _loading-outputs:
+
+Loading and using outputs
+~~~~~~~~~~~~~~~~~~~~~~~~~
+
+As mentioned in the :ref:`tutorial <downstream-example>`, output h5 files can
+be opened natively by ``scanpy``
+(`scanpy.read_10x_h5() <https://scanpy.readthedocs.io/en/stable/generated/scanpy.read_10x_h5.html>`_)
+and (with some effort) by ``Seurat`` (see :ref:`here <open-in-seurat>`), though
+there is extra ``cellbender``-specific information in the output files that
+``scanpy`` and ``Seurat`` will not load automatically.
+
+Additional loading functionality is distrubuted as part of the ``cellbender``
+package and is described below.
+
+Suggested use cases:
+
+- Load just the CellBender output: ``anndata_from_h5()``
+- Load the raw input together with the CellBender output (this is great for
+ downstream analysis, because it allows you to plot the effect of ``cellbender``,
+ and you can always look back at the raw data): ``load_anndata_from_input_and_output()``
+- Power users who want to run ``cellbender`` at multiple FPR values and make
+ comparisons: ``load_anndata_from_input_and_outputs()``
+
+Example:
+
+.. code-block:: python
+
+ from cellbender.remove_background.downstream import load_anndata_from_input_and_output
+
+ adata = load_anndata_from_input_and_output(
+ input_file='my_raw_10x_file.h5',
+ output_file='my_cellbender_output_file.h5',
+ input_layer_key='raw', # this will be the raw data layer
+ )
+ adata
+
+.. code-block:: console
+
+ AnnData object with n_obs × n_vars = 6983 × 37468
+ obs: 'background_fraction', 'cell_probability', 'cell_size', 'droplet_efficiency',
+ 'n_raw', 'n_cellbender'
+ var: 'ambient_expression', 'feature_type', 'genome', 'gene_id', 'cellbender_analyzed',
+ 'n_raw', 'n_cellbender'
+ uns: 'cell_size_lognormal_std', 'empty_droplet_size_lognormal_loc',
+ 'empty_droplet_size_lognormal_scale', 'swapping_fraction_dist_params',
+ 'target_false_positive_rate', 'features_analyzed_inds',
+ 'fraction_data_used_for_testing', 'test_elbo', 'test_epoch',
+ 'train_elbo', 'train_epoch'
+ obsm: 'cellbender_embedding'
+ layers: 'raw', 'cellbender'
+
+.. automodule:: cellbender.remove_background.downstream
+ :members:
+
+The posterior
+~~~~~~~~~~~~~
+
+There is an additional output file called ``posterior.h5`` which contains the
+full posterior. The structure of this data is a sparse matrix whose shape is
+[count matrix entries, number of potential noise counts]. Typically the number
+of potential noise counts in any entry of the count matrix is truncated at 50.
+The values are log probabilities of the specified number of noise counts in the
+given count matrix entry.
+
+The information contained in the posterior can be used to
+quantitatively answer questions such as "What is the probability that the
+number of viral gene counts in this cell is nonzero?" For help with these kinds
+of computations, please open a
+`github issue <https://github.com/broadinstitute/CellBender/issues>`_.
diff --git a/docs/source/help_and_reference/license/index.rst b/docs/source/reference/license/index.rst
similarity index 100%
rename from docs/source/help_and_reference/license/index.rst
rename to docs/source/reference/license/index.rst
diff --git a/docs/source/troubleshooting/index.rst b/docs/source/troubleshooting/index.rst
new file mode 100644
index 00000000..893f7bb8
--- /dev/null
+++ b/docs/source/troubleshooting/index.rst
@@ -0,0 +1,426 @@
+.. _troubleshooting:
+
+Troubleshooting
+===============
+
+.. _remove background troubleshooting:
+
+remove-background
+-----------------
+
+Feel free to check the
+`issues on github <https://github.com/broadinstitute/CellBender/issues?q=is%3Aissue>`_,
+where there are several answered questions
+you can search through. If you don't see an answer there or below, please consider
+opening a new issue on github to ask your question.
+
+FAQ
+~~~
+
+* :ref:`What are the "best practices" for running the tool? <a1>`
+
+* :ref:`What are the "best practices" for incorporating CellBender into my analysis pipeline? <a2>`
+
+* :ref:`How do I load and use the output? <a22>`
+
+* :ref:`What are the details of the output h5 file format? <a23>`
+
+* :ref:`Do I really need a GPU to run this? <a3>`
+
+* :ref:`How do I determine the right "--fpr"? <a4>`
+
+* :ref:`How do I determine the right "--expected-cells"? <a5>`
+
+* :ref:`How do I determine the right "--total-droplets-included"? <a6>`
+
+* :ref:`Does CellBender work with CITE-seq data? Or other non-"Gene Expression" features? <a7>`
+
+* :ref:`Could I run CellBender using only "Gene Expression" features and ignore other features? <a8>`
+
+* :ref:`Could I run CellBender using only "Antibody Capture" features and not Gene Expression? <a24>`
+
+* :ref:`Where can I find the ambient RNA profile inferred by CellBender? <a9>`
+
+* :ref:`The code completed, but how do I know if it "worked"? / How do I know when I need to re-run with different parameters? <a10>`
+
+* :ref:`It seems like CellBender called too many cells <a11>`
+
+* :ref:`It seems like CellBender called too few cells <a12>`
+
+* :ref:`The PCA plot of latent gene expression shows no clusters or structure <a25>`
+
+* :ref:`The learning curve looks weird. / What is the learning curve supposed to look like? <a13>`
+
+* :ref:`What is the "metrics.csv" output for, and how do I interpret the metrics? <a14>`
+
+* :ref:`My HTML report summary (at the bottom) said there were some warnings. What should I do about that? <a15>`
+
+* :ref:`The tool failed to produce an HTML report <a16>`
+
+* :ref:`I ran the WDL in Terra and the job "Failed" with PAPI error code 9 <a17>`
+
+* :ref:`How much does it cost, per sample, to run the WDL in Terra? <a18>`
+
+* :ref:`I am getting a GPU-out-of-memory error (process "Killed") <a19>`
+
+* :ref:`I got a "nan" error and the tool crashed <a20>`
+
+* :ref:`There was an error! <a21>`
+
+
+Answers / Troubleshooting Tips
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. _a1:
+
+* What are the "best practices" for running the tool?
+
+ * See the :ref:`recommended best practices <best-practices>`
+
+.. _a2:
+
+* What are the "best practices" for incorporating CellBender into my analysis pipeline?
+
+ * See the :ref:`proposed pipeline <proposed-pipeline>`
+
+.. _a22:
+
+* How do I load and use the output?
+
+ * The data can easily be loaded as an ``AnnData`` object in python, to be used in
+ ``scanpy``, or it can also be loaded by Seurat, see the
+ :ref:`example here <downstream-example>`
+ * We recommend loading the input and output data (together) using the function
+ ``cellbender.remove_background.downstream.load_anndata_from_input_and_output()``
+ from the CellBender package, since that will create an ``AnnData`` with
+ separate layers for the raw data and the CellBender output data. This is
+ quite handy for downstream work. There are several simple data-loading
+ functions in ``cellbender.remove_background.downstream`` that might be
+ useful.
+
+ .. code-block:: python
+
+ from cellbender.remove_background.downstream import load_anndata_from_input_and_output
+
+ adata = load_anndata_from_input_and_output(input_file='tiny_raw_feature_bc_matrix.h5ad',
+ output_file='tiny_output.h5')
+ adata
+
+ .. code-block:: console
+
+ AnnData object with n_obs × n_vars = 1000 × 100
+ obs: 'background_fraction', 'cell_probability', 'cell_size', 'droplet_efficiency', 'n_cellranger', 'n_cellbender'
+ var: 'ambient_expression', 'features_analyzed_inds', 'feature_type', 'genome', 'gene_id', 'n_cellranger', 'n_cellbender'
+ uns: 'cell_size_lognormal_std', 'empty_droplet_size_lognormal_loc', 'empty_droplet_size_lognormal_scale', 'posterior_regularization_lambda', 'swapping_fraction_dist_params', 'target_false_positive_rate', 'fraction_data_used_for_testing', 'test_elbo', 'test_epoch', 'train_elbo', 'train_epoch'
+ obsm: 'cellbender_embedding'
+ layers: 'cellranger', 'cellbender'
+
+.. _a23:
+
+* What are the details of the output h5 file format?
+
+ * :ref:`See here <h5-file-format>`
+
+.. _a3:
+
+* Do I really need a GPU to run this?
+
+ * It's not absolutely necessary, but the code takes a long time to run on a full
+ dataset on a CPU.
+ * While running on a GPU might seem like an insurmountable obstacle for those without
+ a GPU handy, consider trying out our
+ `workflow on Terra <https://portal.firecloud.org/#methods/cellbender/remove-background/>`_,
+ which runs on a GPU on Google Cloud at the click of a button.
+ * Others have successfully run on Google Colab notebooks for free. Since CellBender
+ produces checkpoint files during training (``ckpt.tar.gz``), you can even pick up
+ where you left off if you get preempted. You just need to put the ``ckpt.tar.gz``
+ file in the directory where the CellBender command is being invoked (or
+ specify the file using the argument ``--checkpoint``), and CellBender will
+ automatically pick up where it left off when the same command is re-run.
+ * If you really want to use a CPU only, then consider things that will speed up the
+ run, like using fewer ``--total-droplets-included``, and increasing the threshold
+ ``--projected-ambient-count-threshold`` so that fewer features are analyzed,
+ and maybe decreasing ``--empty-drop-training-fraction``, so that each minibatch
+ has fewer empty droplets.
+
+.. _a4:
+
+* How do I determine the right ``--fpr``?
+
+ * See the :ref:`recommended best practices <best-practices>`
+
+.. _a5:
+
+* How do I determine the right ``--expected-cells``?
+
+ * See the :ref:`recommended best practices <best-practices>`
+
+.. _a6:
+
+* How do I determine the right ``--total-droplets-included``?
+
+ * See the :ref:`recommended best practices <best-practices>`
+
+.. _a7:
+
+* Does CellBender work with CITE-seq data? Or other non-``Gene Expression`` features?
+
+ * Absolutely, `as shown here <https://github.com/broadinstitute/CellBender/issues/114>`_
+ and `in our paper <https://www.biorxiv.org/content/10.1101/791699v2>`_.
+ The results for ``Antibody Capture`` data look even better than
+ for gene expression, due to the higher ambient background for that modality.
+ * ATAC data is a bit more tricky. CellBender will run, though it takes a long
+ time with 200k+ Peak features. You can use the argument
+ ``--projected-ambient-count-threshold 2`` to tell CellBender to ignore all
+ features which are not estimated to have at least two noise counts in cells.
+ This can greatly speed things up. Feel free to experiment with that value.
+ Anecdotally it seems that ATAC data is less noisy than gene expression data
+ to begin with, so some users opt to have CellBender ignore the ATAC features
+ using the input argument ``--exclude-feature-types Peaks``. There is nothing
+ wrong with doing this.
+
+.. _a8:
+
+* Could I run CellBender using only ``Gene Expression`` features and ignore other features?
+
+ * If you want to, you can (though it works great with ``Antibody Capture`` data):
+ just use the ``--exclude-feature-types`` input parameter.
+
+.. _a24:
+
+* Could I run CellBender using only ``Antibody Capture`` features and not ``Gene Expression``?
+
+ * No, it is not a good idea to exclude ``Gene Expression`` features for the following
+ reason: the ``Gene Expression`` features form the basis of a good prior on "cell type",
+ which the method heavily relies on to denoise. Other features, without ``Gene Expression``,
+ are probably too sparse to cluster similar cells together and form a good prior for
+ "which cells are similar to which others".
+
+.. _a9:
+
+* Where can I find the ambient RNA profile inferred by CellBender?
+
+ * This is present in the output h5 file as the field called
+ ``/global_latents/ambient_expression`` (:ref:`see here <h5-file-format>`).
+ If you use the dataloader from ``cellbender.remove_background.downstream``
+ (:ref:`see here <a22>`), then the ambient expression profile will be loaded
+ into the AnnData object as ``adata.var['ambient_expression']``
+ * (Though it is referred to here as "ambient RNA", all features are included,
+ not just Gene Expression.)
+
+.. _a10:
+
+* The code completed, but how do I know if it "worked"? / How do I know when I
+ need to re-run with different parameters?
+
+ * The vast majority of runs using a nice dataset will work just fine. If your
+ dataset might not be so "nice", then we recommend taking a look at the output
+ ``_report.html``, which will have a few diagnostics and will issue warnings
+ and recommendations as appropirate.
+ * In general, if the learning curve (ELBO versus epoch) has huge spikes, or if
+ if does not converge near the end but rather dips back down, then you may
+ need to consider re-running with a lower ``--learning-rate``. The solution
+ is typically to reduce the ``--learning-rate`` by a factor of two.
+ * In certain cases, it may be obvious that CellBender has failed to call cells
+ accurately. In these cases, it may be necessary to do a bit of experimentation
+ with ``--expected-cells`` and ``--total-droplets-included`` to try to guide
+ CellBender toward a more reasonable solution. It has been our observation
+ that such cases are relatively rare. Try looking at the UMI curve and picking
+ a value for ``--expected-cells`` where you know that all the droplets
+ preceding that number are surely cells.
+
+.. _a11:
+
+* It seems like CellBender called too many cells
+
+ * Did it? ``remove-background`` equates "cell probability" with "the probability that
+ a given droplet is not empty." These non-empty droplets might not all contain healthy
+ cells with high counts. Nevertheless, the posterior probability that they are not empty
+ is greater than 0.5. The recommended procedure
+ would be to filter cells based on other criteria downstream. Certainly filter for percent
+ mitochondrial reads. Potentially filter for number of genes expressed as well, if
+ this does not lead to complete loss of a low-expressing cell type.
+ * Experiment with increasing ``--total-droplets-included``.
+ * Experiment with increasing or decreasing ``--empty-drop-training-fraction``.
+ * As a last resort, try decreasing ``--expected-cells`` by quite a bit.
+
+.. _a12:
+
+* It seems like CellBender called too few cells
+
+ * If CellBender seems to have missed cells, or if you get a "No cells found!"
+ error, then try increasing ``--expected-cells``, and also ensure that your value
+ for ``--total-droplets-included`` is large enough that all droplets after
+ this value on the UMI curve are "surely empty".
+
+.. _a25:
+
+* The PCA plot of latent gene expression shows no clusters or structure
+
+ * Has training converged? Training should proceed for at least 150 epochs. Check to
+ make sure that the ELBO has nearly reached a plateau, indicating that training is
+ complete. Try increasing ``--epochs`` to 300 and see if the plot changes.
+ * This is not necessarily a bad thing, although it indicates that cells in the experiment
+ had a continuum of expression, or that there was only one cell type. If this is
+ known to be false, some sort of QC failure with the experiment would be suspected.
+ Perform a downstream clustering analysis with and without ``cellbender remove-background``
+ and compare the two.
+
+.. _a13:
+
+* The learning curve looks weird. / What is the learning curve supposed to
+ look like?
+
+ * The "learning curve", a.k.a. the plot of ELBO (evidence lower bound) verus
+ training epoch, is a record of the progress of inferring all the latent
+ variables in the CellBender model, based on data. This learning happens via
+ gradient descent. Typically, the ELBO changes gradually, increasing and
+ approaching some rather stable value by the end of training. Ideally, the
+ ELBO increases monotonically.
+
+ * If the learning curve either starts decreasing, or has a large downward bump
+ or a downward spike or spikes, then something may have gone a bit "off the
+ rails" during training. We would be concerned that, for example, the inference
+ procedure got thrown off into some local minimum that is sub-optimal. If you
+ see a learning curve that looks strange, then try to re-run with half the
+ ``--learning-rate`` and see if it results in a more "canonical" learning curve.
+ If it does, use that output.
+
+ * Examples: 2 fine learning curves (panels on right are zoomed-in y-axis)
+
+ * .. image:: /_static/remove_background/PCL_rat_A_LA6_learning_curve.png
+ :width: 600 px
+
+
+ * .. image:: /_static/remove_background/simulated_s6_learning_curve.png
+ :width: 600 px
+
+ * Examples: 2 bad learning curves
+
+ * .. image:: /_static/remove_background/bad_learning_curves.png
+ :width: 700 px
+
+.. _a14:
+
+* What is the ``metrics.csv`` output for, and how do I interpret the metrics?
+
+ * This is a bit of a niche output, and most people can ignore it if they want
+ to. The idea here is to enable automated analysis pipelines to make decisions
+ about whether to re-run CellBender with different parameters. Several of the
+ output metrics contained here are also contained in the HTML report (though
+ not all). But, importantly, this CSV file is easy to parse programmatically,
+ so that a pipeline can make automatic decisions. All metrics are scalar
+ quantities, and the intent was to name them so they are self-explanatory.
+ The contents are:
+
+ 1. ``total_raw_counts``: Sum of all input count matrix entries
+ 2. ``total_output_counts``: Sum of all output count matrix entries
+ 3. ``total_counts_removed``: 1 minus 2
+ 4. ``fraction_counts_removed``: 3 divided by 1
+ 5. ``total_raw_counts_in_cells``: Same as 1, but calculated only in CellBender-
+ determined non-empty droplets
+ 6. ``total_counts_removed_from_cells``: 5 minus 2 (since only cells have
+ nonzero counts in the output)
+ 7. ``fraction_counts_removed_from_cells``: 6 divided by 5
+ 8. ``average_counts_removed_per_cell``: 6 divided by the number of CellBender-
+ determined non-empty droplets
+ 9. ``target_fpr``: The input ``--fpr`` value
+ 10. ``expected_cells``: The input ``--expected-cells`` value, blank if not
+ provided.
+ 11. ``found_cells``: The number of CellBender-
+ determined non-empty droplets
+ 12. ``output_average_counts_per_cell``: 2 divided by 11
+ 13. ``ratio_of_found_cells_to_expected_cells``: 11 divided by 10
+ 14. ``found_empties``: The number of empty droplets, as determined by
+ CellBender. This number plus 11 equals the input
+ ``--total-droplets-included`` (or the value used by default)
+ 15. ``fraction_of_analyzed_droplets_that_are_nonempty``: 11 divided by the
+ input ``--total-droplets-included``
+ 16. ``convergence_indicator``: The mean absolute difference between successive
+ values of the train ELBO for the last 3 epochs, divided by the standard
+ deviation of the train ELBO over the last 20 epochs. A smaller number
+ indicates better convergence. It's typical to see values of 0.25 or 0.35.
+ Large values might indicate a failure to converge. (Not many people have
+ yet used this in practice, so we might learn more about recommendations
+ in future.)
+ 17. ``overall_change_in_train_elbo``: The change in ELBO from first to last
+ epoch.
+
+ * The most useful values for automated decision-making are probably 13, 15,
+ and 16.
+
+.. _a15:
+
+* My HTML report summary (at the bottom) said there were some warnings. What
+ should I do about that?
+
+ * If the warning comes with a recommendation to re-run with different settings,
+ then that is worth a try.
+ * Some warnings do not need further action, and simply reflect peculiarities
+ of the specific dataset.
+
+.. _a16:
+
+* The tool failed to produce an HTML report
+
+ * Please report the error as a github issue. The report-making process, since
+ it makes use of Jupyter notebooks, is a bit of a fragile process. These
+ reports are new in v0.3.0, and there has been less testing.
+
+.. _a17:
+
+* I ran the WDL in Terra and the job ``Failed`` with PAPI error code 9
+
+ * Typically this is a so-called "transient" error, meaning that it was a random
+ occurrance, and the job may succeed if run again without any changes.
+ However, it is worth checking the log and checking "Job Manager" to see if
+ there was a more specific error message.
+
+.. _a18:
+
+* How much does it cost, per sample, to run the WDL in Terra?
+
+ * It depends on the size of the dataset, but $0.30 is pretty typical, as of
+ the pricing used by Google Cloud in July 2022.
+
+.. _a19:
+
+* I am getting a GPU-out-of-memory error (process ``Killed``)
+
+ * Please report the issue on github, but there are a few things you can try
+ to reduce memory usage. Typically memory usage is highest during posterior
+ sampling. Try setting ``--posterior-batch-size`` to 64, instead of its
+ default value of 128. (Make sure to restart from the checkpoint file to
+ avoid re-running inference. This will happen automatically if you re-run
+ in the same folder as the ``ckpt.tar.gz`` file.)
+ * If you can, try running on an Nvidia Tesla T4 GPU, which has more RAM than
+ some other options.
+ * Currently, CellBender only makes use of 1 GPU, so extra GPUs will not help.
+
+.. _a20:
+
+* I got a ``nan`` error and the tool crashed
+
+ * This is real bad. Definitely report this issue on github. You may be asked
+ to re-run the tool using the ``--debug`` flag, to get more error messages
+ for reporting.
+
+.. _a21:
+
+* There was an error!
+
+ * Please report the issue on github. You may be asked
+ to re-run the tool using the ``--debug`` flag, to get more error messages
+ for reporting.
+
+ * The following warning is emitted in the log file: "Warning: few empty droplets identified.
+ Low UMI cutoff may be too high. Check the UMI decay curve, and decrease the
+ ``--low-count-threshold`` parameter if necessary."
+
+ * This warning indicates that no "surely empty" droplets were identified in the analysis.
+ This means that the "empty droplet plateau" could not be identified. The most likely
+ explanation is that the level of background RNA is extremely low, and that the value
+ of ``--low-count-threshold`` exceeds this level. This would result in the empty
+ droplet plateau being excluded from the analysis, which is not advisable. This can be
+ corrected by decreasing ``--low-count-threshold`` from its default of 15 to a value like 5.
+
diff --git a/docs/source/tutorial/index.rst b/docs/source/tutorial/index.rst
new file mode 100644
index 00000000..c26316e1
--- /dev/null
+++ b/docs/source/tutorial/index.rst
@@ -0,0 +1,224 @@
+.. _quick start tutorial:
+
+Quick-start tutorial
+====================
+
+This section contains quick start tutorials for different CellBender modules.
+
+.. _remove background tutorial:
+
+remove-background
+-----------------
+
+In this tutorial, we will run ``remove-background`` on a small dataset derived from the 10x Genomics
+``heart10k`` snRNA-seq `dataset
+<https://www.10xgenomics.com/resources/datasets/10-k-heart-cells-from-an-e-18-mouse-v-3-chemistry-3-standard-3-0-0>`_
+(v3 Chemistry, CellRanger 3.0.2).
+
+As a first step, we download the full dataset and generate a smaller `trimmed` copy by selecting 500 barcodes
+with high UMI count (likely non-empty) and an additional 50,000 barcodes with small UMI count (likely empty).
+We also trim to keep only the top 100 most highly-expressed genes. Note
+that the trimming step is performed in order to allow us go through this tutorial in a minute on a
+typical CPU. Processing the full untrimmed dataset requires a CUDA-enabled GPU (e.g. NVIDIA Testla T4)
+and takes about 30 minutes to finish.
+
+(Please note that trimming is NOT part of the recommended workflow, and is only for
+the purposes of a quick demo run. When you run ``remove-background`` on real data,
+do not do any preprocessing or feature selection or barcode selection first.)
+
+We have created a python script to download and trim the dataset. Navigate to ``examples/remove_background/``
+under your CellBender installation root directory and run the following command in the console:
+
+.. code-block:: console
+
+ $ python generate_tiny_10x_dataset.py
+
+After successful completion of the script, you should have a new file named
+``tiny_raw_feature_bc_matrix.h5ad``.
+
+Run remove-background
+~~~~~~~~~~~~~~~~~~~~~
+
+We proceed to run ``remove-background`` on the trimmed dataset using the following command:
+
+.. code-block:: console
+
+ (cellbender) $ cellbender remove-background \
+ --input tiny_raw_feature_bc_matrix.h5ad \
+ --output tiny_output.h5 \
+ --expected-cells 500 \
+ --total-droplets-included 2000
+
+Again, here we leave out the ``--cuda`` flag solely for the purposes of being able to run this
+command on a CPU. But a GPU is highly recommended for real datasets.
+
+The computation will finish within a minute or two (after 150 epochs). The tool outputs the following files:
+
+* ``tiny_output.h5``: An HDF5 file containing a detailed output of the inference procedure, including the
+ normalized abundance of ambient transcripts, contamination fraction of each droplet, a low-dimensional
+ embedding of the background-corrected gene expression, and the background-corrected counts matrix (in CSC sparse
+ format). Please refer to the full documentation for a detailed description of these and other fields.
+
+* ``tiny_output_filtered.h5``: Same as above, though, only including droplets with a posterior cell probability
+ exceeding 0.5.
+
+* ``tiny_output_cell_barcodes.csv``: The list of barcodes with a posterior cell probability exceeding 0.5.
+
+* ``tiny_output.pdf``: A PDF summary of the results showing (1) the evolution of the loss function during training,
+ (2) a ranked-ordered total UMI plot along with posterior cell probabilities, and (3) a two-dimensional PCA
+ scatter plot of the latent embedding of the expressions in cell-containing droplets. Notice the rapid drop in
+ the cell probability after UMI rank ~ 500.
+
+* ``tiny_output_umi_counts.pdf``: A PDF showing the UMI counts per droplet as a histogram, annotated
+ with what CellBender thinks is empty versus cells. This describes CellBender's prior. This is mainly
+ a diagnostic if something seems to be going wrong with CellBender's automatic prior determination.
+
+* ``tiny_output.log``: Log file.
+
+* ``tiny_output_metrics.csv``: Output metrics, most of which have very descriptive names. This file is not
+ used by most users, but the idea is that it can be incorporated into automated pipelines which could re-run
+ CellBender automatically (with different parameters) if something goes wrong.
+
+* ``tiny_output_report.html``: An HTML report which points out a few things about the run and
+ highlights differences between the output and the input. Issues warnings if there are any
+ aspects of the run that look anomalous, and makes suggestions.
+
+Finally, try running the tool with ``--expected-cells 100`` and ``--expected-cells 1000``. You should find that
+the output remains virtually the same.
+
+.. _downstream-example:
+
+Use output count matrix in downstream analyses
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The count matrix that ``remove-background`` generates can be easily loaded and used for downstream analyses in
+`scanpy <https://scanpy.readthedocs.io/>`_ and `Seurat <https://satijalab.org/seurat/>`_.
+
+You can load the filtered count matrix (containing only cells) into scanpy:
+
+.. code-block:: python
+
+ # import scanpy
+ import scanpy as sc
+
+ # load the data
+ adata = sc.read_10x_h5('tiny_output_filtered.h5')
+
+This will yield ``adata`` like this
+
+.. code-block:: console
+
+ AnnData object with n_obs × n_vars = 531 × 100
+ var: 'gene_ids', 'feature_types', 'genome'
+
+The CellBender output counts are in ``adata.X``.
+
+However, this does not include any of the CellBender metadata when loading
+the file. To include the metadata, but still load an ``AnnData`` object
+that ``scanpy`` can operate on, try some of the functions from
+``cellbender.remove_background.downstream`` (see :ref:`here <loading-outputs>`)
+
+.. code-block:: python
+
+ # import function
+ from cellbender.remove_background.downstream import anndata_from_h5
+
+ # load the data
+ adata = anndata_from_h5('tiny_output.h5')
+
+This yields an ``adata`` with all the cell barcodes which were analyzed by
+CellBender (all the ``--total-droplets-included``), along with all the
+metadata and latent variables inferred by CellBender:
+
+.. code-block:: console
+
+ AnnData object with n_obs × n_vars = 1000 × 100
+ obs: 'background_fraction', 'cell_probability', 'cell_size', 'droplet_efficiency'
+ var: 'ambient_expression', 'features_analyzed_inds', 'feature_type', 'genome', 'gene_id'
+ uns: 'cell_size_lognormal_std', 'empty_droplet_size_lognormal_loc', 'empty_droplet_size_lognormal_scale', 'posterior_regularization_lambda', 'swapping_fraction_dist_params', 'target_false_positive_rate', 'fraction_data_used_for_testing', 'test_elbo', 'test_epoch', 'train_elbo', 'train_epoch'
+ obsm: 'gene_expression_encoding'
+
+(If you want to load both the
+raw data and the CellBender data into one AnnData object, which is very useful,
+try the ``load_anndata_from_input_and_output()`` function in
+``cellbender.remove_background.downstream``, see :ref:`here <loading-outputs>`)
+
+You can access the latent gene expression embedding learned by CellBender in
+``adata.obsm['gene_expression_encoding']``, the inferred ambient RNA profile
+is in ``adata.var['ambient_expression']``, and the inferred cell probabilties are
+in ``adata.obs['cell_probability']``.
+
+You can limit this ``adata`` to CellBender cell calls very easily:
+
+.. code-block:: python
+
+ adata[adata.obs['cell_probability'] > 0.5]
+
+.. code-block:: console
+
+ View of AnnData object with n_obs × n_vars = 531 × 100
+ obs: 'background_fraction', 'cell_probability', 'cell_size', 'droplet_efficiency'
+ var: 'ambient_expression', 'features_analyzed_inds', 'feature_type', 'genome', 'gene_id'
+ uns: 'cell_size_lognormal_std', 'empty_droplet_size_lognormal_loc', 'empty_droplet_size_lognormal_scale', 'posterior_regularization_lambda', 'swapping_fraction_dist_params', 'target_false_positive_rate', 'fraction_data_used_for_testing', 'test_elbo', 'test_epoch', 'train_elbo', 'train_epoch'
+ obsm: 'gene_expression_encoding'
+
+How to use the latent gene expression downstream
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+After loading data using the ``anndata_from_h5()`` function as shown above,
+we can compute nearest neighbors
+in scanpy, using the CellBender latent representation of cells, and make a UMAP and do clustering:
+
+.. code-block:: python
+
+ # compute a UMAP and do clustering using the cellbender latent gene expression embedding
+ sc.pp.neighbors(adata, use_rep='gene_expression_encoding', metric='euclidean', method='umap')
+ sc.pp.umap(adata)
+ sc.pp.leiden(adata)
+
+.. _open-in-seurat:
+
+Seurat
+~~~~~~
+
+Seurat 4.0.2 uses a dataloader ``Read10X_h5()`` which is not currently compatible with
+the CellBender output file format. Hopefully Seurat will update its dataloader to
+ignore extra information in the future, but in the interim, we can use a `super
+handy utility from PyTables
+<https://www.pytables.org/usersguide/utilities.html#ptrepack>`_ to strip the
+extra CellBender information out of the output file so that Seurat can load it.
+
+From a python environment in which PyTables is installed, do the following at
+the command line:
+
+.. code-block:: console
+
+ $ ptrepack --complevel 5 tiny_output_filtered.h5:/matrix tiny_output_filtered_seurat.h5:/matrix
+
+(The flag ``--complevel 5`` ensures that the file size does not increase.)
+
+The file ``tiny_output_filtered_seurat.h5`` is now formatted *exactly* like
+a CellRanger v3 h5 file, so Seurat can load it:
+
+.. code-block:: Rd
+
+ # load data from the filtered h5 file
+ data.file <- 'tiny_output_filtered_seurat.h5'
+ data.data <- Read10X_h5(filename = data.file, use.names = TRUE)
+
+ # create Seurat object
+ obj <- CreateSeuratObject(counts = data.data)
+ obj
+
+.. code-block:: console
+
+ An object of class Seurat
+ 100 features across 531 samples within 1 assay
+ Active assay: RNA (100 features, 0 variable features)
+
+Of course, this will not load any metadata from CellBender, so if that is desired,
+it would have to be accessed and added to the object another way.
+
+Another option for loading data into Seurat would be third party packages like
+`scCustomize from Samuel Marsh
+<https://github.com/broadinstitute/CellBender/issues/145#issuecomment-1217360305>`_.
diff --git a/docs/source/usage/index.rst b/docs/source/usage/index.rst
index 295ad11c..bb0d5e95 100644
--- a/docs/source/usage/index.rst
+++ b/docs/source/usage/index.rst
@@ -9,8 +9,8 @@ remove-background
Use case
~~~~~~~~
-``remove-background`` is used to remove ambient / background RNA from a count matrix produced by
-`10x Genomics' CellRanger pipeline
+``remove-background`` is used to remove ambient / background RNA from a count matrix,
+such as one produced by the `10x Genomics' CellRanger pipeline
<https://support.10xgenomics.com/single-cell-gene-expression/software/pipelines/latest/what-is-cell-ranger>`_.
The output of ``cellranger count`` produces a raw .h5 file that is used as the input
for ``remove-background``.
@@ -21,6 +21,8 @@ analysis using Seurat, scanpy, your own custom analysis, etc.
The output of ``remove-background`` includes a new .h5 count matrix, with background RNA removed,
that can directly be used in downstream analysis in Seurat or scanpy as if it were the raw dataset.
+.. _proposed-pipeline:
+
Proposed pipeline
~~~~~~~~~~~~~~~~~
@@ -28,6 +30,9 @@ Proposed pipeline
#. Run ``cellbender remove-background``
#. Perform per-cell quality control checks, and filter out dead / dying cells,
as appropriate for your experiment
+#. Perform all subsequent analyses using the CellBender count matrix. (It is useful
+ to also load the raw data: keep it as a layer in an ``anndata`` object, for
+ example, see :ref:`here <loading-outputs>`)
A few caveats and hints:
@@ -61,7 +66,7 @@ Run ``remove-background`` on the dataset using the following command
.. code-block:: console
- (CellBender) $ cellbender remove-background \
+ (cellbender) $ cellbender remove-background \
--input raw_feature_bc_matrix.h5 \
--output output.h5 \
--cuda \
@@ -85,15 +90,22 @@ This command will produce five output files:
output for convenient use in certain downstream applications.
* ``output.pdf``: PDF file that provides a standard graphical summary of the inference procedure.
* ``output.log``: Log file produced by the ``cellbender remove-background`` run.
+* ``output_metrics.csv``: Metrics describing the run, potentially to be used to flag
+ problematic runs when using CellBender as part of a large-scale automated pipeline.
+* ``output_report.html``: HTML report including plots and commentary, along with any
+ warnings or suggestions for improved parameter settings.
+* ``ckpt.tar.gz``: Checkpoint file which contains the trained model and the full posterior.
Quality control checks
~~~~~~~~~~~~~~~~~~~~~~
* Check the log file for any warnings.
-* Check lines 8 - 11 in the log file. Ensure that the automatically-determined priors
+* Check lines 11 -18 in the log file. Ensure that the automatically-determined priors
for cell counts and empty droplet counts match your expectation from the UMI curve.
Ensure that the numbers of "probable cells", "additional barcodes", and "empty droplets"
are all nonzero and look reasonable.
+* Look at the HTML report and note any warnings it gives. The report will give advice
+ for re-running the tool if appropriate.
* Examine the PDF output.
* Look at the upper plot to check whether
@@ -108,7 +120,7 @@ Quality control checks
* Check the middle plot to see which droplets have been called as cells. A converged
inference procedure should result in the vast majority of cell probabilities
being very close to either zero or one. If the cell calls look problematic, check
- the :ref:`help documentation <remove background reference troubleshooting>`.
+ the :ref:`help documentation <remove background troubleshooting>`.
Keep in mind that
``remove-background`` will output a high cell probability for any droplet that is
unlikely to be drawn from the ambient background. This can result in a large number
@@ -125,16 +137,19 @@ Quality control checks
difficulties in calling which droplets contain cells.)
* Create some validation plots of various analyses with and without
- ``cellbender remove-background``. One convenient way to do this is in ``scanpy``
- and storing the raw count matrix and the background-removed count matrix as
- separate `"layers" <https://anndata.readthedocs.io/en/latest/anndata.AnnData.layers.html>`_.
+ ``cellbender remove-background``. One convenient way to do this is in ``scanpy``,
+ storing the raw count matrix and the background-removed count matrix as
+ separate `"layers" <https://anndata.readthedocs.io/en/latest/generated/anndata.AnnData.layers.html>`_.
- * UMAPs with and without (on the same set of cell barcodes)
- * Marker gene dotplots and violin plots (you should see less background)
+ * UMAPs with and without CellBender (on the same set of cell barcodes)
+ * Marker gene dotplots and violin plots before and after CellBender
+ (you should see less background noise)
* Directly subtract the output count matrix from the input count matrix and take a close
look at what was removed.
+.. _best-practices:
+
Recommended best practices
~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -142,6 +157,11 @@ The default settings are good for getting started with a clean and simple datase
the publicly available `PBMC dataset from 10x Genomics
<https://support.10xgenomics.com/single-cell-gene-expression/datasets/2.1.0/pbmc8k>`_.
+As of v0.3.0, users will typically not need to set values for ``--expected-cells``
+or ``--total-droplets-included``, as CellBender will choose reasonable values
+based on your dataset. If something goes wrong with these defaults, then you can
+try to input these arguments manually.
+
Considerations for setting parameters:
* ``--epochs``: 150 is typically a good choice. Look for a reasonably-converged ELBO value
@@ -163,8 +183,13 @@ Considerations for setting parameters:
* ``--cuda``: Include this flag. The code is meant to be run on a GPU.
* ``--learning-rate``: The default value of 1e-4 is typically fine, but this value can be
adjusted if problems arise during quality-control checks of the learning curve (as above).
-* ``--fpr``: A value of 0.01 is generally quite good, but you can generate a few output
- count matrices and compare them by choosing a few values: 0.01 0.05 0.1
+* ``--fpr``: A value of 0.01 is the default, and represents a fairly conservative
+ setting, which is appropriate for most analyses.
+ In order to examine a single dataset at a time and remove more noise (at the
+ expense of some signal), choose larger values such as 0.05 or 0.1. Bear in mind
+ that the value 1 represents removal of (nearly) every count in the dataset, signal and
+ noise. You can generate multiple output count matrices in the same run by
+ choosing several values: 0.0 0.01 0.05 0.1
.. image:: /_static/remove_background/UMI_curve_defs.png
:width: 250 px
diff --git a/examples/remove_background/.gitignore b/examples/remove_background/.gitignore
index f631f062..ad6bf5ad 100644
--- a/examples/remove_background/.gitignore
+++ b/examples/remove_background/.gitignore
@@ -1 +1,7 @@
-tiny_raw_gene_bc_matrices
+*.h5
+*.h5ad
+ckpt.tar.gz
+*.log
+*.csv
+*.pdf
+*.html
\ No newline at end of file
diff --git a/examples/remove_background/generate_tiny_10x_dataset.py b/examples/remove_background/generate_tiny_10x_dataset.py
new file mode 100755
index 00000000..fdbc9554
--- /dev/null
+++ b/examples/remove_background/generate_tiny_10x_dataset.py
@@ -0,0 +1,79 @@
+#!/usr/bin/env python
+
+import sys
+import os
+import numpy as np
+from cellbender.remove_background.downstream import load_anndata_from_input
+
+dataset_name = "heart10k (CellRanger 3.0.0, v3 Chemistry)"
+dataset_url = "https://cf.10xgenomics.com/samples/cell-exp/3.0.0/heart_10k_v3/heart_10k_v3_raw_feature_bc_matrix.h5"
+dataset_local_filename = "heart10k_raw_feature_bc_matrix.h5"
+expected_cells = 7000
+start_of_empties = 10000
+num_cell_barcodes_to_keep = 500
+num_empty_barcodes_to_keep = 50_000
+low_count_threshold = 30
+num_genes_to_keep = 100
+random_seed = 1984
+rng = np.random.RandomState(random_seed)
+
+
+if not os.path.exists(dataset_local_filename):
+
+ print(f"Downloading {dataset_name}...")
+ try:
+ os.system(f'wget -O {dataset_local_filename} {dataset_url}')
+
+ except Exception:
+ print(f"10x Genomics website is preventing an automatic download. Please visit "
+ f"https://www.10xgenomics.com/resources/datasets/10-k-heart-cells-from-an-e-18-mouse-v-3-chemistry-3-standard-3-0-0 "
+ f"in a browser and manually download 'Gene / cell matrix HDF5 (raw)' into "
+ f"this folder and re-run this script.")
+ sys.exit()
+
+print("Loading the dataset...")
+adata = load_anndata_from_input(dataset_local_filename)
+num_raw_genes = adata.shape[1]
+print(f"Raw dataset size {adata.shape}")
+
+print(f"Trimming {dataset_name}...")
+
+# select 'num_genes_to_keep' highly expressed genes
+total_gene_expression = np.array(adata.X.sum(axis=0)).squeeze()
+genes_to_keep_indices = np.argsort(total_gene_expression)[::-1][:num_genes_to_keep]
+
+# slice dataset on genes
+adata = adata[:, genes_to_keep_indices].copy()
+
+# find cells and empties
+umi_per_barcode = np.array(adata.X.sum(axis=1)).squeeze()
+umi_sorted_barcode_indices = np.argsort(umi_per_barcode)[::-1]
+cell_indices = umi_sorted_barcode_indices[:expected_cells]
+last_barcode = (umi_per_barcode > low_count_threshold).sum()
+empty_indices = umi_sorted_barcode_indices[start_of_empties:last_barcode]
+
+# putative list of barcodes to keep
+cell_barcodes_to_keep_indices = np.asarray(cell_indices)[
+ rng.permutation(len(cell_indices))[:num_cell_barcodes_to_keep]].tolist()
+empty_barcodes_to_keep_indices = np.asarray(empty_indices)[
+ rng.permutation(len(empty_indices))[:num_empty_barcodes_to_keep]].tolist()
+barcodes_to_keep_indices = cell_barcodes_to_keep_indices + empty_barcodes_to_keep_indices
+
+# slice dataset on barcodes
+adata = adata[barcodes_to_keep_indices].copy()
+
+# compensate for lost counts (due to the reduced number of genes)
+adata.X = adata.X * int((num_raw_genes / num_genes_to_keep) ** 0.25)
+
+print(f"Number of genes in the trimmed dataset: {len(genes_to_keep_indices)}")
+print(f"Number of barcodes in the trimmed dataset: {len(barcodes_to_keep_indices)}")
+print(f"Expected number of cells in the trimmed dataset: {num_cell_barcodes_to_keep}")
+
+print(adata)
+
+# save
+output_file = 'tiny_raw_feature_bc_matrix.h5ad'
+print(f"Saving the trimmed dataset as {output_file} ...")
+adata.write(output_file)
+
+print("Done!")
diff --git a/examples/remove_background/generate_tiny_10x_pbmc.py b/examples/remove_background/generate_tiny_10x_pbmc.py
deleted file mode 100755
index fff6fdb2..00000000
--- a/examples/remove_background/generate_tiny_10x_pbmc.py
+++ /dev/null
@@ -1,103 +0,0 @@
-#!/usr/bin/env python
-
-import urllib.request
-import sys
-import tarfile
-import os
-import numpy as np
-from scipy.io import mmread, mmwrite
-import pandas as pd
-import operator
-import shutil
-
-dataset_name = "pbmc4k (CellRanger 2.1.0, v2 Chemistry)"
-dataset_url = "http://cf.10xgenomics.com/samples/cell-exp/2.1.0/pbmc4k/pbmc4k_raw_gene_bc_matrices.tar.gz"
-dataset_local_filename = "pbmc4k_raw_gene_bc_matrices.tar.gz"
-expected_cells = 4000
-num_cell_barcodes_to_keep = 500
-num_empty_barcodes_to_keep = 50_000
-num_genes_to_keep = 100
-random_seed = 1984
-rng = np.random.RandomState(random_seed)
-
-print(f"Downloading {dataset_name}...")
-try:
- urllib.request.urlretrieve(dataset_url, dataset_local_filename)
-except IOError:
- print(f"Could not retrieve {dataset_name} -- cannot continue!")
- sys.exit()
-
-print(f"Extracting {dataset_name}...")
-tar = tarfile.open(dataset_local_filename, "r:gz")
-tar.extractall()
-tar.close()
-
-extracted_dataset_local_path = os.path.join(os.curdir, "raw_gene_bc_matrices")
-matrix_mtx_path = os.path.join(extracted_dataset_local_path, "GRCh38", "matrix.mtx")
-genes_tsv_path = os.path.join(extracted_dataset_local_path, "GRCh38", "genes.tsv")
-barcodes_tsv_path = os.path.join(extracted_dataset_local_path, "GRCh38", "barcodes.tsv")
-
-print("Loading the gene expression matrix...")
-matrix_mtx = mmread(matrix_mtx_path)
-num_raw_genes = matrix_mtx.shape[0]
-num_raw_barcodes = matrix_mtx.shape[1]
-
-print("Loading genes and barcodes...")
-genes_df = pd.read_csv(genes_tsv_path, delimiter="\t", header=None)
-barcodes_df = pd.read_csv(barcodes_tsv_path, delimiter="\t", header=None)
-
-print(f"Trimming {dataset_name}...")
-# (naive) indices of expected cells
-umi_per_barcode = np.asarray(np.sum(matrix_mtx, 0)).flatten()
-total_gene_expression = np.asarray(np.sum(matrix_mtx, 1)).flatten()
-umi_sorted_barcode_indices = list(
- map(operator.itemgetter(0),
- sorted(enumerate(umi_per_barcode), key=operator.itemgetter(1), reverse=True)))
-cell_indices = umi_sorted_barcode_indices[:expected_cells]
-empty_indices = umi_sorted_barcode_indices[expected_cells:]
-
-# (naive) filter counts to non-empty droplets
-cell_counts_csr = matrix_mtx.tocsc()[:, cell_indices].tocsr()
-
-# select 'num_genes_to_keep' highly expressed genes
-genes_to_keep_indices = sorted(range(num_raw_genes),
- key=lambda x: total_gene_expression[x], reverse=True)[:num_genes_to_keep]
-
-# putative list of barcodes to keep
-cell_barcodes_to_keep_indices = np.asarray(cell_indices)[
- rng.permutation(len(cell_indices))[:num_cell_barcodes_to_keep]].tolist()
-empty_barcodes_to_keep_indices = np.asarray(empty_indices)[
- rng.permutation(len(empty_indices))[:num_empty_barcodes_to_keep]].tolist()
-barcodes_to_keep_indices = cell_barcodes_to_keep_indices + empty_barcodes_to_keep_indices
-
-# remove barcodes with zero expression on kept genes
-trimmed_counts_matrix = matrix_mtx.tocsr()[genes_to_keep_indices, :].tocsc()[:, barcodes_to_keep_indices]
-umi_per_putatively_kept_barcodes = np.asarray(np.sum(trimmed_counts_matrix, 0)).flatten()
-barcodes_to_keep_indices = np.asarray(barcodes_to_keep_indices)[umi_per_putatively_kept_barcodes > 0]
-barcodes_to_keep_indices = barcodes_to_keep_indices.tolist()
-
-# slice the raw dataset
-tiny_matrix_mtx = matrix_mtx.tocsr()[genes_to_keep_indices, :].tocsc()[:, barcodes_to_keep_indices].tocoo()
-tiny_genes_df = genes_df.loc[genes_to_keep_indices]
-tiny_barcodes_df = barcodes_df.loc[barcodes_to_keep_indices]
-
-# compensate for lost counts (due to the reduced number of genes)
-tiny_matrix_mtx = tiny_matrix_mtx * int((num_raw_genes / num_genes_to_keep) ** 0.25)
-
-print(f"Number of genes in the trimmed dataset: {len(genes_to_keep_indices)}")
-print(f"Number of barcodes in the trimmed dataset: {len(barcodes_to_keep_indices)}")
-print(f"Expected number of cells in the trimmed dataset: {num_cell_barcodes_to_keep}")
-
-# save
-print("Saving the trimmed dataset...")
-output_path = os.path.join(os.curdir, 'tiny_raw_gene_bc_matrices', 'GRCh38')
-os.makedirs(output_path, exist_ok=True)
-mmwrite(os.path.join(output_path, "matrix.mtx"), tiny_matrix_mtx)
-tiny_genes_df.to_csv(os.path.join(output_path, "genes.tsv"), sep='\t', header=None, index=False)
-tiny_barcodes_df.to_csv(os.path.join(output_path, "barcodes.tsv"), sep='\t', header=None, index=False)
-
-print("Cleaning up...")
-shutil.rmtree(extracted_dataset_local_path, ignore_errors=True)
-os.remove(dataset_local_filename)
-
-print("Done!")
diff --git a/readthedocs.yml b/readthedocs.yml
index 91ace26f..6e320be3 100644
--- a/readthedocs.yml
+++ b/readthedocs.yml
@@ -3,6 +3,6 @@ version: 2
python:
version: 3.7
install:
- - requirements: REQUIREMENTS-RTD.txt
+ - requirements: requirements-rtd.txt
- method: pip
path: .
diff --git a/REQUIREMENTS-RTD.txt b/requirements-rtd.txt
similarity index 100%
rename from REQUIREMENTS-RTD.txt
rename to requirements-rtd.txt
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 00000000..3585b6a4
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,15 @@
+numpy
+scipy
+tables
+pandas
+pyro-ppl>=1.8.4
+torch
+matplotlib
+anndata>=0.7
+loompy
+ipython
+jupyter
+jupyter_contrib_nbextensions
+notebook<7.0.0
+nbconvert<7.0.0
+psutil
diff --git a/setup.py b/setup.py
index 9078221c..106fa57c 100644
--- a/setup.py
+++ b/setup.py
@@ -2,35 +2,58 @@
import os
import setuptools
+import codecs
+from typing import List
-def readme():
+def read(rel_path):
+ here = os.path.abspath(os.path.dirname(__file__))
+ with codecs.open(os.path.join(here, rel_path), 'r') as fp:
+ return fp.read()
+
+
+def get_version() -> str:
+ for line in read('cellbender/__init__.py').splitlines():
+ if line.startswith('__version__'):
+ delim = '"' if '"' in line else "'"
+ return line.split(delim)[1]
+ else:
+ raise RuntimeError("Unable to find version string.")
+
+
+def readme() -> str:
with open('README.rst') as f:
return f.read()
-def get_requirements_filename():
+def _readlines(filename, filebase=''):
+ with open(os.path.join(filebase, filename)) as f:
+ lines = f.readlines()
+ return lines
+
+
+def get_requirements() -> List[str]:
+ requirements = _readlines('requirements.txt')
if 'READTHEDOCS' in os.environ:
- return "REQUIREMENTS-RTD.txt"
- elif 'DOCKER' in os.environ:
- return "REQUIREMENTS-DOCKER.txt"
- else:
- return "REQUIREMENTS.txt"
+ requirements.extend(get_rtd_requirements())
+ return requirements
-install_requires = [
- line.rstrip() for line in open(os.path.join(os.path.dirname(__file__), get_requirements_filename()))
-]
+def get_rtd_requirements() -> List[str]:
+ requirements = _readlines('requirements-rtd.txt')
+ return requirements
+
setuptools.setup(
name='cellbender',
- version='0.2.2',
+ version=get_version(),
description='A software package for eliminating technical artifacts from '
'high-throughput single-cell RNA sequencing (scRNA-seq) data',
long_description=readme(),
+ long_description_content_type='text/x-rst',
classifiers=[
- 'Development Status :: 3 - Alpha',
- 'Intended Audience :: Science/Research'
+ 'Development Status :: 4 - Beta',
+ 'Intended Audience :: Science/Research',
'License :: OSI Approved :: BSD License',
'Programming Language :: Python :: 3.7',
'Topic :: Scientific/Engineering :: Bio-Informatics',
@@ -40,10 +63,15 @@ def get_requirements_filename():
author='Stephen Fleming, Mehrtash Babadi',
license='BSD (3-Clause)',
packages=setuptools.find_packages(),
- install_requires=install_requires,
+ install_requires=get_requirements(),
+ extras_require={
+ "dev": ["pytest", "scikit-learn"],
+ "docs": get_rtd_requirements(),
+ },
entry_points={
'console_scripts': ['cellbender=cellbender.base_cli:main'],
},
include_package_data=True,
+ package_data={'': ['*.ipynb']}, # include the report template
zip_safe=False
)
diff --git a/wdl/README.rst b/wdl/README.rst
index 22325678..b82743ca 100644
--- a/wdl/README.rst
+++ b/wdl/README.rst
@@ -8,7 +8,7 @@ enable CellBender commands to be run on cloud computing architecture.
``remove-background``
---------------------
-The workflow that runs ``cellbender remove-background`` on a (Tesla K80) GPU on a
+The workflow that runs ``cellbender remove-background`` on a (Tesla T4) GPU on a
Google Cloud virtual machine is located at ``wdl/cellbender_remove_background.wdl``.
Method inputs:
@@ -20,9 +20,8 @@ refer to the `documentation
Required:
-* ``input_10x_h5_file_or_mtx_directory``: Path to raw count matrix output by 10x
- ``cellranger count`` as either a .h5 file or as the directory that contains
- ``matrix.mtx(.gz)``. In the `Terra <https://app.terra.bio>`_
+* ``input_file_unfiltered``: Path to raw count matrix such as an .h5 file from
+ ``cellranger count``, or as a .h5ad or a DGE-format .csv. In the `Terra <https://app.terra.bio>`_
data model, this could be ``this.h5_file``, where "h5_file" is the column of
the data model that contains the path to the raw count matrix h5. Alternatively,
this could be a Google bucket path as a String (e.g.
@@ -45,40 +44,31 @@ Optional and recommended:
be called as either cell or empty droplet by the inference procedure. Any
droplets not included in the ``total_droplets_included`` largest UMI-count
droplets will be treated as surely empty.
+* ``output_bucket_base_directory``: Google bucket path (gsURL) to a directory where
+ you want outputs to be copied. Within that directory, a new folder will appear
+ called `sample_name`, and all outputs will go there. Note that your data will
+ then be in two locations in the Cloud: here and wherever Cromwell has its
+ execution directory (if using Terra, this is in your workspace's Google bucket).
Optional:
* ``learning_rate``: The learning rate used during inference, as a Float (e.g. ``1e-4``).
* ``epochs``: Number of epochs to train during inference, as an Int (e.g. 150).
-* ``model``: One of {"full", "ambient", "swapping", "simple"}. This specifies how
- the count data should be modeled. "full" specifies an ambient RNA plus chimera
- formation model, while "ambient" specifies a model with only ambient RNA, and
- "swapping" specifies a model with only chimera formation. "simple" should not
- be used in this context.
* ``low_count_threshold``: An Int that specifies a number of unique UMIs per droplet (e.g. 15).
Droplets with total unique UMI count below ``low_count_threshold`` will be
entirely excluded from the analysis. They are assumed not even to be empty droplets,
but some barcode error artifacts that are not useful for inference.
-* ``blacklist_genes``: A whitespace-delimited String of integers
- (e.g. "523 10021 10022 33693 33694") that specifies genes that should be completely
- excluded from analysis. Counts of these genes are set to zero in the output count matrix.
- Genes are specified by the integer that indexes them in the count matrix.
-
-Optional but discouraged:
-
-[There should not be any need to change these parameters from their default values.]
-
-* ``z_dim``: Dimension of the latent gene expression space, as an Int. Use a smaller
- value (e.g. 20) for slightly more imputation, and a larger value (e.g. 200) for
- less imputation.
-* ``z_layers``: Architecture of the neural network autoencoder for the latent representation
- of gene expression. ``z_layers`` specifies the size of each hidden layer.
- Input as a whitespace-delimited String of integers, (e.g. "1000").
- Only use one hidden layer. [Two hidden layers could be specified with "500 100" for
- example, but only one hidden layer should be used.]
-* ``empty_drop_training_fraction``: Specifies what fraction of the data in each
- minibatch should come from surely empty droplets, as a Float (e.g. 0.3).
+* ``projected_ambient_count_threshold``: The larger this number, the fewer genes
+ (features) are analyzed, and the faster the tool will run. Default is 0.1. The
+ value represents the expected number of ambient counts (summed over all cells)
+ that we would estimate based on naive assumptions. Features with fewer expected
+ counts than this threshold will be ignored, and the output for those features will
+ be identical to the input.
+
+Other input parameters are explained `in the documentation
+<https://cellbender.readthedocs.io/en/latest/help_and_reference/remove_background/index.html>`_,
+but are only useful in rare cases.
Optional runtime specifications:
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -87,8 +77,9 @@ Software:
* ``docker_image``: Name of the docker image which will be used to run the
``cellbender remove-background`` command, as a String that references an image
- either on the Google Container Registry (e.g. "us.gcr.io/broad-dsde-methods/cellbender:latest")
- or Dockerhub (e.g. "dockerhub_username/image_name"). Images should be 3GB or smaller.
+ with CellBender installed (e.g. "us.gcr.io/broad-dsde-methods/cellbender:0.3.0").
+ Note that this WDL may not be compatible with other versions of CellBender due
+ to changes in input arguments.
Hardware:
@@ -100,23 +91,40 @@ Hardware:
* ``hardware_disk_size_GB``: Specify the size of the disk attached to the VM, as
an Int in units of GB.
* ``hardware_preemptible_tries``: Specify the number of preemptible runs to attempt,
- as an Int. Preemptible runs are 1/3 to 1/4 the cost. If the run gets pre-empted
+ as an Int. Preemptible runs are 1/3 to 1/4 the cost. If the run gets preempted
``hardware_preemptible_tries`` times, a final non-preemptible run is carried out.
+ Work is not lost during preemption because the workflow uses
+ `checkpointing <https://cromwell.readthedocs.io/en/stable/optimizations/CheckpointFiles/>`_
+ to pick up (near) where it left off.
Outputs:
~~~~~~~~
-``cellbender remove-background`` outputs five files, and each of these output files is
+``cellbender remove-background`` outputs several files, and each of these files is
included as an output of the workflow.
-If multiple FPR values are specified, then separate ``h5`` and
-
-* ``h5_array``: Array of output count matrix files, with background RNA removed.
-* ``output_directory``: Same as the workflow input parameter, useful for populating
- a column of the Terra data model.
-* ``csv``: CSV file containing all the droplet barcodes which were determined to have
+If multiple FPR values are specified, then separate ``.h5`` and ``report.html``
+``metrics.csv`` files will be produced, one for each FPR.
+
+* ``h5_array``: Array of output count matrix files (one for each FPR), with
+ background RNA removed.
+* ``html_report_array``: Array of HTML output reports (one for each FPR)
+* ``metrics_csv_array``: Array of CSV files that include output metrics (one for
+ each FPR)
+* ``output_directory``: If the input `output_base_directory` was blank, this
+ will be black too. Same as the workflow input parameter, but with the
+ `sample_name` subdirectory added: useful for populating
+ a column of the Terra data model. All outputs are copied here.
+* ``cell_barcodes_csv``: CSV file containing all the droplet barcodes which were determined to have
a > 50% posterior probability of containing cells. Barcodes are written in plain text.
This information is also contained in each of the above outputs, but is included as a separate
output for convenient use in certain downstream applications.
-* ``pdf``: PDF file that provides a standard graphical summary of the inference procedure.
+* ``summary_pdf``: PDF file that provides a quick graphical summary of the run.
* ``log``: Log file produced by the ``cellbender remove-background`` run.
+
+
+Cost:
+~~~~~
+
+The cost to run a single sample in v0.3.0 on a preemptible Nvidia Tesla T4 GPU
+on Google Cloud hovers somewhere in the ballpark of
diff --git a/wdl/cellbender_remove_background.wdl b/wdl/cellbender_remove_background.wdl
index 014bd882..4efb8faf 100644
--- a/wdl/cellbender_remove_background.wdl
+++ b/wdl/cellbender_remove_background.wdl
@@ -1,4 +1,6 @@
-## Copyright Broad Institute, 2020
+version 1.0
+
+## Copyright Broad Institute, 2022
##
## LICENSING :
## This script is released under the WDL source code license (BSD-3)
@@ -6,93 +8,234 @@
task run_cellbender_remove_background_gpu {
- # File-related inputs
- String sample_name
- File input_10x_h5_file_or_mtx_directory
-
- # Outputs
- String output_directory # Google bucket path
-
- # Docker image for cellbender remove-background version
- String? docker_image = "us.gcr.io/broad-dsde-methods/cellbender:0.2.0"
-
- # Method configuration inputs
- Int? expected_cells
- Int? total_droplets_included
- String? model
- Int? low_count_threshold
- String? fpr # in quotes: floats separated by whitespace: the output false positive rate(s)
- Int? epochs
- Int? z_dim
- String? z_layers # in quotes: integers separated by whitespace
- Float? empty_drop_training_fraction
- String? blacklist_genes # in quotes: integers separated by whitespace
- Float? learning_rate
- Boolean? exclude_antibody_capture = false
-
- # Hardware-related inputs
- String? hardware_zones = "us-east1-d us-east1-c us-central1-a us-central1-c us-west1-b"
- Int? hardware_disk_size_GB = 50
- Int? hardware_boot_disk_size_GB = 20
- Int? hardware_preemptible_tries = 0
- Int? hardware_cpu_count = 4
- Int? hardware_memory_GB = 15
- String? hardware_gpu_type = "nvidia-tesla-t4"
-
- command {
- cellbender remove-background \
- --input "${input_10x_h5_file_or_mtx_directory}" \
- --output "${sample_name}_out.h5" \
- --cuda \
- ${"--expected-cells " + expected_cells} \
- ${"--total-droplets-included " + total_droplets_included} \
- ${"--fpr " + fpr} \
- ${"--model " + model} \
- ${"--low-count-threshold " + low_count_threshold} \
- ${"--epochs " + epochs} \
- ${"--z-dim " + z_dim} \
- ${"--z-layers " + z_layers} \
- ${"--empty-drop-training-fraction " + empty_drop_training_fraction} \
- ${"--blacklist-genes " + blacklist_genes} \
- ${"--learning-rate " + learning_rate} \
- ${true="--exclude-antibody-capture" false=" " exclude_antibody_capture}
-
- gsutil -m cp ${sample_name}_out* ${output_directory}/${sample_name}/
- }
-
- output {
- File log = "${sample_name}_out.log"
- File pdf = "${sample_name}_out.pdf"
- File csv = "${sample_name}_out_cell_barcodes.csv"
- Array[File] h5_array = glob("${sample_name}_out*.h5") # v2 creates a number of outputs depending on "fpr"
- String output_dir = "${output_directory}/${sample_name}"
- }
-
- runtime {
- docker: "${docker_image}"
- bootDiskSizeGb: hardware_boot_disk_size_GB
- disks: "local-disk ${hardware_disk_size_GB} HDD"
- memory: "${hardware_memory_GB}G"
- cpu: hardware_cpu_count
- zones: "${hardware_zones}"
- gpuCount: 1
- gpuType: "${hardware_gpu_type}"
- preemptible: hardware_preemptible_tries
- maxRetries: 0
- }
+ input {
+
+ # File-related inputs
+ String sample_name
+ File input_file_unfiltered # all barcodes, raw data
+ File? barcodes_file # for MTX and NPZ formats, the bacode information is in a separate file
+ File? genes_file # for MTX and NPZ formats, the gene information is in a separate file
+ File? checkpoint_file # start from a saved checkpoint
+ File? truth_file # only for developers using simulated data
+
+ # Outputs
+ String? output_bucket_base_directory # Google bucket path
+
+ # Docker image with CellBender
+ String? docker_image = "us.gcr.io/broad-dsde-methods/cellbender:0.3.0"
+
+ # Used by developers for testing non-dockerized versions of CellBender
+ String? dev_git_hash__ # leave blank to run CellBender normally
+
+ # Method configuration inputs
+ Int? expected_cells
+ Int? total_droplets_included
+ Float? force_cell_umi_prior
+ Float? force_empty_umi_prior
+ String? model
+ Int? low_count_threshold
+ String? fpr # in quotes: floats separated by whitespace: the output false positive rate(s)
+ Int? epochs
+ Int? z_dim
+ String? z_layers # in quotes: integers separated by whitespace
+ Float? empty_drop_training_fraction
+ Float? learning_rate
+ String? exclude_feature_types # in quotes: strings separated by whitespace
+ String? ignore_features # in quotes: integers separated by whitespace
+ Float? projected_ambient_count_threshold
+ Float? checkpoint_mins
+ Float? final_elbo_fail_fraction
+ Float? epoch_elbo_fail_fraction
+ Int? num_training_tries
+ Float? learning_rate_retry_mult
+ Int? posterior_batch_size
+ Boolean? estimator_multiple_cpu
+ Boolean? constant_learning_rate
+ Boolean? debug
+
+ # Hardware-related inputs
+ String? hardware_zones = "us-east1-d us-east1-c us-central1-a us-central1-c us-west1-b"
+ Int? hardware_disk_size_GB = 50
+ Int? hardware_boot_disk_size_GB = 20
+ Int? hardware_preemptible_tries = 2
+ Int? hardware_cpu_count = 4
+ Int? hardware_memory_GB = 32
+ String? hardware_gpu_type = "nvidia-tesla-t4"
+ String? nvidia_driver_version = "470.82.01" # need >=465.19.01 for CUDA 11.3
+
+ }
+
+ # For development only: install a non dockerized version of CellBender
+ Boolean install_from_git = (if defined(dev_git_hash__) then true else false)
+
+ # Compute the output bucket directory for this sample: output_bucket_base_directory/sample_name/
+ String output_bucket_directory = (if defined(output_bucket_base_directory)
+ then sub(select_first([output_bucket_base_directory]), "/+$", "") + "/${sample_name}/"
+ else "")
+
+ command {
+
+ set -e # fail the workflow if there is an error
+
+ # install a specific commit of cellbender from github if called for (-- developers only)
+ if [[ ~{install_from_git} == true ]]; then
+ echo "Uninstalling pre-installed cellbender"
+ yes | pip uninstall cellbender
+ echo "Installing cellbender from github"
+ # this more succinct version is broken in some older versions of cellbender
+ echo "pip install --no-cache-dir -U git+https://github.com/broadinstitute/CellBender.git@~{dev_git_hash__}"
+ # yes | pip install --no-cache-dir -U git+https://github.com/broadinstitute/CellBender.git@~{dev_git_hash__}
+ # this should always work
+ git clone -q https://github.com/broadinstitute/CellBender.git /cromwell_root/CellBender
+ cd /cromwell_root/CellBender
+ git checkout -q ~{dev_git_hash__}
+ yes | pip install --no-cache-dir -U -e /cromwell_root/CellBender
+ pip list
+ cd /cromwell_root
+ fi
+
+ # put the barcodes_file in the right place, if it is provided
+ if [[ ! -z "~{barcodes_file}" ]]; then
+ dir=$(dirname ~{input_file_unfiltered})
+ if [[ "~{input_file_unfiltered}" == *.npz ]]; then
+ name="row_index.npy"
+ elif [[ "~{barcodes_file}" == *.gz ]]; then
+ name="barcodes.tsv.gz"
+ else
+ name="barcodes.tsv"
+ fi
+ echo "Moving barcodes file to "$dir"/"$name
+ echo "mv ~{barcodes_file} "$dir"/"$name
+ [ -f $dir/$name ] || mv ~{barcodes_file} $dir/$name
+ fi
+
+ # put the genes_file in the right place, if it is provided
+ if [[ ! -z "~{genes_file}" ]]; then
+ dir=$(dirname ~{input_file_unfiltered})
+ if [[ "~{input_file_unfiltered}" == *.npz ]]; then
+ name="col_index.npy"
+ elif [[ "~{genes_file}" == *.gz ]]; then
+ name="features.tsv.gz"
+ else
+ name="genes.tsv"
+ fi
+ echo "Moving genes file to "$dir"/"$name
+ echo "mv ~{genes_file} "$dir"/"$name
+ [ -f $dir/$name ] || mv ~{genes_file} $dir/$name
+ fi
+
+ cellbender remove-background \
+ --input "~{input_file_unfiltered}" \
+ --output "~{sample_name}_out.h5" \
+ --cuda \
+ ~{"--checkpoint " + checkpoint_file} \
+ ~{"--expected-cells " + expected_cells} \
+ ~{"--total-droplets-included " + total_droplets_included} \
+ ~{"--fpr " + fpr} \
+ ~{"--model " + model} \
+ ~{"--low-count-threshold " + low_count_threshold} \
+ ~{"--epochs " + epochs} \
+ ~{"--force-cell-umi-prior " + force_cell_umi_prior} \
+ ~{"--force-empty-umi-prior " + force_empty_umi_prior} \
+ ~{"--z-dim " + z_dim} \
+ ~{"--z-layers " + z_layers} \
+ ~{"--empty-drop-training-fraction " + empty_drop_training_fraction} \
+ ~{"--exclude-feature-types " + exclude_feature_types} \
+ ~{"--ignore-features " + ignore_features} \
+ ~{"--projected-ambient-count-threshold " + projected_ambient_count_threshold} \
+ ~{"--learning-rate " + learning_rate} \
+ ~{"--checkpoint-mins " + checkpoint_mins} \
+ ~{"--final-elbo-fail-fraction " + final_elbo_fail_fraction} \
+ ~{"--epoch-elbo-fail-fraction " + epoch_elbo_fail_fraction} \
+ ~{"--num-training-tries " + num_training_tries} \
+ ~{"--learning-rate-retry-mult " + learning_rate_retry_mult} \
+ ~{"--posterior-batch-size " + posterior_batch_size} \
+ ~{true="--estimator-multiple-cpu " false=" " estimator_multiple_cpu} \
+ ~{true="--constant-learning-rate " false=" " constant_learning_rate} \
+ ~{true="--debug " false=" " debug} \
+ ~{"--truth " + truth_file}
+
+ # copy outputs to google bucket if output_bucket_base_directory is supplied
+ if [[ ! -z "~{output_bucket_directory}" ]]; then
+ echo "Copying output data to ~{output_bucket_directory} using gsutil cp"
+ gsutil -m cp ~{sample_name}_out* ~{output_bucket_directory}
+ fi
+
+ }
+
+ output {
+ File log = "${sample_name}_out.log"
+ File pdf = "${sample_name}_out.pdf"
+ File cell_csv = "${sample_name}_out_cell_barcodes.csv"
+ Array[File] metrics_array = glob("${sample_name}_out*_metrics.csv") # a number of outputs depending on "fpr"
+ Array[File] report_array = glob("${sample_name}_out*_report.html") # a number of outputs depending on "fpr"
+ Array[File] h5_array = glob("${sample_name}_out*.h5") # a number of outputs depending on "fpr"
+ String output_dir = "${output_bucket_directory}"
+ File ckpt_file = "ckpt.tar.gz"
+ }
+
+ runtime {
+ docker: "${docker_image}"
+ bootDiskSizeGb: hardware_boot_disk_size_GB
+ disks: "local-disk ${hardware_disk_size_GB} HDD"
+ memory: "${hardware_memory_GB}G"
+ cpu: hardware_cpu_count
+ zones: "${hardware_zones}"
+ gpuCount: 1
+ gpuType: "${hardware_gpu_type}"
+ nvidiaDriverVersion: "${nvidia_driver_version}"
+ preemptible: hardware_preemptible_tries
+ checkpointFile: "ckpt.tar.gz"
+ maxRetries: 0 # can be used in case of a PAPI error code 2 failure to install GPU drivers
+ }
+ meta {
+ author: "Stephen Fleming"
+ email: "sfleming@broadinstitute.org"
+ description: "WDL that runs CellBender remove-background on a GPU on Google Cloud hardware. See the [CellBender GitHub repo](https://github.com/broadinstitute/CellBender) and [read the documentation](https://cellbender.readthedocs.io/en/v0.3.0/reference/index.html#command-line-options) for more information."
+ }
+
+ parameter_meta {
+ sample_name :
+ {help: "Name which will be prepended to output files and which will be used to construct the output google bucket file path, if output_bucket_base_directory is supplied, as output_bucket_base_directory/sample_name/"}
+ input_file_unfiltered :
+ {help: "Input file. This must be raw data that includes all barcodes. See http://cellbender.readthedocs.io for more information on file formats, but importantly, this must be one file, and cannot be a pointer to a directory that contains MTX and TSV files."}
+ barcodes_file :
+ {help: "Supply this only if your input_file_unfiltered is a sparse NPZ matrix from Optimus lacking metadata."}
+ genes_file :
+ {help: "Supply this only if your input_file_unfiltered is a sparse NPZ matrix from Optimus lacking metadata."}
+ output_bucket_base_directory :
+ {help: "Optional google bucket gsURL. If provided, the workflow will create a subfolder called sample_name in this directory and copy outputs there. (Note that the output data would then exist in two places.) Helpful for organization."}
+ docker_image :
+ {help: "CellBender docker image. Not all CellBender docker image tags will be compatible with this WDL.",
+ suggestions: ["us.gcr.io/broad-dsde-methods/cellbender:0.3.0"]}
+ checkpoint_file :
+ {help: "Optional gsURL for a checkpoint file created using a previous run ('ckpt.tar.gz') on the same dataset (using the same CellBender docker image). This can be used to create a new output with a different --fpr without re-doing the training run."}
+ truth_file :
+ {help: "Optional file only used by CellBender developers or those trying to benchmark CellBender remove-background on simulated data. Normally, this input would not be supplied."}
+ hardware_preemptible_tries :
+ {help: "If nonzero, CellBender will be run on a preemptible instance, at a lower cost. If preempted, your run will not start from scratch, but will start from a checkpoint that is saved by CellBender and recovered by Cromwell. For example, if hardware_preemptible_tries is 2, your run will attempt twice using preemptible instances, and if the job is preempted both times before completing, it will finish on a non-preemptible machine. The cost savings is significant. The potential drawback is that preemption wastes time."}
+ checkpoint_mins :
+ {help: "Time in minutes between creation of checkpoint files. Bear in mind that Cromwell copies checkpoints to a bucket every ten minutes."}
+ hardware_gpu_type :
+ {help: "Specify the type of GPU that should be used. Ensure that the selected hardware_zones have the GPU available.",
+ suggestions: ["nvidia-tesla-t4", "nvidia-tesla-k80"]}
+ }
}
workflow cellbender_remove_background {
- call run_cellbender_remove_background_gpu
+ call run_cellbender_remove_background_gpu
- output {
- File log = run_cellbender_remove_background_gpu.log
- File pdf = run_cellbender_remove_background_gpu.pdf
- File csv = run_cellbender_remove_background_gpu.csv
- Array[File] h5_array = run_cellbender_remove_background_gpu.h5_array
- String output_directory = run_cellbender_remove_background_gpu.output_dir
- }
+ output {
+ File log = run_cellbender_remove_background_gpu.log
+ File summary_pdf = run_cellbender_remove_background_gpu.pdf
+ File cell_barcodes_csv = run_cellbender_remove_background_gpu.cell_csv
+ Array[File] metrics_csv_array = run_cellbender_remove_background_gpu.metrics_array
+ Array[File] html_report_array = run_cellbender_remove_background_gpu.report_array
+ Array[File] h5_array = run_cellbender_remove_background_gpu.h5_array
+ String output_directory = run_cellbender_remove_background_gpu.output_dir
+ File checkpoint_file = run_cellbender_remove_background_gpu.ckpt_file
+ }
}
From 02dd763fdb3b740a100cf3251cca3525ed9d134a Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjf11@case.edu>
Date: Mon, 7 Aug 2023 17:29:31 -0400
Subject: [PATCH 02/19] Add Nature Methods citation and update docs
---
README.rst | 2 +-
.../_static/remove_background/v0.2.2_hgmm.png | Bin 0 -> 203431 bytes
.../_static/remove_background/v0.3.0_hgmm.png | Bin 0 -> 187301 bytes
.../remove_background/v0.3.0_rc_hgmm.png | Bin 0 -> 212594 bytes
docs/source/changelog/index.rst | 54 ++++++++++++++++++
docs/source/citation/index.rst | 5 +-
docs/source/contributing/index.rst | 6 ++
docs/source/introduction/index.rst | 4 +-
docs/source/troubleshooting/index.rst | 8 +--
docs/source/usage/index.rst | 35 ++++++++----
10 files changed, 92 insertions(+), 22 deletions(-)
create mode 100644 docs/source/_static/remove_background/v0.2.2_hgmm.png
create mode 100644 docs/source/_static/remove_background/v0.3.0_hgmm.png
create mode 100644 docs/source/_static/remove_background/v0.3.0_rc_hgmm.png
diff --git a/README.rst b/README.rst
index ea39e2ca..d5895c76 100644
--- a/README.rst
+++ b/README.rst
@@ -149,6 +149,6 @@ Stephen J Fleming, Mark D Chaffin, Alessandro Arduini, Amer-Denis Akkad,
Eric Banks, John C Marioni, Anthony A Phillipakis, Patrick T Ellinor,
and Mehrtash Babadi. Unsupervised removal of systematic background noise from
droplet-based single-cell experiments using CellBender.
-`Nature Methods` (in press), 2023.
+`Nature Methods`, 2023. https://doi.org/10.1038/s41592-023-01943-7
See also `our preprint on bioRxiv <https://doi.org/10.1101/791699>`_.
diff --git a/docs/source/_static/remove_background/v0.2.2_hgmm.png b/docs/source/_static/remove_background/v0.2.2_hgmm.png
new file mode 100644
index 0000000000000000000000000000000000000000..62a76adfca631a9ce43daa3c5d76b25f304c3e9e
GIT binary patch
literal 203431
zcmcG$XH-+&7B(97ps%1J(gZ0Q6p<n-Ql%&cQRyX=&;^2Y5ds7VSm;$mO6a`=2+}(!
zT?m~JI?^$r2mu0wz}<Sz`;Bjm`{(|-YYfUJu=iYZ%{8BAK68Z-O?Bmq^qlk{5a{BI
z=MZfW=xi7Wbf)E>v%pW%ndbt4FX!JpH*^Jo*qQ!(oa*K$D}X@PK`$UrbUl;Trq3sE
zL(#vtD7VgE_kR4JC(0M{Yi<nGWR!6{V~KxqQAQ?G=2hMLMt+w08xBL`=V#Aam-+k(
z<|D3#hxCL`=7-dLQ~}rAxPD!rQ1OF<;=6`f*U#08Id#O*w}l%+loRh%FDFW4WwHzc
zL)1-C>s1Bp`R}-wP&HR`=HIIecj*DU|2NKF4N&^`d(fHi7Vy6p?@m3l(4xc0m_hXm
zh?=u>O+CSn(&dvizKygWL}cd<{m1qqC@~M?N`pO0S_M4N__9}6@-;?A1=4b6QTW6z
z*AL4cT?w!9`V{#UgWcL6o5qb&9)CJK>vC=+h+IG^u&Xia+oP@&8N6%Vs(S{}OBxsd
z$fJ=EPQLa+E$93YC+MLDR6n?ji%HsjU?vR&dUvc<@m0V%_qRC`gBD36cFvd`6*EvB
z;ProQ8=9B%(q$zej`{d4V$QsC(JG@VQQ;j#G@B6Gwe8OPOm=34pFX^`XFV2?LQWa$
z&>h|+WNOAtpBS4iEm`UOYP^e)=^S>8p2)JNwVx8Ty*8pjxxCHUSlqo^7)dpLgv@6n
z-7G$(!_}|Z_yk(AM^d081FLSc#Kwk<XR$xB9{tLGK?#1)Q&)R@qA2zVA;`O7;Ny?m
zE0w#P-8ASl6#%D}vX@wur6_=~+vNLp#rMSa9uP+a83T_`+qU$(Y#~Z02<2jiNv|K%
zA&#n0nwQA7V(LNZjv>+5NwAU7UeAUO4zRJ7dX^d)L+Q^DSWg<an5HMEujYCGgN&Hb
zT{?skP}qWpPhe+q<x?~2MD_0ZMWELeR<}-pKwkRM_ro%eds?4bt-jqz9Ajx>pnE>p
zx^#4EbyIetxzgm4vVxuD#^`PNbE(Y<rVFCj7M1PaAkPE^e=d?sRr;&hG5e#X(U*Q9
ztfP(11X+1pl%-Uz(1g^bl(A=K=S8#?C0XgPIVgk(z3o7~Ok!~JbrD0&5i2j}B+T7l
zRx6Pm_6N5E_0Dim`U6a#w=qog{nu=sq2=8|EY(=Y_Cv*4RaC^x$;(BPbB4+#l^@+y
zpZkG8BZaCKTJVmJdRj=`Kthc8EX((&2*i4ai9mvEtR8<7k|At$X{hwaD2uP@GZETy
zUJ<zet=VSC4{T#g*ydQz!MI~X@@AXFp_=X)IcKh<rdqaxQllk{OTTs?Ck4L^<hY-2
zI&S%{u28rBp)<&1YgBiUBPICO1$`o{GKm{E6YKi54E~dj$=&o*B4a-a&p5ulePQVS
z4caM?W)?U=Y0DRH6!Jh`%ypUMA*-uT(3=Vfx>gdRG38v<(qBDS<O1{I@?X8BTi`mp
zAme((?l)%q>qJkm>la{B2zxM2*ub*U8LL!EPgn1)p95`~dF|yKgn7nTP)Jj{K*;AA
zvH25*a+%X4{RoOQuyJ*V7HK6`Gl)-!HQ1x)o&G!LK#ytk)xz^j1h%L0|J)PT@8;XE
zn_u9+aPVzR$aR!6rC#k}Z&ajpjOg;C(g;%~?kw=ibF&nUjGm#fr^W^6mU&)o#(0h`
zAG&KSK4BG~^BsRWQR~a9yC_f4O}t)w>a{Yufpnc@_5J8NH}eoL2qc~~V~5yE<fM}@
z+JClAN=aFz$d!-g5H6>oeLn5(w$(QpndE5pP^-B^18f?DY#qCOK8G`&$eUiP(3FC$
z5!MWRuoU*}0@3TtWK*4=N_K_6A!o4tp(yT)*A)t?)M5JzE=-&Jf<hY)TCUrQioEr;
zmx;)3?-g%q(iROunUI7o5P4@QbjL`$J#C`^ohuf+Qg;X`0w&cBaPgF?SESGTwOmxE
zm@i~iVjY{cf8KBLx({V0F1qTQgIhvgeU){6*|<kyY*lmo4541Xn+z!jBm1ZlTxskT
z<j1zd>d+WL%HpMz*_=(AX&Lv%IDMj`zGCHKMt4#5OIVMLy0nE_<6VmPv6Kz3UC$#c
z|H|9*SfX4eUVhP`a7nHyh2(02%rvb(sZ3J9eDZMI;f&tBkXY8&v0crB9k-?EpJ2Pm
zW@sC-c4szrJy&>gzgCaXbQ<)I1s*4T(5kv`3ASl^^VDpG+-G;+?BTM1j;g+m2uXCI
zh1d1y<i@6gyrY~*Yofs>7q6$3qnrgS_JQBN<r85p!=9BuzACeFOH#88pHtwKp1z27
z<lC8XrG<2JS=G=mt9`kPZBgrq;Wa(s@neU22WKA7EnQRPD>o|Yd^IP+wKC`;bpOP7
zwlB2egh5zgj@&O2bN40I+jp9TwZA59`#tki4?}hjY*^2MMs^({<dS6;eDarIEJn^b
z_HF^F)CvuyVZZU64HvOriC@Z##fyg^FDDEJ%nvs!sc@4O9pz1LC9+s4l40fI6>4nq
zC{$`$EoIhmy1&kI%?n(@Vq`N%(hosYHS|{|1rZ^zFp+rgw66L;z8>RjILrQztNsFZ
z+8Z)$`(vU;o{(~zSV=W|bdVK_Bs{}1A^X$SxO&2W!;2LtUePYNmnPmAt{9oB$q3it
zNR;)fzoR=9RwXPrXzbM@X2(!&_r_B#|1~ZocIKph|HRc&^!}cPp-)fE3jhlU<um6v
zBkhtdAUiCu7%atiNe5Y+###!C5kkokJUugEem`h4S9mcpvtp3g*66Vky3+51L6(&Z
zD_#bA5SN8B1<G?qsS+iajWm7NI}+wT6=ppsg;DB5v8uj`hnIQ#1w;Gsr8TM%_2FMt
z_p9`5+?Kw?v)QZhh1?6ti;itl->Tu*u0eAbit9lVJOtYOzH@2}n$xsAv+jxXCp5=o
zy5{%k-F^{UJ!j&tw%<H>KYZS&s&mO$L!lF1l)G0E`&Gi-p&T|~m8v0b5;oEJj9srH
z&_~tim)yvh9v&|rIhb5D5z%zQ%tNR>^O%adNpEC#anyZx2&I!KohE~v?NRf9lFycY
zRkECfSS09_QgQr!n=%`JHib#A(5?B>0w@3N4cGPMjIt5*Xt$e9JVB;#mfZOR^%}hS
zDaYPLe^^eDsHy@j>A$!X;zc|rs2@Lv7A~9EH$PjH`zVpp9Ko%!wO*p?f8!;!wz0AT
zxnA&Al{n^95viVPA4_2cfmlB2^zdOgB))7>i^gQpl}$AMG?<Q{wNVUAq0O$wp}-<l
zPe??M;YXQWF|uQUs7JbUnU`VWqVlWe%!&Jk;xDt-pP>szJYNRhw=siv+U1M9W-`-q
zBuI12nGx(#!M1m*G>V5bO<_fkB8ozb^Pu*%bj6X%<5Fs958ortjux|;DqUe;*J^!u
zXbPcL{7QXkguzs0-3Bf^CuPJnx340Wc?PW%<#>jWS%n?EMR9eCw%1MhC0FHVDE`gn
zpwhu)a3aUrNERBLms;x#6Z^Xf=r(S|Zl6`l{wUMGDYI!~BI|C_BUC9hQ(4I`{5H*A
zk`eC)3!|A<jQVKgR*r4G7V=9T3wlPr)Of{1e1^|X<&Q_z0FQ1R6P%9oU(Zx#*h4nN
zUJoBkj%^vnR35h-S9s}RNfSEz$V}|IJIPq6zC`uEgKl$BTJkrd6leFacT3f&A?!VF
zBO|oQU!{pf<Gk`i*7dV|h1X)CNQ=9dd6jr1i~{+rQN{Cb)P?zC;|G&t;yI$M=?Sc>
zb=S*#KqtN?(HGod1pAP917v5VHNo*>PBd7w+^Fdt?gdF?*u2QQ6QR{1SLJ}l2I(#3
zv<vDuG*+@HTKhMSNL*p$MNX&Sx{>Ggj3xxHE>E2+mMKv}7-y+$*u=IBRSkRV$=;+t
z-|cpL{P_(Blg{jA`&ZK|I+7dSNXn;|$d3=EiC+wCQdT@1DdkSQ_3F-&mbEQ}k}D)!
z<DcwdCsQ2zYq_*@7$$PCFed*n)nhfFrtpAA%Ls;fi*kCPG15wD?mo!ML=w(onc^rH
zCftszRG*+6{K~NG2vhHFS&flh^>m>X82aDTAsF3n>tdP8(K7&7tEFYKqqMjSUBnc&
zP)8YHN``#iXwL@>N2o;0Ad#+^6=&Wb*_%Z9W>^tukP$Q_z)XPeXCVn~O-Y73xAZ2q
zFXl!cSPwlHzS~owDRgKqZhPUQV8Csq4rZZ8?RKAHy$1E*b~e{agC-RNJPpMqNvcM4
zaU!r_14?CnF_3F%!oxNw$6s=LnXi=A>ODnPw8dW*CR1;XPK#00>^FRfJE#)8Ty{#s
zlY;xJ^hG6A<oZ6UAtP6Pq*HbK+e)@eY_)%LtR_qd-(81C4sRYTNGwzpD|X%OFcF~y
zpB>WvMRL^vV65@PFzF!c3X%|M#+QhjjYB=V^QJL>X+mc^VFcMr>aNvuGeal*9o=fh
z9m1i9MG=QH5pC4<v~}9eyuGLmlv@x|FVS$>lTTj#8`p`Xee(4!`4ln)HSfgAvl-^m
z<Y#6j_X64)UENeRXIrg6e(h5*r`HMLAtw33pYa}U?RPN@>!GZvTc!)YR&~XCu5I$c
z<O?_FWQMywWOlZ#Fp}zWaS7*2AIb~apEaWxn|a9k&aYVCjUu6M9sVO))~CHS^2>3*
zhQZ=sDMMlJLs;>i{eb2?6*i>zM5(hSe;uYfEM8Y+cx%WKaOqC_CTgueB!XRHZt|an
zbygncJEbX8mg}@}8}6Xuj)K)3YVJc1dBkB;Jd&~ChV$0~(UU8d14{)<M@~O|N7pS0
zzt-V*TxoU%ZTIaY8DpzW3vQShc?#zlz<Ls^f%G^u*YI@0{Sljo_$*)ca)mDOaOzEJ
z{>p}5v;<k%nv6x5^4%2jO_YsmQjMkB>q!wC_MQja&m)aFRrgiqb};=1WS-6{DBRt3
z^HS$oH{&c!P)>t+7x&Sm7%%g2dUTvca7$^~;=6<&Um<5S?g}5gQs9VOat!3FZHlF9
zVJ@~uzDH)A>#K-=rKSJDu=AmPeenYurAQT@$As6(g@*F~xDw4Vw_3C$gtf0nB*X4S
z=ys<ce)~!i<N3@NbCwh?=fL9iWLo`dxo0jaHfW)(jy`?fZtU3x;<i#ngw-ii6WhYz
zsdI;r1Y!*v-X-F0(a)z&71R7f)VH(o2ZLcuvw3Brmfa0I2^{rR*)KYjM5iSgfXI&~
z7w}OU+v<iQ_3M!zW1Ax1^vajEY`;LH%pzv{>kw;omD4Y}jhpF<3UMgxw5)EN%;v46
zQRvsN>96V1eS3LTxC!POe8t-#^8xi+yB-TZ1!SHZ;(0k`gcW`=q|x&vtT0hQFxJEI
zCug*{Moj11rcP?XEg^Ze3CdZloc3bE!>`j$DfUXk0K*l?cn%aD@S9fGq0m?8k0O=!
zNAKOMW*Z2drbkL5F`2UqjD@z9u2n~-Z>r%hyx<QpvD!6zp3kI0<r8cIvbqnuN3hPZ
zwiDahMc=%(N1?ZlNR1L}stPko3uUV+McuY-yxl8@VZSm-wt3h2g~b388DrZf_S;qV
z3~y9cCZmr_IS_k#vXOdic4F1!&zNY({f|lp?vA~cN0W`m6^m4piViOkLMCU)p!Js*
zt(AF{H%{IntE)G}lPwVqKl>L0TB&(Q@|~MWTO>`5X~}D{e*iru*&YEnHy|}tF=7pR
ze7ZSh^ps=T>_vcpq1I#?ryoMGFJg3S{$z%3BWG^PXT4&OHeuM(wz%)8bj0mq>iVzj
z*ts?d8BPzxEFb4{m4K3d?k?utd75*W;NM)TZpJ6`MmS4K(A&@PXhLTmnYm$a+Sy<E
zxnCV9mi?>f!=@Db-{7El&x@n0In%<MJu<ys{jV#Db=NT=y1D#b6Fn4k>Q?dLmD^sQ
z3b`WvTd2^jj2q<kHvzCW#W1$m^Ex9|ez`XM_QhLM%cUVh_Gr%5^#&;zX)-ys>?m7L
z3Hws=7IaUjhu&<qMWU|J*6W(}khQVi)M#{7sDRFI143m!*4~Z7dat}-r)jFc)ZWhS
z{AZ!{8>3MXH|H~LHj=h>-4=W>3r+D9Z*Jrn(F?LTBX9p*lXVlaUgPqpi?5`M0*P2k
zYH=sQ_|?X##`Va2VZ^jbcZZ4wH~H*#r}nV8C+dCtU_tTAN1ja&-$vyNiNrR@O9E`Q
zfNB<42Kb5Ow8%LK_T_vO$4b7)-Mm*3O5(4Y%yBKodOrYxo<A264#kgjafh{wowKI}
z%CYp+wA<6#oaae$%!z9<H-xyXO`Aq(_m~-#Csb%<bHODg4i^hEP3~Ja{tkHn53u{>
zbyZR%Mmx++WuM%_5>_vudfRL6)x*lQnd_n=?$Qwz!*;hT87^(MBD<0%Zxf7&OSK9Y
zE+8jt^GILOrSV44x;5gW1jb$Y-W`*Xgep6us%3kvNx~|9H&@!kuy;|rk`E0Ev3N;R
z%*(H+`I8Xc4*t5#gV9#V`SSVBHnY#dzjrFT%|3r|<S-;0n2=~iJ)PKvNyiN0!PghD
zf4z2^wBCULyPD8a72-HczWeeD`CRh2vtd`StZQuMG*i&FTU#ot6*Y4*v@DXOAZTQu
zOA^m)mlsDj7i)9y2?{vuOfb*t_1(99RX;675D-2MO21(!64osb$+I4vr-|sJm~<-n
zwI6&O#H8)22WIUYzKq<`TFfyR#JTKz8`ubYFCtouTQgv^8PS#72}_Ywhop!vXDf}V
zv}+W>4bfY=^O?Ud-D>K$gX$`cG-@wl?XHMc6ITO^bS4G7Rfl+viSnC9cf5Lw#Z&>_
z($no_j(r-v)6~*G87qG>^?5-U0*k2D)P+?)ZFHLrC#@dnE}%r6>(`erR{eGiD&^Am
zmPz`@FTO!xhe#`8k1~mw?A1Ov#CQBnZ!=z`8uc&I>g8U)69{O9xJxVCf{xYi$vKhU
z{v*k8RRW%WSqzyl5PFa`LB`&~GAVT+?dJ{Ryp*pzN=2|#teAL*29|`ZI}Odb7RO;N
zPJzBzO<J|R7*~ZjiW*N`P}YM#q&LH#E5&R@Q?47r2PVNB8DP=&Zuo^sult|$TF1r0
z2hTh&d_NoJP`C#Rrt5WfmDBR9W?@5`&6PLUC(6LwzO3tb3zeR#$=Z0EC@6q&y_RDX
zw*PVJlR`pAl+Ep7y+KDjN`#moQU6OQ)7ncPzpAueP}(xqK++Iw4lIrhGOdXJOx{j6
zLF7(4rD!0gDciozlFs)+2<ax=2v)tfTTUsn$o~;MkI(K{Gsl^giQcD6#nq7&di)z%
zuCFl5Q8{4;Lr!fbDU{%{xo#Dem<lAyNm+E}r0O-FpRQYTV$Cx*Au}V1-s*|8x)M3<
zvU|mS>rOdOS70V|>wr3YqvUJ8>c#Y3j!KcqU&K6zz<WYs6Hf6@p!VlSqX!-BW>iZR
z8460j3XC^)C@S0~TUvUsm^QsK$c~J+KYm-2a(6W|UXB2n+lqfdmQydY+g9oXQ<Amu
zwM`p%R{JOY+wUuad_Wugb*c8Ru0i{DSKmt+b`AWffC4FengF&XbDBIG|4N9Pfg;{3
z6a`|_;=C;!JU^xUBGe0JvZ=ZX2`9ztdCfK5F_BFpCRgJ(5pK5=_+MVLCQP;+OlyCq
z8jAANtM?k+5;hun^QU@%+BLGSM8RIL72MOaDNHuS@#>H3nFKUV2iMDNGHJWQx<7lM
zoWL3*<%9u%asaZ{blo#ILBAxyJ%`)tY_C}W!Z9tfui_x7HoO8WrUHUD&yc%?R+y?^
z#o8xb{oDZzROiUOReS#l*4*PFl;=M(Lmnl~ZV9|b%7q(+p$>WY1|_==Hpd+NM(LfM
zZzNJeFbc1JR@@GKZ;?@1jApH^SBD<qEKs_dB_(14S`cL<cp6_|=XzJKU&VxPx;rw_
z+Iqip81mo(BB!{)x<}R@j}*Zy{P7d}tIBK-(dUVqbT76}C1ov`<a|{=O0Awbmnb?}
zCm`hbr+TeJ$m4xx?PE6*<<E6jRbfd-SgP-6u_ZwTnKLB1V_)|N;4oz~#TBl8Dn&L=
zQjm2#y}a=gdL}ML?)Bmwe<G=cJ~u}ep&(Q>6h>*T+E%_r#s;v+VwEDjV>%zalvJXZ
zu|t{)d#x8x6TZ|eFBQ;??ZJQ~`0lD$(!+ofB`cE&=dB-N(Q{FSPjl~VO|E5<ei%@2
zuczr&qN=LFOVt?|ivlq&RzkHC*+TH)CZ`#t=*ukIzBAc%Z`UqunR-+{dBLy8lj0bl
z5D06761T<kqNR@~G0FReOPy9~dNnH9#j6RpUBvQ+a^Zexuh<+}_;sZy$E@<!*Sp9y
zl?@}}A!+Jam}LdFt90vJA3{8b9Jhk{ZY?%!%4}gr;WV*O(ElQ5GoiLwib{5i+Md-v
z*mM=#7@jhK-7PmO9NH_4-$)vrCKq6-rZaNcZ)dLGRuvkW&(P__{+~+X+kFwYRD!2y
zaq=IvM#cYzn7rkdMWZ{8)Xgi&RDF?WSf+otYLf|(my`S^+cYhyn|b#)t)%!!Yp}sS
zrL*$2u>`s~BISgcZY69_P9)O4e(#GkKf<}XU54`!p*vfF|L@4%9?p(an(2P>HnFIG
z`#9ouLseJE4^9_(^EsGvhkSJ9-5)&U42OtP+#7H5BA%}GbJ3Evk<=AZPw{!3jHsg<
zo0sl6PHXA>J}`~BUt64EaB+;;Xe2?9`4}wz3x*Fb&T=<2E__%vM6q>;RWme+4knyw
z)jFye-;sm4KN-Q~e|w<C&@__R#<N#zs&Ws3!qoku%OLzzxm(||F!vSnF%n+f+~;WZ
zB7Es9xX?S*z$*&odaG57f;5AeinPO8bz6b*LJ$?EufFNKroS&o91Dj{Iz~{L>KwKZ
z{+8AJ{<qf;6T#wly?!;_=}p#1Y1*s|k8<BWYuITBU1_uHi?40k1u$Z=6FWTVSgAVr
zdjuis2-Fri8&snml%0^jfltv6(tACD5UUV(cZoaO$VGoa4y>*6x7fX>G2%*TP6T%P
zwk&t!vxmve5!*SgIx`#)lJU=e_dGAIlK$WYnG!h{BpF{d%_Bvqo4{WOfpqXYUudBZ
zUX{y|*_@r^dc<<Kf|8hxWZ$H%q9X*JR?6PLq%^|)q)T`$wd;8-sx*P}O+5LZp6<4W
zr#x{tpwkLf8i=5*Xx~Dzsi1%SV@%x;#U&PVcL!;6;NfaX=V#w&|DmIm*R;w@hQUPo
zqV<K~RQ7lTcM>y=ZsK(RA~sS)@{w15XuDB8BW26gqO$pCPx(|7o=fzL;VLPlt}JPN
zPX4~5T&j9zIOQn+T{o}8Jk2iXT=3Mlo9BpqFu#;ZfqN&4jGbJ5@0$k}tWayRQD=>~
z_t=kLAo7a!s@yt%u1v#9-Y58=had?Pwa?huNOm@kKM=YvH%m_~*&K#_8j)gfDZ<#x
zZP@4|3$D=@=pq`MuNwR$M6EuNYMOrS(>%_DEpiG1mhF#*JV-VDj)?8M1`*ZPMIeRm
z-h5d&*!a*-U#E4#o>Fcq@2=(`>4*kIip0(Tie@RJY+iLy9U3rnjr@fJPl4MQVqId#
z;ji|iQa7nhO@v%!s;KxhAJ{Q3M*>iM76A>76K45G@qyO4MZCW0cWqe*uOru7Ef_b`
zPT+{hs2%8!IN+|^dp>BXO>nRG!N$G-S}5s7Y-%y(xd>Bgo@TXN>XyOqE0lho_{T_u
zh_2hCK?xYbPh4oUNZuaC<a}uAA=O<}@?30dtU=4;Ioc%ZR-*syJm}Wlo3G|n-D7n5
zj{=r^i4Sf&9|R6{X};t%T$xNn|Kr?KbN4F0O4y1-UqQoyf{?^!;uwYWwlA#0R&8*X
zP#bY&YlH}^!a$QHqpRM=r6pbIup*>J_!khx>lTkIht<pqc6chcw`Nz^4E&tpO>VVF
zYh6z%*{T@FJc<K^GxtLu47o{vo<~>E`RA_#B3oy*8zsBZOM$d&sZVhii@W1P!;s|(
zPCS@aTSxVdrjjshYw6JwJ1MH1-<0muC`vU+#%!Wic?2Cd(fC(7&ar<zeUJr466zxD
z+sTEcM~-1Tz6h44vEzycj}e#k0#DTm)a?PhYca0r!TrA@HfhAU%4iFwU~6*TSf^|<
zQsj!rr8fJhqFz`U!6(ts&D(mJLJPC=2qXBsSbvba@#f0~n_ozyhlr^DBw4n?m`bhk
zs%i6m$(-u_21JK-``7-LB>L}3xc*&Eq$A#3u%0o*+NRTPyqE$*)`~B@>Jf=r?kCFJ
zRJMbu82L?43Un`cxUYFt%?x|@>^nVhtTa$QYA{YKLQT3Qbem@NY%iD{c*echv?atO
zBjOXAe;&eb4NZ$%pS!>LhorCmb)v-y{mMnFw~1dWylu7qxG>0)plf=k)Hs8r-%A;)
z+SU|VcDz}fW6V3sKdf8mxkfp18Gi#sH%IP1L^kO*Dzg}96ja<4H%=n~`FujWV&h25
zbVmG^DtGO(ihr_3Vw-or_JyFJGAUL6SQgl8gdvu=W2&krd;fS6XmiYKGlYREa0!91
zH*?=@Ju=UO21;tufF32#;_Z@X`;TTaB}`UKLP`JxLIqJd{`RATv#gmkKS{PWPncoX
z&Gh4;AduS?irC-RK0$iG=^($-sur^%k;QgSP&%)R!hf6HwgSQ}-xy4`@`-}H>8@8!
zd}A5YInJ)jA_f*NfCjw(eJUoG+4bFpgRt^(Or0r2QG~S1Bl6UgDr4XTZ-uaf(v|<c
zU+g_N({D{jlJxNd$GB$A`i{KH<5v+_FAzvB;(wOrfB3&!=p70Fr@77!A*TO(7U2KY
zXa|l9{C9i<WqtkU-wX41K<n|}i+kVxlsNwzSIt@2{(9`s^@V%?U;Fw0e?7)Q<{Y4!
z0OO$=)Fz~BD!jKyvOItbu>IYJrWhabuc?o-bP<ukMO|YZFtJDBe86uHaJxM{_xJCB
zPyc_;&zg;`fT#*h$-*#W>Rv)i^q@L;#4)ncPr(do6CM0+=I4sd!*%_enwZ1su?NmS
zH8bN2h%=Z&?lko!bH^&rUu_nc!xGG48pbOFV~WLibd;QLybMSh?_}q=KQ`MNENx!T
zq+GVwMV#8a4$OrpS@88a)F$WxI=F~Anic#&9|8}1{)HpyS6cZIl@5G`K2p#BG=C%$
zDBWT*tV>It`?-~~gtNn9E%<F{aDyHsk}djna95z|Ki5E@$66Vmo<m0W>jTG+FM+)X
z_ilp8H14+3U{+{8(H^=>!gv*6j$c3}16#V<Rc>G6S3Cv!=up@K7F{!3;wURy_P@X_
z?gIji!HN+5zb-!OH34s%Fi9%|+ys0sVVs8x;P(Tfiv9ei&vmoGM4ysX`Y*fpm<}rQ
zDp!KL^n;&(K#<(dg>SggLLxCVKxw`>LeW0EHP{FQlIg_tu}70b-U4^X3-l4>+X;VQ
z0dM7UV_7;_7Foq|fZp{EbqK>!m?TNeX@-Ff<5KVLv>7QK$va0W_JkM+g=T5{M)^{)
zC(VtRJ{^3R?ry()M|aMQ@<C+jQj5>+0wSaCE*q3qh&k4SDqKzz<x`D<;;RdY?9g2g
zj8{Cy>G!VxR^m(wAtD!J`iYC-B{b-WI&iV05#yB<l|?=a?d4*|3PVwWnAAK>vnJjM
zy6d}HP0Pc0VKMS~N;Mz#3yJKyyE0MLlh9oP=1>cYdfVsgjA(efEGI{*$tzZY<{^y(
zr0T~PHgzywnjPE92Oo`PxtrwgtLqbMj4+47P?HTRDv0;wYp)rG7!Tc5LQ;XCewX@C
z(k-t)0|@=s{X1WB^u2VZc8@&-WMTGj0z}N8J-hVy<FLO=B->jDP(>z&8rl`Fv^db_
zph5lxfy%)LKU-(K`>71x{?g`S(U3?9={k?`%0^GeO-;?`Fb7e~Y<cE}qrux)&HPvJ
zS>d+>y1T%-S}>v6s9yDB1~RvWO=Y&Ww+83V74n~kPPZF{;$9p`vpC63=lSX3;}ggA
zuk<LuAWaEF6Jp^9<%PliA2MB_G<_yy2^vxKECmy|F(|+3!b*{0LPlVkW^nR4t?s&z
zW4Se(?tbrs9bJtx^;Pw9K(LRzD%K!1R%)4A44n3ovJF%ymz!RYJe-=AbWcIA>%&_y
zoT>fCFCdYZqfw~vUIieglol=F?v81YivY>&%Wlck6v)Q3!mDKMZF+^RCL%9Z{%R!E
zW*cTa{;_iA-9xd4uh&^vOzwcq*rOG~x-69zq2RLelPb5Fu`Z3q^i>b=ws@Me0&l#v
zjp&YnFxW4xEQR|>EPUl{bg(f$I79OeOSpf_7fYH<m61S=Hr^(2JV#o2trxt_<M;JE
z-NR;<Xib3~RVJi`=h-XK@u6L0p9|<AHqn%S=rj~|K7@BgHU1=3jrj((#F<<P(3`ul
z#?*8)yRYxJ`=NH$lXtqGJ=$jf?dINYFG=&|MM|ou=c(R{JVAae(LgR#kyYi@$BV<=
zl^kl{^h>uhbk8iz27HfFh&MRtf92QAr>Dshc|bX$Og0QS@cNzFstKh0!Mt`=FWxzk
zAiavKJpvW`X{;rSp?e~xGWRDtE1is1p^qW*f?&(z2A1*fY3+)er=6jA7cuJ1AV$!;
zJ{5bzo5l2@;2G;-J7=XY@Gkq)ryeJ^((1)pt>8wmpZRvL(VnHjvQrHHsd4vbo4roG
ze4Vq68AvgHsa9bGS&YrdO+-(}zZKE(GndwYRPE{sC@*rAQam$ydT?vRwnNk9Vq_Is
zQewLSh_9CcN~U8htSkYiV+V-8g6_=C_8IX^C+yoA7NraL@7{v%QQOK&7y}-dN-@K(
zqFQlHdiNWwA=AqNQfa*0@j%8>Oh%0|FjSL2WW;3dCK?K8%5>JB3a1bAwS=!!jO|U*
z#8}oEzu~l_f-pSO?`gnUv&Z%f!!ybW<3!B@JVSUHE+Z&rBBxOd#%^z@`z(J-a-JOY
z(s=etbebm}mi@52Umd(10zAl9LSDwl@I;9n1fq#}s7Cw=JJ*)<XiQyq=kmS&>sG8R
zFTyw#Kr$=67_YZ8iy{tnbIrK53eQuY9DJ`f2(r-cH_ypbmVWcmvQI4zkGgcyjcfX(
zr@2V_ht2zj)^AxTN>SZ-iizj-XF8>(rjv+DH|ei-$$i(ZNspd17(pUm2nlSjPusvx
zY~czKO-2oF;0i&R?<bAQkVp+SmV>g4jv3q1k{L7c23DURxEUUWy9-&Z*J!7<c}|F%
z6XWkNfh`@%2hFoML4nV`E?X&ZcV-||ZdtMIC5(3`(-(vr={7feW%LEn4(10<@7J(E
z<?iZfx~1t`J?cx_Qd}w>z^tA*t_KW4LMIMUKEpjKs<Ru-0_Q;K-r7cNFG^iPwlxj@
z<2LK<b#&F|YrFHw3L*U`Rb-iAnvk;Or%)X6)#vUg{jCb0+RXNb6>p(<VYA<NxAvU)
zWCx#)DdX0<KlkvkP`zu>CWKcrnqzzYGcqKOjCA3_T_YWJ6{8aLG#0}i+7<RRL3dXR
z?v84|*Cqs!EjmN}LlubF`{ldRDkTnD2HzhACq_hF@)K3vs`L^%DU)YOI@zh1xqw@v
zVL0o4_`df*r|dOc_EPUE9AmFW5x_v)%YT6xu)vs>DmYKSbDw?<FRo)78=BqEywwBS
zmE$>CXDV*V##lI{eME(4rZg6oKW`8U=v-NwlyvotqW5sep-W;{Z>$Zo#zj=KP+t{t
zJni9uUi|o@X{K35HZQ&Ck|cV(1Xgt0REjOq7T=P2uL|F5`kLYcMTN1a${CoCw*VQ#
z(@k%^lO@R8zDZZw<PYspwMNn@eeR;(j&<6)l7~qW<V25mu#uC6z$LqEk_!I>T!*_$
z7d%M9@T^)6p7rqZa5I!&3sBOyI2{ml_xd~;J+9SQst?>pErvc!QQ!69?AhG#&fM@h
zu@E3b*G8`WR<E{0^Ucs;i;*A5Ic$CT#F$3b;VtlktjA;SJqI*K>E_v`v`MzH(VG~V
zR)&&|`~l4KhFHzTq0?luD?-9D=`Ryh1UdcYX>s~jc<w@Xxt!%#>+TqH+6iEY1lrtj
zW+F^$T@-A?o=h=FD+8hfz<LV|r<cG!vP%&|e1uz@jk%$T7lSN1mX?_EfkhcjiqPYX
zQgHOVNk9-bW3xMP5Z6?KsXiR5%=qoBM)`0N^zG-k9S#f?8CstrYKj>_)OR&N`SUY+
zJvZ&5<uu<ky;%4P<hrf@`2D8Vf1p#+POtwEXF)k-)XpG)35H42YUj7L_BSQVA2Hzk
z?AJC0KjPz{50A6E#=N8z*ZakktUmTv#RoTNL5`n76cv^w^NGsrqA6WM*8$@nI|~gx
zxYNTRyU7Df_j@26g^`CsX4q-hk`(nj)1Y|n=hx7eLTGQa-XDnW^*?>P43rY$liOF+
zp--z-$wo!;EEm?D;GK!%U(q5vANWA4hUAWCrtJ1}WN}`V)sPvl*OLSLH&tiEO-f2H
z1>82x#}_!d`#ZWPRv(|Y^h9;mUkq7ewhIRKU~RgcXNPi7^R=9;qB{~)4!RZKQQxeJ
zJ8s$Gl}u8zXq2?XaWn1ZlojMMulG{cJyeGu!@&=NOVI<l`+KcWy5*>4<f}e_+NHy{
zoK}}B40gLUA43+;oGIz*Su-t!Qru3Cg3*mx?LPJH0<ua)WF#!|`-r0M)dx?Z9(M=o
z1ZEsf^dfM+vU^4ehF2-8MfVEq1Y~7#_MZaJQ9A=S?(6Ob59}&M{RaRf^8uEE20Z5W
z7kBzFp7q$=yFd16{=ccf?vj_2muV>2ZN@YK7*CEn5M_!Ek*HFLUFV*xzQ(Y&a`Cv<
zqt7R{fS56(gs>V;s092aFAk7h9$w2`|AyT{yU{TW)pPTzpPe1=yF1HFGh#0vEy$nA
zf70+ZA8-CiNUX^jT6pf;$XgWwO~VE^#ZKiQ9iudW-!Q?1<V!mD4(}fy%q1VFuRS)P
zzh6#;ip#F`Y2)=1$5&La^OVwNGZ8RvPom4#P>5{^>O}ojoKeH_;_S{qso%Xlk6)uw
z@T%|i-)de1uZt~5L^Sl)7kXC*|2NRN#h{2Pbk3f6CPvJaK5Vp#iJ1R+ROHU2T$hl3
zfX{Eytj(vX+R$b&*!gnw$LC_`dCN~TCg#W9N3JVpTPwnXWyO}Ld50?qN32=(I6R3X
z3faQE=jbZ;?QzXUsF2k{M+BEc`Z-l+v44;Y*G?tK__CaRlL8>?5rC{C=W39ad>P(>
zHoIdy$_v~9P8YP#r9$ypD6T5gEAyV%VB>U5^TTo>760|1(q@HB%<O~mk#niQALRvi
z%^h!QX!G<*a(gh1mjp%9)=QgtXZk{cfRhAb?;!d_MT8$nx*M=`Oa_TsF}lGl&8k-}
zSj)y#`Ds?ddPbkRg4${UvqOg_V%%cx9MAmvewwV@@4Vhr(rpEwb+q&u7s0J*+0F*B
zDQkF6JfWKIk&6QDK=Y|?t2eeMRqp28;u$y-dWlP3@*n`y*ZZsP(zlUqg(k_B0%A7%
z4(2clyj>z!)2Vgi2N;^Nb2$%_Ayg*nE^kmQ0xJ_1{BoY|Dm#YgI>e@lU)SA5U=G7q
z>j9^H`f;f3a8Ob}^u5z_{4v<)`X2t=XL4I){Z4l&#c9`R+b=Bx!>E!Z>L!$jF}v7l
z0M+4@fFj}z3d?zOdw~u1F7muwYLOORI^$NfMmZD#?V|?a;iKx3Zu*{c-(9vupPCv$
zo;5%R0|4yxxD_|qs}M80aAPafw4Nn6{&i~6M(HWg<EQhY{i9V~J^{G#L<6-HY3b?F
zQwd&r{(6ELLqWmDSj^msf*0>}M(o`8ncJ?bp>t|-6_gBoR~Tf`*5a!!l$J@c;jv@n
zVX}vkZ!jU9<6Ak|=eHh2Z4RKmGeUP?&8AH@7zN&@nm5#YDiB@x<nQ)pp_*p%iSeiu
zpkllq68AW)cg_CJ+&`TMsF^4aQEsYpaM!g)>%n%+;kZFO>zwuXCDR-8o@t}VBa>yu
z)JF#fN}^lMt!2FHnq$|pFe3HBss}U?_}<F#qz+Jh2lQ^szUElpyK=b>EPrz`4Y~`3
z?@6B=qfD<`rJksU)``@|{Re%AI12L-PHWgWorRs`N1O+wibPgXVS2hJ3@``B;L~99
zH|TK`=I}g|^f-0RR|Wh4OF13JPSL*kXHhR`AF7LvjkQx*q8PXrQu)Cr<0=74fSV~m
zM4m4Wkzb<Y{9cZF>FRg%NQ@d&{p|3wS-nRvB7=#&?t@fkCkfpaMHOl^YET9?uYwN5
zlsJb>)QHv$c0+>I-J@8!a-=2O49~UBe$ihnl05L1O1tw%GNCzSpew~pqxY|-Etrn4
zmSZ@#J2G|;PaS)|mfjiPKRrWSH8lOS!=SC1+w;m&X5GyE_;qTJ3>Q$}pPh+wT5b5T
z*TTt{4K-^49S#5qPo2Ec;Y$*^zp>+6-%2@`G@fWWSqOwCsj*{LxxqWcn`R*zeyT$D
zf`c8LckO+;`YU*K(zr?^-N;)}P;D!4j)}eF>Xm3eGAdz2unVV1N*j=ywr?0%_1d(j
zUGm$9vCtdh)^x?n+2;0RgJpB#BOB2sB#}h<`MDZV>(l{yA-85pA|s`F$LXHpsd2N$
zv#(d-o4a9g!jBqb=78Gso8G9`Qg)eQ*mhMXk9hef$z%4-P<7hD+`F+FiM)jIy0%xH
z-VBJXxEIeoY}Y1R8QQdTr8kC>4*+!ReR$wIKs}vm@<yyxK2-}AmK6(k2cNqD@JtYB
zf2@=F4igi7Xbbft^p@8{Lj&a}TFIC|y)kFX%NLWL)>-Af4ntS-Fo$wX$W(pEgpFq1
z-fpR;-{y&_4LqjfnHr9_uIxDC#5<@HKg~gw*N6E2WF2;Qgk}^2ip#Es?BY+9DD~hL
z?W#}u!oP{6@jH%3Y3<B(rP^B(+`WiGnY6q2{ZsTK94@YU)!Ch`lkKl@HAOtnN0ST_
zE|GaOwr7;Nj69x1JNWW$(}{VfqSqbZj}phrhv`G7NIn;Eoe$1hNjO)s#wPkZeeZ<f
zcn}e^LSGR9;0!^Ung&BXT0|GWG}H20Pxz@vge19%Ut;qsW;FE$kUxN-qr7I)s;Brw
zYsXLEKic>Jr5R<n7&PsDw&55Ra+4@Qy#S@<piz8HUonAB8D@(ozV(tGXF%@?L@K{N
z=jkpg7A~V~g>(PC*JC{8ob`-yx^!5)`r17J6?H-`u3e^jYd~o}-CN$)^Cu<VH!M~I
z-*_{fKk*yxG{1Pl%if_K0XM6#8efqr0)$XgxW#{ky>IDS_!P*Z0F(4XmSE3QKUQ#S
zMrYVqCqui0U`I_)`!-#&<eP?#6j{7a4Z4INTMV9&0XXm*_5$S8^z;8y6yLqi!%T*B
zp}`Np;77OS`B=;y?g(q+&4Rm_Q9dadPonB&;cWDw&!D?v;B9^;X+_8$Q}jjf_PNbm
zr&=_vshaX1V5auhB1PV10*rX<DRdTtNz=yn=<Ys&ibGLnq78VVDH#}NEqoe=D1z}y
z!yMm`H7Foj=vXV%t3fri@h96AKJl0t45kpe%c~1L_sBi59MIr|j_R0lo0Blq@fR4c
zC&KD6%x9d=fHdD>Hs+=?Q-r~fyh-t|U%IC1f0HwSGNy*Q$O7eFJ=%7>058+KrVQWH
zfV;m#ubac`(c@3xdrtJBAkgYv7~3g<g<r>-{&4O)U=^jsXMX058jwp@q6xuWg}MkF
zj$i7mqj!C<><v8?H)kek7$j8`d^?}WB{n&>LLGAjD1*QJT5)?AyDo)~saCLbC#CH+
z2o&J+gFH1)8EMbJe+oX5v7Eb?J9K3ev)&GVduI^XO@|!7`>wk8V}g@KzzV<{rND0^
zgB!kFEiw^E_sm15Tq_<ks%wBk<T-hap}X~P_x6$TC9J2b=e}gQi0CkM_Xl{}kqKG*
zM=X?=Gte3mhEAJ66iGDzz)0&L&FU9@#OlqjQ@B9yB=YU9daTr%W-Kj$A2`70oPXH8
z;o1I#;S_Y2115U=7=W|ptZOo$0HESJnZnIj!%rOH&E%jvBRJgMaCVj-%bOp_4Zp!G
zJ86)H>Q3C<$iF{R6v|RM-W)_#smM73+^I9H%VWlEXzFq^;{`8(lfFycrtl;rwODHC
zYU+LyZK&K~mQDuZY!=Y8cTPd4CBxS<Fub~Wv-Xaxnkt7?=<XY+oEfx47hexWpsprE
zp$gd;>czKS&)dk7(A_G`)*9)W{q<Cjz5e=117$)*uq9AH&#zmzva0k;;+Z-zhXzb&
z(ur@uZNAcXdRxZBh=qv<TSXYLm52x*eF%^H?;AmP9nIq(Vb8cU+>#SDl_@p)j^`@I
ztYC<5p|p3HCiAVK^L5lSpzE_*rc4{d@Tv@m5ABBn1swT57r8+_$N9mc<U4L^74<<Y
zGGc|rKNX=*6^1(Z4Q^UYrJPZ+V&`RsM+EzrLXJCeUf8O7%-ze;2jayfPo{&HuZli#
zL#!k>)2PK5PP-vgS>*a|OyUAE(Su!*EsRvPFkb)s(G0Iu^K&kZP|JcVz?2xl+tuLh
ztDA~9b+h-=(WsorukNs*dFeL1_scdiaxg&?fKNu;wQIl7`<Tq9E)7eG_kkqx4uB(l
zR)c8;{^kTpiMPx9tG{g#{;Z&PcZ4y|z0H1@_+<OdsLul#h!)ubc2i3corG!OEtm{{
zicWvtqugomXAp5WwTCwoOLa?8^C!)L69=pWL|t?EL*9Xt+fBzn5wD|Ub(`q`l8<+S
zr<L4Tquu1_|6XTs3eI71Z;B~u#d(yeAH!+Nb&?m1`-hgFG%gxq8S;`#QT8M#cuMbn
zWZEro3Ay>{`z>#!MLyP+5uwEJmDRc}7GA;~$yDfWJ5Mt+e6O)7tk*3icwt}MG7^Jz
zh>jE8by;ljG$J6agP$rbKD9V19~kPwIlZ<CIt5CYZ$Z7)Z{Cb^shDjp;eQXrps%u|
z+iLD&PZi=~z|tp!Cxg$J4r+siw?4!;?|aNMo}Go_Z|@pl3|2%L6@<aJJqz=Pba$nk
z8(Hmo8ryJ@V&%g)hbDAD+*RJ+h`^TY0oEa&>B%MR^A2*Bl2yLLW7*U!_U1G!Uuxsh
zcN@r}In((7h%@P6OqBrPQvP+2jd5{+F95*I5b$PhiqTRj<qnv=fXHVgLTCl1D<IdL
z`CPWpu4Yd?xN8^Ca^)lX9|E)W^m0PS^{!>XtGaM^#mY6`c*LMlYLL_$U{WLNRS>9a
zkGpx6ClBMz))3V!ETK0T&?W+;G7lEEjHJwXM@y(#MeLjMW?}CN^u>VqR=9g0x^aL(
z^LVw8zOto%Ge_3~4rrhZxV1+3o&sQ9ebziielabNs?xUTKC@YD-e!+i%VDpt{k-J>
zC+HhPS^CRf|0s*P<tKa%SX1n9+U4kG=IFwfGY*~{Q+-YDD$Fj)ajAvbO9SGGXqrSr
zI(xKE1GHn_G^`cuC7k+{Vn+?`GQt%RD+^cL1S6L+Jex4xi1AM2{rABQ`FVUSKpjG@
z2r~C}7N0I-XM<ks$MRDAcz++s0F1;8=!lsxdyiLDM6f$5`KulQ=Vg0vPO5#*@OV#;
zm&;~yM$j~gSnWVSLlJG>>_MJ7o*SDjfkMwd6TQLK|2+k~Jsh6H&|dD>JXQaj{>(Bj
zGwgz!{fXC;ms~bEEAB(vbiQ()1(?3o*)a^Z!}C%=a>loLD=5wBAyTOJ=f~seg2uAV
z6m4*1JA+U5JmNr4^eeyvIB;t=GoQ^GuZ<0K9|d6QOnm-GjC*=`^3hB~{R&4bd9Cu_
zWwGdH;P4Zp+85FumW#_jgDe{A2%DM`)AW<ZO$^1*;}lx^tFO+j#f`a{@f~}=lK!ZE
z6X7d40Kjgx2(hrx?prp!k60+$nNo*Ils9{x>f8yn98jf0bdamSZ>?TU;QhSI{LBix
zd*$^7BCTD8+Gm&BT!btd%5eyQcZkdggBA8pXT52r+0FAD{wWYG9KLNC#1k{F;C?`E
zfItcKKy#7jO36hDyy<1$`7|Z^Yj@QRFD=bW*W{;ongT*4XU`(PsKB9Ir+Li8yi^!Q
zUKwbV4ozCp-T0Ur=P6M>jZ=+@3HIr?k^b)Iea2;KXqTghm5F}Qcl2ljP>g1AYo<)n
z6P?x5s{zl3U&(3WSfV+OY0pZ=SnGm)KRsmxfpSmNhrJ7GJ6GQ97}~D#@HesUG}$z^
zjH|od^#PAcC%K_CniATT23b6S?>(CL3xmxAD6fN;#1LDr%qobDt{nMY_}%K~Xv<OA
z=5jQVqS_OLA(C_v)HVu$m)zhhLp#b+0P-%1l~ZvJsli-7r`Jsvi^E3pG!HgrTWDjQ
z-y73j(*g8OsprYH_fIk~QJ14zE$gYRcdAI<)Rh#v<;EM<wQ7H4zD2+o(e@^<2EXGl
zOyvG7pbeg1r>k~uU$7tAH#m<5U?;;Hv*HyNalrBPi?I`&A$TF~!GZyVhn4;#WKk9h
z@cCdx{nr{>vnLvm8N%<TowB-lj`8lcEpCzdeP_s`PDZtiMvr7z7Y}$_m$z`$2)?Hl
z34~ocdK|IUV4m;KG?V5@;sbKrosbr=c{!t3{ORxQaQAVRu=tcCF~o%wfA3c(8_D~Q
zk4kb<ANJ(^tdlu2W8)r5N7+bDuMS<wzt(0${lf3IARZi`glON=P+D~R!{&Ha7QWC7
zA&M5OHn=qg@xWSMJZpcK9)ILQWn7HQQJCiH9dQ5e)BP)+9?}`4It^i1*;*qPVA+Hv
zoH4lT41~{Xy7AgZ{57W|S37~t$*)hGjH-QmPQ2F--k>EMC$bOrrfoFFeJ43$n$Ce#
zQF0~*6#%)z0YN!OH)Ez`rB?Vo?L9_!7hp=CMVnkgFKKcjDxujb#Z>LMNh=?>{Rsfv
zM&Nf7Sj>=gu4o`du?Kh2L2NN~^HLRyQd#5adzG+x=`b|6_$rg}gUAKKh86760=ltr
zCLK^@(|H3vWX8;*UK-s{SSc#EKB4W+?VtGuBEGbUrWBrVo=_y=dof>-dppPPvqtGD
z7$Wd)?(*`CcHLZz#ad9m=gIJQLwKCzfTsdI9rn>72<U}`(T?@i-)U!hUaIOf{N#xZ
zl-js3UyHP!m-1cb2gn039<SHo%T`}ANzc~5kWM}7P5bvOz}<r^C!X{LAyZA3#x_QA
zTMY&}HTYg?-#HGQUZB%gzh85+ejkf!ug^&g-Pn%*j1;~3k^7Y&JIBI5jI~00+a8_C
zo;OkpZ$v)&=;>sNmWrgDQ3RG_;4GBZ`Px>DVd2<u*91CP2iR~3t)>r~S*25oPZ-C1
znPvByr79RzSQ)Mcnf(#Ix{UZzGXDrlCmTLr??eAI5WfPelExC<8QjGSS;Q%+#F^Za
zmod&xs`A|-4csaXUNgMDrV&D;*<V53(ULu<!4_R9!p8r9S>p>?bDqBndbj7(5`Ss)
zX6ETshNBbAJn3{Df4C$|U9^hpLT`0f?Mh1dQvC3Ycj8*ZuzJk$1rl(oA=inOjhmP@
z`|>y)^9V4t3V<M2pkC4MUM|&^Cz%9t1F9P{6MdLw2xfyHJiF(1^=xgLBV4C3*InQX
zswr5u>03kR?YO50o|jCM$={yK>ObZ3IJ-5~@W6$u*`2u)1gco^(}_FgooCVczBkz?
zAUiW)$<r8xn?k$=Dk}3gh{CS|m>b^S-Ol}cZvqmf<8S*?-?OuE54^wkie4L^qO^F;
zRMRacxWODmAKF$wW^}nPMf18n+#|0y8`B?ADPN}r8FiCLK#7K~{G<a4Jls`Y?k0s=
zoi<mas!$-Et8!IG0E?C(71vu_$he~rVhL;`E4Ziz_Qv`nU~4}qSBu~Cm~69|D!24I
zk>8pWIj-$lY=1Y}fS({JEu!TYq~URT=F&?E<JK9I%Dln8MeWCpkM<}wy+tY5t7p}P
zQ`<{)`md*wVy|=>{z6MRodKqQ#f3|@|L?Pam>38K++0lV@60`tm{cg?01On`GSaQt
zdHm6$l_eX)NQO=I?>17`0pWxH?3zJyze2wze&MqMhWHD-4L+e<;(c@hzaU+4Rw~5P
zsg3$dDCJ9P$il?s{$6{I&TUOug-EGC{~rScQX}d~x-~OJH!3Xxa9NCY{M0K7*<+4g
z@nm<KRWaPqIzCwGJIhWJ&c4rdkQvOiUhk3<%q1o-lRPo4QKE$(XdgZaqqBduwkZeQ
zCI7gdG%lpmM6JUFo*D^KWDtVVa)Xa<y4eeZ=jfJ(GYyrju;^)-=jRn}rGDf-DtC|-
zWpoh$m+0dgLsoj<&P%ai8!dnavg?8ZTkyiVoSdIA#Od0ORNUz;fXb+}xw=hz0!J~x
ziLn(8FQ$QjfnTtK{dXcM!^JI`CHv`p-XUFng+w&?3e4AM#OM*ac}jzGd-ne!>bt|5
zTDq^JUgd%a1`tG=MCnZky(>X_?;u5K(t9T$A`-et4ZTS3y@OZ)A@p9Of^>+K5J)KB
zLGSzf&L2MV<eW2;J$v@7z4n?(B_!ssD>221$Gl#tXQ1(F^6||k#r4QtCf$a~_g%tu
zE{BQ_F?Rzv((C1-&|?|9YuWx84;^a4uh#q79JM?>`p`5<JDd;M&q1_@`7SA!iBbo1
zoYzx#^?qFz{jg@m#XIKSD6k<$JN)fX)mYuYmYyh^X{4SMB&$%o=?tk=h80Hi6%B4b
z@SE40^m((U3?@!8&(0P)x(@`Qu66zATE$n+dBJ`k5bdZUMRroWUa=FMxvU;8rUQ0r
z*DDvzI&tFOnbgY?6c7+%EzJSx&^FYigAz)~6)*m#`+y1*DB@4UtXhyPy(Dr<Hq>UP
zv89{VSxbnuw68ypp;=jRIQ|yI^1XsfGycmHD5E0TR3%UE(qaRw7SX$FfF@$9%rZ$w
zgS)w)>aCkA9BBZU(lr|Bq0E*t&_lQUQocRZ<O50>^mp2d55b4`nwTFP?F@zATPK6I
zO^pWu)H9%sctEx`wHa$W+dlLaQrkS_byWBEXX;^wtfWXB;B*%KCno4z^HihUELGID
zvff!3f}W~%rsz^&b=c<Iz)eD1rlO0Lt!nV7_L0F2XzI+Oc2$j5N`S4=XSB<-xx<2c
zwLCFM2Kl)^4s0bl!&PYZOyD)|5u+v<>x)r4oO83eb92q{yCspOzybXWb6hW-^&&o#
z1i+!v0sWUT<q;bw$K0!;T2Z*`OV@bY2QC4OLmK&CzUJ<u0K)E)rh8KrZrCc0Lw5Pe
z<I!&z>4JTmq2m$Sg4O?9pFUpwNY!|Y2&h{F7wMw|Q$CYiZ{x4OfWQ3703PA;vUs<+
zNeGhtCW=_?60(&BqDPj(&fg~<^#w6Q;;7@7b-jq7ZESqJx?T!kzxXEQ*64{)=ho9#
zmMc>`K2Ef+?dX^iO+^+OeqGEr&C3hhA3Yi)@51r9f7hdNe|E*<m7tm7hf=i#BNQqN
z{n7u=y7tlHL<Z~!tOU#uK<zM3vVC?nTZW1no_YA%ndG`W&QSy^@{f=7_Kz?*B@Z#s
zUCGRsMxIuKTjPeopZ@e4v~vu&qB)F_8M0j3nu*AGVdJCj&z##@cS()79QjZ-Mf9C~
z0lv_>PYew;z$0&?Zs@E0M|r(8S@(7Z*oXW}8fNNAZ@}XbI$f+gxJ3)Ld*rmu0|}4l
zYlPJD*RvY~d|4Y%I>P*9&lz{V7+^4IJU&bh;w4v~ubjHA8t}Wb;vC)H0T&$t$Z#Pu
zLqUK%x^YQx1J_|G06DmTwMRp9QVzBf3fX@G7V9f&E}m}*=C4x`%UMqjILA-(m@Mr!
zZrvl74-5b(uoB9}kv}E^9ZEvm9@g_}Pr~Id$+?kUR+1+w`FyFfx9F7s>V5K0tqEx6
z_VV>7+05!AWsIG@yCzS(XRq}3weI*-2z_Wn6KMR#^k|$<WICllA17e{eSOe~LSrVs
zrHVxp$L;OEqbDiYKD9Hy1EJG{u4z<4|J**|`#ipsarBG-;5I?6oMth<bed2D*JZm3
zZk#T0!oFAUJ-AOdAvmS<VQuJsR=UC^oXaV_X36sk1S&42<caDW|0o`#uXSk)P>y=(
zy)+GgcQWIKgh%#eL25%U2W{<XI#6^A>$wMU4@JV`LV-K+rfek7YUk_!$`%hndQ#dj
zX%>i<3VI#>w&2a~FUY<!*r=+MkUD!KmKhje+e~(D@ldM2_xmT6cFMv!w=qgjm`5`q
z?kxyOsRw;g@>a13AbD60G|I8OM3Q0vh`Z^s+mr4R?=dW4J8)6dffJndD*`4VQzH>5
zb?zP?w*d4%g`_ONE-u1W>L3)YIc*BCC1%LQa~tyF&7S&4>9c^;%mzHG0yC3+4i<MJ
zINA1{6Lxg`%HsSEr3E~ue*zQjRf!8-EdL)f+_kV=dcy80g0qL3Fp@-)z<R$fk!)yE
z0GZ6XPv-htvOjm_7%*Udw9cS9jUmwn_%-fW?i=4i$(<7&E_@F@NOih=zR}{ei08S$
z)DTo>o#&m{25DOBZ}Q{==P8AKg4E0FVU7y%!KKWsxa7gD_D(qkYYdrlS&FB47SU`@
zVU0RO>JOrULvIWNi_>6020FaUz+I(@uvSU`fb_k@ZmWWN5RM}WCmP)_p-$rudS&C;
z0caED9)EU_U-`P^DG{C8^mqJvRs)g{s+93`sB*^}WP`|>LE(9n@27FHxuzbTiIz&U
zIsGcy6*%VM&lV`+(og8zzw!lf(1|#6k1Ya9u4431XODz(<;nANKbc;j(tg-hd8xYu
zL~0yhKi$Km;KIprIKWtH!vH0rG{CNfK9+Bk0nqqBP!O@!KOk9o4e`U<Cp#=DQ=`g2
z<VtK(haK(rms!)VADkmafx^aFC0Cd{fTJ8`0FDbSzY>#MrUO1nL!9~oa=G9%5#d@%
zCnSJUGX;Rp!=%@(?gs8Z$RroD@?bV!A3g!-%1ZJDg_b|edyb3gkHuX^El(m{FM!7~
zsoHK$mWz8_JD@J%xT(y$@qZ50gz=sKRy({Dpq9oLZS)UlUCNY$8)$1^9W3?@s>k+)
z=pBla`yLa+`DY+2)O#KT7K{IsSBXa=o^4(64O7qV9DmgVOekRHT);J9SsW0#($>ok
zGXc+ZayV68D$N#`Wz1niziT3Qsp7*vTp@jLH{Qr0xT?v{{=yLgT4!3^e7oy=(JpER
zaH&bs1Ld*dcY&JyI;*XYfmkVCw(sE|ZySk2ZO@mAb3*{k-UN_>ND06WDv(;vMhQaM
z<~IlhXP5M+wox>IgZS%&=zRhBhiTO_JQQ`&JM4Hru(+GuCVU&H3Io@pSE%jr($+Ra
z#tuf=Y(qz|a9@!dsCkTD%cZ1(62v!vL|OIvqCG`zUGS2YdrSDq+L=%|Fh;;d`Uona
z0n1{8ps`nT%QwOy?jMi^-ub}H7>NV2pqH75H{jj=*q1&z!RIX@kNyP7S66IoWWe#3
zx>yicOc|EN2?;nRByhp3!jNZg1%ZfAm<9mVFw&_dh&ycb`rk2gx!Tz{`~C0pkF9M0
z5TSQzh)3+LleZQxPS)O)GMc&~$s_xmBYyo2#rU_S(laB6pc3FV-AMq)#YOb(oa1cu
zup{NPdR`KGQb1JLtcY}`h`-o=rup|F`*C~858iCRlB(P~b}JKEVK51-No?{fV936u
z-r0LZdE{VYS7{bE1as*3<!01s+U#5BH-K~(xKy1=#dTn-h`u`53ZTjv_Yw6}!3lwl
zXqO!7!Nh1B?(#ONd=pFRu&enRWt+PY_r~z$Bx!lt!uXl(TyNEv>Qc@&<{0gFfBN!~
zs-RTDm*)5f{EIh7fFFPn4PnxyrIe6;X=q!Uqhso<)`#3(tw~zI<_LN|$S^p*_qdu!
z=1zRDaGg--JrSk9m}PG`Y^`f=)x`MR{7WZcn`7QZB0pE^DyV*};S*#w<FUAgUNASX
zqAQl6#>nNZI+3Oe-QSr*OW!7&1>md3nF3uOXaODo#5zC%s|Y@6L`Xi~Z}XU6YU>XU
zDIF57qZ8YQ|LQXw7XC4F27~QV#^-merSn76?;a|W{14{ruVwkDZ&(kBZ8^>ccfmPb
zC_$R01~;RuJVLaZIhgWGiF`vw0h*bdeEtrk_7f^-e8=?#S3bnu`0(pBAGF-|Wd}g`
z_oZ)f<QHo^;jc9x><7ahdZt`(fm<eTzROy#h&><beMNQ<Y^p~Nts@e}!ynM2hokxc
zsahWU2p}b6AWcf7DckfJ+>U%4#}4{vvElz*>kF!K|2%k<ko;CUO)3$c)wxU-;D-9s
z7~I|vMz-&~=kjkDP(nILG$rn;(4UlRzrW$q{>i~Gh?uBLCUrIp;{IHq+J@TrP6c)R
zEW|wz5oj^8Oagjj=wDd<@7ZV6`9sKa5+N7HRAHy|%J8!gBF1+nyUiW(UYLdfX92i)
z(d8k7AXTlZy~VlvkYER!l%_c~Xd6t=oE?%YYiDAf^(n992L&Iu|3r=gEG;3+vr00M
zu$}ua<D;Z`Oh7SB?_whqS(z*r?7G!4-*=MYgtn@_0=n+?>(((qDm-sv^R3%K1NEMw
z$r&YH=tD?uSu2tJN)$jmk|u`o00XqlCs`8zZ_6Qqr548vRpF^yZX{1|YW%jfdhfDp
zBiG0d_2tdKaj{3H&bsJf2Ot7$Q*cw}Qv7+L_bh6tJRq~QwF)6=3L}ex@9xT&+u9J4
zo}qt0_B9~xZxGvCVV=Ra<EtY3dQsbRt+?Dby(i@&Ui_p<i)}y7)s{&MS^(#4#2>ZF
zEU|A*VFc}l;J8_xMzr_SO>v-iW)jjf0=@)wGZs*ehG#+bOCXa&NA<BuLA_^=*InbS
zKlKKm2wY~f=g-(a&0%zfNQX|Y4x<=WL%rTUX%$JEy{_Yy%_n_|wu%FguLDEVA^@<C
z<S`3}xNbBPSI22mZw4P>_xyL`UzECzAOt>0Nm?)zyy##_kyu1K3Fc0Hg8T^=$YXrf
zooOru*e{z#^twS*-|*?8ZZZ8SBUtR@g9v%o!D*Ztx3kRQuZ+XhANleR-|Vdcq&p!v
zb18`7zo{^ic8T~)(AgeSv$YpNgu^L-H4@J=fje~c$yAXJHbrq;*Z^98pRAC7MO;Xt
zBK@Ke?G;Q{jZ-u=C%hjIwldB!Uk;0GLg#&m<<i;A21}zyISoFr#A+sXH$UGm91*Aj
zoAT`*x)+>sND$!Cnp}ph44k{LX9R>H;XM2r?CbcHxa_Xe?aO}v(Q9KOHV_EIS}(>$
zQfCitzHv#58$L}R`Dq6baJv-qskOoj!58{%)UDBFChnQvoUVezi%i1g!~NfGhO?v)
zLPUre?b)B{j8@67IAT=%oQfYpAbbpmir*-EI7li&`w>k~|G>P%82Kfhxw=caNwSHU
z5Q+^F9>A5ZU15RR4mhLWw_BVmu`h96TL^y<oSntZzVFmwu9y9zBfnh*2AuV`=F;Aq
znCDLPS!+_$c#jzD_XZ*P6@_oiVfSRFF}_jI@;`vpwcW7x*?O+tu4D_b;6FblWGBJj
zI8x{&y~`itLIiqb1TbSq*x@Z@UK*@Po6At3t4*U`M7aG>-|HDG{Nctg6SmQN#Fr+<
z$0AKjN?rDM_hkx7f2d&<P>$M^M0%M&I(!0WdguIfM%`j>BmuREq_?3G9wgJ3a?UEf
zNu5E*GM~FP-nyBi#ih@ye*oB594t17!n>E}Dc*G`_M&mu>jwlFp?QE6wI;)AuhWd4
z&QH=aUbYFW#=2N0c$HjkcQGR4j&VcrrU+~$CHDdG#hDU>Vl{IE!^R4;nxEM(L2c_?
z%c}g@ksdgX+Wk*@iSuRo!HhYoRLb3I(bV)k1E`$yj=fX>kc0yi*SbEKSdXE<kt=4^
zrTkWS>08oo1A*eIFmGITCA3ak?~yLn)ulh^jjiXCiv0ht-Vd4l)3aBr&)#Vxg{c1R
zm@5*+`c2+U-hbma&&j=?ESBq1?SR^TE3HZt{^h*0_~U8tVdW~BT`d4VhbiO1U_kK`
zd&xI-04YdWIDmllnPoPVaQ&Z~(1u$}Z}_8rF3z(w5rft<fR*2LH;k>fWYZHGaNi~5
zQ8uD|1Zv#n0MRUMy`QhKpbcQC5rC__*u7?D?>Ku|Wd(~V!T<@uvdz&q)}w+1#+l=s
zh3eg@%D2qXGypgS=8Ak;t>K(cMQ2sar3Ca-PBR}xyi{NZpq*<mxemjkVOF-_mNC@O
z#%Sopml?Fdom839`=?(rPB;&6GA?)GWg_|x2Fv4ds%Osgw9v-6P)*JH=~!~%R)hbW
z4ro#lOTI_qU(<cp{8EOQNTNDNHJjf*z*a;xaC1JeTJ3?K-(GTw=+g*1%l*h@h4&`8
z2FPv~ZH`Xf9u+-sM!U92o_|Wf`~A)1L$_i&>eou~Kww#H`yEG8SlhZsJ}CL=+6h4R
zz5(`z@SBSOs>`F8>z~|;f_F2a<}9W7F+=$uWgP?EE2X?5b=NDVs!FGAPwR0VxFxo(
zW3P)8nZ=9m*94^Tjp3cC&h8FTxz(89?~jcxQY0ErWjv-$t$GgakOwX=IG2b0+0Umh
ze(~)`nWGnN`bITAom|Zk@dP|*8Gs*EXZu$s)qK*zzJ=WUmD_#ev)ZZZ1?K|4^ps}v
zlwbGI1+c+IDk#X@n+4E+=Gkoiu*Z==3C7!N?}&*B`+Qj&66Shc38;}0`<I31(E{_P
zK)xOh<m*Hod@EViB?USmLln3<YZ0V$he^?&V3(8PfBO4%9yz`%m$=uy%F*Aq2f}nf
zt$2~$K>)>-&in)gsn>3Vy?L1UhC+G0Ly3kX50Uq}Lxe)d@l}7F4Bb(Yr#$$k_lDl>
z$E`WJVW2BlALg=#P>p#ki9DxHu}*OO#G1Z;0Y?YTuP$Xuw1q4NtY)~4Ib0j8{IY92
zF#aJ^Nb8j(d`Iu`8Gk0#f`U1!f+de4RK7v$sdP%JVaj{=<kg^y!LI}#v>(%CMN}=Y
zOKLLAbVxf&krP@58u2_ENrwq0BiATK&0(I#!)Y5&V9OEgGaq<=2g6o8uVc5m3USC5
zRR=2M-8`30dun!p&Y=NR`PIjtBZ^~((ggT|r(x6UJU81;$+wrc<t|9Gtm9G|6|Y-t
zR}A;K>0>P}z8g{LPNR3v{vb6Fi11=nUp!yYv!yXz-U^dw2Q`m~{I~8dH%0szsyz;2
zyx{@X<2PKfahI5mJnAVRSx~k!;DziLm=$d*<Hq~OmN}RECMQ{nJ;L8bt_Kx}-_3@(
zsQc$H@QL9FRRisV1eY!ap4ZpP`u>%*cJ=b=7QLnm_ExdG9gjrO7egHr3t@sasjEG9
zHvx}NRNj2ivL41pw^_gO9*<8ZqvpUpIyfjq>OlmGHyJXmZxAW`Bl#!9fIyY1Z=^Ny
z1st4pL+*3o8uc`UcXZ`@WdM2Yp_47w9Nn;<wVnk~Q2~97Hr+Ch=;K1Q>s1*|<i06v
zi%`6-ldS&&*T2Quu}A{`#_r4nsf|Sr7h|trja3&TUYaCH3$VxcznpWiemhB1<6);g
zhPkWxr{P!jFZ%0|Z;KI+l6c><K=$s8Y%z<My|kmfD&G%=JMoVfkn}wwruH)&WnhD6
znkh>ZTRaZAB9D}wQWkf6FMWD^7FCh$=Pr3&xY)c<y>uD~JA`V&3yxORc?z3d1sG@g
zQ|r~^jUvm4bzAW4%Qfd@d4t8NG-(&@y0F8Y%73Z#T<mw49iZjU7&6K)E0dSkofza~
zOMD34nu7bfeX3Cs%tQ_*g*}5cP}d=5+!rk@tJksxoNb>G_0=VMAGvD}R!vk@aH1Uf
z3YdBtBGvUrmrI16WoA|uKg=MZ%KbnOlErFS>rc_d--6ZWl6d}7ZkU|!HL7767j{y4
z`8S^f8qJphSBx1+HI@9b=4tQO5ZAQgPN=#nQHPwJfBS1chGlNQzuA-lZtCboDG|lb
z5kLJYvhwb<Wo$`aFOd3bJ`nrv)X;LnKU_(FTXb~AJCtHBi4N*xxjNVr_06#)MU^qz
z_Tu@UBpC#%j@mkmDzzARazOUz;9tnST+Hqy%~(jjPwK4xFSPo`7OV9_WqPNybAHf_
zas9V*ym5`ovhUNBu}jLuvAExgdh)(BUURHm^x?R0=N<lk5{iZtne(h!$$9EVwx1~x
zELPu8WL3Vq|2L|FWLann>mq~h#0QS2WC%DF=9Bzevc}}+xlLCmoqiG-)UKcln-R#o
zesBXU^)j?_NN-Dsa&$+haPSl&AbD8z{+<R>c$}FN!u=ZYcJth%3E|9=F6q@nqXmAh
zL7^qrFa(8?Q^d<~V_ze)r?>8vc;%=_1_52I=J_YeQ3GD%$qZ58I&u_0KVhloYU=x~
z*|5IqU{FgA=I~BVBzt&W4~G7mhk`gCdz?|!tn%Ht11=PJK3A6AW7_JS?h+^|An9oW
zbdnrKr4E~(FOK2^kKTpY)7}9$0{wD)9oe9oCB9ty7F9karT+`#zZKpH<Uo3RpDAb=
z{lz&U+gc4~4btYd(MT6?{I9~D)Du03c6T#9q;)z-3gDHBGjDU(c92Tl?D-ri)t0`X
zUd2n)hr%M*Wzj-ZR8+ve`b*CQu3t8%PR=E|!AuGB;e%PwWO$jJ@O{X2R3IDpEO(eA
zP!TpPFaIWUDxfPFDd)-2M26?rBf)Tk7frz~GSD*g`S>8Cuv7~8BT6RSg$-hf+U2ue
z)rfe>4Q`=`Z|!=GXj~_hQd!t82Ie9%N)XF{U0(fUHna+0izt=`UoS;U6>KjGK3KC7
z67O-9nlDleTJ<9w72Z=Bl2%C64!i*B`0u)3$pC_9;Ce;!LAj1_r%r#w0-i|)SYm2h
z%+giKaZ(6;l8)5-z9Ip!FIVZBp4K3T+V$g&^$5*+u%DbSeRW6dJ&A|gQuf<9ym)h8
zAsDC6x`wPe*z-}}bZpzL6S4qDmPLN&uHX53TZ;{1|NLu~+WZ&Ylxq+3>|lopogH)E
zf0%V89%)8P)Ac5a3KEEh^Kx3x?*bh6Ww#>hQV=)RQ4G}adEeuPN?V>1g-#jyV&CG$
zXOpEU<6}_eLtMpPL5y65hi(1s$dcMP*uw_#Kw;h0M3TGw2TicUf5EpVD4jea){A*{
zaCTO{#eO#~eO>n|;Bb6th}{Xh9~WHKn8J<i(aDYb|2daE;+MD~l`1UzQJt|wpW1a^
zM4u8-6t$t~1TP++QxX0crrS6kJQ;?ZJPGs0)5nsi?oh^OCh;FaC|uruwJL)<-_N_%
zxB$>3pl4!0LoQ&j8YJ%r3aAVqlg&bqg!FPk0t_}^I?A6%x_IZ_{6<h~+oTh<PoDC4
zQnK40Oa=l8AJBr;MjXh)wxZ%ms2G^@%+<`Jk$SVOS@;u@GpQ<p7{tp5OU^(e#ftw@
z&!!^2JUVd8jAeWb!%!cPP^vi(3l+8)P_oPUgl#qlP?>UbKEL$<xEr{tdB9?yQKe7%
z`21rSbv9t0afp?4smeKtO^Z<`UlrhvoL0H&;uoitovIY04ybk=SaNzN3A9TY_TX?)
z5aM4X&X9q~fIz&odf#K#*ER7^|GXCj(n~~?TiNJsI+gSC6$P!AYP$g2p%7ta=yGW6
zK4M*@y$ZNda}jXVCwyjNi!01N3G-J_r*$X75CL*}$K3yE0eVhr01pae0;czC`AS%Q
zBGTaYC^h^-9J1-S@C**tt4rnlUF{f)+(IUfK!KU3U$pyQ(0uk377HF#nTn1^kk(=U
z74WzQnJa)MOH;**l+QfH;_M*(EO|$e_l{dZz<A5<Sl}$cP>BuhjCk0D8WFSZJDb;t
zz?WmYWM^R<K-F4b(?_@UWc@{!2j34d4;9am4Y!WlD_-~)$`#gBED)KQQ+7_qss9~F
zP+bA6vlU?<Y^;Kue`VQHj+WsY=Yj;5E_WxglebNgTdtNxu&W0Xw}v5J^Gpy4w<aMv
z^@ZLbXIu6IfD{)9dH=+d6dz8TFOX4aQG+wb>)x<}wG_Ze)1<ikf9wbI$nfrCz@*t_
z_|I8o<C*dPWeEBv8%5Z|#|8pgm0uqL`J}AhEKeN*8BRI+WRYkp>S+&!*6<5q<P%sE
zIKFCV%;@EobcAC2E^q>9QwTzM#y(i@`V@Dm<hR*1gf?$*3w}x3>QPSQC5L#KAG}y~
zgB|XMWq;BhmH8yT>5erSXkIjjO_*q`{{t%aqKM62rDW3xpp3t{1b;>Gt5{Suj(3N<
zlFKrsZV<IC2lI$fOv=#_LS-@M@xdNuBY7@2P+P51Zc2)QZ&e{Iz_NPdw#%YXQgbrC
z)Ym7my@bDOyyl*xa7jHC_Hdl7zd5r8ex|(129iyx;Snzs1FJpEtMoMhmA?o6wBZ1s
zE##}T$raMga0xL#vHiD%7U~n(_kycw{=>keSK0UT$Hb{ES!O26{8Uti(U||N_X@jz
z_`$eIAePo<?3!#vJ@De2;6Yb`yRYdP7~ov|&E%S<O3!nx)@W9U?od-NwepCwmMf%q
zkM>yB7*wxIGqB?;NkbskMP$?s6(2hvZ|IhM1h*!IudG@2ITq)}F=gHroBm0r@u5IF
z#S1)n*!RbvpqzgsbVG<C_r$E{P)hMXa*YaZO#vuW&tl^#Jl0wSh*9=YPNVz=loIW0
zKey=@*sGU}_%xcJs@%uGQP3KSq9P*096V?3Ya1(dgC6c4GK<s<&`zLVTYa(Jc^2@F
zvgmo*{Lit%yX1)>2yW&)7Ek7Zmx<k9XDOF$CCJEB%_?`PaVg&CfmliQzg=P<EP_p>
zAX;JaZ{W1R5};s52mE(%v>*lMLJrb2k^H{0es^VSIdjLP7kUdL{>qEgIrjRlbLH&G
z^)T%F1f9_R{hLe=^YSd>Lo<6unyrFmId)JDqc{zNXVZV10D(9&Ir+g==ANWbzUQl}
z$>O<;wl)n*A73)Ci$8^V{7$|i>r~An-h)Jli;7k&F!13>+c!YgMd#iV=y!ONv0xqV
z0q!{k*e%4%oCdl_b^T^7`oY9=W;Kid&3#a@B;@@t3P&+&57B0Gs(g@lhB}Sn)hZea
zpBOLr5wNw%wfSBShKo{UT;c^4vrPECAhHn_6_faZmVV{_=q{<>ri&sQRY_L~RlxPA
zWbo)DY$6j`1&$ZecLYGs$vU6H0B3HTCaF!O?fW}6?Ck7c{$s{v)eb|`zL**HvOwsN
zC;AZlTSRnJL)<8hPLIG@iMVDxS++E&SU!5K9?6%%b>o$QyDPi#^1w0tldLz28y}6x
zGMH88985z3+vmD?gsg06ApQx6tdWqsWxNXIjD{$<<3%9$RJA-`XZC&F@R<p!<{xYa
zTsN^<Kl%?F1A0{vp?DR{qPzMrvQNNf7Y4h~a&W&LUm))UjaXNKq&<X0_+G+<@5-~k
zxN8v~1P&5fvnu<wXHlpFATlS|@TB@lf@8*=6p5mGA{8s8@V98x9^_>oW=`H9$Kmg$
z^-ls*w9L%qZ!*=jN-z3nSY>%~uE~N?r&QpW`*~CN;)IBoeBiUNS`Xk;3l(n_F@W7f
zw(x>GiU`kD%Y+Nb-XaAJHZ)IQRKPBSoeAp^*a{UmGw&wS{3JG(?Cu{@6<E_oJ^h*`
z5bRNiDaz4l$a|gm-tuctZvOrT#I{$FHq&@r$UZ;Rg*C4!;-x0B+GI-g8q7+E#hW~*
zU(5QJZlk7>ZiKi4H14nG*-O|se^6f8dY;MxF-MXUL(z&bG%`b>B@U5gv-%wJUU$Ct
zwmnUQx7TS6WeS7GSVRKLk_ZV2>3suRwm^FjsPt|dpA7_Jf_V7|+B9(`KNhx9L}-jd
z1X0CH={bD|Z9lG%(X}@c<C+rqL*{{AMFGxt-a7W}Z1wXe)D{07tAo0Er-*eLJq~&~
z;szz-9>04sZcj=SPJi@bOaf2On$Y>w-66*SxvtBvXY(rn9PuEpnRDb0`*moU(1)v@
zT%l=Ne_svKR5LAWV|W1Bqb6eplsSm?yWp7VxDHuZCZd=I8b<+6j0EOG@k6Ip&y8l6
zUO*a>DY4e`Q~~p;dH!h<G4<mH3Uyf3U5SirKSs=7S`DOYaGB6K@Sk<5m}^kRXRm*H
z*e?^oZe~lj80!@V7@+9WlU(zZQ~kAJtV74eQs*R;pJQ}m`<u=UK0ZFTqkWXibfuhk
z3hyG+nchv#K5odVt_=U!d7<^Q(}@cgd~#>&t!esPlffD{0Y~Dv(Po}<yuC=bT<J}F
z=V;~|NGAKL$xdfKQB>+|48k4u&jy!XuOI}kgtKYr{@fAd#B}@bNESCnCqZM=m_Sew
zmON_CR6)PZ=&qQ!s6JvFGi8_vrIt@SRx{l<pp_=+Jv1dDUF91`2@MmrY)gfAXegyY
z$AS8a?j6eq`DuaD@J(x`{)t2@R>)fOVp}H7`Eds>_yWr8y52_z%?dvH`iA+MzqS6P
zh^sA%5jSlf*qGbOaF;4E)Iw{p&VG_4o3Btqf8cl|vuV-0@aaEMsVyy?NLGj~x8!i^
z5r+NFiuLLs?C?F}RO_VUy3XZ!HtHKV4)o~gmF54!)nt;}b3+wtC_ZgL+Tv(I8Xoi2
zAV*(oEhN(1gZ(bkjbiHkgPSiGZt@8U--x8Ar!U5QXI_)=*-bWTsx;5?-$o(F)!V$!
zmurHsrFq=3dQmifmB#JGgUCdWi{lgj2Q^ph>Dw+l*3=(4ZnqS7do+bQ4PJOnp7Z63
zBJ6Ae{Sb=gfXK|}&i^$oNb_K2!`)q)I<~y4Q{RaKl_gPDWx9t1oXHH*tM*d-AR#W2
z^`<1nIc)!pZ4X<c^Qh?X*yAru63<ijH&8dw!{^-h`T0k4tE=DbdC${Kag^RYQYNM?
zyCzS>s*$-DbCPfwC!|ziL=o%qaHA@ML^v{hrs)5g|NL;{^M>P(>Zj`mH1m+patDK1
zW%*S#@AyEnLINH`K0!UQqY1;Rq7n|9y?bA;M<#ynxwVfDK1+I&)K!HZ5AM@EO(z#h
z?&u%Ej!vr#b~*Cb^+WvCrmtlISq(lcD3AioFuJigT-M;eJ$4zOF0W!GUF!s&h?k<!
zE-2tlLYsj~KnBSG`!L0=UGP@A*0o*E6R2th0+FdSykSeEO(W=Hr>C?o=AIC5;P+0#
zVBOQG$x1!f-Lzi=xYGAtg+b<ROO|JNY^TD{WhpZHnYH(x+o_gMUR({Yw>O<B_ls&z
z=8>zF>X01kjfG1VXnh79L=e?`PG*~lkkZX4F9vPI_{7P0qC1L;)Y7+8;^Q|O0;biQ
zj^&6Lpu@KAWm8T!=NcWd;n>M*8j`+fsWnyH-hMjf`n_pR0sqt7fY*~>4{MY#aL@gg
zJ&ijI9zl<<sAU{KxyD8#()8txC2vFovHIR#_$>u8rd9td=TsBTN{<hynOm4EjBSN(
zI=k<N5O&oX>gknG=e}!473p5OO-w<OoF@$_X3D9OzbL!vJ?xeOyfL0pGJ!KSk3(;S
z!3pH3e~Imo_wX|hIBqB?h$s7w{r<*waWF9S4LjH15FYR(u-&qMR@Zqt*D`2r$h7W>
zaid}cX|Uw1g=_n^(7flym>W8lVW>jFXQ(03T5x%wC1`Ud8~OWmHk{G_;g4~RB7(I|
z7A|tOZUC#fa76)&ACxm*YjQv{&!f?pg=X8H%*w50FC)8dr(leyX`M}s`3vp$4o^O5
zrM$81&4~TV<SyxjClzv<QcV-QHRQ6?F0bx7UWk+nzooLVSeO1)C{Gog&C*pdkHIV_
zb@}c=x{i;1hi5thp8fnuChFfDps1MMcyfk6Qpr3%^@!MY=6%r`_aooguEYIDpvHJ_
z^s@6rfkV^l;PWL$rxNn1YJ=+HtWIgCA7TxUKBSy~YM4LHK@@vMC9ju%@J^ZrHX8o9
zic4TCa2x#uV;APUJ%A@!#^&A|ARVNNH&-RkM(UNEX5kmU2AYuGg3^Vmhs#?DO1pwU
zjks@zB_Y5Fa7Ic;w=VClRGNyfB>s{|U88XcdeqV#wVeI0z!GM(EBI*8XEFWfjvm{^
z`l#=;j6UyyQN2YLr@<zaFfbbuMN_db-y)}2Q|0$_#jPkr?na{uA_{uQD9B!hepRMZ
z$sool!@<$*w|8p3={#HC4Ye7r{660sJ!^MQ^{%MJ$Cf7P4DD*$Vf8o{nTrVnDevJ2
zmOArI0HfM^LAEnz)}Ec7-iztNscWIzjy4h&TRfT$%kbp=ULHPNC=|*wI%aXX(e<2f
z$;imaevnnh+g0>%<P2AsEGTV9YZv=j!EBm!71L#kH-jDrc3qi%NZxT?_&PJ#W$*0x
zGfua~Ap1>%?ZOkLpyQQExhiKL&x@d<yya~fcI)52<4?lL7+eWnyCR)*Qx#ZN2xj@4
zttr}X=j)gKpxZ+kV#}V3)sJ^{Pd+k>M5FGC9=1ZK<L>FWK0JT76&coruQbPnN6Ps+
zp<D8Y_D>Fu(p|dts&>qVV}naWTufy)ge+vA4tHJ~3rG8Ubaq?38!u89cUr3L-a61I
zqS;w4pviLiJ^uQao2B!;pdoh`UPHo~F$RbBEeL6a{+!~s`|R9&a(Ih(s5Usbe4Y;-
z^sCNhvLCNhG<;V&`$(yM!`o-k4_)#&^lg@Dt5uNcvbgbZ7k-u41tU&#?cLDOkU@QN
za>r|VGrTdk<n;W?5p&&Ibjc4du3~eG>o)q`ouI`rP0d%>g};85W84;nkiMU%5YZ>k
zgy2*XOhEJ)iB|WjIagJWhR_XsulTA!4ElI9A%<G#N$0b%IozFOJ5x4W`8~R<Ue)!8
zdBpdkf}D*_5(gZdO%xuSA}F4AnD-;Mm(l)&Fvh~j>9HE&7C$(}ZNlg<*E3<8AhVC#
z-%qABABrJ|f3LI}Ut?AG<+YnwPd9d1Z2#u9JA5;csRMIFwd^5BbTp((lcZ%yN{z(?
zV_T|;FzSuRPvqBp8P<^zeA$8PodgY^m-C-qfE+L(4~8t6g6qRb)ak{ks`{&v?{ec$
zw#1}44>p%+>*6HA`0upK%WmP6?}N=7N-gb`1uhN;Q?78{PP$FAX0@c=g2^IMUv&Rm
zJ+-l6cY2mFymf{PTYLX3Ba{0pSQ$IHZN$i9{FVx3Qj7mM_~1qnvJp+Y_hpaHxwA8d
z1XW@92<`AIEhcN)Y+U4Xy;?$#nnxjg&*{`OU!ef+8(=6&d_EX=GcY@t6=Hb(gIVB7
z3v1KDBgd2iQto`83xam2jnn+<;K|SQ<|`d-rrJ}#e-JTYn|D^XOoGKcO3}cg%F<vr
zD>!|;5bCpNEO_Kbbksr{Dvw+A!r3M5>1Wl{d@*yNroFgLXY9Ls4-RJqc-QDzkL&n*
zQHWWC<{`Xx3jPGiKKVB#08OSKRw5Otz$aVLxn&c;HrXdI?h4DAL<FPunjOq4JBDn*
z5dQ0-v@#G=Xo9!jjOT{B(ddM>dTzVQvk6U~o~PTYtE-RtZ6e%;Gqa98SCkI!YbJbr
zb&6RH^tfN`SC_XP=f!2<516aJ?Za8;=~MvmY6xVx(+s{bXt{Y*iZ(cYwgGBs5g#CO
z=}6XmmZNA#HztTH<LivQNkYOI9TVf$WYa}yhQ|;efHVheV)l0fzvt%llJzjSW?Y;t
zh!1z3;;nZQFOk^N0dV439Fr6UrrO6m!=r^vGvHt`jy#<xGEsfUWL1)|-DRZ071S1}
zI_V88pFZwxA&_57^9lvh*aa1Nf>yyEb=7hH7u@vMuVJf!dpp)4kg?vNf%SRl&6_vv
zHu}+<hD&l^YYc2ggU-G>RBwc$$^~NIZbau}sT-lmRT3s*i*H`2US>o(Ww#i2vHX{{
zk)tLcUwv%r+V*`1)C}1q^6&?kk9SYVH3tv~&f&U_`h4p6mOGCGh`jEQpZ44U6@zTw
z>+p)vm~~~Lt+$AhzRuR!IhVGir>9HGI0tIGrB5B;I1rKdf?xa`2ggjDo!kuGqk}J3
z54@&P&QIn!c+e|tWw((Gc@R#ntmDA5v()gp^kWQ<pu^bxDiicMEy2=KN_jq$bNWct
z>I<>Tu&I>A@0Jaw4kimnoF}aLt9O4+h}_WNF`7u;DUM0QeRHhF3E6-B-Yt4TQB15!
zC0`@=!aALcBK+qk{2Qva;Z}=O7E5QwB!M3h)p-rojn3#zvHmBn{VyjneX<!w+8k;^
z`U}<4OT6MFD{c+Vz6T;w*OrBrgiVk>4A?(?XYz~*+_Ggc`r_t1txoOW!1!Ypig+E^
z2VTyPDdyV7mjXoHc7W;_oW$wq?Hj^NzP!6O1M6vAzndiX+XJxWZ(tiT=jSQ_jj7-3
zMhznrTa<T!6wumkV<RQw%VEW`o-2!7lKYXPR?k4r)>A7%`*rgd-!}@u5472&)#SVL
zDojM~5YnASa+vKZSrVV=tUXU}BIQt26<zgYfj3$8?q1_<JKFV9QBvZ!vbGf~-KNe5
zl7i1e3>TtV!~HfsBGWGj(+xhD15Qz!jTRfV^cBPLp?0U%S%N+_kDPJRT`LmD@8%=y
zqRp~N|4PE~RuMJtqfy*T`#4^lTsV2#2ip%4tE;QxhWILcU;O7vmYt&Dt}>Np<9+i3
zb?bGB8G&=F7&p7^JKXq0L>2@h<mv)FIIursxr3M-E?eIqFQi*9&o8@#paX4}v{<oe
zqkOL?a~>@j8ClQ`h0v<dYi)T}d>`*5e0uV2h_8>FlbgHU%fwUF5HMdJe>tdu4qEn&
z_H%^d;ISzm#7&s--7tb(J9A!JMW^@|pFJV<8ffNy%349em!U#yar`uv*XJYq1F4Gj
z1^;y69c3QdqNyair9A5p&d{ddHTAXA*H;BZ#m%O8tDkF@Yz;RzxAu28UlT@lWTzX*
zUZkHJb?|j%7zJZwMhlVyB)nG3EnSyye7(_gFL@-GoE-PWVGh|eI6T0EtMT_sV!Z8%
z<HeJT+E{%sKYv>so!|Um2orQ1^u0gqAaJr3R<Ogzvy<`l?5+@rqwdcdI?`BqaPh$!
zluJtPGjBHpEcOXy&)>&)*s4^B1x<!Kl_OFw;ulL*N{hRV4ZfKIH5*W|J76$D;9Sy7
zQuUY&(+sK8ttq9wLst!VSBu1<tIb!#!hXJs_gh+4?QeOtknyak{j?p|B;UI*4={=^
zRIGu*;8&ErQ^+13q<i~i!AbWc4+L)t#$|to9zw!nqn0l~882wl8HS*y7~MiGsPZ)2
zhwL5uUI9I80bQun1*>O>djO8Ou&|J)<S#AlRl#Yyw!4)-cjr2Vz_6{v#9nJ413gX3
z&pPkuR41s2rgi^<H%5wiU~TxI);s;cXqM&1qh}eWk!L*ibC~yl@p#%_!%P$u6v}$*
zs(<enc3regT5f;Fy)E`iE{6|hWMsVW?fW`gdX1AiQ91h``zjO!%d*a8v!nvkD8_#A
z9C^tGhChUu@Hp-#cIpS~4sf-!0cHZ}jjEx@Vo23*RIRFABX1LxFx47#0v)HMmw8k+
zr6~9>{CHz?a{YP?jKpQ4?)*y5Dy=e%Z}VHp87eLoIjOoO?=QuyL%u#wKK}xidU6x)
z5uWB-jw3vRx{UlfR!Jw#<D8ayJ<JA7-tHK>m4=tJ+7^;~OdSqQ0pY@fDuXLENh-?8
zv+UnJSj>Xg&SWif*nt)1RDtQ1%w&}h`Y?Gg)MN96uv+HsK}JS{KXLWd{k^@Rb80`X
zz$Mf5>QjD?ut<>E+l5C0mZ;RWA<wM|F=(@rW>!j3KyLb0*vB{$K{0W0&!&K)s&|X8
z<vo9pukb-yhx;2%xjO%p?Ix97Li6(t7+k76{-wGylZ9{74mLmWJ2CS+34`C}zFJ)K
zcI%o7Mf@N&nf7%j;}YUYn4p4;cU7r#{?(@_Ni|p_kj*eoT;UREv0?bw%8k$?{RjAc
z;mF)6!p2OUK@TuVymX@mO@#f?0rP6M0}^HcjFW$ST4$?MF5a8vb+qu+o6(uYc1l-J
zFKBnI#r%%iYt3JK$=v1*O=<&$UvCVJjgOZPm6vdM?=Ee%y7b28jRpEh9=mQX4BIjY
z+!#i&6eU0NuWT(oZ1uSuP}g68;}`Q)FwoP_<j1B2+;~fs^P;$m59?ESz65;q9Y-lU
zI3Z7DD@^U-*+ZJQ0inp{Wxc{WzUCA%$ZPeZ<|i*ut%+oweipzlS|8Y5<>ldi(0_(0
zDw+<X@UE%h-z{iZ4ET<yGp<rxbe=5HsVu-oa*hv45uU}I(FUINO}xJ<FMoz^>m96Q
zXbjEojWv7rc_Wv=B&G)-ffHL>)tI|Xr!$vad~ufKq>tDiE_T+c;vZ0RMz=+uPOvS$
zNv9%Ri&#@-Vm!PBed>xds;F7ext;*f2{lW{%|o(y9U>?E+{@H7(;o$5($&BrkO9X_
z67SRN?ez``N=Y;(wT>;5TUgD}G%2z2{say$cwppJ?qF{7O{D9oYQ(Jk%@kg&(9-T?
znMRn}5mfXZ?^G)nJ3F^<YrOvvS~XSNYcsLgdAz4WV&TXUz5C*`v;IvZqcy(R>}-up
z3NAlxqs}i)wXxs6ZJ=?@)j1yv6XXF``SXpUWuEFbZ0%Pa6N6ZQP?M`-i@PQ}2*BI|
zBhlI*);0Abq?9eU)HI6v0R*B@dH8ym6(sv@Ve#3D45hi6ps;b%uEqg&J+i?T{e12D
z1DBTx{=6ORBR`jyGv(zUhj9n3g_7T+ZdwG20I@oSfioX{FQ(F6R4R*`z6}g)z4*??
zv*TDpnHdO&^EO6B%$hpd=steT+`+$r?GfI!Y52an8dRhEU8XIPDc?!rsP%k*hyJ2p
zoYjYXI0|FxZoh|d^P90AgxCa~DC2<rAq{FSAp6RdZ4Vd+3zHe#g-OVgI{PmrSzHkv
zVxhufrCRqE70O?PZD5lY12jGg>!Cg@7YD=USx}jizKdmY^uae2a3bu*MDv#^o_i+c
zGq(ZI`MkdfN*5?YsC)J0%a^EcG}EMv$v=MngoufWN&X5|!d?x%_e&aQb*wb*Zr%G{
zosN+V{z3l9vuC5<nnI?{(#1CM-N~$MGqChzp1K&qEPmRCFRw2l`DngDZfpBSclQa~
z&*fVKt7kujPU^(Ywc?~)d+Gs-Ot7~oD|19w%A8JSX)vv!*)R>Wqf{ld8K74c4UitD
z+LbPgg*&_CGXB9T<HPbShg&}b+T8|hs>%r7GA?EC`7P4DGS6R73ov&rrHy`FWwe%0
zXyhZg$qC!#W^F=4@NaICQWLvwm&=(5dR6)~lm4dmfTH30bOT--!0mzqmqqrHFJ1rx
zo3n}B?+#GN$TTHk>yjy1D(7G#!1<kO%`bB0fM3wwPBxkzt}0%c*%&X%@ZO0$GVI!H
zQmIu2D8a}f2N8#{N0_f7OG^HBpL8}aZsUic@iD7|gl?DbAG__A%)IvFSm@o(;IkA=
z%X6|ZuS0LL28&`{l^oE|N`@GoxX<}uvhw?T1}}E!#hY$7Y>%Fml8T5l<vAvjIT>en
zdH;T;s-Qp=8A<Q+Gi*;QW@1?*8eY9o)ph<IUEwesE$$V#AIBKuV_M^xt8q#5q%3vR
znKap@U_Ymr>l&9dl;&8r>5EAbk@n%YhGiDXFpFD)IWEIwA`9ZI4Mwg}8K<uU=UNV!
z;;avX_ISCuKSWe@c%AMToa;DxBYAEy2(G`V-s=!CBNh)l&T$w>uBauw5h^Mnrm|v*
z%jUog%yJnX79{OVY-b$7CX9y<XJuh-cNRq0GfSr>pET{iJEc$&e!_C#>6LFlcZ+XN
zzV7D1OF-J>>+ok_K!{GZygiubO=!4Ambl!Rp#V_!T2z89Vy&i2sC!DPf}G<jdR+ne
z6Hi2Our4gM)j2OHpv;2<QrlO)_);(aY36t|DY<DZjso;i{6^>l`2Ny-ufO+BMxzg=
zvz3@w!8G<uuSAtm%@?8Fk3mE}<mC1X&34b13)K$fnBr7j=F0t&Q^My@=$^(Z9~@vi
zMLo`eGYGQQ=j#ISZ*S2BE|W7I&2$B455$=V)4Fc15IfV3l~y|u$WBCJTZI<q=O4hg
zXL@z+36U26rYtcJ-t$^`ef?l-CuVlX@At<maSdF_eno{JsN{=TJ8+{)ceu5i*rX}l
zYTd4NWh@4~*xX&}7=W8^n>As}^jn&o#|s}-Yc0PvaiLB4VKB|+1iU-J+!s4(Y!In_
z9ZP*K8H^5Rg5P#M+7`p4rY=^h1f7+zTTGQ3MZKnGX7;=A+h|xJW0)&i^4=O;<K!}&
z5dL7d%XdJ-G}CUE5bWp}J>Mf#7TAUNOv~<f-5P5!=}!75%J9QxO0(^%vpvHjJ^m;J
zBgJu`O!)t_0P!xRvQ6JV1tg3#I;7;Jc3pJ4>OP3zuq~d%#8G?QMalzN*mR|d!5l%u
zGY=tH+X$V{v(zobXdBR^G&1v~(Yv~Nf7-tFisW5Fx~-_|@0W_+>O*E;FPedRY2STN
z@$=N#2!wa>rqoy6$3|?B=~U^DC}#g#4<^rrOp^L-ZS;C<0KL<TKt=PiY2<wfQ(8l?
zQb(upavI{b{KG5R8iY@N6t8N#U6?_HM}$mh9`M?t(yvcgAp0Uvv??q;;w3Bef(5dA
zQ{^&PzM|FRJ$mL}sN>gL5SyZq#d*1xP*W#Dkr8TM#8%=hAfb8();jPhSH{tuAVNnr
z$Obzs?u4t`N(l*+>^|RBG2Jx3p5zS4vpdcK-Y4s)o36JbZ$$^Kg5*7iQOU#N^>R!u
zJzFo0glZ7KX>vemt{-rC9^gHN-K%XL;R2pf19VnidX!0zC*<|Ane#F^Z$E@2Q*U~)
z<cYx&*&ykT{G_ObP{b!M>QVhWr{MT;K;jk1!U;jWs8)o<@*k)(thC<fh5(JQSJ_i?
zk2S7;WvOZ?6}oIzw{_+$JafL$d1-WPtF@&?DHD&qIoE`Fw^(PJ9UXva-|{3(4Qfu^
z1cA2IiNf@^R3cz28zk!(LwIJvl!kT)1Sr`TP6<vsNMBY-5>>ozQvX`tx-WstHy-gy
zsbbBeLliTCg;*d=5OF0~FvX|>r)61VoQ_jU20WI%;P7!VW=HOZoEL!&cwRMz)@$4>
zTYSoz_^icE<ulyP?Mfr#$MSOSp~yIq3`896S%+248cIXD{9<eiLmwc13U<&y0ElEl
zH`G&q^eR|P>#e2g#B|8970?@5KfNYsR}XEeSV4X0%}U{t%#07vI(JWh{36Vg2&#bS
zQ;jy*`BNv2h5uJ4Z6v1w$u<;Tm7hvZ^zari<ZRKgvE1#r*nU%#EtK^o?+~Th`+0AV
zx9b$&aRCsG?b<r;Pml<O^nZj&Q60<4Tf;vnfh?xMP!7J%Z%zN28@>}u4(uQUB^ops
zRTxh1cIXY)a%|BQ%_(QDoETZ(iBNopDE9Ad0&CGh{38+c5id!gO_e*>E_25_pXbZe
zA6Gf=)WxJ+aw^jP=p7DVZEU@mO)>8hb-)K_w@0De06a#K+Hwop^$q2aYE(g-_zgvi
zYL}BS)9T7?$(TRca8PXp-uBA9Yz!$6IEQCs<iA@v_SntuZxhI<;Re@R4Sq%RVPJ=C
zelqf~m><0;$@hr26r($^Lw_fxsySFvq7L#&ffmbV`Jg1=#q4D%Sqv$MJ8RHQ7FJeq
z$Ma?KrhKKMTWVLY?eo6_>d54%XDNunvM+{pDIz!A#McK7P^S;TCCUu?RM^HePFf~a
z7*Bumq8?tnCea({8R%=RN;Lv`v%e16xugnXPq(4FShvv1Pjh`_A=Pm7PLJi6<4GOS
zrB1deJYCRXTv&pizoytx{WLy*sIWV4WkN9AFDixLxNum6+g+4c(231<RTuT1V9t5|
z`O|8Ju{93kqc%4a&LloruXQ$XRrT6iGKaPxEGO=gO*U-s9hE*<i_2DZBX#^f{na;k
zvX!TXbL(6E;O8G&-E~D8dk&)lA1urr9qaQg8<hHqK(7Q|*yvq#?jV~=wLQLjk<Bj^
zp2uo>WVLz);=b^7xnjsi@Il_z<h?HtZTk&F8)psJ8TQMZm+z$_GN*kT8<jHPrez-3
zscoJH`|!yoL%`!UG#q+BVxw+jGn#VjzFo!#9e%NTChjCP`OrzK`9*Aw>Jv{(6L`&~
zU$9dG=_G%nkLNM2(fu=zNQ2+@xkOsYt3mVN^9}DuyRR(O)ydd^b*$2)Wo)^^fK`~@
zeGu4AmU6!9be?8_-N>%26glslTXr;FwMP3HRGDC`{ewUE+%j8rIzK&#%j<1@HYvdn
z0s+0c=aNM>QG7r?TRKAwHqB2yr3Bl0jj}8jtqZ-@ZCiyMmfa#+wS}9;A<Q@`_@UFx
zc`j<NP0-7hmpj3JSyu^oZh8JL5(FZZENm23FAIy+vnopUqX0|!6cCUCPOJ)mf+bIT
zY|!RmiBX+2Y$5`YWwELVMtT%f88G@Ejo!k;tAC~ZiIHV1Yc!gH3s(#*=k<9)J=B^2
z<o|ss^oI4VPMhOdG6wN17sINc(F0<o2;MD`a%`3a!9(VbtiQl%JoRA8DfM^>e1Lwr
z5$=yw(@8R@tF86L&c$TqyU-`wv|`lKr_CuW=Gv`8MMlSIrZwm8P_9ul1vH8CD|S?E
zro{SCfpS(~*fcb{@SPD1t@C54;^kI}xFH0r%SjyKWds=`_QfhcWd9(E>>UDSxw->r
zQ4p*GbwuC}aiC7AGRI$+*8M0Fu{6oDrP~z@=pm0&|EvhJEQSvup{45*tM=aJcYc(b
z#GXap&r^h(!0k+wU?zv>Rcv{&rTZk{L48_<#$#<GYE=1aee}9ejx-IRV?4oMPN4q7
zS@bcJm}j5XBBR*qQ9V$hS5dwbEN$tXVFy|_ml{F`5Q-;=Faf_O7%ZPQi1|NU`jKFl
zKsS_z-}|cio7n(2=<4PMPKv0v8@PcPcrBmKYegN?f(faU36|=$CE#VG^nUkSuy2x8
zELQKmDAc-;zRa;f#k0=S3Pl2Moq;2#SQsk&vyhW31B-Z{^GsHw#|Fw+>x3a*|JR65
zd5}-)8}OxW)MrnNf!-Cb31fNOGW60%46b4Yd6V+4NpL<WuE`qIu>s0zKATPx-(u+G
zSAZY!kcVYaf|K+PTZ7E38A=4af>x`#QRe&}jz+vdDC5QsFheCa1-<n0&96sfmA{L>
z)MHl5CcT+Io+fKz%O+?n6KmzF>E3K$9XVKLV@+*pq-Vz)U>GxNojbOc#y3NRxRUWy
z)aU8VyUROTGd{oH9avG^6^NAa{`7meP)$DJP1e7``^`WNNz4mR#U$!EF*{@P|A_kQ
zu%_SleH;f16@e!TB2qdPl$4f6rBk{Y-Q5j}%4iXg4rxY7jvk2gMo2SYN(~qxj4?*-
z`{4Ed{EpxKFUP^L$6eQTo!5Du_wY1jcfkn&E)uq;`6>6Fk;%qS>u4UD8XD&hD8Wqf
z!)FK$Q-3zjsmJK`Cp*p5rWto%%gx+Ps!?_fpVFwx@`0w9{V!e)nA>hZpVAU*9O2a#
zX5Sy5oaoUSSzTP3oFJUo$2>sV!k_aR-(6_rNIIJJ&G7J3s+2-U*$c9wpC>?c|KqAJ
za=l#s>R${K9H-@~o^20&7V#N1orfpmVFtc0gdZh)oW(sB#=DHS%J(&YNlWwB68R0Z
z4HEcIX(ap|CeEy0v@La<mR1=xO<LpFoEJ26#IS~=q&Mrow|p5?#)KWGINfFUsxiSw
zY+qV{-sxy01x<5<>r|z~DIPm>|9WgSqa^oYcXvn*(4a4~+}iA!)9LFGXrlSC<^dTX
z4;}gA26|ZWZCOY4G(+gc(}Hk_(o|`dP7+oe_<q?Pt!E!DBa*J51ZvZTmHwZ{gl4%K
zo8G-wmn9c!WU=A4U`1;;)f5Bll3;4MAOdjWZ^p;mj2*{5_{5Lv2Ys~(S(+|BNe_HV
z+o2%GGu60dViVwyNuLFrEwU6;nMuD5VLMQ7+@JdK^7gh%)MekV#=P<O6RsyYZT?%C
zs-IfYcFcP}gq_%$=j5~<9rtZ7ZPIpxX5s!G_dK)fWs?aU8Wb}0!EB5V=Cz;NMIsPt
z7t2`p^W1Rp0*7h5rVNtjw)vtjCe(fszhUAoQ}sdK^OVpsfuAWM)j53f4{@pFoKb}>
zdYv-1p8}TrOO81m6V-7PPA=i|(^reG%7V*>0W#s#Q;Kji76n^miEf(rC6V)8=&Z2e
zFT|GOF;nYymdP#kdf*JP-y0tWeKm+S<EF8`)mrnU=PjQM?z{ZwC{{obV^TFVNt<-C
zT)M>)u20LN{r<Tb_PANa#GpQfbJ_QDn;?i&?&-vzxRt~VnNC6%Ee5M#j#`iOM<1N~
zoY3X1RZQ>fpek^^+Xb5Nvk|`S?`^Sl?I^#UuS~*!KYgqSIvTj?xFj?)JX|<7HCO%|
z<}i;zAInPyfwkJ&+uKz?DCf+>hBifQ47H}m0KX-wcijcE06@%jAHg-m-Abr#i@oxi
zfEH%QBMmMl4t2ep#tGwjVdq*oUkd+-xE{E5>>I09larMZ*Ei7!e*qkbme|?w#Lf<n
zBJoPWEzRW<0Bv+M>5?Xv|7Ok=06X-Hr9}Q456Cg=d1kucTi%2ZYJ*l*?V4oMhYGn)
zG=212>q8_TrkM*T``x?(lpbX7=o_dUGf#JmNGw`r%9@%hGfN?j(T=dN<Nd6HuhK}9
z25!S958_g351ZH`%pOyD$#t^MS5j8h<nlbw*;_h~(WrJsH;?0$ek%8{VMXae(_DZ~
zc{M+!@($oxWzN{2eAjHURWZEMNN>`vE8_L$F#vN--ga8HdoMjEhB#|5-y;ZYdE(oh
zO7plX%cAoX5Zj{&DSUBpZUB&GlUe8-0Te?Gz7eG$_5@=ThQZ9A5J%12IAf&v1RpvT
zRz1>J4t=;0_RMG;E|-j=QgJZgMT8^|8Y@vux!T1F6RhQJv^{}KGL7ezS<>aRGZFKD
zfnJe}Xz(O*6Ics^_M8ly`ve($7~d1~K|;~mI<_vPm4MA#0&*j(8LQcP`_f?L$SIgV
zI+i?d%6TT(eXefY>QR~h-o+*t!)Of*zh~ygk;u$*k`OYbcmBUPa9z%ZMzBKQ;ND^2
zR|0+yBQHm%E;krNHz^jQh<~V9IcghH)&#>p=dV8@?T~|7HS;L>sm#{|>c#yJ8aZ{M
z8X?1<GuWn_)&1EEPFm-O#Xu+8p*Q5@T!=xyoB;Z>5Aq1fY>2`JpFXr{&w+N%;(Ph+
zl1iI(Y4I;bSMEKTJOY&V)Ml{HrEb>9v_>2|j0<E9tsdNwKCv1!262k~??t+ITDft|
zX$@B0XVXrE<kFD;8~}LF-i8k8F*rEFD{}8*Ea_=!D&;!y+hk+KDIReL=z6NDsrk&d
z`bN&=VW;U*^V^^XXw@O&;ogMk&YFGRZR5I-j6FE&NPduwEW-GWoUTV1HEsNqVx@p7
z8}Ir?eEmy%=#Goj8$?Leh`y7}QIG=f7F-~dn0Px(_pkR;y>T7RK|`e2;{tWLl-F{c
zJw!&3Qy&5rtz_xS2HXQ>q1$?&cUeQL?4r%VYWW)=+PR{c!$DD0C#OJlhwgTeJiC{g
zTUvviP77op4fIpr^hT<{MSM)#f8)`dXXe1%$<q6uk6keKC@z|&bl3Z>-KI%3D{`ZN
zx#~1{_yaI~ZkyGzDI#&%#kl0HgN-3hr&QBOElWiE*?Q~YZiU8`acFUDC{5oly3}1p
z$seR=MsLr9e8t!%%RK-3drKSC_m=CGYA_5>0$#&Ak3g1kK7zEaT*bPbL18J7kEkGv
zbfO=~xbmue-;B<bw0`+hqnynY*8|i4XnCyGCI0a@&RaWqRY81clafGQrI5pC1q8^(
zG-ucd&=k-AISW7%2KMwk_maXZ=VW0f1@JrgieuhyS)PCwESH}}(DIXtkA8Z$1!t_+
zkG2ZxCT`z;`4`yo`QopFc^6--Rr;naw4bS*=9bmP#=-Rbl3H1%PI&dcr6;gP`!_ot
z##f>94jens7qhFiC!<xYY-;C~=5Xg5O(u;2NtiE83U(<?q@@nNlw^qq{`qAeH8B;{
zLvBk#FGm~VyZ<PewGjq{n;6H~u=M0)lM0}^cXC0rS`ubnyDX`%i};*))|$2gX^m{K
ze=XN7KqsaK@3k<}7eM$10C8Gw&`iw0OWMtn;ga~@uepBj*l+H9r*nE!2QfV&ovoG&
ze%r|EdHET*vzX?bk)RcE(eO0Sw(vuSf)n~N4iha7Zpp)9p%USviiumdgoC%}S8~n}
zxM#4-6VLe<bDa{U-|PRzmccA|0?FvK2ud|;kXVlx)=8A{a~l3ewFDnIFVO(P)5OQJ
zvA%Vb)ae11BU9Kpw<BoKN^>~4Gk;Gnl$V%01>2L2MYm-NIYR3?&B`4*Kq>iG$;x=r
zTau+G-_XvfGhDe{>lBJsVf|2g*m6y{<;VmSq$GD6!4$33Il`{p9X}Uj_Qb#AT&Hm@
zw1TWJ<HwxvR$13>oj2|||7Jt4+v@7NwwKCQ?`fEC=wL7LfQSLqh@94}?sioFelzVg
zNG6Az)D7J#-X~BF1fKU|mWNsw<M=QqQWs-V^2xWpoi(|UAm6XPd_f&!TVs*5FJsZK
zJ)a@iR5Mi0&Efk}6&bT=u7Y}0zXOfElrimMog$-x^{AJCQcPC(pUtG8G8e;$dvQ>!
zRqx#z>Ydyn;(mpg=fn*ve4gl@BB_vqy1BTh$V0X;zumm!#Ad;_U&?-xh6TS;+EB<e
zTN-Q^`|a`dzRK^N;u8a==QpVMzU44YqP>`a!<$s#8=;UI#N^h0mHB&(-dsyR_w^XH
za$jGdOs_3c85W&Jp@c4@e7n3RXv@;e)s&40fuy|1iq1IQYRtc}<qg^*EfuyNF4+TW
zrQwS-c>vPnFC17mvVPZAE<1i4eXrW#&QkcPloKtpAu6g);obX!qGS#;%Ec|i&{xz(
z7^l-8ub%<p2c^`9w;mTr+b)dEUIbh}9MrdhC?Qn<iPwc1Ao2QZ-TLwW6hh#`0~VGY
zJ>-AD&7q>xKm25G{B&>5Xz9H+D$O@bq$+bQM#(YRj`#Z&nRQgz-h)ekUcoE|#k<|z
zmONEl6k2iD8=%1`rcFmN<gtb8nhpI1AoVg@FP|60DnIFT1#t&qu$)(=Lrqv7<|QtK
z9{V!uOHgl#T9#M}B&@!-Gu=pi_Lz-Ue&cSi0_T$<#XQBp4;fDY_bi^o!jIC6s&ZQo
zat>!Jp3y9oLGDHZ{yMt(8-ym9p%i;0VWj8ybumEYbEq;FH4L532m~zbwdQXiLa1n|
z#=+Us`Wm9c@;Q)Zotoi8xVE3gLvAkoH3hxz)7+o}vPG6kz$J<7-KX*hRWr640Sdqj
zp3I3>_;F)7^}{)x&*rFOImS~9mpsRnP7`ta6sxaBy7zS^nQ0cyb7qWyO3jBjgN-f&
zaE`jV`ckzlq5+VI+Nzc9sLRo$FBpUd72Y)m)NJ`Gb$}E_HC4OHJ{bYPl~J(S|6T~9
z3$Xv{F0}dEyxL6Ioo)<0O>G@Pu!5}ihaU<OL~k`e&EQruVzEG8qIke4d(pIMk<>AB
zq~ol0ociSL!0bhS{AqIW)#pwj7r$q=B!B;{+y0nQBY2ChJYvBCC@FrgirHP%7!SP_
zPK>$P_5>gz(NmrhV<17Pe~a(Q+bDTz#wb0dl^3YLZ78gx8vNh5f*tD%kX}mxm9C&u
z$74N(k*`JfI5v9{(ePu0pk_j8*NwEAqx>UsvUlU#cL6E3I^wL+{NuQ46%2vktY0_C
zv9#uX_B$8C%280w{=UF{@<<P#{20WKsO;1gn%tb)yFcMmFu2nK0;NvMn{;JfCTGBS
zr&CsEZw+xMf);&OtwHq=#8GSa?Q$Q1jEeE^GZuA&LpxU~Ope3=<!DuFdWU(&^z+|K
zLM`V|qx>?+W!2Pin5p$vjiW5n;f-9?@W0mTX`#DGE}(e{K)hpL2>L04<t`hy4Pc4*
zDgznO(613Kx*~6wb(3tnH#ZhphA3DQzgfg<1{{v3MoMx_hif{?vb38=iSxCi(<f%I
zVD-`bmTQ7=r#cVYhqN6!wu!&4;Cqeaj>=+HxPNBKH@@{c5>rtV)7gF|BLi>xe};*=
zZ^Z-l3v&@Cl!B-8o6jf3zL*6aY%&Dp<XHjv5dXy|x0<FbqbY^F4Ag&GqM@Q@6jm{~
zZ(;3E1nUEtXtjnPcrfzXQn|=u^HJ)EZ6A3GYO;6o@_X}G@vVfmHxHve`WziqPgJR=
z@v`*fhLmjX&b}S_7Ibz1u#5$~{e8u9LYihtX>qQsx5Ag}!h@}k`Xwqblbw_2dRUR*
zg;NIy+_xQ+KNO2WSA`OGS$^gPHVwXidn@e>)fTE)iCxi5bpKtYI66X=C5#OrJV_Kr
zpAd5Ia$I@JEMuL1Hpd*^NU9D!X`m@HsBw<`Q<78D;i!uhl}TUtqo!gUO@cu$m=fiA
zIu0&UG_h0<{9I806{l6mX|l|Dl>F*E{IE~r{R{gE=ha)D!KvSMNP`$?#4SOY@7f9Y
zw5;VawH#%TKnv*f7lO0YyLIsGE<5aMI?<l8pa1`_JM>G;?X3f3hv}2cdfkO1eaL8s
zN(fz>gSXyC;aI2AOTB8q!vE8&HiiNgBEZNGYR!XaM?X)2SY-r4bAPmI%8lst(U#v~
z(7U2;aL{a)+H1Wo3Pfst8Q{1qZ%dkqs5b_`hltaT>!&i>s<VG(S+F_y$|d)-@Z$=&
zAF*`?IkBAUs2j#BevVoP`ci9KzJ-&mY8+fRJUJPsv8CnX4%`R|dH>@0yw`i_)Nk8k
z31%Hdy3DWafix`q9f_5`b!k9?O|yoSXAfd3?|e{P^0e*O*|~pLoo#H$I;p_}jG=xk
z1`Z=s?8V;8y{Vz5*8U)cxEpKKiPh(>z!7I;SRaI`gJVf~cOK5ddH@c?I{D)$bOJN}
z%ys<Ds8M~%)Tz@E%=2oHKNtQLI{VPA!@xXycJ{FL*DtvlSL}?aX{S}~-iN&P-WwwS
z_@3_>>dO03-e&}C!~R#Mt7hKQvOZ_4<9}1uS+=8t_W(*zt|y?sT5q7WZH<ltfI2P~
zDAvm$KTLe>bop`=s-z{iRM&$*^Ow?(7U0k0O;R6K#wVT%4!!-NrEMM9+B#kl|Cu2;
zVD3|((2GsNLzzoCJa7?s^Ve@ivyLY}0rP)*Q~zbvN<;4(d>O}u2vLG}?msn9)HCJ<
zB;X-HM65N1w@JHOZKpV)BOue?CN*gtElWi`(g$~X^&1R~%xN}uTv)j?0gZg-R5g3o
zt_W4vi$v-B;>Hv)(}2Rf$GrF^ut11@HnrQkGw|@Oqm9L=&`|7s8#_C37j);8EAe2B
z3Qjt5Cfpj8`-nJm@pFNV?09h&&|qiI)-juFddKTU{t48bb`A=%9VbM9AAL46t-ecn
zpZjJeVOX<zRQlAXqDvuOn}#h8*FpNL__dl%8r1Q+r>mW1JiLv^yuJBv_FTOrhekkM
zF{zTzy3OA)&-LOV4DC6S5kW?(^xKhG6>|759kaP<l>YVWc5^U*ZZ$_nVby|g!tj2!
z;2%@Y?rP#fL6eY%uP00l{#}^Uys&s@bf<MwX{q9C6_psvwISPF%Jos2`S!35Y_g`z
z+*a(UGzzn}X>Kj>Vw;t_fU-OFO~)7OY^l@7pnZ05bG4F9z{XwQTJ*Rsb4~%+#|0rd
z^k;e7=w9MI2q01(`?!?IAaNIZ#{A*`F7J29^YTdEBYo33A)3(Ik9cG0XSJAGU!j%`
zRLoHuw~M_XfbBU6>`Cg~<pbpXZPhj_-#K3TKPs?wD)0b;Ww~`9;#RNZ)FTOrD8Fs#
zceGJt!id3%Qa-W#$f%VUPoC84cObX+u5<Pi+ZWCZ%!3|xAFldGHSu`Q`)_BSpToc~
z^wErdkEl!t+gKGYJoc2Z6SF<n(&K1yxSJDy^(ybJX_)h(<V;Oi;fFx@*`(82-7_e&
z`8f~De{XYY2W|0A3D+0sI`50f6%5wz4{vh!K3k94J8$z3Jg*Du#P>O;Nlk-ZJ+HBc
zX*hw)DH53d&NeIfLYL0}<tP*?#zS6g2Dm-+#YOXrq67Hz=@yzT!+?!aZ`QGOupZBp
zd3JVgO5pcj!20$hXAD$sc-fZS?FN;Zzx#i#n`BP68}DoAy`rNnkH)g4Yw$L%-)XMp
z{$g+)ub8^6gH@u@#jr+~!Kj<d)e2_hR;O+UnURsTwe6RTzxY=7G|0y-Yq7ux_cJnz
zMvPhY)t0u_`L@jVybw+{NsBPT!uRbHR?I^7n`8b!+1sr3mP_h1Yn;TNFJA&GC&TXF
zzg_P8Wu*ALO+(8J-6~%Lw;sL^+g_0L^Y!EO^O7aFNpXMk-<fW~^^!&tSSp(L>nN_p
z=1cyJ-#k_cv-|X|?h<+hM#R=2vS_A7pZPt1Z&y`{-x>&{%=z4=FjTniuzBc}m#^=w
zt%b)$AZl`_^vL}*cB&?a`)Yvyit`j|d&{_Rq|v~UCkxihO=kUzb(a&eoSl(m*}0U6
z6mnjwo1Lk_?0MHdbs7FL-AkC5DxM4r!>wFFvxQ%b$k?I4xtwHn<?2B9BS+kOx>lLL
zo9o$+_wZiK*2#%^h-BfF{$W;Nr(3O)%L&Cm#IdclTj0Upb`ua5?9P>Op^iT#b(bCo
zY}S|P;pMQaQ&ZUxGvR;LSKoKv-uo(<EU~4H0p65Qxbi7o@%8KX>fan{e=Rl^&5%Ty
z|Ew?TSBKfzVh*l;|Blql5OY3B&Is)M_28o_zmXy}HOWoi(9qCw#g+!RQu*tly{E#o
zc;WD&7-Yp8p1;PbavGa-u6XJ7tPB22aQJdKFtT?SEKi#WtW~?0T9c8J<#R7x{9C!v
zC#zNS{%SpI$jigkAFmBcuZy~Uv43fLQ=sGDNl_QfJa_gBZ}v*xac1phW9>ZJL5fyb
zh?4pl_F+>%>Hqy}pt|N4#Cl=3)U+!hmE1eR8G)~88`%paX)_Z1%MCfE-mIM{>zOzQ
zzu1X9#4Af2(gOhEt5>f~B{ji@2*y+IO~CUe`46@R{@mCa4&)7D6=}R^?6{Z&;0N)D
zOj+)=?&e=Vqr{1koC+=56&Iv=hCJUE->oewiLyODfP*Q4ry#3-2joiyDPIX%XjIgl
z8{Db-^kwoFzpG5$5yL$>%*;-@`oI%0z@93%loK+_D|}$PHc;l}NYX*Q#F^6)uRd|o
zx<l*xThp*UJEB%y?b#76z~KA4``6Qx=Q|El&kVURGZ}E`-!(J{P6PuFWhU&`)@p1W
z{ji+LYovRw?YB2)ss7Em%C@GyNpuT40chmqzB&_6-&Fx-5+<Cogt>R`-nlR;m&u>F
zFdJ-j@3Kk8e%Lx$vDUg?COeLp1Di}GjYaJiiklz9=`Me#Q)L$ZeYrQCYNg~rZ)?ZN
zd%;Wy;6cKDX<jjUI7Qy{^>BPt-XQ7XD%qryI+G}ppEdkvZJ`D-QwSl*mIz)aZQ1Yf
z4-J$2sS2~RU1Bpd!rd<V&lA^%Jw8-iJ5r$)w4*<dkkx7|`i=9)s-A~%)!!q__p-sP
zLh1f!^B(_?a!fWBg!QI#?j5>*M@|lzX<Dd?CDd9a&pRazdeYdhTWS90_YRlh*Q+vU
zsA{4CoOJZYYWzplL}8Rh;uYZ}<V3q_tp3?wVcZJMmF>kT1*)qDSt9P{DfkALIV-kQ
zN1Kzgq>ss;Ir5?%>|oUT_l}0RRi?4@EH!>lZVX+@K+)!||2%3EwaRmJjl@!UbRX&!
zjASC7{D_yiSdf#jJ~Hm8k?4;r5jWpWlPvk5j3%8?E&&S%)HPI@SnD{RmD=5j8KkPC
zfpfId9YNAtIkV`2EC>j)v1rw2UC_B?>pBJ`37q<cpD&31(#J|em&DTPDl-ETN+i}R
zGD~PgEIE6*!3*z|H#bpwOkfk|_QrqhO}qY;Gr6$vYWzcQ#RHDLPBYEa!nTVNTG#&1
z(w+tCN=nluRaqS+nxy87=;J4!ovYcBTI}YuZ5Zf#%)1L7uyEh5E8x-Pzto0K$|KJP
z4Z`EzNtENGDml3fhKv;~z?|<s3j9x}_%4>wUL#fbTv`>3!e28I{#T$J2%kw0EQN53
z?uh(9XO5}H&@SK!T<-yKMWA%8KY5@Re9(lhlXY)D8>%&HE0CJ5yg>RQx6r&s;nk!s
z-!b;#C%xV5ukqrjs>pe+dmQW+=LC^ad_|Zynm7aIO$uL}nLN&pjg4)WAx;#Qm%q@s
zd7pmNyc?7n^HP_Gc@VK8_jE?%9GK=>ZreLX%WGOeOk`!9UI(}rFYs|Xfi#cCW@JWt
zEoy22M|-kVkY0_{WPTi<z9zVI>KDurAXV)G0UVqm<#Mv}I3gol1g#<0m%a!dLgd{Y
zHw0<hOH7^9%QeWO;R%p94o2I9Nk$pn0djIOUGj@CaxvT8PcUa<Q+-m`jf?AC_jwkX
z?$a~2dm^{QtKD`de>hHTCy6d)9>{GxDP1tlsgM(K0Lt*6^pp&}VUjELL9q<TieMX(
zv;jOK=(+*b1E1aZ!Mie3dr)2@J~+|bf_}p()wTwFu{Hcl9Q06DE_v!?bIPI;kb%<M
z+#!!BxFXLo{n?ju&tkhXsBgrMoQl`v7jS0tZ-^Dl47HTRMF|;#IVQ(wA_ZyVdH)6y
zv~xv^r=|6sl_jfZ0$Vk}YPNT^bb*A3H48BFkfq0+#E$mr*(LsDZI>`IQ*mZsV(F+T
z!stC(@2|lybM6Mx&Y3TG?-*ald^>CxKK;>@2Z#^__BRKTx2!gaq|TR*l>yk~0_T|b
zt3p+|#&Tz0SJCg*{qj$bGKlOaGQk9LR{6#w@-q_q<6D~_@8A5L#QD+cy)!wN#cT4m
z*>{nT4XbB&?e}7GKxGD#6^3M7AM%ALWAd+Xa;gVuJ}UZBH+U+=ksGf%`?Sq6dGRR;
zjk_Z;Cno5<j`)lik-$iG0hcdBk_H)TW2Ev4kBW*MFgiuu|9&kCr0j&49QAxHLuQFa
z*=;Y~yUW5-^|zKk!ELdXZF>T9jQIO@M00PfsrT{g@;PuAOXv|J2b-Dxr-#2pCuiB?
zV_}~2j`ln7^@G?6tHLL-&;vR@6&3OI(NwOtQZG(XdEw_)ol395MKo)rLPlAT3#mK5
ztFl%$+!pI84rQ!KdTD0qySuN3oD#o$37RjHx3Zk6Ui3<oYI2{ZPcv;Tb?z6D#<sw%
z4TAoq%#?@_mt*HVB}Cn+hliz}Q)P+rJxI-NUFl0Q&_C{-(5ZprUPmU%oag_I$qSED
zLN!ShSz383YL707y&Yl6^urgwC8WYKz~NxFrjXtt2sqq?o;b|M^<<{zg2(E-wXwJ5
z>t5(;Im=q;<ENlH$I%nAR|-I#IHk$uL>H{tBN=CzTVmd&z?}!pl+^<C{ZtlDBa)Fl
zTVx1y1yV$PaTt1H@Q96qmR8Gk9hMD(2Fc2ekXi83@shu!h<Gc9U>I45F_`hAPx;OX
z0;zB77$_)NTBa@mPzeacW)yIC4=zfbh$~AQo0LJj9wA2L<E5S(m+Yn0K*LEr1Xwm<
zP%$cg?FXL}Qc{n()ouzpMpM#3A3QpnWn$`S=5-?Fe_`*unXhGp5IHI?of5p7K9ac~
zhEV0}`0|oa-vWoZCH*OL%*&X9Up%ZknE_kN1o<0p7p3L89}TURV>QLUmnvG@zhb^#
z=e^SRC5K5=A>+MmY5TFuL@-;J*KpN~Cj@54j?S~K(}AQJ@RS=aQ#h(fV4=y?WOk=Q
z$C&=r?MOG3&UMI9a#9A1&lG+Jc7EL9Z)JuaJzI$ogVnNi;QY4x^Gv_(uXEM7KoVkJ
zZU6KhNkhv~Nv{+oC~8xHpg41im(zDB$^WR+iYe;Bc`CHzR>5RTKV&j3#wX`zBg$V4
zp6ciT_dBX#$Jh>amzFk(bo|Uv$ntl}iM2vvD^-^-SwbCy!RHqKcj)!X;x6CH;jZMQ
ztujx7H3S0LdV25qfl#sO>dBaj4As|88NrQxj$8*}JiJCo3Du^k5Z}2e8xmsLx8I-V
z@k4J|gs3$>E)R@>A<6&R!+K7}^1_W-!~R*AA9Cb{&=D+WcIW&Tgq-|0UCpHrJ|A+h
z1_i`ogIfZLb@qhnFV9pvqfHwEv?h$$!s#xy1y}BZLv}ONr#!=8E~v)3_BCXt2qzA-
zxw)?+0+u{H`?@3$e6&kNv{*l!t?9ggZW>TBd^Go7d&a^RbQn#QW)|@Ja~s|a)uUwy
zS#0y0GHLZu{uPr77Yf=*9n>PRhMEogI)~F_9WP9zYQ!H-*7J$vN?s7}XNg|W`(S4!
z2zeeqZv`ial_<I?DM8#8jtw?a29u2!B)8_jslrL@%Z)FRMWa4V<HQY!np|O2;eWbM
zQK2W}6n`T}q($AkBFH5u?!4luuq;&g<QPo2!DskN#%cWLr(5h0KY4R=48?A^N9ZW4
z8BO31#I-MOY8`8`J)fy4msglp*=F;0Um3omJ6L@})4bCwFQ=Ocy#RdX$t7v>8wZ0x
zpbD~d$J68J<k(08u#iqpeFq{^UtBB(4?hrL&~I6H`*rTIt4Dh(V~=#4R1O}8+q0`>
zdL5>Z7b2QvI(8nYUT-oO`LGX|9jU7;kE{+p1Az*L4pe_QAU3oJSDJ@!M62a0;Bd@(
zYoFAq@}YTiyvys+(8rH0zeh>&r~wzE@63dz%+&EF2V1Nt)xV#s<ypX94svIV*RpUs
zqk;dC6nuHS%cM~Ju^RP*A7R%0W9+5tc>ADbciNGwG2Nk5C!ssREcw7L^lW_G1m;rS
zguUyJ-`BO#)lI^l*mc?y0D@9({l7oJlKL3J{nP=9N}eh=m!WQ}lrh^B!qYUbDAH(-
zVgfUAA>@c^a($hg8(y-R14Fme+v57nXK+z@VU%!OQ<h04vHo**9&x*wW4euZZ|tb-
z12A?3OX2>94KrZpNz81RSnB4W4{p{36(UF){+FNsND_aDb8hu@GWjZ9qdvTTJv7L7
zlqKZo%T;eoh#!K3ZRC!o%$M<XEFb)M(U*xMcZ>`3>t0ZmZ|G>zw)GDHZMD=MjfM;c
z95lDMl0O>MSZ7t(<jEjam|y$HriY)KNWMIL14UN#1eN*Xk~6{fdPw?G=dfB_`=nDz
z0w$hOYU}r8%4(<NCNR~yMz6!1T5#JH#5Z#WdT9H6;ABF|{05LnKI++Q!Lx#^Yu93^
z0SZ^enfHifvG&YIo_qIVBkv?UTv{QfgtMfrWox2ak*<F#CAE_9jq0T%&!pT%+~&8(
zk5!71nSv5a#OZ&QX{}OG{Ve`2P2OR<?M77Xi~YF`=iO)o>Gz$4g!dR4sXq^&d)Y3@
zI1gO!&7;%_!)L~)inD_XV$!=3p-a^HRiR~5akdS}k|EKF+hm0FfF|hD-(lG#-;z+X
zDrhmq46F53dG;@dED^Cg4MV&#Ba(>0<jb9ux>jSM#T5J=mkTYyZPNl{ij_<M%#@j|
zc{g*_Y|8hY3BW9)e)Q{CAAPHd+m$<@-zc@5uU`{xP4-b&F_?ZG;t}Mk&|Z<dMnS5N
z|C4V~qtnAE!)>Dd`>k`AGDj=bA7Q8ebS44iqog3*ru#$L*%g>NxWpGLnsO~PL8cnq
z^7`jWBFUZzL-X0z_<^9>z|)U<Y<g3Iy7=dM@jrC<^g`A2{B_4KBxZx3$ut}Jm4^z4
zn`jzLZQ2O{?jRF`&5*o1Kum&i!;8LO_2^1a)zxD#1inmemEx4fh9a~hK&u0ggcQ8C
z&-Tg}cY0@N_;JQOyf^O2`)Swtb^UgQKPO>C=uP{zhcL+@SIw-0SInPgo84RokMk5^
z?mKvqIB=MeV;`-2+Y!rcRHK{u`YQt#3BQr^oqo0+0YRDpzKWU8gx&GMl_PYcv;Cld
z($S7YN}Ks<6AQICW~RzS2R#w`AMMo3cvoN1HvQnAMK5aC;WkW4L8^MH7n-_LIUoV1
zE}|dz6A-k=3r7i%^1o-i#~t93zL^W?vOnw-l{G9oO_+~HGCh%OV3m0z=bHlmAYVZ)
z#80Bwk5s(eqru$Zedo}qLfM4fr6(K>=08Rgu(EavKNuDJ@g3ROk*~w6O-#a1;vD1G
z!$m5ZVajPEEQHh+S~e~-HiJcR*l=ta+pb<sZULzG*Wik3-<;TH(S>>GYOv#$NxX>b
z?7M}3>{J#jpLc_eQ#mjexi08`2K$7MH*Tp9-=Sl?bqlHTLCXZ@N`q%NOsNeA4^kA#
z<LuUMG#Wn@;WO$C?H4Ke{-j_Ynj}@7t=;dRGJNJD+J8|R>N=u%BPqOvayrHSbmXnX
za@xh*y=rpS`n_6zwE3n;z{$R+!%QYA^}i{-i`)0lM{OEjS~ElsxI&G~Iami}<cgw)
z?`b|>cQgmqAM}Jrs^j>i&a$v??FM7J@Jf~F3PN|<>70UksB>!w>u2V-KNhzoPTNn|
zoo=kOMe~}~wxT_ysJ}I~H=r*J+5AQ$2`u>Wwh6?<3M3(6oT<7atds-0zD_@&S)m^V
z7hUO4&@m5aI(~>eyxHo!y=c-Q{ajH0_XXnXBMlqi^{(!c0Z7@`wTw9<^Wx?aB$DR^
z8zZgJ?sZ!m8@a&)plU|TKaPf6i6x>OcaOha1ke2L%v?jamG6h08n3^-Bd05sAY_Q!
zrqm4XpuQ?8x5@>4a|ih!fn<Azt^l$@$wFb3+YzTZ^nNyJVH?qx2?7*>fM2DeUpyl;
z)~qAr<D#QoebytA1+zGW#DND%c@<$J22*TT*BTZAkZV=QRksY~zl2})_e75*ORUX}
zu<)bb$Di%IV2EoYtj&bKNRqhp;EF|R#ddb&GwWwllvF%0{}im(%%4P#+qZA49NoKr
z-*^}EOp?3OrUeFovC!8|S%@sMpcH7GH+SJ0bgP=ezA|(z&HM5^3c;_%>gvhcq!V+U
z`=sw*ru0_<tMcJ9Z8dW1yAhjB0=T!JCz%J6UpRhy2}T^oP>RGyzg2?3t>)z9e4Z8*
z7Q_Z;+F)!>4?kZtiY6$Z?iDeJws!iz6a`PaFBrQoeYjq$DTVjIX)c9}mKd${A=xzp
zJXL+O<-inq^S+zf9_vPtfK!IqLA7;jL(QbHCvbpcc;VYyi=~HHvrVSY)tk862e{()
zy6WMco}P}||4mmgXxHVnx1V{*@U;X6J{#^>q@mTKI*6pYlViD?T(-qlvIWEr@ov_i
zX9Cp4g`ritiP|8jl$wh8#Kb_VvMXlZdJ`?0D37~FfE^%C<aGUsYi^Yg7mst=u4s|H
z>bG*g5rppBB9_bQJJ9A#UB8vR8y4)|yet4wu{l|C!hg_}HPys>t3oNxef7062y|PB
zib>@8?o@J<zh-J;oQaI2Zkw-D7P?9Y1yoUEqV=Tcem|p5AIJ6{t>$M>pZ*X`l{)@a
zkGhfTY3Dhg|A-t&XXu(-rxsQwLwjnPy+}K-rCKDy-qxOi?_#|o{e@1Ss1e-FbpUT*
zzOXL4{{ByRsVOG#X9xoolU!QSx8dqXYbJgg`hy1>1J7$dyt1?;T`T6i&1+|mN#RZ`
zdnC*X%#WS+5!h{SDuf5*o+aZ?hUjTpC(Xi!UNziv87)Ky97JCk*37x_gyU%?Li{=d
zJxD|bU`)r2J*T^o8qR`xE_yX*n#ZFJ5oe~Wr6PXsxz)Tl$g3ZmxC?7y5by}3&&m?*
zj>LC7<W9c~y->xPT{D``A18Y<^|-JY20BoKIsTe!0(MH6ux-g;m+l$Ajv$GjEYzRa
z-#B>T{=2~u0KRNz5~CA%e9TCb$qB=qHpibFua1h<Ppk_&Z%KK~xGLt1VP?7A`_#Pz
zq+;(H!~C`vn)z+eBW7zi&i!^5Xt%Q6lV%qxJe5`+{z|(R+T_N!q><6_y~zz_UYECy
zyeqc!f_6#;Dy}Ifea(YBB4DC?fVH#*eDD=o%@eu+I^ZKiZijw#Ck{Tk%}w#sI*H8`
zz94QpP1ptca+3XWy7FiA-1CSC8m&dk-$NNoZ`Sx<Y)S3}0$2NcE@|G@jbAz%gJ2=j
z@Yo!fJ3%b`UpWP|qRti7z(<{1d9gx($kN=``!vF0ye5PHh0Ee)6mJf-TPF|vH%?(Z
z0(jiWaMf3^N`0l?1->c>s2#hpqHZ-a)s3g-n&B73;C+j&@p1kjwKQt;>+nnw_2+Zz
zt-j8~EGeOjpBta~8v!;^;LS%s!<WkAMTEWDP}sz+34MYD87xH$&t$faIk<KYC?jo&
z&%)E8P5M(JiURv__okTb&K@W0SGSj225^r)r`6&aFI22qW(ARQ#EXslg?(NzFP|7Q
zn1yub#{nrKnE81YP%>%>%AA8r1dtvJwdX1a-?`6DvK;(vk9H;gJ47YQ+(!>=OZ=I1
zaSu*vbVfB5MUDMt`_{auICxM@;+EO}cyG(}(HF=TwI0$yZ=!I(NBGtKuKg7b4c2_9
zZ84|ti}M=tmv8jA@DXH}i4HL5mvN(Do~t0t!86}-I*wp%o_)~TA1Hib2z%i(-QK%)
zgK7web*}&+m+fvX6~C=;b<^l@ZNq)((+nulzO^~Z`pO~CJFE}kzmVI*7G`*VGg}5*
zSwGRSsF<wa`RQCks_NB`W1=G4ud|c>&+q)#5I|ym@#EOrbJ?*J<FhmS*>M9>Hmx(|
zqIa&Pmj-lI(}<*8s0Q#1GlENo9$0^$rw{)3D?xd!el}jM%BprvfzCq%`iYX~Z{E0t
zz0T+eWlngbcCu$e1vJ*YFjL`Y=3&|d0mQM5Tji$mJx0cI``n-*%kn+_nz;v@pbq5Z
zix->e^C5i{Ibl}c>ix%;1HGScO|)$sFD+8%h7Z~@+J)`*v&n0Qg&Oz#w(T@IQ1^uW
zxxa&wsP0()E#!|919DK7oE_=Y;TjssYY8f>^%L#)>FNJg2k$!Rop@V?b>c|7F1Fv8
z21Sc@504BX5YuR-R!C7w(4LLNvy=K-P7@kh$8}TKmeXbK(wGD9`@eN?62Xu9XUSL7
z@Eq9GfcH+NpmmMT=o0zvW%o|7)cLvpLq5`ee{--q{MmkwKiVyGvcew_8r~vW{{8#u
zT`FLVkk2B~6m?POC4US}<$Ckk%tu+Hl$wRQ6=gSH&61QTy?>}xy+==9zPf5d0eE-u
z0}bqwDVkg5Q50-}A&Rkf@yx#G_4qTLQMYz6c>YSeSl%a%5KGgFgdO9W&FSK|xuU3m
z%}@K`&xD+^n;k9~s1y^K$Abo!iSZ^dw=^a?<|4pN>vqomZo>vYEDqW1^$Swv$)Cg;
zTCKw5IsPS-F_^<W8%NBx%R^?f3LbCAx%IOKoBG+<q&$(3;If&kV*@dSiShaTYk7Nc
zgXUqwN=)qMZEsjkMK#+LR4ekvGXQDXLm3Fbw@TR_F8K8?K6^c2ohao-dd6V>N2FO^
z&iBJDpEOnu*CZALq957lNi8j>N#ges!E)lnz^No~o9xt&!Xcvw#366E_x^+L<Dcwl
zBzv$+XL3t_<gWo|LW^bp63VH-Dgj@TjjBI=YeMFd0rOCvY$xKD5zT^>o#u2oO|HR8
zYHAG%A}$*n(^aS4lr_ykn8OTADa>`My>kD-zkLe9pUBk{b3?;)5x42evc+*0Z}(0t
z$qy`WR?ra|<TFzB?dlzmoCLSza2$MZ;g$P{y&_*WNq!?8<>prJF|sYwUhexQ`S@pA
zpA0<kbfg2t4i{-0WB?=FeynEIta#C^D@EENZ|{b0m#e>+G4uWPIt?h<gakXJUF{+h
zr8DOKEoST~?#$IaE85pm6bm}q&Jo1#Fq+hKx}>8XB~ARcMfabzcon0B{EobJ4(0-_
z7QD@`BZ0&C;>QX(FcW%hM&1|^RGV{fz{JhbyvF&|dmk0lvoHdNa(eOIpo2_^5c!|>
zIXZ1a!*U8`ARQ}vDC9m?_Gd<iy#~4saB38dhXpAPhXs;hSSsoBA~iwto}^v$hL;aE
z1nh!Mq~{7N-)^NjNAIAJx^wsrR~m35823@(6U{m1h{f)}Z?T0fd}H4N%l=4L>s=YQ
zqL(?OY-EncA|J9)3T-UrZ~{y^_PwPbz#HN0?}VDtFeF=<bDe9q=YHr?J>V`FdP&UL
z^RLDXtmB^p>zaR?^49CDzFUgS{{RFJ`b;GK`<z;~k=!?O5B2b%T8K}@FH0yi;m{g)
zr=iB=%#h+Th5LK?f6}j6w3@kxDcQ*BG=a2sQj<1~WM7F3TzUQ4HUX>&oM1V+o9kC!
zk;^hkh-FaDN&7+X)O?T!dEJiC88Qqhf#?eP<umSqSi6+T5^AkkLVGnITu$%(NcmM^
zy+x`nwe>4Aw9L%E6rMtRmvZ<g3L{lJ?XCd;Em7;y<@FpP9JaiQ>5pTbV(DV6Udb>@
z4Lwlis?`W5BeVp9T^2|FS(#Yrj7b?p2mJ^9WBL`du*+*57|?rw2y#N#7qa<)CZ92)
z>8M$~3!|0Bs<@=)$~EVw?Zk;aKVa9{`$Tr5Z0liIOf+O*YIUYmm@a9_F7teCW?Y6F
ztQ9_X{cWP#bcuQ*@K)-f<WW&_tMm!t_#>h+`|a|zyb&|H(7%w1`NqBZ#^8`1g%0s&
z&scD9|C0XHnPxlFQv(tq?Ci&|VuSlci#H)Xeb&4R)u<}Ic_=78dc09TTQ_aeVP{=d
zpFofjuMRX?Qz%)OoHwej6u{l*jcygORl~COv&<lw!5H4fuLyqg)a>8usiNpuO>~}-
zo#az^{*UGkwR4o;=I;()tccVp^@FjskyBQUg&&T=9j*Vmtww=~5RX%SVIOo`?^jV7
zt^or~(&dJDrzfa?%zKu@-(yq}3T-XVgdU4W$n(l%sc`@3&Ko3Ks<8+aWc^D{hSe2@
z#b5y-zEKvS3greI?2F`uA6kZD#$8ZNJRw9;^iy8*Guv&1<6dQaTp8G>Ka{2^V11l|
zhSARk9AvyfIV!!lwVJFk;xtuIUg+cOGT+LzI+SG4L~>ecf2%!WW5;@PsloE?uRa-)
zOY9#5(K|8(L>$y<u$CY_yPToO`pTy{#TA8scMrpf>}SX{aE#`0$lpxw(XBkApv5(=
za2hx-G1HGLjaRB@X1_$82^$V~&;ocDS-)44Jo(((Pkw(tg+edhM|A&hHW-5;FV-%T
z$+Co5mq;N!zpE8OOex||U#{J4Fs^~d68{2L8sMNKV(HzLLhFCKQS~#zno7@S>s(DX
zRFlCjN5D^z3)$b~&{{E1PydgRUeDo*d}I{$X#46i*HMGpoPfK#anQh|sJY|G<{`mq
zH~6@_a4B72lW)dpjTw<j$=CL8!xkKF#{BW&O9zi5`ZRD`+uJsgmd#UvH|t^i9J{Ie
zqoSS5KXLQD2%{hC2=eo~kWef85OKa~#^+nJg^CxcPpVO&YA`pkaV@P+|EunM0Rs*@
z<jJh-Xc<_)hO#YSHZVt|63hTtkQb&UwukNQ_v42pIr;g+H|uP*YTCtXH~VBR%0v%h
zC{ZGDd3z$Wx+X6Tbz`Fke91Jfj$Mr)+D6whz`3?{)Z8naY`wUEB=2%e-G<JK%O<Qm
zJZ(jc;vazuj2*@?-XIJTe28Hz&9#aYmh!%l$N#5mp97EmAMo#31ONS0N=j;@&%_-q
zxH(rU($>^9XMZH)VK=|Dy9cOiA{E&J98`e-tP;c^CB93v8<8ESpLtdamk#_5uC`vF
zVHEl!t9E|Uttox<?hNh@FUQC1rCmI{dv|-{#)Zof2NlBE%F1dyTT_};s6PDh3a0LY
zw3QrL!K)s8wBIknhCoWVg#1j&(wWbcO}f=&=al>Z;{pgE1MJr`-EvTCVre;k2m5md
z=6=faEpJD6nwL-adiGif3Hp&5rz03V4IOiRnP1!WE{|j{+Ae=Gwd&`|Jpp79q>TIu
z$Ui~Z%c~GTo$R3}pL6tafRG@mCMRHiQ;c*i=HfoXJtLuz<6_74C_HxhM!$&TcEEbA
zVpGbJuS18$;noy~1RtMf=XscsGH2$(*%jiRGho+?L!z2I+h4jZu+PXQ>1b;w`Usu^
zY4B%t(kjlYUm6-G73{~5_+oNvBcY}GJ;(sx^Hvxr7Y6VhG%{Ii)<OCiL=LuKzhvk9
zfx&Bi7kj%DO_^*4JSO*#h#%YwoUu_8ds4A-Mb{&62>Tnd_sFzZuhptWrB6W1V}$4M
zWMO(!Z(IAKl?=#m<ae%nv`{hs-wvQ1N29YO$TleQr6K~dO5CI(jG(RGT!D`lMv7sm
z)%^|Dcyn6UUtvb2FlV9cWy{s3rlujD{{iwuq~!>q)9l3iSG9Upto5CqC~&5^_=A;6
zm^<7Bb2fwKPH6Gjm^%%f3%RA9ML)BXwl<6%C7nbKOlXMnwk<bC1cFCi>?Tb%=&AO9
z=`Lk`Ua_!wuF2g3?>y^Y)52s2s#K}TV_+6Zko21ilw=A4b3Mx2@(=Z7_2AwHZ}b3p
zYcv__6^|Q{^8Z^bIk{c}=MC~^nl4XUq5cWbfmB449v%mE6CegZdV>4q*G$oiB%C8z
z>7IZJ#LLW3el5ER;vym?V`AtgGYXh(+}POIvg|v3eSKx>CN$}|u+zKKUVBfVALKL#
z>K?CFM-|}1U2huKuM<veQ!6U}_VCW1qLukX5D3Xf_IaLOrLP(848K}SnGhIZ67m3}
z^pPyKu6|QPd9lIt84+oa!y_94&}M;RuqPC^;YE+`(N5&!$V{(4n!OPrnKK}o4S#BT
zL2<eF$dN?Y!X4QG=cfgn+PX#6G!^^bg$s<;<PcY@>Rs~7+Josh(*M%RtrDRQRJ73y
zNhHVs4H+VNT!QRut-IlUo+b#ysVaAV;ud^<?fUsW&`lZ|f0G-%;dvKfJe-`I#ETXg
zYp}9?N=bZHvpVU<X}_Rk=j;80hK3I5y)J=&li{`<LinBdoq7M~h~eQL#XXJK4t!q{
z&nfFn{pn>040k@^@ci^Cifuk2N1~njY;|pYYTnZ9Y^r`kfvg|JXCl@VrL`X*4mGyB
zRtJq%0!3K68n-`gs%WVlM3M8!|MD?_-|rf<bK#M28?y`kON@cK9j<sbk-r?eI4*~8
zy#{7M@t3I}eL1x75j@{Kz@bBc4-s8<z)Y<ug$uR|LzB_fy@+<=pbez!dJvrVb74ZI
z9*qva@nZ&9v{o=hev<1)e$dw{C<#-cK$~<ywnQ;DM!YZM1wgF#l)l)9*U3TGHBn_1
ze5y=-yMNC2LjNg8y2DR7*RyIE6y42WysMFp?!B3tjjto?(Lp=$jmUcsQ%xrrfTu+a
zO8ntrEy;cG_NC-=p*Z(<&hUH3T!tx5d|hFTGM)!>K(`Qerki*dbRZOgGF=i6xQ2R&
zspG*tHnBs^uZmyLMAU)~=76Ttp0c`f^}%~C`Bh#r29&uezeqsinCU8Xi$Sj=9EfiR
z9n#8X2Wd5i=Yf`%T7`?j1kt!Mwp=D=X5Doi3%S*fpFay8;ZF4o4Q&qEMmLiuKf$f#
z8oxkH$^U2s=PmaDeW@%8<FulZr)++L4gg$DNUx3Si(~Z_jIboz{Vxfvhp&3#%1GOC
zr%N}O)FtV1G6?)C64?(_u_8Z^S#)iTRlgmSI?OYuh6qW8EKKpN$m);f@|e6qhMZEf
z(6sEa(lcs##pigxEJY~&o_5W8%O$?sN40Vr=!$#QWFST>!cGNU`hdV`AAV$1|H<1<
zu_UgucI9+W&4lIQiGNhV?qcw&3oza6p{EIB&9mu03sEe3rEEVB4SqZy@9bo|`5uIg
zP>+W(cl94xZS%Q{f<RZ*Uk)592{I%=PmIdf?6UKN$xUC{9hT`_W|kUy6Dp~J_1R|w
zwsR%W(iW2vs3uv!8#%QE>xqgDCo)*3Y#HPX3YBP=(kOtYe9IzlX7So%Q@39J@9r|p
zas~pq()(2Ldm&g~->+1dsj0!0jx7sdktDIuJ#GK4?sF|BM(8=01c#S<kV4N{`x&?5
z_nsoZplDYWUFSu|;D2ZN(4Jw50l+a5Cwi0x6f*8|Hqq+J_R})U7PJu*<mqh&Cdhyl
zsPSf$VRYHiZC-JqgpN>5wO3-};X%w$Tk>o>|IfG1J1=>Bu6@o9?R#HRWh{<Jmq})q
z%JWzms3836m18{$J;b%(3jFt$%$RRJ^t<>JwSafdnEf?g;EVlP#Uy%z4}4_S&^FIa
zw6r}B%nRA}GYV|=-$|e8z=pQLw&z*)R}R=7hMjfCouK^VX&gg!-|WAAU|<@C%$;LH
z#1WlWz_0z$ggkvg^aZ-L^(^q;OwM#wQ<{2u-ck4iz0d)?6zx9ePIj!~&i2*ftMbRe
zGK6Tzt!p{9eKdlakLM^=4S2|X4!z*?)PChqyO07j@<Pr=f-D&v_#yOR%yWxFGcS`*
zdPe^TkJ7#LmDZM*k3noRC&KQsFus38(?+<fwKK(0K0LC<rJu;-L8r#c&FRu?ZT~tl
z|B>TzjVEySlL`tzNN+mvBRYEejMBUPnP|~mivr_{Pf|aK=B$6NFmXHmeQl)qbbd))
zknuUFpzhZ?+q~z>97FCAw5pxQUGn_+RfF?lHs&MsJyv5)#e{axQES5DBBCgy4L>4b
za^b&x9ACT#SgN!@1DeySJAyG(5`@b0V%P*>^&(?Jw38JUu(U8R$`Hb|Hvjg81c!lJ
z8}6IPTH}SY`)C}S-%{6k?RXRg^f$r&#`(V<e=WOCvr}q$;37#oobEV2>ptbk4l*A3
zdL8#f4R}vn+(#v3yOZSN?LP`M;y5;nXO>)iP#a8s6cjTfH48usA%j(tg>yWZJwIBB
zgRH})rr_4vB%Ys%TYy*8aap&~=7*`?4w$N+C?0`99IJt>4t{uNUj@ksyf1j6L&A7o
z=Vc8dU$!b-;N|l0cy+y?wMbJtjegjmva_<R%)u~B>te2iK}Lp5PTI0FKQjt$cci@a
z^Yu!!lPD&#a_NSH*G@$Z>qlf0%E}`X0&$8XF$E7N@O7rMz_zh9SRlc(>ftdF`DErf
zH`Rt}bj^avJb7}A3#?~qKU>D_TUuJ4Y&SEAe*Sok&CkKWJPg0P>E6j|zpj&9;~2iU
zVd8fvwR^0lX!#u=rxmfAOjlE3?IXo^=hw<TX*i^hx)!TWj%0jEwV?^40-O6*)&-O^
zB)q{aC0~N7>G_0aZ+OKh{J2i)JN{x0e(Sb+#!K6t%}9+Mn;=Cg;vcD7;PA#rcC-If
z2}fH=DWU%?+{Rkr(@}ijPN5Ujb#Px57$&mujj0CB5_0@_^7duJ=H?$`veH;30QoI@
zwiB2eY6ciP;slrXEXE8#bu-j4j=9WRt`lfICP2%U6cYT{3DmYkc>U}Ez)D9B?{g+9
zYhsjZ3S0kl&`4Ke)n(;)x8t`xf$0wG^O3$&w1koqP6<5bO17q>onmCCI@el%Z_h{5
z92>g0Qux`J@=pNSt+NdvNjGov`ta{i)6JUIE#n}*U874s6toqEJ#?IyF%>%)IszlJ
z7{o~RMp|}WRkE@773rl!H>arVsW&i3xkNGh+^>QC9W*;3TYpdQPS=<)g%GL=RfBd!
z9DmD!v(HXW0>KE!cltjDUcS!;^4UH(ek&!we^}8}hV%Sy@P+s6msgiv#nS`aCpbDi
zb741&2v*=WYVK0wS#<3=RD&X;!_Mw<Yfm)-FK7kBLbySg@-3C{$=?2;d8F&UMLnAQ
z)yxQ?au-PC&W4wi1Qsuj^f8^$3;ZiIdg=PrA3WonVpTuWymdMJ8-99ZZ^N%7C%0B7
z+9Wu8Nf<LXa+~PRLNM{@`FhK9#VT=~c9jj@twSgx&#Tp)MF$cD)#QZL!n(gl2U~dw
ze!2G*I}r&%j!Lg$oW)t>CT{Z7x4*r9io0lOV__({nN&CY|8eyeP*Hu~`|ur7Kw_k%
zbyPq=6c7=l)In5Q7?AFemTm?x5Cp`a%aOb^NY|he(g;I$Djg0T?;Y^-{r}c`)>4-<
zbMJ|}_dd^l_TJ~Rc{>NKEi6#VSEoF->Mbbr<fRy!c_9>lnD+_*QK9MKq{Z_GLasw^
zcRJ^M&ecuLb6s+?8)EpC<7uL{*_XT7__o$_f6cj5V7qGJ1MO&4u6#t&po5C{m!A&&
zej&FDWG-d?RJ%rL&*{9l6?*;rN8>;_^W6RQHkIS2JNSN~asOkxwq*ajygKLk4!Yvw
z4*e?e_u*2L*n-c@=>&4At*CFcFU#&|%5&2%`DlH<(0Q&S1R_hh5Zm^HI@bk+@6qd)
zbSw_#$}88^@^GGaTfv@nm-=E&1;yx)GiJct)Z$gZJ(h|ezIz_4x9|7zI5cid9H*|E
z?^60wsv>P!;;H7mQwu62_%F=yZN-2<;^I?e+lR;M%Mtlme9oG1QaXCxkg#ycoO>1t
zm>HV~9(ro(Ndyl&nC~VrRu#a*eE1NS*$aI|n&;p>RDaYrZ<J6oDQ&CK*PS&=t+2d;
zD9c(9k#t3hwvK(h??AD&X9~UEK<4(tuA9Hq>|>GKAkBH&ap(EUg`8Fb{1Q?3G*_gP
zL;2@e>AHv4m}kef=-4-|SJ?lM`(pvx`d(*q>+ln)(A|Qtw*AcyRpuNJK<FyeZ)not
z{P4bmB&|(j(5~0z=<)rsW&_pw+E)Wq#h*Fz^8IqWJrndBTHijYbX?xhWYlj{(yHkP
zsxKhsqJ7B$=k1EzinXuib3l0phWgFJ*kg(UZ%H-0kpI^a(#VtfW^QQU_30A|P3huS
zTkArb8Gv-%V@lLDzqf|SAy_#aOTD#UQX#hf1+1M9ozueU!C6e26uh4x+qRdRG+en<
zu0u{1O4OfB@3G}}Hud09!t*^J+LpQUHZ8e6sFq(6!2X?DFsdpFZEvs@^cdH~#i9xf
zA=n2sOkAU6^<Ef974Nt70Dfgs9z;j=6Z}&4=F!VXcl`tM!_|=n#l`*uXD=V$n)V58
zCZcU((L{0lRN*A_`3lM^pzZrrj2f@?df(SB+G%!3c;uk{HtJ5cb!bu4pn?3G+B}(4
zCs55*4R`HUV$~lt7r(60SCv_>JkL0P1XtFL8S@1{y(D$ZnG7(n4T1<48SCD)uIG6F
zg`4NaFB(hds>Tx!N0xg&``KwhNZc@7Q-H$4HZ}#nQ<g(ke)4MN`Cp}pQV%~_O(@<M
zLw#_TF)|NmGx~gJn;w3<Y98;FgDKO%m+^sNRXO}g7ABJ%rU)6TT&>l2l<a3B<opU8
zlHr|a+HcQfEz?crOT%{`TnOJ+|6o376j_xK{@9trbL@jqIMFMJ!tEO9<4G+j$P`Sn
zIaxvk_|o>mZ$TM$Z8krG05=<+7udM=;eL76Yg4)kX_ARYE0ePCi6;Ckec1=aMKQnr
z`H=ySQDW;e5CIukM^ZTJYh5z>l$iF`P=~pntN1KY_hn#o=_~(bVazL3^1q1@{z)Lh
zSVFGqK^^mu4|W;$V_2_>n`=z~=s~&q0~#d=en5|Cf|s^s%Qe<;GI;^YApEbX>Cg8Q
zyj-<$PS_1rbXE751o3ThtTRHG6YC4=OuPoYTF$!czM$}U^Dq^(sM1wID!Z_3THZm=
z!rU*<nzBl4RDB&ed!c4Z=Vqu)aLNUhrXVN{4Qd+^6d%%maQ+F0QWr%$_MlN_3Lh{*
zm9*J*DVedpHn~borSxc#y)F`LArMaHNRH-5&KRJ4M-B0U5nvuuq$Mz7;dNgr=H_fV
zgk)XwvK3=)rpchAmm@s6*DxTfD*2ND$j68vzR~@WRm7{er7N@9LcT}})AGA&=2k@C
zkY~qD<ZBTW;$*I*r9xlimy5ME`4pf5iQf{1<49?|-_E7;CDU<(FE9wk=*94<@vC)u
z$lzvez@YfkAhA&l=%=gf`Ju>;8;t~x86!DBiT);6VJTk&C(l;=ijyfkR?Uk8#j!G1
zQ;0Ajx)%hwHzgKO_Fa0C0_+9x`7<f6i+R!68hBGfPI=e!B3>G)3Z0oLm^j8o3~SCa
zMfSdgJ?od`MqV7Atcs^{LwTY4&m*Kpp@4I~E{cP6mlXf){vxRpHy;K)<9K@^cKqUI
zm6?CWL0|E2r+sTak%uUM!TbWyj|X8noT5;mKb}#K0Js6+oVVuiSpnyHSp0isTeOu?
zmN2)}D1@Fn-EsMviudLH;OEW-{@8iIxH=T{*X30DU~|;8AiZ(ZqZQH%z0$#Z7txAI
ziU^Q@@3QMp!&(!3zMNcKOk^xQBL`!B{0F3gaNQj_xrXn*53Kz%CYp@YWW4W+WME_?
zKV0QOB*V@VO&H>yv_5`<({|ItL;}iWUF_&*YwbXrP<0OIPr4zbVf6k{v1Mo*?VGn&
ztOd{^iYE<(Y2aB^Y_1}xVI#y7-18otq6gB%@W`@ScLt`d*4ys}bF@F~)YRACm9xEz
z=$x5leqqMm{MMC$X$4qIbiU$G8hUPNQN<u{yO^43SMv5}kj-~+bT}QyFH&Nllnz$G
zVt(qSNCXB2Sx+1PQw7q?C?#UY>_<Jc6BGc~o91;eu2tlLM%ByTpM@f8Dr)PlYk9@*
zlEPwVm+5hsP<GPr%-lO$mv27nHF8jWcf$y&5spDnx3qM;8HrEo?oyj**0?rgEt1N(
zsa4Rt8_Kpc>(#Mw(7<y#LrkT8Lc8@ZYFS}4^MY3p9nLM1l$=)lo0isVtzG!4k^;HJ
zSIW$&<qy43trcT-6Vq)<?$z<+$e7qzT;tnZP5^a@(hN$_jhp9nS-Q%{H$-KWv#ama
z*Kb&lSq3D4-;0Fq?tm%EGZ*2{h*Ao#aXfk2Jiu|P^lGl^D8|S^z@E(UD#TxDME`xX
zP-L<jf{SLn&F&8Usew{0B;T9AZiyX)oFKA(MkvCqb9TZC$Ow7pkJ8uZrbGw>*k^7z
zUm9@gN__VT*+il2gY5VJ+1h3mx53fUd5Q~|%EpnUu}V{wjShWp6X3-<v8+a(BoOQS
zzOsDza5CShdlAShGbM?wuI$fTcx}$e6iEdrmdxThJ{&6&-56c*g~CNII-A~|E*k>0
zNBknfdIIl(m>gC$&Qwuh4j<*s>7P!1Jo&3l5_D>!Kl9vw=kK}dJ9YDplxh37#mMFT
zS!NY|OZ&TL7^7`!*PZkH0if|R^T8%zv#6twJcgZzU0}tYg=@nmh=DoLcBW;@i3fm`
zK9QV5EPl6elj}6LbZ-dG-sY0RQ0~S#@S0fMeA~~4mHzcRI{$`tLJCzdFJ17%uS%1*
zzO@QnPuS)=xQA@hpEh<IAKo!WfQRf(zx~y?ReNt16+Di(^VLAguPAr{@E^NaGZyYG
z(sh(z#^e5@cDsTiW<R46;)t&JidK&3<cs4`mm9@E=rAw7CuW-O@`NsZ_``nE-N9Yx
z{#~9xxO2rZ-5N5L10CIdAAie8G@6SL?1GXQNkcV6xv)m?3JPo@5cf&lhaAnL;p!u*
z8QF3Hp=6uVdFHCXUpp-=%upjqNV?)7S0_I~Jk*SYS(L)dVA1Tu->uR;OBGSYV0I%E
zrsQd`wU!j2aS@B&E`~{>UgDt821ci2wIW&rF9#Pgok*Zb;X=~@Q%h5LKw5NqDii$P
ziAwWN1<J>FKYR(~q(|b~Q=>$rUbVcBiVvt&u%GrM)Bqq5)93cBSMDT$pv+DdDsM|T
zEwPyS<D*&sNYFkD<Aw7=2vcLz06|q;#DbSM9b9dcfsWX9K(=e%4TH4|^zN~{Hvby=
z;O4CO7g>|W3w&ZFdX>j~-aprjgTumk%M@MAc3V%@tT{ts_(D6m-&qE6&p5J9Gn-vT
zG=*(-D)@{u=l%+-o>?$8!>--ZRcKzaShdlgHee+iI<`&j)1})3@r5#P7H*wj)11Zj
z6N;YU;XONjzjf!tArI9Db}i5m4LT*4hWcTMbF@an=uZG^FMbP(YE*`~@1ArI?SF#0
zwh=YCGQ+YcG0z)zw`zjV^GY~D8QwzE($`1|+zx*H3<4Zwp*K5%Jz%wEWGfH`f|n+D
zQ^IDi4l<o8FT5LG&Vr=`wB9t6p!Dk-58{I_6WwN#<%;h2s2U*2*YU2964At$K-Tu1
zRnH$gvKv%bez3AIgeJ{JO7S{I4k!ZHljGhe+<`c&g#l3l&jyCybkVA4yF)BFMl<xA
zCP)qnd@2DqOTB5Y3e0>LLz0-8&1no!T1cit+j~#JN&E&Sh0^Co`~(ch*9wCZGC+Rf
z&wu28oVD^|LT_zJw@3qvfPgwK<T^G2`{I$hGWpuei1172v75kyXx8YT?1t><6~O*n
zqz20O3tvR57fhB24Flt#<P%7&C_Qj+0Ve~Jk2EwKKS%G;Nc=Xu3aq&IVI%d^`ZJIk
z_p9%iTD)Xbg2KB^wgd^npb-e3;jwEfu<}+C6Xc93;Jm!gNJg)L=Qrg15F9h{a};<L
z^?bGQB0>a^%mTB)OsUajvC%ZkNHK1}(QQ;{+hq>+l=1>Q`Vq|iESeWk=TVF+5)i%O
z>B`9Q;n}b=@D{ishhp|FpoA7IQI(?};ZA(hOW5iXY2G6seR!yyu!Xfa|BD8Q_9CKR
zKf#Uhvsm+}WL{s*4{pu{ILTG>6Ugn_ixhyf^pRB>_+37$y`Wo}#UmtLOWEQbb=8Bx
z!Dn447-nN7vKK4`<yIIXUjrjHArVj>`4ohJ_D5~!m4GJC!bLmKODD@N#z0XWuqAbT
zE!G4v)*8bo&w}Q&9lw8iYdB{v*aH^A%sSx{6Bm-CFcK}PpG0b2wwV6v@%*(&Fjv9N
z5s%5u58mGyOf@E6drK`FD{&i$;Y8LAC9@GdByz;mWYAl;ggy)amH(RSL*Z}c_$-aX
zYSu5!UcKf=0Z#%jCZJbxFuSGtO5$8x`wuw<fTL!>p*((DQ8iZ_?To^s-Jv&=BQ>f}
z@`T7SnAWy)Uy+}XsDl(82?|j4g^RbOe}fVtYplJ7GZzHK%-e5&qa`k{kMeCv?NG<%
z;tV&7PMM|v#rW_%frvq)ZXxJKUd?nQ;LtcZN`gyP8ltj=OM8VyUUgqWI8mWHQFscj
zrHF;{n31zHFFq#z6fV!CE8Nw-B4hdD!RH&i<s%P5-LI-3XB7M>m<D~Yjb5nHxdyLx
zz(tmEhM1cmN{}#B%A<%8Pe~>o46`+oJUYs+HhKxJ2xW~TXRgN=36A6)`{Q(HDHG?P
z80CRcp3%c2^YF+TfED)Ju4fEI+QYW^QMhy%S8JaW)%RH^zn`nv^ZlF5#x<>@8y{#r
zE}}zg^c*tv-S0BuLQkFq^enhKzV7QSKg(xA`Tn{hH;M%$ZKU+)qCHfP?uae!z6GHJ
z@yg?E(Z7kRmKMMP$Z=A;kewCOzH`O>#y_i|#`w(6o3VY}X|rSF6N^eF=;}6jBtg41
zE6`IB2!$OWoe*_}Uc8cW_P-01;O2`J<E?uZ1O35(WCuy~?lR{`cN)|?)w|*t^hbyY
zPV)0fXhIFwIx{EvPX2kHeC)gV^rni{0!{*|q<hj6C|9^8+SG~9>_Mhdpc_l6?@>~P
zE_i()Bu_viymk@s91tYwqLuX)Nvig>a)qLMKzcS<2IEkPc^(``Y&A~+DbHXYWEj1<
zVkzF$>T!B~Is7^cd{C#c^qiCPX2C{(Nr2xO27rcB(&OQkwoYyB*<-LbpDVogW4wYN
zO`g^Z->KpGg&@>BuWvFGSzG32Mzi@z2bA8O2eou+Mw!p*5tc`z*ZlJ2^`i-{7TfJ2
zmMX}=5T0_J5^@3NL4vh%mg#G+kLk6V2Rxv1PvUYF6qS!hSGHmkiHLo8@W^M7k<1^N
z`uVCukajlF8cBJnX)XC|qI=kAiT|f3xLddyDBgvUeKK2*f&87X1V;pAE+DOr#I!t}
zW3Wx05~silZ4v#>2D7etTD|vt-@=s0b&f^N@A&D-K%kSExDn|*tv*1F)Onu4RPQQc
zY!<?)613R+-7HOXquB!GX^4E4X|1XKa{2J%Htffs_Z2P@BdU0N`banF-k=R*!9~)W
zD7(o#zw{<G509rN5%wHJhjhF^&iHxHhgXJ5o`t#P#z*TG;@QfkyLmXT?(XIph+8pe
z|LzGI{KWB=Lgw(p;IGua>`a4CjGUaBSB9@3Rw1>;^ySs4>cxcSi!9AtE!2OXJYBK%
z4-Fxjp-dZwl>ZR>M)MtDZy=C+ZG#0J3;02Cu(62W*4`Y6rvA5PGaw31Ug@Bc`d$fX
z7Z-LFu`vUMdBz_qJrbmN_~s#IS;-|sNq1f`a@z|?S`dgfQZpkppA=qspHJjS5ZuoT
z<*jw}ST@0)_$^=P(3I+4VfP*l-~&uj0!rw~Zc<6``3mMuQv9dBm#!d8VQ0yVINSXR
zJ8CWpFNyNKkJ8Foa6^rzgPR(z%~1ckFptKu{`0o|GA7(<qlCDV`2t|{nwq4B5X`AT
z6b&nL?-UJIJ4DAjb;jT|aBwu&D?|OcDlgUA_;@^|xtXjQo}C!TF3UjvWWFmSJx$<$
zxd6T@$b@Hq#?*0^o|u1of3wO1ac}x8>z4t^&?UK0kghEBp*L_J7_ln4xr<qVKB!tQ
zT`hp+z}*h!9tbv+674WybpeZP&_TX*JNG(}U{U4~-b(z#Tzt6pSPtMK<jUr1a6J1x
zQqfpj_Lhquu~C9S$XU24$LhAX>{@Ee`sbdC`ocS-?e!1Q`+I|Ap+3^#`P4wHhDS2w
zSYPpgT-6SvZfR8eHvS0g$^eoI;$;+&2$bmX=n8TM353~ER?dq4J`c^pniL}W6t(RA
ze1e;G8A$^YAi8uCVwVM+J;P^++K-YtrJt^EotU&tP5vTs=B3%Y60OWA)s<i<1jwA8
zet8vq!y}8*3Js<J@g#JnO+H@W+bQ?qenX<_G6uqIX7ZDu9n?VNBxKO^>Qqw=hiFoW
z(X|HlJ=%WcoewHj3Tsl|lFR*gON_5rq(p3>m8vY;XcIts(%PDeS=vXds`W>bS9iq)
z72$~4CR6?)Za7XEviENuCbY}gBz@aW2BqigC637)Nr9fxy{|=;*XZFDm3u`BbWED7
z@3wC(4ZV8K&s3OrGL`11|8snD*IEOE&lZ@jXX=mhAXBqzwBCOP<0UiY`GG=!QUmX9
znA8rL0OVU1De&+4`*6X{FgddafAAWv9$Cqjq0Oy;JwRvQsC_Q70*D$)ZjVuvN@#$@
z_aNzdL{Mg_x2x;zP9ga2i#jhReWH}In#{`P@2(buuuJXGJ2pLXLlz$*VN35PztgdC
zs2P)*VlG`w8w+h-Te{cM^VOa4XR{f2=ftuBTN3C+r6p3VJi;jqjEe2oK?(vla8X)!
z{C^-*@1Pmo;<EsQ4WoF-;&tbk5-EXu8MN`@4A=svD0z*75fml*9w0X4#%9H%HTQW!
z2olY@-K8_oax|Mlk%`1-hFai)go>srP83me4qf5F`ojoH0(eM&=XWX`N@Gm>6Z>Wa
z3i}NFZoXun@b^H#R$+Y6O<Pl$W^8wj_=OE`AXadI00fXFkp$j?dmMiT=mso;gDHl5
zjDp0TZ531v8Of_e%XaHp_;QJcZW<5i0ZHL_TO$2w;03Il0ZSa<5u$-333-B?u>&r_
z63(Vnv~v~@re}F;qmC)j1cD3q2pa2^!B9Sj>zTF1`X4@knFcc~o5-%ij`uWNf2%KS
zMECgbsqA{&la17VkwJVMtvpdx>%zt5`;i@Ef{$iLydAl(-<V#}oBz%&-VocTjFQ^J
z*gg3h8vH28oe6y;okIm7qwNeR+giET5vhUK1;+QB;*qXOCYz0fq9aZ&{8nPs8dQoC
zeQ>ZE--PkJr=G9VzVsp3TIQ6cxOKgM*xE3Dw_JKAkpLt`=O9<cW1IO^<|`Iqs~bbN
zl>R<S`*Mk2Iq?KKA#n;QCEDnEgcCjb4uAP-0Wp>ylo)IJv}9$d<yA5Em`z||W_piK
z4WOjWeGOTCtx6MaRhP?)^lHmcfnMa0mcure-cNk<YYkU1KqU5~OZbFQ`aC8s%~(V4
zxx>ZjUsSIUeBXQl%gv+F2B`?dx9`61H%{GpqIj^D7cmF>*#rAa+y#?O$o=aDB*(qx
zM&Z;*$)`p-{h+yMS^E-XFn^c-w+Y#qXRwr-=tc3?DlVfJP<kmoFi7E7%atFu1Vre}
zPLs7EwADZsC|#!R+9LeDc<3iv$A;%YCiM8AGJtG_=Ew2V?kT|@A3r=owvCR|mumf5
z0P|1zJm@DSgU0SI3DI|4qQ)E43DBERgxJl#KSYO}=FHoCam}GE@6>@*LiQ@p_Y?fF
zlSWB5K0t1x!W7w&r11z@3EXawep>Z{=RG9ie60|NBBqz*@rW_<C*s_4UW(_q0QozU
z>wdq0YMmNfE(g*qm<7{d-o|@+F9hY;)nT2O^|*2RNU~YnW8{>Qc=USW__XI`2p6<$
zw8qGRaUxKu_7PO^BV6;cgUnG1oh>&*3k<`|hA|Yt9&m<K)y#__Ecy2C3bZ!}Bt2gZ
zZk`3&E+o>z&I1#vaT@*@a06GSpmzQ8z;a*DhVa13srm#5AN-bNMJ<Qbjqm??$OXFC
zFP<qAGDJ249t^Ttfd8B3z{vj5g3DQ!`TC{dk<b9FX;P{V?HxF)rGN=PRN{2KJv8Uw
z5vXRk8QPGKL`1Lb62bpaOu&Jlr3U5gCvv#-#40V6)=pNXXk2gt)x<SGC2Qdj;q+MF
zfgX`D-k)=|_?pvYW~;{ZljQ*AyoT;Sr>42WF*&&_;Qbm@W{6a=yM9T8seAX>7e9M3
zE#wVUxdlXLE^r&-0-(MZxFM<{K?85uQ}qye4)&M4>!-ObqNm3y9kkC5uN;tIr<#xi
z=1c^}1cy?=+?o+h8K6H-?>o@uwh6QS&%Lz{pDOT*v0qdNazs^FQBBfsi1b~|NZ#kO
zs2N{#w@5V}*$$=CmI<8yNSJS?o7MCj??}zO7+9%|+#@<qGePEU_&G_z*J)`7IahT=
z4z_4)LTl{u<A77ZsR;JD?6AZiE_8K$<!^VXqIHdmg4D9<%B`lsBz3Vj4?(bYGV~BN
zK^ouTfClF@zaIq?8K}Oya7&`-o3#ZayMP~KRsobe4K*CF;x$2HlF!Zt#MyJhpOK{G
z>VLH0vU1(22!Z<3XeCiB;00bpgZC;GIuQ<1M#OS+Q}#*AOq}-W%N)cPD6tywpPD!p
zl<RMV6AjSgYZ3wq(sa4;P<lM8VDUeWPz^{IdVYpt(m(q|L}u=QmCRU!TUVjJQCG%r
z2lAEFO9JF;Lg-P88;?%%Szv!*@5l*E_gvAAGTgKIZJk-rqx_KOdnhITk}xqldlLs+
zbzF~LDcrIu)q&KpDe@MLntmUl#Ej!GPTf$>#OIQqCSh5Ft?A(6gMDr_V$ar~{n>E{
z=1=ZYJ0%XWo|w3|^#}zNm(!s(lHf%)d~w_QpBEV;+ocVd*j>nQ*{g~IB)lrde@T5w
zBJX0yR)Ah9!sPc(rh?QU7buFoYfrwG!~FUV160Z=S6xT=N>pD~uPwOw51>rq_?4qS
z!)4;jP1j{PdR}BIJ=&mDtNVo$h>xLWVu=oI*3Tn;0$TfXvH#iyHehcVBMZ}_t_=3e
zBWIX^<m?6dfW9Q;k;VV7b_(jF<ea=Xp}L;6Gaq`(CyQAS-`L-&=i#Ib!5zh}!utKi
zEnxDWBtd8|EBLAP=}bzQlK;HTXF&O?WxW3KVk2Ouf~XTdi8*%!6}~-LtV~U+UT+4%
z)Hta&wA9};p8=rsx`%u*6Gf8m@Lyerr|xppImfPQD_1;L!`((Py3?Z@+~S!BX0^yY
zs{Z4x<DTj!vnR1Y$_M^|BH=|rvdV?^$$9`cqmD>C*ZEuRdha;O=H+x7X8^w+l1uqu
zJ#1(taXFjEm%6V)JK|&{(1UvwP#Tpad4mx2X6QQj^h&THT*0i;IzkmkyXH>!=iNQ9
zvXA6F@&`%O9BfAhpa$MjbUjci$dBdxI67w`_0JT72#CyjCJF)Q?3RJOSj0J0eNy}P
z>~%ny$nkAW{yJkPGeHMj2q?nB(EmAENXr)xifxhtwPFF>@YpOcl&D}ik@?du=SSD&
zbWVahNCCNn9d752^(rX!y>=x(aCRGqgfNh54>mZ|Z4H@`knYNeq<zVxi`?B0ohO%c
z;=30}R5JbLm}JWH66MV~hpT9ANg-%jsE*Wcb>r>IZE_1NZW|v5R<}CWze1x?Q<{e=
z*S+`uD*ZvL{7iX=F!|Lk{+KCKkU4U9>)W!9*dkfrUMbB>&fY7X=Cyv}vp&kG{e8!H
z9Sx0{ox^#?s_f14!9?Ah7ReVdEf8`gD&-<<O}vYLt~exF^A&AhS$TpJl_X@fYV_>#
zKV`5qbpPlzE{Z-!ap6nv2A6F?rUPStei6-(m-nxd3vK2{Mn)WspEXM4F5q6DKAz0?
znlY(>V0IkPO|-wZU<Rye6=-idHo-(_3c<OJ%aF&dW=CMt@5UG98f0zj2i<>TA1K$6
z^dLZ@Q6-9+>7|$It#9UpSmnRx{re0A$t|Hj*WE?~gMG{_=fMS&QbNQSCwJFRVvcLs
zrvJ&R|ML65S+ago0z+97`D!4;_ZOHEXRwdU>lz4Rtg2``)=?J!-=ha5dX<w4!XOZp
znYW?UiE#&Fo;$w^e)s~6gI#PVg<!ZRGxiK%iZl~JivwOT4|Z89Hi0<Sl=(#}0CEzC
zf{HAob^en`!ay^JcC#%bJ#{j=!iSxWv&^^4Q<K>&y+^B`_55Oq2P6&K@%Y}1Qax|@
z8vz>*xOjKkcDEbQN<4Sw-!vEK;BV+7%@G?Pf=3p9Mus-m+3l{j?jI4VO*GV@BUDns
zDnO6y0x359Noy%^{)|yEht)4v;ean9sFJ@=AHeRSw1{^G$?6|Ot7lZ{&t+tRWK-*>
z1P2@kiw1sEr@U^xpTZbRdVvCpVhYZ^egSoDE#aT`StQ`opIHH2K13pLdAQK`*Rsb0
zXN{6fFPjC?9+3qRfHtI=4|O==N{rY5(DhqQ#6Z&7wdqqv909g>d~GUyg6j8lN&qV#
zAT^18(-_gA`*$dyA7J9z9O^M4nbw?W=4=~(NMOI!%xq}BX!ZF&;mAS6#us0{pKYC_
z1q`y~%R;#C+6-7~u(Iqi&A-J7mW(v3`{F-uZsTiNH_cP{iSXpo32_mmeNnMExMz+2
z`uV4j<o^PR!>;<Lfy6<eF>}GYHd`8!7eYhb3)DA{kIuF)$!{O{eJbeLmff3MS&8m2
zlE($jmKv^6!3k#wCkMf9lD~6kG8VqNV4#Vsv;H~x&yeBFI-b9%sgXYip;N*DiNfmy
zt%PAn*p8GcD-tC(lq6d_G*$-G;&G=wCN$W`F@M*t0bjSBraLwiuM-}ssKL8r8lx&-
z>RI{!W_e(hCK6U1jP!Uy4HwE<A|Ez*l<m0zeVduS@4vM|q8MJ;#<jDZnB*<|Ay)75
zhkb<Jq%`uWs*&7XwRfY85%{Oa+=}NI`(3r@0zFoXsq$XLlzS|LzeAQN0|V&<83P=z
z2+c_IeAtiMU+_Eb)ZNU6>Mq?o|C)LI64^YIR8dEb;&3?BcQdf=u(pn#ktx^FKUlvx
z`5zQTNCfyT8hG7AJN$JLtNqg_7O_49EorxrJLVf^7Lb7HT}8~VXZ$xC<#$}f1JLh3
z2Mf?gZkSb>Y!6QZ)_SP!`p>tSM4$_)Wq2+_`A{gp-Zkj%(F3I5Y$qWpgV3lUl8Efa
z_)D3&W7xnIgbX~gGq{;tZiXAa%Y@e1E-%O1Xp}_qgh73gRUOb59wVGs(9wD?yCs3n
zkY?w+zur>lFM!FP($8Qg^$6#`GLCN5G_Z~dzV9<7Empg61`w;zMNowd!F+U3Tr;$!
zgnmh(ArY-9PO#lsIO?iw8}+)cGEw{A<#b=bapp)Zm=7biYQ&Q9&C+JDJZvNn%-^&`
zLnC<=LeKJF=mS1`7UfHgEm4RPEE5V%hn*)*>0O52IA+7#*Jd-pLI_BGo_p~^4)h}R
zCt!I(<7KXpwGvgKA&i6BbI>mty7>CJNJ{wC2^B`@!~dF^{ZB85^ECw6$qIOJl!A&T
zDUdw<uieBd_r}P*2`|5kIHtgk9$ftQ_%QT$k$FT=Ir9gY1`+Ryv6u!bf95A-6EXDy
zTHpTJMg)fb0rdL$(Z4XoRwJ$-zPq2Gfmx}Vo}4a1=;C`^un!3~Vnpa2XZXu46Ar@l
zcg-*SbCIC`wLm9C!>~@wKnc-~0Gc8XC9M}PzJFjwgI+CnH34d^4ZopqXF3I-*-*Qk
z{f}3+x}E)e`kZz{m%IiZ!hvGz9+?XdP$d`s_yAR~5wZ!Hoe%e+-tWV{|2I(9jy#uD
z-x5%|x97PkP7J2E`YnwW?#brwMHq3x{~7V10P}(52gkF$$?(Xom`<?r99BKgb(CkP
zhj7avj{kE@zESV-M8~s3M4RT1!aEi0y@dYVu9XXWn13$v>R~~+Jj&h7`}K(f4t3OQ
z1G_;ad-ofg6l+X!Oo6S$sEedO@?;2X%>}<T-4*1#3d2>%J;wjL!U`mHq(m1|tlFCz
z9r*w@^qk4xlL@tltbAmZ)QPF(&r+BjLD$QZ&;LyoGhNc++u``&o7))BwKONgvgdf;
zj1|2&YRSUpNp{**@It~ekhC~&XMMfzpJ7HU{B(yMP@@&(W_(ckZ}PUba}h;YDZe+t
z2TWLQfF97JNI8Qw1`KE55Fd%D7hE*%{ubvcQ^e-~1iC4q@j{_)XMzK-w!)MF$}mN`
z0)8|T*6y3RFvlR!e(DhcXV!;LAN|{Bx(m|5Iz{{u3JS9p3Wd#8Xu7@zBxxtgKvuSu
zCc=*XitN?Dz!MVrEYq3|y}CvyjqFx!fDSP_ZLhAucfWeCwu4?W{^y{NtZuhVLvI&7
zX&B?BVPFNe=5|B}b=mFsf37Mf+a4RAQK|3ZLCBngD?Qg08;PKWWX#a^{wW^;$*wjB
zB>W5;8AeZcFQJtv96YuiWq6Uun?6D(<=9Xcm`PhwXv-IbZdBk@2Oe3KbCEzFa}9U>
z)4p!UQg&G$G*pQoc|<@n-Ht#0&&W~wq!XF|9H@gwD(v<UVz1w+hr6haZ&_VJM-UCb
zCRE?MS1O_MxmKD^#o-l?fe5_Pk4BJq?S1x*IPwW?xF-$P0hUAV$RG>7%<Bnm{?$}H
z160Bz2sru~Nwi2+>!UKv%WKMrc1VCbbEV5}+sES$z(EFR-`a8K(ZVvqi5m?Ewup8L
zLGEy4XaqNHZwRn*G8NMjFcSp79#>nijwKaVH*vt?L#RXq7iZPFKkw|eyiaS*CtBM-
z8i)#hit8(b2{<<OYf+}0L%+P0%oO<&+?zh~XgAo|e}S5N)BFjn^g<gzd0o621=RnN
zHQ?0l=gKb{SBRB;=G6wGa)|Q9qJXEr!@s<=i6Abg!qyS3eP_vD+nt1ZQfb{N^cfAj
z3S?CE=a8-018RUur8RWy9>)0fr~f46W`b|5rA!tRpaMLlm@pSB>5MG3D1yU1dC<G|
zh1A%66uukFcPxNlVjs27zOHlQc$KNIO*g^p&N1o4x3dOb(V#)BM+WG3XEKi)hzMym
z=?_`c)Ph0xznODGpe0aNWqc8+79-!=<=&b+4Eppw+j=Oe6<RZ>4PHPnm)oT2u@JxP
zSN#4X^m^{w>go-~eSuU{<95l_1CZ}K<}%cscb|Q|Jaq2-u}rjpzBCz@b_7Ez2kA}7
z54;AR242`yM>)Z5LmgY*QL}xmc;2pKVm88}D1UpbY^f}i$s@LJoFIId$zWqO8Tv?v
zp)!*_(`<y(HMIQR=VS*HE`wrCq%_u``eaZy-|a2NWhFPsPKx&2Sl?*Ts{LwJe*jf8
z^+1~I&g0#Wcrt0Bt5ITF&Q&~JXjSh+<-}92B>7%|cOz>7I;FZ8un2A51uKPjcAyIe
zIknoj%B(X>86p1Xai}R1zHaKpwXKrm{EA|+Lm#=BMutum-0N*>Afu4^RC*GDL)Z4+
zMIlv>b|i}g(L{kchoocEasqUYXm4m3bZGPZ-vWbM2`i<$8x$UP(7fowKgk`gs-|SD
z=EIg+zb0R=^xjJvkN2wuR+u9*P<SIpwr6{5N1W(S@9SRHc${d36?#y#Ct?Ku3;j0W
zn+-RHey$fw<*?A8yt9}v^xX@h?dB_0e@BBu3l8^@OWR$Rqf=4`&o2@d5$a44wfY2?
zjl$nT`Tn_GdL$}VBQl9;GClMe6`OB+2pIleeuct$dH8?7X7C~HJ{0(cn#)=D_P@_j
zhcc+#<!Shyj!{y-BC^89$Eb6`t1!B1#Knhp5sza4j!BIof+Zn9!WcO|=<y85D^1Zx
zadmbmKoDW}mL}8e1-LjVwqk|NN#WPhWdA6a$lQJCC=?K^@sVE~`RetML-NFqvHluC
z@6K0l54`&+s{YFK%6i<J-#cMvY(Hl)>)N>WYzT08+X(G)K|x8Q<Ql!Hp{F{I`d`3S
zud0jiM~rSNddH1?ZMT=d9u3>D7Gb)1S!rCDYYf&G9wQNve!f-2>#R+t`+NAUrRU<s
z_mSVvf5(itBo1e{gF`Xav?#ndTpYj(M&(3odFF-PcX;czm^yL_kkx47=m*l8L@F$F
zid;PLkc`x1*#r5p!FzI#j4mzL3>d}7M?p_GIdh>;wyMIm`>8-*1-@Nun>v!~VqJ?(
zC&gHV@#Na)An2`qzS7rOsvBhsQ!+30yM<WYHhKCn{vd$eeamFkcCBf<eAij{Oz2Uc
zlDOlg6T?K!UB9JK!*OH0&VqHkRrBeDD7{*p6-;&Rmc~FRHB-eig*+~Z2mDA}73u0n
zigWa=3nJ~Lf}0aJ3A893$%eo7Ds-=hA*P9C1+^=EAXgEI3<=^NP6<P7hy%%P?$0M4
zoUXlMBiWYd2iPXuCy&~t6OykSD$Ti1_^F1+qhA9uUO_vKQDG6U*zWkp#da?FDx%GL
zhk-Dcpj%`7i2U**4Vy1F`tc*-c(;kho4y{Cr8o+s%#T`-03@C<SR_R)x4-7S9}+wu
z_S!j*nGF8&t#g0)BYr1kY*}28X^NS!k>8Ya#@~KXdU)DbX6&PtyZgrrpEzl7NU_n@
z`w*#dv*0|_EcY={I;4T|#GCc73)`tIU)Ns)E9$tUqlb&sKU-V*W1Rb9=_CBrA<_N*
zfjW>qwBgmP_xv~<S#h1*U!GJ~+x(=T9%`5I<bK7E@gUOdF|BuBysIQw88k1RRRVOl
z(6%i#>>WwikUMfgTK7r{GKktop4|pZMwUj>e$o(7y;D?RVNiE=C}t%-P}9|nYc0nY
zUvlI@$T~9q^VPYtK0d!jS|Hn{Q(}H3T}!K0FOs<G9`b7WxLTEymEH@ZJ9w+>eY6H5
z4as+8iAU8}DsiJvA`Nn;)%|MU|5^m3Cpdlh(GJ@MM-BFSdlUm&g!_h>GR>x6vd<48
zJWX0H8ro>$6AeXv*R+`k+vc{V$?J);kYe?_+@M5kIHp(<epyoar_{OfZEN_WK_@Mw
zpzCyqw!+aT$t37H;!tBX`wj2SsBM$gKs#Ns)QtS#xAc)c%NFBFArD?#4zr(#gv7-_
z;yJBt6Y{`*#jt06Unaj~pm~Jo9+z>ZeFpoKI~?Rp>X!p=X8U(5X=PhGNaR8rznaX6
zK#x_UY6hx|I>!SR{^$}Lg;gc?4)%_OW<bAy8b0(eAk0jdjPhKP!uKvnoeVW`)qb^d
zJH;RSqKRnzw2@=6<we0G6cemvJH2nl%{R2Eez9BMe@t^C*^u7OY|kZ*gw|Buz9FM3
z<r__O`}^qDT=XZJuGUj;)L3PLuDXQz^(Bw?uRQHuNQMHlKO0W)U4=EWi8ok2+?$9h
z*LxaxAz+nsO&DzgM7(K!0Qb~ECNW2j0!+kK<ZJ%^e(ST|o}3WWBI6`CYE$a#*^UP+
zv}FuYc8^P=1N>^!?Nhw=K#&?+)46=#^76N~)3krGwpUmx{luT5gw4lQQVCgO!Kj{*
zrd^Tpx@lSiZw%0eHeUgbU|4fN&<WN}1IQ4qNufqt2(a75WKI2uo#FQ%-QHHe+Br*N
zTu-#=sq@tR1`@lmBk8vRCVq_{9`t%s+01dMp??02S|QcIM}T`2sKhid4LT4Hw1f&k
zS)gz%Jn`w-P^akeo-#;j7GXX%z&J(IDWt>DU+oc6CEJTT-C0zP*(&|G5q(n=RROtV
z<bPTC5wzU37;r3RMtX5_pAx;PCo|pztNy*WKlHLo`v?T2(?D{vHM^l0xjL6kmD%V|
zNjK5?N+&WI^weKZq+&5`b+AO$=<G+mN{dqiqj4VENriz`l|tOFn-N!>GuY}Ljx;Dw
zmCfc{-$tMg8)R$mhHIaSb<^j%yx5%SyA0U&3b*!?&u$Jts%Q-Ev58&Cmf10NHCk(u
z7vK43B(*^6yJ}&<2OpsgiqM)~L%2a5XSbqjcleFWpJc7+vy;yWkhb!ehKvl<)K-k2
zB3bAopZ6(@NL`m<4O_qPkI8_(bW4ue|1TE+PW^=4_SN1uub}+rWgjHs<2>n3QR>P4
z$gaMU>t$AeB~03TZ?-=E|9w~b`ODj^1!=&oILak4@{EZ5$x@vBg{coo-uUN3L&XB}
z$Pl?1JH(T7<<~pyB_US)J~MwPuPH*92K#io`oQXDV-Qr?bkc4HYoAXAw=Q<30ZG7`
zH{6pBdkikXl7>g1Y(;RWKM@Y!S9G|ECIh@RL?wPeo--a^2bc9#`%e+u{@<<2-wf{$
z+Iw9=`xCEa29F3L=0hJ2LP)d$#Y)k+jXk63iWlXjw4Ig!+=+K6&^=_GY8NoJrsi;b
z!xNy+rPG_QF6ZxMPPu>c!qoGt8W%BQlVJ(ODePF%6*pq^YnYEbigrVT_AH!*cw!n+
zR3CQH9%TT%fAS$pSA7|2Li#?NRfL(yBO`uZ4Eu&xIsA)tWb@HvmJ~jrm?zH{FUolq
zvVqbP%OlQfb2j-9j(PNb`FxW15vjbP{;rc8%@QR%?M8yXJmlBNYl_c|Z9NKQYLn?K
zNvk@86%Z4+OOKWZB>h)?TaU~pe?eMlcl5bv{(4Sq6D|Axn_!;~l!tbK>13ChdoebX
z6Ki0@GjD&Z?s#lJYVdD??s$jNJaml^ys)l{%LQe+ui^Ub-R+0S-s|U?x}QTb1gY)8
zc_s+1wwMzf_cQv*FeUOeY2f%wiiD%;E!a-QOFFPJgghIH=9%W5%N*sfS!6&_7?GfV
zF<iH}+x-9-9?fOKdI3Ga0X>?<twRGJ6ofCQQ#&?gllY&!Y>*5`S3G}H-ARkW4fTv(
zLe#<5NC5}Z8#olZHn8qTNPoVF@{N-(9)IdBYjaWS%tVj9XZW41EjgNO8rEr|%lqdv
zS<rm__AB48$~VgwQsMTypYHl72J=##{~tdCkK{2<V2fW}i<Y~L0omltE~4WBpYxEd
z3IP%p>@5MKr@geu%i|?ct<LktVc<}?TqkC6U6PqW$z?LqfAf24r2R;iu-G-)gibB;
z&BBYAqo#sjVQ$2<UWCpEuCPXr|LG2yI6QVK@MRG27-%ak3>6S_y7=7#9=izY+Ad*y
z1>0hNH+B}bL|r=&|1Cy(i!5?@00v|X2)TdO*Fc%Bun%05e()j~0%^*wx%27el&NLF
z`B_a}5f54d`1y`l2xSFA59eKPiEjQ+uc@B|^f_+>x7pBPD!r5w!Bi1}rizVvX$p=k
z0j^2&op+B_aZK<(E^Ib5?9*LhW%^aJ&l`*yztzfA#@B_Ddeu2D6E8vqImav4PFR}E
z>rq5IY2IV%yXI*;>%9M&4{GHo!QkMqH_k!c#A;gF8dH2aeU*I7*^jkJjYQB-_GyIB
z96)K6&~w|7gz=506hpa@MvKZ)r?sGq$8vgI8THJr*A}X^v+p_O5`~WVI?#ZZWRb*y
z#&GjC{DYGC!|y#TZ)X0VwW7Wk0AhNGy_4q~zwSB!`NWweZdRbqs!oEo`D5}<_pPjj
zT!d3TRp_m9=H4WOfs#zBR}YHyVk`|<66Aw|i{2O$%;8DvzA+n!=s7d-T&%DCGDPX`
z0BCC<qEhIM?O_m;Y;Z7&)pf6P`#yzCYRb_Us0?FiOYHn@%h-Z~XJ2)XMUyxYkAJ+U
z!$^lTFz9Zt{@KlDM}AK@q#hm`3#dK3rf_P#i0`7rh{~C;_W&99<xWAOH5EV;C|5z|
z85X%zpb!l?MZD*A%J<JR%*J5F3pHKX^cS(wulJu{#F#=lwh(k*DXe^^TV?Y7Jo6-Z
zz6|#M!;kmmZ@$9EyO~VSC64t*of>D}?Q%C4rX5KT<bvHG##-vDLQnhwP^a_N=%^Fo
ze&Fj(b?Zs3>T#P#<X9qri)bx+((gv_@23zk(o6p7ig}E>UKq-5U#QZIf%c*r<*sMo
z!NC~`iJ_V-DBFn8)K)0Ie%Nfh@dT%jC}O}m19`BuyZv!1hfwj3$5W(V9$4aQ8~T97
z>2m@JfB8X5UWRiDC=cYgh#@ehHwaUymaM1EtELa7U1|7QIPH1n&69#UZA<0s2fFe+
zA(R{S*W*t*g3p$|=xWx9s#&o&WQJ3Ptq?h8Rrx9){YWP)fh1L&Mbg+2^O?1!kr($q
zD+Wr4<$V+$xr8x0&!pCfG-F2>u6XWt8xLu=5PkV`7J8qe8BFcDF^rY~5Fn!F!u#{Q
zPrFh;2mvh$-teMDcLRCkjzVY|*v_4#=_AYedq_SYfdVU7aP1?%Tk`VIff}$6V=pXB
zwbjd#qncBgI{$=SJtX)SWC;m7OyO~L2e;odLKKD2tP0LW`Y<1SteW^tj8%kk{uqd?
zNT4WXx~{t)Kn`PfG~^B^ia9eOd<6WJffp8jsU?}va1XbrVE*GWKY_l(H4Mw4j}%*}
zaTx4@P8<3dP7F8ag0!1zQEOjeyE<c|^|8v78sw;-G*&$%WVY0{#XDnHioX>GgfM0c
z7KKE)*%VFLa|o*P5p|JF+(0t{GV^~{Ws(p4H@W?fk5&gQ9kqyq%bg3@u*&XBrYi?e
zhvT#IDqA~r=P{(4Vgv@|A`p?epm2XShEB~*NC5M1PnY47xD+YL`D2<P^lB;BNNCm<
zm5rfcI5lHm33~b;H$BVE%utV_vLjL@Ga^Ul<~?cnmU2$4WVSu;F!tqrHjR?INN~jR
zt-MFve&_b;@z^yW&-K&&ZQ>nKv;?Z8pgFH_xq5)PJX`;%3njWj@MaNT>hKab;#N|x
zJn?*7FfRl@kUTNIq%-s?qi+m3eTn@=HUthPTJEr%?|xV4smr0GqoexdNnmUb$U@*0
zp*DM-E%XdFBiC9EbuFe~KL?)>=GoKNGX^dgC-vm44ezyU;a(pwJI=Sh5-P&l$5$O*
zT`MmE<H;vB1xXGrplg65HRR>z0uV37rEo7;x!_BCn1Sj|@w(o4otJ9<MPOR){<eoY
z#<sXlTzB|R^|PC-n}LX*t()nZ1m=y<D&$b6VC7ICf*CpByUfHeSX2Ujz=|Ee7$;)q
z)x7qX9#HvKau&J0`@}uD-1*~tGBZlb1KPk>NFB>;;8u%%PmqF^sz~G?NRK4yA4G+(
z@rAg#-SSqqk?S`UxBv}7Bq?!KiRHZ7t-<jG%|~CK_qaJ&GdgLir547Bm1zDem;2~V
zboahMxwBwQFxh=FGMYgiMA}v)99;tHL^!r$Z!e#tB710I++IzFI^;t;lt%^30)i|~
z<RQU(j4~CMB<h#)g>0nCZc~{uX@7c;;z!Ac)MV8({)Lpe<tl%U6YtYUy<^>fN|L5y
z{;_uz?SZe1LWP+*cb7X%Bz=kP&!IJeOA|Vs=3bmR=N_Ssm5_79F|dhzffrA3a~Ge#
z2+kZYEDE*)T-@BI8aN!hdd#5d(Ph*%@x|$hl6~2rRAsiqr<W>zyvB59_Y=}hztDBx
zer@04-#T}`gs-PGqjHce42d_kVBKnf^C8FDa2Gx*Q3ie$wr08g(V9#S`H_Y{G`OgZ
z2=j+_{Kfr3n(GCFUZcd#om?@;csEYziJFz2heBIO-I`fNzR(=I(JcWD$?h&QSohm3
z+(3ajxx~<&uI}#eFkPGOvA(&N54WCr^JJf9Nd!BEy}Spdx}b{zy(<VDkaIZb<?h|X
z8Ux%mnr2xiv525r><ewa2=oUx&uidUq-%{{vK2S8x8b*ykvvS_5f}mQH2W{i1Lxz1
zHssWmIo^G@zZek~WOnU$;Pgf9GcKuV6kl2%r26E^j71*n*?elk2EQ-7p1hkoCvdhX
z<%~gt-VdsfgHo6JiRPgmp&O@`LSkJH|E7d*-x8Q!x)i`L8axEp!XqtVlFRo}bpZLN
z0lof2rnEIHe4#7nBDDd(GZWM@ImOmDv;toIa0|#IO$>YPB7rU*9nC-6ap@kC_@&)%
z0$5i@DwoBRhIjKGb3UA}UKlixi!98##<P&6lt?ND?TPE@cVEY`MpRsH{>DM&*{4@4
zz`3<^v8#mNetk~eZkTJ;%E>QFti&9H+!9-wvvVCNe^w<HCTGekD7e+fDrKE2=3w_{
zgNpKAedJ4pyQPVJTQrUd8Yff21WXgXR)OufGc$-170?2jL<94IUo_qMi5x+yD8I>X
zZ|wli*{h3&n)*<LsFl;Dsg42L1bgb*eGMYV^+aQA7(EoLWcvf{!Q54%OW9AnEY`SK
z5G~xAOlpRDM&<=8p5}3giSr9fmzE)g&lFJ7Pc%jB?(RhNNC(I)X}?rxO8#?=oa5QE
zA4^TrAr@g64TQ&^<!$4=%oWY&OlPzUifg9Nq%&zub@Q*R^99n}5}gR-k$(9+QdrV!
zR_eUIX3}5$;rcrtO|`GJB{cG1t_^q@(Mlgoaje@NPcLnoY({d#sEZQ!x7V`DBzlYV
zagDtdM~Mhs`|O)|d?X(gE^FfZu(HX@>d)J&5&}Dsd308reTnPaWyjcyNW`weOEpW&
zUfrg^OY!!jny<A*c)nbC|2o|2Y<yWS=1Va>>F#cfxaoU$s5XPp%$p!5$Q+`yt9*`a
zn2|Acz)mG7eE|^-+exQ-XF#?lZ=*s!;lph7wfy?$;n#>So^gB8*@9kX-GRFbfX@)A
zYjpXr7;+|sYH_Kwv4kSWH5so=W)$#C1{O2rwxk6Zu=6#&4lW9UP@N)*3tuOSK&-Z3
z=Tl6MI(xUC$&M+r+Mxq?b$q7u(0eUXw6*4WP!yL3%;rHdJ*5Li2p@<r?db@&`n~)A
zkO~A$;xC1T4Y1j*+uvG!EhiNwVs^KOEB`QwnU%CmDl3zcOW)bQ|8WTx1Vz+qukVM>
zPS+VGe3jET*{|Fm_p1`qJl7<}8n>}YLQ1;(dui$6-VvsI<wiw8W;4OC-@O~TtygY0
zGcR8)6mA>WnN0sr?l#1L4o6s*B?^CayV5OEYibTk4!kZm!?r^8W8D4PAI);rDOgf&
zPGL*RI8i3dv4Sn=j9^ODuhBr;gse+|vJ;G6@7-T>l++Cunp5zOkv>yuRooKq!mSmy
zKGeDVrqoMamj>ixnH+VPW*Ky;WYcP%FMF?>KnEysNcCkS`JyBpI0Z9zcRsjJ)bHWS
z3rz(5JaGDth*_&Tx$>LG-7Ez{nvx8=YTl<GVaLE%f!7w6cKv{;f!e%h4R#(mep1lP
z1D8UT!<4QW$f-&6h1gHMD9jvo#_IOp`@*7L)g(}q6=pmW;r-)Z@>g(CD~d6;tb+2Q
zjpx>r3Ax;tgCt^g6#D&WY9`3XfzGrdN9Pmm5WTKw=F~F&TOMrd`bmG4%rVk8;p3DX
zKT*Ihm=?TL7PUyWUw8qtG=1Hfwe(3Cu7{0jNYC)O1hkn=&$Rq`mTZg?^`kQ*%-+Fn
zV<kVFfyY;J)a?@D^ZdYmVM_BrZ<O~`Syu5o@K;!_nYs-6aU=M7%uuM``<RJHiJC<3
zmc#a-X}?C>dLTmPwpWFlesZ|jERDbIMYe3)?@>Co!)pDe<s#3&okxCAyUtrA6q-|%
zIa-qcx6Cqxit79+NQ6_<!Lvwid`D5#OnkdAz9Z{hfv=ZxGX3+lO2{Qdqa<lwpi09v
z2oyx&Y(~V3Zv{}ne41R3wJpb&Neh}tk$n<#WT=|4K*VIjEjXU&JAyUpK=YZU<el+&
z3s%Fg@^{u94|N=B^R6C7i*Rd%Z#~`-=>6&5diPdd#7lnH?mp8%=5pV|eV)3R<!q(H
zf`-0$164*ZWroT1FXy6r3bB74=6tt_cE7A-WuAY$k|8r6CiiurD6L6@YWI_>jiBKx
z^1K}#$=gK*OCiy<0~ymw74?CM?ZXQN5GtsyFwSP)!;U6*%;gGdv@Eg~XQrn9$i6^n
z-<d9D_b-2I4`*9ec;CL7EdqLFa|$n!Q8ajq{QIkyqMLn{nwDSmN~p|ZUETygOaeJk
zI2m9kk1hK5A?Otc6_j@C0bCW^wdP_~t_O>529ES<5fE=sFjMxJEne%lQf{sechXSN
zIW=ZaTl<HG6n`66l-8Tmd#;f~uWoD|+q!blIpV*S{>E<T)^dKsao=%6d|B4D?=OWT
z44$j-=7<jr(Rt-v<H=4b+P$wdxosQp0Y#-<B^l5CN>lod+wh%mk53Ao@1u*_6H)xL
z^Yh;OWmDfAJd9%vjXy4UaH*6qo4+>PHJ$(AAU^+6n&l|*SwABagHrsH<&uPAbDKN!
z4xJFTIK>jg7U$61+ubi-$~=WUwers)O%;hxYam{A%4h(hSDn~ydZz8`t)ENETj~*1
zLxt^LpIcc}#cg4I$x*Ju5QWlBUz~xw83W;G<DdCkgIqS->9Y-KEF7=+EfXTnRNi|Z
zJ8SwY&)WcRn-4ZZf*rJUe<wD=ht9uhPj~EVvB*Kdt|5)RQWd+4^0(4XRvCVcyXN_`
zS}6UhsVLKXERSNisT>55aXPB8MU;HceW}6)V_Y6re``+HRQkkQI20P7X1~Vo-)BF-
z3eqfu8GWkK*BWqEPzb3}ZSKf#?9CRY)MPug{@v(7W3I6BDChkpx&M!;w*ZQ3Yr2L>
zfRI4YV8Iz6Sb%V`5G=SmgG+FC_XKx>6Wjs>8{91r7~FMW(BKU2`W^0l-oO4+Q&0s}
zRMERn_v+QFdlRvw8swDNR#^2#8347TiHqL)Mz~a*j(@T@|9L1N|Kuz_O5i0)#w+#R
z-NiO?GTmkO!kb5f`<W`2g`#7WR%rPn6pvS6MGoth8eW6Kb&{9fw#!lS21|9%h!h1_
z+Iv$c2F7k4suA5>ioS&OM2elUK=FZL>P7)0(!Q({AZ?YQ=*Tla{@u28&ux&EatMB<
zg^Huw^Bs&a-^|@+{7B(rI{lBm9)`yq`n!(2UwOrI-Z7{WZrPL<v=y<Rt($x7?oFbd
z?yVpz?Kig)^dZr*_z>ENp0#P=Hm>SCvgl!!?fvbW>-sIbx#~%siv}%S-KhDRLN8%l
z9fto(UCd%0Bspg`j|vvF61AnX<^^Q)1$X6VHZEWmTac4_PB7e>-xQ!MkparGNYlFZ
zTQLK#b=Uj9O?_R3(**h6J_=*V)_z0=Uh7K0-Nof97B@6JtDD1vZI(tGN?L}}7M^m)
zePbtA@#E4d4HR?H(+ISq(wpg}R^BnUzp1kwRYnyW<IefkEIdtCQ5viG;PXx23(d+K
zw7lFSvlfgrIq>{TpOYQPpVF`U0&Ti2tQ$8kYYSy(>anEtmoK7Ob`;8_Xxa3y1NEci
z6SpmVMfn*5N;%PAND>m))z|k@0>UHU$$MsMIK~C&u1<x8UW~0J`?G?U3E~vP`<^2;
zv>ZphXVC6ZP4*Tm7-ISza_dNyH&j_2<`iPU$Wi3++;5x7w7Q4M!N|`P2VYV<?L%`}
z>ny~rG2h!J$IMo~<sjRwLOlOAeXy|JXa0<1Jys(*(US#HrqW-SAQ_DzAT;g7Idkux
zA>=<2lpwnB$^5Q=>|Af}it_Z3-=~vQsE;YqW(l$+UG~7vqc+qddS2RQS-FuB&5Qp!
zBu#xW1iCxxO`S-2Q9yFakhHR=qhw;JR?S35LQMGDwHKeGP~EsYnS9&gu<y$CP4`dZ
z<axjUl%(51j_Wx0BXB(1XGahuLF}C8okDZaXGjqUZ-`x4zJ2j7b=L1iY7j2z!pqJr
zq<sQhKJ(FEy9cX&Heto)a;c2wngrqC0Zk0f^wC1!@=#G7dmH<a?dcH;cNrf)73rIj
zg6PnOpL+F#9)7ZX>DrO4IU2h&^m<FA#h{Zdck(4l*o?I|k4KVZFXVGI4zw%V@;V8M
z^6B?(w&wc?bCOxl%LPutmqj*1)E(&CX%W2oN+tmIRbpnndy(9Rkk$XGTp2MyIofwn
zgI;ZNH`{>2te1pmnxe-xrz!pf?dr<g6or-T#W^gj|0j64AB+}t8QKbDm8!1XUa;e*
zxlLZSAi|$NDKEL9Qj{Fge)%Pw3;L>-gcj)j8*CM3;EG;`_Wr)MDSpp@VOvMr@gZNs
z(#AGow85?Lk;>O=`fu&vn}Ke_$;AY-kmLp9OzD)<=e8|(*To^Fy|zO?a1dDEw7E8)
z=SW%XQ!fv??zJ63f#}AT<<#6@V@+53+{?d5SB)ULb4XHfJG#RpNkRwWjpXSWcjY1B
zwp<V(b9W3oJfDjnfxI86l?jhg)R8RWQOBjl9%ixC5=cYX+FID|eU^4Kd&K^afgC?}
zUdJI975fuy;8GOk^9I<xV1`7Dx&j9K280S-HkN=?3Fewuy6*vw`S8hyC902Tbo3W%
z-qmhop8o*<BTuPCo=vH$mh1-4n&R5WE<f&Fiq74T53Q%u6SQ6|yYhXw`(1xC)*RQ$
z4~ckL@iPOAXjN$#$M8?%+C}|LxOQcXjS590QB`{jQ(vn#Kuyffjs*}VUsmgpwZ8ZO
z?<m_z0I2bMzy4?!1BPD-P0y0A4h;KeIpK-UM0t~s7DRjO8Z1k;T<nj(x#hk7O4xX=
z&D<cnf!E1I95Pd5RU-!#{r8mx<+bDTSd@M2=U!uqysI|@j<BG_PZ8YFpTp1)9#)5b
zM|lfqZ0x|MErg(5f)MdUj=~s{P?d0xBzo01y*oShG<Zyr=rpKr+qx3?3?fD$)^e1b
zZOvP3mMOBQk)nh7=-tNjbPw=NHppD_KYrOlhp}2H5dC87A<z|NSIds)bWNMF4HWjA
z3d6B?lYnCKz;<(PS82>-uwV1Dagw-3n+qw0m1dxHT>Qahyv9mfmI0$0P_YlZMOe$K
z$0aIM(&^`Y`DoIh_KoOVz9pEaYK~pV`^P7*+IFb9CPSUABssUuuZ^bvBGdp#h_;%_
zqJLt;bGN{|=kw19pl#$4^CHW{;oCu18+K_)T?Mc8OY#?>fq$O<*LsS@J#dTWM*j>M
zuo7I1YK1ZWOgKSgl9k7;HYoYBjvSN4z+qxXI%P{uK74s=^iXhu>p{8cuY0tAt$6~`
z3URu2r%}>p(&bYQe&?MI!)lC;3n;8@N17mijXElrQ>fLbiswa~{AzEK%-!c@2uaD%
z-Y2ZOaY6_=Jb}muZY^vM+0sCtuP+PN7yCn+^3t~$y}C~D;xD=WM-x?VRw&P|0gxvR
zN_|^P3iSE)1q?HYWPaD>FErLA6G;3AJG6|(zlXJg#UA{C(#k!oj|FtXjr(hVhX7ev
zH9brYC(aYpwaQXW%T6{C6lv;T_REAio|!nZaf_Td5$$W3!pjGhqNwn`I;0B6QPTPO
znI$%~s;U~g9UpZuyayIU?ywd}ILDM?!oN|V!bJ_XVbK-3KCbDpm^*#TLs{c9^GC~L
zeI?||ms`mE4*$=IQD0z=FkhRcA?=#Hv68*EvUWtEJZ9s?FP%y@SWt`NuUEFWT~MKK
z<%sB1osTcDbx*)PM~V%`-cVe*e#-fC7y_*on&mgEl8EmT%|3~5eNpJQ*F?{D#l)KC
z^J!Nn9jE{Bfx!6Nc(AHMkD`v9gr4Xuu{1y$I3gl+T4z(WmBJ+<jc8Adhd@?nI?3_B
zc~brmr$=#XDe63w$5>%nPL7msiI1j3|Iy>F{c@y<WZu^1Bj+5khG!f<*+&;ninTNG
z<}GzF7(^JE3oKZ;y`a1(NxAFr<SN{bVnoM#eEg44&07b7xQUN^kTA+7Tdi*l5mZ*s
z_~}8wraAcsYfT(nve7<N`NL-ptY1572Km5xBQ+Bpi<-*2B^@B30mp{=f#R#L=xn=Q
zt0I_BC7LVv;0t5gPd(Z^>L^2?e>jMA_swOA?fQC02!8l;t@VHsy5hr4X2**Mv$B@7
zmVh`Fx3#pX`<AgFn*A4VVeEsHJ$1($u>87Q=%dmt(-Qan{?$n%?kmFt9l@!^HtdJH
zksDwA4=9Ha!5-O`*`VZE^2L~hh|<<EXd*#U2yi(it}X+>$Hi!guUD1W3iT31|53$E
zUytQ#Xz1zA)5?G6-68b}&X!oyo4sEbSi?kVEkH!(-1p!O-*;b$mlO(4$%5@iyK)va
z!n`rl)sxToA5;P4v8c&TggpbfDCFps>nw9zCG=FcJ<FY5J}-%On-ToRl|OQ#4O4cy
zaQ_Q1Iu@3e!iLIm8XE~Lq1f*C9UAsYvafPgXyZfgUI*&@7u0r>*pRQ(zKFM6mAEu;
zAZhlDqp3~bnw13Og8ms#`+;<Y^L4OnJc5hPanwI8wC%A;$u1xogaj-I4rirRAv&rH
z3&Zrge}`mESA*4w1+Keo<EGdNpatc*JoN<BrX^Z0CE9?im2bZ`2(?z4%lj~M`|k!>
zYVq+#lMm-^Tv(V*D|>$Xlq7}C#f(ZK-O^g>>1}Zxp6}nAauj}bLV4Lr;<cE(%)a+L
ziizZH$7r_Dn9Il7gWen4+zrgodTM*Q%WEquTbZWiTkY7DAHA+icNmxpcoYZ{=05Fq
zt>Cex%LhC~Tdi;a_#f7F)#Ma^QrVX$qFf|Jw|d`KB?H0;Bcv2u36^f4oWry@6nU<*
z7pegd=7U?)lTP+g()44O+5Awk-rGJ409j+33oGl>255468;G2ggxkm!eEwo`T3%kh
z)Q(SN#vSrM();H-5aGq*wtA)gpq&$e_{H!b2mJn`e0o$ThZZ!=O4ZyC&+ZfS@rVy?
z^>3>8o9W-mex{pYwmt*v@JEsTKP|w|o=mh`v%?zDNmxmH%xBjLDt;6(mjK5h(7;<&
z35Bpj83#eq&6vzLu-(S$s<&-2TEgz)3d;I7*!6t0X>LvYcy^dRRu5BSPU6Cz6B|-r
zbW^1w*o}JeLrde|X(;@e>pJ+}-}{MV<H8@Mi+ZNLS9Ix0xsW&4%X4eo_iFpl{<0VI
z1|T_dKE%gbZ|O#@1w5ClAX&GqVTJ9=+s>Omn@m~OjL5ITT2%6r;Qw<yKsxBSdCKEW
zAidMeS`968%|sO={7i=tv7)`C^|EgB#58xHQ1|-|u{?5aF%=~p9o;b!eLC=-(7n-Z
z`Q+WUwq|6%bn*Kl_}^wu*v)jPE%CF4-@xvhA@4(-3w<O?q@%Z_19NpOEy`j`s4Lz`
z?Hd7k3YYvC^b{|S<dJ|ma~_X~8#quCvk~FDHp_r+>rW+!(8(Px=h1ixF7&jJG#X2)
zh{XTSf}CCjf<U{>PkD<38O$1O$)tbN9YfP1qQeW*&K_5ft>^i(m7cew<2Fq0=>*#d
zkh<zVUb}O+Iy-7FN?=V%aS4@rxZ9p^IGQ5EjJ;R$cH#MClnEbXL0kW(WGq5uf;d<c
zjt>efh=trSQCMx4pwzfI5c+kZQbVR}&+2&v0UvayII!}Zou3b1w!5q08Tk%2u9KPF
zvs*IQ0wc*sBzmZps>}4&+Zx74J5HtN{4Dy+m}C>e``}&qy}am)O%6Dg5)n=B!f#A)
zu6`cmxRmB+I$p5rm?~h+WCpbL;;VT7Yfcu&ikYB)ZD`>(!v+jg9`Hmol3to+mMmd>
z37`G=vy={g()n@P|JhL(rtlZNEtVmrX`Op?!d)Sd6T#PDR+Q?9+^uK!uLEVD4LKXj
z4?FW<g!*lUgdYV8-XcK*+7H9}ygjYLirKCR>;~R)NGL@8>zFX)Z{guqW=Z(KwX-p!
zhx;EB-QaMR+X+@#5wr9ZrB!E{Rn9V-1xjsh`+WnVj?v9!_ER4+8?4wK&R>9@Y_w;6
zw9}aH0pH-EyFx?d;1a0UH2<%-?fzvC9UgT<74z2<B<--DBi3$E<*W2Ow1Zgnbx~pb
z0GbEK4-rlJrsKVBEl2f2euR=n<j3{CvC|M}_uj)-@xAtSmRSj%qxO?YmHMa}oc^wY
zI8W3iAMReNlkN<Yb(V9qM}av_nSSlFbNXA5vLXxo&b1y<EGp?OseVBd7c5NI6Rb}V
zSJz%+>8Skd|M1~{CDQK~<bKGV|6HOkg;H_%Mw2dY(qFVGS)Z--(%tkw$&Xz%Sjv9P
z5-*n0$)*efL`9kIj=w=otj|78(UX@S13Ztu%sKkg)XJ!Hq+X+Wxu=3w?U!~ijn;|_
z)>lgZlgHtf)rUvjb|V~0Y|y7?K$sozsJ`gAs+ha`@y6WDcD9};h&ggsC)Vuju0cKX
z4Wx{>uJ{w8IeeXlj#UE%*7d{EG(luA*z;OCCBN$6>tOb1LeL5$^>*JV^JiCS_rbP=
z2+2^24x>MEWfqJ3!Q4R6*u4+b3t%*O{W;p)RaW+D1r6Hx+sL*3Eux^T%lX+y%u%j`
zFt*!i8A1skqn?yDB*u1#k{m1XH#m9p)fmhl+ji?$FJ_qMAzeVZ9T|;~T)5?H9`&J~
zka__v(!k=P5A+`uR1HK5zLCiG$J_#`OP`pd>V`I|<h_ttrp5aZqM35xrerF6q(Bn;
z<@y<~@~yJ8KG?{$a2%yR>fwKYHE-Q1)B7YHX@*`uuj$u$WkWLc1!5|D!Sn$zGlKhK
z82FY;|AArGAM_7$Co0lDY3F4DM#Pva);{T2G|0%-a~G*hZTiv8SYFF?&fuwAdoR!p
zjMLzE0}<CemXBa=zK&<fhw#41($!-9R>$4^%t)yC*45`>rx4R}R$$JP6LE5|>ELj_
zMFUyqxN5m_a<1h%YwTEjTO4w-rT7ay)}1ro3cZr6RfvTgzOXP$QW4=*=orc4Td*qM
z+G?%Ge^#|)?Mlkt?vh9ZN|Het{x*yKLR_$v^#qv<!@$kWE!f2E5wdhbpnG-40_86f
z<~u&Az`Q!RL0fy45N&Jn`x}eS&k6e_KlhnlDjA~9IfHJfj_cpppS4#XpBF7vcJ=^C
z>~v;v!c9)oBzG3-=O60sx1{xn)Lrg=>gx-M!n(TYQnM2%1NsYHbz4qrFSk;R|Ds4p
zI0t;8%-IknKFX4#7e2IJvWqHCc5But=gwwASzAk^d!7`}Cv@ywP~r-5DvD+#n0tq-
z4EZ-%CCEVkxYK>i#!N}fCI^yL{E%R(UF?g%v@?Icwz*bbgp>Ds{)#bR(j#dxwX39H
zaDmi%{qp*B56d3Ayyg0f2XCQrt7+Hb)yX->5`{g+mcYS7Z{==w^NmAG>+Q_r6sh$V
zjpA}h%$Yxg3VRJzj_@^HiG3C$!sg~LssVInkD~LJaZwLc6utVk7eDqj7^IbBI7s->
zbDjJgr?P<5d0R=<c;fbcix|ThXp`-ye%Q8fHZI;}n4Xgn*xA#R`B90~`<#drySTPs
zJJlT4zTilntr4}%W&h{f#SqipO(5#mnTL_*ppLvg>_!`sXfyx)|AE?EIWZX|ITvcN
z@{&RlD%i~1txbb<?4GYR_R6r?qVoGelgqWYsIC+6tnPMHn`JACg>Kcbu$bSKK%uN^
zuH&N-a612?)JVpn6x;&_Qao}qlz*4;qo@1zwPgx4QSK#+fu^5?MH3?|!{s_dxd?iQ
zF@lD$nSZ0L@B4)VtDfsm3J=B74ytH>S?Mxf^QHgZ<C;8W>GIc=mx?rWx;a;wZ?nTf
z6$pvRrw{HsESxlY=0<3|+v?qxGv~o0f{ZF_@*27)wnJn(2r$ld1*_1!PJL%TlNZnh
zmwNm1+89f#*mQDbo!cKM7~+>GePW~uH|O?Xd$+f7<%CXcajZCj7)C%kx`4IPSqpDz
zC?Ue@)^1KXKGYZg!+iT=PG-b_mErx-bC3}6iO%TtvlMw6+=r)K3c%x%vDsL(5K2r5
z%w&Z3LB;<9qx5f!$*M6H(QgGyOT5ZOHNl)Y`lo_cyH1&58Rk%fo0)9iT8u1huPRnt
z<(CU%$|IWX7yfH4MyG@CmeT6<yiG=HayQD)(;b)oEH$)~W@-erjaPgwrDcnEgL`rQ
z)XVw^;@=z>UmrndIow}Nckxs+6u%Snf2dWS6xoaEW5Zyfw1hw#s}CF_u<C<KQv+D}
zcjUf|=1WsXqCPK-957!Azd2C%$zD1M9BWqOasjIdsCAHJ!<t@DdTg-PTD_i4&fWy0
zYjG*t)}p3><u0`swWLn-OAoqIE@sy_Aj{EEQf?1pg7Rh?#=_Vg^qD(12e(IZ^W~^t
z`Mri)tH?`@5&&DZ05o^&%~!*~lQxGR0|4QWs4?z2r$U^f?zD2fPT#dkbnM?y9f7*;
znhwTDjM(y}fKt1ghos13Z2JmfRGIJP?(rVBjwfZ7VG{Xs^);o4e9G6?#`s#D8)*OI
zty~sa9Z18P++56Ga9gkpK3K?0BW3IzEcA&A9;vxzhw<A}?ln4JVB%g&g7q?gkBI3q
z07^H^=k2z|WRxJ1k9WJ=8&R)*!U?JYE-TE{C8(ZcZ+~bTtS!2Z`V;crYR^KprS;*c
zdF%w@#!sP5fy*r4>{FaC2#_8fMduC{@aopJv6F%bt9@I390S2c{aCdn{#xD*V?k?U
zGht;T*Ek6O>Vtc`4`0lmPp#`)4auiaajyzLW;3A^VV7oS#|2W^Bi(Lw{Le>U?ydex
zXAwDFAy!Ig=$7uy1vLTz<-uR@3;6ks`#M;Pkk<D)`sdO;rS}OI0B{3;p(a7qQ0MQW
zzikzi5|e}xI!UP@@S}9`y36HYt6CnD6R;Sd7hYRem#-p&=2Dgy()-Q^0*@Z)1a|$c
zLtW}}tj}Tpct6uvZs-4>664i?9$!s7En07T(+v;O18OIqZVsocSQbItVPv1V*k2~x
zAJm|q{FPdl1W1q;qNagO@xK^A{ng91yVC=%Qx$vv;U*E&au}F%y9qE@rjP(G{6eyG
zI=u1%-*@PGT}o=lzRMg9Ng{X6E$0?f4Sm;d6YJWbUOIawJc&`BKFO*<5QD`s%7_sw
z+(;=I&w6OHNDh=z2L2xwKOF?q*asy)I(?y$Aav0nMcDJDI&~<La@CYN@M2mUyUF*E
z@vC<%VMN8eBa18@*!ry2dhCd&B_4;$DNIP#rqY1BE>qHKn}$`~V2k0s0{g91e40Z#
zTih8P%L%|-c*8EFJ(6wf)hA=+jQhLi_T(6agp{z*Ct?BUCqsh-(mn``<00qT^YG+V
zy9708_+jvNR2MfOj@~oCYF&x6bpR#E8{EDr1XgUIDPr3Vu|&B{uOsN$n^+d=_!KC1
z{@%F0!Bw{ZwF^f&z>3l{NfJ*%_DDe$eU)DIIS;4Pph;&87z2xRx&3Z92w<?=9smi1
zdV+3ZOSFH^Zk3hvrL{G_pf^A|;znYN467@tu_s+X<3$|kinqqfuhdmG5qA0@*}L<d
z4=H_Q1;#@(wK(Mf)C)&d9J$D7`!e)HdaGvDcVjE0{!_b&iUb6YY>R%To4S02S+-@7
z0#k!zx`8JonZ7TzV31F{xf|r)Ue9}H=%bdOyTNF>&Tc(5b#-E(*iO6Q*HL+So*0XI
z|H~1>jzMTWH3z`6q+8n~&Q?c2D#I4w{s%&(udK!zH|9K2nuS26A{6-@EL@L3gkE^A
zpQSF0+^F60`lZ&=!*+w|)vwpVkF*b7s=te?agV9GC@;<TzQK=~=+n(qegW+X$|)j0
zDtxk{e)C)V@~tO<h$%+1NtqWE+w^;ZxGwknoS4+hV)V#}53UnIV}h==?q7fs^%)$v
z*9erC^V<O~CYHzuwpE@6c7lB`@sj{6)|=5H*(jNk%3Y=C6ZsbGtJon6J#u{=Mb_Na
z&ICmAU<?mI<ds)MFQcp9u^`zs96TV+G<O~VgZgMI`aB(!o&L`oCwymG)&V`VK+mbD
z*bgJq@efp>i`c<`F0mu%aDTlCcvJjgLZJPr_I+qs{W+#?oz77@LLTlzCa&>OWl?nV
zM*B<&wjo-+`gN<W&_So!a^uanO`fhnq<1M%XlV}<0*BGc_!aLdS!hHDH+oEqnf9?L
zu(*C!tbF=qAvXYhM-<gL5>ufx2pbp{+CXm1-N6xTy*N0qg??&gGLVQy#zL=ve1v1f
zR<H>nr^7D3xO!{wwZ4d6>rlKUQ(z(uj(nE|s5dJNWU+ORxfW~JxFKDA+Jjn=yrpC}
z{clf?Ysi^JZcm4PUgl6is5skWQ6>CP<K+X_aZsW6RZyE-LVQz|UkQzwF6(Hf1K2@#
z$8mTK+YQ}=CPI5L6rY6+sXxIMV!a>#F*Sbk8&z#-PHd7szp)r8`PBvyeh_S;uA$EJ
zw|dEO@>3OEZCa_z14#1GFuo(T?%EM@&K(N3P&GI{+h0hxcSY<xFiD5iM$%LBsZ9%3
zdZmPEesnJ_MZq9Bv=e40diYWF%*PiHq)nBZgoUYm(d6X!f!-TG+95q=l^$NjU_QQ0
z>o|^TDAIfMP?LU!WpNq0iOSf!DE)J+#88(`z96)c0UNJGa6tBslsT6?m6rcuD@fhx
zp}NjJQV2S%L8ulq1D|{QaEs{AWXs8G<Zr*bS`sQ)Jd{e~HhuIu;eYRO%BN4I_Cw?A
z#-9IbFpKcz@D=UD#hLp`f}(#9hH9BrpZ6oE@i9-war`tp5^^6`=k9Gd*pqM=wQQan
zhbl3XY^SCiJgYKN;;+wFi4{9-9|kc8i*raQhK0%rVyHdIHX0{US%2fu&uy~IHM2Ul
zWnC@&ioy3*#c==UmE%M|Cyq>y-^tzQL8eYYp`Co>PYTK)whvBFkJEa++?>9|#KfH9
zuaBHSCwKf6J3~b}VDg^cwr7<K@Unp(>o<57#yqppMK!U2i5IWLN=iD2;xH>rS9&VA
z;4ZU%BVE+D7m&J=s^fwPBu%=wI{EwZk^^AwujxH7Fu<e*8h9xbw`cyeguwmrt9pyQ
zQIs>p!38)Hd-D@3w)WcUwT&c*rv1b(JEgATTWIG)m4_?vTbiA$j;CHZAens(W!Ca-
zC~1!$M25x=FvrBAnw-w#zmAae#$LywgF)cs>uKK~n!;^&@lr@Xb*<C2jZ@eKa+h21
zKaABg*V?$fI5^!Zjp$AiGfy7s%GPd~<#V1QQ^QYwZr4{hPgC+O1k8BBe?B@8;41rp
z-Jh01;4zX|$mq41;4wSCEK2#edUuTa#aREv8&Nf)<XnC>p(VEwk)K76XX1%6D3oQN
zD1~@yCol>rei=lI%3~6%E)jT;+gNqhDZ<ylRKqM}fqP5}J>BQ_Yy0fZ4l}D<9{Gr`
zml!rO!gRX}+U!jJf7ow-5I5Gk)aEzl;7s<$K>vT#^w393ku>)Og2ua*bF>eDxO$0i
zYxk;|fX;O3)?!S--fHj%zfd|Bx#eePSyEb%N1G?#03Bs1AS@)a?GC%!e|SuiGu@D4
zBVu7>y(kI!l*jJk_5wyLL-9ngSoU`wKDF>MTY*Ymqvcs+DSVhMZuF|z5@?wuJj>wF
zeR78vX{92tzjlv<Z@vqZkeu>Vlq~4=l)~?*`k&|jn33x(zW}JylsBr=UjA(>Zed|w
zz_^?cUYz~&t1Wz;#jJku*f5^&k>?C;3_VQI0MTtwTq2}U7Y&@O&f^63!irnT%{~%&
zh~UE1Jb4<*R%kPBoI@WwkH3wDr@@)U#Kv*(`r<Om1HVHfmk&9AVKGzfUGgjQFOsU4
zL|-pBf419IFqKJN&*fUAyIb$t^z@4IwE;`f41H)-nLIs?$uHLdO~4TV9ov$fA5P0b
z+Yh2_Zspr2HR%RdOn#2j4XxLhXzqV>rst>Zj8D;zLj_7(bNV#p+(+jdJdAPaj1jZ-
zEdwCuo<qVf3XXNNYY1QG3We(jJllm1v<7?6my*WhnjE`j9!OyDr{}Zk8MyYM8sIb^
zAr%239qk3a%2GQBkNWqzueZl#3l2H=488a<{DxrhVV3fbp1lQEjpC?r{6jyuP%-6)
zmjsiNlFwu@gn;xk;Ct^1|6B3H299dkzQg(V=_o9&93E%FF#_Ks!S?N-IJo1ifI_`k
z5ttN%{DJ&wKxwB=`9;Udph85_wq5_os&QeRE%T@(mZo<8`;RS~R`cQu=DU|)Hv%^T
zil~)xC&xtVMwFatImlkL%~jFY^5^Hkd_CzKU;jnZ-nu?*fn9`Vd4DuqSeH;U9R9{~
zyumFm;5Qt15i)kt)MI>XwddCgx#Q#?Z=|p#=?1gGLrR%cVy!Kw#j-uT*eNW1PE(;V
zR`k7+5modsl8#Z>U|Zv<ypQ`mvOa2jT}6%UoQGsbg+b(fm75$TJ9<8rE8IRdma*ha
zB==7awIhG-!f0A3f?To^=snm8Rg!W*!9|OQVualCEX8L5g)6BN;l>z0(fOY2To#>P
z=WK0h8LkK=0Gr*rPQHo}H!0t?@iRy0*Vfm2-F|e|Ht+`n4c~F9L=U)2ALd#nV~bhZ
zVb)~P9<ZqC@=J>yfCyGax$cH_cWz!3KCT9)yr!<~q{IY$qS!&x$(18_B4U`CJolPa
zr#1$iD7cRshXb3(&ofJ!HNlTs;8Bz#f(H3p=&B#?+srExz|VmhGM2Awuaj$t$p*Jl
z9eqP8*IM(&($h1D9>(tGLuTVy?QJWffi_~cwjf9C!AuBj>;w+EL&+a(77cVRRZ#T&
z0!<9+4pzR_OOim~iR16WhZ4?7C$ln4gyWhD*ZwLgp&ZHx29)Zza|u4gp^7N#UvVw+
zZxwpnw)>Vx&jE_$rF(l)#|WO}xL^@Ntlz<6fTZ^K406l*I^H}!Q)>M3bm3veY=`&I
z&4bfFgwF-MdIa=8Boe>eXsU`S3t*h)o)8cs-M5#t-r@G1Q5#By-739LzpDz{C$^na
ziHuIK_ZE&rlm=@NK5g6?uQZ&M7jl}Fn-Mu^m0R{$Nqc2C6cPnV7yI+Dq9-`_pZ}gP
zhb~61;^3YG#Pz+@|2JfR-vKa(Gqd_u2F{b_YUksB(7@+dNU<$Fag3}s553J^*Y`4L
zfh(GV;p;cAP0Oq5$66NNTEAj35Io5!7{uFoSf0%mTBN9;g%R)(YcEYYQQ>zFQ*f|s
z#Q8#p8dgFgnwIXSH)@Ve*3BT8i^k`B<8eoqeFjO2!&k2ftf||%H;{90NpR2gX)JWd
z8KRgE?20t-7{&bj1bb7WdI9;!y5V=q8BGV5-kQ_4&)fF|&Q&!v^QG^rXY6NV1&9(r
zbRLy{=j?lE$VFlg_x%A`OLpDb8|UUqCjM`<Tbk7TNwr@ZgZayCf0aUCsBjXB%3AYr
z+sv<m-=Z#u)(tWbndbtxQH(pRR-_Bs&B~Q%FF=6}jGcFxzO8mQ^`vZ;N_qeY0v=#i
zE02r@q#o^3u}~^hFy8_}+Z3XO#&lY>{T5^cjiT5ijkv}7Ps1PCL@&Nz;mRXcj~Lp?
zFoRI+9)`QqMfNsn@U#W3rx>?i*H*xNzDN!P)_NThF1qg!Ia;54SKY0#c?vL>V`hMT
z=|Vhvs{T@*rO2Yh1u?OEx0l~5b(6@k3b^f4ct;PkKDU491sLi+4rd!z{zfm4M{9j5
z<7s}v=huec4dBMVm0b(wFoFoP!GT;UY>^sgLvOj>0KINqRd%$5(%wG+EvOKF_hB1G
z&_IhB;8hd-VD@bAi8ghOkyi!uSdTHV{X<2?{n-RmvGj%mF-dP|)$QBJ+_r>&S9?tH
z0hogFO{1BW17cg{l^zdW*{L9N#uiXUJw3oUzxeF+?UQtSrVu6s_EiwGN%U_~O54`v
zDSKUJ#Q1x|aw&V+-$ti`3>kyqZpJM9S+|n97O~x^pXKr2v6GVeVd<2j6#3`cGp?fx
z{ZR;va#C>*xo%a7(az*<__J|Xh(ja3TAD%Ft61N$maRS4ju;^$Vdc2Lp2mdNe#`hv
zlKy46?upqUKU3f#loejYkJanu`Y&B*XI2qO!4m!*2MY}1D8t)Uzpax#6CTNt3lQ>Q
zQ-;MNn^4QwkQ!Lp-U_$swqOy-ecDJp4tS4LJzEh8=z13v{Bn#on&%HYxfV1$rj;Kn
z*9Na9v`*V*YwO+pRucQL10^;IJWdJqQLl58gt#X@M@gHoER?yXga2N!ES5zu^abSg
z781I+WHb)s4GhFt*xH`LhUu!2`mvkt2wfd+GxzVG=FU-vUJVIVltll+Qdo1H1mxYN
zz%YL)Mw!aTe_!P%X#Oopyor9!*~T2(`%>hK;8dv`SXK>;XV3jNT4%%Sat`XW@lp(V
znrYl{Dx~5U&EoMIo2|FV;5vFqJS)@)ZWt}(Z#s>d>pIGW9&6~*xWh1TBD+3FZdt<X
z*ZMHaNa?+up`$8zQ8OuYh$^mbX-qu?&DPmR6(>Z}Rx(lT_+;q=cvwVM>&l*{>%7X{
zN<@$Tc0vl#vB=vF+w^G0vzCx^hOarFQ8Zf5#Ai=fc9;C8bj)vaeV42k+-$ABsa~?H
z*zh$%fpG_Q7w)CFwt!0*WAB@j?i_$Wkq)HtjuGI|{GDh^Sf<%&M8Nwyr)Nup!gfJ9
z<=j^q`aD-_lHO=0ijjf!fCAc_Q;AfR#Sd>;?gdvLC*2+OK~iD<M-~S9PkXsSLRT8x
zfz>ANWUY1)LK<M^1U-(8r-C1Nl({Rb87?zO=JG`a_Vp^?|DEF8QwGLM5<}e%hVC3T
z2HGz~rBfD+;fU6|>kEs<hyrle;7nO%#jlo|ZK?MgDIaiMh2{Q<oblOuths#W5D(sf
zKORmkP2aX33Xs+a<KXbPDJdzjL)hExb=N<1<Ge|@^NsF&zp`h&lxl%b{2Jx7ot=ar
z?_Cbn>*pY4s`nh7@8Oe^lN@b_)`+b~PX)O1?a~5xT8{%Rihb&HpSYEwvz6sNQl3|I
za1UQTE%1fBZ;%R(rD5NM1x(9ie4|FPGLI`{qfikU5<l;Uu)X^HcWfMs2C+7OJpeC>
z5-IVK{`7ZY<&A&1fxKIq!aH=ZyX$W&er`vp;cpodv>@f{dHTc7Z}V>o6T)&FEwcE6
z2k||n)OM>IlWK`Mx%o2HU*;hcpyN&=UFfgg|LvsCmBL_YF3*&-=l)IR|Brk8a%6KF
zx-<7wAtKR#Ai{MiP&A)`nX&WkW%}#x55R|^AtT^S{d&rGfaA=VIK&Y6q}C7;MygZH
z@Qoy33>^gh^t(Aw<B%P&uo>cBgygx2YK;*{&D)JY(4zFfnGTB@qwZ+7h=^8|pmk>>
zd1D2x5BJ&w8)}=0>FaMiyP_H2O6DG|Nc}ko$y*dl_@_cwRO9hLovyF-=D-_A@^vh~
zbL9Le6F%-`t=(@*yH=TS7TVqIMIqWv{;MH%x+|sP#?aS}lH~@C;H*5KRb>&}rs&%A
zK3Ht4dN!;@)>bw`JyeaM>9N6+$gMM253*PuBkD}|O=^o~B7h)V6UHb8iC=6@+}ary
z3I8Bb%#G-y&JE6xsGGJw5Wq=Uh)(X+ZiL&(DGcWn`D<>x&ug+IQf89cKDoXxvFE2)
zcX0XFOa<LH-8Ss%R(r0K12@XhC&K|BW)v*ZhLcMD)*&F%nD<T$7~~vE(*%9YI_LpS
z|5_g!#do<g_jH}_VO&7#-8{Me;Ar-LS^$(CFID}_@<Z5jZ{p^G&+H<{mGc+P6MH?#
zxrRF48r{&O#+plVgY4OqAQ16Eu2YI)_y^Hdb^?cydhXTwx(2t!QZgahsci>)S`^E^
za}FGmI>h8YOP4o#8arbG^;uqsak#FGx)kli8tyCc*}sr34^RZ3{(Qdf=jAr126Q1w
zbN@*Wog}F#-YIKzb+!2JZMn`L-3)}0ZNFkN#Q0-Ij0G%&Cdx%t<E}(c$FGxAax;}}
z<?CcVoOt=3cI~l0S>Uu#C22Kn#l^UeUXg8WZf5VDL5Safrg`l=(`7V1Km+|J|ID#7
zK8WQ!LynwFWk5DMbwS88G6TXv2^F{xG8XP2yS5#;ART9QPOnJWB%Ht5^hAv01<9>+
zm7JU1u371nlt$eyx$S$)EH<6-81#7zpuNGR`>L#;%V#j3pTYORCW>R?A*sR<O-@-N
z)Z6>63o>_;h=!h&F6}Je@5L`L{;I=qjdj>w7>f2F{z`~)WckZQOZ}H<+YBLt%U2uk
zLDaI8qjb75-zU~^3Y6^L;guE3CFNVteeRNBzys5f@kU8~0DY6_xHQ9!?MdX55<OK0
zeP6Im=X*cGFaVzNEF&sS+Z%(aK&iL?eSM@u&%@*UKqK+z&6djbrtNk{z>gJ(L|&%g
z#1vWAh!G!!_I?=p)d^UqJI%r**Pr1o`KFqVAx>rHr?)cf$Hu3F`GC_x!6xUjMzY;F
zJA-&0`m=4XTZ@$PLkOBdZC&=>Lg;L0Y#PGc`+*su)n0G`($T>XNN{Qo`LT5+FKV4r
z=|KBO=^(d_$d7=QRIT~G-cR88J4*wOqcWVguv07m8@NE;R4zE3hC8KP9jo9&f!CvC
z%7cCD#6nDOJo?=v*uD73otDh^@dQ#_^=iOWZIw+Tut*6}Do$0J&e+tp$bV~JF%?fy
zS}8);RNWUhIiVR?7(zlvT#nB`MjrLu(=Jw$H)`qQl;wz?|7GS#vDEt)kn~5-CrhLf
zNEn1z2TVp!m$VWWeLt7n78xgfGTde@=%<{NWyie8^9iG^&%QsO`8&Pt&Oj$yH{W1K
zrJ7Bh7Ex@|DyPKQ;4xBXcoq5{*Ucl+OzflS?awi4z<CrU_LV;l_|Vl~@Rjhaq9FhS
z2}>69GAFF|h@DA^G)2ki>L&rmXSiC66llosY+G1Xff}>?g2UUL(J=f^=E5jE;!xL(
z&n&x@AXbhHMgO|c>9qi7lR>yBrM_>b7lQae(p8yJ{r;A+WPRe85ArCe&0IQMMcL@Y
zV@J8$bC7cQyJ>VW${Y6Z^3iuPG_f=+=cIkmu6rh{`(uY0L9=A&6q)N%fnN7LlL=g=
zi?NLCfZuKtnTeT2TiMyR1k!865RmZ|TQcSy&D5B^hPBkE0qKHF<xg99q1pdNX0vBy
zc6p298SjB8WRxNG^rF@56Wf8jZZ0L(HARtPkUS)0coTQ8LOz(ZIEvX)|M7mNCd){<
zq8MYCnh;D0*Z0&+D0rV3T0Em7WZMr8!XXd|J3j;cW~VUVRS4mT%Ne8)c8DI#a&DLb
zCs*HRt(^|rHtVNbW2e{GSI*B=uvK)S5|9pG^lWSu3K(zU^Q2uT>c(h#K8s8U6ZzxI
z;2(%*<I>oOy_p?&Xn!TUc%Q|_OdaO*JWlH9v;mRkPBy?SCW4EQ-6$+-SoJ3s!3|E!
z_YtMW?HN35$9TM592Hf%`k0~w!s3mn{Vcu_DG+LF%qKd09&#g>8%CvpwIP@H-4h5^
z{Rf=do`&0H)LRT?=>upcFD2VmakTJMjHvhv2Y!Ct<P>-#V!w<ObBP6$Pz#6~M9rq>
zHNcD2PPeKoDOXO~&oF;-IJ0Mk@9k&bubF5;(hL9jH~FQSiW3vHioEoiIHyFqL+>Pt
zo_&0FL?z(Q9OQAd;J&Q0n-NL<8W&R}tO0}ifI#F9*iHA$CtOGXc*9Zse7rZRQce)0
zHCE3Z7`k7xRZU%H^(N_l#_fvos)+ijpgBtWsx$gK-Mw1pTRo*Y`w=LaLjJ+rmx20O
z(8{*gWvu|yN;~`QY>{y-DH~?1UA;fwnH5ZNXJfV7Op4!0Xq-Loy9|TX1)-6i0_KC$
zeUWmKM#P@~`BsU+)b?XdKhIIZJ%r{c4TF098<EqCB(i>SrN+3!liaS%NZ&Z>DR1X-
z@=W#I%`l(U-^WKU5BRoJuY?dn!rmsY$#b9ioHaD$v2h3UX$wY;X56w*Uj{VSUG*M?
zjtK}v+$<VRyG^L+2OrulMg-=QYwAr!LGHKUnyQNdq%*BV`5dYqt{15TrY;|*0>d3%
zVgpl6%c(y^3_pziP>>O3CL29^a1!lcUP!YR29*yYx>m-BZ_whg$y~_<991N<BZ4dI
zPVL9uq>Dt;V_$MUqMq=7Q`LSoT;uN$q(;*hBnCF+?TTY~M_&YHrHNe#Y)q61E2wLP
zvu^zJ6jX>qv`{jWSyP^0U81^G>GQanZU3922hmX@ImpBxQC0>xp?RUVLJot&Jib1+
z^(_7n9(3yM)lgETvi!?uoW!|W=Ky{?DuVtmGzVreU8$x`v+yjRT_6_f`ID_eM8e#G
zb;SAe$SXn$voZyV*vKApsZdOc4mSJMwh{fA=96P5qaj_Py7H;NoG{oSn@`KVF7F>&
zxS!P-o%5UNM-)oYZ%~Th?)B<NdO_T*9$20k$E2EP4oa%j{kZPJ305Q4qG?D!4R!J2
zM)wy`01t>}_k7I0{j2qf2IDwda;t<FC0>No-4D*797+0%UxIA9C+>~s!7WpMTWBQX
zZ+t~xsd6q9EnmA^loqRE<-}IBQ2^0q;8=7dJ8&J6!)eZ1OL2a^ekvGu+#)X0)0-zh
zr`d`tw1363=((2C>A13y7kxEQy8QchUopD9ByS9V$mCCk)Ej1bg#tyPsy9L;{^Go`
z562Oi+(i>RK{2vxvRcyAFG3~b!l0H_^jv{C3$bSQZ?PnyYHwZsCVzS>b%*=i!m<L4
z1sV`llNc!~Q6cbC2u-s-<Tnh-HrBxX7ez(>N%J_Z5eZxPnaF6&zz0dMw_k|gG30+b
zR$g)&<RjKW$L*{oBw*GZzo%%Z(fpPg$)7DK=jy5fx$$);lWak8(0Hs}bE{orJm3oo
zBmgMu`mQoT>0bW4mAQMrY-YkgF5{0~kIFW4rTp;x!4E`r-J|iEkS$_#dNH*czMCFd
zn7fCn-_o~=A1QMDa)dDxL*Cu@)YMXOAuf(WN04sdm6(Wf=dt<vcz<_e^A}eqTM6P%
z<n<L}I1v^G3>s5W1A6$;bZS?;YxD__^=o(EH$sb7;%Cc8G-VYf4IB7*4$ib|-#^qm
z#tSqW4s-1a;MsjbyH)#v<#nhUyIj$=qU_N^iQWF?f($*HdAw+VT#5RS!C+kEyMy?x
zk`x~kU#J(+qB(EQ<04<o#fpP>iqo4W-J*-4MLHBWZoA*3+V<d|d-$lN=&@UFP1CJX
z44)=`Zo%@G=+SV)MMmyqr_BC)Nx}eEeyv^cV;g!t{MM~XkLS4G6!^`OTo6t$(t{Lg
z)-yAtJHek$$tr&8GSWPj+DLc7>^-MJI2yz;x#<kf-yGT8hddnav_w7fNa+q!R9OGk
z0)L?YWu!QBf=;xSo*8X%4Rz3*<tt>#_V19)5wOEZ-oA%`jUxSq4bd&L4I(nY+6M7$
zTWY{fOae$Q4{LpXI?F{80IV^f`1rM|FKXWJYL3Hl_T&=cKRQ#G@tNxYNknRB(C}<F
zT?i7CVa#)f<jr<hg7-s;DsRk*mUz5fyerZ`*C7<|li*FZK|SISS6{bb(+H(?r~7-H
zgBWj+L{X7k=x9ihor>vuLRm(#xW-ds(ZuSLxPa)75VNE-V*~yI87DRt6!Ppptaxo<
zaJJ<H<uYy7GLS3-zQ}tLfeo6t&j1AaM3*bo5@5w7JKdUBMO<iZHi`aZ+fSkYx$zLF
zaeBgZR;Kk|z?X`jf&3LuarF6BOyY?!<3<inm-y;VL1LoL;R2Ey3APyjmx5nkSIh26
zkrFY!BqW~TF>X*WZX~D-tou~I4dDe}jV$7Uf*8;oVok;NG7p`o|HNiWCa3^A`syWv
zaF(gxyh17wlG!nhD#F?ZS=1cJ_!cB$=zx+5y{rsFDcHFMYG?xuHtc`1bxRK2%m+;3
zZ${A)y4#E)KnA%@^0WRB6M+vWwN3ZrsOd*tW9w!u)DZOaQrOUicvAYer77Ik_Qf5~
zA8%l~#Vv#A+Q)yzLh)p=g1nBjO*Czbgh1zX1Tx#y<L^MmGIF4AeVDl@`7#ZZ_s1WW
zZtrZZC<k!Ga+mwYMZ4sUEt~MXrJ|&XX5=x%5_Pt-*%ob`ujCHzt4RjmlEn1pAJ=_x
z+1q^#ah4rbU67WVc&@3)UyvdG*JuDGy2bdE`>70^YI2@x<O{teTC``(Iehhsg|?V>
z?G;`7_r7YCu~X_}dY;t$0nPVzGp@zF<bGHojZ*+3?aqYaz6wo34i_DNs#b<{vhio-
zg=Kw}XXRt83p>%<{yivSdLm*(_otsh)ObD7&9X9YaYgtI`H7V!e;Xj(5%7i=B%a0-
zU{Iq-k{J=3F@CACu*?l&=P7*4p1Rtm_MOT(>o~U0J356RqTe?%@x)*`$aM*GMN;Da
zm7-mx6XOd+1ZF{5_Ri~pN?X(sx7?1g8M~p<s=BRkX^GpuA$9<_-Y&YCE6B&d>s%D1
zIWErKQxOhRf<(;Sx$HW(lde1~^{q~cR~SbX29M!QnLZ%MS^fH58rb4t9-4@hpUW?n
zve~erksyux{SUd*2B%k~S_~DQg^M7_nj%LG6hq{VRm$#4K~4TVCg|P3f|&2Ga&dYV
zGDKw<5ADIC|23AJz_IC_(Mb*-^_e#<@eX@&!G%%Iy`ou1La^YW2k2Wq^hPc;hA><`
z{>$Iv#&J!J(62&ojbg;IHAPBBzoJk(5yv_F#SdUe0m@AEE!^N=|3cMGmSGh=9$Qr*
z-f!h+lNPA@JWodul|UMG!HlFYdeSUpx07%Gyy;n#%Gskmyw3dc!d3^?;d|K@puf;y
zTLot@$_+D2{Ta}cM}U4WdW>i}Y&%X)w}+vw2mt#x4@#d?fmaN>OJLV#L;5njUuc)|
zNjunOOgixK8v{T4$V3<H4u%muSC$o|__m+%`ugY#UcZC(18N>`)*egG;JBgA!ho1Z
z*UUlIK~0^vIhsm6<d+W-MaH7}O!%*2_=c+s>_9~DBP`fjqeGR~&*EHzvjdaP-Xp<q
zO;=xz)`%&H?B1|3$B6tx5%X$x%sqarqANp@LQMnlEvO%j<JaNlE;rqQW-y4@Glrt|
z&iW(!KS2OD5sgNFvD%GFdfYRdtpdV%#YT=SV0J*o)WJU&IGBB}%0-A+_PoCY8?YRF
zuPJUS(+-!Rea+kRuE;|)0RNRZ?K^TZ0V6wlT+!IL{_($mO)mM~y4dA8&<2R{+eL`d
zHja5S_iN!ti&5)#;RhC<H7A_XLchBa@@9&D!q;Qk`Y%w)y^Bh>UI@a$7rN$P22{jP
zl#hx-cvZMlGIr!&)PLZbGE&Iv_uA{L5%%#)K_(-Dguwv7QopJIBGONOEB^6%TASti
zpWXnB%JyT<NbgbFWkdo=Q;SWf;Fc<(oX-Jt^-|V4TaSJ8qUgrK=iYXDBh|MHehter
ztVf&kF#T2e`3lC2H{ve$y>%(Zw{f>a2zzF@1%@nr5C)tE82~J;6U!B1Z}F@c++#yb
zDrw%d)vw1ZiQO}t!L$vkU=q(KH9;kRsGNXUo-dw7nbCnvG+hA}QM5MYYWLo-TQqwq
z+;Meokav4SS$yx->XbqvyM;yaGCD2qO%kn$-c;oXzGbgjk^DD_mY{#adDPLy)7I1%
zyr(WrMEWxK{UXIfGiap5zEQU`<XQv=XHTL$|CT5c%hjR6wDsMMdfW|`EjsqUm;!WM
z<5dK$X|JDBY$`z{PTn+5Nbe4{4dYLRU29juUzIM2(nz1NzG_^*F(uw~ahwSkmx!Od
zBxW6ER_{z4`3M}uLfK+!13iU79_g&(kLLE*Q;}pYgIdljUysaPabh!^-!u_6j@+ws
zOCASTJX%}p{0N)xZgUmex62XIycgWRMyS?vvw!3=vYR|lxm~cmB}7ByEfC1tcenrM
zb;oQ9V?e{@^<y2i?pgZ9%dD(pST27TF8d3G951YPh0E+?d|5@Vt@R@k(Nr0GH}uC=
zKivnstPjCCq?dl<uN8}OF9(o=>(^BnJ`HG0nT83|HMy~4VubOi(<@C{_OZB>d{>%U
z@_e4F^in^Yh++VhAi4yBc|LvQ+4fR0Z(DGwrS(mJz|5xCUlbq4;H;jHp%y||0{>BH
z#-R|sw*dZd+Sc`=ec>zo(veXdUiA3`5>_0uy3D-dh!!=<#(Ux<6$#8^DH3dLsr6-4
zx3}!CYyM_LkK~;rvDW=TGOZ3Z3OpS1?kO}+{)>+l-k*5B@kKrd9bV%+&fw&GiW(+|
z@7<4@)T%}4a{bQ*A`3A%$9l5RJyDLqy#O0MYv+Xc3g2tHj1EQT?;(t#Hon;(A2Q1C
zOO4qKZIuJ(Sx-0CkNbJl?uCkCU^n%#*_FILVI=NI0>?NrR}qdM{;~QLFU62|Ns%@e
z3g_h+eG~B#&!_g*LgHPblG@HizL4E$I_-)|?v~34C}l7jz|oVhIW9ZFE!BRnU7IpL
z&x>s3ITsUs@AuS#E9E+<y%PT*(xSB^IBp-3t0O%$6p|1S8wcq_=L`K8!vXlV-3`E8
zp4P{##$2Ag%U9D94k?v!#`bbA$3+|ZkZeAXNsVMkypwpzXLzRNNi7AS=ylCB6ON(z
zUU;vN_Sx>`Usid)+){Kl87%{>2)G+GdNheQF(biAv8w53XOibV>v^ZtjhnbaaCuvk
zMRd8|I-BDBl2Scvv2c$_J<E8{TO8g~e*s&2K@g$f{N;?e7jyRX^;)6qfbm(K6F^_n
z;m5t%sTyqmC4(2yColF~DHc~E#_6RNwP8Vi2T``qom<B3&~PKIJy(_w)S$GT7JmaH
z6gm81%PAjJq?z=W0hAl&gr&?YXpFt?-c0->OU0TOZ1MUNQF_)=i2SHcPZ322oV&td
zm`I27)15AVT<2$xsOKadj9XpOK-~|EB~y(lR>S78+`_>S{~sR-006C_#8ui|9txBr
zlHARRC^B(cs<?zR39pg2pObLbc&Y!42H^meexvkD$0nRS>6fcqDr9t;3_#xhwUQw&
zHzUQ!SoK|ol{P4>2`-z@n|vLqGLsVckA0Qx6q&nQfJQqVM5!~0*JtK@IAv`QS*W4T
z7J4xsT(ZW03-cBgWn9SZ(~Y6lbKiW64$Zq#`t5bM@W;Buh8ZHdwUR;e>;0@B1;cy2
zroqtPVkoSu61hhV?BN5tKU|bvt#3ZMv$l})T-Gx;jERme+B8Q2bdnqqb~qJ`K<o(}
z>FWDGC~T-tDA3_Wy#u(t+ENzA8s6}OmVIuHEu5s_-}|~>WS^C<H@EqRaD5{om>Bs%
z#7HW22cl07ar%rPJK!1p!DT}lRq0YcDkZbO-%J!cT)V#eg%qeSzFEH&BSvSL^(;dR
z2k6(-6{i+uuyRD!(lcjL-n<On@c$h1->=Gx7%TKHd?8t=+$VTkBoHsW(Cy?m0+Tkp
zdPHw|qZ!Vmf>cOVC6EWCsDg6l9=eeZ5>(oIi&70!RfBs#W<3u4Q}zn~kE^eain9II
zMn&mvB&53$7*dAr?t1AGgc({w>23sMsG+-4Lg{X#ySqUI;d}5s=lst3)|&OK`GbGv
zW<Psh`?{{Z?{0k)Oke)i6~2AdMvjsk4@2njo15q2;}09NlNa+JN;08X=_!X*t&fH$
z3XRQI4pw&>ua{#CzWM!QkEc^-h)gW@QZs!~tc<9m)#MFk(YCA7f>fj9^n!B)$pjn8
z^&!p6*Up?%5fSz?Yn@2^#i!1+^l>F2f#|rF$VY<k^A-9V7I&<$*LNBBc<9mviK`Ok
zdo41xlv@N7Hagg0lD!)>`%;(P`A!Z$NQJqb$5`u}F6BRb8<wjzcj+4fNt(BlEaU6T
z=(wO)1S3rgE7Ro~N!SK4t>U;m*cj2G#*M9Cc&Hi}MeDK#=k$6dax&OXb7!9uVFdL)
zWsLI@$VUq@g26ZC6V5CH;o~!6iRV7JasHrZX_W-Xi0a$4|1uu`DmCOxTPV#K7EQ;P
z#8i8`@U|)J2s@0k!hR>xx(R5WbOz17N;YJ-XOJ0TN8Jq(K}=*}l&zCfp_2rvqk?=E
zqrf|LtH6s{)1?@)1EJJA82qtt!MENW$1of>_<UNAeCtMj^w}#4t*bAv^GF|HPY21u
zw5QwJTE~D={dHURi06^}j-%wsY-Ll_ep%A=X_H7}+bS8Ir0O2d^Gu&}%*mu2BQ7k%
zrq|{*lgvYDT`&W^Z$YxU`=lh5<Tu6leD{#*C03LHkmUM!$w?d1H$!k3h4u_aT^&oP
zp@ex@A*z_p4qL|!$748zN^Mr~fNXD8n<0(@S_H+6GbkLy&LV`Ms%1)lh=UO7byp^U
z=!5L;2Ip1Qp>;fkOdK)_dT;xPr<`kL!q_{u)YrDXecN$O;%_73kE!?F*S=nPnB3Ui
zd;gAw5GLKCs(6W_sp%W0XtSRGZaKYcu@a?Djb(xoJ{5X^h*n}<b7lE8p9C<~WiPjg
z<G-9zsmnL7Vv&Htr=keshRCsb$)st?>eR9bJA<Yy1Y+&b-QzJ6prP%|=$))DttV(!
zGj*Rq|A~KeQi~@qi}hGP%R`CMPcm@KJ>~(>YVTHG0!=kq-VFx>1=m4)h9Io!gL5UP
z`8W50W4qr=-$h$KOdF2mw}pu>+*ujub|jFv;)=;KQtMy>3A4J%g{(X!`Ul128hGn6
z!KB9=FC=kcYkAO7r(-0P>Qq>4I`8+Nohg`y6ey&SBg&S*uoR4rqG=FA`F*SXF8-^G
z`}<d;Z3ir;9>Uv(r6W|Ny>s<o77@YO2F5rB@}H+3+Bj_Fm&={WkL2MHhzb+p(%Q&I
zXkaJvdqXaWnvQ0951JI5FFQ2L*R@Vdz}M{f4kt=-eB6_);q5(Fy>w5KYAqSn?$AcJ
z+1^S{6Pd{8`fYtop?^^TPcOqTgU?g?=<!!f+9bLF=EN4IFX4tSI8=S9C5Yj#<>?Y1
z=CLFDK5H`N9qQQ_0!XXdi9SpUg@$h@r$AsNsE1N>d$S4QdtJ|R6L%-Z)_WTkJH$x8
zWlbNF)M%xJS21+ddgO;jK|4ibwq0I+ZsoKL(@ky#11~LiX6L0oFD4#!z=S=hRtea(
zLU7g(9Uholu#Yw9np_?{JC!_acM(Qd3?lN+nIjcditdGI&PW#hmXqI`C#WLQ#EcNV
zie^Gm<rNI)R$<4K>Y4hOz#W}_HM3R`^9UwS>zIpED5B*DA&H0vIneN9hOrXym3H1(
zHLS1NApOK|fM+*$xEUUi!A+FRc{G?g)C3vE#oF$D7Tq?fpznPa-)G%YpkHN;ZE5AX
zKI{pkZs<~d4qxOCTsRW<)T?5u_Fz0oPQeT3%~zE)%a5zzriftrY*QpL8aG&)+lYN8
z`5g324DwtQ9X*P!b2aZ)WDpeoZ@%8fd0Sp1;e9j!M0YM2d5!BFw@#JWp5{cgJrSZ$
zye3G7N&CCOa?yi)av_S-%~!M^h+_ilc-<-USq?F#Jea_ocp$debEl07<)cHe{4M9F
zkf;mKfZh|Kx^KI$UVgp@kXTk(B?|Ff6bRz1ET=6~ggO!w&vJ^$bDbWpZ2eva&#lUV
zkF>t-!uytKVS{Hf6im&$QEdFh<P%Xer6w3heCt94>1^_NCm!?boZ4dLZ&E}b-xdlA
z$>-18b?Qhd*;5Af0Bq$tAYNP5{BqRzhI|XVn3Yj{tKAPDMYi$HX_#w!jL$}|W(Iuv
zv)xwXxN3SwotiHtbCFWJi)iEa*T@~)D#y$T7;Yp3e0!GR1&WxLL&3W0z$?L?uk~JZ
z<*LS%_KkcEgF&BGtOH>RpW`YR38UEeV71tiR43YRXNox81wY_6r7t=rdu4)0;u2(q
zGF{Tn37n&;pXf_d;@2DqjkArOy))%c{cji*Vk^=2Oww#=xqA|xW6w6$z+XB{dGggq
z=z7Vh3dBM{VAHKB;zjI7ish@Y!^NP%XND(!sd)Lnr`ZOH;2)jbx_)LvTO<Q$!Cl#K
z;L6H3bR*}W;>R`3&BDAkv%g!mor7FLnUDC99e!)d(2azrUm-@4OaYm1<0>9_8U=!@
zC()a98<F@+eFHPaboZ5&Ok83Ct?|BCV5<g_=0Sgv#O|-}?f=}&Rw})t3K@m=8f$=Y
z61!E_!wVh|X~y-Y-3D0Eg5ngAQ20~9+qRCnP>k#3_r^|#K@eOv8H(KZ*(*N!u<>;@
zX0^~|`@*uWxSZbP%U{Nsqv#Pb0vGhtnmYez8RUFji7VgyNQ+&^`|}O1!iZ4PukOE^
zI+XCDg^Cd6FTyn?J4A5$W|`1N<wEi40Kr=i`OE<!(?qdP5IChg`frkKAJ_#SwqM?}
zdDx$S?ok+WeX3ty{jC-X^YjEGq9ebSo1h2L1Dr~ek^ClAVQzg=_MEf9-D!#ucGjY(
z(TpGE6O#zUf(zyiO=KV{nvCnx&8q84r%0KB=A5B^kz`JWhiuYSx8<ziQog1c1UW7X
zOpPzT#nO6oagK#!869gKFitYvzPP19b<iBSVPA_~L!2GDJqDUB*p`l_9<Y_93llEF
z4*paZt18K!hoQ2RJ0NLhdEY}FZgP|T-(CQ&ZGh2bwpDKSsoY3mn#hG<jeqDW-|$gj
z?~4Tz3~cdos8op&4&lRBquSmzY~Y(_Tnm-AOL!;wJl+`VG5U$ZB@R|=`3HpY@7p|9
z*P?)YMkT=rF}_TA@z!x_uj4ll!trJeAUvo2NTnrA=NxEg?Z5cvOPJ0ptkf3<z_}&T
zSgX-y7<q6cZ;6r5JbxrnV-zTPBjLcJY&{OjQ@ZHM^5!N0br(-v4r7j%^qO&%dj8s&
zop-KajzPa-PR0?poH>2t)f=2t)5l;~IKZSvjVn$EN>CryHx9tsuFyGIjP=9(Dl8@M
zM5-vMh_i1|87}yQ5Ek9%9A^FcGZLwGYYS_M<vRA%A>n!-UY&V#;F}jttw;V|ME@Kr
zNs5YMU9KmO9@zU%5Pgfe8kL)OM1397G*wlnD!H|}S?j5YTP!O8cs@Ykr3cx3S9zzd
ztYm*n_ft$+?)`f*-}{tmmkKq&>#YdvFHbQM3=r;H>4<Q)IDgk-n>`Fu0g+|79f=ds
z<~oz1^K9%ulcy*l{X*k1#62v(hZSc>r<4b&`lk-{Nx8@jOv~QbDm%9?Ev?Gvt4`MW
zD;&-Gq?x&8m7RDJ7zxkat{{q<w)5^4uqs1WmeEv3D80{f%C_-~AX_=zL6uxqSeLnn
zmmt%s2#*i(ir-GUyi+j*Y04P@_R9MjkZ?g>W#*aKE3*5`)ecYpe!%XNJW&K69W{d~
zBtC|K8{?2AEFC#zwjQd~O?1jShF?B^Mu&d{0QVX9#juq~5wQ;~Kl24^`DQrg^clOz
zK~b`tcDAKj3ub2m`T5q<zhTtapa|foVqLU%@X@6I1vyc&Ond#6D`#~q0Om1Y!$N2Y
z$+&t89x}Qu`*tmS0X*)n4cw1z*BXxJFCKjk%0<$id8e!Ey~r|rk$HE$bOUK|2P7C4
zYsC>bpfxPYRA`XP$lv0BD)?fhZOs4<Mctre?~sTN%Uf;#y}~Jz{X2jDr=?Ik-LsFn
zb)WcSryt%*5m<@LB!jHSsbq9Pb`>S{VsF+ViP{re=X0vuMDi*#>1*rqkZ|=#ejCB`
z!exEEs0h;S2&jHmws~7Y)*L-JE(L1-Lt0p%N+l@*#~cO<EKH9@@y`7x2h}oJZmKFH
z0MIYdC=M(x=q&@T?#C82nzCW4$o%%vnZfhSE-r3qt<5nv#qo%*{o=4<TF0a1lPP7a
zRaVoxp=ye<7P#!sDh4qC<B+6q#`NN(IF$NFs66f`D+yCFk)P&*Q8us3o(0#)?y$qJ
zav`KIy$ihj0{>Qm?I8*I+v~g(UXyR!=$$Q=zZgX%P=OunzJZ%Q$_IE~T-3yTpVY}E
z>?C&H-#h@~)E>#@bQu~jef{ZAPYL%EqmQ>>*u-b)11cP<`}P?^?wA?maT?WX6*D9K
zk^9=%U@^8zM3g75E1uN-jpak83wvsPxM1l`(z?X$YfW-#K$+k9pz6w8;@^G2Kx?tz
z+{t84V3UziDU|5Hj8My3O-w?cBoscNr1C+-bj#T1MueoB@BpN%0n?M5vj0vEd7X}0
zK)N3d;}`KMBG;pxmF&rB6r2rykR(<D)op*=6g_Vg8TfJ0lRZ)z1fPh%eGYe3mM=W_
zOk&j3jkO*%wFrGTFD?m+s6E|sAyrHJdo2H!$z2AQZ@YY|#gkNl*Z``o%KES=smo~w
z@MWFu7byvAj(=zoKQRb5{h$jwyQncLr)u4h{a*}zvr_5M+M-Ck%X@(-~gDO3SS
z_l?3N3okz#5)hU%P^jneJgNvLqbIdDuBR~$^w&Au6_gx*cg(3ER2r&U1)Iw?4e;Q{
zqwszi0_DP#Nd_o$mn3EgOM7VWpq{z(*49atQec!MA&dqMLG`8_m$aUCkZ`b7?#hZ6
z=2vdA9g9U#h6uvS%PLZN7R^FFwv1eVNkUvhS;eA~GR|}X+Dd_J9JjY0y06C=2YjH<
zV;1PtkCQVqL;CxRUh43O+K$d_oq4J07)8TAQIZ>p$QM%Mgm%GO7Re%GRAymP+*BHZ
zp*F+HJNSR4(tlslVeh-+j}cKJS4@8IG84HiJt#m#gGORhLiIiiUx@uY0m3J}vL>{j
zlxgcz;`v)IHGRB+?rQu}HW=ElRT&3bv<vgx@{F4SVrGzIIDfJh%Xm;${Jb_mhz?ge
zX{y@@FwlCi{scxn6t-j{%{Am42GAt4Hy6RP5R75yKqPufdG@EvRL>+EnEyM~q=;dZ
z!~ycW1zJO+hi@En=UkTh$xZ@U6HVF5Jx2)``rqgM`Y9Iux-nhcuB?V;B-TjFs1>L)
zWLfZS0}~5RiBeZo2q@TBb&e?ye0i09l)78UWQ#(ja<;W`!++NDy%FOG)uoO@7q`aw
zEFXSeRa}O%Q|hQq7bgC!oxdy~bh{vk=HD3l`^@@~qN@A#MNeIg0`eLm*u%y4{fD>5
z)6vg~S2X}_e=j&)YA8ILxqT59Ue`v3E<^u5M_I;NE!<y<vfpxC&nN=9>kE@bT1tUo
z9+6^ctN>(bS*-KxJ0*Q242N%z=j5|Lp~O2grRolBQ)K8@vyv3W2lwLg|HOB=U_E(-
zg2ES$a0hNLk261XGy5gY*AwHg0I$?M&}XTF<9H}F=ywloZm@?F=CMdYF?|@8u6L~T
zKwi&!$qK<|ORNZy)Jj6?b0Un(ikD-ohD*`mW@D1?NedGOENqBt_z=LW1kRxWcI+hv
zq)xkiT#_{7QyXJk7*r961zN+d+tvl0eS5dIII<O;%^BN0b}pAUSxwdRWC`T|p~293
zVo1$j7-?1zfLj2^Rr2VXVGfx^6`;28zaWfAAqbd`iW^9+N`&dOk~$)HQEBNJ#HI^q
z8E_LZv(M}d&L%IirGo1Twl}k7J-Fy|ife;|4Qv+6cZrQLu;wKRk1EH({RKD2>Sm}-
zU!TZ$jY2A!1dYwOFV$7Gi}WovYN3Y0zZ<_9nvI(-*H(g`I;vAoqN%d(6Urgw#cVQ_
zc@fE-1kUdK@$5vZyDw=j;E05;#oS#R&RdvvVqV^AHIU{QS;h%VnH^`k-<$kM!_dVk
z)f_TqZ5xA?@5tb6?B|~m_79kqOl8nSL}PkkCXSp8WxEn96{EscgZD$VCGr*Ni!!Xl
z|MvAyurm@IlC1Y~)z=G!itz)E)PX0;_Y2Kr^){P+2N(SB2b_ko!3(ma<NJg+Q%nGl
zDEHkAr^bvJM#@!SR>pw@C>HXJ)P=NbxxD;-<1$`FJC%ATndw8b$&zU4cOmixZ@9U6
ziUO(q2WUE5b&d2s9GDY)szm@_{iQATxY9u%+gkRZ`;DB7{6fKA<N9hjJ&gUsHRL(a
z!kJ}vi#K_v@%xQ)t1gmxc<$<Dsf_5`snlO4^>XjnH<k762nS{){1qQp8DE9Ouh^G#
zT2m21DAjfaAb4s5AlK4B7=K&uhPVOrsO46TGF_&>IY#5MN-zYL{qRuSj<hzXDy>8C
zzmiK{?m&$y@+x_cf~lcjNF;qzT=_jv*B^8NKS28S5-~OyxoJ^Fp9e5_{&+ful*kdf
z<pzKU#pAhdd1t=XU)lw)ewVF>)yyLlQTdLz`3K22&g&l8WF=Mz>N~+rvQ~<=`;$Zo
znFywl*S7}Xjq8AZ^k}S|kLt65L?s>@!j60b<Z4BT1uphKcgC1GalR=hB!qSh=b_c*
zUt9@fuQ-MX#h828k$Gl(@^a^D-zi-m2Cw4=rKe#?jV<wHb<wamHppy|OkCtRUkb|3
zj;~#m!V|SFt5D?uH#M?g$*(zmy?_zb6aRAN|Ktr2nrENxIX%J-f=#wH?<DM1a}oBl
z%Fkr74U68BVvCRqrNo_GO9T#J6BWW;ULp-EJ|IGnHU28)W(jK0ap1Ln3Iy_;-uF4t
z!_k(miQ5zM1l*YN0t3UdGbF*$!2Kx!e;M;&&$|R0p!F)8Xlf@&k0})5%Nank(_y&4
zdR9EN*_d4kl`7x-3bZQ0U1=(!zCwJ$t_jQ$&<kb#lx&!cRz|K+=sqp#c%$Zob1e80
zQmz-FtIunoxSXqDY4Kh?$em`5Hf_bCW(~Q`V>NAd{+3@hcaB3$?87qO0hjC$71b-K
z%Io2r1CJHq>`?GjUELJRT1d}iM~tzIj=RaZZdjm$VacvJL1JMpnOsA?52l!Xt5y<w
zr&e8YS=_&p*WVJv85sniyB4av`|!l`90cD{?1*K|)dYk#%vH7KT>g%srakrL>TiSC
zDG%7+SDybO_(Vk|X^5;gu#*b3Zo!n#tgEGwoykAhNBu!HU+as9b^QR0y}uz&wi~y|
zF1|!msqfx3j;X6>PODvGxuj-qcv*9gx@L2_G(C{Lzf3YUv5-x&^mIEbe8WH_zSff)
zSrk9ddH&F{k!+xPh{;@l6_kX>{`vJjhqQz;7eL*hHS`9Y82cc#fq3#d8Z4Wm#3;T>
zNz%tMEEeaw*&0tcOoOJZ(bR6D_KG}B4+&okr-H|n7^r{!?)Li(xY>{-L>OH>wdB?i
zY!?9%$os>FHW}H`_kB2Fi|{?S@2oSTeYnBdkGC_0G~*S8nH1idSJ_AWrA_o3zm>n}
z3P(TMJkmvDUaf|&D_}=-<0!~_ETzT}#xMXx`WTOJ<j5{IJ)Y@MS7-jWzJbk6&7LaR
zO<TJ{FfXYL(w1PndKiD8LgSP(JZ{r3WRn)~i-N_)%W=J5h4hY#2|Tap4Om>XFs4H$
zf{$Aq(-cpBWK*ikqI!wAc)U4FyBzneus$D`r^XZ!dVktzDBDHH{G6^>F$0Y#B;&vQ
z{y_Vff-YO+NPG3VVr4G_j4Ur{Xo@U%fMJc(%BS%z4WA}BFR|0RoFrJ!PM6h8Id~_^
z;$v4-$40;LTiG8!)@Zh+i%lkKJ?BJWJst9hfeY9Xi;S1n4S|L?ix_?tEEj&#o8g{g
zPK0W5RFkEcnPe4M?VgWfxkIB8_(&H{Mn7Kh(7vt<d+vwOc@~RnT&PS&e%R@^(k?@O
zaNyJ4@h}0Mt_t@@tN38%M#{roP6k*OV(sj!0DW8@l+Z^aW$8=^jsvD)gUWLZrD)N|
z1#QXzDk$C4zTj~NuDbJ2CWK=AYjccPFy7UP=0SFyulI_`_O(s~f3dgX9)#R7!fnkZ
z>S}$1{9|I4tHa>5u9jBXxQIE%g7gRl<nDfDy1`UXFio8iGNEb`h1VJo+H1Bg<e2=r
z%H#PY4;7CgjD17)4Mhlve0fnesfdrnE+w%#X7)PVdmkM+tcpuwHW6h?QwRQrBZYf3
z-uMNIx&8dUntEb|G3oplb}_vFBwtRsHoyU~4)`3v`IXM<10)-wuUoG>`s5I^Fu>
zp4pV)j2t6p8V8$xd4Va>YY~+Qc_)8M+S=!lZG>!%%W+NCqf5Pq4nObeQq`_0;B=v5
zCl6+?ae}8y;ghl_)gVV#+7&H-JE>l%yy>VYR=?S!Aon*wjs&hfipDqHEmT@ZL17o{
zdX4@p#87m^b3&9Bf6u23+6dqtJ4{V>ZHlp`5SJ7X@ojJ)O5RMBDlus1D=Nm?pqFtN
zV44chW&Ofky40W5S)qj(?(ABL5w%QRtMzjPPmA#}=88WP(3b!Y12GCy+)nyO;=V1n
z1*@qE%@DvjUTU&{+P-3q7v3VuA+La{wcFU{`>~vjg%%&^YPYzngQI9d5~RPy*<gr%
zIKNYfep9fTGlLz~t+^C7;@1WK*VphjGv!(Xu7q=obF_ZiCG~+!S*DXj^@nGl(9oal
z{qY_EA{Y~hU~Tjgg|EX!x&ZBQ`?f@}Dr8{Zzq&Mf;w^WUB&iu~A}Ej;y&$q~=oPN8
z=-noH<Gs6ydLH3BSytO?!i7L`0NzK=0iae3SNTj5O#mNrggIZ}rY@X+1jKr+qg=cw
z!RyFJbqKq_{c<G4_d(-I&^;O=5{I0^G^55BYcIWp5|3+QaOWuuxm}qb3FCRh3bLhH
z>)Vk#mRZM00){LKD@uYFw$#lq!^BXkpX+6*x_dT(?7S};gBkInB)NEnzqnwk{Q^m3
z6_ppKG<<auK(APE5b~uZ)w-JyV2i37XxhBF8~5GRop?qNJ~a3vcjw*Op3CunQIdbF
zUbO`BSqt{MT@|8=fjY4elVwce`7U)c&~{p2L_9%c!Gf;85V<QChW8Oytxsx@Qf@u0
zDeX_#OA2hNAO3-nz)VWg94c`UA`!6}3djZ4F#^8ZcQ6{S@`TFmoI6KPf@aeu>NZ@s
z@*slg=zH5z3pR@zXL?%k$c^~D0aR19eA56GNCU8_*#i74*KO0{l9n>o4Y<+KE>~H3
zTuqcoQ^%sqC2i(gmfiBT&XCe?L`Xu>bT12>G>>GIX>9wIB-djIBm2tmht5(@Y<F9W
zQuUzAn!kNSH3SAwekDSg@KIAQe!}ClI4A5sI237IX=47bl5aXe6UfNZ(N`u$;tm85
z&>B4Hk7JaiZ~vmY7&6}xzWvYH<)zWw#3cL&8ey_MZPQYDPU8P?;;7Np)lem@f-%Vp
zUi9tL7IUztHB?Z<bQH5`D^`ts2DkrNgEkDKU08`+Gf=jZlmBV8`zY8jp}y)u?kSu>
zCt{rMWIykd14Xix{S?*}f9nmn36nT;MpT5$C39S`@Zv%x{qe&10i;t>7mF9xQ&GyO
zwmv#vFWo3YL9kpdapl};9+)tRjvmeYLr%y)Sf(ebQSv!H`2y$z8!4!BSk0tj;q9n+
zcl!sxUGXp5Pi&=TD3`p4VBBFyL7mjwXY&CUd|!iY*4uE>=I$wR2;g8={|1;YZD6A`
zAPj;-?{vPqq_94rTp@rT6c;va=gNo;D~(oP<KR2EW%ui8<G3P|?r>sPlRJ9fhfgsQ
zHu+4lDH3+Ke^%-lK-eLOg!;#gl~0soDD7$%OZl)f-@5**XE^s3t)LdHJSPb9#umn2
z_<o58aE*m=krrz(3gBu-Qyio=P)3>P7$kA6V1<d(k`|fs&g5jhQ3(@1l8IS3CyD!x
zA29+-lKn;HEDuPD2=EpOUgNIWQNa&v(&I2DZ)0W{X0awH^50*ba=3b+4V1SUe=OB^
zuMz4iodH0Xplsny{z(kcsT;6&Ya&KubpfMHR;Su7gboO&avR++aox@)HvIV`O`PaM
z@wo>DLjz66pHQFFZUEVu_FS9kmLkLt1SQ(7yuWz$gB!^!j-`JOQcdw%nj%|c_;szJ
z0J-j&<a+KR&%9@Ca~`Gx8wN9F9IjV%hH6K6T_!=y+HAi-J=KWI4SLo_pR=Oka4x#3
zb@x65?qdn2zrcKXze_^N9mFMa@-ttb4j=ech7_$Q_(P@2w9KIYcALh3pD6<}-UlDP
z7b%EY9Xq+mQtsJm>Na0~@IFat%CEg=V)VEq*&H)%%siLSl02_fojltXORpcd;#N|x
zuKRyP`jmC37he}~a*G7PVTNfX!DR1*g71{FZHmU5;!At$DDLXBjSqcqDUvexG%sGI
zGt>II<OvWdBmFeFx_Kk~y1L0)R#7elYKUtRsKCe-((~plsH4+%CT2jyV)G&U2KxNL
z;(;onQ39+X;Q%Z(?T8vsqYWfQ>&QDNwS+epKX3JUVTFbJ`MEEL{#U!RIR<y;1|lB{
zZFl`HC-Oa4N<oNX*=I>A2xXBgi88Gq11<iUgz~@m^XXZ`_!Gfn06VgDH|<$kWPC?_
z<QWs-uo?~qBwLnX2@R-cKq(+VG54|)Zdo%CCT$Ki1wcen)1_1>e7MYCGlAi3ul7Fz
z{e4=pkWCeP?)$2d@l=+6ufw0=q{2>H$CiT{wYsolssj*2uu`AJe!IzfA}CLRzkKqu
zhK>W#%PWD)hxd#y`~bs#u^&npZ*?c<Km8L;$=+blDswD2utb+2#MWrnu#gvxCc&sb
zZCDVPp1NC#8(z!#bJ?#07U4QFr6ZIKTifQ|`qzp2@2lDqjSUZCpJuX-OT3QovO}ZV
zMVYp`>SA;P%Y(WLVu3dOHPEJ);(#@gseVpIEZUxVxZV0roc-t-9CCm9!A-Tebmh6w
zz=3txA-LK|_1n+V8ALSh>sI)dLw?Yxd2QKPg})12n!aytf`D9zEqmEU`N@;>R)(4A
z>ly?_doF89LhJZ@3v_en#g<#skg%vG?q0_P$Vkr1(4PGbGc*8{Pn=6Gh`5D<x{a8a
zl@pW)5LOpH1LTr23v;d<<tL4JLea-R%oKNhUUb?x7HanvRC@egBzkqBXnN?=RKg|1
zZC@@on%c#DKx*qq<{16oISPo|uTeeG`}v<oq%8X7YUde4<XG1yhd77iKXp@YTg<-t
zr~RxrQfZX6U+z8SfimD<0_bwi_)xg2hb9Q;>&QN1t9fTJE5wtFJj&wm^{LmmUuVo|
zXOxL1U0qq^sA3*IC87T!R}ln>4Jlosbs!zA3gSC*QEk5!SkeQ89(m-$h!9OVK=lwG
z+|QuW!DzBCHQnRI$RqgW^zh2Sm@&!~o%9AaJuDw;>rfuT%WbZeNR)G-TY>s04Odh!
zj~cF$$X<KZY3}hppFD}Zo!J^U8?@G+^m~*qs3pfxPEZr(COzJFV^Z~^UH2Q>lQ?b&
zp8YB5BUdWFCI25|2l009m<lQjRbPBWpF?6=RP{=qp5)A$cjg!1|4-oS0t1m8zdEHc
zZ~x1Qsxb$5nlk}O+EP4>)YXvwL(0x=fF8LN8V)UQ(k~)R(;etXCuO-B=JeQ$jUN9f
z`M_*QRfkIRqWGYHZA?W_Op-q)hKyO-$Y9RPAAhZFv38S9%C2p@{Aq6)VlYRg;o+DH
z1$VA)SL+=6qMzB!2#G3|V$T;3&Md&pRKcfS!lB)8t7EG4h0rc4dtX`z;q+z_;2uiD
z4KGA*es1MUL(a{vHKIVOQV{^4qsRHL@`wS?@SC>V3g$r;!>zd&xZ`KG%COAP{}l<O
zY=l0t$l@xC^{*?`5bzL2-YyrUA-?eb3ZOWYO5Lu%@b^)IJ}^|RJTdQ`en&+=E93@-
z9^)?hVd3EPl&iGFM*mXQ2%u*|-$;W~I|uS{xna)(oDz$A#`v$w?$nR7y01kD`aWL(
zqX56ox<s`I3j~gfoQk1t!NCirky1MbneMp1mj7Io>Eabu`MjcSwlNI?jN(cz9(-&q
z*l{*w2^y2dnF~kKrqrP}te6d>4}Th>k!TJL?KRj?(at$}g}G-q;#A=nVLU&5VqC5m
zlgCDS*k&n6WKt`zm51q*R1qaU!Ou!Z3RCAe!_h7j&lH40Y>@x7(#AnXKRQZ%-_huu
z*@2fexygxH|4$_k1}Th<WCO!5vxSoyS-pYS94`b`W(mjRGGR?oPr0Ap`l*iF=w>NY
z<$VyDSn(?m>*)RB0vTfBqDaRx0+Pf(4sn0%HRBB8r`62Dnkp@-Ug8eq;&RJ+bIl0y
z<Vg~HWe%iA-^j_{PHi-#frnq{94oUFvP!26yebKm=v}g%vh4!gOMtoj$%}qU?BClT
zDyqO`UBH<eqhfBmCcER|^;X#C4qZJ6HkrLy=n^i*8AwsG(HBmkf`=O9)OgVd`D4=(
zoDf<cCy?`!`JfG-y}ZMub{CqpC><8KhdQyY%GoQXau4SVI-FclfYkY{NBI=86ly{z
zcKpxdEO7(UqgnqeUtZs_PSzd3D-3@871^!v!N_6=KIXl@Z_?Q-(;{_-JdDs*)u|Ah
zqA?LX2)^4RU_R+qGQgA(?zT#%Y06}G>MrZPp@rCQb8p_xyLjNTtB7rX<}3yfotBBT
zw$}8<W7#8_F@<O#R6!?)x8Y4m07eMI(C_!c&=EL52i#u-r=3oWJ%j1gYuA<Gzq~kY
zjWTd{s(k^F0z0iMps9i<?r$OF_n5v~`nfLPgc2a$7tbCn0VG_sP(&uwANhXSW65Ip
zQEz$=6ogLV_LvMCZblQogPY<QH`57qdn6`pKK_c~q}f15<u`6TWZ%5I*uGNP@}7{D
zV2ohcYynvVcf9<N_;nNAJKfX_3>eud(V~kuO|yG<Rw<_^BL`=>*mokSCYab}Y{XT`
zZM9ID(qW%crQlTt1qR^0qBE-!m~)e*Q>R1Wn9|7{1PF{NVBwuZ6ZM=<Yw9~aTa+G#
z{8hJ)s;aRepH?FL?}ckA_q6}mycZ0fNyjf>e7|_9h#W)a^E{oWcRW!e!ky40_S!pT
zXA+xDDCRn=tj@T}ztk*-u$|WhbN8eBKe;a~%hff1MI_}^6jzh2$s%T}B6jBB6*97L
zi?JLu<#{%^XT0Mj)S$Z7H`re_qE4pz+aeDc1a<?rGr26G{<Q~yYGT!c6GlIUTOpjh
zIQ2z%{Y<yhUH7qE!!b~)=b(Kr>WRAhVc^)dXDjheQ1LTl)~Abe2m+yhSiQe|e$r0#
zT_%ch3hl~L8~f@d^wCQHVDQ-z{^Q))3Vk|bP0p{eX8W3f4yfsoK7Q+JcS<Cji=&1m
zEcz<Je3U+2oTszV^y)1(=FZSnNjV=mXvC8@S!)hLswPBETP^%?$$h9QROJsOWVlWu
zHX$UCWIt`BTcXK_(c(488KC#*SBNCvH74pxS~`f6#J7I`A1{C*yZijZ#YmXB@o38l
zMUsXMS2QhdIK}SW+b06!4_%rPmn?CI0dwugz3+!8-X}-jrmb<KyPe<#CQUYdaT+=X
zTq=yz!tB|I$&wv;fw23<EQe?rq`T8IpBu&y%wb!Te)>(s`+-yms@?F23pG9CNMM#y
zaa=jiL)a=fuqKF*b)eV7h16vC_Sya-NI>-3URbRXrwMTJU%<0lWXaQe*H|v$MZAC9
zgQ9-{lozcSR)j=o%GZqU-CY4d1>M}*`e%XXQNc$Wn6&?NVq9m?lo9QALq)2Dn&;vx
zJebZO@i)hB57A6sF`6_yX`$V5>TaZ-?R!#K&J#g@^bvLdr)Z&}%wFjvEbMKM0#)87
z122&DwMPW1WaV#FQkRLqi6mmpM2CZ=tRp4H7|B#gsy?9c?6u;>XIk+zbm?QUsJMKx
zjLh!0fRnv{Ho3!R4kjMpV^ao$zG7Ol_{<4K+v2S@bDh+ch}y&UK%DOHpR`07-^RCt
znJ&UkT#X-nc+kmFq<DLLfOR$Gwvv?gq0M1qwGPG&7+CCrmt;-uvWeqs+k^oREw;me
zD>?tV^)Z~>s*ulG7GxhP0{NrL2<kYwncJ4LXTIJ^HKT}9c#iRh05vE_U|cA)^T^1G
zJy(^sIFOmZP8pdjb63Ww(QxZsbuh%9vFy8Rll|%GrFaX6`&(vIc`gx=*1|ga+^%HI
zKOz)qwmKJ7y-62Cf}Op2s7*2pj4l7HY|5Uu6X+mdDINEId2x4EUk}W`isH@#A_xeb
z{O`LH8K5|m<Gln}y)J&r&~hXi^&G((6XPA(Ti&-&j4^!w#c$dg*BdIC3w>4#I$mfS
zFaa9JFR$yXd02vv<DSvRqUT9nX(bKrumioWMA%?N8(@ri`S}vKB=3%>sfAckQgt26
z>X_A{=X&>2Q<^I*Fxmwa)hzuxu|#GE9TLx+j3w?0AsiRv-a@NJRZ*6GQ`{#PtZ%}&
zqgvb|>1DjBeVbX~{Q{V~*Mtwc^K4pBGhUK`!mIluEJU!N9dqAXghe?6;oF%$=rm!4
zza3kAhZpOFP(*&dZS^hG2&3~;dz3q5L@9>m?MNPT*|X<n$2nhKVKJ@s9bBaPkGT5S
zpKb+@Qu7n@LW!izo$DMFEXXX|z#byUQe_Fm;gKcISAu0=Mxfy|BNrO4fZ*BUC{!$c
z<q7-!1)nIZG*{I}nijv6zY+D8TP8ik1^oJdk#_sdUmdBt$_hHD-l!4b#swn?Teg(?
zaPW1^8lkN6d&>in>p0uiS4#Y<BWGI^D88Ry;<T^uQDd0<_u?^(Is=FmB&R88vWG_C
z!1e_?L2OO+h46P)TykjSOI0Lks}k^@g^U(Gl-HUnL2gJ#?G`ai>thjQ@5DTI$oGYH
zUna75de&mkW7Mq`rJYu_nixBwA03YS?0QdnKWX^LVWGI;$dDZu%&UA6K@?uvay_Zk
z?yS*H5r<6Gow%-;-|^?3fYmM<j`L@Z9`50=TZqcAbLOGdMl}o=k8s1|DN`lmIEb{N
z7*Psf)T?Z6vaqimaV#^IsDhgmq|wTSN4%8!8&gxIs8`x+9O_%YHmBe-`XsE~A*eq~
zT#d=!OStgee$)IG{N&3A<7CI!rt<{#OiIGJT#p$q7jt$Fdyb+qHQ<ElNVuumb#Ldp
z=>G$C0G1{p_8oqXx;s8tHJYXYyF<ctL+}B_v~{FD;4aJb=&u4;RR%G{Fh@#><y)Yp
zt4}RZ^_x^{8vq9<?yzlQLd@`7EPBfIQ;MCD#_+I;>Zi*zViVB#5!YjjOWD{PfU7an
z%dGkuHlsXo-!fWbw-)-Ubv8>WrFVsnih71X2-V87G)Hn=?)Eq#K=ICh9UhorkEXp5
z7baob>m&$64e;E@A=O7avX;mvq^K2Q(gOTFixXk8wSWBHKfKZRE*l_+t~w<%ZR)<n
z66~ejx3wM*J|{$c;hO~QV-8o{6<`r*eL+l~RvV>dU{ns6h&v*qRf>e)xwLE#*j;$5
zHa4k$5T9_~bla$VL4L7aE>#qih3z82)A{S-t5`?D07DWB6PesLL3ML5FEE1B##_$o
z58)`PYc}-QX^y(+s?!8z{1*xIbg=!HOj|8zew{(N?B6w7M}bCq3P97P#MfBlTm|4O
z*J-6ezB-dyF>8BDGQsexB({iUlW#$ZS=fE7@AJ@J7J#_83InMbKU8-qw1zl2cY0m;
z(8V7v<r(t-<aD65LkxQTY)$=2GZ*#IYIsGWL#w$(QX95$DBxFe^yjhe_Tra)vd|In
z$h}N6(zI(Er9YY2UdOWD&XOfccd}hn>QA-MH-SQ__A){~vWVjyNI{o5-ijPBfxr$M
zcEK159k8-NjR>GZd@nE*n9nmOk3b|yzmthjqm#Nte)S@M7CtKDphiMpV5O=V8`8fw
zmv~`}PWxnx(gA&8G$d>KWEfnS;L;`m%5{05qb3bom(_`3avEQ}@`}G3{zEwJ`2?Z=
z`=jR@$?2AwsR|5t>4E<;$Gz6ChW;|gx*i>Tc8r8T%Q?@Mi<slIjqRqK-Nj_F>2*||
z&+Qjiw!}4S9?qn>%r~2J4vH&0D{=pWp|7w%vCRqz_M3!`fg&!wCs26jt0eMVVLE<)
z=rmYqstW2_tbSBF6-DQOVMa}_g%~ek_#~Z!wx)Zkg){VxdHUwfSHRi>o*B$Ap?N)k
z(vaAJYFM&ld^D$Ie%@W|C!C>rXJA{XG_oOilEvm1<8~&{%~&!EKFD?5ZINmHBT-l)
zcXFtYoL*8mK=7Tz3e()c!Z4^R4F(gxJ4B(wWg0KZdlqMd-`o6gXILd9d(s~V*thcE
z$MJ9FF2j0ah}UBH{jeL{9^Wa2><{(>J6E1TrRjO`e_FWK?!(myxTCw`g)97iJftF~
z#xHs&nQyPq8*2avQycJ{?$qgIYoiEc-X7^x(yK-xp*O@OXqMBn8LKfiDL#*xu9Bd}
zX!51_x)^h_UGW1tDZ`l&YIYK2(*4pE0A-XG<R&4}_GNwE4@DeX+!L52ENGwxSZqXa
z7Y=GWVc2kxR!$W#L4XOJoOEY)jy`|zb~xHeQq8>shK4UMF;aoKQ1d7U*7S-#<~MmA
zUvFeV>4=IrEO_{$(T3vMjGJV`Nr*W&Tjf1cB{MlS2Z(@W&aA4*TP|?rgvUsl7IRP4
zh4%*VTL>(A8lp)5fCaTL{D+o$stmB(7J_cLGeYm~!#QdSTEa`u&j`L=%a2__(`lMM
zAm?v~CO*kdmae?01O*bnEPgb{tY0=v(6RA-UQ#%0J@IJdhv;DhdeO|$6KE+__ACWX
zhgku_6EZp(o&K~8ias$pw2wfP*8{bRuYivwyHAi~U?V|kw-Hc{zwMPnfZ4IWrmZeg
zWu;^nbhg;o#*G1J(MFs>SXXkz&PPCn9QQ1pA?#5?<D|vcGm-%QM&47zk&Jv_M#GZ;
z#xsbqowAOdblW5jt7#?)=rkuLCR6?vJaDdlse(Zxn*NUZ%TS0M)y~uQ0k(0Jq!&z%
zjrs6#;=sYOE;?YW5G{OvUdjSPO2%Ky+=svdT-*dG3r|*y!RLP=-_uu?@WbQS%x8c^
zlc3Hg$3&F*DOe9@4CBtg<Qvw4-ybb4lYU)^?w3ocOGNY8H9&FG+%sJGHH7c`9vM$F
z9@`Aq;zh|mTUhTsi+r1p{lz)*H@J7;6+|U2moFvs`^(QV^q57U_|S6>qLnzcRlQxZ
zNM)5daPQt*7NqfiD9x%;nCo(6(WE4E8f-E?$_h19r^ZlZE`s$Lsbyu01~}BHwcX$M
zxUkf@XH)_+d5iC8`(k+*@WbX|Dzq(Q*MONsq3o^zi87rZ>nleXu<H1&ZbPwL7;p_&
zm*(##g)kMdZ+JUmH=c|0zR7m>{vZ(J4ME3|TL~wWH*+E$49v)JYU*g12KVOFjeRZ4
z-Ypc{(idD@4vNm{o&KuP)b#H=m!c6+Vel`BpOVyLpv-Fag_24#2-{4gwJ%+|$?(hb
zfnf=D65LmXZShs80`?I+CMzK5zCd?Y$3GOJ>Xm9Jc$)dBIops#NprcE{vBlyBsG!g
zDp?yjItn!qce-)x8{<6(nV~|4H_o=cHB1@2$vV~qULzom4yGR6mw5o<dha-bS+$OI
z(#%t0o-k`l0I>p_ZfQ<zj0V#HGY7GS@f$ZxT<;(0K5CLgb>l>Jzu!=A*K&wrMkEJ8
z<DV<UC>d6j->urn4BU?EpF^OD_F+A5*1`i}pHF97>N?}o3MaO7=<43;56K7D;uVX$
zmPZB4naVOcM=Tepx^1#oZ#z_P>%)&~t>$~4!sdTILk3SoXu|)pM>14>aH9q6k+OY_
z4w_6gcrClH9$zG%sDeH^>i$`d^l?=tFX3F5q!gKEm35Nc65fB3KWQ)M)r=?)q=A(d
zxavcA#kGu3fD<0v6&7r_O2UOBGyA#h%}jZn0F!hC%2A7A?IPE~%w@{y-LQc5Md37J
zi5(3z7Omq?`WjI}S@ZbDM>N5lGL{m~+b~J(L*-Nobj+BICjRSGs%@M5wtgJCFYnOa
z4ZF(QR5d5b%AR=%%NyF-+p1Pfv$-+@eZ}Y6?{IR#qMxn#jc<>H%cwC>gx$&RkPKIx
z8Qk2Kgj6^un!<kLnQEyopbnGw%!eVPSQ3hdcCqO*G({D1lu3nTO*VaT(G}K*HhU}l
z&r$vT(&D=M?bA@d6aOblk~@*Lehl`9W1m;JghjMfL)qW3A}MO>kgTQEzUZXXx8vbT
ziPpF$<9=x4|A;5rNZ<NTt*SCPYX5aCj#bn^`j>l&N@PsI_87vn!uQvN=1N&T^zTp<
z=5+N%KbWu$2tXeN&OC&7TORH_s&+W}bQ;$3O_pq<p^pSYiFYBA>A+p?1h@7{2oley
zqlXu^g94!$FS5`7ATLyv9nd-#Gf>DkrwkOIeNXyUHCcUPRu#@E--jZjSQTUl;`P<A
zC6Lss=0Ks*e_sWMVAoACIY$Gmqge6X3tyL*o1C`F3Hox~nWe;SVLbeD<!_#SOqUmX
z68TV@r@6W>l_9hlOXJlTnK6+$v*`u%z>Yj%>0cEi1H-4^&GOKZHJd3y-)K>!Rzmx7
zy7U^x>Dzz)zw-Ao;_omh62e%7BW>^EaUNG;lyXc#DF@O)_-lt8?aeP1uK1Ag`&n!c
z%xE%D9to<zqCm&8G6SF5?dg_Z2q3mV(<Xyzp)NMdZTg8I;(a+lJ_7rXOJN*Ld6i-2
ziJgcQ2U43UN-Gs$CSi%zXhoe2*=|BYC6r4KN;%jy9GMQhc#hIAj}ocW0vCDLRCU9+
z+{@>qD~NzA>vMY<l2v=4Yi?`b*Vr!=F__h4g<@qBnri+C<?CKA<u6t4vjA6!ym=NL
zNUaN#pv4Ea1Lz~d+wwB*JvYWkOPJ*~+o*v>%vEK>Udnw;FC)I^f_%;1wVD`cgvG;e
zd-YhEDK%G6YPEm9d$UJT<C}l!OCR^Y7<}NcQYK%7WSeWFAIPWu#Y*A^Qskm3?Z#1x
zmU52b+A>w&vL-?x0c&9SBz{?qlmu(EoBRtby{cfz8uj=z>ji24_4L~q*qxnPa3FoL
zfka;(8p94cjZTt@-N(m$|LJ`Z_(*yJH-EyFckW<v?v6aQ7O*TZ!fim+kqXz{m;VZm
zdU9Pia_0aZHzOvc=rkdJ_%i<(fz(z6A^(BBTWY4)>}uU{kO4i7)v&SNAiOWV($>;X
zm*AGdq`^Ru0u=buJD+@H#J4BT3%<W$s=q+RLQponC0(AHHJ1<re5|%J&XJF9LZZKS
zAG;V>J>k1TtLQi>whsUIg=6J7eDWHAhisF7{)TW0^L&nWI?oqPs)^(D7kz}sNUl#v
zn~+>f1*_sPj?B8dSfOI69Llr}=E&P6Zx1)U?Koq!vUu*<Q8sWkR(j+^{CqGhjB1K7
zakz;a;eKOqMaEVc1sK4ILLKGf6N(DRHYPBOmbcUJ%{ohE!hdx!gzd=t3Pp)_UiBCb
z*fU9m<&XxeW(7!FrQ{XiqRo#lbdEasBb~wHp-sNbk8A4&Yi=mQuJ?LmWFqkMi(J%i
z`7!(yncWo?^XAch)m<AXZyPcxywH(kgr=N7==!{s?kz|h!x=hbu|@~kOZU)o5u|DF
zY&$h#ZPaMQs?GkkG~*)_w{}5p>Ihh#CGI#w_CGCg_5JIe+WT&w(E7{tE<l%O!pfH(
z4JV4hsjwAIqe>6a=vm_;JEE|$CbB6SLoy05F%S0_CVMXy>>>QBVkRW`H!0b{u{bNC
z*7I05qH)ixw|6ZRU1rYVNQDE$LFrJm=@>Gpse;;Hn?_Rmy4hT=ZGzMRr)Ujlt%qtW
zgZJ+^x6b|}i9g|d(4W7}&npS@ZOy_qQdgk&*=fFE$_~h?wl5_*U?x1VV#eIgW>YP?
zX6x@~C5#$YG?VVh#*m$=4;QbALk3^A9g8^-B>;FLl`<42lPD7k;|B;gr-itXC)DcK
zZ?)A03|W-Cn4ab5${he)PMIEu^-b8eI>+@*wm{Vv2B`ATx1a6aO3f$NkAKv*seu2Z
zzSw{E`}Dy6rqEmZ$!>f>1&D;z6fuY|L_|?cM@}u!vW=KFKaupJi$44o8JL^&x#w!p
ztfW{QkA?IA7yDqN(;M+Ic>O+mtZ`>Qb=I_ed`w?bgv8e>d3~6;^nH<wicB32(Q<?{
z?jKxfsoB~_RJ2qii&5u3v!9;P8K0Oqhp(1|R`|DlVthj+jze$?#yM=Yoxl7-zem?k
zTl4MWQ8)ow&yX;;G-ts-r<8HBQ<IrG2ZC<H=H@I$Z--Aj*@^uy*;6n0SW~ojf_xVy
zW>Z*YZasT6zucgNSOv?afm~4^oW@@r3eC^=CegO|nrwoq`#={vBMvt67ipW;@Pj=9
ziQf@fno6D5S00V#c<e=np_H(XJoP8D>B8*jblNDGWWtH8a-N|*+Fw4dZOev^dnO)I
z^nSw>6pP<s4-d5<7Fb$2(=vYr>>Rt;(-4b<+K(ua1Uru3laK}f--LyV03Zo}Q^?%e
zBI~x7sDS%V+wWr0I>PQ9m!%`sNB{8nI_Ea{M}i^@<J=amf~t5V7IA3%cJJTS<~j_}
zNG2R_B2b@o4~fu5ZebEoM`c<13>pumsi(0XjD|?^24XuFc4vC+6LKsXs`|n;cQ4{<
zBtyRER9ejaG!K@4{hFiqV^2n8CZ*3@NkCv2`L*!}u`oJ)#_X~8^$rrV-$x|wT1593
zMKni8jE0x_f@~%`!U%JB|J*|JbHHCr%S%z-Ctyw&+`%gf%0nby<I3E_`^qJd0E=f=
zxrzIG>5XikY;vhCVB>hDy!=>81zl=d82EVHHMcERHgD8+n+2Z7S`&YF|7Uu}lymG1
z%|1G!?O{==R!pOu@jtEbQ}1^Ubbg9IVx6Ld-#JoO40}O#(0u9}>W+KlG`0gp+wy?x
z(u-;0L@NeM!xGfUtHv^%TId|okWGMl;yIFtLGFEOx#?217<4VFg#wUWo6kY8*Ei&I
zp|)y!Y{~N>cYHk@h9sW;ybmaU-Y}Muhf&iyjm3B6bN2o~{#>vi6ziopo>L8f?%sZ6
zwdlS3PJQ#MPzfP<dhJQ0hK=-+=KHzk!9o5@-5qU}vD@XBSG$L`BH5vcG{q_*?|EJ_
z>8tmBv-!l+cbLr7vNtLTOd|Xoa&r2bidR6}UEZKG6at;yODK>{a5AdBB<snYDBkql
zUub(#w!3$H-Ngw(y=sfu^yR5d8_~&K5>|Dj(Z9LjFMRS~{I5@>;&~cyPSvY1!><#9
zhOb&y`5i1$4If=Kh-haD`kZZn^it;m#z7GPA_s4P>2ldQqSf7Ik4S;B>vd4Z$8(I5
z1gc$L+U68x7|8~Mr~gD-;}M%mnNe)LV<D##J~--&Y1p@N9TSbxNj#(7mbstLwtB{|
z3<VwIbX16IcdPz+VEsT6CE3z8UiT(s=(@*B1p%Jtwj29qg2qcYTQ)*tbDty!n1R*h
zn+(f+$@Iq;7)*3c07#SUApRxHc^&-;1FR-Lm!<8xmFiZM%#u4|3fPThDXflGl@Y$L
z!{_*2sKKY(fPU(Bmy8xa?z=>Jk<rI%^Mh-n3<<_0p0INAy3S)@<BmZ+pO?at@}ukj
z1y5F+oi(2oe{ec1I`>93Wni%zdk|7rpwg~VKmRP1(7D4x{m|?Oh4^Ip&^pX%uJ!ZM
zVtAk2d$DTV`uTTM#=Vow-r*Jp<AY-Da&U*8Uw>S{S}RBU&@+Iay;Y6!q5z5(mj#`}
zoACt28G8*LWoX#Fq1xlspZGN$r}S$x0mJxXK=jN`rikZP_hI!QPKB-5?DL2nj;<u7
zh}KwdTJ^YqC|-g|2O9L5=#`*^_dF=#OPJS(BE-b?!;}+IFlE}x+Y2g(<u){Bm5Kx#
zc|F$w`>E7QkvMz1+#!SVT@Z}*Lz2ERIo+2e{OAc#z)-#CAI&YM5REZc*}FLsP`bbl
z_WubN{3nEf8qDA(mi#^(I1c$7I**yggqE+iBw-deT3N9cUnO7Z#AB*lpMwrw1mUMY
z?%a!+oUjEff=rjhXpxOTj#Ui!;}aZ;saJJwWCy?`dc8&xgWBF^LGpTl%JQcu5w_N*
z*uVo6P1Aj3_;_EDr0Q-<*2eIKSmoY$t?O@c@%%9w>gVo8v#Vi-E_5b$L);a{zme;*
zEZ{Gh+WRET&_Evd@I-nNFHBift}mmhl9sZYC=ybVUR3+n49%@{LGS?FaKfR~wJoM_
zP^+t2yRAXPbPQAe-uZ9Amsk7*zXffq)o026zg7gFg2tE?7c3cweS!z}HbmxXNBGK1
zyIAkoX$ubh7=&GcOofR;MMfnVfxYfb)g>eOY#xEVP>97zg!0g4F2aQDHbjSqZ9yp7
z%JoYj6}uEyqr%?k;|7s}Vs_w{R74t6R=h9)(T873%T=|x=B_@;BA)3xxD%$U%W2Yv
z+){K91UhFa^(?sHa?YZlY3KN!`P-uTe=__u7V5t@yUg-~LkekRu)=6+NAdXUPkN$z
zS`$V4@QUtdgd-V~mEX(Nce}Uvy_TPX3rmp3Nmu56U<oSZkpS*J#|;<!_I^}tQzk;-
z!CrFAvuLSyD*Ikh=l@al)lpS$U$h57=}t-MZlpsF-7O8$U59SeLraHrcM1~HN|%Im
zHwcH277p<~?)}~O#v6k_7!1aNZ|(i<z2;nV&6WJpfOFLIPqfv)XGPx!oK3AzE@#cH
z+4E4Gxlc`XPyLp_mlg?u6NxZ25`EN7*5X|UtTpa=a(*ulHb~F1{TdOZWHg6)EgyV(
zCY#}p{3Z%z!6R0KBM`Kgz=62F0>GG>Or~T?udDg@u$%8EFNUvZ+B>}x81ecx3Q7uC
zob>T1#ZUjL<T!-&;smEPRzz{z^-SNaNL+?rA@5TDvf39-rix5e$`qi_7HOz*S%{8^
z@VV5MVdHC}ln;Ba#E6-ee<8oT;CRwUqo3p$e*DuV?taQ02E~+<gP(N%v92|y6WtGx
z*dOb%y=xDGZ$%>I!1qQ-V5z{N-Hk6H|A|}bhe~Za1|!kOt&4XiKR*S{i8{)Q9~l++
zDLAHU2sE0K@atRq!kOZ5*zH6jS{IpD+v@a2N9v@z(IW-l&kYk-aFE8B!^Xrq^7o+f
zQ6dkn9@`%Z1C$;ne1OE^Vv%q=Loqd-S8T_|#sqoH4UqNqa4F-S>%TS`UHUde(iTS9
zZ0-^B<T&*NQl90d*<5oV{jqvIaBuynN-fPC5)NX!IJ!+UFnP`Qlbjzg3X+=hgzH&D
zacc9KTtKUc>pj{f>ZC@-+#IeQICl18>!>ol2^if(LORD<A28Wnguh{*m(0>#Pic#x
z;$JeKp;-<4T7p&cQE}1Xn@xbaa<?sZ8!2PjaE+mDVTl1Z5<)&%PaTR=EVaMS-{kF?
z*V_Ke-}E{TiA9)mCPiWTY);HY(W#%u;{gDBQpT6bTPtqH9-i0lhNha=_vB3b`v}op
zr>ZKGIEbfiz|_C167iCgr8P|6E}mp)Npwd*9*krqu`_@9(=`;aY!O@8m#QeDahgq<
z?Rf-FCgRmn5iF`kt7_Y5hjtY<0j;l_@$N|Y%)y${bTSMps$GWi=qi%X7Oa0at<3jQ
zg!ZgGH~bi5mqEyKlm_3E57kpB(L&6zq`VUh_}gmW10nai<y6ajg^w5^O9<pT?&nna
z;4JapBMQy8Y=-qq66=yDMLc5M&2ODLtK%1!7bhvd<zVAEx2shzh%@<ocUG$L?!65M
z@hbcU7WPQFXvA0AH+(eSE?J`N^@a}+<>xeIu@|idwn7e@+m?aYxsOgRnbO79^c)HE
z1YxKt7Kuj_=BY$wWi`qL07PY{abtUbByvrT&D#9+<#<QyBz{zh3?1eE@`u8L5p=k<
zOrgRgG_{zXK24yR<0T|!v9|9OAX_>S&N68YP?Jv>Vanw3YMzoyUT7*Df{nKe@xEYd
z?X{>A2TvZRH=2-$%k9k7D;DTA!c>hqFNs_fmLA%rJbp5*C79Zhj{fidW41J!3v@>5
zSK;##BQ;<4laDc!o0$i^WS-*3j1je<?-6w?c^^(tHa?K|pDzIXQcKU+VYVjtI?76E
z5wP0e8=iU~4lAn;4O8VvT+?&WhYaGp8@&wpWA9**Yk=+1WcDlVVq;sGfh;zA*mgh9
zKwf}*2cO0UJCeW~-aTFXL7HY*q(Od`L52=bkZC#_%U7;}F>5=qdn_S3d^mAkiUu`O
z5ipc8*NV>!8?$OUU2?g7)~|bkCR+lFKq9W79i)D(*8mH}!%82^XNHKbns8Iac{=w!
zWv>dwGxM#1x|$agn+j@dyQpLjLHiT#)3va@AT%lwKOEkQ5*pK6WK=%?E6PJ-0q~f{
zV5^U44!5qlaM?J)hgfJ%6eN;e#LK^Xq2Px13v2>Gbh&QGU3D_C$)SSb3b)!rPoyFo
z`fq74O+&TjDV9!(b(})Z&s|GktxJ<|b*#Am-fimQv}_q+q8aN)EZ{J`5|sMc^Y{r$
zbD{d{SCUbemYQt`GgS;8SeY1XL=F8Lg=1cb)2&mRy)f&!)QZUq#qma?=0~R%hT_oq
z+#93RDWY)uha6=hzAIl0uzlsZjqV;5us*f3A!uj5HX=}L4t_g8bs(N<bfR2@O-*@P
zj`eVK6<Zj1hIW*9C-%aiOz396cvVv*l1W0NaOxP9qM=41^Wz{UmRmFDN3w`T%cj#K
zwV-}{eGkJ1rP87zI;++E9|i8)fQZ0gZpqa+!1G7r-;vDDXw1E3l+w}x;4HlWK6sHH
zKUr&F6j|opUfd5yYlJax=~xv_v!x&7a<|kd)fcRy$3`DyXA?){w>w&;fF_Y#7&%rh
z;Q&$HA2%?SYLzvx{WTZc`}*JGFn}5=C>4`VNeCt<9R)nlPD`Fq38-k)P(KRYell!S
zhf<S1HLgFMsh<?MMWP$@Jm}u7Wv1ZlRxFD#Tu9Fvf<}itKSFU#5&!g&&kz<PAc`4V
zw;M}N_JZ1pzS}@?SdHfHCByv03(PI2svy~<wiAsvQLxtM^h7ckM?8r(6z<2pq?~MI
zEl##V`R0yq2XeUssVhtm#e%GN{x(?cDOa2tu`Z@aEirHebVsQHNagh<yT6z>j>^?|
zeTA4!0+Fx$T^^6YT^Vsf+|=L+>DeBQBJ<hb#RdT{j7Pq++j!e)+v$z9vFE~as%#q!
zeYlXWc>l0o%mzdiW8}n^U?8gP;y_dhSbhRg&BYXol4nw+=h%71KA}#gFv{bO?YXd$
zCgCq^vd|!Qluk3r1jIFzjkWKpf7h~xmi@xRYoTKf=?$Mj;wUr0g+8oZil9VoE@k}T
z-|EM<{c=&ia`Q928fik6{FuxVB(d!@v*oCO*#GTYl72Pr@fQa_-uQkOKOd?r4TNJ=
z?4*e{DBL!+X*O_TG&AQ%yugY>_?uMmhzE++fu9sG?itN>aH<5l@PWzRglgt(SFBM6
zc)Kg<ES$#m8jT~zgkxQyj?b>nY=l~&qYK%RuK9ABIIe!+Jy4bRCr}yIvTX??*c|#h
z%x(8P8_LO+zFOkPbfWI*Yt?lpNf$Jc2rX~pXn)B%g(6^moTEJvv?7ak37CIUsWRmr
z&)9`Q3f)(IS7W{%424zo9E83{!e;er=YVwCY9I~k^9CO)_nLDbF3&6v&~IT&hJ)%p
zUC`mFHhfFOLL>fM?ig=dBEu=13PI{<wX)$cGO?2JWs*~)p}Yn!%*?2sLb=-7+5|o)
zRfK~l>?G~aicW{Fo(IkfpWt9@FeyL&dy&6*AG|!RXvdALzTy-HFO45iq3N3!S@}lj
zOZnf1Ri=v4XY+gc9fn}!2fW~>rge^y0u}+lw_4M20i<kq!=5W|*W4CoH-dvFs`T3)
zh!6H_t^MwgYJ=_ItJ&U%BTXjMy#mvDHn5CF7^iWa;}DU>HL;6f^G3VFXU+~pV4Sc7
z0r=uu4cP@9sfuka=*wo4Cy*dcd6pWr1?=0*;Cl(Q*1$WAOvB&vtl~uU@9DD%ACkkq
zIt{dbrTm!i&SQ5JpEvcK5`X!Ogul_Lr(o9CXRaXnR`g09=G}Eq6i|9jQo+*yH&8ZB
zT?i{$iP6kLExNhrXbXPZl-?WCwmKVP80cAy|FJAETq|p-YR<9n=;Dn*vB{0pK<wrQ
zEvJQJQLo{H#iMF2zCG2hO1XlDyLDM0wxeIe^=(XUI4hBHvS>+R;fVV}yboTaG`G`%
z^dJv9cO_EeYpn^Z>%4+?HgzVN%~Pa;N%>sZ7H8#na9JgxZuXgE?YOGw`|734Kcnv@
z^2%*EO-$WsY6LeD#ph}px6(dqMS-6uA|4J!&lN0d-c&=s(~9JLFKIf`Q-6*8d8`X$
zRopR`1kD&qV#q6Q>`+sst!3v36>HGIO)!b29sVhOPmD8UB)2MTG@8pRw$Sw@ts6Ia
zYF~Xx<AhYTY-}v>Uk|Y)wAaplXx|w~1&rGQm{Da!>RaCRr$zdB{9L@8Bl`0#vLC>S
zhRykG(Qd3~LjNT$z+h9Zq3LjZ>Ruhtv9?6b{(hYc6Y)4ZLAGK?a1E@ri#L-Q>ht_N
z4XOIK8h<<SHdkQF%*=6^E|&0<8PEMB;WR{%<WJF^*;SnLB-2dxuROh2pak(Q$5WNy
z6rh0`Syib?jpkHxJs1Ox+tfrallN=yc4`r;X@*X`I-${~--ea~wl6YIQx|F}JhGX`
zXTISQO9Wd$;8es_1Ba*kE3Xl)27UyZbW)>X%H5coh}#4PU#o|Gak@c)mB3ybOK#T>
zH3hA70SwBHxZi#?%=LWDSxvb3n6&8YVDSBw(JifZ(z}_hcO$d?;#M}}Red>u$Qu_I
zVY!Ba%^VZdQl=?RB5~nHv~n>Vd)^rP=eja%qVUw$t2T00If0(E9DX|t%F*(NO7r*P
zx{r<YxN6Oo@E;v|cNeeINSS*cqy;F2POY;7f0?`FIuq~E76j43+>NUomqP}Gn>UXL
z&+8zwi9t8DWgXM}!MQJnk2~0HB?5Z!tBFQ)Q#3s+NUZ;4c`TOuT-F9JZM8bD*tTVr
zW(W0ebhZDv`-UU)GVCS8*}Z@w;Jj$A(eF0p1mjn3^}5kKY867uoAq3=khHqrDsjAx
zuXIZyZF}(M-+GI2`d?aQ22e2!U3WO&B~rLQb~mnIugkII@ZPTnF}z_i0KFN&fLpx3
zf@S+z0b8WDh-JI=m=|O)z3Z`4j}{7R4QDq03`xbLTb44J55Djj({&f4Qf96^(1MXT
z<V`-~EtKj>n2=P;HkFqWV2wnhlb=YGejy<opvMtPQ${XFXC03trGzK@n{fICKIG?b
zp4MPGg7o|oc~J#=g7g8K?%?~u1ti+*xG*R#npo*`A~cD)Njv1L@ZMM1a}lWxm{QU-
zxKXJM)#GZ%ukIHnTgHSpEIH|@P-P@@E3w?<=H937>ZI{|ePH|xFgPP_OapbX9OJ0P
z=h*3B+{s_=^Y~Hn?+WOh9Ju8b_ff5;O@QblKHX=V<o{az5RgQF+7-=EWCDNxwhx?D
z_B~$o;ad!(ZIoyt&N8Q;{>!9rr6<Q8Cbf*g5z#s{>Rr!wPC()$uJbzmw=QSPz3(9v
zZbn*R!a0bpEd~-gq2EU$a6aIbr)soPTCDROCQBC%F{Ht6@*4MSa}7i+_^lL&50n>=
zc>T{Bk~fxc;)w@1bJS4;wZ7y-(jh3JcDf|QiRoJrMsF~KP4_}$mG*cFZkgK`Y4kJM
z@Ta@915t#<DelkE%ImF<bB~WH5A7E-J!ansxryJy4gai9>L%Xf;=a#DzBb$KWgf~B
zFTug1qe;jNi!E+{Yi(GVfD-hl$<Xy2bS9bdIb{>Zct`e@d^Jv?SjWVe&f_^QJF}!H
z(fM$~7l5U#&29wYR}T0J#d^R$-U8*69PV>|jo&R%XhZ*NE@S2!rP2#VndEn0=2Hp9
z9G3@DQv1VOQ_x$uS21aGfqegy=Ns@fic(NUMd)qp{TXMI545MDY`2r_7a647R+EV$
z2*t~O2%%i`Hok5#vxSh1(g~)UYq`9hh0~4Wg4NSLlp4()opq?GeO)Q$?>cF97G0x2
z8wqt^12J)?H0Q{vR_mh-U*Ltq*h}Lwvu1`yaV)OcI!n|jvtMaOzCaSPYIujgtu9IQ
zOufl$N12*ogW=DjO=>YgwFg;y@l@>RR&0imB3Ld}MNe=5gXT=F`=RaU1g{++q=2U}
zms_huu*<!Oc>Af>$zs6C2S33&A1dnJssRF{-**pp>Vtn;?18+T7fP1x@^C2J+f0w+
zPF@uKN&$?UV#-yDg=x<G)}z^?o;(g!$07-w;M_$tUXSZN>c2be7Ja&Nqg%Lie08hl
z8z*={{!gf$N8D@$Iz;TITO|&lQL7@qsS62CJaL->Zrv`oqqAa8DEa>%*xy@d<y3QO
zWx6HxbC@cXFG7XR1Nfw5{6Bd<e!7&io1lgI;C?p&rLZtu4cR}|A_VxR!5710kelcQ
z?Cz`5YloIvt{NyVpfo-RgU00h2jBk=|3SU;{M$5?py_D8M6?+(J&_zFyr;qcEMB6@
zh=f}fozf#myCB6>JxyOT6o0rB8O<;~&w{ClOep#VBmk#mWH3q==)EPu`^mvkqZcA4
z-0ZWl`P)chrRnE4e&kzYvadNt-|R-0<)<7<{ZuJ1QN`#8LW`>!2+sFLzi0WdvkcRF
z9WG`PH13X7sG!L$M2YN*N!8THQLM(5@8Gw$ZarPW=CN;|Y#V{C)fW)O_(^GyTp)cw
z^P%7b@vmy<fC_6WyPGbXA)ftm)aJ|ok|0_XRj`In+Q-kzyhGXjt)~0m!Zpu9d3tj$
z+8MnFpsu{Y8rUjT(;hft#m);`)FdM=(rUonFcWt3D(5%+-MeQSXM48x6^sB?`!SO8
z6Ei!R2l|wdc0F#$=od)2k>dVhykf%{I@);0zM-Jq@F>ELdVxjI*{&g1;BwK#fP=TH
zFoIm$CloxZ#l^VI5`En*;mMsH;sW=dWCHjqZ6a>^J2adG6wizdt;Ig}VK~e!%&f?<
zok1y>&6b*)ol*>P1d}4JgKbl8yL|nV7&%f>Q}VqIdc`-+_+8d^b_8iQfO?0cG{<=k
zHMOhk%yiLL+dQyFd}h=?Y)^rWL516USgWVha*L)#2d%WP`?4h^+y=yo(Egfu>p#gU
zvvt&4fSIIg7YbTOuYm=zMsKAe{&hY{>u2wM_)G>%2UB>I3dRawD}lO3ou<yBzn@0h
z0apY0fzYqgS*%iYd|Z<}zdBzn&3B7SZi2N=#X|wZj2>$VF{?b;pk!tVO$N=#w?UE{
z2gIFGGW9P6vLTqe$`wNgqI6>c>kDL3du%)_Lc)9ot*woJ@V9Y<-S5(UY=bZv47S>F
zJPM>Q+?-Z`blcY6;pIdldN+oHuEwga_62zIXF66}n=bxpqm2aK>GbUvFX~GVZ~vQ^
z_B-Xs)iQcIf}3-;&|*_nCR&@xH}ZQW|7!Y%aCfm&a&UAqwo10K>E}F$Ql{fiuw`P5
zsXUFZ6WQWeFaBdDc|R~+QDq@!|771*J~C#xK*0T}KARPItf=LzUgWQ-E@qtOch&hu
zsL%_SL7J_Bf!sx3z**6Rj5NCMkNG(fKtv)aD%L+P(g%!qSSCp@u?0fN=#AQU7RKLg
zB@iuNkJhWVBq-Pa)VFK}bP;@big!X;AG$AZWRiv<kwe8gLjlqvhPSOv!a3~-UBdOu
zF!5sG>51vqCmtaBXl(<h(s_;(zOlbs<87rmdB-*R|DL4(5d83~rTkxx#r_KO<3}b!
z?ydV@I6FVxwEhZy50#|;zk~rU5@R8gXzWY&VLl5(vnrcm7wtda0ALiBDDwgfmEVx#
zkvpjMZ}cjUp~m{O9eNT>Yh+3U@O{rXUq^O($gRCjZOF!4!s6fBljHZ3n@DU5M`UN3
zh=yR{8dy?qFt7>RXv}l@E@l_@;n*Wbf*B<`rVSJ2p&_n;l1?-m=0EQAb3&vH&?%Hu
z%S-eaIVhS3@uPlGs6E3(6)`A7l}h{KC7Hi>7ylOz$(R_;Mj-fjv`T7X2l4#;tnJ?i
zG=o6*87eR|iiMofwxs{3*i6$HuqB8tGFa?cj@3@%old755}**<uz?{*;+RL_Y`(sr
zSqaUeY*|;oKAms{Dxa?s)(=XZ`{4pXD~8se>rno}AJ~U35xs7k#Mjz~+_`WYijEv@
z>)>H6=PuYt2f+>lbqEwn=07>d;cqo9U&=cdg^xda+3v-9?b#;mx$4U0F%?PeSlQ#F
z3KU(&Ph2dn8E=VI-25_}+gLr<Ef&i;PWZXxZ80({%Q<Sm|3QPnFN*}rqnXxhdY?Fu
zhPk)EvTL}?j5ul?-afDMFQ1#k6+Ac_CKxK@OgbO8zL7sM@PDsabH0ADpbi5{l3lN&
zO+rwu4xGV~Nn~6v*a~Io8^)WC)d)|I5>ARlEe9f+XB#T3Q-0zLll8>ctj$>LT~lrZ
zTdX~5j%qBGsx`XOQ*Df+^^!~Jf)=-6?JnDVv5h=obKj+plHm?ZmI7M~BgZ7_WsQ**
z{hP<=F0wr!cUqM)@+2K#cx8Um0KPL%a3b@8Wm4^s2ek%ASqjZ&%g<jF!DW}OJlaO{
z%Wtf)inWjOd2(;I<~rbifINAU%b(eLb|!C{<wuPeygfjA(iU}7TyC)_5`O|xHsemd
z(3I_2QxJgHukx=`DlZBaY4{krMvVUMy9`LEZ2te-%F>e&GGG)u^$*qyd#Kc_8T+mo
zhR<EGLatepzN7ovAH2IE3&AcEwg&yR$haNjnSJC3Y=4l~ZS3zO6NjVzG$ryK+Br#r
zeFJO*SEn<Un~f6ww1Dd`Wq)QGMPGk1JPZvLF}quyu(DHFT6jYvyCrf$p*=>6&z=m_
z3S0v7Zeq?7$e1Zg1~1!`oaF)yl_YQ(95WlxndMy-J=<3TMcV`I^!59f-#m~jdUsS)
zbFGU4t(61`u9raN_HM5iawsu>fT|^YvAnIYHMtm6q-H|PPLSRbHFzXlfLl%cVdUQ~
zQ3ka`-iko{5JTRqesX~$B`xO(6l&{dX%y;!kN?)&vi3&&D9gitu1SoZPp_he^gSwX
ze>j4_<}1yxvVrf{SGgR;6vm}MowD>b#+?_vgd?d)L`2{$7572=uf9ykI5jZJ3AJWa
zYV%23=Yp3E7()yRds#BLrbx00R;zk=gh8$wKfm2f1dWcpQcbf$-t2yl_h~H4*L-&1
zd44D!o^d@Yl>kAJ`(_%Au9o3Q9<;*INS<`KYif3um4H1>>$AqE(>?XpMgZBf-ilNx
zZ$yFNLUQI2LLJ=%sFHLjPh^u&AWM_;+RUEbqJ+OSJ0PF%Um~E4Gg6}ic)Evw4L#CC
z>lVtcXsD%&VXYLBrDt)O0MVX!f|QpFa@G0{FGciW8EHieJj@V);t)UkrruJyTk&~U
z2DCJDW>W)mrt;Md`G<$L{>|LP<sr$NyD%1}`RRbOr3UU3yO^v6FxN6Dr@B$SFV+wF
zJWXAy9_N2rJxej`CUCtoSFK4GH^VEyiRigZ)!8A`&!r}_`KjLO$`dL+1y=Ca9SDI|
z8ojwmtQcX1XRV#|p6>AtHx8y8*?@1nSy(J>fkJDCwuZs6Q1I;@&B48>vv318{&0VR
z-Sug>aXA|*I6RJAN-Ykta{8?Mm!14KU^2w;f><mgf$yf~AeZmI33lHgK#*gZv;~w?
z^2Q*3A^$F*=}|oX%`T!f6M*|qkB$w{l1Q$NYUF3<ewTi^EW~$E`hOBmj3h!CTB2%z
z@)736SZN&jEMA;bZ_s6Zben4}xU8cfX$a!0)@$k=QV)6q)F%;{u}&ri9spj5n=m43
zY1YZ7t7{e5Hx~|YAVq!LS=>=NdG7iiXEXeXw};YvEgyGy06z5rJ;Yt7E2tg#T)$w7
z8)+1`mmx|s%oXM<VnRZOo;xJkQFv<K%g1<1RhkaPASedLg*%hS9BS=@FNBX>(V7Qc
zC22;j)J(!<ODVoq*^34|G$WoPCcH^-TFo<(bP|ONC}|L~pG=A4)ykvL>;d(&B}dxE
zba2apsv1**Qn0uVp~VY69&8%2Q<>(^z~l=8r<GM`M8vz*Wnn#W0h49V$*JQ!W-ky9
zZas#iVx;fa(w@k^FHU=uoSc*2+$<(3A`9BvF;cC}&Y&!6meJDB3nrQD9ZK>>j!%u$
z7tmy?kK`s7W{xP7x=H?tt8(X`xJ-;a?C#jiPu%#CFDjup(dcH&4L|_e5bkCJ)MAWt
zO}~2G6N9xSsfk+8#k|-7oebUW-SprTK`OCa&*i_hk(X%Fb)=dAAzw*#e+@8bdI)@x
zi_&%F-j?p4V6E@Z=fNmZI``xuyqe^SW<)hW)>4D;+6xDTx`YI=m653Ps!>be{{G}+
z&1Z&=wYxtvdUOu(7DdR<*pG16vneHRA<ZZtnLTW|kkgM8FG1hU>6=@Rk|Exyu3m^A
z%GP3XQoa<jfyh|7$AVk2F1)>_W*LAA8@Kmcc0cpLbF|4I3UKU1O9Vh??4&Gt=lvYf
zD59227_%#qn%Q5L0sO?}YKqtm$>3R^jTMwT^WoyDuJFjx!jvRc5yIMsOBa8mxoR?}
z&+83S<AQgeiF2rQK`Le>49{C0$Woq24c=-5SCtH@aHr)W4U4(Y0E0g3+L#}Y3yyzt
zK?d4Fo*4E0Pt3$1KwwDHN(UsZ?Qb#?iH+uB)_{U~$`e}!3c<^tefA0+o1Wd5<U2(T
z%=-2BVZxhXed(kd3Ey+}G;wFmRmLlYPkDD-*HjC?*(>WkkoN=lrV8*f7n^pfvopTX
z48MV#V!WPafDqjDOi+h!+}_=M`0a<GL-PHVd^@$AhBk_$%VDnW;F3h!EIP$pgMZf^
zO~E}-R|}M;E9k9%7FzHFnoBDgHCDPLv!oon8Hv<%<kH_>-WVaPw@!j#@G7QnQ;#qB
z#opXg^6~;HCLP3v4_-!GPYUIC{`Kji4)#txw%|RQjMJ0(-qR-A#>&ARrLKV;lfPp-
zcmo`IU^C<zmfp@?L_BHnKZh|=n`|FWB{8?zc>C45p%I~hZe;kw_2{Ob*%w+Z>j{k@
zJG(<+0pZ`z_t1xBVhewOk0U*NLv8D`X7|R^k2}<WTBBILd}s-++~nQDRGr1#s-;*E
zkUc{wwW`oElk6HK!IQ{T9hBJKj@?!%(XJ;lmHJ$yNV^@?Cn9Z2?;}t;kpYHU7NLbw
z060t8LTqkUnHBk=I8i3rwfv1LJDf=0|7TsLjAd2___dmnv?lt9CL~46f9S(1#pLv)
z8g0Xej(Gy;A?F;Lx;Dgft*&K=Zh#+4++Wk!M(c7Uonyp9hohnT3UW>0H{beRy$q|8
z3yYGSG~v9U&d&7g-q+na&nWc@*K5#Mn)>Jh!j*J*+pf&@G-Aj0Nwo6osyjo2n4Qh$
z_sKJ%=-&@3t%U-d#M~3EQU8RPfEXNc%{<9SNQcGlW_)cU2@l0Vbfe|QNRoTY&sQ{z
zNsRc0niq>D|JsEW20Ir?@8(h-`%+Yd9B4hsS8TEL^=3HMNEQmX5%}PTu!6Flg+MVf
zF2^qhn_;cT?bP<0@)j>cwy{I><ZM)1<(w%7JOSBQ`(?nUca9}ln5sxL)0#$SwhXXd
zJVK!jT6Re*pQy4I;@?J)V$EsP4C=_Y86^(=kb0J|U!Ujd-XVe)Ax#??JJlT<JhY9B
zCMFk4bHrBQWW5eN9Wd!TUkhEP_e~G;QyEJ+pZB{PtL&fo%L6_*FfbFqFG6#kOPWUj
zC+VZt=yy3B&nVtxe|a!2*d*q7vkmOYv<HGl;7RfpTB)81&ULZ&ZaC*;ye8E{ej~AR
zDXm~7qiP=Tno<~LDaN1EAXlJ~!R2!z6e=>_hC9dS@&o2TY)OImaZzrJreJ35q=#$a
zn`Jk&YD%@ZBI5Zsnqr(3mA`Es1&Pcd1g%=H)|`sDJhZ(h=)R3@fbn-U73r3@f(sNC
ztHfJA&3kfRs1FLEi46!vLy%W`-fHL3(X{)odJws-KA9p|7g0m?u-4l;okQ;>1vdL@
zh6$6$bS^x{<a)Veex86tZ9Vdy0X_)rNjbjS<~P^u%xx4)58T1iKv+U0C&&6xs0YVR
z4gJ=au4&s+a@5~(GX(?mHwe`6+Z{0GZAs^g@vqr_fZc@j#;A#TDrus$z??7_P-VN@
zJI;bDLp_IVsmu6Tw8DUgqlXYpc4~vQ=l^)`OVewCm^~6`*PNW3B%;KA!nI}rb2;?~
z>M1M9=~?Cenv$m~+Sn0~X#AVr{5=TvN>6g-NPBI5?~D;t??P(L6T+u2bi4LX0-UaH
zF_$9hd?Y2DAMwp1VhLDLXWz-XBbQr;hN2>bnI4cy<?3w{{SIPm_Fkzg@FJ$E6Py>e
zJz~A<ymjlK`mO>x_BL*A1TFSsgh*XYZc762=MOq{vvvAc+p!{2BOzz*V~Kj)Xa)jR
zX%q^|S|MkDrNyhZpa!i-2$7uzqdXEC*#k#*nI+n*P)Y-qHK(J)169N3^zlI`!{GI?
zgMu)pZHvvxk1A>QBU^K`NK_?~&)FQj^Cyv=63mRm0p09@Hp|j9HqUBmG9mQr84@46
z@PP#wRhP7wLj@kJXIU%JvEe^003Y0B8^;}km&eV~znOl=AE8}we5*V@%LVxl`okdb
zbYZ#XK~+~P_6n%(V_wGBT{p<yZtndTq46{aJ8T7)0R^G)C3<@y5Tb-Z$aSc|){gt^
zmT@!(M;dH(dUi8(Ad@I%HL-gq0k#Kd$NZG=V)GUy@_7Pt8_T2vm5D!$1+%MYD4@Uc
zt8b2%Qrc;&Cfu>IgUt#V?T%;ftHe9lLY5-?)OAIkf3=zQQ|c;9!GiEPEiC?4iErZ|
z<47O1m84$@v`EeOcg#at$t`hJT1(IKUxAzS;_K(?%&DxEE3Z*J)uD%|`Mf-RmKVcM
z?XKMZ3WdG=A=uRM{Cn;v%AAdpYWf`{rD0A3&4<}U2xn;JSHQEWC}1q{Y$P?A`7umS
zv!<cQV+Lk)xe8-ro&$buD)B60@p|FCk(JVLG&^&!nZkmjf}Nozzkz0u<WZ(m7W^2;
zb4FkDVKDJ5gx)HSekVw2xU|eEHJR{bY=>dCP19mfho4KP{T#K<+yFbX)Xzr)&CVZ*
zkQn+M;BPur<S-}=!xgr_7Li{}c~$)E(t%l}CM&Tq=Qb2ATPVB){y{ajWz^K<>vDLW
z+zN45eNp>xBF_<0Y-^m61$-qn8QV}Z(PzO*p=BWvl1;z!Da2y`{2|GWy|cE~cktn%
zvNh-Ez$m7I>O{>`jGdXd9M&MPEd;gCtgQJdz4jnbcGBmfo$EA1B)t2`#w?M#Ps&pm
zZr+Kz%d9&&ZZx;@7IgITH463bp}3)nF8Wgh<@x6tnwlSM#vT}URm#q1&xhW*6H!M`
zxJpH?VM=eDbwd-$l&PkG_O560+}mu!-iD0YaWXQY%7y|RYG&Uqkc8NkM&vCab!|6q
ze(4uc$frxQ;~`%+QnH}g5NioS-4@cV24Blvdz0M77QY(E-mmA)mbo>d`XnMhr!oLB
zp?f->Q1a_w5<bhM86IG~9xmFCvLF{tv3=gUj{5xf;At_;yv4bx$b(B+h$FB}gLOx8
zJjR$x!{Z2gV)H`k^xy6TLMj8=tTl69D=GRxyi|xT11aAi7CXIO%_%UUgWQMH?-0g1
zb&(Ws;8k{3EE{C-=wNrqll$gcEQvA4a_VkKNZ@^GQ|$ocT-S6Bf=w%UO}Tc9${$7^
z5qylpAk(ETpuYm9kGAy(c$WnZo-Gzjc2xYpo%a677$knU)cB@D{LQmMizQC|U+967
zzGrJL$%U{#s35v5<KOdFxl>uuFDd{Hgix-&MTwtr`J&3)*9Fnpx+9g;<cpRhkWZq;
zQf(28IK*9hDMF57N!+qw>obT$$t~-_aqao~d1~^>s0U_x0-yv){=Ja@6QFJ;2YEo3
zt!s6$V_8PeGUgw)bKjcfSDhM(lxMNE2+!p~h{}bTkd_XIre4tT?;v)$N<mTS1@wMJ
zD)J>5vOSaOTp}%Tb}z_Eb+?%*7&-30%5F=#`Q*1Ex9?80Z=^{k;=KVs?}A3)*e-Ug
zjbQfVH>Y|0J^ArUqmNChriBY>=Q;4Wi+)ssTqkPm%Nat^dEi-GlW(%PLk6H3PpvVU
z2g)ZhS_<WUN~o8(9eclfv>06juv;I@`}JEUSMRX$ZS6xVHNhnGh8KkOjHHyIZ(OtB
zOe!_m#oG~H&ZoPL2uhv6dHLN1G_fs?%D`A5FSGE+I9L55Yq;3CGV(DP+we2}PBX};
zlwUy*UVhsqs1dqjSjBKUgzFw=vGk~{wNi=@pl}JJ266h@{dmut5)Xd-xLQ73g<%_w
z^at8nqkd0~>s)A%TOM0W9nKcQsLRin)l}aZ#FbB4geE2ro}>iP+47Z-Un&<mO<z6g
z%zb{PyBG+Ht++C(pd1uL{eHs6tmf$mf0J<@q=G6Fo~+&|6&hK5Yg2wM@lHWuBbJ{D
zK&JMe=lykr{0z<*6$k6IvS41GCz3+@IhxnY4%3%(?~gG-IXby@rXg+Ny)-kcWIRUW
z*f+7f=|!_c#TRm~<AO*)k!I*KVmu~|Fm6xG^Gq@*vn`-pSS)_{7Ve*wK12^2oGd=@
zFDR;~VMObVP3@SUpW5s=FuXEfO<rsdwEg%~zld>jf8L+S^y5=Szq!d8s4*+-+?t``
zclAWtE>6J>!^w#*L{wAtha!))TdXd`ObF{Hez}tX@*y8Eep_^!@;@(e_o|v?Y_XdK
zYkR_bBCuy5aAGTI+5S3iec$@A%$apK82#x+rMCL)*PSV(+RFHO4~^OcC{Oj5?!OFU
z7Ef;5ijnQ{{!OSf%$uSI{<w}U(sci=>*ID1$dyyZANh|&YVtV+AXjpuPAzNYP}_Sp
z1f}0Goc=ha4jvr<vXKOFM=PnGR-H=SE?bm0%ymWWznGdY!(g?+u;JTV@r`~q=GwRB
zJtRBM`^3N;V(Gl#c#+<^Bd1QkLGCm69mf{nRv?RYg3HuD@<9vawAZfS2l?@V%?N*s
z@ao63HkaN^ewgsH*CqY2vxb@f`-ZHh>!Bd!@^UB~lZ|;_c(shuwp5|F@nX~?gI^=C
zi^i?5?x;srOrYW+ekq1r?XeKrZ(0ujVw}>Zfa>7!IEL_t6BM*A5!ou&sO|xrG$-dr
zjPGG)py=rvXKdSRHP9V#2zCj6bTXo%X`uy(@`BL!%_UrDpb6fW)?8GjJ$GcG=mT#b
zt{wcO83hbKky4}xRzhg7(%xu2Mws(t{`BmgH%g{Zs5Qn2w><}0Ua$qSY|c53#r_pw
z%%fB?7tusAphR(cOe~Za(n9QZXUlG}B7Mv@xXC>qq(^{-_K@#Rr`Ji-c-_c!Cd(cA
z(HFaf#*ooO-Q<fcUj>-Gq)S32Ke8Wp%7MWG-~_yopCDFWx72}x<*r~lG_Urw(L4=z
zWKXB;^-qibTl>ZUxp;NxDvmKWDMwCgNW6K_ssu^C{HY9W_oS|JeuSRB=~_W|L$L}$
z6?kF4%Vm4@+XN>;esi36*@Q#yerobwY55gOi_{%Bp<)kVXY@V9`m$Uu^h2AhQ)d}c
z={>?_e45m=*{{DngFA$bC8W-ZESBbq@S@n52bic)hCMI%wU{)IqC}cHG*MZ5{pCfK
z-Thi+Pk2LVbks`6UtS>;I=0njF{5;{_W63Wx%FJm1uu1J1-|{vyho(1*OnAub;UyM
z8B4#T{~(OPp>fQ#`FLlZXR%ZX>^-S?^`a|prh#SCX~SGacMfX;^-=RlQme^;uQ;3C
z&up=I<>e>zjD%57D4KobAqc2zJY*NrqLC`m{o`aQRAQ)vB3Q@gsmfmQWm4vG%*RCq
z-r$ZCq*kJcro;J_@$&phYSduQAJuw5g%o{v{4DF0Xv-J|@nVeSYP1u;qfD(Fp~?ac
zq`|=lCxBxRZzuy~PX8sP_TaAnKONA%Vf{5pzRP~(bDqsMIG;EOz>XvbUH~SeVq%P?
zr{6JcNoshpMc3u(2-mqt)z2emee`gI-(X<gbmAFCco#8Wf3jFo!!FTEO(wuQF>hQb
zUZ%WPZwaLwQ(t)wYDjYnw^)j?8x9)KQ=1zQp;&TO{Q4MVaZj+;sYd1-^0s8y4!Zh4
zpWKGR60q>@Mloz_NJFc5wrhjfC*ZdOp+^3`=4my^CgqqaG|J)i>TPXaeXP^tB-0${
zS|UBY%nxB>Cra)7wGU%*A11T=4XH4ht`ZG`nBGx6ySkB@2n7c=34hT3IUM$4j6yJq
zc~Ne3uF6>A24L1Cs`0YKn4o`e0IDtb?J(y(!VrmwdtW_$FjdAZk6GT3Q^tG^N9|}E
zdPzqmVrJX5FLkS}oroAyv!WE0%kGN|F{JLd`I9>r>vwXgg_CvzY?Y)>@?qt&(Xs5;
z25y_Ej!T_ns%$SRi+bJVek+{#h$3`p#W9TA%&JJi6D&-Eh%4z1Le~lc*qyuk8p@L<
zSGIiMcO2NQYL|P7F^ENt_e{cVt=w*RddW($Y2DkgO+5qc{?W&Cs52ImQbG~a9pv7|
z(ik_ZkH+Z|-oi-IT_^H5`<*r|vcaYvxleVf{y5UMRsIP8m>$nxVx{2I5bCnNk2(7+
z>bxyPT6-pF^t<&qk#B$(gEt`HZ77G*fNXB`hEd}vNoQZ3(?H@`E{o`ron6#}FtYBt
zhm^Xnu#`A4B+(?t`mv?|*1~RBgX&seGjX0vNQ#O|MK_Lf4o<KpzLEn)t4&2^H<^Yq
zWyY$eM6*!j*QQKzSblq(4&FQBG7^42#&0V0cnL+bD)qIs3Jj!C5u@9#MiEq%T1#LZ
zq>x)<glMLVq`(!|T#$Ji(ue1metdab?8)u<UMuD@@+AbFhc>N%F#3sQuNC*-3xE3N
zqrUb=jb0OUi<=N<TfOBXa@;(>>c<N4YL@Pg2!@H=jzcWYtmQY58v@0#g4tLLDVmy_
zgtyE@Lk8{BffI(BWNEI+(nyMm1tzva2EV-2H-mW@Cy#d#O&0d282nB+GbCu8#M6Jg
zQT+rUO|fP?4P9PQbKFb+kxsERfkvW9d1r<jphW>7p~md8m3>D2E34VxN1;VtJ*}Dl
zkLA`Nn)t^yaIXG5T<o_$KaS=O(iGBVr2B`blQAg3V$N}(nxx!9ch2(!2W(Ug&D9=N
zVPrz3e#nKjistum_4P^86V*Z-+eU0}%p2R66L~WKfRd&Tmy+#)40Z1p`E^&A;Q)ri
z6(p?5a<OY(FL*kUN9vSr|MnYePu&0@V9@Gp72_j;jTVS~%tl%~AQpedRXbBcZd~bs
z#a2LTsc+2jTbPY&_AzxMAQ%vggpRju+`<HE431>1=`;?yL1uqeK*KY_D_`GZ=jmyh
z)SOXfzif~QnIO)5=}Q2|R(%}B9MC0w2F20(B+*H~9X9w}#Nf&t&=p=+0L>r1TK4CQ
zFSAE%UJ-!DXz;2y{Lh5ds;ECtsiVnyFdS8pM$GePOy++ZMYYXvy$Vm7!XGmkE6{UU
zEy&w`Y?0jqiu3VW=HI?O&V--V=^pp~k0C<_M%$}j{E(|1=%L7jyk$(H*zK4h(;QRT
z-CAsQ$0mBQ{puI>c0DFN@r4+AG*JZ8#mZIxJ6zqZ6Z}1-r-22(-aX)NP>m1xN`w$6
z8`{1a#Y}hvhS%ipd3Q>~n;<OhVCP7Jq(yw>`Qa!&_^(h7_&5Nbxq3g?Z`~_zdN_pL
z-~6_CH%5i%&Xq!rRO@IP!Q?&khU5?`-8wFlbncMCB9Zu>tiFna_~LQ7-4v@><YnmW
zS75}%%?&Do3_mH13JoWEG{E>h^=~%jt`dl|6(Y8y9CzQ{MsNXWk;_2;MV`sE*eou|
zF|WVTmP?i4{1#*O*9fG-4S;U_Kp6U-7MLx}FO{%2WXWZ|BoFOv`=a`~Yd*=jJAIb<
z6HyE888_~cw`y9_Vb6a2{IAuB@&q2WWom(m?u=H+897gN<Cdx<Ayo!stY=8%AI{wR
zwzxT65*~93oo)kFX;|35b}%B-RQ$S5gS_7tRcye(bJO~7*D|lyIiN9lC*XWQD$K=F
z=kgPamch`9ZwFwDz*omK%0G$f=UqYv+8jB~{K4DU{L~dfWsmkm`JLSntPAGiVhbgU
z_^F5=MZMF;WlULhBPq;k0~xN>^YNy;*blzuKgOLj6&V%Cuqh1>u^<!G8SiKwe5Bv$
z3sc_){dEC#6}7)R;62p5i7Go})aKx_PS$ELPk^;bFojrudOQ0vay*Bryecgt8uq5U
zXY-c=K-Hs}KnMA4H2A?_y)uqrV(Bp&gYg&k*gH;}qEanCFqGtJ%!NyJCyH8wyVI2k
zL=3O!Pl$(`!DbB*G2CpHJ7e)|B{O|>F?4uln;0=pXbc&fHu6kywoFTzJ0FfL5{+i4
z#&p`{Z~y4kOs*hutfqz+t&p?*#r!N;4eAMekyQcv@!g}gb#+SmJY8VKJ3gu!7forx
zarn(W6Z8Zf;GnhU|IEJ@tAUAgKn0CaW%t3qo^yLME!ryF+)FCsO~pk!XB>6_Z{>BR
zLTgu5Rqfs%UyLbY3Z5D#5Ljzgu!+Je#3?LEh+1kv8D1Xeq7DlN08y#Xf$A*q%W>5P
z9U0J+m&0Zk76VP0;gw<tj+z^D&|$*My=V4@!lUq=*ZXJ&njAd4yYNtU1AS8}Q-HWs
zNqzwA>P(-xCN16*Zq38AvdkL|npRxzFKw!;YSmt*P7;g9z3lr^{b`7F&J+80DGC&)
zj%%pD=S|XPSL^yL^~euiJp{(8uMD5!x#FADTb@e#vvW$#Hc2aoK<IHG9|wTYk7%pb
zdbC7dK;0+q6Ki|#`1XN_ueiGJM1J0A+)~?5GG(fA#YLI{bPL_PY^0+tvS|Cqy7*`^
z8OP#i708P3AXy{;d4#HOeOW0|<-oW!KI|}lwGz>1k%J@c^wilaCnB5##x+YZSmnQr
z9`@7<D(FD&2vV-ijzXP5jG6y_0awODB^O^zxN7}WUA6wG3mnX}BoGn}JYVWwILN3P
z!fR>X=7;&5YP?VAZM8}D^U@puP}Bm+5G=ryR~Zja)wxLeF*SLoGeEXy$b#=M-;6#4
zxFBnQ%*i(DDu72yq&$vF2WHw-beC?i^wE<Q^MyU2d%iKlBd=kPU|mITJYu)cK8_v<
zZ<D0<MfonEJ*ighX#33#xFcB@XCM~$pP7fNp8t7He-QX1@yu}8&ZBYhagMKGYskKZ
zJzL#UK|u(mhSYH&Jplpetd``~RwaIf-Wj_%3<|q`Bk~#4Pd=Vk20y!JL4hfr^2Wo8
zm$V4$+C8!81$;)oKpT{IF3@PlpoBRPN$Nere1%|MWh1=}8VWj`P*kzc(e5!w-BFPV
zwsg#8w#cyx>S9RM{<)2ipOb6Tv11(CZ<84xUsywqEYJ1rG=@e}2%;0f8!5#WDsl<d
zyBBu0t8<A}nTzFcps@GTTmwtU<2Mff@?}p0A;(RxoW^UfFTNg6AP_v@X&m+Ll$W~K
zXNM;P=8oOS{QkELJr^b+=7A`$ww`1@L)>E#-Xcttp0Nz|XBpa84*`&$W^mTxwRH+k
zHp6bGy@LMc{1r<l_sf3w+jd8V8-I{ZQWM(xT8|Ntl463|4*-05DJc#Ux-*AB=xYE_
zKVS#q7*SuL1$jM7!+ckVXVaNkj_G^2MFFbOqTUyGQn-<bb&|*pKgWRjYuk#F`gHUp
zo*4r_iZ9OzSi?>S7C0BkeictWc0LK?e|JYX%91iqJe_?+(oqQnz@{#$>_yj>B50RR
z<6RAnOr?EK<Vz$yJT!N_*HO$tf^7o4ET?LNRL7#JGgcGkQ$uQ;eTbQNH#@flq0kaN
zo4oipd;CjarxC_hF@kQ6r9Ym^T*7P{BYRb9tXFn!vTZLv^ri!i=j&n?YO20n+jvLI
zdR)P8f1uNpBZZqw=u+eF28HgiMO~xGuvn6Xdhw!GFycx)ywjnfl>5B$QJ4SZ`;!$w
z;zNJpH;Ze*RUSg1Lv8PC!BL(CFhteLf!6D2VQrvV_HQVSy-f?-(fK_WTtmZzpVS%Q
ziDZ4rZ#3sr_{Rg#`vLIsk=MkOQPM6VlqRE!w(hqb=tRRM5cy5D#gYvGucqscILnET
zYITyjF1Bu0KsmtM^EK8<>o|gxP?#zIYRytXNL0b~WeL|!!EKO{jlvCu((rf226>z>
zMFG+D^g=A+Tu5R!VRfppdDxjE0PmuV^vB&as4xgP{j`!EGF<dEbI7z)do7iQ0pfIM
z;i826dZ4H7=qG5lQiXI|)Lg+=rs6C$c?&@N1`89eD&mV<4rF5FuF<En9jVFh^A#0P
z>Wb**h*AGeov$?E&@FWURU6o|MJ4cLk75AbpxKpuhP{g9Z>fXlm&P{=8CyZ6sTc9x
z!b-!MrXmHY$+^1uM`2SN{RB8bGqx;!5Tm?~j29M#mgC~|EN7K1)##n9q}lH8Wiun@
z7H~7*2k9oCyNt`gW>)gWfX}U?(2Y}E_nMmx*H{n+O*(&gNvPCG*lJ=`kD(HNZ2{;I
z+gtO{7TjJZLopG(w{1(Ci{QG73nzZhXYjefto*O3F3u3`<Udh|Pd<EfTD5uOshY`?
zAGakhWwwzQ%goNI6c#y^zvH1w!wv%P&}0a@iDFUMD+w1UTLWtTwBa+$;GZYq$H;q3
z!|eEssux*OL0D8ry=up8!Mztrz+y9-urWL6tgf=4k!{ak-D+pMIc5gs1po|s5*pQo
zmeQ~m#nS55{1rDnP9x&ql8hw%%}i12G>vGBrC?wQ^@J|O5!X!t5~ro}H*(<XkWkPV
z&oUwQF5FM3>AnFd@1A^S*;|ieWB!cbqqsvy+dROA7mOjNy}sRMu_W=e=&oW|LdsLj
zQsD+Y<!iqNb-(JfKeA{{g;;GuaVu@A*(XMxou+!4ofWsURi9oYA+)QuSjuZxj9d68
z3O3yXZ>h&BHmli=8$h&!4mH6;bcFUXH99}&cDx+nl1jsv2PYztdgJA=h4M3!juD)w
zLAxe4sX$JB&CaS@zZ_~sO^#AAJFI)o%L%Y_sFoN2lIHl5->&5<fkt)LsxJ>r-pLBd
zh#)dRNe+{wfu4^$q;aM`&UH%>3YRDV(c5&!{6a<bIQB5t6wh<V&>55{;oR?iDV9$W
zBdd{87J?s5YcTvm-6#0V`&!2iM4);e!8*X+zw*^Zw)2y^nM~s38xrmWaJ}86mF&v&
zci2v?96Ce{P_V3~6=0?mpM-c}EHoVn#s0p>Cmwr0mLw+x^>``V1PEJKA;VS;Rs+tp
z&g|p~X+<*SP`uB!nY~OO#-yxN@&|iIW43$onKU-e(|$<hz6?(T3YT;}7IU~i(EK_P
zzN7C$Y48V{T$+pSplGGd-Je75-Vz4cmt_34Ad&rt4{@G3m?>ptq>Pk0QFC`%oawz*
z{{mL|6ReyKQxPA)4y<(%=q*SV`b@s`RPkuTi>Ztem6bd!9Q=86@WiD9?5(boe#vf>
zns%)=`Jk@<vpff>%^A!CgFqIQ1q#J@W~9MDPwi0lfI`2U{P^G#FUK}|Y5yVPhQ&P-
zCE^dU;!Ar>14JRO71OftUnwVigwe7XQS=Zrezk4V-0vBkMoO{pUtwbWgil9UGS4T1
zd}H}yipYBVdu-}-d}bV!267HGPmkX;MfJJqps33?H8ox2AKckWa^%Rr+p7s&6*D!z
zG&xDz@@e@X-%E`kCuRO~_ee$1>_Pd20<vKT@U{FWClnGbVK~Vc^z&Qq9uD_?JkPOe
zDIrpQ7O~!{c!|GVT!jayjvN45+yOhARn=TW!DD(CnEeP?d@vl&>%?jJ9vb+xYXxG(
z*(}trm9-Oo&NNT3SON?YckFj$-;v1H0UYyr#Z2X11?r(v^19*}5h)k>Ti=m~XCa+Z
zae}*f)Nz7MD!#+wHQ+UN4UWuHMM3?&`73(}E-AZUV$JJVr(m@fOD4+&!eM%Rouq5S
zl=wG7z#y3o*cm{4Z`<Aqpxu|A)n}`{CrqA-7z<QK<qrn^83CBg;FlF}ogFFzD+S+D
zv@PFsyc2vDM(ogEaM}=4gunN$yR1AKM}ONPND2N8sx9?c72{XIL%ta@{=roc!TmLq
z3pz`HZ94c7D#40-Ud<wayOODLq5o`teg4W>X*h1LXiIrgYm_&k4g6qRqaG{`W$AK2
zS>!Lk1uJVjmPG<W25y#GM08Q*8dONbgtRByMt4-Kb9jP;2i@M%FlNiQE{1<fOVV}<
z3bF#J`(vlaE-0hv`X8%BuIY>=4A%3FBmtVGi9%&)8B9mFPp%m|l~g5=6LCEx!dK`O
zMVzP_M`jo3_tUbLvLY8sV@#Fpzp*T=Mn3JEIC?_0Bm9f{uwe?7W-~ii>emxyr$*)V
zs#Z^eXf*z`nCmt#u~+dLa(<81n1%bLR)M`zw)?p`$%U%&-nIe@J!qxeWP!3Rk*_o$
z@D;#bDWc*_X1!~lBPpoqVOd_ac`*?iS2<yxK7536lP8u!Pkk(VIgF0jAxFMcg0z-P
zxnB3)H~-97H5-l$3gU|Qf*&jtcMoB1&lBq~b$M7Fzi{==xc`a_aW47GWKcKgxiTS8
zj`Lwta<w?;_l=l7k@4ISmKziPC=W0qxT;W{JVy$h{+)~V7)?-3)(~Myp^TdLfVmr6
zhJ=N-2no;511Za$xg2*=G+fCP8!xI!fS=LWO){2D%ZJ)q)%b|uaJg}A#MCmZpzyO7
zGtia|$Un@&S4-U~guLz^u_$cNWYnn-n8p)crpAudX8T)NS^a*yNsw(AT+$lYiY?4n
zw7@^5-+L?pBo7Q2W&<@~6f~C$e#M+Ii@6W9HKZPzshCqcAA7-K`*Y&p8ep2%_rCtu
ze_2hj!_1;$G`L5AEJFf{lYDn<0;1Av>hj=LOTVbn8WM|SRAgb5U7>xTBQU<~RUF@O
z9p7|t?JI~z%hMT^r%g{bm7f2h4ssUPI+Ul6Ai)3F(KQMK071^XBn&u_x~ZlQZnyiM
z<4Wg<h=93j?ww@v^&r6(liXy1FYn~!LTD5uouD}JNdCgR`FQo~xM4<1eACJD5vrQ1
zXe@<DW|CF;Coh?OKToRPfnS}p7~%%zhtYbCKLZ1Vfc+u_<h8eQ7kN0FW|VD2<c7&1
z*Bt+cr>}siqwCs*;*{d<P~0gLm*Vaeio3fzwAkU`?(R-;m*VbLw79$bKfE{JB_t#z
zDQ9N(?7h}wOJ+Fb5aF!uhw`{`PJy;d8_9_p{_qM3phZHEkwyo-iOqGW2jK%S1udPp
zoc8AhmjkJ6g;MuKbwsjChM%Tblz^m^6f#oRm?1Q>`~1hd3@7BW2r!zcK`<JMlk<RL
zzAK~e>rax^I)^Y(btW@mFj8T%rJPntzbMN5PYWQMyom@5(knGrM+Gyydd00>^%C3=
zSM;1PcIwM#;JDhprXT2P9otfS@U=^lk|&`Cg|h|a$|eh=&&~XuBS>Vz<~<e7{lZ_3
zi{24k+OWEnkrYUNND$PrJXeRa?0S09eS5&-@=Rn5Y;%sEqI5pSKB?x>X#cBXpxrC5
z%BS4%#wsioZoirNX=wLdczQm@?*X^)9;m^p?=5A99sSGzY0zx(nRkk(Rr~OJ3Wthn
zc<DVYaOqgWiqCy-5cVR;qWVY}&_S7>CD0JWk|tWCtI*<b%)g{kO$-g~y0LIwKsVT|
zf1<{|qI%jJt01`Xh`9X3ZBn=(FJ`9BNjE$ax^|Fo`icm3<98S@wdnU7E8d5OILu2v
zCzBX1t??`6*%+(QFu%lPQ#DqIcy1;=H9FXErHK_+?<=gU8%pXO$z`R*P%`4#qI;Wq
z2t93eo42GU>9j}0^zDc}#?*x-F8UPBQk1j1!JRjp{@X7#XJ<m3oPAIBlfP$R4c)Bp
zt62Kb?#U^<dtNSQzN_Ob{RGjE1>D?QD~;(e8frYko`4*MX`|Ze&xK9`eL(DK7W_Yw
z#Q3+e*nrM*^yyYQy`k^JI@x{<%E-}QO+gH_>+t%&cy_0zd>!x#B}s+$T_N5qq#iI|
zh#>4SyWDpmG74IR{4hGkN$3DpX3!j|!LH1tp&D>UBFL4QVtC5$apI3_)5~6<J5MOW
zGYqE)OWj&%bl=TIv*dV|-{DIk?b=OmgTI^G@NTXKY*&JpIT*@jF|_vA^3|U<aF>>l
z^Sn=Gs|dlIiQoMJ_%wF_oa3<vsqoLjN`{mtkBSyki2G+Gp~G~;pQr<>!!6R#fvGiL
zu6+nXe$VnK-r`=mixuUebI)Nh(P)g&N)3g^p7lRu+jC=S62_h5JXg-xVYtCkaKKW6
zmTs_<@<0eOfMW@eoq-U@)fL(&fo+A<&W*iEYlG^_nsqG|eL(8Lq}rmPDDZsn!0rOX
zNxRxkz&}yuDKWH&GHL}k6OGdR8MV9aC|&T7demSsg#9R|JCX7|z~(89)yz`P@rxGq
z;hr{~tibuY(KWe51h(|GK>OOeB(az57tC-%UXqS0it8z09{k|3aa2BQTOr9FD#|Aq
zL-=g4J;=@OB>t&iHQs2TvOTjFEIc(QpAj+~>R3l45DZ|+?R?^x0#Th#55F!&7=cb$
zLc4R&$8CXMkoIRHu-2i^7wdo8TYHsJ?$APXApFb2+f~^=NFw=g?{BmWdItEE*DpVr
zqk@n6Sw4rnL4E^z$M4*uG=6u+H$Hx7U5&u+1z)=W64g{Z;s*t^$y7fbL=kP-T`kNc
zgDIx|E@!UEOU_GAYJL_TUNro}5H)@*`Su-C@jzt%nv_kda~5Z7B6ZbOuQ}FE@%pkW
z$6f2d=%kgYxqxO7i`%275;1vP|LRjLQ?3*S^|xW=CfnRMFxCTlsB%WwB?geN^$_3U
zMR@$zYjPw_^i1goKklY_>C;b`l|Mo#YN`<x47*Iw!gs*XQdu7NW6dU@c-hz$0M;rc
z#2L)Djcp*|qTUh1Mnyj7iEw#MaH$HXC!B(C`4sv*+Y{JS!zKT-S3O%T=F-{j575ko
zU%a|q&awv|Km}#y4SDmIC@hbjfBSp_csQ|4xW2vR*Xjj6PrsTvyZz0S381<vv98{m
zr4?01pvr@6DJotsIjtr&2uy4VXQ<_)j>P_9IDUa}l_*jH;{ijfj<%y3I<Vz)N!6FW
zm2<fgNZ@{K-k}ccYRMgegsszpU{(n1QdAQ@S4RDK)H=0x;5WdxICS4JLyOkfi)lb+
zLGzl+svwfyOv)5A7#-8IGSdnB6EEPJ$0S*E_fiUmx>n$M5*zeTv8wJ7IAxg5=}KOS
zm14mgn&cO|Q=8-$mv4N}G!d2tn;2}Y;H}+-RVr9WA$S_HDyWu}W%yStyJ5|JU#d6L
z>ueKyT!04aQTSWP{;_6Xrw2|TSwZ*z`4=$x`UKLF$7Gpbxctz&)Bzo^4e>Jj=6kW_
zpaDi7KIzH_QOR*%ubkJG*k}PCEdT7s<onH|Km&{D?j$+;4@nM;D8z3P&j_Q%8|^2^
zEyppDXT*j4magc?mH8!N<EA>uV#$VJFxzRNLX}%n(>CF9&I<-ih(z^whyc`y0ssFU
zCjN!VW=sNTHH0sdRAKi{z{Wo0lqUXh1Y-86qo0RI;^{^o7NHp(AJpYhZ~n~3e&Soc
z-wi3Ai^i;gDA%7BH|~ozjr7pB4FW!uIq5#}XI4O%NyW30jvWe5O{y+KfloB;E&5l#
zAtRYT50)JLng5CxJSf-qfeH8=dGKtm1rAmq{ML?MppNT{twHYMg-rjTABsqVJ2a`k
zP<`@I{K{nnK#m>-CZj5$`sN(ppDQiReq-p?qHA$rml!=ZAjl2)VBT+TLi97Mq(e|i
zS>unCdorYm@YHvOa@k;&5i|EDH^vrtW(9YhsURq8E5Sp#{y2S6TbN8|V%;C{M>6dA
zQat9c!a0ayzWYxSWP8d!Z~oe5-c>9qT;6p}#7%fw8h<vN@k}Ak7hR0?rMT60q8`oR
z9R@{_NRbT08EA{3xS+ZY*k>W_Yj_F);o{$bVjff(>BE612yf`FBZx`9E)EDci2F>V
z<pD>ZVfetW9@D)C+?7fHt<g_#TF7KEeG{mW;BMoHqZYX98GhjqTsh1yqu<Eh4^<H~
z;z)5$sEXG5EnL8S0Wyg#uPs}4qL2RX8oW(|3D`JM5Xb?(|87W3mD2~`_3SUFp#b%J
znp=qs{cNP`0v;QSujBfQn9EM0_hw!(KEdHHcHrGX3!<caf$&4>BB(*4O~wI85ZHfV
zBJKj{!M-2<X|GqtL977K9dM?dE#7|Q<~xBz4|P3T%jls4z(QP8XC-xY&<i7Dj!p{$
zP-d}}M~{FA#g;TwLpC-BU!4EOEf~a<kCk1~VgU+6b1ND9H4M{5mOirX(J?l^ap(BQ
z(-A`q=}|@Z*&=Q>tI5A6dAZu%-~o{^!DEW|z6=m4e1#ZSMb44mVIlAHb0Q}EY7pJ=
z#rT40VgSrQOAqm^%$<DQ+|~ibrO>_$;ocj2$uQ8+xSdA?yl_Armxac8L=%{;k(vTQ
zRZQY&1Nb=Bs}29giiY$6X#|Fn?qraKzVMFNpD-N-zw(^8mptj~W+;>-EAnGwVa2W@
zeunR4D~ir(NZnLCsgr7N_6R&rtac~V(eZJ@{(7>e#ZH3Vd@G+U_>o_JbVMRwFs|B&
zuI=SG)5KBEqGX5LITUP``GTRl+Lm$#1E?W|w$vb60SkE6zbIO0f$SwzDEkfom?7Tf
zsHZE!;P%LbJIBFq*gnwa#1o!Eh5+<bUi3EQ*5>z1K+{R`!5-}=7@`B^AESpQ-23IR
zc$kX5o5MW)clG9u*zrMWPgOb{br%4;t`2n%kTFm)rClttYox!-i(!(jNztqlzn^f2
z1&=)R+6Ki0=y~T%ngr3|DrrTzINyYd#Dp|jTM9(4NDy)_9RwBR3hdv7!Z>2aC~oiQ
zgeGEtUryN!>Nx4c9v-6uV}GLus>C67TNuEUoh|j&3bfAl>BXY$SvWrxp#h|VqY(fH
zRjV}HdG_|I^Yil;vsl+2e`ANXUCl4>0tzxJvXk!+?Ag4_v6tur4SpF{|1S<c&Movk
zORx>>D>>s`=}VuDd!nVl>Jx<eZeeXEv!&``5kmrN07U<Z!H7mLCUJlR?36>{z3_4m
zBkU5td}l{;ZXz`_CHik^MrOA}DT$(p0MJsHJk9$EqM#I`yZ6Y(I6GUdMg!QNMQq_(
zBM@y$?ZDah!`#4bi{|JqLEOES8qhMi3Gy_o+Dduw{)D}mmw*6bh2L+un55D~Bp|3>
zdnULE&??3mGkp2fwFQx>c;rxT$JEXo$khswIswGgvF11;eH04^VJkIQ+gvC94W*zM
zj?N3|O%fgynOpL$Q_$NN+l-;LJP;87{IUu|l#%fQoiE&v*_$7q4i^q=3V7)6wE75H
zf5mFQFVL}Iw|=)87T4%x=A8G-EKF}cBT9_7mU@qDOD4}ZH<JV!GTDE+TeL}j5XD8k
zVU<U>_ro2qT~Uw>YCC6G!VJb*g6~((7~ldM1AzE@h?HhZX6kM=%&v_A{9Y)YM-=(d
z$zn(N%MDLNXV;htmM{wU(iAK%!xYTaF$}d_DDO_7IJeow8z^aLX-?bv*==urpaeu7
zMyZ47A^<*_!q@q?=t9EP1rY5t-A@PU<CQPafd2ly%LHN8*6yc4SV9uQoKJnVTId{i
zS<QO#v(LMuJ@*3=dNO^t6*)GH4x8TaGfm&qmHJ?t?Oe}nE)pru-x;FAQirwJhMQL=
z-t?YeXhig6bnvmZ)P+0wp1SSkef3aB{13{IzRmnn_q|kpBYf=9w4m5nT~-Eku;&N#
z6J~M1g<mL_5l9<Xs>CFtLG!+|1c%y{vY4jyS0JBkRa*&N>OfE|>X=J;44d!&rZlM^
z7RmpcSXyqhe4XdodNB|7up|BNjpHs#HmcIjL6=?IYB}XkL9F`GK=Hnzc&SF_0UCUQ
zu9_SRpsGhwDF+~|cd3H9w3gNfEG%@4(g9EA`&$f`uPJ$P$qzpzy~&D{cs9tFFaIhs
zjXc(QY_#F+?;l?VVh!)9sA>9RV2@rqVw{+-%nVqlT01*Cdsbb{87ZOf?9%{G5ftUR
zdy^Ug@MO_dqmn`WmCQkSjs63bFkqu)EdacQXd>u>0<@<9ynpt&ZMjg<d9?cEGv4qS
zYr5pjmv|y<2HL0ec?A10vu-_!LeJ?8w(mMZXKbIL9uri)x)SzBRdt{8d$kjmZk3fe
z99k3^Jl$59+t>{Bf-zXkPPLh#@45Ff1F-1T{teSt^@dG!{#?j%;5qGl^Yy-KUGH0U
zTMbgtbwl*M&Li<$*wt!W3pLC1g|*|~tGK_L@Wqb#d4Zgvko43IaF?|cyOKpnJ=X=v
zeYs%EX8MsW=kkpUZd(-KF0y@~Wu%n~X3hh0hxZ8q=aFhW=%Z1V4NdEx;Es<U!OF`7
zX;FK0(iDoCYQZp0hm+Hu?!Dek5|OS{eFDJ--B)o>rxQnWnwQ|HTuW<8Kk`w@K;T}X
zWa|M{A}VmGmF^OEK>h=y!fUaYkPuO&cWkA|0I#HCyh?0T?iNU6C!mCS2uh^fIc0DC
zsBCYT{O6lOTO$eoW;;9MiOv03oB~DTawK8@JP~i(jfC$NZe{02@OH-lu^*vPS!wLD
zR)gKZ1f5?bJT&@#gyzN30P%?vl7W~EUk**=Rt%%A&kf1_^_<N4^^C&SgWwAukxoW2
zLsb`^SqJ1^hMR8ozx=+!{tVuW3@y7)nEA@A$wnZ26;&DPj!#%`S?z3)ad~tLZ(8l6
z*Q3T%UR1TrDX7qxsL{ZztLz%^Ki4YHgdHkVw0JQzL-b!^HDFFiA!ya!qwd@fk8l*I
z@L@HdYm8HThK$ovMrEhj+SA{zGD=jqbyQY$_^s|n@4PbQZHnYGc;-?nRqWc=lWb4$
zKJ61SH(x8Os;bEkUX4(E(O@`)X8HKdAoOXBwP<MVzi<ZJYfw+3qNs}Id=NfAT#Oc^
z1WvNK^vGuo-{SpT%#1R=w+;d85-V|s4OwTmJL7E}f8hzap^?7%C2<ei_#$xMyAjqD
z>7b6^-R)zP&J4(GY_E-_$eHby`)Ug|=GTPY^{u#eJj%ddP9}9dm${0I1|7eiPl7B?
zq=0HJkOm#Zi6Yh=Fh9rnb?fkZnc@76+gXA?rq8i|T!!9?Q<OGFBr$wK=FZJVAV%Z>
z0yQ#MqOR|O*zDX~Zb8AP7fpiG_PY+qG%Tca`)7xmHh%ly0haUULOJ)V2+o~LGzr4U
z%66S(0h<<QQU1H<E4T(C4`|=tmI;JuoiDe`jyk@+0=_<PFKT#>PT#+O-=f!hKrAgS
zot&Kwyu5TI^jmQV2t+_<JZ*cj<yP2SCIb<MPIy16e17~&;jmqhq$h4<ls6ZNqvwgR
zpzFnWczC4rT?72ROl@dqLC>!T9k+y^qNzfwalCp>E{$Wzhlhuk#pRG4`1;|}>=HXA
zlh53C?teV8wx-MXupQ@qQINEwpy4p3t!=*{=6;)EX3R^L`$g6VumP+zmX%>Nt~r*u
zFce+YF{`{C`DE@tCr?qw^^u1r@e*|X4K~nb3Y7G`GQ@?=IQzRBu{2+80i3`{aYIdL
zma53)jr$)k=8_olwM%zY$S}d(d#uJ8kIjeC`rB6z`6D5_CpLL$L(aj0x!)QDP2hkN
z1d4&@!JW|Wcq(t20CUN<jz|V64t8?U<(C1jF6WjSAip5e;6<2dR!L*FSJGvk+1Q1!
zQmc}u!1e(F%R)-eIYv^*f3kK$kXaBs8T6TR=R0>_a0G!-BmW&?=F_$Xf*2dR)**fx
z?1@YzJ#d{_zbE!!KH2neIN02s>cHRLY=ra(`aeF3yZZ=e-?X0%UtC;G@JAJgbQ`%F
zFLv2gJY&DQt-evYK*p#@;%rxT-fH4$Mz0u(>WDyjdU>5ksp}Ec*=cxi+ji_=4;Zbu
zZ^;FW#~cn15BHC}-Gk`ero)&!LZY%fOSd@C@%(<9j%D9CBf-=QfyHjG_)cb33UEuB
z00oJ|gGmyQg92r7cI{z$JY-zGZwC<xrHJs43S!-%dLjx^HHCA<k0^lS-Q0!;0VcXg
z7s{iZ%@;29S0$f)P?l`BsUN#6{IMfcT1k^APmUhsOjexJEVuS;S;{Sk1{Gd1$D9V4
zCD>r}P2#V{C16B~25cJDwBcHRj25oRF_GlOtBZBgvNlnmFGc?7yZK*II^4<;1jptL
z%9pxO(2>|JOV+Uwc^sB>Zn2qVaFb0$;Q7Timu=>>R?>uqVBlyw$2`-JZ=lZWx~*o%
z-v0iz{Xw(V^pmuVdHKno(^W4PqkeW(^#%Pcr|K9fyOxW><nzGZBfH+h{xm(Eh>uKd
zs8i;%Jjg@=q<_SM@xysK<Aje+o`8jexfZD{*z@a9S#gyo-b)75pt7vzbJOMNlZodA
zY&R`EZBvDf>P)OidH7vlg+qj@vlrGK7Sa<&l46tOUg)OYI<g{s;;3PPOJ~*531mjc
zd**-yn(xAIA41S2q(eUv^~6QgDPkOd%qg%b+EG&$9{cfaU|mfxI%zO#M}me>h-TIf
zOye9-z9Q0sig}1)GVA8)JNEOFF3!F9CC`6mL&V9&n^NN~wX4QM7SBJ)1C^uh7>KaE
z&%e4{8QmuNAnbU)j*!4lj;Yl*9O`uZj7tmd7oR0`o-{s#4HnX)3ADL^)8>qilh#vP
z3?erNLjQTPJDzu#je3GVeVK?KubmK}PdaViBOXZ+-Ekp1PVnnMubt>OgFCcGk0$mx
zt`@Qxx_sEUaaYvl<1Bocx%zlsy%pI<TZ9BJDi;w^gA9WpQEIzcY!n&kfK6*aTMXX*
zgFu&ThKNUxjX>r~5PTsAfs$M{o*QL=Tfh5h)Ygfg662*Q1#>pqknMNiEy=*H8P5nr
zx!NA;=_;pzLcBpT_!F2+nugrd^UmmTmNt)SWxpF!=)p;{4o{{rY*+?b4kFhz8b}M{
zxQm`Rli-Sf6^_X?LzsqL4-K#y9vDOTuee8aI+j+@9hdoH%P;KpZ`??oiYWW6SbEk$
zn6DbXi%fHn%aNqO3@OJRH>TmlVlo;$R7>z0Y3J<}<YFt9DeEhkVK9vZx#H6kWs<fZ
z47v1H)tKcg?wkGAiBWAb*^0uw<S2Bf%jdf{VfdQ$y4z21ajQo%P?+>eyP%L28-}Oy
z@jQsx8x7~>&aY?df*e2MU-!q3UuJrbMc>a-KqQf2o!2I1#8XSAlr62>&sVGkc~Ecn
zdxG}I1Cf0&INCqG&e|6=JI<f4&@3$bxg|(~BA8KTo6q<2wp)fm8_k*i5p;HD;k_hj
z!C&wi{iB_nsJm~PYv9zx41Q3XS0?Kv>F=n=K=fogjoL-vKN%TeOg|<99JZ#y-Qdgs
zv$Hi@>m!ti(3t@y*&&xH^A$eHJQyN+s=H`b(i)(3fTl<ME1t8wZ$q4l99Uo_?PZ3R
z5kSll1<i!a{hpeb|MAGS^9n?t#LPdo(}cz6YBC)t@wIy2X}1=FdVFr49~%)?XJ;4R
zN|NS~=V9x46`wMa;;V!FR3IN`aEkqGBhrfF;@gC+Rj7_c6eRIIR0|20J+C9dOI{K%
z0c7eXZGGS*hdG(Vx>91o&X=Ti#pd)*rG0#<Q0k@8^YCh-OG}N%cHl02IYt7WtoIF)
zE8fvG#lhhdNP+pMRxu%hY>MyrRs^ac2z7J)I%m&8*RJ0W7|xiIA8jve_(_QDFErA~
z0h_9Ea`WqCwzhz)K`0^b*SQ^pAlmN>uljExJFvM%V^wqsd%X71#h5|y`eMsgrP9W@
zjUEY<@M7e}C4`*>SdEk1gnXUKxAd&@<o{)mQ5XV>^0Ow~M*Lk4AVBn>TLcqr(0ats
zkJP@rUFY?iekq=h=mmrx|2wEOBumw+f=LEFl#?_w5Dv*t(#7uNG|;kq=4}G}FGO1_
zvhU5z7ywcB;-fp#AT96O$elm5eZeFEvxzw1lz4xm!$FhH89ckuj3h2iN@tF5luW43
z+r##!4UTpfK_=x#nNU1;M^a710YY+q{pe9jLltvT@e*8|r&&HAlKALMUaYV?nu**w
z0ZRXtXw3<rsISDfBtTh1h*<HLf7HHLXO(a~Ifajcpz*^ZU7+N|xlSCsAL;Kud|GIT
zu!6r9T}1jy=`+sZKHNaiOl@%r|M|y!BKHv{-~_l-qM!7#tXGY%-~QDpvNT0BRB5y(
zu1!BMiV1vi^I*k963DDv$3{F6v9o3?x!RK;?k-qwP4xn1tEf*Zt{A!hyYobqz5H@|
z*Xcydq+1iuJ1-iXLm)x5Lf^PM_H|wVy>F`|6%q=fOH<bgeGFE!uCS@=%~XX3Nm9`D
z5haAtEOL^<m$+ck$?ZT)k0|ap+N%x{&<15{!JEcPVD-z}5rCzEVU-(_vkBjXcM`Y?
zd+8q)d6A;Xf@&B#QO;~CYgc4r&L=a}^2CIM1a9F1%&jfM&~~qnhFC`Tg@vbd?p7z-
zM_ELE4pdl-39H~vtruz-YstfS>GbrIz>jkDT+m;qW&H)NU+9SyxHAtTe5WXo(UUhy
zY|fjVqmP_vFS(*D=zPqwNAyQm5mx3|vfF<^#NA16SAw9qOLg5>P@A6M@p+W&ysdHa
zP*|7*LR9Ashv)0L7yInOR#D>+saw-c!}9q`|DOGDx5{~j40F)H5m7x;xH!3yXPYx<
ztM(m4zzaa<5g+-w$ts3Nxt*VS72NGO{Y{cPOwxM_P;JF3ijw_PMb6%xCvub)p7Brp
za@3|8!bgc4GnlU}RcfOu-V){S^CeyT%P$!>dZoXq!|2|>>(P8W6&eD@buvK{oQ4Kx
z;$ixFopT_OW;NjN0pAXaoXWQZrOq`Bdyrhr;|(j&%gR)kNVoS;zt|w2$1$gAg0IwV
zZlXmF!gBP_Ut)O0Oxqq-JoYxG20*5mR>c)!4i2n`M@Ou0{{kK~zaU{G0VU|aH~yRa
z74I{j&fabn`9CTInTaEQUY7`bSEFq1*AsrkUm^E^g0}y=x}FVDRMT_K2-OLUQT2P?
z&k+2vutaGk3Fl&8X3ln-lEKMrN2vfM?v$g{l=Xpbu8VZ`uMIF7-!-bol%y(8yrj7@
zIdTxVc10qXT3j@gI<@#x8Y$0P^vY^@YPpp}Zp=ol+1VNWKoP#Z@>G`|4wE~dOAxIm
zoLP_OjZb!Lf=G?v@+!CyfIvniQ4wL`D0f^mIaA<a0+wHXL6{J{kwes^5DJwH%DJsf
zKt2QYk?R$Xz2W(Y-YYr|GhmOH%_8>2HS^wObIj~#yxLpH?kMr7aii{GLCB`&(K<RT
z|M0RAO^HdH`TVr*a8=Q`IZR&}Z8eA>I?~y8CeZwRW9JQ(<hREwf*prXKwy+OWYBA<
z1PheaFy~J={L7!+Qc*WsjLiVc2pz<F;aO%1i?dr{=KIGr?qPBcPmB}A@k;M4Ha?~D
z!e!^DjveNNmJ`h_L83GZj&UlbUnJonrAx7{u!|Wd#gdknY#;g@g;7<}q6NOqt7+wJ
z`Q7fb0DrT#|4p7+r|a^;dvFXPJ1<WrbKtpDd0e5FWeW*GOy%n8{tHVu)b8R95kzD2
ze{g~<y6_+U7fG%R;D-sKNzeh)BavPfctEjdto^2+fNNWXq$Uvvb!KfWmW=cQKr7-b
z?Pmmr$9r;XPv1DCs0M=`SiYO<8H`f()qXZI14%HDRio2xnNT0GtH8Gbo$kH~7O@oZ
z@bDP@a&u!$6}KKq=C>&O$f)V|l+oDy*I@O;?KOT_1xXL0O#8SSSjpjN6p0l6qH(dC
z!m??aq(vfFp{{PudCfM|smrbg3TBOM+*47EVhId_u5QQm!!*J%Gy;oDfk0@>`Pwe*
z@ycq#wY&8Bg*$9voorTWDLP?WHa^ylnB!}<he-f53J*v9uXMpaF(($EN7W;n6iF^}
zk-?!!XW}SloQF06QL)(-yS;B(jbDdo>+4OPY%#a@hIq^cmpXk0GV9(#)e~^?nu<|p
ziWK5z(!XJw%8m?>eJrik*G)OrdOu*KjLeS@U3hI_-pLm{BW1ME*Rys3&$9n%0e-<r
z0#iG5Xe@1P`ojO#0!URqVdnyfp$cW@@?)^GdC#-)!}AZHv9HN`I{d{*L?WQ!=;XNK
z@K<u`IUSYn3CgXy9y?g>e8kTOF`$B(9fhc8grW>jI#c<KV2=c>%EN4<!xLOuJ|n^u
zmxW5Gvr!s}n8;P4c*Fq`H$vN{S5NQ9x|yALM{GG$pw{LYz$7KnsJ51hN+SaDSZb2e
zt9RFbI*E6JfgMHm%iv2*eI?A9uOfXB*l{QNo1!MWjC@A6j|hS&wc~zWL%ocIw1fn;
zrsnd+Ya;B?;s!7Py0{BCGR>DeX6P{3own{J<5^&raKM4p?LI~IV`Co58ZN%<Hs;nn
zPb9O$=bo;A0*0rnd0TID>uunG<{9`41Rz3H@WWTEABrl-QEvJqh<5549O}D#h$@3T
zOI1$H!a2SvBQ%7_Lk-gpb!Ur|Hp|cz3T69G$cny2-x^5Ricqw@Qt4={|3mYPA=q!d
zCr<t=mT*nO?|VpIA^b?Aul4WNe1P!N3_o3K>2iHN5;}T%rH%8S8F+q@&rPpORs~J0
z?fP_udpOypxrWidS<$0^#>6EsMme1<)$gAk*7SZ#J^kCPG9LGHYHo^_WOSxx{2=Q&
zU;C{@A(KN=TcgdM%=6B5U{-46$<u3cc{$+r#D;ZG`pNGSBP*l?6o`TDQS2j42!DH<
zD~Ky<r@&d+yo(R???J2s6%6!ZW)c$N+2`Y(uR#OEFEXPoU#-*zjvLS8gXW)0zc2jj
z4r7uoN#bq28`kD&T;n*elu}D1i4wfUZD_O`XnpftP@3hj&-DFU=+}9LtX8E<)TmuL
z$y%~@NVm{Gi)Juio`O(T-5Q|$OSh1c^_1D~p*B5I%{E__EDq{<g0*N_Wkn%@nQ-Iw
z{ag{z&+i^7Nf+O+7Y(J7HyENp(jrCuE!?7n7g{7YS<Pu@<G-+1@3_&Ghyof`kH;SJ
z7MXX*>qfjQT(UuV+l{vy@?G1uo$f>7H%`wKlnRQ=dlY^lYd*euV5MXpSSrlQvy6)5
zFDF3wLY9=f@FN$LEQ9?Dv@L*f2p7K<nesyLzc&es_+cgsDDHh|E#}<TJ*YcgZ>o50
z^Khm@Lqky=&?G~(NiO^N7KDm$le@*qi?DC3{EW|C(E}s-Pq`Sh>wsRAoyr)8f4>NB
zvnb1#hr@b>X2HXtS6c!dT9O3;)bnG9lwCC;$>5Ja3K@Dtc4dQF<izGq1et>fC=fYl
z_@jt_h~ZSOm?OW>akGWwpprir!`T|~dtrA1ZE>ayQ#&SBLc;XdSXrd(FSB-sO>V0$
zNqm)-Ns%3qks|V>!yY0zqaF2^T203_XFmqMPr4rR4)^7eZj2mVAZtc(0aAU!ef$HT
zBBySIfsulV%qQ+Zn{_nwn}#_~ieso``I6`cX0R<q>{PYg^F@Tz9f{QbQ$e=J3$L6%
zX>&kw3>g6Ec$SJO>^Lf~Fnh%>#i^BdZZS;^@5tZ)v?T<L4TT^1(9H=1KtBQD>1A;_
z%)nM<w+SXM$?e5K97vS354G$D>xhCy^FpqJD)_#!7qoqAk&G#*xBJ!`qA47W6sOJt
zqji|N^Li1OlezXXI5Gz2wxh&6$dzn+mh$x3rYkL_scD$0t8<qR&o+_32{nH+XXhv6
zk;Wvy2o>da#O%XZ{=xDeE+1xq>s<TW9;si}@;H?)&hR)@3SbCs?IEZE`0LY02t_73
z>G{MRDnCUU+-Fho3<fr`hi9CQpMozt%m@l6hh>$qDu}FiK&NLgdF%5OaKGkO)jQ$)
zde!9OeOw#`Jg2CwhOAf2w+E89Ur)X-N33V^PXV@^(f1#EqNQ$KkV{iQo15s?-lxsD
z&0)Riv9aiN?2TNZ9P*0A?1;?rTl1yE(%W(={?wW7!moJT8~0o1^>2^N_iefd@uYt?
zTaGG4bbQ(wyyi8`x>@)x`?|WDogDVbK!!cRMKKu}K@ucBqjZgyb9Cv>(^8lF^y-!2
z7Gfek?dh+ZnO?sS9`<$|_Jy8gm^&FnJW5P}75L2ckRP}GSX1+*Sx-d;L!RfsI6fgU
za?phwEhw^xL<>#fSt<ws12>kes>CGB;5zR6m|GkpkZziWb8<iO9?e&T(dK<LxOhyT
z@4V<gy1#h=y7;<-rc3EP*L@!e!(+SC&h<xRZ|{rECShfHdD|5yuG{^*-!@i^t}d>F
zZ-SRkJiws(cIB6(w6uW8cDU}#nYWB8-;*-;{XZc4-WZ9lql<qBW?3$L!Or|I6F2V9
z7Xv0<wD^cGeq4LywMswOadH&l3knK~%ZtA`IbCI+!GL4q^9I-4MhEpN5bxr~af6kt
z)<y*>0&$;@umEn;SD`ZsA~gtA<gU1Y`{=Bx%yZ|mV0l!n;e&fbCi91T_?PjiT49Kh
zAXK+cd$LG9=C6}667g+Mj|3R?^uW;yCEopYbDKf==CNIuyO>PKR7=OH+fy=}hBlQ^
zjMVYTdc9g=EqvzjFw{^6zDA^bY}Vh~gV2XJC|pIJROg@4%~YBpm{8KwulOoQNXn+?
zmt(??)6Q5jU8~??mLh}lfHbwV%<k~^{q1Y^6(x;A{2}k*OwMk(v^w&`evJEB1gBOY
z>SwmYVNwClYgV1A&d1@&58(>RG}@+aXB`fL50$4r_p1RYLl2be_odH*qq4G8hwC&w
zEb=ht@y%SRMnH7?Eh3pf`lj#u8$b)x>C2D|U(?bKKa~Mh>oWyC%=52rP8;PKl-JqO
zc*J+>eu7L*_lD{&eCJ1cyojQqZtGqsAQBJbX-`S}me*i{{+nS+ryZL4XML=iS}6D1
zzh(wqT|-6b40~Cgj}~R&j$7#tyD0%Hq}z^DIy~DQi<;_t?hKCK_->Ord-)dU@4gS<
zNlAKVvzQK_Z1!ug#c(|xJDoRDXGNQdZfw%bE*k#}!qM%@RUmYi5aLdX-#YMHc&W<k
z0e1u_@IM$&pKV8QuCfoV=Dki7w+SA4-Cd3_55UW8YkS?Tx>1k0*|(8xRQN1ZCMvS*
z9|{P>Wr$q|<G@Jrt+IsU35Vq4wS4Nyh5%)8I1#mTbBqyk?h}%*uLpZ69YM8}ykJ8h
zx;L+}gqs08vFv$=J}y;+)OVUt^~_x<Z3Ky+^deBnMIaGOkkw~{;%bvtq@LXi_X4qH
zraBr4#j}o=fJ2pBSTOGDYWtTYrO}4{pEM#ipCy!hjLNhNAI;B&b3w!Pm|)S(;O*Lv
zQqtzRLZ~}GE+5)1aqaD8V?;{J-8qCe;58f1Vsq~HSDr-3=>4YLg-_DdqN$%o4!lb$
zK5qyZC4)a8h`Fu03F;{X;|WZ$;<HUH-Y_(7smXR!TP<y&K##+o%T;7t`v8%CrV#wQ
zE{4f<y%CDnp!FBzJ>T0}hsMT+hL&NeSw*d|=V$NHWPY!eR&w-6qCLg^3VtshBVtVq
zl;>T@hku&<q;Gx-{IpkV47zPx;`(I79`|VJPt8sjOY~Lms9@qq^n*DpG&Qg{#>o@g
z-vPC{igr%nxbq{ad{gzt=C*5!tYRm9b;vID$L$8PDXL7#V&3#)!%c?)Vk@7ln{yAI
zBwa^Z4%-#;7rn`>vEsAp$HgE@{(QnXU*pReY(zy5fuFPxzJQQ~>GmN-Bhz*p{mGUs
z9Z1%k1=2USuWoBmp%hU4EOGl&Aok^WJWaNjz$-WagvBf`kM9pn(&=VzUS~W`kb36@
z*PCW6tmL7AhFG}WzfBe@{yF3p1zQ1##r)H~clBxbEU&w8iU-lCsMXb6`i9P70C_dp
z<t{Ma3rwh#up_3{LDouRh+<L{Sg7v@4<L&<ik%EID>X)>gjwRnog{xXG|X4mOn1zu
zo$w>5O$x-4Q?qy_X=W#qmNZp#kb8{m$NyuadQpX~A%6eSU8f08F+cXyMFgH(VYOB8
z?_O&vohTUqOn4kxBa?mH>Q{;%m4aqF*8Y;SpV7G1?+*i)(<vDE>>tAxu0q3&IMmR&
ze%qb-F=#$h<{&rhO-tZDQ%jXklEq)Hj?6)VvULHR#0(PdDEV#od#9QVJ%}yuT-W6j
z)PoUKTr#T~O?5!`Ss#V~6}hdLaFOHfSuF-d2%QAF+)#7=Z@(xLCVIxC8RHh@6iKQu
zu};dyV!Y5(?;aN7BE@v!5KXT?H*PCmeSobLCHmvgGrz@$^Rf#I*QN7d?(-Dq4O`&U
zCprEg`I|W~u_1#sXWRLBXKRNl_XZ4%FW;+_r^QB#Cy)TMyB2?%ODLF*5$Q%r$Q!(-
zK%Vl}u*ku}j)av!(<--|Q=`-^QM9CBnG#?}x#E)kL2e+rbEGuMs>(#EBbj@H=|6sS
zC37*vg!u=vg6Hm*%a*{-_2e|(FowvBiQdZ(pnbcN#sw!X-X?#<WpTdR=qD@b1Cy56
zU0<YX9Oz+B;r7ADp2V$%L=tHWTOQb<$+BcIikOZMn?If+GiVr@WquSkS0l_6$@Q3x
zb3qYguKhG8!~;2IIlG>@ta@RNpHwuiC889ZG^;CDUyh~od`i;u2-+!cdJY{S(m~#$
zT-vXs@C^~Tq+t!j(~!KMy1*qqFm^=Wp=xbqoEiL4EeVgtfwV;rk;#D}?HkJQFW8_B
z=^vISviYm;D5O18dz*EK_csNUPO3ooX0Fj_K;#D9#fpiK(Zvm6nta0!Pe!LwGQ$#%
zvCo<5TLz)n{sMtN);M#DEj-&=+z)=(ZE2V~gadRP#)~kJj**f=J?C7ALbsO}-7yC{
z>6cgghH-QXxhOk*)ofLk`u6#)+H@mf1tXF;EQ33<9ko!rZl6Sxa&81V6+)$@x0Qs^
zP85oxompUr=9muQ{Ix|62@7!N7O0QNlVW+8H5DR@CG~LYZx|iC+A4h1HGN68z`k_l
z|LmPKSVu)cf#l;@y~e;hKxh=15Xnl?tFXu#$L;#^`Da*MxK;^sB-=oR(HB<}ZMHdv
zdn7l!xYI=ufxpdnr>!~}f9BNnA_(l7DGgE@*W5)A#en$G{gBVl6&<_yZTHhMUyKtl
zT$51Sks4i3G-XzlP%`a0E~t)A55u(v+PDCl56v`*@_-X}c#`>6f8QULgJ9yds2@${
zWMM)Wf#JxH3JwLIdhQ8-u0tf-z6uc9`5sR;AYYk2HTQl+Ap&X$=zp!BRTuCei0dgQ
z@)tOYQl!*7O>p}W$wZ~?Qt^TvXnwMdu5W=Pc1Hm%oALk?JSxRybnT!N{PYNrax7o$
zNGB1EZgKrA?3`Y>>q@^(fWYGGmx#1iNP4G`xDLAGKtbYBB~mRS=1{0-EZ=#H89?iL
zPDLE?$iF;`o)o<)brm8841ZyHo+WL>wo#XKwAG=|pRxSqxho{K^8^oEm1HG|%=naY
zQ;;k2uD*9C^OtAGIlYU2hOEUcM*+e9!YF+m-|B$uJF>qEP+Y!i0J3ky>T05F*!u9n
z^3wVNiEX?WBXp$&jLEs~XtiP!pH|$|^h2D3>-u*OVGubAlC_;x4+1se8O^5b1oJl|
zCkz44?U+SnuP$p1ZFWwo&QSF5Opo1^WYK3H3e36L(Z)y-glmhxQup9J$K>MN2o-sJ
zueE&V(xf4z**AJ*M$QY|bX4*lRN?0_LTE+VJ)b45ACI85-j<AE1Ttch9X6!JL>)?&
zh#veC%8#*Dev-^!i2Edsbi*I(X1M(iw)E^spBFF}_79fZg+m%aQ-Dk(8CSAA8}tem
z{RrxmNhJNY2?yr5DwV<Dk)mNn39|kyBqoSG1|nDDE@-)Qbjn{O7b8YD4U9JYguq47
z1*H4zK>*!FD7U<aR8IJmjbQL95W#hK-5czY_xx>Y@TVjON6*X@YsraAgiUcJ4c<Of
zlmw#K@K;frsh*QRm{akC1)%J_FQ_6o{>DG}KQ$s}7NagsEg&=|wV0n0jsld!Nn>qA
zhgeI?@xr?WGv#;qY%;j}SH)l;>kJT8-Oz&DI9=q1G*B7t^=gs$jCzL`2X1n~5DEsb
zMI3*VU$J-8A<9Tpbe`wl172iI$4Y>9*LyfrhqWnUMG7ig>S+J;>-bn`==3NpwE#Xx
zWaJ`{96e*r{Yvh@TJgQ@R~ye6V$xS!SHcx^{53hd)_9=C9EV3UmGG7&$M<DPBOg^*
zaV29X*dRLmffWKy<%hD<l`VuHjxbUrBRkLJw>>F{st@;UKZ)Y0ndd&^1z*sz5kj^5
z`()Z=pX>Cnn4{JW6DmG;1NBoNYHQrA)SDBn+YB0nd6&}=?hQf@OB{sfU{$MuWvR4w
zmNVT`|A<!pA*p)o!V%a3Ddt`{06mrN4l$t-R2R0m!WqGMlC1aLwJT-&qU-DqWtbp;
z2@4U$_itUb2`^I5XJ9;UF{+1W6JKd#6d}j9`Ls6Ir!kx25{AB4{VdEb)hDmJ6Osla
zn#Bc@T3LfT$x&%&1Tic8)9$#ZhRwuW=FQ%N0Sf-{>k^fJ7r2+6immOsRhZyeZmJQH
z_asqS8uQU>8K<`EQf?+Wi&5jqq{P%PJXTXjR$HGa;Tea;=DI*p0d&HNOSliU0IKQP
ze`i8PdAH_<lIZQ^v7(~Tr-XTYB%%bq`Vsyb&GD6kjFXzwsh-&PUIKWX04xqf?1_4`
zQ00c{j?JjKjf*Xh!X{x0(;m(*7`n=xzakOOT##E6tAmVx+fjRG(=s>0GJ5}otd~5<
zL*>Q)0k4YVSy7Jz>2rbmGX6qZtyTIO4TkV@7fX0l3Rr>G_c8%2%#9lyuH;PebK*mG
z<B|Yn;69PzG5~nM8sIpjcZc@%bl6d#C-_HUP(0F*iQ!mZM_cD8viGyHpSz*5k_W2D
z6J{;Yi%ifOfC^u7Sa!cB<_coQBCM1U7xD4b$v%O4sy8EaD3m9;6~{kY_Ppu~S`ujP
zNAad=x&7#l(FGv}Nt~p8HR5qSISA2e)CvvT&91ir0jBe(9S_^JRKALy%2y^@Of1hD
zrOr;u${*V|5BbGnc0|ZrC7~7e-p(h+eowV<gcgNOILs(m?;0)#+#J@Kn<a1Yh?HrA
zmEnAG8eDIHk12=xjox+k@~ewGNB=jdHAV^!9YHKio}01&o7kflEnq(8er9?l{9vX?
z<7oV4r8!bqN$xwAuGh<_bocHHa|;WXmDj<ZsNQlFWEg)!I+m-PeS>lGDalcRZNi$5
z7^)L3E=S5d>q9^J8<7~Jp!2>n_poez6pWD)LF&oo92KzaAZrv5`uH|Lc#ryKf!lu<
zDDn)9xw6=ufb>ovY=^-9NEV!spI`tGJ%|^(6p8!t87+@$_b&pp!4w$b$4~tjq5F#z
z9MRFa!VM)Y--)gh(-z#iUh?)RYa07#spZ8iVw^95v0-FOnoF*lC&QOd1xEvGREH>7
zw&Te4GggS|!)T&MG!z8e%U)!&^PE~6C;q!>ex3rKi7eHZZ;4osiW!3JY3R~vk81KF
z%S3FF1-_D;@ZOT5BaB3Qtn#;4c6W^pr-sfrTIK7jrffa$lFca?JP#J+2l6@s$O~x@
z)NG$X<0g3z3*7$U+FlE%<T(|Hq{o;C=%=j`xpm1~RPfmPFv8qNdnrC+BIyEJko~7y
zAIFRbs9$Mg#pDsdbIIb3mZDZ3QvfElOG`_`0eobe$djj$ixRp+rr+s<kW0GjVaYq!
zlbCa`=5f~Ew7eBmrff^<2shP+=u~)icH_=1C><KEW{Wqj(Zhm)fgLs-S39O=gCX4O
zqXmx5zz-I>b;D<vnrX=!WkGg5g|H_pcwm9!my4^_0{VESTYHC=`&?cZQB{2_9k%X`
z=(692u1P4sSlnqMNeb8)E6J*~?!lPzjSKuQ>G&sQVOT>;HAS@qj1?!!6s5()gL)Kq
zoJh@`r7}Cn(-~hbXqeJzMFvU~-7!taSxDwfJ5={K73ns1N{7SH2FGI`5>`l`v=D9L
z{~725b9E8~DJPVEC`YS=G{pU`D>Ot<yLK~hov#zxiN_VcIU>&<CM!t1vHFv<hihv7
zeEJ8Zeb%^5pw~D1%y(*GO8)yBFNj(hkFBC}t8x?Qevo`uGX`X3KRo~XX={~DK<U_{
zfeq`|l^JY>3Q$g$S2$EQ4(OpFiHVCiQ#7(V9u|1FcXxf|)C1Az{<>Z}T}CD8<<@wv
z@+dfPD)!2iWIkHClTH+vO1_v>UsXFJokN<YN|OwdFM{mUrsEV+E0&IPT`+AoI!7NA
z4h^Rsc~5y<fZBfnk_)LTYJSys?|rfqp5^5afSL%*6W)c+RkUL-FzH<QTcX0{1?w$P
zWrg-_-EE*!zV2)V3gDzY<_gZ!<UcRzDr<-MJUvCH#Sy7+){8ZJ4f?38F9xR`ei=>u
z3Kb(M+CQJ@0VY>+ZzI;2PxVB?vy}%kZl8k1^UN*(l1+Svmun&;>*MoVDys5356f5n
z7b_k9&F3?Bc@0awekhX3!g8zMx5-O)v<T#DN}?=aS_$uLLI|x&At^QW&&R@(ns(Sg
zmt)+Gq2Z?rc_IC)2CD6yoxxxDvJRFb4dTraOX8w78^nd%JGo_6Rq1!Qu)D?zwc|<T
zcg2?*_PA(d|L)w4Uul7A%O}Mkd6o>cC4uaaWyFse`LBJ<)|MWR%BXlA|1mjC{K$xE
zX#xHH4Qc8fN*q#?Gu6JItMUyBOJt}<e(AfK7&s&Gp499(GHYzLRpXODbGV8S>@QD&
zvY3;k$VGLL&)F-)k?qs6YqvoVk|1)bODNmlR*`qbba!5eMl>lD1--*|{RDRSn4Xb6
z>F<^*0S4>>Y7!2u`W~^P*MhOK7xT$lec-7nq=Yd;(r7k&A5;JmkNUH>4%AHvj4B|w
z6YrB`ora>yE}ijWx<&vIwQbBzXfbA*cSE!kTj9<~6p`ifpR19n|2l({UkiJEa9Xgi
zY}=`{scbWjmo#pQZHjg*PJ}_94gMk%p{scNl9bZRC@w5of*EhWOs;U}C<?lsMEacO
z2V$YSOa4BLBBs&%uH;`hPc;1p1I5xyK_Lp$6nc&lmQOf-MiSA}R8UPlaTsnS%_`0C
z@be!_K-i*zB3mt44;E!duk!yBlF$Plrp*~FsdYAx+Vz{ROw++N%n30FhT}#MbU74W
zZ~<nb*Zk;}Wk;;P@)+oAJ}{!P((7TN$qn!f9WtV#LYwGo^8hyU5GG7C3d{n0S9vrA
zn%r}zB<~}1EAC81QSFCS`j~?DUI}h+B>$}u4INYe5Or2KpsO+i>ibjtToX!c9j}x6
z#orEj=}|bIOqpz8ps>Rue!urBqiW3_ecC9hRf7CuQ@c@|g7TBh_T9e#)*%2xpHSa6
z_?i|-CJI!^%!(1-dHY-6>ZB-7-aC|(Kd%(RiFl*N<<;WNB5iP3{*tRy*Kihz6>5D-
zz`N}WZp;wHp%nduL%*FyB~6JDYC4uq)<?}h5My{ddSGs|Y;g0bx%5#)eC6?YPx3`s
zi3ex-FxB{l0+aTSq+Hw_8Ujcsh~(YI`z5sbq8~;$WSUhVhQX#|LCnKvH2;iOB#mS~
z8Jw~h=>JE|8=L}`<jWZ(_N}ZQ;{LLC{p>H)Y$?g*xfQ`_f5q6?>w*%~l{f!q7r(Tu
zs&RKKC{t!z_CNd`xpOHklP1UP{OB-X3aU@@!xT{Cu3-w}G6{Ik0M$&g2s!Ky#kA8_
zvA3jF0|%+>2T)4fparj-g@)eU=`IiVYICd1sg)bQi>vFW**roqhs#SEawSL9h^WOF
zNg9779M*rzxlTFH(gN$>cr65|qp?&Pnr(`Xwv^>_1NPLB;r)O8r2L?yo{)r<)7Ja&
z0)Za$?D?AWt!tS7zpVJYpRb>{c42sZMvY(e5y*UjB>~)$h+2WGlw|DXq}B8GGo2s?
ztx86GejV3?#wnZ?yP&Mka1(aK?hv(RWB%($al+Ge&mX++hUp!a85^_f>nSVqii9wH
zjz=h1-jlWNuPus3N|?BOUXI>^N<$GDbemAsbv?v}UEVb=BA~48eCqg<^cs}bi%C<f
zs|&VrlmJmX#8f++L}N!YAkMLY<`+m?fSa2;S)KbIF*$YU>2r=v%fD>9B3ed_+ux4c
z(*CE5)k@mza5IR7`U{y%$BWIpr_G#S8zZDUX&{H0;N>#Zeztpgp|@nDhamoXxRORg
z6hj>(E>45BrsK=xK}O(m#fe4M2?!`1F@F7gsG{?n!-Jo;l)pT3%9P9EW>aVt){ulX
z5##?o7d7aet`)EGAtV&NfVVoT)s9~nkOi*~>H~O>6HwW17;v5^F8h8l*Xiu}u~F_c
zUVVx32kGP8y^*bQAEaRW2?wC}?hg!Ap>Txtc`H2|nNU131(snjGB0qN6W4yXg%xOC
z^1d@ZAO#*ol6x*nP|RmC|K8Yc?pg*kgu?UBxx$E&EzzPR_3{YfLxdu{t}fY?)r>&d
zni_p>h!ZNgf)}bN8XXoYyi5DblvHQ>mjEdwNHjk-_piedDdb5s%pAN#-h4foP)NG?
ztx-&C$RS1XS#3>ixsN&T4)6Qs0ld8?&YVh_VIsh8!6^?UlPA}0)Gz3Q1W|trxV>$j
zl5tkFwi0-#!0!3o01B6<Lb<f_V?b3|?sbx&zEcNabZp~%7)~GehAlcpiY@WCq9H^L
zw#`d<qthL+3xso90E{3xr)NC(I)7ERjprW6o23O9>CG_S&l1iZkKLnv4W#@62PtQC
z$kZ&~d66`Z5A2FKW1>;t(M><R*f3c1G#TKnm47IEgvP@QH+#hHJPHw3o3<%2Ws63C
zF3GJ!jbId8bHDtb7T_e{5_WFBar589j$=n&8LzlA9Z<VGeaZZ}b!TSKUk4?~Is8j>
z6J>`OA2mDh@rfDZh(uKfKaMxFGS4yeJ#GJ8MZZ6VtOPXvafHewY&R4%11#4puW+R}
z0APu7augg|8fKupGc`aQ@z>~YXhPF{ZWw4b$|)>#ilm{vX%|o(Hr*~+8Fl+(ydPfa
znE;5)IDjT+@wZ|NUdr*mT!k$-!6_+)HJEd>VXXaJm#ex5w}E}@M93c%*xl_<=t;mr
zVHL{)IKf0zA%9q4JSK<YD>=XD(yKn|pVwb)oGlpO;pm>?hMcb)&*@a=vZ8|<B4nF<
z#3EE4pHk4ARhydaxdQd|;2CXrO*;3Zf2wJyQ54C?V^4tpCTWX^IqhZJJVgq<NMEt*
zBN;#JNC+{p=<pfhQt=-g%6m^FW+l<TjkCzYW>asc?d3Lm;@u5BCl3u<p0p)&)dt%J
z1^*=hjEh^5I)Fh;A_l`nlbF;~$`0Ui-M2R+wrd=5X^%&S6b-hx!T~v<tS9S6FHlbp
z%Po8i!Y{@WuFDsyRo;<<(EFgcYuLF}|G+s=>;t$w^SIyYn3)^x!y`vnWP^psqa9Oe
z+@dq%BzG2bzIZaFS5^`OjvlF`k+xy<krQE+9wS-0oDfnFO3*kxV+zLShJ+-{+>Yw+
zi3cb_un1z7>ozWO4o8(%AwFTz!b35N(4aO23JPv+Q2v%AMilTGEBN016G2$6|F8J;
z=0t19HC_2z`Sl+(L?fj|w1DYSX7)-lt5AOh23A4pvK<R3uTbGQ?K@S4g{7^yi(G^e
zMy>b7*+v;1z0$%Y>E1%6j@1sP{{hRi6CUkAjFu`ccf0)idU{SlHH<O9C4v3c$&v-h
zqO$A^vokg`ylNQiy37#R<l(J(gm{Vj2LCO>0M|n}M7JmdsfSDIBb<c7M-{IGqYz+t
z0qrJo?I(k(Bo#f@QDcxOOlz&Nf*B{E5-Fe&MJCtzSMq;My#-K}-xoDZN=monr8}jY
zONVragoq;DAR*l?-QC?FC?MU9fOLl<4GQ16zyJHb^Udh!3=G`oJm;Lf_S$Q&&Bi&9
z+(ll=rixv`F}U|*APvWsE*k0K9`~helmcI%4c+?)c}E;R1_(<}(#T4Y^dF5jaXd$b
zjpc63$@t0D>J438(Xecg<$<R5Is@Cj!}92KAYYjgHa+zK!;r0GOc)_l9)gUeZkW6b
zS#Em<ZJ<0>C9_uVXQC9Q3TA0JL_IM&%~$9QO<)NM`wy9e<Y!0&DxKg=&djtv9V775
z>U2~n1Qv&bqKTt0`)<46q22uWdFkYRB&$fyQ9xY!yszyQD!2IZcoKE6=OS0M0312m
zYoUM_+b2$+`JarqNpsS8FnLRu-l{`BNlr3Dd%puA7U_V^`lGn=PMO2|7fiOok*@jS
z;d02{H>W_Q#S;v{J^8W5oGlBLxo~G4#6eTsNlgC`h<x?p$_{Xm;puVV{XC=GR-Q$#
z`F8F@;;-Zjr={+)lH;LU+G^Cco{YU*V{MX~yF)K`kU~n7_XI7{TrGj9RYZ@&tYSbJ
z8LZL3vGxLx(hn1S{svgabRPHYaDXIYu4kPM2e~Z2YsmX1cL&jW^9fvK!wT2iP~8*>
z^H|7;N0a%s94jZTNcjJjq$M~_dfKhzKvrP(lGRl!(BzE2bf;0wCptq&yN3J5$d+WF
z^u&!Mw0__tZOjtJ#-QaqS0C3N2_R&B%YM-<qEo{b9Z#Qeb_D$iK)R^q8CExKpH9
znI+97lV(?Lo3iVSu&lX5tq`D&5I3S#l-IHTFHbjA%g4cm;;$*EvkQVVD;kOWIYQD`
zRc%vUx=E9;dv_PRq(&yX0%kOu3bv3(Bd#923tJXPbvNb;mPpH%B*YUCWSo6eb<jE^
zNf=~k;B$LNh5ViG8hFx2UK)1J+K*Yi1t`2yi=-9O_unXJ?t{BnCQjgvgPBBXRn<7e
z6&=(>mt+M0Q5W?*r~MSY)HfI>6Y$e;aHkJwa5pj1olO1TCYyBbQ9rhrlVA%s9e)gS
z^M8rC{n9%E32uBgV2cXw0FC8&FqWPy{NGj_i$#T8WrUPNMTx#I5lQsCm`UmJyr?A#
zQjl1J0B2{e+6F;6KI;7h^pKK~qyWBZYV2yeI_XJf5A<E~0<T?t8hB`g#QVFA4CkG_
zXI7B;AECN~#-hVRbH;LW_RjDG+WQ5i9zCCvP)iQ>`p42|>wP@JD(}CD=Z;xI7elfP
zP2*4utOYfW99W5y2ZOp+$?xLRYwE?u{!>n2eoi3aP2Vogoz)*Ar!8Jd8}23V<Up$X
zJ`VV4KKD1^{%wCC_J9Qy-4#>6P1!>ZrO($CA{vI`f`sm;3)Mjjvb4CX-Ya&ROQdQC
zf;z7Ta}X`@D`3)|VN4z&%r)o6f0sMJxgp)XeO&^c4eob>#(Uy7$m0JKuV(36?s4*r
zeT)5PClopQMHBV>A;Gqr^$hN}eB;J!ch_n!x43v8m!%~2gJgpi{aL(CAr#R>H19vN
zr2KFs`exJJ#r${YtDyYXt=#cB`2gy+gp6bOlj<aCg6@YY(3GJo&^Zc91(A>?V+KNF
zF!7?2+hlVe9Z5d1^<XZR-sAAYM<aFoN&F<y<63=vih+eUpM}&o(g5+Im~)%Hn3#4d
zR);>w7CNWb6r=MV&=J8pN5*hS%DM5s^vemrudDQNCrvoJh<|os9e53lAymi;z&xwM
zI5A<m`Z5(jE@4~w2ZHqex3SyRytAKR{`RH!o}S)zAO<IS@-G<TQ8WW0z5C$1&?Pgi
zu1$Nc(&ECo8eAFF^R)E%D`{_1M7ZxLM<`RM8%}T%#L{yyGM(;ldDvoHn8(;;q@n48
zV_DpUo!72PHxYOKX2L;3z81vTL%!1HV+yPJr+oDFGHB8emZHM&YYngH<fWo0n#dpb
zL{@L#j&_MgkV|(IzMERFu(1vvnv;d3?L+?+w3LgV93RVDYGv;JEoaJPebCp?F%6Eb
zq4Z)0*ZZ!aPfz)jI}btLAxY%2y$g{nt;Qvd=t0NfAO#xJTb;f$ckMB@<kPp{79kNp
zrBPrqv?duqJmj}-05g=>nPZI-6fO0D@C)0QF8+j2*r!a;5HyB;cYz#U&CQMN8_R+P
zCY1wqPr|*+d4e0Q+LAb>6nh%z7k2YQ$73I0qDQN<u2A<F+xe$S9hcG@b&N}8Pg*D_
z3u<=XBUW31x$-^}7d>qA=1$!}BRU)hHH?<POiis%1HNZ5zBhF7_&d-pHvBzFpXIV0
z(4G@$8h(?z7zHry6csuxV|>WOguPI1>I+FTfB$HaaA+)0U>B<M`7B@pk?6Pm3-gNw
z?@G)FWiWhW=NwQQ!fi2$8?IfO!HtE!`$Q9Rfbt7!(J;IHXi9n(XRHla6m)vJ8PXPE
zrZ}zZ<W-W=6pqW26{=)}Of3<`o&0X}us%)w72QR;#fN_eX?i9swb}PrQ_WXizNiuK
zVYr{K?*1;gjgAjfI_2BEa=@Ng`fbM-EUdAa8#+k*gv9%&B>I4x#JWJzpRt!Vi?yL=
zU8sw`FPb6Ae9Wq_7c(xErBK-YO=zeDPT4a@_<*D_u1~W}JT8eNP$rVGoqbM~XBIXP
zjV_5crOV|GH|4Q=tNiUPp#&R==CRFvSRjZ~={8ulLnkWle&ZG>WT&7R8_@?8XSGu<
zL{~SL&U)M`&!l?;wLH0-bAt1bU%vTM7?xFdy7OLRf<TNeebQ0^o{r#XaAGZ+XCEs9
zCtsbw&%cErj;V>X{j=f9GYqE4z`&0uYb&R>&sHZ)e0IW&c{&7VN`<1k#Ig4#veJK(
z#F7kr%~PrrNKh{_7DQ^VX=rS$Ywa8~lz*^ySQoO9U=i`z(_Z!f9LZp{S&!U^?C#F*
z<=JKYS+x;m=;mHHAUI;rSUc2Qzg)YzD>Ej-R(H*C#pd&cr<kshe9$`~FMS-QffZko
z-Q*NdPTVVh%fET*vA+G&JZ&s!paxmS<|Sxf1??G<0x|+qyiTXfB1Q%2xcjkJ#}?Al
z*#LRS2t&0jlg13J(k|2v3a~X%DTpT{imq+TfVH#>Y&*ls__c^kD2^Q~z|kV_Pt<rS
zt10=~o~n)DTiiYsIE4?6=$s`BNiGW2S_McU7%RprOy4Q-L_w*>LRdOP6-3kScb!$G
zB^|~ch9EGO*$GmATw)e*IaHr00C)OfO8q`>L+tjsAUZXRzXlIC3uSq5&Bfin24A`c
zb^X>8prbxLl-ASxK%r)#1*-h4=dwL=nqVWKDYnKB2A$>jHq;KC&rUmiMV7tG22A(J
z_Q5}>T@6z<D>60xWkMtKo=E&@E4hkmHWSsn%^nj(V4hRF8-4x5T3})1z6*&k3|2y<
zEf;0WvxvcRCFW+_mClWipcK)-m+{xmUhI7=Z)Z~(G_vlOvz`f!;_T4(3q@}aodhKT
zPmsDfO-}m9*7}~>I@vZviQCHz9%4akI3Q$kP6r?ZN*kk3`d%Obf6HcY|GZD1Be+dW
z!@Q5{!1x#MWV-ziO0v!pHMakj67%w-l(#-_b5545cNs4pP0QR%BcW)fT=hCmI+>Bf
z_&P3nPWjO4!XE7LY7r}gHBw7rZyTd;$;Je8$;#gfT0Q)u)vd6R<pU#ha@%A28H7vF
z8~J~zKPY_Eke+V2C%O!iX5z@er`0aME-x1gHilR@qrUT%9R1}`PF1$J{VgA(V&le3
zV}0rUyqvFEKGJD#9G<0BS<fTR$DfEokvr#2qKfYOX;`Nccz00>0bK{_w&#O2t-tTe
z!?u6Esxto;Zl3tN_+qke?(Zt@uaduA3ALbDi56or>BI}uo6q8bmM7c*!jj(?60ol6
z`<In^tq+Ls-YkCj%|Zm296xSP8V7h^2ynpQuo|LdOC-+^O7v&`{L2sVq#`gL&2#??
zJmVja+Q0VF^#`EuU((DB$L;E>tfDG;)Sh~A>U@wQY+yhxIW=|E3%pf^3((gZo|tAT
z!zm7T8{sAMMGckvpgJ3-L0RnjCO*-%?5B!){xnd!47U2Al&FSeb*VA5LtS$B1QLh4
zpD%u@XP^-b!}|hFwSap38gCK83^_a_#G?Xwk;wAeDpwA`d9ZC3HFW6<zfabRI0APF
z-O|#^CjYv+!8=?-Lyw~5<8UNe^kp!Z#Rx)%ChP4foV(7rqAZOWvf+aFx-15<Jy&#r
z=Sm<#%v(X^L7Xe3@YCumA?|(Ue@YzrKpuE`w!+zZES-Svuzy8~Q&i?V_oV#_f~&CJ
zQ`E*~&LqaDK>dDcEp}wtWj8svfXA#RtLEogY{jyR763C0FX#wV$M@r7Fq5P}&*%c^
z8O2=fBv+Kthsjq`(-yavK_?cA={6CEAszU8`Is?ZxqgF!;14Gm@p|C_@za?)*KOwT
zitm`drm?0)gzO!1gv+NKfWJZK1f!-sg;2Pf=sENbpaF9RtpEBufmY*Vg#7~CPP)MQ
z`b$j9Al(x_0IxH#iiAo`y70gHM_bMUE@LLtvjDF8#>W!*`C0_h=fXkPAtoo8Nn7pp
zsb?hh2=WH!<g7?!D)1wUq8jG&b514kI9oy+0i-x`)Fxsq9(ePzTS>7-`8FGZ25G!L
z!Y!@8l<xinHzA#Os24fV<`0zDo(dJ9()Zum?dlwe-~~$OPr=Kpr%-)+-NU}9Nzhb=
z@t%eAGv&|Ne(qg$h}7uw`$vKS4#N8I-~x;sQ`(Sx6!E{RSz=FohfJCi1(E6G%gNv&
z)AiFV7uPdr#_(c{nB}L<_k_Vh60_+2VQtYh5HNd`q-KQ*STW?kcI=eE@l3|d-*W?d
zkQ5&#fn&v7d%4&@&kr7+#2P34gsmX+Y!Lu-f`BL9)3r;MKdH?xT3)^XerVlE`gK1T
z$7ry87<LCook1ak811^rTt)rQ=@rz|)df{rh)3OI0MAu58<Ow2r6}dc;cUxNwT(&6
zOST>G&VCu!MFkH;qTI2?Bo-FFtb|+v1-6;*RBqjk))QZA)_k5($nFx%(3mr7ED4Fp
z6NFsRo-3nqb3>BN^SX+F99?_Q|L7QK%_A^T-(vNWoGQmcO1rc>e#=i3Frg`DG%}Ii
zKhm8P2cFU0=UHYFOapTpp-2U$EgV<yiN~N;pl{}9lk?q-i*YQh<GV>_S?E+C;LDG{
z+HfLvE9);y!_*S}EN0-m6Hd+tDYuos6^SJ~+%BxxC$%>+{UQtzpj#vZ47>aQV<xV&
zdc$NKoIYaS_pbI9|AUve$MEDF!JaRlXJk(Ua%Ye$TINtD(2}nAPMFlXAumgz<Djrp
zC7D{v0$~>$BDI-!<UCeeE~%=cxcXNBonF+g+Kv>5Zfu*$DG({<Q#pyN3h8~#nC~r*
z*omo*S3;I`I?a;B535LRt`+$Lz0H$D=l8XhjuUKNHSaUU`g8DI?Hjtn3qbR>>rI(U
z@xOc|RvLAqb+!w<FRjI<p*Jhtme|_!G3+ar?->(fgXzAkC)*SUOT4$;%pwhH={yu^
z_OGBhP{4tpXlQv-5vk-dE&`vt{C)lZY+)E!Kuzo&Pzw}Ka)7Nxzh+%{GQ1|-bcY@m
z%9`+Qg6@!!0a>@!KslE0SY}ivzpxO$-Al`gg0FZtUmE+D1F!|i^1T*XYI>JcC2J<H
zWoVt25D;8YS<_7hYmE~seiLdwkQ(hy#0(63R=2_~xiN#BNqkAZs$BeaukHLdu@9pv
zaww~_ao7qU$lM^CKYf$FNRTjC)`dmAKyyuv*qx%`&hG%8g$aC5F|lV2`CfQx+T9dl
zJ$Nbt@qQ}p9P#YgJO~Jsaq^)bf|!4(ql)5LGHGX{?>}pN1=My5t+8Y0S_Xj~1N=g$
z5Zt$CS)`hI8d!yi1qPaP)aS}VpR~mxZaGckb`rDK4KvaomKW$I!;UmIugXjb<<#Wc
zQxA-<uTNy`?bV_&BFJ^=W;8^y|7c3#KOQ8drY8OIfP6l#C=`ic5fXsq#=H><{PxJd
z!3=e~N3|mQW;pSRkV37$Mwt?%0VW$FBUyj)(n<TD{vL#;=`?k@j~^wAR1`H4<PS-@
zB4^C}>He8R<JRWw`Reit`HeWOFNN3~BS+5(Pj}u}9hl6G#eU%t1=bUdhUhSj-?!x`
zjswQNc?FmHdbMZ04m9V1D|K{IB3Gb;(Br3F{&w{y?Gb4Anm28AURHmk>{*QR5Lh&r
z$dSRSXy`}#@hZe*wv_ItO0F)oOn<_XOo93vg9JjJR}q5GRqHvj7Fs|@=Af=u+)z4i
zZ39%;<MIj=Xq<)u-+O+cJp8D6(5qM>9qDYLkxd8sR10e~++?`{j414>opf3sDB&j{
z9<T3jM6_q1UZNs9n5DBLjXdBbPfi+op1fW5NoEdaQ;UyD*=kK*o8CK_&=T5%5@65v
zPNnPP_>bn+T7xf!d+b2ALcsH+&AUn+6);wo9}Efc6(YP0y8&_7c?A11eHxglFvH{R
z@qp7i8jK&6HZ6}T^-5<qNvDsi*OVAL7}N{>vwFWLD8#EkBU@G9GI8LSc8vG1{%LgZ
zpK_j>N+y@P(|QK+S6@-=%6G|Uj(`HZ!zLv!bw<<YDj<bHEZ-Y7B2iTM>s4TkGe!d3
ze9OPD{LYZ^EK@FvMQx#arlu@sRE?e~uaLe?lKir+DpV17cYV9hi@Zw`LteXk9mOJ`
z1@QMJS^9Rk^kM>S1Tw8PU;Zy+4>FMw8FQ=$(B#LfCW^e#15dt*jf5`qEic*_5B*so
z=-Q=(v+68^Y1t`Z5QRUDqv&ZS<KDL0am;4(=H3DcVo1qy-}eC8GlX_g?!k~+p~%<h
z@hku+gS9Ba0Zg6XKs4ds3LFa?Q)7S)G3Bv|OP!CH-o_<`2O}ZOI;sx_Glw6zRk{&`
zP8?qu=C9i*<sT5-x0iK^yXT{ph=fbGGl#|GJSkAFFpvP8nW5J^3WP{gExuto_J-e*
zai4fRRX!;>h=hzchUJa^d9Wncq@&Zg6*H+Tyr4Kd=AW68xtsX+B`txthjSQEiW=#h
zdTPwB*+%kseVE92k!Kyh{LVzc4T0RpRRPb&<#P6C>E)N(N>ZiRpT->|U)>CkA5nme
z6blR_y+GteAw1^`FjL!`u*|x9ys3#qla&WoQgQn-x~#l3E;}~izs-N2^-O8@?>>q1
z0%vp6Lru$4vc4j002D|=rCtd?^%(=EX^RT$jWwbt+*igi9VwujWDCCCTxv;y(Y+d*
zTSL{-kp7hF=bC)^TyNFx{BanmrX8t2Nrq_Cau0C-Qdf(`lp!9Pkam+rgCv9#M7qkO
zWk*rw#2CxcksV@@;~9-8?->gDyN1O$Z$fFONd}*XR2E)ASgNMnkc60+{};H(p34_w
z+Lofd-lY#&0y!3%H5W1)(5rIR4q9Hze=wo-U6Ui*CAp_wiX@)6^)hb7fw>JQfN|7d
zqht12k^NYj|F4fvl@pXU`~cKNU8_Jin1Aus-U#!xSiqz$1osQ_B?eq-NT5{0Ug=OY
zVsfqxlThS|O0Kxw5cpS6tZN_T9-P3WFco(q-_@@3g3Ov%%K~JGuWl)k9i!bF<pkU-
z53Nm5&;QjsfZsZH3z6yaiY8e^3o&#V5ei1S3lBvQ^1NIX+1V&B=SV%BJ}%joEzrmu
zO5!~T|7Y#VPY7J18#Nv&D@p-ybw4{wHebt=Za!U)1J#kVL;-a@m(kcZ^B=DH3Q`ua
z=whjCaay8@r*E<8OEJXhWg9Jj6nA!CpZ@BLes@>sye5aZOOkPPK6>YAPlK1hbh-L|
z!k|AI@&7bDtB%%8PAX4=IfD4aWSu{=dtptEsuDo)X4xed68O#Jauyt~AqHor05LdI
zEaaf<q_$BQ^AbAD1#A6W%Q}1Z+av7M4B<DomWp?-^97i-={1oG%tKgd`cW)b_t1c6
zj|<!w|KxY#2q>dZDkw2-?g%R;yVx+@g=bN3N_#xU`wS>1Cj=5j7n#qU>~qr#$S)|?
z<ei+b+4?*0mZ`g6#Z;&K-qL=`nXqqaYGE9^*yfK*@}=56P-@EkQoD^Nu5G5}^6>Q|
zU*u}miPRu@`7u8k(;$veRAPIG0`=CE%b2I6^fXLt;{%Ld182vloQDUhT??DO8Bba+
zV$u;{b|Y_85C6mluN7<(23jilw<r2mKh>WUmw1wEK@Xg_@U@J~Lyy_xAnPC+&4fN^
zJEnRzCC3?_w4K`1C=>eA+sNAQ4^&W8DGVG3K5E=g&Og=lUuF#83%j2jC2tKYNcGAp
zI^|;~sq=a8V`b<O-7M{F<u}Zh5Dk?l9pvZAzqvF1P;MvZ?~giLq9*0=%g^(RxF;qy
z(A90^WuV_OO;>Iv(42kr%hmDU%rel7&Uab+QaZ|lA1f@6k$=@>@|l)c8!|0Q5U~|h
zt-55m_+i&lNxSDC`Qhq2rBhq`75dTz?&sJa#U<|VWfoeX2X4`fbZ~1o!_7;-%U1^H
z*y$X|#61yiy6*fPUtolexSbt4ndHFI1@gT5V;`_wKwIo09%JMMqaYz2ot+5+^i+?M
zl4|eLVc#Ve2u_1e=4|!#b09l0QQ)-76j+lGzP6PAReS}Y(j82j;*a#3-AL4i0XY~a
z4iPYnUquH~vtRy6{|SPP8{Y03+I84ew+r^NZhYno5^f+dRGRH8rVor2@jXf|s-Y0Z
z_P*Y_@WMsE8r=mkHl>n@lM`jxz>lqrIm+o!ea%Psx&YuT3X?U&8pgh$k%o32ZIiC;
z`!AH0HEg+@491kF%BNH3UYM13$u4D8O<cQEj;>-u=g|GRX<S)*)sXpTfq!1%Qc#5Z
ze0es{9k}SmG_|~c+c<EUb|uofE%9{4OZsT<G(4CjoT6<{|Cu*JddrDP6qY8Sf~<}H
z!T#Fm7>FRMsxE=63JC+{e+5#k;c5W-eo_)egJnwFOd7V(NU^JM>EeHiwlnG!VkbF$
z1}I-E7CP%&UvXyVvcE!>5t)vT!q?<tCO$rvy(iyj*fMi2fWn@zmjiHA>TU^eg<*{=
zhKmr{tb~b&WG2^*0QEh2LGnl;GZDHE%<wTKO`^+r@gDocA4#Z?RuV&df>Mmi$5L5K
zEAe;q{<BL(vP#FI5s4V62D9P3K~vE|i9MO@iX(nAy~YG>s3J13C^#8~od#ZQo;yWc
z`sgA?F&(TGwU?#M(&79o{PY}7Coor7L4{9+9oiihp%8n_R3P=j+hcW>CJtH+=XHlK
z=_mjiDotUtFIW%)ERhzEw!5I@4X9Tf=H!N`{6(cQZmb-Bj)TA82Q*o*LayWQ_%e#{
z)YO32(gZ{R0Y6E2+J=DL$Gy18%={bQtuboO4%YG6+J48+8vC0!AUo3x7^s=0h-4&#
zIT{{*&uF&YhqZp1!8_swAj9qT<HBd-kUstmQKapD`tobDB-37w2`_a0xQ|fJcp?=X
z)KF4fo;&N^^__0y-|zak8iH~Dk$BqZqHRDgtUna?`4#WE?l18J>s^M%-;AhQ0zY>4
ztVY$}MP%_rIMPlC3i*jDsFKVc+q}0=^$NyoPBoV7b0TSOI&JD_UvrYh;>x|Ng8byp
zT2<VR;gab~K;<RcTpMD(ocifM%;ox=!&rKb0bNW!wxIs~wpEeOx+!_nW<m1A$BQkq
zJgb4KqFn{1j^pkJMjv7z6?p_ZY>8kH_QLKru#uR;<|_Ai2|8%<*fkPY{;w7w__IU8
z6iS}J_ZtloYyB#dMmay&U2R<ia}Lwsb?5!E_aFikQrtFJ0V^sPf@@V`Cw1!@do5D!
zv9R5F8SsZpLoGW$lEt31W*fzj;(|rF2+RxiQS4vVT|Z*FZj-{!B%XBod_Y2ri{(&6
z-HH<N(&5Yx-^^=&f_oIDC`+0O$!NFhy*z{_e8C0AtdUr!0pB*K$p1{+s@pR7XivR;
zSqU&()UMTF7YB*8MZ&Pfv>03sk~-1``9)#gi#C5SPqX3|n4^Lg&^P|wKh`S>mWwYI
zLAEd#>v@e2ZkwN11J*srnTUBcQtW)9FNT+CnOjbp`~?Tylrg}Jq(AJW7B0CbQJRiu
zmAUOk2%v?yAHDkKZ;yXs52#dGWDQTUe|3rSPiXRRN9lC@8<qOJow&#K(F?S79d~^6
z%Xw%3n_m&!eAsLMCC08+`Ceaucp*fcizFN7#st1R)=z8s9YO=GH79us?>R;g?P>`3
zaoW7KHa*3uW7x<0gOR8c7iN$J?<D`LA$jvu@UR0IotY&i6;@+L1yC{)O69!i5GrY?
ztZ{wVh+KijoPd4Yv8eMlrqZCgi~G>9%oa`r6sqO9UUvcC?u2>>nRmc4id^Rxrwu9i
zKw`<F%N=y?DOw{vG8|3S2a?C~JFkz(U{kw{)gQ0HokV(cIk9@4WVoyFzPNl#6bW1G
z(DX7Z*FagC{ukh4S8=$86YOm|?|Ht_DjjYgW+MJ!AP9>-9AljBEos1fg|=Ww|1@||
z8JdLcfwK}C_!SYQyOhLTM95p*@fBn?L#$^(s7MHDW?Hz7JuEBhRMK>KY6qf9dzeb4
z>MrN%`5VumnYEkY>GJ)AUYdlT@W7CxgrgXfUq;dzR7m1lb+7d;V9M`4I!nMqn0^5~
z#Er@Sc!;3jFD!7Pw=mmU66;<bzvWSR)Q)x{=~dcbedb^oKg;^_IY6Xfz6ceubQb8K
zdb@R3g607yIQF;cK~M!uX)FxK(*mBl@xWhFLpmm0Zr7DBq`>o39Wvf^q$%V;T&KWP
z;FVGF2VEy6OV(g{dBF>Fp0Eh1*p(pi^&T8VYCk)~W2+IxU#Y`u3(Va>plER_&zfn#
z??45XmG06J+B;Ul&I{k1P0?hZT5r9vq+sS+(J)%z5Xc8$OlpGO7{9Mgg9lO)8R!HE
zlE8_l%Ai8`_4OUfOyoSyd*|V?DnXHGLl>Co(2o2z*B6W5?gxR#3)TNcR07gdHj4oi
zntm*}%lxo}&1rA0Q7%lxR4>ouOu2qs0yzYM)89TQNGLvnf_Q=V5BEilLmUW)HAQzi
z#Ru4N*{tD6Y5(OA0&nKxie-9L8sZ(?ahXih1lP)wnX<C+%FA*98|gG5RXTbHt)YkQ
z4s(*E+zcAT$#WEhko?zEcc%oQ_2;LfwRO5ve<KN5wayBRfcL~@n2d8y2`sp+;G1!E
z04F2I`FXw3yY6o49M3JZK|6mR=hy#sQ=f}5(!oMV{YL3uu#Hp<Ej_TlyR7oHo_qx!
zo+pPYe@oWXomQ}H5QsvuDr=ZR7q{yd23eJB=<wJFMhJ_tMBaY5bvIs|@GKDdv>GnC
zqa8^csK%*$<$8IoA4rG+E98OA12pN~m>c|Xm!adi|2_poS@~e{^29eP$fL=hd)jZi
zETnYY-=}Hb^Oping~Zyq@4r?}hH9szFnN!ECoAh=q_cnI|0ZjWqZIXiLw=A!qfo4C
zPaHtsT}_AKW-GPadN~gSdY7j*Cl)FEA#v1}tq@+n327K_0j#Xhkb~C2f|<b8!2j-+
zz_}F)D23;PM`fsFB*5f%cHotv5fKrwu4Oa(OJ35?jPa`8SVrgbXQ$S3*vARgu&UCx
zA{huZ*QE)@$tB6(gZ)gfcTKnQ!+b$mPOVbe99vO7{|}Iqc<vKuPEg>fZMC}(yl92Q
z8_k$dTdp>WEFqsLns&Wi7eC>Bma{(J(n{UR`r7W|p^=;-Fa&va<Kp~58|+PTQTJ?h
z^_g@T_$=&=&b`6sfo2QYS$}u-GON^G`6LLsFHRk{fh|F5ECfswYHeny+7(GPWhvbu
z>-aOmx@AwBufPryHe#AD;fK6M4>gC5zo?<fZuROr9Un@1Mf?A6uTT2I3T~pe2RCEG
z%}MC_3V?^&WZF^xw{m2}J{@K(XOB!Z-vYYjkb=G?-%zm+u)f8WJ6O_|mqGUmn^`t+
z9g>gh_i^Z2h6FRI2JJ;NQ;FTD)hwF%7E>>C8(p9DO1<D9yDG6W{fm)2ws9bhS;(9I
zBo~7Oo{;1g5gzY^5+RT6q<mE&T%BuZk?a38!H<uRlg?Vh)?HCRNm+Ti+g4ARiNw{7
zD)-Q#rO4HVKiC(Xf;2wK2=<8H9(orldYMS6=5b)&7AqY~dBkBV*bWW8m(gS*J9;dH
zkD%YC2aXhWtLT!Kg=%g<{!gws{D)8q!%iMT7?ua~v>qU#?6l?rrfTp3gF}4wIml=M
z4v9{LlE1k6C1D=}$PZ~?Q6zX!wH*{ZWcT%(VRS4!8?vNRpg+j4R}NYe!S|7J>|E2}
zG7@0*{kkwWbz}Z(swV-@j2nv%l_l8x{k2Oh3<phw{(A@yBVI^{3+v^<Y1pQe)FZ*p
z?z1UieAVghG29_XJE$ov1KQzaO%ncbpB0FG5filRYn{e9f`<5$QJVctYO~Pgngdr;
zyfYtYiXr3t6iq5FJ8Nz6klv$OPRTQMGy5bK18RA<r$N}8b;uZXx5@ZBp{G6~j{Pp9
zcPzlzpQVkoGh=(99PYg?hgnDbV9mS1`4V+H6_06p&?aKn-~k&d0qCnJG%8*Xpewoz
zemx2xvppZ?wjSRr2P9A*1r}F_d^SKnhiSP*pcLENo3O%2aFnXG&LkwQ5Rf5JD{4<7
z90;Xg1gto!=$C&4el=B|Eo_}{Y=c}$-CUswnJ)qsk5xTSB9J0p?l3iAhywX~9kqIN
z>w~+6-?g#prR%7h8b=<vKT*!p<>YU}V<K`L(RH0z(p|%fEHS;kSS9*k3IMp=MzD9V
zIw|}-H;GNHeE#bCf|j#w)-fQEvWPYEn0*UA{u6*>Kb#7f3BF)#07ta=Y$)t}^r;BS
zJ<gx-$)QeTZD`C+^d$F8$n*jICkVPqRNh!*5m4T9J`!t9nJu|GiD1_eqf%#tgmBQ0
zvSRpZ3oRrh(@cx|$jeWG3qr&A*>8H~OQ+?pJU_Sh5+Sp+n+niE=ZxJ8#T(rDr1Ewo
zt?$E|g}wb>d`Bq$gx%9l`fG(|&?bb5yNR@pGS_#Oz99afANBpI3!PZZrMg+15_C4?
z`XBiSe%RLhd-$(s*!!L7{3Pt193p|3h4tC!m(>5m3Mwr1Lpvre_w`N=ihB*1Uq$vx
zxmHIVws}+-@}pi@(yO)MmR@E_5*(?@@6r|=DxjNOcPJDo6`Lto<@suNu6Ud{t(rMV
zkR#EdC&`+F5z~dkm}qj8eL*NUj~H*djnk8CRwfomoWg>0d2{cuC}&3Em6@j9?*D{V
zo<DU+#E;hg@_u!;ZEXfSDx~rN%!q&BGU9OnG4+4jD0P^Ls9f-6$o6Wi${rS6SbDCz
z+0WlHFmllSpkPct9N+#s^O~KE*T+z;MfDX~$mbP58u+cXaW|l&=T<j!{rcw6-QPsA
zz@_w~it8hv#B%<W<38a>64lL5skq%Ydrt%5BB1Bu4;ilZ-h~!+>O+4}e_qeozpp<s
zj>l6T+Lr$t%rSrqY5>?C8634Pz}z)!_$TTAwt))6Hh#J6UXx4KN<ux*G!B=26n;-4
zw<Lldu?KQqv1BHhzyX^m{F!Ea!k~WU(p7QjyJ2zzpmpDETT?uU)YIe1I)WrbI-|B9
zrvCJOVE+`(#6jk1FS(^r@sA~2ZoW7P@-_jJv9<C!&a(=PsMj{fSLD)_*k++&IZq5q
zUxa3GFVh2&HGk7Mg*g-G+?H<UowVP6!(E5YS3w55WGW^a4jUc)zJ(PVMchtBd)=PZ
zkJe?l5UDC0ba*g2ig~y(x<h|qxx@R+6uw2rc$Dzjho_QFe>`9RpvdqgUR}B@KYRuk
zSxOsvd3kcMEGGI3sy^xL?W`ct|5Nt-7DHG4oBI%&W-^1$p%{_^>ZeQb;gmc790A@@
ztH#U^(j)eNK1z)3XQ$s}w~i-fd-z|7zG`}P-bNDouI2ZS`B!z!BV?68Mz!${Zr$k3
zop)w|Rm-Ip?A9bH@tQ8s%}u0BsNElJ`kHL?FHadPIjD0W#Ps9h@sTNIQQv%EoPC9{
z2~By!6uG4Xov!}q3SViIdQEv;^(uc<%f#6gNm^$2{eH;3l8Xn=_Q@|qI>GNy*{_GL
z5Vk6<?7exiuf3g}Q(7kN4|;0f7Bg%vbXz;QTYiz$zdG9;dLItW521aVul&)726I|S
zY2SpKw7lJ*hM(wpX2f-Zlc~P~qn_K2cPtV^gJ#C619=W@(tAbe#RW9!q5}SjH=@!0
zz>0G9E_bfoK}JA)S72M1&zbVa;J}^KEAaduJWr3z^xw4826(`y!Hc53?$JqkLVQh6
zDm*Kq`O<Bvgo)8`sR>pIz{G!zRHaX+ID={yl_>uL4jX~QTJXsa(G5sz2YyYP<h`Vm
zOSLD+M9GlcR)jjW{&5n_`fb&$pDUf(+EI~7Tv+y|$hP!4Z#1qbEBLzw(Nzn9cF>ee
z!9<vz+67Bpk(G5@Fl4KqHrM<#-6U<sp{q@tBoW6$Vdu_b4x@mh@G6euEI!R;Fm(8U
znbGsZt2hm)n?^Rlip_8bTFs-*O?kepF9pn`zwb6M&Q|Zf0S#pe@#}b(`;|!SyEw6_
z8GX1(n@r9!_PIt3EtHW_ySBTO%zEEreoV5DG>;E!9$e|;qMyNoss!N=VY5oSpj`MU
z`_#NIrlaF;lVJ}4NfYL(Nrw*ONt@A~_WsWX@sOH~qnYaP%f}b8>|kWZRqE=VM$Wi>
z*to7~wU6kHFk<%K!~&)BV1-qu)}?87w=73pQug-H7kVk%&b*}<avg8t2EB(qXR{M`
zj_`=J$lIS?c8oO1BRjv{ulAQKd(0T<L&EZiZYR{h&9P*C_Tn)vu~VR7>l|h1d5(mE
z05_%<=kFiQUK@`a0}e@C&w-x{t8NW}j@JEMfo$9US*KS6NN3ZYF;1<2C_n8p;0Z9O
zsNk^4dxTxWWcA{0uCw4RoujOpY->EN$^y~g77dD#>=X0k`+C$-h=b*C?YDUtavJ%Y
zpI5(o;d#0$Q$)Uia%>LGFe)`rn>Ew^u=-xcp5&+ubufOhj5vyp72(Y^5kK7<wmK!&
zW?dp!BK)BuQa?#2*~P(OAoiwT9HckW!xe;P#AoMA5*|!F9p_x1kkMo{aLr#{r>~OU
z$8-uvM5Oj}seI|PUHrONm?93ZvMog<E2s9>>h}o@62?eVN6F6mvdKF*W<RTw(wP@~
zzRC!W>42miTL4+#OH(^XkC;*;3@?`;Mx@%Bh`KboT7>vkVMJ20^3F~{S~+Ek2FMS$
zL!UJJ#NeUEUt?&PYqv#1wA^8TOnDRwC_ON8!)CBMVii9`av8*LH$@1_$&FGuyUCx@
zrOCvfF@4AQM@f9g`!QMydY3cM=ky=f=JGYQnYG#|s=4_X%e3>Jm|vE3WoAF%o#@b2
zmc0SA>D{RA8Eju|<J4IsR8(|F{+Cc_*#@!^rFIG_9g<`)0W~+{w;{ur2T@PQ)_qQ0
z?fuAM>fdcq%x5XWbgKt<zIYwzwzPO2O!Kl-6{RMwtf{ub788#l9F=rRH!SbkB-IEY
zSa#Zs#j_q0xVwr>9l`}7!p)JjR|nDO2^%!A(|;>V8rNa$#=u+8Q^2g+-R>PbuzY9h
zIF1$Ze1zdcxlf2}FimJQ%E+mEsydXJB8hOSy)Dtjqjizog6LaNnska#2ldXlru-lY
z^N&N<HskSxD4_;&$XL8_io8}ax`0cC_U1zEuZ}PiyTl==MIs^I2le`{-&_MV84P>r
z6P-8XdaKo-!<<>#)yA`Mg8RSNiaE=Jir%mlR;+vnJ7iojYFD+Au>YuF^zJ#L!y`On
zrN#ptyOKG^!}HJ9)=XV+BR8Z}@-6Y=hv>8HuL_Nh4)()Ww;(iI@uPxDtvRA}H4e5Z
zCepZ@Qx@20m2|^&YzZ@VA<~jFD+E^+YUZj|y!|+(JlY-wLjCgs6qZz#BJvPo1M%V?
z%22ZJ|0ml5zaRV=kG5teQcSc;pkp>oEE4>269kOjK}9OIjzl<vp(@nItQFztXSEjn
zF^$97p&?I4r4XJLQ?V!1>fPK!suN`1<a>LrQD%$j7+HHX?N5<5^>A)dELnY`Lu34D
zd<^*syaN1rRj&{=kpj9X<*zh1F20Mgg9}@hpU;9#Uu6_(Cl_}2zJ9KwMdha53e%gI
zPY{~6KI70e=w4+Rt<`_y)vS}^Vf035nPt?U8*7MVRHnM&V}wqMPs`;UjDRI!;6lAc
zneo^gi$Qk0D8wHyd6q{NfCX0&{f16a82<I?%CGL86jL5QtG_bss*9Geh*zFZ_4ifo
zuvGhaEOjNao#9}wn#HH1;5DpzhnLD6tkkqvIw9Xbd{S+WN{&@!7Y%$KAxJP7w%V95
z|J(qRLVE`HO+zGKZIgboi-bo3HQ{C=sceNz(G%-x7-&Y{eyP~IC8Jy?A%Pe{Gp!JL
z7OdCm8?dy&RYbI3(0@p9m4H00Z1NZ!J)`CMbomUL#-^^#*h`0z7-+%ut02PDarA@W
z2d@LGm;u|q9H;iY!fW3{%Mbc(3fQS{!Y`6pTAYLRgEAZkNMA9%*}e5ztF2q67`OHl
z7yT_{l{6ogc-6Ol|1+#?s!q-}&fE|$v-JbUa);f6$;&FE3-}_0rEj)+PyH;TVsIZ3
zHl$@x=ugjiMyh9L;i3^7M_5KF>>{^59A3Z}7_Ml4+QTz<1YE^L?*vpRCzdU?jj$J7
zLz7E29Q4JXgTJUcI@rJ8UaW$>Q@t4QczW~i+9wY?#5|1%YOaA6o2}sSyv=102^t)R
zF5pU(q+J*#cz@;^Cj#APfihhGMjv+$Gwx2HSATDxGFq_TM&mvmmASW3efe_8<p+)f
zx^zh+xetdA6)C+CaLY0Kd_=qGYUNavN^2fo6@}_7D^Db6c74ZKbK*^-w)>{@N8QTl
zQjQ-<5MD>$mNhbi@x+_#z$7FIMd^n$&%b_0I9q{*2?V+7%q<Dxl&_ppkAf~xu)aR7
zdt_<rysS4r3xDGI-TUpX$8LqptA=8XlWhP%3T@$=sU;s=LR|f=!Qm{nOZ|VZlbU>O
zc0`-vVr&yzNS{jRpJFOF=@h-uzG^k=w+1fOJcj6`pdjcYQm;Bv@xu4pT1?OK%X+E)
zY67qWZUcq=7g=V(-?^FnI)42(Le;9q#t%(RO<<7Se{r2OA|gUC`?K3j36r?kAWGGn
z1nZKgm}_;_EO<%m4P|W|u`}IkGPn!)DNdaxgh#o=qMsEN{JDCsUVyh`Xp7w+{C#tS
zxQHYajjdN&hNV5R(6U!Ae{gL}p{-6N1*>!IQ_h+&U!^D1Ed+@WPj@K$#dmbbjVqO@
z&^TddIJGhjUgQBrHWKw($XfFHdi5Om3AFjBhV7ZO-rEWW&%B3D`#E0{^o5VCc0b6j
zx)&5syw63-L0KHDn(OYI;<X&~+4AniRP2{8=+Ly+M@>3Bu}tJX!`psZ<=)eU$_7Iq
zY<wGMx|u^ZG(Y%=mX*Xq?LStOhmCWMli8gL>*AAC7`3`1afL!rU|~krfD-F<IRn8z
zkD2}m2gD^lyQ4XaN;V&3oL3ow=Ys}~Z+{_7p?pPz(`WndDj29mvl4K<K|IL+Qrw99
z^{S#`&x@t3;wUbIaOMUs)OK<Dii+Q_H4}LoUS%G|f|KLj;l(#Dl>jM;zq#v$7?H|K
zA?nb<ZvAJZ?2)Lxxid2}nvy&bU^b~6J7GgVO@1hwKT~#=8@ggOzHwuMmhMRdiJ{P~
zSO2iaQkFZMneA3gW!9|j;cZLo;zppZ$@I&bY(dIgcR1Qs71w57okqC(tA10btk+mh
z*#hX|b3NZfTgVNekoR#_2^?ojmBs750U=;=OC!?^4Mv+>^V#sbt`d?GWS5@3=aw?h
zk9D1MA2a9_ss8+!|CI(GQ1#klnP8W}Ojy9~NZW1UWSJH264}vUG=l@&aiQi236j^J
zT^dso&Gs|Dh&dkH+_k<hd7QdV@8Cv}bg&b#L;Kq54&o)4aJ=Z!p$<)v8kV2k7#FKF
z-TR%eEtu#$)BNgP{P)24cpQCpqSW|)NmTw;xocdGtvrMC_~tmZ$uv3|f|!Vtpc^a>
z=#v{hm-mW`j`dl}ka+mQR?F1M&*mG#)y8*`f+Y_Eg6E><LeSHL1;v{uzGWS~jzWS~
z=%ssm?`*#Fbf_nn^>`ts0a3yVOq(29&S-OLxHsBC2#~Gb#G<LTiN@&TA$O^T;Q-(N
zaQYC^A5O!6O<~Or+B)NZb`f-Rc<B}I<y_#WYBxKVL^X7-qW8o<8PhL~ljjDRr@;*p
zk5YfZY_HZzMo+bAdbJPYKb@NzPKx45;29g;j>@S3rSgJUG}I-$ZYw=8T?!W;UZ?kW
zaQ+hE8ZrM{p=WltKA(^M$wl$)cJ<R<_xf9G*Bu@3Z@!ORGRX-ckTJQaNsBHw{6HuC
z<hGfk<hHhOh+X6B?J1a&4qtjiVE_9^T9DHZjZnMP?vT|62n|8HiZxxXBC%W2OHG-6
zX4u{}#B91={dN4`Gt0ZyGj^u<&Qh_w7>mi%N#TU&qJS&jg;TTPxOlYGO18Y3dp18i
zJh<iiA+ljG%=y1aeoOj^jp&fuC3fylhZVb>jtrnkNJ>H^7bLAnq1<6z2G0)PD3B+<
zF$k9i3&0c?TWw85uu793Qsq1JoZJ9MbzKNWnQzgd!D`^;rr5dNrg+@qgjgG@bz1Nc
z)UC_^&Lw=!me78S1NFU5IY))hV;QMM?nI}{eom^B(R<m|f?wFRU)MU<@fYKFXYr~Z
z?aw1>^yU3cg^HEd*hpZn%`JGL1`JeTt~L1=CdwYFZO0y**fjSxe@d^a;ls1v&jQd4
zMcD#Z{h+orzbG3={?PKv-Z43r{8toDE1`@3UI00jV(&;<JlqwU!UESNEp>h=19-8~
zrD?JsRCg%w^cC{*J*BLMXFX0m^$aTMya*%19nh~Mwc}(jf0xrrGeMN|DlIv`+LGFu
zyiruoNbxm1;dF}fg7#+y8aqW{?J-0NA7+@Oc847@DC*N4THt)=;aSLEw2SumZx&&5
zz(S2P1w%Br<oh&KFJElg##J?tBfW02o?FF9;r|6kQhGY8+>~pZ%ga-GhDc5R^<yO)
z@e08kC$VNl`>UMV{IPQH_rPu!21Xgc18p0|t9#?zzh$rKlv6(bPM!(A^p5h@f)x5D
zn&U{cz0x)!o*oO0jR28J#|636=-^;K;+p%sjlTgUg^(O?kYDHfbKL!HAI`w(;!O%_
zWgb5k(P-RFiUh;>PnD34sO`sUZ*!<3%K}kRc*COHHNk28gW#H>LYM{a8{ttCrzr0|
zI+~YuA$26qz0Ahrq|=t{r0h|v+Z5baVMnfzp1Rwt<5A9?b<N4oALyxu%8lslAcf}j
zj${0x=Lpp%$9R}=u(fc8BSs{#|GewRFS~pwGnrE$5tpIg)Fm!&@p;!4Q(9xWC<4DW
zo8Rro{PZ?g*MPz0csS$xv}h$;Z9TJ+oymb$!PgR|+PQTF7@i9~XGHLlmH$MK0@o&)
zSj4EK<6B*VQAQr{^Z9w*e^YJPaqp#lU(fTIH{;t@IO584tTBa71=(8b3RBrNM3PAL
z9}H2ZiCE!1tNu9k(w4TblUftq$>2pbtCHJpHM{VqW|W^*ZTwv#4N9*t@|lzxHVS@o
z=<7PFH_7X&P{n3N&dewE3*n`^6P41MHE-{41vv^+wAJs=Lv&}FvR$V9*N`+ABV;i6
z^jw;zALpvH$Q#Rmlg$JSFYrA$JwpH_gwDPxC7BTRspYoej4QW_U0(AsZ+DFmz+V14
z5^y;SY#!ZpBY&7O3yQps8Qc94xnw&1cY5um#NytinEPi)G2tx{yom%yAHnfu0Mq^7
z`9Jjk+1Z;LQkTF{{_%k$uzPx=nD{f9j(I`r8vwu0^FOkbX0ssv_S<&*4BKECjCb_1
z(m3yJQ$nbdU#EOOzrO5)(BC+?=4gV>Yh40zX(iK+p;-fyWX)<L;l*azuzIo1EqYtL
zPeU^b9y+ZrPvJBh!1{#$pi77t!I>kvQl44e_O+`Mhp<9AQmXP@qslm(PUGiFvWH#`
zkKydYQm+!MsUP#EvmJzl5DhkP&apEKvgkDA8s}(O>&0{p{1TwFGH)%b8bk+JMn&Md
zn4>dUZ^KkI(2s4^bq&O^ze=9gsk_MUqZM|xbyhcwIdsEdD5e%Wysbe|#SxlfA_Kgq
zj*n5;&s8sl#3DDigH08I`jXdq0&oW<VoMs!5w5=Y*gz!(qF+wgl%u&~kr7U@2!YdT
zt%-n>%GvsXd5~rU!c54Ek%xy^J@Y^$#)(L-v%#OYZ2t1awwkB(1rrHI43PwenQrE`
z*#B1xFd%&RH!m2SQ*x(a!sf8Z%;)%%_?G#nErNC<g=uaBj{46U|7I$5SVj4+e|lH*
zLfD!PQHZF)Nozo_0Cuit##IR<CB=EwCVIB#wgBK<GjdiHp#WO$3D9z*x&|e!3X~`~
zijer0Lu`&G9#=zueb1jIOc%#qTmmH5sVfA<Ls~SM1y#)>&O~rL4<E2!;{@3GiGKg@
z-sj_nKh-zZnO~$=EXYgXZLXhsRIqcApZU3$QB`r2sk-6<J5%o{r&3wbn6%a`h!Wml
z8T^U=t`REp?I_BcmS0e_&3ATE=;Tf(n8qv^;&G61lStXyH|%2K*1351Cf9%KsOVcw
z8q3_IsLWGYtyOl>AoC3SXUYY(qDxCM#`a>%BkW--X>BOUqz%5*6XVf|cunki3`q(7
zpX<(FPY=^*+clR?Yv=&bF#;ew!!r5;ZfwC+;p$nsqy&g&-=A~E!-xKhd71o=@at9R
z@q`XrWb?=$p7DkVww<A^BY!yYdghVZeSLxd5>{0e>dt)pvWk0101aIC0%aLJ_KSai
znt~BD-|xb<VCXdU&HMc=>lijW1<8D>Y`6R}KqdhCMhlhODuEXO5!Z0L7~BMtexXR*
zxoNzR%HGa`SB}w@91Y?H<Jsz|jH&3#G^pzwgzti_<Sbl`E}usdx*ey;3NJ!0U4g!4
z;G{r$DH}`gsf^PS@_>3ww=I5^$(F!0ANTI#f4kL)lJsfn8P!7dJw_}*FqR)8rEf}0
zt(oO&em+)4@Ea-cP4}mtJ*OkU{et-*|DMTq?3a$JkXZ!WZxkMtdA1E<M9;JVhtIpe
zsvAz`zgxI$=J=e^=mpfo#2zcz+fz#S^~#DR&(t$!VWF6MUFobk_QgPoVR>!nJ5+=p
z4GQu)^XqWYASomNDw~15X~Ql;{FSq!g7U+C&VTHOzto>8#LANSsLYpm7MC3gbu_jV
zcveH7GiMJaZ=QbcMr!N?<G*t4q{Kd{JV_5(olEc$T>EOHtxsIWfUir)kZXs+;*BJ`
zjnFRE&Ab$aHy?4_`pRP$8~Zqx^dF;S2!4`#H|-X20;|wk@hUd5k9vVFTZ5v)z2lB{
znhF@uA%2P%bm^OcvF5BK0W&p8>)7b)`0Tf|UQkgX-)f;UYyNpxq1x&57WOhhNRGDd
zdm~FrJ}0%>{L`e+e=aX$O1Nj=dR8uQP?(w5#nUF7=vA^|gE&&dNoWWXs*Ddke7lbp
z=UXg;xqkQWvX9>9rlDpQ_R8`WyBcd6U-dlIX7ZbF6)9Rg`Lm`(T~fWOf1wY%x*p#V
zStP_>tA8<EF%(UtTT}hS22W>Wly%+mLtjUSxaso0mLIt9y0Jk@!UZb2345Q3X!+Sz
zkH@BJ5!NY;V}sg<4@i@OGrdhuM+$ArV^NckBaaB)85&@rj}Q=~5JtEN)BPbAcdk(S
zXT@GLhzrprD)l@W%>N@=Kg=Z~wcE<7M?m>i#600rEL@x3C^`nID$R{EYA&)daB{fe
z!*owto#V;aas+1*>wuf(WTi>5>u7n#Ua#kKa4_&m_`YXEELOuS=3LMBpna`lpbR*d
zs%{Hypbz~Q?qYL62)F)&pg6IO*Y5t(pwW7g5@o{;?i-wUe3%8CK(34t`!f@!87GMa
zA*TZ8rH1C7syF|Q4SI)@@marLn+EACB^(vvQjOTg+Sgea7&-tfMBQcwcJQ&F8Qa^~
zH`W_xYri5$#LM5YIyUriLBn+aT1!EaOjEi2aOV}hUWE0~TM!+dFxQg(?$;PilRcBx
z_%6hVt`rsW^<p`m^q|wdOYlkR+B)G!9+;+7>LjJ2rxALey2W23R{!sPwCavHaknw-
z2Dqi@vbOm6t>m9y!veN7Yk!@YsmwAD=i)qiW<l3>`QN5;t@(FarQL5ozws_JD$4F2
zW*H?0k3Mj#z%Q$KnNn&%(OheriwIx2tf@#{VEM}RWT}2;88<b?W?~j5N8+PlBYJ76
zt&{zhJey9@4MXl52tjdW6_Grd6yJ2C>i)jYc;-whU!e9@mF-+-hx$Bh>z<8yiG8L_
zeNtCpXF;`g*1px!_Fk(0ne1y<E9@#JLG9u$%p@GI_Ye{Wb6@Q(DY}__ekKOon0l=b
zw*N89P5jJ16WG$Hlwzezb<a7R#(a&t`xy#&b~;{pY!04Br=_U0B!VUz{}<8x3DXcP
z(!MNK*fLlV2YmAT^Fm1+PHe+c|M{C3?Dsi{TXTp2I`^`SrXrG*&R=`L-SBJss16rQ
zN(`Q@b()ABfN43|K1oSQF?yzcr0!veM2)0Rp7-zKb~)O2@cwqa|6Iof*|a%`r}@P9
z23%<>aJXEoY>J6TD^N2XovRV=8ki|-`6D)GVq9@yQxgFV4UNK4*4ID{{MXU!mj=0E
ztnG;x$p2#7*M%&<c>kr1)S26T5HYEbjE)x(%(PbAQDlOcmyMlv`K>ZDkv|<cqS&9f
zlL~Y8yz1wM&>W2X=8%pythW!u@jc*et*QgK9rv(t^f4y2<xAM~N{-BlgKp7k-PBc3
zW2=)_&^9Uem_exh2V)PxFu4hthb6&o#OU>DxMN)qVmYY=Yzfl#{3_Wd^V(!pm&&z+
z^yf&wp1mJ&ckisBF?f#?^^D)qNOhpPbrsTbFNuXU;4&$uXQ|p`O2+>oZ3)~@BqcOx
zMT_Sgm4h90c`1JO@3q?hKepZiD2|3{1B8%-5Zv9}-QC?;+}+)6gF6HW1Shz=dvFNu
z1P|^6ceo+%_gCFj-4;b{ZSC&t^z^hm{d6~3e94vJm|u6ycB<qEPi3j~9u9*?)e%Z0
zze4gr8VpBjYW5^S-%DQUVNEDD&{-h&EQ$f_VUngulseEvsJ)QmWHD713q)v9%=6@J
z2&mN4t=7?%JxC`PXW(=l@TMk5-jTDal*|b)9*zwU&aq5vP=O=O+}aUuCOH?em02eZ
z&W(Ifds1_q1%;LeM}G}>Z6$(MXvpQew()r|(p{<1NOF<~3~gcq$gaHh&BoSAC9rAg
z&?<Wyv1+sG=}*{<ebj4X5vW7@X#k-MX&3;2Q>LnD6OiUOS$+@?))i!PlWELg1E||*
zL10xhU|%-gr=5N50-I?h<YPiS)ny&FBK!oe4&Z+#()Qu<l}pH%R0BO13oHRt4J*r`
zYdkHqgJf~nQ8odITPuDO>xY;emeb0GX#0J>7PF&`XwmZFLuIDV$#PW69c_tGUq?l#
zaup*UtC;-)H^rV$a&@|E+eos1Ju`2<^VMnA_%=l!)yc53#e?DZxy%#$`vW^wPTN8!
zE1W9-_UBkSUmRFxw}nLvIk{dX0~-qIgi(k!2&Qht1vy*$rRZwvNb95><}|k{c?frH
zvwLv}!6mdEInp>+=r6qW&$9KeP-|F+?s}rrb4ng26EOecjBBAE&oYu-R22a5_5l6#
zLon;>>i7=UT0$B944gwk1w^%0*(>4|Z|W}Cvy)7^A1FN8Pb&kc{vAY%648PSkQ7)k
zw<<=#!JwXQMKjFqbCI(`eN}Ufu+gBNSiRP^cS6n;?dkL4cc^uzZ~4KSl^Fr*)fe@a
z=16c+_#tZxy`6+c#6X8*C#p@{)&7a~X1sqgyb~+clUac0ivB_DRxAULzVrFQ_Cq*A
zWUw>~u>hUAFReMo&JR0Hx>(b~KlKS=<WQ6*9`nS}3V0;%pyFtUgNEt;YzAM7We*nM
ze^FJLQm?$@>8zcxEDy&HjzXP4tx2*hXT$K*#W&szE{Nw`aQ@vNGDJ&*f+0TFk~Om>
zQWM2f*doG!hdYO^OFAHCjFa)8&Dvq&=(FQsQz)bf=#z1h*{ZA7bff-Ni-@>>0ET2S
zPoC9WAytEq=pgTv$N{YO>72S2LLIeCjgJ-Jx1o`$@C%_9=YQTV@2A_zLN)!HHzx;$
zBxS6edNT<wXtbLn81Ga9HFUpYpAW#Zix)vOUbk|X&7W<=kxr<bT4A1Sb^weLwCIf>
zxpW$&VeQQi9>s)w_R>;w&Yhn?MCcpvp2VGh;p(fWFv&nQ;Y(<ib0Ct;J{}C7iG~7l
z<>C*mEjldryhuYa5tyV7`)kysN=oxnZv)MJ&w45rU(!$+^Y3O;hq-8@JlA^S68@|H
z2A-#Vm`Sm#^ikmeE^6X6<v)}F6SfSFb+Ttpwp3WzVrYT2TBz`T*EUkeL{W>l+cp+c
z1I!cwd`rEf<ygIt4x`=7d_jXYtJM*?9@Rw=J8Hy7Wr$IGn}s4Z(CXw*J63o}i=cFN
z{<*Hjy!r~B!!hx6dF2}|vwIH+C<rm@WH&}HDxRt8S8gpy_Ss`ivwQV1@$7L<_&JP!
zx(gph{be_Kv<N)AHXqNnKJNmNkwA6;3&<;ptYml)FiUVCP?b8!Eupk@?}B5Djy`j@
zaerz#^_^jhyA)#KM24u&o85;EF4vD;z6kyy{K(V?_^f6@?=0T6%A&3Jb;s=L@d|DX
z`-f16ZOT+6ozyF&O9h=3JgpTf(d5hoP0g~WCPgN|Xw`o9QThK)-L#&428g7(Kp^&t
zSXJhc>N7<fiqEDV-@_V?aKPG4EDgKCVJDvSji}drI}_M9MT+SH9qG(7G$Y#u+uA2z
z>9X~&)Z;<PE)}(M_BSub<Y1r-^a?->pKjZ&W{PChwO-10vlREB({pW0m#g0_cc_)j
zyMbCM&|KzKHL0`;FU0M;X2hTOXcrJavFvzBrBV3veSyT1fUxvhmQ0$YiyNweb9=3%
zFm598H_&<8L?%5^TKY?v{4KG_CZbJKn71^-&L7+-#<T^{l)l=)piRxV%Pp9#YgfIC
zR9CJHg#Xovy<$N5CL!;2paw)v83&06@B-NTEezDV$mo!}lJa583Qv_IjD5*yl(R?h
zcZO&ff6^m&gbZy{yCdVPwv(;aN=<lpa1Szn1x^Ad_N}%xGcPM<WH&E&jczcHTUt7e
zV~=b{Fkd-ly}Fc8O)X;q#;KH7P6jjt3Ia@uNCyOX`W!Yl#}n)hbf#_#inKC_+F;<W
zo9AYeJiVk6T3rTU7-UOKubsKRNm{le?o3k%0`a#xbEwf%_LnHS%F;vXvAzd#>4YKq
zWlae1w6E?ABeUW(h4eJSSVOl62sE+NT}<69Y{RMrn#;A`HLf6&X5-MNYk+1dMt(wi
zD{qdyM`8EJx^Jc3MS>wftW2d2&b1-&-VAw(Bc>MUE_-Bqei$;uMl8=_5W^%L$8r4;
ziWV7;zv&VGD-HrTidNqcF%$-BUXUi3SPxI9@r5~Fl4K!PcKw>moq3nmpbD*SD1C(6
z#Fo<ijWDcQb;YLE*v?^72#8DG6R7qb#5nP;%vT~>JDI1=v?f_;57^Yy5PY1(|Ebor
zri7O#X%~4R&VS7!h8kA+H$l;!oRCk_DDq7LDHLQIeI%r%r3p;6@(KBa>;XFyOFAYP
zV_*j`M3q#6m_y>23TW<Bu%jv&Qro-~(G2{-v7=G&3$QUA2p2f|LPOeQaP^9wH9e84
zQbyS$@m&36%q7rs!iQALJVY11ZI`+Kp_y2f#}}Sg;x_xRZ*P586IRhf(zk21MnxVV
zg_%HQ24p0|>KC-RYqRk{$XBiT&YOC6n)$_Pb-z~^YeCMg!SFREva}S9^&k&tzG7r(
z>C?|?2<<i&`CxfvDJmQwft-LCoz%VkZo_akM85^iAztG4WQwT7mn+-wFl=1XK+*Y2
zyS%1gJ9NB|s_W9?KiDMQyhcO?bZ#??vP33rZO3$dM%umtv0cW4^Apv7QqN$&AnDu;
z#<#C}y)5k?@T^Mj=9URi!aeLI0Ov0Eb;H(BHg`$LPu7E~Cn=+}pL05=T^MH7uQEn&
zoLO#*)pbjCZ+15){&G@)Blb*p1I1zg9+^<Whz`|2zU5<onA;Kj(v3Pqg<dMfKp_?G
zif;E?VBe>v0U#p7X+A9<e5>B>#zL=|WBnpk&^6o_4XZ@50}`MnrUXY+N9a1CamBgw
z*^df5U2Sl14S-}rakqHAB?Q_$S_%11OE1(Qkc0hSWB`%AgF#!W7&Vuyo)ZC6rjOY<
z#igoZ`Z{G6Hi<Gv#_nYZ=X9WT=y>!5Dzh>j$kA}*pNTcI!?88E7ePV8syIrw1q?A#
z3l%0rt$IN-^PP&IaYx1LX9O75MWeI@5+#1%jY8a-{`;b7&<wQ~Ms6hghrqSAUN~Mm
zGAi$JQ%U%<JC&qkX8|gSh^V8HtDx=J?~uC<b*7nxcI<sZ4}#d8%wSXnv?bZEM1xyW
zQ!sFLoDd%hLYSmcTE&_qtL8JEXc*g!@qVYKz)>D6;Br_T_n+(;kkcWQrmNoaxxa0w
zo+%E-#kuU48EQ6X(O)*Tp<q0k*Vm3t6#~@cEvN%-(O~lfAK`@jmo%%_%rr08A32PM
zzx?(+B8&uk=owLd_7hYdl-l(<5$>`Ane|N-;1qrei6~#xxPD7cCLtqxFP({+7GfyY
zMdK9-2(Q#7(>!|2q#h^t`FX)V5O(2pweDJJpjIY~aL?g<sg5uTA>s<$kAA>2MtamF
z?xSbs&p1(fcQc}G(D2S-T@zcd*<F;JwKD2B#!EvEgE`HT&lJ0%bRXg(gF{Y1>r3wf
z;A90~m$x#!(;ePU5gxWBmNe=|t4QV1_#8<|$jtn5s}N|ePa-RgsMo&+aVN&6kf^EM
zTZ?t>1LBeTD|7e1MhKwZf(J>D(ZOm_DP4zic)K6$^vw3!^2|=ysxY?t^RSEWu4(5P
z8cz2RTJHw^x$~}8lj)(h(p3-ydQP1u+FMF(TK9Wpckqf+z>;Km$sKQi2du`x&r>y{
zCSHhPG8LZ2G$;c)726%I;(XFyrs!7sR!#+)V%Hz$WUnRPD*k)FZ$Om(f(>2g(E5^O
zQpR|cu(i6E?o!-4IH(1~+*bN~^)<WcMk&9(_OEld#PWt=0UIcbTit|7O<36W{^nS1
zr44)9byTEUZ)S44SjBjoxgDuP@UHvcK4(HSa=d)|-vy@6{DKz6#`^pSxsuaQ68Hxa
z4E^Odgn$N!y$q(!i(z|eBoQLeub6bYI(r!@(+e=0a}nZ-z$|DcxgUn?>63LjVdi_G
z41GAk@C$xQ7M!k~t<|K$Vh)4*p_MkSyuLnvG4a|SOnkw27DE`=AI!vu>d=4mDTfNW
z69g0tf`yUFhZrWrvxl<oSS~5WxPp)@VgrV4djh7nKGpr&$8;xX+&SyUmL!N>mSv(V
zCUsT{7jQ=5kt8#AJH@zapy<Z4m5`C<k;yc}LUR%*r=X3*C*Mb2`+!^9lqYJ*+ozqU
zv`r=4OpaC&n5S$=D>9+3_Z5i%v8Xo!Val%h6!FXm$cHqdb1M#@r)aoP%!lbNhQ#s7
z(5Hxvt~u6$8vs&9+3hG$LV}qM^FVZFfJKIV=McYlDTyw8+q}{Kf69`E3`zQ?EcSf>
zza(aXpx(#2=o<QD20{SpFF<!{D=VQ;;;M)cKegZCn_6=`bv21rgdHe_4sGDV8^@^x
zrv9E2N)aUlm5-gq&d$MW+vzQ0FH2izLbUZp6g$|L49R63ou_FiXt;}pWw>S0GC~8;
zkQ&dms#pKHyEDH=wbpUfobF7^uWxw#K8_NymmP3KAHI^aqRj>9ka$N;^d<zZ8Xs>b
zpFTrHRX6|GR1VagnUbTMzld?^wSOB8miCQAbPMK+U0Kv3+NLnq`qvz60-195>2<Zc
zo&AyeKevxnh>gr-JO!NL)H|YKBBUh9J4Sv>)o>`F0OKTxM0E@`i76C~_$<$Id}fn>
z9&gDB60FAju4a%1xXg2%3)#qErP*#2n(30?MK<Jt$J{Pl8}^mC{4BnD&64>Ki-Zb-
z$3Q$tzT~gB_jW)lCKh;f%KORi8aZ?JopEgGgk&hJl9o2sN<E8ro;9fzz%~Lah;Owl
zaQo|olq#xIHKNc!O&J!$=F;akz{m_e7nFl^Rcy67Fqbd4EGs=cDJ^}Ogd=SBu(X0E
z{DpAss3*$$ze+c(UA{TsHBufZXxMZ$$MOLll#KnHwdmRWD%5G#S2JYcWN}me{J2cr
zDWIUeMI-5D0(ak#3AA=$6?P9Xsrg1A#sEGSd+u8qV&yx-`9tPPy=;rOA;<<qIr~yF
zpMipM{P^<*R2JvkN0dM1r-+v}>OpLfFO4(L2bY$Xb`2~MN0neo5hapQ%Yo*M6+_6R
zaqJzTzatpxn7Ar{=2%<v2&EH_&0UDBVgbF3Mt%U6q*$vSVSxvYIfySG7;8M&X;!9>
zK4*+igh07xs9*Foxi5d!tdENwyx5j^<ifYapZ`pLqvGpfi_R0$AXzLvU`39bTELAp
z?@h!aQ|DMq1@&Rv7_Tm~C#IrOs6`DSYB=)ITQ96Yay~LeDgycZPGHC4eR>c8y>V)i
z)j3jQV&FxxQ*!=^%@^Fz6mH!&sahN#q8pKPmlj~KocmCdqYU^zcP$us&S*cN9GsGn
zO1z2Y7GXnqQRHMx5^CjcN_#J)uoYV>Z^}lO`~?pCC=QS1>fmsxJQaY*mApx#i^wa0
zwag_JL%Bmz_Vz)+MgQ<%n)gfAAz9VEaF1@nbCsZUd#pJiWf720@GruYfsmWTz{*Oo
z<dO_fes%{IQ8JCT{BbSi(wChQ=_(YJ!&7moE7g^~_bTWQM}R?nq}D3ZeHY=Q<KBj^
zGlNLNz=eZvtYflQLb|_Di97AgN?d|+dgy`<xch*_PW^*y)&BUEEvhJn!9&#acx`7K
zgt>s}na|@<$C2r6u>*&*F+((C{q{Rdzm9jL1mYu`H?W4r4XFxQf@N5ecghFdn7?mo
zO5^DNzCmupI1x<&NhCDzk5aIPQo{AsJS9re3xts9z@UgM|FF@7e#-KK;B;++%O<=-
zmY#~>QvDW$DvrI6yb8%p{$WV%Vq?^m1c;Q$aU!@KQ6u@rxcz<;jEovWMi{+I$-&LG
zz8WX|ugG|3qKe4J@{%(uNf|w}ARFJH#ma8XsnfRvEK_+1)18PHk5`b3W<>rB*aEEG
z6&PUNWv71(%r^tKEU3Rw9e?Shrx^eeFq^qd0vr6#h9x7&`!4(PSm}h>rN8c3{eK)e
z)Mo%m)1#m;6KJx__l*&ISenA9B}y5!Q_P!+#{#4mZT?rG18G-8`=iKp(aMEl^T^e_
z?;bq(`xnodxd@-uc~J0_i~`k&5g@}kwBn0)7NkzSH=4hTG|qq&?D!grFCv@`*<&MC
zZ(Ziq$47YE9v7z&Y26j*Tr<-%Zo~^Vo5XIqCG*U~pKEA~nvF45bH{p7dinNUURHtG
zxjwf(0_(7Ex1R9ig6!X3c2!v&AjDCPd+IciEJzVYd_HmD>j2I`E)R+6jiCzp=SVJr
z3`Kd3HG}2jRm68PtC~juGKf8Qv`;0IUWSBwS1kwoO-7)~3}OLdzX<cgUrrFH4#!j|
zi@0l9^%~E|Nt%>n42ZBUh>kE&{CG4)#AO(^Ty3x%_7})f%<_5B>e)w<D(5cq<kXKD
z1Rv3B#ZYCV?J}vy++f6mMNz{uLNPP_Fe0C#t6<$tB-}%U3Sh-2$J-`?1oog^5^dao
zeZc34Z^VpSHj3@jO6u~Ir_2Y<H^AWY_Bp3r%r+)~xM(%8xDC>bL_pX>OcFwLNMWCf
z@>v3Ci^fGgYbOyjw8`tsGoQx4)9zKMJ0f&ucP#s^Em}5w*~dojAD>U}Ug<<D>fb^L
z04T8QrDdGE5O*?Vwn^l4-$?y|433Rx`!u>-({+S|rxIq1-Bjz_A0-;Svb_aUEwLRQ
zt}Wep>0-nY62U;a@GEX?)Zc?;X7NtrJ_G+a1rxV{Mp&<3Hb1bp;x<dfMxhDKE~Htt
zdxk-JE&{-u9BVJl2r|Uld&ero&9BJr<U?=qsGQJG<9I^gZdaivIx>KWhWM)V#}}HC
z?K>=CacSE0BB>G(@k1=_R|vKJ9%Y;Wf_OHv?cL~=O!*4&QIak8ETG14?>(ApWnbf6
zh`q#;N+;D~WYoYM^+pvA2a3o_AW9Y)7xB^D*eMvPaRGl0^Y}axbLEA_SQ7L2)jzKN
zbS%AU)aJ@L9q&1~XpAwatv)3OmiE!LmiAWDFh}K<|EW%+uS^mD-b5JO+fqVeJ~f9Q
z_S3pCGDW;%2Mf7QDm(E$&)hI_nNRMPX^-=!9mmJ`(P<RO3sWd0Kp>RJ0ojDB)UJxu
zlwE))SOK99(Hy=C^)FbP(KV_m&H}hgntlR&rKPQqKR<%834t(uhV3hrk*kLUb6D&G
z5y6z#uUqxlSd-qKet(s8_oRzr-$;rK*GN11zCr-R3p1`Wn#w|Nu9jz{6RgF?Cs`&M
z#fvsjIPPJjY!c!4K&(0}VbQ{$jBYDrY|h8?zNx>8yzShuXDlWC0sR<(Hx{$N#N>}a
zpH`3ZHdV;Bn$xIOpe8lSGJe?9>Jma`;PR%L7KM5C!)M}z51I7w``9bkEwR32-!uC<
zYvJ%1`9lO7+uoViedrKLM6ymkKS^cZH?)emu~;$^I=WQaR&unvHCjw{dXGY8JG>ax
zZRZ`~RA`#%91}NSn*p6>;&3@u`4+JpB2=ZOKm?aBFwD~h%do-bkwYN-Mk@$+05<jl
zdNKxy5tX71066z%u?v-TTzCpuVt|Mxf`g-=)#Y7U3ev%Iu5n$>E#FCHNet+W_FLb~
zhHmWv!@hlFGsHT^X$94{<0H?RFFlbR@{()l2lKV87Ih=E-}Tk~HS?)sO@+9ZrQ#2T
z3h7aGig$IC|Bnk`mzwnJwF(&N?{QTf=8kBx7HjR#$)=kdH^C)2(Bk4TwrAJ*mY)KB
zH!C^tqX>ri^2&{WE8hGXR$)u`OK(c6P-pk8r4Fu-6O1CznkPppg26y$>{=Zh$=20(
z30rZxv%iR0obk44CBL3L3SW=P+I;d3Wz@NVn`IoopFa1#*y{fdaW!2Y*=hc78G}*E
z?&m0RIlUC9A5D#!jlm%oWPnz&k!*jCq0<GJ4J0w2;cpk4eYkmUKG1Md)W(&OX|0)Z
zp&4n-;^mwd@P3nVG0oturTf3HFD2C(extl*bF{6Fy9_hEh<}|=U-dP3JCRVU%4%Gm
z!ujG-nU|noMB=x&h1KRQY>9zsktWr!bohn_^QR|Wmd#+qtJS`HKh?9N`4ftoSV(Z5
z5H*I|*Q2S`B1B^1TzF@hOJs6Mq9IdZe=lT(?V@hFRAp;giD356CFInfn8S#BP(o(R
zD5q!(5pT@S#J)A42hw$OsV2k{xiJ08C{W6z`IWg7UN4rns2r~3!X*;aqokrA%%@XX
zticgR)Pob_9QF<xkE#(fax1^seu#cTvY8xC*qk;N4Nsz(`<<RiYmas9PqNfW0TzOT
z4_aU8g}Fe!A5VIhg%ad8E3mij$a?Xx_F_Hjk}m=KWMaBMP~@8xCrY}P<7bOELD+kT
zXa&&<`DxeZ-y8!umymx5P$-O$RCV}GcTfL<b5>;(Jg-lteptV4&zIoKKL7M^EDe9W
z+^=ZMYZb{apu>}eN}nJ__t)Sof(egkEBfhq<SEbv-&p+P%A6Si5Ci|NkBG+TiDUe>
zI5G)72~HL;gWIYqhY&(!#yrRW$9pYo=`sG-LOuIsrX&)JV>>OHPT?J-vr#?&Gt)<M
z*RXz-b2r5;so_xwGIYJorHGozmI^EA$s&0acXK>pTJCq|nv*3UzF70@p3ZW#zEJU;
z7`z}>na*Z&rM4`tEWgjV3MV4IQg8ssAQDTD=p&QJ%ysn+Eg93OpcWb&ixYC|^1YWP
zRjGK1d1VQbT!p*wo5fG!l^2bExM3T8s?lb$W>QlZB$<pzt_5x?$RfMnRGiCBc4f$m
z$!*ZYg@#wrhmq>tQ$tOdxgaI9mq|Y?g<Z2vH7g7mJF|VfDGaTh50?0NLHT#@l`&Vh
z5SX2$#>PsWtDw$Ve=^C<h=Rd>4M;R<TnJRXceIHV3K={A^yvxF1k`#$q){n51ptEV
zdY~{1ecEw`Zu|3%j`3##zboeY$e5CASD>&PgUh+u7lFFJSVJ0$fc-A<_sVn;FpLXG
zKgb@YjZ&6~q)|H3@T>G{OC{Gk?GR2^hFyl3<m!UO0s&*wt8vdO-kye7m9cwk3<-1u
zcXpw2E2M`HCOY?u0ECBVJ?;-tur#_Ss(4u+^2|40-Tk8n6T3=OB)^#ocP_CQo3G$a
z;ZjKVyd`m+mTS?sOV6xArw50RsE#i`As1gCV-rhDDdt_~{3m|GLhrpc+S!=O;&S#j
ziyA907kEU}z3hfo0yK1EETH5|TEA)o!6zH11JMapjrS7<LPj3dO!Zp>UuvgG(oQ{x
zlhet}^z^%e5rUdu&IL0Z1p{$1=p43-Uq-6hr`^17%C+tWRXf`_-ZEcphW`&6z0M3-
z6>|V8rPRQJakq&G(9F)S>@X4U@#%6%GSmEU^fs3tu04ly!`~wYs_@w3qimr>enOEb
zw8)mQcB&lY(FQio_^JZAAdETo0m28nG3NrGlNanCToT6#v4FIVw9y0;1Hw2AL67u1
z7<QK*lE7f2!_Qi1*jkSx+89y5J5oLd5UQuNGC}{;SaK_)Gs>~^1jedaReoJ&uBDi&
zkl9~cCkDvOA#eq*a>?QVJx2{NXf>{%E{{Lwk?UZXaeb3{lP=P~9vJwylKS@D{KtOr
z>D3q(xS1c#xEOT?+J(wM`A$MYmm`q!ZETYjk|(Hp76$;lA%4W@PX8V=M}4!~V;Ih&
z5~o9jAC=m19;UwPPP7ctr3qSpENt$ldywsVoNSsC$E*8<l&lWKBrV{Nci)KtEI%y=
z*AHy6vaUPNEXQibv#wdlaQwRkw&U>{U1Qo`?+2DR(C0`8%oX*%I3K*%e{*&4?*{3N
z#{ur!v^V_zy8pUE*QMB7yda1FCj~e{xKT87l(ncy_`l~V1BESp|H$wug~28m%yRU1
zS7m2TXD6KMEk9HjD)%SZVK5P~LwN^QVqqVzt=(rH#xriu_k+V~?jF&buOG^uen`sw
z_`BzX`gdxD`)*>;V58f|Vej!D@$y<oQz;zD-#1`oc5dL|Z6N-?ze0*P{yid|ptqXs
zk;Y5#LUwil=6XsHnD0TUF0J$`niWgnkY?p#P}zSHw@~1~V@iCn!Hrhm(*)t={x<B$
z&z=zXwHPVNoL>;VY1I`!m!wDg_}PUAgy&#*>pCkTGf&M00`<)i-#P^sxp!Pk3Mw+p
z$68nbT_u^uI#JI<lS+8M6heT4rbKM07E2UIh5|H4HXl*@`EA5eBssJ?@d{4p(4^Fp
zzc*v-Hld`e5?So`Y_ULgf@hP7!GY0rWHPO4NonEaj!%Ob<l%{dUkj2Kyq+ke9C22r
zwTS-=Om=eV`lA>L3j?dJMWfl@`SrBsxmnWU4Ji)%Rma>^lK#B(p+8TIzE||wUf5EJ
z2`IlB8Qd_#$sY&hc~MT$P@$M7{@zKLIC^SpIejaTf7(Bbn!ped7-0Qpzu&}oj{aXn
zqhbSbzT}1x){Ff&k<06Wj_-07)p*_9V`IODtz7F7cFog?LLrMUqKFVT>{NQyX{&)W
zy}jyMSgQrTOr!FrAf<fKBlaDXF-}r|-vsKsIP4l2)1l*)P|m;A>|vAw`A-07<}c_)
zVFhDOv9G?wfjD64?(!G4rb_j)+v)s}0EzkdX5ebl()B#y*1csSx|c$FG39?bo_I8|
zJwW#dg*?VrEtQ=a)^xgH4K45247Gd`H%3@_Fb|6CC+DR_xH|n^SJQfxQZ`C?uS2WH
zN`5W7g7h199rGA-u+T4QP|l^*a5kFeHg4Dc_!qTtWf?7iRsRd?-u|QRr`#d1xEb>G
zo6k0K+-xRDq>Dz@fb3m_sGB0{JB{sA$IOGnLc-}OXz+p_GAZdI*0LVs%SKncTn-E9
z!5d``-@EPW{uG*%_%p{$Jl{TP)!$F~IFMEW{Es!r+F4@A!XR?BeAg8uJFF0rz*y;a
zM~4L+@YSW=SgkA-iF3W52Xp9MtD(o3i1L>lSn;d@IhVtOm#LQi(0d^8e$)7rWFU6W
ztEhbfsAm&Rzb<&n=*|GjB0(7)xF|DjH}1cd`hbsb@fguu{(tcp#~s=9^6KxF4_vNn
z@X&}w8$g!@z}w?dyoc`|ThNQ>1_p#!>v!0eR1aAp_&(gfgX{GCq=U}vwp%Ov@lT*|
z5U>^t&A|84FDgUI`C0;lujB?Le%eX2UU;5PV&7BFG@Itc`zX#|pf?BTZOzoQTMsS;
zLl4zb$yn<T44P`?ZwZ*EsVd&ajmI#Lb~QQwlXyKGz`X~4dKAPoL_6Y!0K|6)E<^}^
z+b3n-Z1bd@-)dQcIRD=G{_}BYYzdGeXF&?#M!>t*wHUEW0_{ZGTL)B(v}FmBd((=U
z!s-eFtsc_0NDD$8-q!>3ZbwS$>Y>LOzGU?k$FeU{GBg7E&7ZnnS`pej&cwP?<@)Mv
zq5`0xw+^QkjE+~ENa@rnRnANl75kQ}+P`?Lr$?P{y{J>i8IQ^N8V4@EE=2!MVZZ-P
zv#|dBy{lVUJ5oFwtWtLV)OZAu48El|@48A`U^!)WNUqQgixbrOGM3JfOU4ULDq#@h
zvF?dlu~iB-vfrA|ytAFM(Cj+3Fva7I{ZXxS#r67*?|Lz@`Nn!eUH+kc#dT2c&is$X
zV#DpH7uMI;YVWbn89eTIP+<uDWsL2i{vB-#)+=kfn|0!_HTKUq+Zn$g@DS<`R<KDW
zV+Vr3FRVP|>d(rHe)gR@jX2p1PoH!eZ3E{@a#iC9nr}~QnqS9tbd7brB`~sH7R1U8
zI#50#^z2V9Y(_}F#$Nik-YhcSa%5%lpl18>5MqQy#;&$)0VCIXdl6T{CchfSxfER<
zZ*BD#ieB?yE^@&U2;ePq+V_dZ@dw<lU%hbO?LHDZj>Le)zUy~58t&P@*BYzj!&yFQ
zS`R8&ZrYkr+}p}|JS(5!|75y<2H)bHzVwATs#w9<hht9kEpvX;qI14Mx>CJS5*|+Q
zESwzRXn<-N^KY5XK`e=X9RPow<BSer%jhH_G(Q+3ZTzuI|22u5Gy#ESm9(SGAx^!G
z3y|gj6tk<h40@o7IEvUoBp&$}C2ZcEy7ILEUeTqkT;YCkIL-apSZdG$oZuvCfu!nJ
zf3^>@nBWLS!AB%mLH%F`{AG0-R?{J-c6qa;QWMfY(iYDT>{oT*0^dl~hJ)|y%?HWM
z4FnUf{twdm<l^@NN!b@qBPrYmsWN*Q;2UNHLhkT_bvkk5w$gx<Vg8LPO#R*~@N=Fs
z(0k#2j5V#Qg4D6$qU)^OWxcF_v7Y_+Hf%7mS(&HIUo-N^{V7%v(=Tcy%wT^8i|^WD
zSJuJsuuf5Zm{c}3#zx)bkK}8Z$L{zbJ876##a_0m`DQ_?Dvbenn@OGV5wh)8uDfhP
z5i@W$^Y#TM%>BeUDb-^i$zS(@-e2<p^+A8KVyRQ$BF^Icw2ZMeRuNAy^qLmj)3ojz
z(GS}tmVTF_9g``R)pkli+ed=3m6a`kUo!;~T<8nEwh{MU?*NyOp&dkHap)c^gKzM-
zBG>lZDnJHlP-J+n5e~~gNOL$6v?<lN&r0+ho|&^0t!KvF^w;C!iXWP0Olu7_i`FW)
zYM`n-P7E@)?_bw6&si#7r}N~YYlI(>=P>)8aCf7DKI(O>)^loLJI=pp3hI6eCRl~z
zbnkx%zMo2`6P!%=@be6<P#N_r0|b+B9aBR6MK)A(R#E1^kN_OEI?3!WZ|`puBNDHm
z>|_-LxAQZgg9P7om>fjM{fV0dg^>Q?W0(DEMBbp{A31df_hW%zkC#j4aY8PnbnyKL
zR044RFSlL>0=ntIz+a51N#Bx<R)2jKrUZ(PHX7wZ=I;6hEIQ!zpuAy8B*OL9`K{me
zRvl7484G0twFZ9oB}kD3KIXYtgj$aHG7A08FmXnEyJMa&JKy>AZ+BPj8FKd1((MN#
z*>RnCSTTnO6*gDBvVWA<qj{`)!2-NSAG^nmZa)#KLqUKN=`iPQnlUVF?M-p|IARdu
zd38%U2s>5IFzS0R|6S>7K{ESiX+ygBv+G%M)!LUM`wsbV0YrXB#G7Nqo5dbX17DKh
z+@kFC*c0B{s1{<Q{VZrgl>4f&tbkXBS8Nl)c0}Yqe-@i}Jn}6}w-On<L{*wjMcC5W
z>{J>$4C?nyh|xN(*Z54$_o61|SQ9P6+po*iH>*$yD$$yb3ba_2G}QvvFxAVAEIOY;
zx#pV&3H_Bax}JB_k?`=MSyL_k&UwgoG&;b4Mm|VMceqIr^b~SXMy&CAbZ~3M{Ab;n
z;I_^H7Lhf?f;9=U*KaH%)o-}Akr(mu%I?3>$X9_88FrDiJ%v=9z=HHr7Pjtv?Gx|+
zV!zi=f6MKdrdIpys1~3H*jm)EiFdXGT4h$x2l27~WG$PnvIIWc_Z}Z!B$LOaQCnS|
zp;4tkSAJqhW#Oq-*l&|?BJ!#2w8KE}^(a|8uJ%W@W{2(7TyE@#Y{tq2@C?33Nz=NQ
zM}Z5W|3d*4cyF<*c<bdQe-Z3A-Y#R7<$1&H^fI@wUJe4Q%uEy;85wOZ(J#Au(l2=Y
zHB>g=_|bi+Shk=(yxxAz8L#Mu&*j{%vKBdaG#4H&6)*UFj|oa&)!%t&c}-=!usRm+
zFG#u|nA_EjKC9p^?fp~!4SeX@m^RxHHuP%wXQ2T{@O<snkFFBmsb9+N<UyWbL@Axs
zyo>JFs`n$lb%~0)nf*sS&-m7d@9?0?f{6{g+ZD=Rnsn;dFI};!QL;2~PQ8UEmCTje
zIt_^4xNLFR^ra6iTAddb4Gvw9xue_Qu8r#l%(av)H-aqij=Zw@?7mIk-fCt>yU2Cz
zh-9gvz+=V&Pk8!^Yf<r@dD8?iUZ5eUQx+JoPcD!MhvjM{0M^ASbf2t~5w>zlI+;9R
zW4|iH?N3h5)N-g>N_dqxAHcOs-q?tv@9tKRb#zQ6=kol*cqy_&*WJC7ykQ8fH{qhy
zHim+iG2iLu^WqK^E{#%LzPue_ZsE+xU3Bn{o+AtjA#8Q;v_zq(xsZYRTcKhq6^922
z2}#X<y4l@~Jov9SeZN}T>%DQgZU-K(|9Wh@Dt3=?Z)U@hfA_P;n)bmlKDZV7cEz3a
zVl6vre~8KHw2a&!{AyfqH7LB=5n0uI3Nw4-JnWTZ{QENA-AmL&|I6P$q}J@?q{D;C
zy1H~XCdN8*n*Mi58b-6|-_^MVtHB2GDtbP*gd!99v6bx;op*R4As>2L&3^S+0Hq*Z
z<_np4|CYb3E023!n{j!JWh#ZQUx__0rozfH9F}=7q&vnXFaC%wwq=)h^67MVZpAd|
zqjQr=#14FU@$1me=fBzrN?|#-qPurSCh*^e(e*kI2kq`gB7awFi4_3nNg9Mz2|Q!D
zx53X8$%$G@?WBOd>+&h}LCXB@BJ{aA=^c_~yjcvhalj91dvUc%y3DI;ha0KtAZT-c
ziJicFA}o~mEu42!%kp|c_PrVo(`?z4Y-ha{OkppmE$h7%eEJlIcse*;utpyjKCYJ_
zCgW!`j_=&zwSwd_^sEVFL(Jho(M!x<#{45*HEA6iVMQRRBNZchUaNkvs@Lz$Co6J9
zfU`2(GB`zFm>3Qw>kcNU!LPpGCLj}UkaQO2k#NI~-r_#cboV7C2^4WMdb+<my~EK2
z(cYjF3b&Kv;EST1>zz*Q+u8@sUdSCcndKk>b6_iJ(W;80sL{eQ`u9yDL=am@W0SIb
zzWx>1$6F)#aAOR>L!>-h?}C<Zx_@%;|Gf^F+s8}%lxP~L#bTr+qWJCGcVZ5@b{WR@
zF`%<2x=R?GN3uOh(?LT;{l}-D3BK3z#H<Pt7)LM@#FzxMgYrRa)gS$XR*jx;{tx{X
zhyu{x**SDg`AI~d7$<qRulvj<yDs|+kG>fT6^VlhiG>QivJ<au1FfGwmkoR((QqZB
zDqpXc|3pEi`}DNMk|IX|*XD^NvTax&>p-vjJ07+X0p@(M-S2yNnW*dON{jl|@c5Y9
zNmxketY)XY&4d#VU2@&Ijf^a96g{N^9;(%uz}gmagPosxv5c(XL5-ZTv1Hs{#N=co
zp(I+t1tJM~LF8S{0;*9Ci!BE)fE`ggzw+N})!d6z8m+l5)t?g-NhD81ibkAl@+T87
zxEjOFHESJb&Gy>ptEhOyC$#Z7t7<MUij!HddYlEN)L$z5J`KMb{i;(W<{F~8utX4i
z^@Wts^P}(SEq%@C6B~^Dwy89Q`w0y-q6GJ9_^nUUH>nG?N&LFEot9#&o1Il}@U7jS
zoakW3SYCab%dVNd7BLy^xN_-qjmOn2Zuwjp(Tr^(Wmy@7<sfCN1hm8-H}08uDEFr{
z?Uk3{tUPa{#PKr=i=1t6**2dB(Y(uWdbz&;hj54HkcV@?H2B>Oi|zd1wZe$i>YEKY
zUz&)hZ>YiK9L0@RkAe~I2ARC~>a}eQw>j*o2kah86KS{k(U}kP>3O7K_oM$KJQ)9$
zNG9omhCNZK(QBrP<l|Vm#@YZQZqpx4>yNcAu)0phSKh}rr4Fl<jZwYT$lr}FljS!(
zSF4so?ws%0!9upcRH|(vU6w<d_$H@Ij`){WvpT7^>(iKdr_K+SRkY?oRNL_AN&TsT
zPmu=kz5MMBRe4b;ljVtgMm4kLW@Yw0lanJ7L?nE(TglGA{Jar6w=@{Vz}R86%M~G1
zvr<tIf0T}xrDn=)`()kG7_bJyMH3a=dX1q>ax#QW_h?m7c6E52USqM$;d-=IH^8%U
zCaCK1k0&wp#*;WX2&2VM?1vSm#-~g%BSvdrAIS<K86Z^v*6d|%r77b3>96j+bp!){
zYcbOlMwaxHqKx23N?;u<@ESO*pQMmsl(-l9t(V^><$L~+2r&c4ue}RmV&@J`_OBm)
zg*;CT`c!oVQc@^j?9EqcM|kYf7po-q<rt_}^|TSv{p`x}eh)yNB1L^1BKyQqa7!}Q
zj?mM2gI=0V_A!g=?F}NIa0H&5HD0Q7(Cs(;SKD()R_pci^Cad3B&^F8>*gu@$+K3g
zGHK(L(ch^nJH;=g{j#$NsjIerAOAqH=&hdb?XgBF@HpDJ*NBT#5ho$j{>du%+yxz5
zxrtluFyjj|;303i9=zj^o;ItxP)hoB>yW;E*D!LM6QGW$h#Zf9>*AbBux3O)_z`@X
z6g$k01Bk1L7bsf*wa8#I<SkrNtyy-edG>6chxXQq|9S-B=$|bkJ{#r({j;}bBIUKd
zAHLACLtPknqF{Uc|8&jCEW+cf+5odP^3vC36HKdQqlpY6A|i!=Z1z`~(_DdSVq1AD
z(`5_T=y&Z|$Wg6_J{Z~LtP`rjm_(2LM9>py)%&V%PD3q5FR#u<<C>vjxl#7#lb1^f
zZ4Epw?L3m`M?|L`1HNVw_3;V8DzOrl54ajHYECaCXLjDC<3t$L1@NR8Zilnt`E4F2
zf3<gK59k*BjYS^4ZG85U>DhrP1@rQn)5ptar#2(6Tp7MYHrWY@h*96lmo|0gR(p;W
zo7Q9cv;Cv11Rvj1nhV5*S0mBs=^}I5-2E-B$zK^>pc_U(L3#2Qc*cLY4$#e(XYU(z
z5)8#C5)pKW)?!OC!I+~lk8g|O28wV0fEcJRYj{s(H$yiSpyWZIRRXtag05CB0JLYb
z1G39DqB&o}bua{4=r(6x)#)$~1hc#&DI)--v|<5(z$cuC8io?pcL{Mt8njsPZ9IA5
zss>+^rs`fZ0&6^0Bex{uHiJG8nrx!RHCr<>GKS@~m-_Hz!l!!}4i?JK-B475#GkSV
z)yoGC%>;?KxUhP|;9LRm6;~m>ceh`Q?q7dJv|yoKF-7dKbgYp4O<$519w6qnz1s?X
zS4_x&=gN$+p?AF6iJ#td*4-cEtx@7Jk)=mcP@~T`<YSu`x)BF%M?UL1r%Cp|L@!i)
z&<OIiKa9E-q{n1{Uwzb6x9`1uebzokY&X?RB5ZKGjZv@|b12jUU!yZX+6_0Ylai3U
zqg2cilal)M@#;)x>spe1s#K$@^ftAG-<7egFFPh|$N%`CJ&U5DB%FC33SZ1yzfFN)
z!eFWZob+HHj>j+!i$c@4rhmc{C2wIER(f8S)uRXL6!yyZHx4vD{E=3abPb^X{ll4J
z<dyvac!m<<wBT=^X|w~-)5aT1@QZqEo?`zU1R@lcAyTA3!I91b9t#LS$(LmI_EY{N
zAyXT1x<;>?)TB&x$daHeHB_SXhQ}W8?QqWOY<R@9AU}txX+Dg49f7vdGu+_kb3~!@
z@npLvj+#I%3|-Y&i7(k5j}URW+K&(~g3`%w!6@%?Jo2d{iNdKc(Up|Za2I_@Rmv9v
zi>1~2oGZ|TYZfMYIxvoN7m>dqsC>gH%?VD1N#FLJNX>vx9J@KTcjtUvzdWw1<9+6@
zt7~)FZQpxoduHfbw9=kX6uiU&CdfKn99U@j9XHQ9lWnD-{$kh`INugKy8o-+F<D@8
z^y9X_ct@x!AgKJ$_0hsdelSsJwqGHLxPb4kBZC~`^Lu)I-TAIxI|~+|5|A8KWyhwq
z4Fl-j<4DiLYek@Kr)*x_?eSQxU<@jbPy67RvWPE^U3NrULXYWbQ)Y{mb!6JYZn<&p
zj7RNNSwqh&ll|$2w`%is&94`S!08_3uX9xL(&3+ZnaRFVq*#5h91$K)mEfN8VX>Zd
z-P`k!H!(c0P#eRqk~UygbT!}-;!E7DU*N-edqwUi?}sBS>kHGL5l}SkM=xo>2bVOp
z#GHgd?@Q*<GkNPpzAKze9aJ~Y*Lhd0#hJDefwWwfE6kDvypHDo+-vAF5^~=4p-$F_
zO6qc6$h2Q~Q))Ij(Q-J94T3HFf>_51LJ2!xKKm(NW&SEFvyacwMxY)fG`NQb)}udY
zr0i}dD#CbyA8z=&z4cq1#fL`z7Fd~!M*I9IS2>zVJ=<!JVi;Ihf3+_7t<z>iTi9`d
zp%3=@xa7Y7#p$(^m1KEC_AE<sBj=URM*jQ(cE<77WRB-nM_l#88Q(F@Vpu!<J<t<r
zQPdTv8kQLopH~_8z|nfOzl=E{aKl8Q2p+aJs`vbjvy$|TblOr;?zXH&%FuUjp)P>o
zeMM!Hdb8%^V6>BVO~t@aLr_qV^+?p}c=zeZYYkCw<494UF;p@GoU?wC-sOlL=^EA5
zIe2x!%a4qVEUws<?I5}LdqFGC&c~hgLJYLcHWs8X!w2<~o~Ev`VtQysG^XpfRhnpK
z;lbl6TWO7b+YW|Y1`}jaws+cfxjrf~GKjJ;C$Bk|XHp}HQ^;A3^KtWXysz&0(c~pg
zaI>U{pg(3o*Qnt|x+**L52jwEB%wdVcO=Wax&Pw=APAQA@JUNcr#x)pYBt(3?~b3Z
zWm|<I;M2CBF5EWD%v{YpW3=FT*zQ+wjq{zbt>md3&&B9QWu;5%$E>!w6W?Zmx9VQ~
z4f&5--Ffk<0MYZ*S7iCytC;~aLa@M(nooDv8uPvB!lJZA)se|jjGZWy8B7o+b(IYS
z?$6Nh16R_j6%)a2vW)GYs=Aoa(_i$XDl-uZP<*xzN}7*$acSk(J8YmcoYox6oE5pt
zeLUc>7uV19{a<I|=3QR9i!?`wfXt0V##Ld)cisp~b*>}mP+t%t6QcP=_xXd+s^4|Y
z4BPMgIA0kVg~sV|=4(7;<o?g<7|0)l3VKv4q8RSa<!tqu5LoqnaGPRj+B~z9l0xK{
zoBkYCbqP&nby%)d(kpQN5d$(NbgrDUspf;&egTcin&%u>=j+3{!E-;FTgvIj{oaaX
zE*iHqXW#XWM_Om@i@&|)kpb67q~G=BL5ADWZTLk91ks#p{-}aaf9V^m*1s%WALg1Y
z*f*aPE;~NeOr4Ys`90bb`tgq|S46&?;oCf_3CgrJ*^AaC(@(XEN?x6=Viv*iphH2c
zRklYJQY-R>@-!T6#Bdsj)5FwrA2Vh6jem-6y^)zM$vkwDRqG#nHTMhIGw_AH^nZYa
zeJ#$Qm|7Lb)K}Nna=kZO&>U=MRFP>fQ=Qjy|2;Q{1Goednwffk9(DK(Blt2l*HSlR
z<dE#jPY=3+;lKkWPzm!XT!cJ$=xF|OKY8R;^C#8-auUl~L<FMYnu7~zT2GwF%+{hw
zjprI)6@387pXa#FpW$_3nzH(h&h!XpE8~0GyH8KN3OqTSJdI~x?qM-?DvracT*Gb;
zBhX<n>=RTPORbEK=h_sz2`~5N;sWG}WvEK4NRk2-l+l4%<2YgMKJ52+;$m~XMp$0*
zxSaN>-@gx{09XIBa+!KXHN7A>afS`*0e?auD%|?VpI!yYidpTuzjeRq{A=4eFJH?o
z<`5IQB1v}UIQ46?s)lezBHuLU-y@f6kj+(${p;ba)y5Ci^X-|q%?#F$-p9A0*0JOx
z3mdB%28}nj(@|AdZR=pV7gR<;PVAFb_x>?XeZoX18u}#7E)(1(4r}X%EY+%ompV{G
zk@#{n#bPrpVFfNnW7Zs7WLl@5IqN6H5v7+03)YJ$_dLZY1}SN27;C*7ryHwPbWF^x
z+p)5#ynCa0{67U$^w}gZ((`(MU^%KJO!6SLE!etq>1`b4q7<YQq`TLbIiXa3*yYJR
zWp;gtx?C{SP=@3&s}~*NNu9rPxlG=QJE3({DO<ZUnmKE(U&c;pGo$u@w4iRCJNoj_
zaAfSh#<)LKP5VN3w|tn@0y;4--jnKmdMYSBVVW#=HyK@R!5YJhIX`+KA96XZYjnt_
z3$<EalNv%R*-uD|*i104ml_kB*@k=4ZibV@#@~=~{^6BLVX0Y0fP)S1B$jux=;cU=
z!*gi7*0_V^bZ0m#U9oe?Mf^W$mzDtDLl{+rO?lz?%9om)JidcVDt@Y63-!%f+-)RQ
zF+G*5V(;Kc_}NFv%KY6cb!&b;T`tc&n`&Yl=bw44`FlJoJ%7o+<vjUA|G0rhWf)X-
zPXT>~PS+Fuk)LU6`_b{_Shu~)B7$9jbT}ZhVEkh5&vjh=?!DmzO1sx~3|G;Jl{$H#
z4ffuG;B;yI{aKdT&Se$;O(V$2G+UwKs6d#z`Qhl1wqm3Kf5Gq3FvRKv9nD(nM6h)+
zO#tRy0dP0XzjQvlW1K(oY+Yt=Hva8;^z2Q7?PE5alRBTwPcHhSuVFOiOjB`-qutN4
zmA(=E<c>kE09pTah~&xSP>pQ7VrN9(k*Ukpu<3R#=HbZe$z{H62VElTS1h`On3&i$
zQ}lONr2HASidbvAvyf6>E#O75n#d$^v`7Jn@yW}A;Z5U@%m^{qNoM@T<~aVpckb4C
z9dbZr^ZRRg(5EA*Ut}3N7p>iPCunC1HRCkRco2;XTCm|0@a8WI`R_HY2ufj?&8QC}
zO^;giRb=kD*q>v2GUxxE_B`*5XqwrVC^_Zrs*c#EPnGw$9=062^}8ii^fut57mZ1k
z=?I*x1woe|K!L0q0utu=ydDui8F*LIn8ZF*@00Bef&?lMn?m`MGCTMH<54_N%h})o
z`m2i`nyp$+2WUZV3Pt|+C>gN;0Jd;D@^T&70n-__hf|p|!Q;<N_UquZrmcASR;~Gc
zen}Cm@`tnitSq9uPMT?9U~w{C8o*k9_El+a+XZco524p{x?x6Prao<Bc^x*%Mhguq
z%xIh4Zl&2wU!Q;!5G7i!dWE);j{gfqk=4g5-mD~XHS&KyqLfD+qoy{TY6Z|nvGeF8
zaX`QzOj|Q$6cdFvVgK);=2A29KP&M*M<vaExek|i^c@DZatarm$rt@;1BICtw0#x)
z&?!72%5McZk6Amyz(;HT`$FIuj0gy=WqRUmu>YZuJ<*pDfSdaj@`sci&zF0MHR!*B
z)Mm_a(vigLkV@Q7XAnqoui|@<Dfz83+t--Dfg|1IL?!?GEy$F?SS7R}0#J@HUkXdl
zKYVG0;Ca`K^wS681xW;N7$xUDiDn8~>2w0h(_RE6lkMVfAR?0oL;%xX7xHGt!>J&j
zou5!GP1%HI9u8`1UHBoJIm}tVyuMrfEY!ep2(_W+Cz@AE^`P+SoRNCx!g|d-*YRsK
zNyy=M18{%y`Bp`X;Nu1C(fxYU`nV?wvhre$84$_<59N0%=;DBtCM72?JVjPZ61|m6
zW2q?n<{bU)f|z>eS>@LZ<kiP&WyPZDnX%$sJMz-fkKQj((PSaky`k{vyHEZIZC!w3
z3>l96&q_pwVnVLa6wqDOGf?lm17x?g!<pGwmW`Af_RSetHf(bispB`U))%x=U{#7*
ziVF~(V#xyH`UE)l|9(=RH1*!S4)E?yBx8mzTUR=Vsle^S&fU^tM=6)#S@8Wj^N=aJ
z)xVKkt10^NyFy8^%47}or;ze>r$M95U>Nl%5|Fa%GJO+M3Lrz|e@gSFyqdM+>RR8r
z6LP3*pox}yX+Snjs#l~+B1J!ZB15Ed-Xe_$Dq}%FpU|w1>;B0-Jlt5?Wx4U#a9D{Q
z$sQ5&_STPIEh^UmGUJ$S3`8gP?>gJQ-Q#n)^=ih?v38cr#ncH2Vg>TmaT@<Gmx0c6
z*r@1HosYF%Wf1;S;bS;8&u;lm8*_Q|-O-shD=8E{tB&T8^BCaEKbluOAdlQUg-l86
ze@WuyG<*!ubL7cf)NDq#n@-Eo>KL$8D--ZvtNql5cq+bJkAakH>m#ciDG%Ma)rT>-
zu?4@L=d#*(X9OUTUw^}v3^#nu_!exBeNyee&M|uDL-CR0IH=GKC(lMgr<{5*xEH6s
z?{L+Nt6iD5e6<bP=d(QA$qJ!YThpHjJzS<AZiv3s&W$p|L9y`5LZ1z<V;dHYcem~^
z27e+U6w6Dy%nk_bLE5f2p-PKvgiGL;+pVZKFE8D~N$QQRwxTtt!X3MuN(O0Pe@*xn
zcpAe?#7zzZ+f*)*SMR!gwjP3_71;i~?4)Ln`PAV8#eugN9T?;j@8QU+&BIbN)`Hq5
zb88HiCT^XzkILN?v87py5X|76grKGpa#}<vBKuU2_VF+RjZJ<`DDT_HMXh#dJ)5Sw
zo3rApO3h{)Zf|y;cNMs7lG(|6HC7e^GD^|l(RXajIJsut6rK9Kh<Y>28`TsQPXsnu
z1E|Q#SD%Q}Ah^~VdkC@pBWpuS0|y&ruD1*uE1yz7Zs-Z7Z@sUE{)tm$l}5++cHuG0
z>2iM@2s);EcYaGqm=6B5*$8(3)A8yTO$g2kLLW&s1iv7pKu0MxIho+CUSeIHxbTg0
zyHG7PS*aU!9x#py&BHJF=$elJp<=a{%N75PXNJp-7Z<?0A#HrFj9Lh|^PBp7#Zy|w
zz9a7u-JAc)r(#vP)L=te=Vo(t>;?BR%l9)wm*smLj^dQ8<>mI0k`l|=+7I4W`yt@x
zBQjm@KOvSgBz4oS-714inYZR!nI7w6G7(~RjdqSpB^Lh=Yi}JE$I^X`4nYIKHE4hk
zT!U*ukOXI7a3=(Jcaq>5!r(50yE`OUaCZyLfx#iTbDMMC^WOXYzI(sFZa=^SOwV*z
z)vmqQUVByb`U3D=_eBlGlGa<8Kd{G%p*QjB(ESr#0L$0*Mvb*78|pdSOOuK3@m8GJ
z&fVgYu<3n~mM1kG&snW<IH<V1so<Mw1QZc?iBV)48@BSiJL|W6tAG7&`yx#?o`r5e
z4pm#DYrAGVAeU&pk5w^GPWDdFg7bW*)DqdVE&2JYJ?&(NMq51UjCA$u$&PE^;8!Py
z<px}xlR2?=GmR}(pgNrL!y2{2Gi|S?jrESxdZQ*JIB{XcaF>61M)~W4l)ZY8LJ_f}
zfG71QrrVaEgACrJnjTam;dEN@jsz4Cnk@n?gT{v2yZNBq0W>!s-SbvwrTG(|)j^R4
z1NT<9fv?2TeExlO`1tYGuL6%w_p|J*EQou##85i3L;PdG9On2cyc-|QN&ot2K3Y{V
zVN6#3qa!U}C<Ij!nDAkt|Ie6yWI!3o4?+4fN2qLkJgJ;z8IiE{P(e6;*NUYg!beQU
z{8Snp*F@G&y#3CJ<X-F?P0y`o`d7a%-u@zQMJ255hB9K1qoeqzx7CW6_xbNxL%*Ne
z1?hoUobAsv4ScbcBsTsE-QR#|>P3mS-CU;SYsd9culirl7j1MA5ZqkQbPh;p{@yIL
zJ~}?_7xe24^Ej@Ojnx@kb6)vA0p62v0F~V9efGAah#<}IB{0vFcp@}^k^8o-_S^Yx
zkETK2P%_O61|d{|Rgao!a+AL3^vkWghCeeUuN@|`et6u*%CbT%P)&SS_!TllpK`Vz
zQATh15e27<Kil5ME<M(*@?n-q6GOojzhye0axdtErMr%(uw@9~Qi-_AT+qizuKM*Q
zYLdj7w_shE{s+y}c%{d>749h@-Y#CU6DJrCfNd7yV(Om{v#^=hExKfsax7R4muwc%
ze|LP7P#I|T4I-_}LS=m)Ee)hR4fYdIOJx}r%D4h0dr>!+8B>s+L?Hr*UHeYKnI_%S
zUavQ+-aH0QRRAh{c?DB%Sk-O#jYT`6^r4J+R&)BGKRuw2x|K#?<=Xz%{O7wuHIeDv
zT0?&cQh`-r%oaCN^<p+*jIh+cl`Y5iFKsuDKl>R0QpOMY1ToxGZ8j*X(HQtnCNAHg
zp^L`;WZFJ03`p(5gRt)3%+h-q?VV|Ar1UPfvM;)K=QdreZ!=QAlbPIXJdD=;vQp03
zarb;y-#3WwcGz9Q=y%xNV4uFz_68Y`2go3epSaFy_Qw<>4X)Q!t{wGPW%KqJH~p6N
zy9@Ii%THaHa0*_lFH1LvJwgZ^*^6>2mf?q^z=tf~MgFasqO$3qyDgcU@O=RTEH4o&
z0$1+)g6vUm#XHENYROLzj}Pw<?P+f&$C@^c$%XT?Pu$r6EYRd@Diam?-Yga`CtgWl
zR-@yr$_h{P)6}oacGf6@;7MLXGfM&hDdd=%;r9@`{&2a;ctVtPGuBxc@#>5MCEjm%
z8Mou!7($fpyW*hWxz61&NG_Vi*?za4nl+oNGpWb54*vLXFLJL_L+)R$6GFW#(xbR#
zPvsm3(ayU)J-MJ}jc9%D+l8%R7R%#*zw~Y9p?9%Dca&YYwlCoRjwMnpULeun^#n^_
zV!o2(!R4P{Hc;xLY6Pbaa82zS&q4iomT;(uJx~PA@O)z5al!Fw$=x0H=biF%yq^ZK
z%}lG}HOaGrjWLH)wG6`9u+i<vm*`4jrKM*29rYo)qv>~S-rcdt+)t3p)`qTSy_PfR
zb!soPKlRlfn1~sEQDcqT4R#}&D5~h&Xvz=~u+<k=!fr7zTr4pm@A|y&eZ7R!G9p;3
zV7RzirMl6W5GDBPw<{!ZP=#CiV_&`*@zuF6&CD4}rY_=8>m>k`#+l1I>LEl$G>i0k
z&{C$Hq&2Q*XXtQNP(P5!Px59SH`wBIH4a}F<s7i0pb05DOUOe3LpISS0I<{BwI5$S
z{I?qfZ%RbQ)=EmoPC~f<X74oN@3?(?U98Lf?vyYZCyGA6U3T$HN!W4pBiVVe;5&`G
zbLFVQ)EXL1b&g%vS(~=YbcD7d(7N9F`eFYRE_Zay2rk6nXqkHmZU-52e}8<wkm-AE
zjqaQrIv=5%kgSA<VfWQ|TwD>;&{<zW->u{%zFq&f_7abX&GS{DMILK-^VU=K-5g`^
z09|mT$&mEU$_tQ~%V5JBV7GUR^xiv35+>l>9ss<n4lQa_QY8AgbC%zpJol+7Fp=Qe
zsq-?7S~dKLR>yjSo0N2=p&bz)ynRMb9XzbdpNQLFG2S}i0nvt0BMf{jyBmN<cEOtM
zNw>4<eI&k1&llwzcSU9)^xKa)ZL|z$l6^;W#TuLh>t|yD<)3a}HL?6xo%;z`L@9zp
z%qvUZ`j8z<r_S!Hv@^n_uUzacd*&n9V@VZtZO*R7Mt0!ay;EF9ehCxpbWnZf#7AWa
zFnj^RjIJr0EPQQ#p2Ar11p1aAp5C8y6O!@o+1y^jW7zEJas;~v=)x_FdsJqYLjG)I
zK*R;qqB74eG~ezIaGsE6S5}q=jT)zmG8AKG`rcr1J8z229AD#-N6rjW^i6+m5Vxn8
zdAMufTgri0cT6rmM%?@RlD!A@*>YKwwu6otYODUKW(aG{<P^VEwtP%(=^CH^mAJ;z
z-4%joH+>|%<vOjr{NTuMW2f%lCQ*n8?~;SsPD4>aN^#%=vqC1BA`;1+@+QMCOQL*j
zV<&_RwTzj5<nZ2@K669A&*(Hh-+~mSh(dteiraRkIP+G6sOh%S%u%J~dxNFgmBT0a
z5KBn5x?BnP*S%Ja8=?69x=5Py8mdMK)az^qp3zcMGBx^LqKAPYzP{_8iJ2MOD>*J-
znU&_iUGE`lGo8oR>*~mi6m`|g3UQ#n(WOZ4X`ejR>ZzU5kuGu!j8bnD#Xv&({1n2a
zDA-u+kcA1u>gq9zQjf)>!|v3-9%cY3#exq$B{2mzf$?lGod{1(oj0Zkq@Hfx2(2a_
zjxD(EYPCv9IWh$W$yCY7h0XGhVtVXu?itz+6Iw_Q9ejEBAgfP|1gUbl_n9iu?b{{b
z=n%)Eub>Jey!L7@QV9OGG8~7<@LKZnI2lK*u9r?K9k-;duM?DBd(*FYAZ$hds6?L`
zMBwZ@_DGX@q!~b@nhO1bXKvf5?b~kn^F{v@o*kT2w0|}Ay0mM9e=uIlR(M$lR%=tz
zY;n`xyEwq+P+a<)4%8rYKPYjC=u-=J4lzBpFq=SfYR0#{7OC<{rxEi84jR-XZFf|s
za#lr;s{D8Xi=h>X5vNjzn0bG(&h#@Oeh=-vQFc<Duku@%c(~71iA4jS@+8_eca!|8
z-woeF-EM$^7*4}Peh-}?v2AhA*1o{XJNUEFI}y(kCN*EAw&y0&_)9F~-|6P)SieNP
zdd!fjS5tOmj}$<R$STBA$S~;QbBb_e7T?bUVZK`5Qg>3s8&A=t9JlIq?W*{7>bROu
zMntp%fxdo(LN%B}zHjXpsGa>g<E6Ljx5E``28;!Zm1)?n_Zu(|saz6_-mPf({28_p
z73W6BjqV^k*}VI5vFd|m=~}#!(jQm#&`vQbF3!z1X!U&_1z=92etcExgCrURk_>-2
zw6@&Hk@mQ(dg|fZiv>jMOW$%e#xkF|Q&pvMlJ_jrhAM-j?;CXJiIb(p3s{+%XBOdY
zXJmlKQDvJqXv9*Cn2h~S@Y??-A4HZD0;s2ZkUpYnvA}Lck5gs-nxvZ!Di-#m3Rdm2
zO!1q<7JT;ZE>`GWgPhLD5jaO)PN}!UzGK-74`0>C1PFqqXfn%L0%h&7B%SJJI(0VM
z;<w8862;gGbMI6KV<Wp38oG^V^`kk`Mc`1vm@dS_4QGa@9oj*Yufp7{CdB)K>UuOT
zEsatgi^&@+>o(o{=q3l|v6I~Qrzhm0Yp7x6&F+ltpHq5V)V84K+t`C44Tp2vuFdSa
z0}~<UyqM*ZC9b>tbM=16^$#_}2$yoqo?vWplsccsA>5gGNEE5G`u(hS2;)9D6Jz_M
z?w;fI+Gq&~n9|exuR-QdEKuGoC%Zd?7EeMm4DM8~p$zW<^Q+<rV}LdW;`K|0rUv3+
z1>)TV;k5+f83LFoG$h>)+<}BPjoTXUkC<|?&UkaSyNq0CJzJk|P~!#3bIV=&jpXZi
z>S*>5@^G=%f^TTZw39#0;jhjB$&iD%FZB1BZZ6ux^Join55-QlP;QTQzhos&SNSd;
zuSI!!QWr4u;+kya^voO?Q;DG0-InXeRAy+FT~8uhZq5_yTy`gld{qV+fd<m$G#0S1
z>$@jLIg<7Ah8Rzs)T-_H@=<p(>8s{}IYxfdzIkj+biDLatK+HdVjBtd*)62nNq2Y=
zr6uE&B!<AyE0k&$h8|+}U^G$Rhgv1^n<dABiOzKQ{*9TI7PrHqolUenv%#5}kJbs(
zk!&a|`%CZ`QPunFDn{z3Ue80^w?p}D2Zj6e8i7Z&P7di-qQGFVnIrt=X5hvTmc6w`
zS@c(~MGwm&j*t0sfOT6IQ(TR_?!aA1jZXnoz2Dx2heEHOoac8a80D&w1&u%5N-*7x
zLtgWqX%=}7U=Rf;gB@$SxHy0o_5ehg`km_-JSf8tI{A#YPx#TKp;gsPjzzE7iu29F
z<St&oL=ELUQ~V6?RD9Jj##v`B{45ZJ4pw)aI^nk^SE_G0AA+szD>MABHVQB;r+nr!
zIu^YGP42GiX7uVNt(Ah_9a1klz{?)$gC#s4^nFgc?EMf`C2AtXj2Su!3-@1f?{OIz
zVDb6MXIJJk&Q?||N>rB>SqTzbd2s_F64yh{!Erz_?s6|KhTCOJ=q-EdBk*2Oe~?nN
z68N|j_H)QGW#`jStd)u;N>S@JDY6Ga=Ps~K;*tGbJGsA#CQ;jHHBb-8@#q}LLE&?A
z(eQ6G7m#B<(f4MC-ReV#wrMKd`>t1J%OO~k*DGIN->Q=p{dp&nMsRwJeky|25Hs2v
z@_MyzYw@uA^G~|WN&mY$y=Rm>WS*%FXl>=%hQ2Yd6NGLTV>++k))jm{sdQXJ8mCml
zeiWsRF;QP{<_VkqF!|8c9|R^mp+YA_z%cC-w&80lHXaHXdYug<+4Z^!c)POL8Pol|
zY7^bjpMuQV3{Gmtx2>e8z-RykmtU4CoW+rw&A9Yi57P&@W(m!zla?u?uV;WUC*lNa
zf!FvvcU<Bd8W{peioFS^&!P?9{zy`#&&2s;=Iozp=)ZEi3aLS#wv5oR3)6|-F+lC4
z0YLfbbSq3Czz_zNXo^YVHCupI<<=Y^In?9jG(nx@&!mSs^06oWoM*`%ncwS_&ec7k
zIJf!v<Gaa}oYLFHYxeOvpb~GpM9tivJ?$(0-J6DdMzO0#`3~K=BlGJ$vXWHXD@9|n
zqW5xgU$kBDMBM)}-N<A;nlDv$n2*<_&o$n_4Aa5$_%z{6yb||~i8rFsQ;iyUfFU-o
zRa%&J#i*plOjP`)J8Vn@kt2TuFpJNWk4(DP(ie>fPB`gz7g+yvxUc+7oHYnhL(il<
z>P-h$mL>f;E=dw{ewE>QZ8?<rAo4vfuF4sq<a6C0!nKpPwos~?^euhg@m^r%1|!kX
zf&YNgorc=?^0OQ~)7uU6mHR<Bsi6lBBGaEkiQEHof`Yg6$5-?!lPjLq^MMGDqcz!a
zpz0Vi_^>ruQROlI^LQ3m?!iKHzY$_zWv8f}sZ|%L(|9tIYSLQ!i1<Q(Q50AfM{9)$
z1OEO{^I^8-iA5fxCRMiIZ{A_WXD@)E?V->h35vy&bp7R&n8;9$n_vANBlg*Ed5~|L
zLbd(gRw*}lK(j5<9>x+O8b$Wn>aTKOj5Lr{JQt%?=YP-<eXmp{e!HCNR!hBR=*H8#
z>Wg09bgk3+Fy42$)^~U`e6ea`|40(yWEn?nZaKPeKBIJc^-9qX$5xl`tTFBC#IJ_h
zt(Z?SH-X*ZbiD6(Ld#)SPgh8;{_dx74DF$&^YDnOk5@dRk3N;(-cKc7dZ12I|8ogP
z$20dOZ;6bQTi);KO0`>KNah*tNaZa%EEO9B7<`9-M+;=FXh!2~JI55iKBc$0@@C)R
z9e8sn?)zsYB-8Y=a<Y}fvS(+y@sN6SIY2Sp1=xgz<q04@2(=QX4Q&49SMOJty2Ru1
zeny^oar*R9#b=CIGR!rR2I#bD3%W5sonI2bc`QCPwm&7~KPcj&pY_RN{7EZAR)kpc
zk_I6Cb^*!m^iMTn<w4_}USQFxx7Uoqpch`+?N*L1q5<-GQ2P*_R>{y-JFytMrrQ|Q
zY>qFFH3^YcP32OpN{2XTHAFE&`rl&1rpu-*5o)U2X70(}ZYgmYxgat?YinA<)<k8m
zT7CfWK<bAhS@5-CX34{9iQjkrEec6Phm_4`hu}$LG8U>{jCf|sH@F~d7m5Czl<o+i
zIzZa%y78Zigh3<82LOY@Yptd+7mrT_gcM(At^V1oVQO{T2~Q6}N-IgtMOVb&meM35
zMi!Kp6OxikW{0)P%N6M$ixnmo`4tZ1N%9_7;$*YN{cU-nCP8cUR3h3uQa_x@WPoj}
zUm;!M2ka0v-~_m&zm<Ze8?JLjvUU;cgcuj9<J2JCW;QD!Wkr><f$tbfC@`$vc>Svd
zFeepxmMakP9dNmR+}QeLADCHPf_Xf5PEP?#l+^x#a(9~!bxOE-53u=?p{7V_`Tl9k
z+iM4{bRbvXQhG&+!zcybf9{FJaoOl1d*@p9fl+^uOH~wzKSuf=CbTuj8a2N($0<!x
z5^d_ImD=+_W<VZ;>=a?mKpx)3buOY<1Lps3B={^h+OF)c;%8uOb#WE!=SX_6W^%MK
z12X|fG8u^i%koX6$D)S6mNt=TAYL&an6nKcaXr@@;%XM_RD`!~r_EjP+YVIbJ9J1+
zI`e7!F=9$dtu(G0F$Nkf2!&JNbP4KfPM|n11&i&5ba4gGv4n^B#!$G+D7n>_eA(mx
z`f7CEx7_9>PS@p(Wr|2~<-m<Cv2?N1g*tHc@x|k;azXGcqxwu-BIvfn%KH)aOn<iA
zG8yHXB~eV_v2AK7iv#=NzA)@ir|PDL^Y?2k#x!9D;_bUKR->zvp9Ds~di3`|1SS@G
z26tL+&QC87{BssytM^i&lKpj`hU!Phl>FkQE6ENQGy11tU0f-k#(w@oSa~>ITlpP`
zV`BCN^u8G!1@!m&Gu8+bW#vT@+tV7?j=5m(<|y5#S!H-RnITskGozVu%($;4$Nm|Z
zP`V$+WWz`bv+)wI0(3!z7YAb{YL=T*XYm}MjOj_}6OZ?1E{-XmoErHDjmge!UV6LA
zGb&Tpx49Q}aY>dn3e&u?6%$>wOzCJ;{>`fxKmV7|qHbbPBL{$|ie*aK!o`+tbpNWT
zndPeoss_c7@2o+~&XS=r<rhJC#o%qb<!ATf-B^!P`?2j>Q4@DaMGxbNB8n&41(tTs
zhM#vz0N~uP#YUqGibOwBNNO-kvc1GAusk^7jh*DZ1Nvf0yMFuAbyA=BfkKFJVA+>Q
zpk+B<eMx#UKGm}??3(ZL&j+VAMX9;p_~eT;=&=VTVZXo**ioNEGD?Fw`Q@O+GEvIO
zGt3;$1=mNV&@nTEpT8Wq)R`T*yZL4yGVImS*hqH5UnU6@-Gx8ZSN3|RjHcL~$hizy
zL^T)H{TW_8aWq-X(5m-kkG34|Px*9^FHwfb7#21iub=)^2*JIQ_8#lD2u;{dzO<f|
z*<)pqq|ePbYq-q(iu$W`8bm<A>TciN>DFg7e$L0l9&4e)xU4h7?<|nVCLIQyQ?JB!
zr9^8stF2qfZ1Kd`*f)%6{!qtVyZnKPX(<9IX{c6cke`$6%YL2mVsc5jJUq8|)QCeW
z+%5ci`%_;_uaaPS*R|fxUaQHReI_||4Cdxt%BC05%<MYWtns>(bKx27_purS&C;a{
z0k?#Mu2{?T%8!REd9P+Atz}=Aa_EKV!*0g<$H_a0%B2c)6u-FxYFaQjHUo^SZ&lZo
zzg{+{T8qw*LxhE6_YM<fLJCt@kf*XP4<swRU{o*{;Z^vxx!+agC2o$V)2nugN>Fk?
z_BR&F!gLcOL!kOnKou<9Ustu$DV+9goUq32HXv14KNz9W5v9KU<`8YU%Ha^x<-=~k
zT@U(Lt<t-x4{~2?YRPub5ynQqk~YK2r=IBbId)FuJ3<qcZmTkSGgoYT?98?6dMpfX
zd#gddBa(SSR=MS>q!RY`RNiSfW36tl7N)JkHqUrslNUgPZl%(1#|>)b>_4Bi@E50|
zk8$U+Q_{_|sC|CPTay0yJSxRxKiYQADq*=|jF*p!p&CWP&LEUZ8uiy=2dao1<~N=7
zH()seuoOY`=vI`#eR9wV1$H;7h@VWKPFZE|Qou69P1Q&f1-sf#j=t0DTb)x<oL>aI
za>YTtK(9y$VV8YRH41ZPzn<QYj?~$nxob|vY@mJk%~4x1iMbu^1wg2u+{4Ye+E(v?
zq8G|OMZHXfa1=yo<`~W+#Qfa$b&3(Z+-LXX_kK;bxHUk6)6<vPNl8os{s*dpww<oM
z&Z}rh{E7tdQVwLjO&~uAu_EBZIxSOHgP32Om(oexB%hP&n#H7WLajlc%k7^BJV`$<
z7Z1SM(UHQ~1yn#eRzC-F=1e=hHUMveZqs$PN8cMst}X25l5njmNYV)bjexSiBSj>#
z%5uySPhCcX@I1w3S~S7<3@EN|uTt!tjtqX#MOus@b7p?0s0xH1?0Q+MC|&i~^4V?e
z>5lhUe70m<j)<l){IU!)D+mTEH<q@O=-vtHNy-qSx`!>3JvP_%P@SkF#E*0))jBcx
zftlGig)DfVmnT-4dt+ePqr#?Z%Hn<gO{4q%W<OdnIn7gdZ7VdOxW282V@H}d57mk%
z=2v5f=!15ZdiLe4-tgZBzm3G(g(_#@^?Pv)b8Ze1e696=DdVa^%bcYhd#?Yk>x7se
zuw&itq2eZVT#Wm*m3fPyJMG|%&Jt$axJLI~u+9^A@>07sWxVe-&?cA@t`)erHO9WJ
zesJAwGTHm=gHU2~qPAjd(S(G?Q-jiq*#-ujekkGIcz2Q<D$-moM@PO4ab3c#@XvEy
zthh1nI@8v*wCh!Vml@;1z=f~+#d34s|1-3j7%P=v$+#I3dxJ}|P3}MzEGMA3=e5r}
zd<6_6S|>B8{12CX;hwhi%%yhltj){M4rDUTi~jc=t;3tVirc^<6x(?D<wrq-6p>+4
zFvs~+y=`nqiK}$|+QI=n49Wl=D1bp_<(YTiGf%>x$X|<?=!%QfBTyiolftm*Gd+h+
zQ@4;x2g5|K{hZ)g5qI6gl9{(a)G^!4V3_G~xVZ9LZCz+KDrXaE%q`9NmH&oBvn?l-
zK%<gax=M=bg8e&o>)}St+I4BsDD=+xZzYRtGj&^!EzgbrF)wb!rJ{uQj=Z4y`9SKh
zU&0zI#fKMrCwtB*n;94toH;2C7TGMFr6J@vJ;%hX1BuzxJ9V6&RPmc#cY{AErfPS&
zq2mg3pANk)bKKEx1E{8AB^zxbZ%kVWp1X}p#X4+K87pO<W#+;|4x8siu7nWM_-47f
z_YQUEmMV-+%q2^+8*<qj<g+uec6DszQ`rI&i9)UQY)-=k6GKfpTIv?sc|m1n_zZa}
zIwU=qu1oeD+^ltMPQLtlnS9gLzwU+T>2d;W%&_Q+e_*<j;DzX7VA^!CFB`CyC&;Xo
z3RXRE&mWrWRLsApt1Qo&MQ}wHe&X(5ZJLBTaCx2XX&YOeEm)=)B#xbI<yL*#UpQVP
zt{;FokBX2Q`rXy>m1a!uYG)p;7y#`Oe+;HKHu2I25>m_4Rc+_99r(b*)nNtTqLOOY
zOvYFRjg58=Z!0!FQS0Ao8@?oEH0N$G@d&5oP5BQWe(3N~-LRQtO7pdj(p`)O@7CmB
zvzcHD17RBTOYzhe=8H*N67Q6{mEyIq#Y+x+jB2rxUvDKQuZLs^HeG{EeV>|l^+s*9
ze5BFv%M@5eSObp1!qiN4Pu0^=uR&z?WNc`8WF?2u>dGY$Ptiu=yPBMwTyC9JP9c~y
zZqZzZ(fdn<>%7RiCh^<51!@E1)`F|j?FCBJdoPQ`>LF6!MM^n{pXB1fMpwDz{==sj
zo0+PgJR-TSR}P7vYzGp?^IK*u_oj4|t^8~I91qGXo3-RcwFO5P3e<$D3CDdYoU~|S
z4&AL+M$U^HM$KHT76{Up$fGpUI;u-tnwCRP-Yp)1s8({;-5P0}&qu!&j0Rtq&P%8P
zvqB0<tJHtO%{|p6)=n{(MW!72D8)Lki)$(-!P&J#W^J#fFC!{1A=9!XC8}FIro$Dr
zO}xO;%eJggm#W8?B#ysKH?e^Y+~wAHx~&gLC$9MbAma-3bd`C%ZG?|uGO|)F_UJjc
zeq#}9Z0;$&A_=Li<MiW&fkSoUx?_3Nm~7*@cv^nE{~ofIZLZ^&E-pI&x(kFD>tX)(
z*4^k2F<#R0HZRV&{!m?F1;F8%Agxiq2|@ZUz=}xNwN#9)E@lsa8hR+e56b4EZO>FL
z`<vgB0s>Q;D7Dd@e{KjEE6*Qw(x$TJR0!MMK)gGo(+N&UL4vBh4J#~-0Y+Iy&BEv`
zXZ1qhk=uMcdNTjOd)rWd>w}7)|H8Y$iKG#*V+&wz?#SswVw+}<R4+-|2T_wt*e%f;
zFKxr>*XVr|&>Ip|<b4$o<Cl)`nR%AT9pPuWiZ5qLEp_?r>#s?{xO|l1T`jJxhQVHi
zv+ZO>MH&*)x(&8<!TVOzPFlW1MgN_~fVp#n<>KLUuGVt?G`a#K{I>~E@O4tce}Jpx
zx487SRm&7oppca`$!orl`1lV1EEfMP6bjHqYAs}q3P!S6;~m~h+uO_G>b+xKJWBS<
zr`j=V4*=@!B?)R0Xw~8&#+1f`(K$NA!`N=Vz=)`~0ZBya!L+VxHJ(Z64GAW#Q^2v3
z*g439H^n>qnvFn`GMoCp8Xha*0!tF1zrpGXklDAS=L-#5ro>*>+*a9M0vw`;0xmG;
z@9W!xA+J-ejI~xh#q&jG`ESD}dF3Vb9$_yejtEWMClGHZWX`!Bbzu?D9Fum%-7p#I
z)EIIhPAn;k`x^~pb&2Jx8J(Tldf}j0oF5C`l=akv7)sBVllI);N_n3BZA$<NxW*Nk
zdK!)+sOT)?8oX?oGDsDiT^)OhqCyiOhx?Ee@`3duJpkql=ijSmDkyL5;v)khwOPv)
zVjXOuXStu-m3|+83f+CALan+7a*n|6r)h^~VXi)rCm*5re`aMh@tPXA7Al(gzJu=#
z`SuX%pu_Dod+@{#Z&|SI-?Hl*LHHV)h<C^((yEaxk(u&tDj%R?D~p#-8Ia+T$3<-Z
zwK+1>KVkqU2SQUoO)Q|8BMoo?X##+s_~H+c8gFt%egGY)&=^Dh0DPrddj3^nhM{lp
z>V)MbG!s%%TzSY1^qMp!km5k0)gJ*O77;_09*<=dFyQ*^_lp4RYuU9E@Ic*k)fsqT
zzd2Xw%)CNSD0%+bBf${L^kS{`M>La;I*q6+QfX*VC~#f15aY(`STe#_O1CaQfoJl1
z4$7?a4i|r-_XKsm0JVW!K$YFbWW^@#gaIWW-br9LvI6Upe-fNmrEHk=%p1bK1&Ugc
z&-t9-0P|MyA#e^4xJBUTVV*`Pn}h01EKVT&1k;>jQ6S-YwH|8?eu4r^4Iv7^3ND16
znHJh<yV3<LfH>S!74&hN#w=6(HC%ml;{1=@vmhlJz`ovqM;Xp-m!EdG$`TnF0LJI>
z2oHC7>|=%+Bo`zHn^Ke|!1Qx0pj})mQ|$iNP3DgyJy#tI7(+=Hmy!mtB!~mmq4wal
zW!S;21yFpy2)6dMiidrhOUUdHiySNsQ!l9~;}gSpIx<}yvynEZH4?@NHY;TbPy4Ji
zG(d-$f+tD!@gGGJ7%ldnSx7ML0-7yRw*m6!ULXhT8}JiWJYh_y<CwalK;teeS<QEs
z)HlT2m5VqPwt52SFohK3hHD6tSWKb5pR`7vMsB2Cn!qCG=br^h^0^xp(<nw<&=}$U
zJyxMiw7}Hn4)6xOM7{(@;7)r2uyd%zTy>}<H*u%8SNq?VXGZj&v+;qh1O=?RVCBYw
zHU*!-y+N5sT2He;0i#d~rODe2bo9p|;Q^||V|h>(^U8PuEezGPv>`OBf2I$#vR%zT
z=xpS}^5G2V4kPeivE^;mqbTR`_$$?c;_uZip-km~_~SpvuNJ-DVBmLx^l#wmZ_G@R
zp|3zZXQEhf-@k*KPX)u@z64_I_<(eTs*cVM6g*RG$*rKI#BneIFZ_#ajwmNkRZr*3
zGP=8NOPm2K74rwkP3WVN6<<{p2m%?Cyt-9*go!|hokjFIDH*XLx=C5hh7$f&{fZgt
z=rr3;xoR)|#P5G(aWerr7nv>o7ud1=yqRK?a=B(UO=@LyC(oo*V-zk_O@`PqN~3=y
z826-BW$l4gN*6KB9IQ6(Ho!cB*;zHAc236SD=bn+Z?`EDZr!vS>}u~#2A`d5>de;7
zE_1DIWwf1Z@=>?<dXqOe{GhSq9P#jLy=!~1enKCxvL<?g;_M>@p9|1^-X;jY;<c|r
z`WD2M8F6+TlM39KtlulhB7~s)jG(7FZE5K(FzYBWeybH&-v4E(=IrFGc9!*;QKSWm
zDpwUwnm|Z1*!amvh_PLZ*|eT5Cy6fbR~+~lh0K^`1h_?Sr=h}F<yWWC8uY}bb#Bz7
zu|0ya=x|C;KB3;V`vX!*N+cuZfCKP=*;>-dBJbz+!`-@!9&uZ(7}VM3SBl9qAA$Ej
zsQ=Rx0f5&U0{X8W5Dn9b$AB7Jo1<eaC@CCR>GeI~-29;SH%X&}fN~}en%h;UkmHdU
zmzpYOye<WC^beWj6}vZdA`32l@x49&QPdtT@PUo{?^*iKkDQbU%dJOyVT!%Y6`{LW
z441+B=w&~Na^2oB`@_b*Qg2rnFW=*m{Lpd<3HK0{Jo9~?7}n-xXm1r#1J5yOs)O3#
z9k1lzV=Wbs1N`K5AyNatyfE;6oK(*${Li4qSZ;aPG##C|{ranD(*ChR&)=K#ECxGg
z%ZOY!hX;_kM!f=Lu0=Z%DiY0G&!qqMy1~r%0OKSK`U*M)jS;=^1)&x%egG9muDYbW
z;-u^o0;sR*)`>oYQkRrZvnkXC9)O3A1Nl&Na=b6FydL81OL&w?|8sntMelqa`ChU`
zTJuS<ACW&&CGHyv5aX^~o-<W2_fN|d7@wHR({)CWwdmeJLJD;fN+oe;iMShF3bYUM
zLD*|*Kjta_IouYJL~(xOf;x~daRZDaRV16JJHY_)Q+EjXxM<K3cEAa!aK@}+_Vf)y
z=yMz5q@=X`E5?8l(fzb8F7-R*48N=I9iU4aHz^`pg%W%!mJx-wq}nXKq8aotj?p;1
zLdRB4k)Ae??$Jfsod!%(M~oP3VC3=vv>=g|?^!U(yXu&-X_f93OAtEA>+e{>BsqUm
z3>zo9J0?mHQq>cASCG7f{2SR{<mbOG`Ehidm}f>`(MuA|>aj<w_xNXbPyvdl#}|Mj
zC!8F>Ls6MYC_GU;mMIiuhL`}-)Y-ZJ;<bzLM=HcNYMbO;L?!G7rp62~HHOf*Kmj@j
zZ$Ud+w<w)hN|Snhl40TvJmM5@+}rTqCRPC*f$pFvI5=lawvY62JCUZ(00sV)15tDK
zF5WeG<!FZm5l0p5NCfD@>j4_LQ*$ql#Raf~Gr$f6f4Q|s2RYSqauRkXUvCV49_spU
zJ95Gg#&PYXK$DH{4g9m;0NxZI<Q_z?h12yO?T9*{y}n@=^d3EngKQND(%edur_;>B
zLE1@!9);yY_z-4O0J@?XRg$A*fmf(2YE;Gk;#CBy9WV4YN+tS~k;(L@ozNVDLM{?S
zp26t~EIEpH>Sv@L12EB>hvEaTq`~<E@LiJufZYs{lRrynJkFfQgN4w*#N@E0XOXvj
z+b4?A$=t)ELZyKXbLJyZqK)q2L>ihNl%VR2ZFc=Jh|@K9$UA(YO`r-l=*>>3_?|n;
z4Pl4FL3)<nW#F+W^dt4iZZ_(s#1kdmh&UsK`nqGxx-`-!msFhs#dTN#*C4N=W*m?h
z$Y-A`(f~jv>XTUPC<m^r)7m2$dS$EL%wyQbnbJ^TL0~#ZAfCWeS`rM}b1Yi7r?l-P
zl4(D{I5f0Uq_o}ilA#-tp#-#nM3Q>MWoNS}k{BU)CY5;W%(UJ1lA%v%yU}R7Qzb)J
z_{7B5=mHx7p?$r0f;$Fnnq(+UGIVgN|0VZ1&p8I|$^-&i2=G$nd}7pVF9JP!h}Ae_
zB;fv>U4w3v8FBA#h%q=3?LK#@MKb9RqCTs3gpx#0Y5fYrRvWGbGOpS0$o&!c+N6fQ
z>A2FM$pB($qA%&!bZcFp5!5mA&Ql{&I9DX?_0QbqE&qf`-Vi~>*8r{SE~KcooPD2_
zp-z?inlgAiG>Dk*z>IRjpH9Db=R!^uGB^joj=C>n@fsnI)mLPYpBtIXNmBHuJ*W5+
zi#s9lS99#c{lj8I8lUl;gaEi@SzxK;l(Ojr)W^bu*_9^KP=9NIit9eafpq)9X7bA;
zvu_h?_I;=vgb;siUY8kZ5C=3%f4`>o5F6V!9;A~a12UMReAiLX9QJQ#eYh}Q=L_2I
zbnrl*_JXk^eLyOqk%Y;77>j-BO8fBjK@2@e@5ygT+HMPGvoU|N!tc5u7o^H(ot&U^
zq^*JusZnCd$ZudCcYCs@wCd%X5uS;*-M1)4Q6WM0AHLmTVKu-jUmd#~sNLD7mHNG&
zt~|-Sx!BjcLs#rspC!Na9bEO0+F93gb1V1aB2lLT98kcG6;*ISuZ*AK7@<!AC#EKl
z7Tf?ZoP#J6Eek0?q${30%Ey@#4$Sb=r-Y0eut?{hkY1%pUPv<wUwRi8|8A{TaZlG3
z+YGRiJW#6tveP0%VL^eTE3~*Nw#D_#^~e_^Kwu#RVDM%a5tvnJjB~-r><B{j2rh|1
z*E*K#Da(|_C(C3&xRi1d0>zuw#6gdJm^Un&nai3njtlW}NaaP?)XT1HLBhzFOi1U9
zBt=**#UHkxf(H`711G1;GCNN)2TX;vm+y2ZzXBq7d9inLp0&_p1?CNDAnNXxiFk=B
zBivwI_>zVBbCV!8uzMkQIX*XVR(wcbBg)e3w??WE3b+I5&+LgNWjW8*_TAF_^oOMm
z`;ZUJvRg_<*6SB#*HIUZ<q8{zqChy?;EF64?jdqPKVK2G@qPe@7$1Eq7c^#t)9EL}
z6(Iwc#J<kcTp`ASRyvWUKLW!DW2RE81iozS4M>jD-O~|0ERD<}C7wLKa9nqS(^Uxb
zp9CH_LdXPQJxM&tS_u0SAl&n-%OxK!Ffzji`aP4|zuxuqs}(TTbeG7;6>-RjGr;up
zza9=mGcj-2xB}`fZNInr37#?yUaDo174k!EG1SZ(oZx}BbK>ade8)7PxYNNnLXV5w
zM=8io4!i-Rp$tfbU6!Av6gJ0nk<UAgO*ScabUZd()OU5~!oTdFd1w{&fCJ1!{GK`r
zJqiiGCBX`+6kt|(gRo8q6F|Y$bMWAj$OoW&2#EId<9Vnh?9z5-^Q9i#v5Gt?Ajzz-
z`N;S2uLF>-JsBcObe4=LaPrgcU@a{`hWFv}49L2_G$R2Ejno+=Llan5JMh+NX}d`!
zLy?6Wu9{U6N+x_UXf@`qDRo{3CcOwenvGP3LHj9QTYZGJ1j3*+J9^o&01TXFr<PxO
zm*&1Dc!=rf*@s2jj!8T6bXYbGvdYpBNQQWRowP!jFE~3ndIPUw^XU<Uo&@Vg>i3;^
zXZLyl9kN}|x%&`Y9l1jSBSZIL!=fdRrb<6cAMy9y0Wu^)ItHE69&HhAj6ZpJ>z=gT
zUnN5)1M%M8%|wjf6;~HihKQRG?5_&rGXz>39^NJv_Jxr8b=X@EQ~}GV9Q!hwt;#Z^
z9!B^@$NF4|boI=Lf7#>2n5D&~F8(wo(*$ak;8jMPt}PMc^=|vqsMkOasl6>jA0B@0
z#jqNzGg7=t4uq!zE|$$27c!Y}fNZHb3v)w{sMVPpZTSIp_O>V7^6NdA;k57XAHmsq
z_2FDwik47Pl%PDVUNZD!I+kl~8BK&RXu?&8XnOvDjCrHPfFxUog!ocRIzz7MNmja2
z6~2H8+sIQ^Ovm78$6t;l0B3NBr|Nz^64_843XYuqBTCu%?Tu-vunRhumT$q1c!Um;
z$9uUeN5Oi&lpHenv2!ng;BI1olrlC(x`@PaSY0|8IuV&{1o;8HfWCMwRP_A~obvY{
zA0UdVGoG;w&k^`9KJU6A#-XJO3A?4*b30H&w877!K2neRb9?!Pg#y-Yetpxvarhbl
zImW{_shfv%53d$0eZ?WkV?IFQdrt9Qco%U)8N9X=sqYH)=fDjvbdHC-h!<QL(5j>K
z`f#V@M%H`SoH@x0q5`?hPWqtzUJ;B^r7JACFm?vrtd{}ikylal>`8N4!ww0f?9P;P
zG2I$Rfu*17<X3&gJu_kjq1iL;VPav}xhN3zvD_MfT(0=1+Uu#Gc!>p0JkjG_J18*W
zFWw3lZ&B`OJ+v{nmj-Wl4uB^}By^1O>$m<PLZs6kd}SSxbpJwalSZhWx6XFPedhpK
zXA4vzjanHaJu~<Y=7{d94t04n$NCqhrPur|1f2}c>z=L6TIB0)kOqsM##=3F031`>
zW8q1r72U*n8^{SUZsw8-Dhn)1@s8|xWQe+6uNfp%spu=9IRs|Z2OT&|>wnZd@nH`v
zr0J~hUhoT&CXVTP?~{+WZlCT@lX;`J&FGW<V6t;d@BxTV`edCKwsbTU0O!8CM~a{B
zcm7v^?cAN$2NnQ2$-cYuLtpe3^DI@z(tLi?jI+637C9}Apmcrv{EaLKu^>FE86Vh{
zp6>}H>F<DjydawLoV4GDcvc2HW7{3=nhvmTd_asl+Ym$kJ|NxZPJq-^WmbG;MyS(F
z>cKJ%G0h4d`&IIe_MV{A=WZh>!+tjX2wee#8zo@396AewYA$TH$zx>SbFCI|G8L*b
ztjaOFH?m8_ikO@dA)ug?IL+TAh#VqYpPPTR03X!y)42iV{`9qLmZQ=~{IHs9x&-jZ
zL`7c+0Rk|s63bQfA|j;12n~TgOCc%xUKWE)z>)BYjFHJRPop+-OJclxc%$KN7o$R&
z1cUxb;9+Osb`vpw1}yq@imAe*(-*MRWEVMnHn`Fg{nIM6MuoJpFOocYznzkKW75vP
zh$gv4y;&9H>;miTnDX}HXU;zdT5u2`!XFB;{|?4+(v^YyJtC5KWw{9_Q6ulD3+1>S
z9YrqcX8l>Hj~wqKl25RHy7+j^w|`vz*|&OO`;+={fG+kdd*p8S6k=@~C55X2@b4-#
zKOg2(#*oo8FkpD~Q};RIqxs=*|Hw$$_UA}PoHozT>X=?^=0HUa`&^h^I{vPeLP5y8
zY!aYbrkjmo!N)&j=cqv$6YinqST4yRzuA*tK<}Mh@C+l$EkU9?#<y`Gzdu#cL0&7b
z3A0;Wg~j+PS)qa@OYzy=^IfM8lamJVp_%l$Y137n@_fBlqGDymOV$~M+&HJL_+ex%
z!`?L+hseEi-jU1ez@99(;SDF{BQ+)Bn)-yay{eHXF|;<bTw?Zb5S3aBoU@Ju-MJIz
zS3)?`GzHGZcyWdIGgY?E@h0o$B433cAhTBtdVL`wX0Nt+sYE2N+7)X}>Bs#Me|`XZ
z?e0N(2EMZ*=#+Z$_Z0H63wb%K9dWU?p^<*eJ@W+${mJ~ie7_+$-F3`fTAGXQa}?d)
zxF<;oW|YV=(5WdmTBvhu#>(Ntrvzu8gqk}as*geykyGNpf?fzKZJ^bYd3pP<19jJ_
ziTj>&{60l^Kkp6_T!mMQ1<GtErv<6kMTNQr<8Wg`Yp8)W^)1NoN|?Ns^tP6pzMh7A
zt;mxk@oW@)Q|~mfo=Zw!^P!F)yw!P{=J+(Zv(ps3^@D*3v$W;Lb)SzIhz*G|XZ4B7
zVCRlJDB}laX}lOW>Y24&LG`O_nu!W}H_G5Y4bkBrfak$>Cx&J`{Pgg9I{U|XmA@_;
z;X3fsJm-%ug?S$u5VB)@>)FpPd=2{eDc?T`5K-uv>9Gcs89(pi)Rd2_1)T?MP-gBS
zc8vy_evP`&0*hK84aW6n{=mF>AES7X2VF(}HOmPJk1}0)rJZ$SQ7DCx4KDe()f!!5
zWf7(sS_5^;EC_oN`dV6IL=A~DpEaDH=RQ^<0R~0Amemdycc0=#>X{=!Ad8#m$s^vm
z`SXs!Z?;<d`BW<%M`i`2QM?W8@-mSvC#Z487b=WAL~yW|M<>JZ#c{gMbQ#xk9PY6Y
zR}4#27yETAW@~xOlNnJzukn=sZd$x;ebkwI*Eg}CZGM_bRU1dZ@e2f2MZyzHa4BMD
zMV2em;BjrGUzibZt_3ikvPKj(&x^dfogb~gHKA}jlu{kmmB$z#?eZbrMF5auyjhFH
z`X{mZF?%Tx{X}zbxRcmm&x&yaQAqxnYGXKlIbd^W*k7kK-CLHyJt>Mv-CjAcn7!LF
z0v2B_=Rq}5E2E`|Y|`}moGTF<Sf?lXOg&sWH^g`nec4u3p5=a2{ctY3_OH1>lLmn%
z0gFV7`NliBb7c;q|GKuhya6;IbjD4#$@g~mTNV@s2d2W|?B&%~WSts7lV&Bn21U>+
z(b*7}A2~~m2Cir%CW9U^)AC-rUD=-+_VYi)!n@zW30o5Lqeao*^bfIUr!u!&^o3Te
z9V|l_Mu^%3v1qB090&vUA8DxYY);YyGAvpzEZTAsN%gI6*GEjq5tVp0n0VR*%YXW+
zcPzbCGo1U17--UDBh6r2AYLLC?X!jpwLrXv?UT@BYiRgOwwrh(lTDGi#U|yLU5cO+
z!=ZG;z10Q~+6QJ$Su1_&ql@PN@%>YBkw^DNyZtVBF-SZ=cW#}9)qi7<C)!>Y*xj6_
z)8N13ROXkDr7eg3RJ5lY(7`P3nX~S@OADNE-SgdP0UDSxXEB@!A+J7QMfB`%_zN3H
z4boSx#U!vAzjkGeSXfBId*;!yh0Wfw%5H6p?1_mbnBhI3pq=sWjUM@uDglGWaP^8`
z2`A_lMPwHnV$yaWUfVgjjtJtsAq05WMghzl!iX<n+%{L=#L^S{Ta4g)!Wql(B^mz<
z#fiRSjdZVr_sfxO>gmdP(9e2FxFHVvnsiSQ_UXmPZU?)dtQG^emjG~a^#eBSfpk$f
zBZvQWVarBEanSLnbJY`ofxD)l8TVNL^0xr%D7}|(LyS++zZDr4A7-cc48L4y9JmA$
zhCCs~NmDtN1ji-ar$$@>lK|sfa{G@XQh)QPS}!ZNy}LhVr>?}-92gX~9H<olng=HR
zrOuN4+GXU921yhU3g-nUy3DPY-Xv`8o*SUwsmz2lMuol5bF44G+(Nj8gbq?Jl46?f
zC0CG@D0Mr7o`J?bXc9d2)$9Xu3G~>dFfvkwNa6xNoR^0DR0NMkUE`fbT7>lgik`kZ
zNXycl+}zfnQ_F*P$Z{X7EmTMmj2!-c>;|<mt9z{9!x)(lCi}PG<<*8N7n{yOd(k6z
zYGAnlz)d-Y7_*I8%HQ0~T{_mUHKgS;Yft_8Q5GqS>@hSLToVZg{tr)3ND{=q79bxK
zLTd6&K;}RIMLmF-;(cuXa0XhBp}!=c1(4fT15_Q%K*+t~4Jgnqjt^uX0S_FW1~G)I
zq`{ySAdjFwXH0tB8t#jK$+)S(HMqDn!q1#xCqV+>fjBE07u=`(Bjxb<J1vd7<Yj`-
z`&baIb5`;^=)Nh}EAfu6B`jzozW5RyDD-IU_NJS2Ab&VMJ_Gr*oCiq&z)*2_Tc%_#
zUI3L)q3c(=X3E&gi!>E_W@)=zkv(kiK;FfAHGufuxMrgdb-HH;j~hMuR>|~xsiF;Y
zU9fFe5L;6g*U0z}L0M{3O3xqNWc8~Ak&X^uqH1<MqR#{wcl;ozqq-qSw_yB7lCQWE
z5N7s_IosLOz%pM`1^;Qk5d}1e2WxH0z#pOk6rjR2PS6l))^!u*ldO;o^}08WKP*$O
z*uu7sRC;k19V*RAc`zuuc%sFq2b!X71+$z+fTGqT9XyMeGZN~f23XTQ4f6$lwk;<&
z5kqwK5dc)6FC|iN04uw9>POERg@#A#=Oi4weHCQmx&N0tmk{_r+__UhJMPE@LA<bd
zuP2XO0sl@eCqIs3stoK6)h_SwgDJBVL~u}0_QI1+tCyP*;M>Ie7(jL(h5GWS!!$<z
zzBthvwKi;7(9v{-r2FC#I6+V4KkPAgbgPj`Dew5`|Ka+@Mz@WQEv~3<;)N7zK90Dr
zV&A)sq{SiCPW!GDMZ}_lG>Q(>0T8v!_IV4_`MLjhAS*+A!<pb{^Hfu)WVEXI{jdJv
z^j}WVV(`8wwe9s{${Q^fb6SR*)|?}JNm1C1k`^b-24Vk0GPK1e9j${A*4o)uTYR#i
zaCfZ5ywRGX<06SctUImTW5;aP1<E&1duJ5!8p!V5N)8AIS$l;%&dJc7R*WLQlJ=AB
zZtJ`Ep7`L;3${cf3VVfAFG^T!xQqUDxO(^%g+9i;$BIzR%LHmjO^j`cge~B=ag<Z(
z1GnsNzLpA`XN-Qqk_!TSy=cC$aau{FW3*_m;PI1iZk!&WC~lk&K-irsDlicn32Qcl
z{8$zKuYCbjZ}@Ydv#kJD0mta`O7QiL(g)7P*N-*qk8zDz9-$`|oF}hGAf5jh>|4NK
z1z-2R#pE85|Nr<nV26?%Er$wg3cx`%kNA2A;<5ArdH4o3D5iQ!(y78We+hu$G@lGH
z;FSEtKt5DBMoR^!BoQ|i1vtYvGC0f0b5nKEga|VDH&jCNjnui7bf^Ef`+f=(w<TW%
zs%&#*IHx`^0UKvgf{v1Js$tMBu`*>0ZmLdfD*})qC<&_*l*RGl)L|Isw+Zvc!*m|_
zga9E7QquIzLc09rn*mxcE!A)@;2ANy?bg-?I5ppo0gi4|lMkya$pKe<>HfFA$<{cK
z>4_YRob$47g6tWvj;_3oE&F!IuU{Xb9AuWm3Srp^CM2(GrCn7;AT1cQD75G7p}wT@
zzru7!9z5aR@J8o`Rh~HQmC=(WM8lHvsMtlC_o&cCYbr|FiYwEs^Jn3MaeJs%pK6CA
z9zJ`8Yiq!eu-)!fv&U^?6)EFi@sK{h%vcKdkCo#q^KDY^e3w_{ip=GjGUW@+XJ^Li
z?k^opY4*!@C`)G{PlZOOzNPF!Cp%O&E@V6Q4Nq58h@@2?zsxGLf6~EhOx{Let);>9
zT!vZ;OSK8cjhU}}m@pGZnofdsA}qx&uysb;VrCo<)s(0jTwVBLXawpR?)^UtlZe~w
zZY?+gxAgkj(DsyiKYdwVU}|=l;|%4)h4Z-ALBx$`rGX}cMaP-Y#uO^V!~Ya<FMWvp
z&&8!x(I|F;m9|L*5#o*`QE{ZkbY8)sn*`(D1bqRaN_07AXQ)k1tzHRl`r~RdyhF6a
zGDEz*qGz1X-WS2Q{Mr?E8=ahXJbtu^PN7Io=Ua&3*8X$3Ej@&Dt^XCBmpxN>&fi}X
zG%AC1|6cV!ZI@Ps(nji#F#qS2e@=Y-`OH{dEAQWT{`cQJfvU!_|NZOVUk31Gf~5bq
zPhfgSM*RQ(2R~zaV)_ZSSMX{oL5)bdD2=@A-T%_0=N>4%&lqzM#p*{8No#qMR+L^g
zMJd_eh{&-8)-w$-5d)XZHR?9a3U-;8=`YP+pf~^X)}aw>XaZ|<SfTG9U%?Cf)fKws
zNe4C_|KCpK;qnX#^-Bdiy_4$m?wFjAZq2<2$YB~7ll~u{$pd;?=9?L+q{!mwEI&9w
zsH;ng5g}<!IQe(KSH72E`ucBQ{@>kL{Y~Y6^z}32C#J!uy=g>b&i+P2i=Vj<QF?C$
zxHf-xCcP;7^u)9tJJM0=H0|Qz?XjT^MV!Olug;{+&ZN{QrUsaiRiC+?kxlbad*4(0
zu46wn`^J+IvPyx{>w?lNicfYTeoloi^69EG=^e5u2WF%v8`oxTXOaN2=>cY>5G~hc
z>T`rHiYx1tc>4EGKBf1;usrLodo0${udYu_n~_a3q(xdu$X;tNuycJ<%}Wk)9v<ko
zehPi?s?aw_?-hM<p$RlR@pUv6T>pgZw!y?m!P;mj*8;a6(b$>9Am7*?|HSm3n(Rcz
zXsGWqcLjRy63afz276gOW_k%T*C#onp_0Y|*d7ACH&V$vp%vbM5h=vTwb|d9B=nhk
z^!*K5?`>9R6Qk@|Vw|0BMMf(8CRJCbxBKL2<duQZ5Cd=ydhgIXeFF-zlU%5?RG;j{
zL1m!d+xipJK5X1bA#ScsJ5fej@eR+kK!mHaXY}#TJ7(N6t^b%-e?W4`>zXdNd!r)O
zPZ>eh13ufD&`9qs(WQ4{B;=L;PN#bo3@qU}r-n~VqhIDQhHhEVk-wEPq|d1!7$1>x
z;lLN8dlF=&|Hf#DPl`7xG$qsmolYk<)C%9!@+E%1k2TQjhMVe%C#HVyq%QEs1?hC+
zLcK8fpO~K0<|Kv|riPv~<kW>(IMV4Pgv#KLqtJavy+@84!PmKA##j6?It>nFT->~k
zCW{_jWNF-d7q*2&_k9#HhBJsi&M)=yNi7AkX#z*mtP=2hB|qIlzqO{2-@B%cSM~Mx
zb%E9U@1=0@$GN3yaTeYkTgLK480EjACMeitrn`QQ@2h`Lmt&t5{U9Yp2c2)UNah*o
zqU)ynziRsGxF)~%eI=Ao>6RLu14gH`fKn<gA&!t{NJvR24I&5(=@O)2bd9c&Dj*Dy
z9MUj)#P{+2`Th3q&YsuKxzD-Jbzj%@oRbF!xv}R7MLGDPe;A&GMQ#{^Wck+4W|3*6
zksG=o+3?6IwmdLJqC7s1YT_;Vg76DRQ=L{}3fC+xwFdGDOSK~*klDSr=c=G)F_HD{
zHyNc(g;0FZyVyv^C<n-qvcy?3#zXFi;h6zg8=NB=Lk;~z;m=^zE0fxCixYg9Mnt6K
zGkQP$g9voL6D2kgE8b!!jj8x)cYV(;g?Nrz4N-F16o6Qi=_hT9)K)5r?L}A0ViuT-
zU4~FfeZ%XoG3eDGea`>lo2&@Tq>UN8ll&Ty`-E64@(JTAccB!@uMty<YlVRyBZFdk
z(Y6QKMCRAbE%E-g9jU#TQZ+9rTI^ys=xCIjzMNFKH*$QI>>|f<iUU+_M`1JiRLzTw
zBJiHgxTd8nML|^Eh&-rShrm(oh><`L^xlLOFO;-LMF4TqL5!hX8r?k;#><Yd#WloV
zr5vY|QXZB(Hle1km7;h%_>(?ZK}YR|Jx}V>DT~^X0_Xx3(yBo<p}t^NOT_A?!32Jw
zJbMCi2i5}{LUN2zo8T_98iIg;_nCX}&!vuEksB&R<dZ{QvJ}iw*<(}_y=vlLBU9M(
zm?w-t?vybZPgR|=c`f`DWXSMxiTos8dyXrkBwsR|>b7PW-WWeR0}c59=Cn8^y0i+j
zS1ogDXEKwk46R8C?U9$uh+O}(;QJ7o_->Kp{AZN9wL^JF%9(ipOU$+M=2cHPR*^Cg
zJ~J!(;FLt!L7e!wpIleV-z5a2^Ox&?d&@oXl#i0cHeF?1d1JKg&`ArLI2LsV&?ONm
zUMNY-e>)U|Co*rDV6Qm=l=n>42_rxOq`z48Z6P*J3LlgFysO+gu=YC-<~8CN79gQZ
zd=st&<yl2laSaLNH5IP)M3t9gUM&T_5R_$|zUxv;9M0MDVQL{MQP*Pkik0FerI{R0
zJ48gI41E1jeeVRBwq^Zvb%oTYolTG=)|8HPg&?H+J8_Bz#q#pjm+@IOA8%9f2E^lY
zSp{W-7KE0A%f^>AUZQIa)%S^ytz%6%uV}VVAQPOr!;U)f+gd^b-pmkDsgG0bSq42E
zY>zIL9d_<pvb-dJI6^l{*yqe(=C9e+q~RZ$2d83WYH)RZ^G+gVUB6WoOx*dSsBnto
zu;i3L9c29wj3;@s@bQrYlNJ0&ekT=ml(kpF_^f%W?V3y*Ev>->1UDBR@Gi{OqhpuJ
z3G7kFv!+Pe!Gi?3D;lt}2@|smzUz3@AwrVf|8c6m=w}@f$!XH%#l}U%LjZXELqK=#
zYRyXQIn41x-9eU$hxJWnP5;}5@Y6N)5?eVCC)B|qatNbvo)(ozm8!GBxXjNHf&>v$
zx^bk@;p;_3<XvgFgXzvdlTO?!QFOt!RQOa@DkpE(wT?k0f%5nR;z5@8Y~ihG!W4<q
zse1A;l`rsYjZUZt`^s)NIIlJ5FfEz-MlJtbYR@5>J`^TU2m7lOat7(CCHf=qKNF}A
zxbcT4fHF$J`hEC;$`psbvfT2Q9+zxAwNab8;b6HQLiK^IhoB3NAXkuwA4xJ_P(|;O
zxA5|fiqlx!;@qJe!rQaemArc+ZRVh%Cq9#fYUbD7zJT(dH7QXhuXnOIXGqIk!gQo<
z>X2#2{PNR>zQVe?I^<?J1BV@}2`Rp|_y3!Xg{cRxGCBLaiYZ|tgGw~%lBF`3l!^Km
zLw_tV=n}&EcF(}7af;N6<vt4Ygj*y3Zw%v*Y2fZjGa_5k1yZ8&%pIX((x__R@5F*0
z!pbQEM+Vz@7*(g?_$#VzpY8@|08fQmy9BCdYHF!^K9OeP9;jJ%>E>+Db9fuPo&AZk
zQPp%!0eggaf@iC?k9~HT!D8m`H>JfRE~(o+{Sr3h6YQ|_iw`G^ZFftn3ifX-p&A9S
zhvJI7{kui%P;!rh-iEiP!;X@+yk&Tlz#rT_A}0S^TO|X9xkSY$!ue%u?q&`yDKrU-
z3tBD)B=?NN6fBPU0wbDp0kmb1to$CfJEk6>CGh`Y3R0Wbj4p%H3}>WKXmO`dJXELo
zI|Y3$9xc3sSO;GGgrgN!7UC=0rZrfBwti7_n*eO3Ro!mNvF1*3xc%AMl)6`dQXp9h
z&z40KUEuPEUkFt5h<`Zq5a>V6!Era9-{Z1$V1IGZRjWG-vJH8rNz1=T99u$Mer2WI
z%w%{?*E#qP2O%Yvr6@n2h1Q|OK1{XK(qp+TZd{({T$*BrV!D&VZ{(pQv-=M<@PuS4
zT4J$ZwBN|xQu=6WXrA1)<q)xpOylqjf36MAO#uKObyscdK!#f6T6n7DL}f)b8HCBs
z)q1@Ib;%70Mxr2T=MHw)hTGCwbA6~v2-d<bC<C65@J&Ys!;cD(g(!`nU}@*Js}MMx
z34UH`+TY;geKXTiu`~d)wNuQ7XWQ5fque0aa+Lf>c<EF}y^nY00Loqr8-}{1g+!n(
zA3zqvoGx5P>-PFlsiIhHaf1*1oWf2q3_d*~Q?K9hETldG(~FA9t0)+?RAp{QqJo|F
zwk?I1&Ri(o-^lFkg|!Cc??i{&YKmfI8h1}8u02O1($3%5DQ3Wba3262^gn6n-MgRG
z6b8>mV!3hgk<7L75Y{0SohVjkWBT>I#u&Js^|~8vIg9~6&F>-P=5FAC=Ehc5()D{d
z85ZYzBsUsk{4;qxAt0AEDQFRQbJ-IbkmXH!B&GN2366ZF=L0I^Z#`ED>z8(7+r?)C
zdYc=ms8*8+S$;yhg(P?-bUhXB#{%gXLUj!CozhpL;VH%_dD6djYiqxNI*z^KM(`&l
zVxz_F^*S?E{1<Q8I5;bG{WQ&w60(=*b1r6Q8!Hn9ck8Z(xswelG~b2TH(5p<CVDDq
z%Gf@H1oMC^4XxMT!+G*rVdJjgJm%mLl)!6fKZ;|&Obr`;svlvv=VVk{MvS37v0K+M
zxK7YFHoX4p>w03pu1&}KZ<=S7?Os7RN`q9Kpy(hvr4AbUX9x0}B6bbPv=*-Q2g}uw
z%X8V08`RD?-@(+~uoM3lhb|<234T<z0A77dj~Htg;#sQ@sC;`u?gVc*51Cd+5fdRF
zle+=Aju-06lgX{qqZRUr9f<qAl3SO(T9B!wCJ9}<SUBxN(x=uvb#`f^ccfDn{3>oa
z!1rwRj{9JS*!(ABZm`DBX{y5^TOqfbHt0jEGnsH)SLCmnB%`>zokx7Ya!u4}$)=hp
z4G}d4@UWPO9fU&gWa9XN-@om4s7r;`E<W5_pGw@FCETdXY#%HTnwbsEzhTgT=>C^x
zj!-2<{viLT4WG|<A*`RgnNa}`M{AY1VL!;p4eqq<u_^Wo)ZCg`_?=1ZEHbK5qyIr`
z$xiJEbf>OC)J85LlErp?N1$ZxWOUce)Y_%ElFDi%!dC2*lRG8>$+Isl`qplJo+>_-
z@VV&Rbt=JMwOfz)u&QZQ*M{v7><+3G#=3BK@+88h%a5Cz+EZaU!Z?(MWnACN2j*|@
zgZ?lpH%^ic9Ck)>jcp*L5)_=(?S30mxdmRjD15-Gxvxb`?stdc>1GSRezNl5J5&IE
z)>>u`gEwo|9^v&E{G9DfVeb*^#lx&Y`|(O0_vfaF_SUvq|KqY~*|Sg8%UhQct6YkG
z_W23f(bqxLc&8UZ9ab8_2P8pCN#hA<lQ$@^ouYm+GXQOmq$Kw1g2^hZ&@c<0m_?x4
zSGN!-<0svq$8!GJDW%70<c(3@{@MPkIWf(Tm7-Xc!?3T5M?dutFOUSX_;cYmvrp1w
zsrb+7IOS~Smp9@L9YI!>m+<oxVBp+<WD%n>(3_?@;Csg1m$sv(ngsub08&J{fsvNK
z00nb&wYTi<OXBfX9)!I5r-^YO8a(35t)jY?NAUAS;uQ<6f35pa^S)E@8h+nwGXfY$
zxxW2r|Euh8e>wyr4`s@5F>rh4u_zpEf|rn+)lgAT#)qMc^Dqr)rQeaHixGUbZ*gb5
zv0Y^NDh{Yg4JF&j^CNS58sz!SeSe@QLDf%-kiE0q8@QVg4{VQ+6zeas+-d`h<(F&B
zDb(3@vjxPemo?J_eHo#rhxi6O#wr7ETa%ffZi522@V-8rh(;yw6iPlMOZb>Ex45JR
zt!QE%a00L>mPQ@lw&(5dXcJYw)MAS!m}O(JzZA3qZ0*Mhwal!Cj+7qGa(I4CMbz)4
zI%*TI!LwkaoR{YoG9F&qeh;qfMV>7+nT1c$y4?N#eAFF#JgtXhT>NEHiBnwgS1XnW
zApR0Etpk8>+Y9K)&!q!}$x$#}sd+CZH740jKxA5-&6%KNhH!H|MFe~5T8>!Gy1))S
z)6|3*=Q<c^1s{gWT0Sgv+1gk(X>O?6`+?Z%#hc%#kFQXtIvKnLVZkISx#|t&m*;Qw
zS9=a~I{XU13|d{|3OP^VJ-rT1Sm(_adN%n}g-BA&W^v8Z)L3f<Of_-iH;vF~+OaIy
zu|39!Y)tn}4lHk~&)i^=vG-07rIo%t0fXjS5jcv!YFqCt+>Qw?{l5BdJtQ36xwt?#
z-!CiwS<SKbJ6NGaDY|fxc=nY}(e@%~%t9j95{U1whHs#Wzedn2iL<Tn3-4zhw{9`W
zXsIbbzZny_uvkBqBj3=(sf%<3s-(vmC}BT<2{#SvQ&$^J(hbPd)NEH9_%~SXdPjM#
zk5i)yZHSPo{Udzmv4dYVIWu5JA+4*jQhJZEI^dC)1K1H?Gmz}0j*pWD+)Nz^SN2&>
z$LKKTM(mIe8mq&MA*r9&HbD=5?HE*<^dP<g$ECX8y*X9Lg7EIi$B7@7x#(}gR>wrC
zPBNncLG)&PD`vpkS9_jY+CX@IxbV6?9U0M?{yEj`fZfU~s`oBy*h93w`(6TIsX{Bk
zdbh>~FqXNzsQ!czytB*jH-3MTD;_3MS65r5<Y8&>y0oO+MS}%HQykXTD)!RsnfWl^
z-^+J;fvFWG<qC|BC!Kr7pJ(%dpA%mOKM1mlT->3t@L@V^Y<Oe_w8;GWd>VL`rr-3e
z$sm?{Cu4eDeYPEFZ3fO}fepTK1-|t0#V{K5N!$#}1vO+Kn0y`gp3DS0P1a_8a<%{0
zs<CK&T~z&0*>cWq?ZqG@8s|EnK0Dj(^v3DiMWE?Zmz}+LwJ#$Mivai`)+nQ9(ZA}V
zhB#FD{vlWflb!*y67lmM9XYyW2Ztg9Z;}xJ_5POcFTGibg8S%Tb!r?yL`FE);qIBf
zmM0B%d4XJu79_YHmpOf+5@hVxDVQ2s<Uc5aEiw_|66Bsw^mM^yz=C~7=T~mVP|FZM
zGKbn2@{Oi&Tk!iA#yuWu$#qXE+((%y$Ug05B=r`{Cgb|QWW0x`_P$DY4}%H+`jQ3P
zeSxvC-+Nh+caJwq$76Wi&2hE>*MrNK9+mN9S%U`X4h2;gIrfY#5nZJB5LX3#QuGYa
z2-1_mOa_!0$2Q*L$g^z?P)l>yk+;I2y$FzBJj$}*9VswZhFb4_k03=K{}5j~EhgO1
zBgDJl^$c97+y1#*WHwH3F&>_}S>F8Qj~E91lK3W@a7#aQ-<f`hMbH0^U_rRc<gPg;
zL(o}1xn@rYVg5SG+dae+6X(IYRH@z9q@kC8^buIFq=rK`HLzVW;+6*q>%lI}FL?m!
z^{Mm;u^zETR-Kix{sJe+Y4-sJt>YL~;F1ALEm;3Y&Ls~&*=s~0k3vMd<w+yyRGgG5
z{<Lx&oI3+}o8j-&R)mn;Brb?s(DgH~*$JWWI-c!RIA#7#aDvxa25Y5*itP#(`m)y^
z)rXl45%n|LD}|c)oV)ISXB_*EcTw{s!GBRoXyvSH%*Oc!6@JC4VU=!fR%4+qTqc41
z{0Re41ry`5wJJ`9%)x8SR#LZwZ64)yfz%n$51Fv=ul0fc?7zB*8E_Q>X(wiTxH~k-
zYl{n~DEdbxX#t|PH_(OC)=tH^>WLemzthPc0jnu)%`jmnJiqVuoANacwp702g*)z}
zZKO`2&@{h7gcsp{e`9ZHFmZHV5BV_aEXJkjEbuVpKz>|m`=0S<_<`5Od*)U;90R#G
z7hdPhkBy|0QNGz1qfNMHCogO>qG0mj3<?fg8}IJj2;m7NDCU;uJ84h)u_2oma9b=G
zi-bSDkLgBuq}MyzMMj$v$Y)Gv!r@M^_z|WRP{TK994yk9ruXA#C+GbyjwGJEpW7as
z205;MsqbJ=@)4$NDM{Ap{1!g}R3W4h{?;=5&NTG3&Ki6g{)^wZx5?V}{6Zrh|7QPn
zWorGToNk~)X{qIU$aAsYxKA+~Cz3tSO?j)+X=ta~D{*;cGep$<UCdRKC|it3Zndc4
z&oe_6`2y({=yuEfXs)hPd|Q%!mL2tg-tsU!+jl>p@XF+LMi%xkxwe;n)son(?&-EY
z)7^L<{``zJ%H#X$Xm^Z95f8b!s0g=IP?Mo7<3nwO<PW_9Chs4wnC=B@;e129;GGFv
zck}_h!j3#XkU<y?3b?-~;kmi`G2=>)M(Ej{>(@dCb21T+{kh8*);Y|^PyulmCcl_8
zT3t6bBY85+Did*4`)#QwYx~@<;F0TxCz0$Py6Dg4jE|iQLK$_;UlT_AjAo|S>ls{p
zqEdL;Exyiu$x3&K=|aVDRG=CrczD_dxQ+)6SkDV!yKNr=U}t03YYcxuEZq`=9`g);
zYx>gP#6lFadn~`NoL&DY^=QvKs`_)h-+%L!v}jmJ4iY0Vw@xp2nL7Kpn){J+Igeq@
zSq_q(&aK0+zbO&8&Jn$n;oqDb2KRg0_>GGDzYO)lAU2|^xawY6<B(-qG=-4ErS`J6
z3}%_s7KoxY7$_l?tx+#b-ap9M9w%3_1?UuN=C!LYuSYh`xiO}C9bpH0f@ZX2<xkU%
zqX~CBud1#BZ%=4wdECagojOSt6Z=;ZTF3I{RdCQ=JjPi&xX3Ghs{FJjQ26|n=~3kA
zl~0Ih(r3X!Jlm95Q?G#Ad`{M}jlZuL*0C<4xf&o4a<B^ZJ$Xyiis1n?_s-bYCw0SK
zGF=>zRd0upY1tpAu4}_r+?1Vl4HTIu=pnL*lNw(D*N-UN-n0l8iiC9I+9VI-Ftr9W
z3RKkWn@ryUW|xBK8F4n&b5a1B^WRE+oRN;EduNxm?O1(8W11qf`gLr*;P^-38yXR8
z^Z0YC2ITP?;Z^{<ea9Y&;~*Xehv)7{O}HEs3`k4rp>cgtp<NwDY&-Z_RStDxqc8aV
zjy1<C@;o=eD&j-%^XDddqM>LhQEVNNr_3)IW1-HOc+>P%mc{;*k0de4*T^RIzIXx>
zB|X%oQ1$DOsWnQO!K91Ty8UzQuJ(4jw|fB)k_@@cRP4oES$j>M2X{z=wL;Rzy*EXS
z7=a-i!{6_cdhhE{%I>vBJuXkC<Z~nUYbH<3hjBq}4E;D-BPE;l>}2=t=~6DiH(ML&
zv-(z2EHTjyqSzG^G{t@MgKn$W5AuOL`n)g7U50-vYz5y|?abW_VGXeUZ%xECL%;fe
zrZ!{qjiznYlx#r!6ql0VG80^R=`7y2Hi}HEZa0@R8GU$n>-RNnPf#?EwA20ACJuym
z$nYsL?P(9Q%g5n*PP?V6VE^r{B}E0q$rohWH_^SLm*U1^+)B4T<Un&G-x^t?5<$qr
z3XFh>CUB;ZLb`Pz=Wg_A_<k`v;$+0PYds}}<{u}1dkFIFx&HNYws1L;gUSy-@NHR0
z%6AMJLPRBeHI*nu>-Ge_Q^v?`?5&?bUKNvI2>S+bBUS;ORApZavQkOQ+Xmzm@5Rat
z!XQHZV-LzRSDEGTR7#4k=Tx1N1m`ML6y>Yy-i`spnbx%~I@{Q2aOfMao=VjPmkhw(
z1UmK;_IV9rY0TzUm~T~3CPXYZh>Ua?f7t|NN_7KTTFI?elZ}L+qN2D$9k231REz2s
zXxP3~Z=pxHb~{ZCn!<P2imJ4K^*bAuu4Y!Y`7;8s+`4u(`U9`@4lrjdnT<AukG-Cl
zBq-y+5y*?0!L|dE-8e|nvMgV;A>{|bfHXgdpjbem^}1G|=c>_CcX(i699WtcgiZ(-
z^Sg3Ux3!9*{+n9r*Pi#f?bd~@i_tlx|FvDpjWB0kc8Y#QnV(8x>XbHyxZOO_bDh}K
zRqJqoYB!~8ZRVs$)X6rfm(h`_di+9;;%%?3<39!L7clzQ$h5_nKllt;cnvS7ZEb)%
zhb=T|znzX%wKlTsnGWj$HL?3Er>*C1s1_N7WX+Jm{oQ)n5c6Lyc(y$%S5T9a>wC8T
z#^T&Z92H^eqwmOb<XPxpgy?$-p-9aNr##<x0w&vyZ}rp7F>|l9wcVtoB-#vIe*u$h
zBy(n#Uz_Qs%j)nd0;rsA<E-gt|8CJ&O7#UR%qk&H;7EGM1|6H1h)R=6f<CmG8kwqr
z9iPdYy>S1{)3uZ~qTISvpx%kxzXx(SBb(i1Xw<;*3pQ)`VGj$W6Pm)Oype|(dlMh&
z)X4Y;u}u`wubdt!`$LIf!qIDdQzlBEE+Rx$0h!L6t&uYIM~ko1*^eKa!ibX(NP74*
z3$JM4AG1Rn#GqYP&cSOwE4QhRh$GU(SDz7`N;w-eegS^+2+2ctwk7&Lg1*R@)w@v}
z_Yw0_26q#oty?=dd*1wilje>Ds`Vjl2yNK9HkEjX`qp0#3_9SZ1l6jIdhs+kw*gfX
z7_gJmE^U<^;!E@Dsh9BO-C)+FPF}^(C$d3WY7wH?rO_tARQ(cZj9OB$jEfRP^LREC
z5iLeP2!nEsg?0`baPl@&fuXOM4{~9<wioD04ylgYqdA*ycSR!Jh31h(Za5334`g)8
zV%8TUP-r|M%(aHhY2wWT-WX|xWd*2yUDG5W*JFUa!do2T7AraP;DCqY_UwfTO8PVW
z)N6G9MLFyr?U+qv=5J_?k|&CN!%h@sJ{v9a8X&_3r@sZi_ZKkGo*^^u8RMX<$gT4s
z?2GNkJI{09lO5HOqnpA6{=KLt2ly6(zU`<382zSG$4s@o?R;)=)j-1o?Pz0ab|cZd
z{v86|i~O|0A)+8&8T<+&<EP}~>X<n63J-t5$0<uHdX5xGaG1W62bxSvcuY|xW7(1m
z;jLngs@?xn?WOx9cbbGhsL+E=9{=*r+a|MUo{*}3)~RMBJq7B@l%U2W|2Mmxx(2z?
zH77!H_-b~ETsJOs+||l0o5*I1EoWQ6CJ|Lirs|gx!PPNFm?gKZcksj|PKek*|KS#K
zS~CBS)btQQoPThLLY9_znB9Ny{!Wa^??{n9APky{as@<59g@em8yMmSC>#WSL9K}&
zFPcHh`GNOW5Yk-gs2iO592hBc`WaU7OXgJJ31vr2D%>HIzpJNqPOc^5IBc1J$lA{M
z>;X@s`0pVZpM03BHzzNiU!dm^Xqq)N{o=}wA||f55*0%#@oArs7T+aHgeChG1)?0b
z>wrPu^BAThj(!xJ&hwm6bx7%2CMwmqD+Qy`_&NjjBNNu)J?cUrwD_`WoYquesH-bD
zO=j$A)jyVjJGU-4^y`nLT#Iwz`mJ1(YF7uBgfn@?+`3w8!UG<`^l$v)fTnO?Cn6Z+
zZ54Uoc!##Ta;MUTIOI1`tDYiJXN1Y4$ljil+`1LD;^*QW#Bp)$(8NY<ov$2#TW(VD
zlQsxZPc@CTC#qQQ^fV1Oh=}as8bVMXKCL^sU*5)K)P>7+$qo@4y%9MLyhY0{5I%LC
z&$u=QNFQL|8;3CUVT(qSAm(T%d*pZfweeY0+L&qhR1l+?zhD*zMe9O@|7;iIPNn;~
z?-T9$5z_K6#{wSRywI#De<#}A@=R4H=<w0{`>R*IUJup{X?NIC-Z6cc;_iL8CDAj&
zLLC~f64<d~O4IxM9t`Q&cvltih3Jr58CmzvVW$Ik$oP@*jcEa|wt-?#o<)}1)fb#;
zT>MeHy<-7wl)*SpBzX+nUWF7|0uV^hp-6Lasu+P`y2mXx+tS&2JGxg@J5-&ah(@)v
zAr#M}A5d{xX{id4$HjTPzX3!Kue#O&RqDlCuD1Us+nn2zDpb1C)4eFRia%|@XINym
z2COA0oukB-A1XQW27=vlnXps)htqEB#0ytoXS2S8pnz(j@A_HaZ~MS%9{CDEq-=KU
z`3b81ZP;}HX^HPK?Aec&k*7~TEMGsGfjyah<=rU{=oSLbevaR|U5sb1u~8eB8S)?P
zcfcwRE=8%4$=K14$NF#3v*WWW#|!A6M^e8@GO&dU(Mtr%f7l&yfRKF}t`6TDCQrU%
zGA~&haDjncmg*yJt+_k5X*=xHx(E6=%V3Dh+iHI{Y7zSd+)H(}eRiTZ+MJsHHGMR-
z{7;h@R)vw2E)_2NoqUW}q9}1vsu&8GTGaiR{~7^QI`Bq&n#}vYN5RXbmZmk!r`x~A
zhIZ~+Ug|RRYbM3i)liCJUyUCnq%rE$<jhk{vkPUnJNqw!%JPu0Rl>OZQ)4cVht%qW
z{d-ACp*0$-$Ab;Z=5L?c2~RI`)G(HaBTH|@8Z<+WoK_$Uiw|B~o({EZ^u664@4x*P
z%j>1oRk)Kv+J?x6Rr2I8^`f+db!&VFVvHLt0sf98(Bz~`8nX~?<lK{fRh<E3!)P&r
zvs|AjGKVmMp+EoGPN<=GCWQ#IyR;&+m)0&!<1S*9qGH%ltuLcl=RXXF&dA(9e%q2_
z5Z|!^2oj^7<=l)`!^=2v3_7X$DyStG%<*4$w%cYIH`Vcy1fmTUgR)x<424mxlIY9=
zPz9GwV3kSMN3%PLm3!K_M04`Hdi0&B{2!EJ*md%v$ipr!nydz;gHM<@cE)yqCG*zJ
zRYmp}|9a!y`xF8KQAtZmZC1BPe?OjkHiXV98b~4#ytwr1b5)NYF3xTT>HO9Fr${DS
zvWeIom;`(+t;?c;?*FTF4Wd1v$rZWnu48ASr~qq}N`8k(uYeEP)3<gDQIW!XZXdiy
zqyFv`p81JDf~EMFA=6^m+bDT4@7A-PHayaj8|XR<_bf^<geu|mT2?Wu&&|h!p%*pW
zM|JdeP7b5EPiv<s276_}ZMTpPuS}HGHEUF3qzKL2xH?v;q&721j^tZ7HL|-7J-iKZ
z3$?}-J&L<xxvn6yS@*Wn^$PAoy$YlzE6}z?r<C)X5=4Ah)>^Q_h{DL0WFZMiUr4aW
zqD(5(R3hh-^m{cgfg!<_k)D}4Fk#2>TH})~2~J_a9A1Ixk;V8fM$_7s^}Lkk71-@k
zS>^=q{?Y|Pc9)wb3~mfuute#d8Bb5qC2AVY<@-^$SM)^~T(1X~=)EBEYWLOdhP_M9
zKi3w1Ix-CqS<UcmDQT3Jf_C4GHJ?t6&-NpJ{ToiePNB{>)H_a;LU&4QcP$x;A9UXp
zDygQg8TD&FCg^!7T@G68JW)OBIBLU`U}j+!`wL8JPDSf!x)M3|=WW$|C(5UDK2gv;
zh{1!lKJ1s?=dFBM6MU^^O}HtyUxE#V(KYK@;g=7?H{7dB4K2wl#CGkCtcW%Ovn!dh
zyxmvp0@l@RF@9sRn2HZ7J;E~!%2x^>d+WGt*c-mmS*cnaqUSZ&_Gr@1C(8noP5`$Z
zvvd%8CgIM{5B&J>E%HAvmOILLtBupnEeNb=a!*fuy<s@O(SeT&5XxG4M_drI2w_(<
zi;C>|Y>!XCsyL$54$u#Fs%>6LoGV>BsDL&)eXGVUnN!csW>-~fG;3aPI~?w&-1<e2
zT-=uq)n*lv<5cAlY9I2B8nXBPBt_Kz9y%Zz`8b)iM$i(MuJlgvcr9!0CG~xgzMYSa
zCT`ZbtADNy9jYtBYWkr1_hcgAKFad>evxr?e0YQ%32f-=12N&~!?CCIz&+7ePNJ~|
z)z&T>U*UYhypHGRVyoxQuh&h9UJH)xK12&_^`mJQKGpzsB4A@wcterDDHS5>W6iS9
zJ*D$nkalhMu?+0N6g{;%Lr-zlV}oQBY0SE4t;st2iP|DtUiYj1Hi8g!8Q~xA?pLeW
z*zp-7_95Kf6?zoOhkI8~O;$7f>1fq`Y;)~k&RXHb{T-bs9;ABGlamu2X<b@Tb@N8i
z!g~kaRz=x}E7m`<?hJwiZhv-KD0fRPWzW-jq_vkFL5vDOG{~NvkWm}XTZI%{)tr?8
z21-5m(ZFRPeX?%S&9j-NF^!IlIi7z;Ny(qnE+OrY>hTRfpKNeURF3qL?xzroI;s`F
zcx=My+9opfc>n6kMgf6eFWr&aUI*T60eg(j525;fjf|LjE$&;!U#}ko8STcv{Z!o^
zTCke;#6;%SVY>vNVxvt?*6c#c)yCX;ZO=Y)&);}nwpBf8r%|b9t}a|Vc>V><9?c~A
z(3}x8Fi*CN1u_fZ8C0xMY!#~Ke68E6>EKd_!2^6-kbeqUSi)SZKn-E#uJRtv$uTMe
zMiwJLIU&oQ_jLa=co*p!apwewYRxkK=<}?mS;%`-JTKsI(yj<LI{(2lhAA(Ooz0K*
zz_%h%S^?GCnmR^B{JV8+XROrtYm%{znl+&Ov^@nVpKJ0+L42VdGEti(X^NywSX7J~
zz{!1+_&KfcUTe`z4IvKrhFsVu7@KEhqy$RkEnD_Tb;{O;oN;+O424Sw|HnhIjodhE
zQBOj_vmwz0&`74kw1LT)IU&>vFUGPqw^z`$n!EGMb2o~Y{&p{oL*|ACR93|KDAaL{
z(zWi+X4yA{88E%>AsCp*!6+P>Bi}MDgIRAOv|UFl+5%KV0uZqKVQjwC1o~FWN;i^O
z2aZ`W;x{v?Oq*L%U3j=V<dM^PthqDCt0!CJx1!9)lWc4d>=(`&!PqLIVgTro{DfX!
zY9&v8D-ykr2~gmff4CQ5iVaV)nXjWO>n{+>qWf@<f_uV#C*5asx)nAu$os=dedsx7
zCT!Ql#B!I|3ElfA#fxN}Sgf!7nR`&ui<C&iSs5w}`n8=U&D6l6Y9>`y2_|wwkw}Be
zV^ravW^4)!tlos(u(f<<Oj8#m4UFc4eGZq;>ZvP;b1ONuH`kn21d^!Iqs(AZOhqsD
z<W5%ISKzM-lyLlK0W+jn8Z%03alOAr0jIiag6~Hk-4J}g{MSrKwAc)L-pyUiurEl?
SSEBnC@X=D&RjW|34*froKx=OR
literal 0
HcmV?d00001
diff --git a/docs/source/_static/remove_background/v0.3.0_hgmm.png b/docs/source/_static/remove_background/v0.3.0_hgmm.png
new file mode 100644
index 0000000000000000000000000000000000000000..d5c8f404e597a2766b818eabcc4a53343520921a
GIT binary patch
literal 187301
zcmce;cQl+`7dJdfZYfGckA7zeqKyy{ZA6raE{tBHjW9}dgCTcBv>>7mL3Bp%34=R&
z7rhg`8w>_x_^#x6o_D=}ee3({J8NNG{hWRF+55MDdmmxin$M}oS;#>k5Y-D6#g`z^
z^#~B?TJz28z?rXf3&FsjTW?g1+(95Fn%_THx_R(&Akg2S7mD(FK3}$HZ(%uL$dmoG
z7tM{=<?on^d|iA;{sU!wi%i0}P`<EwH2nMiTq^@*zR)Y!XH}i#m2}o}u)YahZFv48
z9oM$|GzN-ff06z5&tH@8pB|jcg~-tHxZ4HV&8;*x;nu{`33G(hm0q-%nBA(Q1N<t`
zn?LRC!lU4&+y6Tf!SV_IpG$YG=ozp4?+WEZa$q?BZ6EFiD~acCUWJ(zW5jL<@5w69
zT<KfzzsdvYcj2y%=TSj`enMfXOXE{UiV`5ux`6{cSYCIO8BELZx7~@sFv&eAv`DN{
zyv5@^%(J+o!yI0dLH+KLNT~Ez$uC+YEm#Z5En!~HuJtNg)c7d4qP@3-<JhNJ@yoo5
zgMh}TG0;~OrtYO8JfgF_<@n++5Qs9WB3MaDEcix`Zl7bDjok6Yq6=!mPHegmc&)aK
z&}g_cq>7D@SQzm^nunLXdi4AXx9`0MO+|f=Gm*uu)D_5QFv5R!>$)W;PnLWWV{r2J
z#ag%&gqJ8=obPbc#ABD<Xbp1P=f!lUd|k+=@pTu^5{tBpRc?NZ;CG8KTl_%ka)#U3
z3*&1%l4%)dK^XIu@j&v`E1;+I{DOi9!VkvIdWvWucf))45<T6|FO}e4n>Nsy@{^nB
z9c$v>D%Y}xzftrZTSLsdY@`E0raNoaBgU4#+qNSGo*a}U+w*2awdVXI1+1rHv8OkJ
zlc!R?fe~3}d`G$_zJDb^OJJ_L(??LI5c|f{tefPXJFn2z8?~+x8q)FitU@Q^Fi&b6
z0{O@s-S63S+i`D8-y={+P{jGSDBpHP+vU$iF!WkD!EaFc`GY`jFJSQA%zSZqpHND(
zelzxzX73Ep1K)3^B7c!mpA*MlM}*%Sw4ZTO@xLt__^@Hj!ZbK>dEv_5&A^+IV@r;f
zeU0T=iThj^uaKX#Hs@d5*qXzvpzc1&P*Ppb&1|jq_*$a0aDp^=zO>Jj%|H=;CRLoj
z*=#6329|Ttq%6KJx|<$>{S?YnHTzJ7aOZ#pmEE#6ZVcfwULkunNP1MYDA^qC`eaUk
zVr*DH`S0Q@u@;4dCUWGs<=7}<M}(bjh!X_r!@M+%LznM62WARCg4^{;yx7w=v&eI0
zBInc0sm~R{1}h!9ZA{o8$mJke2>}!@^>w(5a5n=78#Mg=5+w)g*7VHX3CEUql*E5V
zw3jm{A!3;3ijGA>to<`D^?ZxceoLI1bXAUH0V`R(?HRshG)_4r`yQ*D{`%Khb@T8{
zSeLVCgTQ$HS3iAud5K^p(53_N!r(DZs&{dg$Ghdf{iNd8HKfNG$ZEyyqs>L;zb<9e
zHAXE8%gp#__%8-T#0NA++C&a721^a9<*Tmcc6iTSv#sjWba1V%qYi4PP|Qs%Du;@U
zXBq~-Yrqx`Nm=5it1>Zz3Y#o}G%ey7YO0ApeOf;}Td@~*7Zh(AD?JE&Ab9oXx5v0Z
z27F0N5n_DGQi>-VutUiPb+$B+rkdo9Ads$mj?`kwJ*acDI_F(K*bn0@_u7rzQ=Jdk
zF|||ZWJmzrjs@D?=4*f6Ph0!#dnH5H+~x8ZsHfTHcf(}zGw;nnh$_6}rJpmGzIe+y
z{aLXdnW}T1NuHIj0+OO`ZBT>We$ePKF~jC6hxH{i>JWYH?w_Fsiaa;OTl081ZBVlN
z&bYCW2O|d;PIm3hl|Ndlgl(ZO+Cu_4OR0?c*(8bL0*vXb=VbPq6rHVDgG8U0WTv#&
z{~j@ymu6jA<{cS(K@lh;yf~|9+h(SGmwP8i>4%F`UQE@u?^ELj8!Kg{h54hJQ=<DQ
zi}mz8IHX$R$$gq*o;WAZq?|(HC%4(KAy4O%X-$)I&ll&WQ7FROIJ|mt9i#30yhHK2
z4;VAo5lw-?eF4Az6*Ie+tEn6m(<&)sN+_Rs$1{HiQlr7jglaHW-ikEc`XbfOEwLZh
zO{g?yuMg-EIQlA(p?shX8#Cc8wq?(ISC(s@ctbYOw2V4(g<Z6HipkZ0z!6hRO*Y-#
z(BbWda^}|>aUnEm;-8BjloXqDRiw%5j!hb0jJ1dh>|C%!l9?=>mUdSnbF7C(ymS+s
zNTvt}3&8M@i}idtx8ym0Xf9&$Eu)lxv|H|Fij=f`Fbp&5IqUhh*Wq{>m`mVHuSuIJ
z;~A5lzQLhsqLrkd*UNcfc$eb&*g^0@=+1c2#93)~XbL*acCB?$MiN(4;i@gl!PBZF
zdY*QDvBS$rdB(O~%@I{dy7Kk#KqxoHNOG&(&_}HU=`q9KE7wt5DY|n>+Y?PZcgJK$
zkZ+l3TThnkD99U^z5e%rcB5jjioxXa-Mho1f03!K&_`}U;l;IM;&ts%t%<yvq418r
z_lTE2YQ|P`!s4pR%ffu@50a{UG#oemH#bqGjsZN6T{ZwYH+>#HW9h&a<Lk4CM%2+}
z7R&KgTY@MOrmTO4XrG77ch0V!a0@1cWo{jOU5hSmL%azSAw1fZYIy!AHK8bI%Y7J`
z8Bl-e`B@Hw_I6rf(K)1?YA^4-=`wakQOd;JQ8ZMY*Z<Jos=ndLc=Ch+&!3jN{h-!X
zCwO&eDNV;D4i?mcc&22nS#)qQ-*%Z!=NIOx-MwgHB9g!9Xfbe|^Ys&`vufv)UY${@
zxl(`r(xXaIeS}+l>6!zzmfjZR`@G4L0*uM$L%!D&T4J@x<0oe-MPBD@)IHsN&i;QN
z@%al3t-?lw!X1R$u78X^Mg-$JuBYdAjG+^*FAAw;o|w`_^o)G#x!Lp01)-sAps8eE
zKKSVK%~&56qxf}Sr`(RdBE#Zx(Ppfe!g*rRK0VbKXRrkMaQMelleING+kB$b!g@E<
zhskp3!o`C!>q+#nWmI0ewV3{}&y=bmlXQ)d&}-3wybmQxspeT~0wxO3+^i~E2gkxq
z7S4{>Uiv%A^OvsCTg_L#lAyG~Le<w1d1o7q4|cw2&0=v3x%G9(Aj65wa;Hp@-jYcT
zJJvOCx<6|oZ8e^De?nj4nND&~_j8ZrknSNH<HDg-Go7s>`2kK|LPjK3At9~JOo4FI
zUQ}qb0a_c!yIDxI;63Gn6{mPRo&T9FuZzcCf^D@;{4^i<c9@RGh8F%iV}nw4$db`H
z0}x32nY>VRK76A0<Gfw#YFE+Gj>PFm^2-O)$v@rx)^D*zP~oeL?aPf-5@3ivF%E;3
z;yguv!R<$EVWe<}4x3TF$QoAL+?PuanA_i>&<uqcH6px?nNuZ|V$DrZmo!oP2Q^wq
z`+Pp3$EnKt9b>WM#?F>=`}r(+5AedFoCAg@YU2svZ7vJPswI%{r6W&=MNb5$m20B+
zESf`2x?oK`*1qs>>~K%J8pU{KRLP+Du$Cp_OYV@W<7I`z+kw49Qz;Dtr+q!Hw6oPl
zdqA*e4Y>GkAtT+aRWY}dWtvj3=|_7LM}nns?pRiY#$D-gys@%JT!@2oaa#by1knf?
zHs1E}h?CRYRwzV<1ccRJ$I(RQ<kXF?)f_02d@i>h1?oM<O$~`>?lD*GoI7ow6!jSg
zS_K}Y#i!IA#|5OlcBpTV8prGk(EVw;U49-vsQgial`I;Svazf<tG_E*F%^)7&yY6V
zW*QCK`X8N}%C??5NouXn(II^^Z?cOZklS`0qA!)0`zDpp`-Zu1P+=t$qV<%a40q|_
zWy5vmYS>nOsgmzzwbgOT-PqoK)D@#oWA4MRiu^i`blPQW9OFwu^qF$nzS!EG(2oX0
z4*i;{Gkfe|%}GnmsWh9~%ja`z%y5dIdA0dQF;A(IhG~|-u3c*LP}H!hz6>|{EsO{6
z<m-Em!yVZR2&>^WUC6GV65*3EGKhb6Q_s*YcFo%w&`ZvGXBP?U??dCI<Xfg@U@ZGe
z#*`)kR)^fiSB<G$JgK2BUSmJIx69J_{82)}rfK?gFWu2TsnqKt-*K)w81bG#HmD}G
zsraQf=X83}mQw6^8vg*DMg!e?*LQte3v|+-#=q?D=G;4M%{*09P-byh0YO@eJ2P@O
zM$75ZLr0iKe2okx6b9v=9GL$!?xH*{n!qYif3Zb`#{2R`in_9YkW2P<+OFTQwXn(&
znzv#x(!L+rx>%t~QC&f;PuDKOhkyBi_als`KTvnwW2z$i2o)+jlfd8FgZVybGsR!-
zIQgtKw^WJUEDN!17Z2THRTC&@Vi2&NZ$n5LsE&K~lgPJrQ;50M-cV0sW1=_p2Dgj0
zJ!Kv37GnF{`5<8IMn_5Ebe~blVDk?{r)?}#10bflxH10fihgOA>#~6}mNQ;@Hz!-Z
zA(lA?pBZEPQO*KdBc%knEs;RId$_41gks9+wCMP59I30`{E^#EfscAJcJ*l4RG_|8
z{XcCq7yB?G6N<=c^yb<RAlf12x(Y~Hew?$GX)t(#WLiWtoyhcM9}BMvR%;Eclm2b`
zeFt&T?ALsvk=D!EH}u7&4n9&l12QZY4&1n;&L9+CYhYbLj)POlaw@mfQEm<)%bw~0
zS6V_T_ShZC&e*UR86GE{oLib17%=1iPH4BY>HD>nwzL)1p>>WGTy~Xhc*eQlI!u~S
zCNfij>wgl*w05Sk@AU~kI?Qfbn$w(FLzr{L1E`YGCyg;xu$PnNA}sZc;vSJD4$8rY
z6NbjOHb#jP>Flf!$C+B1_GtlGo|jhA!5ve?$^og{(;syQwFnyaIRDLK6a>#P^mOjm
zKaJ-CALR|QWAOFCM5EGZv)bB}obDq5gJA0sdk)O&hM}j8HHU_p<y8sdIKaQ<^dLj3
zcD#gsYo|(!eZkPt_Z1nP`PoY_`)V|1)hq?8tfGHjg%m`gr3m9iZyq6cVpyy-CCX>b
zQxQ&)TsJ}gY74*&<6>U0g;A}4(}zW?Ab)5-IjOaSwv*(!gj((dTf^F}JdCvryPjsM
zx8D{Z$|fL1`P>avsQRAQcs!gIuT1^L9@3Us@>gY){g@YC)Xp|a&Hw56>o0}O(3|e)
zt09^4nkkxGc?s%V)uQ=Ok?jv3LzX+=|B$k#X3a6z_NG;)UK{Olmw4>s%Et|@SziDb
zs1db&W?#_cq}W_B*R?PF*nHzmF)8^Qqqu9NaPhK&bJgtQp>Fc|M6$sMn)y5o#2SM=
zOya1o$$rtUBsVLe7R;-J#uo?>8d~dyq73qoL9oWCH+|6Z_JbE1sq-511HPWyzSXn0
zFs4naMFI}Yk7n`DkPGXNsN-|G{JO4f@f#Ux*IfPibu~IC3{jRBgZ0n!S#b6zH9WCM
z#{QA>=F#^)86CN~%f-#rX#N#5lW)p4eZ#BergvcZg@*g~&5gBhc49<3vV(VyH%$dn
zE<m8CWVYlHqGDY4t`%Ttd)OR{&6cRED$iFN?j0{e9MJS6jzB|=!qmzm!smK%g(0yl
zo#MEm>JG7Ho#P)e0&JdR2030R^IZEawQ}062gU}edoFuZWy2(&#K@isNV!8EO(0+T
z5+t{$BrkSK-4bndkL!y%SK2@6JIp6OZ1JdH5h<hccX@BRF0Gw^)4BmORVpO@qGq&Y
z-C`+Oncq0#P0VYm5Fcax&M58^S(4ek?;=y|b7pqOPlamDkPdPphFo|vg%q8}rR;BM
zTg*zB$fVaH=VRX?JOKeKrKLTK+#bES+7vr>g*|;<Xm-FCwO!XR`^dB81>MQM5i^2W
zugWELH6wg%y0wA3l;P>7@0nn+I9e{@L&=Xynbc+G7mitluH%QYnZZ?hXJ<3Pf8<+(
ziIw9|@!mNW&!p%O4bQ?5#>Yx!8Pa%@8&^F^riCZl04cz~o?q4}bU@j`28$ej6=^=z
zx7S!S?6~()-+Y`*M<bhGd~n#jQH$`=Wj682V36s~8}7Kd(qw8XG~@$f<XsQKNbkP7
zPAq974eeOf?jF*?qgo1cp!7aie4XmzuH1b<XUy<jYBp|UvedNvMmO`q3}#{|dk2;O
zS-9tYVOSMRIsRQoZi$kHSyo|wMY>Z?JVQW~*{rtOr)T!m<pGY|G)yS>`H!V}Ol!Jy
zgoe!dlulM3%U?8eQ(Q$jeq3_xu6S)2zCMy@q!n#OT{Ar*YOjCx&{(g6;=3kZ+bFp%
zQ7<yFQEF9aIXWi;vG;)%lBH`B@3rr5zrr!7XN>VdIZJDe7!aB$Vfg2^<Fno3VBP>@
z=>%~+m08pAa+@dX^v>Y-u@PBj5QsaQk+IFefj+8j5gp_i(2~a7^Q{d*YIRv;m7z=C
zmbxdv1~YGbOcJ4^em<o}!YqWAmO4-sR+v4sZP*JFh37bY^1aJx$dkjkF}BAP{D9G0
zCrR_Lx$SlL&fHf%q2ni!9zzbi)fBgJEy!GnTCI}Wp&s|ZyGa}NdF9RE+IGaJzM{K!
zPB?|x-J3kw5!Z+Ht$QB(3vHgTm?^|B4f_~}ReuXDZHnSkw>Mjli)UmadCbRgq-$x+
z5W;uwv9C~p7m`lYJ_n~Y>J!yIs|v}n$TYDwq#vD3c%tehH`A!8L#z3cVHJLPQF!h9
zW782^imP>igA$}N4<bSJ&)9Z#X2iXs^S?JXV|9bK+!4r3Xv1YrpNV*&UUgaiAVRSp
zT8(Zh)@A16=W{%oqMO(M@zCD-cR28yG5dPoNGJjWRm`&;TO<jMqiA$%d0vVJ#f%U5
zO0?)8Q4Ad?fG;j`Tz%sEYW=Y7-?2n}xK~k#m;;96tWB7OH?xby{WIOM>ONhek-n`=
z{5L+V7~Q;GH9v-H)4X7~#9_#z@O$7aD&JRqh~pv|io?e>tUhMW<tZ72*~*GYWT+b(
zZLIC^rw3}bxiW&0OG0t+V>{Y_)my6bg^K&k3rsl!>#^8>+ETDXJgvXIWu5^?P4bo7
z>aZSP^{O0c47g7oj`2YL$DL@H@?zNXm{C5xDZQL(iTK;qHPg+fL#?vUSTxm4;<3<L
zH18%6@w)F^X9+n@J{~H%gspZJ6lmZIV;|rlpNc4vwQU)1_u~Ay;$9q6VtEDhucEH#
zVzvWMW;v5E`v)!oF_z%c-5|r+6F%QhMWX9Y=#SdyL@Sus2lj@$DuoN%%<bz}xLtk`
zM5i6Fj8-owq)fHw=Zb@sbjF?M#5L%GAHs+uC19h@(F6=p`+Ie$TxZ|mjp391ZTJ*_
zeKCM(%!Vx`P`1mt46sP^ku(;8k3w}9*$0RTDde7#-+eL&wbZCh-OlLg$Ox)*ouG-o
zPUk~4C9Y#&$-smgJvH-?+d3-WI`-mSUp`xbco=llr4@?xznGQL8}v<4LBJNFQoKBf
z3q#)|$d1yCI5Ik)r6jZJk$lMbn&#mW$))i2!E;+eVI)Ga^q?9x>_2z$2(=Rv9&E4S
z?9FN>HJ;4>`Ub!UD+MdT96(N;`A(BB4CG*9?~Txdn)Q?z6ScCikIp7|csPTxN>H#;
z*RD9@sjJZ)*7~f&1MwEH2G$gx*pliY&(yH}U5Y%qN!|FD>t|RjM2azZEe_Ma?^$YJ
z!0>#wYU~&~A~~LQW6|cGg2b@?B+HhDMMQQB!h-?UGG}P4Ga7tR#!?qxybxTzAn#RY
zf!ho<7U?S}choW72qESk?)f{o&0*(L)-^>m;vQp4Y9sN|23<rWIgWbgJ`-L+6XXTF
z7!v4!E?Mh#tbN9T8S+K+b5?&A)U_<?(w8J9$(2+V=aQrRXuLJ%xUv_j;&Rr{W|Y(N
zT}m?mEYiJd=I0r?u7B|uwmqpSCG?XIKq9#&dj=EE$u$Gj$^-@wNG5-&IrEEpV3i}U
z&>alD@rrmG?vY>)=d!BTxb~X-R{<2^Of7`0+;iE}0tS~x;u8v(O={xx^Ez@zu9cQr
z%N)aanMd+%o7DrKB=16ED~9^skEuvneq5{y+ui=WmChc~ftkHg)j+M|qgs8u$COiv
zviKbAp^;OZX<I1sUK}apu*hPjG|FTZnau!EG5$sNV*g4?){<FX+4J+X+8;NPOQ-7i
z1e}hBQM$gKP;vkH9yM>L^lLB^6A8}%36$K{X!%wi_T>uGSpc-?$oBcm{xHqj=@(m>
zx%lef^BfOrWF54R(_l(p*VV+co~qp}mJm9mnwtaGkP}$l=pb4=Kg|k9j=xmp7#Hnm
z=(A?dNf9MukCcHVG#o!dsq>rS!yg5dRLA^=6CcqUNzLQn3g}Ml$uhcguLE{(IPSzr
zfAYwt$j`TG_cQ+av<doPy0^~KUX)FIyf&c6l3yHWY(e-KH_v$HLP36f#GlT1a$YG#
z8R(29)--d9xi9j2o_-kabi8!(ls_*zwyt<AFvBpU@1eeWB1J*qf;3v;KN4jq@W(mu
zq`Z!1k^26#l&O#}d&LlU$Mb*OQ3@L}1T5scBYX<Mr4FzMi0!p*%$j<h{r#F~x?P<M
z;$~)*>7B-Pj&F^N0mYirr2%y=#x4VSxqm{lDP+()kz)0T5NGy!9-JNO&30YgENw^C
z2$Dnk`c`3fP~6c-zk(-M@%`pc%<1JPWdYKw8Y|J$q0@}f_OWaB9fZ)_;g|AHUX~6x
zt%RMOT#W1d3DUZjj2{b6_s+NXU+#a#D-qKeJ0i{eFl@28wJDWRWBt|q8_{8ZTEW?)
z4BMDmw@>eQw#2Z3ORl9tZdMh_UuvHiuJuiHRF2)Bw1(;xKTKeDc1gqL+8BPWHf5oB
zZiNaKw{$p+ksGeRrf%XrLht|2xQ)6q0$y}|J4{kM7xTVJU84@0T&^)99MPl3Rp_4%
z&+)APMWl_C6rpX?3elL)vSm^(4p$YVtNx`{{9@ZOeX05{-SIpOHTLgA7XN*mJ26f(
zJ4T7aE@ymx8EW(5<4T!mp<miWqhj&*&qCRNr|99s<V?IyV)>?Q0LydPpR#X@CaE?(
zXHU^V)MvTMJxde<Q1JHHhrMhw3*GELNI54JUocH*m^oSJcKPJ_g(M|auiVoU0Jzh$
z2~WRzWKSH`z}2SX_kh)te(QY=nJ=g@Kc9)yHM!TiXdOQ=?b3lQI4?is!~NY5{+nWm
zQ674ky5$~$%xtW`6#pW}AG@}(DW17<?i2?}jXmi9+j#9+V2$+Xl8=Hin*?$vOdz?m
zeCPPb_p#xU>sLTYMY%cK25~eg3+P%O_bukMhmmqKUy|_uqB89sXHt%aw?3iQ@&4;x
zWtjn(3wG3IHF#}Y1uR%di|;=y>(v_#vLcRnJTcy=ty|A|=5s3>Q|QlWP`C+B_x=zl
z!K=Sl+h5P~TK0UwwlJ^qXZ&#Q$18&oixmad?wL(bP1eyf;^ixFC2Cl?&(`4rG(fnh
zi{Jl+;X;Q)4O6$^!I7fbWZ%J?^rV`yyjZ@ahQIm0hI6pNJcC@U%`yeDjLNU5GOf@n
z=9Dybq4;R)9`C>DJ!U$I@JF#)ts$!*To=u91shfIb_j3EfVpBkOK4Z(1E==|u?P?Q
z@)EBxnCbK4Dq2?UfV!reFCr`NbYSZGmjd)t6YCaC_wYQvl6Oit2ZP@D*Na)L4IO@Q
z4gn~)$-~o;QKEU5C#;`U5%N;cQ%O*mTeWbop~}ZVw`IzOP#FOAa&nY#LJy5Y0$<$a
z%mQOS3e0>!M`rua+H>%$4cj}#<Pl$=XT!p_TGTnIR}^7p^FIyc0BfEgyvDv+k~fIY
z$?#LmUaA^dIQkwbqhn#Fz}m|h-&e01T|!g5?9$?o@x`nW^gD*?Po&mb`A3Bgw%jF3
zWM3X9J*uANNJqear)l!}^X%Qms^$=vaho+AVjKwfF>jc$*t+5PGuv7^C$3UrA#*wW
z$Y{9RXPA)NqGOA)O%Cf=jjU7X-l-87<Cc*uj`F@oLt$5ulFh#W5qHr~lvL`ejZ=o_
zX2%>g7H@>@>2HFCto;%g2iU*pDs==Z1KbSrh7iKoY=lgH?zfR5-;b-6E5t8iBdcI#
z=U?JPc<PTG@2h*34l)o*W1Y|VbK>*kqc5R(t!EVV+n783`-T$IM#IKhDahIK!#-5y
z^qOugp@E$6RTX(nYrtTIZN7Up_IQg_G#_m?Jl$uQdw5jXdYYaLmr`$vFR>KLsh(se
zHbU!ecO2X(SYSImcQtXHK78o3<-Y-NBsRvIhi5Y$Y5D0T85(QR-OFYvQYGR8rDF%R
zk{cuppMN>2tHJw`>rMeuT?)nfbJ)GOyHWJvp9V`AajBW!GnTVAyJc!tJ{?s|D|EXR
z5na?{oIdWKZ0_X;TP2Vp9K0jO{a@^e=5D;vqqJK=x+m9ltZ8sWh<8op8%;YuJg3gC
zJ;7?U+xC?WsBDq{Na-EeXF;-$zQWW|huPY-IZQh386$(mm#lh<Rr+xY==dj>?gOLz
zTP_Zq-Kvx6RB3bWgq*&r_{yjg<oA0<#F`xuSc9^`%9_u@(F0!poMF|eqwH6+@vD&U
zjm%*l-kziKy+nX1Hj_$E#^OdNnud_q&`CV^cg&6To8|_tFHBU&8>k8zY@WrBQW^Zi
zv-tH;yr{S;!3|}YsIH6Lf3kkMgjv9d1=4NV<NXW#qWQ|xRPA4|e1l5F^KRB6UJsvU
zWnvMkNp)_Kowu?c#Wx+5wb$>Je9?4oHr$wstt`p39@_WvLzciMf3w8q4wf`;XFl5J
zlQD68IR+<#KtbU|AVvaOsQNvgw+C2@<&E+m@>f82pUe#WIscSRDB|sBEb|5I!foek
zZe0-&5~O@f|C2p`%Wcef<^D-Tb^L{u6AfexcwOj!U#FsAN%O&cJ^VRLGR4H?f3*Mx
zL{KW-384@L>JxdrR?xy>2jpa1NR)rRG$J0Ebj>EW=}AfshE<V!q+;&Zpy)8d;)X!4
z<NNiipu-2dc7JaC%GfS=m7~Z>cG@DhZ;#2rfLPUUQ8Wmn27N7cmAm%m9#3sUAWr^I
z6KTC$dGmh<uR?YH?<@bGivs@lJV@bVGx&c8Z|^Bv0u>McY5(5ID*ONcGa5_3{*3r-
zO~EE54LY_QO$DzB4FbCKlhpyPDEp^xuj4WOj4|=d(}s5N=Wl)s#y7}${Xg9Xe*Z3d
z_&*2#@8e6d(HpfI{>S!M69x$r!b=VPb%w_5=3(adHuP?fF{&j*a%NyIb1|1W7z=|@
z0llNB6Y&hr=39SW|Mo&G()udsu84GaXAH+OLej^imotTEN+WrH|LDv-7>FBwx1gW|
zpgUE9V|8mE8EZZALQTP7yNJ$0#pPFGem6T9|8tkpYootF|8l>AuZ49Y+0_1XQZIh8
zLIsvrxakOEHy9no_{NWio8!_AP6-rxp4T5zt2+nGU6#D&Rp<#b6bR?i4Tuh?Z@rw?
zfvpv867-JXoYpJ60Zx<mYfO`H+%QrNW-c2Os55F&u-VkU)^l?>3w9)gIg^BuiZOEq
z7*>?5B{p}sS-kWW`x<Y@yO*Qr3)DA`xhzcW*%PT!-O+YqXUyg>?8p>GD!|NTVhE?z
zCm|T$C=8U>Zt@olb7ll<7+czT3?l_&0t{j4hGjcWQ3gM-b~fyUD$jyVz39R_b`rkl
zW?wN{{e~ARoZh@?w=W+))HK-qW`a3WhnX!9QT3dcWqs)dn@q4HB_$#t!?(LPad*#=
zud3JlvIut&Fwc9qzX9<iO8`g@(@z^i1{M{tMhsGnGDCw;pTV8tLq8@>UXP-CE2?({
zqB+eBWlKvUJ%Fa>W1d6I>tyKf_!}su#U<hH&>*8jvFb33SMXXMiF%vKlS<U&BkH3K
zGVnpacD@yyDUp&Wt+ad%EGm9{bQ_wWnPgrA+%@0N%08(l@i2dLckZQ)SNFhz@%f|l
zxm*ly(qwSh3g<m=I$aDGZO?~?fyZ}Zw6z~;!7eov=bYgB5=F2Ahjoq}vGVc|a4AC!
zfNkWGCwW;XqQW}^-;>l&krwH<j~<mhHz860<92&?FQzG0rLb3gI8XaiRcEi}ePf+?
zOrx@5b5)a;Uq~qv#4KO}I~bFGCj<1>vRcD3X1vJ4OiZ}Br&qh@!ceR}nV#(Hzh35s
zcgVv7xWE>SF)2#RgWv;In#sOElgELMyVMO|*K5IUF$3OTqkAe+n0y6kVw%(@m-{c6
z+ucvoLQRFi8QKr%KfZ5{vu^n1H=z9{2Qwg2nR(;AM0jVtqD@xl`>V<6VJdv|z8^xd
zBBq*3%YV@z)1vR`PJXss%J{V*ZT^Yow2{ewmSq~JJyGG8`IFytfOE4)Zm8)jf~lAE
zQqhYx##(WCqjAQ*vGLSQwQ6AwZv^@JeQs%w8hoHmgDeQy$jVsFN0X##W?*2q@TMm|
ziGj8HiqKTU%V?{>CRu%L`lwS062YY5r(<B?!d_AW;kOSUdp5uW7R|Gf3BH<zhI<L_
zhle-ge?k<OuhJN;SqG#&<-Q8~F#9-IDFgNiQ_^J*)8odzF+;zMU<Z9p+qrC#P>0u4
zlWFGP=L0ADQk5DY1+uD6FZw97p9#bpeAr#?-A-=xa;A+j2<=Sl-JY}g8@tk#6#7K~
zJnbhCx5mU)?2XLmu}GPOg&mQqYgf^tfVlCikx?W3f(Ie-`Cs|y^WW=h&A+&HdE4AD
zP+>7|rTksyow3B^NgQp@g*9K?elz-K^LioNOQ&y7pT;wVY0+Qa!!$vVmD=_A0Yq@-
z2gL`?@G%5D<LQPY?z#xS;8_kjHMMf5_RNlNLb|{AWpoIJV>Flql(jvr$sg5bsC;%M
z79Dw~l&`1bY^J-u?mFnzdQN0PIS<#%$K24$-b45j$QKuh@pT(q7IO4mXzJTO^ttlM
zu?uUlQC;6!s+*$Ln`dAB0~o~lOfNJ=Uwe7$W`fT>ojn;C%2eKThQs*IeTr~@)Fmro
z!%b!ONvhmq3CulYVklcGaPM^ec<)6|Z@^ttt<9v*YuKUaCBc~n>00r%Gm-Qy^a}@Z
z(_-^O?T3MsNt2}x+qv~lI1PHs^7ch;ELFY(pQHb`OFvhA`54Er2$!nL6kv*9PqN`O
z0eA{xPZEL0guX8iVf0=2$q$h+X2KZ@KRP-_R*2eUOrC&mQ~q?vZQ>R78o$gpE{>yr
znMmS;cenZvn$m~l4WaCaU|QqGb1AzyJT_h}VZb-hhv<1HEnj~nq<33_Ara-ajbng&
zIrU>R0E;#!wY8D)RdDdg#6(leq!c}yi?LveNqjF-u-VeLU6*VbY7jb6KtTlj(pz{D
z-acq8B7!`Kaj4~mwEvgMJmk4@y5DJgiFoC;LPtZG0yW*`a)Tt>FImtz^zg5HV9pmC
z4}i)(kFNL4nc95W7%q4XSl}j{QPimevk99!G99ACN&B$hpndv*wRijAOXsBV(*~c7
z1FY-M#U|NDkP4Xi#~r+0lk?81S*p@;IEggzF38LE2To_9UziNP62?&n)v=!RQ7a!<
z>AP!uobTWxLVSL?IY~xlt8tQrCdt^z?l+geQQu-@3@$HrK73Yht#_nh?o(6HT9G4U
z^!Ot5)Y{#`$nC<;57DScX0{1{C(58(I4<nUPgTx(QH}Klo9j!w5{RyfeIxddhz2Sc
zm6=80L_KJH!bO^RQl8N~odWYrBfJG>u1>JnHfEWDUP;fhu2!m>#osM={I9D&rK5#)
z#nfO51XIYm8z0_yHPpL_@q5Fve?hPq=1kG0?a+#5vP5y>Snz~IAZwc6NS{ue82bxT
zm&#;5Yn}y*@e#=A)KcvYG%6}v|0QKvlJ9s<RzEhYvTmUb8#o@*UHHet==ESWCPRFx
zR%zLfsyb6jMG<P~HfN3Me)by#-GxUC%{@<=Ox7+KrTAFUn;R0_p3@*^oSy&MV2l>$
zCUw*LFH(Bw`BOzNN;7|=Pw_q6W%>P{_P|Fj9^9o?_W<pbpGs1iz8U-sBPC7AQh+H`
zK!7>VT~M0xG0P(-SKj_g!y4X%oLOJaS-3Uc8m|H#ZaGnqJKlHcrl${+D_icOs+T)g
z$!s@e7;G8u6Hkn84opYS>R4TvSOtzF_X;tN=P%Giy?@*lmmB69l*&p&vF3^wx{4**
zqg$|}76803G28)E=_?8yEDNu}e0|2h#3+t~nCIiKaMB&ol<!uk>5eX0lPOvD4gq#&
zA1b_aRg9uI{9-h4yhxvj@%;y5u48#N^-2k-4Q+@&r4{zzVj>U0YQN55qU`Hl_Y$T8
zke09W#l*~byYY$Xvh^+5ITLCpPL}ld1<lpPDMf12;Y<#mh!itRlP|l)fz5Z#=l4S9
zIxoNn34js}l(MEFv$zr!H~0|YeF5;UQ$Qm=^o;^VrmxZsR?gd{LqGJfvY+On+3n!C
z$CVHdk%V=kbzr(Boaw%uNM&G{Cjq72?#uB299<~FKGJt`74+(74xqRCir|WbWJwS$
zP?s?GeE_^;e{p(G%r6X0ibk7R!UG>y60Pt%gIRkyM!5Nvs2_1${FIZLFmHi}f!OJ~
zq8o?uGRG#tH%!E&dwQde-!-rKL}bmM%JAlI;)#Ff1HEj)<bPG6U5e*{Iq%cvVRW1I
zL<S@=zOfkdEKIugsDeuIsZ7*u0(y6EcfG>@-zGUd7_UrV(qxQMd=>5RKW+B_>kxB`
z`JRhL+;{`q+XW9@K5e54N14i>uYhIqMgwV{-qxRWj=x5RPn`%Zi(3pfT#)zihM|-q
zVduifgu(;3X@#S|gVlUfOwtO^rN0d)J5i&&Nuaj<VPzzov0Xq4wLi3C`7yVa=&Ol%
zoCNgw?q8U*xSzr9>ve{(`7ew%Fj=uG6{eVm2I(4u5a)fp>iM+5v%Xv9d!~+XZ+P%P
zZ{v!Y4Q|eU;Wq=|oYD-2nHgq^53<nQx~7@7nGX8U;;j(z5tzryy1#CkHTE&Z*c}@=
zDlW0hLJ`+tvfZWjB%8>P6@M7YvF#0t_m~i(ccfz^*3X&Z7al4zeMyqPEA!-Gk2eYb
zFI4Kdw(VlBK9&KP?@6}l4l^l=-@w>6U%e+=umiaQppcEFE~MU(HVsni(8Y1;g(5Tu
zgPLw`q-JBJXeOxswf6E-rsx$E2o%bI^aF;aearAu<(alfRU(=l0SBxo-?1p&zsm%&
z`4)GWXETK_jC6zMv=;g?7(sF`$Lz+Q>%dxq7S`%Gi~J)Jg`-PKGeApQn6occWO|-A
zV5njYRQAjn@Ji{5%lPsY$ji}NW<bT*e8=_A@vZYQkk*Uc2x}L31~FXjtVzpYG<9f|
zVz#?rGqL%h{-`}Oewzc~d7N5>1V;5!;-LM0_f}>}C%~vuD|nAS!1oibX5u0`8=<J7
zKnm+ujS1YJADG+Yx@qyx@izL$Kuoj~+9Z>bS{V$>cpGXBqO6Y3hzF!7|9iFXU2;CV
zS5d3tj|qMMESy4VnF>X(lQG&Ep~&qa<eGyC=;!2C*lAkpFdls`Z!kZ*RMsBsgPmC-
zzSaJefk`~OnBfOo3TE=21v&BU`50N9eh0g~_l}`iB?8Y5j(^Vr9Qf;7RoqTrjpTMS
z7+0)`+Y^B>MRS}`gfMJWE39+&VmetwINl)Cr(A1G_ClCooL3*-nWeOxbKz}x++s%V
zhtN9`LvCrFk~H96_+L8@K%k`e@VJQ_Y?`*z13_{moGzwEG=LiEWP<fL<pmIi%qQ=c
zqn~iHbNJSKT9R`=j(ICdRzxp_FcNl7jv$a!`WQxK%_kHgboo^4HojxkTqMxZ*+&>~
zVE5BYc>oJ28Zb5Q_w8huTdwA*ZnIAx0nOZi0?9H$1iQDWy<qJGWVJ6y9Srv>#RRA;
zYBI)zyie>6YpCz%%rqa;^84*lKP?1?$h8JPiyyxazcu?QRTz9Zsuiq+SAM>W2B1w9
zPu;(q&>*|hxYfoJZ}>5<OXMN4sS)?(VSvuM!_s!qzVG?Bre>4R(5mPtl3&9mKEy}<
zm#$S)@Uv_ITGtALqF;P{lK1-S?q1Yjv9B7s=37qHfBT;G8^bV9ebw4dYVFYtvS6hx
zsfmz~vo(z_PQ9b9v^C1$7jhAe)Jv~0o52_U`!+ANbIq*fW?)tq`Bx{Lg+4rxVS;k4
z5EV%HsknK}&j-4>j)eRU+1<CtURRQ2N=Qegw)oXPHl453*<L-m@`ESJAfA{uwZwEN
z08!x(tWO-%qjOC~r~p{4Z(AR>Q`npdSn2UTVBO|7(Fewa5Ccj_y~0BMTYJZHJ%<4Z
z%Gkat`Qt_($<O6hbraM1j4-(As066`o}`ZbAzmrxL+~g)9Lv(8i}}eg-(dbPzFPa*
zP;UUbwh(^o4nuwJu(V+F6RTH-DMaMfMyT9X+i(R}C|IT8&RnTFNZ-M<#|tqRzrGJz
zTtDZg&lfVUJCB_-0hYwmL_xjLpGT;v6NM!*2K`F@?+G-Kl2xc~tWI#D-%_yBuJ?sx
zMm@rjArMzppjf47-S2hQ$RCR~Tbw2Ls!I#WLvb3j$?iVTg8n+-v=i66c8Bo3g*w*4
z?3m7Y@e1l`-}XP+KDEz`%m#ynQBjeB85no+WFv@ZWdN%P2e6Pqpu@MA{$uAT{QKQy
z&B(|Ync)tD{=y=ug;XU542$u7b{?>tXl1dGX7kBcsQxMk_<%udI@c`@b5S3P74o2S
zf-C5}T8=Nzd{%+z6RaId#muc+%aA=!GIFL6fuTYO{(JNA<IVP~b4nRy>-!<H@y+0`
z+z`P!Xd#I4DWMhNP}K84FgXW<uom}(GkKW&hjY*#{YXou`XW_dkCec~$ew{IeN?GM
zBv@|u%QjHdmF5aEd75}pJ(nETC~wXrsd+Hn^_+d2q4pFoO(i~-I(#>clPasx^mqP4
zi$N>$^Q%>Fao&<OjOe7NIgP`7*q8C9jR`^xO1BX9M|}#j?0n^iv|ijn3r9z8cxZ>6
zW+8=fZchN-%~v<`rby+_5kH+En$P2jB7cPT)&~(QE^cL|K{q>}V#JqpV#nE`+|c&(
zN4cw$=EKUPur_l+aQ)f*{N^Ld*a<DG`r~cd`cuKn^A|sWZ4X~9br%cq>iY4NQ!1Kt
zR{!1Y@|1f^_Q!P^$#O*|Th|CF<~+WxUPj+U;ygo4z0xuV&FMd(*ydjSY{3J@%Ne$I
zF$aFO*q_1WW>SFH@0wDAp-O=GA$|8wOmz!~D*(Tz<Cc$QDm^iFV--dF|J*WIwpoS0
zdGsHF^au)yR*JAM$fnUZL_etsV^n$pKQ<EcOGEF#;q^(AmA%{U%#}_0iX7}bV)frQ
z%G4R>9JWN~Q)iWI1ntH(@dN28Xez&%E1;iFbmS2>J$?Q$1UkP=CN`mi2UXlseDr)u
zsH-VPC(fmD1?4PI&{+!*O;hWPN)&FAVSK-1oL|CpN5@@5ld~|>1)C{)M*s^Gi<v9L
zG#QL)VK!-EZRPwlU|KInFORGI^L|q`fc?e6U_|M|px>}Ok?6^#9w@fGKlm3+R*o^R
zD=g6f^tN#U$HIe^6RD)j67jvKJxC0dG8+@`?hSB(k0(!hW|uv~h4sD3#c-sXK7|e3
ziD5czve=4#bqyq9nrpk`lt0P^F6?goFhKF@z%M!UOOLDbEQO*CI&|ijnZ?XjP;;ZP
zqIm(v8vYpCwe*#BjvpLN7h@6Lc^TFj0CtNDW#a-vbVn!QHOPbK3f_RVA(C_6OuqVk
zHQ3p`M7`)Js;kTz^tLOjml(v4<NBEf;AQVLxyOWet=GMxNDa}L0G^eT2Jrr6B24^H
zUF^A{7k$j$W!ps&-34{<bFDDA1>jopB#VH_wSWU=Hv1$t)L&aMVB1VPO$f|;M<<ek
zj2<Ua>7L2pr<N!Y(HYX>o`NyAf;Xkf)X&hl(8pLLO^zjQ0P#m@*GVn-fQ9C?^^MYt
zdOh(Gj@!jcjlAFibT4s=YYjbls_=ac1bU<b^Ja+QQd)k)slZGB$$zT%!liQ&=|_V^
z_T!|P&N(Fb!6}N%8s_%vKIx0Q<em|~L+gt2-z;ZNj+ep>kp1ry6TNfx@JnlW)6v+K
z*GXuUthqL~Nh3x7QFDa$&p_J!IcBl?NB=oTof|=#&${lIcs2!SOf*A`3w9w}@?HRr
z^Y{dKDd{=PTYL29%Ip#Npb6|&$K>|jxi+1ix(sun2a~`!j6LZOeORdhXIwn^4nBBD
z^IcW(0t^nkZA~B6`HV(#q!+4;1(cQ**lu?1DUAB2wO{@wQSWH2e4nqRfW;Svs==h?
zVVZPC>-9W?<kqnkFlx5d!%v}W5wN4Ny8Z1fR>ZwDZ~Q>Ll;QJ%<<|1w(c~CrAv|YI
z?lDaV=Ik|&gX_fKLfPsTa)0Kh=Tb8Nei3GPExNBkQPInH(=r5tuuQBkp>V~D%R-DE
zH`;ZLvgKgb-aDJ?9a+GRz5yz+u?uWjq18LOHzza&_{y<~xK-2A)+0R_YSZT965z*b
z(P;hDgE4#fym)1xXk}2JBe45yxHruao`;^q)zx|f?0%T@`cK^j+mz7rYOUudpgj4z
zd{yP|C}325YthH$D$U=dLqa=up*+3DG`C2Cfj@fBN^0scA6WD$Sd7Obl$M)Y4ew=a
zy_H3&e*u=Pj2O{}Zkm;wsspcw6LuD49B1#S^o}{Uav405T=aRyrk46df3%4yrinJ@
zd01y<1GtUYyWPywJmq8-zIMj|hq*JbrfmHTfX3egfmo|CtW>IwzQd%H8NfXpGA<s=
z@%Qx!)LoF(=spw<-~)$KZVG@vu`_G3oDZP4f6=<^A4Ghdy=~1ebW}2B4hGBJ`v>4A
z<N>nZH@7R%390AVno^ZF+6(Q>+oKh(nm|UPN!WpxZBb#D3c93JhmtMIKxTXB*T6?#
z*C`ZY18Hud+yezi93(1lU5nRr>+-oh)<G^(c#}H8T{aEk3{HcIW1n0=XeI#=;Q*Io
z{kLUMnJuCN9~j<QEWUJt1|Jw-O4HwMGEs!;ajBCs`+TBGShWsZQ^}W%U>1w}($4UC
zwr>;a<B?CIEX3f2o0u<E=X9LUg+JDM78`uHX)TDA`xwY*%z~agO5Fy0N-P@FI8*<e
zg0L7J-l?*F&A8?~p>8ch+8Rbh*j_^ha{<HsM_IySJ!iS6=13@brOZ%%5KePiJ`SQF
z<Fz=#wOVwVWn#3}EA`6fEI0z5Dw?krXYvEVal_GA1Nh*jnDkt)rbH!K#$;3Id10xT
z3{~pdtO#YSjf~tgMfljqwe>a}Z*r)e3v#dPXIn4;!!v6s*In(oqOuPCU(P;xT!{o?
z3hLlD3p3Q#N(}~zjX-5UVRd4|I9%rW#_5Xfq7kg6+S^pIgf7PW_G|y#4P~~DVRvWa
z$CcD6lU8E&GriA#%yRZ^gK3a;p^kg{uK+Ebo<D^IQZ|$-kty9jAHfYr^Ew0dTVH{+
zO#^*{BWocsv-lk=$##;f*Aea`=?J$=o}_31txz{&v<1$w0jdn}V{<WpC<Lf!3q!y*
z?uZ*41zgIsSb%wI@sK%ko(b#E4HH6fAMM2IOI9(NJ_!Iw4CFqw9`(^5eE<b*&xdx#
zgK5pVsh53}oLXIteM@2!6vkh}j|=zSnJa4U>oco4d??L36HHSeI1nQ`Q+kVbs|#@+
z0#PeJd>SkIJtun~hj#+}ixP@G1He*qF>{!8x`0c;EiLGLBbx#UB(lyY_-vrnyhdp2
zmy!u{2J1h}aIXaP?58jDy~6n2xXdZ*ftRCs>bY?ln2?hmoF9ex+rYn%35Z9917!4V
z?r%au&&y(cj&TM)Oy-vVo@XgU{9_qKihI(pI*^*Zf+zXDd53Wxmw$6ZKWLwgT(X5C
zF&1|8`SpRm+YjP-y>8gj3<6(Es_geXWrUOgkaIHAn~R|TF)RQ6nNzC}5L0{5Gg=Ob
zRdS*yEe?){d&kc1(!|(lR`^K%kSi6fl&*w}R=S)<wEcRUVB)8t%-edru?QHUf@)fU
zR=g+;GRS*@fjs=;)5H0n?XLR1yWB!4lj#j_t&RI0zUB$yS(2nYJAc<&6>@p3c-g2=
zNWNC0$JUdeAKt_tx+G(Xt?c!@26}r#(VEC^{7@?DKvmg*CK3dYAV5}7W<7agK6Zy}
z-}1sJw`^f3Zbe%tEePjHI{+|4&#lqtgB0Nb9S?n;TDyKxf~9kTrG2OdO}d38x*pLE
zz-#i*_HMK>E@y%e{Ji;dwzWE8>+hn^{+WrMnr8>LOixMpZ)YsQJkv#`GckeetcB}a
zYytn7u&W&OoV^!gFBl<%3cY8vcyn|G8>Y=<Z8ce_s9_cMSa*{jhSL_R>V0wgcn)
z9@A(J^H%l0q~zq{KhYAK9#P4^p=E8)jK;LOo=>H!_k<K~0%@O)R>H8}(GHk)=rH67
zkYimwdw+R%H;<EoFo;FNJafIhIc7L}<N(5pQjk32tq__N65d&9?Yf@|^n`f2#(sab
z0eVab*yjW|^uQ!cfLyZT^6AB}Tu0-jzxxl4C;KPKB2fKjaG64f^|a0JbE$1t0d!=`
zQuol@l*?5ROidju4|a>i1Sr4^3N{CnxU;u9LOTr<mt`YA_?JCm80gf$ALTL-w)8x_
zlML*Zg$dAr$rWHM7-A4g%kxS$)o8O`09yWSvR-GZ2;C@-Fu!X|!faB&JVQ>N0ds=O
z!>pb8yz@gcd00};Awtc6i_;C!15#K<foZ910t?V)&@L~O!KFI80`J1^L{t6!faxEI
z0^Yv87w2RvWmDoc^WL6!26QAs2M!)+thXZ;?zFq>x}Gc00GchsxwI5FD+tag!^gsp
zC$G22geHgE*%Z=t5LR0EwUoHARlar6q(inqSBAP0;>2Q3&wSrQ0B^?uxgPm62&GU+
zBR4q3nxYW_IHAgO+3+*(#L&(zMF<egQ=Zk8y#|ow2H=uDe4MgSv~imIHSp97b4E)e
z*~K(3^FoZMV_;%iH&Q2I`~5~Q;L6~0>!F&ae?8cT>4_vV55T<|L%(Qc=Q{95DX2w>
zO;c??MR2E80+d+`#}9_3CjmQ}xho=RlO`!tECB!3(YvcNnx(kx$SLiC(OQjpD66kR
zpT?RP(K-F$+q4xRx~{|QvQFY?VD%+YO^hhtEkitcJD`II{^V+61IO(+zkY}=*j%!n
zrTWVbZ8Z3FDTt5$+wLN@w9=xP(0MT3*Ux-uK|2(T+-Y0)qUwWuoGs!3Z&V*<4FVZk
z!82%k{b7<hvw8ok1(2OKVF@M|4rw<&zK~bc5AP&zXCMPSlSZjAWAK#J8j>n?DOL{S
z;Nu#Y!9=eF*;mk*8di{xMK=@@<=bs4fyCl2RJd!m5y9kJs~0<%dhxPiJ=mtqeH$lg
z&9)%#-EMxd`|=K?)m^sL)%(uC4lr-iZ5KRIS3s}PXfsyMrkAjm`dIsnBN|s~5QwMH
zoPn`njv`!Gta9{06Xc+g5Lg9J8b7I<e&?vcE|DJLn2X-ti#+0wcAIkG>zcsK1(+&w
z+y?_lgdhwA+2??vimj&j$0D?|5nQ1ahD?t$ygj-k>5~~+#FfVSoXz{j{!Amk>-o=#
z`Bp$t%;0cBL?@|L`fOH~6&JhZ@=a1aj^I4QS6eb)c65m%Oq%3srcRu8ULoSRq@9~)
z-{Y=;Vl7I^BW@E0F4Zo+V(NY7i8VkkE7u~UiarZj=lFNmQ%b8V&Vj-7j2VouSzX1(
z)eR9!H@Ti(9zTs8Z9cz`p)xhCScs-z$mTxV`|>j3zyGhEu{C+obpW_m1+cN2(4Dv_
z@Y;4Il<x-nH#r(gkiu)#XfH|3JwKCfmw|8wE3=rBo#_kYn+VGm?p#bp=|-uq{2-^&
z^0#|@e6%<#$dd^Q8{L9UckAM4XCO{&xV8Iloe1OJ5@1fWildZdylql``;}JLj?lH^
z_;BmVp}Mm>m$6TN|6d6ZsCXULlet5e?m7UZWa?X8Bu>|M>6oq^oJM?uM0ASzw^#NF
z7XRxu^surAb|mV8mf!(jCX`3#wOw4qp_#idSVbA!?yP-)ge_UCyT$|8FY#$%OSgY3
z!kvyE`19YN`Ko|9qaI}G-To(`VIQ{{J_abDNWL%3uxgID5$MtFQhAQw1SZv+Cm@16
zab4hUeEz1Uc?9u$<4h?d#0(0xDZhLDb%Tmx|DkJBBg>`Ef1|uw;)G>B=Rk(|zb^pp
ztLA|qIE}SjJZgAI%ys7O;&;?{S3&A_yfr{ef>`hmEn?d0{*j%(EbSq$izBN8f0UrF
zi&1lLiInJ)TuW;>N*ta?)os^omC&tHZUM$R571N~4}44)!-Uy9hS%uSVDCt3E2b?m
z7rW#N65d>@a~4yp&<rJo`fL6_WW9A*)8G3)K1IL;M~ajV1w=wx8f3IIqq`-QZd6nT
z2#Dn9k{A-wIRsQ{bmyc-Gs(fI?-PA}-oHP7=eoEqE*+=OeIEDYew=&Is2#Il7KlP0
zU_Tsh&jS%6waxRNsBmxR>I?s`?jxcJvedv+<Zwqi(AM7K3z9aXgB)tD02VQYrc?BA
zK-|$A3A$bd>R*1imsJGoMlMs|CzD%bhIvZO`$oKdO7`BdGzIV}>$aIyer!jjr%Q1_
zYRLQ<RKk4JRQjmrMe1F|m7u0fJ<5hb(nGHpnr&~|7y=Xj@I0k_lIQkB(oL`FpoYPP
zTM}#F<L*5PRvI8N#|J%Ni*xVEq{*&lpKFfb!xin1wk?<yI!~5@2A`jADi3b_R`H!k
zIL=n(Gp&v<8#b$zd#2|RT#EEnIfEE9G=zGEHN~k;iY(U%*c~3_GqJU^VXqO}@AhmA
z1(xGxhONZ`*vks`rQB|eY<x-+{OBw$6|*q^ieYe9DGa03J%P+!wiag_;(sAx68nlF
zz*cdH4+M(juV_{+;olSkz_G;;K=Nr2ouol_o`8=WGay+{R+h7>odYhxB<>whY=@eV
z^G^0Jb{~mHE1evSQgjCxLkVBX3V3Y~9x{JAy9L<^0AGVeJ?lnc{Ph$cqd7%%?{M%{
z{F(RGf$`w?<i&+KY1%XP_D+DX-J7Q_y2|1TW0C!>c<mqV%Wm^N>w;E5u);&?AA4mu
z659&lZ%~1maX`#=F6Y#)%Pqz~iZ~4!tG)GJ$!y&P{$`f9!clxvxE0-cM-;R4$)#x3
zCxL4n5Z0(98!z8;d=#bBckPG!@x_h=%9Dvbq>ZM}BR2$+2?AjY@6c{KuFH6z+-2*f
z&;<!4Je9}We6)LaY3NsXwScYHgBJkM`&(XnNId5J%*C^f`@_B3g+5+8ko_ma`VfZP
zdITwwvqr)LTG5g^mB&cJ!4sMiZ$|&;D>Cn)Km{ZdR8jvp!aJ6w6fp=%wVn45bLLcL
zRl0`w>IZ?L-eLbQZOdzm13eT@O0eQ__m2?7DJvRxj`}owQHP0oL)O2}tbFT#)&X4{
z3~#9&He!*8CxW{)*ewLQo!$~ooL##mJX_gwK6W-oV$f_cHrUNvb}m`Ia4_{M^^!S-
zV^f(Tta-N8p$8rH!lLQ#$joaumi_$$e>S0drI_XYWNO~O&5l(U`*NV$7jTY<$4VW4
zCGV0U-RgTq@>=ODJy(Qj{~yAz<H^5y@!;@l*Vt$598T^A7_q&yQGNlyy$N=`QS{|M
z&6vO(4cNC9&S3n?WZ<OcB^3y8On`pa&#ov|c@_VwWoB0!?E4PRP>K{wnV69?(j75@
zF3OoA_T8eW{Z_+6q7#g~rdfoqXbZc3mJph)+drT38aELtvh8igO9Us#mN;2iUjs=u
zXBP_mt?-Vm88`+&B(l&&POz^^SgrobZfI*ZPpn_RN>=8|7Goamhu=|Y2i%OxK)=8L
zLA3Jz+i<cQ3jnUeor%LtN{|g-<P;7iPq-mDrWzha@Kl9)V2B}67aA_0nA^X1Sq2Qg
zlpBD7`VCipLU!cAzG+C)`w(GxhY6GLaaac!Qr-l2BI61NY--RX;E#(bNvu>)odq0^
za!GDROC_jkd)o^D!<uR6IX9}Vk!q&{v3?u95uvkx<otCEjT-soHz0=eOJkl2^=7$k
z;XgB}E4At+PF`W5YBdA(+ad#O?!|@MBnbHaV}MCE%|nKUbsR$%-$Qn+Gv>2hmd_^8
zL9G_DrPHU0%}z5rWHqK!e(dv8M4;ztGxi31ku;+(D#HP&3=~zqitcCY5SWBCG>R%l
zHoz(SUZnuxE5bxQ7w|1Q^XQ4Wm65<4041dL#yiQ*OR~+A$M^l6Z%r(bGIOnaPv;DL
z+VbqnSh}Vc6u5rAy?F`5n(snIlcqHJ@6SNV-%h=zoK5=TC>&GkePcJiZMV!-xlCDj
zPkf!&Z-`z0=t>THB2s0lZ*BX(Oh#N!AB<kU3VL1Nr6c*baZu8pj;9zbFbx?b4=dFj
z5!3OEMWQ57I3z(C7FdNuz19JmV-0cf0En}T(m&Std-*c34MX|Ya=NuzPXW&EI(Q-z
zd7LsKAH@wEKx+V1ups>s;qRgx&xfU-Z?PGT`kEw0bp@GB|Np0))VdK4Oc|hFYY*$7
zJ(W;me^4(AEx8?&Z{rjg21nDxtU}xxd2V>{?AD%3x*ZrWb4r{8C-cE_s{g_0xq2{u
zwS20iL3fe~WbU@^YoTdB&jl&{yJhDf6kC#F5gm_gz0{mgdLOnAp^JX}$F{Dr64N^S
zSwX39H}jt(_%>gu`|e!2C9fR2zLxL2pJFkK2iO(>oL+2+hs|fJebzu^2e00hMg+?2
zD_l#5>^v&H>RSVU$Gv<ugx=B{>3NYQT;W=@76~aYg)=Bvfa{&xH)~=;$flbQbeZ?z
zr#1wpr5h0da5YoE5n(?sQ6?|@Fpm8XsifZr<FkeC4-#rXFnIG?C^Ktw{RW$A<sg7D
zMk0N0Wc}fuhuI!{`RrJ$1F$F5KxPvViJ|)S2;i__UIB;Y{C3P(1aM2$DMC@r0LGI7
z7~TemGvXrfbfJy9AwfjT=f=Di)G^l%sdP8Jdv-zY=1<}3yW8M*F4pt=Yrk)#$Aj02
zffO+(?j)p4ZPr71J#IbBE~#bh{~yi_-4HbK8V@?Yw)>uY^I11LNIG~`$ztHb^Sm7s
z+c?_WJ&*YI^>0k8IYUawyJ~DwB&Icc76|EiC+Z7(XbaKHa$d9tA{n(hIfagJ<6L+9
zB%bcwVXeXV)}#hZKB~e${a7`uMC%87?e%<Bvr)^@uI#<bO3>U83Xk!`m$m=hfaWhZ
zvG<#s=9xAYGw`oAF1VR>m<71Xi)0G${6DJ;t4`T0TSNnB9T?khL)?8#{yY=Ie1@?5
zzt1WOdkjy~IoZ;X{0=g|?E-5GUK=)z7x{Z?`@fk*0JQneN%>=~5fiz^(niRJDA}bC
zom&lCC(?RuO2X4Vv8nD`6GqK}$3|)ir}!!F)&L7M1SEaF3bW8GsH)(p^15glBpuzP
z4o$xg;fHy4Uif@cfKiElA!7*03y>G2yq#Yun>V}zLU!+iQh|W6Ejmj><HtA@^yDdk
zETP3|m-3M0k<@#tQ<X1v%M|`fW6h-z6Jt7_P^cs~#5Eb1*Ir=uGiYN*zDbBG!qjPE
zyy1#Ith(T9<ZY!w6;%xW7oM))NbtDc3IFh<2y&4xE3Yp-Y-xZ#2eNY`<Dluhq(S7b
ztorpE8{3TAF*RWwtfl(L*THo7>qRZbt};~AR5Q%!kv$vT44yz`bbH9M6&8qF%K7Od
zn1r~aJZwe=CL5aXdhp!P{`#5O%nc4LPRbazBafj<6ED`NzY{^!gYKn{0ra;IP4C4C
z1}Z>J&21k)od=A5`3kTWw$(B)%)Dx&Y;h{C$U&3Rb-OiN#J2j?Aj6wW*0#YEF;CvN
zHXh5GRsF5(`P=0gW_ZT{(yt88@CX7>5wa(zN-J8yUl6C|<ZprObNvOTkp_Fxl&-KD
z6{zGL$NAH6C?!sD_=i7}D!{T|tV{}eLm6E6SrPCc+OqB5%K@#!_KzWM31j1(6v;k3
zPbySGm%jgP?{%jFOt`92+tyYhu<rso==Je~1-6;Qk$3Bfr`w4nH)oH9y7t#yS5r>M
zz3~z@lL9;K7k~L0mCFD~E&5#v*ho>;M$)bVK#E6v7^UjnVbvU@g?V=_tdPM*x>4oL
z{O&LKd14P}wJxlkU3iC;oOM0!nUCG882H!gR=nP2MwoEV7GOvB)7l5U5lwiH{>Fo^
z>Q4&&CsM!32b(c~&MsTA<35!?1~_yT=wcgElpBKMt!O}Ql_Qos+oBT+*SsJ*fZhHE
z7gxCONBzS)$mJHDZL1Tk%6|UmS=Q~H06ScJ>)eLnKd2*E&kRUr{j~He0d?W{RyWi-
z9Dd>j=ry8MIQg}q&-6UATrqtY@{;Maen=<!>;c5$I^ydwe$s-cZtY>l+6(a+pJNi=
zwVGdkcmKL<-_5l~LVayI+T&MPT@=8+6>x@pBoNypAv>17Ml4q(@qKRm{Tz6UItJh&
zj*+6awhL`cN;2&1qDPh54m`oVb#{8D*#8{P;7#YtptbdXhA5Z~!i={aQvhg?M-|Tx
z4_(c$BfvE-ITKlyJx&FiaBffYzpGt?2^3biNK**XKp;POvzTos$jTI?5{SV^jqnp=
zsLS@{+;;z>oaXJWQj0U5qBXFsmlUtmlh*CikMMvFG>Vqk5GGp_lK6qph4cfoPk!Sp
z*Dr!~d-kqoiy6&;r`g46ZoB^M@q=^xmrdeZ7fpd~#%P4>cbbaqivm<Lq6t8^FKnf5
zZ~;n4KYk-h6m3j?SO-qRk8FR5_a1s<-Bw8KZ|Lb91^>DoKEmcEehJi;r)NdOxP>Ug
z=@@TKL3Zwd5ffUqe{28#(zgn?DGxq+%-;ZU#mRorg9T;**VBFtekt*@E$e%YEZ+pU
zc)&Odq(@1;R^ecl*Ix&Ith!X~&%E*TRUhN?yuziD-30%<4&OxFxNX@b(8GP;tb@)Q
ztLNRu9waD`7p%$3Su<mHDpgbNNdBF^CuL;Y*NbUs0GT0rLk~%i10xvMQ6*Hc)5K~>
z7kYMTPZYKfa#i(&Zn*QD9~ttCmCX}?nBe&XNFN?AM&I;hG!eSUV(X<EHd497i<*@>
zcyhC9reYgK`ioBa!n1YW-K+|(l^Jy%_iRli-Dr;9s6F*6T)#>zcf3{pNZ5z;x$#4g
zxinyv?2*A(J)R`?<ls;tqR?kK5H~-8p;K>O)7uyGV{SgPeyO=}o(#H+(SB9O@6bs8
zsAvFCUP%~1XJZZtJc47?AE<nyX`d47g>^)q%nY|ccDTVJN4dgLjSG#@A9+^A_eN6p
ze#PVjjE67-diRraPyb_8pd)M<Udn}nX-8-L+Rgum?lER!B%qaWzyW{Xa=$HYS(7?3
z0&9kOmZUMu2R2x84|bm97Q?9kfoh<T4BN?BqZ^pTa$30=*56BJl-5ou+|m%30c0_3
z5XUBSyW(aQf=N6;Dr`e`<j#eudYh)v+{$Rm>-A?xIm^X|zZLq7wlTI|j$MA-SFMy`
z;~sY$*ODmaQr2*yL3D=V7XhxKRw35nNc>DqPsw}kh2#WdJGT_z2>@_s4N%VxANs)3
z0G5fC>oyjk3;g6aYG$sn4kk?ink@Ao<E;$Qsw4D2RBKW`-XQ0nfJ6<VB~2<Po@4-u
zGouL|2VyHAwiHK{GFqI5)LrxHZIU{`({;XuU`Cq(k?~bgHR$Xk$PKO0T}803@%v30
zvJW}Pxi7!k0V!48`2&9X@weIi?~U#0;;^%k^h=}r^wnqOI7$ER;X-ZgF>u||sPnpm
zx0O>CY4uSO(i3oxR9Aqm|Ae@Gq`mY0;G12%J~mA<n4Od0K-+rtVp>70BK*gZNR-UQ
zWl0ODakWqH^QZLxkb$+CbGw3L$Pe|!T-n8X2g<ZYp5zV9p^m+&s7eCA--rODhWW`8
zLNoh_i5OM+93&F{6~I^@whnrn?zMA-zx(!Kr{PnPb}uRD4owpv4==!LHvl2cBIJjZ
z^K5HYX{N?kJn;PndIz6?++&%4ELChI@+<^5c%3;%SjQefx>?z4eb5JMWSiF@6xg&a
zYJ+w;>*MjALiEPl`|1L0Kh99&=q+^e02=iL>FsYbhnU?oNqIKro|XBh9zb^I3hTT;
ziaTb&?!_RSHAXZ*RJZQjNoMKO98vdWt`+wGwO(lPQnr+kc{bC4-ms+ro{<B^{`=7x
z>OMEs)uyie+Lf7iel1!tPh1t=%gyq!?cMSMh;*sfq$g$L7|dhdbLPgPy0%G9POTAl
zIcq@0F$(=l4;2?X-$FJqmyZeEms-7s@0b;Haq)0%O_JN*+Wb6q1>~WA!OULHc?G-W
zY;``-e7&bpy*tU|beG6hw~)vDW`+^&sG@^__43Ee;NxBl`FSD#_H8o%k&)stH}3dX
z!CDq%hYai+0~h}>Mqj@zWtZ~0;H&{ZxkJw{)}TJrxnLD+W?*@?5_YytygQPI1eAF_
zl1#h2w!g;5A5~*q8k|u#|5}K`qQ4zd$GlFN0Q3d2uJfn4sRqIKAz7u9Jl?!McQpD4
zvtoB|9lAzIs*{a7x_<sk3OrXU#pO3?jd+K3&;gHBbe{-jO$E^=`=(O;0<<5|h>2pl
z-TlF!Z_XGE2!r?GauCpqY$V&^OVSqD?2@?~OfPAowQ;{_&67@Q)4NXs-=!YRd$^>I
zcdj?XMxvHlc=8a_*#=6Qi>PDa&JR);)c*OZ{|5JzYD#i5SQ~*ylP(0eWt?+Lo&Gdo
z{J;?MAe9G1knKxRNt%!YUX>svm(MG_nKd+xO&Gr8-hc51KNh0$Z@KIL@fVe<C55W6
z1o#>6h{ursDTROd!=D~T8xQ4k<Re+f%d)cHQfoXKti$*&78{Ha(~R3ZI>(6kS2rPi
z`5U@6roC}25J61~qgYLThtC}uazBy<Z2qnefM@u?F$S<D_^R{(J)ux}PG1Q+3mvXW
zl9L>xQ{?t@$)t-PH*xe$sjOhV{wW_9xG~H;Bx)D9*E4+=!ojPrLvOR_9v0?VSZMd+
z0FN>D30mIt)=m>;e@qI}FwsFw8bZp`;iGzQN~8sIm46oAyl15OWy2{g>mM-$(qOQ4
zDdrD{COS4~LcP>e#W$Cd3a&=9K)@62X#X%pF0tHjk_hqVOv{D?l9*OEAr=V#zBr5;
z0%=J|(4$k>+H#3(@VHr}fPXyl=AG3ZR?IcdKenRzydK>h@aN!oxZF2~!|h5YUu{Ff
z+*2YQa1xxj>my=)wNsE<R_2W}n=mK2VE!jnm7fzYds+UQlgH52bpehFtRMB&@}0~=
zAXg{6z@1v6!vMPYg}K3SVAiLYqdO|_!gBIorIGR78;*72F#Vhn9rDjafdnUibPnC*
z)dX{xJWzpBY?PJCdgu<!^IZv#Ya`5B=BZGBvY_DHDbE(;CP+5Ld70Zq(i2nL5cJo_
z!(#Ua*qDv&Pk-d7Yvz#bcMmI7se(W+Z(fPWM1S0UKG19+&vMhHWcF$Jzusyv@>bLf
z@o9KV3}|J3K){qLNBsWe`lUmmJmDRBLen-@Wy=|S+;*v?1;{fYAi&nX-i}0#k7`ZQ
zFLQ{#SPqhTG;LzSp+liY9Il~72FfJ7Y_$Wt)W4-lKNw=D1bqxIu;LA!V`F6J>-|%D
zS&TnPWb#93kA5~+>BsJ))#ID=2R{?$pO20VhVPFK;9q^Iyf|?5&&mouSYd_G;t7am
zOYBoyZgX1;?6Je~x4#!85DLn6+>_C0Ng&3OzC(l};w&-6uDK!dI6wl})toknSM9Df
z77h%amOm}4(vRKWud<x^wLoX23{bG3vinye98e1I6-r2b3Vd>o{~w6~s_oE_(CcRa
z*V${$hQ;QK6utiBcis~EtMNXBaoD*vx3CG$E(cqKWD7faJhfW~4Hr|gsjnLtE?j(=
z*Zn5gr11D$4tAo?_B>g1=D{aVB7Sm?5-xtJqGdrtR!A_+l;@)NZ$gq?*}#PpS8?~U
zM-$|g{u|M-<$b>~d=m3pU=zn*8<LSATc31J-`-#8_`?O=3nAMwcFd6#PKVEjvEO6m
zu$17J>O#|$PCaoSODcd%!B>Lo$&KU|7m$Zwd4&%dtl6P^+QJaSR`gg>jTKf8zTyn&
zx}QA|kI@DW?qrX~6uGo28-zBJQ|sZ1!!dAm{4G!B3YiVBqxX1y13ES&tg1gUPrQZ~
zlyBEi4&X!^uq+TiR#uMrO4DN)ngOX&N4Ko2^(MOl>X9!YsbNb7$$n(Z<DV=}W{J({
zNu-R?wC$ti-?f7Z$lF;JE_cK2V}nZGDwufe5<@;aMV|s~zMjATzelxI6C#l7oVF#I
zErc_D59>pN3jD{GhbgzZsA8HuM|-@X!|&jCZ)*mg9%#hoeZ^6Rt=NLYYKE5c-|Dic
z_+TpH@TzhYns2xh`$VO!(=FmeECz>bIA)!Jfw4c1%i?cCM<PVDB41iysp7+OT(2$Z
z>zUb}y+z=8-C6lFEU?b#6IOoV8a(RXRDh%lz1Dxf@0#W%vG`~psJCL?Z+PSJivpas
zlP_G(nh5O43V|hXDdBrg^7ve5btaxd%D;#P{NQj`hU$DTkOKGI%O--xeF^atgRd|`
z)~Fcin2d<)f#-%A>-Q=;uJ9L8#@cbzrvsbk)pMO6`A{kwzRA!%IDK&7pl|{rN6tEd
zi$L~YUzW<_;3T(W?hYS4;xDQT5r-?rtGW-gWFOJ!GUVaN#|dQ-@avh4c`gq3IKWk7
zV=r~g9ZKO&r5~dWp%my6$8pQF^)UM;uCWW!fO`3lIT&cZv;n>%3R%lrn_0{3dI~=3
z$enS9E`F6Cfp^>$^K75l@Z|uHJZks*P$<H%Z{?{~H=RpXntWJJ3HB|D5#U%*gmM?+
zfbsk-@$>GUx1RM$P!+n$uL+-8K$5K{!M3_K>wata&E$z3{z#$UMHinuSnLN{H&cU}
zfi=sv9|pP>&f<EYxvYvhU#9PdDV8wS$03Qt9Dej)4h%rEm2tv=$=R$|jY7^mhD1D1
z@cn=oS;^e?3x*-;M<8yRlcM}Z3xh?RQ!vlK9G13bj=m__ZLRXY)8iiY5CZDV`pLbE
zC@oMjk(4pkwz|y=!2|AlNy>Ue#&QoZ8~aoHuXR07=e+|ufX`^mh-^3r@tZkQe6k3<
zT&)m|{$>pU&pd-(^1J+lxOyuV-e^18zSH|y0E%h0q+#cCS$=4%`zEt!zvD|jS6s6;
zrDOG@1+VT{V&X3=6?2_@%-Mk<&6&SE*XH~Gp=BK+NO>v(R#?MZ-<X=bU=Z(@!T|~8
z+GvRsU0{!f>;C2k;Nk1F-uo>OTix9Q^k`LZALRfgRKO&1*aG`Zr5oPCj1GAA(0S!v
zL{ErZ^>?)M{)5C-(eJR-n!O-)Z{FPsh}-=fWhe>k=jmVU0<ef-(8Yq1%UT>@I)`=H
z@90}_`_Nu}f+BRb%#v;d(;lV>Gg~z`t-82azZ*t%afT1lJ>z)l)&S?MbgPpJTVa)`
z(FF2$Y>n}OJK0O3ffqlB*f7N}83L@?TL&fXH%GN?7InNd3<}&lmpY&Q8z^9+#ljSO
z-sh4G1ie_ggvdfB_Xl^Ye}4nC@_Of`PT7hJBRiqw<{QZ1qu(6*!@W^~MeGYJ5XS?R
z_<{yKDW-69&5<lDJ%>A89bd2Gf0w>_tG?@J2VfWf`;LIYAb={U!G4D;Lik-F+1>CJ
zjw=AA3d(F`pyuT3C9cf**!>OVUHvN9sW8@bOzrXLi~934WqUmr^w;t-))Y>RAtu@m
zd-e3N7Cx;rEtDkk?cJtunL_h__K+bG0Xu@@&mz0wO{G~LgG7$PiLdIK;RV8A&r0oP
z5C~cMV<G<txUx~u#SiM&God05gj6PrO)fR!#R79pWV%0l!N=~Z5q$n;Kbz6^=z3?R
zWgw7R;&{AJ3g!1MrtWqii^L-~pvxD<IQ;Qy((n+Tt*bWmTVNnPo)nbXHAO+6U*amM
z05#cuaY_H7m5S4}C4^9?pnn{acjGCCex;sW?#)WCOST82k47BgdUTJyN}q}Ewbto=
zebB!6P2t*Rmnt-<bS119euWmF-Ye?h6pi|CXADDpp^I8k>qK{y$U#+u3FEa-AjxZ%
zPXxTYyvPcE{FapVeE9zT=QOiVnA)<kKUl7BufP(mlkIo$KQ&vjA3{Auz@B#4$587k
z>h4Av=$WW~WiACl-xPp5a?lc8PFKHYiLffK93QlC--E|!4GnW%$A9Dp2=uik5>B{1
z;0au!{%oRrjl=6#fM2!Wd|BVGqHye*BLWRyHL_gbX7lpH^o&&bzRdBn3v875jL`qH
zdYMXH@u{$I-$Sm`M*M1A#eI>wx(w&ANPi?SEme!(%?FmDdz|1S@PLyo6E`H-$e{zx
zk-5VPjvQf=a`@ct%F*1d+uh#+;dxB+N5D6A>vvG1+yiivlT${9?=0)%|3FN^Pf3<+
z6qJgFlA4!al<%Z^kW3CRN1P^pN_#sXP}R>S?%v)=qLaW1qLap<KZQj4?Z4-Yf3MD^
zr<rN!^E_O7zR-BmdJHdM*I!(OTd_mz;47MGcMHmOg!p#J(I=mjDiY8qRN#lV2Nr(G
zh1q|E>ub;VXsWGk3YA+P=PgaCCG?jd5ntJQb!uWsc={ZM0m}8FyykdOjUvhZXo;0W
zGgubpta$>);?~FEG36+V0nUK0APst_ut;s{G*MBJZ;2Vvb|jJ>ugkU1y@s3b@4G6U
zr?~0c#$3M>*n{?juLRCSYZ7dq(|&yNqU5de5tNDx62UlZ<M<P-X-B<Lp562!<!QSY
z{lAt$406yNgs%X&c<{?+>HX|L^hpG_CY#@PXLZP0acO*6I{6V3xU(&k6jb#td1^Kc
zaf`cz`>>fjli?<bY&t<Wmoj{~xlkw6nuac{gAFS18XhDp()rE9C?ETjB4+KAmUHTD
zWGS`tvA%wZD15l;(@3#oSO35BD1B?62y7w)4S8&ZhXPmd%yX#SeifKaqU^v^B5_wg
zkgG+hxs;>kBSlQr;pNNHT$ayYo1gJ)gFs#1I}7cjct%z3bi^R!<RRrsX&$`pk0E!V
z&@?!^2Hu$me98`KV3_ISJB)|V-p?*VOsGKlxVXB|r(x`hQWAN5>_zjvRI>3|$4bpr
zUB8zMn!EP@GD(27c?*3I1E+Yr`!u|R!d91BC0)<mEvd)bDYrxrtuw|9eQ+&<ADBL!
zxV8|WZ3=bhHY`VYi4F1!9gqsYOQA^xZqh?{rStoW8$pq(RK`#>a7=w+rXH%ILY8v)
zzlnFCVg!X-V%y>NcfeHx$BECN4Ph%haX<hH1TsSMwPfGL3Pb86;HEa%O-T8$2R`YD
zI<P!H7d+4~@&b>jOIOm3%gW5`Z;|-VK$r2@>MqT>uw-+ZaGyJZJHIW<drI3&N*V$c
zmep`Db6VPx3mW9W$0Z;wXbuqH&&7vBI=FkbXA=Q*U^<(#$#+WH&Cy}u8Z8iSh#?yN
z58$yDY*-qn!hgXP%T2sGf`(J&c<9HXL@ul{P%`T+vF5OCS?}{3;G?bpy`&>Q2%&dT
zv}*>s_mrQUI>zeoPnM8o$Op=(^jhoa=(MK>^X@snBHwn*j1Xh!fQ|pl0=OB;2FQmv
ziB*zDPOSHCfFti^KeL@xgf7S}dW(5=N5JK2cGK;y;Sff-A!^YpO4%g?!rDmG%a8CA
z^WCO0DKwA(<ZlstPiUGoBVTE356Ia2KIjVXI9c-&3!KR&M_toM$+bz%&rhe<TVo2|
z>m;QyJO90jFSOZ}KkNmTl9vqO^a0_&eiU{<IpFzI;MkZnT&L_O0!{!ADBGeQ#O5Qz
zqHgMG-<Dg8?V*EuNaq)3uw*kjTvmpxJ%9jDt5N#dSWb!<leIjGh$nUfUjWRaspa5S
zSaWi6f*k#K&R<VrIhvYZK|Y&XV65OBq420RK999(ljgAqgu-unVC6M@CqKORw8U<$
zF@?@}F~l1P$VZP)NCD8-6^$9Dd@7kqq|CqH{VPVYBI)QO{A;;q{+`nPsh5ElaIkn(
z5B6mDWvocz$&S05<0QAPe#R|z!>aKAFpKVL;J<TMF!yX<FudpI29}W!R1ODtb91Z|
zoL$C_ekR=aJ#Y)=w?4}fiO@Yf{pJ#h!x#zsvoWf;iTLltwY0Y@7zZ})PSxO&7->c&
z76=w?eS0f|+zm+F;)k9o@q`k6rc5R?{yijrpucCQTR`de1XM~zM=*Y?Hhb0G62c@&
zN3~*RUNj=cp>O?sz?6ZJE=bdXn7@cKTW=)nVh>_~iOwi7)0=o28m72ZC|T5%A(FeC
zdGbX)JU+(hiO)j8?GgMw&-F8XD?<9jUhg*y!_0pi#=Yx<?Z4qi-^{VL&|xjJoWrTZ
zdU&ci_9xVX1?<_*`)=~b*i=P9TxwnFw6U_1g!|mcchO6W*fl>TF+n@9G>n&3G;cLg
zo7H6Ji+F5uHpO-01@A2nlRpc1RH=ki$w2&=NyhfZ5DzOEUH{_>WYFuW1utLJ+5FJ0
zkp=lk`sI82I=f$u>(If3xNC15+FhEL5*Cwhb0YI`>P9}L4Qe>u96ztSW>jE7*8xac
z{4i>#{yJ8NP86ag>^c9D&m+0Tyi8Q3W9Ay=zi1^$2mFiKZJEm>zqE7V6cAsI;FAJ~
z9jHAS3l2p>lNY){BhP|@E=e4<9<u(Dj){$>+-DSD3`pdj;aeZRhSn~92(Z@^-%goJ
zJ&SZ{Y+ijeJ}km_y>G5oRg=K0o4<CwFZrV21puyg2Jn`1O|z`g<jGCm&oN`)F3`~5
zlIhhsw6u7_wc!~YSUULt_-|cI!@}`pkVX?!N~C0>QeBGsF6STf#qqdT>f-Ukc{Z`)
zDq=A&_ug5|+YQ|fkXqQy%*}m}-`e^tMb+x^QU#lk7b#iVrU=Kw2PPQd%0IKz1W6H5
zKJf{_Lsh0elM%hmKW2p`Tg$r~MNE7ZkWyR}=_(tBq{*MHWH*P1Olp!62lU(z*ZfuJ
zlz=lMr<$^FlF6&Fw?etl8^pUB`)>ZSoboQ~tsN<!)1@Cv&BaqA!5#63Qp^sV3N|#G
zEC8^yFI$a3f8~q%utxdKlO};bp;XLqPNG0O2D(@N%JB$sC`AESH;!=%`sG~y+L=FZ
zKst(1cYlwve3g)wyEIIeK{?<L^$_H576Z=a2(W(u9N`Sj#d2z$(*`)rDc&3y<2n0<
zzXHni)30#JS$;=8>yj->&oG_)XSM;y+jo~s_jt6&U6044-YEVn|Nd06r|J4=@r|dc
zpMF;bpPSZLDMSPLS0B3BuM55BdS-~Na-dG$++2v4v_-QNw%iIqlQQ+s#|IzYWEL%}
z@)JLc#7AL*@hd~IZ0c9h9UXfUbyf`utt~{oh-Lmd&lE4>%}MVL)XB~pwylQa68F4y
zpVb~bJ>4<~pWotHLvm5dsw1xPe7n;w-S4Y?#_o;lD|O4J_P^(`b1`u~!(5LZ4LrO*
z_i64=maH*P0k@#bVCB7B8+#X^dj?#|!V=!NnVdfT<fsQOW^Wxsg}6iu+sVZ#Vt3XK
zx5Fq|gYG7IarbzS^=>j<3ndc~w(rj@b3K1Umz7oA`%db_9N98~h+<6>2q3JUU#@U>
z3-sC;$0CwG{Z92exLJc6f7Gm!SYnV>Y?uec`aWAed1_YPQ*B|9cHlCFoy=27G*&T?
z*S6~=(>Xr-QES&X8+;xTmJulQ@CpBJqxy6S;}+=xUaV*{^j&qQD<;dzYo`9qa(cjg
z=aA<~Ah!q}vxE1-j}+;VvNDQc)02FDaBZ(>CgxltO;b-(%d_P&R2;JH*_8!WLBMsC
zbGIu?R@)!HVf!kIeW=fFAiU~$jc~lkV|?gn{_NRXeELko^gPg>zptN?{gl_yDd9kD
zce<kQfv;b!iBA0H8d$i>0A-$Dv+DS3!G&tH%BV#u*fp|hnA8z~|9dH9G*x!ke7Qvl
zTip!8IlMzV!-=c1EcI1fw6CvdfpKiu3R}}p7yus$PtSmp5moU}i2EH}es!R4_K+Rq
z1OYsX)a1sqWx#FOJ|@Z}p^PC<sv`%_K%weCp>pD1@T2-8*2(Y_OTQ~gb<E)X;V~ZU
z#Qe&W3f>hDMX3C|7%9lyKSK>MY(Lh^C^rxau9-m$3VyXxGZ3<5s`c4zyLQVoz`@CC
zVVCLj*bd{<zM^Bs+<d~KKrtO`hWc?ncH|DL(>kZ)w-YctT;G{75&b~Z&_l@<-EfnE
zuPo3qpfL-dDPwR-Vw~<b;>ofT&8XsG@)VaSnsR20W2|vh*fmm!oOHSS{LWMJ?{zhj
z$@*oEV_v1M%_p4FO<rRmR=pl8{USpir;JlI4S&Rig`=f>&S$bw-kU>g4JwH)tBL!}
zaqDxdnn?)6YNK5rcQo;C&~{H(m)6)i(M0F!?^FCoN$tUHrEW!c6rtX6^4IBUjS=QC
zNd`{%cX-_6(xxo)V@%3r$)lA(*w%>DwspC}ArX0M<4ITd6!Sz)-xqtSV{&E2fPK#+
zQt~Z)68~|^>wU+uLy8#Oc3j044n?rk@@>A_+1bfGc9RiMo{%%}oz#DPqh}f$unE^B
z4y7ck=P`Q1W<E2C88KJMD4wc0;z|?t&N$kkKbu_J-P4X1yjnoX-b8Y<V`aiynlLhk
zNpgL1XK}AEmYZI<7=N@=@knm{&UtIat)S#)fuP}AdKnedDZ#}@R*^VT*T!jS2|BOc
zg37EmS8<8y9@xR=)WUXmjoDGDX<g&lx1r7tJZ!<q=JnHzK&0HEKl5=Px2ZsB)GMZ9
zJ2T(iccrD6_n_ROBFFb64I;pfqytOjtrJi^=sslNeF*vK*$W7v#p6<%rw5(<=F@a|
zDg@%GouPvI0tdvgAe|_{tF&<V9Tyv`o!K&R`}g9HE%u2U8Wvjm>=Q@CN0B#f+;E+x
zpr;*R`5IWKuyRSd>Ln>SP!URwuwFrah7Q43lx~Rqfp;$02u;c?fF=YFyFj4v^kiig
zuc2QzLe~B8Yq6#=kqn}4(@XobzXbUtgV%Geg#llG;W^m3+MHpf5Y=Q~dp>t;bXx5b
zlh~C?cWmtuF`|;N#p`UfwtJ0vY=3(@zL4LJ?H$HnA$--pFVY`74v!J8oM_xu$4^N0
zk^8J{8EHr2Ohn4O^Xdyf-nA+#2tM!blbiOo3;%hG+GNZ#n=Rdcb@9|2<6dvokU&GE
z-+*WKV%+VJ<$}GO|J`y07F6}ae4x*)Z$}TIm+?I9m5<Y5AZ=b!m{eTS&Qw=?T5oRN
zA9W1+rq-K+ABR4)yIx&0-ZTDdd2_N;x%v6q_zKil{N#;326uN)a0ET)Pd$FrABPE!
zO7pF3*s&{v(9@pE{eyyU?V%VSXCvw)ASqdT-_8hqV`+W1>xUO9d8)}v&(qJGTNzmt
zQn&H@<#*sm*!H7KeUji2W8z%XfK^1E@Ga*0oZR&#T$2z=Cc<lT$5!>6HAougLL0x~
z_Zt=3=u)>mDa4<zdYu9PBes3A+9AO)0EZHD;TApEKF0gq36b;0yjp${zpa~YWfdm-
z76@G>Ax6SRJS1E{uRi$9K_Mg$MLVX<d{rMGrnPc@=zWo1EVF9@U_j(q0)Jb+yyoTU
zzK=9&=NXym-c5?7%<-U+CZ?hZ*OtfnmM#vv%hZX~G5H?(G6UW({S3?3(`yWRualLa
z!|xOBNF-GiqjbM1K*Jk874VVk6}<io->4dp(Im19fv-G3*@8gTi2w@eZXw#H)e319
z_59jDeYYssXf6%dnN<5>emjVvJ_)A4&U1_^J@Hbo@OsLr3D5fI_IQizhpO&G{~nXd
zLGQ(V*&)snop?_F^;77sbj&N0(vQKrE0JxX?@CUb5zQ~k3dF^yO_thfOu4?@G;8R$
ztus?S7Y&leFfrmc!gecd`4@gQd$o##hLLhFXE*V)8RAO|<ZsD-;JhVoH)GUoekRpu
zf4I34?>iPk&Zd5f36)VGV|%(eOOLxdT4;jc`~JM?XqGJ$O(dthOjhic9z^(=`^{=g
z`=Q;7iEx`UYzsuC8$lnq4(Df6OMgixER-F_5n1zc0*7Kz*6;1<i3FBa2w2vV#DMc*
z6h>fY6up|EzHizX@ZM}nmpqLd7nB&9!xu3eHm-+mOo9JG2B#39P7O#RPDlzD=S~tz
zuO;_O{Im(Mc(0KwB)0kHArUQXxdu71E+yQF!_6eGceY=8_w4U)D@RMkjYFC<@9F3M
z@slXM1{yYpWOwa4aXFLASywYgwO(T6d{ziN6c$rdQ)*f<^f~o=LrpmKIB){;^Kp`s
zllvZY=0{PFm6)2GzfDTYR};tk=#=?My0-AV`AHn{)~Kp0n#n!*+oiYbB}{R7Si{LW
zs}$H9dL#RRN*U7{)G7bL=EMSQa!seMnk_qrldP;h;QMcPz|(L7Dzg4Y^o`^egPR(0
z#PGsGdvsl#u)e^{{@;yLF7?&p<u6%EnS*RMns+G&)0;+#+c@bw7vMV7oLa8Vv-J}V
zx#k1#DV(_DBMs}}uLd^+aWOSjEh0v4`epDli8_Lcjf%lUb<wH+l~Hr7AK{8FZvhpc
z_yoO#&yv{LF-w_cxL>8ZG@efI{&E($w7hh@xI}v(1BkvioXEh1=Z+tEcPSxSC7*$a
zmbrnUA3?6F<?>0ubq==zt-f^F5VAkh@q^73;jc#XxNr0yZUKm5L5GTSvhk?Tav8xW
zV6T|$f9>Izj&H>$-7I3IFa;BiL2s4n&1qb<72YL!cvweP*DJx@VlH?rpf5M?PPOZC
z`%!?dkiTCo?NoC~lYfR@3$F^U%6?R!J=qKEJk%puQ*KW^S^wjnb`FW>jkhVnJQ@!#
zt7k|WSblxL!XF`g+SIoDJ4@DOm2S1SuC7ik;Y^2FGN7X=cYhD(?HlHP_lV#_8<?o2
zr8afK&mAcnBB$XjM={X-t9RK>m)(jIn42yO0ICoy1_wIBnoa`5BN%^@fT$Oi7Z9(o
zXaiUTY9V3#X+8jBc(tx)CRk!`L+X>^!*Glzgu-=!CF}Cwz}jR&a`H3P;5WFtpx2v+
zK4#*}=!web%k`cC-HcDV6xAQsSI<^EA8YYMA$@J08CG*>*!($g#%v{<X-JdUEvrll
zI*iv2AK6V$&n!uw1f@KOWTfaxpQr5jHX-=)M=+=AUmag;sz1!|beYYS`S#}9hlhKZ
z?(T~Ii#gieb?a^_m6ZHKdPO2I0!7qIN4J_7bdO_!{j6-=pE<DK{^)JpB?r^&!O_G7
zuikmTk8kOT6KioKA?})2)=rL|rv&`6d1DTaYqy3Flz3_Y7AlMiNOSbMD_=fGgtwyQ
z=g33ycINMrf0;J#yl&ef=Af%ImxiC9H&meNt4&%mnM+vz4s_O<1O(~0Gsy|LP76*s
zVWf!>utkNzX>Jd!qo<##R^5YKJM6s#(lOSUiKqxxOI0q>C}S5EuCYf!g3m>DVZMKE
z{^=WAj*at`oa*drU;n(TEL<~R;q@`8uF(Vq?~1Fl*)z#fS?<~Y=7rcbn0QF?WZHNp
zTtZK1(_3j1Jw>B?Z(wv{Vq$H<K;P%Z5aw)rLlFCBAj={Bp<@jKu(mtAZthhMq1SKx
z6k_(EyQhD}h82>pa}RNL+8$Rr$ea}buqEL5C)6f^X2G6sE-TQ2K%Bg1Z#43&*|mwx
z+0@j&o(0kc7r+SQLZAXFJ&J*dJsf5z2up8mD>z8di*?(O$FCieIt?h<rk60gXcI;6
z`!Sni_n{vB;pSCHW4yTyqW}<w$@7gR@#QjlPjiRoq<w3Ab>f?|Vj)7j+D)^QkDfSy
zV}9RT%HH}WmEiRkG3WiVG5Li3t&~}gTwk@O`b$U1>D1Vv=F@<Ce0fJ7A5juH4kb!T
zHNS>f{d#w&nMd}6St75)Fk|P8sCTE+Y~`!;=o-7M(BtMe%+&U7US4Zkf7J9o$(B6h
z-&LVD2?ZigDYKxbY5+r%P`mkmQ6<WMTwYm5(SCAtScm)#BTbo<Q3q*3aL-_Ar0$5b
z&UZM1Euw>les4>X;0l6or+b3dKi|{ky$yUv3;Tz=e`X8qN<R{9AD`+(M5YKGMxmBI
zR4hrDmXwsp$(&T!X&2Z3O7}dNe(z<<^w4f}JcUCFzq?TyP9jsV;}elkGCflbtFR=p
zTU~vjYd1PfHRZkKg=ug$(tLrM+VcO6-r#9|YJW2wp(3K~wpzE=s`=<fsm4!*?~wn+
zeW|JQN;RF>Z6$u)%zgyni#lVxa$Ke4zmMX2E2s0EB&VvaFXqKfkVg1@J!4O=S#Jyt
zaYRCIhBn)B@mO?yU47j8xw>mcu*&%61m&`U;nd7Ac?0q&e0X6rO>EgPxwL@K8B;Yj
zGoH~hX6l`nE3K({?v2}dX&=h?yJLWb7W=aPJe^dXnwDMbY=TLTB`P}StsH-`Q@+Et
z_q}W(!lL3A;6=w}rPs=|L#Hy~R~yFWZPjzrvdXG74>bUr1QZShiKh5KtdW&%Fl4Rp
zGA|{os)=2IiOoIUUFHjdn=o0B$o0pU2sA!Ml;f)6dmbVTh+}K|<l<?KDy<9qdRR)-
zE4*@+E8NCMe%>Pq;UC$?^IR`GL3w=hp5L~Xec4w3v)823e0v_WuwdFmupW10yVd*Q
z#+%Z&Z(A*M=<RyX60a>RFYC=Ws?rO&?v7V^7Ny*3FyOMbNo#(2a{2UBIW^H?U{LN-
ze3d#k_j%y2_r%0P#da&7%wynjh3P!}-b0k)2luo4po{d!`($4a#n%wS4nkB^=24h(
zH}6+m^#-54&H@A{FN`o?ZyfsWaF%;~(})jtA2bm2@0p#-4v`bOwgd052#5!PHoAw}
zRMUgE;&b80IHwY!O0P|3nNbn|F-k0ovhK_!>;|3qnrUPPziDoErJ*LKE<c<hmvVYK
z7g7^#Xt?FkWw0DZDdk`u5VRXvlM4(sFd%^YO3>h7FHH)H>h|y8EED7DUQ7O*VNZ<k
ztZ!3TBrsMVA0N}&mD9ecv~-Djua4?r9M_ar!}-qUSLGCiA4A>8t#*lSd53+q;;9@~
zq#QuL71nHePp<(1TNiCcf=LK))u;s)Wn_f?1_H3surF#`+I@Kqarh51_W{c2#vO$v
zLyr4a$%58QXpv%E)AqwlpaRKOTm#F)fQ^MtWLrkS9_qY}H(C(m|GPltmDnc$O3oE-
zmy(!CSWAcAYA&w7tzT#Pr~f8kN3%Jj!<bRpxXoy`5gyLQ+U%JDtNdHuv)<}OKMoJG
zb0gpAnZ%3CPqygV@x{ZYbc?yK2nd~54dmDUZ18f5oyqjw&2~aObPwJ}{5ks4616xi
zu<YN)d`%NK_N&x8f`*GH4(orMrWXP$>u>Ctl&bNY4O%W3RZ>!lr?Go6kiGQ2&!;VK
zUTOh|!_w9BX~)X(?J7g=B9T5|O%Kmi|62kGc+g<Ozc5ENGZ5x@ja)PK%j>2!3!JKJ
zobg{tZ>S#I#&@=n0&fKblj_aK^;KCcENDvy@vjDA<_v6cbRjAadv}4&0_*!hj*;3}
zThj&B*Ly9HcRTj`YAoB8gXWzsbZ=U{&FD2Ci$Hd&WMG#~dz5~u=VLwNINycEMI9=z
z!$76QY=mp0ztc%=`!MFV-&xqo?>hJ3s|-(HzDzESHai>m(YDphWw7X)O3z;FyZiPU
zv#O}h;IuIH$QLVnwa~2@cR-jSC_JPJE%x(h=4@+g%VDSgVl^eSTzXgh-q6HEm}Pc`
z_;#x{Z_C;~0FoVr?0U&P8$34=bIZ#a!HbG-kIR4FIH9DZNNw}O_qq&K2ab~S=j$03
zRv|OBk4Hr$NN;u!)}y7w$u?&<`7&MmqS{=UJoDss6GQ;twKDKn%2Wk=HgBY_uh`c_
z7~QTh3qo1gcpP`hgm7P>#{@4d{R$;xeCaxZ^h{_mzpY2B7Z!AwzPMA;FnJ}SQk4MR
z<R~%>F@bkH0w`bDXKu68<;w>#p97=V_XlBP_<gpJtoqYkVnEYl*@u!X*_?dmQC``=
z#KbhO*Bm4RR9)eJuo24(P)cgv_CuFxCx8%t42mr~AHKqB8<qOx#hU9sGIA3-Y?7Fn
zVx1KnxbNTKG=m3(DOHf{GeC;AhI#~Khd#=XThy!YVu8us^;89-p#hIjav;#CraM_6
z;cQC5-)r-8wLgAVG%QSEnvCs1ZHZ}hheXi1B%3-9kHcWDfkmus&rTb2n40S-<Ux0P
z_rl4ab@^Okl0#GW=0Ee=(djEw75)Um{yGr_8RO1mjce&tv{~OxQO}(#HP@X@#!F<|
zPu3NtQ@7UF|3_+;W<~b@pB6wl_4pSJD*ep!<m6~~Z8K_WHPHM#s8p~!;V@lEl`CAM
zY4@z14OXTo5wsOn5UYCk3U*_fTF-8|moXh+=MMC1+`ibQR6F0GCl-EGdz7b1_r~*N
z=+{Tz2TSkuObb1~tmtUCw)E{D1nq5y%GmX#??y&Ut)lFE)#WZ#&~O>jRNlPR;~#WV
z^oUW+>#2Kpk8pa?S98q^2d^1zR2#gI5j;9m!PaZlwibWMEKo?~2^(8zTfNug0>v`8
z+>dZu7!kjeI@Gm5>l0)i<}7L&!s~9?P+}3eaCVko&3rvYNm(^-v{^!De(m7a+1tBg
zg_T-BvQcxp%?8geGv_~8!xYP^j3xnJ8LOwiv4tJwuCy6C1w1aAV2=VVw_oUNO%|8&
zO`M%cLO(OTkl-{uGqX-QI{WimIqc=ND_w_U7zu*UVn}}R3v!JKH@DrJ$K(vxdU`u2
zin(v%uR2EmW|h9xhDPr73&ENcNQ2SE)G?2rn^|EsBtl~x&XE8>iO6Y1I|yKPAb$Oa
zpHX^KbinHiL=|&sbt$%%H5&{049DeL&52%<o*3V)xecP~+wkr!1Ms>!tY-Mif?OyA
zDgJ`J&z#hJ192phD&Zzcd9EwquqR>HE^+4ZdE3{|Hg7lU-IW?MY=D{fYGa+GotE}t
zSV3sAki$4TzB;8uBGeTSHtPt1JFT#eaQjPOM4}5!CRdEpWA##O`(vl^C3P6J{^?<A
zub&pJMkw`A5G%MTJ49J{r5uP7gFrUA&Tr+=(;;w|(VQ3aNa;k-#<Tb&ZG>z>LAghu
zka!poASN350GJ+`QWwT@o;BQSQv~@;pyL`pPxHWG)7URz4W3in5dJR91;B}dKopYd
zQj$wk;jKFRbdbrC%Q6tl7ohiI;2ub}IQVFYW@dCu0<!NJ-NSLn0?7wF@&mW$I`4i0
zH!}ke$QAAIBz_AVGvF&YV50g0uXY3y0l8lk(&Y(QLLSOZ$X#vV=$8OBXt;x^jV#sr
z%yg!W<*`p{z`!KfEMx@jjMtc{6(`2UEmm&hF%nS>V$SxH9cO|U&z?93WU^ocT_DlX
zG#TlukElMaF&UlAj;^KZ0+F7uf)GZ+MVPeeoooq&-(AL+W(ZFoW}h6{a7Ck7KgHx-
zUF1c;WA|X^6OHqf4co(qg<*A%=Kg^7hRij)5CZviHxdc1hr8sMG_e}rJSM5RWS>2_
zwnp2B_xFp9?C>~mxydBbEvXUC+Ovn3@-FF-IoG(1DEHlOEA7^=`!m*r#W71E)}{+&
z%#*f{x~ftRZUP~+p~@bj7~USc=ZHJw<uLn}SqXp3w6wH87m8Z{ySbzty|giuHsev>
zrX5)VME}lPl(LBV<H;m#w&!M)1Nz_=xHj<5ukU`Jm^f;M&iAVI(yDSup5)O_x)@gc
zY75}Gu=T)C<0!Y;)_&XA(0bsX&fySYGWt~8No={~=@6{UzRk6$u{X~~^yt{nMCY0P
z$fT!aUf-CjKF*4XlduqtiI{w6GAhRS1_Fq$uH4#Z?K1R}Qo;Do%L3hez2l}+W?xKp
z7RRb!LVhn>U!ok#4n);@l#}~9uI}CiRQ#SX8D`SgwvT5bXPjq{z9lclZd|`!dqIMS
zl-3fQ+N08Xg}>@Gublp{QgT@zE!|F&Sm^!^fcP}#$bd4Nj%?~zX=0UKyxddydS=w(
z$c*^-pGBiif0%qv`y<{e1-lq8I}!pG^7c{CXtO&%n^?GKxq_C}^!97<Hm8HzUoL|R
zB$Dd5z%#1Qz_0>p@W{Z}6SCU_&maV4D0in})tjB0UtyQ`C9r3}aeeLPu2`lBC7EL>
zLM7EPfr(ccpHNZhC$Cd(Wc@h$`0;B}H=u3K<VfOMNU0AcXY$ZBsdE`Fej)Sbj5eO*
z%x&$X&k(3T*e{j&f|A0|1Y!!O<Rk6<7hk>|%8z9eII$dLyY`D!mK&EchsCmrh=`ml
z13gc?5^;7STwy;vgHJy_S~oZ|8atgUeyc|!tLbguL%2oD7fryghM0pv(p=xs9P9#e
zr@l|7jI3_w=b$9I(NYweaq#gS+Hoa_gBp~pyLQ@!AXymq_o3#`XA-yAqZJ{Og|7R%
z{Cx}eE>;buqf~n<*m>7y53!u`G2)slGC|Y*vc&PQdl<Ich>Yqq-#fb3Kp4YYZ@+^Z
z>83fRr8!HiE$Qk70<5_!4}RXc!>9bzuk}w~Mlg+yKxAaVo{w~zw1_8x3Mb|BV`SWR
z2pb$Cx1S=|gBvN<PI%u#1d_hhibh_UHWStOVOtXr$RBvB3z__ySP>uTH#5dp5h1v4
zjU{t14(SarjKDnrc6A)>TR9D;SXFtQGy$J1&I@*3UnsEaowV(iOC#~2ulMHI<FBXr
zxF!>Io70>1*hw^xU0_Wq#>V0ci8Vzu8@-NiE!X*m8|DPxEX>S|$#%*OA)QZKIAC-W
zbjCDzpJzK|M>dxph{p$vP8<@DRp$wrj+6DCv#)1O`tTc4nf4{wM5G4|bW>lAYVz3i
zJ%0YqT2)h9D!vX_l93K$h(lJ0-~?1a^OR0wfgJh+HUlpx4w6u*xp6;RL}>bF#(onL
zyY2QO5-|}CM+tTY5SYxK0)t@ZQRp%7hyf5uVNBX)SV33`u~TY?{ug(gcB~j6cec}e
zk9wv_*P;#0by-^!V5XD=_NDYI`Dt;E2D`rQ@V7$GegJfXTE2Rkhe?%Nou00|s5fp}
zF!8!OkAdN>=(E+tx}YiVjqC8Kx|zpb>ojAs*Fu%WSDyy$#HqAb->22}I_`*hl!3zA
zhM>^^O}=0A=KoRk9nf%fUD!lK5M&|<f;S@}M2QF!bs`xpdKkSV(Fvk=l86{RB6>>@
zy|>XP(W8dJXrmit^ft^W|MkA#|F89*wXiH}+<VVGXP;;9{XBaQ$a2En&UVonmEky#
z+J(c0*pbt&6epP`zu(eZNB+Jts}qGofyeGiPA%y&^4#@Kb*rNmet<Slj9ql*8LVkP
zgG-EM&>igZ*DqHdW|DEMr<aRy5X>RH*4_)K9GAm5?}CzgdW}N!q(CyAy{54JZh-3`
zY}eVlwl#MZ?4yxTq<sn$Ltx*AQ~qmIY|4~!;iw;LC&`x)0)N6n!xEpughQ3m7krGH
zRu+6EKy<vB0kC1!$gm;zFX#VcM$22MI6yB#j<28}4%n@lx7W25`yge$=dUsLB@G1|
zAzMvM+vDFt&s6kEuvnyC0`0hbM&F6h!}&9*kw&`V5_^5e-ddI$!uWA+?#I~%CDGA(
zKN<>EwEF7)m<2H2I{1r~-`L&V6{6vB-~20?JM-m7694u6y@ARgzyNBTyvA7WEi+p(
zM>a3?XFmNa;~62jhwB-DNX5D2toZHcN<f*_u8%c3Og5&JqWCXPo16ytRt8b(u`!H3
zt_rp0Z^c=nQd#)B6q#Z{3IDd%lOfq)BLI?bgUP23;Tt29Le_r8qyBgd8_N@BjV=U=
zOmk)5u3PE;lo@3l)P^nE{@^KFRcku0D!7`NK!MMDI!FWdi9oOw?vh|tPcz77@(U^+
zx;}}z2Yghg#^Xny%Z%@GV91CS8jU_0ftcCaO$2iVobV`NL#p1w%Pf8mVWSmgz3QC?
zgbV=lMhP%1zQ(8P&d=sfH<6eqeQO-YW&@t=WX}=Y@!|dVCf!!Q+V?(tiN{Y>>UHb@
zw%l$yYeuhBz)tklukUVtK;a3y;B-kTYdPBwU~a5qbF<?hT9-<>)RPR<zUE$dBmXyX
zgA?Y)1fond7S`IhvO@C@aHnKTxGv)BnN>!PQC>^h(3Jo%4Vw2^!CkBvY`-ud!<{sK
zR9mc@rwBP<jOhTo1R<s}wU^*z$|IW@UA+zke=d^5Pw8AH)iCJ|K5dcm<g9V0dp)HO
z{33llEe6sv1u`~IiMbHkk_@vc)2)f>rnWzfq{B_$JF}H`eUdY3DdPIqE+1@Sp{KJS
zypVl?TiZ{(kCe2?)+&047j%D3sbRElU15Z;{aY@Jo~*lIlKbIbe^`~=c@2a^voO8T
z#hszh1ME^8@AvQORHWy7-r1|iIQWp`vSjg#_V6wrS~WQTfs57Va~SHn7DaWcp(!k0
z;yZo>5F#tUP)P&>*s$~Yr=(Ku%m?ntBz*O$7l3!3F8Mwt{xR>t4Lo40ee`-hBvQc?
zT6LfCz6^JkO~q#F`uZCx2#2&^ujTPc9r-<X;%<P9C_BW_IPn5LdGGmTKu1f2Xp~Gw
z3%BH#mVnbG<!}fcEoNHfru-|g=TM$n5^%%jsyM4AnlBzMaAH&uNIg9bEgg24Uj!U2
zJrkK(LSym_>z?Grxmhe>*Pq$2-z0flrwcpd%P&Wq<m1VmhQ=S-d0!6a-{j&Dps68z
zzIdYob%)(&to#i&<0sy-_VL1dFURHiUkx{!Os;F^YIwIfPOMxU;>Fq|U5T9V47DCN
zn&_yj7l%uN+beizUD>q1=+}VbKb}kT!Q@o)ghE$kHtDi+d4uPtKEY{D7c7$o12g7G
zAt!9zUhpG;eWtn4H?l{8?e9f}bpev)hoXhvVVi%0UM)}Tk;yB5F8ob<>(|Adc`W5U
za`Qi+9{D<SNnW8I3r=&v4w!P7<2rh8|LN^Xln73epSjUpUeerqLMggsj#IUgZS9>*
zJXW)Ebj*;sAFGKy@?u({1^(68__*t;P`-6>!xC;uZ^ApXTj-Ij?zsF~z+#uw*OAkh
zR%&%fL}POw6(h6|IZuE#>7JI4<6D(8FnG`QUxyD^&emisj@(&)7Nb130(<_jQplSK
zH3&!nnI~zqI1pEJ-<amcj9EPm?9~I7VK`9biqgl^T4Zmbcm1s`VFbVCKcbG($xA@8
zTb1<m${Z9ki8jpkTxce0aCyC%SBPg@kA!kBDR3k5nmLHs>c0S@=>M<b_!(rRP>|CM
z1-a{H#V1M_!mQ#N#YwMl^svXqR2#S_nHO3F&ANbe)OQoO^df)5N&&lLujU?e{(5Gb
zGYQM-yEvzvB3?{L6D6^uKFRnRceH<kh?s4zCYVA8Gv$>I3*>2=R15=WdWgn1k5(r=
zzqdct1d4O$9DbEz>eKsl!wrD!!HD%YxM+k;YP0(8hkEhv6%mO0@ralEKQvKUTtlw6
zS(nYulmwmd&)a*~h`*a)Zpd`h_+ziQ9^lt5Oo3_|!eYM1?t)J;cU>$1W@h%%@~j^;
zsCvNh_a_P1Np>6nEsAN)H=Nu3#q8Uf#^Sn^V&UpkyxyLOd`hZx!B_865I}dTdXCg$
z5wm)@cVB8lrF>s`83M~z;8tUa*@(mbtGUvS?vQrA1FY%_dD#-7{{02L%6aJNTio1(
z5{~$B0N=j`Ad=S7oC5S_dc95ig#QIq;;R^y^Q(+aJMRUA^ID;3WbcU6-|c7Pb#kYI
zNOH1&v%S9t2lltT+-bp?R0t)y%cr_C=@9CheCT}_@%COl(4eCjGd$AlFS&qtUEL~X
zgyg&5`1e&)Cql_2Tq}@pKJ03-#huAJJjQ0_F<gCtqC#ABtOi9)E){FACWzJ`^~N7F
z-8NK*?S<NWkd+PTuKx-+Ws?Wb#U$Oc5dx3AjUv@we#qJ(!ZhOYjRe7bIrvvrVlkGb
z6>mKO8JQEXe*)nz+0$N>s=Krzlqf?v1$)GSt9<v_+1u&92(ER@JcJ3Fjap9BIL15x
z8}(H5mOY@#P7z*MUx$e>%%l`KwhOX2&!q9K=?>Kc`}^pW^hO2W8a^QF4#Rd2IRS2m
z1nlm|B$h%bwPZ@}6C=cmJLq_1RBE6Xx%NA)pGPhq#&mONfMO)WF!pyzWax^XRkIN+
z;1_0)pMh~u9niqNx+;n)ypeDqJ#MCl-308_SO^dpw;yi&ZRhyfit{cQl~6+_YW4&D
zNnY5=6}$`RqkE)E|6cb$4>S0lhd?y8_3~fORyaRejRoxuDyfC$sj(uBa<d@^D~SLf
zmK)U0I4Gv12dsm<ah+d3qfi=26tci1RL3ClK4j4*))l8aN$4cR>e1)@iU1Pzoj(RK
z|1>|Y)6IH8Ks^FTbnQT(F_>p#n1orA%ad{Q6}#>pi}>c<yfmQzJFBYdxroBwWMukL
zGH^})qE?n2w<2Y@)mqGLJDpx>t7bD8jv14f<TNBlZ*G8Ojhi&VuQGOjhJ7p1ZVBIu
zr)TE7(9-0%JIh;qi6pg4{xltt(&dc~l5aC*y<kXRjFC&Co};G%H)GQFx14rmfSG=j
zziVSktPcSIMM>PvL?<)YuDkdk*x!HAQATmsoOor){92dg0??GGDI#zItNVv97u#J=
ztp6Y<QHnZJ^)pCFk#zN#<tEI)zg_?F!n-hx{^;G0C-l_J`C7MoPpV~q><o&%hTDFT
zVCIST`-G@Zz8u!f=Xg_z$vJ4V<3gS@cdN+K3jXqEi8}jMiT3!0<s*<x9vTDxzGA~9
z`}q|&-}v&2H1ERZQYh=SsIUtTF)*GO7d20p4OwXBa$iGF>YygMYHr?tbPpWdE3Keb
ztq4=w-xa502GpXagWA>)S+^qLYV_esN&EmhGLvB$e(Em>>toM;@k{7#<^V?$D-)oX
z(1BKYd&`lLaWmbQ^w`bz{SlHe(%Hs`za7P;L!)SVCqQ50mUm<TXO2_+E&RUFz08o?
zXp=(s5ju!|`Fk24lTt3vYP~W<5>)Hn>4C8Zzorh(JMp{E8lb9rNosQ1W*M$kH}Pxv
z_x3f7@{d28gxXXF5uo~1#kf$VC^6bRk6vkwc1gGGb}JtBbdP(W;<9$){>h(n?#%nK
z_R`s+nL0_(IIP(Jw04B7PqL9b;Ob~0;uwKD6S6MC!bJ9-kLf<K7^|ks75lp(T(FDs
zm9~16WGXn5M^-$_N1{Z7IjkX{z{M0Z2YxwXzllT>MzblCYi#wNHMp4LJcp@SKxr>s
z`GC6nzHk4hGX0+XDBcFi|I}QuW9=>=^r1HQH5yQUYx~}rulhmQ&+dK26rD%|1(@CR
z3IWh*nRXe)jifyBHuPoRQgrrV90aU&Zoix1&yPTVBL@IQ@OREt=fW1z7tT@a=j&<W
z#Y0KhgGp<IlG<(WF1JUW{3ZFD>0<fl9EW5;O=6sQEDrWj5H3{ru|Ah_AG}u+rc};q
z)SIf`a>%Z;5uJ5tyTYZ!TascjACRFW_58JcmL2H_-~a>TACsl5q^2v*@gLYU+4Q5X
z-7zOoh;M2zg9I_1%fkMpCERiOhSZ{FS%qm|n#Dz4!kCjD1A0C9tWBdGCnJ-y5uobq
zVIJLM0v+(T=ETA*l$@Qy5HzOpqLuYmMxLo{{3vL9xaXd@0MLHyviL=`Lg_EJ8zn_*
zkFe6J8;WLjx~X#q8f_(9^a6v9y^{Ie-jqwBEUxAr66WmkX4jb@;LKsWJg{*8zeRnb
z5+}nh55tJK%77%6zAx-H^?THXCINCcU0d3Rl_7SemG1l}q!}IHw%FW7Cry)Z_lKq*
zd!PJ21J<;ks>T1yRk=)wp~6A?^G3bVpn`)`3gTb2`6^)ZBPD6=ELdxR>uGX7DY|At
zXC<BFTK$FvtIC<+b3-4V7i(>``Q4F!R$o)o<ZJW4vq(lJloP}Py%ECzI&JACr<a6f
ztbTn<7iv?VzMQMyTU4^DirE-(K^IIY_VhZ0=A{HAmr#hbd#j+PLj}LdgJR^sEta)$
zVr5K)L5>ihFj26_ZL;rSVc7CX|27BDPvX>c8}oQPrr1GWKNq2xVN>HeAQz6QF!R|v
zRp^~E!K&w`3Lb8KE$<78F&92YK>!ABp^rwfii}yEM;2x!q~*Dx`^w)s^;cL&j(X3n
zTgPo^H3FAv<I4GT@#JAGboWSGt>Iv0{O9X%W|K|7?YS=i#^-nb=P8h@Z2dG~Jp#vX
zP2_5XeD^u17q^)%>VZQdZ1!*=jBIyvbK;JW@PnCJ;k9JK1&o8dT7hpE*3i(=($jBF
zxc+)izcAy4)LI+gXJggNzmX&_7SYYTJA@oc%n8ia%WA!DE7{EpWi;UbZz?s5u~{l}
zA&i5ZR)A_<1#asHlF(z7%6Vk0L;V;`3`ID?Eq4-+BdKF^`op8H4M1wpmS~*@7DS#X
z$U!$rYt#ZP@r#Zp#h3@o+U_m~8-3*Z2sCmpbF+Tk<R>627zhmw>l&f(9l1%(TR6LO
zne09ipr!i&MN^Cfj&?}2m}~;TD6aRMtf_y%B7E)WO~<-_5Rp)Z51#gfCcQjA=FJR(
zn?(TJz1sHEQ_sI&@1ckTB-`K9LLARstpXstv+d62DM0Mh<Z8Dc<TU(!HIS|FKR$>i
zl{ROVwxf81t{iT7{S$!le%fW*KF2-%Ai!kyw4Egu?=QA6c|_jS5Lch!)Vycflwjsp
z61r*+r1&<CMvLuh-NswqA4B`F=sY>>$6e0wg04A0H^cc924GiO&<I*CehBOe1`9sO
z54ZfO&a=iC6ZXwf(9!_C>oBr<*P=q!8?MXRD2<n@g=&%mXGT$SX4SuE215AX-SxDu
zr8kGgN5@L}3+F!ukCiGC>jKsC?Qb%jmACa$9UMlo)Jf(Kxd7RtW%Eq%%YFNF$uF<h
zzpd1L!h9RnpfY{VYPak4gJ}aEk7@FeNTCV`Z1kv-neT)x+k0(JpC^9V^f!ysAKDgV
zy#QSz8|&lcYDPDjou66PbZi2hWi@mE5fw$-;z#>9RT6|8j!uzgzEQqlez%Eb(L*=<
zkZdQXzdq7O{yQ>SV=iT@(br;|Nzdx?2$0|fVxfT!XK2^6ljF$cv)xz{S_3*=uk~&_
z={eOr#isY{Z*z>Pj6joItAh9BoAX2c+IO#84FEP}fa6qF@mhe0s|KNrejGPhdr%|w
zHR$7-RrS_ikzBbT1@GzU>EQf)zJ4jsw|fZayRNM?h#8&V;vC5kMeLBGQjEG4ulq7^
zVSzvjJV&P<(Q`;z`uh4ZW#{Cu#S5Fw%LGn0cCcfLT?6-b6pb{_PxhK*r_;r9*5g}C
zPSzxmW{g^9d&nYv<Mm-dBLte;w@3mc>cFVy?n!M|A<Ld#Ww-Ue%ux8&e-?ZAaiC5N
z{#tvS+Jk%_YJ8pFi}@~pyv43pF4KmmzJ9;en@sR<qmnD!9Y`x&75hb~m94o08;(K5
z=N*S-a<{6;SozRftFT#EFaNDcBW;w^l8HqT376s){7fCsdKS!-9yc9Lmzxx((x=s1
z1)gtUIBHI+$G7_!W~kipyN1r^D0o-N#-wpg+ve1Cw$@+y-^4?QXD$?YDxf9pQtoI^
zg22nA?KP)~dRM(t$8*Y{XU~%ts~u+ed`BaMB~L5e!}lOiKQpB77(QWtuNY$5v{)VA
z$G&{9K59^m?>%b`Nev|0<(&I+HL**(w3Xc`?VNcy-@%yW*m(BrX})*t=VNTWTbBnR
zzm*+Q<V4&fRLUq~#V3|)*h`DQU0dv){c@p&f6$M2GciGL<+Aa`8ng$={nQe3bi&A$
zoyQH;ME_aP?{u&0)%;BVpV6_qEXJY^A7cRJhwgECrAYTzWj5_x3W`o6X1zFUdIqBy
zsaCwuLaOAeX6nf=GD$hGko}q~LA|yp_$d&{X)<And<%N%HTzdCHw(y>u8$0yJuiJB
z%5am_eQRj7g%X!eam6C(N?3$s4c#9HfWo^CFwJt?va^03R+ech^#!>JZJ@lv)-Dh8
zO`O%JrASw}?u}tbevUnX5gcU%;ypJ;fA$2UtQDgtw!9=h0!&TQVati(5a*{*GcfpG
zP5_({bz@NKLXF}ZifE6$W3^eTJsD>nXA+M5sQI}hS^M6SPL8InPN<~4X!QTitX{7~
z>Avdu*<0Gzq+i(T34t?9P>pt3?vzG8CYHDB8XCklheU=s8-Tz;Z38CHegke;_$5B!
z@~eok9&Ji3V^PsZz00KXB<4HCVc3AC`ZJanEj2U_+I~<h_Ql3B1lrGVw6nam*5Pp6
zV?_B&pc-Jq|Iq$potDDP*!Yr7EYE(YNs^qL#dM%-VaMe%0Od<Cgxrz)^dM!vpZ?W%
zF0_=!)@t&A-`T!@DiPaR3IDKQR~4PRpU3Wg<J;D{F(Vap<gMN2ox&I01|kVSg6v{E
z6{%2tu~KK~{3Sv!QZNd41xU~A?F$4ra#$}1nsZW8QjV&Ew1v_)?>U^K@I9Q4oJ(ug
zx*kF^?3|pL-Mt2ARB!LzJ=0;22R!&LfK1*_Lrc5i(Rll}mfmrsfjo3L`)N$*%zmq0
z(lOf}B+1~X7z0`ZE<Vx_wPN870F!^w!KS9Bra5l=zb#8fHu>kFfwuFWP}1&jG<pZX
zLQf4EPX~GFB0glAQczi`U9hyzf<h<vc6X6sN`o86m}yx@V@t7aY~Eu)y#Hb_3DQ|?
zV`Q6X5Eq>Crb0#|wo|V6q|zH`74YrY9&n3JWq~3dLUGE<u<#^rR*=Kc0DL<y+}Rn-
z#tr@hUwxua?g|vs{&#vJfAtmxE*H#o`@$@|ZApOl%&5f<?q-bP6I)OkYgdL-o0RI)
z-JUk=R$#-EVn^OW$;s^Q^77`=g5p}<ym_!ZBw7d?&o4!#n?7Drx!_~y0$t@MCwmsY
za^|K<>1B0yj3)zrrL&~|@K{()oN}ZKt3mIfw2Ifj6qv)!2d44OE3!H1r7LSYqKg5(
zeoLR?|J|FKh)yJFm_kxfzfPmTc;b4*Cm>P!UtNIn;YYoVfIr)b{0(+L&8-Z``pL8r
z=`hb<4y<R*YZ}6(w64@?(2OP6{>;=fg-MT1XV3i|DG%_t#Onc0mIJSqzxD%$G&lLA
zJRK@|da6f(l$}yvQQ%KWYI7VfPy|VtzA$p7>m2|QqAhSvp(}D=wmikZd8g8uLCP7S
zT|d@p0~0GR(q6=_KLfeu)?ZD!C+8H%-S?)hMM+u$*^s=OkhRFvQ52@QToYMZNEHv(
zw=fU@VM}IK3n|{-Tg{pYF9x4dk>!mfi`mATPo1}ADH<C-c1W}ikPb+K@ctd_yf1u?
zl8A6|sSq?RIx}B@rlW*SKEBeTzJ1VR7UycD0;AYpNqa|iumA>s{Vw|9ZFTbjtZG^p
zU&s%LatnYcE<Fz50Fvdk=zPyxNJjK8ZOrxe<T^TmhG)NWU8H`+l$_GwRSHF^zn;Qq
zlDx%0Si2ju?^RA;z_wQOO0BGfL4T@-za)|H_H=g5nSxK^frK}v)N?=;AZGvR?w*IE
zinlrcZdHWixR2e(w`9?>#}B&(Ip`d3eVY-un?1YzJE&hJUSJ=3xG$<4$u`{dh0^^Z
zIzGsoY51fj0iFRU1c=#dSHQ{7VGN?KMqnT9jKn4Va+-Gn|4DdtjhcOe2)i@7N9Wbv
zL4+7?MdN`qNr&y79grt<<|K+k%KBi%*?^b_g5-SueA(>gH^atVhK#xAR5e15QsnNJ
z+DlC}2PHpWt0S(TdSc_}mlrHCS^^HOEs`79n!b_bnubHpiNR*~JzCa%;ilt&ER307
z{g0nD8&V#<PS|o+98=Ke^d>^T(M!P%Md!}eSMOr#e@D_vx~(K)E3Gpk29Yz)_0Ln7
z#e1fp0gpz!C-zwnbc^)Y1LwOy$kcTK`nSXPg~8IqBW{N2*f)h6y(f<9zfR1SFL#_(
zj)>k{O}ta)oOM!Gf*pviJ;Qw?ZPp&ED)($o`TufISr;EjYpC(v7%T6?Kb;a90MftV
z3OAzy7n&AIn;vc^789w-#9kt%^JOfj>Y9kM+21^>;}LBldr;VL9KtX3ei&l<0nST0
zcgs_s{i5m;iN;YMa(+nSR814Hd^0;CW{(e?Jj6`l)KV7j=4RImm{ku4c6<wUCz(k+
zb{)wKo{l=Xud`GL%5qdW-fu$U^IhC9(+`WViQDD86Q4g@oz{704SeIv7>35Y8++GY
z-TFvR*dSs(^Uaf@90Vd;q_Tkaq*z_qG;^N5w<zG9RYG&EzBZnD{l6`)^Sn^l*V^+-
z0}5{NO!P2dX1kRUrDyYLtV3{8#x&q`3816(HY|p;CJo<=>)WVzB6(mWDp5PWJNb%1
zia^AWI4?TDn@klMmNn|kBi1f|dQ1OkFS=Pf$8YG>c&24T3KgWQ)G<SSsz2>>ZJN|6
z&L2AIwEOK`?8I*^f`Y=wv)T@(9uUBpp``RMRY_@E7dn*VPt-cyDMk72?{raV=vA3L
zz&P&vY|r@`agHOmGGBfM@VDc_JUC@(iRDQo9SZeqQj2kZnzY?|XD{%me_;P$U&eN=
z@*oB9AOiMSPBNNEWYGl^6J#gFckkXwyKc84e_d0v%(*yENlWgxX^E-I!^=HPlTOb|
zIf|}+^l9&glQ-zdU5CYhzh4^44(9yl)2DBLiq6NKHEz$}O`?wNy?Z>p%xd*0ccXIP
zOR5BZy}G2|j}QqF@92P?3L?ZeQYb}E`=ME_gH<08c1`yu$(yICy~kv!_AeQTZGSQH
zL3P3`eio>EI@KT|$o>wTJ!z=%yagJ0dnZJ6(EcFTbSo=~%lX3t+Wj`}Xvfo~gr&im
z{&c3}jdcS><5=kF)Se4Z<JstEW}u_?m7))$Sq#!&n%|CpO{ARZ4Ku9sa)wRJ;|=R`
z<YBOtyunhZaa5og$CdCGEMHy?7MgcrIK<t$uw;%`{!Y;o#CwB3)RSE@y&3+ZGg%%s
zl5zYqC1cA(H-CFu@~me-8Afh-Nq#!#Tln;K(>ahNiQXbu`m|S6koSZT>BH(v==$;l
z$)&2Q+9kd8O`8Gvc2E+^BR;z7Ch2@1|7zuPBzsGsY0cf;?YYJ=;pQWOIdp%g^^7m+
z3-%tyyH5?d>5JPrf9oD73=uXvT%TsDc{lfQrEU)=PF&j?0GoLQSnNI!mPOU^6TR(;
zB}|rTPDA(ehrf^H4K5&!omYR&4?gle+@vE}bY~xJ;Q}!$4Nl9SX1q@0I5-1$lOpxZ
zJd|^Cm$0N(Eqw&G!DGar%I5X(E)(i*(?LgVVjx3PT|3dekh`PtiJg%=VXKbC)b^p{
zzDi%sTO#*j$sB>>$-%HT;kKtSW~!|0A+~BYdV6P~L%hMf3uP!kbe0=7TIh5Ly;{1k
zXENIpf6gW`U#>VAa6HZ%<4-dC=GiZ?ye_?*>dG*p@q)j1tO4TIkE}Gu1&|!f8DS0L
z!dqjYXxW;%PbF#5wdK8+zc{Ki?7ccA9?;%k#4mA4mkl`A{$f2`XXKWSng@e-U!jU+
zhQ~VlZEZ3bM%&c=-Pe!uUOx|MG@9SIkHpZyEL{F9>A1fO&BzVw!jm~xV|;rDM9Bmy
zpak;S`mVgJ&InT8OAiF%{?)r9PP4wc#kE~wg3G)-X_FDcxh0&B1d2ymoUA()M<bub
z#tXU%H?9o%BXg&!8XPUWG$XF4sUqC_kLG*Cgrj}4=Jd=0l54#;wqg$09}7q?MO2EO
zd-okQoIAfjw(?9qzy4UvhVrid<F_kcLdd6_O7%)0<9B-1J`^OwW$KGP%mHTB+SV&w
z@eU2Ifx=Uzx*f^3&fOeZ@a}Y;)_~ZpgLUmts}BjCYGZ!B-E_Vx2DQJ8kGTB)wBx{&
zn?@nopG#+I`C8BspgGV)k!>rA(rlvIQTN>hAC|#$WlC_XZ~!W0H91-g_wYpK9(p~H
z<mp&#ZuItV*@?yl$Zkxw&iG5^o_=<>e{RZl`o_ifcjnjA!2)l_&Ob8s9J1g1y=Efv
z%{){zG#FS(Lwk{%4a?1Gg303Bb!kvRlHhK0!<Nl!Q?rM!UX9XnXN=FvHu<OE031s>
znX5`%Nep#D#SJfHr%N!g{&DNvZZH&1JXNz$Qo5SXelRp4#4&j@cfcN-X)71a7(K9o
z3H<e&72Pbm8Lp8F+muQi_c?=9>Nb=v>}zZEN}LWeV(L)VFAqiZfdb@XZ&3^B3+A~8
z6$4E4y;Ksh9`2f~Y7zX_^-{6@)5B4<eWPKU=3JYqf?2O^^Z8ICe}Tf#ErFc%PX6%2
zmM`oj{STveaYmJX$HJCskPnH`Grrd5wrIG!tu0#BZnGlCx02pzk`)el##Q^kA;H@#
z*>OlaAuf}MGB+Ey;IwpTojlEtx+HdY@eS#1vu1iu0hg$zPH|#vquE9l$ewwGp0zGX
zX$FZL4r$fecyu<bU*~<y^HW2D;S6;zvpY5=#q4HT-Mj52s-CMiPt`g-1owp3{dI%;
zsg3UPvQV-XkY=(f&zyJfVDSEvLo!$TyKgX>=oxuR*LW@V<@6af7L@7?>7jFY>ZjSH
zSNNklw-;kI_=PlOf%1C?e%Oxm2#O!HR)N}XAwC%&5;)x$Zq|uznLV8SnmG`h6gV=W
z#x%hT&U{oml+FPZ=JQ5t)zI(l9P*4wtWQ<s+X3aX?5T58-Xh7w#~ugZa<010ZLy-y
z3ncF(5`qwjJRVw|zRc0XPix$8Y^R2Pb!)PG^<P`sEg{!j%%KnCpHSuMrj(B@Q^(L(
zO9dmQkfrY5LD#w1-P?S568Gm8`NYaR5qbOqW@oO0!f`21syki~J<G9re_t)KJ3_KQ
z{-W2^-v4k78SPBMaMXMB32#MJJ*6&XG)!%|ga7-3#*&<l?$n;+$DJQcFQ|9W^TF-h
zRs9kRZqek8**?zw{n&=((;7G1wiXH3H@mbnYNV#=4k8_Uc#v#v>QrU#KXZJ!`@PAB
ztKanZh$JqSb~ni!Ce-umG0-NuSG7$)4oyIHnsAh`!zp-|foyAxYvc;lF12d-8wIeE
z0{WC3NY%s5_rT!HU(`ThvG08<4MCmF#Mi)1PHw*(mS`1)I9dAjd;QDt9&S`wUESkQ
z=DdhMf9}=G`R0Q|c6mr?l^vD`d2;U~CR#<8S#gWK%4X9!Tg~8h)U&sSsFK>VWswcT
z<3{?01T)=4$xNHc693nex@s!Lt|LEM%j<abzs0$01{hQeRrq3}ObO>DCQ{+3@G5&u
zbMMkbwe1$rf%gE989ne|hry;bM`q*~Uy=zYLGO{_Lc3m&;mZB@6P@K-cSGnnhPt96
zB>Jz!pOI}l!4G=qu9ik|a4pA=Iv5z*b%)(RN3)8%A@WIR?5O@DR0c9JMh&o?DPQ7~
zTOtHuFqqEk4>Zih6~X_+daocp>Sfh_$^<?CDU!e1&F&2QuwdKvg|~cgX#i;4Qhcu`
zuo(aDW|v5(XWx$HH)U^nE6dNxl9HR1ON+PZkNr>U>tlH|RFh8+^yIl;OL9!U4uo?Q
zecf=zEREH8Yq@j#vyCjc@pnY9n;2U!M~L<x9%s(jGfrzvG&-AWb*4OiPq*gFX$y}}
zZS}73U*M9TVO$e))Uo^CWiwT)?EKKqh<NzU?4x`V)gepZ(_}G+msXmE1}^A%AgP@#
zO2+{@2KtZifzf2r4)f2*>c>N<Z^TWd9;fZoyg%25qn4l-U}1^O6qyw_k@GX>rUoNO
zTido=!e6*N(RQJ|^)r1Pzo6LC<A6tqj~b2v<U>ebp^ZeK7*>RTU|RF2Mo|c28<nwk
z7vk`~Yg<yf9;TCjiAlg`fH|sBAGWDGLWA8=g2g=$5GYz7c2Ol`9Dec!fw*+Xe5N^b
zuZO5i*M05Lk=-t7iziXhHararbUc8TqP#I9ddA5H@CN@E1fevQl*7}^r9z+dse^a6
zXceME-)BS!r*HPLHJ5q>B=(K@!v1I=wgR1E>AFW_G2RW2;Zi%hZ+p@soERP|&;4<e
z>+ESTP05zj7x?YA&?unA+}KSHCT(e%2c$0=t$nVpuHnnkQDGkOb!jVk^Q#gP<7eT|
zZh-2_KV<j<71+oi$xz{(?DXSGt7Sv;v%5)Et_6myNiTgSm*EA56}5*C+3!F4kj=Gm
z|NerUk<_XMi1m9x#$MtLN*2cZY-hh@6i0HjG@@a^O82%1&3j1+=we++rgXE?WLf0d
zUgg&)sX-!HEo@_>pOS`!fz=07&>hYA@OfA0j5O&fY7OgbZ|6ObUG6Q)iHtq-Nbt-x
zH(V-Q8+s+7gv5HOJ=oiR+?T!WiwhP>QGZhM+|;%Qbuh6)#Ld|zQ>R9H9YKxGoL@jn
zgW#JLuJkOJ+&o53Q{2jZ!s_O4E~lHoaN-4n8+!*-EMTx5H!~!#svp3y^s~yOycZh|
z4_9~@CL6nDQ+(bp=i3c#d4ZA`eqE05{X!x7DD*cN7zJP6Uk=xS%eO7g`S%RVGJ{-m
z`?Yg`(nRsZpNSt8_}?p<L^@k6Z+Be>fwW@A62~%Qc*Z5h1Dd{{mrj1RJn>wApg3je
z{{&(t{&qsw8kZ`u1C(~o$T9Ofd2Jkfg)TY5_aLqq?)><K!_RIqUt4aAmdH`Nf%|6C
z>29a@!*#x9V?wI5j`$gCeXwpi$M~4A?@f7pY{@L{nZXHYL$%+Rsi&3AZ!9h?B^-PI
z>FarrzKI7wQUM5~mSi-Qa<VqIV>N7R#IEEs!*stTBQV?}tV@TC!a<N5c)Cm)>I+`3
zc>TEzEi5^fJp^HqxD~Wi)Y6tRq3a@=?`BS}TA^(nBCja+#w;e^OvltvQk&vWb`3XR
zH46hyedM!z9vyDx?`N;MbGKHjbc!FFNY;$I67zE{)BK6bY4HMv-&Wq;bflxD-Bh)V
zmf3seld<Q^^|18$WhK=jy|Eli?NV(5UGpK4f-vRoRPXJ7UrNet`s3DoM>H4qT`ST3
zHPSdx9v@^t_53`|r*e7mXerN3xaLW<v4;fT4DimGG@>J;cK8sCpdCHUg2vf59v<pD
z`*|)e7eFAmdCBeBdlYAVSW}IIGGElBH7*ZK^DJlji!%4}pTo3<?UtUbu-p?;)SS)4
z(N@BEuOtG&Z}h!nqNc*L?bn9Edf94+slb+Y`Kw)FIF{}e<m}ps#Ixr~)Sup?nWs%0
zg^M;0tj5K@0Eg>@)Xq_VNNda5tBr0ZAQ(UcGR?L7D9)QPCV@Qv-t$P!B+IsPxzp6`
z_qf6DsRTN9G5&`RB`dqm<vYQfJp{>x#q|eNX7Pi|Ez=yxA`>I)wo^|Cw7OF{#4kT(
z@r?qrnBz!yxhp!S>{QA#j&*;s&Q;Nw5@+>C&A;GzvZiqPla)#b-$9=j9CX!!{=PdN
z$>PdZP1@jse42Iln6F|)!?EQ{Mx}Y#6o<=7|JpkXFOE_?UD8vKJtIi#`{gzMb7qF<
zK3M>(l+*pY^xuS1m-nLwBp*O!WMnMXioBiSbqC_blT4Hvy2S<@aQ0|Y`Gye=O6XO=
z8M7J~P%V}7DpuVO<Qu5>IIWx>{Ct3a26;PSYWHh}Fr&v}k{A&giC&PbCh#4N<Y_I;
z$LQcOr4RNlEfXX0YxSl`+sTp^IdO32KS->2sePzby&bEd7F+{Gv9KZP+QtZMuwcpp
ztC-F=%`ly~o7Dot<IFi9r`PT)^{f2;AK60X*PjAt{{3+#0382l^v#<$kF>xAWr
zl`s{emHMrR5A8daGur!4<|`BV%=*LF-3E_kxq@Of*1WrlS{u(i&OVOS6bqdjRg5>0
z7(Sl>XylW(|Aa(Sg+GV<C=f=U2R#c9FuX-ORb5%-%_3-c3-99(to17gc_QG<tnFC3
zYiUk)QiW@5l?8xjlWM0Y(eMrKSvD{^c6@R0YV~p{z?iJo7?P@7mSolxJ1S#C5q2?o
z^>ZP;(*a_~3$=P3s15hCK`}feBv8!Nw1}nBXCLqMAghn0#zu;#FV3ms-f+aEt=>9a
z+hgiBA3|x&oR(l>b58~qL{3(7+XqW9EXIFPsolYWn1_5n%B)LW!#6)>oQ?B3qe{ka
zZQ9E2oka{l1&PL~0h`_ufC9jZO%hlm)0HGk$MrN}uyAOX7g-QlV*R_Vgo>Hmi5NxH
zPOKbo^GxlpKo{e91m*Z*+<NG1lKXU=k9H5QLeta}2<rRrpe()OosKiYDbYH^6IWh1
zT2x{bQ=@wJMxl{anURB|xXCx|02%YRaMY28C1w*hSZ;(|u-gS~`A)Bf2;F*EX0zYd
zm?WCnw*of9I0Hvx2$iX~dycn*!mUVO-#4dU_fE>*$qi!v>6@r_J;6^DK9Vk8{z#tS
z_U?c*?1VcPy#C~_rT(A|I2RpC;;Khk{uu~8K5nMkvbW7)=R@YiNfLo=TCY8lL?&A;
zmd4ok9ZUw9oE^j$)R!y@g>^xD{cjw0ZXMtG2e6A_G)ioe1=Fx>==%T<-gyr1#=2PG
z(~^nXq!6`C$S0{L+qIL^LUaCll?_A4gGA(PCp=d6%q-}hN&=6x-$2@*jDRD#EJt3C
z0f*r{=Kz?6#T}h{5D3JgGiqZFRiw{Z=hD7gCYNBAgf90dwWg0ImD?-s#|}i;ZJwp0
z^S{Q%a@1S)<a9-!L&_cX^*N*u+t5<UV#+SVhYX7<15IVCQ70eiMRWZ*WN}L|f>Kou
z)io-)hu+Oa3)KmV=Yvo3<#wmWE7usBR_k?Xm-`(57QiQ`zWlw8t@PD9T4dnupl@*X
z!X02tkb-xFCJG1O$`}*bSAn=`XQ0kY#^@rRFLs`IFtF}2q|9rrsMz_WqLLw4(DX!m
z9|=^L+43BG-x%Nz1RD=`$w@NB%NfXPwn;jrraPV~6eI!*%Cv4h*T-{)^wjzKZg#A?
z?7Jn)1g~q3AQ0t;LNaNmAQ-4d0!mg}`7H;FJM_82@DH%G;5UQhbwJ_dZ!3D}zLu5I
zn5kYIQ3ZP^{dIH&cILddfyQLiozP~cU!zGgLJ>(4zE&Uer9`dIrp6-nH(onCm$R9m
znzK|A78uMtHt3{y)mUpw)p~L$e->};Y49i}hmd@M%S?P%X0)C(mt357J`ZXrz3q2A
z&1n*SpmWqPy6UBhKao1~pUFuAJX+l3cx6in)?X3tl&fa4paOmDn5cg=v64{pcBT&E
zKV`(o)}Jg^J>5dN$LPYatklG#M#Jg9R@avgIp50*5(4}VR@}~E8{dH}>CKB%hFrhf
zmDc#tvMYTlQtD5B`1XHd;Bji0Ta)7qf(eB)pWh;o!k9SGW8(vAJJ!odF~Y=~kGa7G
zFAD57J8tuU2`OgFJlwr2A`80NKXbkvLR|RY8<5fH9C&gHTp*SESHioj3+mb5dnals
zME%vY61@us)99+<?=hYDjMm06Iiwq$*?1NcBxD$pLP~oMq_!di;3hs*lMqP`H2(}2
z-tfY)W->Y4q1qaE<YY3L#++^Ior^i7>Kq-NV=G&w{Y*Rn1M1^HH+1<^H!V#Z&L)Z$
zr1kLS$Xl|dGkT!bqV#*fR4g)FW4<sIBd<WOU)x~6?`To^ArJ_=OQ3P8**kC!`~(L`
z0^L0Iy<qjjL|<6VbC_yh=HZd3*S^Wmi?bAA_xMUU@Z)0S)gU}I)NK32oSDf8WL|gP
zpK1<z7}LC};#=;-OT(&-UjJaEVpzI8NJSq)5mIr^Ps_$=!t(lcdLX>59rl6s6F&xn
zw!UX&X&Fqcj~N>qYdQ1ppV8~{NGM=UhXJ7NVrV(MR+!}3v0>CwtFpzbUkjW~r`a0M
z%T>&DD9kx|tCLx3)oWAXi8<_2&{8jjR6O9ltQ?t)d!!GGi~JWQ{(D>s(vtyQ&Z6@Q
zvOcf7Hv0GTS<3?Re~Cx)@d<1z&@vimaPo}vNx`N6Cg<PFmV<urk&Pm{-jIc04}k`1
zt1mwUXu_NL-k)MRWu53=v0uzUvU%CRQvrOl3m+#_r~Er1Ykc`B11axn#zlD~=J%B^
zlBSGcZ@LYF^F3p(&R!$y95+;)+0j<jX}AY*(ERqIANzm7&df1^1BmQ8b4Q<A#PeoO
z0KG8&Hwu^(qFD(;Wj~%rQN5FW-*tl|z_d<L{Ogf5g;!fE>CL1^GXwrc%E!%inyMg#
z-50<iS{l9m=mU5ZaH=(K34k*~m7bH87}2%R&^8HoDO4LNIP%QO^EE3tQ(vWBC<Smh
zFm_7;k30hSA_u=wz}+R!v)4u^y&upfn=Wb*5*B>2r1H1h?@k8oYyjK@pX&j<8#<BA
zjnD(9hMi<xE2rl41MEchC;iSW4cTeG<J}DWXXcxq9@}Ci180Brh&pQ2*?PWJ)!C7s
z7Pjt<^e3EB(_KVxeP;SiUel#O9(ZjxQz1<~M=Ysc;8|TGy;|D-grai-lWKtgbl!y#
z38nj*+Gm9P)4;vazY#=rIIv$eRZyMMBIRl0pzvpYjyHXB=O~EqmhgB+Egy;W!B-C+
z>{y>+ex~D+jyGBo8F){HZCAg_i?B{2F`<zve-nuZG?MO;Ld*Mssc{l3^8r*h7~ja4
z_Vx?kZ!_ZZ(+M0cvONem^S1pay8O;2cC>i6Uph<As#^`R;fSuVl~X$p1Z@|VgNLsD
zYA35Fe<*iVbXYPwwOv(q*0nG%bl)?djMKdd-(^Low%ZQ`a}MpAtDB}_E^lu5#{Ii&
z_=5|Q4d)uRm))r`JcosAIR3RAq{Ela_Q!9c=nNw%ww)z<QfPIl$cywwrn;Sk06Xj%
z931S08_xy!CF|(=5))+&jcDVFS8}_J-ZO0)Nk=I6K`r+*g6yt-<#LAKI>I=ZM6mpX
zI%&y^w@!D){fO^llX&6Lj&o>eDM76gJ!bk+SvSrCbVq;x?PG^1I1o<>;#d7$7iPnK
ze6QU)Zru*yvOhdw<c{hsPfT|_qcZ`nwb)RKiXZbLBJG9iXV;6qrWjScabN7?gVdOm
z+YExTq98y+1wv<?D?VU*^ysrjgr6mvPS2zs#dEHBRUdF_-jFw5o0Sy)_*_xI@b{hR
zMd*buW1qso%&{af$Ojs^nq)9d`N=#h>-vqCgS{Mp-Gaf$E89)2^Qvg|?~Q*S27TZ|
zejm94PwR}P0B{eQR#+w>^%i>w12nC^*><6W(g=-iHDf2dh`q>MBp)oXh;W;6qMu0e
z<ovxHMl_Lpg71D%f<+g|hc`B{l3I8g_Y|^(dHrt^x&k@v&-#zb18X4<mfw{admLDH
z3Zz3(_piFE&jgHPO!rE$8<REtSR_>RZAXzUwNpeu{_pAplPPh>iPY+*fmQFW87hz6
z1knYP51I5P4LGwvY+pk(e&`;rI8V+t;Q+rTWY+u`S6C{mv0b$;TsW#|EV42m1@K!K
zA^{L??^tiS#m7tNwM$P;McM4kCqkxgqixX1V(yOhhZW)njMl!_#7AzDQ!|GuaftF?
z`}IdLhB5IH4DLxW<`Dgp`|)ssKG%o>PjM=%pbm3vX8&Z2qE5AJs&sqd#?L7)j(PK}
zro>ppNt-()fqx(u!5@jv--HzI<cMXmDZz~>|CnNJGFYf;foa#F=u49o^uevyp)zLh
z0_;T+cXDTrmu~|>`SdhQi7~6}&IW{Sa}p`MSktgME_xWeTvJPUdOmH<wVqZ^IEtS3
zE}V|#=A4-*QhvF!lF6&sJlw`~zy1*S&EF}SofJYkm?!rikhrCXbX=*_kV!w<@^y5o
zKG`3s1(bIGaREBzQJ;f@iPb@<bfUb>UTSLUdg@{OzF4u*V<B}`&qJus;%tVVW#jg~
zwh{d@?sTKUb&Vt6LG^pr<s_bSr|t=%&YC_~H7P=`c10<Q*c}lr)Rfn{%M}>mw*n;o
z4rK^bC?r(>!pCwTQ!g*=6U1v!tdRes4u9V0HQbver}I3C_?)NM*dUr1vm<>1gEfor
z9;w6KZ!WS7zn%fX>4A83=LxjZU5;kF52j<FbHG8;Z&&95g9D{5GHPcuH&6Kh94Fzm
z4QIuL(BUz`9gw%PUT68dHStq!UTK=CUT>n;8ylsM&wz-FC(SaBr7Lj8r#P?1A1z+Y
zAsl_XZXEsVqa(JS87^|}usVcLHd7@$X<4>eAFpg!k#g<<08>3x*wjemHO$w8FxgWd
zk5pCzOH(Lmm^KAVAuq)9@NkxbyC#WCPssGPh=2dbv1H=hH66D4$G&>g!axKPpAnz@
zmGisDunqXlXO~oqppU^GHxvFe*nD(&ph~|uXx`zU3dAU-?uM>hbuJ&QHSbbe^636u
z^j7s~_PGbA7TNyFC|}^8T{xPSHch#Kz?sJdhdhEje%xmMLmga@*z(L8aeNNfUwLfi
zn0v4?8hYM|u%iaqGJNNK??HBBLo9-htoh}L`^oL=AT4*-Ir)Umz2QZ54k;E4SM5s6
z2a=j*JKB$$I#gaRv3RluRy$1!7#z@Q5O?AW6S>CeH`4s|_pJo9IzDN`M<(!hV{M_(
zM&PduET&c~HnZlm!a^B4z84rKONfZA>@lNBJnpN?1Sv-v*1Ws60yv|#dug7BtL^zf
z!)lWouCKQ$?JW4~pHhZ@-Pr5Ab`xH3eyLUk9(}YOrev*cfA!Z-AUH{_p`zwp3Ml+R
z9CVBZQCEcnzdx&G*}JawlMR30ZC~2$e#62<q|xNvdmYwoFh$FdPgjL)7&$0&wV%^l
zF2lC42?kK8@cEalzW~s%U)Rn|)_@T&d9~&A5iGtTPOi>E3nZH=q+n+P7Q?@$j|_z1
z*!ey`U`vv!9MIwzw0dV}CsTB~6S}6U(_7Px?tbHnFYd?lkg&>tV$;f;68r%*wm!iy
zJWsx|BdVr&@(JXmT*P8yV#Hk~r8M&jL%rWtU}y77=mY%sedu{_l)o_RH}6q{Wh>Ju
z<iMG@V2b1mBOqD=Wr^f}L@o?k-R#2=*0FGN;m<%Av8r!6jd8BMQAe(T7?gVtegFZn
zQp`G&EmOR@z(aX#YlFq;{(C)obi(g$Mm%t19-S$#{BEj*czo?u!RyTAmQ^>qbl}N5
zSKT%#A5NVFE)3o)7>6xdrchHZ4eQB{d{ew019XJVMM=rSdX;kSkX^a7GT&OwNH^;a
zG^VF<(Q?Q8H0D3NyI5Kp9TOAtbK2|vp>P&@Rp5->`yZgL&P<vD9h;Bp`4zqT!Zshu
zBMJ~JtMBO<vOuFW@V1Xb3ss7Oj6`OgaxP&!gmIy7Vf8azN=AU`R#3vV#<yGW1XHi?
z7GaGXQl3Y*kN))Z6_(E+Dco;}Fg_(c+8qATy&RSo&n5`$@XT3EmQn7KN_=Vk#Ht&8
zw<nH3b-lG!VUY#b-&IPkx(dvLf(M^g9}igL(k&|rAzX5N{>(Ph`wa5+>a8RJ4hV%(
zMVg2ztO9-@7w8^^NOh<vBg@<J3Fndl+y?r)Z4FQZaNJ(5Ll>Q%^|kuEnGeyj$C=Xv
zs2>aR)mh`110ECN|7hpbYtO6D3AIj`@v$~RcPO`W=hM*8gt4KUtOusbv&?2@`#jco
zM+?i{RVjcVIpr^uYCRORnp4n2F&xDW)cP;paTR$_E-qYNZU8qrt}RLT;2Nm1CG3*F
zE$2;f%3I6n#-==)9G#f{Y2nA3=W#ubha1AUohW}G965jEI&((~4uROV{?%pN#aU-@
zT(C^u=CUt7s3jWgP^q7V06q|7Pd;T$A)gV&DcT?L&@eGRc^4Tj2V_UizvNq9<Z`9R
zu3dUpiYU?@Aw<_y!WQkw7}u-WTu*1i0qavtUN*V$Sp%5x?}N<_q}Gy}lg2mRRjhLh
z`3^}`48iN;^<q<vj+o}w*92RUL9O=XDTxr*7m;@zRAz3NP6t5pGBguCdJcU1T2C5x
zku%&p8pQhnYZx=gvB$*l%J2`Q`_WM#qh4J%|Ls6OOe!#~og0(Q+|qYsOSThh`cyc8
zklN+N7lnLR_`rN*(^t0Z5XMCX4=#<*fOjKgB=$S&iIe|CkU}OGp{frv<j9cWO34My
z3(*_igPgAOzMN!xsP9RFL$;rLWPTJ}W7hgB;V_3?61R0Ig~X&L-gI-|1=HO1eJvGB
z^#OtC;w6l4Z9((bW-FLgRn$ZR?O6b?Qs8<_u=u?+gR{IKc{0w*nMt*^GxY+Wv2aUZ
z%U!a1dcMaXZa3cY+U_iZgY<F%D02B)<C*dPjAL!Gthv?tCbl8GB#wd2YaE0`Ae9_f
zY(@*<b3owiS%SdXol$hNz=^FqHp$Fy)7j7<YHRAw)=_`-`gFa$=kU-y=*am=>>qdJ
z78}QCf%igY%EOW-T!)M{H0cJRGY9&N(`)hEQ2lte>7e=;ee=tS57;HANd+fC)<-VD
z9HR_&sYL9HqkDXHo>jhJ4i~UCex9T-#;E41H+0R2y=xsW>~SZS)9NFj|GKfJh+90!
zAS<ciPP-<mdKSt*A^Wc*2(+8Bvzio}>^$r%AL#2&8Zi2EDA{uJO4vp02XU=53xai<
z<V5Sl?rxNjHtpx!)@BTw{o1=TubHMbcn$#ZL`td`>d=ke%F5ooi(ppn!0{Pkcx_yw
z&V}IbStC>^HBEf=;*FJVO{acmGljt&=;{9RLCMJ)`>nSVwQ@T)*aue*rT$i8>LhU0
z%5cwf1aGWN^-P!KH^$wm2McV>%%f9_6d$_h9%ItUeixH1{5+hC?7lxyJ23jv@Qs4h
z`0MGwqxu`5RddJn>ih>+^4Bnqv0IEF#wR{wS7N7c#yG*nvf3o4i=WEmy1`{I=*w_1
zN=Y=V?Q&ahU=jOUGW?^9EUM>GeFp5zYN$JD@R_B^>S`mayADSZ6lO3kVTzO%m!!F4
ziu(w-eZXFMR5yeqvVC_RAv@F~V%|2Cq)8hCCy0sklaqdqN-H%red)jeEV45)ALWk6
zA?!<=Uwcl9JM7Y1SbbK{kf@%<mo|R81p@sL1`xcqzW&P0;X2XYaZF|vi&WDU%KkYu
zwu$@Ux%p&cQM=-HaPTgX-JIRg1S!;V*`2vRiDN0>D4qQ{V=-4j`tNR4r7y>m{7$|Q
zRiDPA+OZb8z!T8U?gQ>vJwYSgeWcGUlin6}=^o~=uey4tG-}4nbjFmUZza!~%QtW5
zWDD0u$L2RO&Dh}4BkT7BVsXmF4o(m-xJf!qb-E#5WX9sM?0{USKP3qbU;Q9(q_xX$
z{j}7lU~OR&8jjEq3!9R*HH*3ry6@t!3<8OK_)bgKmU7a;IW$vE_V;(^M*MBo14x0L
zx`p24_Yn^ux%;h0eBhDRQvmuc<-|!P?*z#8Uk<yVl;9O|{sYnWjv@cXDBx*WalEoM
z7jo>6?ADI93Wod%<x_J`Umy)*ucv-hL2&QZ-gZ}S#-i(`ngT{R<R!0KYSkezrNoJV
z?M)i2M4j<>NKdi<Ei!Oj4mY2f8Y88qbK0|+jFF0pe7OSrmR--ivv)26_FT30HamIP
z&vfi-4;H4UAG(fK`TV2|zsA@#UKU_7PjZUV*nfOlZ}fScH1Pa9!F2cZ0{fDzh{!@G
z6&2ftSBDTST2z`jr=WCkYlLdSAA%zlQ14s+eT`8*ESvj1<3?VlUeDO5wz$59F*le{
zXt#4D4wq@5Nw!~;Z@N1B29%#wP+jlEqU9c&1Sj6lzSKFhwteH@FwJWGSc-TbVS%}z
z{%2$JjhJ|=fRtm4@(kkP^Av?<;r9(*jfjGbsgq0DP6XM2M*roKM}db5F6b*fOt+wd
zBe0jJS8(+reswP9m*h*9KEivfB5%9qs2hpCo?mP$&GnC|Jz4(wf5`e0cqqRw?q{r(
ztwBVgMj?^xvQ~yHA<G~;p|WS+4U#F_PepbT;<1x`8*OBl#=d3EPD1wgj{5!o|Mz|G
z=hHgo-uv8p&bjA&zvtYC=<w0neia!Qc!$ZYHJ4xMm;IeSYSr84x49PDQ{ygkemD5(
zipZY*c1d<HrI0+Un4_*!Q&CO1=_BQ_+YEfX_5+=7eXC}Mw$~^9_tthR#{x_-9<v?E
z=GMDS?#mND+`r~fQW|^TRcu)AXy&!w1_?D&or%|6HZnTjHfbcTVJ5z2$|<ajROhRs
zU0-rl*v0f!srDtMyiN9Z>Qe9%l8Fe1FMZ@!yy39ejM8JRZifw)l{f**`Z$HWNCnec
z%c+cRGo|gj`s{^{n=)pr0XH>xemMd|*lJJ=D7ZIN%<ZabQnq;iQ2d|l$G`l19<?PU
zhK-6`T4-T=suN*pZN(hLSM9ud?qk!zzT$C)b30By84EpNUPX?oio`d`INpMur`%>6
zovmsrcLNC004^u81~CxVAilE@+R2?S{~F7L6h86Al77zORh0ooP7>50<V-RTY>ei@
zo$IG{<!aTnYV7;pv3=mEe5<l+_N|AryByP__4bvWw%58|*)v~egA!kiJfzLZiHfuz
zEB{oDTa~00VB$oHdu@}OFNoPaOig@sUnqKG0XRPnQ3|Qq@70T_95yV-<tpoBUx#n=
zhT!%-P&~VRqqt6b3f=gD?Ah(E=@q(_P#(Y|A(7D@I)v~acWv&A1tt(sLM^_IOwkiH
zuhAL#UL1HwF{|X`jQkt7(vAgWMsh%A5miVGTJf1uj)x}k+uPdbK&lYBY0&RyR!NI@
zMPE~_plJwBi4JOtH(!Lk(b4Ki!E66;_j8PCVP^7BFongNH_sGTJX_l|j$j6pDn>ks
zVGU19R^Vk7a&A<?yQKnYjRG6;8O3hZJwhGoXRKe<XA|z<6*=qtD3{w`P6+6l3k{Eb
z{GbNDf2z?eI-u&=o!ErKt_UB&{1zUQhff?57}{Scd>lSsMvATCgd^&-TjD6jU;os#
zd~)ZNV0?)9Q&>p}ceVY@ks@8ZL&)FEYF&dUV*77ahVuKLFvG##rb!vT;{Dw!PCjBR
zH+Fs((E=1UsJSF^{~rb>O`G7G$sxGT!`#sTYftJE@+}3S&tE1!ja-;ry~+?hMy>!L
z7kbEu;%~6^w3`AlMZ{&mltF1Z(@%?g*6ZT>>LmT_F20-QhB%?5<7svL&mS)P587+k
z$lbxLgr)ukV+m$rt1daXI7XMg0b9nIUDu!yTbGClRcIx5-ZADI!^E^tN)eB!=0Dd?
zvqbH6Ra=Si<`y_^#NAa5!tttJ4^{xtrIQ@Oj0e(axYk~@K!vE~LA^mpr^DpzGgM3k
zeyd#hhABe}C5JXTSlEGbH7>OmV-t8yYMy6cZ0Qe5i8?`c=2*2Hl<3e-st%SskiH09
zwl-`=s-U)LxB%sm6!qADQ<I0d5fT*mI;h}ObSP$H3A>>`AcXjK*$U^aEa)Jo7pkfK
z>4-tvob|kwB2JofgvS9_LIkdkPnJLGmw3IZk=X*s;{IwKw{;G)Umw1$L3;I5q}g^M
zZt)I*1v6w*Xjj&lP9|XOFumz*3k@0lmm)cRSdj9!Q{o@6N}3{*@x4?wXa}C}v}(_T
z<@&q%!fr2#*Z5wg2y1K3p0vkjOWUjR>Dgz`I~eXZll?sV`t^B24w0Vx(+FDG>&^4b
z;ErBi7jYGJ0Y`t0=ZRw10a@&ChO?;z0ju|*rb15(6=PN0HhRhyB8B$es2I%EOhW}^
zUBh7)HC%k^r1jnLXHXVjNyUmvpPXWn46bSX+x?AYLh<BeRCx#>N8+VB(4vOE2d_Cd
zS3+Lh+J4ti-{BbU?R^2f6`)0fNt1>dydLr1Zf#rBw10VHx?)ThSHN$2yf8HQROuaT
z?PKs3m$b}eH$ZEq{Liw+LoI~#e8Nm~$C?-U-@$&p0xaxeg?O*-OOX9TmM12P)I8<f
zua-mdK9c)qa!<~L`>5ux0Ny*?$b#2AuVt^yCdl!GmSoTKA_Woo3=lQ>s?nuV+3`*9
zbeO&Ui`ltw@~dAQJlyqNb&4ofvME}d+sQU#8Uobl=Q%Gto~1b&R)>GJ6}-4?YDT?E
z=^5X?R*pP5<<Em1!t;}_kG&`nP*`rKX-v#T1#e%E`;Pew@jRt2)ILdOkAij2(%xcQ
z4`}{u@azd61<0jQ?0F{;{jB{=!M2Bv-;27<lhBJRD}u<oIdIPiGh;Lo+8N-WBk>gT
z^a$bcOfFw_C>$(bfBzI05k3(O=pW#qnQcWfFNKxltZ0y(bXYldz<8Z0>Wb!)C5b{r
zI#uiN)>K>qK~$#Nt;Mv&mr$twmg~Ot^v92%(Sl3J=%}_swzs7`!h?k~hm0hr`4T9{
zPeXO+DLYKE;V22pnHKA^^t4MlKwJrZ+n};<c+|O}diXk;>y?orhSJB=5ytV6cd=o;
zWJvKSi0kYa<#oNjJjek<s5=;a#fukl{ku{A8L+r_jTkb=(NVXK4POjh(ILorO5?Jb
zw!1a;`-RudbYF1MFl;?}@}wwRGWylepH%OK=&V0UouFz@OX4t61$e<PF|6U@>O8og
zSP`@T?p?gX6g`OU_H0ad4MngwGV_Gt@P{ekMH^usDR4V4<C@TmPjcpO#T0R10w4b1
zImi0xq8SL3T1Va~YtAk;6SHH$;>{~!grVV!C@H%oOBefTMojy3<ArpC2`bF&bR@~G
zwsJNfR`J_#h4=;Z>8mK-;GUfGp(QTqTg$C^D?Cn?Ig%9v5{l8p6rv4Eq!%6D4YEMo
zW{RGMWD#a5!i#yyh+%76sER~kC2=}<J1uu!59KV0O(dGfOOWtO+b}iFEpNB_$y)$;
zy-LT>?D81YsMKrfTz(u9a{iiN&Ytl`dMweT6N!?mUrl^|kenmOw@PQDb<O~|?BVS5
z>D(vG5+^>hN&FL^6yugH3YlbF;7#nA`zC$s^}VQ`IT4o5Cy9ru1s6#~OOCl)YLZrH
zijORQV9+}yrhPa&)AAr3cKCV!yfS8=9lv#8V0sl9{baMB#XKa0dhqBd1$==h27`=0
zF-{BV5wm>Dl%pJyV+HQk&(D*DP9A_?Xm(#z;n@;WgDm2l-t3YfsF-qKmWMqW=-d7y
zyO7AGdz<Oz1X!~AM50nANUr|V(#MNh%hTDSD3JuQ8H+8j)>`h%gAA1Cf;gabp=m@j
zK!9E}b#nxsy{m)9m?Dwdq_fhH4B{Q-KW5yHMc3A?b74-rT4qC$^`9j|5+sgz6tVTE
zUtr{<f&r&6es&nz3|Wx!8iC3KaXe<Ak}9n|L8N&Kc}`LiLC5sHvy?YfO;B9sSc3et
z*}pS`mC+Y^vz%%=ni7}`c3KeUQC!zWnjNH)IaI*Z;q0_yfB+>}s=r0$Cb>q$-@QGn
zM3i*KV<;_3{B4MPw>0*V6+?76WFjt37{U^${jtD&XQ%&R(E_v`Tdpv#G`HU8rjKbn
ziU^{Nk2Hlb$$`}A8BOI>KT=T}%EJ)kF$4>wS8L*colWgM$n}wr+!?$}Ousa77T<Yg
z^rzfQ?m7P+C0o(JwrJH}If45Gi4D-u0e1Z~I|XXR=$hwA0M)l7iHlV)&K(UgrKdL-
z6VJF^r?c@Ea$BL1&3fk&5!3=XRlyGV8SVs6ln@e+fv#v0I&FLc0Ft_-bBZ(UU~Y5p
z)cD#4Su+Dg7|CPNFdQ!kWkA;h!AHF|PU`zaOuT?7o*C7u1>e{5-8+FN6Eo2!gts^v
zv|m7x+NJTBmAwUv5Of1W7G}gPdGYWEyBfRM0pjn8j0!kbvGDsMzHVSJ%<EfLl2jTT
zg6Ya)RYTJw!byuruGbBYa)7|I2BT2a-18+;ocsFG$l$jadv+`p^zOK&M((^d`y9Lu
za^kX<doDl=fbsPB1Q<`I-dH~{&d-dm!3rtr?w%z7e)Yx#E!3*U>WB2gU|y}`_mt7k
z5lLiep+hd|$+br)R(%R|Kt6sO`XZJ*$^={<PkD^ES8>7{E|fK6w-FQyB|QhQw-f-M
z;C>OHrCso|XayZ`rH?#BvuF|m4o5HCQ@}-Cn|8lx=gN5VW+WHYC8D@S2jt94B9#%V
zV%ScFDGQ7-l6r&>rRn}nTPiWgK@PhCG3DB9bC2Kp4o=ig(yOf+yT}vc7a@7omlv<@
zzJkU*{{lqX@w--a)M<iP7ah@x6Xmyn9R4Dq2k(^>UTUR*6P7<Xt3mIYQn_8mFb1SK
z(nLP7eX_aW+d4=d;{udU=J1oh(WsZwQ?Nrk5BhBxtzBDPpk?vE)5z_!c<NpMsz$N^
z(>EuP4ANc@J~Pv(<`J&h66%EKTU6#%HHJoJ#Y&o)3rU=4Mg(sZ#zRa999b7L{t;vW
z-Sz~9AfJFckjwip_4b&_v2jUg660oKWgnVd@^~&<+<v8<W^FfCr9ikL4cqi0o9WoM
z=Jcw~sd08l;CE5dHDy*B{1;74glO-~hNv|$R}s$K|GW8w)CE@ZMVGj`1E-u<K3NB~
zisw-t2v64Q$@FiWhT|XVk(|B&V$gZj7$-*Bp|lQqhU=xB)^dldRu@rUoA~d%yg!%+
zo7S{Dr{bh($IXRzfv^-P0aX6m2k%jGrxLnEbv=TMYTk!%3syR_Ggu$L0lIWEq_NX)
zTx@L>Z^>7~a`Af{UG*2`NYSE3ZZwrwLp7nCc?W35T>Twzy$%|J1933r3*C-zCf+b*
zjoJ3>qmw5c;`K^Cf3A=1HmexaIRw3Y7)**~x$!G^ze*PC0I65>xxR1TLusYK_x1*2
z0iEtOqnYj}OW-L<arKc&M}ymyV32*lJ@nFNw)(#CqaKp$lpN&?b0?~g!=mI)(J=cj
zdvUOrHs$lId|CoC=TYe@ATLcY|144TFfLK#^6Ls2RcaxWXHI+fgqLd)jM-;I-5wv3
zj_XRTxNy0pzz(@Pv9Hw8k|Yw{9sgMMbjQqGI>RRkyk$&O<>njs{_bmKQZ+?>=!1f#
z=hKH-Pr4aU&fM5Jcqg2LD1$ooIu&M`wSLyoXaAaeNgJ4zu@MDT40K=RQu=9jxA6pK
z<epZ|(`F7aE8`5~gGSiix$260qWKe;o!1_ry`_xaBvB&i8p*25rM4F7y_=a=QR+9U
zbFofd2QlC_&tBZsp6ndWIQPp|FCjr;d2Z?Xyk3S53lbwv;Me~~%&>5h0*i3e_HrtW
z`*<C1m1WO^x1-ZwuWTkL7m<glJNeXgg>461y)?1t>`%v(V{`LikKOpv70iCy*ihym
zXG7XJ&M%_19j&fkV7k>{*`EIRn8KSM?^Tv2S0&we-S*LpXMrL91#AvK$ld?hsl~O$
z6dmWRi#&mUT-bS4K6;5wYz%W#o2XjEMv@s*HzU~Iw4_nmnEvb=18mn+ogYxgJmlQ&
zVs~vFNuCMF<j4oTPi?bwy@lzfjz4d$sa8Qxg}|rpeuH1Wf6VoEd3*CafAOUsCR2?x
zXORz&r3Knw*EfG15DJjNqei${&{{|7JG^mw_v|Tf4RQa{hP%KFxU+(|{g^=b7}FBj
zYUEHS*Z8`IPenM3)}A3q5>P_3W5WSZ9xrvfURhObZH6qEw^JE-pk=pmKQtC(rHRhW
z9zPcIk5;0`!df4vEN=oOF`Ea61MsOb#+?P50u2t>anG=R{u>+CQ@GZe&z}?}@9J%u
zHd?<=_8sZUvGUK-r@aDJLUGJD_6B6MeP&+nBL@d|-M#87{T_k5jj!cLecSd`)oWRH
zU#PSf>7Reit0c8iXnSc_Sh#qiuVp}d{V)U<B%etc_Ek&UA5du9uXcS<6Dq1((QrM!
z^X5i6Yzmwy7e30z3XffiE|cdUW_0qFKGfrIVF?CRhoD*3k0qp^)qqZ)NGk|s7tHCt
zA~K+kwiLkvk{cf-Q$c@S2&;;oGQseRgB!Uw+V{lb%7l(ZrTfUu_y{&Ez{PJ53U%PY
zYPdpKRHN>(gpuF0+o#=W>acfP?YmJA&S$I^9;$Cl1k0AIpb2P9IN#4H#obX0|2w$m
zvA>R}s<AD|pHVmT6yHt$cM+=wCC^;<5GYtoaCO`w`Ed~ilGbBWz1M^xKg37U-~Gc$
zIdG^-;y^F<p#NJ%MPs@~(CUK(mA!(R_4jpga_{Y{(z4)o>9WpniD%*AO3Q<penl%J
zm>`<O^2veORDP4SVF$hQa^D(?JFVjryB_W*O{A{K2~uvPQ|(tr7!SPmbQ>^@=f5Q*
zLvL>=y0IW;gyQ6^O^n9z8qZb5^Md8Q$C2xad)hqq{{8!ZlHzpc-}qZf?<O71e|Vqn
zI^7u_twe{8;U9zTMv3HIXRmWWrh{`9v#CFtOnis3s)ai!iVV<SP1vLkaOR|_=XRsw
zJxjg?j9!kE-;r)~?p`mxHoCbDnj$2C&G@NxJ?PJNfgNkSj(##!T|3je#|#ejJ{62W
zxp-B94q*vSD1e1AZg`M3XF%dPO7U=KU^*&6`JuRX8>qh*gZj-EnCooM*i8JJ3m^@S
z>N-@(knHVMt1z!Ib@6o>$5?HG86VGH?$APqN!!?`56)mj9ddAp3VlK;SQg{>7@})x
zjb_H&4>%MalPexQFQ8aBp7i58`(R0Feo-siy9<D38OVaS`ZX1;A9vdyNS*+PUd>W<
zyHZvdNKLsk@O<$@c&&29+-%#g8Q9wq6!WB<+rLoB98I)9yFU7npkMepF~onk4tbPu
z>y7?yp*@s6jz6{M^cOsd<`5~*)t{M!{mVNe_WVx%(%|xRn+DH5Px`=uBN`sOi771U
zA&d=e3UArUGQV_NFT^F$rUk0qsJPI)1(NQvKZEM1><T}V*YHB!-RqdXgzcp$zsrRW
ziE06wPME|}!IPcJ)`IQi={64?pkgP$%`kpbdT&2i*kWpeY#u>RVg+|eH@9EeoFoH>
z8T{xy$vgRx#gi)>zDCQGQmi?K7s~*!6xWl}+|qGvo`St)+m_ABk9@6fQ>Zl3|4$~4
z&ckxj2Xx~i9OB~%0hDxV@$5pE4i>4Sp5k&yMMwM)NEY)2MV8Xww?wbC>{eQkLc#gb
zIoT@}?Xu!)E<VkasIOhLG~yr^?eYc}Z@LASld%7LpIam6uRS;1GrX|tG#7X_z&`3|
zr-BBhp(Y5Z)BW7};h(IRQL~Ul-{>)^PE#pvRPquo^EPH=@zWkZ&@rAn1j51MlwMP0
z6e}n#s<}8C>FhId6!EHKC|~{d0<Sor9oK!KNQd3_qQEL_&dP3n9#0>*1ybvx<%m^y
zxLSs6v*-mAe^WqoO|WIV_oNbmi=D>rq0!7oOAOMGeu7x`_98j{KFawG>~CNI?jimg
z%yNh#ud(BuYdS=rLa*Wu0B%{t3I=jv_D7<OPt%&+Vd;6(62IJGbc=_~DRHSncJup{
zGx#FlAdKymQ-bcN%C3_eZ<0KqNK1}tt(qA#JZ_I#i2$%hhl~`TWXsAY>8C9&)8jXA
zQ4#PRIs4(5$mrJAy1PaF-FCSd{SOX7MHI2^$zy=;|3YhA(xrvgtOOUA*3K>F`NCZM
z+W>`>x)T*+on!2lw)T~Z&MIhh&6$Pa2kY^3Bp_~3B%5ryiBU`A!JuU!CkCuALWJDR
z=Bj=1!+8@uf?QBq6>&ecz<yqaM8rL!aY`@YSSb+&bh`b4fs0T%Y?Gjz0<bHnX6ex*
z5DXWkX1BzTtNGfwhKQYsVVUdMyH(16<WRs&4#ji5#&3+PV>AWsE|%uhxRi2uEc59Z
zNI4d~81Gz}9RLW`xG^+tCs$4mRFM+M^2e~dby@Hj>F?3h1tNTMc<dYI;Nd24y580q
z&9F!{2$nq`#7sVIMk6(MPwNlu5_c3dCCoa8Fr4Pijb#%%r!)BBc7r@-H`gYRL#i;?
z&*b~0bHTP$?+&Rlc(_wc+0hEt5*wTlYeFIh{gL`r{VhzXN(T?B*j;IZf0Oo%@ZN4a
zUZlpGAsXSXUtK+N{9h{<KoMbrM85Fe|2N9rwI;<1s7bnG+;naRRF%8no2EBk=@TKr
z=KBbiyOvM1qW~379d+=YNkSwrO@@oR2zUEbwDP9lSeBG<Pcl?5)S5GYp50~)*DMN+
zlCMf&$q6}2=$M?eqLTn=XXn`7KB!gxbBe>H49Bzxz=6ZTCYi9`5~mvGzK&K}vy<@x
zZtKytzCAC02Rh(C5Kw`vcP*XA4pL$hoS)+2-v6Tk{h%86+;w^@wlw?!u1?}mAA8!Y
zkoAKLJFByyiM*Cuy+o>M6~Fgs7N<341IR#P!vwJbAh;wd*S)<Vj4}i~6ECk>BZ?Wb
z&vHEgsa*c)v{W^L%YfAE8Jwl3U}$$noy18HdQrmPtqy$ld|#tiECmMw(0!7t3+<Fe
zpTZ+aR-eEQa8liGL#cVhX*x4D?5$jj#J?&xCH#VXO5cU;209W$AWY|9M@u3kT1Ow$
z{9{Z6fKQ!THH!|TIp%jVdnv407+PNM0c5=N=ViCRbh`sHVdr^wq~M3Xv9+enmNg;l
zXQ)GyFwEP}uGtk<a0WB{G51OpW5ndbtAVNhAK5~0{{Sc3jZI=S0X*2K+qx$)Bc_8k
z0KXsKiB2|MvIU|CYO@l|G*9xxhuWkLHKR-;iULtCC1MPCRE_@&n-vL25aPlqmBZ=<
zdB)7qR0yN{icC54wda;R(s{WZ%!lK$c|JBNuAe(P83-h(l5b_N;6b*4`ieIHfWoVN
z)sfC@r3#j(vD^q(iq%kDTh-i*A+x*G_m9Tq5#1pB=%I}xTBd1uh+;hL%Y}5f=Qj&6
z8ReNp(GP)LJG1-8v7LxTF>dSn0yKID`e9C`N~=XE9N`+{Mnxs=sEWn?%EI~P!Ymhi
z*Br>58Q<p4S2-*w70u0-BP#p)_KCC+XmFr<QOy+l*?ISuHC{S_S)95Vl8Wg1cW=Vm
z+yVm+{nj@S7p+}7Ei4?=>$EW9ySYS=RYyut-!kX_qg{E^VPMd&eXqeY71Z|0@6<3s
zlD{iQo^!r2Au(GPt5(9$=~~_*sw-$ValP#BO+7w{H?SdkN(2f<4BObA?|Wpc5t*-H
z>2Pu%joD|)12aq?yv##Pvu4NkAkN-f*m+6q`!Zws&4Zv2oGY1AIveJf4NJRs3q@GC
zEqcm->xC3XW{9pv<m7c(?01w<?MIP5V{Hvwb25~bJ1^YWo2pYIk^C%gF-d6ansDx4
z&;@O^CPKmu-XGR)OZK8o9{n78SN!6(_49Uk{+JU3hzzoy90zyzqjv_=o(a71{Fo|V
zyT52rQ%wVVX9JdU-UC&_jJ@_<l>)><4eJ$7(EXU=j*d+)?US1q2|^$P9h`mAi+@Dj
zL5E-@*WW689z0qLBYpPhsptjB*NRj@tAo^`CpiA{)|#%s;P2p}1%{Tm!kapxH!z=i
zNb%A&9jc4J&IHw({tZ+B!<O`r<4b*>DLSDtb~>8a-~5$xoMCOpW7_j=Nti<s($F|3
z)JJmq;Vq-kXG%VhJYtyS-1{s?;WBK}jlq?s!C60(wwh7kPFcdGwCs<6_h$>$pomb^
zXiIZD-&TO#n@Y0BA;lm6uvRppeb!g2DDhOXixZiUVq)H4-f<!!MleoOx7M;ZM!5*N
zgL>c~xnQMRZ5i+aG<*rgRV2cI4~BrkGiB_L_RHS)5_GH%oyHyGIl=g;Nl{N(FR6TW
z*?IQv(1O|pZZ3?n&yjRw4<6s-<3Z~LFSYYaJ$m^E!WWG>g77c<zN`TGhVh^hSh7#{
zaN-1Gi~>e4tob73FA~Uwj8BJMf>&`L`)7RxX50$DaTqEutx+KdbU%AP4=^pJPgeis
z!zBY-C8;Ykz6LjX`!D1_I(9s#cEynP{`vHEI!~FyYZ||9>MK3qmqHzA$27%eR|^HG
zB1EPT;6#FBgUSlP{gWh)1<$6;xlj$qfV-9~6VJ=CdP%mCKKd%i8P&LBj6#(nlIX@w
z*<)!a-0aVssC%FaD6eX(wpkC@yPTshNBwP2@y_gYO|CNs4*tqs*zyk&$##l?hu4f5
z#E6*7Yvsk)r0s~hHUhBqmu9IRFFnU(uoljDtYYI1UeGWl>wa0v&Hhk6GD>{|=j7uj
zQRzOZ1P}7X4HAhh&=|>f`ngt@g@{VNNn9&C?Q+7UwR#^Xo&ouyX3?)a`nG##OB&Oi
zh|AQmq89N7b4Q^hPk@xIG3WlS^M`y^G<6CWvXaUY6Xh8YNr-WY#Picq><lQFH~#yr
zgM^(w)eLX6D(d}auhg#o<SCrpx={6yl-)Gl;M|?Y^h~IOq^s)V?z;U6rk(;r#PD3g
zDzPM1<SXQZiq+gu+iWR>+8hlX{e#^f`RoIF?e?GP>lnTxS<MdhJC*ol%9ms3hTVy3
z=Iy!j>9Cf5dREEty?6G6ZwvXml5$7_JE*7Mr^DtX3SUR3>y+^{6v7Tn(R2tSO0zhe
z9pWkt{wb6~(bYx~#zi)VOduaOf|@)99S`YS|D1G4&FL3f<#z^mQzY#9r*|=K{822#
z1JDEX>xVGQ%O7%K!Fmj5$&ig(K#w5B!O)s^$mGQB($<ZO*!$2shUf>F_RV)#z=nzK
zJdna(Y%Hbv^AKp!UT0{waw+Nf$@XILtVm+Azm!AC>$7A<CK#7(L88tkpw2a-Fe4XH
zdPC9;QYyEHcu`qO87$`CN<OA@84qQGD)3m%J9h{e)K6jTUrubv8^5%H#t_VgIl%UU
zmY#5dU-Y6dh;veQuRLwOzd20}^5&1Y(=tsEvBcwBGK8}D>7yedEj#;Ezw(rB5saIz
zy{ezJqM{$o_+Lg#?;MH~`c*2IP!pX(^K&0ekJ#X1oeeE^Ak!##iG+!dHwvCXcxJ-o
z+$b*oVOs}Bg{nG-bEp@pSKh$EGPU9ubS%jaT6}E^!VHO3f~-I4{7X8;l>C*=KdWPM
zj9C2aP$A%=v(bL;+MRBQB1VY+|DlP#39*7(LbeP4Wk{zkBbA}6RM&a1TRIz`e(Ih_
z9a1>)V|}67Sl!Y%U_7#`j~%}wWlFMY@ihIp=o8qVeB2dcfb1Zr)7YHSUsPbn&c1~i
zY$oderB8(Q1{cp6FGZw|ULD)5K}djrPI??nk*In!#!T4LP;PMMp^n=Nw&PQU8-8@-
zJ^}fEA{XEjDLB@AJq1^fCemcj8zAo@0PO3BsQn~Qd{V+)u=Ms7o!|F<{f!SbA-7Vu
zkbu?V<8u9p>NRWQK+ZJX0MFWe*5{_1kQ}knlh_!qM=WLa_}RS|+Z}#PNgNEW0o3nj
zQl~i7Kd7(CsYe@wLhS@>E|W<YR1b6f-fFKGd01UBr>@%CTM}&aGBb^br$+BlO}dF;
zPj9Y?-klT2o*mRVy!?vF=HcZ@55k_?O}}}Ahrf&8xW`pBLFD0Pyw_o;&yLg!;T}HW
zhq1~tyIv_?9oK~KcU-&1{W590WA^|@fvX<j8iM6YLsy8_LggAMJ#c!?Kn0Hi^!G97
z(J!V)rOV(WRtV$pO8CWaS%W`De1WKFnYD=R%`jXN`M8yyxd;co6!r5g_7)<Kw20$+
zMlK-4UG_ab9z8u8ErQM`xypj~QOLNCIa9MMsE~0IA5y?O$A4e}R<hen{7&ZAPO_F-
zLWHSntj~p;cV~zP5wX?b>nnHsp*N5(Gyd>G(HFzvoz3l!oVCZmXCnPHmVhAp(?q(K
zDV*4_#Skd5xPmRrw(;A0^i&CnJ)|6OFutQhGp=rOgZLba6c_lk2wr>bf3+b!(As_$
zZHd6jRd;YRk@=y~>|{mscoSyoWC-7TW~R4vbK45`#xh5PU8AT8ua`Z@Y}K6tS9s$r
z*DHvAG_h$#zz3fqa)xJu&QU!sAX$}%8YrCTumY4p>AK8}Bb{NyP(}VmYN!HKPn$-8
zpXktaGN&g!%K0PC>Ll#Vj1?X}t?2an&fW)U>Tx#Mo3fA*EHlk1Qc-KoS_8eW`=_?@
z*}xAITe^QHT!^TAFjsO)Osrc3Tn@olc`ut}wKu(FJ2c#ixFjO|nJWKTotN}R=iVjg
z5yPvm3IXq4|9dtZ^~Ly!m;wHu#uHHtt^)bfqB&8sq@LWMFj57Ya&vLe!E{b;gD<By
zhMWKW5oD$e(u)I|31wY;bmimwbXPeK;^@BxiB*(G{lDWvaZs+RdeCMhy1YJ~M^q(|
zY^S!rAEQ(ZQ9=m(31&!?^fw9QvwV3uNk8%0?w+L>HnrLf1v!)me1~ng0*J1}gQA7b
zmUJXaD3iH&EqlJaUe@eyH>^KfhAYv!g1;VN{2)sCjJQ-ojJW<k7f?)mGPL6cyLqH|
zi-TuJ($<%1-YvS^&7y3^mRb@5s@;#;M$q~&m+tNjO+B3rdt?<gzJc!neY6}rFI;<)
zIODhtk28{UQ?CON=cNp9O_=;kq)8eHkKCtU3yeCv7J^~<<ZF(>m?Bm)c6}$bYl*~{
zNdLwyw+>yXbK|F&XrSA_%UrEJVKT;2VWMhQe-eLC*=I_z>Hyter3O~4+pKTayc_If
z!#m0V?Z02(As0oQrZ^|H+!ge_>m%*^^nVf~XBCJS@H7@Q40Zi|2xR8-6CQl`CL$lo
zn;pEQjr0&JM1hZJPEr3`0m_qH_x&@w<dx;Kc_dvwX5<)rZ=kafm2G8T#_(exQ6Ff?
z>7#N}@HI`z8n)|XcqxSo1~)R`c8ICEIIM#)`j*0!F0zUPZwomizLET<_;+X41^*oc
zrxgRsa-;MRr{~x;VeblD9=VeMmTLGz1qFT@u8wAY1)hpR>dC*6_{(0M4Z7t0J$UCi
z?ku@ky)CAK4!;huV+cyssZzvAvHoXLNVgl9{8J+uUPS!vLAJ$zWc;BjT6V>MQxQ}5
zqTx7L_?MAFo`k>NHBZF;X?zrv)M`osdQ;K+pXDIbr0&}vFj4!VeySq7)!A)(_V)k}
zyJK-i68tE!soal&tP#!M9d8(TA@QFpIXv%Ir!s>7H~pV(WXR3Lusy#G|IuC6A(&?E
zDuamZ=YJI`v-A{n6Y5}yezBcMB-7?_|J5IUj2|gp6cR%05l!wG1#w7mJ9+WP18EV^
z0To^h+LipRT#-F*HRt~)KZ(b0t^O%`iVqT~Iq_y2N;wrh3rU{<A^1LoGEU~f_&M;T
zC?VsIjq5;f$+9>{S`9I72G56f>7Y~+v)SSDk@<t6-OJe2ocY-LZDmQ<0$dyURcgEg
z<dJK$cUlq5?_sdHrnmlk*;?>-(2OViw0FEe|770+@2W70d!!q5A00X|ds+2Wz61w;
z%;%mV$^v63M^V(3Y-)<Y(nCyxz9wjJmW<RDcar&{9AdUn&M9!g^4onN;F*l&|5~7Q
zSMET{iC6|NY6UiOGd9fFMT<*Vg=J}j%FWZ*GYDS}EJufaNT)#T6~jljIZ|8u_~S^N
z@`mnT!fufsBlUcb(FSsw$)@{1A{(gJAg>0oQ?8!C#yHrda@wm~bE1BlKvcsPRM^Wp
z8{hG~rVL1`$a<?>0h7vc`LLE5fjKj&1`;O&`kK+#@a!kK99S;W2S(sOph`ZLp=v*4
zt`kQ&^+dt%TvyMQ1@tQ8y4=`r*;)bpOI5J@fJ^+5l4cj;nsb)@FJV64`0l68>Nz*9
zC!|;Z6ZT{JTApoJ?_rOLG1sJIifIq%WpC6&U&uKWdv(Q7kQLuKl1{F`49{z?E#%C*
zp)3l=$i{hXsaP61@+<f&wqm%8clW=XT>$?Ga9z78(6e*$15rk9QHOcQ<AyE9p3Oz%
zKWhdR2J<2Q^=p2f?S`nI=9?3LGRhBn4Y$<lq-jsAB!0!W=$Y)#OoUJskgmokSgi;2
zcWdyWJ~@;tV}`!&C$U9-qGHy^Nj2I}yp9Y)rNX0pC{xHO1DBKz7m}Hp>WDQIHZiU@
z<+H-}<Uq5kKMIC<o0<$W`3!dU8Ko-8nIfx7UTk^vhI*XB-s=w|YB|tzCC3r*i$@dS
zIbH_Y@GXohOWAQ#h^<DEiEV0#%M>SW?7S~_NW4sGHU?ykG%`k^&e1gO?TRV1s9c^Z
zFzB<B@*H#yc|2?GfeqM(6`B1!4{7l#U&b0u{s^cDv{*m)YB?pxl(_Z`-B=09qMnH3
zXR!?(-6MH@oG9n#R3QgxaQlH6`r`r3PLSaqG5`a}cedh^>{mJS5wLe4Y?2ELaD47{
zbnjW1bnuk!?d<cj?VzSyC7{Lg;NFJa8i%#_vXZBc6avlWyrrg>JD-XvOAoz2H-nd6
zNyKUO{KjiyxSSqK4ZHe^FKm2cgc?YSRM9Dn(Et$~v&$2KnDv|U;HZuE0<e5qXD)|n
z{R0f%_=T;d%1q<4-#vyoglFEMe!5`@WC7`A4hE7L>VM8B8ZU>gy%)oscT?N^w4dL*
z2=xEQDm7JY&G|rn(dvhq9LXNG0^1DGTSQT(Nrx*xAOE{6+Jxz+4oYZAs%q{6!wrf^
ziJ7=jDH~Ru_urj<Ifa=9lxfiJk40}r9gg~a-!?9=ZPGR+nCM8Yj1_{L{)Mg7=X%{1
z?1bFMqaNc<Gi7FljcCT?&Ifix-od0YMDs%KZ!BfrX}I&Br59|;@0r{DV`{a&E#l2(
zV%z+BzFJ*YVUB6n$as&abwo+pZ3oi4<0>j<5^rCzu4C6Ev<JGo_$hEkD?L~ra$I3S
zHOTtM^c5gvX`c}$7qm(XB?HTLV#pQjYwOXZQRbR9KPjDuQ}Epb<53(=>)c#Orv0~x
z&f|(i3yos63*v|A*+j+~99dJ%r1)0B@)U_8cwEnXQ8v3*<uLei{;dgyI}>mI>smt)
zud)AVV^CiRZil3aPetwEdgLy>MD;in-HkX$)W{oX*()fo#u)o9MdUU;wzKA~gcB#Q
z0N@}MyGPQL*p{|^tiF$;X4a2EGxdb-^eL{QHzeZuvsS>ToHxxQc3Ck+zdiZjmYMAD
zrRwhvU7rmVrZ9T3HvEh;mHtD~XiA60bUK0I@|(~T6>!Tlvitg~fk3tmiX)x+8p>jf
zUA@ZxG@9{VUQNJa5$ZlvKXCRe<;D0L>yu+rnH9^4xNP|vM7WDUe}MJ~&A3pvn#Z~;
zz!^3Ko+#veY4B3cNOintCMy9=f%iGv?~!ANbjrjP2VR7Shc6B^btcE*tAP-IL!2SI
zM*lQpDd+nbJvv8#{4n`Md8vwZ*WT|fchP=7$5C<3d6x4i?doeVn&@qbnM~5wk((9x
zm@ASj>@L(2qMAS+*2_(N5s%e&p8|sf{*IZfw+z_xLStYm^eF5X=M95&-7GnlYd2Q<
zyzl9OQ6J95y-@!+R9Kf%j;U@+_0;;%V0!m%5=rPybh6W-`jXP?Vq_3M;luGfj#|fO
z^%bcUHE+MUM)<fY((5<9-Sf@VE+qY3<g+FR?NugjbUO9)QI6{a(fi6u+{&`$Y-a|r
zm6Vr7=bVF%%nwh#D5d{z^!u4{$#&zsTXPDMUJu4w6O;2f9~Ejhv7eCk3B<V$I_sfF
z=RPx*#_koa`u=ArSK19<;=9QW%YUhAm?Bv&qnZXk)XSWred{>%C(VNOWaEu>RR#&{
zxL;4;mEJeMx{UD*9LL<s*oj_Ba?;qdk2i-xjPnj?#EtXer>+J-Lefn>f`sesI<H$7
zoLn1I8ukf271Vgvt<%gl4{ne1|1mFh){bWKU&4cfDz~%}?&^pgQefJDOyA-Ph^qe8
zjt>*ZHTz8JF-0#RjD9esp8YM)m0HH3_L?bUYPI^Lrm30!%Y?nr+81Ccgh|%H03IyZ
z$>Dn*y5=|JmST^VrMdL76YBwIpf-K2;~?RaH}EEm%hQ$6qtvpHG`oac=2~Hmn??3K
zCv*i-M1P$fdkbwDjFaX^b%r&oL#og_lHcnMCI()Y3lYFkJe-8WbG~@*jFOmTZLDJm
zPa-3Qx;p!GHuW@2F=dF#t;&1HtgMAx{^Zd5m&q2DQa!J7Gyr1q{M9p_^SXj=x0x?n
zd#roEgk_J7-#}htz`sR+*wU5xq&?}XD&&zpZ*9T~7B8?-V_W-9+FM75pnczV$KBkw
z0!@D+jr72G*>(q`?BT%{{;gEEUdDj~&K84t?&n$tVggN!Tqa<y%dTTob=(SI|KJ-y
zoxAZC`Y)^ZQtF?AH(ijf*7@%#OZ-i4q<LrAbYpI17l8C&)EmL}Kfqb5?-M*kYDh!-
zLs${L95wa^HW_Gi>HV;Ufw*h@>!T2!UpBVN64NI4F$iB>OzzHF(>uYigYlWvHeSBd
zeNq2*Rv^BdI4sk{<}Sy0n(|@{&}U+!pn0NDS9lVCD%q|qEv!${x$aDZ(r|?Op7?Yg
z_m%~wu23%EdgR?d@)4Q?%s}44jnnv9t)==S)BE76x&BM=&htWttGWTJO<(cjQFJb=
zw}|rxE-Y+wLo*(?_hjo&WG&NFL_GJHwj#<+%!B1fJRfBz{RN*!#g9x~Cesszmm)|O
zTfdJZyCv_xAJZU45`SCwxlfJO+1z;ozyHYXl=9Ke9*u=!%{a}Zsli9;l;pG6!J5q<
z_-`%%FW>n8{d4_KBA;i$K2ZM$gRdb%!tvMsCe#o|q0!xYnR>J^#NVV*&$&^vI3@q_
zDr%oqYVfgf9$$LK(`J3>-thlDz3y|?=KsKhq0%d4fPiZVYu<+rGGI(?c@BVMFlBEk
zIZ>POqdd?C<ep>WiAx4H`@BCV);s-QCH(KhunvD@+mCkeX2)_sp89BT-|1VN(syvM
zeZ;UHFZ4J(rt7vL!gIoI(gKnr1M&Vniq5;|((6CBpyr=iK?mcbTb8=!jS?vee-;MB
zYM1Yrm$2X)prm7I$M8AJkFH@U5qmnAT(6wnP+T~LQyDJL;~vg1MlT;jkz?a>pAu;r
zI0b)hdkBp#%xu#*p`JbekZo|G<Zw;rtlz|OhOwP3TWJ;`6Ed!C%G?u6eI{eC;TK=m
z&ZfAw<m?{qgLeWjN-(14(bb2bF8if=o^$hl<s_4UkyDASs)r;t-a$Wz0-{5B9`p{y
z9)5VA36#cgPup?@ID4ra5}WkM?dQ3xC1y?UU+*uQ2WqlmM|#yeWbFJ`h_30w2*a@g
zwFjC5dpKPG*zi^jTr2}inkyw;K+S`Y^GDHDqH0R|&K;cR`+!RS_G<Z19i(Dzv2?jY
zuSkBILYaZuB}jZ7)6MMpGt!tkRW$pfnCtn=Muu;*>7fr0qicyU)iaP)GS7Yr>P3JO
zoXb#Tku5+FS@LE>aNZpM4SbQ&-2L@!5vo&u{@HME%JMDLPd&(#e3b*rs(!{vb*|uK
z^jk<e4c1t<+hnz+d%!>C4Ec`oce>Ax)_*i7+<0S*wKT=^l3jgYmMF#VbaJCW@?o5z
zzuU=F2Gl(1e*!%~1v7|uM6_3Ix1yfsz-Q2wNNg_TOO3w^g5k4E>`e$W{clw03$10b
ztxkRB5%6H8Lg&Qz0?QSigdguY55I>he;BFM4skGGtV}s%Z<~Drp5g;aB>@-iCS#+f
zD{=Pwe^i8Z3T;_J#RMa!NF~00zEU#W96~**U>TX~P+T0gS$Xwd`Wd+<UPH09t+YPN
zq0?xGdl)FP#AW$rO%bd(ELwESHT)QoWx6+c-k?+M(J7@B=<3(}a_iL;Ce%C?byzlc
z3&giJ22P#-5%>&}{8@jpaWY6uR#O;jz|2nG^eVhk<<+<l$%1s}STKl>)KK)!XKbmH
zCgU6(YTN@So;8Mt_z6^pHbjzpZ&hsuy1~^I9#OGJ@%=Z3R%oxWZskkf`=CeCVSj8F
z8Xf#1lL}>VZ^CHVA6t~m*Wz)C-?569IakGt^!ha{$u6xR)*dRa7JOtkoOEm3wkua?
z<g;~Ux$|o9`2g-KKYGU1jl(1zd*OeKa`8r6!$ZOI)F(9?%=FR9E_Iqq=7*EVt*otA
zFZ^pwXl@3=>Gw!)rGEeXZd(-FhtScN62auV@OR`1Wk&SXtmMp#qMUQ7!g_1iY&@jg
z`R9XPci(27iPxi?9RT|L@xS;Ok6Kk)^I2#$?#zNUJ@y{td;0n%Z0JZFy4=5v{3OaE
zM)CaNzCX|RP{waQI&R$9<O2%QK0X#Eu?XjdhR?0?=rw)%n`lOEdQH30n%2ax-%A_J
z6Ag#DN*k>H_|MsL+4B?-GeVzi+@$Yd(G}!CuQFnX&{HJ<cGSP9^^rF$=yixs<PFVd
z$Bh_l6)GCbXpQ2Rfi}<*VLTo)(+`NG?bZ8_X4b>%a`?47%R)HyD;`OSy>#5lD*wyI
z<TRDNDaFs4WA-JmoRDuLxk5WAH(G51;otCOVB>Id2N7E)CSPA}`9M?%Uq*c59Ju51
z#~nNQTWH2R&rNXyUYVY6tO~`;tmdoosBh@%at#g|fZ_gN@S5|--@RD>3@$QbLzPhD
zU7#yutH{2D<r4t7m*zoj!8QtXL+P5&jd-{lPh649URY+$w&BO>4F}_XoFu9miI5PP
zoZN~xU(RSm&YRC(t>0WW#TUXKW<GdGbzWSGYOjRI=m5XNCX6Z~&uU;{d5sNO-1S2d
zT5*k90U54{Jtq7lJlNbi1gaxaj1%MT@WiKeL|o3K-oj;bQit(8L<uqBPjA)dT^&;C
zK6ww({Y!At>#omJa<tdA0r`phf6!qS=C!~eavG4Mfva66c|!-CCLb4q?nA#Rv2<(9
z8CwHoC!;MO7eqfzd5E*ua;b6R7@e5_*5Hh4%jXz_TBpE|QRhDiX5z9h|Bt$jzxp?$
z4M%Yq`Ng?WBbthR5Nd<2vRtQy&UATX;BHsYE<-Rvr%He!6mgr(i75c#or>%4*Bm2r
zdRL<;JfePcnN_E1q4*RYkQ8nU_N}S9gqhc_?6~X|LB8~qBLA`grxhQe+#<8Ho&O%&
z-Qy@!Jk08zy58oNdMfg`PM*G7H0Iqq15Vnl2QGnHSFjY2D3kyVx3-S?$gYpAbC0>r
zvTd>qlS4uyqW#CsYNWQ19)5YEEjoH5R(<^MJiiD11QG*i>GK#ruk0jXC=H2A3s+NO
zWfr5J-LbZ&&_HceE1`>HCMjsVd&{tsr;hdwnaRb>*odLBC%bbDq<Oa&XnlI0<||${
z?c8_J_Wtz-Y#@N*9KIi3R@+uLTOhUa8-)!sn(j)54oK3pF#LQaK6W)%Hp!Q@n3(`-
z6bWSD#GC=Zha<Q*OapDz(BIAjXz`^pAJ5LW7k!g-yL#V)KSVWkcGjc>cAvx4XVR;w
zsWo?YHUWvE4*%n0XoW{8kd3eFmZif?xBFdsy*obFKAc;AfGK;@_fF%t{XSFnJm~@L
z@J5MwIA77w1pDE!T8I4k8PeY+R0uf|GjZ(UDdba*uxGgDrspRCm7=R@d82zOkHpcY
zmGsN|Tt6Dh-(k<8-Ne$Wvi8I$ACLk(&%>!5ZB7x<Hqc%g=!mAnx%@oNHy+FM+1!Zc
zH_Vkh?z^t>n;(B*LAVKS>mw<!IyYGOs(SXL@f+&H>TH`0a0WM=!qXXtT{qkkMmdWG
z17GF@z=l+=f%)@3D-UG_I!VXa&@z|)w*E)E_UxjgdhUh6qDQ&nWj0rHdA=%rOf<Os
z7w1%w2~0arFD=1*`*HAjQqs_RZWKB6gw#_Nvv6`q=&B0cGcwW(rP+4=b?_+Jcuj#7
z-~)w2ho6Fg=99)Uj%QlG*(`rH_JV}9rfH<(l-@O(&gRXTw^1yerg7@}(|3xVXP5C$
zm8DF^{e^vV+Jl~$g$XgUZ!dMAYJK+|=&{z&hjD|>i#}|5^valkb0{k4EMnZ4>1K}2
z<UbPf=x+=9!2b=74M9)Wcuc&|FbbVaF&@W28}1uXrqd##HI~;eCsB}6J>FA6)fD#E
zx^e&Z6F&02(!b5V$33_+#jE?C`WZ!dh}$lrZH*4^DF2G5#vk<@aQn;rnbgmKrsnaX
z(h<z2EK?Q!rO()@Z1P%o>MeLGBiEY54ZgM#<osi_?X~3#_4sK-nNG3^mN720RGk0T
z`=q>F*RJ0vus4=T0A12&AfHF1>OoTxWjgV+`6(1h=~Rz<Nm8_Ve_?pi-nJf**78|+
zf_zcxN}eG%Y8Nu%L78e1I5wR^KRL9aw3FxMq>waxK#m_42pCpgOn+n6?6}{Hfu;(D
zF9~y;`kp1rHrCwH+Wg+Uu72e8`+DD*iiFau%|WP@S}SYs8_{F2S5BY3>my0=SHMU+
zF{%2wC54F7wiWZquADZAB4u<6`RY3=faQEEA<5Yy`^@e=f?*orK?x)FNbtOJd8Ia2
z^#ULNb`LBdPmXxx%7<kb?kS}05Wb1t53&D2G@`ob$g)uK$nRGV*-z)~rOh}QRu1fn
zt&9-fY&B8J`#V`RYWPByLt}kT_{6Qg;<O(v6W2T^^Q=s|EGMEq+3Hoir6*9%)9rd*
ziruF%1e$q9%;D4LcBwR^g1SRNMMv}V@6D~P>w{mdtnN3F-D0{zKH#2UID~(s_2;<U
ziK03yOErFIbjMPBqy|w-VirXenK1IfVKdUb@vGmJwtkXvbCZ%$D}qABX!)MM#{m^&
zM1zOJ>UmHu-{D5!K*$f0S2MY4+xsz}z8mvZIQXy)M9>yi$c>LVx)ckedfBf}YD}E7
z$|v-S>REUymIE^t$@^e&a<?JRhkC)LMbPIXwUOj*&<|TCywR_3%*EAZCtRErCrbpm
zEjF9uD=jz4l`<SG2jkW%XDi7)^KJOm&-%Truv^*vRO5(`IzwmX*?84);3jIQ^T3J0
z#B<9zoagKlRj$;)OWT~aTWTCKh*P^;6Y>H^VfprDqo?HqWVS+0Hd;b7{2Qg}AH{Eb
z?ksYA{~+Lhxw_-Sy8lP`^@o$scn0}?M^lT4i{G2fvhR+-U{==QFd-o!<cW|vH$I7y
zxf%YLxNG->N`!^Gm?G1BwaEu>p0~VjW&RoYqIUMLE8++sftvuYt{B?)s5$q(_V~CI
zyd#CYMuq3Ua^2wgxPq+)<v1g=nMjqu>miyA9+RxmX*i4>9DI#hP^!m`k@)2`V`PrI
zd9qAzrC8vp<<ikF5Z|4h6WB}3_=98PnXjGwaoLVnK_Hhi|KXrQ5k2)hZ}_bLS6QXI
zN<SwQJX{B#bU11<cOR3j-f=J7x$pC7vR`ofHskCm@4^t2xd2X3Y@}wvYjJSRBag+)
z_vSsyyBAwhE9!8&3%ZjnpKXkU7D$wycy{)QoL$|#7jT|%HoGahYO?sX6|>4us$#2^
zE53H^&MR>Tj(x(FPD4GO84hfU)|U4IGH4v%Eq-D?T=U!W9i?w;NgQVz>TkPwn~yyL
zxIUu`VrBN4omN)IN=i#lo+xvxj8Kv|n_R*B#_2+Md9N6Fcc=GC$dID0ZkG7xXth!a
zF(Q)-+?9asbJ$5E=odN0@9R>Z!_S{1ds>+NV4QECU*PX6eb1LKF{|uF>g9IoTXnsG
zTA-w*&*y@uess7*ojY?MX2MG_Uu`5;7SFaJbY0^|e`5xuSmre<oa70&pnJ#fCRSX2
zHai~WmRr~B$TUu7T2t#u%bsnxL@rwMzFj$j&2nj+k4H$X&SB-sif7PJMN#vn{dz$1
zmi$Vb{cgG4N>!0(IWwb|xo&p%P;C2<k3)9Odx6ddq$=#5o{@I!?%lhU>%)%UU1H8S
zInPCt3aR(pF*kLU?CP2-i&=OUf_|aWl%DP!#msm&v6?fhJnF8c_L&dO99y`1f4nv1
z2EQK0vg6UIo4r35r}66fAp>OM4M@rrM1IAtHi!pijVQ;tAk*oq%zjOn(=uGBL}RPI
zT|^SC8Jb&)pO#DoQ-HF24eQcwy4*nfMIFXz56*W!vZeJfO|Ba=DBZA3Y%C6tWH_tZ
zOcH;w>6Ll=Kx^Hj!1;tTwSAxO&v+Gz`k3;Nbg;NHrbsMV3xq36JVwe`4vbYa_)+=o
zO*T`L9Y2tgza!9AlGdE1SM4SDuH;_4;HH@bP@2{Fw+#_q0s$Wq**i8gnu??LibL%3
zN;VRCi>wT{0vy$ot`17A^k#K@<|`=<Gcev&3Gm?(;7%l>OvIg+Lq^j>w<>vG2UK5e
z7MJZZaWG!R4eLwhJ5VPZVe%(M7YyNQ6AXZ4LY5QmCgK$!|M}{}Jn|wA{1xVtBge;`
zF?DT0sPFA!$cxw*g_p5G7-1%!{@)UQ8v*UV{qFWZzb&eIf2Cs9o|?lpGqz1^e`(;f
zLT^w(>DJw5i{14tTY~ICweH<1J!Fif<z#gAhDqE#f&yJ_E$b_0w$KF)zV*+@`Q=JG
zwg#IoS><P+K7KP)krQ%<t`g=t^WoT)wvw%>o2#B9-nA7E#6~4-<G1`>{ETd64t>>H
z*jVy!mZaJ%Z!A9^oOW|(_whES<fL+MsnGF`L2r7@Gzv(#IrhY|jaD-;4#X*JZ*QBB
zYdZOCCSjP@R`iYz$#j30jG=B>@zID|=$}S+ULH3Kg~bf|$6=#9-no3>WPYn&3(p&z
zmNhA<pMIZ0t!HQXx(mD?Dt*_kuOK4Cqddj9BHhr@u-*?~5f4l~)qYW=07&?97!N;r
zuC-ltL-7J{;)m}mxgx@yiG$8U6H$bK2DOTE0#}_{fKTF(O|@?-Ifw4~U*8^%$z*6f
zSQ_Q;XMVWupEanI*hkUfcxcU@Z5mwF8RKQH;!9BT#k{cZ8~k{e@LbK_%VopCaaZ@Q
zWzRyS>B)SXpHc6F&J8(IeWMCL&~2c_wC}zdqJCo7xNdbLI-vH;yViw8>CdLGUVkp!
zD_)qXNLs&{df}jhnnIx`>!qUnd0VMj$Ikx*YWpC_q-Z@2u7uaGwhO6LWWP6ZBhRGT
zPQm1=9rzywBnX;jpWXlIydfjQlLsRM?k`t~$?-UMf-Kfpvpc}8<N+vN>>kV_cXc%v
z^Za|oxrJrU^)H-?J_Ny-0+r9PlWv3#z61Szrw>`$rK+^YrK@ctW;lqiGrD)&fxEw;
zeGoU`o;Wn<v`J;USzm1-_X0yYZL@iPh2upu14DZMX3Lh{MwI<Id~6En4ko!o5g
zZ}JtdTs|T30t5DZdM)ve=<Si(#5aM_`?uZmQhg_?X-Ro~Zhv82P8RS<<&fQY$C_Qb
zAoNXty7Nhd;acxbyGiRcxnIiYB9-RRdn^3So$EKxAF>X3DmL1H)DvQx7axKa-S}rW
z`_-8RjQ+`odt)6kq;PBhho`p;i|YNphY<u(M5G&(kOt`vY3c5i5*WG$h7#$N?(Xhx
zkS^&SV(9LU=kWRdp8vezy13@W-1pgM*IIk+5M0KD(BIc!e6pzhne<(~mEj>VQ8u=d
zXl^NK&*|Z9TF%vt3G!&`Z@8duSIOMz-MQ6FJC4hT?5g9m_$Y!P?zyXB0hW62Zkz5c
z_1PN`R5*;%&T1NTchW+dNzwi_Z(7Txbo>EEGuY3wq2co|h%b0y4&Ku$r`=A;+Wk58
zTVLV!F{Rv4Z2&9f<e@9!dO?sLLS8y<OAprRy=rZ=pkkqb{ZX@vh%LV;iuEZ$)Ya_K
z=JJ?e2HSQU-R$gbLL2XHTPXV_!t1vLu1R<!0N3kggw>NCy=dkMPnrzTWJ}>8O4W_e
z{M$<ztcKC(O+B8`|DK6a*h7T8>*+=`5}5-84X>N2ONeu1jY#lm(FuP6U4YNj*ufW%
zq{YqebwmY9u8Y<vuToxSnhAAOr9%RVNPik$6EG2=kN@i!B$JZH`K0r>aj8ccEE{dL
zL;zx_FUxBga$=Bdr~-%MV$HL)!ROXxLQ%L@eLkQlt~*VsjU4b*;Fhjk&FG%KieTtd
zC%DCCo~)JtpLfpwKF5?2|BSirC%|#pck7zx$$Cc7jpD<YV7K?V?{3kk{XH*PyI%^T
zQfHi1o-Fokd)@m!{^m~$MWbj=PY##Yf5yMLyF3lex|O%Lmh-ulsi~En=fPMUjMQi5
zt{k4)TsV%UEI|2Xv^D>FobYK`SXo=XQYCs*|MYtN1tk2{8OXn+vVR-Ke5L+B+wsCd
z(nVe+;3ReA^!`%sULLr_NS2O`eqXm!zXTS?!D{W?k8=~v`A6G1@yPd4(H<x{%nRLZ
zayiRBEZy)3S{L*cbY=O<i^*6vE_GSYM*Nli%_e|1E+WWQ-|@pz2Qj4fH>B@eUl<9i
znduhctg5=zyRUq&a<^eUl30y^fp=f4rO_<Ap18ee_iUXT8adZedcRUF7+4W94!SGL
zDg?KA(lOEr;Or;67^*?ROeOfS%0vz&eW58wtt%_}q~W(?u6<>O$1V4;M|!)`J&>o8
zsQ_)4Y~XhJeeOebR(a?sLy&bn1^HsJWR(!32Kz>UQQhbPI%`v_nq-BrWkI#`RA%6M
zU}3ICySX;o#&>a-&zOs2{|i*8M4TVSSC+`3U*J~WoS48l!=BCW<uaTw-ze<#^4ygb
z9l<DNh{>YNjR;@eV0bT+{g^%AR^`7)9uO6^yLn5To6A4K&MbKKYo<4+E30%Gl5I71
z#m?#2Sf-lKw@MoTa(jg8CNNtIUpQHkn{`H8jpZpzgax^_BjOU*7DDVN=>mmX;Q&TP
zk_VF7;SzFtcz2$>jVrsBDlC8nhgufABZ&TMHP+>cLTh@1G+R}Az9~Ii2j79nhG=e0
zHoZuH+v<K~#^Paj)3ViwF-m62at;NQ7eO#<2;qQOd-iDS`TcQq8IQbyJR<36Inq0`
ztHS=it^NU@n}HHywg$vSu3|7xIH@aE#XdI^RuXd_TN4w?NlEYkTK+-H{g~!ZC*=`4
zq^-HmuG50zxrNwzv8nU#5iY1COMyC+Cc5{Mnp#T5AXMf5jLppRGRTq{*W_DNp!<r4
za|2w523+>iK1bX5NRgXhx%!~iYXjeAPctotBu7DD4?O>iyYAgUjkv8usXmY9L5&m&
zQ^-mF;vCuFmUXk)wB--jbiEY{5~dm=Hc0upv^AE?en~Wi|59{yb@I)2Y0!ypq8<t-
z@0a8>-a}(|TWwJ6nEvc@H=<N9F0Fn9cidW(Q-u~SIU(#!`iI_PdFBoquLA03fGG}(
zY7=wfy7pmbEvT9~piG1JO>uI)p-DH+@vY!&dyruBKlAQ4o*}nD`u8+wnRN`2KczR`
z@AN84D_#s8(z;p=b6NfnC|<vx&XkN)?DODY0i%DkgI=Gm`uClVPuII>Ghmg7zf+UQ
z4T-D!GfVOz>Q`@GoWNsYa&mI(<L0Ju{xtW0Xd9Un;_2r|dC|WB8?-fW34S#dE*$PF
zMn}8qrJY{K6s#zC)rV{)&ry&`GzRagET<+0KR()&V!oH1{A-((%kg`$g(e-wLAsOP
z*HsGLd1|+qHPw|(-J8>pg+8gXZI^@MV1IRvWlilv)oDJb1(()BKk{nnFA^*2Dg5-N
zi#V@ujZo*gtE9ddm2R`>3K>t<2??408a5H@bNv>ndauDgHZ8BS9VTI{>@`f<9t$Pj
z)_Zfmu4J|3TH(=F=fE$Q?p{+*d=AZ;zkx3G+(<uak6+etPOUE5(DyBE4`#uL5fS^M
z&wkSrnwj^CkajNYg^gQemX=-WN*wTMCe&;lY?{yeFjCPC5y~SIbdKsm+*(rCx<fEE
zP=1=5oA~|jO$Aa)&YvY*#D~+h1?P)N?8SSSMQ#B_(5?*X+MvMNcMq)|zCTFG3oW)u
zy+5KO0!3DhF*dd5C&Ni}zwRg=R80Sf@KIdjZRf_LcIUDqd)NCXer6+t(h?oQ^SidZ
zg%9yoEhOoJr24g3;>hZPuKZiA52!X4ol^eB;`*eVAjjdmW5%Q)-8EAtiYf<VhGY7#
zjcYez(8{oZ{veX#TdmzUCs%&$;r?b<gIjyc4Zu+T{M16*ZRrKQ(A~}C><?XO5TVdY
ze+n5>LZc-VgK4$txFs1~FlX|l<RJma?ld#PPZbwu?5eAn%p47oA%%97*zYj2*SE;;
zAvo6&C0NFidLxW#YJgv)>*~0C-&u9BukMhy(o?zTb0_Kl?&y+4kfQ3-3riM%&$0Hm
zP;Rgn+U#Y(rzt9mAVNpnBM%u>tJN)-t5_hlBG9>H(1fd4U)n#p=hi>Z8}Ouf`kNP@
zeV)8`4crkrdA?&K2u6u`SQt@Ck)+C!+igNvzCD+ss5c!kk}P?!Vq5UDqzmEFvc)x7
zOS(DZ6p_4z`S);-SYO7+Mr&SAb_G00lEqiMjh{Mtjzm}DT6TKlp*uKj>bD<F1s*CB
zI9!X1IsP7y=U9OY*GqSJIIjNKXnbm5(f2Z<5zhQ5aDGwW>OGVPNxtqYakwp!Zmpma
zxTwr9>^>rH*ekR=kzD3s`l|KZrZ0v-UhJLZweFQ?Rja`AFqoC5b`Mtl{<cMr#}MR5
zTFSsMSJCM6|Mwj|{)UNMkdmDR0ZFzF6`4L(q=RTqPBTA}PaY0>;wQ3mG2d?Oi^v;W
zNtkt<RSC^C*G4z0`q#K!s5(2fJRCYI@>90lcg)BTIGr6RC*Rb>f+dxmya=lJCYppP
zCn*PMne4su`b9o&reZg>MsyG=E;KCa;csz>Y1mIRTk&lir$%eVC9ve_P`CZNA8~yj
z2=S|R#w)h}Y#Mz<WtV-UfNd=*er|2!sbYw(#a36lnAEy&6t}XZ9?pDwxVWNFjU7SC
zr7>VUX*yeUBY$amqH;^uuotG?;t&%cb2C3Ehz9B`Nx>+w;n!GG0lXd8_br&sGnZ+)
zdW$;)NBvA6-*&mzbicA<l6okiGf|0wSBo{jUXe~ntt{_XK61kSBs#@}RNVXXS*z~x
z?q0+B1o_pJLNm9qLrX4il=KI9wR(~V!hWsIQmgqOhct~HRes_~<<R^w&0wc#!c&D?
z%!F1OB*sLKZ)_Ab<SQ#+xNa1GN@61Cocj7%eE?QxVPRFMwP{z^oPwpdhg`?SyuXIe
zY*_*wK&!`!>Cc|G!hFe*#jxfR-=F=jQ|%TVILFC=3#aju)C4^89nSKNqM^v^kp2i&
zr|foV6fz$eeq+Wi0-PND$!9`L303tIvP(uzwV0;8@?qd1&jpBcC{&|fSgzzPXBSze
zhOzry8PF_A1W1aC;B^DQbu1jP5J2~{rslw<?Y^jkHlZos9nEp~C4c+_iR-)v=jx+>
z60sig4^UHmAxB}wPI_!{*10x8(^FJ$UJ`KteL9<+nxm_R4w#LD=?aBE*Dsq_r;Ohg
zuDRQ{1s&!8=K{z=ySEtV7qDlmj)b&9(@I=Ky#+!I?2SL+wU$<Jt%@qfqY2Q(BaO-E
z=y=`2lMZd(UlZd%M$v-Z(H-?rF-WmKS3GqY2^`5jR`j=QZ`!z7op?Gt)VFfjsx2F<
ze&cxh^Z5I+O}%ZhEi<hBOi$Zo+@^wjO7EAW=eW#+Gt{MGH8Huu`kA(c>KHGSp1w=h
z2Krn7!0pu{9Z6qO<eTEC5?Lr}ahz+cyjr1SgQ|qunz0d$W=IQkPS2~NoqVWXC9AOT
zEq7z1UaY+SyuSLe-*+dq#tmUgyYu;Zpycks{m%s{>NmXVFhf|ZO)kf{=KA2RO5~xT
z+G>-kh4+~Z^VKlYi6WfYyae9NIJEWITrU^8E!_APP7bL8uR|0UqY&jLN%XJQQj66J
zI+DQ(95k=!s^xlQzs!e%EUWClwpmeZ6L^!63}bNkAeA<wkFyO66&>`np1m?}hP$ob
ztQVmPRa6fmmynaXc~q9DDd%M=uNZP7x3j8UkKcza{5!NL>jc%PGxE1R>NA%wp4sI}
zfWHOTW+umSLX$bY7$b;{><@p*G7t5IrTr?}y0SCon6Zw)XDXZzbDhdS2U5+Wvuz%o
zbN`W(BF`?n(({)_OXmnG{<<GMKECFh`|0sXe{a0)bkg&7Z5i)JaD<;Ekk3sgFF>q@
zbB|L`5Sc(V8K&Pat(mZ@tVT$fyHSGa|5jFpJt*|u7=4<*d%8V%ODICal{c|wT`pKt
z<JXeWFHDFM%u-$T@q^!;@z|4wBy`G8^r?mTH?g#y+zMjP{|cU6aZ&n%#)-$dkt;jz
zVnE5~%3=AO3ht%~QHo4y4?e4S>HHM*X`SjOWU5(5TX28loE^^0vC__xZgQhcgxlLp
zY~!=)rNF^iaWv8z&8Mtaz04uhCPbPOBpt7}@4Ip*M}y|~6v@p<di9#hd<%jI9rBwu
zW$)|y=p#lMc5NHkiMo<Yr*D*(lc6oHvvbNnX83cB<(4Af^={M#2xT@n8+vM(O}1M-
zO5_VFJFlA6(dLAbuxV@0S0Fm1({(p8xzpUG)J9ZS^pS@27de)k^*wm*XG1T9bQ;)n
zs%)X?`QT~ED?JC;em&&4*1O+_puMV3YwXqt$Ev*WtSjYv;v2eV(EA*^kZsXEp<r$%
z^h>Q}$}Z^2c!E?RJpiHteVFlodF24l_nEVplxTBv)Aii>y)*s)u5YSXz_fnM`Zs6M
z-V)GQ@qFuzhHyg)2hm-uq7&QjfIF%%4AxcHSox&Ib^6I$8!jckXl0Xf&nX@0gbo+W
zPlqznqZ|Ai#jj||;R;py-qx(h19w(6td)Yy3q6t*9}8Vk{#2P;3cuD0bzJ1_=(Qgz
zd*v?~Rb|v(+z1)4eEFd#9Z!2}<vxcad#RHyqda{uDyAH<m&~yJ?XN>(M!}{w{^82o
zp>TdeSV-UBlH_Rg`}CWmxVbAk_R`94stQ)xy7}eHyx03T7fGZU91}RKdWtN2x;x5*
zcxD4d$!R=?&0J0siqI<kTL;5adOoH0w(imu$b{3JVm0nc%Dv{psb3G&0d1<h(+NUn
zx6u%kN)k!BH+S{O;X?d~<*Ch0D7f1ba<o&Io6<_6KhB016HXfDQF~cXSa^7z&5-&P
zCzkhGlrH2;Y`C9sATE~ER}QNG%S!$PYV1cj=n_`nt6t(18&ulayfoQyOfAK7{vv+h
zsI?1xusLaa+hN^%{HkI>KwM$kLN5dg@ShMgHiSPF<5kgS9xqC8A2n^jRm^aZFJE#`
z@aHvd(+gE)*hB?te7cFxbPoR?6dc2Bb`ZLg30SPKtTae+f=J3bG_Gs-%Fh2RXFpC8
zb*+~%zL?PRM;zyj3fOAPz1m+k#C2KUI;<vu{^_C3e-7>Ly1pk2Hz!yjr<J9z_v@aF
zBCd$yA!u$NQ4A_jb6uoR!UKR#C>k6GP;{&9D@k3Df}`^|8=rPNA|?Jjc%^vObxc=-
zkJmKSQtk2eb?S3#Ld6!-K1sKP`U(lLbU#UYu|~*CXLT`R_KC5~DrH5@`UPc4mz3si
z&ocisj~movxof)5R$r5uWsjDWoP2-y^ZeYtg_g{c*K87ql>Yy-SWyp{2jkBQl7pdT
zk`^=sP$(g1TbSQBbp$<m_xrD_X@C`5@3&O~0H#cG7(k}z2PSMvbap*BwPA#WAv*y7
zNxu~!5Dujt8UhL9m8*3cN+ifVI~yQN8B9y!aHS%z9*sR*KMWl|%m?Qqo!$aIp`#*D
zsTVVO;aTwKc|3`OS~~GU4SMvik;5U%+kvpLNepC*g_`q@o?+^giA{SsHS6KCiibJo
zKLjGO#Edp&i|8Js=@Ea!4DrIX>x`BgY%Cpbr;|8&`H#x9h8}Hh$Bg1iv6t3-#p~ss
z*`VCJuZM|P2T-(tba2%56a{KIGkOTL5Wq#ArRsjhFgAWL6z#-f6Nzc#lwYa-gBiAb
zpM%pIrV>P%<YMO^y>k;gJOEyWhA-XURftx`t6?~$xX9HGKR_FTBqGh-@Ug&MIv2NB
zhMqVel2S~!EE<{6vbe}eP<;FSb#D?6(PlDT%Z!7>XbOR{?8Y0NlCS?)<Fey6qqbVr
z_t+6$ijWq|%x@N{93h_gy)enF4w8XCM$nFog6p6_52G6u%QpB`30*`nE->l=LIMKW
z4=Y#NS1nmZ?aoHBnkIA2$4%$q@&Iit%h+z1ObWj<IDXY#e{SXD;*ZishQ`6H{9^s!
z(Q@mu9sA{b=jYO)&1F=`k2v*1O~PtEzWRdhP2CYl_So$?`d@z$-X!NfdYb}h=$;T=
z6T*V4f^AoW5sybHrL|xxq8MdV>OZh}2<M95x!yw*=T%ggVojC3ZjJqbyX_gGo&wt!
z@;B<C+`uMQO}&02qxZK!E1Jro9jn<eZ21~!r+fhJ(S~q4!USuDW7O+$%~g<qjbW4C
zw(?{DLJK>G(?c_{PChRVmTtQfA|qq_K(e@7=3}KQOKokPzL8+de<=KaBN_mJ&OB?C
zGvSB6V<3|}eq>0Ztja{aKF9fJ+O!zyfj9F`{c?JF^fDvRD^+E>?DW!FCwq6fH9e4;
z1ZZ%Zgfb9e^Tz<U6i8Z}&z+*9!A5HYc5G8PYu0g*@~<UdB&lC2`&9hN)Yb7QFky4J
zp&c!2sI#Sa#D>3u)pR+`?%T}h7VUur2rW^ItTdJ6I&R@GwJok-6)d`3$^4t~_CEeZ
z80M^_<S64v^f@`r>#sT9tuqHu!6-&SPTVzZ&qK~_{yITiUS%Z+xWSV}R?LQv7KPGF
z2HPOI3>=_Tm~NZ~gi=oFE^<LUj19M~Zk57Rw{EPU@6H@}`I=UFmMep^`Up9H=!lnb
z>P|J08aYUR*w3#U%w$-kZ8b_@E@k`DQf1CX^v(23``F9y6wu97CWF4g!PG@o%q&Vp
z{b;ox3P>-V`d&BFUw4ihUkhdI(S819YZFaqW-;9HtN{$+okVj1)>M>m^DYs}VK_IJ
z!0=&+QPKhp(|++eotHLy|Jt8&Y13Xz7esx&e&p3hkGp~6*ua-r@rd_u`&e3utvp%%
z<uAQH-z8BB|5eS-*=_dd&P7AO_FpwN?C)EhkIm-8%zzJ`yIm|?311GUNc*m>VEm-?
zjN|FCUJG8IIyN}(Y41L2^3b!5dD|HrY=$00^pHFtoPrPgBC0*&v9)Uh`4)*N?lPi{
zfr&g1u^{*W&irRk_;HU0<p;fYC5H`hZN_iyZJ=+Lz>!<C*s8JLc#Spu_ee4XK%_<g
zsx#b{ZdL)^Mkz8uyff`FP1~HvKu!|?19<s3&OLN`loIz-iETj0>}vy47|FU3Vnt@$
zy6?{Gi}pAibv#5p`rvuI{_Tc{Z5Z=DPgFl0Wb2&7XpFg%Wx;t8NYEI`KeqiBNxBfU
zNh0R=?mYIR7l!!R$3(aFE#_x7o5*1T8WH++MF|;!f00bQR-S|dh9>=Gg9Wz7a83Hu
z+7Bd#xneYGU5Hs&eq*zzK4ThS%QUZeF<B>`p&tNp2>p)FI>v6<)V)aJVb)7uP#k+V
zKU@o^{@#B#({`+Wb2+Xbswr51Z(*Txv3J>EemM3~smC`;QIEIzE7rc2Wks!9J?@jn
z5D(!>e;(nr3K;ew)^rQ<Yy2vyXr~UeEYI)C&*-)n0&U?RbU7BiO&k1q!ZtOx5(6;4
zQlC`bEVdn<6hWGdjIeYCodifyPGxP9b`r#J9$t?maiap=bd2*RM3E;vG0zSY<^Q=p
z9#DQ0s;CZcQ#o7%N{8U978<hapnewsx1Cw+P=-FMO!rcOp@(y2^Dbh>a#6ohpTCVn
za=6!pxqY1*LqWe3tjo-fsoWKU_{rL~G4Z*4a@NV%+v5LzxEWJY2(GzYT~WYf<tbY_
znXWlFws!@61Kn3gi~kMBjzJ$d0CVH9G~OE8F&s*<_BAlmFSosufE{{lj&FuJY4h+O
zWsA#oZ>KfmqfM3ecC#~yYY$Fa-WbeH4y4mFsg<1x6N8<Of{WAIamJ=>!Zq{7$n|n3
z4kez^Jlh{_z!gsnDG96u<+W-(C1%ITr~f2Q?xW*Bt=E~~*PZltk)i}VI6HX?sp4CA
zZHo1Y^--|kFwQr1vsP6CAcwX4qw|t}{Q_j9=#MAov8fGFPUh&-Xbmk%SOKRE42g0O
zKC9b`UtN90doSq`xwf{Rik?`IV1&Wy&-+v1>@s-9|6Q+?RHUWZjN-HVeH`=+!VnJi
z2v(%oAq*bQMk8ZmA?Z+QT0HGez{E;RuKsgQDf)ZwS+Na5I8USOcXM;1x`#TdB4m1t
z@nZ`TJXeu}tAdoGQ!5&pF#yY;ex_@RpHa3py`{xdJK)?SlZ}pJbzo!%tP?+~@*T3O
z%#<2&53}NB%^%u~vpu*6QXn;i$81B@fT4Tsy~P#S7Tt$~V9o1Bgz`RmLFHSFy!V#n
z0NZeI|Ds{`XmK4;Z0q)}w6brhXN5Y1V;NgEQ<|l!wWAL{6^z3ZM#}Lnu}rE5EjW06
zVQE@?;p2Fqy5`0)WO_7A?}n9;bruC^aMI|{kE{38Bvr5Oyl~g14lsl$39#^sg`0q8
zPXNd-^4Ne<k&*+n6daaZm+9`Xf>cF3-)M=ZrM=MiW@vPvH#`;63af-^wHN;A*M9?C
zC;^euBQz+G8olUFM71{gK-lsMZdD4dzUeD<nuqvLQjHdrj%hf6{A*W#1Pl-+1iI^s
z;s6YgAFr8&j+Im`?@uW$4PfvB_X@CLI8P*lpvQJe(Qu4N<{9CCvV9@_)aV14k#ukm
zs2y*XegV_pSvi$e)UttqkW1OS2Q;e9&;dgH!+(1t*O{f6b9-r9w)Uku^)+?vUS0+J
z6+dmXW#flxN2ZyF2=|PNJ+`(+{}$E7G7p*R8@Q8sd3soa`f-AaH3Xb4mJ&7xZYxz~
zzfF(a^Lmu>&b+^^!+#E=ix5DQ5Ij5mnSrT7HjkI8?ISqdihthXC+S)H=lFT%kycC7
zQDqc4MaL}%a1#btV^E02?m>bX1J~WH>uh#i;Va5pKLFv9WBWSj2mTeQN(w-ZBC)Di
z^EW~WoxNYG;v8lyM$5Ct21a^ff?z-$YW8iye*`<h=`Y1a2|Fv0+>fsjzP|<pt_BpQ
zzu?}@!Ct{Vlar}gIkI_2>rxIOS}$dV%h*M@G|z%6Hsf8<-QtS`$jJZ^SR}aZ1qmX6
z;!h%QA9sQ?YdR|6nYdS$?Gd<Mk1+52!tW%4&-V?y!>0?G@D7}N)`?sE!>*PPRU~Ma
zx#7aKg4-(vqEw8dn7u2~Tl^iW=G&}O&p0u+dG<>|ci7(5@&J!zGMLJ#p{3c}=7QVn
zX$d{Qx8#29^f*XUz-FPaz{-3@U)NG3zmHh(Nl`s`<KL32;q931rGqrj`{31OSGq3|
zL?J^!C7E8WX79y&xd?^y8WWXtRPC<pFScLASY=b07`v^F?H@Odprj7l4s^ltp}r%y
z&t=_l_BeQ9^h8^-UUpAmPzV!!njg+iXHm8g0G#xawigx`5m}y^!VsMP3it3c)CTy5
zwDyoTzmyc=pd~UCUwenD$XMe0l{7?g9MDSvh3eaBe+)#l26`c-X|LkFF_lY^BCU1W
zkT38FZ&s6MDqa(FX0CLi{Nq%)G*R_PFaGFO_d9j2UT5W1N86miv}k?<$8S%C3pQ_g
zCfa2E<=H+dQ4zH6tcw%n`{Cs$v$Zy-KT&#?KS?F1fs2I(PxXOnnWJ=i#8?*=D{7rM
zz!B%-=VgrX+TE3~-h&T4B-#`u#EfIqNd!PLFX%QLn&3o7;L=O~ym#3C!_gB`dz03{
z?BfG1V>-U5&A7%H5pVU+S_`+aSNo-2iFpAaprO_8&`9<@QP*ly_wBB4IoE4lmV;M_
z?@bOGH?Fyi*2ndn{y`;+)$e{PH~i}dy`6Sb6~ehQ25#JlCTwi@yiQ%~D^F%+z(K1-
zcyi$V=rhSoO6w9znvc~!5++I@=ppZgCMnzeF7%5Wo<qL_&d2|klU}JBY@E}EE;TZ)
zRFRuLkQx1oL)3mr7!ZsVmUw}YLPWxAaaAk#6f>1e-Xb?S9YS6dD(m_POz^8a|3Krw
zOCi-qUT=A=d=EhJqGI&E?d;oCrb<{bOOb=BSVY5Vus~*!Jfm)%Pi&=V44Ws}><}=?
zt+>(VXNb_P!BkRdjdl;r3qISiKp?=lbTBY3y~hn8I{k!kh!s$S{n8B=fY7l}25ono
z4{iavmk)nqi`(>7^<_6_<mA~w2ezISLvf)A$7|rBSCrh9q^s@ZUTVYkx(#LBo4pE<
zj?J-lU1HLczk0{>l#>d7rb+f>2sz%)%|11Sp%@=RL3iQI4&J&E&fu%<2m#{BfUr)9
z@F)4X6EO)vPj&*fx-JVy0lGTYSZ8kZScm1bWu^Od$W4E;U`Hf3C0<(MZ<@p&F8d8*
zrkq4}>#FOD$r{&VIhkRC@Hz#CJhO<7v*`Bl^XFStri^Ns7R#ZMjKkqN$ZgCH<>c`4
z+7N6ipjYlmJlWgAxu*BI?zj``HmbE>YFYO8BoqYjhSk2+me-Ml$udy)kBJ#6v+nH^
zqQ4{7+MLF-?%b-Q#gd8xGCxDzP?Kq8%I=WXA~N2zmK0EpA33rgd^>B@FZ@X<#Xl?e
zFJos}Q#HU1Dsdva6hdxH{(WEJWdJ-oi#O)vCJ++}1bb>L?=j^)R}`&CDYR7Un=I5J
z;BOe-yH}}(0A6Wf(_H6>u3CCNi((z{A6DW9>~DczwM+%sS+a2QkwK}SIsUBfrk2<n
zoOU4s>luzl>nZ_$I4_+0a8q=lg+y*KO)jNEDP{`k!5-Tc!rhHYaroQ8kx;84b^vcx
zz^(_mv{@ZuOFlgWnCW|$+dNub+`5GEV!O<=;pvVJ4EMZ+Ep!)De@S}}veaK8#RW3F
zE1Tns;T7Fh48(hu(?(2z-CEz~f}4g9eQ0+^%NFjI8}KsJ3I3!<Zle%u5le<M3e2&V
zXXgz>tv{M#!&K=FqmT$cnRbCEB;!DXz5dGaL+$1UL@Ez@0;Ks1Uqvbd-F^G=s9ht=
zIHQpQ%%~6EpcS=J<I)`@ZNErF5hguHiQ!Yu`Z`02O>3r<!onb8I_fk;5tp$R9M#f$
zW*)N!m(hD+pp+Tb5S>rNPvCcWw?|0O>Vbo^fwr*slHf?`*(~!m-ha_{F~o(Tpf*AP
zMG+~qqi2Wpu21(BB}@Dt)@Y+*00P-1*5bd*vYYb1q@zF2q;(aH(@zADR>c8;zG9)(
z8e#B3ShVcmWCR!wAv>{yb2%}kswMupNb2BDs7?DL?s)mO{zH6#366OB`8Mg&M@Zqb
z(yi$5Sy54~hWauI{InlIt3_(~!)zt4s}ZtOjwU&@LHN4F^L?b`oF(k4w;5>ndi7sV
zs|t4B(v^SvE)MxRRbp7OfUnJ8*JzOs@c8{?W=rm9wglAu$M_xr6!fS*>u`NNq8}au
zC9))nX1+3jo`QCzbrPiz!yFb@^ze>G5Omr{RTXCgtx2A-$C(KoVb}$D8g@qL^QRga
zCRRIjRx3fv5X%mxBVdFlH}wXR@4?Ua%6X-x-{a3rHiQb@Rqz;)ai+#Dw3w@Xa{>Q}
zo4VUZ9hcAe@MtSNuz(N{Mwk^)N4+41i*Jm3_Z<`6(xal|vj54us7+lpzF==w#fj0|
zrrB!Y8S4luQt|X&1DMrVM%LUrneKF@4G6zs4qK0}1~?{hA*=yC^j;PdCd7=KzKhk8
zUTsb6b<N_|6wXAMg@~B84*=T!J#*Gxj(prA3yed1c-P_4;L+<2YF?csE~~D1FiUt^
zRsD$?(2}u^dmgtMG-(~KxqS-I@22Iga!@1E-J&^36n`171H2@onGjOCzFyv{YfZ72
zq9x%gMSoiL$CnFysPx<tc;;qGIlHc{`f-@?KGf&nSj=+Qd1JVc>xYH86j8JwPGsuh
z3m?%CrpCreN>pB4SV%)a0R&9Ng$hW%e$)LI%g-+MRf6Evwzql?J|s)LL4l~qZ9Xpv
z3me`|tjJNPZD61qc6XK@us^0~53$ItwCL#;=!?w)K+seEAxFk)_bF4+p3@PbgDIGb
zlb!VS33t0b4Qk5G*Ds!plv8-sE%~KiOmFLXNseq?e>~0YI6OU^lw>71$>X^?4xd3x
zQ^Hegr}f??QsK|d;%QHBsN(}iO|fG=X*6pO8;dRP@bQ9f@auo7WH=91#nU~Vnr#_O
zkr#AA5Z}as{Q6UAL1sy_t^Sz8e5(O!yKw=phaN)?_kUawt83XCo*wuX`>oSiVi;O2
z(i0`cxK^;Y9S(&<phGKoSlHs8PI6{mgjxp1MhNWnHO@C~Uez#=g>`4snU!9w;&7#@
z@h7SfT}6QVJQ39F(zI}ttO95|2)dnUOU+)NKzIABTPUDFNCYla%Sa>Uhv?s;NI83n
zOJQ7z>(q`137E9}D~<Wyc(IMd)$TKd>pR(_+j6w|=tQzA7(spy6RufAsHOm;joa{9
zE^N!GL1(^&>Mli4KT4aZ`Y0`K&YEP*M&Iv+n)f+;nu(OvUTb*Y^AEQFdadrhUArZU
znW_K|1-r!@N>uoW`Ll2ygP?Nz0E#Z?5i?!?08tXgdfsA@1<+o5R30@XYgVVVKfE>0
zi%45Vl<v)IPj+#bm=+#5S)5h%O|~^TB`PNy3|}n&TJ@DiRr5rtEq9`Lw9?${S0sOX
z>5o;f4=BG&f(zUo5QVL*0$QvO&wzP`mr>5Fzb|3KakZ@wTARHuGu*1Y#1pU8r}KaJ
z(gag)R37VjLSJ@UJgev7<x%yJ6D@+9lbsNqty_!jpjE50bH){1t^^jQ@1%B3+1np+
zJrtX$UnGFICD+NcW9QJGI5|~-z6(9oSJTMkvBALQxk_@=PnKhyp8i*)_pTBtJ^-nx
z_2U~ocKJNc(SHa|#ivJE<fqqnsHthq27jhX-W03$8gCfb<V+XCR|ySh{4{<wTxMI1
z+xu7jK?AwEwiM@u+8wpG4BfY#{?^X-EWV`Bx`<-`V3FCjUSF=geq|=ma0F(emU2GG
zellfPEOAQq`|63pSxok>Dq}}T1an~lsjR9X$CxWASv%`+iw9O;aQ}R;{MJ!N)0OG&
zx$IlD$@e;A?cXK?Rm%>OLIIzB;g*L#3uUmj`_2yZ^or!g<ErUlEXa1jz8rrMS~#p}
z%y~=qc}aTmdxAYe=I2zMPp^{l-ZV{;P-s^fOIL|oQWyF^&eS_Oo6jn$7D^RGNliTW
z6*-L-@;L3KWM}E-KV5ndLOt_&U*wy}(nUts)oW&y6XpIqFYHe3vX1mckdWA_(ZFi|
zo6fUxDxs(4cj1Ce)el2G78704sc=POwJbN=UsWK~wM{|}_>(XA;zEVUYH54h-7`Kr
z9i+J&Ks*NRV`smNw7W}8jttx%xO+J14mR-?6Q0yr)hD<n#vzJ$V~T&@ua{qb{HEM_
z-q<Vv@*)2`8DG2fr{u!*v92uUP%n%jG?L<5<ZC)5#EXlFM?cLKwLQMay-D+89|~V{
zwAox#W3>r6YxbpC?cF(v{ck?lZ#8nr`cSp#wOt)6T?}-umEVyx^JopveK*+*F>)Aq
zkPYA&pKhGx`vRjIXtM5^I_Vdku>y`c9=H8ew4=?P^8R{(=2izlkec?t=>7Ndj;nPK
zW=+68u&_1_LsMP0Y=)~yO0&(fx@nk>^wUzLAJI29_tll+%wP`Y@%g6ZRL9lln4=z9
zE-j>9UXPlSTC#ac)05DD#=ToXJHCnX6zh0N#NBC4!~+*x20k$cRTB~<yeB>BoK_1P
zjgvuKdn$O(XGIY|yo{>fs$Cr~MC<5lEnkM|N<$`0mYLd*$P0=~^f>-RayZ4ONM-+^
zb7oi0443}z=<pIgWP~+yCfcqw&?4HQk+uPW+WJFrvNWMjc#=mBCKdE1Xv>czZ5;lo
zsI?jkhS2~)p#P_%T=y-JTu`%|QQYN!jDb1|HJn(I>*T%tp#$A^R{aN?26M4N509;f
zp|>kw)=E(_liRwx-{$f4s@Bp_vu&G#mE@|Uh=aBsm6MAe)2LWJjCIK!a>zI<OJHmI
zvu4W?Z+nP`{*5N<)}K&z3gx$~Ah5lO`(M6%ZVkgc2T22+Ar5#N!+_x-%ZEMAr;qh~
zp2h=!Hk*i(tlYOqKEXw2jx1q|+dta>KNo;F-j{l!Xhc7fT=}iZjPfVEzfa*DQ!46)
z?)C)Rkh_sKRGrqi_rVj9eIP_7J`&co|C}%q`{;6GjB@j=9J}Vsf5hO0D7=@0cA4x-
zo)ns<<XR1tH2;t)97?o)Ul=z_R3fL@*Um?n{9ThIO`3rkGSo>b+q}ft29^bFzafXU
zS;^9ymwr(dR3fIZ>x!MXO!$YQ->*Mwr}D|5m!Js6cf>93wGaje>R)oObSb7@VeYp8
zUm2F82o=@=l{ob*g+&_n9xldbJn>VoyKV#zdIW4@vzTjTIiPJweK<kv=tgRHZm^)t
zgW=29cL76jHGnJ;S(Z~h?~qfcmV<Omz603mTEs>3|2;K-vD`l54AVv%_vI(!Sj`#f
z39a_t{rX3=<PG=OmO5LDc$8;*hmo|H&G<u4#lf@9)6SY9-OIz5*K-WK=hWn-IX9S<
zp@er{TN+}JNoK-_-jf??sC+rW>OQ?>CXzqaa`fC&YQj5mBTCGBO+Vj;8n<MT7fU_m
zEG7FHu))}V_bHh*!fr|7#7mN5$-h~O%EZP4ZHXVL@nq&*xQyt(Un<HUV-xNC?0Fhh
ztIK#f4`@k0c6i%1Z^<&>meDM37Dz5?aAK5R<*L}SGv)F<{6-(QI5Atyl|VYE$fX*3
z6j3}H#<lXsD0Q;T>ub?Rufy7}<^26lHO(IDX>*`)sw2-vhsSs+QOEO2%ViK&<!-A^
zs5kG7Ihc-|Aw$E;<hhl{xq9O#PchrN-P<f%%vxv=X&?+wN~qTI5x;}5S<l~pyZZ~@
zLYeRYEqSESj{%26u)s5TR)RI0o8o$)>DmM@S^~b{bl!Y2m@#>?4z{wT6>e2Sa7(1l
z!=VvLOZ+yFXEB+XG$e8NVi@%OwV2vhJSimcOk2Ej2M6#ID1sYl-UZ*~yTH&LUG@ez
zr{_hlDK@2ez6^CHLAxwPI`_~0XZOYrAAp+LQrxmNC%)yi<nEC?Rk~A9V3nbUid$Yv
z#m`0?>Ee{z&0fwdCk(WlC&;0`URONvjThdJc!Xb{KE$!F3iXh%+CV>Un&5dZnw>Uo
zlXN%RG@<;C&i--`X%}cr`^xXN0EU{Rfin7s>}@K+-&9DWYghcb;A2jKKnbdJW>pau
zlE_8I5dmCif(V^q-+pAqUb4;OP1Uy`VpbZ$W?9au1>|4j_I<HIf*Rbj4)eZWG1}OU
zO0+|k$L66k%Ynlq*+o_Faouo=mOx^WR^O>I|CEMSy~6vCIjs0$$s%iyXvU>H)xJ;^
zh(ffq<w-jq;EvDqwh7>FdrsQTOUrX1hmM_p=;?ByuMHw*?xG)x(Z{7KQFV9iAmI$0
zA0oBKt4$G_sB@EIlsJcE;rniQ(DL;6^r{}WV)!qZjtffUH3e$2$xUv<?)B50HEGR9
z#U)N%qQ`PVA>_ZtnU+sYS!-dK16-o0qg<jGYI7Y_Snt+MAU15B<N82UnzWG1!!Rw$
z(M>|Nv37k3OdPB->DzC_B4(Py6HRskWy>tZIZVvR0zQ|>Pb3$!O0s~Pm9R6mSY(F~
zPwx)=$(zN-%S5-3h3DUNS_#KXtkvp`Ik$EAKsD6roHSSkJbA@&>SG^u8cL7HXI^iw
zS0|pwSI@7&DzvV#N)c&19;N|af-MG;2dYx7Iy6q=G?xr;>4+!I4?{06#r9(E_CB2T
z`IQB!ceUo#Nnh%1dHGtTWJFCp_<Wit&a=Z!%hq#GX)y7GN!?ee*>oySHJ4sL5l^$K
zrQICO-&Er3_G5zR7H>t3l8HT|&x|c`);Ke3?54lMR?#P~o*^oIFF^qwR=-14_9vhG
z(3~INX#?usbeZner7X@i5T>_Xdf+7zv!Cxu)?D1;w)x=pGOj8-dbT|NBdl(#Th7xJ
zoOI?yxz(AU9{17U3uj8ZH{)ZzA<`PUT8(vbS?}t3isF_kbS5MsiB&-c`cV*d@~y4x
z22YQLnmm;6Av~q%-afBKI}N-`w}H4sP0ByoRQ~Fu$0TiEkN4786OhF%_*wJbzg|J{
zf(FjT3`~4$_q-@a#CwJUWH3(-kbEU6&zloxIL<-X#J_IKA&d5vN1SJS(HesHAyN?!
zMXQ{~<u6}9gGr!yu`iR*=ISA?_CT}VE1l)m@?ex+#=ml(UyP^^7#vcYi3TA+bT<Xu
zJ+892y~`A?;48HEKfJ<xj%+#rE&;GKK{tFR3xPj6);2%cRKKy5{gE-zoQh>=1IE##
z{)jKMmOAp0YBrbAwhP+S&Bipk5V;x6Z0i%X1?z*im}GJlRyNDm=lvI}>p(*$$;pbx
zJ0Cqy=L2v?l|e}Q1w7t)XA|ZkxX85Uw{_LoJt`jov*pcgT{<GpdrYxncWKKPoc9b@
zeVLSw2{`60ekxMhL(tmab(Fu>=Oq?&<61;}p#ugVQhTd&mw#T_MSV4gCbTeqaM4!n
zLFFx|E;SHOMhbVL1@|UJPVub`|J%^C^P;H0K($b-9jVY6ttZ%1PQ7~hV3mut`R<AZ
z??>d)pdp74t6FiGgQk8}raNS2GpT9{XTQ1bP)75Ju-VinYs<h7cxkx)P5CDeIRg`s
z)-30Va~7GkM153?mQvbH<GaJDf)oZ?-|w4CMW2pInn}6|i<*XA-5f5z9ndf8v6u-h
zzz`9>y>3x+SdK@sJ$j(jY2Y#-iGN4-*6#tvG{EvwTrq}|^79?&a8AQ>#ci6?%*ycr
z+iq+0VwCJppQP>0-y0xx@WXkESw0qPS-$<v^Xj3l9^yisdvr$YAKu@O2z6<MLDBl!
zcM;{!P!`hvX)&prNQ=f`(}hS)9z$Yz%VS^d)yzC^9OWwQ<xC=3V|0*~Zz}dDh`|xG
z_AqdpDnc|dN5yfP)Y)FsRh97y2u3<JT0t;<CW28GWR5{?pD$}$PDFcQcHNk7e?0A$
zMP!y$H5q*S^O~EkRCv(oo+Ui8CO>*F+p+xfHNX4p_a*hek8?W|Dv2S(Yfs_VYhdHy
z*J8gg3i>1ABz;P@ba`znT{>zqlG74(F_7n-9h@I18};GysxSx~UdC`aUPG2Cj?|_2
zbuC_Esi*+wfb9$*N9wL)X;t(qMhOF+rt~sKX)orX4$u36@aj8UDfCO*F{;REi{r_x
zR>au`hT`Yb<VJWUsqfxrxT84Bueb<emD11KlV<Lg&5!OSa;N%UwcvUWdm(msf!le&
znBI$rMOEQ5cMAlhMGqEh0?i+j(==i6`d+cWdb|sCOrF~nsXq=j@o})4_>g^`gpL<;
zHCR?E{VeUmB;1%&kkYIES5EmY2BhE18g(mkxcOj2H7?8^`I-TvfC7*#YWoyPojBw0
z%iv=Wc*$3AS~b>`AfzYyyly(hO7IPn&K}67w=r!fAw8KDTM{G&EVrv+&uMhFe{(dv
zyD;FgtfH9p-%XV9w%$58OTNYLmb4_k-U8kZ*W4*1XapGr^&Zm6vm|BYjN5=gj}%Q$
zxQM6<^81i?(sx0vblo_`ReW*buUTBc&w(ui&5b#v*KbQNp8+iaKGaooo2N<zr|GzT
zOq{6@Vxno5WuoK8HR}hh>u#E%vdqLOQXf7}duB>{VgD1+uGD(ZM^3wftj15B;i&e;
z?}-wj@sz`wG}@Guv0@;mc*5HjTzLxH%4>@%VF@a!@+5iqw6f9G0&4xA%v723^qC{~
zO7NcxzecV0lfE9%5XZpEEq+-xnk8GRFe@i9VpHaQ?b@vwxA0QBMeZL$AC%!d1O4`&
z8qOO~J$%mYf57Rxu670pfFE|&&hzEc743OFFKng0{>=yseCKDV7!>72Ete{AUWVBh
zzJ#_q4VSbDQN^EyTGoEbE#lv=yw=+PY_rOLyUvO8-nuoj+-+Sw>byH>7<L?(@6SlD
z`Bdpq|0O=Th(L*?-Mp_Y<?35ks`R)WA3vuKuoPCgxmmbMRS8x-37i$-wuK7zFg!9s
z9vMjWn{4r@H0G}nbAh*%|2G!YH@H{IwEiHI$aedFE3xhJmqU_ig<m5sD&|h*n+yji
zH9G_UIlHy9QS7BBRYbw<KkLyFhP3o)w~(s2%lMGdhtBpX01hvBWf1$%%w5*CWt|Tf
zaMqR6?11Ct_uHn`g=p3Y6$<DorB&+RLWe`mbgc^({QZUSvV#f|osZ_KjR|*qF`{m-
zY*P#AX+<ll)eeu0n^C9P#pbU}4Yfz#zEw2nLy)47s;i<_{zu*yX+BH_?xTDawZt@L
zh5~40WR+(Ao0Zerw{b)803JsyMedX*6JAtjNkGjEvyI^LAlYQxhYr2QDWns&=Oob@
zwK^S*4OH7Zi*l*+GK{%gR0*1C3Xi|7j~{wqxz{~TmCGvQK{msBEaekKp#D1xo4MyI
zzm<@Nq41!28ceLWiFyRYT~B#i8}AZ<-@W*JGcG79eK_ia$g6EwlHu5?O%s{1<1Q~q
z7Kz+tV(8mt2;}Du&uy(E$bP~J$q`w{XR%T9`;Wv{1trwdgclo1+!SL*``j<)l?#59
z60Iqmv%8>pf5(tJWtn65-WwM#B{o>2^fE(bgf_QMnvkc|X$mEx=9(Pk@G-fAmYj^X
zp;SZDk=1*ai(iPwYXbK{;J#$;V%VS8uN?WXx86uSdT8<Y;<=mW*wQ-Kr#on%WM40_
zq0X3_^8HyWHY_NHG{%@nT2eS7{c(#F3pAWnTSt-^ma_YDl%QcTSp|x$^YIw|<~d%2
zyj(R8R>DMWWcQZSV4afx#1^%%_1ZDetw_qx!gG4>kMF<TYRM+qgJJuCtj|W5ew$=d
zPqENhgn7(HZG|$FNBcN_NmRvF+ol4a)^`#rsAW(1URB+yAzBNwNJjGm;@OZa-ERO%
zW;sWQ7{+KLh1|sFXIIk<k)Ic?Mf=hmm$hPhRN}MNGHf3th+(+*yMMtLym+tAKB;6V
zlB`W4_#@A<)&`SRnWj^_?z~P<#lU=SUEM<P6F~!u@A-0T9vSU&+x*~Xq?Zs5d}{n;
zU1LRmL)FO*%cpfxImIJiODc)oF#;=X57eYQv}rC8bEnd2a>J=}_3~pD@?+jXM2INy
z0G@0nAeBCtJmV(mR{Ud4dR60V!zoLjjkud~T%VDVEB|x7^^*zT-Eq6s)=Cv@rey!_
zdE|<zeu{$(*^jFx5n8#uI$?E_*KaqRCBtwqlPp+MIjz)F*ILw@@&+RCj(J8ov(aj?
zBE0k8Q@EPWL2EFWFYfBpfZ8`N6}7?MiP|+?MfA4s|H=Eo-g|q&RVa(IpY&2%Tw7aV
ze0cTeTh&m?<Ry!Z47oW!3)2WcTz`3JK&;a8S+Wrv5P+Zm<R=8Gg&3k+eGl-d1+%o_
zeN;uFJ?BAvLwNmVdw;07x@^B3>x@Z$R4zFKc}w3b?5T1shQmqe@oZ;<_2&n*gxyEI
z3bM?xc>QbiDOEMI_ueeE71}oY*bDudQ5^?uPDIs`UFnVt^9PS6nIE;T*vN1~66+jH
ziE}eB-?I8gD!m#W8^QJv1cuX(3c6SS)z(Dx88NG<tLD7c#>U1r|9_`j05SuXyl?h8
z2)QbLvr%%iA<ke!#SmPQ!*klOtD$45R=aTHM=iKB%E4D~Oz@u$Jq6mQ#~9GDd;M?K
z<Y_K#e{oz!!CM7zRP5QZulkCu8EERgIj=oBrr2&SM>EsC9^Df8iR{aq6qr;<-^}Lz
z%TG?%x_HGrjq_tfLLt32;Ego(wc#U`N^&&E{)y{Zw!*U(UhuqBWBeFZ{Tc!oNZ%D4
z*Vet?>H<v(pU-dg1_y_5eLj~rNy4W*9*&q2R{nIHpVcl+MO|d|7sJqrjrWU&q3`<2
z@%|@kIj*MMVX1y{7aREV{13B#&7^>+B?AJEL<@89sF7hR-~z19eX<$hO7*ZO1j^wH
z=Zod4V8SVzi@t&H49Oo~Kg3HLkp5gFE&Bo&25m&{ff1Io?@5BlOk5e$^mBVVo9Z^)
zjc)5o_WgxFr$820)+>*W*UCoYlT6e-oWk%!b{wL|5()+TcV5!Tyj~%Ef6Q}J?%4_v
zpUBIyoB?Bgk>txCzsWv#{;`_>qRdI{GdodQYr@HRtuK32%6AQN0TjzMQO4=ZF;gRe
zdc)JvsixxaWyz{y(e<a=-@rm>bjJ@KTdk)7c`b|Ba4XTLK`jP9n;)$4grafrz9UXB
zmR`b*F`{F4iN6cM(6bpWFR|wpTz`BMt9B%ZU(^S2u4ViO(A1HC^wt>aCNSkC4_>X^
z&an=t&*78}?N6tXw#THj^Kd3QEsyWgv|R4D2LaKrCcTuk6t)!nDaht&57jF$VninK
z3iQO4Uspw-N*jobVM;r6I~SZOLlad*R@jE#X@DHYyC3~oTC(>siWKcufb+|5vhz9N
zK$VgV>W|1-nV!7sL9c1)1{F0G{WC3LH06hV%)ruD0GJ4Nn_WPEaf6ZRO0*kc(o;0#
z<N(hN)uJ>X&`$1F|3&d!CIhbD__U-)*leF>Kt>xAhJ}L)-i$vU*D`x{7C_0JB8P+4
z;_$X;nwPe}7qYe&SRtWh8iJxQNGQ^Kem)$<?am;Po!`^>!BMV0ui1#ls2Wi9+)Quu
zj=1P&@1#Swsa8DDa`<Qna(h>cJ+08Tmws17&{BI9XH3VfnHMG@+$^GNW!^{k4hjBC
zMy1hnj_dDFc<-?JsH+%Ga+a5Wg{$Ec&~RHT84Cm(*nXz2Iy?bwM%QFS50aVZ{@*d<
zZb-b<!TPJGfT0GyXF_(v^j<kx2$i~FJ@VMTMaZ3cp3)KwVPn{zuv?fzN5}pvjuN8P
z`kdgI{rW249sLoAQEwyqKB<ZKwB0P)=v`K$n~mO>m)O4S&zq+%F40tW-s-Yu9XhVt
z_-AOJ+?Vc8nzOq3G1>#;d;Qx2grjp~-kG*shhWje(&llGQ+va)26s6p0mxmStOaz^
z7eV*X^tp+Nf$z&JMKcKNmJ!N*D&xaTLbI^mw(lnvJX;PdhHOCVc$2b;WaIY6yB_|A
zKX-}-u02h9h=(joc=4wpdJE3ahl<v>s-L|JjR~8C6wywou38Da95+d$h60=&w=UgX
zh_X0|!Shx7W|oMY%s{H0Ri_Y}C`G-YA78q`@rJN>uA>L(Yn_sWFu~hkP)|g0y;sHR
zc1L&wg{sITYc<Wc$b)W>@dw^L%MqUYYxextA}mf35+r5@FVV=@eT*Bm#gh{F$uUt+
zN8n~$tbuZoWxXRM4_bzwKG5SJpc-&`8(w$kN}g4#zGSn3G8Ff#CFVGKgk~%B6(QXj
zxB^zW8?!k-r=vF(e`*H34?|nePro&CKV4E+uVrG98*jF>5z`0t%H+BQa-@U#BHtSs
zTXf@p{)XOiBCgf;U10-F?_pZJDpmO1Yj~9MLqxGEG-c)Lfqzm*(Un05p5JXZNFwNm
znkL%91-)OA7KWqd(Y5mzli0u=>}#$5)V}c0#Mb)iL<%n8)k7mIPGqQ|W4V&WN}Qmj
zXw4RfjLA3xn!v?HkdZ{%Rqip}=gky-t@P{+kcA2|I={_KfnpD!@7}Ym9A^~tK(KPD
z)9mxoC_85g)zi3q`neLrL5lw@L2C=5lyg!wsmkiNuw4FKWs8iUUY;AO&xRe6L8}q&
zjrr+tW!55ZyBUIFUj_={eH*%vlx&-sk2bFH9-s`H>gOJ&ELzKLEyE(?981_<%!H-X
zUtdAWqmQsU@WoUvs^RodS82VYwe};@c|3aczCOPJmI2<Z0zW(_X$Q|6P&MRtx*I+Q
z%Kij>Z1nM@`LBlKK-w$#xqjMO!V=JM>_e$m*N#>y8n_RT_RaZ`_3_HTZcWSaFX;PL
z$gVbmd>*LPtBO>YSIgeB$5)xVav?jO!}0&4>MEnE>bmtIL`mt8l9Wd2MoPN7rMtV4
z?(PzhhC?5^1Ze>Y>2B#fG~CVW`+fJ`HHHiZKOB4SHP?)1KF?g3e3FDb>C0b>P>F5O
z9@T8Uv|Te5`u1AAdw)uX)pvgTU7rIev-|Hn9|?K7c^zsWsu82;=;&~#kzXX*Y#<3i
zCAB67r{IO+xpfYRB|1V-zpSd_&ecS2x=*OSk)(PazB#T|0P*!8L(6=A&&=*IKw;7W
zTE$@S9{UsW@eFx894JffOY!I)!WspD_nMcm*)4G<dT~<Wwt$nFLpu=Do86M(9&cs6
zJbpT#v?apZa^~4Fjp$}DcUfqoJ9eVVdYQ=Kd}ZCaRG*h<(Zb<8mm^P@SZ}%C$&%%|
zQaE37XV{spm-)5pqcxI#Hn-h&;q5CuvGH>=uQRj=*CX?R>$HPJ72ITL4M-Kif|V6+
z|9eN_x211++#VAMoCb+|B)TaH>HdKwfVaU|`*e)0lb}Ge^hZI877RgA(>Y#ZAL|MW
zn|{+d@-3Om#t#u@6(}E7g%IW}kY?b<$9r$cGs21xbTIU_TehNp%2UszNgDfA!t-4T
zxMIf2@R@Yyh92{TfBam-pZasq=PDW{BX_`uA$3RDA}+TV{6?zZW^#IdB1nQVfrxZq
zEJnM&&YCwvguSR%VmlBf8S25#&YqT$p#{~6iYfMGO13w5Q-&*}r?+35@GoUZK$qyw
z&)Acak&!7<!3DAaZ6V6LjjF6?BLvp`5KtN_A3tdL>CFH7lcZ0|JQ0a^eb@fy8Za&+
z?{oBVE#P()k4df9FmajQWJHNI8b}7z9bBIxdG%2|lar<0I7o$c-$tE^VtbXds#}Ih
zQv_1)#&bBav@?%!IF<Vd;C%Hi(zb0Qv6$HnVZdx*m=Tt1AVh&b!6WXp>0>}b2n0$K
zPMFvro)l{{J87&@*D*5OXw$_Ui640VJ4nGjT))OGyZZ;co%3E~W5Kdd*@&pB1{1^%
zG-P5OX^7drh$mvz{&#Q%2Gty?#%s{EJcADJ%eECXdOLipQo0D*La5jGiL{61=rD6e
z;5wG%sL_-4G@&E!5Adpc+WKUjbYUOghP7JSzCVvvoeLx^#T}Y9l@q{`*!sm7OZq|C
zue^CHM}D7q$(`R9v#iBtp7PaDVQ^N3y0)54A6>$we$y{2;q1>e#(nJbjg17H#0T#e
z1nd2(5EEG9AL@+xffi?}i6zCmwb%-y@=)s+<6kPF#x&G-GzQ8oNVI?VA)m2gj@S9d
zWD8ZQl&+L%LaU$ZZLP9hHlD|%0Wklso&=acLPb%8OHZ<RW9Y6qj9PN1hrhuS!GrT@
z0b10iCsyV|lI@KU{e`9N{dBgA?oB?wc4l*UdTB+n$S}VjYn=bV1cHD7VSPb)g9GHi
zLX-iQ^|#Q(86zHMA0=kwlV5eR(5VIh?K5Tm5-$r}ZC8gj9IyZ?VrawR_Y=tJ=HsBR
zN>XT|P;8L462I|PRZAI*6(M(H08MHc-G5%eGS+SOCFs6zF-d8H^f45a1An5MKBFbL
z7G>j%(dDF3A&lhS$v&O_Pr!!lA*UmV-a_rV3~v8$E=HLwRJGs=_h@@Yz^dwjfbF&C
zn)Y;#@@`vEIi)$FCxQ8t6%w<cr)oqOa*!tahc2|1-wzn7WW>VAsTT`R7uaKj-@%cg
zBRE4x#Otb9bXhpGt21UobYLwzN9y))c2-+|w}ZfME8P|mfdr(C1r@1(W+zltYjFO(
z<|@91dBkECZ8dp$<fIZpv7E0$vE;LNkv?Vt-pZ_{f{9m>vS_$^n-6P5$U{c}vvY_~
zbEaLlGmrCJrN_wYkL};c>vF%=rr_gBTyc8X-0+8|9DBSLA!MP^WPD1Ksnsc~g=p3N
z=J*ppj%aMHV0^3{v%n(9JV~tx`)zK3Z`4(vLsjeuDw1c<vROkG!;i+f>!7_ZMjmfW
z&)@pAGpf&`JSi|{!JAs@^Pii!){a5<F#h5I<-<DAd06a@&tJO$_Ro(R3L~jfvt%W|
zZ%MbPg;V9pN~p^ao<N`>!{u8pPDN3Pr)|@G-WzxFkeOkJdMussI+qN+&zwA>VeeM;
zc+IquNGpaokyhJ8l@_14Gv@KjcjIz)59F;0o-qU#CciCt7Z;qOVl$;_hn~#OZ~g6?
zeg<)+UqJQ6oCPw@O6p7YXInKCDHl^59FEyHA@|8*t`E-1BB^(wO7o*8<z;`Zg?zpY
zz<BTPpJP~1@G&eUN<ww9ZXYwOeK{V=;Jse+{grLXeZ&RzG4vcNM@t9$s{lf>Mi?1}
z6mEdAB};p$2C_TNY2@eN)Q?MaXJkon?ZDqFvqbk7hMkw8?=kUcmwJ!gEpFF_w?gFo
zrpnyIL-ER`E-`m2v8t1lG&`Mz@kWZ#GOQo#6#`sX@4C{FQBlv#yAx>tlbudP%r}0W
z(C5Cqc}b<_S=-0={(t^xPC^ytkrw<1p5Wg4DQu?xLUgV_7GirNepTqpY8yXS)aIuO
z`jRV!AO__2ylzu@4Gr@_oIMm+_(UXPT-d@cm)%WLjzo%8juP8jHKRU^uMm>vOlg`<
z=2W!crP!6E&zXMy9L+1MCCdA_*BnO0*8LsyU&m&W2T4J`5;}$ww-zatNp+KgvO%^`
zSjy%lv|PFo+?{8P-~%8APrk0?oXv4hMGCe=a==x6`{4Z{i}(+#5QdCueBa3h*Zi<!
zmH8i%H(ksRCsFAuE_n!zY{Ya&Xm*Z6$R|?ayg9Mp+!~lDWByROjMIz7>hx(<wRcyG
zJ~=8-iAXMGNTShx^YpfE6OS`vp%e`^N_wujC$(V!dD`1|wq@??;rQzgl=Okn!==D4
z<umWOoJS7ae3m$9iD8rauCId{6A{^t=Yr=m%Ia~1p!k$1JX60zLRxILO$U7(nriv~
z!v(1L57IDn1aMdO0E5<C=(7NuwH?oGLS5XAFk#}phwHBgi7S0^&h%#Nos<*p1yq=4
zMGz9$x12Zq^wg?&!PUi<c&n&SM%$4zF;g6FO3m?jhrU9i%)^`%es}EJCUEB+`uJvh
za1`n_UfOVrZ9jZ+;Kc^J1}ymTkU}vcnL>oJ@3i72C>3;7!*4VgsAv-C8d_SkfE4rV
zou+aCxd&sZspD0vt2U%@?8~OWKs`g7J;NWmfCTWlO4rZ38Hawse=R-#0<iRoSW^D1
ztGUzeLV;uGYYl6o85-RkkCyNf{0N;+p`_MhxySd-l|WD(Qg2DJa|w!KzBroK9=Xp8
zk$X($BrDB`A0k3DFy4DDw#As3CbcRI)WO~H95(ScG(*CrWBc<n0qr(kGVT-xF^h}2
z5~4;90C^%K#T)9f*3P_9ga1jqf8<<OHkvmj>%yEZ!v1~Q+3oqb220cY%D5@Cu&r$|
zNWKV`j-|6;Z}2FvzPQfl@9+YkuJ%n*S^>J1LO8jMq4hODDqs#~-md1sP+Y!de48A_
z35)`Jnr~I$Cp+)0Z_Ot&*<TncJNIN~9UWaF{V^CXbYQhO!Z}V&`E{e<GRvlM8-RXW
zeodN9$Hf)awa=-%lC*bTGXx;aFaqi4)W!3zl2W@*(<ilRRf;PLI=xiHB9SA=&l(xo
z4mTJlG2J+ZoElsMF2d}YnM&BpPoI+*^3^BImO&85*)k~cXSp-F(iWyzFdg%Ch4eFR
zV9Z%kHe~<94Iw8V%2u*0p80fS3cghL*b6I}s;Lc=`jkv3e-r_HdPbLMTJ0shczq3c
z=JyFm-*h^7QgKcBd(=mgIM5bs8FfFdtC2KQ`p4%=kfe5r2M}Py_D>*)rKk`}t~vva
zEUfo0vXSz1Wl_0){f;4Qna6mAAe-5O%%3V79M`#C^0a3!)y2-jV^Ee(>7TH1a@>d-
zy~)j_$w`Dl+0U#sTlFQYPw!uwAzwU1-8iAd?I&Z|!1ow#q%C~>MC-CrI%&q^-V$cs
zZqrgNL3{bBaH>bDw3Fe>Ql59k+WU|C`;T%%XU*}28{e@b;6<MkOn`XvLuT^DCzQnL
zuR(7oJ~7R_|Dm=b5k_GUMRh)bBo}~FigK11_etx!2?b0dym&3MZnjsAk2VC+V^qH_
zO4Qq-jZt%qlut6wI52xa2YX!OCs3$^hC5dP)NHUOrT~eK*zZNM;6KBz@{`vf^cdqs
zR~*1?ol5h4*<7xayCWSqU1M-f`@k`Fv4^E1&%dk&BrLWtCnU$82*VtgW6B2XMhM@`
zE2j{azhS9t)k^(@a60fq*&JCl9YDw^Gc=fF^4GO2E~CJFJhO9_eQ*iC!~)XK@$O`+
z*DH6BHc#f{b0`cH?4I6^y6E>!T~r_jI6}}*XjqTpR4|*t7YDw_D>Dg;J4e!R+wD8A
zf8HB$I-oooM1H_OkGt*4>%R}`t3Fu@d>RfRJuOlPp_plehMA94KPQVrAH^KifvjtH
zvF0V}m+HA%8N#v#oI}XG+d((~f}QwteQPqu4s_Ysf#=k=xr_>8W|@u)VyZ<=59wap
z+&I%@<|a{i>pv7xc(`HggP65x+K-4?%V=ScS}8RM@tPS;GC}-8IXlNJl)_39Z?#;`
zQ`_&;(CQ}nd0UO#WFGCg?f@2}!vJz*<sC^iqp<CHk>hv4(MfU9{b&P><S0M<ThG4_
zN!S`PE{fe6h-1n_kUn~+YFLB+`0eV(g#L@MHP9vD2{-e*Cu2=EnSqvNR2X35=(p^p
zdoN7fG>*5tbJ1aFCHC;<U@<G=c>hl->wGAtYSN6DeP5M6ED+RVaNcqfkPk-;9&`w`
z?%EKfcKd2Vtlhl%`e$wQ(le&;yh@XlvOxt)i*eIDg??=S)ngN{h*+;D-4}d`El-94
z=7eJ9x6u}66gXvGwJGp*I9ER=@1k(RaarXwzt`u(9B3<ts`2sMM5Rl8PJE3R!P9Z#
zL~VY%JHUIYe{{W7F_z+A^3B0*zJGJy4Ke<Rkbv!O>vdwuwe@8bJY~a3hRTeDIk|xp
z6At^MjRXmWK^$uE!K^TgoPw%}SnWfmb!SlvOS2CF?*N!#=y%s|1ETp(ym9x%whli(
zswuu1C+?dJXqqt{<I&;}m3H#qcM=t)cPzgu5IW|Cqhj1X!0r+{KE*(cbdLxx!LF_i
zQ(LF;H)~(0oV}0{l3RDTW;yh(nXQ^o#~k7NU_S~kmT%{2{oBjDhY~?-zaN}4I1LD3
zdgEJBZ?u=4t9w*X0r(&T#V+H`{>b%5%y|i#$+?_2qLNlk&I@SpfrB#=??@uobma9}
zXlm~)uI3QLPE^qbTQJoM%ccGkVsJc4uL^t%V#Kg7iW<Eph3tNkH>2%b)r;0wUQaq`
zg-?P=PIIm_N?OX^`b@r|ccpA?%D5p<aQGs?d4uu>&cyV#CY^R_OHubCVH?i0f&zdr
z%Jy%l=T|pbGr_MC;(j!k)g_BH41Jm&E?gh5m48={Xedg9{fPlr%<k(AN*d9EL^l~?
z8bc`q5&$<du2qQg<t%cv#r2cY4LTh-%syCKEC-jBMHnjtu_cSxswTB{94NQ!{Ws8u
zqed>5CMDRlj(Tz4^w!)9--<17oKby8m#4)@h6M)Vae^=79ztbkagY&^&wD}vq6Tcw
zQyZ7Q=yRdK`DAanEYql}vUO>gr|^nj@P<Iyz}ajc)f>y^b2{u>Kkb$qbE28zqEr>4
zW}(;K3xPAPwoe_n)>Ys>NPI6Hn}zhdyyzg-(y>E*p#PfKP)m#ce!+?Uo2luJLT5WQ
zfp1|sQSDj^pL$u&LG$jO`P`7*zqcaiC)Xy6FasA&+Uea+UeOi$zU>r&9%VLWf4w&%
z#J~VUQ1Nhq0+gvvs`jS{?e7%uaN2ZBV=k|FISWC1j&}<XcI0(DU>a)DWSHeqzO@T<
z!qOr}=`~z^dK~9$a%SAIrsqz-?ICp+iq&@5tX-%hkC%`_=5kNwKO;~$O0@4oP&w1l
z^eZ&(1b8+jG_T+F<dqGbQLUv4($w~w^()ry2=`G8<09PA*FK}iG{0G9s<`!A1D??s
z!Vsl0B7fqh`3JjJozKVTwjYk6xS%RUDaqgCez5Lux?Y<o9KR&-F1foN?{T^x#v0d}
z3riN`=dOJB^r0hd%VA`wY=Q+dUE(Mfc#T$|L)@U1X=Z|9gZ6D>Ms_P2o*oCsq@=7u
zLx~n!+Wd#%X>Wra-RA&2P0sANAo_-pYPrulF$gj;a%A{Od9K>*bVI`mtzd+m*1Mmc
zwKg8n_V#31mdka>taTVNdS9g)a)$eRm+5!sS4_(S7OVS1^nIb%#`*;|%SXD=X_;Te
zDOqCUYb4nYE2QuC#@%7@T9_%`M3NC98TgsecfTlOdA7EBQO9Oq-J^M&e~?eE`f1D)
z_#`D_&@#&EyMk~uP@_-i-HP*#(jLb#bP_rWfKzSu>$CBGm#_s+ppL#hpaohKeQj{Q
zY(nCP<^T0mK!&-IDLIa_z7vzQ<<_H4ghgMbht-HyyHkKO@OH??o7b8Q$kA?JBFN%I
zt8CUcuoa>z3&Xsi;>7vrwZH^4pvch&?!S!zh15{QnK6B1%OU&hZ;jZdX=<68WWG~(
zhB?g9rRA7go9z^dgCCcF2y_VAQmmy@ZL`P)3Ghnq3<c3#8%)7SwZ>o3Tp)$@i<ucY
zy|c6MrW8&kV6dE;M)ST-8=Jqmv+o*FaKJX-b0itZoU4@dI0X8?;TB-8HiynF>Y0nP
za%F<yRWzW3mo7D&aY!)QO?GT*4!%K_C8sL`tk?o9n0)9^I3;%scU8NuWO36-zaoDW
z`?ItDP+=TuXXarE5p_vR7SoOnB56?7E=qAEXTg(gh)bLfw5?8*FqawoSp+X(JKWeY
z90)LCGl`HTVd+YjYiy2<TPHZ=x@j7dGl^Rzv?e5TRg{$jOG`^l0?CK1IY->3QaXKf
zT2j0mzY<kdY>>{Bn#g~f<0PyfWchpd6JzpWM$$|r054{0H7TIE;}kf6As5+CU)=5r
z_}!0J0yEnkq9{c@G5ncUt{UXe4~0d0`&eA*2*AmaMhHMD{A>DG5C|Ldm#LPz<_Kpu
zZlVnlX{uv4*%op{4)20dHzuO2EzfK`CjNKAs)3j%Rz8+5-}9{WhFg)&(`4?dP~M5)
z^SZ{vtGFJJBl;F7(UR6K)hY}GDcgRg%p0>dCh>t2L0)jt!|FIqF<v4@g&^OtlFAay
zNR9>qFM<DjFZlDSmtekMXD#e(WB!<+N-5fcu<CHUU3rI86}_|s>gZLL2A;Y)OscQn
zQy_a0C%toLdIra4Ca2AD8eUAe=pnIZRy$SyBbCAT!hdq^Y#&N3I@*S^Pu)p38qnNV
zaxW61_BYf>CB@?5#omixQN8DIAVzgrNLV;iXBrVV8m!*>sP5&+3qTE0%XU}NeGHDe
zodVhCv`IDb1rs{}nu7Ie5MN?qS}x|!Ti%Z)2ouU~lS_HWkSx=XU}-Cm0@X0)PR)&e
zb~C`jEkus1Mn{<gN!lV3WJ*{)*nk}GK+g%lix4pt7ClWT!W^Q9kU8RX=Az#Xd5O8Z
zKp$p3KD_wkkM>F^K&7KtksuVU9m$vx9?{p`sQs>F=2b`FG@$zqAEl^**WTon(Z-B&
z1J%Y`$p!AxHiFEw?pY0Tf`-}C&_#4=B9!%Jyq47ZOlis#@0g!7KmZk%*FRg+#6q)S
z?O#I>BiF&@SW}^H$fs4tQ7htb`@03ANxX=m_w-(>65FtROB<r1!T)^m?C3^VJq_dS
z$P%M|rm52ReJ)cn`}9_4o*znS5^_rzRe}MY<cVy(&<}v&kUIe0J$*^~=-}DHqoqVr
z1u!p4yB!*-zC_&VAZiutm6Vl{Q(4Db^e^+{f6WrswLvW^(4y^k2(PW_-yw;VnA0yr
z&yLN%oA}7Gn&8tI+Y_V`xlq0F=K0_POYzi0RigYU!9R>f943c`4(E}>x!2jkdEnUy
zu<DdO+O_UtbeqiX4tFC*p1GSHp&VhK-+%5VB*dN@B5!QZGU-3AA&Y~T^##!WymF^1
z0*Zm9l&9uelq$@Auv`KXWM@p5O;wU5c}r)oVG=R8XK({6F>#W*DZ$n4`$y0n?|ikL
zT^(ZF{hisuFRHQc-p-MoMt?UQ_}PG<hO0Q}uDnElviU3Z=h1+-;-9G$Bc3tEP*|AY
z=M40&?L(ybM)l#;y%FwQij(V{pj<4xtShkvqnpR)fqED15+X~<IklD#z7wahUX;i5
zt;{KYni5=FcC==_`HEQA_3WJKuFj{T<Wx-whmv>Pt*Dwa4f%=vXwLQ2{k&ym4Dxqx
z{@qXDR5e@oRhz?pthzVvgog&7Dd^X5#l!*eia5IU6ezJzn+MH))_3r9ayXyBenUeL
zBaB&)8<8u_??<3K!vA!{#b2XA#NV#2B>aZ3OG06@!F-slFIFFD2T4h{+5A>*`lIkJ
zM7-|<8=v8{+)PTPgBkx&H<`2kp1cL8*_C7r1xak?M+qn|I!fl9ar4tq^_*FH>Mw~S
zH|_Dk%7%ZW9PeYFRXK{|C-XpKD-1Z{MaYr$(_cX+;h3{HLzCC{#%G8CS}pPGe6Uq8
zERiCxs8Gk(MXWwu0^cZ*d@aVnIFL$n2cQAvkKZIwG=g^uSl+Ra>aviGI~v^&+?^B!
z5*Kxsl+|_z$<;Cv^w4AiVuT4bB1vQ(pg+*h1Vu~pSs1S-4iuS@d7)_Kgf;4RmGN;f
zL@&KhE8$M$J30_oOi)^%AI*Hr9f%shDhvYdezB7(rKYt#$-$H2Oc|E)-nnHRGy}Y8
z_fRhQ`-<Ku>E3`52Ms1>+4Vx-AkVl4!<Inl&-jzVHU?`tC+lHu?Yu;K57KEBb!w3}
z9zJe<ERyODl)pt!Nb1NG(u4H$Ao10Qr<*RYjsk6Jo{u#tl&>Q9G+4O)E#-moZ3N(x
zfPW~0uI6@ni_DN5G3ix2=5Gwcdkx(USBdl{yg_;BPgT_flIDgY_=utnGt;YktM*bQ
z8!^n3FcDF68?U{Lv+V)go-x5+;u+muOXQ{gCDo0?SrD(#ul8zdxS=kw)JC3@lk;nZ
ze}d1EwVj-rlH{sOZaM#<hBQDDn{=1{Uqkaa(f;^<GHpwBveETu)$HquDL?9+NXdZw
z80LGTPgsg<@ZC)pF=H7>ujh3Z(fi=Josexd@4?$Cg!%&&lXDTdN=2$!tvQ6%vq;(A
zlKb3*XJIlz@izQ;0Gwctlda#9gBX>C>yn&|V{WySytvbBJ_RpE3*OC|_JHcleu$h6
zh<V<t=Vg9w^uOJkc%npRF!I{qpPfW)hn+|vx(G@dOqRp1Z6OUUpE?LQ2>gD>+4OGP
zv8N7S3+H_XnK56a+cYIhHV4TL$Ag9t#O7u}G#Ig&3kg0bq2n#Z<5B9y$-yGD{o6TM
zr3?W!+B*kVTaRSt?f{g@Hf5vzJF|NQI<{qp0o)M$@7n+Ne&q}c=VN=BtnH6%W&{g<
zbv%u0SluE<2=C3i#USfA3h;D6yAGF5O3m*@$4{T$B^V@vWooTw>nFM>h)^nTH-{aF
z53E@kai~xN8uh;!CDLvX&H=i{=9!;gdHn|K_v-eWervGS4aa6DjqF}uQY|!EM|R2l
zbHe`dpX2ieCP2l^pDf!o)xDesH0RPX#k9?N`*y0*|LE{V41FY6K_Qgk%lX(c#DPdl
zxm!}kz?u93b92CFp>gK3n~@L8)M-_!EO`yODSRLqAdeXH(#e;}KZPI7K9&`D^9mk-
zeQ>qLOZ_jBwOt9%vRdZfwApvyow}psDd68hIWm#HnL7)Z5K(#kRC2&h?vj3i9@&c_
zZ5|YeoR_MCJ;=kmr>?F_xo~MGc%va~=9BezD}kTWXwN+puX5LSZoS#eA>*Zze5`&W
z)|y4E=T0RktPJxs)~XT2?*<xz0%xEVexHA-^hHHeDX2nzS;`(Ieijx^9*U@?+7A{H
z0<gg%m6y#n6B(}408s5I#yDxri)=73@tGc5frcL0$d$viEbwm^d(3P7OelWt6IW2)
zTKr~hkSxH~EgH#!|Io0N<Mnd_Do0nj3Ms1FAR0WKwyV-~x#NyM%o=Rwn)=<x;rySC
z7p3xl#wpq%BZImQJK^_>Y8DMtS>P;({>}JOR4GcXa8dHzP8REKDmVVQ7?riA?}>tO
zdwb@ZitO=AF+m)dG1np(Za*oqe@Se<ca68QTObJ%AwvN)VMLjd^9L_rhYsq3RG{pW
z*@Sl*4Io?g+Uz==zKH*zY2^9qz(JteV8!dq0#YZx{XF%PSP=^oq@dLo=-G!OX*uca
zzu7rXfCClmgm<gdO@%00RAeR#5EOFmzo-<YxTQ-hv;e)$SqONUProc<QDM0DH=j@H
zOW8CC5uvfBNRQFB2WAMs;78v1e5%s~URwM-1m;M%v4(l(CjDJDwkVM$#?Gq8FDeDC
zT>F3G#ecHA{8aODrt!NF5VMQc1J(LTJ+HO<UG%lI$SC$DU&zs5%OFrWcGl-e<MG|V
zl{yQ>FBFX~y}a@2+_56<`j$AhstajcziM5M6xAsX(sFHg;!ph`$eDV6bJD2}VAlTP
zMePrPrrL9Kfd|d@aCDmbunjW*?~(2~;-vbL)AKefo6dXRvv8)@{!X7|@So3bOM!>|
zw?!T7)vyu+gzmwv5)cdI{jds|5p2<@8tbPx!hB<-oC(!64a`~5?fjxb-IfL$Jn_nm
zTyyo@MVaGrv^=AoN~*l+xg$N%op0ql1QV`uu=xsH)<wgr2+fDvKOH>x$NCKk;F@Gu
zf=&KwE%h~<%gW&;1A0r&W@j$|@U01A;Wwn_HgiBTZ}m?CYm-L@pjdPy+86~`;aT*?
z)+b4}e89b%9M;?`ORbbiwXVOXHc(Y{@nwFd|2dlX4Dv~qtbMc$%(vIX*S*t{`HUUO
zWGx+|tu;X=(wjwCG<Ugl8aa_2>U`kA1$6$?aa(x+G0MLy!fY~BOsdphX^*6E$}34p
zm#SLYY8~iJFaxwrK*E!yWGPCEqbjAVEad~VrXtJWhe57097Vt(?Pul(zUWcSV}5%|
zeNJLTe#dN7`Vb3FPAT-X8?Q8u8xRsH9U65<9}SnIanuGg0@|`4gV3Jq?#C5QWld<}
zrA<j4rjn#{?l)A{x_By^15B-VB1EbDwjR2zD=ee&Pke(OI!8o><-paBKzSRMq|~W}
zG$XeR+22o6I$p4RF7$yc=#j&kkf$)|0_D!e?$pl;+?ipT4F?i<y&C7VQVb%R)~Y&s
z_fUAe1zJX34Fz?een#>~fxplKxxu&>2bpW_i<S}%29GZ%-L&%P>(tM99(JUH#u%8~
z*y{l(v|uZ%KBqO-2p1=ML2INT=)CbWYcgLAK9G+@G0r=##h6g$No(&{QrXNJH@B1_
z!4%f2aSl3Rox-o5ay{oCHk)sZ2Zsk|WlVpP?|3`{$(?(5RO6<gI3}iQQgKy=RORG&
z!s+FM5syREbNrw0^6hCgWtfK~Iz42ZL+Bf(FX^Ap7`QI_{&0B-<g4GY0)O)D@ny1-
z38>@MYVL8KKD6kna>TYWlNoev1-^7~dkIL);M`HV<YBH~Xcnuux;xvFA9{b#NpDK2
z{k<P=F{N%dyq|2$KyE?XSmnT(66r_zcF%lKcgkoUiSlO{m2+)jCNZn{JLh?ah{X>s
zo49Df=twWrajcEsY|wmlSycE0TRgp8WPFl9Z(;E6scs=L4q*0C29j%YV_%FWcVbF-
z7;!f540OhzW~h0RJewLv*c0sP9JxO4;N|}Dmd9ZPqccF#2G+hcMPr`VH%?-JN$Zu-
ze|#xmEuD5z1$q=_fw^H|B5g`mV%FVw&{YOhQXL1Kv){W@ht8~;U$*zoq)rIP_Z3Zy
z;srx}JUXox6Qu(3y-0b(0s3;lSm#Od`Vl<=$LDx8N0#H5^hrMxt~p=_T@yRwd5dCZ
zSLQM%97mXEaIKr84ijyF27ohJlH!mG!r-;|S;if&n%)|<5pJ{4I6yhywYK}j)^P$i
z>ukh*(@%`_GSKv%!tYmo#`*KSMiqcoyP_9p^{&~sJ*Nk<ALe%0Qdp_~EN?M?u-UBM
z6<ZGdV`t*N_y}tS{qmC3vv~S84eNb#teu#veY^xMaqG}~eUwAqH{<9U<|5RSEkk<%
z>wzwUB;6ee@C9w}UGoOKr0wa93X-#a-j#nYD9o%^OI#K1js-koD$t=3Q)<W^K$TNt
ztyJWw)kSPvjUuDGqsfb}|E5igwa8_Jp{lBu=JdITT^n&TW@}S0bcY$5*(C`}ftc~U
zH6@>C<QyO7Q`c%XA@u&KgAn?vzYh-v9=>ESGv`t;H>nKHz8H_1-w{w^d|>eYK;OB3
z%gLPTZt}rZiuQMYl4q#WUTl)m7)@clTWi{J13<k2E_}KTQrfFfvecqwFG}*1qP-+5
zoa(?i(tqhsvAm9_UK?#Vhg&IWI|uj@>2D(VvJCYn34eZA`zFh<K!5%1KOlteU8PP~
zEY@3_MAawM*=$)CMg<YFKx6CG>efPoNG38yzy*VSq*()wADXI+1_dI9A{?C<nvuLV
zn9?oD!Lq!(h=G*5HQ!^2zad)CeYxNHrA9sIETERP0nDQzuT4^N^PG{@SCopb*q)c9
zy&S^qbhSJy3%Rd~cRVbE95uv(d;gxlcu0IdL#@HMaYn!I6moYxr)%uK7swaoqj;Vh
z(67>mY3uyBfHc$_a9{hgxfRJ@M`jUWbtXxvA2R(_%apfs_p(35E?K)h5`~-7uK>bS
zT1~SSg_o&uGh5TFX+p?D5V*V6Dz`2Gz}s{J+5{9p0a+~~j8U2sZ@(cd{}qw%U{(;*
zQ^JZvo2pKyoQ*JRz{v?QRc*+J7}3l4fXs@6B^1zJyB`GuqSn{&fq=;hd1>I<$m4
z9mNiGcu9JG?sw$)C?<TkZi{)f3N$9#y*$3pjurc4ez4W{md4+x!}LX=>)}+=QE1Gt
zPHTSJUk58%8GUb)+m2o$bDnlrUf%B$i|RFLm{h59onpp%pwA=++#909=8XgBs}}rN
z{k7X4o(k?Q<*Xi>85X>nt4xG`5^7vhlnX9)^YAub!wI`};Pbq3+gJjk`!iwM!pG&&
z?j>Rr4?Z@S7Ll|2d+#x*;I=n}+|<K%v_3zw*RW}pH%^I4c};cAcpc##W)cG)lO&aV
z<$lZJO#Fv{96`5n73e^Rn!4(?%6MW}|M-GnSUY2@3_1rVDZ8p_3ZVo$kjYwDQ^k*+
zK3AjQgr|866&}&5&xyrZdta40T$KnaJL_~?^9TNYVM#)622b2BX;9l15DERf?{^K(
zqfdbjWk%cUYsb8kSgn1<m2;@O^3y{(v@7uSL;KY*Gl8t%>PvZFzXN8$)6?=BTXovY
zcgruvZUB^dml1GY0y}*7kH?c=#s9+vPzDt0P@r{4`tsL{)v9ErbFF>nPT@r6`dBzF
zUx1$zcQxFC2lN>?IotGS9==T9G9&wts9CBo<AwvJ=v)1YPNU=4{D*n8#hUYw{MdNW
z-VglzR&-|Tr}(g6-nt&o8<bfYuAygKu^Di%!~)QHl#n8j`@e6RMH@wxs}1kNGG%y6
zASgM%et2xnzCpG%)FH<}(vSLbe#AFP(u<3E*J5$bKYkpM5;ML3eFFM{w_8;y3sQ;y
zTuLjmE-|BYj2-vmMygbCJaE$+5yzCxMr#`2PLTRqU2hTm@iyn5`8t@`mz#AyJ1vsl
zTY&51QS5&9HJ&HHXWjDnKPJYcS|PL;DNF0Z>@V+5TCtUC<hQPRdY|W9e0Jm)J$<{P
zgr-rt10qU@1jKCG5zU!^8myGp;b0)G(OuH2-MY%Ua0(TabA>B!2|gaHUM|np&$)y%
zpX7f0kgT)-GvGVOIj*X;Z;mEETn`@RTt3ZD+39%Nl;VuP;EnY`c^%`^!$4B*$N0Ee
zsjq*@!TN`wBjP!XY=3@%u#R!jxn-xI-b=b~Lpfo2xOgK?>&2idmijUTJzP{lQPBW4
z(HV#3%r6-k4^O*NQY>1RIW;2BdTcD1@5x)fb57o%k?g>=&}aFm2;l;%JLu6$KbLCF
zq~9EJ8ZN1PLuo{|g;{_xKD2nT&U<p7D7ZVL)o9zRe7&B!%zgTB?9<o)?;0G&LId?H
zZgU8>)~G*#OajK}n#1M#Lv2at<*y2e94~t5(>(20Bbq5lqT9pm%6YSE?;Cx9N)JHd
z(<yiIqtavgaM$uxS^D!3exy~qHoeg^gFF36c(IqD#Z|^fuW*tX`T6u`V{Mk9w*Wuk
zbg|pj^p@8Sg-?45BVn}-)h5?_(nLJ&Dmc~wh)w<PV^Q6=)&b<{(`qWILl(J2b81#8
zK_Zylv!G8!@y!=AXP5|c4$;PR!}h{&5%Y{QX#F7IP0Ej&l3Wy3W>?As0y45=8B#Rq
zq3Yah_Lmum(JVDp^!*(ZK7*g0%lCEs=bL?10?@cyFOp@1FeY?(vw3`qxMzB3S0RZy
zoiUr9hY`_Ef0KyJA0ZD=dEfEEBw*mwt6iIp(66v3P8K<?Rjf~AV#~eBEj=E3D=Z7w
z9R7_S=WgzdCn<$V2|_Acgu4zq`4&qqDiDyeAX}>efP<PDqi=)<zZEJns#Bg-9O2L|
zl<Usa5YFVVI+_*a$&<a~M%2;y!1^dZm+;N8|H2_K&Y4W9$TCAmd%BRK>K#lhYLu}y
zOJ9OM!y7T0tVed}fvxZZ^v%zopMGkHfB9%2Ox{>4r8TCis%rmFtDa8mXf<zT$lwW>
zDY2xte`12O2l!*(2W1c`lt1auJPzUJGb<cOp~6uglpVk>RZ4rkAU*nrf_ACW;;q2?
zX`c9PT9$!7#l`~|;LFAtgH$JnZ*`-NKBZ&4=SSz8`Jc3~y+y3}_gnRS2ceR7rv_k`
zZ6U?!S?vY=z*rQY-)$v7@g^b7r%7OWakFYTKyIE6z76kP2kKnYZe+jD6nWwB{7R@4
z%9?Whs2PiIDV4+O-*9+4Tsf^>YTGxo;*H;!_aRXOLq#i&D=CIuAiuf%YF1R<g6nZ|
zo~|DbFAJhc|KWQnhFdzj?t7!P?p7vxS1Rolf^1)$2uAoEUOYR|L4Xn7k-7n)=NR}K
za&0gPJF4QYzTbX0Zdh#TZ07KS&>b;vRx;P+`I`AQ9@jjHFu5lw$%EpA<jdN9XB*vP
zpO`_NoumC3>&j+N0?jn%Rpy<g=2b3c=VSmhg)b{QQV(eW{vZ}?Q0|)xOF-{i(6S9+
zh(M7%qG3Z}L_FvJqcd8DY(lDeO&Y*@luM09`VFaRhJp#PUjzs_A1no)<1xScYPT@^
zTSa}lek{FD=S*F#cLNbVkg!>s7C3D-hDZM2npF^+XXJnxHuw3VFy0ljGjc!xm<Ilh
zKJVwwzx#=ZCambWh@86BA(`W_{b2IpE5cjEu#0BjX3mC06nM(;pT&^r*lo{EV91Lm
z2{crNUX4(P&l!lHz=(WL-!s;KvB2EVjBO?E*1uy_a81gu<2$z+02*3sbd0|EQ`hHd
zsjT1ispn2Tsp)ggN!0^jL(GPyEEUW`{>pzO$||zDyOIx`1vW;hR>6uzu3nErWyuI)
zf67V$lQUPI&Hgk0(Q2us@8vG^P0px+&(-{(&k-;K8vbNYSm&RSLl`zHgd7Zj09!16
z(H9`>n|bgZ<NFKhH|2~My5HNbRBX)?{WNjUAzDlYd6!nVs_DdEude|mIr*3IVd!#G
zh`7YfNuy+)svwK$xD>rFW%Zy5U%s$IPbv_zlE5vhGU4v9K%WVKhN&^JzpI-V5BdP}
zOa6V-0X5lY4Wi1&qtN6VB949EUsApr6I#-$6A!(QJ?=AA`r}%i++#b*xBoS&hn{Vz
zVIisJDC21>ueSli75lz-e3>Tgs)JB^q2ttG#$fmCrikKf<`fNvlnRD~)i9}0cr;*C
z55`}2-#1$GMUkuN?CDPGGb=|5&8)lVFz~d_Ffm}mglYG=fPCu%d|Aig%cfe$g0_u`
zYJ=r7k$C>3vQp%QBK3!z$J)QMMD}0k@XR87c|P0`F)MMYIkfK>w2lb18)v2+Hwcs4
zhU33)GPMA)UP!SY&zFhYbiB~bI|aNeC1K2D35e5`H7jXeSGBy62#5p-2sp9gn^mqu
zVoYn1;5S-`h~bn9RyMXQJ0)><5X`xM_A?A=08CvrYEX9sQD#d13tnmusi8#DwOJTv
zUr}I5O1gT)+MFY3jQjxa(%QDHD@YA$NJKC5lM%poqPvFd!RWAQYo#%hJa>mA2I~FJ
z-pud4iM-87;!UP*qF1r1OGcoaxIeVMGEs&}+MoT{Jhb&B!cr9P$4BAeGZ`Qk@Rcth
z-%zVE<5}@l^06pugl$;--h=}fk|ArNSv8gP-)*8NZ1*<u3U9m;q4LBY<32|P#vKXv
zoIlTm_zxjP&(?2bx&FM@{w{|JzTFm$#inwpCZsqv#i(34D7ytbGTk$<D_X0XMNC+N
zPI(Z|5$0%pGCu@qIOqdLFe!hag#EIb0Ep61m<Y;CRItk}pKG>(4+lSw$JTx{?W36e
zzXu`&WC^q`g#ImTuMgFw+C7_#rE8G<Nnf5!>cw|mMbwv<U&4U|_ietRc6FM~Ez@V+
zkXEb-<RF}caXta`6w4Ww{T*YLmTNdc-~`|HhhT8p!KMA#JYd@^=WdWHJ&7%aV`kGD
zn>L+58l$6UW*$fr$Ar#C&?1Joa<b>v<`dyDp{9KvQbVG`Jl&iXO{oWNKc1_JP@ABB
zhI7mVTlNm1^f2De={@#U*_Ec6cE6L9&W0O(EE*=`dLZ;|HCBZG$Fr8AE{V6l9%j1X
zxCnC_PCgT`{;JtXF?bXSdMvMS0C)j}&$Lcg6M$u%rDhLj<#Cz|A*t=OR+j1w_}-0O
zqMJ!7&8v^ia~bJ8^1(<trjy5{T{}}!f}NuQ9h-R9oKXUwqTh(V$k>{?^nEQLi+VrU
z5krWr>9_(E)Fq{QB^|ZjmGc%h0hfu9oLe`_)UHUY3Wns!HD*3q6$3L)USr?bKgxqh
zC7hn5VZ15EY=9GNr3g&IK(KFA*ZlowiNS(<_E45|T+7X_^4oJYr$tH#M$LA#{c|+D
zMNLKDSqXWZ=OGdQq-ACC?yBJGy1cgHwsrJ=#qvhYBhZRL`=DiK6i-Ba!}ea~5NtZk
z0<Nl@+by>1iEp$9<a;_RE&EIJGJTT3K%AJsZ1AfB0%Z&Cad*SPnt4~R1|AyRNUA_`
zE+fpjUo`kqBUeDj4j@lD1r{q=Ndr?IL{-cIe^26CnP~|_G<=2$O!~08=s(fVENZpu
z(eK2~P}+NtDn$l85_H|Qca=<Dv0%Pk+ISc^dmwx&dA38To{&-oPhk9k@1%{08?YgE
z?w!@1yFeCwd%z+;v+SQc1G;#gmD2L_mfjJl7wABO24C+2-*xJ5Sr_jC8Mffz$`L?9
zPd&JMR*JE{iy^*2@o!C1@_Tp=w~5Yj=Z|{2&wnKVRh`D{+mX7T<lRL$@TJofE+efn
zF83@LIi0_ls|=9?dWygp1TMVhfWt4<>7Vk?OHfjAF4-*<$@3hb!ld~%_}zKvP`3th
zom?b{p_YXPmC$%Ia{H^2k3%(51CzhU4TbA+BH@~5-Bb}66B5YIo1#>wmKzB9+u^7c
zBKi%txgn4R1zfZ|3$u3a)GxWh@9{;?bEtoQ+Lpyp>A2(J@m-2~Ni~-IIqKPKq}|WR
z$!)!MY<G36%%tSIj#Jv3R%&v6{#K7jU7bx=9y3)dKF?q7U8w9p==Wwr{mvB3!X$X%
zuP%M&L%*>y9~RL8vjgKm6p@vH@3G~x`V=*NTMl5Kwel0GE$_bWn1d%%XlM!x<S!dx
z7Sq>CCk)qufT2$G^K+^SQtaG#PO-#ohYh;Se3w&YjlfLmOTV@9NxYx;gk@J7$1C^6
zV!Q}90T2K7{q4t~3#WrKD`TIWtq#29pHeK+sh_fa*1SoD^WOe5e@mKJeSo-i0EY>g
z0{@kvtu0d_o#avMsCXg(K^ftJH+7^mT#oJwH(d(?Ecwfu%&p?CoEh>sMHSj`Z9Pwk
z-5&=@FRZ}YoKNFSm;YNfOhF^VaB*%sm4jPc&nh!XKqnmx^16*y?}l4_eYAi`y$^X^
z2;ef9RS6}&FIt<v+Ux?~Hw?*$B3d&^*y<q%b3Qv@QmbG7zcV&IA_&<ayrtr62X&yV
zZfNbdF;yZJY8c_R*F(3BH7L(}DQ$kpb>3c`wEP|66)7~1G=SGWvG)^||IDEOt=6A&
zPE2lrQW_+>;3H||yg1@JiRH*`hhy&xT@tnpyul5d5!t3+lc|mcdDYA`32H8SpN$vB
zwg@j~UEMj_5>2!$A;%tC#&8oF3^9Nbm+vb802?|k!LODm?M;Gw3#TfxD@J@DT9g1G
zoK#@`hra^~f_v=XktM8rFlcm+?15_Dg4^DXmS{n1mPE(eXA5uwly7=-m>=1URZ_9=
zMn7K^KcamH`7PzV)<&$GuuQom2{Z>wUIwY0%<hp9WIoGpSP>Gkki7&77_(6+$(>4x
z#VJ4T0YvqjW%|L>a6lMBkLbBr9U&y)X?r=NRuV$b!P}QouUmi2o4yzHNz5Z{Cnd}r
zndSnLIDMN5>KMEl4RN9l&D}*j=GbPO@iHR#@XvkDbZ?*ef%=3k3Vb;<1Crj*+c_ZI
zooHNLeGYv}{nH#Tu=s0slv5;T)6TXhB3Z4@TI?*i2^f`z{T(^#erFVQd}tWAgDo}J
z(63+kG8niy;BzqpHVfn7fit_%m>Tl&nMn_7#}0CjTAidxe4F1AM-;$)re-sFmeJNX
z(-t0`3V)TMB;u)n_);!yNR=k5P+%xxWVby%%4~EGsxzN`@tZ~V^s_(fxGTD8d^?Zd
zje~PWL3*=(p~Qwc*DtuR4WaEOFPx%>X@l#E1WCjxQe^j~cw|g#9qWJ7U9^sYmGA|H
z$&O)fp4cu=zmV~EbLa@uYdxmwVvSrlwlrU8@f!gKq6AuNrwHDhX={!B(;<0u+G*g%
z0%X9*MmVLQTZE8=vH0o%1e`}&oR`{{I4NM$#1D~_yUzXl2^>)<8$fWZ(*pR=9DgqN
zz3TX=g~p}pAnB7@u_m5TVblGg0{3f)+#M%V&%fwLej?xJNC7%qjF7y}Lm;bV-s?m-
zdGjQ~&8NFljaKOP#{a=(#wc&EOgETIUY&x%VkgqT-f_wr!i3)EP99O(OfaST`(!Np
zWdq~zS3tOVP6gfpw!+4V`nUKAU2H)DRW9tHhZ>~fb;BM#A9v_lt}sn*f)G*L5P66L
z@YM-`DH$-=PAhp^w|R!jx}Bb?LUUyik7*k~ZDIa|u!=kiqcV6L0CQj9y?GUI7TSY@
z<1{X@X?IghO-_eO2wlK0RAKk`i$AXLA+4r*k!Fam!UxR30XZgiipZ4702;@J>tRPA
z1^Jid-A{ofpZ*7m11}aH)c2QPrQ2k>PxxvfWy{R@o(JIfrK%uk+GesY6v+)}h-T~B
zG%X4yRanM1=yJFln%B5q=P;vjhSeG2en(bE9yc7956~<%N5b1(0p@_`>i{YEI}d=x
z#z=!ns^8;7mf!6~{wDmM^GOYOa`3Ys;W)004Wg$LAdMECO#p6ogSX!Ndi0p${Q6u#
z<2CFx$X^^X{f(;Ht{zra6}9WS`~0Uj#$Xfk>lGKwXG!TQ%}Ui<Dw>w7nnY^r+Y9iM
zpw|uzsklSKsoAmZR{Z9&jf-YqT+lDC<C9=H2??_A=!X5Dg<r*uk-g<!O#_kn(-u)C
z#W7!?A@=&qN(-aLp)aa1%%x;fmb(|Ww?6nDn?u_gTj-V~-CC~3j%V8$>sXz<&PMp|
z+K%PzZ!55_Da;HfAQTO(0|eVY5u2lSHOzAIA(#6i-555poZ+M0n28Ldabe>}gm-CQ
zkj$YE7O5M~bK`Cy@ROp)m+rGB%9RD0s#N#cpYcNs5U}2K#_ZHsg$?1~32dS}p|J9O
zK>enOw+?uT+*en0L_4RTobA{w1ANon!R>y_F$HO<G^Y|1&8u34q$PK7_I#leKNgud
z2BH{M2dwv8E%l5b4>b(ZnRpO&J4Ntp?Glb^a-n0u;<B2m9=5BFTP#6_Xk&1(M%*`x
z-+P|iLQLnRA7qHEg{Y=|Huyq-?jGi!N9UG<U0$p|dc{T~AK(8epEt?$^;<kf>v@mT
zQ-J@bGlS0+^#fC%0GeR``yTws_i8ZLq2Ge<nzhT3Kj`7D<F%mM?Fn1w+Pk!%Yde!I
zPX=EnGl@=hlC_|-x1gtZ5|=g{<{>+TGX3y#A!{i2Nre`tx{%3D1-onE8MZS8<smCp
zAv+Y#s~<H%;SiHlE09mCnSM~~hLPxond*jdme{tSzi=4!`KVj0)4n2Zqy23UqMGcI
z!z3zv-r0hFw=c{v9(=8XmoJuFCx^E#iFxBphHzh>U3Q*n=@Z2CgGe`Q#)T}yne4s=
z9jrLIpo+JS=t6o?a<@*U8+PtOW++Eqb~oVICMbfpE~fhmli6)oxMC`}Nf|F+UfAi<
z<EdvH&#(TD?>qGI<E2@7i(0#7e`Ee-G1Cu*-|~c*Z6QT_rllR}tMi?mw|Z@5^VTIT
zzqrSCh40_%k+L+4aICyyO)B4Y7_N11@I@bIi_W=EE2!K>zA)-$!i0er*3RDl<DVr|
zr{9VoZ~cU_Le$Pn!u2>`uqZbA-nFg|wUMt*h6i8J6&ndjRHhigW`<A8JRmE@^eqVK
z(wvu3!a6^TDa1e&CnxKN6WfO<!eR}q!wv@$kt_VAmGD(p2JU0|ZwsOw1ISKu+}UE#
zb^&y^I9-nHVUKvg9MgnC5Ky1CYpjHSU?$*)5Rgp4acMv(tqogE*MyN28?z>4TD5qD
z+R-Q*i<o}6$@W}pw=Mq)ZX)Vh|1$c8?)#P=Xs6XRSt*BI_tnr&J7xG;C8!$xV~5eX
z@g;fu{!R;eac0P=0jx>9#wFp_bS**NP##`5g;{3EIM&v^H_qev9fm7j`Aj-!nL4kR
zEs8O`f%Q|iq`e117l}j6H;yIHYH=l0_Jw_dr=otg;hRrCLs6Z{%s+bx`X4vDj;K@t
z^Lp7(H<We5_>$i}X~%4okeb51T|3a2eE&$Wp~a@qPSUl-RQIXWGrMq4_Mv&%qtXIj
ze^2<=jvVZ&x=hcte1$6vm+<ds_c=CK>VTPGcbG2D{46uShGLi|VZ8Mzcb@?izgmx?
zA34NY+&?Y{#I|8d8Ws+Z52PgV@YURYTskYXdVIr<c%}8&Q3tnyj5VgciDD$q9f?nk
zfr(+6g|!8RZ(fjPt1Dx0wr&e|bYz3ni@LFWkRiYNTw%0JA`Nfm&dpn))3Mtw7+!>$
zDU!FHWq@~>C8j{OPg50-GE?G_fcx>Vmevd-G}q|afe_q=n$f#?hG`gV8wQsc;hVZE
z?3#;dFhIp3@t&s1u*ZXLkU*I*QEG|nFKHETJaof0WP98xKE-CieMaMFMOjq^u{FyF
z4BFBuDY7hXV@511H&3CmpZaO$oBR=8L|HZ4V?;EIE$xiNQqkyG`b{aoZ3tyW-RS%f
z(Td*1tlFCM@}$%ofspGXBJIWTSm|C`#w!;LP)ez9u;uG;`>&;!N-mFVj{UIi?WSwp
z?N(*wFEPu6N`3`WVza}%6*9`c)p>0{6}|UZvZo)QV<8VRSHHX=y?o!N4r(hzu1U4P
z&otDz@}$DHQ59ANBVkByrIzjsmLu07`?n@5i1yDj)!3D7>{<+pC&d|wHc#vBEeASI
zwC5c)SCu_dezElf9hZ#FGIl+r{~%XgDPO<Xans%5PA{1rEWB`Ik6r$x`5JVPe-!Uf
zrdL%WaNu{1PQ5?4o!*0R+0S3opPsJb_>Q|p#bekz_=U2PBGke}e61+j4;5UT__ljV
zK}~Ku$%F-CN3@w=NKboZpU;(t;UN6Tz1alQ0XThFSN!;YfYX;)qJcG7!{+KtHf{Hj
z6b~!eE>&8(@t5cZfg=Pe4l1m5VfMrO=lm55@o$Y>VD1lYNy$z^YZ&N{CNEHWj7f)<
z9CE*{L!c2%KVYO7@(`bn(?d=(TEA@if==)}IWW6KTF_OC3`D)w-DBwXlQfv(+;@eE
zib%3TFbzk6?J~Hl&j?07)~R<t4Ml^Yp=&`m<Brkx`5o{gbVfxzG?)-@(>0icpXl7A
zzbg2kNx<K7$`+=F)by#MaevBD8cq2M7{li`A=;F!<7C(mvJmi`Dq8SCD{<EgsR16T
zaJz?I+@KVtcEeXpsoX*NS5;cot3erid%t=RTF|Fr)yrSb7sEBdb<KU)6WwYTU5~K9
zw~2AAR68VVB;?K^Q4-#e-V?A^?brPl#NOgZ5Ar7L3OtG|@tVSo$a#;tVLjuE*`x>m
zURw9K^3|m;eAh9^`(27g0g-|$sSC(`)McQqYzg=z<7Q`opH{eBc%6LW8axCZ*rb`u
zs@wG_j3G^daT%m#7@jxQ#0^%+^^3a4Ca1^oZn7vYpJ>iO<EtDTGuHlSa4@Q`7Ffmj
zHPWopn|sC%k-B#lYpcFs&}nk<5B669$W<P?qsA{EeC9@!DS&7uCd0yY{N0@EMYd@5
zl7!s&<i_P&JJADgy|A}z$=LGWqlOGSst5qbYQ64p$;L7%%}2KDdGs!Xpg|{37(Cau
zuAyI%@=hEx;um55K6hwjF_*B!2hNo@VG0G;!4%!&E^>KcC8d?i`Q7!s-k6%fwRKM&
z@m?DmjS4}_9LwK?t1#~Ngo%j1%^q%|IUtLbWtk<jz8p%&QHw!(afNjW6`|r{7pDpR
zL4;YoEL3&3AE$9B6E_3+ZArE~2|U9B%z#`S@qX@La@8t-3w(Ad4YE<n)AxN_rLm)~
zNWg!`l*)QQ)L#Rx+{g-8wNR`ynjuvGADX^7I<D^fJ5AHrb{gAu<D_w8H@0otX5+>U
z8(WRliETIb1n<rBeSd4se=}?Dxd;1vu=hEPnit;p3b$ImV!X}yeapUsuH(>@#p=Lx
z+hywSl^G?H5ZJ-~DBrJXga4;JF@&tKTZMOD=v%fQ=8m=;a>9hs47%xzfH>&4F?+ty
zID{oTzt|_91(3+#fF+Oh%~`#Pb{wPJWWIWLqt-q>-6huM7&bEhBn3<)A6t$(rsEKt
ze6DvVKlQYw(<*j^`W)AtqZNx>iBWqZO-vz?UG>jKBEhw_Z!8rI+3Qlzb=~xsag&T3
zM~=j*()-?j01U$t3=wNUeYx0(Ggd$U2Q4OUj^p21Zxu#A%Tch?_MxcKYbsdU+Clfw
z62#K=<s2x7G~VkjG3+2IW%E}xLU447*-_63tMmDc1PaYh@Sv_}=1*BELZNG(j&_XV
zhYnygTQ$a)^o=pLcdZ(-+VL$htctFClNox-t@g#?yMmY#AF~Wuzx$Nh4ol~=^re+)
z@^QBeuV~)h0YQN2wCk#D+GVDNH>hFMjwCc{Wq<LeU3w@h?9zS6(f5k+NK^^4%6D#M
zg6>AA_+-Xp+c!@BK$7QFrSaJ=#5baLiYSPQbk^KjDj`d~!GZJudEBs`3vn%pKTRRa
ze`XayEvVsuL*CnJgh)1>elPtsVBiO5r^kAZKazrNrZ`Wb1xbG0y?wH-B6*7QVed}O
zpsgMet_lnKh)}kIXVhy4Sg}~iWfYP4oHf!II!Lpj39V!UW$hsO4A%Q^cAr)24=)Z+
zeQg1(Un~43cDQ+G$s>=n1_*n0nS7>6<Cfy!BuCEyBOPo0Ha8dJz7}P+QUFs%r8OaY
z+RwavB}Mo48lUbYtYz>Z^p9QJNUazpvzUJJks?Do@lLZkJ&#_a3NyqFiwwc<USHp+
zqYpcHxFt@-Aj20{-^Sl#_MAvOu*ezRai}uYayH7DVi})yEQi;_j-JwG3L~FmJHu!#
zG0bQ~eN)?7x{Gr>G*FrT2<C%I-E+KZO^Cip!3D>=RCS~8@L26g0?3bmYS7ZS!-0EJ
z_dHn-eEI~X=F#Y}xIOqADlS$ceN+M~Is`nb%sg;Z=}~ZPMa1+YVv=jY@$5AO-qifs
zN8Z(yrH_@1VporFOMu`h4q^1)(f{QF<S3l+JP9QxkG86`)|`TEopbm=m!OpIJbFJ}
z9giWsdFQDIMi_??jqBAJ-iIw-?{4yZvfJX!J&C5T&10Q#@-A1T%6cO+IBx#FzuNK_
zp}b5SBkt0<vtpjB=;+A+qQ4I@l2D#TSXvQfXMD=TMrLxE;|rCi&CD#E6S%4(DdW)u
z8j_baYo1%S8C_Yx_8~&z@`|7JXwcKX@i){@)TQNl((`|rse``apP%fexj?5_%ALfD
zeR(>?=URMcc<$LW!kQK0I>2z4d2Hc#Q}MeEIuLeir@6+LSPHB6DFXNQ!hOvud8skg
z5-1pCSsFB>#AaqgG7c}=Z?D{=Z9^Rifyux$-bkNuI~3I=-szAcvLiCrY?#`arCfhq
zBRXL|5gp(a5U=9m8sKg2uRF_yYB&)KuB)avR>?RFRMz~5Ekw3?TFI<3t5HX&8NxoU
zTxI-khwQdrZs>{HvHHHwxKr_=qY|6p!Dhw?FEdfR+?X6T*c@V5<HTi3(&s@OXB#v1
z{xCoqOrGFi&fXty^Hz71+q$*XE%F8Qdn8V{OUOyAIH-bEqTpN331Tws^uWcJP-uS4
zbZ5w{QP=R~2I?HsNiu)?^+M^9H<j2Q8ezvG3Xxl)3?gxOggN)vjGXb<qeG2t)o|_N
z!y+c#igyiBH-fU5I-d)t-K|fIH`vd(B*t0?-b2!zuyRC4?wAik56mdzED0UPu>Szd
z_~vyQ{Y+Ou4tz@fQv*%%GRBkESP(AQD3qz4^-mWHp;52q(4EKjO%o3oH1j8d*PoCl
zw0RoCixUCIE2X_y7|}(PZ2rI~`Kl`_jbtP+b+hiyG4YkreVtjyST9EO<Vf^k?)muv
zHr!`l^UV&bZaA}s-UyaYFXBXP@|VeGpe=Ip1j*|uVwqSq8_Ji)H8+3r0f#)9YpWe$
z(bDOsg=n)~FJX2(sL(7j7!kS<=;lQX>prc0v&+X~ej^sD!XL?}{Uf9z|01b|SwB*v
zEu`u6kkXK7eBMm=j9$$f`6Ar#8d1@kPw5=E<e6fDck=WR?dLlR$_XB-LCas2$SO(v
zexoV;smNE8!6+kUR}92r_qE>WE+9Q5OG;(7OjTg*^f^&@I;3n;<i$P-_BJ)5C5d5j
z8kw;|D6-x%U_oKmxeh$~%}A=ymt;yXrCi)<gYI@|x;x(dP;|Ubf<Z4rh7Jb1PJQni
zJizwE!<9@RTX1?1QgGya!}wLZ$oSlvMnC=8<<X+Y25z$W4y{VA(F#g@u*s~c{r;8i
zsi4Y18-V;<_u$K|C$&ZMC%Fe7hAj5Ff*gt?bbKRuo-yClLAIFr@ABfMpjHmm+K<Kl
z6!>k-xz=M^alZC5O_w{8#@A`KXu`=Zer744;Y!QH%X1roIrF_`=!_I8xI?Tc3vU*h
zZtyNg{mRy}AL7tC0zku)y*!@riN@;nK^(RQl(>}H9ZgOw$7KkYLm%YOLGjUcS*M*=
zy+LN2GWW{zuAJdh#u*N$366{}s98yWIAX+M7AFb6US3fStSH*98vJ@2<A@phHk(si
zUS562tlg0jl8U~^@+1smnMYCcCA;_u-(6C!tNb>e-!|Ci+3mwQAJ~T4K<Iq!>*Naf
z;xO2Z`aVcMg;euqm_O(ixfvYrxQnF7GN_NG0jrOvtfY<4N>ftfL4YcJYl)4@@*RT-
zOXOQ-2`8HDL|)uR^7lLQEr342NnlO+!mI>Q{qc60&3nZacp10s-G#<XlHb>oNuEZv
z?&H;qd~Apf+$H=v|Do?O79C|xszBP9H-Icvtg6~JXoKX7nknr7te_EWo48Pei7>rZ
zUSsS9%=0(cLO!)Lc`0vm4QU#(8v(tpQnwjL3I<Lt+oXRqSwf$7hVHnHK4M%t!$Hr?
zuj4h^OJZSPbc29}*^ldltBmd`51NF&Fiy=TB?hidWZ_}PTG0IMQOC%0%b?xS0(9xd
zmG1-#sWeY(7W&JH=x+NJ9uj5EbZs~!IRuXm&bASVWw~hfAc*I-m@Bt(-v1}UWs4>5
zM^}0zsEVnnxn~wQ<ZJfAA@71VqLTTeE*{e|`jpR8A6rC-o{9k2UQ=CO9k*ZHHeB~A
znqqtg;qVjvwi=Kmw0L9vs_y&=#ay1Vw5ua2F#&Xd9CQt{>|o~6V}zE>s;ZC!&CGEW
z{PWgDG>!}%pru^48$n5RbxxbCt;q6-Dpy{}oSYoRWCu!x$f%R{v$BPTn*#cZ(n@ti
zJeQ+&LD)H@`Q`q!P`c&|$e7r>;o^ZFd0K3Fsf=_pjkGAZRrvyGldgPGHeaJlj9&gQ
z7StaJv=i-enAFiU+N`x<xo^CcmYGgw`ZE2A!M7mR19l54P#r@wnWjLs<ucD6TZKz&
zxOU2yn*#I<=;+X6K1#ac{1U48w*vI#70<73Y8YAiSY}5bAa06ZEgH{JQ&E3V?fVrW
zY_TI5qitj$lK&9OI}z<b+En_h{y{fQy7^Gl;s{e$uHRxVG+&n$IN-=2lAaUoxQ;u(
zet`DO9^x1x<>R7WBzyB<r5OH-2TDL*d4dbhv8aFv{vWVqi-P|{1AMI1$C_k7kpfuL
zlE+>#*e@d7688=GL!~G%o%!qsi1lctDR}f`!ffIwRgOjpoNeZ>EQS93laMTPj?x3V
z5>d7l$nAq@-(_gzWM+Zm(YA)>GGh$q_9dad_l8g|0ELLT1Sdy70*Xy%4kLZYSsV*w
zPF$j)ISOoVGMq+EDUvo5n4qt<Gh>LLC*3y0oo7Em2I?0`huV<C#3g&|O?P(nANKmI
z&lgr3cIqZMqzt?F9&+^y(R+~`5$dRFH?|QY$xmE(2^9QOwQp7uza7W%{N|ITrAC#$
z>sP-bq9FE;p$z|qn`$v)=1$gEc{7YX84R5?Q_vvQD~Tt&QeZQ}IFu~A&<7lq7!q3H
z|Ir1p_?7Gb_HWC2{u$Vh2cgf6wDS$bLgPNbfyGBtc4tt8&^zLVXGYzvv=RSK6e|$W
zF5wu_p<GCjL7|_e26=<0{keJON6>Br4<>;LO%35Onz~VjQeyZpSKK`@(S|A|$(J@@
zq8E|uLOQ(U;t;h$&WdDA8SwL%5y|Pb1i{hk_l_&6<cVmGL4(7V;$LMFjkUwA0BfS`
z*Fpp|R#jR!KwjK^BKy!yv*s!rpwx-OsQ=smdrMhxC_yQNzRLn8ldMw^Icz`vbgSn{
z@yFA?oKmpZp_EBHw~|<2FM?jxf<HIJh~kvNh%GFA;otb6<Vm|cy`s*X05NkMUMqd>
z>T9O172$H&2*lzH`j6gm-hszhFBuNR)w?^U)?%;*>`ZC*b}DH4H>+F~DQlQ$pY8fz
zI^mCW&d0l#hH>QC?_$ouZY(b8#!4PRXj+3a`V!=4Ri;8Xj6a4a^Jxj>qcWaHE-?(_
zD084AoBEYr-l%U=A#fPQ-or9F-Z;db+arc^3hxVue<n1e`syyOpjM>BkRf&+6~uA~
zGy1a))u7@+B>qVB)D&Mb*%FuN>u0M^4#`=`nLAkmW*IqvoaXLIs;`RNDGTxZC6Y4b
zB-(E}{!=YNKrrE1)WP311K{5lLi5jGboU5-G;xfG1GrTc*qppm5+8OP#3gckfKhkU
zi`Xc8W>=lRZbMb9%;7A7XT8IMfcchy;ksM{TI|RxKOriwzW>cMfKy?6k~$M0dU;x)
z%SJ(`R$3=SD%anlj1yE17No4|C_%V{LLyI9Vp9&)F6zSP#dZA(1!Bnsg|_It_)?>t
z2+rl{(Z$bLA{RvDm|bt*^q01w=2bQk@1auCrmLvLsQ8x>YhRQ(R7e5RYHHZuW3V%N
z0Nw4HcE#Bzs-cl(JQjHn^b<v*sTQ_yZ?XuXh7Uxw)gjh^rQnZQd)qWo<;a>ZO!g<o
zJ&QpxkM5x(%IHYo)M$!qXN%8>`;BZr#6|CwrMYxmXlZlN>$+F>5@f^0^NB%-V+bZB
zm0+KObI8z2@*y<a4cm11MbE1?Z@0FnLtXR~{CQhgUw?p>el;A%j+2%vycbuBl)dvi
zGGaq)J;RL3;c+I<4E*dv96Pv*DVj$Ro(X7nJpDThRR>01O&@w->O#n<OAJ?kBCzU&
zg4!2T1D>(5{3=xCpIH%$T7X@B44B#-HBQ@!JU72^eL)peT`<AXb5L(6bB)Muynct?
zzUg8NQCZF-95I!E-b6R!QAHlFOu_YScHgCOiX#S)(RbAJ_RCwh*JkPQv;V5)!76Pe
zc(|q>SCS0!G^xX}WkR&pEUPW{eViLZ3Vr792Xt7t+4NP3XuY#WP13`|2H$umn-7|W
zAtO<+eVDv0T{V4kta9r7(W9wf^M>W*i+2|F(+v;Te>4|43lgQsX9P72odyoAyi%qc
zYtk68LAd=Mpqj}MTXOql6E0fc_<I0Z?I7%_xuOsjxJSbkiC|Cg9OOzn4E5}Hgj(N?
z!1ddHZ`pORu|n0Z{MQC5dX)6T((hIYyV85FTqb~mUo0CJR&>A`R8^(!B@obUa_@1{
ze=U_6`o`g?1TV8@mSH`@@iOg7(*E`iE={Ceo~`0lJfE|)aw>#^7m}*xD(UP8X%*^*
zIb!lh1FwwtB){WIDy!&qI?ZekrdmOQDS%{;hEQk=TRQtCDa@3xzLLyIlQPQz4#ZIv
zw_JE%ZYiKNn|5ng0mOox96hZZ{IAZ+82=ARifou@+K*~oHhoUdZXiV~Wyc~D-XNVz
zuK%+XNRjp?Z*<I!<CD6Q0(sk=8RcRJ!e%HSnZ-k&_^CaHV*_!tw9ms&Eb?9?H91`L
zNV%*q9il@p<uNL!%uBa`P-K^wQN6lq;<&*#C8tzo_Wn&3nHqfmK@bj)8!_0+VJq+J
zfPn5ERiU>tk#8Er8P|WtQb!6$&WKmNCsjRZFUk0)<M@b@ID6cXHY95f`RuCK0xF-t
zFD#T3YrxIRzs|{=!%JYYaG!TrIm_0U8A8c(kB!*^rQ@)(%I;to5A_vThl+DmR}FTC
z;zUIPzvNEgEX3})91MdAvM{aH7dS(8g`2X0dTk9%ZIyPWl;C0etOCMsB}+Rr+%%tv
zvi}yAe_O2^iv6vx%hAd$l{iEmO4csWcE^-D_>J)QH3e_Czng`0yPcEXb<a8X(#sG)
zl_<o>b`~HxSr63Cfg$8gsmOAn&?-`BjNxd}D@`-5?{vof0|2rVxf+@eheEes?^f~u
z=KQ@a$OyYkLZ>3Dt2EZ}g*0s}ckoWXDIz{7pEkOTtCYPE@#aelX{nr}XS}&51}r$d
zp3oNoZX^9N7n0KIm(Kw(U=@1}UL?94UH{1!rO>Nr*YS@K#1VDEzOAAeF=W<*WMlQ{
zT{4{VMIH(FmDFDARAYsvT5CtYwrb6C*@eox**TG?MvQa;x*!syJXl<A*^S&!5e%9!
zQd9-Md>F5wGlUU?-o)mw?_~Vg*?s!@xxm8K_tZz{EJqhqsvs_A+%S%3-X1WFCC<F8
zSa5tg?Yv7vW*FWCYHV^peRdCzbKDmo^OEaElr;~pDYAzC7?_}AIS_$%sJ@M<xAH2d
zvA`u07%94&cs+uL|5vM7-+Ttv?Tr<HxNymb4{9I|ct;mPN~MEKYe_Vv)W@$$srB*k
z6qw7&e!u$j{c*#N_Qa)xea3zJSmc@#F&YTdOkZ?B9AEwJaUeF-nz}Vv6*CuC%agp(
z|EHAC7+_(Pe1{Sk2641b7Kd$dwEnOTNxHHoc4I+I_&%M~Iyx*PE$Zm+5KXj19AuM!
zk8>N&SiSh@Z?w1Kbg)Y{gsMlN?(v#mXqhZ3iYCGvWi~-0@eV?F5vt0qgV_Iu{Nr<~
zuXYUfF>eosMQ?D+)<R}S{bIZJG(DqAf}TH|X#z;>UczLGD08uCpi;Rnh9W|cX@h_h
z1H^#e-?gxpwDLe4|0IaJ%`hudJlwOG#LK<<_^H}ZP3eB8{G4<*Zcl|~zlZ5(u@uD#
z8rZV0{L5XYWDN9|mc--$XFWr7E>zo>!3F_Y(C-Ged~K~rQ>|?2fFc~c=nh)=_(r!A
zXczJbH~P;DnSsWNb~V~Sl|UkuTITrRq@no8>r(HuzcdY>zZOzglTWqoU+^n1_NNZO
zc}!o<nXS4Rs}nLrJ%ffl#i|KuyKXWJ{hckc|4qXg-txJcsbd<OsbykrO4B6y+TILb
zA4!O!w1%j+hZK77)F>|eTb}W*-mnj|J$K_xHp!M#(O<xhg01)%sfO7!g<ZJ%&*Z%e
zRC{(7Vs?Nt+u^Yul9&c!G_htw%f1;_z|WQpkrBGJpeL!LoL3uXYy~nwMGgwjhYq`Q
zmjMor<uSjosI7K|6Cp5nOt5%4!4_KT?Io6&{+9<3)n6&laM?X@dw)Dq=bRh``K30y
zx_+1I9X8Bvh;T|&iYQ-y#{ZC-O>wQ${;QElkH5`y+h;U9sdCzcBeajqF_;MPrPNi5
zx{G7pI7TWR7{4wDlfQe3DkT<ADbt;q{}gHh*9~T0N0O<HK>N1&r()`;1>Kwm)ph~G
zr~y+agJsI0_1%*3bO8%Hp&LQV`T>r{lD&s#y-)kUu5dSoA*-}kaM7JR^hw|muEltR
zPdg-8SN2eg1@!IhA9zO9c5e$^z0J=z1}-paJ+0}N`T6>rlp_R|XeKb+^w2A5G)VLe
zCzd5M{JJq$%~Y|S?dmw{w}1G5x%Uh|eJU?0Zh<veYqVPY9d4L$^WMMPT&eWINoVKK
zn;UJY@fEw9d6gkIX`>oWT~0WQ!q~Jh!|3M507!;nQTcK0WarM<Z-PVtyY^{Q5-_?C
z&mjWbAa+u~4RV|zuT;-z!s=m$^@+MZs%w4}yi1z>a#>4T(&}friZD%}&C4eLty5O-
z&DR}lWZB*bDkfSS8LrTRqSBk}d2Zlp)U>pGxtzR{t%WB7uGpi{kOKEE$Ekq+X*3Fb
zWbO~npI=x#`*45%wn;D8K0HT=6i68hAHdVT2M1f&uB?Ogs_BA(r0NylE&)vWA2iMg
zziL)6pSrwdOF~=yAU_B<+1o2tC@37D@6p2BqCibAFJm7;+1Q`M@+&TMvZ0{2{<Jtk
zG?!AKIc6W#5&y?Ek#Cb8-(70C7oAEUu9pqSJGfLL9?w}Hl<Iu}5<r8<Z~1DH1%2Ur
z8x``>pO4lfV`gA|GXn8i0HlVDn&VQj<H?iK&W%m*Lrk(D1z(S|A||Kq7HYaAY1`p#
zd$BOxWvG8b;qdrA0x;j?KNOg$5AeHJY1POvjMD3t2}siv=3J4CZWavu%!uIz#@{~3
zpJ~Wtjr#I_S>>HRl!(LEp{ct^OSvw!?BAbu=f#B$+k;v*bSbJfG#-W%AKE>dc~ifI
zGfW?Klb~ttaA_M)R9&eRY3j{vDEs+KENU6ksivZygJB3gnGcW>zE7SyhpWXmHM(yl
zQZM56u*iM8c;FOcK{9S+8pwsPgLJUVdBlv;c&3KX$O@~Q>jZJc7<_n`b2w)Y!!r|r
z3P{R(jUP3X9BHC1KL>|H_}SzCVSg@QM@*jB9cQR2QO34qfUIq<_v%jmFlR()W3d(m
zIT4Hu#2yfDo$MK;6Xw!nA#1-X&8Xz=uJhwfm6I52Yp4#(2-Yc~Mbm@WhKf};`%RWg
z#w3oX;fASuT<Oyvvm-*XKd}VSiJ_0Me^7&^rycox{=&6J(g>qPim;DHn<$W@cUcfb
zN>Y%CMCxZrMSGF#9(N|uN$(P|S2%cZyfYOUU_>)B!bG<AnWjS@C<E*v^BtUSigz^k
zsI-f6L=`i_W^ztm23h*yKrC$#0;T%zf=f~rRy2n^Sfnt#qNPB%7$%fx#a>=zF2he0
zQo(fHV;mi{=fzh5S0+rX5^RN)7FBw|7H#7UypR}2F9pu$nc!u4WD@Pa-Us&_Ni5sX
zOdvOO0~`0mja59r8~Ul%watnvnut+B4=3T*EzDq_bX3-k6{B0K+okay7P}~T{C@h=
z9MU%Fsb;VNn*+^P*fz0F;49!0_G#&q!Rw2lVJhMs!vy45{W?)w1$*M>4+n{2)e-=i
z4^Aoy?f(YMMTSeL<Sn@>)Wj4RJ2l%P%x^wzae8#JyBnIvc>;Ng@Jzcsv{_tm)>aFq
zYRq%LRz8O^mR8DCCntY}#S5DJLz}U8)!7H29r2K%8lY2vkDO%kKf}W$Wz56y#}kO^
z28NRfj(N+)BFkkHtn}22Pmtw+(wZE5(Ba(Zf}MC0p8Nz%56fLAg<qPaWVQvcqpUZB
zquPg_o=({+CIzT=OHgnIO&_C{zjx>Hc|ro=u>Yy|9|d5VO+GHXzb&_bpXvYJnk~&H
zw>9I(rsEBHHpBM0R$8OrgzG5Qf0+-*@0kw`C(T?%>o$X>YuTnAJZ>+5Kxw(_rSMCd
zl<X+*gAPqS>DIvdJ|fY9P~3$_CoDANQI%V_W7sJ9sBDQ$xGF_>!ivo7&b{+1C-B@q
zGGsd%Pf`sku?E6>m)2&52+D|?su13g8C9Y%UfZhkY)Cs9y8jwp^ASWgK0N_46xeF&
zy@T*RzNoXx_e;}|I?P8&q?rn4e3%<nJA=Atu>!Lh7>ayFM^rw`XY)PH-vy#g*b#G%
z3O!-X0*2G{BF~!m<*;hOFj8wF&8{xT40X~;%sCbneu;2E)7@mud7g6Ur_&p=49wo0
zg;OaUqm)r1S3Ah<1dLxKq_8_c3!%OXLx3T9ka`wod2dS$Q&92`BLE|j9}o?FA-o~-
zMZowF_Ye=k04p!=Du&4tvfK_S#7rN1Z#QrMKBEqr+_pLFkho#epZY5KJCI;j$)jZ<
z3)ZyBK(y2wvGFy6r%@2Fpn@Iu_en)UoX~#hub@^qIL@e6S+BC#2#KOc0jOZe1!k}9
zBt|BVS-m=?M$R0FI;>`@qA2;mPK{{oLV9=<uK{tIXTsC_J0HPOFu(wD`&Js)3#uov
zTF>Bv142OFv8WHA1*3fPSds_QGq|`)YNg*N{s?1Kbo5Mzscat;V5Vin82M~TzkUH`
z2ODB3k2sI>wW=B>=~QYwHLjdIQ+sdU$u;Tb{aB@PHr?0K0_*U=AfFHHsM4CxhSX^9
za~V;`4{{S?5B2yPauM~1D!JeyDzuOj+<K3L(j8GCxeG#H37HD4UH=av?pqB!gg+>h
z2cTwZM`ZvQ13K);d@&NR24)!KX^)VP!*$01b8}u_z+S;^LiZid_G34*QzsX+@ePtX
zp^dBxWKpHReu3b33ppUO(%#77Bq7?v&EhP=6f7~=C$ASyRwhT-eN}BW$AhM;B4~AG
z%WS%161W;je)eP?|BPyR3ZLIw&yWlj-r<o_i^yf-UFL}KM|S+wOk0Y<>$whdffkx4
z3u>%(gq3K7h~q{w=IkujYPom<@^6GuHlZLeAfC~B3@t1E*LX`a&4~BLV{iy7)pXJt
z#W=q`NAJ&10p6Yw1d2{tBe|Ch@D7N{`m5Urp-f|{y%ERDGp$7$gjXKB^#$d;zvCCz
z*D3o|DU(ZSR6f5a{;{1WaWnJNO_*?Enx*%*vC7$D*aMB00&!$L_y-Kt;0@Y+o%W{t
z7(xbcAnOA#puY_<X?O7KEzE@G7cZvsaQRMQNA0>+R!#tIPENS{6i^O<JQvCXC<xsp
z9c=H-=`a3K<*%y_`W^h?nV{zf1adJUh%T)5S`GR;ZNL<VJU`9HiV5InCae@uxPoTg
z187c|yn{Gi!n5L5(TrAeoK@+rjX@%Zo_t<y{d(e<)wZsBD)Vp>P@%a+EY9ywV8>nc
zJ6NNHi_w!R^A+%$V6ag>Zy>b3*NVGvs{xY6a={QF7VQtKdy6+Oy_LlO&4hA_%0V)D
z(D3?m1$yu)$pT08EnL6?1cS=ShsK(uChIIW3ux86cf<G#i-n=ifl`Q{BBGACR%4q!
zQ!<te%{U%!l(bL59;~m0eAkNhO8>GEf}vzH9(p5#6f#*eLPb$SWqd%&h?2JrTfy6V
z8po>PtAw+EYR_WhN?qPiO{Q2;#YhCuO4_*FQX|5HRrh+P5bju4Aie;rRo>(uVq}SB
zirgl1T&eCM+%K!I9ROI4KUu7Sd=Q3cuiSYnv|MJgIWh;97{mW(103@^KkQe@#kfrI
z6LUn!6ct<2D4LXjD%}yS*~TLNS#0!(EG`@5&|wVG9+sHtIy1!$O!&kmQ#v+PeU=DX
z02tpK_Einz4Zbfll-Y+{LQ2R%{Ptt3psp-nE_~X9$%6a-1~gee?$+IyQ0m@=Ab=G@
zlD@!s?(=N?`!1J(nKUt3m<=HZGmX5tO6kt9)?Z&$x+>f>1IB#lPG7A{@!M6B+(LF;
zKMyNe2F6Z~L)+cuYe5-ftyr}={BD=!YOMs5%=dR=#ds(tP1ei)e8Z4pzm_yG2#N}Q
zD2Wsmmde$KOb(ZuZWcT5EO>+DGdNZ^EnR+yNzHWnc72#$-@b-^^k{N~3LgH7)pb-{
zE|*9BP8dM~c-Z)7r(}3jZ2MEIze2$FOxXP4bKRMyUWA}?L2nzXu^KU1Hs?$nCsMHA
zY_+S+GUGq2ce;a1wP)gdnhH`28WGk>yFKuLF#rO_#kQyOYn2n>3sest#6FUO+WQcT
z;~MaM<BO7}DZG^n5UuC@WuFEqBP)yg=@TR%b0jgh++hrYwxy`+#FmzjiJkBV(V3D2
zh48KhBWn51*@~ck=<zEhFq;UGp!b`p&2W8eyp><9Gd^@?xWyT*nO#~6y)>4>m0bD!
zf4KntClgP<V-b_xmY3WAIz<9o4c-x<X$FVb&VnKaUV9KQXn_9!NO}7Z9uPeZKnf+0
z3{qpbZd0l>Yl#2nj%SFQH%D4!hmfD*P&7nu-i7B7V^B}B!^udGU0W-$i9I+P20u<<
z+u^ZxRf@zf!uM_kM^X5e!$%**uH}eo^|?LU6e>&%g&%WD;05vZams%)jKFSSONuaV
zjHwfom|4om1Q4A^%<3$Hw7DoebBU}ybfZq*m#hsbjNHq-8KCeb-N6ab39E#Vkxd;S
z?ENdC>~uFe_)JDx`dPS+<!&Ubp1|U;U7S?e%J$`{R#(E4T=Kh_1Zz{%KsL#+%_*vj
z_{p8ZQ5ir2xZ(vaO|-o%{<n9L-U}sLPe0sV`~A)Y`XPX6V9V-k3mW1LVH2PU<{vt$
znugSf<oY!eO()!5KPwJgo%{5CE+ElvLycv1GcVp8*o79+^zePWQqnQ-{KpwXEQqh4
zha>p<17)hS%?3|F@SiI0Wd6?E)99EJcu)9^cY_uPVw6h^*o}}Gu;0HX9muPXo#x&G
z3mT8&R6sPo`31o|e(9sXKnYc8E^KLU03+s1(mH=AzBzKC_$R{B9&uT(ywM$Sv`R{S
zXlm}6T~&qczzr=%t*>|cD)w&%n<LBqfezLS6A_wbOBL#(+ArU4m&5dX<--t6>LP&U
zho`EomwX=Cx3e(5DUY=TWhEp0XHvCBytdxqbx>>L3*-?0zXQ0&+lo&Jc5zMifI_)Y
zMktM#j0W|cbn~3R)NqNwKBUvYVNUG|3khE^mVqG!4MjQypWO;`|M_9a<Frm{=C@yw
zKxJ({g!fq^HL}nH>f`JAd1v%;8=q=Q`c6+Ua+o(PHgK`gFu8x?0gilLl=I}H!tk2-
z4kvyC^z6gr$>s3sZ;ldnZ?B?)uG@K?f-4mMQ<jv}^devHb7B6Xp_{nJLyy8=jcQS5
z2UA@T@q(|oS=@F5Ddj4cYtEB!&Z^2k*1_P7nd#|(q$Kpm@7v~izF{LOC7a_QMa5-b
ziM^M9sT!<iH#CGUn}L?D_Pi(h+S*aUC_*06Vj1lx$!B*cpMoDB+3J8v8L_>m0O?Rs
zVwtX!hIKy_U1B-At}=(xI2WuU53J_XFhYs{vA=pRYj!%Sy~$v|NLWrcV6+_J<y7@M
zNODJr^AOEG!{(K`Lu}bpVxY%#=+`{=?%#uzRs6a2kF;ZV5GovO_c{`4Q_2%|WMih2
z-2zzUm_$jphaZ!zXo~gPa=s!aQ%*m)rj=N(i~p(KytFQ<V}V&&J=P@o(Q`Rztc~9K
z1xN>Cs5@i>;L{fDi-`UCs^z@3wfzuMH-e(<!g3dpeYGa3fr8b^Nk5@tVwtUG_#pS{
zbWN^ou}}7+a3{y<%P9hr^ph7wN^kx2{0!;Fi9X3!*(itg)9XEOII)islmA6OiP2yb
z*545chPFHHR-esRaGmLxx?B#=hl7dS*fHjI1`4)>2M5pu_n7f>AVgNyyljC@FEh?2
zj)pFpN=(aRnob{7H+MQ?KWR8&k_*^K<hJOfO0!EejJd<)m9k}TUI#ez-)brlhA<%8
z?uiJpH^8NNNeL`Stazwvekj(y+P2(C!$?20V&6$b_tr$&EFnOxFW4{2OzcI2EO?44
z@LU?c_j(lt?;)s9==NZE_0kpwz3Mf_C=)HXuyLWI%KY4@YoO_i<kz~HPJ3id<;gMj
zm8sROB;Rs*?ihxEb5nh+o+(;`Qyz3!b3Qf8JC|mvH*`PVGtKeKvAOe(;b}^W#N3%3
z*ziS|c%lzC`ZOm$d;8OcX1Szfo;K1Jua9z`Y+iecc_tiQQ&FSAiL<$6KmFJQd4tV-
zhs<)?_W9&4W4<`o5{#MYg^SuB?^KkIc8VL`xRo;?H&W`6K0}Xc+^r)R_&m!bn`#Z>
zX|#%bKli=21|J6BV`N4c)u%MwDXjXw<(cg9BI>y%o%okRN8xwRQ}DTciZ{rlDAQ_j
zI4vBEwhb&<=ZLuqyvt*P3Kd%2U*rHl<sI&j|L(AsOPeL`6!OJWtO}I^gOyDTR(aeA
z6|u+W2VEX_CVw-z<uV%j<U*zRjTp~J$0A9w<pu@>3@U{llqr5iJO7J7nCaSTWxip^
zMP37aHD%KDG^~Lp+h2?JZ!Z_S$%@-|EA~TI-09>NpSETMc(#YhpCVYAJa&Yba1pmF
zDr=C5+@I@)!r(;3B`hN`-&V(~ss<_=t;KY3n6)wCeLX<uemFBr`%D|2aRe3Dw5317
zP&VEmcK*coA1^N#3$eA^G=G-#{Yd$zRid&lu}CyeCl@$nLTc9EE`x%WiAnN0jZwbD
zkm80wDEo(mJ7kNTVCR*JJc#bY^JsmQkYm1m_x0JxHFK6Xb3pPO)z!Ck&IdJOVv8&;
zt=2YJgU<;|iYe@f>@3o_WPxQ~2~OBZTndiG7EhQlBrDtkTJ@6!nhD<xU9v5fYk7Co
z>qE181FE}As=1lb;J*0zLzFCGpexgi6W@NCt>f}*xDO|Y<ld#J{~_@y_Xbu%;r^Ta
zQcm%wqD(bYrz6arZ;!n1;8!jQSlS{Tb(K%|$JRHSEWkh(3~<C?>qW%It%|Ti_@I}6
zZ9}~=uKZhfB>cR!$91w$Bc)|(TJ8RkeK>)g4o!R6UDEC6ZDnPpaqb?rth7r3oQG2=
z@8e2kO@3P&H02;i*ZuG%HE&2xA?mCnMFfxMPAoTAx7qR~EM!@%FPd@k`4YuH!UGXb
z_{>?lERtg1=lZlktncr9(|p7p<u{U2v9Za6PgoT*t+8>gVIDD<XNh<MK<Z#UgLNg4
zjrZZgyzvOr%QCA;{Uu<jqDh5prbty(Nq)cl<NoxcnC6R(IkDi9S#$5k5~mHpD2p63
z1uD|ev@=E6qA1$TOzX#;KQVII4m+47u$jW<LvbRm?Tem~o2UV0#Jg-{Ro`277Ei82
zUo^>Q!UnRVQ%|Ug^d#Q1O@=$gW!vU=(Tt6Acg7pUXF02es_=e_;p=3*!%Q0hn|^{Y
zAH-`vJIZi5xTI&Foqq8wtmyBZ86aaL;S>Ecb9%H~P{YT*2*e=5p&SRud;VmMmQXSR
zr=hCy6<Y1qP_hF@h4C*2fyG0t$kg1kUEuTZ<wk4eb0-2Kuf+l<ne7;m<1WP}Jdyte
zDU9rW0wa8)gP5p_gsCY>xJrG(1r#MdZ^2cw?HYE|1a{i!n5@`~<jMGS;>0A`Q}0W}
zJU@a<I=PVD&yz@O0(rqi&V~A_A?WU=et0jsGsn4n)2_ZC+SYYa05VWh<Sh0f)%kk!
z&);NNIzF|8(zY_<#$>_J>P<J&aAda=--x9^(U1NUgHxm>W@V%wD=`Kr(BdTp?jVLe
zbkItt`1#IV&B;*BlChc1Fz-8L<<@c<`fKn`4KLy1^m=zRs6KkElVblU--je~3^QIq
z02We8y7hFy^fv;oE8kSnCJy@^zi<*PI)^Ok%2GT$D(9oqz61N4{p3Pw+0BRQHRpe>
zEuC%Ljrz0a-_3EZh`_{Mmy$N;Y<+R(W!63Bu1LP4`MQc6c_Q(tY4Vv7Nq^ZoZl@ed
z(H#c@!DFpKQ~F+T2VD<jR>I&a#YfL-KH}z;YR#ug1H-pMGEIUkR@=W&9T(&55525X
zJD{nz*c$(Tl`ZG%;LW+~zjre}H@(7+2YT*X0WT;$pK$bH<EdMHWXu2X+tO_kv~~Z~
zdec~~uj;nQY7!lEdL<5!?)zl=wVj}cG0l_{NrW>TE6vmv@(Xj8+^u(KuI@QHpoU|j
zD@a#4hM5qQfvB~a$eLLs7U-EXsRlDG`cjDM6|OC4$85z-$E#@mj5Z>DQwkhCF-B1T
z-Ca>glLCqmFUB(kT@H@!ZeyZAHG9)0qPkGnsHG_MAGb*^Rrg|V{YM;#ai+4b9|4l?
z$2qx}YEj}zNZiea*OLrE<BmN@B&3bA4G`1Ut*SJp0bDHKf&Hn>r`|-RS+_-HIKmMM
z_}C1gOo3Q%G@#rkzLc>G7HKU&i9fB^xbA7t5w3uv28$F&ery!J#b!f21m<EW4l>>d
z`E%5o5%I+-)EGO4J%uCFpG?1A6j^qvlf`(RAbH4!F-aWOW=JRt=QEe4N4*JP1o(0{
zF|-;Vfq|xi&`bYo_Rwd;5MA*mmU^cNzC}nLIL;R|^HZNMZcJn}1=9CkVsXTD>7qT#
z>ttYDZkLz2M)|*6VoIL7SEQF_`F&`iSPaZR;?#f)2bYri@-BmF%Y@rvRQq6gvNKBb
z2r5N<n6(43uHod{l4<9}0Y0D(30hF5rL(hP!}NEf^P*9h2&A#b*I2-23%|e@bs^Q)
z&}f$H=Yif@d6F%wLbSFf5~M-?-eif3(sB<Je^nvWLG@sbUf%h)p$=t-)rH+HFz8Ba
z_rA>pZy%=sb!Hxp(**B&D2v=zT(|hu3X=Kfo%bjr1B;GngM-Fvbv$pILKek%M?ua5
z1t~&{xzri~SfxP;ow1@nWnhxgzn>Zp>+PawPZKG~qYKy+MC2R)V{w~oODq$=41Wpy
zE(#g>&xg!$9J}UF;x=IX1w@beaxnJeEu~GY!xXYeD!vc7|4V-xl3`!Y^)<kUTyJhN
zTEMTuj`3`A@LP4(f1E<lC37MP<6Sxxq20AUO5zwL!4KK*JYTZQ00oBkoIr7znQ#r*
z8iK;*->*I@zmidd@~1)UUg^`0&&k4?@c+aVZP+8P@7*_g-OW%=jeQ?8OM%120*&JY
zp-;H%Wj7H4H0xl7DL-gk(;Z2o$3if+?pHJQN(tN=nXD+-Y0&9*cHbS|yz!<~{dQF^
z>$sKEDz*$}LtI1cyVL@vL`D^<c0{wPh$-J-u1Wt)BK-)mwUjKnevT*_qYTa8>l-E5
zmOr;_6{VP2W@@P+%M!M~onb^mqZ;&N+|~|pkASmGWM(4i9hBdAe<-T$Njx~Yesoz2
z4=dSM$`Ck5{WBzwS=3HF8~zcAX!w1k5O;cBN%vkwQf1QiL?O&2mNZa}4gLa>e*wZJ
z5=1-=DP!LLyzX9t*wIRf;WM|Zyxd2jjfAkf5w5Pc7w5}EjQIP?d4dJ|o{-C(>Bjyq
zj*e){{NQs1Gk_zNAB=Qw7zl2O3v_J%5_(~Eoa%gfx%XW!lsIhJk|gvS*m}79nclZU
zw45U$h7|p-8_20|V)G@rzEA||oE|Ao67X03@U7O%WyGBU10Hj&mu>JWwiAkFE=zC!
zB#{ZfpY(X(F-*{^u!}&d`~jatu|;8|ZwF&oW*~%HrgBgm)5jBh?vOZyhZj6w*%VAg
zF)1Lobk%8J7vQijIoA^nEM}@A-8JkLwQmOih$toiu{;fI#QtVX2V9>bCMZhqW7BAJ
zBXz*YD4sh?02oMFJjizTxHzt?+45K5<vnU*@P0G}-UQ9*2XVrR={&>iycDuGps^(*
z0%)EwzU%a_EE*#8WzWmK@mSig3?9dzRrl#Yn7eji6u}#%M<cJ1<xDI<9812%oA~#k
zWvRaE1fA!=Dpl?`OpgZR*fJka<Ga8Q9i?+9ot`{Ednb-wwB4G$YB-(4n|SIY<`5zM
zJEdS(6_?jtD<yj9kX5WNV&b@s>Iku(NGfeMTi}Xs=J=~DxXp!_sJ+pYpNq?0LGAwM
zCw`Z_SX9p|7$aHPY@4C+D5FWm0hVA`I0<7(ET?pKIrOs2lLH7l8Ix0ZAds@MAQdrT
zDNuu0Mw5Bwznv9IJv89?`8@7@^|6hwp<b&_H}Ha!BJp)O6b$_(_!0xE^LcdQy#sc~
zdt832-_I}=KYhJdn?OJ@xcmT$--DRX)d;o6(O|ks#tS*Jur8nPy6=HrD@9_&=Y^?M
z4P8<LIj`aUA%_}BmJdB`fVkC3nTV0LT6Z)5h%@nm266Rs2IX89WB_nST(kT|(XTT%
zw1>+xu|dyhFW6B{O~1!<folRwoc_cTg?|4`qbH#RIT0a1q|@9eqk%%i?srPlI)w;T
zXJ$(-*8rnmgOW@Y*Ix{0Y@|O$OU!#BS4~P6f8j}UkNN~l3<!><-ZCs%LbmmY+zU;e
zAAu_=KKOqLOF)2n?yROnxbXCa)fV~P<TjFkLPW&>{pn@eS4a+57<x^R3b44v-4)|8
zGrXoyLiCbHkKplXE?`ZfUcc$H$v3S%E`g${Dm&M6{HD`^C^<rG@q8@r@?VD|vPV#m
zWBR_<D6(45Gx;`w9*nmKB3V8%ptH)+<!^a-YT*gs*H$|U%5~#$e#nKv4}vQm8oNFk
zp2D%PKDUO}>##uEpJ=B!_3k)aw%)q*sC}N;bE@tw@^o0-3g7AipCLFS^VPgXzrBQF
zy4!W_VFTz5@-={@hDisT@2<r-F3}mM2{z@#aDxqL0@*y-6>H%)G=#(r!E$|7h(8A)
z+R_I+LL4)WRY==2h6`tvaGo_S<Vw6f9`2+7S8ABU{MpCp+u5QXzzdTRx*;Sw2ZDML
zEbQ*IRwJJPX3#(Fg6A(-rl8&UJhjGHLxo%b%zw+tzv|Wnp^m2Rtm2?FkCiIE=0_Dq
zxHk4L@Sc3xQ^IXawO$A%<lg%UVy$N}ICE~E>C&e-y^72?Ehu@l>-gJ<KZK)NP-25D
zAXbH|>{qJA{x!VQ`H<fGJd|1JwRohyDhIK+`@Rd@=zG%(E@IE8lf#s@>2f_H*`4HV
zm@+k2M*cr8$OV+hIIu}vKBSJQxsc+Q%H?RzE-nt=f5F$P^Wlx--3n>J@pxS_cQi`b
z8J2fBE-(GQlx=S&-=4YJC)FA}m%&Zkwm>}?tJ$!px<1^O$Og3xFfSpncag6JP&|;s
z&9h%tRC^lBO_nAF8F#J8O|h-}dYw0RzKmy{8~Fdp8bUi49E*a;e)9tJAM2W~u)Q&j
zo|DV`G7CPlJbxjk?o+mhkG-`{W~ypi!{qUZaNPFYFI;v?ygr4Iyrnr#bgy^E^ac0S
zgzh;EMP1d%&FfCGi@Z0rm5hXSUpO8xad3(}$I_@kEa*>2d4QHYV%5qza(%i}THFE@
z{J}vp`t0218Kz~#0UmCD5a0>PgyuF3;&clK00$$W4SqtmI<2);0=DxerL%ly4bGlY
za9;?Sxey^=SCvQBI5%dS-84KLkK5WdoS2&hpZN+TvU{@Lg7jnZse@DfuM*k~uumIi
z)Tu9pPpA_q?-k)v0)#PwzDr#7lZmWiHfB_bhve|yprCDrfYE8fTErx(^JU+ExlnRp
zfb9B8@dnV5eimA4<;26E2L?TkLqr6Qd5FC3Pe&gAZcQCu5V3wzI%%!c#tDyeh>~_L
zudJXs_>i#!w|;49g~kg#qk5k=#IH&k^~B|#Wnh(8l=aO2S}{3{oUCTTOC&?xOV$TO
zj&;8iA_jvx-7aQYroODU3mDv?Sp3*YB7Q)ja2)Op&u?<G%~(v#;}(U-$4X7T@ap(G
zcr|KW6PgO7YCxUV4dr^=-u@^YZbZQ8JGc(bVdhXTPS|U-UKn*9HNT#ZYVHe(%}h<D
zVXo<cIqzmZQT--Bq{i+J2S;VF;_F_e-OhE|R&q4t2au-HJTH7tt5y5x&k73@Xi~&z
zPIC+`$8?^nmqWz#Q!CC>f_<Znyia9CyPm3($8(iWuu2s537X;yzL%jmebu%IEa~Ij
zP|b|3j5n`;AJaGMW&N=)uSc)F#P-C=5($M=4b1btZ736wi}a!Xi*CvrVW`()DAZzl
za@(i&T#w+Gqb_QBnQRvXh`(&0arcK!R0z|Xb7+Xsgjei<4M4CLp_*&;vt~b@Q1@Bx
z#GBK}R%W%df}V{s%qSRO%$UaD+ii)Yz+K!%;82uP9yv;JjHcPGBApq)PO*^j49p>u
z$Q6S)MtZ|WgcHG}vOhgcBsKrWVOUn=3<HmF)xT_pAfcd4q5Vo>?zsz3CY92emo^BZ
z4u{jW%bBEkO0<ND%E;iB1nL~j(K<;G%Qj@n3QuMZ#Eh^bos2?=)ohG=-=$lbPk^z9
z5=IVfxM~t_kW5Mar)EiXqA%1Ng2x8XX&WG>15$0$JR^5e{`0WUh0P8hXX2R#f>GXR
z%>=8pYrinvFI(3{h@uJBy8SDZXu158T^N?be;Dww_OR{5<nQB8cjiIc=Cw3eQ8YN_
zclGx8<YcD)*QZ}`&-D5hxV%;r9Jh0&JL|f2g^GAsrlw^z_5qY{2&EmxxW9g_i0z$q
zPp$Jjrl=U=u6f+AxnGP7<>gyTX=t96S|DSh?^Jh{;BSB@-FLzFudnVL931(-f2&3b
zL}N4d9o89}z+P9@bm7MfJ<+7Cd)Elu?ItUTYG~kCog6DE>&K}m@Zx#@Yixc#%{fCs
zOa)An4oEX|3E8r5$&S?GRsCG?pVxnO{GPwD`#dq3QO6#L^HTjg_ZT@2#y{NG@N
zdv+9s@Rr@qef@hPhe&WPJ2XY1^x8c^k}(AQ$Q#1<x=V_6IgIWbzTDOAXKhM(tKJ)(
zG>o+WoOKz{LZHYibXM-%w9NVLa=p*sZb_^1bK3K-{4b;FfAWa!Exue`4v}mX2*1e8
zKaSFa_YWdn{`F#-oZj&Y!<c!zS0$U>iupBDOf|%BB!%&rqNkKysj21NLcEGxN8fPc
z_qc!X(E<IAW*BSZS2@<@GKBx?Yg`+hLA3ywM$)LYpD@s)=I5nw5N!-2b0nG&hVbfR
zjV!!7*I9<-NviT)_7WCF559Ggv{C&KGu23wA3AO;Nf;Ivoyb~VpT?zFo#y1Sw?*Pj
zw&6}uxhdQ^#-9%#Eh2fFd}@<GD?K!A4u9%n<dC@>#g5tLVgFHZkna~E24{)sb!HUq
zkZhMza6XCO$)862O1~RFj1VbKft=(`uN;ook7zy|y0*T=Nx^t}6iJL9Y*?Q|4sE$A
zzOxYd!o#fbrv?w0fkY7d6<$<ea$5eKH}-zWtX<*b#o@Eh+++HFc;O_D;<GOt>v-ID
zy5p-f6!pNBE3&MXa}}v*{VpPH1U`o&`OlJeq+wUpix}rk(L3?}55X>B3H@PCU5i`2
zg}icyo$NKyNUd(Kw>64CF$Uc7avb-ye5s!7pVrPr!I8(u)+VpGPVt0*WMffQBjzb#
znpdkIPUsZjgI1dBg|Bb0e)3Sz7K2NA0r&$W_ojP$aN$totUG(@1`1`&<sBYpn0$Y<
zrli3neut!|o%idc%`uE^1ANTdYj2P7{!~8ChT%kM4mw4K3n0CCk3GH>_shX+7}ao>
zzj;%EBgjB3dinxKMh&Qp>-AN(mkQGywvYkIvnY*<(4|>o(2bXHw-l@GaYFedjp-2i
z0~O8IS$8F-S_Wp{1|PK?0q^!N(JQ_F<L}sz59mlaL4h;2WCpvTL{i(M`*&PUfVH&N
z8orvqqRL^5j-SQ#wK-O5dF_`_H6b_)FbfOQq_$;O?&1<I>YcsMzeF8NzH4XsAr$CQ
zrvl<;pK0=*qNt);LMe1C$hIu@vRwa8DlP;NjzjfBk=9IwUnN-{vc4=%cSr~!Nw;PV
z=bl$H!-)suc}*hwYHZW@B(Dhf_uXNuMwI(5TZhj(hwU0?=mv)rwx7qpaIgSdF9u<;
zBSx%&qKwOF)G9%@83*7nLOw(m#ktf>+njubAYy=K&6Fb(RTfB$5Y3kV!6j&>pSlw#
z1U)5ijR=^CZfmv+WeQ~IrlYEwZPKKz-#+aDw_SSb8aZC{!fb_LD8P<Y8X33qSmZTK
zx_Q)UB4fBeG+Z;ce5RP5PIBs4EQ}@a&gNf`WjjOTJFltfAiT>R)vPy>S$+~yb9i72
z!}F>4Yzhsc_rvIshH#(3Fo4T(PUp+@8y4h-btPnD%H=M-1AN>`H-n91U&2>)d5{A1
zmu}oH2~r{Uf=SG<>fVj8sU~|7r0@dcjF``BTis--iVwlqE>f?=GzSUnIEtXK3EyKe
z|1AX_zfWdy8Bw3GgyHwdqr@$N1PEd+)U@yj(NPKXU}={l*9(t+C*q|=-hjh${}nW2
zvAyN(I0AI5<z??$OXp}8@zivMXu7Btz#10Ch_u)ZV4EQJ9B=!xFLXxRHVN>A6K*@W
zTbZgZJh{8fI{L}wj&{tbN#X{nVXb4k^YM!MY2O~w!*{Z-YN$1T0-^7YlOFayI#XYb
zPNQ`3yUjj0t708BJd|7FHXnF}vt&*IAvx@gcW*c2R`>FWfah>Cc)N)*O{v)1cd^Nq
z9!LOpT6G+?ES12DqOq5k1>Z%Gp@q6R9}E--aB&D(Ev;iW+TC2_u&ZXrPMstp<MY>y
z^@bD1Sl4tiRS~(^8H$inDdYrLS9fXieKmr^NDN8W6vyzk*^cBYN0lX;3j~N5KH!$V
zkc1+t7aVq#TP{aK?CYe>9T?=SoOgJy0=gMcI3;~NMD$X5yE(1OZ~UN%<Q2vKQ<OtQ
zvm}0yf%QlKG^<?LW^S#b>qCw79yd^v**5}IsIw>s%*#fVexEyjD!X;u7Xlj&?Pnlo
zPUvX=)D%~;BP>PT3#nGqQl}U!=WIMF)-*woq&Q4qj|{<u*H~Bsk?RSbU6)rkjg7cH
zA0Fo@@$gJwpr=*P$%;LC*;stu>8NI;m~pEa`dFiw9^4>poBvpy))cp%qksqPi+oSg
zj*8H1m*fu?qzlRAH^7P(M{G!T?()R`6Z8LgdhcjBzwdh-k?6g5MkjjjF$_cW5+q8r
zAbKy+jWT+#gJ_Wuy+s$jixRy=?*!4?@A3M4-@iXBmNkp}oaf$i_ugmUb2-SXX(@C?
z-DW<dGtld<vrx7_NH5Qlw6=S5*pLqare8CYbAy&1FhLSmH(}(keqJMu2Kl_es3c>F
z#zqiJD2OsiqhC-$C5s5NQ9&%=;$TPS>;Jp}bB-q0pB^^#OmGgFi*b^>`q_z9$puK$
zOuR#hg`o$|VP*xVh`d9#st3-B*6Csg;x8klk#-3id1!Rlr6vY8>;Te#WV(a1aJE<k
z3xq9+J)M{M&kc7iq<SmeXsHVxAR0gc&R(t{w^W#cC))*+*#EHx5|XS&)1o~FSsK`U
zDttRx3qrcpwAMf>MkcTD`28+bH^;&N?U`sUtv=GL%?l)J4fQWewgzvEWJ<-fxkUb*
z-6hKr!FXuzKShu5V5^eyV++*EXI82+ADtd93Nj(&3BRoU?jy|}Fwo)ufY}stB&|j#
zib{=Xsd(-8mHJ(T3vls8iYy2_ThR=sF;OA9AB%|B0==co-g>!KpeT^*X!i9xHV&Rc
zcZxZNU8Xm@Iy@5F_>Ex)!ikj@X%KGaW}{jua%}ovhkf%+U%uv!Zk2IST-)dS3*84v
zJ7`(8agaumvrrR80`G3a*tD7!Xsn~XUUWZYaaTzr^6g}Q9f><yz@w;5R3pX+&GK%L
zshiyW7uBMX$`_(B&3_80rp$!N`#Zw|W08A%3~3rhav^TSjjsE%)By)oUN_&0EgfhH
z^-+HQ#h(S7Gc@xt^riz8B{;Z2TZz|cutC>$)}WokhUL@+_KkWY)&(55<%78&dwZpF
zGkKQ-9i6W|Ot%yBKdldPRzzgT!A;YRtHpO@dXlFF4^@!xCfl|5`u@r8ZH8UceWjb{
zoX}-eoh&bw^kV1Or$Wl{EIP?k6zeO2m8Mx|dQ@2&C`nm<v6<IH_Lakd<|spqw0o{P
z6EbvY!Unu2q)d;wbLj~p5Re2I`>ofz4d->@>~AF=vGnf>tDcj-Wo0F33`K|qx=E>r
zxz;e4h8fD4d_VAWQB7Sv`E0rrR$1xovBbt2PvwF!H_F&VYLsROX_}!*)$7N#RNS0M
z@*e%`yJ1{Sqt8S>$FO5(MGCk5$GXH7DK-GBboHRJ|Jr2#ra0Qx0n6M$#DF8)?n#x~
ziZ5|a`eI|8qPLT;LMOrnsMxAc<v)*Lb3N?ds}!a5m_6jF;ep;UUd;#=V&*-T>c)$C
zTSiPsva9ua*g3h-M;&V_)!yc+P>>6YShh^<(*8-9m@6W030*v=b51PI6;n_RdBETC
zxcxMebK`VMTOErtDYmH6jW#Y^u6h$)P4BH|Lr|KDzt~JO%ycdBoFboV=mw~p2_gZO
zWx5jCAgavkX0nd+R7*&}fw&O59{U2ddO!2)7)rKZe_;G;_i`!fN@|{ZC#A<Tf@PD{
z-&qOS0i=k+Tt0rVKnI+y6&$($$*Yo@bcFG#*_TipdPhIb^4@?il)borefXWcWi;V)
zEw=n>fQiVdF}+J?s0uTbhY1wYVDKgjSpUw<KFUwrO#&oPbRtu`pZzp1;$m$akFiUO
z1sZ_f2NDgxmX^3GY_W7~KT>Ua5Ng`Z;Q#KS$iY*Vca<oC>bPL4-T$QVcl^8wG2csu
zr}H{sPfrPjmR~yT-(FpZDLPLy89BM{hO)h6gL)gke5yN{EB*cf2M3Sn?o&uP0Nq9q
zzkxSG<#?d{3ojDwg$=MkdSz1dYpE{Dl44r^-C<$O%<mla8h<*XAl(}YgXNH!?7T~r
zNE@58$_@>u^MR^OH_LS=ApKi(9Sd7ddN0TQ8SSW@E+$5Gc6?GIta_}##ipGL<tH=B
z?2LP}A83=z%u2&ym6*?V|HhX0L;50UzX$YB?bkTo4qyx~P(-;vtK<w7)opG-@fxZA
z-c}Vl(IqVO>KX~@e!l_!z{32A8NdKKdE`4L12r#)e8#J8$$&;AlpVj7RhrZ%Yej$t
z1c2`1vlr_ka)=T5LPCb7P#U0ba10DCHC9z!h~`um<EB_Jp;~{dTsaeum@YOWQ-A!G
z8LM>N5~rJL0}_!FoC!E$oo=8>xTQkLXD0vEt{A|vt9Yn4Qw<JFFRuS(9)7I~Wo_cd
zd2=8hFVXLTBX&kaa2`XYp=%sXqtHOyad6}5aY$ZSY3_FHrgbjf5NUZ-r~}A{*1LC^
zn#Qn0nr!Q%*|7O@)$}!rpmvY)Zb?@ONr~Fz?VcxZA%Cj9GQRhlm~7vH*sQ7a0n*57
zv&H0&*>fyR)jrd0{@7-=72-f`HwG5qI05s}-Wawyo8h8nA1zf~(13U(2m^BM&G@9(
zDHY=zuC^X%I5};e=2fXN^${F;g8LHi<twaL51>$&NXU`PZtFWFW#M*wfjcM6Fp^Cr
zNP}fM!V<8t3h$wvOY8~hxL?(swSKX=zqhT0a@iU_qrDZM0g8C-%09{`Nfpr~y|R-Q
z(fPN?2;BEVJp=VGO9kdnfVKau2OD+u%o?Lnei2239zQu)oe2TP#NgVe%qsFCBGePa
zQ<+?S*xki3Q2gx_iJt&IxNcdl)Uc^7{N=**`Cs}Ml|6^%)J2BL9jR<ObXa&u6zPL%
z*pErSYFvqE+9|Cg2CQl+r^dvuv>t4d;&IgC@9OJdp>F$llf|mNH#k`v8PY~+hlY*f
z?&^l)R<FPY9mk6HT`$-d%gSxLc^*D-K6!tLusme}KL{(yQ832MPVL73g+GZnCW};T
zy5{3zGKRSxi?d3XmmHZH2B2Hg0>0T1!D!qgFbh!xGqK<H!iri`7|#B5=(J7kSr#M;
zycVY_Xr(Uci6@q11{bca0Q56}QtM9{DkTiMV`==k14qB+Ekpn{i|%kM3x0gS>#|$j
z`g_v3q3`Re<J9NytDG-3kG2q~e%|@$hvjE~41M}(I>u8iaoA)5(W)_uhMW5hF5@oo
z+XeL8u`CUJK8~T?Q#cE;X5$7eh#^!yX;+)oO!crygPU$1a5=4>ontB<wa6be$zxXO
zy^&NE_l8C|=jImg^FPKnt0l_~TrXq%9w>!zE)d8s05<cXf}O=W_ZSz6Ly<MfYsJI)
zN@wj^;g2%RsoCP*k{|bP6-uir_RD8}@D)@#PrkXfpohN&<}=?JE%UP!ledD9rewR-
zScJy@d}fh`d~Cj-LfTrdUqDzjhW7gsNnPsTSktx1>E#qBo`n8Gc>H>dX?6A!mdvLV
zAUGU|&dP^FBpuAmOeuTS-5KU}$$tXh9JDkclqy|2tE1-paEm2*wPFoZe`KHehM_|N
z4h}#`go}Q{?^y!&yTxB^%&${F7uFK%8Tq+eAA8qW^>L+{f|k@@lJl#$RO%;ZAWA;g
z0Y3b%>YSSSfH$T8LG2lH<XmkW=IWm<H6$`5Yn^8~?iq7-z&fPy{PDvYj1N}=x?oKn
zD6xG{U)Yob4A68Z9lNhU<K64$FnJb2kKj~hH1X7#5DF6D3QPeW&6XxRTYJUHn-2zW
z3^oNDN7y}P_la+~l3n`~LRblb(~9wV>xjI+-``izn_m{G283otPVC|{o?7@9$!BN^
z8#NV5W}?ZWTFAsv8jx%1S2J7^nEfcSv*U{%^o$Ig!eci7^5-g;gPq*Z(u{<I_57kn
zgs1vdKd*|(Qy&4?C(HlYD?tL(8JpIVLqJ#^0O!pvP(ZGc?#u(?3(Ch@jTTaoN$$sJ
z=wAMe<%wA?t^dXJ?;TcBm1s7&`-f(MB`h#b=9<YYg~<8hJyOznnu$k7i@zXW1Oy3S
zF#kGa-tZ@_>y*7=(P$>qb<z2H1~wZ%^j!8k07!u`+8^6wYImRwnPUsE!`O>WgacJF
z<bYp^VnJ+ER9w5QM~9rG3G@P%6JEP7l?5N8CpT+sUUO!OIEEaZnJQTFxoryO-$N#g
z0Rc+{=y%(HnY<FwIjy;s^V`;E+pV8OXQmQaH-#A}j7g#dAz_|E9F6l|X+%P0w<ih)
zoLuW?!E1QTn)cRqM<nV9?;t2t^RU>}Cm$Bfs0oiPv(?8l8tY0QXS8H%jveJ@x1P%<
zN`Z_rx@e#OgKgtQveL#-ixFO5qqy%258iy-acBLhYxVj?JUY+*kPwa*H74p^i2>=G
z@^8oPqTZ6#h|ip8ll`42we<?PXapbe0t00N11)7Qy}-F~holQd)u?f4_oOsS61?^A
zEru`nu!(TRsM;S=t{Q(2e7o!!l=oNQS9zno`_&hev5>vaA}+#2dh!FV6bL0%t+(fw
z3zkqEKXz;)kZJa{iYOB?{=~KVaa?~ZT41ar%dw539=HMkw-b<YseE#wvH3ckX83@U
z`d5dU=7;L$lHXrYKVD{-FG<I}kgVA<5xmqrzmP(jeEMk%sEWDgKAH?wm(%$h0TW1^
ze*D85U8O}{x9$zN`K{42gL0B&v!pQ+X`8!Q&C^X73?1`B<^HgiprgajEv|(>Cm*j|
zbRBc(2>HLL=HR=W&US`r|JiwmMvssEkHBuHRsAxrx&^%sXCEkLf^v>urcu9nSMLzj
zvVy71$Be(Y`~BNX9)+bQoJy30q56fX@o|Q(BLf|@p+pz_O~ccy2Mrl51HqpyP37z_
z@d$CQ=I86X*7I_(20iB7PkmZXmD+M%AsOeYZ&By(Lx>{#oj2U5G*={KL_GCn2M@c!
z1YJycV(99U7ZE?U>*_qK^}9Dvw5jdyB!e!*`&9E>fdk!NaWA6T;<D1MzwexF;B;XV
z96eb|nW;QR!(jP(porHF#Txlykx#7sAgm{UxI0JoHPca*fOV_uh2Tn+f4mRGOiW6O
zY##-vW%bTV37s`30R*3reN4Ti@{GZLZk|IfLpL=vyfG@7XJGk5xK|dx4FbXDj>*Ur
z{M~VxlAE}ZMvi-D;LiG1i5&33aa&&d=`hQc0(j!~hYvyp!D}3;K1J$@-U46m3w5E?
zyzkewd?9IizuL{^q1m2HqK>l}49bf@WplBedo6NQ>E@;aT15kN%k=N&A8V_E&L0n2
zG194ZTOS^Ca*mWnR6Yz?WV}q0Vz46sEC~469NaRp&s(`?_T3v<SU$6H2gs`in;E`3
zK`|wetsq-`)yiS5@k)dExrI=?U#<mz#=wELz|!RQqwg5v<45r&sU-p6OX_&r`{r9o
zxpwiW*zbX<Zw<@Gdi%YSo?jLlmXj2`A6o_j>{=cDRL)qmnqWqk)W*hIk&-8~e^p-a
z7r$-yzvolSb8{so0Hd=zq|<-qJ%QzHsAHORKmOxk!8dC0+0=e(o6;LD(h7@jVFZP3
zFr7wm8A~$_@DWo`3$BGl$iOeC!s8_Zt^f5L3p*}ZrgD|hZ0Qrd7!BW!go{N^;QO>t
zOTqg~G>H2?SFFH!!3iYIOzm2{qEl)HX@7nFFX-&xn@~y&j7z7xiRW^`pVzU8hPH%9
zJ6puyD|q;(>eNu$#N7TNDF`|PYy<s~F4N7p`fM!AEZ|Sww9KCtIKGyO?;pQoy6O3=
zL$`PhkL_$v+9sc{S`Pv^>1b9Im575ND1I30&iBYQw<oe@the*}Y(%Rbg8`OI>zk;Z
z4y|_`kT*<K*uPVMeyf#CCG>5Xu-1`ulzlmm1-3z@{*NFnE&?-(47)MB8@}fG(2a%r
za9_RH=2}6(E@|3N6!i9OT;}P3p=^!S$Sb1}3u>GFj@4A)j`Ee}<M)HSp9*@2KCdl_
z*0~Gq92}_adYrDqbv4VX+?*LZS7k+$!uBpC0$a|!%u}#;_p5u(R2~>9R^rD<nA{H+
z0)9;qn8d*vC1w{twk4TwiH!0e_V>&J?~1|`Hzp*6jp9Lsb8wRFI9G&OO%#gG0O_!A
z@LTci?G$Oh=92vTORkD?)X4d$k{WV?@ZQRpt+gQvpoQUUt^G34Nq1X#QzWqa?O&Ov
zYBDkiM$F?7;FZ2J>3Fu^zhDQFkI<jsHGiujA55aqO~Z(i$oi%K1f=2B*WEc7^FCk9
zcZ1F2P%2%E?X*Ai6SO<aD-U{;2-nz9?0w}zt1G%GXa1RB$N-g-5bh=5s3tcdF8@lx
zc@89>KR#PW5Y&H68Z0%lI47WzU5T52s<hb^tBDNBi@*m%rT;d=OD05f1D2CxiQxEP
z%wVZcm!J*Ns<AHV1XgXITf^aWX~tB~enx3UP3{Dz1w9j2oKcxalo}Nh|14sbmM~Z{
zw;L%5%ZTvqM|aAyvu2;nvvI1TDQu+4!w(v?T7+Pzc*K><pZ8JhyP+CHm+`*suIk#u
zRUoqv6Sh+XRLg%L-+Ok&ZFXZX;jkANV#d-fZsuS6x5W=K><Rh;%Y}@Da7HB;4Db7Y
zHMV22FW>Rc_`22>*rxXjHJg4^K4Br|Q1n}h3UIJcv9xRTg~*{K#~P6-RR%2P_nI~q
z$P1n#%&=t?(HWT3-EP$c-f1~TJH35u`6dAo<Nt{vHk#H-fZlwR5^?KL>*Y?6$lkEx
ze_P2oDs?P*Efw>Tnrt{Lyuc{C>d-<g(q*8bq7~4woI}MzSl{wJYC4}ZDAH{TF%`<M
z5@mN<2n41fr6(%vlip5?Y=09eMgD#+43cX7_L>MBLE_i+=_V{RGizb?efNt26Q-3<
zKklFH>>WXViN9+xU;Lq@84%Epp}5VuWGElXLS6)(|FoFjh>l7O{`p?zB}dKv>fb<T
zJ2?I}o#HG40^&ck6-WHVcHJK4?Hp^nANk<RoV++3W+OV~%jA(QX<^44m#j_8Plx9m
zblt-UhOYh^Pq~vWYbu$zYX?IUCe~~()&!gW@JuG=Q#085G5*1j+z0x9`N^}9CO8)_
zu@YmfLYa>7u7PUAs(C%TATimeOWfR$GfF}&@nnR4VRELG5f9;q_rSftuaE^fwR2GZ
z(t3xb^*UQ!P3I13-wLB#H5#?c=<8-k))O#TfEhvY+kz*HvTzr|fr(tK;?~wK-#?+6
zH4bEA3Y`e=(X?@xJrSjqYTxp+5r}CT*$D*3g^@iYYSmlWHkROjAYnZS0VSBjK>Dn|
zV~uVTr+H$zuNL(QmDOM%CMQXFd~&~-1+}8|Nd&%@=-oznb5gzKX1H|V{GMCBCxKQx
z^{LR_Z=`aXyo3ku$vooCQw8|2(Q5*ANE}&sZ+^LzOg$fi4L>h_uf*~-E|9U;(Nmuj
z(E9Qz%Nn_>Ds-`(u{ytii_+)qcP9|Dy=Gmyi0XHK;Ptx34pOY{YVkgcZTtd@pZLtw
z4O|@Jf&?;cf8L?9Qa%>h^;_?Vl4{<*boyrR#K!*Z-HVPH<d>Stx;AB>2x$tF#V5&d
z%P~6EudNZ&Z;jPu>=y9ZInDPt!7SoyqHxdDTHv_-EA1KruMcAgVHLKwypF;wdb-_%
zgD$(B3U&*1E9eUT4AA;kjEI(NQkrO7qEv;hqgGXjYmbGUrLV<ZON**8RN9%soBP|d
z@V8;1(XSs-$7@3yu48I;s6jw>VzWE1TW?X>QYEV}7IP3m*&Q{~G(4NpT6f2|N!pPB
zpLT{z|6?$$GnG5gxpbUQ9=Cz8H66zVMB9fC21j5kL~GPaj56>O&Y?4GhqjzysTKYZ
zUDvJX3@a^P16BqqyH6Ex)$ns$44$Y@ldR*D{N98Yn%);*lbeierCdDDoL||t8nB-;
z1x_M;NB%p)RQ6-B!M}JE!r}pjDkVUtC^#PwV9;8`4x?#|&uy4I-Ssk7(AFmVDHa4Q
zSTa7j5YCLXMK38Dk;-_;XGL&8W%swy;;t{Y$5iKKKY3JXu_ST#=fhIbXvx0#=y1_s
zg;31y@}c)`f75;L)Of83m*il?&j^jx4CUbQm2Te_6=1|zKq%&11~PS*5#LynyeGZf
zPi|sFa+$}xClSs0`dbx63bXEXk;hv3$LIrOXWJuG?y(?QYKZ?w7WLk|t#OK%H3VQa
zW7K4G^7`*yt3VXo*gw9GAc<`I2;pf=NS&c<OIeFQo-2|(6qivSq2@BF?JEOrXJc=}
zEt(R)70#rWwQ5hO2xBk$z*eM|7#Ds2=e><gA#<V5U-@kNh0AnH@tf-81PzyOgOOJo
zp)y1baX&%|eM1Fo(j}nsXg`OvHy9qn?Ut)2fB5CBG-2gc^Xt43Wvu6{Hx34mFZLQJ
zc!iJzzwhd@UuxvZD;RX8vy4R77)l-Pn_xbxvRI{Fr#@xGtt5(jXMOS77}OMr&W<Zz
znfBKOE|by{S(tp4U%~TS#4&O^QK)qK1Ft6}VZ;Q3;rZZ@TEKN>SdXI>*!Krt7#$<9
zAo8uVA<Ey@eU(@mo_RG--*`74*T7@P-#Y3l+auHW<zW4}pss!0Nwv6hJ1E1LTb|s+
zDcj7g{C5esCzN&LSQN=1ZJ>Jvxt~ZU12ub%<oRE|vhxJHOXlaL_ewI1QUgS+L|{xb
z^{JCZ>1rvh9EveMrt(A!Gg;p@1_R|KcOhMC@vD4+9@k<hn~?fqd=WZ*&K2X3ee;Fs
z+?w~-@98k9TfEqLK1N3g5;?VeFN>CsEP9z}O-oXKUEnQv4F|#Gm*4CC3H7ymm%RXS
zGJi*^!Nx;%L!-ih*HfPMvXp~!Sk9`kcg~{_zzG?%7Us(q+bRwHzWt*~nD)!WU85#r
zFDkjR8#zw}1DIjNv@*AIV}BnWuj2HZbA0bG5Wc5M1rFOh0VhNHqacDvi%+$@FlqoS
z3HMNkImFnvONxEUT7(vRFUlajeD*hyy6$x~s!6(qiDjf8T;J-9j9p?J9-yse=3<gx
z_h568nM+EAe78v20wmtK`0X<yutJdJ&gJP+zgE3ALu{_anwGDqL_ixug~Ro7u1T@;
z1@?iXc=VyfWp6j@H%4P50_1|RLy=J?=O$~Ko>3`PB^vpG&R--28mRlplo8l5b+G-j
zUi@jx4Gde*uaPW)e=8NR@crH0sFc@0SJj$vw;kw#=bteN)@>nh0-~>8uG~^QpMe2}
zpU5zmDKZdccdGcf{#?}Yw5GS*ir*7LyHx>;N9p1r#_(M7n@R7^KRv}ix;}KQnNa$_
z3G^ksC*SpO!SwQjS50dlH<jyD`S-lKhcZC?B5=M$N#`0nMS{NQ{DicaU$-oTdziSv
z$bt;l;#dp#1x05!+(2$;{Z+V*pAfJB?`95&1*uEoUk*4&QK~L&0{g0dqMt(mImhSd
z0_f?%RS#=>;HKhUnv!aMylu{V(9nA}(Hldl=iruEW>CDy^n>m6icDbZEpG4NSgwR?
zx`5qki-G9;xP5s~(3dS(_8E9#Bu5M<sJeqMIL}a`y#`u7XH9Jb18`Q`ueVVewQt$o
z74t^2*5a@C>#?`X*}hzJ0-<>#1Ca(VGab|7;xxE(B9^$vAO*G4&FhINv2|^VNRo+V
zH`PDSsT|8sL*eNCUGiY4GyR&}U5LLkJ~N?Tfu{ZGZ1-`TybvfsR#R+Pt6t7>aLVY-
zo3Jv83kGWV`<p3omF&0SQn$DnhZ?*QVHA_R__3cm*%d#1N;9Y^ar-61%R7uw_?ZY3
z(pmdWTf@i*uE7f{Tk!1bE>KEh$y=dK8nzdtymQDI=e6C!rj~TON@uhhP7l+l#BTMI
z<1YF6Q<m|cN?8TXg`)f@k2iuvwOnNnd;ie2%Jnh`c9`7oxN$R3m&mtNWt)2~O^$6k
zu-3RW?>*zCVT(ExjK0qhp75ii5aGGRBuF0$0-r|RmD2X-v*nC5p;KTHnJyFe`BKsH
zXuF69FlP|CARtK6*nd>0K$U7LcZ@oPUFGkF$4mgeg3pi4ww0CghEw0-ChT|Vq01Qh
zCX|rdpk}Y;0RefoBt^{uj*OSbWl$$iIEEPN*IOL%fwg!)N0&n842Au}wKBAT{fsG<
z@@;HwUTHJg&-WCatH?EoHzu~=EI8HI-9J?*k&5o;7H%FM%!{n&gNlay{?3IftlXhs
z2G??(Z73{4WrD8VDMgv>G)jx`qxvEc%cZVZB_?8}ISOstk;c9KlBDV5b_HYOukyIL
z^%tw#*eT&E5#mKCN@vmo=kd04)613!=(6Lh^0Ekd+O=1i4hzcY12Wl0Qa`RQ=$tn&
z*!?XY-U>t-J3YLmhfym!gb5>EMXT>@;1CXod)|#bVq-x($3y!H8(Dyf<pSpG!@DhR
z<&)r9dX}tphV;FX$L2w(VT99$$!dJyVv_oM38F(z7J8D%)<W*P5=GU2J1V!R@Xqr{
z#H{n-Szjc56+$E^(3w_gRt&?t{{6BY5sRt|p8Y}_8X9gLwd=vtX0n0Lq^EAcdwq)y
zGGub0_3rC$Qwasl`pWKZi6^=$c@y}ja%aT#U~Q3yYlw3fHpbxCz<C~QH_m5<*om=G
zqsduk<f&yU*HN-AzwK(++AT4XW160HHtJVf=QzhGc{>Z^5PfC|jdnt%s#iPJ)MNu-
z3%ECW*d)zD3uC+Dc^_bEnwC~f$YR7`|1l}7RX*N@ba7k18Iv#b%&Y&1k%?>h{>YQ@
zps7mz%XOrMwClOh%%8h6o(_|TQ!-p}23UM8`v<G>3ZxdJgFEGKLT~q}v`F44E15Vp
z_@z=cG@s2(O$3WYp|j)MT?gNpj<g*27;NCVvw-@ivp<=y$>rY(b-Cbw3nb9OrMBbC
zPgtQKI(AF&tNWeuwJgVK?_hVg>pfGV8={EI*pC(@oDv>f#cMl<4M{_{$&KKNsj*5o
zlmWp)y9kEu(!=CN6l29YUG00q<6rkwQiI!^ysCXo7mo54RfCY!-q!zu)h@I)u)kS|
z?=Li)=_2xHn#$G-m%vx3Jm{f@Z28)3)hDiJv`56>o#sSw3qI<7i2B&}R}LzjaJ~P0
zci8CXwe3#=7epM&k6FxDCe3T7Llq_s9`>&N&P+p(4TY_yGg8}6{xYenUvfj`^OdVH
zNZcv!bmr(hmZ(^7$P@2Q243^@sIJ`Rh^^#pg<SQmJpLSK1@b#-&r4Dzlu4fXhWM9=
zoBw$MYTC>lzf?@Vr8~0#4ugt})r!niREx~3TbBjMNRP0<;ydXx*#7wOG)CNkgOa1j
zyw_Y)ulJK5?F+`ORayv46%le;;So@+X&y7LZNigPsBK@Rj}&q*ky`gR9bE%cR7&-O
zk_<yQO*7urJn0DAipDx3L)7W7lLCr16E*GNJ`rYydpgFP6F`_*^Oocz2)1=W$O`kK
z$+ztI3#oj`C+G7-D{<*g#4+=W>Lgt^oZBg`T$wV!EuVfL+5v{zxsZ~AQ`a{|L_rup
zxLF@5)8KiBn8RV0+a&;3_r-rT?jc|<6a|9YpqSZgw&TAin9=L<@K(cj!Z=DM)adf{
z0Hw!;*!E7%nAmoxyX{OHz6&B3d6s2{e?!#Bbmo2qF6PFxGHg;w@6pTdt}Xfi%<YLU
z);*xI^Uqp)QZ6+wJ`XN?1sJoR2Ug5drD(k_du9_>7YU%L6;a`AfUt*I<OllFaZqtv
zNBpG+#r=T2KRbf<^Ivu2sG}!rrXJo#*Yu3t^%_(eDpsQCfKKE)OK7AbXTA6KKmJ;r
zAa(7)#$fC~7)+X<D;XItNKTMM!XhJ;tkaEiXGsKuUk?yP60SDp%70Z7oIVlDM-9jU
z$>hwktb}*1lF`U+GW^4X5uq!SvwZdT{&u%|m_PnAUGAGCWJgBkVVyaerUwKDn75F}
zgAZ$<5aY0siG7z}$k|MA=TOR>zF~oxhzUa<J6`iqU;~ZEk{m<dbqu-Op-`c|{Ap(~
zPwNz7Yc-s<cU$_>;QwEz35!mT+U(~nNHX;o4$k<kGoD=+FHD>**U{RwZCm{inJR!X
z6BU7-xZQh**M~|uHn0Y#%Aq$kgc2&KDTX(~<qc1<<OaHvGWV2qxHB9PLN3GcLoAd;
z_VVNje?1;tl|M1XX_WkRANULIf^c%ivL`^<GT)knPuE}6?p;tz5B!CZC9My{mdU6m
zi1bz?M2eAD`{`Nq-jIuozY?|UgtAj*EyhuE#^ztO=<shnI0db}jNIGdwHk+B#a*5<
z13(do$jp+{WW&?m6fC%1{^zdl^IyAcd*_(Fld>6<2=(E{Ic|a1;&T1u=*cOo0%=Ac
zq=9(>Wg+&frYNgpy79#bbVtISrP{wzrwc4aLgd%^w;*L2O|kty#(*9h64E$?0U(Q)
z@|9nm2F>#Z@wfR;P%#LIz&{yeSARgxD%}Qr83mb#0@hbC)E7Ky$qAVdr|H^CdxWxb
zss6G(62-~ycs)P#t~2*>Aw-70A1|%=+GTx*ZZd4=ikX)Xw41rl-V0EDF$eUUFWBET
zHt?#_(8sEnJ2TPWZ~i2^<>>oVPA4GTd|`gv7UJc<&rS4Dz6q6RDY+EF2@q!Ji+naD
z<B(&$8H$y>Kfzn@kbb`#o#E+QZwMLx+FPC~_aQ@Vu0H4JJ0cALNCRX5C2Qn~X)N`P
z1$I6sx49QRte+(I-sR?g9?$6)G)IXh7>Eao1Q2SQsyMs27v44pd$#32ueVL!r=IX6
zJjHxhYibmP2Sl&ohnBQ5`+#Y!q48b)y>07JSQc%Y*ZH7%E>hxnHD$-!<5{RW4_ab8
zh(X%WQ21}Nisa9q5g1XF1lV5*&Q`8Yn(Zu%U_?ACu`=z1->uhWimyjAh+PL!QG(Kk
zpfDom%>$`lcGM#NXo=B4*K1aU(x#jPTW%Iq9+4riSM|wJ`+G_a<(JHO7yq=Bw7C9#
zt9(O^f%tn@DymBRKIuuiZ+%!5tmi(a8+Tr-TN}4WIntseYA+YF>BW-e3llw9Y@<Al
zHneqh6HXjKN&A`D1ipJ5VL5reR}b!2-hGo^dWY;k4j0|4?vzg=G?*+KeyDnFlr=DM
z0U{WT7{oYYglsqdR;DXT&*=d5ua=3ZcwE5s^cV_0&r$0zx0nq}pVJAe4Br9goN?o+
zIR)O-|5Pu7$0fh4;l%Q%HC>PgNF6fM`RZBz4s+;#wlA<R(Mma}l<Nv$$%OFEo+3uv
zXeM57?xZXJyy}FJQn(9^(gIU}I&f|>tKle8?30?LgQ*H4eqHiscRTUP2W2Ci*xGv2
zU*fLvOsAbIG!0+JlL^NA>B>ikv@O7G1|kWtv3D)+W1Ww@^8UfHn=jY9Eu%Evkn)bQ
z;<Y;^ev4ldt146@__ATREc8||K)C*BkxC*!Qp^A1I|bj#0Yrr84=yYm@tU#L@_W8I
zOT=tVaUF$gu}arFx+~&Tv6_=2?KhNp)(({uFK6n7%~$_*bMuM48Jv0n{vc%_20U=#
zuhqr6uU?&)GQ~YYH{OjYsPwLDLi^Jd?WRlfbu5mWzz%Bhuy6lW7Bi5^3hDrosaCB5
zB-VawHnFOR4cfAEQ>4=bwETo>L0WVoaBr^s>Y4%}#so>a!aBc_j)*T<!!eg?h=l32
zGr6uhZR8z;y{FO4KS(MF>CjQ!Zu&kQAv3+SQKwyh-x#>Iu)(kvy-k`a?T(p=OB;J=
zdwkG4#V%yf{3VCTwy)fePZQb;xgmkehhad@wnjTwQwyrvGWU0Jsn=U5V|deTCH3MX
zq#VB)gh{P>NKUQ<^VnEYLvm`YAEq{r>HDSG8AtPVHqj}#P9+XG=x5v%$xIYj0jBl`
z52^dq`D4E`WoNl9C=F&LLg|7yeun?9ud`Ix?#Frw4i#-BcOl+Q=Dkcy1e0FN#lcK3
zHKcL@NMHTd>G);|T6=C(oLA0eLP^y4U>2qaCuid}nj}2d&<I}2AR8u0k%ZVi^F7jN
z^n4Wb7a>(nPSwd8ZnO_=5lJi_hSb&S;6`|t!YC^)<71WaJuN@Cu~hXD29LW?QaBt^
zoo$cu=JP4ONYVyxB%Z7uEoKk+k2gvNY~>8gV`9b5q`~e_V{o^uo8va=OU~S2kvGEv
zNAws{tg#@1K&{Z!BtNbC){1b39jF?9-p<xSK=)5(STvUMHwaSdTuU>eofP+SUc=>d
zYa#&os-(iw<e3w4U=ni6r2qRfb&ShrS3l)xL=zGF?YApkU;tlRU;Qge%Pi>DWCABN
zr~<3cLU}wAqRj3+aV_0^%~N*1LX@0=`v(_kd%nW3XiB7ZT5qeN4j?iQUaM>J7Wf2a
z3l~z1O=^UjjFecXy$|(9w=p$}duQDlrrzOp<V{K}*oFnp(-9gQWG@O^33EnDU*9lx
z#>+UyMR2)VsrSV$p@m_*0@ZqPBX-^kZt-F!KUblWWEjFSVvPvV0*}E9ps;*HQ4Upo
ziCiR>)wBBJMG2K(i5H&^PaAk1PAzYnEV0H&HpC*z#nuJVwZz0c>?d*9SvU7ft9hOZ
zBMgn5A_1zNN0_4~SmBlX7opU^9fG>gf2}*FFExO3&YyXusWT-u1Ufu-GMIIDNA`@1
zCQ8h;78HwI!$?lasC31a0&<eYcjB6w9rugg*A7j~#O>NRQnI%=fs=kBwl3xkZycuT
zrcL?-k}w!7d`~(ni}uBpQsb-zBk0%R5|yUY0JU(OzIPwGW?K0*D)qqok)#iy3GT)T
z&7j)r3-+#2eW3kj^{{9l>h=Ri!sC$6BbuyQ4VHS3hRcdGxkY9sDL;Gy9*artW2a!G
zA)`|s<+&Qp5c1P+zALX-Zu6g-(cRu?vLqTKNsbMPhzI|XA~TAZ-x~~WB>k-$I!OwK
z_V(YMf1>ygSep`fEVT|$5_=da1zzXGuiGIx5e5n)#f2F)_yyyVJyHG0cr((d(VC$*
zNC7I9k5Wj+ragsZNCzl?Uo32ou@(7$ty@mq7EVJ|{zifHWuXlf4XuEA*&M7KrDDWd
z0Nf6VPQZOPO6NSrM93fgIW*yU*_A&a?R`;mey)Tr!Bd<dGmN0;W+E^q{Jk>09QsUk
zgq=ojcgs0aMq#wdF~D_LDP`x!Rh1gc71U3Q%uc#^FjA=xA@FX7Ya!xQayjsXCkn9q
z`8^I}G%!?HbSx~D%I1he<Tc*xu~+~Xe`oovU|=!pjVQ1sGZcyr_)}-!ImfnUJfy@9
zY53pwW^L}yrg%KI)85~jS4_NZKX0OPc#Heza|=E0d#Qk$0)=I-clI*6;)LuI4*2|h
z1O&3+DkXg(0~R82Tsj+(HKls|RkP897b+QA;RXW5>pyy2iE`edT05UW=bGjzYgCwg
zu5PQ|FRR!sG}1kLW;*p-O~ElYs=AsUg(GmT=EI*K%x46cDOmv4TjM6joWCjMt1-<a
zFfXhqv!|)zPq1FIW;ks<xEzFIaimz&(ngyNT>uVqbCZ{J+9fyZw`_iCMYWEPKN&wz
zNEBnH`=4Lpbzc;9n)G?mOF0}LzF4mW9y}uuc;nXzTjfs>PAj9?<v|%;Nx{tt|Nk$I
zCEEkxVWI+NwRM7#L)?lpASGK3Va;8H-R@WT_6$jE-bU$FEIujBGH8F{KlFbwJ*%C*
zM2e4->}a7E{f2-uqO*gwHkX{yI;|aSh!!w8IJ3Wt=`iVJ${1$SwVNxV=9H9Rs(j0q
z%(qp%Vd#4iVu>D(hL4W=%k8hzurAX%^r67=^%kz{E_BIWQNsD3bvS`~^uE&b%J)Yq
zD&-|mpJfdL`Zs#syyRGEtk*Eo;avNRTZ|Vv-|gX@14SW*c+Qeaxv{h{`7Q%@`c&OY
z%#AUy0>y-XJ~Y~jq!qT}(6$o1m3Hpb3TWpzA2tm@p)p{?CXb{mOIMBiJ*9fRW;_3d
zrTc%-VeYK?Ryc$Cj0DIU11|uon9|2}k63*d!0TXAV>G?d6a<tazDIl}#C#Fc&!?5%
zWVbr@({;zG)LTvsHC^Y3HSUU93Q$4%U~^G;Lt8Sr=n!M{W*u<CqpwyPI<GO!8s&VI
zYdeASU7r<WfJXkWAY-a9ST#N<!D+BE7-Z*-lKIk7BKN??UKmio-xnDuFN4Ag7=ct9
z#MXeD=cP6B#Ee%7U8ozsl_ptJ?aJFLrJO~?t9V`Ig@>o4+DAfw)mtsg7X?Y=tw!$X
zM7=e3Op~aaMq9t|UL!OwCICYTqdTg;j{h)IV04In(N3Q~7?x8MfwX=RuDQzO!}DI$
zpO1~knZ-EJ0Dvbi>PEk|KZwJem=by0C-l$IuAj5gkhQ5@a2ZWSKm{|a+n*&sRoD`X
z?6xO?&wmVYCiWf$&9N}eL==>j7K;?rxVc%r$pN^c{8oiFA+f5E=kx7~(5|NQ{WRk`
z)2{YsqC_41NCJuIf$(3Dhr3W>{M_3`tMQd%vzenM{U7gx=AHy1;DJ`Y-(dx|B;I`}
z(rKLbL6XB8Aw}|vX{GqY_R~4WV=i%J3n<|_6VRz~&0IZCU@SXikwkn}ASihWJWIkl
zYvAzc=5L1J>7WoU7BWo^*3=(?y=%Pl0e_rS78B2W+#{a7w9eIKl28PBYDp)=yAt}S
zQA0q)e5f=7XdvPx-`hCRX;?h?)mBBtu+x7VNl5%U^x<pYp#X6sJOP%X9@UTbxzAV*
zC`4sg<-k?5gq<q^JAp+ukVyT_*M)~l2V^vrm9&0gBF((T`my-^9`w_hr{*?xz+E2L
zTYOLCa|SW{A|Q>VLcM!eCwutYNuo*JWl1Wir`r~hWS=H}(g~}l!l_lhjV8o6d6`2{
zL18&h6GIjt^^t%|3y`z-@u++ZD|I*Wv+h|eH=qB@ezj~Ls&;PI+Me@bbv(aVXAp53
z`&f2-PWL%E^VuKb=IgkYhID|ky#;oVg6avScT?o0=2qYPuRV`!_($`v5D?nkNU4FF
zgjfOVLRSkpYQ{o+?=c8jeti>v!L**<e=k>y{%nSGGP^nWYzDk;_o+<ynV09RdjXN(
zt0%cr%ZIp3zp(B|Q@ItV!?EoB#x_!5z-y=30R5pKiyKS1E2L?-%IxqF$t4dqQPKWe
z3o!ITNpC8`bMmEuDC@Ca7m>|p6uzKdp1X7p1$V6~#X($!g4xzf4XtfW@zKvGxI&6h
z21c-^y*o(-P{K8=JexU2ju->y+Wq&QBE=Wd9}6^qAj9IdioWK8Y|3(VlxcVzFOWok
zzrD&ub}F83;{3@lWdCx<DKh<+g~gj^`ui!77U`my`21BAF<5NB`A!@t2`YzNk{xXd
zeP0=mf&c+tG?sn8jK_#SVD`j#E+~li1sWrJwj!fg2qwc6P#VDp6S^+ph!{O3Rh~kz
zr)mE@G{Of9+3EE4(@+wNc1*<TMW$EOYYuK*R`M}^Dj>-;(6l~PM0m&1U@C_BnlVx;
zv$?_r0}BEZzz8x8F(Hhb&m!gP3_Q^~`H+X)m-6@@+)Zxp%L`AHl6Zfu$gt>XFob2U
zw-W)09rXvm6@!2Tu44ON9Af&u>$`Qz)1fC);K}^Me{v4E4&P@9uhv=RW;@f2|CIH;
zowq&tD~GpUXVp%c+xAZjb#Ty&`3Vlv^cFKv6x2y$zmh)H;7=??tAP;%m#f}&T~>6B
zN#yi4e^mczzg@l_@|OV0>ZmTcyRk};2k07dn@)F=Hi}>EcJBO<UoXlBqeI~G^zEEB
zKv5W5UU`T_N99pX3;ve4<tZc>BcAp@I<Ua8llUG>TX1Q5CMn50bPp*z6WSqIyVMOE
zcl#QqYxnWdf<<a><4fs$W+2S`9#_e^_+SONq-ZMIAJ>qS59|v}LOHzo1>?0#0=u-X
zf8Jx&+pmPwUsxm<_A0R0U!uOm5Nq)I86}q7ot^ysU=0$Z0K-QvN$%_;8EPj|pAx^$
z(*^#>zy-}G5?}@6YYhSpDNUpql=PS0QQ%mC7&630R4ZUJkVq1Zej2yoD=Pv9STvzC
zd4zVSt4hDH(ulRoUZHFB%K7BNOx#V=-ggtY%ED1-NK(Ap8pmnpFXkc0Mq{FKjM=p(
zA<xG$d;8wbT>+w_=K2OI_mrf*))8R(pbyXix|>_SvpAA3K?fE?Un2w*g*?m3a+*L%
z2mm;5J8pJ{t&uE|?*fF8u}`0W4&fwaFQO$kCwXh@=a8By2Vj5*EP3$9GE3`{XmVAx
z?1b2paKF;;>B)V|(JJ}CV0by0FEu#s9@VzgGEz0!8hpe2Uo(H~IW8^INOsR3XTz<f
zp>VCk5pqX=LA0{$;f6D&WNf56@YsOC;0@?wiL2sBtjIsmJGKv5X#`ECv(uqDpn%c*
zai`u=<4_zuBcYd1p}|kY(#ek(?FrQ4_~`gUA*5PlVHjOhAzABIUG#8DEbH;wLh~{N
zU+b=arAEFyTLm@z?|b5(fy!(f(0d6F5sGoeZaQa*BF?M+szhxi@Mg@`l@|TvB4j;b
z-&)d@H2Z2|^6y=eI?b=QA|F~$LGMpaUQ6Kmofl?S)w}ISzT+!?=;?n&^MADo=9kT<
za&)ik-F4Y<7#YMY^B-k|-iQWp`GaE_&6bu<+ky^NSVU0851Ly1DF6}l9%u5)$Xt6s
zz<6y25M*{iN-C;x-;MsD%Kb0ycqRvq^%icf!ja~{L?A}W!B6J6dk4DIWC;!-2=O>l
z<ke5r=syW>O0oo*a*!nS&&Kk2XoY&emjAL@6Lcq>n9cZG`U!-o0wd4@$JStXIQ9rS
ze4<}l(MIw>*`Wx*2Mk=m6y*Ich2j=!9|7w0YdSLH>JPsqkiI7_1|qcqWk#c;cS=tL
zC1OEQ9^u3k43y7$03C?q^}agN{H;%Ak3HC)-XgWy$RtG+rq5JD@LN+JB)5Me29ux7
z5NQ6nA)lz>xr-_Fa3S|xLtBbMHwD+n=f<h9u4^hKiu!eEZ}6^wbKhy4`Or9xzOZH0
zoEqfIrjoO^wr*7M!%B9+o}6cU6m!6pQ>UNMZ{4enlDO`AiDn<CTN;_wictJS&xwro
z`9@o=K<2dX2JXT@D(oS7!!|K(^lzYYpBAwb0>t2{3cq2#W!&OpiP}*Q<gL3eT6^Ku
z+>WU1May67+S-vafmn1WvLz#6p+pVpqLS}Ze`13J;k^I)-#IFA`lh36ye1oaO&K;i
zrKgvzF{?z1!qaw7=<jfJu>0kp{_s=u(9X^{8wV#%a!N{(q-PaltIuuq*~HO*&%)S#
z`>~jR-!Fb;oN1+f=oDjPixEal=W#U8k^3|liKju^e%Gj_=x7f$N>h6hmte`%L>xc!
zr<(v{2o=(Coukdkhfl<SxQUeR5QmM6s}NOk4~zfT*|bR&phoZf*DF7|R=u+`roJ@;
zhDQXPE<(3*qYY@b8p{kSDTy@gWBu6v5+ZtfN|u5aGTQ_+aKiU0thC<WRWO3+j=Hh|
zm^5P+kQFt&yw8vF|KE<r+$t(v)bn8>OYg6q)W<i^yhi;g0W%k59tX_AlJkRMl+<9*
zgW=B@22<Yho!51Q1^Xe1EGyMvY(;5VD`$ao0s`t1J#)0xWsebe*+3_aY_pqq{Xq;L
zIK1R;%v}Hvj{fzea#FhnO;MS&H-3;>Dv-{iew;>=wsbMwELj3FRRkWut3%N2I7!w2
z1<?b)fbQ}W^nk`xfAfwWvV9On5uNkT4AGs#qmn}(pz+uI65ow6W?xXCnbgNhFh|@(
z_?sD|+q4b>|J*8T2B^Vb4Zm@QG|Q&G+viNX%^T*44?{QdtB>_|odbAfgYdtgbEVdY
zo7Q9b1B%0Iroz=6Ckg-MlUf6yyaRT@2H@pSRpF-H=L~%reB)>#PbjEp0EQ`NCDzBi
z(AUZgiJX3$6P;OHp!^MLkO|+yU<j4<MXxoCB+3YL%M!2^*LlS}d>_ID)~l{&O!&%8
z@Qn#52(C{|E|3vLpkm0fMtCLf{-7)#&+knfH6V_Q5PG!prUiB$1+VLggb?(v9TAO1
zlJa|VOD=3S-UKd%#uX2BWg`)LUkxg7V8ah?T|_qtXCrvOO)~@*z9)6&|2_ni2cH|l
z=A2JPWUWBg)Nt{)<1|OFoA3WiysT?w0t^<2)w?<*zyTNn6dy<f$55`^`<<6bJ(g18
z(%etG1v?kZRJ0yDvvbZ^{a89JXCMmz^_MAZaX-5y{Ix^N%#3m3*P^8jSf*odw^&8M
zHmlj)m`jGfl`2>kCAk=?>@+>_W@f%KUxEJ-EUKp*w<gxrJyR4a&%<D9`GlMMNjz7s
zg0?$UXvld}TV$<#)as;sUfXG*abJNrx;y3HgO#JoTVojY9P3*A%WbucusF%e<%2AW
zOuKr;5X5#4$%^S{KFVfvyRooK*9RP~w)34_{=M#Lk_I!i`yGD~U?Zmlav(LZzv6Y^
zAPpk<Qg)mHk4hUTgEGr{42_wZt=vD#ekG^ylqo=kK;@}u%65{AV=NlUm$<{$cV#8H
z;0y4`Ki5NF4KbfEENNx)EN=>kl0!f<X2PN67VbqKi%q6Z_u`L9M{=Ku`QDa4Lw_-(
z(#j;3v#&@kjfZ&9Rmsu`PrJKb?=pPF*yvwRy_5jB<;6gt0cnER4FpI*C{_3U>qkwX
zwPHmfvodpj+eiKiZeTdkP}c{JfaHz1VP~XefS<`$Ekt@$4*4ME*5tbVA`+~|0&+Mw
zGQzCQ;5zB}1D8<?D_8_T;S{+ah`{~b^hxm!C)~L@t6VkgZZSG(=(7HN13Vv<(8cF?
z4$GVdN7{gt?&L*ZBuZ+sE#K{WZ|9ul<g<SA0rigivs-5~S5`Io?ZrwOFj(1~hz(Cl
zICK-Pk`tfoa3!JTrIVsZJ`<<t>B!@MH)N7thP8Z0sw1w}DXgcfV^$atOgPQW)j$Z~
z7hvV|sNb33EdQTuQpk28@{Ye%YytTXN7*@^%M>lJkWz01!A;)X&AymHSHHX_q}#u4
zisEhow}#)a4r%aOBOn7e@n1*bY~<eMxfg9Y>95v3k&77T4J!{MCG`T7=Qmqjl^qH%
zWo|5~P5Am?`AJS*IZ&au#RL>48=2Q#UG6X+jc?(o`4Ajejdw;uHy3p$YlPrxOkik3
zc&s-_Pm8(mA3M+<-7j-<<FM=lQk->J?vQg9Fe|S5yOERQetBi!(@~DOVliWdSfN%x
zV$fgcXd$y1WD^m1cO2twZvg`eiFRVyG~d=lUi9A5<0X$ZQ(rJ;B(@}c;HDjDm$w3f
z5iba}df;%Kp_K^mi#(g@WH@M|%iJ-4;xq9R#ZP5V_a?LdkPRQ;&w06scxg%XhL|vq
zE#W<}GJ!N6M&}_(WK>D~-zc8S>zH1SZfJIMEOmE^`jYi2scs4XzdhNOni~s4{Tgad
zPib2OWKD4Xw+q|hnNsUFgc5Ng@PE9V0Kw*I-eA&57d=@C2zl}0-7t{BU_HQH36;5}
z7nkKn43e?iuUs@oT+V#zST~}D<q_>_d3kn(nP|5oH#|)_T>@515$DL3u>3gbtCVQ)
z1ROtkI72Wo{)rDgMH2uQ&_e&`nZVB(uV82e3fRG8&jB5H<#{q&vLTDt>IEAyO_Yf*
z0a`pdBnO}ImeABtjw>7=E)O_(?yvTV<=F_W66k}IBP34eJpc9cu5MS|^I(@wO2oDM
zJ7dG~Yk4MYQ*)>=F<^xjhfgW`|Nb#Ms`7+EVRE9sY87AZLfv$R)W8#)UFD6<%f2PR
zPVHF)!wp{zbWSEzt;2bME-(HJy<Nx4-|%3+NO9=a*G^{G8Mdc$swwMIba!?R`9PZa
zE?P5@0fw}t+ObJmo^$n!4`JsY-o>W4mZ{=@95peizIdWVh-Ok&hCVU0j>l(zwU8!}
z``@vfU^Qcbz`g(T0vuho;sOdWcGia~p@XJ(EH~*-5^~{R_8V5i8ON4J+75}Qf~Xx*
zF;Y6#=V=oK_ByhTt|yhPi_m>*?Kfv!tVt9=95w*X!Ts;zyr9~Q5HQq)XKUv43eq`r
zy%G&R)xgIQ*drjgi~!n?1R3C^d`rjbANxH0WagbFs7XFV+94c#0jiI{ceQsE`N+wd
zMftc;&Pus@(MAg_ehnJ*&fl6f$@bT6Z69~8uEbk?zUa=R7u}IQI1aJci8VdrsmP5|
zV(ZGLQO7NHE&cQ>xeN!Ua_``UM3Q%^e{Z}aCNnVl`gEu(?v5t!Pkhn=qGmKu;G6H%
zcKczd#40M-CYf4_?$ba3IbRaKR_5J3U7Vie#TC&pmFdwu@_V|#Vjx#bf@58&1L=^x
z8Y`R!m$jyPg}i72XP$|YyCA$xXJXVgs(=>?IJ&i`o~kLR-ndAgmzO_#*FF<9YtQ1~
zxk9<>V77hZhtv`FdiuLh+4N25iM<rCz6@7!Va}hw&4~s)DSOou6f8Gxo|pNX(06vt
zB@?pe8_F+@B=G=xKmp>`A?>JUc8znUh6~t0EzDpuRrCc0R-vHovRj+~a1#lw_dXX(
zch&}VqJ~ROm!hHKr|;{EqVtn6l*O1F+>XDFQsulJYedX^K2vjz2hCot$n5eXns+w6
zZvJ?**j;F$qF7elOpHhouaG4M04`M_k5K{R>r*~B#Kh5ue)}+B#rZg2Uml8aR-%{F
zM%!TtxCizRti%8q0_k_|hWw)s9B8whHOq#;e~@-iAq}HAaR<t%{NrPK3XZ9Etyx0d
z{^P{5239z-uQq7fy6mz$#g8a(fw`itMdf85dy(4oTfA8H{ivtY3M)9yPEPSWPy#R}
z7SBrj2Jz`O)#=xg@8qP)YW1d}pWPxnsWb3BnzQiX*pv+L>j;-9bXiBgKic|R!@|IA
z<5qDE|H6x-@EEr7G{b*5IiGSS%*qh@tHJVh18zRPj7B_47FK>MFX<k{^XsUsDkvqB
zY1}{3Y-q^)g8Tnb_10lge&H4<2m*@Iozfv8-Q7roQUcQ5-AH$*lz^m!#Ly|-AThuI
zLxT(rL&yF2opbJeE`P}LAk56S_kL@=>s?!4d-?m!V`D2h|7K0b_61pJ;_2Xok(ilG
zxRxu}8@R!++Y$iZfe7wO9|YAa*yGdw5^&>=Kw$vzc2tL)aa1B}n(xn`MgZ)`MSYKW
z{oy~-1pp-UAF$lM3)-w)xvN1~Ede>GYN^>i_`)P_d4jkOon7wjZyY@oegh5f8k(fM
z)hJO4putC@+ew9qjSh5YseF9yV%2CN|M)|G#%h<YRgG$dGL;4CCSV%K)y44pU4}wV
zl>%di3GvuBh61p?cTnCdPtZurZwoQ88?X1J(Si*`n>vp^>Ue*Bzd!Q6wrUe$wfP#K
ze__>TWRlI*WKf-j8Ds1F{AbI-T--M(9Hjwz4tFc*vc!_|OeYeg*qce%JJmRn4-mz2
z2w)YA)yVHS?DAi=9YJQS7g|BK@K2<?%cyH~2orhyQqS{Bi6+GWS=4i(xX&#>t}fdi
zbFn716|0+o^y9VUs$IhOID`NP%uWCY8m?RkO#~(+JE_FZ#C0Vnsgx}fwFg|^4>!VE
zE&QkN7cH0u2ir2`FR}Ti?$vUg3A~MgTPIc|Pba8fX?xBei&D7Ky`uE~+Oc@-yuP~K
zyzOqpi&0b}0cP3e2RxeDo1uG<7{K`jKDD$;1AE14hB2hwPg7GdF!H~+aDaD?Tcvq0
zzQhiim-{OqZo5*LdRq*7%1^$dPp8*|_42v}3#-MYa)_Z){TxylB;HC(`t4p)c~5Bt
zMy1ByG{dpPk(x5c3|yrO+~e2nvh4R%(-nP{a2SYKI;LlmBcA6(I~b{NbvR8UyR|su
zL1Z*bf`~kJTYWN}`CegSoiG~NnZ}^u;n*rBoTGlMr$7R{B~@7e+^i-_yxBbmMgV|h
zt6)7!q>8w4w5@zh7p*4|=o<%Z^(Gh7eqFYe+VN*UXl<)^i(la-bQ5F&2t2Nf{Or$x
z@9ZN0ksD6r91>GB^|F=5s{shg5OU9`bLt2&UOwvZG0U0cO4tY^o;iBpYJqf1IBo3m
zy-^1QP0+BqAjT#=yGj12k&PNVR`hmFTZw!FBf0LlQj|DEFZ)o}UGuaSwMJ=S<LVJ+
zGy76fmNB^%M<DZg>)!yBTJ$2K56=aYN46NVX2g_9o02}QQbyutDk9(@NV3sCVnO_J
z-}+wx)7W3gJXR<P2C<$4msvZI_LM<Z%jV3za<ucJJb>3DYA2H`&yhIwPFx$+mF82w
zJl!Q%Oy$}W-^HeR?0zjXgS5>Nk?1)#nBCZahM~YO;QZ>L5GJ678oaYmis``jCpi^L
z_^&^g!5`)L4m?-u+922KwT4gnoL%;YGC#j7-q8IValbRrQ$`W0TDz)n(+8-nE(ned
z*y=9p4L~qtZewGgUbPN1btC|^b%zlgFa=ap7u=6W{tD1lu7YJ5T|a!>ea>Ix^Pcmz
z1Z6B{KFR#awPwCW^Unthz<<2XFf$&*0T5<JQA~}B9C<UdC7;$fDaL8g^W$tCl%uI|
zPUL&QSZty&-nbGdR2N24^a<IfZxZHCk}xDShyENRxOx{~lWF<~+)->nd~~>MlD^C8
zO~<y}v`mHx(<)R;rF7$tj<3$E%>PyNt2wU<@M7Zn0TXj-nh{?U1?JI)X*dH71fyr1
zAN7drUYsZ*Sca4Brv8SB=F0wJOL=mr(rIAI?m!cm$o7*|f5Oi&dQ1mMat`&7d`y9O
zozgBda=9kpFM58%EJ!5Tv2%^V9fIe*|5?!+yz@8eI3M#+)@X}oOF{Ecihl@9sB2qa
zmMek0f{8`;wW;g?VT<VhdTA9edo%q`qipem6jSKsn)wOmPDj<PS*4gg4}nUyd1Ke?
z-%wYI7CcX}OBos*5%>AGI{cuXE%M}60O(QU8s?zdmu9T^BvJnbURtx!h;Lilc=+K*
znltAqUkat3zP|JMq1bmc1La}jzz!D2-KExz&tv8J=}qU@l4JL9*Z;MtQxSxG&_r%N
z@Jw)H@JfDV9M0q0?zGD#1++FwvlcP!n|bg}D7Dmrdl#8Yjy3{6rg+u;C8ytT`g4AV
zn<$R$c_WKZZSa(?vEQaxek=X52I}P+c97w6`6~c{Z@Z@a*#?#Uh`1K%!$Q6!h^$|x
zf@MR$O9#iLLItKQB!}n9u>uz1WEk*9V{lSI8H4>p?trt|e37x8Jv2uA$D3zPjOm@x
z2Zs`oD%PE4obT0AazI9U?01Qo$z?7`z&A#s&WJN@qdiJh{~T&%*JMcrc(#}ErKr_K
zo`ds(J7y%KX|CV~mM_~DKNM?v2t36Epsb8r{HlnFFl~#jC#41Vwz|23#7#2HwPKI@
z(eR>=p<-W_@1){;`My4HZgO&=JArl86?1Vlf>WHozTE1vZh7i07eb4`Im>)9+Jzs&
z$Ah@s)Q&Ill)G%e(hO74gmKVp#(YV*DG=Ob8M#=B+`|RI4t?bDd^R_YxIexU^BjeB
z+L6iF^GdI@<v*aXE!UG>d3Jr3{(0QcNgilTJVMA;@NeHPQgD4!$YFp^$U*k}rC0>6
zB85fb1n7X%eP65cehSsaCw{+T<ld9zEiGA2v)`G7<4?*`(e{#egWnQ0fUr&ZTfh!z
zCaS)4O1F9e7&ifcHGc;3l>tAdAQJ3v2ZO%0Tj-rAzI-o%?D*(V_h{&*6RvCwK2*_S
zJZQy?AlpSls7;zcJmEx0Aly5_0e1Z-0^`2;wjS)5v6$9+OWV%K%s=d*ujnBAi_u@b
zKu<q?6#(c2|9n2~lr0ojmDehNb7aaBom@J6J{Z3m=4<Hez;By);@~~>8bg!WU^_p<
zdd2R>oo#;;VDytP&}STUtv|yh%NX3D+Z?#N=>@CfN^YcHZK3bta+XM%E_i`B`&nCg
zXgZ}`;^s9fc$-s9lxlO>op1SK4oENs`Nsc?bix2R;2%Kedhs>HpNfY#*>v9(FxK~|
zx%}fzk6`SPM`h~E+j!DnI$D%2Ilfi(3d_zL&9~%1lmJQA(X=xA`Fw9GaYPhwSychr
zsh<!6j2%lp1t$!9vc6)n_kKS8L4Z4xCwRg@B91UQ|I?&g{ypbxi5C5wexh)5w3JqO
zhn$Cjsv^QqZg6hASKlU{D|m|3HlC{`DztcR<~04g`Ht@*`xm#3mRde5vzCTN4(t3a
zeStw4o>67;Jl8nmf56fExd9X!lnSOt`yE?DH@4q!JPcZRM=L)nNp?I8MDspjP4`7}
zeRQS8d&O`)-;|&>-}MxVxBa)6q&vHx$acjMy)TpX0aIY2Mk+pqyWOPx<W4!87Vql;
z%`*c^zFh4;4kBD^TOdKhhBHWef#{FB2haq;w61E34g3gHH;I2Yo)Njb{0@#oHM_?X
zdry`p{?*QdcI?p`So*Qpb6Ct`m~2&XcyaQ*%kOmukqgN~t1DN)&4~@J-o&Ur?J~G!
z&W}NHGQ$&UnxjBXu>xtqzk2kJ!w$gcJhZ&ttR6h67wPJ?4s4;d&Tkj}IMil2(aN#I
zXrdzstTIi~!15=GqbDib+z$;z#FUcpnx?9ibWdsv?%U&uVoR!>r}|mhviw(w&|4kW
zslS#t526;6?;$nkEt}G`cltxlhWH*UMgO#ba^-=^gYc*Moju_fr!kKO_muSKGBr9X
z87oHk;4K{Kl4~94T^ix&zR0u}n>om~q5P}&^tCVn6M!PDZI2}#jl`JW(-(AB3t9()
z@DcRxmQL0+q3EI;s+7d_OX6ty%k&Lb@E@HvE+eMJv8idMTRM&hO=K(I%ryx%gUxrd
zPOmcBO3wu<A7Gwp>b$QaZUtlCaVfk!>4F@|O35IhoEE0Y%`;GaP-2YEg#<82CWUYU
z159?%4ukAd_#kf~z0Nly4-w0WJ4xubQ2!-J3*p?icL87!F|_um3?QfK_|A}{18Wx6
z<29De@5w5-bjZ3{gkZ=u5qAasdJ(HOPZ`L4Gk!Ak<?WibF_@hGs8YY{u>xSq>AYMM
zLZFYTCGEC1xuocAd3Ji7DygYD((wf^x?K?-QM++3QI7pEaH-6w0p{X!-L@qu;KLm~
zqKK;E&1gL7*+j;vg6%!kWikQF^rc6V{*uD+k&itUjWMAePLOhL*>jOqTg!J7atV#t
z-B;z<JWlhqdqUSaO&C2rl)}XPRca*XEH_%j&%L)Xks=y$^#I!^yrL2`5&F`edA>LJ
zoVq-K=dPVci>yf8fHaW<zIs~m9{07Ve~7VG`8y051Gh`~*Zr-1;|G&?79ipZ@;g$~
zU^h;xQm_RjRh(2|#;P@n4LcJ9tSA_68Zc#0W!U6#7)Lc|jdX&02Zu<-pP&7P#b;77
z9kj*`U+RLVbEX@gdqGrLQnh8F(|XYT<6Mns1w6np#LtsFtfSF~1)n`>Onj$-Yinu!
zsHvF*I32`iPDYXuQ8d!?3b#Y}V_IL%T@-qDD3k9ct~?w#M&;c;q@P=YlS0b!wM_qX
ztgPG|{1%i{lbuR3jsIoKO|^ZmQUDC=A}7C=cP!J3<zw@vTLnGgZR#YexogziQvkU^
zm^^JT!}rE@^ZE>GBC^}JQAqn!k9Q_5(Kzs!M9I>*LxiP)(j)*<t_@TT#e0#%m&EOb
z()|&Ai&zmAf!*754$b_*v7=$PLIQ{d)<ez{h`&by;0g$oLvuF+DZBZ1gpXJ8PjR3%
zDD>&LctQEu)uk3@6*Cpr*-wJ?V6OFF7y!R8`*dnDcvO-ZRPWVukn>7bqh>Zg{sSH6
z4c%aDC`J(tZQxJ}|F9UX<h0u34_5lX#7L>SW5Unxd@SL^^kUPydADqjQy_B7&paEo
ztMtg+;ql&Ee^c#Mwz+SmCNccBzPVpNM!WBMpTuxX@6Ut%Urq15BLLa~K#`zMiDnGG
z5#z;drqjyf|FQRux_=BvbL^J&y4?G#GBw5Gx+pd3w!S|1!w1P<)ziQ3L6wSkE#6`!
z8ALfekCH)05Y8?<2yWhV8FutHnf~bNSM6*%oZP=r#yofQ+<sZZn|beVX<5r=p0B=*
zT%4*?tlCt7PJ3tEAu<V2XM>d{%2Rzn(l>}FdbT|T**Av1G?yY(@S`tqpbt?OhsH;Z
zr(neobY$wLoEF8A3*BHD5szaWDMo@mZE@L_9$BV;x&U}JGGH1`#&&efE14Af>+Yj!
z#h0P1^Iz40>$s>*qy_T(8QYyja!$?1W!TZky3xIRY3C|LY1z?02~!J9;x?Vr>2|h<
z?kMk?d3R>b+`Lzap+vey3j!lvO90d0tOu=8tm7?~fDFJwJqyLU-Qz>H>oVt~?6X&=
zE?w9V!@Gw8Nm~ulwO{(a&-)-~AG{4mJhwyOK2&4h1i*spoq?w1Fi%!!I8Cvt{k9qI
z$T<4byhB))m*w<xlp^A<spO9ffIhSj>vLy@?czArh*|@%lUGa2wDo}cmO#wUH&p#f
zpS;7I;jySVmkz=!N<UAA=GzsRKB27XA=tilT9v-Dj>7UVAa}`$schRjM99MsnBYUW
z7d>c=IA0zuMQ9(a?pk_$OmUW7c5X3CsyXsYcy889)`{1O6sOqDV!lYsz9)SBLnO*+
zqto`hgDQgw=I5BSg1UCx_{fK(<FJVdtxnwste2;{0~qk(9-`$YV-leyoX(ry(YfYr
zxy5-r{s;h_GI)A&(|@7gh$rw^a+ZSFn|Mh*Cu5_T(2z%zteT$$PhOzF->vdZ#T-3~
zx?#API;_|U`<lI1sWuAU)oU^~oMTbS6@F<ia!7ws;MCBC>z-nR#e!}kF|I^LzBj;C
zrGGjgP$TWFE8qd*-9%@<Vf@@`!0$id%-jBj_$Y9^=~NOz(H=1UJ{A&Luq)3&)3xG^
z{=}<tlO!*ZQsq4Z@0`*O<8@>X2euiYBpoyqr#|Go>tcFy2y5^TO&;uc>9*i5=Ym)q
z)2?>wu7BB{xTzIp(8Tb#khq!i>TBx?Q0{oe*T9ARVNXM+|2g<%q-X_}{U*Nq-K9y>
z%bkN#Jy;H%4cSY3k6>y(lP61P*<zYCk77CBqS~9k$LOH>t7{Kn^2ePnh<%ZCPCrcp
z)22#)uznNBk1_r`BOz!ql%0VFV=Xk)Ner!i+gaH-Z7~hJRC|ft_Nj^7G^aw0A+~u?
zMD61a(nCREyC)g_^v%2X^BcXGl?h>ezkCDA?EhXupuh8KqJf6VLX!6k=0VOsXCc0E
ztLH)HJT)!(b9!tHn{j>A-rbrE>nh6O(-s8>oi+6O9o7!WH@!OrnR&^EFau9ufGiK$
zi6n>wR&i$$T$ph;e~jm;RgJ>1#ZS0|<xkNyHupbkAoN0pyU`^7u>;8J7n|ov5y`X^
zAGDa@1jVz>S8}%g#q+nqRd9SgSZ%wqYnnlIO!kS_?hR*<vWKRFk{DmO>oaI9AOn!d
z&=geO>5lOFK_^*UO(e1HR;g7PqwPy^qB*;9qeB6Sl;&UOx+(cX`9cay^et&WtCDR2
zg_EI$-{&Pjx)oRqqOj*|{pAM><91B$^Em$aA<+MK;!ZU!%_aTy!SGr=-)QfH4D|Na
z9R_AUJ$@q)FU{l~vx2uf!<YOC&n2gbYB1@8?^Bl9(+bc}lOJ-f*u1LJ?=l1+?k+no
z(F9*0A>P|-{STQs{Ry;i^kAji7II8H=-sFY_ozX*bT7l(k?q7H(llIB&ZMf~uGU)m
ziHybZs3#hC9?Nj1rE8wrTnPN?eet8%VL#;A2w$VUg6?c1`+OYd9Lwdzd;PJK5S)m{
z1!9iG5J`Q1Te*MZR48#@U@2$UhJxQ*YI~sWM}{BtD}uUzwtT8~f;yXK*fvAE+Iory
zs!&Y_Eac`WFjt&sv8yDhO@Cfux1FgRKCidqES97id}FIHZou%`7;K;>Z}7J5r~|0s
zUUm;|&F=st|Dz6@-f$s%ZUX1!T$7lQ^KS`28b3s&!mxf$yT0O1HHeM9N30-Y88!vh
z-OY(uZl=S_{jX}j##@0X>A^zuV2(CbswgL~5%?6DU)(i*Mha5>Sf@}ti>p}vF;!y|
zFVkNs#}RY20HktzTQ`4ejIgQ<B|~6;f8g*OFUsVH1cn8GSl5xOe=?2Q<)ZW^!{MQD
zb9d6f*90Kj=e{kFCz@Jn*c|4n>5bms9QWM)#NMA`X%GSAy4X!!kBZ1468CFx?Tsw3
zL!YMQ3fB9<6W3=#2+KK-nI^CnFuqk`A*0;L?5==H<lSnsdma1y2(*ugl_1G720R6d
zq6C1wRH|!Z9EKku`bzwdPymk#V{&404N1|5|MuMZXQm4QdC0t;iM|-1YcTto<;faK
zguZm$QK~5TqByl(!6R+{(tZ2SE;r<lTOIrT+iu9gt+ltj9zZz>jg9t~OCTy30>*iP
zPDT8sTlQ~90Clp8!q_tZ`u<aQ!zLHf3VlOvgQX^z_X{qX@j8(1xllW?=5dg=>0$Jb
zDCl=VAaR1NF=)g*RStt{jo-<p)if=UjWo{2F5FLQtlj3t9KT!RKJMaIfwq)x*U?<T
zzWw&dYfUnMRKcb=knXc)&jEYh8_kE)8rU7)o{|{b1Ppv^-VeV@^pmz+yp@Ukj`;Pv
zJBPwUSK0?RW3Yd67ZPx$el{~j>88%p2aK(psVk13EZ%2)j5I*NDQ(5*E`9x@<)2e9
zYsv(-Pn+WdP&xt6n&{k58Lg6zGf_2A9_`P-M|b1JJpZd`CkMx70!m?RM2!Q=#^Usu
zxQ3YvkatwQcj(XYA>pP+nDIAQ&Xe_gFfS=(%+UOs*_P)h_GBP^uFV+_>`fAXgsN<)
zfu~}{8Gy7H_l`aUf&XI~*()BU2|bVFfQRIm=#2LrP~X*>j=cWi{cod#4xA=?gVgE%
z1aVKklQ;2lyzqC=F$hp#0*zFE)Cu+6YL^4dkp2xYqn0;sK~ongG}wzS(^O}TUXl*}
zQB{f#N?|@R8eI*0BJ)Gl^I4h_n2KxRsjq$5qYz$fB%YyHmT)h74H3HfLAMki-A$V+
zCsmfi{VHkb-BMw$m-XB$uM$tvbXCLXjt=F$_{%~zwU9uuBfsUwLx*<Tp6sbPaNM|o
z(1%hJ6gC|f48aw{0~`AR!t_AA_o#x-w+H4W{!vS8qmT)M2+*Z(cp)Ic5!_v~T)K&K
zHkln>{<z8GsL1Il(zBDr^#5uPAS!Z-hG^_-*}?*AKyrknt6OxrUcw!n@cm|g9t6Xz
zwM>~yngaz>;j7XG5w8AcS{3daY<*F`r`Ak#IcWk_H;y|BorE@_d>+W?XsNN$Po%?X
zr!xt~%Oz=O%&)+`73=ga*rH(N(39wCrtafGR%y6@pxn&ysaN3n?_;(~-FV37N}iTf
z9FRf!BbIw;;#lI8)5i5MKF$&wIG)ZVGt1852gqUU94qZ^-uLgb09V@J;pj-aH8*W$
z+x|g**mKPgnwZ37Q@aPyxKJx?_fnsciT;5}k-34XxFt0tOuwxnN{HLOHT#1-BDa7+
zJ1luZx`(pXasfW8_ryfn(p=NFUgF3A@dRk4I2$z{mAe4pMuP;XFTch>LYKscSj^ji
zo*+4zE_9zF7Jh#674r-ExNlcBO2VffKMha?6KL&)sLbr-F$!|R!$*xVXgOn#v|K}k
zaZhp`KE!F>?J7T-*7G`&L;;S4<$Tsv6Bxr2_6L}(B5&7f?=hL4-8)*=5MHs48S5=2
z@(D$D3|eQlueZM!QV_s1mm8?fGA%GL>_+~;6enoo2a|K1&U_}J(h>N72Ri0Ii4+K!
z<6_@pYd<&U24kyYJ>?wv9xrxCPq%ui%3gim81B{QX!Ag6Y&w+X7mN()h!{6I?@v2F
z+1!B|$3TLInIm<#-L3u|i>G)v``p;IV~NIYZ`V&1-A&h^eqdh+Vw;OlNZzOVF~v0%
zqsD0VXNKG<+j&0~S1lZ2KUeRZKz8LPvzBtjAGcD1*NDief9`LI7d$V}9`AQZYfC{d
z)kuJeuLBzaJeV9UP7kg!<QWUU9|w$yd8K>@#d`prTTKtF(`Yl(h&eg*{u7zBO4PW)
zf4Qf8lp43m^s@p+r`dSL+$Ftl*27ZUaJmksC&G2X{OzK%okgQnqlvnoP75&4;}6c6
z`%K$IZvs#LgMx~B{`8q2aHwHg1%HPYqA(YCD35Q&(d^6rO!PZTV}5PAGuiqPfpVE;
z?)r4J;Q92XiNbdc;bbtc;Ka?-O*HSf!z+=i5U`Q;e8;7MR{b)XA`bJU?JwZA5W(FT
zo!jiItV3IU=(i`q!;t%8Aiuc&dnj~ghoBleMe?7_SWD%P)MQ@&cRlx@*`4a=U8ede
z%}7$a2dL+4?F81UzS~DVA#E8(Iq|J>7c4-H4jV@>eDPJ0<j~+iADUD87-iKo4~e|E
zC}9Ie$X<pw8%h&pl2&q$0^FxDxfH4P(>#NnAJ$VcRj&)B<OZrlD5%k@9Laqq|J;cz
zo1Zgz+vn=;yorhyqD+YTa_8e>vPB0Zki@K=e`99&{~DFh$+U13r7j_V4SVako8J}i
zvYSuQ?&yCg5{jYX*zTrL29GPz_*|g*<9R)(W-uSbkp8H}I#q<G$GOFVCMe~wQ~=QV
zDxCoa1w0aen`{s)68_`)xJcj9OX`Si)K{g6CW&2HFR!T{a{qr^fTf)$k;r=a&Z{LK
zt0|tl34;rlde*Tj#a#moV0{F35^HX?OP!gY+W)CVyDY8ehC#~T=!@mV_&uyx5|
zAh=rJ3B2m&+b){Tlti(vaxd6f!XBIj%e>S4A9XE!4aMAl=%`@icg{JvEC9=z#e3ME
zl-k_9|EB<s<d9k>_bZX6bb*+&;J|i(S!LmZ#4|g)c?Ee2FaIP}j+)@sWc-0dMy?s1
zi@9|Ds3Th~>~&}nwltFZ7-Rlx&)<|Y^mCgxdE`2u1t^iGB`)4vezk_baFg+wBLCMP
z1^PI}GZk_vzdL0R@`6Cgp|5h^+NvxwxM%O(_f||(340m<S-@|*e1wMlUIl+X=BtC<
zQx3Wi01fDj&rL{HqblblHv3p_2+Y5Oqq;vEQBIDUt-77=rzv|A<wre>2|XVqI`P_Z
zzpbOzktR`+gCd%G7Wszx>@*G47yNI#NrYPD`?foeX_;(#E}y0S;;!2TX?>-sv1uxv
z8}R3eb8jO~@9YyyY|L4yrzM_?FoUY`DaFm{5LJmsA9$!QM`;}*l3*<5!6Uv(rA{c9
zMQ*gwK_}+q<ze4qkDyXbPV~_?3awwv_aK;nTc{8R7dO8C?UnK2@Gqwyun)I?*myl1
z<PZ?sjaO}_romz3<Pddj-H`FBHs^`^nJN5#FB|^Er_*by1fT-Tu%#1x6PNn-E>7Oa
zP~aH(Q3H+1={e-<+OGW%n~mG>?8O#<&gug!k&$e!IWtbj5mQI{zWcQQ{C_G+KP`bd
z40LSb?SCJGaFmn*ti&WBG;s_qPt`+RWA&)lG%?!!Wem*Lb@H7DhlEV~<pT5^`EN9m
z(|?S*z*|>IzwIZxw&gP6kSpqa-vJ(C^F7&pM&T>C-5^NrD8}AV^p@eBM8ywF>O+f`
z4ayA4`s1A<%&unVJwyHJT6g)q7Kl+`F?Osp<L$`<Ec^^_F3*Ha&MohmKDRNRr51Ev
z62<$!TPLG39SM{Kzy*FKsW{l%vQX{EP|=w*4bOhcB2Xt|k0;Rl$U^|y09Q-2<v7Pn
zh$r+2rCu_KeoD6fv<&2mv3<*iu*0Z}i=ES1;4Hi*cE*cRqa8ZCv<whYV#JZfi9Q=Q
z$elDXu&sB-KKy^@o~l7mKrxH-6(IxL4?AZFy5|N=8S|uKH_>h^8z}t<vM?9DH6+He
z0V#JySvg?N7cmEOO$&i*%=ZJ$Zpb_!50W7T^cw&R3`cCPPY5M${-98m#&k(%k~?nn
z^FVr6+1Z?y{)$cymcP@IKnjRaRYNH&E6gmsfa8kt->B`R?nOaHP6GSEgY&>dp&#`G
zHM@2aOaC)`(S8h&M&m-G#w?oo8L&3+Cq@2GHER1=HHw-}8Yr_6{Rq2J|6)ble~5ke
z95tg~M9aL4oxg43C5Fi%v=emOAs``KE;*q$Syg#%x1j<Jb%ZS%M`YQNk?Rd$f+Ue@
z>)T=Vd7v<hrTXYgv6yX^@46g3>GgYA^Y#m?wY!-)GMJ5`9(EIj`xz#RGF8L6ZUDT{
zMiXHo5Oe#<OlP8SP%obx`2DbnTrYTKMqyLe<Q5B{*-NN7TKX<}?{|*O@0Lp1iyZ!F
zX+{$?090qPYnZ3TtNN_CqHYB!!wLezJ$PnO33so<QE0cGERRLkis+xjf&%}pg8mt%
zz9`0H)W5_cj;<qnR~8f*1nC-bVkdipkI%sT4viTt_(kCv62lKW-MTBrFapLEp+nI4
z6Z~h3OUEjx<8Qj{O}l^pRx-nI3Il6Ln`7Rwl(aMteuxad(x*8t{F@#1PjjpYjr^l<
zZ0u=}MC=N#Im1_trAAkU(>;x@uuDPCBkDgdYzjVol_FE(4~wxrT_E>GZri~4kxmw=
zrAhKvi8zvrZr#K}_;E})h`^yoepsP@pX+JO$pW}kOE~jPY2kZ2VeMj+)%q&J`oPk{
z_uQbF5gBA6;nRU{Swy;)2y8oj)`&kI-}<ci?bC6;sl>!Aip_&VsCjv?SC2SfEc-b#
z{^R!HI|!A^pvKOld_$Nq>KN6>!vQD5R^Q&(rLVA%o&;6<U=}EKL`)|j#AoBt<?^LO
zMn>AaK#3VLPNt1&AM3dd2ybuq<E6${e@p#BG@&wEBQsk0qJaZ7tg^{}#?fY-&p#a^
zZNqnJKax<6x<=DNt<37X)H1%vv1Dc)u)lr$6mLes+#14!^Z`+yx{Kut51+MZPw<U^
zi$9@jwD5S(C`{UsaT+}uc;BJ$yNgdP2tL_(=BT7m&mV+j2easm=vQDreM#bStkVB{
z44=1AL8_TOn%Oa6i+P9!=Zw9Bev*$kI^LXf8u@FyPxNBLP6BFX5|y~t=fI0Z9$1xM
zG3yPBV*_7vT7IGFBc{AT+qVI*H->p+5807rDA@nDrE92MkEu?V#(c8#OPj{+wg1+a
z)%`4Mk?|$?*C)u)iG!U4LZ*8u&(fhQ>G!ZGlvHNnI#Y9maLo1JtKLgXVQ7AW;~;QE
zrMk6LS0DV4mDWm$Om(I1SdJZYV<j>Z_z?=Ut|mEW*5CV5F};6F$h85|a^L8R`A4{B
zd?M{($gAYNI5aXEj40b@X%v((jVg)~Se6_uy^V7==Qn#<7SXVb$JnBK!g1ZRXCobq
zgZ>KV4V#0Lv%8FI*Jt_3ad8V3L6Vo;(s~{!_agU+DcK<<5^odpa&>hwJ`zVoqE#j=
zT+oKz&NK5~0uM7~4?8!viP-GerQ!Zf0EsQziAI@14iY*BhKYeydP4wwQ|F*|+i;w=
z9ZIHwjEzbdb}pICf4=K1wppzBm>KAl1`dhLfj)u!5w7@1Kbwfi)D+dBq(qi2-Qfo9
zh5Ly((*1ugHI0wm5!UA=v*YGox-qetoE(Z!#OKjR@u%u-+j_4E`n$H)-toD&Fr22M
z;#w*^CelZ?PMuWXg6(9~!Z{Xh=hL2^`&8<=)6n;S3`Cr1+24k=>zP<u&<4}8o3%1F
zPw>2nVzb_R*CftchrPFx&_O)t<oeZeHp!N%KQ&wy(O7<5Q7Taz;mbR!nV&zB91XZJ
zDq;Xj4TJL<A^8$xC*_xB@TQf$`S&lf%)-llqi=X-)n8lKn6r;@^tgl9d4oZdUNwIE
zkdB(V@Ox+vwM!<g0thjavY&SgNbRwsBCFL+z7)#MD>IRm%HPZAWZ9%mJk7S7#MKKi
zbPmB%Y{Ge)Fz{Yw4b8le`BM=~vveO^A)a&KcBFm?Pp{0RzPnEf*f@)khg=N^2TWr@
zdG!*?@XJvfI5IkqxWW)=sHv%$xer%7^<k!0c_=3%4ZyXB4ohvvkJPI!2>j^cOjRJI
z=(;3UMQLg2iP#=$=7d@V`s}OG@gXxB9BhLkz(_#i>VVClH)=6Zt7#2>8c6Wz=;w*F
z&C0=wk?nE6rcnm|)Z%Gxa!5Qs=2+usy0lL1<^dhpK>?vo-GVbLRV`NVmA>W4*&E~c
zz)NP|3wMB4`ROiMTr>(Tr~8GAbtb7@^#rL4>u@#+)96?YB5TCnpacLHF_u)ek=K)Q
zJ|70g`=65kKRaQwlcp}TO}$Q+{%N0W$6+!3(?-x@8ZTZ~T6%l-`G_u^dL7@Oy{+uG
zpsxrhIm>%Kmdb-)fp6Fb*w|lcl_2xnw0O$c6|7AaV72VP*}IlMhbgCCNi+UF_}S39
zRV5KQ;K>>@@xpq5Riv&sC1NSGUE0|G_2G84r~muAc4yO2o_KLEsn1$x!BwGaTRKm5
zmc>L~$g9n~zUWVZCPTRoz?`m@fiSlP8NozuoHz6_Xiye9dM7I_4W>=fP`SWkj+0l{
zp3i%JBj8ILJw!&DS2|a;N;w<+7^b@FjXPsGHh_+sSEjD){<uMRaR9-*dUF6n1{Fll
z9^Jxf19w&0`Am5Y18#?3J3>kaRHlMuW>&LejXmE6<tGpx=cf!$YGZ$rR0REP%+UJW
zE{g9-ak@CXr`AOLH;si%AyU%Lr(Ir?`ELsK&<rKv%x107<GZEXrtX3*+-rtIu`iZ(
zku(D8@<ZBK2z=5zmhX^SF!+FoV6{gOG*At4el^@Y-GsDE%hycwWvbfa<4=nco9*BZ
zq@>av8fF2MUVV6k9XB=S<BNL{f2^%TP1a*pYynJ;OSLSGku+nL=i><&#;>^lHH>~w
zr~ksny$v+dQ7Jb(BQNl?KJWGpfUH^FcQ_+iC<UT(aBz5?mUCE<Oyh+d@A=Qc^swt(
zm2|o;Rom~Ev<_nosO8-3Ad)bHhOZMWe!P|xt6rzXjrq9PJvo4x^_->p6YYRiNL7b8
z-<fqhWCv&-#!uK8cs#N?uC2p0YdPUTjVt&#TIwoyUF1=uufdeuhpb=XB|f*1vc7Ct
z$iR*8kDM`d9FA^j=UBQF-BcYPG!U&iaa}TQaQ+omg|zEz_uoT3|C5VO8a!fLc|D<;
zFX_U?`K$76za~teM3_;4-k!}+8(CXh+soJ6N;4>*PY`5V00&g1j{f{9n#>hz8@R7z
zdAiZw`mWUcD;&A2Y+r<bpXx9Nz3ZB6xSY5hNk#hN1x`=yhmcZU!WvdxN7Z4aCVg*M
zUE5?n;~TtJ-fG5`vWS-0j5ruueT24z&#PR=Ia1f9o=(X(a>ggCyK0uYEhVcfYL&We
zbOC>~Qv!aa!T~tqXISbf^S5}mCufxxvUsXZ479}?*Z-`@gs`J<8HyTOyjbUB4%60;
z5dEjN2S`CQHBauKd74ssopv+CR=19C9nodWKJLe%FhTb}<wG;q>!qLU<Qb^{_D30t
z@49e-TD<IsecFIfxiX@kai~?pbaZO3*=%sZ!ud&Us*^b~PbPRz*KCFD*&vEJM(U;+
zM}bbzm@9%X!VjS!0)#&CCL}m~T2?neJJqrX2e4}Qp0C&fKyLN1qJjbkH|UEtc;ta&
zWPH3gYRD?5%juQC;-XAr6xgI=!B7{mD4kW`_Wa<$j+V9r4-&w6Mn>Opj(EEwfaXpT
z-a9jcTYw$sy!(4SQ{Y(N%#4{g;chkt7xZ=fK){FqLO<{w8`5{Nq<_4)9Jp2V6mfS%
zd3XZ9Z+c>Uyxz79jo@q0myPuXw96l6*c-rI)hnTQ6ajc~dX3WZhgfe&;qTb`88UpT
ztAwtMm(SU)2XXJhcx*yNkx(|kV)Kxd!DTkj1fuee)xbSk`t}d%=t{^aa{%+f(*)Xs
zi(%BnTF9Cg=+0T)^^cz0dg+ZaiQEslV^Oco!+(*{jVrM3R(pV!ED|0UfVDoJZ99ks
zZu&Z$Amh`K+eMU+lpq{`5+2v!lA!8E5Rs5ytIrN|ZRfmUJ+xFj2sG__ytbMbHnT7%
zgiJ1bn6y*dcUcY)Jd$N4FznB@$YCDBGfVT|*ZK$Ofn?u<n?Jp;V#uefeO&dGv$iE*
ze;inEEFGcI_!Twlit0Pglpc)1+n9mJ*~s>OU{`11O8KC}nPA>f>}>`LyVO<{`l;`3
z)9>H?GdffEO!l3vn`UE<ycnt|mwyQ3^1H9QSFP+X2Aw7^t!&Z6%VJRV+OB(k<t#e&
zXUZHd2Ll%vx$muGM{7%>m}unsekguayPrLMTZLhA3p&@jubD9$i}iZr?8(+!7%=nu
zsbo=S<_@>y2TkEGY6%4iC6jh3O^51D(>84Ney))^q7x_dAVP#l2Kr(g`q~FNglzc3
zbMBWzF~~w)TJrWMg?pz?;?Va6HXovws5hUkx3_kUC{l)@HP?6d>=u&)gOmhs_snwR
z`)+2EbUI9k2?DQp(gW^9NO;OLe~!K7yXf96tkk$)YP5%FH3Kr`&<)!sYDho7`jf*o
zspksU>^kh`u%@fQ+rsCxJA;)P>vG<f9e=TIqs`HIV~eBm(VAx;%qsf39!*}!|JYQ?
zA0Adnx)HR;W$7f!iSvk_{^JgQs3?4u@N|c+Dt7am*CBPWmFE}7*JBEn$4ATM<@$P4
zLPzzkrq%p!!!>xA>5+My`cH-l#pqps##*3jjjJLD6ua9Bj_VIi^!O`JTdyjg+}~o5
z<%A}zW>viWDH}@>zj=2>n4T4Q&mUm)z%^g_h;CnFt6^d=*md&S!k@e21io0twa|Xa
zet$p1S*g?dhN=e<!=MfAY+3Bg_+YgaUyStu<5(#-T09Vv4=6ybmW@*qJuoiER`)kT
zXNv$A1{d(U;MQxhI#4qxJBt*TFIp4yNHAKuOdg2_D@#dE;}tB~E?V`)iQ3y83B7o@
zD2CLWtsJQ}<Egi;eyVzfZ2)Pt_g`IyiLLsIN+QK{L8%LwoA82ekRrV}?c%P7UVOz2
z@)B`!BM)-+!K6`O??Ud*Z8rcs%+|p|ePoWFhN52iH|qL}*fY#@!{j14Oly51_hSrI
z%qtT*b4nwY`$q;q@CLKAh;2>759gACzZ%YLGgnK~=(rH%JQ>U0Xp~TG(BN%jqn89P
z92_bEy+kC(#Kg3_KUeLe!#nJ+cb$~T5p~Du=>d^<<^?{jRE}qZtaNP_w?_$VzZI4m
zelIgH@-ulIpD!4}v6=U$^AK`6fqD#CKAvfT;-N%4w80$iZ>~P~4bZL6W)Y2j??PMD
z1-OEJBK+<<?JKU7LLNHkc%ZHVp#c3RVnTMhjpZAD;gh{hy_3Cvy8|{7^=DqN5LI*}
znZk_8);I`gGsgF7@RpV6gIOlQh4)61I=|Oe>NwZ^NiQ|4+QkM;v(^`iU1cx*)oRvM
z)e0gOaHV_?NY0;}m|}JhLP+&J4|;B5Ug<O(5K41fSG*iI%s$hL0zEz6IfBHeEse(H
zzDP!2y2AgGI@&;ZYxH9V2Ss#$qR!mStV-+o(Cc}!UB9`TN9Ae&@BGVRpq!b9`d%40
zd3p{OH-FT59JOiN<!kk}Ixp~q+PnH@U{rxN(IB2r+SR@i&tYNfAhyo$eUQ+_yJ@=9
z<x~ZAlq@{+0tNNUY`pDu1$7gp=(AGbhe%kepA2S^WIgxpaDl8(fbUd>hCyXJSDI}n
zoqXw68WkU(s)!r6#!BFxbXEzHoQm}wzsYo+QJ+j%3e-<~d}3m{goRE{B#^`K!MB$H
zzDQ^|-3S%Cm2f|AB*!o^x6un<`t?+`#6iHKt8RB^aJ5piDr91o<Zi@nSdvD0dQ!9x
z_{^Dfu`H1;`7BE#Zv*|OX@-13t$f#A_Hsi*!=+9@4-2GUZJ)uWmldDeS{7IZ)v>#~
z%zd(k^w-Ac#WL3&PQtT4QBoS&-_3fK{3=;qHo2S3y^3oz_~Pp$`#H%T_ktH(eO$k2
zV`vzVbVz7|I~!Rt-f)PXQSG*O8^+#w`x$aoqVMyIl<K9+ws`#9i78mY^%zfF3w
zL8ZN*d#9m6m+{A1|NhCf-|CoF=5J7nZr3(VXJf|-!Oh#yl`wF#gXyVY<|Nh9Y1?tk
z)AD`oB_+td4MQ)BZ#7`@=<#?w)4qv(_2xvgN(uVdWG!|)kCQSpYsh!?Ip25Ny=@Td
zh(uIUXkz8F*r4<l(P@*`y>iEAXo}n3tU#C;r<P{a6@IxcF^2M>Y0;<2?GO-?Ak_qu
z-j3drXY>q6gm`(Si>JA5mvnih{_W8b`N&_2!knHJk-@d2Jci4zdC1iC@Y)dtISI)=
z^3yG_&7X`m9Luk>_2#>Jm89fUMpH8)+S}`Pt+db2LH;9L;eTfpa6!Z06B5`vpZ;mT
z=^71qO~+C4m+#}}uZOo{zda`%XpNf7;0wPSoo7&l`Ay`LKv!=m!_<3GXuK!BT;KUv
z!_uFwLwG4U{P6tR+c|b;5>6j5V@}3W5^P}o3mm@!bP8Fo4!l#Oz}{zCc$}P^4ckH8
zKQu<K4rVa@9b0M_ZsQLjHm=T<QgZP{-qIm|?4b|8k*%%E^^VgOaE+K7Yn(nFi!S*s
z{yqM1TQJD+JFv#Yd<p7=l$U*CRttBxL8UG<f1e}hOcoIzX=v2w+p)m4S)b6r$_GD6
zz&L~~9VT7&nI;(j?IS$htUEs4y0)C&9e~Av###B?*&Fk(Z(z_sgmm^tf|eF~rCjcp
zc@!>5oGWb}vjc(A!Y;_Dn3S@|q-{%;z2s8i8;|Q!(N)LLQo_gPO_c*^6Y56-zc40T
zBNwmxw2VCBgs_;-fhDoP!LgR}5B3_<S`n_DhLN~~@n7LKcOnU);c?UrEoJDRD3X%7
zn5hWXC8#h~Y)2oE^#|qURi#`kNd6$%Zhq15OR95!2>HEx{|dRqchlrErwP|aQcmt`
z_7H}`8PlX{3+l0YjP@_WPW*6M6E2+JG6J!JY@G`Zbs|UFGXLoK#Pfqlov68R>>s}O
zg2z!A{!ad=_Wkf;`61DGMEc6detq90D2BE`y{eV0G%~XK^{Hb*!ok0(a!oESZ@~Y5
z416%IsZoyBDB<fY%u!F@MItIXR(<v~vqUV^K+}!ZCI2Ptj(8;dbtN}j@}@-o*?}ew
z_J!AZ6?nSZ{e6QKjzj&HZA$6dgXbz6T&LW+e-I6)d{0{3_`IvFAt(y%%l-LXjgzyp
zL_&4$8dY17hm$>PTCRK#!Dn_Ni03hQU=Evv*KThLPD@LN&*tmxld32B)59Pe_BI^2
z3);*a|7~NAh_7sRlY?oXaG@ZG9krrxa&Un~vnv288}?FlGMN1fUhy6wQ<iS@<#scx
zVIHSa$rmz6lPcu)lC;K3Q<qk^AzN2I^5N9b?8e2ZEMVnpwz;&8;30{keEkA<Vb+R0
zHnuXTMLvV!FP@;FM-rNi$cI~w_D07Sody^9^c`oy%PX-<ywl+y(^<r!`o}>{+QVWD
zh>zcvzQjt=<%wgNQmgmlTx}+~YyK8CrwqRV^@9CNQknLQTcB!C-lZl|u@VYKmpAGb
z-M5y`?-(e)vB}czWPj*uic4oVm)S>b7MB#Yds0cGWv+4g8?3kVBd0$c-4cEN=Rt!3
z<JnnZbM;Gqs9E=OayHUScFOvw^>ujTqAFe{_8IZ3egTyJ+Z+Vl68^7;$IEgUDC1BF
zn_CDeX@)4YwL=<Jy_LN0I{qAaP=5u2F;%R0$ybDp-lvC%!9n407P``fB$*#}cU_PM
z4yzLXYXM%nl`yVdHfky+kW&uPxwDIF%vExQQ4@jB@+>;U^Zr-IYJ+_ffnUo_BLsWz
z7zuOn?aCd&8>}w(J2v}H`#}%;)t~ORv(?*1RQC2w%B;bXmt~)gXiH1*Vr#Yzx9vSC
zUDaOr9MkH*3IA)lSn=s=nSM?7+UB7(0B!mr>I#%_8Z`0fBa#zl{Dq+SsFn2DsvQPm
zAOG7n56s@(p3E+{ZiM{i?u}S3N{bZr#_4(l)&9P?9fvqOf!To4pncX<2O3!6RTz<g
zb{$GA3{+hP1nGBC-3e%Tx|C95GC$<8z7+~q*L_w=#q@A~ul-|4HwwmV3wee|bU82^
zm&SX-IWoGsQ2^kI+7f_G(gy9Oou%_l-WsNzC#LQdsJ)-68QT^S-g^CS?w!vMHwWlz
zBfm4Myxi;XMu+oOP$o~HL~R{4e}nOZ;MlDTqM?6)u*3fRt97_kgLuMT5>?5V#@|S8
z+xZvSybo2Zj!<1JwEe~Noj->nt@uB`K^px(_S{3qm)aeT1_$NezHVQ>qot{L$-BRT
zzeoF8Z21$ynqt~Y#(&OC^-g1YIW8katA2-_*>&{*@YJ7l$yv>pUj2NPE~ecnx^3GV
z84my8HjA`$c9sSutTw83Y2oE-PN!g$S7F<8$0xz=F_cM@{uzeb@oi$;!p7D0d-GqN
zBL*LWs63weL9BCL2T+6gbp4)5USS&U_U4&uL!JyzSaWq<BRQ~MNoj94XqbamWpz7S
zSmI1mxAhS&rf0zRUqgWEG^yT{M9?L(q9S8#<(N)kgmscg!W@5XH{`kxZ!0yiVzjcs
zO?Ufr9#*pfHqy2(AGHYJB*#o>$CN<q43iUdIUxGZZ5yxe$Rz3!xA^!#SK3gm*+kKq
zC8qt6uF>4K$1;+%by=;s!!{$h*5&cWlC>^XVWIW1CKwKeh?R<fSb}8~KW`(){OKjN
zK1aH{T4Hsj;_Y;53Fw%5Ep*jME~Ln+*`<J6#W223gZloUJ#D+MeA6SrZ-4fcJ(~~c
zC2Cq%#NvssoJs$NW9?a+4FBD{(8=jp_dfp1O$+#et{)4NEt8LGoezBfllR@1_tHz;
zNxTkqA*KYn#|%m!!SKwQe-BaP95E$MW57|Orsg``THEyL-m@%X5wp`s%|8IetEcyO
zS3vi+Klx_?mONC>p(CxL>{m;OZHQS~IFUj00q5U7O9(%{%mt^<_O^o8^&?fJpg)Nb
z+cg1hKDm+CMNSUc);6B1s%4}3o#Da56%@x-^L^RZ`#he&>cLwhaQDgzxrpPQ4zNt6
zje6o=p_hj`wflTuqTkD>tQ_950m)oSB+3$z{05PNLKOb0q2#WLK`lbPlhu`j9n>Go
zKnTS4d1tNXVd={kTd>OIAaBJRBvg!n9AZjixauee9Hjkzd$**}?h%H`T3p~xUJoEr
z@UFBam%$^EXn#Zp6=-IDzP63>Q9C^S55iJ{SWR$30;Y+H$=dq*Hy~!Gw?0{mSk2X$
zpWZ3q%Pnup_FN7Q*A4oJ=?|ivT~5=(4tKU4)~g!4xfUqpDiH?l>7ViB;h&6<SU?)1
z+lu49Y%kU^)%?^g$4irK=hbbP5$c)(0y)bInV9~^`{{T8SuwZU*2NnXZEkxlO8EY=
zMCjH3aRKJpc*B$a2iSr4T|Y>n7CpRuDdKMtc!`~?$zF<rj*c$1l_h^)@2y??pz)XU
zWa+22qDVls!t@T8s6XY+%}w+6rBDNh#~4u4?Ew|t<$h5DbZ)0g&_nt>@<T+8j0abt
zK9JP7Xla?1Z;zVpt$ouISM)m1EmyOP#cc?DLbMmXJ>}4zeP1F>;lH*X3={Jwne4o;
z8^6^R^t)F|o1aWZTWY^qy)c1^<-zGxC|LH0-H)fR`x{YF2`}D6#dYZH3-n#9V~Sj7
zHnv~fg<Tsjn%H|U>_sRZRxu1*-Fo(lExBZ{#RCkZXDm!!O#>^1$Y{gN7C^o#oB$yD
z*au5q>9Ajp0^wJmYw+zwvCI!NpPNCi_rQ@X6a4x+vf?E+8cxUi?>yBBJIbU5)+tT^
zh!+D-%hHoaw$1Zxe7A?|zw@T^3!wI{4>c6Qn&y9zAaeRH{l1yRa>AhRCg$|odR%as
z!wZYI>H*!rH!Wx6SJ*0z)CCz%t7liaPL>1T1&hW`y8dn2olmRBVZnF;xsb+@X(}x(
zKLRe41<mH6)#c?;izD*9!@c&$dP46@>KzuHU>WF`oY&?+=JlVCX&Mp9Zbwe!tB)sK
zRegotu2YWP2rP!3uNhS2Xqf28rwKPe6vp}E20@P@D4S8W3EJM|<4f*`(UYrzs9d`N
zRd7Hza*x(9@1EDtzxszKB)2!*bf|kn4E=Bk|3XQ>)9+P$Zmgt(Lmuz;E`%b)>Uzp}
z$#1trS*PtbaOvzfRqw|tXmv+KV%60(Y^mI|d>aVOL(VIWq6ro-%J<IE%E}g_IG;FE
z3b!)xwlNSUBgd8qBxv+g79Jabf-MuTc&l^<SntRj<e4dT5H-paVGn6HNb#4*M&7sU
zN=wy)rFP<cKTdL0(X*iWyFn-zUFw)L>fdpDRFc#&OI<=t47x=ih$sT%aPac&<sg^+
zGJoL^%)!8s@pDPGs6*1?TX04f;b6DEWB)Je{yfiZ_@4J|R76rvYG`>)XAm^tRy~^q
zWVQGj%BB-Mu&!;HLyB#eWKLWoB8CV+^2N%R7sMY<DrVJ*I5s3M_MCC)W4*PqS>OjV
zgsx)wRSzmUqyqP%EZ_M^IZA1!3$1(qQ~~u-Tt<FCozBfXE6HZOZ@5j5NPz-uzPZYC
zQUliHJZfRTTJI=!Nc3FUO7r9@sEcxqoY|co{P3E`6xT%9!PG8Ah%Bv{&%z5ccziyL
zdd9ya%ic&c3s1N0J&q$epN<iaJ$W$Yk?WsqKV)*4V!1LA*n%Iw9n5mi+7rc7i1@7c
zBw_xL17CXW_dn3(y(^sFc5@LG?c7@}_OIp4g*5&3*@&FAmb|KtPtZ1-FZqPa`T-07
zllQ+lX#~@ezTN3AM<W))TXJ6eThffei9fxAq}Z%_8!e@IQdh8F9+lLvTM@A%=Vdw{
z{D+f{&te?$8fX1Fv#;LD=xn079WCQO++N5XXGfzRSNyF^SdA!+U?2!=!g64q`k{dP
zrqh;aK=k&Vu*kCgyfm^yfZ;h|=e-Suhr?oM#}mG4iq&D~deQ3;nIbun!iTX8=}(_n
zW;EyHNZ==*?12VC;dlBzFy)}{B>QQFw2$C4EHbhpj{@!1ApiLw*&r+Ut;=Z{Cez9^
z9ZZQ{oCNwVgwI9FXI)LVIhyRoAEtf(lQ2@c0tcHuP2_s`zW5Z5|Ld{)cTbbGoTwWM
zPStlFmQxne#-o985kZ<lb<=~THY1q1ob)!h!W5Nkv7S%%)7&lCgv;Y#nQptwK*&G$
z2eb5CIMFY8;5sMiuT0#(KNo*~crIomE7H}|i==GkUdP49TZ{sbM4INqEgYMalG0Qo
zm9ryaH&aF${G%YmAtJktG~mWw#8J#Tz2;sbJH|Ot5gqRpIwAYi4rXjt_kRuP=S&S-
zooA5SO&a`;<4E&Z^x^b0_FpqBewBByqri6XTQ5Vc-8$bAJf7_IYdLVw$z##=PE3~7
zLgCB2fV0zI^C~wt%!N26*c5mgCYUXC>F*uAeFr0B+Zb<;sfB;@PSEv3MDn(8`h@%V
zumbZJ<3!_xq<BflMQd(!g5fSY;;dC{&qdquABBcr5M;T{i9ZG|{ZU(17Y?(;cWimm
zY-fq7S<#O`yDb+lCL}6pXna&AI5`wu*ZD&$33iqbhNTPpnwtuzl8tDxI*%zIbaMq|
zyn|#zS{CoSR~U^T>_zL`eo%7`ZK&^j*smD$;ErIm!Cp#dMP?gc<agA~6u#;}g7x25
zTG0QsR*``+?d<GqF>+3AMG#c$eQ0#;3^yY3v3#=a5U3W|xZ>qw$COLR?>Vm#E;R^f
zt2S`H>i!!@Ae#Nq!W-zmfp>lLCD%}f8o1}Rl}RVB4zu_E|KaK@!{TU~wHHW$B)B^X
z!8N!`!a{I@Ebi_EcMC~y2<{GnEbeZ>-Q9I@cm0OvdEax+_4SWs8HSmj>guYy@2aku
zTip{<+DA0#Qyh|^G;zvlY^oP*bjDK2Asf*_swap{uq}5J=)XlhvGymw@Ose>_tyFN
zlnp7AF@kfVCd`>>%{8unt3FWpk?XUTkkQ3$sc5_omCA2&(Ab5u6bk{ijh#jBx_wZ!
z_Q;hiV(9Y>R~A=))L_X@!Vx3RpEA&P!zBv=4)>@7z-V*yWh?2ew~j^eU<pZYZx}hh
zc=GlfhvS6V{rv;hL<#xNse`%IFuL)tPl1>kB^6_g3GC9z#lJTSQAm`A+J`&y2Rrvf
zSdI)h=`DHLJ7d<rxM{u~a$7xRo%Fn}C9onJF8A(^ee;V-mDVDynu#1sj!JyJ{+Bpf
z%Cy$iw=+%V5Wc|rK!=U|8+Wt)xwmuG0#{K$--;#Uv=0X1(gNsL2C;!C=6{#Yf1ve<
zz6oHK>^4LfZtpvBWxvaV8HJs}oAI$D5GzD!P5Oxmp_!3WTT8t^(T!k{gx>ZjMH6X=
zL5YZO?^at9_%!r0-%j`$4&0<~U5WT^Ku`KP=<{zV7b3mZ{{%##rHb6u`(tYwn+(C<
zCzw37$4(fBP1DfLmWmc&RRZxv#%GO|{&91SUQNi>OXW7A)=Sy?dlcPPDl{_5&}@^*
z@+qLlj+}`&o4kAD^BT|bSmp8B<M99L*RzsoHvm-eKP{g~`ir7(NB;%-$wfe0YlJ%(
zx0AM?dnbU;STQD{oNS7=KdZn^3FlmHu7~&&aIK+G!RG-48nnztGyQYLMFg%~*`HoH
zFV-l^I>{)<WJ<|p;D<s~4vA#c7^Ng1o#fb66eVgbWzZ<IJ@pglnz;2V(w<{P^O5_K
zEOh7e?DTl`Gr9gqM%zrI8~YJAU<o?7SAbgqO__>P(jC6l19V?BWcp$S_3fvpPmjKo
zlJ6l9!B&1zPT!7X3C_iCb`!ld%Mx?d#MGjd2;s*SK^h9E2|#!J(|{BoXdRH_fv>{H
z&P{y49bSNjKiWnCya%FH9>fg)6=CG1h>swGIPvIv(Wo3^;L4t{doqr>%GGZt9f`TM
zhAIi%wov}<7hs|OX_KJ@(X@a)TvhJo&fwGbMMi77y(#2nQihP0KzHY%P&q}&g)=A>
zE@iy`C47QzUs+&EV!BR^%n}uJ$7a__X%&~tV&>Nska$C6F#K;nDWLSaaIPwlNk&bW
z62vKjQqmwLp45G}*d~gy3OGQ1gaHw)*?|akCx7N#bQl)M*3R74KOd83IrYZIgoyr=
zt^y3MLhKialc_mO0pb)-DLD_fkWqr=fEbc5C6ma<=?NrE2${jO9>U7^wWS2LE_Cd;
zcXx2aUe|}^?u#lP)YaCn{eY!U1E3t5_FoB7m^yS(sNwW9h(aVbhLfz;9kcQ@g}vw*
zB;`L|y|R4^+?1F9IbGqOHuSq#%nQraB|lhFB|7oqHi~*Qd)pRN1HdGn|D1aX3{8F4
z=GK<M)eZKxa>*Yqwd&k^Pu$28{9k!n?><V3#*di%Fz%>UvV&o!r-m|LBw6kac3WU$
z`PGJkzuyGB{g;576IxTdfkK%YT3@byY~zW?i755em|M^~RMtX4qe=T1;*0p9mH>lH
zt)X<`&-K}i^D5}jdJ;YD(B_K_^}U-tf6`6BYE%@ws^q5{#G<MT2DFm)5f!gW`;ihS
z(*Zkxz<d9WpY&jtqpf)c8`=I0Fx?E#;s2E#RDAqDD|7426zFw<kf^GMy&k_n<YLtg
z4dspZa*`jdc|o$DK4tFSZNArS=mKW->j;{;tYROGOD&$_RHZ0p;R-uChF5K5-ewqT
zNO#OB6LViCm_`PXPI&+1Z*CQ6r(f^B>sUlDo2C0Q=_wMsa<By_oQ?kp;aeH|7VR2)
z^ay=v<|R|-Y6mR`I&!-h6=4Gl+4p?AyI(3Fq>3eaB>o0Xux}Gjkw*5i@+8>94>l65
zg1ZR8uAHsm*IqnHUU(NMA4QtKOV_>PF$fk-NkS!L{_;zAr$wNqR~BP+7S#J&dEyGW
zEvS&T@;qFml*G5#X-?020IjlMlEj8t{JZP5C{_`(q?vMoP2?T2B$F0kXNF{>a^^tw
zn#%U<wuF$(CW=-=JHC@G{U);eL<Bu`ItVMA#LKT^Ikc~BNdT3B;G^J0wx6Th{+nOd
z_f%d=4+MquZ-$@Rw~C0rZj8Vgb<l|P_WpD!(lA@j#IASV32#6|6Ujz&3Z5`IY`h^j
z^NE-*)HgGuh2NbLo8D_!FK)Zi(|gj1b*73^zGM+`IpY5oZmnt7FW<z2neKC~Q^rO{
zMuv>`Ml=8|b>uG2o4{-N!;5QN*<ov2T^ARZ7eA&%8=Oh)*2B0BH;z}$zOdA)@qP|U
zuZq`64?T+aLK002=37PmYWd{ndftYFp(DHz3`CN)%{|W!J{4*{Au*?fe^)Jf>v%YZ
zdxchqE{e2h;nGq8TQ4;|g4^e>CBM5o=vq2>T08#}Jy$K(@7<QK?_XSN6D#&Yo~Ve(
zufr1?UOP09q_Xk@0i_VCaHTVrw~XKySLdfiaWJaJYkYiFfm*;_@6LBLcO!ZRY*~t1
zbYlrnsV;0SVlj-P8RXj%AGpSKkA=ewz~$aHW$Cmuk7DZPkfFv!0sOAiq=1BZ{OmT{
z>`0_4iJOnz8v2kqee0WEB?n9w7+Y;!{aepT<O9#kKtA+i>#&=eM!(O~@|~{>+>u@G
zDyP0-PEAuVT0Fe)=546z@m@=6-W4Q_czD2z%iAA+Jl=}?Dy;qX-SVz`UmvbwS#>Gh
zRyT#^GT)Q5IX{l360RS`s>>#22F|_EJ6YXC=TbegKP^I<o-Oa1t&g3Kdc_X2)oL=1
zl2Hf%G^tA}8I@&yPz1Gjcx^RZ14nSNX$_=AYzbyV*^Ey&V(6l2Y)b`F72*a%@_?i-
zKW=}(Az4rWR!*QER4ckXEIr7yV<``#fyX%VIe%p`bdCJ(;?w9^myVg&aMSTumw=76
zZ~G}ElpWpcN5a!9KVCfrD4+|3(kU%}0@N!+OB2RCZ8)(V5L}3T_xtznwPkC~udpI+
z!|Ol=SHN9;^=Q$gFVYqlI{r}ac&KQQ2qeACY6-7b%>L-ly80+CP907v_${wn;G6B%
z5cn$DeRcgRuYGY)ZLQhu=BB4-pyMR_^YN|&_J^a22v*mg4>E}Yue^nHHtUN^@EHI0
z#R~sjv1lEa7eZC_e}jdERc6tt2-Mjp=f2trL0?MNVa$NwDU6mS@SZt%avr_WnyzO;
zTogineDt1s;J3ei`xCVp)wYS9v_#&scWd;p(Gqldv=9Wu(?;rLH60J8nm(v)>>DVq
zXXd}uy$z{zB_&hCz<|i`@)T>)+P~al6FOYt(JzxN6n5M(S}l7KofLsRi@!<(FR9X9
zwlJ(|Ke;8u{F5!VqLOM$I5{&Cd$vqWC!~Jz@-<_13+F4Wa5`9(Wu<ET`Ye8>Z#YvK
zVdoWg$(wK&ora$>MQEFE5MzJahieFuvRGa9Ux^k<5A}&;#IHZeVPHm5Kb)yM2o)Fp
zPT=*hDSEZu)yR^jI;v`Y?Kzb)YUOSZkG=<)=}mqPo<1icZ+w72ra>6d+k<Q~D~BpY
zElwUzyBv*AQ<MTX`){tl&dnd6f34AUKIl4dy`3$SkdmXrW9;(_&Z7}9cW8;JL^-d`
z#rIaLCunfC>CO}lutL<L|D<<-(QrW<*7V?M`O<Le^LUxdu#cIYHouJPeVc*t{sHkE
zE&O<1b2uS+Q$*{pde`+%Q<CE8bRXMJLv-N`rjk2TZ1<~O_1c5S$ZR{!<>cc@Amiig
zxMM#;`0MqxVLDlau4=JJJ04T1L8sLB7f3j>hxp3GN;FO34TogGz}gyr{wdX*JWIh9
zq0)+|_~Q81w4a$#`@X3CqLv7(xBYj8-0FAc*@t@ly;e8p@sS|wDf4aSiHFR-@sUo>
zU6m~k$1;B(S_dM&gSr05pNBzc*}(bpY%9{gGZ_LHWJOJzBY=P+Soz&pcZ64#;y1A<
z%E{^R!5@1Ck4P(NCx+?yOY*-J!t@klel$h{l2%^-vlhZEm(s@z7icq^*`xtEJv*BW
z>6P3q3xNT7R9;kEHuFr(pGZB2N7%?2xR@oA2ev$U<@v?unYE=BY<E4#3vT_Nc9#&k
z85%nWy3>yTGN2vM^d(<a!HkS#^b=Rn1=mKJO%0iG;|Lv(yf*8rM<XIeFH5FtBQ5#{
znW`UhxtAuY8yXr4;@6_B$lB9VJ<#!|j>dVHXPOZG+wz-lw;GK3ww9A!@W>p1xHkzD
z9Z1jYSXMmYzR>Uz9?MjM;XH0*YbQ<=qD(s6=vfEL1r{7n6%P0SgapU*=fR`&4t5QS
z3RdB4JNq~w3VUe#)n(cItIN~y%HgAx`mXU;7auv6r;@hC#gMD<vj4<s`&v){>8Og_
zGvWFu_20V8{8$x%qLtg*m#(cW6WmXSnB7l{ZkL)j0-0@{?qy8^TSMsu2a`pb5o{4R
zN)2%>T)a@7rjfsb<uQd87Y_KNrDPOI9FuXwsbThJ%cbsZ$hKM|#57z}&Kf0AKthvY
zdwPo{)lW21IR666vN4whm0!55{QSf|#mgHbC({pS@~fvA{5BOHs5R&Wft;4!&b5!v
zn9L}hQu3qw*ZeviiF0P69H<3Xa#%`_7l;qeEbq#lA!HIXw`aZ(aAxu^%U>FpqGKh+
zcK8Hy5^z0NDwJm8Lks<fQze`Y7LAXJ{G+!SB{pa8+hUu`-G1{1_9lLBvH&RjiL4T|
zbBOt#p^14jGC5tfG=BFwpSIW?TIB015^yExxsNj>o;n)GUwLFyF>paT-`bhG0)&1^
z9JjBL&=1GA2Is3p+e`VJ`zO&YeZA&u-oK-p%v^uYc!;!SHyw@@>(U2VpTqNaIEX-@
zCgr_7z30;f6~j5S#zZL-IkeqkZV#IWjhVm`Z|no$HILdI1&B1DJ<15`7;tu8a^^kR
zZ!1<G%mow{ej^i-P2wt;Fp(JU0qpIvWxwFvxtQ9bMyz{tFdi;xNd9<URrXW?`PRtj
za9TwBj?&BYJdMiQq1B|+;Kqe$x<YTSWMcs%ZVuaN6-^;G=terO<>v^pu_+)Bl+|Is
zbnAQFEeGav^n6D5<<@^|StCH(`L+uPD=HAk_>9)Z_qjbftQuZ5P6eH@D}c$Ru(tpJ
zk&>o^r}T3Ik1YNrza!l!E9;N`wyJL>wS*laR}{2Hys4p~ZBLKaNG_-?s(Z0(>+2z^
zx`|p_F)o0}<5dgE;*@mK))B^hgk0oS?Q4kL`6`bJ>m1(1e_n*Smi+11$oacD_wmw$
z8qmUDTP&>amNM7V`CY14X4>9S#W}<qTc^cDt&N;K2O{%%B%>NZjfJ*6pp0}VwAYH>
zUU3WO1uL2gi$*%WJkzrg-Y`OGe78~9k)g4irmM6yXWhe-d7v~UB4gjTIh?g6(i+QC
zYl-jKD0o1of4**b+>tGx@mX~qD5kq9ab(hupYE9<ds%q<V7uV9H<(LC!_AG}EuY@L
zb`-ysU6W_Fx65hs9qF)P+QbbrkPUksB~^g*2UUNZ0brJ^s&tv}$x`QH7d*P-3Jd$y
zB^DMI58}>_Y0d5`Il9L>jH#*P-L2Xtt5(wc2V_jd-T^|c$BLqK+g@+x0DXNSq6>jw
z&*X1(43JNeJfL+7tkAc3JfD=;&CR$`$(;Kehd4W*GOs+k%Inkped=k{<D&-?5Zn$8
zKhQ8YA5T$`^y9*$)u?ynwRky4w%n_eArt!YT$>j`tn_tOow7khp_BjzpuLzol#*Ds
z-&NsjU-}|rT&Mt>B<2HBQY3WW+8c1&l_qVShUW@D2R7<oKdWtQRt~>bl`>WKKpTZm
zi2ZD^edpjvwKV?zd|7r+J&6+8khpFK`QFsJzY9=Fbz0}|Q0jc?S_9SbY2)dVOtp;b
zDz4{cTLZ3r;apploi=P}dmiR2z;P-xQLuikCf$8`(9VA(-=1}Iwz9K&&FSIpj?2sd
zyoB5VoicKFICCB8*0L$GJa!+ee{*wl_o8gs<C)6oA*7_G6EPSK{(gPiuJ;uHz7?Q_
z-z@1EiNXE92v;(o&=yti$-w<Nnx9{Ag&V}xFFEL>{_G%zmNN?VEM#6u@eTa`ZwQu`
zNqBaC6?)UBV@wOl@HR7Bwb$}Y*+x)Wm0dM)6pI$mG&`L!<FK7Vd2nC!UG{l=mF~?-
zULqvi{CC(TcWjKsXUopX$tfQx#uz>0x<h$?$ps_b^A_$ob+u5doIqcHHQ{7w9&P%=
zh;YW&8^upqi}v`;I63I(DE5pQxOh{bG6hpRI_ByQ`z>Ds-}H3R?&lc@wE*~i-ThF2
zg8gikv5x+z{OvbhK;y@BsR|5`znCNE1EStr-?yN(-p0KCFY#CBr)&OWC5;~c7#uCv
z01IZxC@m^Ucd4ix!6-^&iKrM#$uSMdsKb0TeC_#ZbT7~TrT&)0hL&o1b*H@rG+;5J
zu@g-bt~zCmv+JRxTG^7hQgJ-o!dP&A78_6#|9RDs>dz?{TCZul6CNuazqU1RGd<-6
z1Pr+z8oE~v#W-HGXbItyT35)#Mfn@vU0U|bQY7&@-s3S9k8gEHkTaEOxr=|4uK}Vw
zy#HDv3g{)cb?n}8)-{`P7|2PEC<?*a>vRSgUWi0W)bOMW!~e&6LNkF401Sr*ENlGk
zAk+(*6L5H?4rodu>O8ALw@%{oiAWQiG4t|xfq9VOp&GXN%3+;PWi=Na?}b!>CUs_}
zzD8MeKyeu33yOmUq$4lo>)VHjC5LN*qNL9mA6-SQ-Hw`{tm0<tU?2cXRBi5QR6B29
z7irdvZ&7ILC|5o#iH(lJMv5&}U%cdr+j{2A^66dc+{~H+81t+bh3)IBRUsE95C|1g
zUO-C5t|RLB_)yrWzwqC6kO2@$oT2TPCpNzE&Ss4G(0A&1GWv0ud1tdbDaL+7?Vq7G
zK*5{T%6bzLSV@BMhYTQYvY@*~x9%t6PW6<Mq^S?PNE~~9kIlw=boPde)oA2TPh=0m
zKt9BL9EJIWls%o0dUD@;dV0nA{`RN2S3Hh3(^&FCci*ZT-8<O|=PR=(ycpfwYqM3K
zzExfdKVBmMOm0Z5`q`cB+yW57eehV`9yz&N@N`%`g)Uo}atAeIL@cX3p(7wJY>t#Y
zTVH5!NJvg@Vnxa~qu;|+P+Dp+Omst`cX$^fzj~5|;9I*vcyMYcD=)p8X40~ef>n;G
zF#z(%sZ~z_l~aEqA?-F+@DJ3Y@rE7(S!9A2yG@=}O?lWATv+&!w^vY`9pf80Ba~q$
z91!87R3qp5Sk-%Ny!krR<B<vg8u4o00Ak+u($oc~w`?ERP2E>vBDoxl#o0F3m9*cf
znyz<;j>GegxHoz29><0fDU2X6-fl=*%`{f^RJV5zuQZfwne|76ruweD^$7@P`lep9
zEC|+o+=$)@l%^#^zSryOp7ZP|^(eS2dbWEMS@vvneX11MbwWQe&xa?cqD#dE^34@v
zSZ1H;Ol55X*F5^EfmsK!rL`}hi;iPzgKg4>7j*00LagnP(bmgl{VuTL#jJUMf=V$G
zZh4Z=C?)VKEdY|G{o^5R3VMP<>pR`sVHH~gNgr$m3P6a`QOc^2$v3Jcol^K~u!vl*
zrS{Ql2@*-ZZxw%7dwTnSBp1M~HfNe)j$koaaDHp6@+=#^`rYgYvSsnS><yS{wnK1V
zns`%X86Xn3xVp9mPsNWE^y-Ai`5f6>vw|p_%#*+mt`h;v=GXBKOczHz?~_<uRM;8o
zs9Lon>o<#+>!aP>dIJeGAqTc{nJ&{_SL=DX%Mh7$-XN7hB7>>;WxauH^flN~{L1iW
za?z*enwUEK2fzy4EHifW#WI}igc8nFS#oP>Yeyy|1nR0j!OxDri3leQF^{I1U+_6O
zg%Efh={0hkpvw5b1`Xz0uC4G)xGeojw!Hk_i2^a8rn|P~VRT+zULN~vjArd)+$DEY
zJ|J(|>g>9}TRY_8?h>zAPEL+qYpHt2HicLH19e>TKCq3&tFQOyd1D+V=B%M24lBkU
zD4Q{4_+jnL{SD}&yB#aWZUnVhkUqjQv+dNbr|>e0QWB--v^@2_I4>4s6lE`NYdWZY
zjxI_FP|_19l)18^sTmn>cJ71sKDH!wR~Rpx$cFVEWw}2@F<O}b6WM|kD^nKN#|n7;
zZ(oS(#_H!XgUP2-^D|+!niJccPJkgA0E;x$n@u;rcVFhpQPRi$Z!Q2v`tUtGkfmcq
zW;#7YI%gz#{`=MIqJYZ-{?$a)rPjbq(a{fuauANhakq0nOrEdT7(%H+sQk`{_Oxo0
z0=FXwu4d=7jn}u7o@@??TC&s=mJ0+3M|JoSg6F>$?-cP2meH3yh|z`4k((c#;Zz>#
zU*pSt{>x5_k=o$cd8+H(J@T<;PH-#|o94KypgI69`Qc_$*8>xo@>$~#&cd&=yh??F
zkQU)DOo9|4Rqv7mQ&Pz=pz88BfOJPWW$E{Ln3VY>qdIE-%4^~bU?onR8wUuT119)K
zCn$-+tdi6hWyKjl#0CcQO;9LazaOJyRFBJL0~A^^&)d9hH$mF^07#hW=pdGouHHD{
zGBe<5UTH}bFcjKn8czGZ?>#QN@}U|}8df3UsSP>zHI}d{_6%|b+#t!&XKLS^hJVSU
znJZku4YJXqlq+Ziq)#c#5*n*}7N<jfn(@~uQzyxhQ%x0P*l}lL#MSY3KKu-lSp@+b
z%*VSW@s=e$fiS}PYXMb9i6V{6(a}*{CkbG(Tgv_a(L!7X1@!$!gcX}6{`z)G$$D9I
zoK5&zCCKEwG9<MSy4w%DdTB|-cBdQ1c4A993dd-jB#P{LqYu#RsmmeSSZ*$XB!)4@
zs@aS_XnO1*R@!13kdL5$paj29D{VizigE>!v3icZ8u|*-iQr3LME6pO(0E3<tT^
z&7sCHz{*VGkB>m!LLeb1KiM2EwVnX-AqMth{QO233N0>M%7^Y!N=6kUeJyz0+NUCx
z-!euxW?m)H0V<5yYdrH$07Rlb6LoG0TqVC4&W*I9hcbtjFQ}qNaf~ByOu>Dzgi0#h
zDj#nSPLQ+V%E7A{#D;-A$&bSn;@D+^6&Z<x-iMa&|G83OL-GgX#gdC5iG$FCGJnEx
zV9m+`7aGi=rMYgG0j%Kzz}^@Eh_We?Z?|er>(jx~sK3}>3CSoOg^(>A8AJ?~zmHqi
zxN}UY|MV)KateJG(4tBQWhMgS8)rd@^`2p?AEpn0W3<Ne%i5o&mTNTZ+vwc_4fX?L
z*kWj#P!S27X&aqIsnteDqy8cotXI<_sUWxO<R*<y5_CiZo{BDkTW~d2l>h{%iZDLH
zoBnZ&$V<p_NZ;c*Sn=N1^&yvlq&gZB8G>iK75t0xPo$8q)o-92L$>Dg{%W9Yza}(x
zt6XaMok{kc&iq?vb>D}<wFD&~(T6BHD{y6q{8rk5E6Sd=&)ym&ufESeMnXF_91FK1
z^uUjtnMbDl95J`}*+u~yZ_5&gN&Q*m3fV+cnGc6Uj;H*S9QGi-eFgyh!9TCUfL)5u
zO_?yXTnoQdOYk}y<cjF4c9rGL7il8r*)m4bYU-Zo39=0dWu_LhE6TSZG@fN6mP}e2
z?~QX|oC@RUvv)&K6$QC&_wUIDmk-waN-BU&qFESK!suaj&(A_{P}*}-S(&e3j?k1=
zsaCxpe=fzAS9<p;D(CH)yJQsj%MNZ6@%?$ywCBH_0pnT}3X%N+-b6c>Lz|w6Lp62#
z!k~~WI!hqi^rjZH!U;-O`Kpbm&iJIS7~`n<15gnSuBh_*@vMbDtjn?mt`XXPT)el>
zLV&ihhNQ@oh7je(DmD;=Z1SnhcZ>_Q?D%V@{h~A%1hIe&zR14!4T3@sp?|_8lv6ih
z1w8+|eUi-q*e|?Uq#ZefN9cn2OPK=-QPrgB(@}nH(EqMnn6g#`a)ucn{Y}qBm_#=8
z@8Bxw32f(pC)%3~wQ=`t4+F?F+7Q?=>2PnqW$W7Dnx&npBGD_+HXHgT|41F_&u<Qn
zH&)-BtKJ_^`sw~kP7hgV6d)N4?da)3(ErM6u>K?U>!3^9@5&|u0@wHUdf$>aehx}D
zj1IOiRZ|^`8V$X4JT(hf<HUS~vYcw4eJC9oycV#abH8)`XynY1<iYg~$dq9=()p`T
zg;o3^BkdjXD+xcJW^TnC6k500ZTtyTQ+AdWR}~xs2c?n8ikDFi<CUfY4_aO$tfjb*
z!OK4<7EoMk>$8C!*H1?KVocdA8iYjp)>~v7qGGxcmoxeM-Kl<yWpu{qb3K^H?ALEu
zSzdl-A8w`G%*?gqLE8||N;b9ZNW#LfLh&uw2$riV7In9w9x~R#p!*mZlNwB#7%=*!
zI3z9&b#-s+jj%tr=VFr`*SY{!h2mEV7^C}3$B)k5T%KV|bB;o1SOO?)1{G_g6Eq~y
zg74V9k#mA_{jvcy1^ASzuXKvZzMR^R{4K#`lS%A}SAJ1@^J%OiiPD7TJ>NRTrD>W-
zf;4htXqp69Zf2$nEedln5VGoy3TbJNG5thZBeJlvjMhC)VQ2W&w6?xp&u7do-;F-z
z-iqTC8i8$YTDcd-)tO!7@q&4adZlIGp-1@)V{6UN@Mh=h=rGe9TSq4;;SB)%)7o2{
zihun=!WH`Pc4RW}L%hg~y1P{o6C5J<7uh}{E(BVrwZCu!_TVYLOw$lJ3BXs~t*q^S
z`_Q5RB#nzHe`5UW|7m4SQ9V)TdEO9Be{Q))pcbQ})5~I|>%+`uTt;2d#WZ8<I|XpW
zP2RVp5w?P!DaAgU3gpYchrTmX5%pmgWRc*s%OQLb!xL>V@6dCMTlJN{*iiR;miTW)
zi+J<0eZxBa`x%;3no#2ROE_|%V+>z@5KxLBt9!H?u;$Erv49iN{%f-L78g?+4wKgP
zxP-fu?I6P$<0Qh7;Y!bn@HqGKcC{+HlmggAE6`DcdJ4f6-Z7ST>>27B$LojrFg6R0
zUD_~Qk<wP%XbxP7={&{Rxk*%J?=*xQmQ1_J-ygkXu+kQFKJjs{R+~HlYm5#zEUP!^
z<mt8sH9A_g-~7E;Ho`F)PQbj1=UPt1Mz{317)nJ_;=*5yYW(A>-N|W^(t1RADUALU
z={<JHrB&Y2o5>c(&1h|UGRPZW%J-08eQn(rzBV5n+`M2(&xclyjPns4E3tFSG(OjW
zYjdYtQj~e@)E+YyA0gp&nm<o!q~8UhDf5wu(*uRx=8-|`pw3`bdY-=b!Ld_1-E-Up
zQp|TLL<|6ebZdD|HS8wUNT=4kW7$ZBII;RVtuxE$tbJfy_&^UJ0RzM3)A9j&#lOS@
z2J@P6Uxui&8SgZ`wuFsK3MCbkvl-K?k(sNEN`W!Xch@K4gRSc;WRH=JSKt2}YIo0u
zLfxRyp)Zgs0Ji|IAHOYt&do;x$v_aMQBi|ju@$iAoIq-Hei42&?-$DZ1osRpKZ-0O
zi5G=fXR>og(g|(Y{YsbSE$^RgX>6;m@W#smQB7~^t8zPyvvJ<)E-Yr2{*}QS#c%~h
zz-X2(&@ATA{)zQ$mZ~xvrN9J@Q<njb6f9O|8hO@R?i#JG5w7bJvRT+y9~G5`HIdGU
zQdH=Gt!)XWFY@rzz~51ym|^;&$%1lWt<;<*2IdScN)y{!#s(Mz4*Xevpz?vbJ03>%
z<Yt~4V}M4}&&PdydY-QaiD8u4G3Cw`k;;GTNQBZA;tDuIdf!I~>lnG6Z;#@7y?%CJ
zz=m;Hv<)U>x%YdE!SOXG^L1h%1LCcs8Pon19<_Ryx{+Mbh3qmnYmgzH-2g?ny0Xzg
z7B{cSGXjS&V@{aCy`9-*j5D9&L*TKhYI38R=wmC8kfSq}9uQY-gFyq&p1~xfLof0S
zlh<(|Li}!Kd&{P`{Ih)Q=29f(pX$zJBT-<>6(lujJLzrnu?-Pc@bBA_TB}A9ZoGMl
zPrG>Li2dbpN$+GlCK+?|geotXRH*py?pwSo6q7G*h{fl$mbtUC2QP;8YMVLw4p{?g
z_B40h`VCPI|HQ1Tm`31)TWT$QRUY!nmkRsF>q++9Z|K7<fk07_T)F&_?0fk45kdOz
z28jWTgVDNpTPu#s>l|Usg;k^2r(@2A)$b+6AKwmRrwd+83wAGgZ^8;tbkI*E6+JhF
zHTeqdiWbxTK$WIfoAes^0FBWrJvdnOSLvBFe!dgk8SKDDT%CE~SYHuG429}-Q98dn
zfNglQ8EZKD%3a4Svc--!;$Ih9)pwQ`*+nsG1GoG<^wM<L`}CsGjVm5lisu8|&rsem
zY5#XSLz_}ne#t0zJm)vG1QYg75N*J#S_13=4RvJn?21kt(lb@|+@L@bg`5_=9-(g)
z$flyv7x*m&q!%ZoGl~R4+Z7B2mn2Aomb`&F8&~wlSxiWbO~o<gae|p7WZUOeLD<7;
zFG3r`$`>-&hE-=f2*jwxMybP90i9JQ@fVX5hd?6NT_bb~(x@hEvN}((ASfyS>?Z_H
zz_kFN93iz>9l(qG1HsX;(;1*}ha4weFfu`q)+q2&N-1<ARQNBzpQH0+(o(ejE^Yt;
z>d_e^&CHg_{FX^$beFSc?!N(cCZ>ux<BS%VHDtq{aYjK1psVNgIt7{lLLaRE8AT}z
z_ch>=4Y&0i+W4B)B$85cItWnHe?>o9HIzd;3)t|19NH&fg`f)br7MRv_C){22d}?|
z$avPoohO^|s*=D`&FK9s1P;bv{&l2*DNYqA!c&s+?;Sf@f0pn7Rd~1!bk#^k%*sd@
zCC6UxMgM{FO?1`kQc2!owLZaN^!sefZ}zyyUpa4OWraeaClJp+28CI@TljT^i4afs
zf7c0FQ2m(AymM^hBpl<%Nh{~*EE2B1@*eRqIA0b}B-x`Eu~>-FdsaFrw~EL>)<mJP
zmWXZZn9X9qIgMHN(Dr-ZUcE3`jy=-?gzSjRs-|4lfh{PyPcmBZ-MTt-a{1nA6ahRX
z7|PaBDh`D{e#I?zc&@Q?%I6<%+5^c)_*WrCxyZ^ZNN`#P&}T6rnS{6PNy`Ye&^C(H
z9`SN<-84Sq-HBLWyoAh`fO-aOrtbmxiv{ulp#ueh{4Rh#cMuuhIDlA!b#*aa+4}5#
zqO(%QuJBu0q|Z4#|Ei1@K>wB`0P06i3aE(`WU?`_0*00da30Da?jWshkUc16av8@P
zV1_m(^8w=<f9}WEiPWT}s%pai&$=`VaJK>bfzWRR{@ui_n5Jgd(OLHvF~$yDSZ>eN
zR``GtEb%k_;s%N_41?mQ)V@L>K6o>f6<@$feBnlH6$R~XZ-Q5LMi~U~fhX3cLcx1p
zR1I&-&nX82gbWWiu+EQ|_+I@?v;z(LyCXq(5LC;na3RKp#mYPq4NB7Z9XxG5D<K0A
z-_M^>HZhxU?!!<Wpm0KHH9d7RI-!kXOtz}KkoH0C^^1uT-Ii}?00?kA!1R&sT%IV`
z;rS<T?|HTu)8D@1&!R6S6LA1d^Gh2h<A}Wo?gEyAq3GvN2dPL2RAN=#!+z$!ktqU^
zk&z|hQv4-UFmR{{B3H0y?BKJ6e&|mo!o~R^VWq5~W%&kYeUth{5=TPG$lFiBE1{JU
z_p|7Vlqal8F!eCi<7nad$xGLV+tJBsl6P@((g`{g87r*W(G>Z8N@ajp((ItZk99ui
zV_@JVb=C!^$=D*%^!CKa+%$ue5bvgz{rN%&i;!-cfF`E;#f4SDo6O!ImCKLAgQ({V
z_sY`U2DGn3sDG9r$DpwHi&f{4O<x{1IJ^Uc{?i8g2rx-$i)Y{sxn37F73u5}YIro&
z2Qpk;cS=hTcUtomES$xDE&fs96|VC3!&dKz2IH`W0l548Q4m3pn{~wDOo@e0yp-?@
zTKU2Q*<zeKER6equPzJpUK{MlgE=S<#(&Xe^->lAxNpX6YahiIv~2J>zTjbS^$&cF
z*NQLx)1Q%0;51oq&GzeTI<TXY-T)`q5ff8T0$hTn@Vh^3WX|%d-Sep#X@0;ng*wQ2
zBDTP|6B3sbeY_dJJYW5}jZ_Z&Qq?a?8BfsE+*5fGaDxAQbb^1o@cHMz&8WSs>16qz
z@2IGuV*fo10_~uPa6SKK<~+lTMCBi)w-tdo8s$DPaI=+MVtsKj^#cPA1oCBbJ05B<
zJ6x#Cbg=PLiCMir8-5zv-7DhvA;o#=`+F_5?r4m`w&heO7Ee=;)grm%e!g&Wo({M!
z7~U`qrAQmC&r$~vq_-Bwc3qHPT5`JkD60ydpZ(1DK<CkQxkP@K!~G!yzO;hxlpghb
z-dr4P56EV@^?@lZ4y_+mOIq4~s&YFB(qKoHrI1ZKIQgYmB5Y`Rx!3dECN+ht52W&6
zg#rIg&t}<>S0GAVMe+$ugBm8|BgAkpz;B~_J)M`v#>rr<;Vr05+JHGKTlTainS6Q2
zk3_<80Q{}|YfM}qU;d{nHM$Fys3n%OTVktg@hLmPCu<>~Cjve#CQ2u_rm=1m7xE)M
zQ%~!=$xedGiin%L!m=u5_eM_**mUHeFp;O$R<v<Mq#By6H=>p$^7Qskts)KHsCRd_
zTc64@blO_&&k#s!5T6x^n=dX_Y^HRwUmPs|iF|SyJS6&l+-GFhK7Wgvt?L4XS9ECR
z?fTk9iJp$7`P-AK(=oI#8C^CO;2FRiQOC;y=HuRbZo(3pk`M8UPZT>3Y%tRJlVu4{
zpQ_z-$X3vVu>7>!#o(Pb9brtt6Nh?i5;>REVCluh)m2qhP0Re$Qrd0v#MF`yHw!A4
z)l8{=`%hHOZ0obh4vLrms~-P++r_HQ{8!lqJKhil-v5Aia<34XtGs8Hxn@b@I=D)v
zimSkp8suha(%ouYAA4;0VC=*B+>63Ww~OKNswW=}uQPAeZ}u1I)CCsWdH-JDm(PE)
z2?TP2V-LAQy<RP3qkv^b?$9{8$nj?WVc-i24BhckO|MNWK;F<Iehb?j-(|D8fC-;R
za3QjRGQ79R3AU5DcxX?Ed~h_9ubXiQ*j2STe*cEc7~LT!P&1fYSy=*Cme~|wH^DNu
z=QFGFMBBpT^#<U;l?a$cpTW|Yqgmi{==V!XiqVtA%qjvNo3HSO;P^Y^iM6(oIZh32
zxiA^MfoiZL#E-sqWKKrjK~4R?Gxzxkj%U9I%KZc<C5cWTyRu*|N`NbZwz0vLsYH>^
z=%F`KW(Nx@_{z~=1u+GqzQmepXn&AxtpGcUi*r`L%fbyZ6W4+0Gs6EXi+?_^1?0iq
zn4<?^M`BD;d2qwWPqLkUU`KhggFWwDc*7}*EQUZcbc(1+;}%P&kDRaa3?O<ASQ>XA
z%!730x>ka+CxJi0KG8HFf)EjHxM|`lIQwFuUv}2M|4tSS?XG;`U0tNk1m(BS!+9a@
zqq}@iF4p-v)Q`WM`hDePR}RsSsU>0p_D|r?7^9{@?=muH#TI3n#lf%*2<7-^(*M(2
zkZ;@?X;I;iIXE-3EZongKc<>vRWS?y&l4rYcOn=HMZS-wFkb?$zWj@Y_LJT~IT4GQ
z;ln)&CX)m>B4DamEg0ht2{A66i)ZFDSQ9&P<Bkinx5^a`^MSt7881fB!+`G*3$v)`
z2d#RyAFC$T=##P{@Pz?WA>dJ<<&koRAU~}oCwcq!y~c6x*qO5*>mJr91^Ao<Q~NVG
z{v9>s<{TZx(+p4tG#w9Zboa|M&8WcU3}JZ&dzAlc3D4|nZOp&=abTjYvt%Iqv36sO
zX2xQu=ndQhrki4J;0_3UYlH?GwO+5ZCLE5;6=rN_{8d)#Zh*1|&}6N*CfH(*hM&d;
zc5oiX?S}2<k7y8t^kpczPu*QpQtT$>rv;fUe}6iB6AC2LA1q=sSu(I$e?9(0F=vql
zDJf0!x}3VZeH@uf4kqy(di_6gZgu4mqxRm9tK?9qFMI%Xu)ca~%4G<VGFC$svnSP&
z0&`nFrqg6OIXbnO`?1=7pGrfOod~TB@1HCgDX|^^J2moR9>5;(@Ew(cwmbMD)!$=H
zpGhwTM=KIss<_1#z!IOeP5WJecA%L8#oQMwQg9_1rZyQtg}&K=Qc-awTj0D3wbHPp
zr{>e0?C7_*CYWYKKR<UTB%+&a@h9A%ns_HBvY2jZXKePh4HU4Su?~2Wv$ymQgeP8=
zOx|e1nV==OeKrw_OtjVCQcb*B9`K}Qw_Vsh{ev^$31WXzDA|1a;LtO0_2H~0ht)h>
zV~Tv>9?6<};dpz7sE9d2s&ydmGn?!znc2OD3fYmyy?w<b(<|$<;`6x%{HpYc32U9d
z>=(QyY2y~dkb@RU;hUQ6cDOIdL#(I3SiTU&r1g(?{B>+bF_#(`N0U{iSn(=!Rvn4V
zc;qQ0vx+}p>hi}F8IcA9`$WV<j(u1f5kbX^7zq0sAM;+SgcIXK0x$NYomo+#iXNwi
zS8$-*+pPD2+do?qjCVDXE~Il{%uNy8^i??z2nYqPh~LXENHjbY%?`ep9e8C{8JM|J
zS6-5XE9o$MEutfEwW!{7Imj0_F{=*)ny5Am{8JsG2`gwfn9>dHzSzvIv>cfO74T5y
zm4`+Y?mq&VRYLYI0dvLvG{fscQmKvEEDc_^{N_HvJ|DQO#<V=ncW1~%aZ8Qq6_(hL
z;y~5%!PgG(d<EE}O<>3On1j?=Oy}?F`PqFsd6AZzaK9&Zs-AIw=&!cWk_Sgn8a>KH
zPm~yR6nCn8oNMfCx|B1pvZ8s&p|(9+8=gqMpt4%T^C-@GLKOLUdv}Xy<loq8grVU9
zzOVF?)Ub-_=aBs%iu42auJY2<kCp99BQ~a2LZG)6-L%=;<l1mxPU8XQnFo{fN>!>m
zvr1(FB3a*>kT=N`SLix}ZNU^)n}%R06|9|Yb>vl&kk(|lj20Rha->D^7ktNtTqHOm
z43!a-KWoOHJj*i{Q<G~Rg+6=LbOsFqz{m`C2!SbuxKy2n4t7%JCF95Xe0$0U8s7mz
z>aci3pwI{l9J2g&JMo@PjYe8Z_#j1dink_Uh%3rZ=){nZk^Q@-wgj#S*)F-=P|;D!
z-8YBBJ;48RPl-f$<lCx#d6%FbXCP0L#c%wRXhQ>64!BKv=>?a*c0aW<lhA0!By9jf
zfcWncoGO>0I;Kgzv3+Zu`^5TWheHOYYpvMFc{>YAF>*8zC@G;efpel0#~gIW5e}4}
z=Kp{gT}SjnzT06UA-W6f=vgmUILGvX+KiKM8UcU~)1wBBI%0yKb!df_T(K-X4Gx_2
z1?7oJw73gS{aD|awdnE`;e2dUdyxchqVrMZ1xlFA@+9w-fiev&<$B{jZb<Lx0+>4Y
zBi9{YhkErkf`;k>G_Ay)2fq$!(;=u2)uAI!EZ(IQ7U&p3w&4a@f_B$Z8b*1q>V2XP
zhv--^Mt?mn6!$^}%uyo3ds{ak;Pc~=p7Y^7HZzcg+~2AhwgRNw!M($jQ|J8_kL0F7
zY4_iXZ-i8<o~RiPhCluD#A*InujLL7WKqq?G(1o-7c+(9sj{@~%sEgZ!)?)vQd3p7
zgSK5QY)d+-+*Fqr`!tpisbx0C_9rZM5(BHJhLJLhC0E}bdC<8JIk;T0Q((n^kC#6<
zU^0&)2o=-&^x>?0Z7)4B!eC1};YRktN2nmhvlA7WU3jCDq&_+^kJVv`O~{Vs%*X{6
zW5@nI)d2=A8My#J<&+gTjHc&Cm>w42KVU0u(p-qb%gsq>A2&lE5JLx3x=8Xa`egt3
z%e#3ru)ZhCj`u)Zrfc-#KHlC~K!C40;i~jelh;j#i!;h@)iN;=|7d47A@i|BczTgH
zmrg5dGC3voAmQePiO`S42&1jkgu0DpozyAozcKbqQzD2JyQ_8m0}NDEZiE*lPEW;n
zltQYBZqX(UNPw-cL}}iGcH7enw%;{-iSvEY1=rJ$u}DtX0f4d_!LXT>fSvhLQg6Tn
z(2$amSDe|IRVCfyzDcD~X&X@c3r0Ijc*DzA1C5W22!?a;hTbck;5$rZKDz4z8#;F#
zIwrp^LVg(^va#i9)v*_n9gOhC?E~G)eR)jm!zImM8X9@<-e8Uz>{5uqpE~oyh9r7F
zso*sw*#*;dF5+Mo<=gNO{p5>w7*|B#jqa<p;$=V6;1dO#AUsu#Ix4&;?7wKl$2`zd
z?!sC~4Rnq%g?2G&ga?M(52fk`l}|+x^#K)h3|$?q3IGjrF)^ZIzlz5b{pm0?6;?{!
z3Mgy14ob``1?8caB``1S$7qquHT(TTip|3HddRm=(*|LO*|;2mR<h|0?b(KVmyS;N
zBJ=#_d({)YhUArQMH5F79>pqBh)%BPfF#XO<1Q@{xqF^st`>#mLSOaE5R;tKj1>J&
z7e);Y`c@{LaldG%J((h*Di+FP4q$N6mPjFcy#Z??n4><$EhPXn%$X^&(mC=≪Ug
z(d6;-rIXUA<aTU3lYMt@)V1x(!|T(qMgL}IR>fD=m(sk_Rn6UYiwxzsgBBe4{G3|S
zw5ZZ!gvWkoPPnVdXIP`$L1P%uT&SDX7C`khY>7b4-Db{4JkIOl{Ar1e;rt-fVd!#<
zlStCuTQW>{voDbY`IcNN#(`aliMOBA3-BDGHzB~`L}R_3E56n%OW1-oYJP}CKaU0X
z<@>B$sb42s?kMJ;&FL{y)x;Bb{WN+3X=mpZtRDl?)|Kw_O>D_H+2~4fQ00!HtdQIl
zhCqhw1<@nIA1WFgymeS#{Wlk&&+C=jzyz1L$&+9)i#M@F#=wQq4lld3K+7~ysSR#s
zhh-Vw)#w}%P5#vI;%4Jz_(YLJ<6wdKL;t}tBrFAHp@j%(LDHx^0+tUMeAoVg`a^_E
z=gr;S(sMy(72geo{t%llV^5?>bgQrHx$(t)gn2uORzsWd`#0iXPMtRZ1c!;#MEAkF
zpHie5kDYtw$tBGz8#;Sy+~vw!FeXQi%*`i(b;1Idb>g#Nmw>?cV77gcNPtbCXv|@0
zstK-A6!yt)`}U@NfZu~<ti@sI#-AX7?F^I6L5y5u?e980Y{AFihTGw77bR~Eu9}<>
zIxFumAyHzpG+-(ME?g9Nx!EC+?3jQ@xeItq(+E!q_}oo-khmrl8EEuj&jS@2pwXkY
z;bbST+8rxqyq~fc_f>3w(cguQLC(a*M+`AO)gf$)t%#Z}yo#`a^$g37kp0Y{K1Z7`
zQ@B{-U^k$K`s2{T_~c0yFf?8#o+%DAosUs+Tf!ZSw3ymOEZ*snPp@$big&vPS|;(6
zvjjF%=HJDA{xfYLdzhUy>@U^xHmkv*@5f@QlG=Pj?77%lpjxtO*lqlLOSEV4m*qX|
z;+A;WJ@4>jJV*XnT~$5On9&!>2LS7v5#GOQJlaxr55MwC3u<}r@~Qw_dXigVZe);S
zY@V*vxE-(-*aMCu6fyVk(TuF&vc9I?<*mFdQ|vbXD0?!$sCb@YG5)G3RSg*rM#5!<
zJMU-$^-P3t$fu$*=`H|1qDXN}izN45mLc|Kx(JxtB*pMI+1(G5k(&DPXcm_E_ogS>
z>ftHaV^7-Shs&W?ku;`tT<D2l-lI(O&e|H8ioRO21_e~Y?(l~+G8Z(X6z27Wl<2w5
z@$otBiwcITNut5G>qLDw^=y4r$z+=Os0B6kV$)Ku=4($kwL(r$&o_IcK1NnUvDkX(
zOl^{3lP$j=_~k$27PQna-dWQPPZBL6m|L4e&X@PSc%Ik0mcNUNnjSi;FP*2SmUQyN
zY%Oimi*~eieV#r$jCpfQ0V`y0q2kmzTd#+Q5H~(rNVn+Au}|GjrNHvJMD1r9%<OjI
z%4EBa^<uwfnrio^&0;chBcmInjAf!Ds=fK^R-=iE>MpzK$fG}IX68|S*O?}4v&oU^
z^S6|M{uuV23?lm}FQtpQ#b!9wdgT}*?f%i$^1_5yuXAkarGAyMz$giQ)J?nYS=fh<
zxl;$_0bJN(yw%=i`i8eVaHCO<@Zc2TR-K)@Ft#w+!A28C5?@-B@HM>9Qo$2Up4-CU
z%%oCY1pS3lp6{Bh+V+?_;gc&hse2|Yrj4a?P{$|0cR9*-2&aE^Jw&uk4g!gsvCbDL
zELC{DJMxIU=FSibbZivy8x!dcldc>yT7~EIO(Vu`X<`-|cCa6d1vn(V^4UTDCn`W0
zvEmnaVa4yV@GtUeDQ0QoHqcN9xX?^=Bp(Jsbh`sIl-JLCW_h{!=@j62{pb4KVggl0
z;fiS9KJDuE57ntwDNU)Ega(V}{R$Q2ZApA*=;k&{t$$Z~xD0-vuA#{7yAeypQLt!e
z{O*mXKBhxw0jW`xXu#_vo!7W+ikdcP7jP+<*n&WPRLwUMZ%Lgb)dpd^yGvJDSH7kD
z$J=TKK6rp*p!I__=D1=1^i(YwHqJq~5X&i`!2X-D-bEFxuax71EXYmWTr4%}5P2<b
z3}=X$V8wMj<RdDlZgbvji!o2r&D=VmNl9{r`p`JStJ>dI?w+M+k2tHWJT%~0-o;Pg
zJQRz=kQwynwrXG?eM---#1+BymmQxxB?%2eMo8|U=QIx=2M3#%ENtQN<<Mwo3_Wby
zow-W;5$Y`j5ha|0+WhFL!!vg3J-luBpJ$rXerPh?Ww)S(ZaFxa1BmEK<X#PUmJE`^
zv+&+z<nr>XhvrbnjOIYcDpx<nkKJMYZ!}J0-%NnqZDc&PW3yoR{~fCne;dIlB|M)e
zfwuY-hojy8YZj^z_j6(P{)~40Nb+mEDQB?_6bl+t6QQ4hXw+FS0Nth|kNSv+84U~&
zv!B%@x~07>1{^S(^g74G4Y9SRMwp_J3=y(-O6W)C+&*3@?3Orgqb;w5*^WnB!!2?%
zeu@QU(OBgi8&o-ML6>YUTUW<IrpqT2h{f4%(IdHlQji%EY@vOUC!#xQ=}PesE|~NU
z@N6;dGDiuU4&hfc3qu1ENV_2ts&2PG!ikejtIjp?@?V}WBwW|iiwzHN<uDKvgtQv1
z52tzZ<uN`MZs(}SV->mzzET+&Fl9dz13OkQ_X{rYx>&xkY%s8&^6O1DdY@Hvag;9p
z#Co<&ue9+qhr-^(#BSLk|IxD&aG9<qLW0#*Mk(>Hb^7~zTk{!jjJFCBZqNrjZ2*&h
z0IsCRl)O8MejV2kNh{X1aU<Ggbv;H483NvN8h=J}esjghuG=<X%2x4Wz$YcpyVxAf
zX$Y6Fw7908lPz|CS^3-5Tw_{sxJTAyc;b=Ddjrn7#`-@#_lfk&*lB-ay^ERGyjV8L
z0?9R5?;ka#$PddcQUETuHRr3h<CwX>{~v5tc*E*bYPy~?Xl%C(QJ6$`!Hce(E?&QI
z;xwcxc>a&Wvi&zYol>J`{^mQ>Qkqq4h#7Am-`J(KbaFQxt>2de6ylSYylOxpJr*}x
zL;lmlUf}i%*uqPu;$OkN>g!Tgcy~w9^^s^_s_A2+Dvs#&LS|4Rk?q!<WZv}EDYC`s
z?LF|RQ6uCg_p2oh?QS~`r7E?taTDeCs3Bie1h!Zm>q^BsI0|-edeUG9c@)WVZVlx*
zaUMk18yI!)x@d-N>il^Yh3BElxBS7o^4i7mOF5O<o);191tVLNN)I6dN`L+iLxg7>
zpH+QEDr&;#eF^%NRO}F-8W~;Rr<U*u90KZH7U|dbo0hJGzfbXPLs<iNYC{4aS#noK
z>=4+iYOW3eS>-(NY}dUCTI~OEApjgIwva^VqWCa-&75(%_G<8?nT23^c9FN!2_PGT
z0_DD`bStp(^21EF-(u%SrKCFKVUO<ovZaWPDjTRK%(qb>j#D&ZY|Ktyp9|qgahR_z
zbbONjV;^1&yRKriQ$_pp>4y4~*sjW(6bFV$KNPm~hS*liGmFW?N!bUQoy6T1!&rNC
z<d3+0{*S}}+9xYFxCUqeZM}h^%qsjlEacFgY02upmQ07$Eeb)70vH4i_>K!PT04-}
z?lwAE9*I&TXBU%Eh|fmXEhI3D%Gr&bZocWr_M-z89Pcz@Pu;CZ*JK_JVf1voZ>V&(
zGVkYUhxa{oiF+zBDz*h0O^-*e%hGj7J(@)T+{<|RDijl6v?X-XQ~N=S?Zei7t@0<!
zE2-<FzpV*MfxzpAqmJV9lf8o+>MH7wX|yA~0Xie3cF98YTlhl|MeeSe0J7B!(rNNn
z?m9hpIugSGY(NK=sG*032!hk(eynSYyKi*bq^2AIsv`&X*xby$exvgGd9Je06U#_k
z^ygo5FYG1KeI)6r;}RK~L}Duz%I!&Lj3@4?<@iT|xWA<&*V%}fUsOXU_l`=a>7QrC
zvkF}Nf_0@44u@2FNjgF;vt=Di?O<v|cCmBV8h12WrkuG^e23EsG$^Fgq_JZa3*ARp
zqb`}SdZQ+jarS5F)dd8`jj<Jqfz|DBr!Je95e8#l{L3yQ5UQpFW9%s-{BHf<IlX{v
zZ?jd8n>#rVRVjHod$v)2*JXf-(g^+2)KroWr}y{!RI|rLC?UU&)5YXG=ICo8;xCBr
zb*ZufHf|=Crvnn4IDh};YtHb|TLF0CcaB=R%2LDj<RR}eef0D&+>3>OUta2H1<j;i
z<d{B-$AJ*@Aj9~Vu}8}w?QrqZRewfZ)PKm^x9ht~!J$7O7uamfVTzk1+Qa{^xbOaF
z>;2!STCG||5nHLEw$zASLXFzBXYEm1V$Z5p?b@VPsZFR|#Ar(>v1d_UwZ$qCvBLLw
zzdzsa`~CR~zUK$>I5{~d=f3Ye@9TM8*YiFJoa;{27W(Gx2shlRBIe(b<Eo8%Mk2am
zdqDF}a~?Kxz@OnB-)D-?%zJ@u(!DcEr_K4X<XVs?B>(`st>$^sD_+QN1a|zCgAlb^
zO4^nF8LjR9)PY8(9|uk3(-xf=->SL4mJ9~-wn8IvcqstU<?5@bqYiB#A{0!Ss^#9+
z239^2|J7k<{?vhsaSJ5h#{q1RR~@zwG}+Z=D<{ejq0s~J=KweXN7J-uNo{`J8a#n|
z2y%0fp7$|8P~21*q9ghjX^rE5i~>Y@ap7d~)e!&x9?RiCzP<VmfG&Tijf9#1eQKhj
zYxXo{H^XSt3xF{GIv_C1EF@&NP|=3F(7vFt?l{JO<SNm881^6u)|2&Q2rZ|pB-7Wz
z<<K7h<fRsU0pA_^v)0-()|F6>hOs&C1Tbwf*b%g^fM%s4wRwBE2}b_)I{m+|Woz&p
zSgC#WJ9<fcIkWZ$@XU7lm$aj*`&GHHV4mDWcrWk^(kSd;fBAq*1$toDZVf1cp+6w_
zQey-2*c?W#!eK3nM_e)@VNZ!6<kVXf#(fZ*q3^lS!|2ZXPdT66PbCv9M(VtUIhVb5
z)MCUFXA=m~F0OsYB;t;8fEG0jH0ni8l1WB9L2_ASsHMgjVc^>b+?+B15cfg%;oc=_
zTP@ciZGe`t1!yTP%^zu9uZFRj$K))up%fqfyTJ&8dl_?K^@gG%jJm;Y@oL%yHa(jz
z)9d|FaFhFmy_%2(NUz2v7<(@6qyb&jh8$ADvoJZbh{;E!{|Pv(Ow<Wa_x5&k#D)Jl
z5i^m$ykes&fD{FNnyQ8av9noJ0o}1`pqu0)&zKHFRR7nAHN4?CQU)wa_pj}fL<EbB
z9;5X7g71)T|A07+ye1T&(<^CzypB%1SLpkfR^Cevq9wX2aBm%RCl%`&q&*y!Bk&|P
z{`1xeYhS?9O5bs$d&?w+YI;do4$FSf)eo0Ph4j+1_O#jIlIs6mCm9f9|L2<j^Y1kS
zsX*NS-bG{%0{ws9MX73wuId8u=8v!v5k1Y3?7Q=yulO3pE$XDo6;!2s!-l`1$>E2V
zE!qz16Md1xWPavvsY&#@3*^dJ6dJ(Oa8#7Rs7^v;PxGHo6MeFD5dEtA0Wf|8GzC!-
zbFv9%VQ<VWR3Rx5>c88DIW;qS19)Pq6x62(FDqo#t@V`@h%UZ>6#bvGy(Ycia3sMc
zqfovAu-S3iEUf2Urhpum1yG+vch)K+#s72czh1r8xc0xUVh-@5^w;bY$>Lk&Mb{J~
z0PiU3oLhA5bypUrhB)8_AVo=Y1p(rV;)zJpPX4Ryk_a(EgjnVclpGVtp2U9Pc6<v&
zRPCTFM(Vy$WZM2CLc6ssTHO{+5+PQz*`Z?BE64_d>Di*qBE;}FP)H^a7SNy)BE+8D
zK!wOHhhWQJ@GQPJ9P{a2X;#*3z4!aFjjaD4!w50^2(e28c{xA~Q#oV+fjdj!#p@r_
zQ>T^9FXH6eJ6Hi6Vrc{1Ge^I)UwF;X*;e=fDT;y=HD7s}6cWY&s#hn&SqO`+;UmO`
zZ=#TFAgrY=+NjWk|E25LL!au7t}J&F`@B(i^Zu|PTXdc)KE?k$BSNh7I!X=%0?Uf7
z>5sky@WG8CdF~V&omLNd2)-AOr>5BE80HmqI_t{n1de0}fuA@AKDR|@8Vu2}aagqp
z+;1SS%Vn4K?CHw-Ok(djc6NV4w}qo-;>?%4<-@@WaX-nMf6Kd^H=V4L;kllrQgO(v
zz>Fz=Y!Aolc=wG*?NKO7$&Cx9(RPx1*NButX+hXfuAtY^R6K>0(Hq>*25FidWoRc?
zVQ}{+z4GrjQI<fqPp8j8v!f0jpw(|fna-v~7u86BMb}ar@0!}8FKI^+&<TwzSwi}U
znl0KSW<!IEnY*xtX7nSpfs-Z_m<f0b%c1v4&d(m?ayR%1(^yq1nrMjSy~xJO+Nlzd
zLDzg0qN8^?xlSxKa9{4yc&I?_l-eGNI@P`j+<`*vh*-D^bLHCm_qPUc5`d^C#gn!J
zgZePiu&F`w4MNowNDQ3B(0BV(mVD%)&|tdJlmkYZ(5H?eTr@HIPM-@M`=zPlXyT&T
zX@bdVJXE1tTM|N!iyxwg9u(Re0GxF^1aL6&7LvrcNN|~FMX#_QGtgKSq2^S1&8KL$
zcu<3>F&nqJX5*q29z)5?1Qm6m;eY__$TKvlDEft+=Jyk5JG24b^w{H+AHVE3>oot#
zo7jj*ERM)U+)iXk0J83VGkk@|TK3xa=3JOI_!j9$YNw1ur$_bfC9vH+J?xoiz0Zu~
zgF<cOaPeot5F~Gf>y1ol{I~g@Tv*Mc3Tybu!0C6$4~W!fs@Q)i)Ov>3?5PqcZ5IVd
z#$~>p%R&#Ivwh7-r?q%HrLa;AG<aB1WXt=laJv^LhE-IT@q~-AFa;H;)TZr4rOd7N
zIOg~%1C<_9fMuRmdf-l%T;)V#KjaAYA*X3apgu{p``D?cPD;X}H`K(6Xh!SZY-vck
zW!yNWRg7KUO<i#Rv9z22rQI3#ZGCs<0%70jh+K<*8}h(G@OJ#kZ5r)N&)N%|2$xKL
zblx=5pSN%)W@F-7;Eo2=InaL4E(B@mi~Phq=65Gj#TKnl&=^=*(EUnO8A@JGylURA
zAamGh6usf^rlc+~P6j;;C7ejp5a7|)4+?prH((V4V$wfjbF`J+Irs#T`#B|~1-DwH
zgeOmHs%5OKYjOi1su$sg%W<504km8xQX`jhhYLxQ68Th!TyR8P#+z<A*{o=%twyiT
z4AqqjromDb<LgmVr}i<5W5YihL+`Y3<i1)6E))TXcv+nJJqc#3@&kJbotTcMi8JzN
zWdE8s9x=ylzMvhW0ubjOb&ytR3l}(u$mL!~QRKC2|5lt`_<dA+^sRHlpF(@zPjyH|
z@QJROvUM*QB-~%3a>u%ayX?@VSX({`L=^_BxPKT<UT%RNl87}y162Vrn^hhd<0dkH
zJL;%pE~`j`ewilRx1m6^^Kj>Woj>E?I@NO7**+fX0-1>d*CnQ^mKao6wycqjCe|
zRU^dQH@l?Ezdc`lo32K)w6wK3G7I(hGXu`6exfF?GWuoXeuLy;KX6y)JLRPDfPcq9
zb?l{2jv0mfA{jzn;IX7d1+n8J(Y3LW*>&;QfN17B{#1ze?g9H&=EMc?&<)fJdh=kb
z?=44%6sjMd?(F<-?X1~@n|_QuWDBB^kn=F8V3o)h9@e{0it-fuax&e-{I+SI7Z!9F
z5XB`HzYB;)@BFLgOBpe-WgB?}ePxBoe_pj)cQTA;ew|JAzElp^ourT0qiG~wO3_PD
zt)X{el$;*^q2hki&_O8bdyQ6@L`C=jv6K%m-zdS79<{CUY%rw#96i03(%gQ7_+Lhr
zM}qAO&;x)`UOp$@9WfWn3wfNtS0oVTZC|=_kK|iHqf|Ggs&crEO&Ka{4p&3s#y4NH
z28u1(&IAPZ3<th3f^(&%2IXVf2lv}lDDSQG_{2mbJCIw5V{?8{r$9x0TeKS{17XSN
z!xA8XbQqiX)={9obr8!On|Bbesd1hb&4qV4yU!H+sFP*lm=6W7zUDYTEp^wGow_&v
zy-gxV(f$jrUorcRy<ZE|u`=(XFLc8_1-$}qr}2kX-<aI7?7n<Eb@4FFV_<J&mdWdv
z3&rOjcD*(<vPVr>->L;f*SL!_8lO6-PI+oXwU5l+^Cjc0q6O6-T_0|2SYYfe;kktf
z1veZ?%}Pc*{5Lw#b3{t>#_Y2`?0&Jc!chbEF!mY}BL4kG{3IQ=dwk6TG0*W-W@*aE
z5`vKwRtI0r<x1Q%AnZLWN1?a1#Cz1h8Go_pG1A)aKclsZ!7>hd!<l=8gX9{Twr3Q@
zLZm`0di>o83l@}Mk5{jHW_DfZN1p;H_|N6>p7+2u^)^Sk!`Jb;{@^t-SNL^Q^q=+Y
zQ*6`7Y^PJMkK4<=!BZSLWVKDc;Ru-)x*o7Ntvk~h*;-DgLxu>ITbAPuUc~mNb?4=c
z?zE1~0-A+&e9XWG9`5#n^?h|-tR-`71V+Fyt&ug1-8{Q7ws+b;MOZUpa`GQ(lh99o
zU6fb8mXYhy>pDV9TK_kt@JZKuBpl!Yq`pkbkSDyw$*o0|pa`*n4Nh#YZ}HP-C6q{5
zRmr|jg~re7c^{*nL2dyjulh0BlLl0b+{4vm0&#@_s&%(lPr9<sBs|9?F%N~f%p;y<
z2&@A1!~kYm>h{XLTDP#gx|if!Zqs|24AQ<0%8tLgvVgnxHyE=@Ip_rGT+o5ADXnVf
zrpmO7Uoy-}MrNH!?5*{t!&w;k!lzmemPC2}Eb79eB-^REBQsc~j8P|=JVw;996tm#
z7>obrV4e$#t|hT@)0fd+nV0z9Hj8uB)7=x`MK;8Q`4U31eu=`U!HZpJ5mD?GS}qk%
z?YQ>NzE@-Ro>N#Ne9><041UrV(3%FT?nd*9Vt=6Jvf+zUAx@$&5wJ*}?>UfnWVQz(
z$^t#R(WPMQXY>pBi6T0L9=tb<rgygi^pgz7wN&_tj#F4NT*WCY2L5tXCRC<$*T#7b
zmUZHlJ|}<<p$6x7qup0oE!6j(DdG14I1-kg-nVU*{-+au(hP8pT0E?s2-3yyD;OxU
z=yGQ9y*3++<u~7(es<|S)RuK3=cJGYAD7uDdcB0dY&s2=cVd6OcJnYVHS~RVqS;3{
zwcT2M)`^D4+zR9sH<+Ub%^~v8Kaffk#tY7`+3CxkU66m^aSa?k{c^o=$IBX7*zvZ>
zA+PD_Y}nrAI34-HVuOWSHujUk^uma9&t4W}+*ixM;^9uJ!X^9IIaF=Ye(gqVf7#uv
zKx(IsBzV5v8U%e#0ls{x(}KtM=e8Wg!kt}P%>n}{KO`b*aF_4;v6LR2B93drJgfpi
zrx5Vs0J=sTn~p{XTjN$+5W{f$tP?kRzb0!H?}XhUwArpnVMx*Z2mK*fB7Ayqpz>}O
z#CFZqeBoC`Gw-txiVkazb#?Y@l!w+ar?+b=C{%YVP3o%1))_j!qn+JP7N41EoV*-A
zTeCpRjS^p>cWhk+-rXtVg|ho9oDh3)4z)$UJ>!fJv!RG@(XiWfkz;@#`*G2*mCl*8
zuf$C%`CsyX6ctr-;M|AxboqU*>Nl~U^;%#^Ha^;5JBv)enpE5b*$<2<RP*=!Ux+ab
zQXxXWxU|?qJ<+?=h)^ls9d4+D;SpjvbdNIL%WSUU+c!n`*%OHjZRX1%e!~^gSIIfz
zRw*VAMARly$O2buqna=C9a(LT6`tYW(NsF{B0#D*CH_{vcYVdbn`1WM0Q|j)3K0=?
z0*eVwm30VF1M}+ZpUR|;9X0ilr#_@ZxiSLJDRCkKQMoUv04D=c-*b!r4#s@GZ@oXK
z1yxLz%L>eI2ol8(AIdF`4b2ZS-_d6e+H!l5SnBu}V5$ZAXKZsQ9g{R=-(}N`(w*v^
zeej#n;VwMO3Va{a=Lb;PR-`6?r=Z&e`)w8;tS1|XFVp=;#|(o;xACxwM3&aQIQu|b
zWJp4!3Evx|+QP5Mf^B}F+a~_X-1!}DBAdhO{$3fw9Tvc2`bPms+%o||w#TZhy{P_8
zHSQwV{1V$tEn_NU8(cqS5n?4$g=1t|sQJdrqqL#{^>N&R5-E+>NrG((@mL)h*=^(Y
zUd6u=L+G>eJuHpK6OQq-1AUJ<9;FHj^GjcV;FuJA&i3k`yZinZRckz9D4Qw23sMiP
z)bu~?9eaDF4TAyZ;ffzeGAzyA=0%(qOdiPH8O6=6HI#P}EE+DB3YWQ<&e_@71eUCy
zJ2ELvujgHF4^_(#&vt9$AuX#h!r0lZ)x_#sIji<#{}QJcBcU9<=uBX!_gED_O9%LF
zLgVj5OnCj!YeE(B9E_z0e$n}?>};tT_rj;b-pbM0jI5*m#I;i4HEaN70Qp>%a`cc(
zMm@i|d9zXA@kjg!JkPFggNdDO0L<*Q0MoHPf}iM;4Y#ITF%0kEy@%Ga&jy0C6$J9f
zzNIW5wlehsVd=o`j)95vT*bWS$d+GN0`IH*{=lu6G>O%7ad%5n&daKxV6B_x)stu9
zBWBZ`1Ktq;F!&@G!mUQ564QB5_F`H4vg>uXWQ#HI<fN*d3-~EoE~rvSk|S$&PC$5n
z2uGzAJ8bSc7Jo17*Bp?)j^htPjUI%Xa8D;2cnX~<S^p|MI$Ee0Uf0D@Q$Xo#@#Dj#
zz#cULyFv|30U#li(tjkXW|=8eY0tMxn=a7YSt*RM*7&u<V!ErbX;|uE7Fr<}3iZUM
zpqdEcRu;;Eo&ykP+Ay2y9BG;2`S%y{S&%<125+YrCP}m>hF+cYFzQZ<bzpsvyY;6-
z+-j@8BE<Y}qRv|Zrm^`VB4S>{^3&5C*p=>CsLg_Lv8Yx5?W3mUgTCN*EhK=w4=@Lx
z#(*YZ*42PMeuf|O%udZ5*O@r4B~)(Mz8n9P7@l}#KwdX(W%ON~Z#c#LGoJ?-rq*?5
zc&W2I<v*NzX&bNuJ33hfntk3F>-6nuSzFU)y)5vq5y6eJuGSlAe)N-g<MHv!UX8Vx
ziRO$5+o(TMX2-mGA#OwAfnq5&;mDMk$y8nopa0Q2kG@?_XPAGS?xUy9EUs+R0p;u5
zID4%INtl#*5&>&7&u6h%_9#|RGys<1N|`M(J<at&&2aIb;m02Dxk&s8;op?}J6roF
z{@h{pA8NV_MZ`useEO(NUvX9hb;3qF_M+$@aQN6jDWguFzqe=XMKx*a2IM4$tbwN9
zZk>+fEDkmHg}G~`ca{HSJAXP;U3$IhTA+q}v${G~USu%g<2w0%1CV=&{5E28WKJWy
z&wtiUO;EK~xJ*sf<sWXDaJG8<Td0&oE-+M@>FKH2_sh7=xUYsB$IH9m+%D{wlH_2I
zc7r|c&V2ZnJ%``CQ3b2wn1`vs9R}4^S59n%QFF(|MbGT)oK?e#w+o<5FYTNK=EVa$
zY>vi6?#wk8Hj7}{t_19BB0HW;n^gN_JPOOe+s>O8-3qG7z&q_WSZTy-ilc*mA3uv=
z+0r8HYTl{vZ$Q;9D><pS2$Rv^@Yva*-ncB?*jwD(`$IMntd2c)j%b*lTDUMyshXRe
z3*Vk>^yk8V+<@q=tNJVgA1%ln{u<`M)7*4)w1tFu=&9viZm?bU<TPvry4L8<gd2^u
z=OuYEisX9{uI_*kxAHmABRhQ>x~t#5JLTs7Qt>PBok=@|+}&SZYDnN_`&o<`h)3pG
z`8@3pB)QDPQi}dQS+~rGdFJs954Ajz3-80a{?nnTJ_}eTgn8Ju;wQYM*H>m>t+J)@
z7Unw(Pe?)G1>vD3uiNOEuAd`eyQ6Q<hj4Gqge&UjwW8osqu&h1B8s*QFp|9xeDM0b
z(8VQMP$AWS8oq6JJxK7+4A{rWq<grZ-yMEF2=P~!lkR}&kQT}`%-iJkk!78HmPvYG
zxs;>jTfCkVjDkp+w_~o2ypqdU${1W6b4m-7Sa{ImHS|-SUu31;@Mys8qcx$WkTgtt
ziO-6_x#Qrk7Df$OA3Lwo1$hyxLEi_SM2r224d@tZnR|4rT5g~<>x3J!p0PgV8I+Sh
ztZgm2a59z`?#Kxa#`_HoE_0l>H+Ns2>MbV184($Rbke;LfBWy3og?&xMW}lEhVA_Z
z-p$+m&$Ao0rXakLEdIsY!TLA`1XT8uQVTW@_7Y=57dV9rs@D_&d@f;XnkV_*Lu|=k
zw*D!jmIlaVzV-R}xAq)=AAIj?Kl(wICKcnNijRiv{%n92vaHYB7{Gwy4?U_60lf~S
zCDid|-z<ptfsCV>mvdhmhqVD5JB~#bZB&h|<#>XFl~*csv@ZM~J7G~L-`wtvOGV+_
zunxu=7XZI6J5DZzT2Bp2kk}s>BWqG>TL=6jO-2KZ={{vRwl1D^XfX9Mc2L`&acIn3
zsvQ=G+j0&VtvDnQXS|&UE`yJny5MRsK0GGbJ5D4mQ&s`!exkv_KVtbUu5-M+8n>S~
zZnu@$&XTw9ZE1fJ3_s{_MiGrY{>dwTl)oH096VDh#&a6(;&6Jyk(HxmwGAi_o{W%N
zu)rW)8<%MV4vTkQ4Yd~vkY(7et&WY+cZH*bOivyX5z!`1!uK=hy!Yo;5lewCXJ6u4
zQ_=CIh9rK6*1OIXwE{E}7m8IoV<z6Qe$}wuqqNZUfT?bXK?2h<2x#y83#t6-gEBH8
zF4kzvd|>_}e00&;1CbJbU0V`&dY2XZ+y=BhpUpgPHj@RPgC!0I52;q`4g^iPA}jbL
z(<W+r^YsjqA|L&NVHQLX$ilM$+`6!;om9F3J)s*0Br0yhLalDkj)5|TCt%GlrFcTr
zTS5U|e)#VlZ69^ta0bx$?u~;Q-jHb<v)#F){+V@H<&k4^422R#V)dhGtIN`I1z@QU
zj3mVDbHs@@HcNdhs(qy_S0rtNp7C$gc}Szc0y2~aa<zS4j~zykp?0R@MbX5>-&;(U
zdprGc^>BX<h4s*3Z(pP#)2d)*nFrs}8`%%3@{u!j5n?U-1CKX3rSB5i)&-c~<QO~7
z#CI08cPh1_yLxF!%1SHk!4ip~>Q-M>kL%k^29kgBK+ZTGnwL<V51B7LU1u%)1$Olg
zriSmM(To8d$<6VD-^)qsV%K+bLbyq$dM9oeLXe-lo;pk&P%$emr5$TFX>!&k-ySlZ
z6&FwkBEfM(DA7dL=k{MwuE_Mi;qxQ;<Y199FGq#)9q+LLOM^Dn##CyVb?e*}^6t*d
z%a6&zOA3foR;i3+;jKW!80UoHwpzw5=?-d~W=;*?vhjY<ap>y-@{$1|S@3{clhEPG
zpUxSzRQGEM2%M58Lk(lVP!G-oQ}1b*5PpAtB`5f6S|mKs$XA!w5|El5=PL?9{+nIj
zX<K6}x4)HxMn9pb?^mb6gcYQ(cy&!cOJoPZs;QK2%8VB9hBX27jsBtiUQJPY)kVb$
zGJRZRvtHM@K@en~JyDuAS+6x&up3G;8l@pdOkzKE_`tydC<W26jeiOyq*V<~H~2?Q
zp_WNbQftr2`032dd_Ld_yb)q14-jd$<IV0j722~GRRLP9o&U?70qUN3=9p+=UVFzW
zXp8CDLjer5fPFaoY4O0J3FQY))D+^HeFvZ}yr)g<R9^P(@2;l=45t(bdEPueQ77`@
z*QXg#Leb#yv-XlTmSN`~w}W{2lqV#gaiPXS+@A6fa#B;-(})3#{z!_p^a0*65AGaL
zZ|bHx)z5m62{<`$<HbV{ZPyxoGEU8JU@{r2u#1qhP$~7eyV8qtWwM|7ZsFYl-7EGw
zO9I`Ex8Y4nv$AsbqYoW(DxV#!JdtUO?N7FOi+<p|w(N_$>;C#o!Kp(p(sJRSoXWK4
z+c2~GK6!;o3uZum4EkfDn#VzQx1%)cgsv_{)!AA&Emf}E%34J+UuVX|-556{Ro~U}
zxd~D;IOrzAdswCMkE)e?YSf-*3GKQ-zdO_B!zq)IN%6Dt>_)C?`?}NHs_~h0lZ(s?
z`0xxbWaoDb3;H~PHqoli096P`a8idpzCK%BOOxo3PO86sf)@KdHbop)STiKi*#EP*
z!VDXT8}&Z_3CUW7m=?xut?v~~Ys`T`LQa)rA^V7$q5+4;@l+d6`Pa(h;Bw3$aF4%x
zsep4>iGcH#P&A=2@6QWS(KUeeptkI|iFzSe?E<!J;t|tYy!0OlU7jy}KcxU@4_}TT
zp?EZ1Ap2rLYDC6}hH2-m_>&IM7vyP(X)mxy^yriPY6V)8p5U_mGTw%xcLCo4%;;D`
zNyT13=bV_WbYiH_J^w-CqVW}=o>&z$E<3P*#!L1#7N@gjrx|4j&Z29<q9r!VrMF9O
zp!(VdYeIjdR(sMyL;xx<!#@NBg2k`fnJ0z=jzZ178OjMh`k&#_4U(>ob1fgQyaV%}
z0gg~Jywo~Byk_KKexY1h)a1L>^Mc0a5;jH<*lM?CWY&ui?wcM^q;t;Tc~oPh2f323
zo-#;oA!SB52v4<Hv@$o$Q{n#OW{hSE7NG|+esITWoo7yP@4wD=30s!*EA5Y~fF9AP
zZ&*TZ5QegE7?S~mCy#;`BaFMg^M$?S)J8KBGw@&<H6G-VlXG{Pj7U%Xo$;|K)=6gM
z@1Hlk>Ta3b)3;yV->#JC`12_`LQ%YbrLl!B=ptP`xv3>(T|je6>(eE2@L~8G3=UuI
zT%;M5(&95Qwkha~+k&3m?r=}`KD}*Lt#Qx?-p%-K<2$GF!OT#K=?l&8zT1dJy3xY`
zK_MVJrvrN6s4FW!_Kyl!R_C{$RTatj;{GW_T$+H0PiJJ@3Z8mWN=bF$3Dh)EE?{>L
zMrK`9cZVl`PlaF%e6xYTk$aI=jrnbnzcv1O{V|(Ya&@1X1tlLI$Op8oC|*{Ra+FU$
z+CMb@GP6JV*i|^UCQ)1OR2Oi7Vpn5JfQ~=_43+rv<qxNsA(tVx=s=+6^$o^VBh>h1
z(PhFoKu&*-5G!o1_w^l}cED-cq5)-~StrHpnThLO(e#AwfKV`vvil`12-Xngd`IQS
zNk!$8Nr~B|)<Sn!jc(rEWwFeWSsJU8`sPI0^Vn!A%J`ONyfKIpx`s>>`ysPhIY)`9
zA+B(VC`%w9zGJTL{(V@z0J`vSts_Se{B7e8UGeEz3K;VP&&i)oc?J8X{5BG+d-ryf
zAMtDWyeiQ-h?VP@=P+jpluCL<ads0G9lRnNvk^G?yVr<xv?9P{XR9>*O2r(R5XBnE
z5H-9WvfLauTNK~g1l)PS7l!b_;;ihCK?UXmb)o0K9ydzc433!6*L~bD!6+Wr&oaf^
zajTau2SX37oD^mtnX6O}>DP16_c6M2!{}ton~#Ku1~u~p1X?)aJ2+LN3b$}QPmP}-
z|KtQ`U$l*JOLdZ#me0<Uu5k|Q9m>!U3~uTaE!yV7_N?Bc<vxLPw-UUoxi97QNWD=#
z!Tq1?2*m*Oy~4ZOKhZ4ZR88Zcv}9rrHnx#t1@PXVoVHqnbzuJZ3kW5>*HLKwNc}cu
z<)z9OJ)pW^7mtj&aqfxs^rz~_t)4WCiSx~XK^+7dN06*wD}jgGiR@Uk9w0M3Uy{?9
zpLyD%*-y7EAFtwYGqs%)UhfN0T;aZY2FSu&X7sSv_~7r=7LM<<Y~w+iItC_h1$aRl
zg;=%SeLCKvs443^AZ{QkRjMIw;DXe>1CL5j{M8Q&8VyP8FN)da7zSea4_|!KuyNh4
zv?jhVact3yWsG<=&oR~71rkH&J|s0``}|9+DK>t>^aLn;7R0;Xtn_k&Nz8^0&he_v
zP)X=b^_$)g#q7BOTBXssf^$IimvxMMW22`=-%2-sUIf{*Ec2n;*NGn83>@eZfXp?}
zZhr=wI;1CnNQS(J<T>|+xKxr2eEfoi{$|o;&OHtjnNrs^d6N8k^kRFd0mP%4aXFp6
ztJnp>7r{J7AkFQ5w2BB3u>MV>_?S+Bt*|JdiX~U7$KJqk=wPN+z(XxsxiFC@QCvJV
zR*BYYn0sH!7yTH_YH1oX1HB5E$rf!#T|(b^64K}&&&jL<M!><l%^}GJ>gA;!-=X&E
zC%uXf#ITw(2bN@+@vvRV^?B-C7z^g<qLuF#A#(&@B>6fRmRD&Dgab#Q=&G<)THct~
zq@!@IN2c5k0R+SSzz}73=go)FbiN~UzTzq_J0iZtKR6zF+LR5Kz-2&T&)YHjluj%g
zXDhjG$4T^e>g?>^kId?bJ)|Fe2^8OS)~Q<-7vrdx1ro=MIn$voJ7rVr&m=4dM~w#T
z;$`$6Sh~R(CPNHOPJtm)B=EMgx;69}?-KGeVL{eMwNJdD(b>50JP?o}%opXOd{hT4
zFxDrK+y=jtYK)megCuwj#!%1*Q*o!BK2xWtX~sx+_)5&aoLF6c&R>^JXD+a8vJ|D?
ze;#dZQGYl)%hO=o^i-Fs)y7yXw~)YB8I+VB#1b|2pog`RdC5tamxN%tN=Rw0;2Uz9
zy=*h9`KIT~6NMC*>}qs#*~xb^dv8o~W#do{RJplwaE=kNai;w}ah4~+(x@Kt<|C(?
zpi>ysxj+!GJvTv_yyDnLBhrh>i>ws=a&-0E4cRaXG;Hw<nYW}wvHCLOj=F5~^Db{5
zD-#o)9Pb?bY`rELgY<(0!OsHN`sQEC7Hi8Z@`g@G_Rh_#q5N-VouX~m-np$DG~6-I
z<IDRa0c3aJ#LDcBH~s|j>@+VVhs}$D+i9hCWo;0q`^Y5tS;J*h9kU&N?3*Cazc)7a
zwt|a0Yiji7IV6r)GbFRDM5)Ej&y|{y4bA0q)eVFzHzFjkDyy;cRzV}&n#*A5tgf0}
zGr@wz+(La9@O5>jBf22@RmK#&64%+YniNHyT!&55PCBZh>J&GP-5}n+yTw9P!4jWX
zt1OJj5QNOu4q0Rl_U)9ljoW<DTe$_}Oix3Dqv5{R48RtqlZp7JMI1Z!$k@u#2~{>0
zpjIxHpyY97ol(-y8~3RMYQuOGZe>A6m`HHlOV(y}OZaUg=hY(;k!lOnzW%yhLKPH^
zd_8pV;p@GO6_R%Mr|%m=+{%N)RcGE_17k3Cm)MJs>J&GLl*G>cApD~4Q0umk#vxOy
zT=;<1J`I8y?viA3`OXu6ZF72JI3RnjPi{Ruv@t$t+jRL;YvA2h_DRH!ajNy~<$><7
zQZoe1+XDHyBrM|eXBOZV2PVF3zw;Vu{{*(@Hm>RQzACZS9$Ih5{3doJ89@RjCnCQ7
zr`pYBr+0Of&nZNVoM~}@b?l~cR$$v*#n#lHH)~<Q!U7$7YLKv+d=@cShPwoOAE!?U
ztB37=n+dbon{!?<)cSokupl5W+pezXP1M!>cW#w3)!S}OVscgE;gYq135Q^kl$KCr
z-y$A5mm=_pA@2(@cn_YD1aH=}k9}%;JaB#3LU(?(CHv$W_ygAV(<Py3sGm@0vWt^2
zHm2>ZpvJW-nfz1(D-@%DxTXfCjL?UM*(}GYCJrR&BKkfaU+I{yZ_|P5Z3R091Czvv
z2Fb5D@WOL>&FS&U+9&5Arv<4d<<Hh%rv&g{0b&}c&m=T1+}_DaEpB$+8Yam)8Q0j%
zYRQG;pRZ}#_r~eYcW!cG6te>7X+4}Wx?-LLg|ji~E^k@QtBPTtER(q*0QzKW;nQ{r
zItP>)18;U(BZu~CL}q+ci?@9e98PF^_>Pq5*m5hB!(lD(0jDSy#r$INZqlso<T@?_
zEr&+O><(v{8FTbRmZvZ_uY&2Au1_1q_6CIKqR`6}+h}Bj0&Ib0-hapuQRT*WpMjIc
zs@O5Zd8onkH>1%aZuRTaN@m*3=xDBKSe?wedN>b4!I|gidze`*<AWG`<X00}oR9Bz
z_A?46>qK~X#+R+@33T^ZGbCD|dj-q5tk`epCtIP(ehZU!!uHjb;D6P!Qv%*8%?Po`
z2(cJC#Yw_^;6ndGrLy~8B#PL@`cC~VBE2-ltrxa|t{k>&x^Az!K9p&}aS~;pyIqfU
z*fekyLd`y54oY@xT&J0&(+jiMagj)KL*AAuuWskRr9N$m@R!d_o``=vxtCD;CAa!i
z!#jAQpfckA?{>6cF7nKW$oOJ$UP~Y{<5Iov$8=lTGO^PIjBUgE8Y0ZXMD=$cI(&Tc
zws;a=WPSdf%so4k-_`~vze`En<2G)$hy>M2bh8a>OFRrWvR$jMKhU041PTlm0krT^
z@|o)6&#+|D!1QC00nKfCr;Bt_Tlyb(s-VlYCZOg5DBM<V&uqJ&Recc#O;8`8Nv>Kk
z(<QlcIj<N7uq52_R%6dn<XEbouf?^$KK;I!1Jk8X!zYpU5|B-!nE-{){BqTbk`-Bb
z)R}LrVp8vw)_fb+i$;FNSmeo`RaWYXfqq_(t}?1=T)UL(JoO&jd5}tZ2*Il+#Nm)s
z8LWCD?=mKRWu7my=CXR<2Jcnu$FI4(sM3C9`q!N~dQQ<gFF~AmLp=0>!gsVu#s0QY
zg4x$M3z?&wqcsASlr4(E@W2dGa7M6yleNHB_dJ6&!tu)0q~z!_4Kli90l{*Qh^<HV
zX#=SSR%Tti*IuyM6A?WU6J;PKy8S07`v<yCeBGPt6hsp$<u{+|_Jyz4(R?z=hz~8e
z8cP~_^dZ;1nzZ;lw4Es~2UfEYTbS*W<V5~E<rZCHcpm(N0siUOud|@<UgEF$!S8|%
z`|nJj(fSsiaZmXELo0I&nb*9xhxyo_`SXd~4}Dr{y3sr!cU@5bwa8c0+F_7#G|<Hx
zulmUea695i%S=WxCw6B}!pl8io^61^v%MK0MzlPr9p-p>Hoy8A%r!M1@&r`A&3a@v
z@z6n#w%Pg=nf%Y$;@G5$EQVpdbgW}abFJ|^#Gf9_v4*Id6<u@tTQj>hZ{xgr0S4FX
z$&*iy3N4W(6f_R3q>D2FX8%@A^K14Zx=<SpjgLNrf&I&g$)6_W`rEM9;?;e<ca?W~
zC@u7xA9W|4`TFKxRj0o0=UV<Thsk(*_}-ryf$&iTI6-G1j4ENh8cnz@zWDhEE#_-$
zt>m;>%qR6#(Q(b1+WhTdn~{(?JC$&<o)h>Xz;9;7ChN?suMp@89*3-ZPI$+W7bZ+7
z9TcLLTa=a#MX_Vn3e@)<u!YTvy48}rHQ%GZxd6&lHXt9pEPw1kmRmrG@|(G@)=jZ)
zxzsfW)SKnweSGH&ayulVUBtM8ES$-IRuy1#VD)wV%bNuER$BLRabKHf`TEQ{ZBa#N
z=JAzEnL91<R7d?{c(;#RMm5=6pB6REkKXxoZ8_$C6BMqkekNrw$yE_tQYFwe>F5w4
zt=(1XBFpHU!Ja*}I~)nyeQveclF96nF*zCGeQ6C4Ji|gD@Ol#q(7U@kn;6r_3EzM<
zu`xepr!JrDVYB5KT`y{NftXZPK(f76u8*{uqm7;9ZT2uBN%~4EZ)^fW)a*#QXTzF7
zWjQb@#w5q{bB(mI>JqQ|JZwvjPklSq5l5dmIcXKVi4hy^Sr+G4iPt#nu*MFZRQXo}
ziwp-fsl#K7U_)%xl_K~N`WEss_c8az8qVPsUHa{1hjgC8yYkLUIoaRP0$Z^wJ4S0o
z_z8QcPbIS&o!5~3BLEAT0d*}>M$TFrV4Tp7h+IWz)K;-gN>Xp6d1iyNB?bopNa9LB
zpOmdcD;BJKPM_@rwoePghJH)ZFagvs5hsn<MTpuN!eXnNW|R7+45@ds4w9hlBD6$$
fZDjk-=p~6V*&TS2*dirxFp-9;4zya?_RaqRovFNw
literal 0
HcmV?d00001
diff --git a/docs/source/_static/remove_background/v0.3.0_rc_hgmm.png b/docs/source/_static/remove_background/v0.3.0_rc_hgmm.png
new file mode 100644
index 0000000000000000000000000000000000000000..61fb1e7518673bcc57564ce4e14bb08fd9d0d983
GIT binary patch
literal 212594
zcmce7cQ~8T`?n6=sG1d{LhVs4v1u!`D5|Z!s;Di9*rUr9s&=iS_8zfAi=tNSSc#fJ
zkP@W!`?TNR_jkR2zJI;vx?G9m$@ARj-0O3H&V9n4YCod6!ghs<ii$@4G4vS~)rD{>
zs&j3ZE&!i=U6!Z^UM{;nHiA=8vC#khJJZ8QR;Hr5MWqgX@WMN79e26;wrar2R?s=c
z_;V5GsxGnqtdAvZ-rl^ra$PqnLN3|v$)f!Qtlf=%SB@=wakN)h%cbxwbV~9YR;@HQ
z_4yn9Q4wz3pKsO)=$;FEO+}q1;lwQIiq8)TqLT{nBdx&)&;-h{{{TTOKWPmZWkqF0
zc6V4W0doCk@bw0!!T$Tto1ZE?fSCUo-_B>p{_k}!&kj94|KD%!ytU(X;Eu>f=`uPa
zv+W#)r~^B~<X`%WBoHMcvZZ|4i#u=0Pz|zT9(^1Bu}foNvHI5I0@ZQO<{3>RN!DvZ
zbm9@&3E+87i2hou{t5;MuCltr>XiER#SPUVPh^bB#rNMgB`ytTKbR3^R2_1lV7|Ys
z$o<h+ywf=I(0{Kh(l%*!!>)-(5i8v2!XKaOuBfJ4V+1vw$xyJT@K9-<!#r9LjR1R*
zv@@yoecvJoYwq-u5Y}sAnLFMOmC9g$4m9Y#-lJOKLeLrZtU+$*i7s+J9ALY4AvMw0
zVd&w!m{;HX@4p8BN|0?|s5fJ8wWKB*mRa2~-yw^(SD_DnI~SI1mju~#0)As#J6Q;k
zbRJX2up<H+;<u{=BE&{C<jc;9{KD@IU)-;G<rQk)o6fn0XeySZ({>(lzUX*<Q}~FL
zD^gL}6IM<vPyM|7eIcV|cPV<uxV%b5_*y;Pu3crqMFIKf@>-vyyzf#eNp|Acq&=Zb
zQd4}WX05v1xy1ThW3JkXr<6CjCGPdI<xN$>f6{J&DP$%7rcj|$1stpDP{`%SqJiAZ
zG}gn<Kp@n|mIquPSkQg0RF1~%F(#+(F51UK@c6tsv3BW_-9WC;I%E>mJkRI5M%g**
zhIjEw9k5YHX)Ug@VXtBX+0igsi3-iSrYRHYQpqX2=X6U;TTEq*0ND!7Y|4iFlF3Jm
zMjUvGd)obh5Ucb3!U=uvss@c#8y1u#86!&-?oYb))$})+`i@9<i;oZvYD_=wNqX#Z
zPxlVnyxDf8AHG?!Bie9K@gt=1q6h-9%&Wm=ZR&^ufmhv0)mpPf%0@{>@@K4m?d`sl
zzTDV}+w8;S-a;~9?JA0X?v7Y-^S~^M&9dwscGYXoa<(Yk>v-aPz_%?_?p9DIIrGa@
zG@&_4;pSC440^-2zcFKAQ>&ce!utotg-s?`4V^nbnHLf>3t=B|I?AypI&w~aF&XVK
z+(t$Fm5?t^*q(@*Pp=-~+P3TSVF9bo#UUhi^ne+Rq(yC$1vmYJ-oAc-n&>&FkRKIZ
z=99E+Qj8k<W8L4hnXl-jJqTCGMfR%xGq{(l#N89=IfN2|MyfphEBxUBh)#<w{EGNl
zjvL?&V|1fk!|it$EB1itr5w~QKV$3RarSszp2>#}Ke~|Dk_+q8nfE>{-g(fs1Q#}L
zK~+Ck5Ysvg=(NJYQ7^Xd;?p{mJOn~&-YqY#yIwo!b5A^(%&U7hjlXPc*^EIy4oy+7
z!{vFc+-ZxMGjiXrl=fJQH@})sG(c@S0_KNG`ZncpKKT1fSACqVHeP_^^ACN_FY7WF
z8kfx+;_{kCs#RXC-Kn<cp%i^mb72dosu`mk!#)xH*VsI9?Y!v>!?UafBaatjs!jON
zj}v|C9&vt~A~JeysujQ-#d7O)7Odl#g~lcAKCo&Sk6uY<qM5BR(a4jG9#N<mN~Mvr
z);@r2Yj6lRm)8Y-P80UMVOcm;g8iPNO8#tTrEMHLveKYRQnkSvG-7ID4|drCuXR1^
zk3qAv&wzGpQI`%U@<;*^!)>dO(l>K0cSD#heclk_C9dBh-O%z1Cg<9lXAX#K^3cEe
za?pILOL#tg=Mgz2`L7PRP<}UbpH4Yy2>aOMoTBSZx-*{(h`tyNIP^D!NnES$8a5oN
zqcX4y*7ZF?g{F=4l%jm2rE?dsd)`r617pt`I0YH4#M93QY4FG3pV8m5|9lyjs~Aw>
zY}%m6H}{BD6Qb;?qUA?jKjrgVCLcyK&4webGCrlhzLY631R*&HC}LDGQ~|jchi|?S
z<EtIW&$1j>d@?M{eT)o0*Yj?!Zh7pI%SK$$v{}N<tM$K2*-P!rVgpu&HvY6vJ3Edf
z=UQ<d#Ep=~?`(7%X;*D!ol<6*i<qK6Ef_ETAh?J=Au#LT>oqse88l<gNN5yoz2WaA
z_e39Z&1a6X(8l%f>*kRx+l8Ng-%9_Kkt9E{NGB2ZjP-cEu?C;_r0uDHux%io<?32~
z`LVB+G@aVge7X@UOKBi{0V_1X7NylqsHaQQTVHf*)@d!@HZGo%5mwCOlWuA38lE?$
z?(cQ@Ag^Ur>T<y5RlHr+T7)<9v9yoVfpnMGA!G*Buy(U^SKP{7ym62<8v|FJ$NXQE
zgo8>0BwhZL7P`=_D=b_;K{uv7hfVenibE_N(pz_qOKw>ua$RiUVz%FI(;)6_;8Yx!
z7;G?6vb|-$+dN;-(m)qp(5|ieuA$gR7H2ei{RMtix8KRAV@uLebKC4Joso~2i%^LV
zO?9f%k|Q=L@JAMu<;=6bQ<P@>fDoB4i}8d0Xwe{TnQbqB#6TB3PKgKp&~%;6S`(~d
z{0CH>Z~fE5exo{yqv&WQ_{;Z~PyY;oG7oL6ez<Qgl7?*z5=D}tRg}8-P4#Qam}h3h
zi+4$a9M5H7^0}7%PF^VO`zduCl*!N4&4)JcS)#Ldh1=&iutKoIB&Lr37q4+r=4rM`
zW_r@J4xR1q>bpZMM&1Ou)UkA!SvGr5PZvo?P7x`|YWfLIw!ebwbkhjA{Kpo)oEmlc
z*k_wQm)!zUZaGf-iuR(_+Kn-73g1jTg~Zcq)J@FDNfEAj1f_sKMNEVgxJOD-;@UC|
zDkgWS;mYSEn*qnk-6{z`XY7Qzu0s5>PKrnT@w!7SbM358?+iaJd7BtewtB(m!Gc14
zxhr+O0k*{jjquv<(&dSA^k_=$Ju#?6A{&oB=dETfdtxD`Fp)N{`7}sCOVQ25fqhL2
z%oY4z7g%|7A<(d@oVsiD(LtH7i@*eSPmaPaZYJ5&WYRG<5&!<Ig^r4F*QjVJt=yu@
z+f0rFU;Nq&rJF+yph)-fx_)CHgUAhLrE?ylbaXuFyLuU*C0Wv+G#nA$UT-e`c9Z`5
z@bDKE{LFKX?-MhpjS5`4gg6j>96?7-D53b)Uq9sdfnuIF-2ZSIzsKt^nBbCQK1yIj
zZQE=5$kFaI!?I79J{#rIHBowv0!Stex{Rgc8|&kE=97t2M}&3~PHx;)%=spM9c@(o
z1ElQ`ya_`LxRH=_>9*D?Nv_t>j_;*)H&gJSct=T*)RFojb;QVT=N=PCME0Q`%Ed44
z;vsq7X~b%3BR#dbeuDqV+~1}5T0iR6=9RW8T?(r7WpYeW4XY&5x{dd<TCH^m{paC&
zqb0P_AQUf=o?~fYQi<+ytX|m1&7oIKc9vmy*|O>C!nF-FpMA08$){ew_tqp6W-*+k
zx1%x$hVo^pkx8s2Xq;WPS$(Pq+Qsbns#BVJh_~yW@$ZQ)Ern-WPJNP%_DF3&o1y*P
zQ>6(>E2mR^ZW0Nbv%MfOY`r!mGsNDRQS}+mK;!cF@9j)$=4OP(e%D!@2ART?(69FG
z%88P#`1tChr*=oVgfI3oYMdwRTIBg!j6v?%W%2KsOn22XxF~MQB9jJj+|qKnFiMCg
zqw2IrM)sI0?B3cIY}C6fZNg~8iLF)+U2k%;+WjdPcBG<50+A79LQqjulz;q-P<nkS
zPB*P7Hqq~%+XTV&fqi3dB&BR=4$M7bB)-(3YK)Cl3{lDeMMQ~gp2hRL%gwFzfyZE=
zR`65ewm$$MJFxrS)0;0SA1B8vh;t9urE#5i%2DH85uAE8=ks86PZDYv^LiOuZmyso
zw0^m*Xlrp5AGT}!fYK72Oj&-1eWobjTm67x)y8JJke&<F7{UWLwZH1HY0`xc9Jwkk
zol#|3B9o|x2VDbc%nJvA91suwcqNEF=NI+qdXX4XH2i8svTAHv+lW+8V?)U4-_@%&
zz=u@H(76So>kG)W5828Qvqb{Px)A%~=O@psny&i#jv0DA@b||0h<81snA+6R3yGVU
z@XQD|6KjNmG{_$|RHucTXg{}9n)9^i`)<`#uCq474U;Ws&YI<9wKJ`ln-LqG-P3Nb
z_P0Xv*YZx&JoM$$n5H#HaAkS0N~Fr4t9wJd;x5PjdSrjE0gH!9!Y@Af@b0@=P})CL
zlov*jKdw&8w`5DV7&!Qr2GE4GSnRl&pH*Jib5o=#Wq+QdFqoUhq9A&{rR9Bsj+)1&
zn*OG!EcSsPoiEmR2=qI>r#&FQI->{9BTG2b+os~<!kScn6%lhOC2?xGKR}5AIelDC
zBoFao_phBDRb2c4x7&5)rpOmLmabNAuTq{Ps=Rzsj1P=TQ?h1{Yg`{EG*Z_#Yh*r=
z8P>sx4Xfgab^5^#YW|}>cM91dBGu?B8Pg3j551J(wW6r$-us?otm7pW@`VOB^;75x
zG`Hgg*I%S#hqJ<bHFy%JLyOuRkXFWnHr4CK1=X`p30@Vj5DVWNySqn^(!GmQYacF(
z-zGYB=P{7-jD1_k(d7=fvcGzuVO~~D->&LYc`#OSP8fa6tXMxP#_PA0w$W(X1oH2%
zsgR4~(<E`>CE}$fK6Jl-A1pkGZroGMvToArs%X1Ph)CN_y@stui8x~)_pZud9MOm5
zy~K9?Yf?uvdqay#vj_V{)~Q+Kq?ed0QtgU0NA;LP%b&FlQkx-9W>wuj+Orn~3>>&P
zapb!{)xnN@S!<*-iZ0aUTzNKd^<e!Y!y~iIx_5v=d^SVOHbo7~w|YgS>Y7kaO=i7+
z6~fsAk8CV!mk+^bf;zc3AL4l)=H)_s;In|U_?|i*`%41oG@ZzKcS~l}xBlZH#~5mS
z(Qtp~R8W3WxH+nP<+-fB3r^!&`9n|V<<k?If`KDp8_^ul6kiMNoIp?5@rIsF=LzVS
z9vweo7F@7P^?dl_5bx1y)inCFw7_KIr%7v<WL2#?+tz&55J&~Ti~5z^O!CxY>N}Kc
zmEF?SbDJ)Bq&KOe-Ci>$7*uFz>$f)pS*NLKxK*M`sBttmqL>B`^bUs(7}Mn1z;u0J
z<)Z=eg#tkilC4_E<C4-;zKcwB``qP2-{!9Pl)J!VN=;;bau8dKUghz+Tom;^W>n-t
zZKg={{g$2xjBBpnCF}k0Q~?5229raH{LX(@n~6^)4<^n<Ch(AF9bBgF<JM%$zVCGv
zvX5eSCF9+nAh+-5`xk`ThJM6)Rc!wZ!{7$@j4I8!Xx`;wDu9(%(Xh=C!XVAe9(C1k
zjQ){ntH0{=*Yw#4{fb!nDBLk@aBJW08PWgYnBujw1VtK*dba|M7i)~OTDq5}GpRn4
zs_wH8T{gr}KcpEJOrz>^Q}4TbaOpH2;-J{=BVD}iIsBxcVjlgf=|Nj;HM-}wwTl=c
zNwjevyZO#&9LV?n*0VC*X{4U0FG3thA*6S&^d|Fadj3pk@Mmgr^M$-yggO_wfl1hs
ziVgB%U@7?~`0r{znDIUo@v~Z~YOcsXB@qqT(u!aD81>e*V;4)~R=J<?n+it`!-iz-
z4!7pN!133&*=U%`bZ3gM)|w@YkA`8@n9zY%{*e`BI~l<czeD_3?W>o`bFk}qjXrMq
zgyK&d7kVwGz793|I3yv?DYfo+Dwj-{+n<rHHm%r6Fa1@zhMx3PmhFWN5|v1Cl-G+g
z3GWXYGeY#E1tT%n_n*>XXswKtZf-Sz92)AZ(Eu8FzoFNnjzcNTc*%iw3^GjO5gb_y
zwM()b%TjCyob-1^rFv2i1NqqM4lQ0c=wlwm84?((HZ8BqknDZ5A1?avG>ku7O80N`
z--~;-YHcum;r(|$5`5t09U1r?R`w!>y@o!gChZ%a^<?QwwAmZ<#2ZE5f4?RF(~2!x
zl!pINul8P|_CrtYxNuQ;TV<h@Ox5)CGaiN2TF96oM3GLF0bQzrvH#(DbiAzPFuvIQ
zhvBgFN$A4RWm)2r8>UioW-F5J`u&o_Hp3%Dh!Jn2RI-_i|9Pd>Nmu<J#Ehkxj&P|x
z6I6dA;U4tl*;l*a$(=uW+v6kQN(K%}aWscLpR8z8494XQ-fmYx;sWvyO;h$yjr&`-
z=f!j;rQ)?G8nV21)sAEw;inThp^lP|>(xBU@u49>UPGu{cmA5u!=5YYP}CMLMzqbh
zfg<A3s6>}uXgN_rXsEReZGGEFs3DT|Vn6Ojp7+-;u$tI+%cQhdZH%QiOM4t>lQAgY
zS+mlvQID$MQ`p!<`1ot{G_taf5B9IxnlnP4MHkldhjfP?_<OSEZ~qh+zzOW76uaQZ
zUi#U?Zc)g6pWuM%OnDdK0TxRQKMb+l+}jpON~ZNfN|Oqnj95)^O4xn-LVNrS<=#*|
z?$qBoGls8nfZq_;Ua&UjJ3%TaLp83_skq+%DJFDPTdX=f`|-}%=`RVq_OA>JFOR3M
z;LN$*f@h=>PmB8A_ov7<O&$jHuNVsM-^8wkObcG=N{m=4Y8!g(=LcG66G`=YyCHgi
zY(jgicIrb=oznI4L9&^&7?Ii2IG;Y?n?+6nE23`1geUf@38U&}lVH-m?cI@Vgi6M&
zq_)08T0AaIA0py37I>?C!Zz$E_p$NMCLLBY1-p#8UzUaoj_>ox@Vm1cqO5yvA_z>v
zl}|~lG_wJ9Ia)dQqj$J-R(F%?NN^9o;v9jgBI@7z32Npukjo}R`?Q|OkShjsl(oiH
zp%_i8vI>QSIQ+0IecjTG{2Ftbe*@OrWV0p9M`f!0S0D{UYWp`!fuDAhl7;A?$_e;P
zCch?)yfSCoafPIa)bPvSJEhoZV^Yi9cJK(usa3*@E<9U+)M#ExIdRI6f-HT>D9^V{
zwCc%7S6Ni~p26`VjMx+d=njLsLoxcQl@YIr!Mw&T8!_2T0|2aO!!w|a%hyNDcG9hg
z`0+l5rdSdgFMF>#CacqI#ScnnU+9Tf_-dQ&)^xftw!Clm4hVVUicw!V%R&`kD_G-p
zcZQIxubPHa93@32UzH)P%3fV;q*K*QGvkzJ6K0h^>%&yd*rq*IF_AH1b1SZ<SC`cU
zjcV<(XC<GYLe+|s?^o;%a+@1ZCG4X&D|mA4C7!Ysj_KX4)=tW@6B>#00^nBdWSXLQ
z7KZfYJ7VP3+4iC=R^bu-f}`C2eZhk=Cy}wLIcgPl%{T?I7v{^sO{>3^d=_p?59Bz0
zolQtN%G0APC0d%?u~wn;aV>DAMjLqDbm;>X%fQ5t8@Jtt_)cjd4X&@7%tA_8eg33o
zj9kYUP*2k&wpz#JFMA4QE<Ef|4?C(S9a^e;Un$2n624le_v~ye*(NYJg~}M*p6*V6
z|EjO&K~77CT$0s=+e>dh?msQC4Si5gl=t4Zh7w#>4n>Ue_3?Ue6BL1A&6soZUb@Y&
z@!<AM#ek2F&D_)7r>gDod710im)fJ7zEHvj?SDqj_B8vj{;OSuC$w`2IqD_k%SMzt
zMZWb)d-AN`QR4<ipL7m&Ty((YQl#x7>4Kg9Mvke6GE1mWhobF)#4Q(R=jP_3+^6kB
zUr~((!D#iXfx0sK7cPaiqVr0BBSKdN94d8YdrEc_V?*39a<VYC4Voam<h!t@rj4{R
z>9U>YQ@bc3(P$ZDjc;6&i|?@X1BjjGi3Z8{r{B3-6Z^ZP8-imyqAgWrjffF_!clHm
z3w31`uW0>o=?*3w8@aOmj+^I--wkTHL}P-GtND1Xi7Orw8A2(Z(bhc}Gq>Ph@o}S*
zFv$^+L;nD!I^$*fdvzXK<Ycum?U>`XN{KU;!r21GrE@dSuCuKEzJ}T>#r}!M8_3f%
zBv_@+NbS79ImbNzGZG-#AE^?`V|{XC+lYIj@eFwZzZojk{v5UOYjv7vHV11(N<Q&X
zSsK-2q+2kgWgFWaZD4<{%Xul0F4)&|W$u9XnbOU<29Re%-4lYPb2BHjc<tKKC=k`G
znsU6$%Jnje8%iqy*LbsEj_-1?Xyu9ctD}$hf@8N}o4YA?*GNG=bB+i8lHmv?ZX&0e
z9v!wT@R3_7LJ*#HQMJ8aD&SX7Qw4?g{Yci09Q${IJ@vyCHKom0n4&wU$(>qdQSG0p
z3mQ}GeBrAH*Bk0w@3CIf9(T>dvEmF1^{hg%9Cx#b4fc%6ayz;s5_{9@j&af#PK*LA
zZ&lBJBuMK?bJ=~$Fwj?6994|IzONZu6ndc*EPt%%RTiK2I}s@{vt;-#v$$^?C|c|f
zFp;xrl(fc$pN?$u_ad^Ljgx0r0%6|$kcfL|wGm6$Hk;K$p!jHG+BV0bz;cR-g?*>R
zW#xIp-Yaq)ni-ZJain`kg?T=1-3?*>+}kXxf+XR)>;Ev9jrV8TE(4YH9kdbYjETp-
z6b(W5aYBFoQ-h(34*;E`L38Z8mc{;jZy5T~-;NPOt0IyaHQhjT%(RhPGE2FT-KYXW
zh|(E2OG7z_9{IRbjw8B1W`c}2>OzS&jFhmVni_-t<_3~&cIAMIi4vFAt8|_2#!%g{
z_T9=^#Ss5eYS39`%XNc_guKgT<r?pvmX6sl)dYOV+5D5L#dp61J~S$LU#cG}rwCm}
zpUF?r^sP2dZ8hEqpbmPTn}_149pSyU%38R(FCZ7gje$YZ5HLI<n!ohu^6q_nXd@Ep
z*H%PF<9KCRLw(mQ))+vseZ6e0=u$VST)L<eNUBS*5XXiJrWb<4tjQ6*Kj-~KFMwDS
z9tSt7sRRpuc#3tdt^2xCKP3!$B4qq%M4}+I8{)c~h54PomN8O>F;IQDEPi*<HduGD
zOJ(ep=>ea{h!>-({nEOTdAr+KwY6CPRj+Qd1S0Z729@%o0uT8fqTmZzKI@zLiT5$I
z?JX_EJ^VQeUhOlsWdMV%zvwciu#$EGZUGQ2f7euwn4?f<NR+eg#?UxZKth`jXa5!s
z)N;%HW<VhaE2sTmX!T;(HDKL~OQmuAp;QIW$0D_CzS52+;gGlXIrhuC@ooqX?d*+I
zUhca;S`tMU^?KC5sbxE1iSV7~l3l*ta&%yC#pAGt$@;t=DU!{etToR@Z3curx^u73
zmCLmjHo~TIVMF7E-n7T#`bkH*-9?{{`shs9LRZ@jjKo7F(`>eIvgIXiWe=BaoVUj$
z%Zv?N9M34?wBtg<YtEFf2D~iVghqT$l1jgS7bcG=8ycKrTSm(zEpsR3cWtH<aZFn&
zoe!2Ipt?1D9NJZ(oY#l%iggMe6*hL!zp5{LIxX635wc%$C*sJ^w~?!sCQ{t+Ks)r;
z*jzfqCL?xAzbcOsBE{6ZR2Zkt>?kWRy36P=If&swAbdOpuE4^USPB6AS_hXdv6=p0
z!*w8nGOKgbAG;1m>`i}Bn6YaTP2$q?abYLQO|1*RP$3D-i|r_at&N4AB}h|tpGYS9
z*mVKXyp&v+tAcAkt2o*ojE7~kl%u|MMQqAR-$<CbHH>I#)|>uXSyU4F8e1{ivWi{r
znO+W&h~HvI_jGuGY}&2HCUqvO5crIuY2jjGL*cIL*XfCs<KkYEP6Sq~ymSBwmBN&J
zQzyI>JvjLNjOnDYiv=8H^kv!PyI*Mj(2_w?U&1F1uA=Sv>XAxG1p{t!)sAzwG0eho
z>4(G5%sN3_V0vg&8P;OMT+%l%b=@|Hv)zb;&>@6YI~OY4fnux*ZO?G_R@@nOO-iq$
zR{Xea&AF{YV^)`h{TSra?n1GmKZ%N6{_&Q~a`{imQ`Dxw_Cm|RgA_ij|HqwDn#Yc)
zGpin?;vN4w`$hVI&Z$zeGGs&RMPgZJA-iN1QShZ~o=88<0-N2`IqA_|jz<5x5pZek
zCA>aH*jLL5yNeO27n|Cw(}ZOYA{Y}K5(&(HW{+|Lf+ME$4ry4uvPkr((I=XEA;(W7
zT_~;EfMpvK?kzTkD*g@5w=l&wj*N*s$J}p#Ey-UCC!*k{?H01~=ec?dF*u9ia+4mS
z?s>%q(`v1wSmq`SKH&A-x&6jncG0>z;=srY$H|SxjRe_Q=RwlAK0DFoz^J&le2w*F
z*c%(rV3DUWFXVOn&E6UnScd`!)~?A;nlo;Gc6%n=C%aW+6!f~h?iN9szwF8a55LpZ
zVKeO`T8&}54f3Q^kl^IzRFOCP48qKOIX$M$s$7n6X1m9jP?X>7WROn0&vo95DOXk;
zFFmfHoU~+ik%{Ocrb%X6^;ALHNGGNPm8{(_SDIaOGb=xPg;MqM>d!<byr%m+SQ`KN
zYyKxV;lvQ<amd}qwKj6+A8o{=q_QObfZXz#NGqzmNenck&*(=CXcg>ia4fvr4hd{U
ztxpxxx6~+97ypE(*UdKDPPhqOL~i@8nf@azZga8K1bMRd{wtrK_M)W*Y~zUjs3hf4
z?TabS@p%|IAM4Cj5da@15FTcbzhNd-{9IV$>aTkKt)zw8VpF9l`%fhE{;6nm-=o}*
zRZr1$<v7tPr_Qld{b1o<vkpG|bCH~`FZ?lsrsdn7-t{7}mgZDATePwCbWx9)`1_J8
zy)Q6x1yQq46q+n*u`^ziG}b-x7CtG{@uzRrY~&{UzNr5pe$ACq-7-N1(h-)cvPlRF
zdtdW(F%#QXv$iw_orZ+Gtby{nqoPK8Oz+hQA@dRM*%CLy1T{w-ZzI4W^)l=KVH7d&
zp21fMJDOfJ`>PY7o!w7Hz`ZW2B|Sb}`)@A*j#L)m2n%M(RHekB=^m=C78KmT`@M9Y
zl4+{@o;`n*w`%FA&ic+Qhr6BW*0EHhi(^NF#m##KCaXlkJ*HU}oK7vX8Ai{tJFO|(
zzryveuC_<O;x4=XW53co`z_uP0&WwddH-?nmFa3UyJo|62HdD(UXIbg*MgU=K9}cF
zQhKb&!{)jz%Gyg`2C*Z9JrYvA45j2O&J2#uD*wA>1hE|@;c@#^FD7$onqigV?cdi&
zCI@@r3x?RCM>8Q$gabNoUtSxR!(%R`B^odF2PiTiO*q-+{~`Z(qyNYbtpW94|HzIo
zf|t^qa_;2^;$c#NAHepVDPjd1Ork$eQZGy+<X63a79;!`s6|MYh2x-VJ(%y(2b)qB
zfM)%n{5z&F)O8e};U%Ua-uR%uZ~W@CCv2rk4?msSG~9EC6f_Csow;(9)6n#DJtD=`
zCGW1i_4~kD->vGNrb5YqzUDj9zF~dMbI)J=9Ns^Ay_aN8aFlfEnHM$ltlpIO({kpV
z(aU<a^Kp2`zjnYvqqd>=wYY||wr1T1?&wPSLbQ)fDV*uR_otCdQU+O2F=Gtx^}gx?
z@_P$i;)!S84kbY77b2xJWp~5_U~t|SuQd-oy}@@nrPJlR7X!26+c3xtzJ;};n`4_o
zH>#x_?yGNHh<&1y;qh!$Qb_~Ud-st28gr)>q-><#C{BGrl~UBU5qc?TMdwX({Ww{#
z-M;{FGHif1aOVvDN3WN-MWHg^PUVsMeDx^V8_t;2Iwa}@Rb)B!D<OD%Wi}Q4N5ttq
zYrgxSKw19NrDVCo^``^Gg-J`1Uj7B&&4;9F>vT0AgS<KK;=I@%ooN}nsnmjek7Z@~
zJ@X7uS7Wr&eC-1^j@8{86;H6TqKKcVhrWDl`3jPw<e{(Q|0rBWQ~CC?(5t$w@>Pzd
zKP$Iy;`vF?LprEWrkQ=-;IC}Fc@?_%3O-7M!3}rpRNljS7pBh=0I|J7_n>{q+7Hoq
z9n$Y~PulymdaU&i+VwTa<@1Nu@hVjn!55{o&)HIX)OnhQWfkwvIhzVjN%m0_$9>@W
zWBa0ypP%Hfi@rCL?Ime&@fQ1z)`X|l?rVOpb)aMVrT8LViPv_%(&19mc-J$6>8=~v
zyEhRhT?-NT&FoCLOFCm~gZCLGdl`K{m^PuMNy%GnuGxS!{Y=$NFCWTilZy94^jBPq
zh(@<dywDTd1Q_~Nc||W9rs~{j65GjOCasTK4x#wM^k#hPP6~yoMETQ^%;IqcSfn&$
z%>Hu=KyQj!5X?<vMGW4%x|DQVj3@vBckXpp9=qiFUkH}>z$aTvwmil*v=TtH<c6ya
zm`@0{B{Qk*H#3qW+{cifc3%x!6;TY*%IuMa8|WZy4}m@}GJ??1G$Su7WrTO$rlSF9
zTYH3()QF!APGz9>Z<;~{dVR$j$Nggyzw*E4l6(6Ce?BG2D>=$fV=w(0#@Y7K(4NQH
zZ%haZ*4jt0Wd|udq+ajN|CyE}I%LkNxkNs(=rRL*;79HH*F)9AO^Ktce0y{^lSJEF
zo-N!1T?9$plW&+?Ttv2~RiE?}A)3-=Bb*!RCtg&oUm<P093uLeyiQ3_wXe~fk?aTD
z=Q(RjvyF;H>hPRl(STkiHgYlZ6!&>=D((FkzV&lnj$Re%e8kn^p!cq(d56l}woOgD
zO_1r$PZuf7W$!A0svCPv1?xW?!@$C2;xNN9f`B9}>J;C_sG-OyMH^l;L614e^HUfD
zn!0JZpjCO2@I+_#_QF$2dqlu!i@|b<iF^P-5j1|uU}MSA1!lvCGsC}j-<A-*NVDX7
zMK9W(o&`zHdWxYHZI_SV{#l>d)Y5}a*=18CU~y&nj_jiY?FSmR?IQoQFye|bQtI$L
zXf{8axGI->`|?NWQ2&3<umd>5*#6=N6auHB`X>a?gHLNt^Y~|W;e({<$LGIDGQeRP
z^{MMwP=*RgRo!^n(qBTm81|(1>{!<`RBwi|{LfNRy%F2v*?V&URIfQOUnlOVH@8It
z;_}8)I~3XeT}A)p-&JLHyk{MT#Xf^PR^egNFV&}~g9H0$-A&tr>BF<_uCG)lz%Nln
zUR`1UTDCs8QdM4hF>-d|b$})*?W5J<IaW>km~=P5C!|bS)3wRR-ge&S->81rz6@Ob
z&{7-xM1dpQ62)A$&F5>D3tQ1zJ@3M=<C}IyGey5~eK@>~BrO&yS2hQTYpJVrhKlOj
zouRl(_0_IRSB#6Z1*S@2E2MGnl;Jw^hw%Bg;otqCR8%kH1u}uVzqDIXKR>ez!8?e0
zl#af`E*(r}QQrx1rQ*8}6f8gbqMPfy9@&_-Q&Um3<@}qeW!C@QgtpP8e(}FuXvqqa
zH-Lx#XFRzDw50rZ@Lfj&kNls(5T*wH@B25Gq5rS_ZNTLIGm`KBe|pSC*WrKDxjzeN
zUl8ng#k1P)4r^MwIv@vY8PK$Med(V>UkreqY3A5|>ca^Mpx*y=4fymLOMCzSm$kUa
z$l=bYDh#zw2-Nq@#DV8qnoO0aq1L@G@6@NA|4)ZtV_0_gm}1HJL6{&ezBqIu3voTQ
z;5{JaC94NI*#Z&zs=76LP|xu0*YrjUQjFnB$>Ap(C#?^lVv*T))ew<0tZyfh-%@1D
zv@gA~65Ag45hTzoh(MMLCX&Th2qN0!>8}aP+F{83k%{D1By!*1sgGi1<?@^S!R-=1
z71ezXV11ZkH;l0r8pK<}jTf*lYxIe?AlCs-5VHKEQX(@Nxv%fkCsiroRZm!&Y-o6N
zJ%#FxI!Z?e?7TQRbysZ0MCrLMV|pJ*k?IX6E4{3)^r02_k2aA&$P@s>`i}gt1#Rc3
zs0Q^JRYAL|V{+qax{^|WD)`!lzTu#YO<?z|y(fN2VG602&Qm?93SbKdb`~}TB5GHs
z<(x}K@sIdF7Ew{f4G@3aOEQ$7e(o_N0A^VQ)%hufqmqW+15KAAcc0@2-4PA3Pa(7y
z-rtj*Ea0i!3pjbDaPJ(|$9pqr(WF${x)Yd}YO|I-#T5|cp2472iK1ZT9x@!o<w*nW
zkfowZzU4Smv7ZKf{l;^fExdL`2g_Qc4%KFfJ*2tt9_X@suo&o|la<}cYd@&S{w;Y#
zRx9uYQ_n=Q5sWsT0p6n1+VuDoQYldmtg(YrA4)DX0>t9$?9}%F@9WZ#)Ya)P2wAqR
zlxU4Y?qepB>-Kw@KCBv8S(&qd$dwbxrxC~o3COZOE;=>FUJjDyA%>brzFGxW{L%Sj
z_0$YT0|JrZ>HLJ;N(rBR2jNPIm}sP|IAnRcQUd#<lebc$>6KLrCx{F@K}X=S7bA#l
zQYpa&$WqxkEe?DI0>NFJ`rZpi`(={#D%YKcd@$31T2I&v#0CP3w!AEz`r5CIE9=hw
zCXdXKY?kMG(hY<m%axT9sXO)MSTbpEKYb!OC<5v8dTa}Do|h}0b^#r-Ob9x`W&X7X
zPJI%U5*Q@X$Iq$HcDF$>V6P>6mD9?KlM6&{nMmg9?3`f)Rh#t65s3rESAlpXf2&DV
zI`b)`sx6>0Emc3zwF{eXHx~c)gK{rd>90+Elb5ljsxq}yfu5*CITQBR<_=^PA<Kot
zynM6vZwAms?v$lymRRY?ZhK(L>LtfMo?O*%cnlUnX3xsZ59`jX?~L!5!i>x6R!Rhc
z!KCZ4$LJug0{Bur#vZs3lbW8=fZhm?K{kYjHR}+3d)<-7t!0_YQAi)pR~!6yiL;NO
zN{RP&@>53$*XiA8pylDn{pa*XfDzVkGrzK;Wr&UOhF{G!2ZO~Uko)M4DY^?2I=tYZ
z&FFgsBF`@Qsn)qr4;&ZZ%I}Pxa%Cw71I?bUoYL7pHI^UI+IqUU=!MD_Q|8QXFvP-D
z7soa|d}l>jV*Mh*{EXJ=URjyP86KWUgae_Kuk)Q6xS^W9pyI6nlj_iwfSmjS>|lJ$
z@j0p%L(pZ@b3=5hiz8t7(y*hVu;!=G&9Ygv*YVL27WH<e6}tGmG0m@iz0N}xJF&${
zzrT^jYF?&{^B1*Xiq(X6NbxmXpvQ)}S3%sH9)=Tjx05@jV0vkVv%uPhcb9_a*<QEs
zfXgISB*5F(M1R+->tud~cSu6yr3q|lEhRLpg+U`1$}|1CHrR|dW9>`asVjIqjxC0Q
zj-P=RTk_*-l(e9nr9>^_p1J&2#<fjZuCb65VJX^E6DX%JxV6yofH=V_oj*X=ygPIC
zGHg`it@P^Y%#1wOfYa$LdiJ#2@s>N`DDvPKgQ{G&L+dDhAy@(Z;`ajxi(`+jU3l-j
zX`}BfAXH`S$H^Vle2{GhcDIWxZMr)xJUk)@J_Um<MhM%sc9Vpc^s+&oE}OHqwy(wJ
z*w0ekQD{Iym7K2`=Vm70^EFnlYg6%!FYJDsT>}D(pWjSJXZ=Im53swq^Ce&J@9sfG
zH2h@-Jy&rRZsDlOIod!zu=w>@9lmk=C#2Jc{XE`UUI&_JUmnMAkQOg(Z`BL0BDZvO
zDm?Ndt=9zxo~T#fyI0B<UK1<me9N6iuq<gL3Emd;h518ykN@sz^93-=`0SICW8K%d
zSvXQ+N?^g4F@Udf@40CEG*nI@Ba)0-^+`!jg}d!>JM$_$1${8jaFMPYm*PchhJJYr
z3U5g=JS_VZ&}tskcmbWyKfat1Gm#t@;yMG^UqOCFtS!(q4xpXTmX$bLCM5I$pdNTF
zu~&Z@MO)}D;BMj6Cun{y*6$=JFmHV>1mtDYQf<aGda%FGGBD8Bx9CK87%*8{Cd}ZN
z2~pd+O!aL-$cfjz{l{v3i!=Y5ud$a{(`UUF<gpWn0cwnGt(qlfzkg*K`q?}6HA#}(
z;D<B1l{3v)KLY01=C$qXllFj!vtLw7(EO0)z^O04zfL4ywDGO5HY4FZM${tL_fCLV
z*gjRf`lnMOdmMwwENR8_f_1b?jx8VjF*`$qQDm1JWP!a8$z$?sdReK;y)Ic|i!x-{
zklv^1geW@o+>uo&!|>2INEujmuaeBnpx`||cg1fRMClt+OrQm}n=xZ!N+zTCzq)M&
zb8|=oi_~ML#Col$=Z=X6TX@0yOLm^Gmm@b9V3kYF4~XblTe7*fW=TL~YRnjVCYOgx
zMBv<7q${06g~bTCgo^6P*BLb+Swj-9@mzd)gtqTf`Lp+lbTXOhy=S4Bcvu6u5xytm
zKx{WJXTXl+mc*_E_|QNzVjOsebUE7Mn1RjV(fcpvE%64a+iYP9Ww9G{DCzMWN6C-D
zpA3G%soO0&I{nR|8B1-8Y?U`0y=rF&fq0S^s7(S0#b27_#2XU&WnYYXzYv7H#*nRE
zZE&dec6J`6gf$!bP5A$m3gjIR%p{uwu)rI}IR)~VXaJK2)v`kyb-Aj@d3@uq36cPa
zd@Ewd8<B127>ZY+db(89F;yjzM`1o1*V0L1I(tlt7MQkmib`!F(94-F;}DOl<CXLF
z&Iz}jOh%AK)Y9JzB?`y2q*LB96Au*ImGo=B5W|5qq)xxDc#K^TX6t@fFIQPx;g!S^
zluf8{=7~*hTFaj@q|{6#8#v@lZb*%nIs^Wx<RTOGxmQ^UvP`3Dp?@tMK(KEz0AFMo
zXqi*1iA3(_bWAB@Hi3z1L(K`tL{;#%f^@v`tBoWX>EMiwfH~hkvxYS<Uv0E@cADm^
z#OSR4_5{+^r9FFKZUqCs-KT5Z_uNZebGzv81O`gkMrTmJvYNSLs&mtQrPp^pjJ7(l
zv1Qr40eQDlf^OVgGrv*U@xw+2)9mTj`NNvQ>XObD<}d&^46Pk7qr;$n)VVEsR{Jnt
zB$NVUSPoh1%v|X}mDbnMB4paUR$V`k5Q@pKtdvS&0fACa%LS!h%*i{7`Beo*4E{YL
zGN_fF@z^J$3=>rs-hEg-+tpdW|8TY>NDXgL)^!kC^XQ~Mt;sZI=VX73t=FPN<*D+g
zzW_bNmh@3Um$BhOChE2yeo;sqI0W(&IEK*&bfiZ^SkK&NsCAm`t2kRXX6n+qP65&#
z-LUl8Oq1%jj~${+%-B3Bds!z@fWkiQIW*QT9oYYg>zD#Lhy<lm4zq6e20xQN;U%R(
z7DLXtG5o?86>|uJZNs`}!@2?JSQdqp_40HnE3&)}c+EnN6F#d-_fk*V*VzEY%Do`>
zM88s{+0#^|=Z%<)yG=Q&i?_eVJZxPKaI;4)jo1X1z5>A3^^dc0c8H&^>*_<e!JOY-
zRRRI>ws{PvwCvif1L<?I>f&<86m--7-fshWMuEe-tiypvoh6nVaDf{7p<T`cVbS^~
zbA_1QWdV2U18iw#rm_p*t_{v@lJfW9L+gD{?BBn1dcsx?4@Xs7vFQGl)5-?<F&^nZ
zyq9vKbRcrn|Bj>spDj<#IQh2T*O!sh03=*KaQ;f9ei@;+Gx5-T?Wn=HyvVJ@U=1S%
zZ%zb&PkglG$EkaZe}I^H6^p^lmZ(@~>RZXHs#tNAip!giU?r$D`{)m{9g{2%&v+~X
zTMSy;^|`-JZD3nehH`5=QASFy9~}}8$hdG+(2J@WZj~!K98Df6!?KAQN6%U#tgJT8
zbEIRH70iZyvqx^tkP-}uS778#1|VF~=nQhJnzqL}-tzZ!>Wjm%u2|i$20TL@1&+wp
z1Nv;dIz9ol!VAe88+JEgtJM@t6%aBMo@Az&@eF$npYo66*+AsW^ge}Q3Q>$E-CPe3
z^4NOw+xKwjt~(9;4qS2M)6g{xm)A^SO(t}Un^NB;!BKba$%oU16W<0advx^`5six;
z*F46KRH>Mi4GiV1Jzc~;r>Fcx!!ZN>4O!naBws)BwK(37j#$#>dT8k2hE9EDwbw^s
z29e)&OvOmfbi_2z25CZ1itN4vRmAbVu9M(Cu<}TjWS~vd!0`s2gbdw*2m5ebJuU|9
z6%N!77WwVNT9KFjm8Us%+!0{ujsX1CCm><-7oxrmD2;=*=lJIA)?8)S7AXA1!Fqtq
zYTic+f`=bqXmg!S<9bXQCFp|o-@p8Bk8`F{g0Sr?zx_Bo`v%fgIgnI;q80ZH_a_n9
zAu=s%Wg}dv<mU+wa3G6GW%P_8TMHOWF;!MRt;x<Y0$z+cdK}ZfQb)CIl~`HL{8_L?
z8r6?T^NWWeeO|UBTGx+HHYbt~nMLaXLq>wPy^qO`07{UCv}P;EfO6ocS4k8<#N;4g
zY8-fp;ZhDDA}Q3<Yu7%~t(n?oAI1~`^^X3ivJ(J<okLR(-$K~Z=1=cImhaR11n+?l
zj({H`Q+9_NM5|A`%#H{YztONHkCbd(S^x>r9jRqkl}&DD^(hvVj$u4IXQ=r)l~>+o
zuz;$+yaEELm+GQU*PXk=mS<1Ekk{X*11cq4x+`O7lns}F*q)D+Dw~@p10cHF<U5lw
zP?6jKkymztfW*1Eks9Mi3<3(l`ZPq)BPf3PnCLud(@YjyaNk7yX)fNyJnHJ45rxT0
z`<puT@#AEpkv{{j0=b*hT)bMGAPUrdZ(?2H49??O$v9g_r0$c(=06F00?OQw<zW2j
z?7GzX06tr^hj*3mW4_=c?xX0kN`E=%^t-H6U}8Z>J@>FuekwmR@`RQ}&sIn_oSkY?
zT?`8U8-`|PPfO!!8r!CLJf3Mw^HF*`1umG1$J$18x2rA=Pb9~zTGzVBg!B8M?Z%sG
zQ|Lp-2=iMvkT8SC2kuUNKYSIGv0DvXB4Ga2RL@CuJ&P>z=FFp{LvyD-<y8nWOhIad
zOi&C*W%gX|dM(PT?AkYg7XFf%r5G*-6vQ8;J@;!nrbxuA*YnqGlR9T_AE?cLmtz@%
zt&v=C@KDA*d`D+ItyVJm7(h(rX8w<WLyr~TWZ!@Pvf&TSR*`H~%j8yO1TBPsJ;ix&
z7Oh^2!AgteU(ivU#ptM(enCNxaB16-V+WB;G5F7q?Ev0_PES8xY<&Gl*CRTtgocsZ
z#)|r!A@2=GSU6JF0*e0<SKD%zVLN^5=~juGTsC54grzJq^VG_Dus&2<2x$68_5%St
z`Z*5}9iz;Yl}WyIW(71i2NrhX__t1z4E>OMN(c3jb#Y4PGfIQR_t{+XmdF>vF}2(g
zj^4^+b5;n98UL+dsw{~0-Z_4f#=a)Sy0m|f#a887kVS??C2pOv>6`bLeTs_YIC!Wt
z;uC|r15Xc$(Jn#rLYAG%lYFLB1k&r4fz;V_w>Z5>oSi0lx@xqpxvG1?u{kXv(WgZ!
z`!@i~WA{?x9&?pdk)OrOR$C{_v4y|6{R+3U6Gbb$mpMhAzmQ%(c6>@7j+*Y;cnpAf
zPlq!koZLYDSRGZ605}!8C=rf=alQW{_R;V#vSZ5DF(zMY>bDzupnCC2fdxbCX--$g
z7nY~5N{Jf_Az6O05ax~1pPuE|y6uxTsWRU2j08h@fy-S%Vcja7t26xL=2?o#_|m!h
z+qO5c#2%VzickQNEHB5|=*W7QOM7`fWaJPHEO&MLHQ=b~w-3Ta{H`?6e1mK6Gnag?
zr?Z3Q7#yWn(jHk1*_sz)X2c(t;Nq)$h&%ayTkDxBA1uQUK6n|r2&}X<+_$pI5CDfd
z`dF>^Bmu03&Ctt=ion5s`LBu*5uCqFRfdhqG=WwH7nf`)+1_>5YZv%G`U0s9@Ud~J
zMH+ak%^lWyHE><ug;Fbl;ut^*Ept~Iz<CX7@`pCUQS$st9c%vi%EQV*X@*TNPe#%Z
zE!^OkQDrRO`0{`~kRJp&#Ey!?Q4QoOu?v+F@-fX=bo68c{M9Gw!Jm~~oWjaYn~i`y
z$K2+7JM`0;_mezVpT3;|DFNR;xipo2xSeLh0!Y;)TlY_@`&ZWJst^~o)_CqHh=RgR
z+jZm)-V(DA09LcT+uOJHn0jFPhU09Np^h*TdjB3k94^eg$$q+`GL<No^;P0Or)ggl
zJ`WoFs&w`^IX%^+1&FR~eVs{uty}=r5xYLX;Aru2|8N8#DQdiu>FFsHTZD`<g(8So
z$axRwgljn-FeuaD{}P;WfF%NDpe~d%7rs~@5=+NFe$x9v^~YU$pTM7#Uf~uBT7z&g
zAeVN;D<aSzD1RLR`)XlX2Sm9q`ai_xn{Z|j{7YD~9@Ly0e9?BkP0DoLnHL}cYQv5`
zhaEk7wK4j&%PgV87<+%GXd;;|qPrKm_!I126E-6RRvwyLf!rQHDVWXxZV3v8{|_fC
zf~b@T4eusE7qg~Deh`GT@X3aUwyUSFtWN3xa-ua=qvV)5{Cry)^uAS}`uV8PV;MTC
zqP~HC4j{c=tdw8@u<i{I93WqAL6&o&<}#HMZccr|R{toW>WO5NS5|@Z`1(kstYslP
zq2NbnvjSw9Bq=(}3WA$C_0>)!&xRv^>R-9_5lGDh$Cf!I%x&|gNMyse$dPrz)Q>=a
zfHfWJQeCv1y1&Cjzwb=%^Z4f`&R@N5g@l=h0)+gJsJUgT(Jdm_ol|1037}v6R2PNP
zyLXvIOIjX7%~?U@wT>yIrUGrq+%9<g7Hr~U4s?;1-sf{z8Z)R`u5vFrmy4zy1t^&M
zpjMQ|ZLXoGMDv~r+ES36m5M4L66r=>Q_#(PeqszAcv6D9x2WOtWE0Qe2n$2HVTd|6
zV!Pm<0QY;P2Ao=yE2EF~is+VxS_90ECNx8<kD9>7cOr8yKDhYl={b5?11O#W1lN7F
zL6<nd9n(RI?DjI7Q^NTG1Q{YMmJr?@sk(SBj_LMs1XA_|)D%`<9?{K^xLq0DeR#uB
zJh_|aB2aYAj*FU#g3HxGVGe(42TI9~X+XBCNS2skf1?59&e_K$#{j5q)hRg^59p<R
zdwV<(54-hA{#5#LiJ}fj)U^GiH=N=E+;0yQwm5LwV6GaJ6_ioaO>1A$KhUrE6x#B;
zqf2r82gx~(6CjaZuBR0&rPA?^<EL)9VplLcvbn&A5;^?zcRF8LMGs}+XFz{2Tb(~n
z++J<CX+7iwlFNfXvj?@Q(`CX`$55z}AUG(&(C^3Byx=Da`O0+8446i=L1Z`aY!?}~
zOVAR<aut4(Nd(*`NCrUBE^)toR`*}fOPDRl03k%x8hNt6<raeh+ibB5%({|i{3~1(
zx;jg)$3{XIJM+`Lhy7apw^|j&7q_C@uFreJGAdns57L2^Xf3Eri34h^!T5`j-5BU%
zFF@g)+IYSWxe!1_Rg^1^!4Kal+bmTk&J+XHY$r1k%6hrf+Alg#POl;%9BIS{)){M1
zGqu5Qf!*oS$u>=4{#Hs-xc~M7K%pnxm7eQ(ml)*aEwJM`=U%<9LEw8V-R|@a`2#+G
z-NVH$m=ZcM`|5^F1_f(5XTI?RLGT{W8q^+>W>;2Y7Z0<tXXpZ`SsrjJH<&V3xoAfH
zeVV~hnI2ry4*BZs>9(79VawlKjxYk?FlPXncJZ}Lu|8ub_1^=?iP<2a_6U}OP?{!#
z{xS_j;G!pC6=$g`I?M#Yp+Fu1aW1PL_y^v+3{-p~uO)9-t~}JQ0;VSW915|sF&RJJ
zSrD2%;Q{#eaMY>9>q2gM%{6h=MHkUJsp5o>=h6#K9UG5r?JT>|0fB)|eOdr(efj#{
z4_+8#%FMjohJi#MYhb4k1c3qyslT$&ifC;N_h4{rRbeNETMc=ifav#qdSwDj!kVE$
zUNk?5&_$eRo21x56hU&<J8WS3#Z3&r-VXo40a53NDog5q2z^&zMx`9&wWhl9N#Xzl
zVR_`BvgQ1i09I{rSvnFq8E@!klXUI^)yuQ~51=!eM)@6`>=}lB&STtN0OdshyHkE+
z;vWBX$j&q<Yp~Pp>9UNr-;c}RuF%r&hX5_#kYs<gQ+{K5FI}g;Q?}~S3Qb{}S#EH?
zy?6m&&e}tGW0yZ4076XhcBLm@rKdBHu6?#3xI9OwFF;jU=EyI`c6X8kpw>n9_a!q%
zuFTSX2(#U={}r<5`ytH#QD}up)Ir#Sq+)BcGdury%WK?DjiWyt+md>9trWaHHj(U?
zR4U#1cLfxcTebhLTeDYA*g9jZveDPA6kAre^?br`O`4&f%c~6w-f{C!hD5gYGVPM%
z_%H=ysHbDOI2F~?&2tB(LrE_+Eb@>@H_A!txH2wMjlj_Q&oN^!<$TnyLN_{7o19Y~
z)2KS+9oKHqnCqBo06>Xd`_CX^-i*mF5-5KZ04n1hQvhdIrde_v3OFFPZf4Z?ShhGJ
zbn>`a)ejAj5*6`69>O*YvX!2;y;`|j>&p}S#p_903$l>c%A&4H22e2s4AQ9aORP&y
zD4@g5+}9;85qI4Vu%}g_Wz|LG)YS#W1?Eshl#xP{<!Cf>tO4J6ZGTheK*hY_W&}ZV
zda+k-buS6Jm}13|1%SA>9aBMAmTK@uKhd1cSIw#KZQu&9krWL?L>*FPsDpSmMSFXu
z44A3?m)J`x1idPBoeH%H$L2;^LO6;Xjxw-G0zzLNZEC#s`f>e4a;SJNBRvS-tnWMf
zOn2*;mPBFc>V)3z&6deX2SFE!X7fg;t?Hp`?!g29hpDfQi?VsYUPJ*wVrglYk_PDp
zSrDY88w5leq&t)cSem691q5m74wY^M=}>AxO1j}4e4g*`^UfcBRCe#V=boACI@dYp
z9ts+1?!;5=3OgAlWHsfj;_yj&y4xwuJ*tn;ZufMLN-z7cgbQ-sN7`)?RI|~I2dR@`
znqYuJKy7bw`B;bun3(Mo*_FlW4?Qg%rwY)9Q3F6fqfqlBW#xNvU2ld39O)d;z|wq1
z4^3h}k<&HX=IW2;s&H=_LPL6Q5k|yVZVvYYWX-PsxZ>EP=IN+#K5D^8akQ$%{Co>+
z?o#UIvliI?6Y8RVeu68HFU~}i-Le}Eiuz!A(ArN4y<er`Rg~`W1cZWWuP@%{E=|de
zBY)%?OouIcb;vmEMi1I;kz~mVV3O3^4t+LIgZx?XureRLUcd=vbB!L6Zwn|tK5WYr
z0;F67brifkQ=DI2vDEAod0wSL|E*Ja^Mjt4sL3|rFsPI_3>kH7(MMy1Fn#VuEW<z@
zM*tNp>EUX|h*cnB$Re`cP1U8ekc92a8m;hM?R@NTD$KmM`QsPG(sj$tH#9JT@+B>d
z6~{F&=YVbgD=oiYu<tQ!8dtQ?Gk)qWv0vlpkid`cU7yOme0mx`e&ZWPy<fgDY{Z|`
z!e=SY7lE=96Np9~9UOqDLIL#x;4T1q<8v=;8L;|f(=xIvyH{BuK(5+jqP|YQ@6X1k
zt1lNTbbPR5r|+L2aMymcE_>{z$+Ah30NK6X49cYU;E@QH%%$_5OrT$0aDW)1{*)fd
zq;EF>1%)*iUTbs5((>agFJQ(iAJB7*0f@z?9F#isC*_#GhpRyS`k~DvQ`bko?kA_d
zzb{A14QnoU7AP3GHC4xHC~VEm_a|=r&9*U?>NQQNB6=#ypz`n`5Ys-0hs?!V;*aj$
zOqUa<NPSp(*qPR`cf{y)$p$n^#AR^)YIb;!=?vEIHQ0GFkd~ZjAdq#y(F;HuqO3aV
zn~o;nWsSf4=3Xo~_m?_v*Gz6*U3|*j;lQN&eDiKKTH>5suhtG$V0kYv48Rno`B^tR
zYt&N5jAD|mFI9=83#zBLe){5ICv_KZyzYCI_V4a9r=CErCzw88S_)z77JC_Z9;ZTA
zD^B0sl+EbXJ_%hn+MW<2nIn48U1=(9Wjpxa`Eyqs=c0XUCJzp`B(^2&zUCH~7P_vr
z1%d6Etqh9Db>XfceEtk%;7KRI9YB~uZRtdoBOP*uvjeX0InOzDBW)lrzeZYsI&p3;
z987E3(h%;w>J~e7%lD;IDVAY9{cdi8XX}(nFJCPv2TVK^*-XkfI!*y%nwk4+CTz-T
z#?F}XB8-Nwo}luRf^()E_OKJMMjH)BHMF8PA?}5tU|}`PDxR0mzx}%k+z%34B<x21
zmJIEA4{hkx?2aQd4`EjaK4CQ`{d#L}GU=XdXTCHyA5tjs+sw<PI%F2;8r=RPof$Zu
zs#;Q-CgscHj9ZEs*7MbW1&4b53Y8pj==qE>%F<F#mK=d{M=bxM%et6kh`wr07s7zN
z&joN-xh__BGsJ!#6=45y7HDj`ks<T(>j02u<ApjuKklC**D<}hZ%_A}ZFfGsinr`7
zP>Yo#o+R;TgDAN3;UIcW=G)PIYJVG+Mm0A2Vpdq@>cXzu8uwQtheK9B{2~O|QUmpp
zrw{^l%^zkSAmzD44hh^Dn&f!k>L59qbs2<ba9npUGBRDY2bV0_5;c_Rd_3);+~U2{
zfa-k|N>%w%p$-?)zVgtWe)Qov<Y=&Crc`(_JJ)>k=Av=KT5zcK7HZ@lOL&{x4R{dR
z-uT}cCb|OzqNZ8EPo6~d_{pv?!}cdYHX$1r8W?{VZO?OFhM)p~WD@H_rg^bVG0m1C
z$`bVjR+;<C;{UtZW%}mDFmni`@uE9};}U%@{Nu~%%KBs9rw$FxfO~F#MFDOp_Yl?g
zstpk(EZ;prTaupTKMCR@X@d)A0{+Ix%NSzhWEkM>tQZB>vS9nd%ApSX=UQPkidTbn
zljIkCP_LI!7OI0Kp=*0vV#Q0!PyfH}%_Vo4qR}AhAoO}obF5t9gA^0DCAea)yok7x
z9rB1dpeSw4KHE_|Ke!eT!By0?8<>|};hm9ktwD(S-Jw9hL%+N1l;c(aMmKByNWhus
z!jV?_4OywGCtASy)|_F62VI8?^F@b~nm1FlD1Fw&->AGW-o07g!Si<&u6KT>3`?@`
zKc_)mUFQyXNRDNn=+;xbP+|Ki!2ydIy|<XCn?CuXVFloVr4Wzt`)ejf&c0&{yR^?s
zwx2MabnGqZ@nw7S?NMo~P?3GKT&hQ3Y6!qxn(~UfW&mRYO^PdIUcOo};M!)$qxcZD
z3ShF}>sCd|M2AftJKc9?<~#Y0_L6EbJCc*is*V|-T{4GMgvlsfZdYcC9@H!U+N)&<
zH{^J4YvDRTcyBI3(1+3}HY;rTC0KBYT@`aA!}xUnb(I|=&OCYoUwI7m0(rc$W5!^0
zD4amBYJq6Z7Y^I+%s7hvkv)kgSg$)^NC8CcupR|(a!F!j0TBM?s>$)#1iEln4+E|O
zy#6b@MZ#*hdd?OLKuqme*C>0Y1vG!z6<e4YyhBS#!jQ~uaxtxAX46~vJx?Oh@R0M>
zG+&lrgAg%t5fq@|o~;_PR84X<$$*~<$honI5)w$uy3vRGOKdo?^DhZ=i3=kB;!=c|
z2^k~DB8w~4g;y6m-vY=U3aBzSmz!-a7_EGXpU00KbMtF?xc@V_VP@ADp)Hy0Ltx>e
z8wRr{@gbg{78E?dexMC?5*Hu(VXTZ^IzF8CeV22Uozq`9(e5LOo%)Mj?QbSFvIc#)
z!Y85?JQlY8f)&TS=7hbNUWpZY{fe#qyv36yc}{VDw(}%5#_75jzl@__H3vtcbi#VJ
zVf)XQyl5?)*!3S?|B2i6^QUS$%-F7x)#sdG5Vu?wPR~;4KYQ9$ae%z9+z$%qF-I>q
zyn@NOUFLc2wZ<VF*%e12u+d&tfE|VOS-wz(uPng!sjl+ozWoTOud}@(0fpTjD}Gz6
zE<MEsmL7}A8d~bEv59#m*x&Vhg!h0$Kq1rW+5)!!3M#C>=zS4}@O=&k1jB389T}Oh
znzFAaS`|FWR^Iek=m$%uL1|g4KZNTL%Yn{I!h?c_F$?(6;rYaly#+Dz`9r>M189(u
zw#5-~U(t%%iHO3m8sLQhw1(I~R07DrAVv;^%m5DZWp~$Qeyf|y?!ctZZ%QG(wqnPr
zM!CFU<P0|pp>6p{H7(q18yGa$7~cB@M_V?yq=Wk8<e#}Uvk48V_$*$qLZRKD5K@R4
zIS+}Tq?zkHf4WI8PfA{E;HbP(R9sflGMOprv2$=jqGJd`p|kw<QOw`u$zE2$0BT+e
zcU73q!~Majl;s-A6MEN_hTVF>c)$hp=&gPbQB_mtHi43+O0Z7v12#{7=I@?P5PIHL
zWU}mEP(2}jaff|GbW+N->miqYOZsHnwOdq@<jp0{O8s*5a>t!^LjJ>Bzs3GF2DNwD
zP$k!IN%-Guyu5>Z7@Iy3v$g0f4oJ6(O4<TY@`I-IY4h;&?&6<6W2@!Gx*)H<EZOS<
zE*H>JG#5#c7l<B~3Z8y&hh^cRQz=0JWzXPj5US(JZRB5+$E`T59^lB<EplEgdM_`u
zVPev-*X`B?{Mz5+x_O4z=t_wqfiT_ukHmVZepQ7KQ-lNii5IkxSFaz#Xn~*5FD)TR
z%^4T&k>aBr^7A+dub8^;A6{$-(Z@QuoT-%Ilh4!AX4_-lQ=I?amDnyBvJTM$5LMb#
z3Oc9{M3VO<K-A%c?SgGOiII~X(Q_fs3r51TS;ai6amSHspD6*9ub{1W;BpCy3VJM^
zpLtD+u4~#uysd$$4DdaIX~_;ZkD<cS@T9OF>N3pH#*K}Q>C{w8VRZ7^`KEvhkC^`Q
z%lrp;x}<z}NA2r&4(etXow}XU&|Vh)E=5`ame;^%iv7Ql(xrf}(t_j#7@`AtAnYT=
zWZ*RUv?e0WbJJlpi%Sn)YC1G-l-2fHG;!{V6JK9LAd0GvQE(7UUGm8O-s@*_pnQ)V
zF@ds04z5CsTz1=e)P2N$Ze_Nen4tvtnuJ5)IjX-OZO_TUm8M-OF*T%iT2b?nabLdD
z>9zL|Ww4URKV7})`I77~dr{XW32FcSf9L*vmS>OHEgFijxhvdw<vx8g(G(I%N#*%@
zu66Ocv)}Ui_yiN8>)o^ZH|9eEW!LG-yFo&;Kb(*U()fx*Er=-Pes*NJM3hw;!`l_#
zu>P4{{Cv!>5<)#3Z>K&s#f{7g4f5Qfb$S44KLxV0>`EtWpB?IzfS}@p^@Di>QeqMY
zA5c?ZOO2+6sk#@kDvxe=>iL$!VK7at@cWZnS1WK60{0oqEsV|Ma>%21z;$*%5QBQX
zMQi~Z8DQVERe+~Px8&UJO(8Ek>%Y&V2MCBpz<sqTZ)nto3)}aDo*lj8Zg~dRj_ARL
z3d>I33t3ywjX@p^x4l2!ao?F^vCO_&YYE8q+3|Ajc@#Qsyts`I`RsZ$EzkF?@!v<V
z4CrWp2p*jjBj(s)0l+O4uj^xAH$d(61r$%}b_1%iD`2Wc00&fS(QhWaQ@^ag`=fLg
z7)&6atB)q+6Li73WPk6E+_HRvM%qEQVG)p{uLeIGv%?BD=48RoB1jJ|{(dA-!L5kq
zq!X6{kwN==I5Ce(DQ?!WN%`6bHFWQpZ$C#}U(KX()yeKRe95PHVsED3t<H_Ol`#`T
zUi$u3=fOpoexO&AjgnE)+;AN-o4cFl%z2I}JKshQza9t9kDM}KZolRtu}V9-=JL!1
zOo-?(V&X*W8vbY{_*3zB_(jE=b{{aLkMjvA<)Ln2s7Iu>(yf8HzG>&^JjACh{aQ)L
z@phk4>H*{%Y5UT%0&MT@KjMb)q=+6$BHNn*gPmYlWg7y_?(E?p&0S(jmw&RDEvNkn
zso$Ko;E6K*<b3L2n{mwVz;reh8s<rZx8wzIXO^MQ{O~H?!?{`pd^|wVnkFNtHWJ-7
zMU0c1k`90mmb41VN7&gVCJW~zMw%a)Ag>2};*Y&x`*)zz)RZo+X4|DQ4_J?nSfQy3
zH|7IxD|s!c^g(ew{68sxpv4&tA|)z0vM{Dv)&z9vFR*<|{m_aa&u#|c)AvEg4<073
z{tg|IxES=gxu~sc@f)skNJ@{#f^ZujQ@*wWNLtBmfC)OC@Aw<!ts!GU@s@n6OaNIO
ziJ^kk)7Vh2BTw1IS9*ftS3}@`pVh}847X`+Excf{b}VqhwBh+?<_o+a813fB0Ps5t
z99~2Gc=kUU5Do-L7ooEhh6~y^?ODRXmDzSdu@|C_4PaG$$ALtq0*mkh_h|*s=3M{X
z^V&fdLR-lVKabGsfb3suhfOizJ=U=OnvR)nBtnj!cWw@~<F@|cnX%swt$0#g2;^?A
zc-UNQBaTTGh>0>#$^$~ESKEZb94Cx+GUUDTdD{&$Op*{ONs6R*<s;S!-P`JzQFTd*
zb$0Qh@hbPB=5@#|Gj!@Cxm`)~7blXJ-z57$s{6hvVdzLd`S@<a2Qk3%o8k+46kRt@
zInovV3T^D^+e6fb)eOIJ5nS{=$m{;=?S6>_oRO5>zz5jAN-49~TmSEMHYUHQKLFR<
z!Tr_{3}6=dO#HyDUKr{XDDld#;+J>ZNVchoi<HJL2{LpYW!az1#eC=Q=8P~l+Xl1G
z0Bw+`d>1ZZLW(S4m~dDsxvl?SPKk%@lu4F5)-TK-S`rG0S<zKYD3`QYI&V*<%pNZ~
zPT5EPot8sjy|snyj{`M4q9+Wt{{kwkw5j4x2TKQh<g>%uYrUF^f;Yie<Q<LE8>1&z
z%NURTb|2UZ;4~{NneYqP4TKMxkUcG4E`V)o*uk#g0nD-EI=7qXIW7(BP$$$FeG#p0
z(Ft=85F`LhD_8K;oe=8;b<CJf?t#qA_;PAF_^RqU1BCs;H4#VPs(d}sZPol7-ddbZ
z@7Q|wTIJuweUv!pF@6O8J~H)r@yQt*R2U>yC=$Q3D<e(!BC7eGxj7y&5(pNUGZE`H
z(6iY`p=1}{<p{gc8()XB{nclLK=T6~egwLDFrtJM*&W`4F8eB&CXLd6N25#$=xj~j
zeq9$&IDt<^27eGBY``p>N|^*+AK4CEqhm+l(IrEcE|{G}M0o`vE-N_c&mlE{QqYbY
zTnCSX9=5Ygli@fU&}QlZ|9i`H%~L6>B5}QJ@Rjxq>iUMbY(^NWtILiUDF{3s<9u|W
zxOP65V-{cn(}(COgzevjHk3yjAGkr!zIDuO$5ROE)h(p;Xz3W~`J4fmm2_>?J^GB6
zzCWo`oEi#w-!sx(419L~Evfk%oWuB5g#&Cyr%9u80jQU<*>-_&#O<IVbilRe(=_@7
zz9OwIZkIj^Rn4NItYWP=W>x9R*(V0)PkQXMGQj}MP&kJj6ayJl#hXF6?Cp9auknoc
zRn@S+%d*>$<jtLbqdXC0WJtd89e`}iFreo|uy)N}RW`k4221cV8p!fc;Ky>Ooox5G
zp!QJmPWXEZq`*0i?s~nvTtoAf#8nblGNH(4h@J|VGhnZ9uWqOtGYqcB8YM{aJ+38|
zyrIAc0x&b!i$7)*?E;`)szC7($_6%?CzYY89r`D=T^hiJ!v!mAIs(iUH|`id{FRH@
z_KAFy`KlJ|Nj%2-TDoJ_;pyMb*&S}p2Cl9|g0&MaIBP2Ii}qj^^APo>2LZC1Jagv(
z3KY6jAg*x1vWGDQDkhF#Gfc2z_=+pkYc{W&1oF~3@-hk#Ns4ZF<RjZo8WlwN6S56j
zeb0~?0C@r83w8{{ZA?c$6C|{4GjV?mFb^;%hal=h@+=x9qTq_(1Z$A%QwlK%v2vfD
zn;tbxl7Q^FN6wiioO>-v5gjDUWLJkw$b(iNc}G(W$hb_M)p!z6`VsXpn{ffN*B4Q*
zs-L>Px!A7j9o`<MwExNNlCsq%sdu}n0vR_vw`C8*iZF6C0i_4bgz<=yi6dJuLPZlK
z&<Q3+W`Pk*4(@LRSNY>lQ)F@@R8c#67r5qP$dET2FjZ2dd9~9H;2?klu()G}?J2Dc
zI{t`kIgKf~KA_584U0YX0-&K`IH0_K30rt2HaLp+8LeL=y8$W6;Tt~zb<5@P_n-48
zKl48<O12}>A&nF^gS%UE8MBMlB4pnK28RY#+<py}@iM|3<V8G~&!@Hq9uWP4a5P}#
z5J|er<tB1@FPu4dNV`(p?wrpI^EsGq3g3bhlk0FN(|^bnO5DBGZM~8amriSoZFUW%
zi(nJk%EY^*A}*UI6dOdcfdWjaBC^pTdHt1)(_8gqM9}5C0&^Avj5Bgk7bK;C(3Tql
zV84E7*<3x5WPEczxq5wbdB7FB9Bn(k1EAj>P%u>kO|wpR<?39@^(%0uWxCD9w2iqL
z1E<=-FVlZWbbr@o>=!tn?tgQM*`G--PyCU6M)du5`U5}i<ihrCs9)RFeyPYRIU=TY
zB$s67$T@8`hF{CVYi)yu+)JOGtHnokbY;L`H+cb#at$cv34hkANZX$x7yV>6v!b9v
z=B&U8ZAU~kynxfpznlcTzeK6r>G9;Rqxt43)hs~V6@HOODlv$kJr=Gmc1-vw9<VdF
zW>k4`7xJL{b9bzF{`KVt%I>-$Y+xY(g0$iyyOX2ah&&&lnx7%$XNummZ+nas+VZ&I
z!tH&q_^E57A8p>=e`I0X$gg1<@P5PBj+s6Y%(@JOn6b|fzTZc2KxXQUBiIiqkBG7!
z8+_Gra$#5dR{rEDX^MOja_9<ma%3+XoE|Di^aN3CO=@xwP%g_g0-KAWU$+dX^X<`*
zl&nKXcl%ua55uMBMmN?`y=ni7L0cC!&M$efA+dOMc5<FA+OvJ|7Qa4V_W}*FWjj?<
z8}nVzzIE<R3)jZfXASXSoFkAHj@@q>&(_kzYVZGWoe~KFDGvhSjwZAX0J`-P>Z6sf
zAf1KxG!fax+r0s-$14tpj53dPLlLfG(U{RZkY$WF-(^s;Qv6u>5C!J~keaQ+J=^q*
z&{p-q!>9@#B*KmmnOxr(-7zC8Os3X1$fzlxITu88*lvn>AM#lqB)i9cjK8{ZCl#4F
z#q!OB5BuHP7IJ(}=k{0o3y=OLdUkyvP!^jm=5fh2^6XA_b@|G{TMurgoACeS7oFh$
zbWB?~<eN_*wZb8N#r3w*=mFFa5lD1}hDTNJbc8>p{e?9=9Ci(d63AZh!0!^@eaPBp
zHQmI{68!yMW<iU>MSzM>;Ufln{W#Z^3$H9OJ=B8F&L9WFi8#uylU(@B-v>r$FSGJ<
za=63ID`M4NUHGQ&+i>&;(iolM4AiXcC+AiL(~=j)7h%YaZ9@gaWdgLX<6#vQv<;U)
z284{O@8A)pqAIm(<Z~NZV&X5*wca~(t}&&rF|$QJtTr|l@(rAkyInPq7s-QFPEk+q
zzT1fxRHT^JrbmCC2(>SNEDv5zMbH{{YYwSSep#OTv9$LZb0#~EWwo*KiF_WnKoBB}
zz`_Y(CFN^4=cQv`T!!s9ej1l0Is5Al|DMOr5F!oO^S4Usz<YZ;zjo-<^5vbL!*#Tr
zgk3J|$PVvA%Gwjh1ppHa+NXPb0*x;>Yr>5^D1s|g;C#{(Zg`2CeyNCsy19_AQ?4-V
zShwuO;*ZSgq7mIJc-4X?C)PR&O*R_jJ_a3PD|AzDSy9gnYf2lH3Jkq_U))8&4&|74
zM?bnD|JQ$jAAX<pHyPrtX00LIL=L(c=qQKtUA*h%+q3X*o266pR1xLlNT9=X_7|hp
z0ux&7*gwJ#KYfY;gVk9T>$j(N%UB(c<-<(si$CE^RnOsO`o7r0noW~fc&{{_T2nZn
zDefq5#;e##A?$gR$FBSH0Nc1PFY^@QnUYVWauY^{qpmunioQrqPv5BY*o@Gn5buU_
ziQ4q>@mY_wIZbYDQS}WDmKKLTD3ZQxfk+Lyc*I)OkbiuHF3Z(#*p}!eTT6CBboI^!
zqrC;*CuCb|{1>4>KG|vxp(c|!`yr#mXPQdWx{7ev7abGP2wzT1DwfLQVM9UpNwuW=
zeb`jGUA~<{?Xp+aFZb$Wn)PPo;UpMgVrs2Lg>PW<4K66_<E`Po>=?=uq2SO^pL0`m
z$S0yXYKAl!c>GGhu_J3_So?5zwubZnya1xCsAOE(7vy6-qtuQ)$=EAobdWdn<HY|?
z>l=>XiV0^7wN#~k0j1STDc`40=f*qW>C&*ogKya8p@>g4RIvt>LGV`67)2^8$04C*
z12Cy-!r9E5FN`+4bL?6;Uz2aRK7R%1dZSUlIuWO(yUHnlp4JFjKjd~EpKCaHEF#jl
zL&Z>UCX)Fs2<rt-J->XOb5<@l(;dm|zU2JyFZZy`al)+9Q{L0H5x70dNk%-L$>CnP
zDa<fGe~PP4j+1<k_nxD<b>TCbzq=kR1ZGOE@=-)8p9caG8&tgb9Bz;JR4d0Aj_6U9
zm=GRW@Xajh3HG}kA&x^iR)O<%>Q~*n&{U#yV_A5Gds;fGqV=P*#F(<t-HR+(j?h6<
zaB#4Y=8@)OoR7HQs44$pn)&zXj9c#AzrUbCL}U>5Gq~^UKH1i9DPbva-sBzlR|PLe
zm)D9Mv5ruNeV6y5Yu5L?606b4A`l3^!7V3E8yi;TEV07oqjX1=51&32WhR-WLm)r>
z9q<W9JEhbzzD{ZD#eXutr+h*<YRif$o|_W?>_`gbh}$h2+dPVTQ}Ua+bRzxh)t|k;
zw<2|w&&^a8F-Pa|<cD^koT*Ztgn=qteC1#ame|^NrlTSzv{t+!T2!CWFccHOxNY-X
z1ezLP7>@Q$A@T5=VZ`H>q@9><0_*lNhNE;Wf=6yt5UKeW^wO|K^2hoLa5j>zGR7Q#
z__(&x^H@h(hMZ1TRJs$*4;oe!3eNGg@$}==`=kB+Lz#b{3&d^K4PLE1wA7G}up3;`
zV?c+m*)hZ3^XizgkA&%|4irv}vvQI#>>gx}<6F);BC&UEFem1v7Iw=&UVt6(Vs~6O
zwg;>?RoFRE0qO1D2CaUYpg|b=O3$IiI=I3hY+d^yT!hZVff-$H_i-eLhOnyJ)8x*;
zfqeLQNqE0o{E2182&T@I0{o=+=$J7FpKUtCOYX;ze&cQZafFxu?(anbW+<RwK$lVk
zF(s*+PU4X?+a2kptIhMntQ_YB={csZu49HAc0kngbt-h*e|R7^Pm~=yl3UpYrrsFa
ze3)QVx$?$jyldFe%x){06s_a~w)xdojTX6&@&t!|sdMX~UkfOB|JY&uXIujv+B`bB
zcz<-Qs;orJYa)W^7Chc;8@TiaQP>@JZj9mP;Pl-oT?x(tCw{`uv!@zdfg<j*Gp<r2
z-xLA=_7G-!cKAqP^)hJY^sKYH(}G|$cfSxHdR4+L(9xLrRkyjsz(@&>-`G|FhL}~{
zHTHk6?v4|*%$1Y|D$z!Ccb6{k4oBaIF~1NN1VN$Nj&j8=?IdjdM`2jlL+5Qlw#b?|
zmRXU$NqFG87*zacxJ&^zsT@38zon$klPaIPcFJ2g@c^kktiuvH$du#ce4Oannue(D
zLBXtCqwHD@SH;FiZYL8``$+!o3rzATjJ_p3JP*Ev;5h9FNx<KL7WmwdK=Q-Z_#Fzm
z-La+Rr6GeBOFY&@;Y~<!26Mljz+&GyJSyxH($^<eRwJsJtR^<3IHRC`U54~a5tO94
zynMw^Q8g^P7Jitwbwd#0{o?gKB6IUa=jX$qK5jRy4}XBCL5@R4AN<M(L<knVtaJoK
zt?i>Z^}>`#5K&=;$#TE(@Z*8c(?FI6tzCwiE@yL9i8`vUfn}ZEa`Z4+`n|FMo-&cw
z?x~do=BErtZSqYe_%L+T98PR)&cZXc#>+D+zbaP1>1$_}pMXWo66q%ck0BVvg%eA|
z8R3Wc&{Bq+={+Z>3EtY<9y$pa-W3th2P_S2<KW;p{jn*$(C+p=fAqJ1$eRwXC-TI~
za9Wnguo<aV*vm&|vQOj>bEA%2ktfeG{Svf>I=Au#JCwl&PqyVMUql3sKT159-#LAy
zHKeLLpzk~(a>I}#nc1KME?EBri%?o?0G1<Lc_q?fyomK|xP=rqRnGLBJAGEn;0@`Q
zf4dDVI?F&9C)EUJ#R>94=|mLZiX4-lKx!O2irsbeSluJ#jX+rON@##<z9G3A)@};6
zJwkr2BiSryqF>uc^$jDG4*R5vI-ST3A@Cr3^Us8ga+~eHLlR5w^UD7^<p-~ZPyDP%
zcm~ha9KzKuk2uzU$S@N)b{n#PKe-zQ`F+21V%7n2dqe<eb{fJaAO)y5M#WpNbRguB
z-?!UXQOx%2d`d&y>S60TRNCNEpZW<H{);$}n0#mda0HX6^8?KxOyLV%mPmf49C7E6
z?-#5n4Juu^!;^b3)<-#btSG&qom-}RJfMf{f5_oQ6%B#jm$1jNh&7?9_=d_P^%TR~
zCGoz`#*PX3H3jB;K{&b%A6ND-9F2H~WJT>7R5$5F=Lv)T4W;FgnvYjY`Sl`OHj@D<
z0LW_4+P^(WXll}0y^N%9Ds#>%;;s!wctb}A;fFjb6yWEdiYOW=in-nCp?<P(wSC5@
zX0EDHUm<7|Lk?AN1y<q#p`benB;)6fiLsLvBW1|2tP$hi3x6UZ?qVVs<#Q^1x;pn`
zQdOka7G_XfKPLeDtqi9&8s~xp>TeIER;5)}^YP$}VP+Fc1&4<{=3>e>D0HK*!we(t
z!XO%TZm@?^_0i`9!N<Q1!FW6QU&F8;<veoM``ZWjLuagh11q%uqDsqfXpSa#8`8ea
zz{prfm!k{&9j;2-{^(3Wp|==e<)oxkXTElCj)DPXVy2wnrAJT_!qEu$;m5uPY16Gr
zEK;KkiL0isQ+j8?mfY{)5VufyTISc7TJXFa7uWoMd1)!}kA1d`ty**9kr|kfn1qCc
zxnRWg=5}e+xVKMeJ<lin7dnZFiAhemqvY0n6rs$FjFM2IF?;48?MgX&&T_FSPyMp{
z&ag3)FCq{-aWxT;-;BAVQhH+jYh}XvV+FO)qu^KxATojWUcCwNDdy8GRd(zjcV9z8
zH4B^$W(Gvycj|qLeq$;#$RA=bGC>=@G!Hk4#`V{OX_+SnM}y;5e6@WkkU&js?R*v<
z^2Eza-<!k3Dvsyjim$=#v=M{biEj#m<M(oLEowiKV%SQO&T;45qLhjvM9cA4>|T=j
z+dbruSXJ03E~r~-T8a>t1n15|Uv&F;N5|QdPW4o<=GrwPNqG?z0uPqoqM<dMj-yd6
zZ*JQkT1c1)sGT+(Rw0XQ$YtOFCCK()h=aKiGkUSPaxF#jI5Lc?*VOsGqj}-rgSS8L
zMuaHka{EnLun}&|{@d~U1!&>fa@aa6jG~acHV%=cCM@~7B^TSA<JHyLmnTxL;fS?m
zz!9m&CM1;5=R_MD>$gDyZuorHAzEZVceAroKSa?zaU~AW0CkB)|IE{5ldCdagENQu
zfd;JrK#4o8wXm?WA096MdIFRBhtK}sZnLB(Z2V&p6HsJ5MC2a?9$O>_vTz;FE&pI{
zOB&v?`cmJCjVb5mQ%13n76`znL`X909s@Pv%UR49xC;yVT9>$~*GY)Wd1?j`I7kNT
z7|+OFkP`)*3ifls70R}0G|iGyR$%ER02U$4G}Zi_BuFT*?Km}cHt|3b0Bu}jk!aN2
zV>_+y>~(O(FZf|i35xxl6#eR8obc@P=m!M$U+-f<XgYkw4d~0s(4|PM`HU?{7nz$u
zm*=Sn2ZkK-;EHQj)F!p=Yi#Jfe#!If`H^RzWmYe#pke7)3x8`dO|8S;9X-CApCKYY
zEdL$3DuA)i;e|VMZixrkBwZpzqc>q`l<>T|QKd++jAwAi;*9LB-rInu^L1W2XE<z~
zABIhM=1$V(0KiP`j~Au`&&Ev|ax5(Z*CVNv|E&}NIO6JUANB;803aZk4gCp-&3FW(
z0(r@|AFy<?jRooGndz3W-V3#r?^gz<TbQu(A34)Mu+6!kqj#r6Ud38v^m*Nv@cM)g
zlELdD(O0UEaXL}|N(3(yd-&d*5UN3;SJH4C&7laRY0-yR=TOw`km}=_ED6K7gmH+w
zC<sQIoG;cc8`r4fRCW2ty!IZz53BpmSTpXwFabHXp7y!Oe@~wwxZ*K#1MdC{C-I=X
zq$!u%9TVzj>x8@d2<9zqY8Tt*MIj;ehe0mrt;$E^tYDrWzxg*&uc1(6-WUrw&j4Zr
z1wm`w?7bDJA{XHa^6)EGD^@vFbCL6PGG)_k{<v*frdUuTX3VndZtob9uCK3t`)n4^
zY$L4hLyx|A-N|BTdmq>pO|7v(qnahwySC|q8p>T?x4I<_YviNu{=Gg96!!(WEepDA
zCOnB_x<CegD4BTRC-lFN*e>gkR6Jkq?EkfkDeF-V1STR2=jxqgB2=@4{VwEoFX$6i
z{|d%jukh;Cs{$82y~aNyBXMH|s=HsD+MM}G8J>VyeUy`s23W3tn`ef$GwG<dsHlHi
zbM{B)Rs`n|=XoYVrDt09<h=s{K;W7-eBYNC72o-vS-Aa7Dh;@>%;?v|#>VM5+OL$l
zfyfjKlzNrZh#%|*{Cu0?FyFX8(;A;H=y+_@{7M1gDdyaP$Y}MuApPN6-kwUzu1nBt
zj#6pSUF`^q|1LxM4|!(AIR*azU@2rgmb(Giu?5Sy-ASNE9Gs*okx|M5ZDr2>+NMh5
z$~1Ctp3|T0@uwGOsp?&NY@D1QA|qW^%R_>L<6~lWD&2FKpi=fc$H-^t9G5$@+?4a9
zud;?mEgYtEE^Pag+VL&@sYX9)bYIpj`ZK%zY=3*N<_KNsxytu|%U)*Xj^(w%+z2Q`
zQ9#nwsrjFM5;nMifVc15a_W%;peokbu4B}Mxk~TN7iti9by($7hOG<08j}zmWV0_}
z&zWI7OB?8k2isZx8UH@p_GSJA%=aT*Z`W=2P7yzRLRb9$DWH}5!>;e>*u)<^Q*9kp
z_2oIJ@~y24>|1pb+tXAmjYx6J{uzDw{#jR)tqmT57cxY@&PH#l#-Q%q=Md8(PY!W$
zqv;CX>0S9gwU!izMwemZn@4r?6UTF@RiaOyzH@4!_)YRR!|L~hHn5^n1r{cImSQ}@
z^#ZWX10uXVSW%R}e;_?LpmgV24v(S`@T&r|RNgZ*Sfh@qU31@=<<Zf9O9LV#m1w>E
z&YrJDWg$YKP<uBs`Q!oSL&z2sEvLe~+i-Lb%3HT}Z|%LeocQ60mZJaZ7e4u)L!vu9
z-lBWkV;ly?r8D}|h2xsLg&G9{*XNt%-@>(i6jzrh1${HFdJ)=l);!n16I`)Sc^yms
z)}<-)b#^}lc^Wj<m&$ER(k0heRqmF7c+ExHt+FbnzUup_?`$4Eo`lFs?sm)ns!ORV
z`Qk;>E8^E94et)I9p!S-vTs2R1R|uEW;o0*FyT>LqEjs{!8vpuDj(Zi$Bz!eoK2ys
z#2Ldl6&=M?`Qlk;1Zn1HGCIOEIoXHT7aIkI3bXOYH4a|<*Mp(nz4N?bUA<ho&p`R*
zZ-15lH3jRX7(2Sc+T8+~mG#UsY;$1{kAI_{{SsqEwLwQ6C)hNF$-4bRzNYk~UJ^jX
zpToH(e;EaPuL1m2=o$(@sp2pHs6V5BgJ>Wx;%(6w!s_$aDN#@^Oh(u;f*O7|v2eDN
zZUEoT&hmY8D>70Qke~Ol9^`!Rz^giQ@HeKpI-YGy7RG!)E`N=NmRv)$`Q{(_4@o9(
ztQ)9;DiW}9Zv2$XPv{*>w}sJj782x{b+hefsugbf*vbV+nm!#IbNV@+yKd^`xYt;Z
zHNKjR&ctkP<`nm*6tw$eQ0wc}L)&(<<FVACko`61Y3Y}j<DNb~$szm?v}SAV`!R7T
z5>DFDb1T#NtbM0@P$X<R7<2q)|FGyR^DcL1d;JO~d$Kl|N!jQ<_&ZPr6aOOoC^0F?
z#@Ej;N~g|hh*>4g)qN`S8>epBa_4;_*gy@U&%|eA)yk>Lce~eV`Mc!#I{lbL#KG#e
z^)BB+gG;hwACg|eayUb4AA7$=>&N?1o0^qejY#D*K9&t<<>VX#ho+;i=qK~Z=9`y7
zuIr!6miaS9!@6Z0+r(|xx_LF)GDMjVQP}}_d?zMU5eqHetBWqf;=<-~F+-%A3W|yc
z*ZcikGLu)eoG%v_uFl_my))(yT$m3<T-RT?-dvqnNS^<p*IVdYEhMF<rx(ra@B1Dc
z8j>`XrggYC5F^vsici7)=kjo4OanEXO?y6j?uN#~^6Vk1(`^6l*~U@GM5)0(n7JnI
zoeHOiwdz<*%)2jv1gyNi(Bd<-x7_Xfq^gpnqWOyzU1m<g+2+*)9a%xP)?eq8I6JKu
zJ6l;T-ZLd=7H4OdpI`O<k~94ka?gYdYrfh#{-!sc+0WCdVxqsln7w^?k=u6#k+Cth
zD7oG5+s8gRTBS2AAnLw-T<1IJJ2!Y3`H(9^vTaP8mUVfj?NVajKk?`>RWQ#U8e`A9
zSjX<}PCpitfOLdPpIBQt6sh<Qj%|)~#)z+XW6r_YFO6qKX{C)7a!=@BG&vBUO68?l
zBVLvfm`H(YuilKdsQpq>rIe!@Ba1#EIPH!k36wXI>OS4SiB}UaQa;$`OZuHg@Wbx_
zVkWOFt+_tUF3jMnF#HHMisyG3jIbo(<v_N`)F9<<t)XW>W$u2WeM<Sy2+pviql<}3
zhLDp&?D~gLn#P=J=K^(&^2arq%y~76@`#wnJQPaO@1XQ@uiUmqKGd`J8(^|CDQC@{
zW_m0}@kz!wP;`dIkY`a7QklsMr+u9$F(ztu_=ZvI@Gaygi$*@&Zmxb$3UGqk4hOEk
ztv}jMmThybta{biEi}uQ6iWwTgv2we{`NPnDSG=U;_%PZvz*Y31`EN1GeN|g{#XOf
zyLa!B$H&E8YTumOB0TN?yq6Z{;o(U(yx9Ix>~VHL7^FsY)0LjzDi{&@)9?Ja*LI>r
zFDyIT#M;<6%Y8U&zBRycn!<BsisveaHYo(4eRKS1ibRdkz23pWAIDqMx!co`53;p6
z^ogUXgrBfIe%y09rrIy#KU>eEJLkSVvwE@5R-WgvIa7Ld)@<0>oV}DnE#}$oag=5v
zBXD!&@xFe0zAAJu<1B1D#~4p2-Cu5FwCb4;lIgQcH{KckT!6`IDXG}KKge%(N56kO
zGKh~u@37y9>y$RAwG+2%>2lkg!0<in9q1Bq+^vkJo<W^#_}CkJe&Y_RQ(;k56wSS?
zIg1HgF=|q-1BrF+bkplVW2a?z0f*#hD>+-|p8s`ca#ZIsjkuZfvlyDp4+ncJvvrQy
zyNxxQ?<BK%3)H;58&BdoBe=q7#g3=UQR*71*(S}NYC+#vo*lXJta}LLNeAtdN5sUG
z&D2x^0vS8`<e|Vq1o8`x)4N^%wolHSzX?kds$yrFax|V09Q{r#V+Q0IFW{%>VHA#g
z!RjOZmt8qbBURYus>K<@ZQr-GLci)_!WxxnP#^h6F(HcjrGl?9{;pLi<;HA8e!m^1
zL?tckDE8B`=oh)RtZL8Z8HFo>f~HvnFL?Ll?0J;jOjd%S1d+Ma&9#?KB{D$eD!2J$
z#|OHGswi)cqo5^A-ud_=sSa9Y;n22xoaA%7ebUdx`aRnF&zH3(_bnl->8|hJX_NUi
zs~0S2RkHofE>HLLl6F6WX5_J5W3Q4a{B6x&n0cj_IlE7wP0(sY`Q8%KYw#9c7{;RS
zuhfF<-pYnYYC&=`eC~(qhnCrX0^f{3Z~Z|M&l`End(~h)FVU}f@~M(;G(<W5U5mU-
zm_|XCul3h|lnypk#r8b4L!+rh6UiT>c^#~UhCwtPjI+$<qm|yKx}kQ+y)7w!Sd9#1
z3QAsHok%K;Z6rKNr!4tLXs1<#cG&}mjOHPauV<OXrnt{>)`sN!00U?&k#yFN;>lBd
z8ef-Ii*!;Q6C*LdnQ9{Q?&!h$SXiGQ2Br!+8uyG1ixF|$QR7nateBqnh?z6rJs^>|
zFjHBMZKreVjq7BYGzqw3+`ySEH9*}u1*-Kb%q-TC^eon;`Zer4Z6{N?F{#m4AaME(
zO6Tk?XAcqZ&K0@Xl{C~aAv+(C8qnG~0zR>1fft44AXR2+Wc8#e%>a59o}Gl2^W@B8
z4_DZ2UhGKu(wwi3$%Y+JTSV)&dI<-Lx5U-Xe`Sr7eU7tnj`Q^cD1QR)Ki*;8tf=9l
zjM*ZMif}|%_}d#+RA@#u+uE1v4IywCO2%%sMnS^Uq5kp(kKNJb@ZUK0bM?K8@)r3D
zcTN_49J?qTn~@IGR0sruyF~9`6G;~kFlHAM6pDYrVO+`gIp;_vgFM}an(j&YX3NQ<
zBqPmD{bj*VHgPEl2?>iT$qV+HoN3KcgBtd;GgP*QFY!lLALqCQ<WaQG)@aql+iXxf
z9F>`VQz|D(B(}Xd-Q0cO-7W5}Cu(|&YrloeTYI;3y!<2L`6sYUU-a?dx$A_}CImTB
zsL{PNHEGP2{jvJhMqGb%q3+JKW%ghtz(HM+591GJip$d`=+%F67_>Hj9oA8#&~J3@
zwLpW#T`YjFzEfw-?NlTcO|ty1Fee`WSas-=wb^pZ^}Yn}Ez-Kn)8~WR5Q2qW3Joxa
zR~ynKFKf-;2L6i`nZ0iQr>njnJ7>!Q*Bdzcs)5$zab&5EaNA0ZvFR2Y^7Uud*-UcE
z3vIwb>yXqSH!fQr<#56o-pLGWSy285N{>?e8sxvlYff_(WpKHEF%^0aH--PhhXZ-R
z0m`-f1g9i*4Gru46@}IE=Cd{8@bK`;2JZ@AKX=ZnImXE^>3m`JB@!>RwieqKb@)z>
z_gT{gw&(eLA_N}b@<#`pr&c+(WLk;*!69YGYC&Z+apCA=Nu!~KB%645YGZK5o63Z)
zh3h}RI00wcyj*@Tm?ogn<h~u(W{YxY>(gfbUfJ2%on3RZ5NqJj@aJl&({#<fyt{MS
zj(N^|vy_aKL(j}6Ju9oM$97wbl#t+)gpf_b{GlVh#9?;EB$Ep95vKLxg7^=K{M_!y
zUX_$5otm4!Pn0RILKoeV2ZwdmCwIFJCXGBY=OGr|c$wkw?GKrv?v2~L`4m`$)JgQt
z7`m;1Gkr~t8&~Cc{kbATSb5${$MyN@C2sMxrXU(h%8S;KgzkwW$}N}&3ZY=X-U(YL
zp|svf3)8pKnsW%|Edhm1B#+vmtY)7gG$0w^Bk<hE(E%G}?A%q0FQFR92J;VG>+EnX
zc9(F31ls7cZVD|HaUbXOdDW<ILIP3WHxi!PVuwN0{K`dN4F}KRB)=NR#(Ti6ZGBH$
zg+|ge?br8@{rqQ9l)*aWZ@hnAU$_iN=WHv?UTxQGjy0(|$QJ(SP-2ucY2jfLeO57<
z-{`)(uw6wjJS|Ai)-Nh4X=0##>?l}SRn^k@;62Q^WY0pv{bDj-dGjfhxa!Qfc%n19
zy2>Hl<i(3^C96@69JU9X{Mh)s=NIP`vaIcqk&)QT48gL72a{`(N1=Gs`wg+_#jDLz
z{o0k)^eoRxd6SIHM&3)4Ot9OGe?o2;_6>-k4u|VufN#2D3W9HwbQuF-k;$^&_wnX2
zg1+f%m-E_O^Mv)f4Q_Bnt#e313I~7HJO38OAo$nE!4;G2ZI^`^CoXFeAmr#NR(0}u
zsu8K)UDq5M9-mm&C3|?QNVseG8t3TB+5vJfgPDb;;T|+9fyHtAdI}QA+N=8fIcZ@Y
zX*YP{HMcJ27UC}bp2dAxvp2<kUg7bge;Slb7S$7KV<*Aa+ZFDh-{E?A*Y%O4$t;~v
z{mk2($P%S7w<8s5QLVIEmGP)i-q*w%J5Q>r+yL28y^w5@<Tz7e{qe|xI8!};ZB^vs
zKatW}8>Jl`f2OM}1)mQOl<<`rynY>0Ypt0uAE*n5=gnVZ-o5g<WV5lH*XB?wZYh_R
zP7G*O!U)60iLfIj6me#dJ^5{0E?1&g#RMOxF0%6@<IJctiA5mF!lTRHEV@aM1-$`c
zrUM7j$_^b8OIwV~7Qz?~37hhO?I-R=Y`88OaMD>mJc>ihdnI^M#~RXCTdXy7u+33$
zYgS;_SQ8;V5Hg1oOb1`3$--k05Ay+W*cUvzFAtL%cihlbl>|J}&(~nh!h*K~YV_?t
zX&2*;dRnL8XdK+}46m&G{=|Wgf<8)a`A=Tn3bUD6<90^@{)>QHy#tM_mzwxXHvfcZ
zG55st*W%Vzdo#W;?>Uvh999kCVJ-3|#r*aF4bzf_9veWz+$o@iv?tn**N-xOiK?<3
zGE|F29KgmL%ST2+@IIE};9);Wv4xS4l6XxQxqKscu_7(x%})zY<w3_hw5{>$9)g4>
zvwRYI^J{t8+(V$XYpK9ScDrQOxOahd;n>n5)ytaGq2GBeVcN0nYE95?%BZx1o%LCr
zo}r=PFP@?;-lX|QKy)?;m~FZhto_wZw-jLK>b@A263EokAc_KK2WmfRPK^QtdVO!^
zRcTIH<=JaY(p!Q$Hm0eO^*{dwT@LRPDa}<8MX=XjTmXt)kSXVcxKlLz?Dy4E8JVi9
zt3zvwC^Gd|XohbqE7i*{-#u+BDJkJL)?6<x3OR|3i=&MU+juQ#DFPYy5`>O2cn$v2
z-^^4aCqg#}h&y`69+ebB>)wU-g8V{VVvh02)7~SA67J)Ifg<(kn>`Cwl;V4JU;w9s
zD8QvPV}82cg?uj68_O>$AL_Ph{n=*UPwIE!zRjB1u*{w$jTO~Z8V|Sg-uQSQx&~2;
zZEG{#N~u6OQ1Cr2w{NQLG&TK&k+0DQ_ZBrhML<U-OAR0DFK4>b^Ys4)MTCv1go%j9
zR;=Xffs}y-aUU|iw#HzJpSvc%=`F7kD56%DQDhp)CzGJ|hreHE3$8$iAC}0^J`R#?
z3R?%{PN@K?m+?Xb-FGgB9S5!IczfzpXNT#0qkG#$%BEZPUXjCL`U<xu692n_e`PN3
z%|*Oq91EJGU0}}_T71F9RsJ|LMpf!iONV-LBY};z5gT%|zqiM7|If*=sQaa^OKt@p
z_+37Sc&QlJkl)NMBBy+b2W|FSAFw8>Nzp%rS+)9ZOY7mhss1{}p6Fuv;F8Lm$8g5C
z9Qn4m|DBFI9XOuuWhj<shwE<AP4`+GJkwLZ&fo9nI*)mg(T(wLvEV&$68=m>pJC4+
zha;+=HA`XH1_s$56kJT<d>PK?8E|!U@>aW4e+D;Yb#+ZdEIz+=s>*{_G&#}Y`4_>@
z0XG?1Y}!=3n1dy&hMz9FySM-9^D|7gD?;CEi_r&dg1Rd4t?>?5t2>vwNUDoTgL#nW
zJ6a7}YJ0k-0=BPj4KCo9JQ8=QvrV;*T^P~cGPNf3|883EmKy!?z@I;3aj*!`gtAYY
z2m?$QJ3G+;IQ9cF+o#FKp7)V0C3I^KWwMJh+LXuw*tkAhNzQ~#OWS=+a^jnh<gLU<
zgN)mP1JSy@Tw<-J@%t$fe9N-FZ?LK=yjHXddZAI&DDHEz&SvX#Djz<SVN+%rguz(*
z8`>~dHt&<tAQ^;-KUDZwih)YRwP$kPuf)1wOeJ~J%KQ2ZH9U<&aDTSWzWAWKU=3CK
zgLchq^d8<8(&D*z=z)Z`G<rn7-%7H!LEWRQ1K?+djN^kCwA63&jEk$(B(SY*UIYPo
zW6`mgd@4#5;0S(pI)C>?Kybgfs9j4awmD4JA0HZ+3J_4}nVazPy2r!zjm<pzu}f@6
zjyKVFAevi+d&^bt(C;YMtiUNMO#IJ2KA*Jka;oUEsuz;mEm|ARZ|@t*Y`PrI&o1Wv
zUG?Ud;G?zfdGp4mEy0YsIuEaQaDep;^{d#-Of8x{to!#1KXZ<)4W|1IXNaPfI~*qN
zXB_M=PkxrR3P$H`yScs~@Ys&exuhJ|#4Q^8P5P@l`XGfqBzmQiz1DN<P(HWsOf1nQ
z3cb!hv4pvAL3!N6#k~!bQtgkf^LO7+>}wNuvUFCW{dE2-pk($gD$yjHp|=Nt_ox~K
z+|TqQHkm>JULU5h*@={4v=K7dz-8uuNOh*fsOlNct?WNgt$OXJ?lp+D@$-XBV~WH4
zOxExquOFA)VyhprUN9la#@O#Dk?A<8zi()CC_}hPjDBpLC%4M*i09rtj6_Ud=gO*W
z{{UQEd))<xlN}>>%co)bqD01NBjdgEB}C-rt1M`B4zo5V<cxrqxiL6w3)z_|Xb%kB
zALbP09kXszBVcgOf;BK@qT5zSwvuX%+=_SYX<i9w7=;!P!)LjBtX1*{E7fY=kQbXN
zH?8)#KdUnsUs3<l8+&pM>=tfaW2%zz{7Yp`Ph<lo?WYKQPsC5)1$k6jTSqsVHbo~y
zsuB->wD_Fhg~rfu;FBJ1PI8$jQjG0Z1TXrL(U8$D%HP#u&lGl<@ag#}%>k&1RR3=^
z_VbM|Pd}U1VlWEHc1Ke;c!{^&>^05nsT-AT44_|LDpi@+GVfxgNcf#wNJnoSplW17
zacv(hoOEz~)!}a{@`cI!`aVs-BGj*&22|4Zgn@&#QBILVL`AjrSpFriah-q5*pq|?
zezcsgQ=2-Cc!>vs8Do?ML>l+5R!W~FjdnK!`${qh?(`)CP<LQyh?>y-uYjsoUwUfF
z0IjL5$Vy+mx<0KJ+C<CJ>SY4@cukKW4r#`r*I~tXFgnUsyR-d~)&2l%FdQxEI){1e
zuaqIF@0cxzva|rRvEi}a18wnom4%h$Y0Zfd0tnhdBBtgP0sC{;y#TD1qvKh%Mb1m6
z>YKo~4UfJ&3d!w`uH?6!xSI|NA%OJA{Hy+W%GTJIX##pDBK@CGvWUYsImhZZXjcZt
zKHF|*{{H^?>slu%DPIB2LxPYH<M+Ru_Q+Qzn$B6T2_q`k|NK(=8zA6T1+lDO{w0q}
zjVLorH1q6d^6op%*V>OZ9&OGish0|@kF**&9lTbKh`w>@6J$PlM@U38<|K84CS0mt
z$13m%<GIShPMy7v^GlY*(fDN=$%Bop;jJMimwfOmic={?opP+J*L`+d$h`ivOHX`|
zxq!(QupNwU^cnl^bq@Tt8=BXuT*~BLs;R{Y`yTcQ3YA4Tqa{{HRqL0Pd4?mF2K+eK
z*-K=MJR(P}-m@Fm=w&|)eu!87uO%T+e_a&j9fsIZC-<=d-XO;%{odPR4Dndb_}giR
zT-`<A$+56c@)8lDKH;bvV4yqwc!kNe@)B=)i{`O}vH0nMJWQ5h2!TY8di#uBdbz01
zd!a_=f^o<M<b|C(n73Wi9QJZL2VU}ddwE&Q?`c1mdHN1lys3M|tu<d!?PO;j8}P;%
zl!Sza5;p$7#|teTY4{#HPRH~=y*^!5Z6jh<IrqIfnl#O!iLKK%nE7n_>4WO-f>#g*
zZlBN3w`$XX?JfIU8^wEhIHt-%JCFjsK3S8z@?h9Kytvrqcnf39(mje_?)n%(hxELZ
zGc@!rIHnG7ySQ$`=<V%2Psn}dhfgJzM|OpWhxY`O65;_pv5y+?uTYYDKncj}{+L1Z
zwp#Lf&e!VLGCR1`u;$619wc$C`}|C7&E`a0%ux4Ky;Hpi*u<4`3oaf*pFhZO)w9Wu
zH){3AeT*QZO%=a52mvn5eWMC9*`7+;fCn5OpG0Nxn7akR(<+~wSg@#(Uwr?PZ`SCJ
z$rJP(!x~s)h#Q6C*G(snzmE@F!9+N-!{=H@@z@X0pyI0V-Qo<@PkE(SFt>QsQ2l_>
z;!-d-mjhMr^O=}-u<>#2ozD+98b{nhGuh+|fTXYObB6m3r4BM^-APkx)urK^NVNA7
zVd3F9hXVs+l?&N20G@RYENDp#6c}x4m&UBShW7_pK|FKf2!C7v>aR51YY)wdM~Eq(
z@-+^4*#P9B&Jj-Jy`G`>6UwZpD-IJNNHW?IhaKPle_XwFRFqNo1w5dD3P^__Eg(op
zOREf^NQX2iNOyNAD1*dEcXvxm4$_j+9nvuL&@p_M_kDlg`quZX<v$K<=DGK|&pqet
zv-iFOJ?^(;)v>EM7;OZ10^b#Er?J=(-K=r>$LKQC`Yw?)AVw`cF42BqTo6Vj7%H#O
zA^R^_PEK9!ozs;_!V4u9rBrqFRliVPyf2-)G!*(|)7h0TXDFD2)qH!BXLS&~tg=Hx
zF$QbS)K|2a`H$WsTx?yC@|lI$?xdYidJ|!yV}{Hr_I1OCQa8<Qr$<jb)w9Rcgt>oK
zDotqLg$={Zr8(#G-pMP-A632)TG!a>gk*`2B>?;PUhP-^VIwrP9V6vmYv+$#{ExF!
zn&Vi?9*Q3^K`WwbBYzjZTRVi&YekI^DG#6vwt8@QfF<1fR_{gAnBt`euW`sd7JM>+
z$&^V7NIoMGeFa{(dW=a+w@NkV#vO=DRKrNWMsSZuMeF2q#)~AFgc{`xU<BpIp12@d
zCm%5qtMzu04DpL97<1Bcs|%OC!^*SLuNNkwua8UsxH$%6pB9i9*2|G9^!E<dN*4bl
z#z3PBpP@{^_Bu+=zJAmFs?exYnS$qOB+NAg@}sA&#HcY8wiY6-F>vG9Y5O$PB@R*7
zA@HSy7x_l!|4WlqdtlsPVqYW)xGx~Zr>Pcs1EoM5S!+`2zJmt9HXbp0){{H;(nd76
zN5}7SHs=>48%C07edL%ox~t4Z)TCzRxl4KQ)hWWdPO)r)=?j=U<#)l5q0mQs#j;S*
z?Yi~U85&a3>?IV?peWG?yh5&&yk=kL{<ZicH0o43QzG&nq3|z2EUcHw-M;NS=U3>^
zJ&CL#9O^$@a|SCeZ!(V@Ii@QHm~y(-O$vb1H(sU?8g^}KX|(iIxX?VNb<%Rm?Z6v$
zwIt*C>F<jelqFeYA=@qU(}S|!y$?@^yG&;cv6UTEbWgPlBq8ek|Cre7AEt=7R#g9}
zjnhmhD=RDaK$-j6ELfq^s3aw_Y?1`?Z+WZsDv};0i?T7ON{tsRYnV@oUW7xc-`UR0
z6sRAR&#;CKy_t@@I$W1Yk&MycjWT4bf9cC+b@29%zuH&~y-Il_+<E9ix+I0jb$
zKealBck(E^@4gsxo+DUCdN4oPZ||sS-~J5&81{!5kcB_IEBfM@`SrJB1=qdlokkUk
z$c!rWC{5+XCa*mSKZhjLYFrq2PrSFPH4R;yv-yEe&SxCDJA-E5HI8VU&>VgUnGM5I
zDQIP%%^xbWyc5mSPY2P<D7_op$D|ta%tA8@a2R{Qzf@Jq%Wp7RhD_=eb6Os`6S_yx
z#1co}M!~%x@dw8L7|EFyP6T78<@#?~7u8|l2_=ez^BCq`;VkT3AO{fw$DF7Q)1Lm`
zF<J@7$@QXh6t(R_IbVAFhrjaSAy(vMWEzKy`6?9lK=ty{CjI$%ER2>hvl%)X38U(M
zBq1AYjz#3NnyK?A(UQxoxi$3W9fp{Hyw4{#0fFk)lp0?|#Yqhm3XN3M+)>SgDC3M?
z&_{60ztHn#E82ol1?4BF{rbZFDNaTBCJKgjmNvACf13HoZ1?rYj~~VVqZhXS){;(d
z3{MsRwOWY%@UhnZ3W<`b<n_^6_h57Uh*qPZOHcG0!0Lmf=K><$kAzpnYQ0HiWX{pG
z6QrGJS?IFHKxNA0jx*zTmdv7}UwYgge)W;A5M=_b+&q%ottuy_(9-MU?yLtX4Ncex
zPC9LozO+g9S>RpikHtl;lhn|oq_bug;MLbDx;|y9?aiY&f#gX7Cc%(G+v$dp5pU8C
zeq*-PGlzw$s;YY42Jp!i11N{KY?AN@EsAZiJu1DgX&I;4<D6m<tMzs&03hgY{!fEK
z@nPXypcBdMSGASAV*JXJ#Z(JOu|}htu8fIqplC<nf4|A+-6J5odX?OBF5y687+Gyl
z&D#k)FCo>*^KmiyE*h1G+A?fxY$Cn1>nq8p7e^!7_}}1^M052nMJJy>ztX2f?T`}_
z59g?@57+=7p>vp8L7okEtcdgWU&W6cEV*9t)Vh0Vz1@GeC__|MK&Iy6J8skN+Tmr?
zkn^4mcqybCTzrNl_R@!)-B_>?9aB8g)fjMDl7ZT+BBd+zaT#^C+5^h9=7V%i_kYTM
z<bc7N%i}kXXcQH|P3AvzgHIPL$_eo>nu5qz<Y;UEDh>QUq@};HDTWtC4uG44)$z-t
z&9Re>;Hc(t@gC4dp%wF<RGMwRa<o00^>BiRQA&tSRP-NDEc|eLv6}CDgTCYI;o)I`
z0h;5ZS{ELxJ&@k8>kH&GBm#S|qTkTZ_95XZXkK)qhAkZ|!LDljz}PLs?{KvmN5yKx
zH3nro?p5sQtECqQu!@)C)a0+SQXd#IX$(*pZW00Q%&U(a!7wA(Q$R++GlHd2QhU7>
z{mR3mgHq(-(ag;yxkhWOdh-paK?YHJUAUuX6%0x@s_%k%ISZxvM@iUXLeC|1U-BE<
z5e~3G2$pu~hgi<Ba(O#4W2uOCo@CO2j_SYcVvh&LMo_i*ZCGvjD4R>kZAsNeiH|Mf
zvGjy*&t=AS=N$*!f5?2a3iK(>_ZNS!8ygutNRp9Gn3r@o%pT+}(!}91M5~#rye&VG
zyg45n13iTK1%TZfp(^fN5Ok;Tx!E#FSrn{+*7yP97UaIqSaQ8n`?7bPTF8cQa4<1#
zZ?0jG73CdnVXB7ky*{+J#e|G{op#fZ@&y83n54nb!;n)z5hJ-VA&fXfY)gb_ztogF
zTICi_+R$e7Gy=P}aNU3xSts@%|Mx#TT%nwqH;epO`|^x0Ni)ja*Df{C8shwvPp}p(
zZb_TMdqI4~ie5dA(_Y}^zMj3sq3jC?%=*l=T0-Oz5?;p}snpZ-E0KmZ@8Y=LmE&D$
zJ$=k)DnAg5s7<D}A8Dy!#tzVIrM|B~&aLCzI=d39`)$SQt(s{zZ-4#5IF=4hu;9&k
zkVULAg4*^4;^Ro};#|Y&8MS&$V@${@9q7u^e6SVW$bEH7k013DU->Y;8iNJv`;@*R
z`-pz)Uoq2@2`_OXA0?<htUCynxiT}i+GO&Loag5WR|FO9IvcoOb&BsjFK>Cy;fJxM
z6G(^s+m+0~csyg6baP09+HX@}RZJe0=U6U6@ie1OL<`1yXYLcmAjcXshptcO(q82{
zJ0P%i$MK0bSEsA3tf>ARcI_c>w7A^Z1;#GZMl5n4$vc!Gd}Y>m^xqpjHxH*3H83Et
zxW2&8HNiac{YEQZPw{UA3zC1@y83XpXem5#CP-trR8D;W-aYl&xP20vkv--!14==R
z!eWz_WCzHxmPoPkY2-%V6@7|6#9iqa6g|#m8$f)bW{<pr_>=q+oqIg`(a|~5A~3Qw
z{uS~oB@-CZWQ}*ncjd3;%3gbJ<})>mS&#V8btltS&+yxm6(rYF5+X8IrKJ-qDnb%M
z0=|5akH~mEAf3Xf5AeGlwQ+sFYy3Jdg2(&sggt@BeN2uGC6%RRgX8K0wZghZ|Ch)0
z$kJz^mGARAT)*KOP9wTDlB-}fa;S?&&yYUZnrjf+BEo>Uf5Q{j&GR8U9iNqGo05Gy
zZj6?j1zB?YjXwS=y&=y%Mo|f_aIbE&JIZ;<?<Bq6x&;uF+IyTzH)Ry&Mcsk?QV+up
z27e!aWi%EIqk>`N9F|x!5=%3jF8AINr(e#ERJ|qf9qAV`ls^Vk+duuiD3RmAw{rj>
zx{`a>%O9bl!c0YdM-fVBJ_VP=_5oB81^^48M4e8G6|-ZCHXzB03K|>%P&lxvAK#v;
zqz;L8;xik-3xkL7;X7*Yk7?$nk6oWG#Ifymh0;cqFLzK}gGwn2u^P+Z5kk)Wu<uW=
zz-o%uSws+@!!#IiAl;|yv`f-XU$6(2R;0YAB(YfAjkDPDwO@Vp9Uek*6SpSvmlWK4
zx{l=&fa<dNdIOr@0r_VJ-S41q@Rf&_6(tv)!7w1Dp>O`EYprN)>+kj6%i6$5udnZH
zC+$OTSE`MN&`CsQHq_WH)L4y|;%g&(rS_){CmwHrrv15frjkU$ql?ye+_E8EAi3vA
z5c6={3IiMOLcPsQBp~UkrYQ0#Fee@ih5n5*@E98{{gu5ENgw$}bK*QvF<1MwA-5JG
z%NM9@c=kpO3v-ewvG~DCR4(Li7F;401&9oK)h;uFBd0^cPHV7;X;RP~3xCXmNwETf
zh_$UBP6wVLK7rUs!g%p!^{PT{|ADd2BV|wBW_9*Bw?}}?;-}_RhX%IY!AK@ZwVHcY
zXWS9eduAq>lrHZ?L)b6zO7U7B-N?ZH>UM35-_DV$JD?a-7JUh+f%$D2Fl1spFcvCE
z&T_ZE6?D~h*KXTzNH3|`BSK$yRS5GR;WKe(LmNZy5jNeRor9*E&z9~Uo}M~c-9XaO
zR&#|We{kr0?H#w7;8kFK^`p4O7Oau6B|&|n5@rK&xqDQeouPi$O{*_Q^4|oM1VpCz
zG`H@R3mNzsm);{zkykLuQ&sZ1skksxqpAa>O=46r6OsMaboD!QFklKn34OqqQK<DF
zZwBlEdLb?7q^7k_UP6y58f6Y?iJI6=$%uv&8G9mSGXp3M>8&U&RyZ<GI}QgVRB!n=
z-R6v*FjzfF#A<g5@z-)1qc6p8Wb+&i{GG-ez~_9A-#{l|2T!3-uz%>$XL>9>Ru<+i
zdUlCf4Ogr?LKaZqG6u9JPY?2@;l_z++E3kqB&V~O68k(uto*oAt&W7nf;bQv#fGqU
z`N8C28#gF@`4#GnBXLA^9G#D{-hKJ;V}=<mR=$ec&Ea>cqy^tWqv6XBZ}36lcyyaX
zN}$bU3=T`@Fc^QJ5I39;8bMT2;B$LOCt;L0HqxM3Cgz2~$RNyC(CW8>^RS+6uPuQt
ziMC9qj@SO$4yk~V-aKL8^%jzEh(LQ(-dt**)!5AVpdR4LFN9qZY{2P>ApSTi!W6GE
z=ozV4Pj&FL+$RT!o)frU%X6%6XJtQktxQfKHK`Wo<KfJRYrIao=b4{e)_K15b3Hg8
zpD1{7-VhqC1nqC96y`C0_i@`c#wipAN*(ypIDJIx<Ojxj4eX$HXMZC0b-!tI&qUbx
z)6nAxj3+#XsQ2Y!%~zxO+R&;ut{+ki%IsD$x-in;vvYB=XjFTNk3A7OT~Lt)K3Pw(
zu(9<|&=iK=p9v@-db`KJ&Z<7P1%-t0^B<QK5UtS@S#TNPJzJXH{nZ_HbCq>fgRcGY
z{o+OXjXx%xG#lr8jk06vaU5&3lhb0y#HD_u%<fu}ZBNR{ZbQVv_1@*-F+QyrWx*PF
zokDxgNgfyj;Y4Id%X{4#lA$E_bxGmNdu*X|Qc%O*1gW(}Wp*N(=s{_lrW)v{)I`F!
zP)W8>@oY!5_<(khm<-gbgm3I=XcH6PRA^*n71(#gz5!ZxSmfyvIPrmfU6kDt-~w^2
zlQz&QY2UZRy-aV^rPK$chs@OenSJF6ATMZaQFyKW)P_=`b~r3p-8p=KzYs2HJ^4nr
z=4xj@H5fuP{0HNuLfy`|E~-I)WQsq@{N{In@7feJA|m4I><{p^c>5OZb-Ih%oGh0r
zH=#fs9yZE^Q_g3$f5?AsQ~ri|Vl{E#T(^c5t5m0sJv%>pGP-abXnPPVnSX(KkZ!)a
zSKPq4QTQ7=C_Xj+QzcnYo~25!dJ7YAYEqJXjt6DTci5NJS1;Hxc>(Q1&p{ag5GhT<
z<P#Q=!B2ewlva55>z8VX>(aF@%)A26G<Y458Ek-@;j>ze(UM412{vGLYAOPCog@nm
z5T8{i9t=BLn6>M@aJ1$87lORfe9Ii5f|0%p)VcX($<jwofMUTnqzFV+*$rw};coA^
z!7>@LY+{=wAejJV0P^Z9?)R(^M}w6b75W-e67d(7>cG{aX{-v6JgEL6{F2rU+;^=9
zX~l7&DRrE>wP!ve$dUcZM~@@`N1z)VEtdxDEf5Etz8;KkK0PFq7PRdvvmD=!G+F?3
zWJ^-=_l{eAk4EyhtxAT6fIkorbB=58ekkZUH>h`~SxQfCT2^GlPfTkxJ;Wyp(-0QY
zu`^tqr8gr8Mm}*enPY6yLwh8bj+s7eKM<L8+oz|J8e$f`9BQVI+q(jqk<8MkihQi@
z|2imibMx{p)m;?As07&2i`S>1Av4>>C;8ClASbSa{_e%cKap*<3$E0@NPh*;br*;|
z{iC608-z#vkDUkdlG!>ZqEdPfIg6PHQTFOj6DrfU*-nmE7(E8hJCiFSUZLuy1taYy
zBuu>4VfkB&0!aCL{`6`rEx1<}tnwON{9T4iIMyi4)LeDM8I(SM!*9RX(r7V~7bl6>
ztw-9RpIXjmdXx*l;ZOJT%TBxl+2;^7`CwYo{X!?e%eSbJVsn+Yl8nCM$<`W>yHLqp
z!}1IC<_m61w1|PypycvT&35z)5(SR5`W?r|G1aNIKMD@p`{w*$=3n#M)T8(^Ks<gJ
zNsy8%LkMY+2cXqJvWbkoe_Iz+_sg$`zfysIXD#B9MhcFFEqv)S1QPR-N*I#=&cIKc
zYp0ZmNhgHMZn}ym-KW8|15tV|uo^%hX?=D)VJIZ#{iGu<U!H`LGT#-xnR;_Z0yq0k
zWN)OdXVT!Yj4wobw&fkVmH-u+<Nv$HU0hUVt6i}<*q_@K5QO);>^TH7$T*z11v9t8
z-*bt^dk@5{+^*?O_ejZrT81gp?9<;43y?I~GG-d4luLksd$S(|<)6<SC;szDnK30T
zzW=zmb#jD#9hFi?M|QON>R{}pSag@cr_zL_?>cIWzt6A^d8gBob9t!V>K@wJ3S|wX
z^;D4=@*b<=+RKbNFushQ@d#`aCHmzRoJYrNf5#0cc=(gzg2b)(3muAk0Xix`lzaE=
zvS?8Jbar=ZH%_a$u(!%~tOB|!R>0m1^M@bv8t8B;#6M?Ho89I-?Hd{V6s#+aBrjNT
zbmHDFE+(in8$`>_77hd9XTDIo%eSIWKl53FpLH-|4NbA1{GOek(hRrJ=kBSAf2&T$
z7rb$Q%HVG6c=z;({J7>86l&zUKYJD!Xakg#mnn_hHAgI&L*!*45(*A&uD=8B((3&?
zO!28{!?INUH+xY?A`r|mGWk`=a=S6+&dU<Obr_OyxPJNRP?-`5m<$;eSfrg7A=V6s
zA?-4!&LrF3<eBMC@!nfD93AakK$bc!cUR}jmoJ5nH$Vlr|9r;eM}Dqa<#!#DjIns4
zPqu!ef3DdgKa)6k*ZlEvz~6M&c%KQhw~Q~JU^+0V4`}M}duV*z#%E0AV0jM!(MBN&
zF0O%2&t}HVRE^0aix|}BlX=jJO@yU_Z1QvA8uZAJak}`qYeSDP_EV`8223c?#e+(<
z_iZ_q3C`aexascrH`J9pFXva9583SKCGFfT%rE!8!3o*ycA6V{`xD8=IQ`5n>;Vvz
zW`idMu5_*>?lR$7E*zQh=)F&ppa+3+$tI|n({}_$-virjfaWC=gZj<BPNFd&{h)Ew
zvedG0)zW6cS?m9}0411E7WVZrk|E3UQG*Y8kBF)ZT6ZjEMe-a6GzJLdL&JQ&!s`EM
zD}f)Q?I_9k02<|_o#cc=Qj8(@i8#Ty*j~o=7)7i_Gtj^`-ZE~nBc23viy4NjI(Ux%
z|3dPA?ijL#K~x1Qe>x%%2;Y%*ek)8~A8knm(qh?9Th-js{f}du8b<B_j#T;rcLx-c
zk$BTY(#g*4S>E&U_t*x3jdGaJ?ShB>jc&*~`PtV^V(kh6_s7A_t;b49ima0~#;LFK
z8#|ESX9JT2;YO|h3GlDZ|0}@HXXm`5Oa__>rb?11`H}0(QzR~KNEO5HS6jzIvp+6@
zf+Edgp;SNkxX>W^Im6Pjk_6Gs;NPh#K2e~gnY{7?>=>C@f~M<}V&)Kvw4dI4e>2*E
zSC{kdQQz5t*?gLW_|Amk#6|y^xlGoMZteLx*@~CYz58r+AS%?|w|kOYS1%oM=;!6O
zJ^5$eXK!v_phz<)IyOdSZ@vjdytFjpd9?kp_V2{hk<<KwcN4w)^*8*_m#4eO3WME$
z&o{i~%9`u3=)^vCf2ClR?=~lVY<+q&PQ)ZD!);jJs(u-}zRubtzLT=Q*L=-&1lW__
zp)y&yZ<l45YZrL<SwW}CT3K1SNv>Rj1cgNx#n$_+dtHB}pB#mkX$adGqS|<-r7SEd
z3teJCpW*9pyD9zGTMK{%^*4kO_wd9Op*Uon5)l&&4*`H}-}R52PX}7aC;VLPZt*l*
zr~+uMH*f%!;s;ySk*r`>lHhzkA{*vRAf=VA-mjvxJaDh>@n~-P27CoTwC8O71|?-0
znW<9%3e5x1H6zxhkQ>#Pb6Y>j@cp|<`aL|n^rRUE8H(W`P>xVFW}EZ7x$Y;R6E74J
z0&IyK-NQBle9aO)vWlxNlc3Avy}jWo6(HCTBPLl0XV3BL!btA#-aq#}$ovVW8K+FY
z_C)!%Te;qsKyo;X8kD6dvOMotH=HQJ1&vy@ze^n<aJ)xilWRAu0MOaeQdjkU+uTkZ
z@Y`xVRUEw}c!@~+%B@}hWECuvGH1y=zx`M4SJZZb)!sYp1?qxM<Han`e069KyVCq<
ztVwCtw)Nlwoq)X&<8+X=TLN4JzRy-Vj)sUc5^T?~uhUI@&jK0lf39O7X<5Ze<rU5s
zP@HeSjM=eF=s71d)K#GrWvR7m{>5xc>LV$@Q*Li*lqrn-Zh-x~?~bLkmx_w7vGldN
zu-nt)L|nWItsEw@D{1!nAxGgnoZI@F5sL2{@p^%GpXF$8!7dPwu;dm{8^G$qgrv)e
zquZB<+~AKBQ2j!<a3G!$SX*E9+L^6ubL)v_Q-CgRl>TeeEipv=`w+)9LE4a+hKoS6
zsF$5``s_7X+#Kyri6T^@WWKewGAcN#MxsH4go`C4_Sn{~#*Vw;(Ii${ZxS#>u_(wj
zpzD!`jEsz9`VDRu%?F*-3Mbn$dwzROr!{&vGZz5>_H4428EPYd0~hR4AC(Yt&hOMh
zw#C?Y&*Dr?OsqhAak0U&qbI#Xk^`THcO`b*kn?P7685jn>DkuaAzd7n-_1@f+s<^&
zG8zNjS=`65r;sFg(Z%`%`u)R@KgP9=-CV`akF>OAgTAMQ<vqq=@;iVUxS*30<Q7+%
z#_0D<GW(B-v51zJm$t@6Isf-m@&$b3e#a)7+FkgkHBMzlRC}_kMqQ5oSvoMlncgce
zPaXl={x%!ZU|m5$UOw_p)b)x>K{)iNvRjapjzV*jM~$*lGphi&%&@D9)EQBvr-I6R
zn_1L(<L}3w{*wKAfk${UiToCh=T3I|Y$jIhdf7dVy)YL13n4FMS69WH>q4%BQ_Ni=
z;>XtUJjODGGN`@#Bc=K?OM*7jhXy{E&fVPXhp7U{sGopJs9B4x{}t$XVrYOxu<da6
z0rRVi8k_0pvr8OG*;ywyE1pV=f!Vq+LAwi|G!vaVpu`&+FPYi5s(0N{@2+zNARRv9
zYzftSasDre&frAH*|~~}WF$-yq``{gi@&EA1a0xc!@|O-c)ZWh-!E>Z5vfg18-EW;
z5P-Uv@h;Bl1vl5Ye6m>G)W7Jg3pzhN^8J=O1iwr=%ZtXee!#dC*U2!_Nj?5YWq{aI
z3ys!0z<NA<`qp>t4t}+KK3O}G#NgV3zv-t?f(j9;JY(GAyklduru?PHqcNlgcagV~
zSK({<hp`Xo+uNwsL2CYcEFBF?y}v0Y?jQHhoz5oNJfi0%;d6A>*&8O#`i(pbCXpze
z-T}g-H60vS&jR4Tie2m8atW-*Xov-(oaf~%f?pkroDcqyyVSPiSM4cZ?)jRWe0b!P
zf&W@z=4O3GvYDB=P$o}p8-3usGu>fIH37aY=Elzz^i$(#xZR3|K<)d#YxG?+xu}cw
zD`)CHws3EKhv(`8BhcsFZ;{_Cj_cz;0;y{RFcx&}G)+&%#lvg-gfmlRp>Ls(30m}j
z1MhSHI&SGp5gjzCj)c;%+QUz5W(>&Pd2sRWJ-?M$uCA|J=#M45FPWK~^KApodo@Ij
z_1Ltp<n{Hlo%Kc!)ax~G;k=KI!w>z4&ML+_-<U)v05LZV_VZB{F_n;|5vu9i$B(OJ
zhQEFwE0$Mn|6ivPh<_!Sc6QH*@K0M5%sMHn76X!gafNxc6YG31Y!I458732_z%+u~
znb<uTMpIiI93PjwmkbFB5s(UcXu~l4M|e2x>Pzbbv;FBLlkDS*+Okg7iQAbezb)O5
zq0X#C`b@Yu-?<;gD3oy({r(a{tMrjWkNkkNF)ZgP#sBSrf~&fBgSW>tDzf9?DEk-S
zCfLz?H!OmJIIu)R7mcBj<*q#7iTrogR3S^fM85p26?ag!@9GEvRHeLn6U~7$0YFw|
z<hz8)xMg)5%M?*Ntt>wrC4MVezvljrWF4e=x0g$NYrG74oJ=Z!pfI_tz}gO&C4}@T
zGa}=A&hyT$cH~$yWe)rAG0GKY`(!sNm3?l;9KC<beLu1bms3#40ot&bT3=K@8!4Fy
zpKhkUQPN^UmBI9()q#mgO#u7(IaYOM?);asW?ltrb7ES9MyK`CJD!HF+K(AT*S(wG
z60BR>;uC|88SwCEUpFBm?TM7eBf48SXG%20IMwFj+0?9}KlgvDijfKmrRzl!xrw;H
zgz8igbnref#!(*-T6{TOl?LnU@Ob>^fw67nLYU9pd^IS3+Oq<JCz#OELhifAN*_MF
z^e9PhZN=n%fG?o4_sP+bxM-6`Nir^@AK;&4zks*f?{yn2SfIYBn+R{w7*m}W%g_^H
zPpG%D>?$fvW57xLIv^h?#0P<}KiEI*6;k6FX2*YL5i%wq&lUFy`iw@e>%o)(ULegY
z=sn`8nkd+A)ZS6d=j!iA5qmqu%-1-Oee-_Zs;7c`HwAu*Y`2X55_lI_uu@hRlrg{t
zLP_RdWA_>sQWAY8h1gqs1(t#RsRmkfrRj`!On3Fl>n`knKbijRDE5xFP+_uY^c-|m
zZ)H}Sn=kvw;RBJ<aa4yd!F$R|F9*es+alU`v`3kycQ<h!Led3)ZRWnhnl~`Wy`iqP
zo?5b-%Q2DU!qxg^|4~c9O`zhfaR41kpx;ra(o^AQy!NMabiwGze3tA@o=2F=&xiJI
zKf1+2-rNnU*jgd_-`3@Sk@<BN9&j@_pYl>xCCBcLUxm8k$dvy}%qP|r+}~U8egE<_
z{L9k^zc564{GTm7{q^J3PWw%JerW#kgUs}-a9r4T`t-AZ*tQ&um(ARtAyam$bDMj!
z?xHmrqMHv>6j?od*N--HY>zEnmAdxhza{?M!8+4d!R+2bjne40o0c&Cua|c+Gst
zs3Cc?X7e)~2D5yR%7)JTV2}zgiV?dW9eI1}=9;yKxRKb(I!7(rqGuL#;8X(j#;3G%
z4=gXA2Y&ndFOjIN*9f0|M{hx5-PPx$ZavZWd^K)JEl<(LcVJ4qVNcTj&C$km_4p-K
zF=mP_m+l0Uh`O6A(f_V7HS#9yuM)Nf5xx?u1C|$2vU<a3In=L<`EvBOr&|aa@mt6)
zc6fIs7z+nm<qipJZBr+7^QZ4r$1(53Xd29&&S0&}?#<U4h9Z{3CCB2BqZvN`%y7wf
z&F4yO26d3>Cs&n&v|7igL7t>qo65&8yk8t<u&GVI$otCB9+8?_BlPT9uNK0BL~Wps
zXaf{~G*9Fn?t2zsM~lcsXxL}JciS0i%@V5uJ#cWT>$5uBWUNUHq~)g}O&(lV9XY>~
zF<sLolc@RCXlr_pr*_FZH+tf|JEEJ=h{b;fBY$aa>zX*G?hS$tuf@Y*h(|!+_?ikF
zrItw#%rn%ty|^=r!09kgm<20u4tom~vSi@4mw@HlF8cW8OijtE%Y?q480e4}cXxNE
zQ(ix!qcglfaU-mK8JU<Ic4zApPIhNlp`WsS{-(ck+HG<(8pop0Q|1TKn%Gx=$I?l`
zh&W;f2wTR!5k@Jnz1!;cWfVH>^TxX}=WxR(B2po&n38|Y^}<@c)~r7(De2$8vD6e8
zEWdOz%j%!rwy_`g(BLI<Ga?-FydrWlmOZzjp4z{V%ptCHnO{p%uYmL=$w=`yzV8es
z7!YdvXmjSyR95h*eBXO`;WW()`^DOx+#*O{=2u)d&qQ0fx}K4$ffUggb-}qL6ZaV<
zqDPzVR|j42{Gf)-ku(pWQFl-Do1d*~z}POerkS@@(kI!VkErimp@v*zG(r}`>k+uY
zK&kZw{(w;bWRu2Bqt^IcP`_3j@j}l^1~X%(r(%xKC9XdoK0_&Qj-K+^8r5sOJNB`q
zbys8v`WU_(b&Sjge1>_9=MP&T`jt9!FB&!TvXU(1>Yl~d27qxlKQRB%9bKWGChv21
zQBVrT1>uEw+sJ8KwXNlQ8ii4E(D8_xpBQ|3bMK1*A=u4FBAMSAulb1?k_m+gJ8sxR
zU#Vn>oI1C6bWGK|?5H~b#-do8b3~SO&7?VWrhQRK<J6-R^<L>#Om+PPb^paA(qn`i
z%~2cD(8L!N{<nL7iic+?N1$#wDk^HQuV)~9KX>QCaN;_5_qkKErUkx%`(_UL?sO+q
zdTKzT!R+_nXJsND>ojh(ZnZS2C$djpKOPz)_iAvpS;e~M>PC3A24?AtxSh<rxh~F<
zZOauG=XChjI!HzH!u|QMZEX<0!kg!GzGoRyZgVbfF>X&hks$xJyE+lIW3@iB31CM2
zd@j~$RPG{K>Ge;+*tUmO$aXfkhrH6a?OS)uGfgevHKB2oC{UCIG!LKH$`djyr`zw7
zJk)T_qbG_G#A&%>;`jb@EgnNDf6mqX$N*2^FGx$$MactVtU<Vuq+F$rQm%Fh16nV4
zzKr%12*ZfR!e7FlsIp?!eIxY4O5dfgBWhV^XZp?%c#*+Y6hHS-ZmRjaASrzYL&lp>
z#id;qM#cw@cKC7&*~48PBfXE1rd9E^UM`g*Q<-#rXFQfmk}Nd_D5Qv6N263YjU??s
z|LL<)CUikwUJhRK)#36;T1qp*=x28naS~_t6Yhz#gLT7ECU8bjsudcLg6%A0BCU(8
z7CNpnGxzu2qY2As2Nl!imq1H!!2!tVJvGFdNW*_Le)+QRf)r}c??M78;Sn(H-%UE)
zZ_CrKItbxz4y6_MbiR!8sPpR)&i~#r?{~8z2!eYCfyCLOMTsn@@kWf)pw>Yf!2vk2
zrMp+4S+|_mKz!c!5+R)CQ)6!CT(ug#IMpg0VpRHwClj@K^cmJ2?HIz)eh)H;VP0fw
zBoL^D^HThwqJjeFWm0lPLV@p<LB7o6P5(%$W8Oc}6~m)K^K$Q$;7)HBYDD@iug6@*
ztB#HoA0;FdmMYolOZ^$mvTsgIOcW9Ax9qpt$B6>(S0Quyt4#1$`JsJ%e;}QA%yahY
zKTFgF4@>{&_3MVK`*+WbI`BmBjL}AbwMrsldvzv=OTlGr{frY|xNNYlPSemwL2;r~
zKLi+K=o3>@s}(*-SzB2N+?i@%H3B{bOTFC9F#VIU3YooTl;UsDUQE<*NQl$-+&q{t
z9hhCe&<4GO?ua(oR?RYfn(NUhC1#oz-dbXZ-J>f)>wQxsu)Y4&6j8qtYfn|9%&rd_
z)5h>!8ZYGJ;SdjyG2<InI9psH(o$oyW!Uj)d9zVgd}h0k1uWIW8_xAlgPN*Hg9vDO
z$FFM!c#YRWbJ?Gs=8sK#f*X$DGw`XN7U0cQNE+t=J{~#krnuM0^jEBQ4LdbibM>xK
zV4;G>x2uf2&nL?sNjjgZ<HTw!PAAjBDQcpRXL`l1{fX5^XZu#uMaC12+LcQ{#Qs>~
z=F0b6A&GesPSgKTx6V<zFG@*bN)KyrV0)+uDT|VvcQ~06zj2n**0Z+WJomX=%dL_Z
zoT;+W9d4A4w7HqT{x{fk_GLyz7Ya?yCw2NpxP?6~FF;7tGIf#hMT%@)l}YgC&SF|_
zv~Pnj0}~(0GsaQDkkum1v^rLaFHl*??y8vBQ$1ec*g*>ftbG2WV{h6Ipq?%a_s0(A
z<>jNpdM=%Il>b_P>v}ymx^ej{SO)pwNjy%bl}h9s=9x8&d`!lJccNJ7P(P(}n-|kA
zTUP#Pg9izp1h7YT!?)M>WP4*dmFn(I)Y#}&?s@oxMC#^}!W+@`iVJ?HDGe&`1X}g|
z8p$XqG`B<$o^zFEN>jj}Bon+e<;LX}-*igg%w7+OHD>8bR8_6lPlSP=_6AQE2}x2)
z$(o<>K=Ko$F`TAUQPK~+X_kTca?7PG)ohhh=a+pD@(7U86D`J%0yGW8i${V`YXaiP
z;>rg_e2$yBNnKTH7@5j}AN~yI=r*|Q=%ligxfG3?mKm&fhj7<DIZbx*OsU)Z%fPdi
zJG3KMyzujn%Oe2labK+=%WCU<b5&71Z=lezj5`tpRIkNwKXCYehNyq{HZwr=*AE<8
ztMy0;{EXrr8KhQVdq&=$ntunnrPD*;IU2>QF(I>Me9xhpZIxB*3}g>0MSmfC<sUD!
zuYZUwL?7P4hHX8Rf<dKVyIUEUp<BBe;CV3SmBn&C!yP>mKjp5AxeHSqDmM;_{Mc4n
zI4=CL0q`@v;0o%d^zxsw#yzBp-t+cbY+<OysO-_o49o&6c9qEnFavfquCu?RPtjfq
z$#iofuEjIPlX*&MnCr-Mx_+x|q(}uk?cPkiPmpvI`rV$Bqy8Oe>F9csXNvwwlkyJ8
zZ*@v+7#R=0vaP;S>%mj&K_sfy@a9x)$a1Uouv$KSne9CKDm`sho;oKDo`1tD&nBr(
z00}tmW?A*5HUH0=dWjP}y&08ys=x+E!N!E$iY{6%BGk1uH0r&<=a%&2rP{i#|NMrt
z|HxuicX*v_bKyR2@9dbGQ_%d3B?X(`ibziOHvLH(Y|pd=kG5gRf1k0xzb`r53>1my
z$(V}g=mxu9*-!Y7#M*5ZOC6SFayW@eLdK8_<94Ct2?i#-9JM}3y5FhjYwkk+7n6t^
z|HIARSf58fhm4@m>K1JR`!ZzV$nc29T+PXaPR~qkwDf>QZgda^2m58(@~IABJ%clS
zah-oW)s&nVO%YPBG{F8VB^j>8LB4AXM!+5w7f0bKww_I7L7CH5+Q=(J=#?{S+f_NJ
zsQI;hsL9Js`}py1tJUv&y-LNziTn<t6}h~wN%&r_du(NUQ#n>bX>l3bljUk-_odaL
z+4?z0sL@j6u7@U<GgULMeRUDP^XfBYE8#JDjl1I|RykcIUbVKX6jBp*lZ30ifFrOK
zpIhHRw%2LBdgvLEq%fVCrhDj50unlxyHzX^JtesGL@#;a@8^u=3ckOMBa0Rc{Q31O
zwn*n2=;!DpWX<DFHVzfcjE^711flq|a7i@fD&)0@>xMF=Ev4_wqdH;XkA;fiNHZ*~
z^9-a=F+5S+OmPGRx#%#iN5YOP!7ph(f4_DZTQ08ro(0WT%5qo#`HX4VaMi_`YkOf~
zebj4E{Kk6(Q(OI{ODpH{o%!(OPIqBX_tMgj<g2TS1)R5U``Kw|M|p*F#O_sFotDC!
z283sj^eG~_0~0nN@8ypl8!)+OZn})O4To*<9lL2><4otb3f7Dd=9bE;552vf&+``t
z8*zkMdI0nW%f7r&;nO{ysptwFw->}jD*Z_UR<WAMczz3BBcn_YD2(bCkV00VdUE^O
z3AoNI&ubS1w}jIDI&e)_>)ke0pF1qIGfn7wmAFONNN#`++?3v(Ulf%Io**a6x@W4b
zXRFv8Tj$)WFKkx{*i#5Yf7*Ao5r{iaCV2o1L+X8*4yS3KXUV459=eP}A%Y6y2n4L1
zvVr%hV&}5^oy0V17C+cJO*{=0!<F0~CNMLv?id>{vvWE=TIOx~`8>_-389LQE8$BV
zi1;(>tjhO8+_}-an1=@!3FnI-;my~<q<P{=(qgV*PSk=6*<Jj|bo-iJIy(9hJa31}
zYI+VWEC|}n6JM|C+6mBju4;JnM-khQQ7qIs<q@m(X#B1SU=`p0pnGgDq>!`ab?}wK
z^%O9kCP%K)d{4B$FhIXFSg?OMNn1}_@B%rjJ59D@Oj?|#0WMqXy)tM1XhOeHWp)ns
zBepl6Drb5pNZ1&kT{pv?sz~U|{rR{JRZqLlzVq5oAaKJ-u1?#=3PczW&!jky$&-PL
zIvt^VAXO&9;(zl=Zu5B{<^4r`$nsb)fj{f0LS}9>tsfo<k;;esQ3uOU${OnIB6hM=
zL@I5B)YO`S6~K(GrHtY*NZ-Tcq_NJ)t-&`}=I{1o-N@NWl+se3xOmBG51n7A*LJzA
zZeKEwu_3){4I)7C{y|vZyBp8VKWu8M@S}~PCY|+mqJJ`*G;Xu00*jd;g+8nv&d$yr
zi+Sx;U)9Io_e36U#y$;E`uuzU=g-4q(%vOG1b4*-yyXWWV>-qg#!QG>f@bFil7E5f
zbz>GbtT42YosZ4GR@`}AOnlh<4gn<?poZ+r0O&VDQT_Qsx>$;>)(}uXSRaf2{4V`M
zo^58rUMM)9_C>PaJ)Rm$ICUgFF(Enm#_#CgN@L_kTQ5Ky&Th}HP!a12r#&`&#Yh0^
zI~tH|aBhxQBZuB`KZ+LmuqZtz;eHdqQr37h-28KY5j`Vu-h4?A`j{uP*cbZd&AmiH
zD@!}@_Drx{V4v*y9R`Q8KFpt}^Qik6B5C;e^n%*%$r>4X;DXmpwkxZXW1DK8ys%D6
zvw-E5EKOGHmmal-)dqQZ82psY-57es^MH+;yLOmfrMcHRB<GOS9SU_xU|Us{f}KOY
z-7&}fkU9U@WoLy<t;Z(sN}Oy76T<2IW5dfA<2G{;!MuJ}Un1f0xz`DJ)lJTZocpk2
z%0uxdVnSHn*YEVo_Q@DI6C)U)P~YLa!8+-!Y}=WW9TmY5=r5J!@=RV3|1}NVe`5(3
zx(<E$n$}_9bS&WU+gm<LO((H%a4bcVOJaUk<R<-9zysTc<?wCXi({lCFG%4LH<y}4
zA=r4`r#W1aYSA#~b{W~3!v_yq;g1W%0OO$reQzD<YBO`LZl-P%fI8ltYhc30qg-@2
z0?N{c-ieVymuKUA%$XOqr2b~=Yh*4;f%D@rkX_WzIk&n-gH}77|2*n+l;@rcYI3>F
zm>bRu{zqj$7!K#QA*C=>wNA`P0|t5&-49*2M+-R2D_nQVV!xe@ZH$&n{iz%jFB*vX
zIBmmlI_GDI+~O)S4mvUUos${4^mDGk%?pndY=8RsW=tHj4yN#339P0>W`0gO;cACv
z78UIod}<*XVMDjQ@sjD^0*smUEzzoDLuzgf0ve~UARD3pTrg6TQ({uoFpVvT*b3&%
zwg$+9NH#{akqEU7Ig0dq^g6#G92;Y!paPLkEvO-G;j(kL<uPP0|3Z0oM|~lm)X_5V
z2dgYDe5t;J;NTYt(bx}zrU5oR;5yb0f}0YRuo2WY{iL%0#OwO;x|5XiNIZ`)_tpA&
z;~1heL?4B$j|Y{_^Ig#0LY-6U*?+w8uEsmQKy{>b&AY7YcRi~Sc3AqfkpVVi(OIvf
zjrWVl{Bj@wiGkm{IGHIMtAFTa364xzXcsXgs6QG`r)8Fbu7J%D3sP9@XC`;f<t74V
zJEauViJ2|RQbT1L0zt=rk31kXn7_0r+LEpA{WexQ4lm@%R_*ri<geske#OdI_$m&M
zD!MEdy*!9$zrIW`OZ05W1@kQHeX;pPi#xHt+I1(f!pO2jmC9A5qN~^fC#8%pwcA`f
z#bnm~8lN?}Y`rga=Y$ds#BHk`$ZrI64swJA?DrL@W<;y)3}cOU<20<61Z}5D|Jjls
zpKpy>^)^$ah*WWJ57+UH^b-wP@C3fFz2{Ji<i%sa-jH>)*`3=uX(P=1>NKB-qyYmO
zGb^4pketK*(8RWywp+s=xbg7ABYz@D=Y~J{lqkYb&u`A9Yh-vhN~5Q;K#k7AY3c01
zzDQu_4K(nd>)pyZ)3y+9MD&l;6Zalap1PS&{w(sQWjjQB?R6DJ@O`}Jpy$&skQbUf
zRbgTUg{PXO6d16wJc)jM3(;J7Rs|V2Nc)0(Iwp)|`j>~3vx^ML@#ak4Mk{Kt_Z)=0
z1507<#<SP0?OrUV4K?R0BT7qSPTtnVJE`|cDzG#K!O3O+^#u^;$g33L+B&}SK-Cxs
zJ8z_g@7y5}vp+j&*cHhleK_kob0%o$Q|;Y=Jk8E-$E&ls`X!iC|70_OKy1UI3XGIs
zbfN6K1y>yMqKg1C-XKJ@vi@2bv1bTT?y(YsLVw#FL^3Np0PnN+&L_}Z^4VXFyYzM<
zkEnIWR3S)EEp@Kd$c?7PV;KFjjL?%IBU0gXeDV0XLed1y^HBT1K>kT<IjFh<-*~?{
z5y>O9t^QfbPuIk1bH`OSGvj)7OP66~%_3P-d1=1G8=GdSZ+$m&;-W-ccAx6j+Nu$g
zP+Pf0$9`K`UH!T{RVixJb&`^)O#apF#d2tCrRNEUV{M;6(R=9y-!7UGJ;}-$8wOUT
zwLqJ>3Tf-98cw-ujYgo?CJ<1PJ3<p!iXSi3F3kBHw1kM5nV8AWiaHo2(VCr1PM26s
z7#Y@_ToXvl#w3a`pEsS(E5%>61lSm+9eY;V&NaMid;S>k55mtj^U{oPPWIAJqiRZf
z@UXDsEia9s$#P>Q(m2vRU&iHNa!GUrd#Ov&d-01@H-lHpfh;v7Q(bL-SO0=Gwv(fX
zIHMxJ1MNXUkZsfPILU6T6KVlH&T^^cdTy7PC0p5J=yx?F8zy8ot!FT@QTS0awGlbc
zFpFIc%m4k{rHPJ_G1K1*SyHR>SiH*ciFTvpR|!?h@1~rpPXeIoQIt<Ed;%3JxE=`O
zwTwX$H{N=Amt=jqH$>2~fRUdu|3_qX#sHFr;ocwZ=+<xjRnx}`XVLZA)z{`Y0EPw>
z0~o*r4?N-WiTsZMEM`#kgz=)Sajxi2tQ}wor5*zGcRr8yT`Wh~xj9-k?7jDm&$du6
zBD|lQ1CIb7-=V#;bNUv9JAz;Cie8-};NtVSE>=odb>+9k9spTyoF{6j1W3u4<#cN;
zm~JFot;ct2=X{Gb%bY&F6nx(8vNf6Usk1MkAqf6>>MfSG44i7rc4xXe$!F8WsM7Rg
z!mO$JG?+D2Z2w42iJ42M>gRzOu7s8P)>Hhq=ELn*r&ShSEbf%ARzgJge(~vwn>L^5
zEW6Enuzu|LPR4VNb^ecMh^GcTPqOewe5RoLJhvwJV&!P8!o-1{e$6)VVmChs<rDO9
z$LB#TkxG*@amxoOSeHB-vNsIt;G*tl?By#c)Ib>{ll5Z7y{QO&t|nIKqDdY>K;`|T
zn{eqZgy7ObTMB~%vgnzFz<S99F~j&ZLAXS#)+2OXF2Q}_MB@K$>>Y#+a<M6*egp7Q
zLYM@|)KqrNY3FA3M65gOw@UT;{px8DSBJfaV{H@X>Xy7}J?yxyURh4-ip*%O@y+al
zBy=Ai&RAAG>at$)s`@69IB5zXxeGy~Z4(}T7YBKq^EDcu(+j`6N*!H4`V3WPa>d-h
z{ZPr!J_yA5<?|oz4P{Y1bBR|CFY0I$(30FM>iEx1(2OqzKy43+=-eK?T`A(Hp;z??
zj4Q0~_dJ(^<m;WW`}A5%!7$Z&>M$m2P>|WBLxVXs0kW1#3@ss|b93O<1`7$1G|8Qv
zecYEQ5Yi9@^CqUXrDtv)v%mr66KpHO<igES^x9q(v3!$BH}+5McFGtb6dv*32@Ft;
z0>ic=mk{)YPCz~O0X`xh1W|PwJmPCwQdNu;Lv<Phh&IIayq-`B_WL4GHK&{sV90+>
z4%@yF!tc1Nr+E+X2A=ulMo*E&9>f-+kjsSx@v09Ug?>jL(n3xtU*Ge=^nL~zgCqX~
zjvk$tDl--ev%gf=U>7EPe%Q|c&fOW@Bn&1NpO|=|Cu5r1b9kUBBoz9-c0VtR3Waxf
zKR#(t%6S33)mHJ^xlni>PwqC%4}&?&jfOWg3kB{MU`K@DUGtW__V2Ds?ZRVFhio{e
z@SQOTS?uZ*6+e|Itz8P<VUuUm2uCZ3XyvG;h~X?sll2`s=tI?eUu6F0=kq;>i^sPC
zR%9u4Pk=a&&Hri27ey@3ROP%GjM?hcFRIOVi_{aouR?O;Rc(*e5FqhOy(Nzlvw&h_
zs(SYvwPvGW8Ef8fN=Kq=GgZ%7n4SHkN*a|NNp1T$CPjo<;F-v@s=>uZl#*na52#9r
z{d7-toef_3x@VoaqYCoU=G{Po!77dp`c!M&X~mb+cfH@nf+*3oQ#m$g!pq_p&M|TD
zyjaF5qY)n7(H{hsh+!K;C=LO3auQbg7b|QEr`=X%HkX@bY0QhKR&%y$Rz-=w_)G`P
z>pjN`<Bg}LWX#{>GqDvUOqJ-uTYI??9A<ss2bY(Z=VIL}5?Blp4lG-{Q0qeIom%|Z
zFRfpGj+f|=8G8R=7ABCqvVU*l;*1Qyrh9UTbK5NN+*E;bzpv<Kb^h_@k@39@#g(2t
z5#F4~l;CjsAva2VzQE`5Yzg1K{bzDvRCsD~qW(;0*Qbq6l}hodyZ8S}#AgMFDzZO6
zih20H>A_+E`<f<V*$KmXLg*_os4#j=`FJdk@?E`7_TY{+Gfj2~cgyLj6+k2F;x<Gd
z*E}WC_M8a3&t}(`5VcZYT)e#v^J1ADS<4K;UaOURTce>j*nExCTK3E6O;EpXEt*X{
z+@kiY&0KSvBRK0TS;ew;uMs!^i|2I-E4a<OX!am0%~&<`dUky-+eTfvbSi$m?FF)}
zvbu}J#_u-o%M6%gLMi9bE&tm5s6xwqCQ9>n$`G{2<GOat5;wC>4P9Z>nk6EIE*Jq&
zG><&DNf293*HT|@^aADU$0KeW%y%|QMrN@>*-XM20(;()6*y7S6;L<C=l`pK-Ur+M
z6mmy%=eFMi>wf0tVxQZ|rKgb{pyK!7P2Jt1QJ1e}9vk4JwBcflKMV{lNdQP}0%n+N
zpu@y;i{2WcYHVVw5!d9X4P&?<JS3#D2%?fU-*ngsN9>pbSB?+@I$>h;6>5<AdVgfl
z;Z#2U?-qb%(P{~?jM|TRHzGf$xbDnpl6!9rRa^Z|ahs0;B0WtH4pC9>`B{%JiTKNp
zm5A$|*pq{Uuo?r<4WTVd(VW1c^CsNGBV-2YRVa-R7m~Zli%c-w?vo+nqtxi>v-j;B
z`t?;24IO2M2!_L@cU-%wPfg;9$ajCOB^JT)>sKNh0W&9><7O=>KB25J(V3BZ;+I#e
zxY-ka1$*jGg#?sO4&hEmswW=hRWXp#;j+Q|&bNCPCfFj|?!d7e=b-B10NK$PX9pMo
zxE2yr-wUlaBErUcFB5_yV6ZEjwzdAD5DKnYUh}`{74-dIDGan;c%SXhDgtJSr0bP?
zd+~%R^Ypu&3?XpL?NIFkde2~D%#hDo%eHqVx-ynSd-!ZWqSyyu485`>If@>?zVf}~
zKP!u)xmj60Zy0sq$on^;7xgT}WRew^oWE|T=D@}ckisN5W)Td*r}JKA<(}dEp~Eqf
z#j~^ZV{`@oxTs$Iv~LZ(Kh@FE;gu7em+ed=u+_<@?*9fWE*8vJwe0C`$ACqgfv3f9
zagyCXyKV*x1D6kJh)=JVn*}xTn}+wU39vJcZ9^<&HJiql4T^mldXc3uz}QF%hM;^M
zDi8=Od;!W$LrCTEjL;FG*MH}Ru~0fU+HR{$h870^p0$cJ(-b0ipo-<?<vBw0_C7V9
zy6UQr$!Sc=o6r}<YS^y=sj)qecQrl)F|B~T;wd7e4kWQ2$dVgDwX$VnBWqv)wjf@L
z!I-C#B@?61Cm9_>cD{RqUU!8KZg%&G80Amt)><;(;;py!JsFayKShXJF)aJF_H{1B
z=`U<pvieI7KGdJ<t)0(pK}-}}7h*LMzQV$h-I9sO9CyCiZU0{68W>h7DH#NFzmXx3
zeaYpX|3#PAH>!iYxN~8D*+QLWhIX_Z%LH`cKd%nx>c=oZWg*uP=f4)y8Y1;H5DC7V
z+0HB}P|x>+{_>uyeH*ZrRNL?t|FrR!ggt(#RKJ&;v-b;kgWJw`XWtW0j%gh}!0lPW
zZTqJ0(`ZE4S)i2Gj5RQEn|n2Yll_z~oK6L}ku&Vu^OZuDpu`vgz98)TMuOf)L$W4J
zKa4nsYKYo`);%rVI{rTEDLeNAB_0p>GL1defiWGhdKslB4uy4daPJZ>cXYD^v8|><
zT!tuk>TQB4P^9&ZrNQ0HO6>hl{DJT@c*mpuo+-KiIb@zECt33jlKH=nvak}5;TyzV
zUVrfLC;p9K;2B2CO@b;{>E;f9pfX~0jAq(^j7A`57AegK5NLmp8M4pWxGZ-n^zbPZ
zh`dfc3GHC)!F(w!T(>Iq)faf39Lm36KNTG`F3E^qiLHVC+#e=x;B+2Y3(Nb25Yb;o
zOYBbO;gUazch&9wo1y6{d~?+-91x@z;j|nklxjXwU)?{lG*POpCiwjgRy`gA@MQ^W
zlYK~y-t~Ka&mD~UHI_^&j(sJ3bNBVpoapcF^VJg;pCsf@0mMz#bAXG|&nflF9q=o_
zdF4*JZgZB~%QA`~uwKytzT3(sfUB2J%|ihYsSk1<X@~~vdl@oZa|LoQH>~L+4EnHr
zngXYc!2Zbj&bewvMIz&)pI&OMR_1ICn-#@{mvy&+CHIXauc&Cl)lEt)XVhG-huZ7c
zuYdgb+pfO3rud2O^zueS>KopG-`c{6tJ3v4Q;GU?<A2+ro15@jk()+|I=_Z<yROlV
zobLtu;WWJ`!i#F^>gwFRr+1I;wi3SHIm@{yP3}>w-%4am8QvMq4#{&A|NX2`_R=2=
z1+&eAH99LbMTrQDEGTRAn5FtdlXe*F3UEH|v_Ru-bTxdw{Xa~71yohf7w$nMq(K^K
zP(VsLq>&WqZjeq%>8sKqAtBu@-Hjm9pjS#7=~g-f-dz0s-+Qx`vev!l&Y9fv?Y+Of
zXB7IRIvI0hHHQ?5_D!SWRSvE68|GDw07_)3&^M4M6q6<P>*f$(vqyo!uStaO-78DJ
znSMCW{REg6QeU(ST6rvUbQlny{umYOSzB2R3OKw9e|`LV?|dDT7_e~Jn5%TpX`<l^
zqo>~=GCsTBpS#)pZ7DKruvVdb$1kbh5W3j9Um+=yBbP5aW2><FN=}b5rsAD6zLDGb
zuY=|ns~kLma${0+3C=#ny^9Dt(D(!Qc~uNF`vMzseEi^E;rh^`3wyqn?CbP|E1yGl
za|nN2A{dV!>9vk=;YTrEUQNS7omk7djDvkqHW9<|lij4)<a1nz9_i^z^s4^&gWVr<
zE>fZ2A<OM;9u|k$LUna@{_U|pRO+h$sgEa?&ow9D!p@d@sdL9esXPt)2L~}l68&0t
zn!S2oq~Gv>#5HTt_%BdUP^jqTKswa2gJvJe4_{BfErpcoxX1p086KO-p+}e*`$szB
zIF;|&X-B2;`tdfuV`S)=AEcTzl8zDZ*FtkP8%U!R9;M0#vE@kA%Ne0>5;;CcVUtdv
zJt4q}k9Mc^I;EOQJWdhH=Mpg}D624^@hUGF&o`D$eOMt61_nhBu+tSl#=5pC%Nkx0
zZ=T1r3bk?IoN9G_*B7m@d!ll@KWJG~e5WH5#MnmMd=1^I2ssW-&wS2YQD*=6$0x1h
zhQ$^DMTMn8iG+^RQf_MBNeel&HrwW_wRm#pzg)uyYyjIyV^6=mw)^{a#VMj3ZAWH@
zXb+Wz$SEj%2D}d8z~iiW{B~^niTT$Dp6Q8kY4#Ce;nnLTL_q*m;yP}hR_IjG*&bJH
z`<Cx6<ocZcqMB>5Dl=$;SDSp#2{iOMeMZ?YL@>|7n7^|YA-vakxDnkkORQ~H6Eu&o
zYXApRU0b&QGO*RVOjMp{QK5v!b3j;kzV%1cOFbUaw)&{XuF;hA=hE<nc0^y_$L6sB
zpg?=CShUQ|bMW&w44$O7jL6B!^}P?PCkzY>Jdou3ZsRPVAa#g|ky}$!bN{8PY5Sbx
z!o5YJJ@p*#VciJN<Y<cX1RMgdPpjLxFa6YapQ3_|Rm2W*89!*4bvgV_V(e8Hq_FD&
z&|PiD<x)t?<B4l8z?s%)6RDMZ&PwIC9QEk2qh6+{-$DDu#kLFuztejFh<cus)5n3N
zyi#AIC??CR=fHpsJv9FsDM)tT)pjmFEiLUE$VAb!L+%00wbOLW#lCKPAHU~<R81=%
zg7iIiDvRRX03Y?a|8>(kj>eAIfQ%qf@Qcjr=Rb;X{#qrrY`Pl)x6Ri3k2s=)MxP!N
zL$B2wQfZ}YnJwP5_)|z4Y;a9)(7=qh;4)lu4}eDMyfN1@xujXDC1#RRudS$vNf4fI
z|A2rz=Q^`I=^0PhouV<-EvEXJUD;ia76tV`u#3~L$l@9mf)&bV<R4=%ei~2mV;ohV
z@rMA-FY*yrrckZ(QkRTv{Cn~;SH}g>SkKv4R$2{!txyw9jhia$=?H>0u4?50Os#0+
zv-8wBD_#n{2XKd)1l{pPFoolW^G%+72&&;z&e7VRlGlfsBm}3s3)DZ0{VuCb-nU#G
zoGowme|V_u=-S^ylKFr$E5N@2tZKlZeYsF=q4-ipM&@8CS9b8aqJ6XAiEKA^uCFN3
zj@QjQTo0TLv8)Jk>|>CemDjl5UUzml;Buzw@bZL>iHQYcuGQwl83-se?=YzUq#8Di
zrIgMUBta*pQ=*1j0-3@l^>d#y^OH*!y-i#%+UJE8ihSZ2Up;Oqh<7i8_U52ZBmBUH
zS~o)hRvzjn&wV31?J_-)X~%}>+L>M3JTBQYf^*q+DiN^?@SSjza@dxL=3yZbY){nD
zn*YjoOj+1EwR~wd%~+5w=VDvVlnGcJS|?`;r}+=ty*i(T$@NxOBs#VI3a6#ybDAS>
zbG!=vc<=tB>&*rm>DR9}clhV0yZ}A0sZ5v0eoE7;A&SEDtPYTli~!C$`SKx|S@~(Q
zF5fUZrsq(apn30Kkj@5>0+?A~HD?j|N@Lw~N3=<X{(r?r#R3#8xf9~{TJ}3;+~(XT
z>lc7O+hS%UJ+T)K*K5%aCJ!Sd_*G@EmQD;fZ~@R?%Fg^?JgZc-@-o@W1N<v<t~orj
zbmmpF$*^EN8}?&g7o?|}!gEjAIFQd&I82gz%;Zbn&4`1e#c2a`>IV>9S6HwqZvnkj
zrf=bz4u%snbhauQ!MN^EtUuC9+o=z<X*#shN>hv*6^2~?R2JSANpE*Vhh3f%G0>F+
zfb3plO%HtqziS}Tt^0m1(Nw73o#eIS8~rGli-<^{dfDwn!RiG}LO@jJGc))1*w}2M
zRkWN;rwugU94K;73RE~Pyd~o{Qktza{m@ejNZG%Zhh?dfy?+-xF?Z33>HE<Na<dPw
zjgoMLU)uiEeg!V~c|RbUNLfBXeV@{t!udOF7SJs$jRoDTGYZcUM2Lf#ifz8O%Xyza
zhifAa)!xTj8$clr#&;CZtG8G6T}!qiGP;<QXv&q1>VL&gP$G){O5l$0#~!?24<T2m
zsBex_eZbhxRg7Gcr!>ZDHb;4YkXK{vF#PZ}i=_=ve%XDt@Aoi5{<EYxr6{OUhs
zlGWOc=_v?lg7Gdtr0^-tcLW3&&STk{RAan1S(G&&<nulc{V5k&ekEY#?~hcV98j=x
z46fA0zklCw+wVZeaF<`T67YgP#{ysZN41O63?wjK=bd=JWD+)|ZnTUmw$M}6y#Hp~
zp?_CcpqZ8$<({VQ*vF;~Y!A*#yo_O9<$n6!9k(JOh65y-i>fkrV#_o?*%1UeGRllk
zO=%`x36&iBSdW*Cvkaf)$l?F2Q}2B-Lq~^|N>YO#8xcpq?YmuHMj2u8=OJ$;V^%hY
z6v)+ZvgKy4K1RtOjF-6q8!$>rYo|(9^Iu+fEPk9yFY>%bgKIdQ$1syg8Z8)A+d9_;
zOZn3_(tqkl&P8YZfrwOSKQJ9Y!w;oG$)Gi4o$XJ8_UPZKIt?Mm$w_<-m~}#j|3a#o
zf>da!JG8|tg`VKCcTaub{@)zXdsQWjqqCyb3u%lC-Lil{|JU8Bn0_ZI%&S(xcqDg7
z(Zo@{=#_M);lPUTw<Q<Ex4f;p(R>o95_@6UtlQWYuPcCnmEh1AQJ-XJ|4|plDm16B
zJHRi6uI`jp1Fx$g{VWP~U)3!I<2h|4r6Rd$BA+15nxp?oV3A+yp?gpezkH_20J>v5
zF4LKB=3b~*cJfl8X(rQ;GygAD=jpYcQA`sBZ2V__S67#Jb22bffik87%pKf;XP6N0
z+l>^ARLUhw{vuTw&lEMP<7M{dx-O-<Z*{1TEt!h3-XWe}sR4`<xL4>_Z4#1i5kF#k
z7GJ$(M|=fJp*)(e45_3Wnlm@d9}>D_otM-qRhSeo6YnC0z3b%3U+L|mZfAn|0zr^y
z!~aGT%M22KyMJW1NG!REUI^)$>5`P9xL~Kfos}GKX*5{DSQl~K@+o3uY^A<Uoc{X?
zC=Te)nNNR<BPw!ORKfdi8WDeTP($g%`s{-7!l0DRlcf{5+TPRu_jQg?i59{;Tmr4Q
zH`BaCgOuE=o@wG8a6{B7Y8fj@kF961JGXY8o0^(REl~gKmUJN-;GbRBCYjc){gy+}
zO)G2}<zAwjLT4nfdr;0S&nCvA0yGLV0BG@vx<U<cvqHd0HOB0wa-uG4BEudv-6%Nw
z35oj^W>(e?Ji$o{nk@workCUbkG|o1p3(%MAr#S8ZaQcLa7(@Z66U+^93R}6<~eo8
z{|;~(mEz#leWP|juEg&&<rR1qPykO!$IQA93xQ@|qW>vkUVT7yIBl+wj3O>F_Nq20
zs5*~G)dH;7izBPXXvEls_EEHgdv!JHeBF)NA5<n6*hov0KF|?gp-&xlEDDgaxJiWl
z1q3uKT6VPG!SB^5Q8?l82bT)dnYboYu$NE;h2#TZm?r2Gy^nw1efDP=pG4aW^`2t<
zVmlyzg6a4;e`Hq52I(PfpwLJf>nG#VL@$uh*)n;b?rB!rj&B|3=bx;!{{8!L!y0qj
z6z^Q_YKEDujs()FxK={(X82Be^(jX~>N)eI#>7;-OFr!v6-ew3`j10yISobNL_)8y
z^xg$?2}f<f4J}bb2?cDDrkmTIZ4eg3o43t94#v9&Q6V{3a2s|oud)LUm#w&e<&5HG
z`cto*XJ$VF%@G(^PJSgHNRPiO{rCTP_-xWbk5w^w!!9<34Mi(h#8>jB$s}3tI`hIX
z@4yugRvaO%_1Z2;)UE#%_&fdaToj#GEHfCeM@YUvG)%@<AG`%N>Eqr)kaKx5iIAlC
zfq+}1msw3gFVNx0nCMCZyH&|kKvKwbP*l_m(zNyTnyEYQj7i26HTfIuF8kU6a#Jzt
zh4j1+S!YP|F;_%S*D~-YAc}@?0TL-%78*<#dOEr+E{!b#uscqLUfl)KkJ>ixr<Xb|
zFGl)3Af2b_MK1a8)nu#Y9K~!KW}=Z^i(lzWvfT|!!=K}%UU#jxl5QkUR<lG>!R2k8
zBPJf`X#4jD>AKJHa8vXu^iDA-(hMY>CyxV{hLjoB*&bV`4vY3!)|~$O{OCYI6iTRb
zFamG)5VArC>;f9tduZhJm8jLF6z{i3|81|F=8e!jS2QXSI$P_pXL}%^XyFw|xFrC^
z{=I{zSwbp~sNe-zDBzNRpM->v8o%I~WihWFlvr{GHj;MI7!Gc1ce){XuZTiUAQego
zbwg>`fI<~XvKawH+&kM>{DMC^Wyj5^fjm%(AReY<)8iRk!lx-|2ZCu0YQz=~W&b2F
z5%1Do_%#@>8);}_m-2!avEe|8H71q(vZdj`f2s4M%d0F$i&V0x^P61T9`av4gu})C
zE`l4=1jqZ;aq&elJ#ZLsdNCL2p)Emu{#~3!c4)mZ&8@9$ZC&+t(SvRWa0YThtSH}P
zjRT7AW6@3{f4oQY%$QrZi{mbMpQdb<EP+HcgDt8_rGG%Zf8+8~34JURU$&zJ15!zb
zE%U;_8AWAzV?b*}6pf@J3XRJxam8#ZRUX_<E%kxF;tnCSUnq=tFuybZ)*!#3V$URj
zUN@t*$b!d@bvhAV21ckH3?E5nH;emA#$jBf*I?rYK_0PBTY$6|tgEbyX$cRFN!54X
zGFFTix7H!`Bu0z{N-BkWfPIO7HS?-+-B^dd-*dzD)?zJ-mre8WF2>l$rSsf@uSq(Y
zpl&5*F~^Ndz{)4jm?o9i_!@s87L3P%+(;-c22mpMz(06GI*>^+bAHsqU{?0~$&hC5
zK~`c#HOdWc{P9ZN8cWTlw?NOOfA}yZ<Sk-m-zsTSrkxiN0|VcDL5n*7Lb-Nb22!h;
zLh^3U)0JS_FXKYu8g_O0;89!sGL*32K3RPQ>iukM{IqBP7O^{>1t#Wke#RqJXF<K!
zwFB1-p?3p=#<w!%Z<k_;F246>$M}K2<N7m-kQZ#8@OU^COmrL^3QABb#-cV9MR9Qr
zVn2c-w8g@TB$^%nNNGBr9N^c)k`5!e?iC%qvUO9j&>$Cbk|9G37XEln1(E^IV&wEr
zAoD7f|Ghhp*hJbsjKo(v*xGJk0}l6N@_~AA7-@c?!N=z%P5f;=3%4*yNY%F|UTmWj
zJB(1hr)eCzxs2$35!D$QtPXN~68QRiybBxd);<<(98wzQA_x1Z($R`YLA-;XK?V!&
zZ%+srV%V<r@<H-NUF20CnB%QwCc4a)!Dd5?!Oe0?8E4_<;zq8r1!rEChPectR4D(E
zBCy#tVq|K5t>E7K9~VITygD@0m*_s(yO7Z`@4Kp6pYdB8%vn!45Is{Bu~KLn#`}c>
zpro%E`jesfgGrMq;Ae`US{;)pOc6#h5r$<dF=N`f%-L*imiQl$T<av#N|VCYq31;^
zp8Yv?>X|5uMpB_Z&zTpFdWA}VKZUJ9?0ot0NIKM@ewEVxSm}nwRu5DU$787{;;8=#
z)6k7tGD)m&xw;vsoX=29uxLe*Dy0LAEbxGDjv|2So)kI}k30%zSGPMwD(x2oNEwg{
zO0%8bv>a0*T#*qZNVCN=rFN2vNPp4AA<{dG&)t=&VLw3|#R_%HDD13L$J*NnI)@?4
zk81vg>e~Q0lqlMA$*r^s=b5Q6eNobPN_T|)uQ7c#O@no)A?+`?B*j{ney`UNNrhe_
z70p`7Fit9u)lwiu@&tY~Q#d!kelSc&IduLEW#p)4dc?4u6$r6PMOcx}@y~ddY{d^D
z*tAq=zl)~l0cD>8{z<Deb+r{GD;l>{)I>qBsEcc?D4z^EK`s*N0TMzLS@i?BbJ)mV
z#yyZ$nhs{!+<*WqR6_qGYRNVWs`-ocwFz7G%?apfF0yh7E+1t3<(v1FC{pK7h<$9*
zl}^J)3jIg?0b{@kBZw-{6EP6uzI8e#E;UbKfZLpaACgS+=zH<8hRJ$74L)oF61uE3
zT3X{KdTw`bkwtyLfI1P-Y&9xCSTBfZi002!O?PS7?$<LljVciTPtcXvY<{i;UDMl)
z_xy&Gt!wKR#k`8o?=U``dIQJ;b0;<lbffVu0gFC+z-vnawYj}}n(i_30XX0`o=b(|
zsSk)*#68J>uo1#;h@5q|6FC?USL0)WE#cK5-Y-!oLWw!^sQUtJj;+9y2lCR#ql1v>
zn10=mSrx+{Qn0TO3)Rp<fs9as4)Wam&cPzURaZ>|h|#JEUBh^0NiUQh9sN~Do>^tV
zcD1Se63?~p1FNL`1MQb5kP_0-#OAK;TXH}5*pO9^$<Y?%YEST{x6<P8q?>8-s=MP4
zV7P~M&^LTu6}-4sD*&Tt6oH>q$C~Ms3ME0Ug)-=w=0pOUc^CmYHs&%C0%X(Wm^dfz
zb;~%(R((_xwLpJ@Dh`;}?>Ob$-VbgQ8PysWt9{-)|4YTdm2badlY-MY=cI66)iJ3k
zEls#{M<lLK7w<p<`r-FcnJCg%!Vr0F;}P9~G)gGYw7!;JaofOOe)%HpY7BghpyQm@
zr4>oOd*;TXlM*>wW0)m2hK7k9a}n->Gj3%$2-?!Qsv!l2+T$Mj|1D=*l}Mu_d1~$0
z+)Or!*Ul*|t)d&xe5}4EuW8mID63*6cDXJOjB06fyuZ%yxHL5hu2$@4a3ayErE&Rz
z;PJoS;5-BerhYIU3*%3B=mKd2$_i^Iu0rm#DtS@CoL}+gfm4x?6dO|QEehD7d7ATt
zy#lk<+PEmw+^$H>-yFSap}RUIDgL8QgQH=k(;bvI01(_1-v_SYXO52L#=^v*?rAv3
z`$3_$#)kI@dbZcQSf+tET2UkyENq+-008x`&}}PJ6958?4hcZ|2+<5E`6S%gy#?Dv
zZfDWSY8QuF1>r0V$fS<we`1&|5+talLhp_np;f(o!m`v%tq5AQW7!Vl%(o}ma0law
z646Y+K>jK7blgP@A&o&Pq5^sty>=qf#_qEepM;5BS}2KB(!f?BdA(5N&l4E!#lxSt
zfFQt^KXf`ag(n#WKZnHW!A;Xl=+ScckQMC|I>)qZ^AO3{b7u9~W92B$@KDaUQK9Wm
zCJT2l?!0>SDqu?*<f`~qUluSJfnr8k_>TXS040PRlOR98u%W+isI{kiW$@^pl10~1
z<=!<kGWJMh;kS>JM#t+93M0Hxv$ueo#Ms8ChjUAhPw7@VlrY{P|0Zd(b>B6mg?T{}
zkU#;l?A$R^?&D*h@oPg!`@DG2)$R8t7KDss&lULE7MxV6Hd=OGGy`~ay3thLI)#O<
z1zoB41C29$lW0`kh#Z#GwEK<U@Pq0PB$bSwX`0%UrjCyHj~O3#>(dI2&(=IGl0J^8
zC|=QkF+Kd6x<GOZTci^JQh+*%pdBb96VN1lQ^vR0Bl?jbML_IV2O%vtNQ?fwnJFOB
zVxh1lRdcHuL^NR{LS621DxkqKM*D_*AJTLG@S04^mt?7zVo~Ac<#pFf8Mnpkl4=9C
zZ2z*i_N67(3*8o?plKK}DOd!>1J+K>t=exg^p6bLJD@3*p1=7*)D9<OV`IBe`|}z-
zTvYo}Jk?HcQ6pxssu%H{_y4JgV$DAUz5C)1-joohTGs6;*XU6Lk{`5a3wa_5>;iv{
zN21Y*l#sJ~&5MqOvHYL@&;s&nicUIp1zEfRvUZ45u0xBya=s+zN+szWtQ{tCTlTrM
z@n(eVZP^B@{>YK*eyZAwb%}>`+Q?`|4+wJu6?XG2Mw?K)Dg{agBJb(s+?DL^y7LWq
zzYqgmv*!J()X!jt9K^dG<diA9Ez0J%>VHhkph}EQ-ozqE=gem<mNJH+KpA5d9Q=Wy
zeEF=-li%=uNkYSJB~8Y6X25dD31<7&uo68!D#m|AsVx2ZvGI*#^6@mQ&Q2oBr67<d
z1)YdpPgjzv;9lX^wmtZ>y*AvV5Hz|DoRzX$zcPSB&Kf=Yt<8GT&{A{3j~`Sig<I<j
zZ++befdFc|k5JWDQ;m3XX)MrVi}rU^yLR(=z1g2@SgowBHSU%aj^0WhChMZw6H27C
zee;mCUEDt^T3h%cEDG}cfjR(WOu3^<h4Q2uDeDqHH3T|?Eve4zJKtCQP-%^ysVS`|
z)wmshBvsV=;pvt+X(A|v*G}uF-8wfy4&jdr8Y6GqT?8t@n}Ybk(rHL4G{U`%G9+Vx
z1B<qK8(%@NAl|2SLzV*ivz1|>KA?0x`v)L_L~sa)^9CCHVCCTs{HXs*RDHniajP+i
zP`KZ@3k8+{7F^Zc&ghG_lBO|re$VGhN`4EOGYIWV6B7iW;^oA$ix<7{zP3|1|3q^5
z{EqW|D8bV+rtohaAwa-qDizki0P@FY&mzZ!fWpC)To5tHITsE%uo);2I-9~X`^3EJ
ztE_L2{^UWY{oH)EaV8+6*Fc}f2V2oP9~-eADnK`?hegl_=2c=bs$z7A{&BKnv03wD
z_wHiKPDAew><L`$Expk%Jf-gM7U~i1x>kaXSZV${8)eGg<)<zgqtIXdQCDObzingm
z${S`2V(3`qQDuGQH<djW+e%i|oKJw0sqxRvj=^3K{}`M08gtRoJ(Cm$qB~i14r6hq
zUvA@gPD;FYHMj$AxIe$Dl^C&OAb>R#UueSY$YF9&?iiG1^4B|Atpxf$7mT+Lk)oME
zu*3_?#Wo?zx##Kwh7#WlLB~-a0e95k(P2eF{7hT&=XjS=+AkH*m?5teDFZk$@yLn4
zs><BTjN@*H=Ie4-x51wmOD{sNpO24|?e~^-^d+~C#KTKx_gH>_U{eK8PeIt5>!$=Z
zTHF0RrMzDv3XHh_8M6!?f<=+uE{=;LsGoUs=3O|f;77N%$pA8B+?GBopcoazYeS8=
zowR;epMOXx-2h~)qOR~G+NY()|9ZUlJZ-n-tcn9PXAE}Nn?cp}%D1!k{A7Eb*nbcA
zcF}?J4}<us7f*8hSKI%T9Un_H7)psKI04I0PZC{49{Hv1O9!1JR=+O+h>?~OyKJ7X
zv>ibxXo#{xL<#AAqJ9&POJgMFyJT-XME@TVJZta?p!^A)31rq>RgrJp$LAFX$1oMY
zP(^1c;o`qvU%SYm`CgngKDD}}e@Xy8)1Z8VLr`qcRHpGn4nTDN9#U*t4U4nKM@)H9
zZ#Ey|k9*cw4j(iI1(^nk+xes1js=rRJMuT8POxRl3*u8t^r~p3`5}_cZN|s1i@*a#
zLXya>lcJNO_aLsev~~G8BV7%ttCCYJ+IikU=Z#)oY`=ipV;G#BvAAy98qHUy*tF!@
z6TDKP4@=S&y{}KbI6?nGsDVh_h7uHwBzM<2npT<thNi@P$il}H7>dG)-RW>q{?Xq3
zP%OF{s5|AToC^EBG0hVJbOBYj6r-krd`kAGfyzm<5g)7HMBa8^gPK(GlZ#_8vxU>&
z>z}qs7yaYE3?^*Y&68=rNJC^$GxC%L`k(Qj^~{D}Id4d)zxzdw^bOhHTV#bukX0s?
zT#$5t?s{s*i>R#--8x1tKS@zhnffCL&XqyK9!JX3$EM{&7C_Gu=Xbza<bdd)_!4|+
z;7U~o!SudVl7Jo+iS(Nb)_W}f@&@5048Xv~lzRSr{7q(K%vAOIw@Pu``hRec-0i(l
zt)0ro&JTk*8<dX*z`zb<i+D7PT3v8Cg1N07&~Sh@X<5$ie~rE<qdp*Vy@d=`8Sm3K
zRPupfcPplh3(ew5rl(zDcH;Sw4CxZ32)TJT$tp2V9<*!%XKN|cioIx~t^oGbGj|7y
z+o=T{qP+Cfp2o{h6Tptk8z?~C2^33h-te(epQLNTtqUn`$QMNWG|&(s-t`;_mp~Ax
zQL4GX&=-e<b@y9O{z(kEF3)KEfjG@Y&ZRRu%oNhFt$*kZ2kg{(QT9@5SjYB#!Ma%j
z=>KV<HAIC+xmBVzvjxE!XlCiRa|IaT19PFNCJ%ygk;SM>LTIJcA*&7fpBF+wWr-~a
z1IMu(Q03FyXg1QP^*X@j_B*Z2Ow3o?U@PMMJ98&1mWtWEg7yp0ew{AN{WM=274Ei7
z#jBy3`2AP^ifloAFZ5tp6!1)PQq;C3a%gB++?E;#<>-v_H=)?(ZL8uDTvj)QHg2RZ
z&_@S!RiJP+6!6XA{ZfGVOJjkl#29VU|370@iDob!^3jFoS={g!yc`tR`!kW4L+ofe
zFihi`|Ce71`QD+y>|{440g5qcr!ho`{QvPwOD`0JLV1Ax0_ZK4+B#qLn+5VYfHwrk
zMOTJC-O^DE7V^Tp<G-yl?m2iIl>wOad)@W#q2G64YDW4%Bh6|4^f|K+s1SP>^LSl_
z<kIY1M}xH(z&Yhj1)_zKvIdd9Y(p=4cs$FH+)iG#5y?u5pk*<_N}n^kme$ZpU#r<R
z<W?3P*oOewA<@ziO-yxMV)#~0Tu0~KjVpqSDm{i5^t%XQCpXk8)#k;fWk+$Qg}rEY
zRPKSvKEs(*<^J4-mxBp2@+SNqk>Iv{kCH*{E2-r3?<mZJpf5+=5C1=f&7<D`I*Im+
zA4Ht0CQ4?1`C$u}h$cn!Dp{ySWRmT~k#qjmhLWzae06K6&&gU|r2W@Tt+&AHF6GAA
zlcXw4E8%;<`&_K_!PoN)^5UmGz#PB$QXK-$mu();Yht9u9EHT6hGP5WZ`{?97l+Yl
z@18T=)<>7~1Fyo0ZRG{nbUDqeiwK)XD$>w*r&997-6DFahEb3vqt&`aE&r`+@@Rzw
zo%(=GMgg$tPJYf^Du}1PbZr}}JQI@qkC$9goz7TN@;UHLHoUj^yF`L&@kT2pu#6+h
z@rX&A#Jm9-4LH#;^pLO#$RukaPAN41a47H2YwAuVs9@90#WWa?66ra}eg%_DFHjo4
zV8(`1{?rtsKF!(n;6KD4w__0}1U@q>KGks&vkQTL@PHB6j+8)X9vc<k7ME1$M;1N-
zEb;pwvqoC6ll@h&@qO@E$sj_SIvRS)rOVCPE1cu|<5FD}0}#dsor2gQ7+eifN8&-{
zd?I)1?s}EB{%i)>agDyDrj?A5kd;30Ryq_(s59@EIeNgFV~;B?mdfzbZTEf#x2q$B
zwIhhBwT|(>9wwtHiDX^{9<L?HUj+&L^#TCvopZGCJ|D}UoHD!iUsL2;(ZF!JQ}I%;
zUpCr;o{^F9BmW!Q(C#I1e+h}7U4W?jCMKpih6vuFK$+ji9l#;saT4!}M^2y@z&E0N
zCPG`1a?wF3PUi`PkZR{^A?D-NmzoxWs*?{9bmi4-08PPU0_cF_O{m%<9^HN3g-16V
z4`1^IN;Al?z~v`9x-d6HftBM%y<>EgKeg(v!MOW(B;F;9I%$(BsM}*@s_K)tBQi>+
zBs0Zd)mS5o6#FUg-UB(5(C2$ppn4dv<C-NA6?cBqzDW5}3W*lw+~2y{F-VY^kAM0Z
zAyF=9K?7ti&Q*d&4Ur#}ILp5X+7#QQ`F_qj+p)z0zrr~*yOv9ueD=AHEWi>3TT1?0
zCK?=hx@Cd@piGv}n4$lv@^AVQe_%&t1E$c{8W2!e;Ke^9F)~|FgW~3ZIi5b^ZC$W(
zI}X9K*WV|89ZaW^|HDJzWJi9^EbH6>G*ne7;te|T(j(VXjLd-TI9p(|xCgMT?a}vf
z?=w#iljhBuWA9)~qoJwp)#D3=>Ho0G?Vs_FARz048qy?LN0{)KEol8aT&<#MzxY6o
zsGQ&N%s619NJn>KfR!38%o1%=dkGVI2e}>O?0&;I{I4QaHqJonynxiwqdzyi`M&if
zMnD_@=nEroXX}iwhV_C3MWI_Ih!+vU3@2Bc%DZYovY;WQAG!AP@Xhsj5hw%GIiRgh
zykXk}9e>w`N-7l7P=<o??r+6!_6SEYD5~AA^ssrzyZLW@#efwwf(nivinH`vS;3lx
z*tqnN400nyXEw%<6C_eB@hCcs;@>Jrxr5J{A5XAgV1VEvq)*>JSKLcDV*W?0cyBTb
zrW#RlP=9f?C4*>r$5Ts?ZKt8+<YE`kUljp%*fHIj1>L!hO&Ndq1I6#oB939pdSkiv
zQ4_O;jfu||H!p!LEzC_S)V&}+!jQxSJr;El<Lcz5#`DqQD+7z^)BQw^gFkfo^UG}Y
zH6h}CGU$Vt_ukyC?quNPHZ{33G^F@3(DWlVCo01e(vb5tlwgsV7!=a3$`2?6sMVFh
zI}9&h?M%J*l&!ZIOpMlyyJ*RPGdGP*II_)c-Rw7Clo*wimMXhH&|J8BY89QQOmu#E
zo9E1=RhY51;bx9zVSQin_&c+&B)E8inc_I6NUZ7<jls{c)`-JpWV*vo&~-9z==f)>
z>%wE-7duz_-i`YweNAXm$<2lxQgFXGgQ6JkS32X@XTF|0e#@b)w71E)I8|f=kNL1&
zzVHRFK9wPdp!jN4>@d3Zd6M$_FZ0Iw49k|T=xt3&y$_H7=qPVV;Vv7X2OAkzzE`q5
z$E`YDAZwpP+*1WXQt|a~sCQ236Qq2pKnqpL#s$L|{IuX!4C&ldVyIQ8d|s=M84WpX
zqVlfd?5Xm584TGrpH-bsl6)@zcf}a<?!l`enqE4?Om@(=th}fb4x<!CrP!mFRakmE
zSIsKt+y*n3QF#kklP9;Y><snNfYWW?B@pCiK55VIAPG-PeX#bRo+a;*01v1kKFWjf
z`	vHT8c9zrIc>cMp!n*Sf%3T-<dQW=D`fs^on8TB#i4;F5Ii^Re4nVNX1A^Bhb`
z<|zbE6Z)D`GCcbC8hRNG$x1>!<X95V6T!3uzpR}ZHl*LYQ4r#H**luLL=&{X{jmwF
zj!#ox)2ZBr|Cws3pVc4jSXZ%Fw~oAZDH^x%ywgD9i+ZDwAkc3gLFd~h+A^x{Tv0Rn
z>8R!B>C>+p^%wpd&&Y(SOKQdDuJT~0?`>9t!bNSKliWo?Ucm5O=`vcH3o0~hFj>v2
zWenkIJ_&hZf#GzQ$U7N*O+He%i-nB4I2~U%HFvJse7F9Mm2+5I2fNE2t$b$8Q7|4V
zQFVCE41*uf$u&MIH3UVX<<2?o?fsTb=z?EwcxyIdN&91+klp??(zWUoSFyRWJ^S+-
zgC3ntVxIhxA-{rmPJ4NI@+cS$Qsu974FBekNr4vKJ2DrH_rT|(VCkyUiY4{`Wd?*e
zPgrDZaknVQA4-oXMWsaujNI0b_^-!;;c=KK!=`!MwGzD0NAI&_p0j(vj^B{gzp0A^
zGcl>)ou9YNFs|TowWA<<a&3S^3PDVoxgw8Lvmai*E2*HlcUzwc!?Te#2{;J@8`+n^
zl`c=Ps3AtXPf~QZ-@=~U7d)Meb40Jm+_n-sK-K}no%J$uDoy(KYp93S#W5q#PV)CM
zW1BI&_1hom?DRB0PnXub(I9sr2EQ)<*Aza+{Ya=_=|sjLZjl1hQ2MyLRZ5HYXtL@w
zGoo!lOMtlQbo9%-KFlpRk1Z}y@q$B?7rbT+AKaS*3e3gb@%2J<&-_hR$CJ>u05v!x
ztwuAK%bvZcAYxJczsH`V`|h#7zBxZqSTSfgPcz@Ajya0>mb2${wgReAlL40)Bg%lL
z!|MfojQN{4L&yChk;Bwo2Nu3A2th&r-^8ri2E1NS{d~EQ-qfBIQe}fVNdK=2EnbXE
zQ4yEDK2x1X_&ztYa{C~|IVen+6pZOKHXff-{T{krVaY2KKJ|Y}4X%CRq>7Z0J6dwy
zRC0v2^A2)rM{l1u`Fok0cb~DXaxq<$rT7_KT0Y>>X2H>$z9z;v_InGePrI!^TjQ(0
z+wv!Euf~l&NBub?)bti{2OTKE-ARn@a^4mzvt*kmX8Ub<rCqbq6`^UDHY>C1^y}@9
zHcjR4Gw{mmeN_cd=i4)sx=biGdwy(4rtTEFtpU0UZ|eY?u`JFqP%&xhl7fCcBhIZR
z@35QkxG_)Bd|#V;;C935$Y~gBd%iO~bG!W=%#6tAW^Uet?c}<>PcXG+Bm7&OV_}}L
z{DsYq9~v4FXtU7T>AzNPJlkojOS^oq3>M{AddYtaN+l5?E$2NIcdqJuF~FEN6)%R~
zu38w^v95e(tQ!v4V35jtba2BODB?}FIQy6KqDafeE3??KuIEPV$_bZeU=r`|f=efE
zpO|Ld6p$bHy)!W2#8Afmx?6w1PX&Tj!~XAjGgc5Xf+r&$sH-~VkH4e=9>3|~e~XLD
zBPx$ieqClHeFP1B5>N4e8)~+H!?f8Q$HF*AE_kXS3bspD+BZkX2Q~M|7~7=au5lhG
zzh6xD%p0riEh83I*3(xbfqhNd@)uG?OwE#TkyzJLg^S`mAYKy3zE6tx%r6L2p>er5
zzR63H6QZw4!>2N$_UU*TGvm{g*Ae^D+;PSc`(}5JWz(+ebX#UE{b(btujvDj>9tDm
zsH)Q-S3)F6g%~|-K#H=>1JU{{?KJL$iWbJz?v@4%_CV4K^2N}V-eu3n$Jp2v2}g%Z
zQ(7x&nl|8E(KPi{{{VzRb>*60tm}SMTgNK_BUgx>LuWO-ugMkc6pH*wG$_mZn<EaW
z7>q-W!CKDS8RVu&SrHUt5zQIwFC+|Y{E@jqww`hqjd?dQ@xfeL-(CIuq`EO>Ub9kH
zi7UY6mw49|o8{D6@@}xl)j=ghEdO^F7yC}FJMx4XMQAp!D{J>-CU7k2V_hqfjxwAN
zVR`($^9zZ(tM6i6iAnv&B=d+KYhPRV85$bm&NRuxli$7GRH`~%C!Qzsx%v6F<wU-j
z&83{Z_Q7B0qSqj;vppU`6zuuJJP@<$_(wZzJ9U?Kk13+t_G>D*5S@tx(~=bEns1i8
z9_U)Y<vcp)sXgyLDsmzsXJD+I^k8PnM>2<MY;t_gJZT5C>fMD?O?$R!VqQf^DlwiR
zp3lu-*Rk2Y7_p#LbbusWzwF9<NoX*Mt!BLp2+e(}p>}rT`klhZTIIE+0rEh~F!2XU
zceA?e*<LaZIotEP>`BTuU`dV6cNw*;h4nQJZZu$mW|{C{PYV9v9ChsyR3~fP&hQbf
zHd`HtmcMAYXl_<MJ$g^SUJzV*X}-VgLS55cWwxq->9iiCGz(W3j}LD%#uq&Gm+u7&
z7YHKplRw}$u^HT~=}ES&^|*Jz+8C1O`D^f*IApqJatt@m%4ua*-`mUIsm+5r;pooA
zF5(CxMecSV;aT*0Az;--!y`Y<xJgY~+wEa%iQGZi`PH8!vGAm3UQU47{_y(AYX`86
z2@|BdovaF+_M;gcsO?9MD0gOjzKVOFC^YlA?3GVlQvP%}`fhi+u4Tiq$@<?rEIdZs
zqy&t~6A-Vt^Yu$uGKrxGj8bk=uqyL7n89VQ77x{r|75;!CN!-k;N3j6qknKT29!06
zM}S>3bs25dATX8DM)tUC{NnA0bm@(sb|3gMIBZh#aItDFq;~x5R%QpB)jU31M<NCl
zvOXJ0)_Pg)I&zn=ke<t@&I|HsWM+-tuxiT|QiChSiBb!mX5|6Fr{Jw>9dY`3>QeZO
z6$u)^rtMX)E&LxBKovhcO;`eMvqbXBc*7i5jO@khp&7b$Q9_?g>P2C17S{FX($5IF
z#aSNI|63Vzzsieq{rS5z%rMu7i@%@uKlNFF>c<EYdFgrfCS(pe7<6CTFSNAM)t1aM
zZ4OLbR$#eF#kx}Xcv*iwj1uOt8kt?1v5V9{_cda>)Nh`jS2S`daA{gfC${4;T6P9;
zO1&8(h1MSEOzMHsJ)2Rc6%$|P_jes|&G+}C((K}2K9Zk!bb0ix+?2*L*xh?h6fqKn
z*~{E5ZoMWW$pA&HO-_zu=jBMGaU+E|m{}s1Ax}-csZ5(1^_C*=>EQdJ<(Ri39Pcz!
zXLWk=>fn!jX><MiKmPGS6Kf*KeyLKT?C)MjeK$i=HLw>(0YWn8SNscjqbcNLbzj)%
zL3yT7Xv_vjSDZW<LN3L}**`9M^CCDigPZb6XTA2Xw0i>fmyAj{6vb4XUmpgpUYk*0
zOkD;&tE8Z3FZ;-nbgIuHzzzlLqtNTc!{y06aZJ{$KZtoK?ddFh<K%GJtDs*u@UTdY
zdE&B{$=Jp*+mjcG=%7OX&QLp!C6}*7rrn1rFyoO$_C)h5NUh)UD(%-Q!Q^-sF|fSY
z<XA>M2>D(xbiBi{RdD=-HurdJk8~q1((h<mb4n{U?~j@SHq@yjH_iHG9(M&Z{uQIk
zh(BBFJ7r3hv9&<gUaH5=Ul)^xetNw?<tFoAa=lFX;;*v)9P+MJQWFx4zT!VN5$mp{
zVQqK<?$C;)LJUquBNo_lifi5;F1=-b&--4t<#BwSKRrd87GGhBkFXp~X<Gcvj{=E%
zRo{Cr9CxFXl$l02oXi{1K@X$T3cQiAn9;(v<oesSe!?4e<($EtefB&$O}IL<jl54N
zZ0d4d;H!wYB)os=O)grcfAh%w+{S1y@FIwJQzgAg3#+pz_oN&JLZJZsGfj@EsxU^?
zONd30rr<x(v<K(aMYZ#iq=TJ3>tM-j*vRfo<E!N{x+x;44gE&OfhDI2x7y*y45<4g
zt$4dBmnTdl{6|Q4K7SL^rw;>%a|#tyv2MDrf3L=g&Wv7gU=GTDH8DnlnDmF_HPTYL
zjFd<_=Pf!kermEK-=2;3=R*|Ka(!ury{A=F1B&HIilT?_hUX{xiJrdCn2T#V{oB`M
zG_xf$(#ADJX6bNBZSoWeA{7CPoq4xIpBvUTSZ_s^(B2m7>Pol%qB`QrtfVd1oX3b5
zu?{pi$*%t*MarF|**H;eXFok6*&#%g@GA=u1*meSF{GQEc*1DBcWbcgpla&!VCOR?
z)JaMUR))>Z+$Qo=i@ou}_pKS9{Jh?73r9u8K`7%EpLhQ{ACf`k9u33Dh~QE;)}F-k
zF@7?Xq#!EmcOkd-IT(O!lJW<F+8mp2oexROI|k)+Q@_*M@`l5f@0c{=6A%^F-t6Td
z?aU2wP0E%!?I$s`aT?Tvo5yY9|Fyf8m-{XUb*J;o(d7D*Dk@l3dv^9imp$d55qZMg
z%iJBm|5U!<uR3i|^!|`{$tWn_AUh&*U(i!~WwwHDy<V>5!x+n^?9F9OreB6RxEFH6
z%%$hoOFy5AM7PM}ov5~j$sb0R-X|#_=Y0c^uMa2YsZO%Xs^5M>WZCX4(oOKnK16dJ
z0RXQdx%)0wsHR+7Ry0wrKh-{g&&_>#v9L6spuCQRBNllM?9bm^;~Zm-#-}dp9sm_V
zhI(O-BpV#L_mFpf+Fkq9UYu|}LR1T@QZv>*{foCI_~d3-0vWXcs6EUlqWn{e#^6A{
zQQEm$_%cs@-$xASpnNCSD^r&iccsCmp#%#z;bfOmZggHGPLwx%KTz-%Iz;Aw9&zQG
z2Ad6FLHVq?S<Z;G8ZA`g5qn-?)6t#2rhTGzvs?Xr^x{aAY4hk?%_)xjX%JPr$DN~J
zl_i?EZNew*uLu_khUJguoUcg0-3%$h;T1p~gt7v$dj-`6j4w*kF)h{vnG0MWB0gmR
z=T1j|-2CjI)}vcBzWc9qlXIN<*|xxT%hjzKKg{3vPXof~c24Kk4SDAwA7E_%uaU(x
z3@m0~(t^XZeuPo3TRrq>jS8c@6^o@)KA559k!hF2w~uyy`**^*FKOF%cRlooKT7?>
zbt}0CDE`&;hlZgm>)FU~mC@3Ym*y@dCDuO_4VuXq-2al3aN_dvk8|=29NtB9yJW!^
zDrV?moP@@4M_tZ8`E1|}sEBDn6c~Y2iy*=_;34l(HR0-zK=d-|-#0*TkpIsz0hx%Q
zdc?X~$&M&Jm!)?p?_!U{fg+VP2u#1xt*^<YP-bY4&wu~?-|OO;FyEN%?_y?XD>SRa
zM{jkHkgC(D2a~_>1d}4#MseM(=+<{GTa)8b7#*V8oT#m&?wX*V5kbMf6y&2r24@o!
ziT`V1mfLDq1~o<Ey26FBe8aO5KOUF8$iAk20A`z1bmNzGqTkySO{yU04Drn~ZT^UB
zXWm4~=f;3!2oj`8!#U(H9AD3KfZ63qQbmXypp+>khsV{$`EzV)$n(Y-5wIsIT!iI~
z_RQ}uaz?-JYtoIIClYL68NzkcG*r^MN9eQfGZWSoo~y!WITC+#;hb*lhb*^pa89eu
zbmv4YJ<8Q;Sg#V~ZbU_fV6-5QXLda=pCbTbI*#T7W%VAXdj$o>&nh27-l`)vGDzZ0
zRSMa^6W7ainhuV>`S0b&y}zf9yYs;<BF~r&z{Lh)mhs7PO1Y5!3}a*M-lV8*0EIH8
zo+Y6O5o7G)kYI+N$W2>1EO+&ki*@+^<kH{6*q*7fDXl2o1~{psse`2COMUPBC<V#8
zyNi|h1WT28Kwow~Lk~j_gP9?8-Ox0Bmd#OytBAbwWF%d-O-oE_pwsA_AoC8<tvli-
zh?SXsDtoL^md70T@c6)%7*B-W#;<WI$wOIDk8X2;fzqOv_2jxj!NUZ?di<Zh1VI}U
z1*}8D@xGp<K28oBqCg&%=HbeDH|6zYclEsu;g=n`QiJ1~?eCj0ghbUL)oQ&OkDDiP
zq^ABl@84=VE?y(){8Fqnh1JH}`mv{_<)aiSUU<Co<!1t_`%w4)Yy=PtW38%c`<1t_
zOF7rcbuus#DR`IW0ZEvUcQ5V*RuYK<9RoRPtgB{rkijznoao;9nO5o3qQgJoczA5+
z*{a_1Qh7F}M#cT6x*r5Ha1#!uAX1}1Y0}!hCc$jnvSQo)4HG=XrUWQLXJs}Ak7*}{
zM-^Vh#~+)=!zx7QviB@IrDyHoG?io|imCsD(wAO|qn(vu_V2SRt7POKO??chEEhRx
zr+<Ll@-(g8dS!N~a_`qSv%Nl*S8j=UWTth!ik><U7QZANR!>6Q7JR*~wwCk1l>UrN
zKg{;H<6gwRJjC^qSEA(^7@_`&&!S)5kf6J+?TZ>LyK{Bj*Hm9wl{iG|d0FRSjcJ|K
zO4$p<$GQJ~_Dvz;6~9Nd_h#?=-2DB$$A++M1ZJPb<?RQrUzx%4<Xy?0-S!LyH~%cT
zk~Zw+YBP0yq9V`pM5FR4eu>^EjMeD43Aq%8{yUND3}v{k^HZ$?VypC=pIExTzQ@^J
zjz^W;j0l8rWcq#o7S66aS@K>|vGd+i`y)l$yHKP6=#$<?fGx>o<_mp`XDILf*>zOR
zxHP37fxL%wL%N69^3ihB)()d)`=^0gmRey1uAgL$cCJF`*F~u%b2OJw4zJprt)C@f
zN4&pBo8eW-Ba7TANBf`o1sYI64;?a+_Y7>ZCsy!&lQ+mPoaCne(Y0@4!Su5f&}oV5
z@<)}sLwV9Ve}}a#(Ag+9e;#^5frXiD8GKB=_8aQUxqG*>aq|23>?kFQ_)*-|eWnNO
z6sHptAO9c>hA=u{#)!<rWNg!h4`&bmRAjtL%`;ZGFk6{r1mZ@^Ys@r3`e!BGtgN`k
zORnB8df?pEyE|{Zw|pl{7#Ikar!$a!+~#pP;1o_Rl9Qs|Y-m35bq#ek<0V$LTgjUJ
zaK~GlM!2H2u$9E9L)B>j+a{CzJ}y#h(h;%zf^n70>M#(p(j)nYGYq4v&ctmuv2F=@
z4e}T2O}MDAJ!^h66N6L&rSE_HYwDU-y@p-51*UR8<y5PUkB^5h3+PS*Z48`<ulj4F
z6xW!-+ZG4~PZ{M;6`IB2#zDkxPKi9qaAmN|-?uFYd^xcq&BOb;T1}*HBR;tRY-mn#
zN8*i;IKSZZ&7!mPh(nsN-Nf4{3VyElOnMWJj)xgadYlf~E`t7xJkvWj+^=Q}m`cq4
zzA<7oXjGXYr{M3I=Kx?gMN7ZM$5Z{sKiJ5$9*_%V8RVO)p6tqw=;|E3k}nKu+4-}t
zc;O`P`E|bG_Zu>Q5z2^c^h24|d&dxAP8|9>VnKdc_|EUvk`o-gk(@?N4GnA4UAp&h
z(khdpW+9fr(_3pqeo+Xli0<SXSm<RXRotrblH=H8kGSLkHc5K0vt-pv2k^Rvdo?A7
zJn`42-)-fF3v8QUWASOn4^z|uo!VgOf$9($N)}MdNy8sAj^~fs*(`<pcb4Wexs<cq
z73HOM4rQcd>-TriWwWWsBa3xaWE@wRrY(N*7uIM9zawAqYZ12t7`-UA=z#?4-T3&M
z&S8m^CwRHa2(BLM8jY|4lu!AHWqd{I?8}*ZfLfLnw?R^4IrOo`;?NS602~_N)7>>7
zoLdAT(oDK_7W4fX`gMQO5{;6sQxTh>a6cRCfWkVx5AU7Bq+)nzHR^WaYK`NDl{I>0
zEncU&j{J=9ziL^SNmE%@EWzqEQEYTHU%-E_Oz5bL-aFit%NyrUvza(`DHa*>2V8(q
z1i(Yjyi?LklFN5FWz{)vC8Sm5)~Qm&+$G><IOfskx8f*EKv$g5W2oevmSPI)rl7o?
zl*NFk{QYhhKXhsX^dPOCc_7l2a2GiEUVRK1tq3^142W|xEh22zf&(J}SYKHq`^!6x
zLkT>_a5MQwazLxTveJ|t+$9jSI(*hf@A?U85id|0`ig=XI`6gKa@2V^x*hqm#=YBX
z@;+hI6U@+wGW`m_%M)-1M8973%O?`IfQ1l0)lLq6`ALECWd6>*mQlT-pJ{w(7c)Pm
zE)5@mn_24jeUx{uaN#8Ay^pDVoFwL&ozETCUfIce2IU1{FD)3$Pe0r=rdxLja#cct
zNhUKv@lQ)>s$tgNb#eJ~TIkVyd=7Gnq+Y~Wb@gBSt8?2#?uD{Ee*7U{6Ij_m376gR
z17g^m7%vFw+Ow;J^*QP*3SpAKDMyTYL#wy4oD?n=IQJ5SUbh}>R-yM6(TT8EoeC#%
z=%0AWXX2DW4s#t5X+FG2_^^3!Fq4ve>PgeY+qKoKuAQ-Q|L)ab5X+kU&f#T#f>fzi
znL;8_LV`Sss*jIPb#0<6q5-g9ara1@ubR2;672qTev@Uo@@zz@x)@l{@O?UD=k$?1
zs?71<tID1<0>c$=?>sO}4SXZ7lz~6~Iwy`q%I9kiL`da?%Y+%b@S|8wJY7vyhcR8l
z+S5*>@kVl-Gjsq?x&~<Np7j2MvMQ~tCVlmEnAxh+6Or`~a-FgnIBdctBbq)_p_M)j
zZ@U>qcxr?c;5Svj0eH9Y#lnXzNi*WZ;>^Zb+UlZfO#1H7VTSMg@!~z?$ycxqPhfqd
zt-NLuO@sLsK0v#gS7$fLqMRy*=o^zQZQuq*sAQ0WFw*yrRWXhYc_jJ5G4-hwOv3)L
zuJ;+QM4wuSB8E;y{%%&@-;GUb>+`EgMz+;YE`4vaHKvV5N`iO(1c+@fPtQCY*H9cT
zgSSuk874^AJJxTZhm5iW5mPtG*<~1(BBS9jy*C$W@knnym4E#*x~|SsMe`OPKKs&@
z)z_rX*ro%wqhDWBtmd;)4jKFqH=$@-dmb*}LH{VwO<8tiH{MNRWmXIfuy-hj0Pm3e
zOAo?JA2d!6V4)_=pO;K4^c2h898cd}-uV+$7tu^rGvg6((O6Z<gdJ7zT#ls<_udDL
z?uS_KH9hm$R$1>>!}@e)1TjP5C$2v``=-C)->$dLDU9)l6LCB5;?|~^T{+T2cBJ=z
z0{0Zd98p(pRP`FA#slosp;^i3K!u!W*%3<mO>g<=0@`%VfS_=CqIwjpvNe00+z$-t
z1bhCK3}|;dix%}&vGi7E5hmmvGYNCy-8|0ctgTgAF1x!9@}aKtY}yQpE$Q_ih5e_i
zrpKqru62Z)?6(D5{C-8Ito%{88iwZ-N5nxJBM}wyi8f5Cqz`}YesUA@wzUb_W2Mn+
zaBKrH<_WCA2b3uSHfGf$L%Q4_wkQ0{y}JS(-^2zf=+>uCObF=SswFZ;=;ZTkn>Rr)
z>xw*wII5Uhk-IXY1`p+gp4?SfxJUa{NWS4!QE-yxeFQT0G5z~U-aCn*^x;JzXgyI$
z2S6WE(|3ZkTq7K=@CbD^;dA+7T#$^ydyW-Oomf{Ay?(pb)2lyhZ3}pUOR!CfJTB+F
z56hg!*RJ3OlTER6W$@aqcav5En-l!kDp8m9k2EEmnywtCH-+~DChs}@;EfJYk19)-
zh;dlnH$HoDb$&0vU#=(os>RA-`x8sQWFbGLr#w7-r2gmZ^e089woi5~VvQS?O6i;l
z%hM4j^?V{VmQmBw(>TOcRvOPV2Rv7Rjx_IO)qdFULQ_*y=wu^TQ&%&=-uyVlim0I{
zyo8R@^+>h<N?cI)|5qc>{Pi4m_vVz8{z_pIR=HgBV?bYXFM@*&fF)eiWzSbWGw$W*
z?o0+h0EuUS@jTA0F0Y2+N&)Wl2|Jy48~t<7_3v)i%b-Q%s#B{22{N#iG82y4srXvy
z1Km8lp4KAIk_ySNVe}pXa8LIB$0P$=bej|F@e?FYt}|kscf>vc`WK3^eNFsL_p(Ds
zoJmuB)d}#D9gLcHb2wNAG#SSw_Rr22tY;3RQuwN58p_+BQTr@z3pt#yf{<~=RDE*Z
z`~jODsf7UzJx|Bn>o*if2VNZ86LBw0_tF$!ymoaFneiSei57jOB{k`>uAEWEj8DDa
zuA!l2D|vjkjz|MFzxq^oGWlBfS*%EJMrmc#*4EZb9D4nN(_1(tGnT37ZHn$cPwjIP
z@!RYbI&6r<Y`=p-ut5#3$Y4mKH+sbth`+Fg7e!in7kFMCGqwp0J$Y#f{LTi|_Eg0W
zFX!xw{;HwJkfVn}M?o>@;;NMs&cuJ89{HMA8$5M*#igtQKVVzX6}(z`Z|V1HVe$jt
ztNIKSdFr#>kfXx%;S`NzZAG8c@NAOj74mC5TZ+@(<$Q5ej`m|g4gZI#w~mVP`@Vn?
zK`H4{dZa^?kZzC?kQll<1f+!_L>iUulx~n5X{Eb+D5;?aqy~v$-iOck{jK%B|G1Ve
z*23rBbI;vp?|sfSO>Vo>;JtIN(-;BIh*b<YIxOZk#2$QU%rg369^CYPHKEeUVshg-
zl^6Kq;?xei3S7{#`acLlEuyPH{yNaRRG@awHfW&jdn0EiRdtOnoG3h#hdMf|zS&KZ
zb&o7T@|O<fbG8u;PsOd}rFVl~@r9`>8LPuz)JL(DzvW>bmM`3l7D-ZKyrD$LJ*3u@
z7(K3s|H_fuxv71*Q+(#4v_sT#;$^1MMPt;TLL_)}*yJM0rm0X3!j0q>;3S9~DKWhp
zl#|WrLe&xvaz<usjP*}rlxtC7$xA4~3HR+?=Q9My9f5hGi<s5)qq&{Z2xo+{w6zDR
z0gI*)1pcoC#y>kt^5DPQ`|Ia{W<zDgu=4L~VZ`iS@};`3H0&-}N0O(X?F^(7Lpmci
z1?w!2^?Fx>IFNUrKa5n;twUG>UU**n4z--CE<O94=fq<23=$NDE&sZI@*_P1MntR%
z2qS{$i3j(G!AD$+x-OLa%O6%R22ro!v5oHau*C~Ri*vyb6s(NjcK1(qwoH6gw(b=*
z?C4?pNEBEzF2sm&a{6w|^W~2^j3Lgb>k4U3#n*ucb5+Ncv_pIAwDPF8&QRNn{jp6=
zgFGbCTP;Uiq>OsVW&Eh#5%fP%keZ@DoNDpuPQ9{xc0d;oW1ZysYX*}qVFY>Qz2$K^
z`nB#wjRKiu<whOdf1VjIZPoG$pVnSdK0=AQ9|`?|n25cmTw3(=UUmdfLQ@{Vj7hAN
zoi99)b_K2y+LWBn-v$0D8#Nnv#qa=QVMH6IffJQ;jF&PeR5g!GaTY~HPG1Z(DT?N5
zCm%bay@5PkZL+*ETQ$^gdOQb)?#f-7c?`gm=9T3tp9wWqG#a-^75~t3=rusAQ)rsY
z#54w1lvhp!1~KA#av$#5A5%}9wQePe+wDyi0i2{I=_3PdO+yoFRR~=L>wjtc>(8(R
zWY}N;&$(uRoZPj?FOfgJ3>B{SA8H#zZ~#r=5DOJNaeYV;D*v1}R2;ELg?HJDeTj%`
zu<~z*U=#$$;rz1L<G?Z0hZ$^JJ$N@SE>~~MJ_g+@=tP(LE7~=VwPpCW+{O&I%4Xng
z&*mBPbxu?_jYTIY`ARgOna|$bMyWJe{Z0UrZJ*J#Sd2mmeetVhGo)++>sC+Kf&>;H
zSG?t$GaRcfnU(ck!Q`;IN!I1KD1?@mov<72B|egq!;75p&9E5H-evU+Q};fs{LEg(
zT(|$<q5D<(Q9Yq?TOZ`m+P)s0%vG>(^Dy3nzl)F-5ZZ<mF1)<P<;TYAD1ZxT#Bh?1
zjNWWG5A)MhzmPiyp{i&}Y#8_AI!iK<!l=gKa=H$7`hM2=vEA$4yecA}ijCS&_A1E@
zLS)3f7G0Uj6uym~Tb?HgXRW~3Oy_*e)jgK{-bHmmC(?dl=Q>Z;D*BKWzxF+Yvjfkz
z<W*hBtgKM>r5DP=15TT5Hzsu!hgT>-ykFtyC|>6q(E=t6bNn7$o^4u@l7e!SnDF&K
zY_7;P$y`>4invY515XCHDND;&0M-aV-xG`t|1^+g3!QGVNys5c<|<M0t1Yy%bd|bH
z!dBuPYbb}Lbz@@`#oq6fY9|cZ9Ps*ZF-~Vg|D)xDFWCIr6@?PckIj1T>TU;)rSc<u
z4)@>2Pd(W>1J3=d_!Tn7nm&KiiOoUoOtP^7zQtK2aq?T_6+W!TYj9hDDOCd#<YZ(A
zR;DCqx8D(vXtMGl{mWfx@Exeh6C8_uZf{)^RGx>G=euSF$+@FW#nu)Lumxd9ttWBi
z9FUs1YnL`%3b2l@H~5<Q(PHmU2R-deG538+YJ+zToi;&mxqIdq5P34Vv050hWDwXj
zkb=TX)eO+y&OmcbnDoGPNT=9{o;4IbK*A*%?~%xYAo@*@asK80YirkgK9f6<*yWAV
zxLn;-bJE~z9JXL{8c>!nKZxhUO>B&;7-6Z}U4p^R&yU7FSp;}|@rS5`huDT;xi=>0
z*39t(=~C&sPrEZu`8~J4tQHWhT`ISm`+cj1AR7JV$I-7Bd0Vq}X`bRYPBkY^tq-J(
zK1O;%q$p{5&?~1~5+%Q<KH^1J&Srs0c_iKA=fqoe9~072|Ej-+*<WT-l26+HSNgs&
zOPk)Y`=hmMpp0hPGh<OpK8GfF!yH-*2Q6L{B-dWH-@21oG2OFdtC*cZDGfaCV=;^E
zwU<W1j2PElLLL~83QOjkydh&-BIP6SL=^@rupDK+_|I*bY{}zuW^O<4`Dg7X$e=^N
z^f^;bEpEm-xKzuTId(M4x8Bn|Z0_hl$<*#R_~^G*Y}aN}Aml{}ith1w(D9YE`;G5f
zWki$Y0+=KTOFrGPSldorW3>Q`>y<dnatvDq^!J|yYE-Lpi^#FhEJi_Yhz++B3l%uX
z(b;)+gXK#E8{@tL1oDSy?tT|B3YbRt3%jDp@Z&)6yXXF{|LB!Hm`4eSV7Cffz$M==
z!)bWjSbg*7U%M<R6Uiq@!1l8XeuibB+SG8$T1Bb?-w`P;7R9yHn$xY8^=;ow^%`uM
z&$$7kE~nR?O)$c8OBS0v_W}|t91|FL77vWwHm+Lk%ESntWK4Z$Bl}l)bD=Zc;R2OH
zFHp7YXx1Ts-&=N>eM_WFZH?Gf&(bVl^Lfp7FsC!m?8F>9`(o<4_+k?oLe}q`&9mdK
zHGGG9er>pV!~8YlrA1d{$A6pmy>)ddUSVvz>sJ0E-a}6naAXxQkN&WwslD{!OX_9Q
zA>qK1FS8pCFvd6sj-YfbH=|0(_e9&7T@A5JFpZZkt_^(0U}bp;u@`?DoqS+;ndFK#
zKqpI;WqFQPNR<X%U(A{$D<jzsHd}pss@z8tea=spvx;>%(x*$M=0HbmRJW<#SH;XD
zDvMc)b|YfXeUB-l4WzA)QQ`9@B|CHX2)5D5B-3ZmaIDCKauIQoUtR}F=N(x3?Y#Qa
zVc@vj+Yga<MtbqRs>)<`)^~^H{$BV*LW#*(L9U|yOBUAbGIb++!^gW9B=Vw&$$iWa
z^Dee9b#b1y_7XM}@D}6T^>(y@&|Kd-gb_cITiPKy8SB-V(+<n5V5?yf)p)7e($69w
z=3q@n2^`JAkLB|kmnkm&%eff$)sPoov9)YnOJzCz#chsRE~tEN3)IsgDq0BZn9Lo|
zHM3j&<YaI}+hu3VCZgqhZaCpDyA6ecP_|FVQQN-^YmuCt=Z;RtN|^cGf^nY)x=Mip
z)NwS!E>R@zzkuuJrNNsO63|(wEBVCErJbG-nN@(@L#-6MBCTJFRudoiKQ90k1dUv{
zDPU>;`86E(mjxp;=EU>N)P78I{LhsceRl$tWk&>-$fib58>uSC<2e`g^d=X<$r+NT
z%ZCSaU^15Aq>A0UkFVK9ecx17i`_i;+ipv2J*vO7@*qRq7%<EOsIe~a!{jA7*sEA=
z@$&BTUN%@qter5ugxy&}C<hZ{!x<G`-U%;`<<<Z3o$G5ILkUpNn(#4@-hYSr$FTjp
zR7If0{@y&IFox3JS_XWrji~l_tmdM`B0_+E9z7t^LiR1L>s{%~k^1JsuI;$L14s3E
zKmNA$CC~y#??peAJ9XtiGd_z6agO&@Lq-P2w{fX);S}~8A@k$l_6$A?mD5L42Xoh4
zVr-LQNB(wG#fSz2{_{E!yS*<O%hlysGbzWF%NzvJm58jUvu8Sjnw@+egin&d&kW*{
z>hSXBN2Js`VR!FH??ygM5xr<>HWF+0G-lP)m{W^gD;*cVmdcQLRdQ^0?#xn?Gt6>E
z{w<Z#0U218Py%qC-^9EfVRKc_Peu9d#&nDzt(BO8{V{JEdmppqMdvH^(*`7QRtV2a
zU;Y28DcAq2rk<ZwU}ALVXQ97gemY@vcd0JCRBBO~;y!rks*R_*#J*>$ocpcaEV~1Y
z&xi--iq2})pK?gJn^Lc^I)<P8p*D5kl7`s>2xuF?32e%upsVO!R>u@TeCaRvsfsa#
z#YFOm$K55SaPg0CXPTf@-!f085PuvCyV!9Yz>BnX?<~}EW*nILnDFt&N_cOKJd<cH
zXdV|5ew8|Z=kI{UIxA8pSo1O>KH(AqkYbmHB-We%k;}`<kNXD}(gHSbr(6Z?&<sF_
zcsmjSOAQsu<|$Fs6@_cg^(d2h%}f=IsD;+aE6FQik@EvkmEAes!z}*%Xe4vMXAEz7
zn?yOaceO_47!ZUQ_dY7SCIXT!a(Jrw&!1My<MC~)SGSc~s#ALjcfX%~2>rm$@VOn8
zaHzMWq2gyerJaxicr;yq=ZIEB5O(^Af2!*qZ+vAB9B;%t{{C&Qp3%0jQzKZt&f{2o
zdEu%5f>-PDJqVJRGn4j|MOaJBHq=<Z8N<xy+-`*Ol+c(>f1i#K7JQ2D#6PdSZgpu;
zG+B*mJr6g!>o^Fc;fLfjW^viLEvI0~Cj*R_v<?N>tK#odZxNOUny=aG=2zG1q!v^6
z_e%@=$?l^+T45_IOxzW>)v6Q6n2s}Y$H)iy0nL$cH~YD8p!tf4+Ry1{=fn$T^zZ|E
z)4YIOFT~aN!qkh}6X8$=Z`wJf7B=^z6hK$GZUH}k@JspxA^ENk_!NEch(o;{%RE97
zUwsP3Hj*|4XGCod>g|12U`w8nZsJHC5%L^b*7jT+Uj<65W>8sqU3p%PK39u^HWiCi
zLfxzSSG0JLYyj!_1+-6+D}Xb^ee0GIy*((N(-3pqK2muc0s`pB`KEv$0srxh1es8y
z6YqtEeyd{*(~kFxnE=zJ4wti0X~kBK_i&9n2P@~|J<#-Xx6kX*xOamuFWe;4=2%=p
z6I_6QH!7udj6Z*i6Y{J(09m0@(PhTHCz$Yzd4)>kauYBm5)(h5##<Ww^W*=^_Q$c5
zx3}|@J*#`io?5Y6UivC^&CQpONHZu|R#UR(XOCfMws654p<J|@iFZHjVry*2ny^S;
zWX1PHFhKW_L213kc=%RJ*c&+q!Bpq7d9#YvNpm#A-^YfHkV}_ueYckQl-v`*Wb+oz
zs2B-k+E%^Y1Cpwpch&Ynimf8w(<TA0p$d?#;uCbk7^2YZ$UELjOBQ6uL27Wk2g4JY
z5ZN%zYgbfmmq#bZ-m3~@K<W9dXz>s5EB|MdC+7^U<lL2FcCL3w0s3zW2LUhZrpNoA
z+_;zEaPz%dVywH5ZQn>f#rJA71@qZjD2N`vd!}3XU=X;u0sMOH#G>c;ApYu{>rRg5
z^8H&(5X;3<v-MHYDrX3icA<b{GKqP_z_?ZTX={h!%moW<XFTQkWZl?j2B(!dy}t?7
z^?@iq;K^WfeX~wIkwLYrG@T~~U&M@b_4LRH@3-_9Fqr^{nm+kF`G;4<vpVy9(1@88
z(?lykz3k9k@0{8Uzh9A>o1aCD9^IE$)g9F_W|>5t6mx@XJI@S(5NO>$*uR^Aam@$`
zEci6P2vvL?G>Lm}GTeWTpP`=YxhVqA3fMF8SwRhx(i3460OBb%{2GrRuRFR?&FR8=
zGf@;>dPV_-#?%DYmJbrk>r+g2Wnf=_5nG*ePDcAL&JI#j#H6ou0C%TAkh@)+Enx{M
zyYpG9WKb#3fnQLCZ<P^L64#>A(eC<uNp4l!J6kh3b*|DhyCDsjCcD`ZrvDJP>HZXA
zJVf^W*hoysc_Hpapjsin_^e#ZCFT4Mupa4V!tOmFFfyFTaKOT=nP3Vk8!V5yzAw!P
zDlP3@x3aQFTvN`a4^cv0##-}Aek=DH!!Qeg@2FqOg3M{_&DVj_pe!QC_3h47^IjWG
zOW0FN>ID??Um1VN!MIC6&60-*&kc&|BA$O$q~lSaY=pmQHD8Hq$5CHl?`onLVN-hj
zy8TH*uSC=5`VTfGsjs1)x~R17GORkQI2RxGWF=JX_uj7Dp&=FDV~D?ovopfvV8LIy
zz&U6*u-9!_@UiCWhPwTH7F9~J|B<zSL`1|P9ym-h-C2V5UA@aBkX$VJ5odnKlyIs#
zva_xEmUPY<)3D8HudkNJfP69w%hF7D#{Bi^>k2b9#+y)VOq$Hso}<}T58`TVL{noV
zY!TC>S7V)G-VFjQ58`XLt;Fbgn;(*wp*t-Q<CS{;ms<CLHbbkZ`qQqIS)kYHM~$j=
z?t%0_GvJFvT}*`IY#diGH+7x41M?23ZoT(2C4NwrheSuop{Nyzt;(eJHE#G6UcNv>
zolU<!g6e+Dz5?YgnX}3QvE=)&?Z;K$LaX@MjDymwr1lR}6OlHfZ!@oj+x~sW6ve;j
znzic|*Q!^AC}Dhk%^CH{t?-9s;h&LfxmDwqv7LbN11GqDt<!2A9O;9m*(!X>w)gme
zo;&@wf#|CeW1Hlh71-Q?D7bknotB3^!;l5xkXS7Rx4ORFG@H8T&Au*P)U+PO>3*Qx
z@<(^MRt&O9wLLpg`QB|qA*>sf*yNI#lIu4Bsq?z$^lwlPs!X8XaY3I6-&L<`z&-$v
zM&l0BDmgGF9oc#3<VOz;Sw2<2|3G$y{)m94>2;qsV~e+f8p?iMDYcia-J2dPYK@$L
zN(P_tXe4<PgY7YDao`G=atsbpA##;F8BZ}wu4s8NLS&S%zCL&8RRUC#SC(c&8YA0{
zQn{)ruZc8TVk+XApzjYZM#pmm>S2fm9wQscHL#n`+c{Pyz2gV*35rBVP;7a&(bCDR
zL61bviiROu)w2}(hu_VrZ_Ic)(Jd|>#7~NQm+k9`0i9FDbhetRgRNrKpfxaH3hCTK
zek%g$a_cW^FOWOPJ*c#==VZkXof3(x`%PK(pX+`7M;o(v^YeymF_HRLPHIRRWhdrr
z$|FelANl~`Y6e6_mObo0X;r{y0q*|(N45OPfAHh20DaS8<>ONZjDip#S19hxj;I=z
z3ihg`An2Tdl=%K?`RIuukn(UvD-(9V`r7RR7%#ywzeo^l2=jkwebv2l)B5CXWk}Vo
z%4pVpTN9Li`IyZ1t)_xxi4@}fTqkNr&g6{GhAX1(K^)j@1$`SLmlIM-y#TqFlMRiG
zl83>}7Ne1m<R$D^TQ$CY`{pR|;rqdF-!nMN?{|`Gc>vv8ZT_E&_k+G;m&PS4@&`$U
zz}4A^Gk!u$R936vl%sM`(s{{5$9t#40Q?9`c~mPvf{H3kIk8BPia_5^L7APO{<@V1
z01zd4JFMe0&eLa)2=~*ZypqlvpsqBUGU^uVU`F04ZcDOmegps3hc;$7mNkR@@=pd`
zjwd)HQpWVlolHHx(JwDO#gwS!Z^(p^84W7Wx4>?ZZm&djWs<Svul!H%)UzZN&{uUD
zdtI<<mGXj|h1p6nJ0+?1f5V0X7ZW0%r;z@!0+KaNw-a#2XHL2z=Z8y;k3;fa4O|1S
zWy~O&8nf$frUL3<iVTfOx`5zk{?=_3%UnFN9o<l<VYkCtr;-m8j3L;v*shc6p%pBs
z+z*3(KueN?a;%|!ga2aPF(CUPHstvHEsv|}4sxram8R1Fd__G>7)$XC+FI_;1FsW0
z;Zw(eZB8MkgF6{5@1`Kk*<{|)0Q>Qsm^2yip$mnPW#2CNOMH$00slp>7FFztO+#r~
zm@L!ekJ5xYT?Lq_mj_j<EjuV?RnPtVq>dgd8oW01oqe2qM_=?qcDJ(GZXcehmwfJ)
zDW$p8CRL)G><jQfI@yVms39~jZiwm2c2?ZT_r^KPi<K(ib7APaCW$pKxWMkYZwb{3
zt^7laS<w~Fh&Q#}o(eW$qDx{eqncY;)A<Vc7bE-?t#kFh(K#lq>WT~j<oDekztYy*
zaDROygLF#iJYL*>oN)IFOOZ`QR6?gZi(!Xle=z>fY0KXQxjFw(qM^S)hl`l|5d({0
z=nuQ_!ubeQM<;t0C{IJG5p7ICm_nmPHvC?#{o?kL8RL!SUchve4B#z9*ZQ;E?r%cB
zuKzQqc#*p5o_=>7<HI_itn<aw9;|I=6UQSY;ChG`^uF2m*mmW1X4K>u-GR8g6s2@C
zf(*t^@5DvvOIuq^R7R;_g3iiPgZRq$>fCDc-#g9Dli{yc`q_t1|9B+tJpu^OyQ1C!
zRdyEZf9iY#S=IN;hJOvHVuI*y%eNP-VwFFCFD+M@C1dr31RMvD+5?#4>+7*hj3j*g
z7|UY@K#>K!d}e*C^WppC6~cbs3S*U8psBf3E*l|ye1Or57~-02N~PLT5<w->6*$c`
z7&!6QR@pAp`?G_zXKG<*RZRs<bKmNLxhDx}a>TfDhBLZg*T**KYgF)~M+P*=_e!#H
z5i~T32Rp$$M9hetuRQ3Vs7aS`<<CW%tIJ+DJq~Q=EvVzlyRo0wFZHR<FFZOEjG;(3
zsx>5$$(i7lw}X|(co5u>-4iHk4jxmjh*v&{M|A(Ed_%;J_WV*;{T9DV(zF}={EoA<
z#bSaM?RP_5U##q~@g=fpsAKW~RKu%ZzGR>3Y|fT?R$;e7K7a}fXkUj55GRM%22*ZK
zN&o7+>fpfeB*07oX#KU9QqY<g%qARrGzmF`&O78gu}c)?(a2EFd|0=?d}{wHI=>??
zSVbQrq)W)HyYk*|dwbf&`apbjXN0KcM{7s*QGJMSBjl(9vTt&^?BUJM(1<W}=y%~F
ztvy}6cT;!dn5|8$>sA_HndJ#w<O9u&&*9P5g>iLASE}1BQb9k56m8NsZ?h-bhhRI2
zixBgg;<ss2v*Aulp=IUoOeva;-B%j6TXTo$cOzR@V#EwD>{*5VY{wlhScev4BmI@>
zO0d<JNPNO2%l;#dJ#|R~ZaZR~kcEqW<~QD8y{(z0NfC|7TV<`P3bno~MGf>jjbqcI
zf@1n2C@ngLnYvE(&nppH$Y%ReTTOjfHlQVFz%=Gz^Tnp~dr|GN#l<Vi;Fw0Q`t2i$
z*4fwL6r%jLZ#BA@U<~-dqDwG}{`7G>b<yixt1VWYX-7hgPkiqH6TZ~jY6&UhsRX!8
zt`I<)=$HK~%MyC<>o1qR-lbgl$E}`KAC#y7GBBXjMJ*8F_#mMW@H?gm@K++O?WN`c
z#{!!PZz~-JHI8h)1-Xb%V<8c;uJ<BDWHBmB@Yhs!g6YyQCqxfe)m5Mw4Fc-}D(z0r
zJp=$uksT#`VDjMdAti`m(HtSm@E|y`1(5`fGY*7{T}+&PyO$qVJXe7?>}qd*(R;D8
z4T~K5A$s-zi{C2fh>GTT-}`0<yf)X2ZwB*%T#N~}6aJbR%1FSEdhN>+Jie0>atTmk
zXu&IFR$yyDy<nc|=e&KDGQr&G3%LJV-t-__zT$(-IVZq?5N@|xfi^e`K#gOs(IZ77
zxeuGq5{weBVk)yuT!BTVc`IG`5`ACt=olE3A(0#QGz(L8oiJm$WIzQzO0$}&TVuOa
z^CZ*JG!WUj0b0u5tbWKA3iDt!x-6I6A=+re#dw#{g@(H`Vb!hMsr{q{=`e2NMq&_P
zBX+{|{)~?z8T#fri7`P3XupWUSk-GH<=C+P5~UVE9tLJFHv?U#D#2*ODLs+Nr%xTe
z2v;l#)7}_fvQ+(XEEhBCadudpa(0?i8N&ow4L39zYn);uKJ$v7UaYuu9?t|Nxk~-3
zeZK^)n0J<FD-qxtsDBWjZXB<{`Q@Ir$^Zv^@Ske(&kZ`nZ+AI}G8t8l<0s{&Hq>t|
zUBhr>2fx)VvvP)UGd`>Ejl{EoKdVFA8X9dDt}M5B9W(^ocyEV`+_(BERQ+vpfD`-7
z-Niv%|6w{S!PNi>AGrN6*x%G%BJ!S?liunRlN(O4IPcZ90*&yGNq)UMsapQK^t3!k
zl4CDnLh{~=e3*&443%7<>R;`YM0UJD-i6d&vZ-EQ{!qJN=xPL{9DFmY-l1e~(gw|`
z$0{D|UyKY_tUQM9y(F}A9ftVL{I=t|9u)~i;1p|$F@?lcQvokcEkUbU1@DzZ<cgTX
zGW3FI#RdHyk?m&=lQV$|HJwYQ{$XM$C$;WfxDgdayIha=->K?#?4E92Pij}Ru8zok
z6Knk}M59PEK2uh2OXn`G5)9tqHmK9@r}N25d0}cy4#^r5w07Q^tMt_cY(G<XDX76t
z%Nr!(01qG~r44?Z71MNyB%+ug*XC=UJ6VIY)I+}7TUl!70Z-s+Ko4u2E?yXZ9`Pp&
zxz=V2atQk5a1IdS9D$P~l0>CXlJeSM-8qb?Z#4+$K6|SShs_B52`&C%(w(b%ebd0k
z5w6$N0^eI1r=4p;;G+&qAyr2ok%ZYnGglEhDd)bhJLUlff_PKt`Nj1FL&uUA#Pwj+
z1$K>UT5~!IT{MBlR-2T<eUQc+i4_FXOev)`?bqEIX7$oktA5F#?F&FP-%iol#h!Lb
zz{|^7STGdpd~YwXT(uMa1-z;M!}7PPhR45!fZ5r2<jq5c+D_eB&+m<3njF~ohmZMg
zl2SQ2xpcsTu0T@<J5|_ySg3~oozt_fo+mnK+pD*bEc_zk3NX;p%^OC|ATklVS;dmD
z_G>Q&$mQV31(YHsT}4!Ps`Jg*N_M-FD2w=A)CY7#AHqBZr-Hz**JrJeX!)|pdU18N
zJ>Sw|6jQDI@HRM-2hk+^uF*H>QYGyM>RCf;Xjl(CEipe5)L5)mifCC`n1ZM__idyN
z*$&*TJ<pwv*QVMld@2~7tZgG%v*vPRLQ1SArLUBvgI|=ak0&#^vD7<Ql=d(mc-f6@
zR=DY~YCCTQy?lL9&=C|d;MJ6Agc{jDdZ0NmxtiDHy~GN5$%D&W?bDJ)>>LT5;eOMI
zwlec<sl=Hf-+O{<=qZ!tiLX_QBG7IK$%nRoY44ox=qD1QI-kh@y)SLL(rsh-@wV{q
zLY(17vKT%ExADt<4Y8viZ>$bhc}&hW7f$cVIEcU18Ea#MkL{<X23YZPZJ8v~AlvzJ
zmvL0@PHW5Bs%K4c>r)e`&yg#_t8)obs?NcpfmqEPIX+*52Wh{aEojfpb^U0lU0!HX
zmF;{T=y#HO@$JlYm-11OhZ;rG6&I_nj5W^^;Gcsy&)PoFRK2g8#cd&yR(<DVBJqAh
zKlm7G8Yr^)&M4@PjLiVFYb+*twsREOz6OJ3@`$^Z11n(mLq!mXabWN58Zx;gOKg$$
za#D~>F{f|D)vz;qC5FiMW4EU3zZJq_gU$C4e$(~en}{$x$`|if>y`0~dBiK{#Vt>l
z3x|4T35;_1E|Xkh?Yh6`xY*H&TnxBC;d<lqODXUcm+`WkGQ<s$F3~p=|CgtgRav8K
z*8KxeF5C!1F?VHRpG~J{p4V;DL}lc4=DXQG9jZj)Xn#JbnI~`)v!q#h%2FVb0XEWH
zxhVM!@}K_DHm4^*!QQJ{eQKGaaK;1yi!Lm>fEle`XXimthvBxvz;Uym()^B2{hsIt
z7}ArXpn5By!BX9JJoB(QbJllk=750)xK4rQ6Jughid-^&NJD*YyHR#?pe4hB1(yED
zA3wY&EnGjkw$)TP_NQcCx~Lmw7je2{dyt307H_5xXpr9H6l`z578Zv<Q)gbxUFL{<
zjDhi~OHl*dWxi`JdX$+lLyw;q*Pr`e*^n~ZSm|VCRK1^nPL=C{ftaY#65@|=U-u|x
zGQe^8WG1qUDQX%z2+U=9Shy5A@)-a4?&Nydv?7&PR|liFy@cE}IcG>Z8^)gTJiz!=
zCnUy?QStaC^`e(8(A))V;5)z@7p;NRbCVda4w?Y^?w(&8P$`uEDATpbtZ!&s>HhX@
zWoJBHip6ti0Nz@iD~RUan>_ZB@R1$!g3a}{XL@dJh57lp+8MQIxpK@`2-gfBeQX7a
zh%7Am`Oc%3!i(h^A@m!DpuempA7S7@eOsf^#h=iM=MHk003aAVt0XPc_X_9Ltn*pK
zqpjf1kP2E&JMseS6cwqal-A|T-XI(8hf{b4v*epxDP1loZ}orK{7$HXQ9@brG_UD_
zjo>S`u6CvW^~-uZv-te~`%#5Ooi65!vT+&ler!Mk=gr5o5Bo1E0!>U8R!3acJXhxv
zo1@jmHfE~(8J4ci;dlVvLlxnVPwmr<dqS#@7uFESB>U(lx&>Y1hc5}HyaURB+0hl6
zpzXwHKq0FH*{@6=w9M5`@Klix0ld}UfgR4em(|opqZfrAV)C1skYhx1)M^J-4`Rb~
zeYX=2qOw|cH1;|EDRbq;7`DidSDTa|kx03vU0`QkdC9xNv1>0Q=7W6ymh+EOD_waa
z+sayEFMv$~h<u?By=-Es?;uhTA4ZS%evwGL-ql=*j_``}*|%v+;C9R52H>s72Ys+0
z?`Hay#X>-@n}x3a9c6Ay+b{O2N;H-7-Z@#x5i9YlF0^ZPf$Nm*fu~sz^g^3<G81Uy
zzXm3!(?R=`l#8fZcM}h?^vx1FuDw-&F_xK&cCnfcFZ|5z%y0KcvxqZVlyRsjrM^7t
zkXC^rnZmHu8<BQwMN4J9d^Lvj04Gy8I)7_Lnn>u#8oaZ<oCIJg#yJQ=4-@?)Gn6Cr
zr1Wk287sywd!R9fmjS2(Pyvq@Dh8UjwnUJDUjsSu?q^9*FmR#LXp!YW3uR_pb4O?F
z1)-{0CjmhxEB9)SaS@2YaAaGIaPg0wPn9wNk5pR&Mn<eI4xV)G8}?V{3LnTz%>J{6
z!d+Z>Dxd8GGmt%s`o^L&y702JtYs!5(0^ZLzjg|zcU(>%5fP|$mnCh<767>O&O0NM
z;PG6#pfVsRzHVIu{Oi3rpsB*IRA*^~;t}VUSmVMbz(~=&#=su^iQOhyO_#(PEUyVp
zYKVtyl$9j>Vu$%1#i^7c371BI{(t|b7@6L8u1kEEM1_3pkkuJ8=pz6|0*!S{W>^Lv
zW~NaRPe)oXChl8w>xIZjFD`lZfJrxZJG@=gb&iZq&bEK#EW%VgoIFTfsEu3}>Sew-
zQTa#&&U^TH8Lhb(Dd`*(`>dpzULsb4l$xDf0khjmVC-weDqn{r-c?)9@?@oz6YFl%
zmyXijY0>`bl9MKB5S>v4<B%?q!55JV#{E6&bcx4EX2f-0OIPX3ceQix_xz=Raf`Im
z0bC8%gIWlwA7;Q~w=)@n79G2{c~$Vb<abSlqks;%tB8zU9ehY#dFVSW%?b3gQftP0
zAiMtp9$k?(bt`-9vx<eAz8urm66paK#j$qNskVii0})4z@XxLr(Xq*aBTpbUX9O1)
z-vXPKEoN9VqOvldJ9z&zm|gu%M0uT%ecV^B<05=~eIz5-aqPGrZLAOrwNn$P#wnZ^
zIyzWo?I^R-1B5L#Kg-MO>hw63_GosJ)o%XL$LGOofl6z>-O$J90{9_^9}D}7YD6Lr
zsZIS3Y%-*pBQ}MG>Vl&iUF$!wKl!>dnk!^`10GmfMXW&eU`;M+$ofow_~x9SI#XMH
zdH#ER3E=oi9*&9wmiZ7AfiAM`zHjvb#!Jgeo~*`DfQ3lBq)52SUymT@lg(w|1w64U
zfFVh&jwn=mKs}8r42^DAo%+JuKEWCIA5Wezm6qJzcFT9>l6rnNW|rN2H!L|v<H8&7
zY4h`~r1cSrXT8vB#pq2=s#khr(aKncmPr9iqF5(6bUQ!esA_CX@-&OywRI@OtgU2T
zBiF<6(phiLvo!z!XKWf$*XwjBES=tgr9mDkgtK4o@{xzoBIIg%%+}<x|KR`HU_b_j
z*zfkE`q(@#0`v|dGy<e>K`!FxM=O``HRY55Y$@*%WaEYZ?W1nP-bJU9n>dE9e<i^~
zQEap|1+kJUiX&qA^+lP5vlV(&qa?!|saWzUic5E+_ZKH1`hrVI%PULYu6CL%!9>OS
z^uhlb4XHo94_XNxp-Yl*<;+=a19BQECLpV|XikYia3aY&D%w8+C4lSj>QfOYqx)D#
z#Eo&<`zt6Dv*>Wg9WL~Nxf0MHE@_r{9tf)dMo#2_fmtVZ8W3t=-UraUl9|b}B5HWd
zeCzV0<^Q|@$h+6d&@j};#LSl#Dz6BwYtuw6%8%N%=1gIu&MU172S~4Q>?Z#Y2i<;y
zpHn?0I&6-2Tn=D}0z2BO)EC04XITOeEob)?@j8G?Iq+`qsW~+?Y}NO1Z7V74_c&O!
zUFo><mS}sf&0=Gr`XaR<=4Ql7Y6ib}-rhsiL+i?lXG2TM;EZgg?#Ogy#;?ZIy(b;0
zzmx>4i|6N67GJwaqK5-?X`b{Huh2B!-W?>MMvl784dS#k4bpX1M@p-}R#pW|8%<-n
zE(-bO<tok5SuuuwI$E!+&j{3VU#lt7zJgX*4p_tWRN(9Y(%$WtOOHSUw4+t+<&B@8
z${%0t0FxEYM^qRu!?|~TZFaU1H*JW7L#lQ*gZi|;0SFdK9KeErK5WPJr?U$<LW^Yz
zEK#Y>xfnEYi_tKv({^bg%}HgIFq1H#aXmcb?9g@;w86#RdAg+|6}>$TX_XER<xV1f
z#z;(BN{4ClB+fh{uD6w2b<T2lDRt7|816gXR**g^q6oLIx0YBNS8O}W@r6qqkHm-A
z=A4Ol8j@Vmko*=37ZU0C%K+{BW0cApJs|Wik5qOCFI#^2Yox=5)L}#=xSh3o-&?c^
zO}F+&Uy#ZH8`I)}r8H-7UeizYnM<Xl=StR2Rj?=If3c^Ko1D!~*9Ac)n+L{L!}!A{
zsGJi?4qWxlrm@ab`T?TwULIm)U>Z?^Ad4fzQccIj+SV_a6z`=;Pz@=3Lzj7Dr%uh0
zU}Kwc!6qaIfB9rfai`sz&Mo!eS=+CGhi==ulLb?&lT3gu6N*a-j1yr}Cv*Yf$kUaG
z%^KF)x(SE3T_jJig%Zo+q<>oh+88{R`IDmu`uS)#T3k1JSwe~3^KhQNNssi$fR2Lr
zm_Wyr@pL59x2Q-5S*#^SP>?Opae16Y+n_rRiizGqf7~silJ^UGgfy$^%YYj2gM?E0
zvIE9>iT1vy%EM)6kH~Ysr0oQW>OJNAx&m@9m{mxbLmLj&o2%tB6B98?3nU4S4Wm9;
z@m7^4g`ggmU-&u_n`lFMY|wuJvnRj8^@v4dfWvIj>H43VpM5#c_B}7Y$SItL_stLR
zhnVRF1P!o}s#o!*u#ob#5tz`)<SNb7Z&iVly***7aK-Jh<Kwu;)#!xfs~$o<o>HUN
zYsJfAWwa)HzwEw>=oqgLq>G3I34GM3e4bswZPf4+lgZZI(tYf};Di65pPZMM*VO82
z^>yOT3;`1kD88_?rqGF;L>7lB)H=As4ih39ezNS|>IwpuQ`|UO<pJn3%@wcwTkYl`
z*+)6TIQtae!U#5VtN#65(fyY9#wj{Pw=61EehWr2O}h@PH8(kV6lGkTg|sj*U8*n6
zs#VmQ+U&B!s7GKed86?-oM00lbjn2X8yQG)Rdk^i0*xLRuFCt!d6yUIfaE<X+sMRg
z^|Z6Jw2WH2>nd1bAtpk)z?h&ezIJ_r4l*3cFJFoZ=3{|AHF*?jXAEZ|P$mT?)hD}9
zorZw`-euR$a0D9dm76kGN9nsUVsUeEozOTX@p0}*Z%c2Wc9P6Hh>sMSb2x4-f+xSx
zC}sG{gQ0UHc593wa47I<y*W^10iqNj8R;!oft_QlK##bSk;rcr4KhA}8rw8uk`Q3x
zhbDalRbzo%v)03~K*WwjGUpsGPr)>V5eR2GjuNY#9#W;udBMR!q_$;7u{-T6_ZT29
zoBt`ht!-5f#?1yK!gB`oi!arH^^GJ$@vK|5yvK4RS9r^B>wNRP_$lGjzQnmyCPw0H
z^7zir9_*GGc5Qn6*3^>4_$0DB_zL-9MN0BhW!Pa)kwuuYBj7o7aNF#)+vgEPtm4}<
zC#rREDJ~F^5ln-Ul5)=Yli*0^+pf{~_CB~+k<YvP`C`2BRF)<;{5lQjBMWTfp+xb!
zM@~LNn<a@}1hr4fzdFt@2<EA9ExlG&Hmw!7dWP_f!7EU@X4Oj>fb&US`R`|#YWsXM
z27_55t`jd&Mimu0?-`JL2a<ZhimgF?AgiZA8arn<Nw|@6we~uqtz!#EOgw}<Y<W0>
zk+0v|$G5eKvz$sFM)g~kdy|GT3P&?i`(iWBaDJ9f8gh^af-vq&N?40)mjoGoWkApM
z&h-Bqidk*lOK)WMsF<&cBGRln12=--`feQ_cS7J}KpEb6@b5?h;=gY<|00!9J~)!N
zj`$t3VevYe+533$Y^)t%k4WO@yCy!tH$m=8(cn)!jVRelNnmXoPl@@y+M7RMh~%XJ
zUNQ+rT*9|ZCyq#IV?~A1Cq8TW4j}9&$s?k{JfK^HPffKwJhnPix+RAUB0)WYPhd6n
zZR+9Oa@Do{eao10{O~{?sB_O6^I+*x!e`5P0gxufMi2G}PyZ9IAeNxp>2^L6)eawJ
zMy7(=J+HPhi$rl>R$tT0j=MxHv#IZ(>y9>5;C2_LT9KMZjJGNa@;lEc@94+3Z<s#B
z>@LX>Yg(#&sFi2pnsXKri7GBGo`HWy)8#$E3fZS!C!SIV)o2eq6@hy_``>Bq^810D
zjI&_$@a_M<^SmoNRjL?B|ALY0B@1C~Khr+}eY4Vkoy3vM2C=c0#7x$c>(f86nJqI)
zrfhvigMnP?bp$MKb+Qjjij+DV@ltKsD~?wxb*515Ptt1nZ)cKcl-=H018<WltPR?K
z)H%O=Gm%5Tr=!lKcKor%N&LFX>9umYrb}lrbe`JD*y<qq+16H%Lu+TPXMy|%Nvo8I
z#fyq@PBT)Kb4|>63YN#gID2ij;s%i23$M-sdC<?WyQ`}pFc3l`!dBPTgp)2F6p*$~
zjk79~!h&s#9<nfA4m=)cnuwZ=g^WtD;=G0moP4VKd0WipQ$a-_nv8F`&jYl31cGcS
z{;~omE`^1ud%+%8w)@|PSC2r!oM^T>h(d}DbnqdQ2Q+>%%^b>&U7^lg7oEC6WBYD*
zt$`q=VFTxBpCPqjrPpJtlkLl2ey_9C85?@9dPT<Lt1u?{m!4Zq7D|Uw2Eg5{v<s?F
z?D}tvL6gl<9@VH8^h|m#ti}SOI!jr``QeMajM=YQBE{k@HWR7Fn|`*4>An^0wxBZZ
z2ihBnN45O!$_lk@0>ZYPfApZ}kgspQmEwqWN-@sQIBVBJO63~fb$uO?J3P6Q94+xC
z!OBlWAoxUHaa>(pC8Xx^{wV|A?lbHRY$w%vcTiHW=}B47^Q0=%kYz18{Vd^|HctO5
z%llNs)br%n1ea+mUv_@5doxAq$<w|Ag~Eq_UV_}o2^WAQ{F{Y_?<-%=D3FpYqu$`w
z@+&a^juep{kgb1~EuR`Sa6l)L%j^?}YFA|td{tZLxD?rs8~XU#!;N=WpHDc7kFF+L
z-jMxNG|KUdlL@NAvD*EDx2lEPCddf&Iv(C3*<NM*#?Hjb3EW(|=lRWPzdB`1Q@##m
zcrm|C<1Qec?7niPtKP45_85I}SywACeo-BCch;G-*{!YYGk$S7l#4nfJMH>a+L9Au
zkq&)XkY2ICKX2cvBaQJ{nprnAGjTe*q-1j^g@i0I(NWo#BNQYW2805!><+<yAHMNN
za4wzP2p!-mmu`9yuxO2H%Zsz{{DKqBW9A|f(<b=kePyX=uAuF903?&RuF`y&HXT2g
zB0i#OGl!d7N2RB{>#1s!4~($cOZ9rGWl32lA;P{nT^+a*mh(E~$~z{`@95_P&xQ1y
zQ_Q^4LqM8H)iYy?cx?9J8-n`4^8zK=*(UID<|vj;3txWS#%XSF{E_JNkr_AMLtK$p
z8!07>#11~=0iEq;xbjTO1Nt*;L~U-d#wI=&1Qx7dXQrQ?V+1}$6z)6Ekz3E4Qwe7-
z-2P++a+~*$@YMHXt(e&9d>Lqd!o$N6_J3ave14^6A2&RX38b`(YakVw%mv*iHnGH2
z#E8i+A!ZNh2c^Jpq}UX;2#nv{l0IOv&ik$;@6KwO@uPx$SY_a4elRxsMCZ&$r*ZKg
zsJkp$!ZH22|E`nl(QGYpPWOunyb2Q(GLMY^y2dH6{XKhTOyM-Fv-Qf`_a|x|d7!6b
z#@e|nWp#POc3o79BbY>37TBoCWCsf2n5Vl?)b8j=!y?56$L=N!k@`BAi-v1lx}nC@
zF1Ui(L4$5iU*^X2HCvWs*F;a%ns4?X{Btyh?6r2)!kdfaRJ$6I@&k>z?^E@;y>`uC
zk|$*}3cBx%oF7xszFW<Z12+D@-N%U^|1tpmyYM(Ry0S%XbdHE#meKZ6Z(a5zXUz}J
z2lACasCr_)N(<vXebkeG26ETv5incdW(gg694&7R&R57G9nR)^Ey&99MtLr^i{tkF
z3s{mHw|vA+b*CO7lD^Jy|7RF(XTJ#Pc`MqW<9G_qnj3gTQ!R-jy#HhmLE^FC^<L)J
zhl_bi@Nn~BU@OB#kq>j~6-h8G(L3lhbzr<I7uU8q`q$S<y-JQmspgH%y3NkJu|UKh
z8nn!tk~iDiPG+n%SJzeE@@sN1wAyu_->erJIjP7Dl1fqKQt83j8h)~NH9#=QX5tLY
zOvN|owtw+!G>Owst&^VMUr|Sgdj4cu0~D&EILsnH%XWI>58#twiw0`*>qbGrsp*w$
zcceWtXIL2a3mJU%d>z4v|1BGECC|R?5jc6GdSjHr)cvw@SLA2poIXf1gymTk6twr+
za25kI<zbnQeB4gRV@*BKR|WZwQ8RAY0e^#68fO5UEW*y%-^b2~s5JM?X6)(A;~C-_
z+Oy88wJudo!u@G@<Zszep!cPixl^Y}S8wpQ2ddne(+i)<H_zV}iB#Om>n0jR7`f;L
zOz^MT#ZJPUzAO(0WpXVUqpm>+OFtddF&B*~(|9bcQu8(Zk!iveWv2}1k-UIc;Yv{F
zt(BJ9hBvtV_*X98ls9h$x0}Mb4CB=2F`i7<MCZnNmA4xzj4CYGg&wXl6Ofm;o7fc=
zNn9uTZ`GS}wLnBY6f^t5@34Y*XUYg!n*Kcuq~MC^q+)8>y_UBOc*b*(xh;RgfvZsz
zi)YNvIixZjEe~X4QoKh9Ke;G*aT_(zeeA`}_F%3;@&1>59tehliXDPx`SUwa9L3-e
z95;9BO=xH?sp?BdiugCfUJa}6w6jJog|bfN1-F7sp(Ck)J@T07MTlVB7TSSnP1P<R
z+v+4)k3nM|my;-pKtaB*dlSF|?lP^0e@)A1xj&ZP{kkwyIS{~n5G4JgtWBpeBW6;Y
zzp4F2;p$0!^<7!Kw=b$iNGX_?X);tv_}iz4b31w}t^!`<CNF5|pFtJjH>7;YO3^sz
zn?*F2SrAMQ!7d^{n*J)cMZHu6t7PI!NC7_<oiVSmRHW;k<?Ylg<rJoN^+h-(CXS3D
zBPMaErvQy%i8R(r7n3tlkrlV4z{|2S4w1l(LU~X;i-Ctg%H7)QIiRA;b0mIc`oBAA
zhj1>EF_d_tmxH+Ix#dPhQ@QtK8U^C??Z8Xf6voh>sVWsYn(f=d<xy$2`>x(VeN0%%
z9MA5u?Ru087oRsozhd#Q%ZyyMr!#5LA*!Sz{TT$*4S5^=>n}LQc*h)&79RSmSN|YI
z5ZSFD;}gQG5H6`K@ApxfoH7iwg87E$*=Z!&F4susDG3fOGse?znH`)qZ_*|5ol3P`
z{13ZTW`3WnweOv&FgI78x|3Y-3bHoyXr-7r&L7B#QPR7$>!G(k(e8^eC#JmHB*J2*
zQ=@oty8LfhAs=RB#&jhqM=y+ge$!Yr`|WgQv9R@e;O|e*Z#nT1uK&vYHwT2adIoR}
z63v+4C68w9PGg#kl0TM6WhxQ|$px6Yu6;-})(vPT^?j#`rx{GW>Gh+qoTm`G7+(fN
z4?Vc~o<o$s>?x4BmfC>ZE0O<In{kSJt`jQF$e8Rj#||Yke)|FYA~*Lb*)G<Lk`Ka_
z#tQ$XM0t|hqk6T~R#I7<6H>mf-@!ii54B)bEwTYXB^Lox6923PoO(1@_gpM0?pJih
z=BO}}wT!+pNE2$)*!J8q<GI0C4dsEC^px>FFOIZ$T|mWGKO@&4&!5q~35>xT=EJc3
zU1hX$=k(H=kRupJdS*c0tyJmjr-ciz2_HU+Of5_jX|b0g>!eYR(M~KR8s?Mb3Jk`b
zcWR){fABr$N*i}+T>1PX^#H9bIA!5}^g$1f+C0aQ$E=*^v<}v*?JE(p0xyMwt;-iT
z5?4!u3}rYX#v$4B{<1E2M+_AxqoFT44fH)(K4zy!Ngnn-RH1dPxgC1AW;YsW{TzR|
z)xNNQ%$IEpBYVby#ztoqgIIQA0u`BHF~c$ZeG;CJJz3y+oFF3?djnp{&b8_<N$(Ru
zAbCdN2u5a^-KhZi9htZHB{B}v>v6`}MZXXI5Ix{8c&ZqF6_K}56l{2tb_0H^SMLX2
zc2K?=g|jFKm~hg%d4&@5xu(2!bqMq{iWNv7e)~e0Sg5+ZTx)p2crL|FM0AM|lLQFH
zkq^KAEu*<LVzW}x1RwESuOI$1V9D4;Si|6=$f95>6V&Fzert3Y+hNC3RbG`_+rQ0~
zy!`naZEfXu;El7Z?Wv(HFIk!oJ%0jr1arrzufl~zBIWE!h$8-p;FzHjaVehry!wRG
z@<BM?_|3J)xa#JRgQM}s$4T+I<OBhCTgO!UAUX*@9YO|)a>I$aIaV5>26h>o`~TaC
z-&qT51w9EF_`mysJEzI;zaVm}8@}$kxX^uX{3X(u1>{N=->UfxDE`rD-=u-U4u4dI
zEUG73^!e6jOt;Np*m3GZS9IHQ-PPPz)iSQKdTda`t{4X)r7Vw9M}m@4K&_bO#(eu@
zZR9JBF1W|2fr&oi*-cB|n%jjL?A)+9&!%;yv&GhF$;%AvRa6n{CyNooSi_EWqH;V>
zkP$K^GR5(D{6bYBr<4*jH@DAO#!K$-!b$yRl{oxJ_(MtCwitsh&%dQ7X7R(Es)fY^
z@N%gJtxs|J3e*jeUel;eHW*k9B<wZS9HyRoM0~<?IJkeU^x>sihIVy}l6@`1Mfmz>
z!_E!^Tsiy62%7Ru&yoO1^639ognV51#l!n#Z=FoWM-Vjd^6&btYo(GF+lS1)N=cV;
zpU8o=Dx}jJ90Nby1f{~Nhx5kkUc~tmg05}qjJ)@7GIdFL2LCcPryHJVca!6VU?3za
zIy~^r@<6Kn%$Ao48t|scbLMc3*Yxo8d*LCBXjxmheWWMlm^`dG>BEtxB3vy8V)D8B
zt9`1cfodcTd(^EQOi&}E`mYbxkuj+Nq;P)t7>64!&vmf$yN>W{1=pMMeK{pVWohy$
zSz<$;Aq*aZ-bz2|vUiP<w5gJDc{wx2e{UY{1Me_|dVjQDpbYv8A^fd`?&y};5@$fL
zYmKqvy}3UO2JPYqVUPd9K?ad0RacEkZ+$B*ZyJk1Ng@+`5ac=YO!ixTt892};H;UT
z>_9@DrF?#?&yhTG)??q`@>ZIJN}uegWdbg*`}jk<v9<|MfrVz(Jhrw`mglo5=CGuy
zM!jZaXVXa|$4#jkxf(O}Q<D<*oXF-r<k!n)NAm1vM{&T9uw_d}^$@`06E{))#sPG`
zk6#TOPlOPR2Pq`6hZ)NFf+DVOY>vHxxbAT^Y5kDg;QVd<ZHb~;z09~^$dB_7?KLRf
z8!lz`&o_Nf7j@0IDwZVGGcxCuwY3Y#$tg!f*TSFhOtqAG{hssL=yBVo)w+m!{r+8z
zLV;#_TAB`Ks@`pI;()JYsnwRJiYEJgo)~Zhs%k~~Exk4$fRD6v6Qkc=7wZC@_LEHe
zpRI`b+-)6Yiv{^!j%B0WfIo770>nn@^JE4Y2An98cp}7Vy<KoUaEzEGekZ#-dmk@(
z5e4pu0T?w1z<-8m;3{=sQRl9uc!<ANy;SVidT6nE2VE{%<|P(uBYbuqggkN`zDvBZ
zxd}%Mz1n2vjZ3Wy`f)iwNaZith>Gqz=z)`%iM|mtM#0DK-h96M&hrn;e{Y37vwSHj
zEwr{}mT^bD)$r|je~khMPjp_Owv}=S>924QC@eU%hjh|FDy*~pNV~lEWX#Fi`d`|x
z1(wjf?86g|6}X_@`Y#zB8d@P@{RvmYdFK9Dy4<@q(F8O4&+qDM8^JB2Vm6!#x3oPh
zGZ4J%08#>MmS^cGKie?gr!o;5#9ib6dVT_^6*dNH3-=%%Ru0XOpVT+b*5I@w$x{3;
z0atvBbTEYymyFzJ;AZxZMD(4J0|lVdAI!OKqMhI<*Bip)73eX4N}6YtU=F|md`%`7
zQcNLYpooW8<z_@h;RXuSe=pe6!X^A*eE9F~c$h8l%3yZvL;uGoPj!RTem}cmX{S3A
zj*1THt-j0n%sGFX9OlGK*FVOf*+uL~+$$dMSM+<)1e6=ttC%dzsxl#3m?d8%6CahP
zv3jv`S>p9z&re>+AO8GwzVlz^(;J~w=Zrru4VBt1;wXo!LdV-nTG_TGaRfgg3rQo_
zU*kOr8LB;-`tw`_Uw+nU-d0~n>pLEHhzbGV9r2lEpW2A^Cdgxrr)+B*1D)gE7q<5p
z6FkN;x3_83#=(!43;**ajdlOAjy`V~2SDwnkOCFV3x8et7tp|KJbZk;ZH3T}6M{1$
z$NE(Op|c$wJpHECD(j<(P)m@@Miwe#RZ_}rRq*w?2b#>4-GSI`S3<)JC64%isCw(D
zsJixl_!dDxYG{-mx}_OVa_H^`38lNGq=#nc7(kFt=?3W%5RecF$q^8cZg>yg&+~lW
z-&+3XEY>>v?7gq+`qXLVtZY@P9FB=3h(h@7w2F;-tB?iipT0?y>)U*czMp`1_d3lL
z6suP}^@?_idl7y#uXPGT09U*K<j>gz+audNBhi-jK;gILgCTR-F7^vPj}IA6A0@g{
z+oBf=S5>sLmwVviwcRt0jZUNip_eU|qi-2N5+8yu9!Et-8SzKa-2LP|6%OxPO}4eO
z)7-OfX#IQPK0JPO#**)umfnPkx>2j3`MWy3L;$iDlbg40kZSYVGx#tBtWpML6qMcy
zT;wBQp2_tS{f#0@^q&s_EsukKou94bEiwYr<xsnhArY&Zi=*=bD#d9SAJ<t|LanKx
zmKA_tnVPs2+Wmf`oz{HCYRf5l>-8jm5mBGQXt!!#S(dCuGf_piK@>%Ec5};{D?iI$
zdxh4yA+W&JxnI38HNxui!-6_#W$8Wu;$3CB7ZNfw{Qa;qd5+Hb6&R;4!o8{@>~-V0
zYMAan3JgSfRq>iekxWJXY3^U<k%Zhw$I0pIixEE0cXr->6H<_#UsVYf|JIr(9ToMI
zfkc%#Xo%%5IW3K4(n`@Vi{ox%bF<I`d$mrYj{@!Q0P?Uyptco?Upc@L4BO9-BCzQI
z)boR5<(#0wgbn)PP-t2ls5bcp9<0J(OVg_efYxzOnfl}DdS#Iwn7#wPdqPd=A3GDa
zYIoR-);+pf+Vyv^;R`(q4LAlPhRL7ZJ_U+5%*}KuM2Z+}NH}~Lj1Qd4*=<5Dp?ov7
zzr8OgEf^+|p-KF#>}De!*)AkhhDr;2by&KM(dzMr!PWb<l7lLvp8_oS1uD}xz;GBJ
z@2+>xE|hDU#DBLH;t+82xwqdrym%aBnVoy+GIOO7DQ@$rzaVtzkdTRiAX1K*DA8kV
zj38R3@Y83tW2eQhU?!Eu;+Rs+i=y?Xw0sWM`qgodg*jEKRR@Q*kB6|7ehjxaq|ih`
zuKztd!1BP~?GpSPc&9k@B7l&4%$=Y7q$yxRogN&@&hJhD*N$x3SE_=$s7H_w8lp4n
zK9$;U7PYFR{nS~ZkG*TkYH-3Ymvjs<AEdE<`6u;yGe~e>-C;$qngCW$USd@*yWDEl
z_UaDL^vrB-ac<VsuFd@q@MYHn^@cF2_B+u-MU9_9Ath5h6|YS=KsK`^v-2tlj+FgO
zIRl!fSI|lh$_!u?&{{eJ8&AD=c6PPuf3z0KKpdo7$HM0vP=n-}nBu6@mY27Ibfe;y
z7N3C_?}^|Z6b4XU7@PI_0FG@tWfE!pk6k;Hxn%X<F2dopF6w^~;)iWf4t#VB$ZFY_
z?hi^Z!v@%yV%pw|;si548cVV(@_ON?TSStUmJ1#Ql(K%|iWvHQzv#i`1WrrOzM-)K
zKyW9~ZQMI?3wTY8#T$~jDQMp<L*j7~vQYc>Hle`oB<9yQgZf{+kkHE~O5e`6b$@{P
z|A<oit#NK{d^a#=z%#mM_<>XKaoRnW5`!PfI5EsXOzisW=7a5{9d<Jc+llt8Rj-nv
zmiA&XHkS-fk=xCgd#Qcj?EnvoWmy83O_NrDpRH6DaVc6e^SNH9Zkj+-chB`-#hMbF
z%+-$hp&&80E{`k<a-Iqd5WlMbu<m?fdD9`^SKyVeYkOon6(s<Jh4_{<czirHIed1E
z9`fk>*C`$ge}kVvzC%9Pp`2F6c}gmI_C{L&-6%t>mi|$iG7_u0`eIQDW*@jKITvM}
zmOHE_=IDwdy;Gmz)6@3;f$WgsOAfW?$w74(TK8i}RHKfg-CD8dz8PB#Ha8GWxH}3>
zRx#X~fYcC@>c;AouG`|F61%Zxy@#b&LJ<m51)bKdjw*xB5yp=(fLw#|XOlKpi*17w
zrL97scmFJ$`yJe}>fm0Bd0xsUi=+0q!2L`miwHzFg|4P^Z$maN36Cz6b5q`qe%~PO
zrZMoq_Gq46)?1)AT@L7q+BLNKLQ=XF^VV&Rz?@J2h9xZ_^tb343*@CJEw#Q^NW0k>
z`E;Cg`=XW*yK%O$0-q~TlC>#mJ4E*=dVH*d26uH_qh5`l+8xTVyI{B#M3vJ0cX42T
z)2*mngu5^E>;J0-fGbK{J|2M-x3GJE0Fl5)SA`eaoK{HB_kTnP%S=QK@p^xu15d~*
zxxGih2SCTw>6f*vK?yD^0L<IHELGcZR)*LuK}C`9xq1~}wq9MOXly9^(O8io)lR-G
zfNvcx&j8xIq#d}>!p?FFTs$tsUmAq?9T!Cu2x-}4@pf3es;YUQ2Aj90`MrL=e~zGV
zkBty7SkktOwk;51R$vaMX)Mgw$l0S#4qtl-;U-A6{=#%icxxL_olq!&L+_+a|7m?V
zp4nxTEuF(EJUo2!O9n5mkw9HuSh`YjO4{(Bub=FoD_Xe(Fa;D$2%)G27WcN19IP74
z*A_o-|0|&CwwFK`LZj|mkF1K_a3R<}EJa8*<El(DCq+j+oWF;)(~tcPkSv7&+p?EZ
zaE%UDV^C1~C<=!enPE!Kn?%Se&&CfpVJcqu<OXnCk~O=Tl5>w8-xLCu6Tr2~jfY~d
zI{+-ShfzjGmE0ZATT5H0qwjCeZ9(ZWZ1}<Q4QwOqsLtfYB5^#`btA<yQAX)YX;*-L
zwi)ax47F~Wbmmzb+cn6HV#=mAQ?z;(8P|HI3LTt2!wKx%&GS<>i?w6F(qmgY`P11N
z{-r5i6nwk&{<rsc-0y7;5yJ7aQ#vik!t0F_Z^>EG%Yv4A`GCW=_>tA_kLDLgn?(nV
z7V9*i4!2t!8)nDkMt%n#d)!eG>cHx#&X79H9&MX|(&r*n*=ncJ&e@asxqikQ-|yUx
zhI9Qj8nsP1n(R9rOx94ww+&@GOUEtTz8CpSAQ6I8(WMKo)G=5^;Q*Xrf`vMLkDT^F
zkm%LZBt5MbM+I@CqpJ^Y*!U=aKakV32!XD<Kzxo02t00dI9Or!04mlJ+2@L@ty`Uv
zxMt3<P?Fx(X@ZF+rJ<o=kEN_>w!2cP`mHmT+~BuexJNWWyt!>$c=(e|wcuYHO;thX
z1~2~QQ665MymB9(`s?P2tIHHBA|7m_G84O&7|yy-x|?>2Hx-&n0@4p;<X1I%u-mFl
zc><NU6*JKYN4di@H6r$8xU)vO1UdG`+5790KQ<4z_Fe!Z+}hg~nx885nB!^_jri>D
z!^G{xi<JYCa?bKrqedO=vP?1SNm4e?{Q;*H)&}X6qn~H!N#+utU!_ePq~Yy%^%8ra
zGZD_>BByY6S6C0M+q648Z8wq$V4q75LvL=rXdZ=;9;grI?}N;@)Ce*6G_CMP<mMzU
z2a(j;9x&VJGA#%C=8!1LE?iJF?&8&a40CIB*D03k?MLcFY7a5C72(qg`|^(2O-}+<
zN+x-HE9Ax(=QLra<_}cZ_d=U~Es3meD=X^Vb(&(1g<z~8-TDtnIJ<Mzx+>Xn^yac(
zBrYPEpp1-6dwN1v<S8YT&mz`+TYU?P+YBq)+S`pFq0UQBp0FeJBnxWt)Pep+Q~3P}
zL>$%D?gu?F@n%)ZI=}G$0n;ZBU;SwES`i&4!XxD#!`yvF9uPpIGe#R)2OFBfowSZ7
zHuOqNL;U*+AV6ua0|!&zQ0wx2l3eRSoWz`!knux5cdznNQKb=lv{YX%FRho4WG}%7
z_&lb<%QqT76^AT)iWg4b>_3<FZocmqT%t4e-akIdOY@&6kyi^W@zp5sbrpA@I~%(;
z3CInU`DSb^pB)ttJs=IhT6^#0Ii&X^ZU>*lm(J0){}Ptnq=o0nO+7;m!a1heIY>L`
zxOF1{1uEd9`&YkSuf!}2qlvAHBOIrL%XQ)nhyq;8vz|Wt<1sX+FV87>O`4R-gr&#V
z>#t>IFkSmvBu*k_^3Pt93pOgeq)I}eGtIAQCV=q)5VVGgI<0IxmGU_J2?i7`3~h(k
ztE#QQ{px(7>J{pF0x?tC$RkVHjmy3&<qdk{TZaGKM6{vw-#hGd&CjchA*MKrT08Xt
zoea}19y86cNPr4q=Oc<7DnkXMP`m(9eSJsw^X+5v7|{SXeZzrCkUW*&=&eomi;t`7
z44_B##StQfg|+?5)Nx9~Gp23I4uN-tw@1JChOOJ{*8R`O?4{^6=+`_AO^vE|o=RXy
zVFi@{{X!W&D(>0c0!i?~DmeK0;~F}6!nx^Hz@^zl{96oup@YuIfX&Z6?FJVkg^eKS
zlRaAn4`p{sreg6zI|f?e3ASI`!ipJ-GJsM;dhU^AIo&*;NwK=Ie>Q=2@T0EX<}D9p
z9#L*utJXFKqNgnIw<Cm;WYN+o<>+$^8|Q61cxJyZk8e4H(9JddEkfMAvR%e_ArQ#=
zE!`@NAm)YYZjlay%4zmqlEq+L`R(bcM`BtUl3@9yrB}Y7pZ<SqXsJ5%BB9T%3HA~`
ztelLaV`MDIB}Qw>{c~EZ)^>8D51E*XhQBuYBdJGKN`JkbbM@CZSow;`@;X6e=h2$3
z(e-54@ut@>&;h5UKi-yfSZ8CT%Vtf*{SbzMI+*PFPq)OTrGt9xXpq}PRwoVmLT<ty
zd1JXSzsf2D95(uI-DG_h9WyGMdy`J_oGhKB(3L)=S8m<?iBC3k8?{3T*ibM=Sh^x<
z`Ig6XN-<BPj;b(Pp<tT%dN22(jlP2CO^vMv!on&OdSj)U#mN>WHpWR)@11W7APjqZ
zdpklk2uTCxV3E?#%;_X?e5{G%?v=jI-g@Rp3tQt$A0LIGson0@2E9;=6UdHRgs;L}
zjd?r8e<oE>A|-8aN$uJiCm^DFWFjCr<Qfjf*H2;Q?sqNu5T>SSo4o1HQcKK$n6zo-
zOqkx*NHpDfZfa<R+;J+KPp(}NoMrO~$PFbNcBo)@9^X2kL#$O8V;1sfN@SU9?9CW2
zg^*yw65l!n4QL5aNcrXjzPlA;L<1|v^qksrI_0=58XXFU^r<IVvq(oN)!3>O!wFAl
zL1y=TeFX%T(3UWB`Jf?u5hSV#H3y#rs;#Z9)}f0~XI`U3hwkTtq;)2hvWr&2!FV<^
z7JMR15rn)mCcgk5vlmEc&P!(ueN`Tk@^Ux)Yp8=$1;xOV;O^vx4Ca6JRS)(4nNOW|
z*=hnHz5mH5nGNpy8ixJ&G)y$8{vLqA{G5S<x7UscO~C4xW+Di~wrGR3qsN=k6}4DA
zV~?yr-TL#D92oxz-n_Hkk5QTS2turBxRea>cjh%O&vZ5JFC_b8FlvX(Ij9$M6nqZb
zcxZI@+wM$$JSylfeb9`#5^o2v){L?^8XKsCxp3nakYD*1QjSkYh^sN05;*g;Hg-~7
zgkx25MYAhJEHPg?gKW&+Ji&x}jO~8oF6{%qZ>jG^$9KsS7w=!96btClt{x&cMDIq=
zPZVv;%aZ_4!i1mnXq(XdH1M+0Eq-1-hj?0T!Kxkm?--g$-tNQz%<4H!(HNb^z9T;Q
zzUO;bK=XsO^2a|5)~*T-X!Tm@lvQt&!;-(5p#i3^W3I)Tn`R^E{-*#8c}&}1z^JX&
z_BC}QWy3wM{{2zvfuCNd$41t(!JS|A-JhBlUuLd{Fd5z6zie5VxgYzE)Hqw+huB3?
zP`#{f)O@u1827^iN?}{$c=P%?j!CSdae9@gi3Vh>qn|(ih(Ge~Vlc4EmLgJVjy{u3
zPijuhYM_B#V1dK!<d^zPelE^}O})$dlkfGV+7MrwKL~rVO%mA4OwG^Ka%>=*`rhL5
zI>nZVfkCFXFi8`{q?-^)7by%U^z}vsN5S0E)M5oHxS|EXilfr7BIv4)=f>%T^4gF8
z_JIkI-fJBIziRe@070WeAL#9#`sAYgazb7I(TSXj0U!>BAl2)koqj8ug3iSTjpFyN
zmKS;hHO_sD*IT}p_rtekv+=yPxZj8>Jx;mGJHO0ywO>Ug`!|Q=W5TV!6gM{U+b3O_
zH~mBW2kmWh`ag|0U0<B}Auh9Ml320?i?ZC2qm}W2P$k9qAq`deVS9By#eTmf9+0u0
zXcvLY3v_5&`_0WZnb07A@P)+&a;$`@z>Z+`U1L+O(*7}nv9z^%6EopyEY8%D#{ku%
zLv*-XCAlCosRpYT(V}#^Y3_Ms(%udQqQ<pu=Iy`43D7?cVGBeL#k1!PMFxQW@RaY7
zz-diHC5OT`%-sp_dDuX*YS9>Ln1Jf22Eq9pl}O0?XT^7;Mwaz?J$BaWoA|6v!mgVN
z+et5tOjAZ$MQQQ2&M(G4kdbGPS*~?wjWYab;h&o3QUq8I(LZz)IsU*VB@Foar0k1o
zj;t_Kd=-}a&%In-2O5x8Facmcm2p+_k=sxU?BWpJR48Na#FAib!d~Pv(N$(_cjNmu
zS4@F(IEZ*!KI<gi4t%LX&Tdx*KxP<zdckBd<qdm>6kmYU4%kw~>+{}p80*Smi7dOh
zRy{lTn5|&IJZ+7{|355-o)=Idy+7E?XH@UwGsZLgiudeEXUc;4(q1k=Z#v=?V$fO!
z<t4#30QU2>e`#YJ)&S>grFo;nR?CM;GIu!2J9+{PemQ*{qtus(H#eDo4?wuTZ>2P)
z01x)iS8#x%+ss?v`z`(~<KykSZTR$GoVTOK^&>z!)=-LCw+*&heWMR1E66#$PK0|<
z3pu9|yrZG;_qo31WQuR%ObFN`2S@?GlL%_v!NRDoqa+8&7jr&a(Q_P+-IZyJ%NYrm
zAq0mP^@c;VCjcZa7X)`2Y&W~>=|S-g&?X;Z!G5#-ZD(&UX%>fLgfe;LFoYS>_jCLF
z8#2_>++1VfN3oryxkF6K|7ORn(74d;qzyMdZMz|Lx011iKq@dk7dP#OWN-0%0KC+K
zdTy^XGBWBcN`K>gCZtQFw6T-$i(>mhU~79lKXe#l0Z34I&*F2$)^@~Z@0u0qpYxy_
zUwhoEt<YZc1Kq|Efhzqj9%Oy7KKiK&9fl6*-8YKc+nshmL~8YAz3>aOW^iI>z(0=k
z_DBL!<Shj!^czB^?ae=n0^4iIpFQ%V9b1TAQjDS0Mu7~zVtO*{p_=u<l%JY)jWRR$
zL3)q5i+qfe?5JQbOG9ZZ(Bi0_l(L^pdV+$~s4<TrbC~tw^ti<dp6C(<!Oo#`*_{8^
zE^K_^pM9k8kL$PR!EMa8Gi(HkP)TJlhI~NvE(}1cpZ*8DrD}}xt^M@r(_03Uel5-n
zX&-9UvX4z&HnAdA`va-uMz?^H3f?0w*^=-zA1}PPh|ccR6JXKSu)%P_apg9z*1?J4
zh5VL4_p7V1rmC+R{<s-`o{9;FD~m+de+_o|E);X(=WHwS{!!j*Ip7{QHk6y=3}#Ly
z#h4%LkuNZoQPHAD5bLIK2ICI?I{vzAZ*8q+=P{1-cnUFA^A6mUIl6wpn*RgHwtSS%
zH<Stw%7hM9Qv2<Cp%SB3+=5`^M^^-roIn>THx5-guDBDcNic&xCjEK~R1{;XzPfJz
zx|Rx#W7hVG@2_qX#3p9Aut{1Yu5W4!&0}3T^Q5WU%MNFbGFAenmq~~n&g*yQmk+7F
z8)%otNtZ7PgIodqsRl4ut$i5M)6)8(w-)Cc@(*9usO%iEq(7JgYHU5pL^_2jwtUfI
zO%MKPv2H^t^Bn1C-ap#FC#VR-YasTr?tb9q<Mm~}{(K;;(-y;%$QX>vN%tkfW81^L
z>gcF8MunEcc2%!Ekdxc@zqkj6OJ25Q3I87jJ95svY1&?I9cWdf3D)krLw9_9Nltn;
z75b;o23qyQwS)!)3caCDz#KA-)hS-fswO!KARPcNvpEp!H0C$5jDD39Aa4?>I282H
zCbYlz6i-e*r$i_Eg;dfsxBEMh9Orvw{OG}rKR<m=C;oAe+&eEnP1{OS7*vZOfc<0e
zjc>}u4FnT){)tPh+|<yj@X(U8gypv;3NyUGkxrB(ykf&F{;(EnCbdiK-$&q$z@=k)
zCprdnd~7>78%Sh1Nwq}sq<PHwJmL?7u}mu?T$AW1%ETl{E;rZn#kO6LuuRbFtNW}W
zqK*f3C8%~GV{@5*_gZJQgAI_Fkx7OJk>45Eb&TIPt_O|0+hwx7U1SDV1J|38Ooa3X
z1mEXT3g-TD%kjp(aa+;FFrj2>aLO-w+X?d+y?a{G6tlo!m9kAVDxsC*EZ>5Wl4L;K
z-dI5`oJ%FlA^7Snf<PJ8E_9RcXa0j}Dz1{fBKb-ZzK^NSaivm}uQ??GuW$<W<a=3W
zduOc>qk(w^Edm3|j_%Ylj<vMV3xClDoSmpe6y?v<3C=HlC@0Ti&YF!xH)UGBk;2?n
zqiiSVN?H|>iKkItm#?#YH~hTC1=1W`{YBR&F<HiXZ}JUh@RB*d6`Jt<xfbB#tAOG+
z`#;5cYSs5CGRiAjUWg(mpqr`I#iE?-(uXBb_EvN-HwY8nz;E})`L$Y$AL|Gkeb?uN
zwH`3!dw@o4@1j9Ww=W$PuHAF0XQ|fonYY`9Yjj&n^)4R?!q&>d)p<t7^|@F!0deP=
zE-;NKDQEI<LuemM^=dQ}@of6^=zFYLwncrFN;cUYz<GxDcZ5sOmJ>Lys$2S3Y-G6!
z;#S5Gxl1sshhgksF4BW_VKCm-{>v5jzTx&SSC_`OYs>!EalZj}Y%tZ?T?yK$+62IG
zHF}YO*30~(0rSGk7Pqvoe(2T42=Oj5<od4Sci(S1ly`Ve(v8+eYADLG<7=Y8nRVs2
z&@VxL09>HAaGn{Z%<>5H17$jyb0oAs56q1+4R`((LhGitwwB}#ZZTF@W5JxLyIIH=
zhMX+RZoWRC{chl!0)Twii}-@Yvz~@K4G_9h&Qf-Gp!c0&S>t5LD)s0L|D2qKe$t-Y
zh<m^&vC!*r1mvR;FGx1?d6q^LGe|!;?*lzOZ7fc#hTiM>9z|A~WFlfl<Cmb^k<pRD
zvRaKujrrni5zu@0lPt}l(h|9+9{hIYZ_KBXb8@IMGDq%4rP9y8ZBN(c6*90*KCxI>
z5NOFiWt8-By}R+rKp>K9r!;=3Z&FNu_mBa@VG9GN!|uv|?>ZfpjD1<36$K1hQf~Zd
z`zG6W>*%<2k5|r#;pG*X($-wv79aU^&FW4+!qT^P%1^~H^m%j;azxx?=sbIP&ty_g
z2hS=jhL2ECqzY*@OtmT-NeP{j7t)%<5|Z3XTPnWC!R)ds8R5jLX(}fQ-r%*hwS77i
zPHp<d?}wtk*o}s)21eD}bAmdK$qqO_ijYm?9bWCGks{*_K4EoSPF<#___SU~LJ_So
zS`1mQ+n;nhZjugkG1G28FRJp!vA`sotXB}a?^Du)kqnHqJh%XP)Y1oPc4i!70elU{
zl6XchossTd=R=bpW%sT`SDm5$%j54P;MQ6Ip@_Jn{MG=J4=n>7FkSCQeN(IoLG2Fu
zSV2b4H}yHH$x;{=?H*Y_J*SW#&;$m?{+5U4VRx~Se_LJ5XW<i3=~Cp;^t``v8v^#H
znVfF#yWi+!gJH6|QF7(Tj6Js-586<6#`=n+H;<`)2mTA~w*Nv25maz+w}mRix|=S`
zDTDP5)owa3D?Pkz`(_TPb~Sq97@vEIq`2R|v8%2(o&M3#zuycw=~$BoKz0B9`S51?
z)S)Ba>jSlMS_}yg41q#`O8bQ{sd@40Lbkwn4Pi_zPaQZIB$ASVy6ONZ1<DuYux|Aw
zKfT}m!tw;Na-F$jbI3}Vv2)VA!j<DnjxxFWQC6Fh20H9uW9Nn`Et<Qcz5DCt^ZS#t
z6IM7{4_+umJt(pMBC$*We_&HL*-AL&+3Tv;pR1uP)fcFTWp#Cmm|ElKL{6^?L*=wR
z8Wtja5Xj)&gxJqA|0$@U6*Ax=%wp>=K08#a%=p~OFj;T$W4hfz$XSYx(NdQ52J(gh
zK!h3nYGqT2VVrF=8^40z3!!kM9oor;$B*0fCr!%h20EJ$%^c(k34X-G-v-*`GrRdt
zpX0IGs(N`qdz=Bg1d8U=FIS~k7pd2UUKN9)xIdhwSN*_;CBjVaNhF_BC_XPXWKLPf
zauYA;=I_2<`jZ`R$uYmt(-h~YxgT(4GPgK>F0)Hi-EK{a15GX&>`r>x96_^1V}SGH
z4Ea`tB}gv$HBs6sT%CoBs}6Z2oC02jan#l~D7Q`g<eRGA^@vrDY1851;YoD#;l&7(
z)X+6$V)~)VLW@yF4V`Z>k|eo$)v%6X7yelE5{(J|iP#PKiqt5=9#!DKoku9WJOhQi
z2571WKVYro6>!dqp{Qbky`!=i`u5BKmnRQM7ZjC(d|60o;=~Ky)6}%sPq%ZL8A_J}
z3gpO3zp{?JQ<=Ftpqo-}fVnG!Pr3`W)8;KsJSVK<g^M@5b18uDvX643tquaxS3@le
z5oa|oX7VXWCt?mYIpQs;{Z<cr)q=N=2I&h#d+YupQuuin)V%I7?L8=kVPoT{^JAPl
z=tJ8Ryd82J;IoZeW!2T1abOG<DXCu8Ol1+L7f;-T`$;jNT%GRT1H$!WDLA0IIBF5c
z_;11SqYRet+}!GAUdrL2jgzW!cNfq9H6g=CqRV_%izkLvbf&GhKy6L{Q0H<2jw10P
zvZxEomoGUPWIVx<dXkZGrgqGB2oJxacaS@AwSe1Yr(vlG>ck&Si&Z<Hg)mWrkEXp|
znIFhfquxy4`<)8dqC?OM%X)U6*{Po<mz8TN=YJ5h7I8LtPL`rLn>++Cwle|H{zAHx
zbwyJ`4E4a%Z>DU58jRg<Ie!Yzl6)I_t!LyWA{MX29|^)o62ZQ{TjG1}L`emsS2;R9
zGO8jkCdZJBI@fECL7>3())WDG*`_2b8wdO_G7_yt<?053L~;L;{npl2VMf$-U~&6c
zeGuZ#Yup5M$NxDU(b1d23V(Qjb^_%=tc^nZND>xj#x_1NCQl!C(eVBKr`1?Lw<t32
z_kqasd+|ut>$aSFyGJ(WH6=#pe=5xfpm!5aU5u(-NOcAxh#l>=1I*NbMlyOX_I-)P
zWRFBx3w86IqHV=okQ+cvEmT`Y!N@b7yFellU^4KIGj1zhYq;<pgxiyyC)KU~LzT$4
zR{<)W=VudRC0@y6$e>J#1z_waC@d@em@Q~<6ZDw~qjiv<p8-wM%9!@c<}00OH4}Kh
z{#;4%9G7xz9adZ&JS{gjCi0l!9zg-L2a&^=47g_gSpp4ZrwCyW+kvZg7dOqMB00y+
zg#TU;4~xNu0X+VUc?R2W&zkpj1LT)m6{oz<smAoMer4oXB@aFR`gc++s4AJI`q<Q=
z0*0Xq$#fN=eGti=&+AKl+$73)-g!y&J<zqwe&v&Kq-q@|2gvS$O)BM3cxoapI}6!G
z-M#`oJPOJ~9Aqn2f{nc}<U1Z6PLz(rc?HG{ljLMaw>z6~@lY9GG3LM&<E6-g$%i0m
z=kr?N<bo2rP;238EM{U@WLE7CT8runeOU}iRlqhAl^refksuc8AR$NLkxD~hRuPBj
zT@b9VXxe?Y-3<Bfdwzfn`iJ|O32>r^Y+kw<Ke3~Qta+5MlYl5_hG^d6?9G4EQ*m)?
z0Y!!xf9$vol`RbZXVp@<J2Uw&A_O=w1f4p4!jmxaAK7x|iQ*f;<Wy=3Et;j4Z#&}4
zeswHO+(h3Ty;`YR|4px~L<Ozb$As@k7JNYKmXV#^7M3837hrUm>kR3L_<cU!pab`O
zCGk2_H8h)7fP*Wcf(=zRo<s)0ABFJnHlHkUclT^vTd4no5;U|;prG43s2Bu4i0Wix
ze_;v8G!WDjiJi;K^++{c-C`r<?Cfm$xdcYI5#jjb9?<w3)|U(bPZg;E3a{VA{IS5l
z5ra6sysdRy5|*c5BhcFjjQD%-%gL6T00A$64!!9H5JPFDV=DX>t7q(C8!*q3Z=}j?
z>Lk%&u&mKk(rZ)KSZJ5WSW!j1n3zw1OqaI6Fn|y2&7BB9X^NRqyD?Ef;y?J`$@JIT
z)D-%j#Ms?ew+1p)+rLZMnIK6jK9rd?!&!LcuC1NopIzf*x2Nq!^3HvU7COVRaqt8x
z%mI32b(2Yw{9>zb*fSCeuClPGNhdJ>*@qcH$Gd|wrqheQPY;uHCX#p&JRjN;9udw%
z6;aZx6-T#++anqy5jd8~${@tzr7jghA>wI-DR1ywL01veP232eev)-I{>6;kLe4Ay
zKDcGe=VnMEfTQTKfoEo2y7m0bI&Yp>4Zp{X7->NmGrG~@tG;dC_N?*V4c|LT>$=kQ
z^*qU@e>qNDon&pdH$xytPrr3sO5zWwlNW%?Jx%sxn{qdUyGhgDXZ827&4jSA64lOg
zG=H$qU8wOo2Dhs-rUYtmmV8^p&nd}Un&#MgavM=eytv_pR5-7UN`--x3OBkxunI7i
zWpAOFNazv@{y?|OV|8>D3QCS*N{D{0CJhbXu4oimufR->)PlPkT&Q3B4F{N*bnbCV
zBC_MUr7S5!X^hK<^cCf?)Id)gUjMK91Bbii;&m!01M7dPHQ;W9f$&*z)W?%9`y$^w
zHlM}KR(;B5jqjGmR->ntBYIr_E+HqfZry6egf7GW?KwZJ)y>tjePZT-d{-L<AH4B3
zpDWcnG*}@ZgCy?l&l%R2r)gD{T{_ZTrKgukx8L7mY@9PlGuyA<s9XIC5t}bG5Q_Zs
z3^zO}mjR&&x~XU8Bh9m_5~m5iaD6~Waj|i~{}A3%ox>wgypT|)svZjA&Wjuq269Mf
z{*i}DXuXz$xCuDeYKxHzI;yQ<E{cJ45OD}2mndG4+tJanRW5BdZlP(D=&rN(i08}I
zmug~|$jA#w*F^QWG?w9&!LJJugj?%7h5uDNVArSM<=(+t{^;ZT%nR=i_w)eSA#G8Z
zg}&^<CbVzaUjh<-{aGihAnPA<jwHN{M@(re?9{D0c++vG12sNqFPy$DE%Y?-#>v`~
z8$qm34?)pKH=>wKi9bN*KYvSnp+Kdm=U<K=NQ;3>QBZ(afB`4{rT_KEMyw(I_#qL}
zt9P%o!%OtUjeUpoy)YaZZ=Z4YTV-H!bXq@RGA>xOKMna~01RhkVayv8#bLyI^qJ8m
zD1wQ+qT||D%roBYT^DcrR#*s9)hvy#dx9lxZ0&33C?fl3JigAM?t`tMIh`b}jQAh7
z{(o0|i#MPu`D@JTVHN@4!F(+3MU(z9<w$qcGp=`wi~JsQ&8P{jYgdT4z#`-`22jvh
zO<atHN8j^tdD$vR#wvZL5F7bieXGgz#CG<?Zi=4f3yxXX=FucFbk=YJV;rAj=QE_B
z`nWON!Bz~FzC-1eu8*80bD}$rjqF!Jr+0EuQvJ3+*n%Hn7zjH$vgAJo0CZ;rx8t%L
z@M%2sY#|qovDKKAQ5nlj91PXcwe`2JBgH2992pS04i1%NvIRkG!#<AVb|_ZKdU0&X
z)(K#aI}_yNlarSA8ApC}k-05w<*eKLYv%tCVZ_b}u#UA7gJTfs_<ofANDgAVjo0_@
zZQcCEJBnV%b{G?}Bj#8YvAs?V99*)?bE^bEsF{K9maGaI+Gj{|02)>ske5={a~b*;
z5`=v5<^j?JXb>WO6he`RIen3s`uPxHscj0i@n)fFj3+#h;y<D#^7TY;H<d@|j@cH2
zs?6xhWpXXw$tb8)wxNma(nU(I9aPR$4S0Ufsr!HsG<Sd-2sxgeG7@$4%rVbdh<$14
z6>Go*%R9)D?)PspBYj`Lej-3<%-Bw;zjv$Ld~iAcw{yu5dw8}-=Dl7{0Z_^AcIiMp
zY<Rwvy*MV5HQnT=jj-C1#};tXp|lUG^qHVxthb_$)~;&Hrx(|S0P)-@zOn0c9hFR~
zHGh5lI6Rb|7XPrcvBz0Nu$@2Ny$i6e0Ci1_xNCyx!MZY1InBGk;lk#Ky{Fn5Pyl%7
z0VhG)#?cPj24H(Gck^t`bXKOJV2PN#fIN#=DRFTYrX}E7vI>s4T@k&4Jt#-AMg@ie
z<CLN=g{|1q+58?474i4LQK@5ss6U;6XA0uizV^5e6N3VzGBP@vtcg~gqN_Euu*1qP
zI@7Gw<yB)nB^+MT1lY~6bjq1XVn=ESnd~3Q&7ovwGPsUoqGlgOxkBFEG`l_vlcd!L
ztgNm?d3X}cVqz;E<9Y1~?<l+2tX()BcYhG(5`w)b%C80Q@)v-gM~|hOh6I$I|390C
z(hm#_LM<5l$9XhrH2!ZT*VZ&0zETg_G<t0AZp<|W8lL4YatS*6*$vdl0!m)MjS_vw
z_Po08T(w{6fxr&gN>GF<$O3FW0W3z`m7Lyh-@X;+4LN;+oZJ_r=H6Bm$1LYcj2%w*
z6aJ|78gB-fNL%#?7VW)!-d<zZ^Xc<4?-ul4u~fih#@ekwjYP%waE_7)?Kv|V-jEC;
ze6mlB!BnHCv`5M&iIdjSUZI7_xa{KN<D;Rc_d?2f%EemSj<paFQm9N?8+=t@2o~QM
znHbeI$L%XC0@Qz=zr9tCrHFR+wi~X6ihEF`vPDqI5Qb%TbvByiuBqE@Q^+i$FI!Z+
z#l!An&Kv-q!R`3hY4z{cOTfaA2G>@+pN7u0>MPOg_i>m#**okeoYHB4?0SV}A%Tzy
zwL>TPMe<_nX12|siME^ez@5E!b=baCNl*v|3Y^3t-^Hskv-41t0kr*G+TWr>;j=vH
zWq6C`3|_GK$;f@P{y;vsUjlC0j^y^GT+W1M<fC}-Q7-$6=GO2MjFI1NRkVcYhy=K!
zl7nYySRsSm9n;Ip`@liLRCkL^o<0_@2x49%mO{Y2BcVpg>g-v+AkiLU%C4R}6=c2P
z=}ASDBIt*h$C8MtgY6ChQ2a?}KP)hdBiyNz+y8$D45$&85^xwhIgJrb6t)T>24*x1
zkps}%iXKJ>%C|ZaVA@p9r%xYJ_ELIgf9REx<%xUs&C6P5Wt%=TNS306t4i1Io1M0Q
z^i(j<ChFvgDk6(Jn*{b%Qmx11?sJex0RXl@B5GU4ok6ksSA+}8HB^S!N3-;j)Ca2`
z^fq5}onGa;OD5>`NC9IFz#9~N9%>4UD;i*gy!&0Es1*$2;pQ$y1y}#vAkN=<9#Mo0
z`m;_{$G6Zi{e8w}4u2#ylzyX%DbUn}f-*@_F=}&E&`;mF!GD&2utj9<H&MX}Kfol~
ziX{E{e_910ph%?RfQX+c0qv`stw_qpH+dKrk82y=#ZvsHR;%P1Y;Ji_VVucX@~zJ|
zNb^6t=KYf{n6vG(KujBP$ChRCCfc?^TKnjy;lA$L=tEI~<}IuVH{g8rjLlL1VeoSY
zpkB6)w)7`R@!bHDm<fMN8GY6RKujF0Ce1MdOlo?`@#kDx-NNaftgG#!??2;6_~`-t
zQb>Skmn1?!Ob{;aCUddNq}UgG^MX#4bQr3DW*4ayI3)jHHQkJ;+i#oFgs2!58h?&Y
zRLx~4mlAr4Ozr>U7=U-osd7P-z)t>CxRehHiR6(^<v<1H(SY*ea346sWPbOlN8AH{
zqSIaAscT=rqpeenzU+Xz%A0!Q+TT%vf%6dH%U>=Lf@ZE9YS63lXKSDOIKGx^SZ;Cc
z*N^yBy5j$B*)vMiME#dj^>y;XK!bFfUa)CvNv(SvS&Y+Cvqa>pal0H9=S*eTQ53t1
z`T+UTlSfageURz#;;6G*+M^8p{YO*t>9(T;g==_Myuv{UJHqk+{Kk(p-V4XdvW)PU
z{}zG=F#YE5TVrkftvvi(B-%I$1N;qeE^U@b7$Qs&(}y4nPUEDPovj~)*XMI4a|*<X
zK-?C?otk_s!));KS*@YG<B&3kBbeN2KhCpj6;KF}rtl$_X|1-QjAuK&tgIe6;CXQI
zey2D({CJi&SgyX(>_^ObUyX`7hNa<0YwjT;>9jcaV9m0*>BxtNiAyJ}K_feONybYZ
zgh50WcGS>x-}chuf>eP9Wr=WUN~?9;q+GXs(oTx9$eW$BM(<0fO3OD4gP7-ky+Hpd
zatVc-`J|DOPuj_kRPb2{;JGqN#}_<$v-x$t6FHRt|9|a<@6!UKS;1CK04M@|{1#Ja
zCjs{uyW%<*&ItTyBO<MU5e~-pYJz+~$RV#%Ui{^~AKL+Cor9J)wSuMKYWxj7!O_&u
z4+}C`YP%Vt=UcUJTk`Xj0z`x=3{D=Q<_Jq!I+(R^+zDCy`Q7PBSm;c=+5xzZWk;-q
zoG8I3nvqaY5F#HmyH=@QPb6f_v49^^B4~A+{h;+|Zn?yGo&hC6dYqbAagmv->Ajw#
zojQJl)mI8A1hw`UpkMkK7~A6g6>#HKORk`S76BRxQHoYVi1z`e^(8YK?kI0|am{h)
z9;P%SY(T7eF1?{ldeu4H>g3-6vnPQM&$u4|@bSZTW@pm@sG=u#-uJxa(PAI)3djI$
zcBjK_UZtt8I8>B^du%9?(|fx`1-i@t=7J4rdXL9%@Z$S~co4eV4ZT>V<Q`ijPLwuh
zu_K_(!P$TQ9$-UQ3oi@xsSk)`05W4lNWfT`7H3>L@Tqm+%q`p8oi+PIm=4M9%fQaQ
zHSX;UrJ?V`5)Ij9jQBg$+gDQ7F3VSlBs*V52R=j-?qmwTuELxe5_?q^@>v5W8GU)_
zBfn`IG_>kh3HZwa)(|lG`d~g+Db@f`PjQ&*Q|@2I`8BlLwNWw_K<ZFVk{S`llTFKO
z7H(zDnHpgc4;4&ktLWtA{QHZ+$7TEY1MjLQi<B#Ejh8Xn?Uy9nT<rgzc6O<|zE)GC
z`fJ9)4!ha&%G%a_z+ML|*btf}gV>C39n6nlf?pc=qHs2C?`O^L2QaxTt2y+_v#LV<
zEwpuY{7ct9KG|$6Ajsnl_^1H3$Ot|dOy8`1ZJbonSYkU#H?qcyplb1l0Dtk7(^c?G
z_Y2w<pJ@@EF5S#b$?03@z^yl))BR&zKebM;H5Ufqe(1Neu`DOJI8BN9>O&v#L(o30
z3eI`Hc-XA?GYFQ<`CMfJ4IRknpj^54_g#Ga<G?EtJ>883O6#X$Z~}_hm9EeeEH0Ps
zda{bf0Y;JPP9TukqO-Ov)83${12~91vT8C-T~jM8&hG@CGBXU&=TYQEyM)9f!<piR
znS}i5^kK*7bB}jjg(KQ65OeQ1=Yk$?5=*o{=-{Ip@QUE-^65Usf#DnCr3ZsAPzAAu
zrZ&VmaDP<lx6u_{MiU)4r=3O!<b1x@gNibNFM6p}dBn(gV(fJOv`mZu>|ATRpfHVj
z=0*5om)_ED+FV&uu1?&$F~o|M=x2ECX`bOmpy)OK<=l!N=em-1PNwv7Kx8qM1|SzM
zzprZaD*j!~Qc(*a@1UT4zcrQ1J}$9xgX*85D*=7odAOSXW;`RfoF4!SsNV=o3_}{0
zzDvIA0f`$Z8rbrBgnUn;19}$Wuq5{Gq{Q9RgH-}5+eN#Kqwcnl5khi18%xg9YVpEK
zh$2Iq?^}Z}4<=p4KxnU7+Ix?ic^3wwMn_xaJ5pQRJ}~$9nJf9rG&WJ%wuL(b0Y~c;
zKmn-m3qRs`@L4>vJ>1$Q7Q6S;pmmUq=zkXQK_gn6`m+=Byy~O=VpC4jzM<~S1Gxz-
zZ_B3&OZYE^UA%4JLy6U09{dk>y(Zv3v%obQZVBW!*UrRugJicwWn~dSzO7U@lBW2G
zPKRqUIV!ulUqJu%nZxtccx_8`NXGG&E3Om3Toh=0>p_qhO4C8{<AX&>FY(7YKnN&2
zdvMG~`>VTYk!z)WJFj9FZQHw}`%?+YnkGAm_Msc7C>3STRbFO5+PI3eua#D%s<Ki!
zl4CHAQdiv`04~X6`%~B$lf*ZOlw3X1-e7I*?W+WEj$1<zQtZ#O(%$RqsnO4L+iFLb
zd>+!B6OXoCxD2r?2;-@ClN+3qU=IEaJ&btRPvwY*ET#@St4#kh|8iJ6%6a8q|03{T
zio*&dk+TMS>IZQNX35b*0*cUg-hg~9B=JUoth42mP!a4s6Z4GpeS#SuaOtHcnB5oy
z?iwIkCiK*eb?&>@!S`t*o){o*%!SXAY@N234Ex!D6JTN#v?SX_)-F7dmP6DAO9$)&
z{ON4Mj^pIMjb!5NE{VGz{|HtF)^&8$6fmsMlM5`<JW_NPkM2xbd)btPf$yxMMvieB
z_K(5Iw2No{M8hCgK#Q$%*n$ieDVz}Fr58oB=JHWCUb#T_&yTyf-7fAAUV{fgdi#->
zW6o)F7*xdUVeb|5KFJTC)uCw=KY2e7pt04V>Grl;-e`oB`b}SjRtBTtxy9oQhDXQ8
zNpzK_uZur_M58tik=GCkY@Y+IpEg9l)hCbqdnP$Uu`RMZChEm3$&g6572u(DIgYQ3
z8d4R9uBoc7XJIxu->y^FmQx;n9SYK4<&UhF<4ZJODd%{^ZZ@;!aufr*iGx`N)QX`4
zX&PeE8(bvToG&E;$W|EL+KrgO&2aAw*v&bx!hoEJI{i(~{o6OMKPm)%^au&EJNCk#
zHqrs(hhj7v0;9XWh|yM5oRJ7%LTg~SEQE}m+_bDL{OCyXe>V8!=8=!Sb|HnZSc$=j
zn{-?18ZHE@AsD=tB=@o<R<9Xw)<&BW*itO8*)@*9bFoW&N9m*<vxeim!wt>eKXh|F
z$Pz7HTaOAECW8nyBO1a0zaBr#vvnV|D0N#TC&a0dh!h<c380-+A?qu)f%me6)70Wf
zMpW%RM#tYdv)qlE434ZK<jZZ9{@#fPWw&ymbz@~GH<%rb=As-Ik_Dn&PIu=NO<phq
z$$xtWGD22H(eiEz4e#t*-CSmgCBGqZ?6@a#|1NsoepZEV(LEiQ`q=Nlu4PHEQu%aq
zK!4xx4flgJ`{6iGR%{f%zheAA(=Un?tPI%rUR>86^=9?Z;2*~63NG@y?*uUCUT1M=
zfXx^l`3Ujs)j_Y|CJ=IPihXR1jAVitNNe7k;g1I{CvP#&ej%o8DME`ivS|3Z74(UR
zgxH3~W?bkYqz)7M569Km=JcTV`qmZG`0&6;tmi#|#`h_urU1%?hXk>zik2w=k5?A~
z4EBy`W9h0nFHik5U<?yD9DN0?`34NXX^J?X$CrWCEYx#tfzD?S(lCl)vARFSAMM-$
zxIEyWhOE68E>j$DB!ed~;V#dcjqAz4sY!3Ai*zFYmfn5W{*5>4rm|R!Ai1~K+2Wyw
zjTX}#c%d1+vhfQw-hdD4GcWn1oEE^V)5YT8DEyNZ1>Nk6T(9z}2Q@SnF_tgL)1)2$
zov<^|m#-Q9%1d2wr*G?e_v#P_S$;PA9lT3TBjA;Ps|_e$Zx}oYgy=ND7Mp=S@4mrz
zj`P&g<2>h;U8UH*OVXbum<_IXTzQ?r!7dPTo-IOM)44Yjn@fE=hld_`^^NPDTQ+bJ
z<Tg5b#;tU?A!7j(r|{PNEos~NZ=YU3w9rWdt^^-#CYXiN6=m#XS;1H?=<qK;0wz$@
zN~dY#7C=TSR@13|K8HWu?v+>FtMCurMVJ2xK7_|l321~80bxn%w3G{wu9e$-#Ca)y
zpt?B|TRkj90+f?eMeL>*IuC+lS;EQWTe^^GM?!m36aO$GB&p)pG@_)hCORG&1c47&
z;UHmcX|481jO&9N<|OMe4?bmtLIR=>S%-XN0q5Qn2u=sA1Tb$<K&S699RYS|@@MU2
z-GZvkX&@_Z;%tH<-tLtEf6=Uf>FOglhBc5Riml;bBP}P6gR*fYz{ShI03`wD6OjOp
z6t*XWy(ytBC#{HJc}9R=C_4?gl7n^GEj4#Dn2bGHMz`y^ZfXFm06`c~H~?FfcsJXN
zS5Vn<t+x9M&5!&zj#S0;T=3Y;m8Ni8dll$rtMCN<$qYX`7uQGUzsY-O-K|bnw0AN6
z{?ot;(|{U~JWcC4n7<!F_TstLuNfEH8UZ6ZlUT&Hj{aY_^}{*h>+EM8@0<tjcuxG5
z!K;tYS4ghwHy17q@xQ<BxVwVc-1W621L>N~;4Bw9eC{!FQnw7?EwZctMz9Yvdw?pU
z!e<`PLxRCv+brOs7N)W0IpIzR&nK9+_|so@da)mpJ6{|BDJGIp;%zIQDn;&;m6qB5
z+S$q6B_c(I7UkurJRp?dqfxIwWrdi;inJ;2Z?q47YuoF`4%I1Le*LCJl-G6Rxg*p!
zb1Oe?xjOHvH)fQldkCn0nO=YWkvxsg9}#ewb0A0af<vf%{yx8b*EjI(c~?O1eX3$P
zr#4Fe=>6^6AXMi+{_(6eJuRGnE`Hoed*0f58F=9>UjyFfPZsAttB&?_IYd=VN26ZC
zyFU#*c3Py<oinZ+;3{atxUUrUD3j`0{Z!xJpb?k$QFf`6l7oX|YH5i>Q|HU#KrFe&
zQ}TD5l$B2;q~k{iXp^@ujK-x0`-lN)rJ`FMAj@Yyv>Lxiyx+_k&D5Bjo@SGhl2Qn@
zv$Gpj5T^_M5<4&;v%PyD26Qh2GoF~EBcl-!5#gx=3Q8CmgweMv%;k;=B$01BoEAE)
zRA+#<TR-}iXj7T42$K5Nan4_p6ZuK)TzR%*m(;DrDTSf)A*A4;{-^bG+|)Mv=|gc1
z<E8OO&v5WB1=*^f1zO(=g^Ab8LmTy$OxdT8TlJk5fm0g2{RCj32IUy*T5(d#S=qov
z5O%QHsJuX$UIek~gD=K11W3wrOqcFuTk^T`PUr%3eG4+X$1V5zcV*XUEZ4?##v`|c
z$CrsenngIPbjR~HuIUSE>eVrM(U<QolTl*hmU;uK#S6|q&Z;q*p#!pw4-#zeRIs4`
zpb(<wKb76fZ$VjolK-@in`2EJJEJopAG}pm@m$OqUpcIoWkq=cjWjNkm!N@DU!y(f
zB2*w7Rz<f?gZLsHGkH)Q^ddC9<?1SYu`2GgQNhyIM!n1M3IT{tr>n#aY3e0t%>UJ~
zv-1jwq}GXLi_^30X8{3|nz?(W>_7bp_l+m4d{<c6kpC&=MyM5f@S!V`IYU7DwyR!x
zj9%R#*l*?qY0TB@+1%>Uj;qP8qc~5txGuw9HwwjmjHYocZ8_x9TVE#hYHGXe_<_3J
zKFLzZN<Mx<Q^iThFU?=ou`HTDk9&s4T=TS2SI%dJ;Xv`bn8@UVK|S{YzM+l?V9i6v
z&4ri3ef&4lA9I48fewci9p3<X?Tto%f_w^n3Zw0Q{nhhQ35$O>scD|7@-0uST0R<M
z^*@O!FBQV#-eyG`GG%}ESi3l^K#C)l3k@5YT`(FV-6`|aqtcJdFL^M5u-OYBJH`O|
zWlS66uZUM4YD=kk|1!GCWm$0MEg91MZv~#4*l-3`WaIx<<X3e#kW=bD^rA6ph&<t$
zL!`-t)O94xW<Zl77^MiW#HzdKQY-3qG)R%cE}g@7;S&1Yhr-bsB>SfqW~1`XWUy&&
zzpm9ZD&2rp-xxu8;Ff)$=dCrOl?nB<M#n|X4zK+ZNxZcFiWkV9ED9^^Q46v?WvXek
zuC(*}Os$4OaQ0D}+P<>Gpe0N#PR2p18UMVxqlqiEIRr!K8_q)D=toyTEcIKc*>}Y+
z{^gJJiW{-Baa7T&vfuX;^ew|Sne{dcQ$@;Y?iSO!o?2?^`p`!#Pmj+1r9okHn;sII
z{3P;aPO)51t4G<g-oNJxW~Lz`9f<>9uc1gb)sIbaIMkif3l86ib@3vVCKeYnT9Upf
zz2%uWKh9Y8MdtV(v>=z|4<KFFzBwPas7fW5`>_PkA9aVqMO<gu3l(WCvjqz%iu6=S
zYB`<TBBHd$DMRpWV}6;pxrrMZW<5>YB8J3#-oRnAI$$#1f7bQKBc0Qr`MrRnW08bV
z?!fB{7Q$cw2hpDh9@^cXa>3%Lc^Jr`c4)w5;?2&Rh;H81E1^D$-WS-tFCatq<q8tZ
zwKb#ACXY~|a9BdZah1|}md?zaX+u%zcK`cZospzYw!c5d=EWSkzL2=eyHdZE*EwVR
zUOLnKX~kagty2vL)3V^2XwY~ZU7w;Hha^<f<I8T-uhT%9d^W}lBc>P&+c;WvXlNVy
zbY|3Nbta7>x4?>7r15gSljfJB)~$%`7-B1En(5-5dCV^6=-^yuw&&UYhs1mkPMG{y
zF9)BOt@4-7_EF2Q_iSOG>ef@zot+0)Zp}Pmq6qSfCpvOD12(2RFGGL7aIJQSnv~T`
zjpT7YS+{)7H3@Ra8*1k;na@|1cft4YI0sS`a65ENsPNbA&HC8hd!Hw<3##+jPoj<I
zKI1LRH^jz?hoa_l*ZOGw_J2G_bcQJljQ8+We0~mAPAZT0-jI~vFT644nySmA%gW0%
z%3dcRRDgv9{;e1EEVRG+T=hF3se@zrtThZ4kh~2<s7OgQ*W}Tvb)_5?AB;Tw5W80i
zMQ9khZ{}r?$RM-Vw8i#ncmF+080@vaPaCe)r)b!IudwcjZ*Wcv9n>y(Ca9NxgMf}~
zhzhqn!QS>cl|0K8m_i`e`<Em`LIvb(y8bA^Z7ucobK>+i`NRrYoqe7Az*PFC=B`Sl
zeh%bK(E-m9>!arAFMe>QT>R7y#}5!<w^CRSpOEB9W5}H}rZ~t;M~Sq)Q?YFCT?CA3
z;pS_JpCrilQ<ti!Ya~G;NM?MO{<*4Li?_HpMucqg2W}yuU%P*5q^U#YLqafuBQQI{
z9m>nflD#dl=mM_0?{1;VS~b0M67CXYf^<3Qe|$h*;`;i}aN@D2)es;rnG1QSqFwaU
zSH&Y?a>};|Xn@&i3>_9O_!}|y?=&cjH8enlo+OGa+>Gq))C)nAuFokvH4UP-$p*f^
zu|9dVtynr;zdRuazw;8lQvP@8v6JFAD&;mXxbyRE@QwEse3YO5@Xvt-#2hN=ju^eX
zXPNb@O?c9X)4hyK*?=DTtNTR9;pm@7mU?(57OgHTj0W7hKR+Yqrj`gizP5jF5H<VW
z`2R8WmQhiD-}|r<GAP0fjr7nU-3`OgU4nFqba%r5LyI(m!jKArD4h~YNq2X5OFy@t
z@A|F(vlef7G55Ofea_kY+Sk7JJ_{F`eV)8KsJ`ZMCb)sR=(sMmersc}_aDEPbVkPZ
z>L1)7&z~n}nd(#{<=Q{GxYKnEe{=c$9O2?dw{;VVyl>cb9(Hj7xqEl!*vPxRi0Du%
zYw$m<$e(?2jAeb<i+uLX7nx8lEOH%)dgpcwoBY)wq_wm)-SAZw8xO3Th6$#MqXXdr
z<Im-ASX{F;+mY1A|1<0P;bF1(?qw%Z4TNU{`~95Wm56}=sn8g0U~o32Duf?4FsMk#
zSw^)}Rp7+!=(K=BKtbZ4-{CUh|F=t39_<CqnT4EOgugn=8%V{++bQLpUurnKe#f@H
z{WxZJcHM6=)%>6FBQr79Zo)V8O*IGC^=2=3!B}QzTTU;nKkoCuP=e^^vxx251S?Vy
zO}eDF;~9;Kpt{KPF!=|sdY!oc(*h)e`naLPW)3=S#&@EYa8y0fN3fgUG%?~&K$N_O
z-`CZ1O{ZMC4GQUB%ipwUxSiYE9QWs5K#&TbF{_%LqnssZ|L8qK`D)v+g?0ygVlK$;
z3Ae&1o>gn)&l^Gyee3T217CdUTWAPc+wVFj2YEHIZ@9D@4%`J6J32c%dn#GZ%*OT6
zdq<xB7khbap=~8wh(KsCcz_+L%nT~DwpPlPHhhkUHySpB={m0!B%G-roEDRc-d&~~
zaV(PbpAC9!y;MRhnW<{d8~CJr8y8bv8pxEl)j0gKk2xu0{?V*Ly?g}&KEn@=6vNY=
z6B2IX#%r(7n91oSi}UKKzZQ^z$DhAutFQku-=W4G{xh<JuwIQ8i_H4$PKQ!uE+~@k
z7Z-VR*KYevM?rY)X**#bp*+1-eOlS~Vovd#pwR&JIC2mN^D4^j&`}S!z~}EeLA&?m
zM^_SEA)-(U>9rnngzcder*6#O{rGooa^Nte1+d;=<L2hE1rZUeY*P-33R1MVtz;Fm
z6DL72z5<LCozwjcE}s%n1UUuu0W(1wyD=K%0*~aKt1$yXtxC6O>*;adBq8GeEMSgi
z(HK?qWUBK(5Yv^4`+o;crA-3~O(LsKAuvkx*`r4Z`xzIieSCp8+xCOkkjMieRh~cX
zS83&Ab&Eb1c@TP#(iv45Mj;`$*5+Yn`>tjEF?iwmY?yesKS7!?Um%21*!@ZKcbiXH
z7Slyt`?1)8lQZ)#R_WU-3vk7lzYe+8$Y7r;m!@mEx?+pp4f}8ZZk?A(zwIkrOA%^u
zqLCz4F_e<;|5)hu%sgh@hjyhXRaJY}K-3T|B2k7c`35u^^Bw(j?d`9esXKQ4x)1|H
zgVdvM?l%dV=1Y4<LMF%EY(9UZse?9cYj|&{NGWpPui)!ORj#qXKGM_8O)c-yQ1oXD
zYYWH!?g+!IR_ct|1fRSw>Hpo~ldbSSb3ERzj_IdErN0OCHy`{6Qu_Gan*X1#f9i@c
zIK}6s>LR{!ECRWUHA$aB!K}Iz{KaEEZz&-B84>sQ_nn34VWHRVRAQzP+~MDn6!W@?
zeYs@2Yi%w3jv5R&KDr21Dny=>H(Y%7RiRtQ%QpP(;(pBh9YM%@h+FS*X3VBrS4VF0
zKzDb(ipQMS?S(FVD-3{%nd!;Ajr<#C;M!01W`5SRN6r#-?y59gorZ$fw;g<Zd@8jc
z-FT}n9rSL1Tdjg%5rT7u3Y&G0!MfdU7{fyDAYYkf-$8~f8zwm4njG{)9D{li`zznU
z3cKHNq)8?E(c+#}j<c`aj31E9a%M^%+3-$#Uddq&-QPQKi~mj#Hj>g(SYhsM<DUbJ
zb+>q<!$zL}87(=a2@*{~OOyTAx>kP&bbLVn!Lq=e7}6Yt9EdIOC??_689zQ@h2&IR
zhC*&_2=)#sCvHn$m(jhyx-}S-G%zudU{bu2lFFW9i=Mfmt9Qp=uOs86qcUg8Ub%M&
z{S7qcoSY=+mFUT<b`kb<4M++ASh7wSQS(El+nkI-@K8v%g`O}ID|qTGDx(ReFJXB_
z#?3(%5ac@Zb!_({0pDxy2#|yMo|0vHf=Ec5ja3>MT}wL&Sod6<!Q<KQr;G%FTNYRs
z-%U^Yxc$#gzZPmAdy*w8|CVS?n;@P6nmlG~&dW0CdoikOY5+V*P9T?CiwSjzy>YSf
z$%u@U$A2m&hXLyrbvK!!>dSN^`QPy>$fTS&SWtvZs(klvT&$v#u6JDpRlY+K-|Q8O
zNVTSVn9S^mZT~+z2DsthX-G29nw_02o-;Ai-sJ{%AaWEfCh&(ZH7r!MNAm!#rzL`t
zR;>!wQf>2NbxEnduc?0n#I2Lf*o6m=vy<*sw90--MyIRW9saampft<mU`n~Km|a=_
z0x2x70g13h!0ckIKbWZ32|Itxb)0ACUf%d-BuKZy_mn82q;BhH4~4|dQAQxO#4cCE
zQr<Hiam~e9^+s`4LRzbA64AXZL3nncDwu;~LF9uE^j*(kvdWyTy?tysZ@m*u!W0_h
zY7MNEQrxYKI<2~FwmCGJsSK3oyBzfSMesx1jWaJ$hqRGJ`)$YXc&vlLN0+w!O}=CE
z<FdfSiNp=T7^Wj1+1!X{cDBg_)MHJaZ2dXBf*XF)9;_USi@^ljE-e4ED|pzi6r$#i
z$k<Vu-kjuG_z#@<N0J{wp$#LOH<rk)W5UbS-N4B>$kPZlMpSb32fOHdQm~23kB|2v
z6u8bG1K5B5hv1UMPuS<O7fUR@(8VLA?^N}A<wtait5%e*8t6LN;2!jloD8$Y^NCwd
zW0o!!e-$w>$AS22dD$BM3);kFX2otH)2i?i<{q=lfdxt>8hlx}8+}2VV{y!7yRWet
zyEuD(8ZCZX5+5Psix;@E^NOCjo|?x`wd<m}@VMq@s`**D;_V1NeX~D4Au|t=R4Avb
z03p00kG`3YG6It7eBV>Mx5Mq2Ee9kOKM_3Lp5gm<7*Cv$ypnox(-na)$Hb@WbNBx2
z*k!#EcKm}D1^K<_xPOtUB@uCbb6Nb-EG;$d%YM8R!nVo9XF(uBicg3ucS5?%&IwSp
z%OGW%H5NVEI2is5fg`e(G93r}s;WhU2*FPh>TtbZ(}60jEXpLbltAim-eS)P==&cO
z;pskbT2A@h!TN@|!sL}Tvj2X~hy+{!kZT10!WFCq(hXPfmY};Y*Vm274pb985~5tc
z7p8`mKCM}1^Oc(ztHEtDn#Ha<SJx+~`pESvBnJHvzat|xf!Uy@50$_>c;gS>P|^lf
zerdT+<kH>Rt#$zReiw3!KG3AqNZ0hxS+EzJ0H%hb@OGNj)BWBFQzB8{%Wx<Tw`Qq+
z%!v=l=xV>gSfI+A@jCNlUM`k7P}P4GEwTI!AIWtdBeih)-1iy~Om%wKH=X_5C(6M~
z9GcyvP}hzz&9Chj$&M*eh{zAmK-jvL=s%x$xajxSZI-;NQI}6#<eieK@Y*XHfof^L
z>~Q2Wh6j0FHoM|APgjTHer(AekoZecO*z5PXrimDtFt0BHaV$GF5nOslz1!Q(obVE
z^1V=5ZXXM+_x9>EwXZNZxV3j0okU$%_uHH%IQa<{maaF2_RGWlJyedUH$2hSh{r3V
z*(Z)^5Elb=PA>3eHDh%}#KX($vwM4SX``}mF%E+owrXivL3#OsW_=VZYmt_-GrhW&
zy&+qFrl?-E_UqRy&+(pjJ&&@fwVfuEZuFLGpIy~wUDRTR5_6cg#3?B&k6s&|HBxPR
z{O@QbOZwcXrj160U3@5$Am|L^PXIgjM+A)mYiSt0g?%GB?;U8a8+f<=*ei6PU8?A&
zyk3r(Fl@@bHZ_WcnUe^1^S{K%>OJ@g>nTJYX?WgopHrT)>$XftK@0Y85KeV8VhNNt
zrwrZWJ;Q-e?$%P(pGI#<8nFaR6!j*c%Ok_f66W%Md@0>@aHb9(dxE3($#0R2b&evN
zGSh>A|Io1r@9yAqc->~8Ik_m-Yz2M0A&bjtQFuf9EJ;T4MeW|v5t#C(^an~f!sPE1
zx{i?%{*MaF0};A|YiaM0ggSd`kE`+1!+fpcIDvy{DVw`{<()<>iNACP0`9`%b@$Q@
zxApe<))gZ6?jF}8c+ADan(m1@H?rk<+LI<<Fv_tcJi^@uf6A^V;QXrUIg~ilw)1Jx
z6pqEMC@+}2xF~57zj+e<QT&kOx_$KlvMO@YXc8V?{xtl_W(-7RNk@A6{mOd<WHo&`
z{@!Z5;De%n<mJep6+4mJp$@dzXA4JfZZ14oX&GnLh{r7NuAhIQ1b@LTXEuvMo}$;M
zFC8e#edltI-myNBkp9Vg^SaXFR|aM!P1mZ6JI?k^=&yG4v5AS1frJDBz1|t4&)PK^
zCc5L}<I}<-Pgnt#K&IhAuCR2m;6_D~7B$E?<<Wkv`efew?gQI8KY?zP5hLMKX;Qxo
zWLRspcSs?HQ^J4f_)Wmtes>((dW5h>eB82<@LZkWx%>nl|5rvRy--fnv0Fh&+h1|W
z{FYLtRu`+W6)O`25!3v>q&c`YOMPRH6t8;J;-)rrU@YZg0!YoCuh!)GTeg4C(|}OE
z5|1@mn9nQ9({S)_t0dsu1pi&#h_dLlyebPHM6ICEB~5vQ?dDmyp^!A!CID>bckv7R
z%-ncst6Hn_b<XASvRNklC=D*YrfLUcpgP{@jO^a{jn(`t>)=~BKZ*Px<0CBJ_!iuY
z6DF<bikxXMSdH}SxIT0+kf}I78(L0HJuB@vRGC_+bfn(B4}I3NCebcJXOKb&NaLE6
zNv1aUgt0XJ?a7-ffr~mUdV70ykirPxQ=*uqN90QPD{tR%x%bRq?Z0Cd)DfAbC0Nm}
zF-u|#%s2{nN&Tu{nLV2f7!a7`%?lacq@X>ZLB+a%U9`&B3!FU8rG#4&AulSWI{!N{
zyg`4@=lNEY>aF7{ivn#O-b(p|s;{rqzd=J{j9bIqiw3k#GuM9Yb9U}EH{1CChO>Z}
z!AP1CvmGLx2CMX94o4tCTmFHU`1;VeW_qEP?oIvI+ME|3?J3s(wh_{v)?7Ly#)v=6
zjmks~l;j2<?CM^{C43f^bF{3vyJ8V6J=vI%`o`Vf-_-kJ@!D*OWFXB4eV2jZSN;i&
zZv@ncoI*ZYS6qBda7sQE<)(suQKMkK*-chTz(n-qqK}mOt(jR!_IIr;Ep55jV$P!F
zBreDK2HimF&or9K%B6=82lMWa3A#&)UOA)#pHrMvf?nPx+8&64<iJ0!sH}h#^|qjK
zaZTxp4<g=Qt7hUR^C)zS;cmJ(UGv@wY7vV3UzYa#n?h$n9Mhap&k-`Xq>LNV9^!dm
znmp5c#cK3{g7yo?S0#O7RBzIcH^Xd{ly+G<k5026R<>hNT|ho&<wj}yr%7;eIALMG
zC4m~m-7ToItOJjr8Dk&0XolxO`SqWfcjM~Xbh58aT}_PW4uf4}I?P$w6Wm3x=jck*
zdF0=_N|4chMwLO?(J$V-^Al6c){**1!vKzeq75ndWb8bcs5_sX%VPp=e&_Az#s)e}
zvkDTO#$hZcdri3fJM6SaD3@!V6YFgZslQr9fcKTv3-?wUD>KH!WLR4r4lT~|{rucq
znUQCRtkA!>%7O#S8>%<hi^hWIDR(R%RoqQP!SX8W5=kR_9>tgHb@g2dz~Meai(ez=
z^~lLzT>pK+r4qCR3MKL90@shHx^bJ<^z;>g0dx+aP#c0_2pr>aJ{spN)AY+drFuSP
z-E4<|1SAUY(cRsht*|)4#r{yft{#`awZu;Ich-KLZ<nU8YsPhyj9u0jcUwcfSbWEy
zA48iJ`#w~vwQ>*kevipm7hZR_NHg@G`OnC2KP24G5A$T08Z@WGh~@2VsC`vr)*#b^
zbY_~r$Qzvu64>!+=l?tLL+`pF0L)tc-Qoq&IZ+~;Sf<{ilSW)G%1SEqu5N9<g~$G)
zgbO}7LegA$2T=4C!?((l#0+4vP@H}7y4iT4I12goFFd~4G_2LtHNV$ffK@7SQ`3Zq
z8hsBD^sp#8_-D|`cgY}FF$DQ2)&<B^{#aSZ-Mz}{44hQPFAJCQP(iu4ExY)bHhJQ7
zSfvA~BmOboR3iXhVhJ*t{y9D}Rg*+FVHnNA87BZe2EWkI6)fZEmV%{5$WH|(^Vip@
zbm{Oy*fugxsdA$PB;m#E!u^6IYEZ8;$Hek|FMs?r-K$tH?-F0SmMJv-R&uuQ`_rva
zX~QQ4o<eQ|N%<SgF5iD~u9+y&YfPIN8h*>e@Qu>KRww3sNdX7%jpAC89EvRY&zh$s
zJodk#09cVCnevF3n77*_3_7Cv%JTays`CUux-9E9`KUfTi2`1}`}>wCob*X3GT_P3
z^zwU3A$rqmX>#K96b(%wivB^_8O2WMKfaiE+6D1f401Q+ZXiv~%xJu*+ZcjhkG%hI
z`kY(E<b(<8_w$R7FlwA&_%-NT-mhar3fd<?fbL1p0H_kd?paMzdE{sK83u&?T+HY&
zzO&B)!mM4VbpVA1xKKvMXcOwMJWguQ<02x77z<zwm2w@XBPo4+-6*iA+A6ePlkz0c
zCVenrK%JdMU4MccZJ<%iduJ}g+VJ-ADeH!&tB^_usmAlDah|`s<*l-PX7^K9U99O4
zK=`)ZCu*X#slVM5g*cU0gCaiMbjpy(<;c~3@>5n>$b)kf&@A_?L_~C4_P9Y9pvLGI
z7s6IENu$%#vW||9c8jsZ92(^nv+DJ->qc4p>ohZU=9YU;CQQF!J|iAVV>d+g)kQ58
zwb%dXVDNWQX!#T3Wvrz~lAH`$q7p=N5{VGLy$cP;MkD&iz*(WsdpoWL+2cfGKwT`;
zc00IK{Cs*+Pp}rmg<)g(!-vQ%(=)w?Rgh~d8$;AbOW@2tKJi7Y@jwlVsJ{RZHx=%Y
z$Yy;qWvZfv7TQq=u79Oat+MQSd4}UL%Ev8!G`w`defFFIyLNN`!Z*$PY#;eL$a<5_
zBVd0vzpvsTKo^iIXup&J!#?J|2Wu)qKKn|(NgXT#nDmd*TBBLLFa8Z<wmqCxmGoa!
zPX!eu%nn0F512{DA_x<GIMNqF7f93b%jLO=7(57k6)!^qiztO%&@uz*=CUb8aHrH_
zNMQw?3~PMt0vS%ui&S)P81VnB2-lM`LZlBqc+44INC5E=wMf=Bfs2`myF9T+9yS(3
z)ORJ=_Rl)c&Wuaz7hiO|$JjMl(Qw1zt#>JO@I?G&hTIXf@pyFiYe#QMw2T1<lR(|?
zB@ID`QlL)U(5Ee>E8d`?H(@KDrv=y73*7u&?~nzEyoMNhi!^VXksn3}ogztZ)H2C_
z1axN8l_*StwO^LbBVx94S>RE+gIcF%ZB>{aR-V;aT~!4o2)`&O8l=btyG<3t{7nTt
zK)%HaPYiWycu=j=<o16vQUFLTCk7}(N1f3-!!CgGOwnb=8g7x+%A07p`cu0#R-Mjz
z%sL!YiQ6@RalDhoL33cXD7RPni&;Z7Sfao5+;u1-dX6@o_+?39LxTwadCY4$u2b@(
z$Z*@c>vKCZ3iLIcWE{Ez`UdAcp1Q8zF^hM$7>0CUbGwIEe~y8gua?A@X~7`vi~Ij*
z$d25@9k*8gNHXjR*9_svOhlcQ6Afy!j8}BRKP>?~o&bK>1|1{lI=v8u?uacHgwJot
z%2Rn{U9czx?F&lU_G|a|_Xmy%U`bDePX`2LSaEZrZ!ig2K~Df*3d5>IVG8oduL<Dv
zUC-FBmWwaUOsj!MZh|WNAQnRqi(<OY)shc%@OBsesaA;ri{^l;d#B{`Nvh_GXOy&P
z@i!tcicQclfjlx;2(~wRW9y2j!&>=B*IBn%KI$0+8%!65837DX1|6InMJNpO_&t39
zmPamb-I!@ClC*jw{-yvA4<!$zj`+HsngDKfM})ksUxZjR({%zb@+_d{*Ll!@U96$X
z2Q0D0XCD4uJC=^)o318J7aj;9{n}Bl)^mw<nFc5Gc>!MW-Nez>*r;=1Wn&VDQ$u;A
zI-|WzkxTa`x%p?R*@tSj^&*NA^)?_H3QDLnRli?;WFZZvarwh0|F`<JokV_tWtrHL
zL<~$8JN*WWV3U4kx-ndixd6^0Org~WBEuS%12{^o_U?!dx~e?Uk<HGkH#>7BQdZe;
za0*z$$;K3m$k$^92zOOgBJT-E%Bfmi+`)DOzDEIBg6C2hbv_AUKIY3@ziLV@x#EYm
z?Im|V+;so;lKN_RWfr&Ey(XyiD|(G#RM_Urmm;=f?ByRdPWI6vq2HS7>#ss$nF;-!
z$y`5Ie^pHlU1d%^ax;;R=_usx0j+%b(U%eMTKeez>rtcRJ~*Z(xqEJR)fGK6(9#gM
zzx&_WgNT{6{=S>|Nss>y@H);l_`{;-WSux~1<Ot1H9XW&H+#Yznb+a(*nss4m`goL
z>@jGAtk}cH_!cVF*{g2jkP_>!tHD>J6twS(Z<tv24tL>IFCNR5Rp5<KBMnV~3gc|*
zN#UcLFokpD{Q;ho?Ykv(%{3#v-Zr6G_2FglhkhH+ulLz9@ZiXwD@xOlGckfOGFYE;
z^p>My?{aoI6vgQ~fG13@Hc4GIo>`uPZOY+-kczt4gr8M>)a^Up3;@NJrv$qrn8&XU
z2Y>YDQF1_`pf^m7S`wd@^tW<@5|J$%+%_-AJ|%AX<%`emI~ubrdLVv!Mjd%)u&wV#
z55yuD@^_z5xfRAmin}3v-XSFjZA9H2B~{?2-UuPK*W^6Op8}|8x$|MWOm$Qd-J8gg
zwrua<;P)=}pF`S?UHst1aGwZ-&U#3?cM~wzz+L&U24S7r*oAM{gr=|h%P7uI&Y~fr
z!((hN-*QrK;%z8$jIggjHroyY8IKDI_dDN7U*$HQygg!Og7kjj9tTrsc4_&#>aTOQ
z6XJf(@H)m8^p?tw0g0RzPcr^lCjg>;d1y}}hF(&S`dWXW6*Y<Vk<XiFigIEv?1S>B
zPWz1p$ov;;NF!`jGU~ZLe=|f!M<2eGMGu#$vK^J<H%s{5m(Cfx;iM=a<pgIO(Cfrt
zQ25B!h;vX$^LL$@_mb(Yt)pgrv_sp(fFfK(9r9D+NbsfZ^5`D9aJ10aKU9z`d-Q35
z%ISy{KO$L~Febq~<Mz!L-J1-bUHh!8`&ocI(YPms>@%Wl{kU!^4d5obAlEN!>#NuX
zD*@03LYXJ<GcJg^1~sfFPXwyTdOJl27SDGtP!?YbemxWjcm^4`>2NIg&%s16p3TmY
zzw>zbrOQu>U8O6zu=&79WXtD4dv*w|E5C?dm;GbRhs*53KECTQOv-wB&&Kx>)JW&x
z$<t@&7o-T0BqWu+QS)^Pz#1JRtiU3jG#F4sY3RFQRK@P`t9gw})9njPm%O`3q@PEi
zM2YE7p*ZCI8pyR^;1gBip>62fMB<VxPg4E8;3SFry96*Ba_b2@IIqk#B2Xv2lpfVs
z6+&?6*t2`qEI%WBS=X|OJGqZ-ejo((GcdKRIP$`RLjlmFICHxCao^$1wr}10kK?PG
zAapc{Lqv=Gn=QYuqmEO_duTk#=>b&Sb43qzdK*&4Pvh;Emqs>k-}?GqWqV{xo{Ur-
zGTAM*>h;_304ehegs==LzY1GlQDK$1+au|+5RIxj3-k8x;FtfXY}Xu)z=NQ|cAp&B
zlQr!3N!&X8@ja7-30ulu%?_JGpFN9*QaMdZ)8c<xwRh*;3r)CF8<uu5I2alQEUAZ?
z)(MaGA+=rb%_YW-B*312^CnqEvlRR0MzYb_Qu<gBkQ&rbfoDaNhe?aW{Fxl-ZOtiU
zS=do=ekCLH*m1T-ZpO(6o(hPd;&EM=`kq-*sAMN2v6aAjIkhU_+xu54rDYo8=yV}j
zt<)FJ1yyxIa%MbE4V4X{-ws#siB$leFr)0dYU^}A1<f#dpE8ix3@^I~Z5y89bgbdc
z^dEbbw)ejQqL~l-d$ZgJQ)3qD<Bz0sxNr|9)CIxtege49hy`nf?`NT(sSTrBAbkg9
z#rRs*b_DHDM0v`CjljSCFaw}?&e!egMijIX7OAs!0Dm|SFy`FmzbI%qhM_VMUO(cV
z_Sif=-sG>7Zc@{g2MNIs#?{g)gkVPNO4~$TEza2R;IYk`(I3=^jT=*FgZ?x9!*SHC
zJ~wRWyhINzZ=FY(#_t~iW(z!bIQ<y~?NGwv2<Nv!I@Blp;d<;H#N8bgDu&HEz50{$
z_GxJ2bKK(D>Yso81F5}s-1O|EpI<stLIUHEPN!7x#e6xM>v*>H`j6+}njm>N*6ynD
zruJbYa?1{S9Sk((@B8pLjw!?Od}8J*25-ppdNnKGQP39cc-<5b&$X1FGV)+5No>*^
zl<FRPuLJw*X~}7M3Mf6~AuckeYis7nViWsh14l^^40RG+CvMmeEFJfF8NSJIT@~(N
zZ^Rpf);%`zS!pQrkrk|c+Gn=?r3}vYy8;R&MpH2Sqs@T7R1Qhe?2T6(eufHZ{mysf
zQJ7;wPOw9q6x<*sG~~_66EgEBn=XfB-ou0H^{**w3xxPs)A10w69dHj=K~$*;VzzD
z2YK*)%k<C;gyrmdoDaf}(G9xJS{J<ZF+>z7i;L0GQczM5l5!;m%)Q-DYD8A$qd6h1
z>Milt*(@@AFzTJPnrm-Qf{2Jn*v!YswFY9@xujLH1S=wea%cM`|3A0ISvKxW!;7WB
z_$V-OHpZNh(sYt`9A~%eicws=L!AZ%+{lTZW?wwP<dKnn`kM}Hdq4VA;?%T83gZR&
z3lR^Y8U!}BN4LnS_j9NC<Yr(L@xZYhrNwiKRnTz@Mx$)|#^@a^vq5X2B1fnoN1yv<
zn08U3l*DudHmdgNPO3B1<(yA;&V-3xWH3Z*L+NVT2Q7&X9zh(U&={vyi3v5XFxTsq
zV=oDDr+L9BBsQ~UWJoh|vwbPlHc;E^<<~n6TZ&qx=o<T{)p<ztE^K*4J)*`wvb>9X
znlfFEIVK{H^!B+lcWv;J30?G)mD1~F)az~iiw*Y+{N&Z$T-JK+5oXDaDJ2mLmmE&}
zLjcC(!P@DFBM{X~6m(W^_?LZkkc=`>+iQ&K5M0li&u979B$RqJhl_L5uF$Mab@tiv
z6<csTPe1@#61#;EE~{=?LZzWd#yKOcI4L43&8vhsj+`ys|I-3YX}>o3JhKlV<B<qF
zQ<YfaiIEs0wv-=_9I<IZyBzx7w1F#m&*;`Cj0-3%Ty9ESJm(!Bvx1*>=GjtBrb7}Z
za(p}LVExtvRS-Cp`=WU+`Q$U&jkrDw_fukqdm0SAXq%-ukPP#x=Nm4Sa*H|n4QI?W
zQPIW}uBUmRJfdb);?-9vzG_Uyy6yM%p-rv4|Co|YsUh#2{<{fuM8|T!A_A2xlQ;By
zx}2ngYa3J88GtdS>x^%4CI$}O3vZcZdtXlCrV1Ldm?~(68di0q4Jc?Fu;}$Nb-D8-
zfI)$mkDKdE8%5>v+_COBdJY!U;c}gI!>K&OQX`3FSMW`rWHxK%u=7j(Q+c0>z~8;6
zqokZRz_xLQH%2Q&v5<2<q+0n4y)D|kDs5^{z#4reD}rGxR(|ZNm3b<oP}e-6r+1U;
zSr{<ED#-fjJsk$hIFkam(Gw}yx+=}w&~v>*Ao1(IsF?clwY<b`Q{DzzUd7jFLfz3i
zKmJXOmu4pd3%vjzKMh1>iF|?#@g<P!A@Dy_^-8QHjiS6VKm^@q<;?I%>k9?a2Cjc)
zvrWtRyuqtDXolB3Q~5zDRzVdj@`-{_+lgIBzI{RwXY2Y)zptmyMQpZvUB|~*V&cWP
zzrGX*gU(p=SAHrIY@8Tv8`Xo)KFM2&fq~<35hc4+Tdz1Mkz62b@4lPze0<#}xb!Gi
z%|cl7{?=}83vCSc%?4)JKp$n&!vmN2)xwVwd#>-Mb-v&W)pe|$S=DBlyWzWB$9@qC
ztiw^z{J`YHo!`2J$mb1=)V{Yb&3<0iIxi?YdpYKY7jqW>z>{Fwf%jqoBS_w^7WJ>;
z)&pvR-shohPD0EZ6Mq=IhYG*5gn1p#6(X8wy^(gClUnNPiP!hF9}bF1YMP~D`$$H6
z+|`PFt50psg00>KiU~&zmB0Df(&_rP+tXI=90%WeIzBGU?<Xy06m}z=k>o-=@`fl6
zd@TAi5cjx?ZNn#+O*tKYLWEf_xN<iz62B;X9M}ioY@gx4$p=dN)>KC4#dEow8(OM#
z1dJAB->rRuOznCbIWXzsCPdM~*LDG5@=u(*VC(3=ja}6)?>{EEb^5p#dQj3X44F?Q
zC~2JG*$SOnzNB66p}#U%A{FOtyG|~<^2F|4rtlVmNm*s&tWi)aRZ*+sUWSY9ssze;
zkXBof4x<>Qrr~G1M%vM!xmT~6nH3Ego+V8>)Vl@bIDB+lxS*%}KJcvu5Upt<=|MiG
zN7Dl#i`~;M9m|&(n<>p&jbH!V)rczQXtu~bi#YX|A#QXJDBzdmRUD9dgP&Xr9nLB$
z(i8I<NUX56yAEdAgU(G2{Zzzu1L2efZXgG(3Sl|llOx-)4szR;$KcST&747Iy+4Zl
zvs)Ry(e0g#Zw~LLC_HyB%ZgX+@etdxjb&nM*_t;}`qlN?FhC0E$=7KRp)Kv+YoqBO
z9F;ZjN#IjL2EXArLEhBTqp;lXPV3>Cp`Zbp)&MDP-w$te2_j<NZXV#K#)59zByMXu
zB=$X*3iml#CMPWSg9|@Oqjnc>j=Av)E^yCAzK}S*5=J^v2gD@g_!}*N;010Kglr^0
zjo#c~t)~_(y)!%f&FC15c1j}%H|9CHplnShk3y+wgKo~vp?9mC&0Sxnj<K#MP4%Ww
z$g|Htzt(!B^cJUzv}*s{@u3)vm*}5Ql9K!NDeu>vEqG$<yOsiY6LJInFEwp#$=D6G
zii4gBEG#~Sf6SfiKlDzE_WymDjlXUMut4Td3Dg()$Aj2ZCtB=*?ZHj>)EI^N$>*ga
zJ$(c$&l%{9qZ+cvxHYRKj+z8WwOYH|SK{KM5NImbqaOri{=5r7!e7BKb)c6f+J$R+
zkx1{USabh%;#MXBBzB*&MN!KzT{D?gfz+n<xV_OX863;ST>Ul-NMp=u$3NKGmBJI)
zhDg4Zv20G__I<t5OQtv^vJcY9Xwp4D4>lbwfO5g&M5g7zowGaIXJQAh7fIA8Xa~l8
z;}b3o4`uAo^Zyh;7peK0=>)`f(lIG$alhaHO~XmF5(w5{Bt0Gz9yf!7rD&SWzELHl
zS#=UvNJRkSzlO$uxgi_1;%S$-;OfK0(v5fxi{ISb5#1qRhXqH~DS>3VZlr=&oq&q7
z<oN_?W!E41_WX#I7+w>b!GHC)t27$0d5oBf^pWUc%=f3ddfJ--f26ILZo7D37Srad
zU^(0Y3>+c4%cpZ-(mfLa=huJgC)LP5koVywb3W-3ge8IQxzdRDns`OM_jNc}$9fQq
z<aj2u9F#X!S{*H2T(ThLWanp>w_2YlWEimwzxeLl2H8k+14Si+H7sS1L7zceF((Ni
z)Vvw7d|+zFqs0<RM4F5p<W_#8mkOebH2>c9Neug}(2KDN*BhSQwwGyHX6zg084XJb
zlx%AtVNd!QM9G^;+AP>}t&*MZtVaFJzl2Q$JYMJLndFrm7jfJIN#AN99a(sFd_w&i
zfu6>{etS<{dzdaBjT;H}O90FG8MVw519wqgCEg(6UJ@LO8tw_;#ZlpEtx(Y_Oce>!
z+gP_;+=$-)Qd$4tR%iXexi{JKJRpKM->7H@@skvgk6qTKgrp^)DkN&LL?$bai6T1b
zkp4^&Ho4BwQ>+!DA~~^>7~v_VZ20_5<5*BBg_gFV1C+YbrYs?6Uwn4_YFQ|UBW*T0
z@<Xim0%^fHLCXnt^3kFh9-mA`8*eluj^>Qkh&WAIS!wI{<$^-v#i0(GrpUm&mr{nb
zCr$$1^F*t_pR-O;?@{kFjFeYBZKlw48ZLS$XpqW%Y;%v+X2VF4{tN!9-<n#3f$Ksu
ziv0SUT7S^cjO9UW`<)>u+a~`vVitq|>eJS6S>(s<RwaBIK|?q%E|<6UmN=S}vG_j*
zmgQBqhZyb9EvqV1azbt9e$>q|qe!*r<f1BVfgV_sL3b|{5!fIhJ<UI0iB->0^tIh!
z4K#DnGY6}WS-f%_ntbxWS?+<T8$;?PfPZ)<uzh>IsT2cHB<N{U2tnRpIXE8lFuf)6
z#!~{C(hPLDTRhEfB8oO*aqbJXBV?U4i0-Fi=j5bz4D4}@fmMd@e#YH1cUb@t^>pH{
zihOx_PGWKW<OT&WlVF{_^BQ~Gv5P$P0ooI!-m#H@T!8dKu4KU9AWtKZV8(<_f?!jb
zq*t>`_CFQxQ^j<JH3g4_tCaBW*hRAj@u59W&fK~GUzO7B@6yfC*)g8nO}+ge8lP>_
z4)=JhMoJAP#IaJkXgDRcB^~lFL1&@$Vb`4H-|~irEU*2!b$?YJ-~M+$K%RUS%p01Y
zR6l&MH8Yb+)f`;v-Yn|iT&X2HyzC33`v<o{2aSx36u6T3Pa_hBVVbBR_n?<OR`V3J
zPGnt6Tu(NR%!aGKv;-AXW&77x=7n~wa;v;#<eO{Z2C1;DDv?$3I&Lb6&?`<`?Ybs(
zZC|wl$U3KJR526C<@1324Xi`_R(=?e+srb3k&tNm+))<c!BMj5Ug8w}?J&KA01zb#
zO}$8_PK_<C%ro}^gdwT$GlIbQ7Dc$;&iI41qapPsSk<q=8A#GZ0f4^}X?lj}v3yEO
z#bBxRJJl`S*Z<K)_llT@8_nXt|5XZ~3aWp>AY!Fyn#!AKg5kg_rq!T_*FNouPX(D2
z*Qt5Xp6;BTYC9DHTo4M@bQR4d7XXok|Ks2(NzjOQHi=3`A|T@`PN+P{3x+h`<Bx(f
zh1EY~qus8qYl2X19{#!0yDFqpoR)5ws&BmGZLzvCC0Ln;ecYZF?)+W4m2HzbYYVh&
zC%m4Q54w5$*U8K!B#qOw=;Vp-CFNOGq2E`HAOQOCiSbYn8I!OZN!{I8>(HQCTU*;r
z-(Jk+(b@t0e19gtrQD16q1>8&{x|Oh<gN(0H*U26eqF^0T^{nj6Q~@zM{nWhgwv9J
zIA=;W%18(V;yMp(bO{Laa(cZF8;0-AQBqu2#Q1+jVJgRKY*`8u%Xd7`cJ!_K1hA<~
z;a;@%r`Y5%g3Z6!;=vJgofUv4YFF(AUdf=|c^wxb@}WDLI5DK5S$S}^MaUh+M)<B8
z?vcuIz5RQ<v~_%H2s|i!?rVSFCS)C?1W&i;K6H~u+E@M9!f(D$#jsd4JoC_o0(W_!
z0Q!-Cfxw0l>m@!nCN@0QG~3HwB_T(JTV!m;%aln|Ebq4$`}Fd-PMxqcpYzV+We`W+
zvxz`f)}UpIW>^x$bcsIfUCP&6iIT2ietG%CoIU*uUm2Q)t>su<@kt7Q?yq8NbY!JA
zjM%KbJ~`MRe`^dVZ^?5#bT_hi`Oa=@@A%s1BT3rKC9C15B(fmCR*RtpVs2Mw##y|d
zVw&{cFJ?$AMuUW<wBfIHKnh;{TJ^8wRJ7{fMb34Lw&f{^J76W5x3yzd8OOqUn@Lz>
zKFXiGn^0lJ0T=`RNuwN(*glRMR$ZMqK|PLt2dWQ<`0%p7p3H<<dB(EJVorI12Bdz9
z&y+LYR4+WBQ3`MsV8wqLjowkEdM|S(^`;WH*n@2WU>t1y=CMVFg`cdo?1D4_=|H5j
zT%;x^^Z}qo2H_r1K3%a3Oj}}kqy}RIv0ks5JZLTKiUy)Oj8e*17f%4xQw-Dd0LwzL
zfX#rawbR__>e+eM=H8TqY(dEk1_E4+mtHRkk4tu?OB8NsRBinrP&NF<wtm-hk-U9F
zOIDuZECi#mTH&6-_7sDNuWAp9AQdL`YH#*mY_5P@i-}{7g0Y^oZy3l1xI=6&c0?$$
z07_LV!+SU{<IhhhMEjc+o7L;@U>725BE^Uotqm)U{)2U~tbhOuJT@}_Z_oux8i_9Z
zi)zYFk~Ml(4h}{8AZnye^o2FGXm{lw-wtUj*dU<2GB@(fRtQ7#1eGJdeA(7C2Q62I
z(U0N5pv93rhWuY04AiQO1&r0RTbW@R5XcHH29zM{`DF=Z`H-_DdU*Koc?TVxQtot>
zue;7$5vyNMpAr!<c!Hke;{LA7_*Qn}<6Betiym3dz>s95z~~dFM*|b@Q&N^+*ZA%p
z^`9UyS-5?ft20-gu_4-+MNbr#4ToRuR0JW~l=XV&7ru+Ace0BAZNx|SD^F41@+9)h
zJ^bLOaiJuSY&2q#k07(iK*PM!zQ<T@hoI#|v)vkrg3+on5R21+a1Wcv54)`F_(@)m
z99;_Ia0dViUtIekP+^MJyO6>SQD@X>^^-pV9LyU`=RgA`7-li(4Ar>vXi)D>bpxs@
z=OUXz7)VJGn9-inm0{~i0+^Ka?-r0Peh1v`?uaq9`uqh@1L$&Fp92a@900>~wMi2q
zr2s`$*g^{VlRck3+WCs{74XZZfF=bAiLR156J-}c6Y9OsG^_RAdJz5D#hwJ>q?=1M
zYn0r+_}Pa&xsxq7tX6j)Z(JlZ718ynJ|`0g&~0#$F@7ZI4hj8Z1ydC}wk|{eI~6r8
zU~G}u<7w5&k*zJ3$2{&g3!7~lJry*v`+oG=M6a{Smw%Yc$g7-22sy%F@mH9CuSvJZ
zP2M`zq2Z%YuqGmRd7;7IkB-+I#j>s~SDn5XwAbX~`1LFI00l_ag>ZkSTG<$V&;1~~
zo*S2Lq?H5_%sVSpXIq;s38J-e!zEiz^eE4J{liAP>tJ3JD@E1Vm@+k8J^EwC-OW*q
zoQhw_TdeeVhj-)ivV!v-1hss?Y2~FAg8do2@+qbh0~#&Pk9YDS@;OCeK%Byu-v<b0
zHtP>;))mB&BGFq|elIt)?2i)7pS}PH8zI=On&?T30-Qt$RyJDSv=k;)wn+p(!?n-7
z4OOYlWLr-*P`UEnC%l!E2Z7iUce5=?kzF6^C-rfeO$B4w)+;D!TS|Xf?Ad>%>kJfO
ziZ7Lkrj&y_GokF7r)znA{7i?tL=~wIC4l1Gj=TM2m_X-Yt*stTXF(*+1XB##`Xwc;
z%UtkyST7w=c%wx*8qj1G%6!TD%KBZin5Irk0-zDX?{5;o5rbP7dy~E~5)RmK!zzXJ
z63Y7x&@o<M`C)Y@#hRU+^0)2rYwHLorNFs3u<@iW4IhHkm{8W&wVur_BpGN-&ho+(
zw7>1y@fx)_RJA=x8EJbE8Svfk7|sHBc1r&LG|8n!`UAv_laM)y4Cn*6y{}Nl>|2Du
z8hx(~e#GF3h`xZs)YdLxZxzyf^Q%&tD63&XmiAjQt5f}YwRhk^u(8V^?*IBdTK3oH
z6YHarl0mawNPb$>%KMcPnYY0YLH|Uqi_~7-w3>O{_svUF{QUf1vajwOfJT}NM+`co
zf?V9w#5j05E$lG`mbGIjP@21H6OZsCJZVmH0`w&vT*+hyjT`)?G*BT-6BT;!be)#T
zCdf4jH~kV*-QpOy1i2F?zSPAP_Xa>3lYktKR!1+MIATC~NFc}Wrzm-VG~QmY7WHn<
zwdF&3#Sb{DPB_(wMQ{6oASie6nBce<e;5}!ll=KK{6{BbMaMXOaGQHY=A4PDkw=%B
zdPv}547Z<E9Vx)sZ-DH<lu(9?i-;}Ho_9Ti-F)>Z+yiNt)<tKDIemiXZMZmolE!6+
zUPlwt?*?G72zfoIMip9SG<o5(tF_^8>N!CPI=NNMnpTuYr2S7LgP0o+@Mib5HDPBx
zT`vM;zg9+a_r3D@8Qn=7JSMygHXwb+!6AIKDKZL;%&BEf_Wh#8@K%jh@_hvCjOU{z
z(F-uZU=~32v;2L0-XaYQX2$|e3|<vYJjcZ;F_Hp1)6BuK2^{C<RQl3oe|OV>mIcEs
z^Y<?HA$pLs_(xVjG6zFo_RoeY6|KjSGwR;ng7xL439oD4YPJt~$x+w=b*l5W%E%P|
ziYl{H-G2_2g`n!t?z&{oyPtAkWdD=hR%4W>Ner)JQg^Bwd{d|%-iSvUIm{=(S7tgo
zRt3j1`cxkAO6c_v_wLA5w+SQkHu8S!4U1+88^g&v<1LVDtrwD_!dMclQtnyk5N1C#
zQy|)N?=(?PieMBqY_y0u&%Kjlr=f`&&If6yFT00_Zz}^O^%3961$nq~^$$?j!q<qb
zR`75*Pz2z#`7ZIXVcGfcg**sB^p`W|(zeJ8pLTP=59*hO3{r;&f9wu`3p7xfPU6zk
z$=X*xv#)KbX$>3Yi@3hm+I-71_{8PgJV7oMP%QRkH)%24q-t-T4zpK;3)X4X0FFB&
zjlkyQV*qh<4Q7Bo$7C=Z)^J<%a?DW+P?o&APkfO=^9dta7=N8_hiWUKM-d@?l<w`#
zER8+t&>$5zi}kPdEC}`Li}@cd%BewbhvP1)cNcWYP`NMl`EEHnb3PGIFu)a{OjfUf
zGOPAN!A)Wbmn+gqJTA3dBDPl=LSis-0jIcoXS~PQv&nAw^}~1V5muW-UyL=B0ZEN!
z61MpkxCL^<qIK?(f74N!&S8uSQZJ!y7tXq#rlnDQR%`buulx8Jzv(~GOfrx3b@{rV
z#z5z*^%i@lZyJ**!Cy0H+n-Hs%4F4xH{Cn9;^Nt?4HmLT+6PsylXq(3hq8D)%>nUI
z1tzv6Ze`0Xpjj{$DO(`v(nx&f397VOufgc&Usb|>d|-H;3{0d-1Sk`F)L@+JY%(J%
zTjIhI?lG&svp+p3Kq=t#a<-*F=7NFMoeeUdWp=kxN%5ga^5hl0(mJCjLTaw-)2^WJ
zJx(l&ZCxrvl`L4ori*s}aF9rTbB*3ajSeT^)9KO`fH*^?58+{20#cM<3l0-XEo{qI
zCFnPDx39JIir*huy7&WFy~^U(Ntp=o@AsmtdWzHW>5nS9lLkp2V_Z(_A66hscoftW
zu^s0p@SlfnZGNl7Btd?vKHbzlO=H*5GQuZt8}MQ<2PBGEz)UJiI=S=L4~J%=hBRm;
z1Sh1^%bQ#yZ?gRW^grzA;^3~jpKT`hty*R#FQ_IN=}z$Pz(|8;yvMQYlFDwM0zM3<
zqX8OKs86fCpiaS;lQkgbOde^Dd-T};p2&$L()=VojC9_mFrXiT*0T@TF46vjDQc=n
z#iD$?N3;#lxvjg`HS@%pmhj_v9Hzqe@35XUuVO1rm)kdq6yX1UnS*|FrO;^&!A>P|
zW&dVrURIziRxO+015=PN$9Pw{VCv;hI{XX76Lj+<5lax)&B>f9EH;N4<Qwk*!L1KY
zWB;6g=N3|g{9Zz8joje*K2_uSVL<>t0oP%!P7_;A2GC#BUrE`4Qn)#of0(-C{#_`a
z%?{ttxFXCO$Z^O2_r2S{Q}KzXuslw@fX{tGUS)P^E{}}3jz2&M!CsEg7PR~MbN<dr
z>7Szlm{^cJ2-!|O>c3bn69fEsjR^Df=}QYFaMa3@Wr6@-Ojn+wXeq?g;#l7&HY5VW
zA#$H8!hVcCKOUBn+)Z^h5HTlyyz@?p^hdOp>^d5we3F~pUm;?~<Wh#@f1hs5aEy2t
z0G+-gsuF*M0i`P<txcq?dA=mNN(`lZHBc7`cEeqX|BSVRku>&_+`OBDj+q-xD<frS
zO(+4s2!}b`G7(|<t0?rbq9djM*s+UI%ItURSp3+fGv4K`zk>f19tGsUcM`S^Lm};t
z?-w8|9pxz>R&)Uc4;f^>qECjtVYYax@w_2EVi<vwytw3Daaw;1jDOR7p;u8pOuN;4
z<~_MfAqN=FR~Cd$5djE_E~B?f;b%ZOUo)UF0Z$$|A^hV_zq6xb|K@17;gk8tZc0i_
z<4g4SjGjRILFhEVxKn7V0KnZ=xNB0>q!>#;TnbCwRLzDuK+U9ze1=Yp2W9k3K;ds9
zm&d_7WwmITzBCRpJrgp<scNoEMmYM~N+u5N=Iud6?(D&IW4f2aVqThZHvWqjFHrB_
z<3H78_H$PIAfQ^UXZ=fOrq1^V{#y;N>#A)XUH<T)K^{lDbncsMk?ODpk#Kp-$_Qp%
zre;p0)@ZSMe%><e&cTlGH)-VC>oeA!-Q9RdA}SL{CLjG5g<p4@UDm*&nWH27{@&JG
z_rfyDQ?-+nZCzbs<D8t9V|IW4o2@MS@>i$wl%9mbYPNa&iJVM1XH@G;mKhG3-NE^c
z_y>(@Vz^Wq$a^)ovWGR0%~2?}KI9SfM|n!TSGN@m&HqG0!`M8oR(=RN_R4ugH5;&y
zdUB2CXb{;$PC<JQOt2VLe4u7RO(H>*F=R8_wy>BL=XP!iXo`78Kmk~uupiyv3wHz=
z<0i8w<KiR8La%(g?>f&EZ`u3FVtX1YwBZ?c_2K<+_2EkQ#J<=gPu9Kc!ABBL&XWi*
zdrrA2`)DMSud6e-^X;^%4VrG)wrY4wOKsJ=s66r7qj?LA_;HCMV#b=I2U`XLm8xF1
z1iT*m6}~-J?`{!%*wIntbG$;D_vG|52)^cuH+|E7e|a!3=oikVsi$WX2exY8lyxV0
z4TEH>|7~vD?`&<`J2~m&BpKOeTtg73nI$33pP~a6;sX0n(5lI2ctT;DkDb|WKivN{
z7vb=UJ@N3a!o0?T`l$_eDTfRK)I?Z%mPDDP$wx+1>Ps*;PI3hyO2MPyqnC<{I4NUq
zcXxLN%-j>tLpwq{$?fk;3Ow4yqSULFI`O;=us-9VvIs#(hcrg$TOLG3$D?M&GG~~v
z>lPd1sLvsHz)~(ZTJ&q14MX7Ck9*qNw8tC$9%e2Y5zf56;qUJj9Q8D2j)fv3<)L9j
zCwS9&2mh`PHK7;krE#J?qd+*^_gmu{g#)Sjj+J<GitkEv*Xx&-8SP;=<-#Fs>vHl)
zI^-4>{4AP&mDxxC3iDsPdVN>pNb1Q`XoCe$ElGz8%t?O63VNN3D;YiH{^R`cjsbd2
zt=GGF5@AEi_=$xysT+cpzhw{p3dh3|&p{uzAm!ikTW*L&)qB4aWk7DB=WS+V4eAL;
zhdEi+bQ}X+52FpKc#LySWh>)8M+A~1n6N)>(kcC?J@jwMlL@dchXDy|{>qD)cUWI}
zW%6QNj}N#-JF{}~u}{0_EoYt%kbiKTH9(;TrQLC3*6tL+j|DB!G_x(Gf5LLo#@HFX
z-y&ybYqf!H?CbKTfEZRAc!kg9<7c(2SZG$Vzs3S2<Ez@1FG!2HmRjr&mX)wYPksRe
zG!QWGA??3uA2d>l3~&c>7l8o$^6bJQqX)reEs2)Ag$HP_3+SC?q4LOQXvDwJXQ9yR
z0wN~DTV@$u%Z!vQ$}L^A;W#ZVZ6uF$dv4yGiQ?D-bpZJf%7pN88`OSWM#>kTR9uH+
zy7`;_D6o4FJ-Neg5u61&2NEEOfRhu8M7;dX(Aaf*pm|dp_s+p(fz;gEWl4%Y(pc2n
z`Ys?W1jHRKBf}@gS2mD&<_+Su0frw0Tz0mz1AxR5H0Av(7@r7U1&g9zZL!BiVdj3?
z0p1w!jeiPgKIOn`6L7)FguEi!he}uJFDKK@13ojCYCz*4mSK_7GWGvDE>o1Y_)>nk
zF3~Gh<v@~IwFxUS&Iqi&2`8Ihe}6vUZRI2MJMTRk;dj#XEvk<5nYkueY-XTqy~;t6
z63}z&+yLDs!$U_%d~LhU!Q$uS0%}{n1xkH2X_~&#IN)ZQATo!%xCld__V63ne87tV
ze<}m-50^Kzd>Kga))OJa!Sz}aNo^vken2^9<XBG{=#O|X98=-qn!-z0`(i$?aj?L9
zpd<_^(N=uA6;s|e0_xx|%(OdX^!&egl8yg_xVvgHZ=1O@cAQxg&>?C0rU4<Z`+*o^
zGAW>cJ>WLtr2{8<ftv3J1yL!B1)u*<3s6qGM)86y;Z`yae?MwThoZ<hywO%|J3G5Y
z@Av~bS3)rt%VV5PLGOdPt@GO1Hez>xywv!1`4squ)n}K+|7#x1$fo}!|B~aWJ}%zW
zH-zM>(sa(U2+`9|bf_op<v?RX!~)go6N`R*BazYGqWJA0A(QWpwU-XX3$!TKsFKn6
zmp_u3{io0ePVF{J#h6JE>U>k#pSc33JL@>P;@-y&#P&dc#}l>s<8VOfz|$Gu+~jD`
zz(#WI9b<LNvuARql7U{!+aA4h+CP9f14NNgQI<+Ff3YKC0xFs(g)uI5XcyW3SsVv|
zu3NPi+bz3`b%$C0LjN@0ui+jTTC6)qAw3Hd&l<H_TEVmSWkRqe?+~nkxj6Yh|CZ{i
zP1>#m)+?nLboz?kt+0t5>x*WIkutn>V)On#B^b>4T{Ox+SJawGQ$t0+L8oJhd%KOw
zWue-=_ys&lgOIF%>cNcHO|5Xpr0>b^&*)5=<szJn;iTzmF)^|E37IrM0S{R6T}Z|7
z(z)+-X!o<AU=TaW;pOHg&#He!58?d0adM4gUoq(1JHnii0LY>MK?JL(Nr^fJIKm+y
z?Ej<btHR=Hnr;aRPH=a34G`Q13ogOk3GNQT2iJt)?(Po3-QC^Y9nQ}Co&RDk=b1gb
zyQ`~Kt+i@4idX{HX(Wm=8xYQRgKn_erpw9`!rAQjQu~`gH@2XguVf<czUZi{k|53&
zrMvv+7%>b7)A8g%LWO_a3taLR-n$7&_-Ims=RhfdrKGc0MkE2w+9JTyXQE&k@Er@-
zFG&T#5+|Pd`Jn%1!VU@)`AcNSL`$b=h`NGjN1YJmc!7977Ie%M2LCgu(YFsCdKR0p
zG_Smq$VGRD#pR4py8uwa8ajMt)UJ)Gw_e@zgoODumM@zBn?B2YG9k#_z8wV&L@+rw
zd93E)h(RU*u9?oYzWg7~`pIK463{Jd1M6ZP?F@3#4kP`@<`1Mi-~+)LH$GSqSJX+V
z^zR|4lgOb5K%uwaq}2YWulvkjXl7|CxliZIztuNT)Y6jX<>i5FanPVJ{KE&eHTo}3
zb!d&6kzuT;18Zi~&3j=3ol1&PS67$ESOIqi425F*CMC<}{2l>Csn74lT6ce%?@Ui@
zZ2v`~@?2Jbf34H|D4}O9F;rY3Or&?F@ny(dE?YaZhc=WaSw<l1zohsJ<p?yU_X78S
zDzW&eakR<HCDfS;La}?4$tOfEbAJ*(6O?t=Z&XM-u!?dZ<RATU+`oBT$vi6FW>6=I
z0fxcmw<c2bTmAp|<<~Miv63B!0c!Cjw*f2dK*x6!C}GJzd^P*i`r$17Uoh**+PEUN
zo$^F%P5EH<uT3HK<~Fda;4l6cO)Noq>wEJhynmCf+Vfgzk{bFbeo)bn6gvw}&FCB#
z6WPbW>?RXOHoZp(3{33FFq7c{LcFj+PIl)XO-pXAWBkEV)c>+6p3OjXDxDNEuQg*A
z3DyW=pI23~rM+*54gtFu`^Q5Y)^YbhJQM8vwOo=WYj@Tv1Y+Kvq!;aPo<ajBhaRY?
zeKR4n1?YaB00Rb)##?~c<H0th*{&vqKza`xIpaT4q-fkC8>2}GH~TaZhYQE^>>SoQ
z(=#(5(EI+kCf}_eLoA0Mk-`DEnZArW)PX}@mQ<fW6sNKegWDYtHOwRxET_c`XMKwN
z2}M0=j48Mc=idZ-w6xwKm`sc?J%>g3BpOnGdUlw1I8Dcf;VfA5?+C&^$`7d{8eT!X
zbm+g?>f=zoOU~I+W(!NeB0jATdC#YAD!c|d+9JV%U|-*?J)7)rkBT{d6q2_ADEi_p
zc3Ik62`YHT9v^SzUmI-VO$y|8y$_NAPwUXW$CTCoZw0+~UTG@Hdw4?+ILMlRgV&<o
zYNDKse6dg`Z$US%3z?3YONjAUfN_R&iBKRPM?ra;ifq-0F>)hzeu8{f?51Q(LxZi_
z$D5?CK(rAK)(Kd36d^~A!k#)@0;hhb$Osa15Vl$fcM8o`S~zM=a!ShVdv14k_YXU~
zrYfsDz_Mm^RO?EP1QM2&lYPYZ7}!;RMgU^ctv|%;+?Tc}Ko1;%diAEOzyFF#AF*mw
zo-po)(5U2sHt9U4_Wrx==wG#l7%z9FN#PLqGwsoU5i$wp_yd{8FE9~lw#4mArEbHg
zbpS#C^KsVfPihylyllN6%k0vkQVb+!Dn;Ej`55jtmA~6hzo*TtP)yY7&po8DnN4A+
z@^ck>75PQ*Vkn<BN&FAu%K4PHvubMS_p~F-ja!x@AaeKkZ|-iR(V`sihHIe=T(Xxl
z{!`ciz?}|i<~oml**<_sWHQdEn13B(D|_@rPpFws+l1|$`_Ch0Q1r^h7(}huI_=Q{
zWoyDgQU!nnnFS}Cpc`p2j}j(dKsy0nf6g}Yfr>{y(*8@*cDJmOB<VQ|3-Ue6&lK2z
zdCFd~9~(qj`>Lht>DNt`@KA!(3Ur$g_JpX0Oll5j!W*pn{+^%jN%!oLMS7GTL)f36
zCEgpjxmaRud7?4F+7I5NWl7S%Ffmm$HHSLrfyhg>eb9dVsQDr_W9xjg;q~!e#}Rtx
zT|*AkkGbGZY%%&yDa^ext^;Bx^kb@RY$8Rt{YO8+p^DtOhCM-MDVGKD1Nkcu1Z*Ai
zTIZiLWl`~OEK81#FJrFZT5n1lT09JkM|a?H63%1!aX%mak{60;I<3_@Y;yF?%QwO_
zpGf~Fe)m3aIxn)AxI1<B{7!_Y2>KDWD*`$A&MIW9jWnGvtJa0v-i>jb6#C)*qTM^K
z*83TWBdC(i_sy%gxEPSpr`&$Y)Q2>HT;6PkevQOY(`bG-9W6{p<xh;;+B)zCD_9}?
z*q^@I9acFbf?$Q)LftJYR}wG@Lj#+ts+CN*|En6hSv0m_c&s$4eM-Wef%}ctC*s~1
zd|(XZ=<u6?grvnz>7?EHZYDakm-@|%Pe&6GYB9j#8~mw_A7@n^n}ee`x65lK*VqJ_
zBC)`S*9j>Hh_v6xp0r(xhj#FE=ix!9L<|v#+G=4#_NO^zCBIbL>}ue(T%R|TolaJ^
zRnAh3f#xxgofosX+OcBBg~MVbCeXMJ27ku?wQR5-Mab^G#J)@7AM^8eqz9s3UQ0`B
zCJ54XGy?bgWHSjn{+y`L2M!L$__^u8(g10g=k7ZB8%~<pt0I!2A;Y6quSKP$f3i^;
z<wuW=dTzH~_}kSz*%Yr;(4&a0ZN}5cAAvMqWyJU#Cs$XKI>TM%PnejFP83wzndeUp
zB(JJl?Ur)DJpeEml*pQYfzJR5#PF}Hiujk>(~w17wV^YL-DQS-#VeRixWA}mdzoiA
zB9xx@u~veiQMk5xKI?3Y1+}?>YmZ~DQ;y>T5+bn8CnbrKyL!N2P5d%GGs?D`u5hQD
zs{RS@`!$Ji!FK{3=h^$i)pK`Fox5G;Xpej$y|-lIx0rN~tD;5|;LSVu0pg3KqJm|m
zm9ZX-^w_;87WY8Tce!(+^Tu9OsxG&#Ibj@lG8$P`7986y&m?UQ6i7D~Peb7aK|nmE
zxivU!<bb%@9Qu>6MMZQjlCwugfkgZto;E5L=-9{o4_)@WbJ|(>h0D9j-^MCJ+|AO|
zEz7x@Fx#x{vZz*+ZAiENK_*{#{g&8>3;JG%8c;O49<zp&fDiCNH$Zam3c87T6DV5R
z`1Yuhp5_w7v7VTjnzf#cJX7QnXf9jHLmfnwZ-+yGmrb1>_F{F9j5N`Jhj)-%QjwLu
zMua!@7(dT{1)J7(WrcaTSw)?ZD_H6&>HS_;;lq)iMuIY1p_#Y0{m-9k5bDZ(rz#$7
ziaME!(ZDMyAK&)5TgI?eL|jY*O7D#QLl(?qAp9&UJCOsU+Gq3+wh_2~u*$pTRcmio
zx(y@?U2&+N&|`l*RvGrtIF<vxJAhS~2+-`9nN3>de3=#l3|3TF>$JIVNoogKTlxPp
zuG`L>X1gH5_y=XkAHY)O7P!IkO8EMtjCI>%x3miz>LT|peq{n8K}BeT%|C_(Pd{q+
z4&t*{>e#rzGXtYmJ$rMj^YCv)lh*ae|Nkd6C`7bZ7C}zG$$RFtY2yW2fpp7jd<yvo
zevPs$74{`OZqQepXgkk=#{d-DwexAxGmK_(bG{ATkwDQvTwdW%$to3E#s2zKP@vY=
z;20Px`8aMhd@voLK_B{_5Dkg&iEOzDd*-s2MaD&FfRoFk8cJlzWVxybgYTe)=mb`g
zAb0&6aP}!xu{XF~HLs`960YfMm<K)p91yfnP-2y2&klSqMz+D#IGpF>ZsK$(bt($F
z&De0X&x5FDt((0}W0N-NF9o-J1`ty!ddSEy5<z*)Ak`xS+N&RUO%4<udhiU!Le)=s
zr^@De&3t`kn28-Ufd39q*ql0>5iBJ1LWVvV_mJ48Q-RG-APwQ60yV;^|L%l!)2BgZ
zbL$0;EjL=!pee|t=ZGO-)9lBf1&F|`rLESNiM^%`a01t};yx!<KR{Vt{E()D^w=lk
zFyYEC5adH|zs|vFm&X8#H80@M+k{m-{I(dM*-EJAW#;dI-ly(>;x6S(96Hpml^Er*
zM~2K-`CGtYt;t>OTv=WqY3(}wexlgcoBm11A~feo8e7*1)%&!1c!q5xlsX;9l~_Ic
z!XXKsHBl18a-8n9B?Zj4&SxAQJ$gQ_Xo*Vqy%cNS-Al3P-WxsQ4eYXPto953F0f<h
zF&|2;q^QW+_A(`{!ePX7x>^%`cMFm!vuy=<_pdj41B6k3U?)wqtKCq&kGAAXzLFA}
zo&&#_8Hqoq?D>p@het%^lrvnAP~x*#s=;Qo#Ioq_bvUf2YMAmUpz%no?f7~1={lpt
z=R7V?`mpWX+m^)7gbak+(*D;bu_}<BV4T~n;dRT`E7s>Fp4hCG|D~LnjWzHhU{Hd9
z)bxn6ctnfdGV}ocpJ*cepUI3lu#v~X2Z*ot2m+HP<q4h-?b3vR-AHXX0v}IFv&Y!o
z-JKUtkiV(rzABIRso8h*m(7u5bH!OQj`gKQk$;)guEi3kWbN4q{=(-Tj8ASftl0yA
zlM7pV3H19p5)zV<=kr}5?;tGuB1HdswMhw&%Rje7M5Mya0-%*bgZjg6A`fsx-+#^P
ze~fYQ-wn{Jt*`&>;^I`>q%^lD<Xk;;Q(RnLJ~l>TOUEXod2o7~V_Dv$c^c-329*;7
zICuP09Ft~OgI-m`61)*cFjibm{3iec;cRj~)%M!r(YGcp!q0%3b9LymQ!f2(iTzMe
zt>*xjuPPuT#o4dY+8;^g=&0jRTq2`VTWt4+VJwv%w*K>d14pyfjnsOnn!{rupTlw{
z^A^O-{Q`aB`;bFS$7<j)>vXg-FGWb`Sy*G{AVI;%NUP|cE!2&vHldg99-QuVHF)5<
zk#OIZq3d-iHd~OvX=^j$TsxdeXxnl6wazM>HJDi2+gT;ygFN38O`2U7T>I@es=gkr
z^H(Bf!zB?4x0hIngT?*y9AfQT@1<$yzbo!fMFo=|kykP*1yJ^hJRUX=?k(G|32!0Y
zVmu!X3o`QK*WTX_uBLj!4VTcY8%?A^R@&NX61`%*H|4G!(u@sO%UIiA5a_tP8xC7$
zDxa2rq|AO+o8mm~IMm{#a#r&88P|#hlOZGTqfZj1SC9s5f9T@_ZTr2n@hnmk>0jX8
z^y|=`r~jmrhX-`Emko=l(oHBcfQ?F%yayVpJcS`Osxo-v{jY4wR|(NV*^1)qrN`?p
zRCEje7(zY-tz{;klC3tVAxU8-N$FJ!Lu&o}J^<1|1x;1)fR7)!eO>?5vuCaiCo+zv
z*kv=I0inql87ovNL@fJq8`ke_U!s)o@m4GPFT$cn(sQ|t(ia-E&eyvSZ4d5hLzdeK
zys6-b4Y-EMh)hyUbb=JIY(Db!R1p+HteLTbHc#H1reG>flxkt$)01Cj)9CVixX#Cg
z9@vQ!8y5;OIxHWO6uBd#xX;wTKlJYFDfp-GfAbkdZkK<)m!=#Mj!EIJNfvP542s3T
z5`1@`2^D)Hs@4(E?d&bRhgTd<WC&eEez+g^3NC+N92ZXmk#+XBth+tm&rScV3X4wn
zWGZ`oT^Jp{Yj3;tW3yc)kB*?@t;>#s4$;0PO!e??3ymm|q04=GNuR$|)uuh4vMtv(
z;|}|+jy(63H0|Pg{gKpjW#hx;ya#pk@5IFM!EHxcuK(Er7UvUsd(O6YrWAxik>#TZ
z9k2rQ=TGOP^P}Jz82^NSqdf60uLw-w3FxUB6kw49$=TWYMDluj>Dci_YERGCEApMg
zu_p6JN5An6nIs4q2lWS31|1J`xBtrowe^86egjY^rr}loyGKl@{?z@Jbc!PScwv-x
zyBe^K3I6l4S=CRJGcEAY*;-Vw{<ksBtRJyHnuA??0!7cIX_)!lL#B|29Etz#z^L6l
zMxErFAZ*tCMNV|z>N#io<!Is04<vLxC0U&f4BCQ!wVwe|-DA1LLx*vxPRrN)o{_`v
z|D{mlwoO4_k1GAvE*u*R{C_(5+;g1`UjlgiVnc%fbK6C>?V`ajs_8h}BK5h1AFxd+
z$XeSRBBOlquO3N0a_VC95p*dbsqj|V=dJ91d&Cm}v8#+2ioBuH@l3qdvbZP5!GTsk
zuDU0sP$Y&1k~q?4tfryEok;}B$lSIE>O;mtlic9>V8eDp+Z&YjwEWf;{>H8In^~Gh
zGRa#wQ)<4=&j?7b2)Omaz;q}95ek?ZdIc-8fENuB<DkL-I6cb-53azB;lsY&K_^U!
z3)CXk`hQgiDu6V+!%o<KJ_J&96RKoS@Bm+9`#9>Bo2@rs4P|)dIH$Ho!#qN<{eIT?
zabCyUL6Ibx)tblK?Dj};z0=#z&y!gk<7rUd=Zj5$%phTapQlV!YUzrnya4x7`h!fG
zYSAYIK|gR3j+Gw`E-XTbrB<*{A#gW~VfR`Y%!%sP3Aa;mC5M#>H`f#L8V>2Feq<sb
zN5PHQ_;iWtUVvzKO2SUqGl{@|t3Y}d19Ny(a#&g;P@1ofME8Xf^0_|R=ws}#7V=M)
zXlT#$L4kO~o1S0QpSNfQqEOuMExA#Xa?zdv30KF&!bBzWpbbtw$xr2j-Bm)0FtTgm
zYotv<%7lud(k#FkGGkA6H^%9(YYf$Bzo7G;YS5=Wa3mSFH4Dv__jK;>X^?}tZLK3<
zb8NlPD8+5jwTU|WR#e}zUPr-*V(_jABTuvCB&<uOOf}gq8BQWzi+j%RCpVET7M^Xe
z-$h<Uh<bv5yV}^KF*|KPb>8R4brZ?iF3ToeptOBOUqpk^Y7}h!9Q=HcY2j!-ChS<)
zLz`7SrS#o4@XzqfdEX0Rs$ko1-*`X{JYj#jp3!c+w#rirfGR1*uc|vM&Uq^%*rtYd
zfh^>%25R3h$p@ZnN<sI(jP?Pa<O>v!So+<vwcGLpUckRi1f(~=waX0H6$0M^#j?1%
zalKHY&K}6V0seX|r5UjC_Gj(SHc#^iOatM;!Nc1{1?fs3w-@1Y%~FN?C=Jcc&2#Bh
zb%c8<DcODB_>NXu<U>F3zk1`AAjA{95gLqYN$ks$y@t%w(vrHGOLC{b2-)O!Bk{`Y
z!hDIM38eJeY}zlf^%dlQ9$|BunUtVK{f2}%5qi!O+kO$mJb`PLhRKNBdG$NRHouwD
zXKznZQUE?krLb)?Hulexq0a3H?hl`#1czU)13w@8iG8)X!6YcFj^q3{+r;ncL5Jk4
z4SnbPvmI}dSj^fWwqmmh`iiIg4x{RxT+d!rL6z$Oj_#fhSvxE-YikW!qcELBfr3`z
zl&Ak%=q<`<N$|LcciV!3d?ZhPAFN?N3=|n^GvI7|h3o`!3X8soViDWE#tng<Htw8+
znOfJ(L^?5(jZ@YITah?Mt>(abtL1!$s|9}@1-rCUOKvW%!a$M{@^j>_<|G0^^{*x*
zGVbT7j(RvKZ=-=F*I}WRCZhXrWDz)*lUQXJ!u%pxWRx+R1~<2yYv5oT$JGs61t^pG
z4Vxo8_(_`vz0q;T6nt}OSWqN>yR1(4tJCSSkEB+;RVeDt$#Ih0Nk@QXRNEg_qN~_i
z;s662UlNog0ue+ujSU}5xhtYDINcJi!J1U~HSb7#3GgD%N_(fzLDDJgCH^_9Hmx8}
z-!@r1*QfFvC-a_;<JXG$Q^Ou+*i`KOn_q-@*_)@lkQl~@$nTf;PJ@X|%F-NSM)*9f
z`%%k>Ww!40X%0iQwa#i1;!Ah@wt`V_H!BG2**Q$L9S>yQyHEHB^!wg2j3GcDi^`BB
zDgtqORPY&>|9XV&=7DWov8W+!f}5UBRe5nrT*Oz0wXfb79^=zDDW`G7&&ub&t#6_k
zmyuR{A7}v^r?aMZACK1}!4h|`x66lkry(Gmt><`m7!v!gS4Uzc!sA^8dLi82qx<N3
zx5ZY4{&{Y#$as2wnq_#?Va?HumI$2n^+Wk}vHb9CBtK!mW>&%qTa;7(V%z<r_~b-o
zSMarSx+tqE*1R8(G#xhWXXI>0^QJSsjhGb{(m0Ya_X*fEU-p#}GUN7gIEHRT;G2GY
zJMLkD3LfMKY)zqEJ}d@YRGc4Yar+-I{;K&HE)x`CAD|E6(;K$i{_%loC&@lTe-SI{
z1jQsMDx+lxHlt5UMNyTtT=nVwBn{4sz~r?rWAuzFp4VcU*OgfzMoaiYKFERA3{zMb
zrO+3dhD4%nW=fSokqH$AS+j^HEL(e5KnqHE_)srXVrk`Yb#II!#rQpAf7QL=T+XzD
zz2D}_=Xub~)x~tp-OMy)gm;{8JlZgVZlRZ)>w=c!R$IwQ5sIu`iNtUwrSwvHeS$>S
z(~F;U9rf1p1AV$&bIem{s2AC2bYa);)x#c~CfQ%Vx@dcR?AAXn-yvNwM!w0&r2W*6
z*5e9&5KyI|u{J*!N*<1Gp5p<=uk2{|hCg}Q8C~yX@pIutisww4i$l79FzJZ-M)Azz
z7}w2T6o+b{yYe7YGir&W_vr#mnVI+Tp}CBZCywbNX)m~QxwkU}Qt^6UAY(vBhD+e{
z#%riDQmNMZb9<xx3q{??M{9Zlt>d3?S;#4$=bPTbZrO~V^CuB8L$A`j-Fz2&=jQjq
z<~QC;WM|D^yE8&s5xU#jBzNAnIU2*^C9Iw`DW=T(NH9VirW_2{?4M;HJoi@uPG+nK
zIj}zpbAk~JYz<3d&6Hjgpdg$1**A?Ib1w!{9$w0c%n?&`D#Kx(n*zV%OJIg^@jD-3
zY^rACPzK>XyMnAs#@9ojrDfu_K4)|(cIXwmVaPF5wanq@a5_FY0VD0R^Yc|TY%bD7
zU2vH1;{$BO^Lg3I3%TLHNkrg>Jk9Y7+lRhOK<B)ntco}@I&XO^%9(qAy}jQ!SxJIv
zWsDSOW7!+eh)zvi5$5&4J$$k)70-d&U>o%bz%Ll@Llc-@IpaWFvoB!WCW8pMb{KK!
ze1-n|EyJSzs3PXyBCrq2PKAMVyEk$m-z7M)EIS^7fg`M?HG$(_SRkk9J~f-Ud6Z02
zaNea{XcB}G(%wx-VH{G`^OFrrS>YUW^<E7c^_}nNgNS+HTwJne!>Ti+>-J(xsXvMF
zZPANm5`{Q2E+3uUonpr`;P6XOD^!lfqBJS>xKP}+FZD_=wNx}gE#u)!>@5E}4V5j;
zuaSaR!$?pa(+MJc#+$5f5j}y4?)IC5%@%SR3|~j#SRO!yf_iwhC<*8|pIHwmmC6P7
z5i~qVLnG<wz%mZr^5jVm!nOa4O%PGbk1pUn9Cg}#eUzzTk1~1vopKOs%+DNgv5CHv
z{PuE9?AtGP{mywte!anb-klo12VKCstYzO*5G|%A@3m>uPh^ktcJaKvJ>5?j(pL6*
ze48Mj7O1AH#kpf<^M;wY`D=b1Fu(NbvfujtDw8vbC>Y!y5H0vpv2bRx-uTIK{l!(1
z_^j=fn(O++5oV^sbzLKx+1t89_%8pryqB}%_7s(On9RoJlhht>bohq|7=;G6$?O&*
zO@qnojEcg-Kgd5_T>%N}>4d9t4lW_$oa{qYFb3g^x3~Rz4^i<FT`0J88AKiZ$~EYY
z!KI!N>OUZqMs8J7RK!hFT$$F|x)Nh8&1$0O$kxm=y7A>Dwa%BW!3xtqsQ<XLM=Cxh
zBL)9*rFft9J8kmNIaC9}aHsKmRH3&C@lHD@2S-*z12%sfKN(%rR$S}%4LN&+s+^|0
z?+nXRQ;>Sm$hy(k=v1;8&++jo_@$y0a}6d)$WSYZb)I^nvSy6mrovgxR`$1_-W8NI
z$v|J+86sv>5@hJa^!%<}#rcEft-NG<x+h;9&raO89QEWmThmn(JHjplwKGstmzVF&
z^3KiI83mMe5})of|7}xf*#7b}_5QL6Beh6VyGu85_AcsJ?!&8LqqEf-%k)0lf|1jE
zcS}q9zR^@vWqkKV?t$@fvG3M@+uk2|&o|QhiTi7fqB=vEI|#IypqMU67Ou}bkX<#G
zVhDw!`pcW{G*4t?yX)Iux3dY?R9f5KtWS>dpQ_@3!zAAxi~I5l02zW^fNRm9e|I#_
z#rb_szFg;Q14IA%{43$R<JS9zZ$+ExiXKi$<I5W<6Auf>?Q=jA7X#R#Vw~NmUH=ZR
zZ~OxTgu*u4XtnY3IjGwb=QbW5K7V$8KKfQPYcfKLMv|bgn?XB5yVm{asdzgK`{|6O
z_w8U{vDQL)z^wGPn@8BHqmZSN*y*saF4J^0nKF>brli_Ix-vb6hxzxbiVpweo$TGO
zG1@SZAH=zA7Q=JJ`g+aJ+HC*F1pxYHF|n~sd!gPc(v%W3DGWvz!EvWL%l5Ytu4K!n
z5|$MRf`Sj8hL#b@!<T#6<LmT4KC=qQiH{GChgFBOSmrzOSD4Rm&Th_dtfM&x-Z_69
zPrH?7a6dL>uRrg~jeu-z@N59b<cNaz`IS~y6qMD~!GtZYoDVXNBdeVQqr;XqHb27n
zKYnN1Gq<$t*Ku8?b$M<+E}xgnM%8C{kHSMNt+H$zXBWhmcl0<Z#tlF4MC|U?d%xrK
zy3G#HwmFtZ7K~u_9c-HUwKy9jiu_1Xd~C?8HqX?E*YUi1URjaHy>rbb$N$;a{f3!_
zi2==;so4Ux-FG{(+o9&Je%ZEvi&DDhSy_$;cXsysLYt?2hrkc^<k-#MyTg|O=e~MY
z&*5h2q1c$L?&2D_0x##j{k4_op>7rH@rzRIlOC%?`yLLDQUNQ(!1aZzd=dP_soTn+
zUh-|tJUUU5)ggATQy3IiWnE)>i;dY>y1tH$HZFddyNsr>Aig_ScG1XB3K@H(X_(U6
zE0&!;@#gYO0F`O(ALG4#vh^G9bHSWMoE&vM^(QjyKxHswY-(b9ED@f(89I38;cMLQ
z-&kZ`pTA=GA`OlF4ZcGf85`(%o$cuA8<Uf{dAO{RSKatML>%s}ZMry;(mR)E{TBOh
zG5`nVX5M$SaZjo}l7W8b@#DqkqAeDZ-Q@&p3$=jOX^OPx+*V3gN>Vz2H1G#IUZm+r
zch#{F{PS^D3Gf$HXr@Nq=NhuVK5i6$Z{+d^*V^i#Q}v-lCLCQFKG%!CJaKk(X7`2@
zCT^m?+^@P4-__NFEuy|dKj$;|Yr3v_yH2p$rS0OVzitF{nyl$|2rR%viq{Y)?w-*z
zGK%Zyd~EiFnKw|=Q~~ep3E>Ux(8kAZ8HR=6uQ#52S+QGn?%XBvJ+V*aJm+hzd9x;P
z9+@=ML;wXyd<h4&w)FA_9om<3Itk2vGF-gcrKP8%lgplf-*C&nomiBfXZYz0@sKpm
zzD|VDF9HPT&#m@j_$PPxr<Z7Qh<@FC&1MIcu_Ast=BSM*iZbw>ROu`pk!0~0amH;a
zG0g(S(t~3C2>#Y=XiS{cyayn7YFK9{C$d#Zghftw`y6<g%_H)n(@7H$c5LeSe3!+v
zNFW1mu@DGbQEzAh(3T)&KZu&4h}BsNyd^z&nMEU{EPaJ%1v8PiJK%<?dX1lBS$>*r
zn8|U~(*sE)ZM*#95mBx#*oedaXpgsz7|0b$1HsLzE>id}_@SX!7+e@`_88!>4OUsU
zMub*beyOy2yy}MVvJn&OwSs76lan_3!Eo#6mG6F~F6fgSyWdsegX!z}atVG_Z#9o~
z<KKx9C0G^Z7a+EIdAqk?f4vW}-d>O=4v3de4<Cs@jKJqa5_~<BXK^2z@fptZ;QRB{
z8NR&v$A|Lft8b`}RPI}s{r+Tf46ScZ$Bk!C$1xd+(Uj7^HvmrILBm`xwl86{z}p|U
z_myyF*s8*kVUV+|wH+=aibb$LY!;CLZSohWX|u7O=kvDcyC|fGEfkZTJDQ`1-j3JY
zW_|-GBKYQr`SZFtuAPkl`^#Qzqp6+2hovBF=yyRPpNGu{yYAqaukw7Cpw`Dd!OcrI
z`%b8RqdQ@k%|j0PZJ<IDYCy9Uw107Xc^vl@2gsvOFFx#?9HI8FNoHxGTiIKM&wZYJ
z<9x49UNFk&Y=aDFr`1eb$We~`O)qCO!<4)XpvSm6<Wk8uW(Y5t7iEZc=ncpU1MNe$
zQN8G^?aWE@F1t)5xE5*hmB>5ufo53@0UDCR!fP6o*_pW@SSS}%DW1^W2@0shwf6Qm
z$XIkAuuHyN8d(pqfW7K>WpZuqm)6`yNmwD_m%qP<e&Td5)Q^oJoh;T+AAsV{sOvJc
z%ZrQY765ZvJy3tF&PiN98A#s&YOj}I(6mMAyI4+^zsTfH?I4+WludRd;c1r?JH4gL
zgb8SV0xl%=r%E)w#5_#9mIPKrch8MfVQB^8F->@2DS6KUcbJH~uD;FYyZ$CCLXOY7
z5voYt^ugavS68#=s2G<%{OwH_>T8o9SAW+X3DZnP?i<7;+RcNb&_gUnda5itA?EPu
zjphUUEO9y@Na!{CQ+!iWAK#J{J1cx&@wJ^tXIW3b<*IY{91vXP=5FLA>INIkq9pTK
ze7R%igxW7C$Tt86V3Nua4UV6;<45ydnX%W^Z|K=}1k{Xi!&s(!O-;TU?bk2c^m`J&
zNO)<gsfE11t-l8`^joG;&zHh!gp0(Avk`c`^`KuKH9Ol|hh|elsf6N%ORm^K^%C$l
zEpl(YD`s@R<PHB0W2JwWzrDK+1~oA1wIedOJs4J2VC#6XmUp~8`B*F1=~g>g*m(?&
zN?(j~zJ=DyQ6)KL*Vhx>uihgtH2x6o@KHu|L9^!w75)<4zFr>~Gu;}@I%byanNzmG
z5pfy8TgL>@9@A~2a8~*|H03*9mARo?3wOUmhUc&Cb41husW#V8v{)D03MXz8zo)Rv
z=mHs0NFYCtY<4p{QY{tgaz#l6LKthYa$p6<rvs8WG6cPrmX@Gx5N0gR!GM_w7f$6o
zR&}@f*Jj_NAIG2-JTchj&=V$trcpV-K-=^8hKPimv}FlO(Tuy&gQ4IdNw&}@X`yT(
zietpk0+**3gYDZB$T*BYjq54&4I~^K->p6(xso4%lU3De>L(kmQ@{(-*tg+cS+9i9
zG+2xob)R^fUBKQBElLXZ=1^1Axmhd)ED!ve-a~Y)Rhx`|($We`Dv*!AhGTx8NloKr
zqahL#(FuW6Wq#^;CtQ>`Wyd4%c=Gyly<kFWU6yuFLS=g7l(+d4GPd+SXEH+gDfr(O
zI@YL)(qg^qD~Ai8XBmUOAk6c9NuP-NOO|M@`Du#cOBF1L-uKl~+v70LU|*SFi;FBS
zzVrAlGpn5l7=})o!G`p$4eqt+{gEKu<4CIiE6tIU(^>SeX(|R38LlF=SQ<Woji$0t
z+%>1|D&3rN;<l6P_=F%v;(!b@W9RRqWY(thMcgFwQ}Ekgb03$DHyo(nk{t^gFa=&u
zH!<h+ggm>s0l{4+y<wQ@4K8UYrp0B?9{}p7;mN~!!`tk5@7(}-^xu;sB>0>N?f5OE
zj&p*L;;rzZM?G$0<Wfz{mz|j=$4xo=5ZJPp_f3pWw-YmgfLeBPuG3FYq?mFg?=X+L
zhUc!0U;k~ZaL`SnH?9!YPrWPJ3Y@veygI)34nYn?Zvs}pEK9S_>Tv4%booAs>i%_K
zx0?(7uQr690b%8&7XupM8U^84O2kz@6+%OY+uHO~@E#83Tp<^=S-1`jo@7s|^0Pt@
z4s`*zZBEW;yI$Wn7fi9IVeeSWB7p`5h6u#Q+Q4Sz6*j{DN}(5X!uLH<fY_?%Cvts%
z;(vSP@TOr2YIJ%*@sG!FKKVo-N4U^#)PiC)^i4Y8*&JK+B-U9;f6M3D!^_Z{X?kK}
zVw}UoB}wwyL%|3RM3HMc$s(F%uCzqK@Nn_`CoU~oF#Qch$*Mf|EM51-YJ0U)Qw-Vn
z(twmktXK*|&Mq-S&~1p$%5pW>qN!?DCsD*)N1D*a!x3&w{hIL7^U1ZK49gp*bU^GM
zNDzT2Fzto;x8yDox${q`(L8f6t4*&uUJDghwHv~CH;2G<t8FPUitN_)^ZGkcqVIJI
z>hoiw1W~8^emcHAZcPb&v8TDE31t|3T18pfQ3%LzX=bpnxC~v|&V5Z?byL>Guw!$?
z()%|5#a6AX%d(;^*!TT--Q3#R*-0e=N{;){SkX4I{|BH{H4r1~izp(~q!1pYEoV>)
z)W|*U;a!*Wkt?@bu^VxjCv**NoTK1_wkRrcdurP6x-;^!A+k|-GrD-b85Bhl&ZFEJ
z*u~9IS6~sy6_3JDmwQ-&+v0tgIQO9z*#Cn3cGo~^Ejc-*n|ORhU;bC`{Yh;kg%kOB
zt_;B^<S>6bT3ZdCx0(>1PswZZZt(csjsF3dtJ=y}{BeAZDr=ZJd0GN9aEHA?QCTN4
zZfXjJbie{XxrZ`IEL%_sp^9w3{e&aJ41xiGQdFP)hkrAV>nSQ;vgk<6^%A}xz1SEx
zr?Iq}eReZ!QO?Kxvedj3d`p$nT0{%qaEza9^jke=>!b_B$EsJQUr{E?D&AK$n#{ol
zLfm-H?jo~VsvWEn{4y1ZMBeO51_v_kA0w2s&-(<$q7hO^%%tBJfx9r*C<`lqVV_IW
zbAJ*a(RB<+K=wysyAqu^?314dzZMtQhux#m(g5S1M7aWHUiZHm6Mz2@n&HVu#0mk6
zU$PIAtfgrbzJgDs?pfH>AEe`%qOiUM-fEpd+)7L175RaAdbBcWi~Gy2oTy;@luVg|
zle%D45>?Up>Ym-#)(n3!0IP{>?lRXc$kAM_pDFeIrK`$^OFZH6dY_z{nltNJ(q1nI
zR~*{%U(6cMj?_Z6D^Hki%k3|R>z<r*e+a<>vk!_;_zhs2L{MZ+a|sFxZha{m9TiD!
z+8*O2w3Y1F65I(9)b<bJWI}}h=KB6}7LzjRWU!DX)^k>vnF1Cx3a#EEV`sO)_SDu^
z0S_~E@NH}<^!Q=xdeH8G6$qNnf{&l!2<CMi8PA&TiAaZTHHvVw8mv0g`91BQPa8Iw
zqXqD`vK7fka`g?ctxbO?B1kMPb%enC(n<e*QHZci%bsZTtuZ5Jc<N?-33Zf1d@Q$r
zE{5nnRs3o1uriY5Xehai0nP<?o>$YI{S7m4MB_f$>PK#We~o!`XyMDR?2*Ovkub@~
z@ZQ5s^}geNHAHEOnyydlNetp#!>P9o2*vO_dEY;8yI&(+OgDXy)QlZ4OXk<@Zp8UZ
zui7~<ZSe?Hxmpl&T214ve^OvK_~u@#Mf@IMZz%<)3BFVm&9k^A9^gRK=4W){EzXtV
zWmx>~Tbwg{<V@dUATHhLo`v{PRxn9)k!0$OILgI3{baQaF+Y!preUGc?GdlIHj^7K
zxjsSA(u0R-R`JLd^QS^`VL5}#J{tw>#)>^Ai%~9E?Gx)(+;@t^w68)NM;7liC5c*4
zTFQU}Rq(~J50RC!m!Z7h<d9A&h``44KD}+B%DZKjl!TFkgi6X@BFsQLW1)e%mpPH%
zPT&A7?pYEw%({}XInMYqxsRMdn4ng@%H45?am$ICKMSifqyxWCg#0K5N5~0_M--~<
z`PmYqE_d0~)T}>)t?ES!{@5(~=;`$GasWa`2Q2ZK_38dz!q#NM2`y<y!J{XB01mJr
zH(fO(hAE*jz}3}X`kxFY5Yd6a--ky}XtW9@t}~(n@GP32w&J)ASLy_oky;@pC&i(_
zlv2IvkZ<t^D?qEG(Z}Ec#^&t#k)QRnvkbGr@bHQ_D2;i0SKSlLIs1KcuhA&F_p+Z@
zFq7PQkTMaRVW+a(C+E~Z!N`tP+2TBNh}p1$Mux&gYI`zmLJ7qlW)CW3?!H9q--rbT
zSLk_F8I{pEySJE&b_Oe&9yYA5$wS2XNAbg-#)u-XGT${RRhr7WDO+=G`getcc!JAy
z?dniEQ^_VyM38cppwVK*{mRShzxHStxo#9L?+EaYs^u)BL7gJuwNW^W>F<eqel{Zd
zDL7$at|Lio_PPL!)}9k{aXWm1xWwy|hw$=^S$ZP-$x*-c2N&Y1)Ua61^Yb=chDG14
zand}C0Nh6Soc>Og@QU{+#KvgVX6qo0?T>0I#q5=dnV@xvVD;>++>0%hcypG@2Kik4
z0jU_J#KbzElk3&=zaBq6lN_$a`lANnHUFg9ki1uFj)kt6r#D=taC&y*C-SN_2zUU2
zCYvC)q9WFn8&-2lhfgmOPf9gv{N>2ug}&5BokBkcy~a9QB@UOc`YIR5!G)<{gn?#K
zmkFH=p@J%5rQqdQJ-Q>0<_vdM<dK?j0m1|1CCYCSst%(U9|5PMqsmMB>-2UMWw!fp
z9v3y_ptPs_W*^9Cx+rCmr>SYZ&U`F%V45$8qUEEM_^kbFXN2~QIeT)|fPQ3BYw%|M
z5wuRyfl+>MIM!-acJ|f_PJbCrR8<KTRlOE+Dv~X#M5HN1AUfF48Cgp8<zcS!+Qnv>
zNXc`+M>#Ea{W1(x{S|(@*vIF7RoE4FyN0YPuNv)HGkJSSwK^s-jeKjA>nsy}x3~mr
zqmVpW$XbUmEIjOD`1!nFN3ex=kB9lHPau*1U?TujN*t6-m$FTl>eyRUCOI60ZjEEI
zr-IN=6w+q_FQH^^$nY%mlnc*XY(%5|ekw958;->-UMJQ)sIg5kBaG_5<HggcZsrbD
z3eCaMrs`5ZVqyIRUB=x%2VlQ*erTM*Ls{62H%5kd2pwm~4{nscLyTw}QntUWw(^y7
zY8_DjR^pAt!ap-q=rdu7b1E?XV##3b`I)>MZT7+K4cIMjW(xPZY$U<YROAhb@xS9y
zsHL%e<h_hXGt4JuMeO@A^K2pr7X{`E4ulgC*CjOd=2olbrM5h#q<ZRUUL+-AF^r8t
z1Y%TuEra8<trMHwT)a{Sil{uJ<qDU*$o-AHHZ6ISwHp-2@1Gc+X#G6IR2fNd>HAAJ
zzLg+P+4sDf>COBRRNqQk6&7Qa!c2zt*W8!d*l5K=%*T=w)pc=EU?5*iCr~siEY7D=
z50T=}*!1C48d$W1j3j&(Bb=QTH8`kFEJx{3i=v_V8y&5~WC$1L5GJB6;)J3|LMG5?
zckytN2{v{mQ6!l%{D<lz;stP+8otaol-<Men`+iOS?q=Zb1Ak}y9DD(LM(t7GS<Nu
zd!X-K%04(?OYBCnr^{woUfZ}()h+|CZEj<Oc1C(jC8$G6*W0_foz>l0pNpfQqD>IN
z6u(U^rKpH{bd2Bgr=~F%r_T8{XOEXVrxbB_x`0=_NJ`-CVLRrKJLBD}=WBNuYx%{!
z7W8lvnv`VNY-l#TECRN;%Qvwf-9@T$49h!JnQVZ>Jc#GnJ#@d(oLy9GDjWn0E7t1P
zK4ppXl5uMR0W&FOKDzdC!z9CkM@<Lc>rOKuzTpGkAPnOl8jN$*)?oW)%$C1kKW3n9
zfopPi-xe%<evWqRvi7aQpmmRJlwWaaM-OMtl!pKUhG!meniVMJvk^L?k!WJ9%8KE`
zKHsIhPHsFEJBm9WEC7oh*OlOOY`3(1`$-Ijw%JZVFYjls^-@rxK+1A{dpXkd@UXuK
zqMmYF(d!Xp=TYK#)`Y2!b=m@v7zrjSeF0c7@bt;UBXC&Kc@-G*l+{{_VXD@E(r)}t
zt4|0P0`c(jiss{1NTAXac)#f?%7g9goj&%$Ly_q|TzOTl+eG94`t_^RJEGtWGbq&_
zIxv2K_G?W$$Au^5ID(?|LAbmguG7a^P7Zxdo2&<&ZZLPldbIAbXY$!qu$AaM{uB@#
zNR28e1Nkf>Q}`~Ul1eY*Fu-w!BN*L{zg!{akKLLAw-QS!p*9#Q3@`v#_Fi3(B)P}n
z_GtEDuc3pmUht9!OoMW2X?=L_?&E?~Y-t%J1iF7W2=S)NwLu_nTA@Aru&$!EDp^A8
z#bU*ur8ZxN(7rEfngch75{89~&0PF|<utAB+h(^ruq}}ohp}tuO!*l*fHb*Bi{6YY
z5RVbs0_~R~DVkJi_#tO$M(mNOq+Q88%hu4es=7g&jw$}h!08H*sLE?W$tK0OvM3n`
z%(FZc{tbW_gZkUzTktHyW>5FWvX9|O{N&&#=iPVCu+~N5lAHmi7lqQ;Dti~2JIN10
zOn>_NtPRsL3%C5i5%G1wG?FZ9qrbLLQLWJmX!vz22NyWI)WW_3i`nbI@aOECl>G{)
zT+IZP<;S9hUQ)Q{;bid>Q8ic2dYe_cPZ>|a(3hW*iV%+$`ug#|hTWg6KfpmP)L6PM
z)Ek*B)mn7@5>uL<Qqa}({~Q)SganmslHwcq`IL~0hVA39=^B9|jiDED*%coZ#sv+E
z6m=|k_Sy}!O4VGJTZk>aN{?hr#CshZlME`RkL8Fs0E&80F6`^C{r2GK`#OVO)@%E*
zQpbP}Pe~WLi?g$~s4)t{pdz4e7cfi`nWRc{9{*z`0d`we*6Km#nghTIk+vZre1Sxa
zTVUF1Iz&Dm7K8?t<38MJ^jt7T2j_D2u*U4Py6@l8pUC?i0DKO_Wt&<zR*o&p`!g`l
zTVc?oM*t(*Vh)5!!b%*abj)wlz%pu}$<{}e^Nj+J5$I~jClFfjqq^z^t~Sr&UBra9
z248;6u((5jyTXG$H+&1e-BmwE{wpOg&6l}Vf;Xz4W!pnV`~^!6Q|duCl5zXF?7LPF
z1WI@g__(Az-L9!)<*OJpJPn2he1R-_IkrYgNXG5W9oQW5u(GHcn)O6kLrMyqGC=#S
zqOnnu%a!JDAz4D{=bgGxwJWz9?iopcN@ysf?NS;}A8|93tn1JyJRnz-CK4^*R*?X|
z{wPJ3&48Ox`PRQPs9b3h9)K`FDiY;AFgn;5mG;3PeA2DV`V=*1@85e*IK2y>C)DAM
z*&RfyvX^_-QBlF+6I7~H?){s@W~MK(kCb4Hon0_btR|Di|3%rEvc1KUGoAN1aqGI*
zmJxM25SXobUw$h>v2~uFoR6n{(e=yx6uO!ZMbSAq)=aIx9QCx@&4_R1tjaZM<;!1f
z-!(owN!=E4Hih4edb;|NiISFlM)`JRu|#YDrHF0RvNd?}hZxjbza@^V`XMO*FrCc7
z$i++c925Ppw(i}bRi3n#24xD>I?j*z40qJLo%&>mh0`GHo5v~?x}Da|gCbKWRgAzQ
zI?m4e3?7&B%vfZjOLfN+FoF&>jv}C@&M<f7s`>fYWqlozH*)~~55o#Yi?1ouwGO>W
zKnwrl@j1Y?ol7$Nb&H0+ymRaoqFf&`nq#2|KPBXQeBnXY`Cx!DLn?;W(D0D|q;2+^
zyh$Dkv0*53O^XLf54xV<q>wX#R6!&ilOjQ!9Fb=kQ5dsUNpXOpz1-G!W#t;XIR%33
zpTeM<WcrOLB}<$^8s56TD0Jm~8B|pJFIT`$PyF?fF**-)S`jaV5mi0qkelZLc^hxo
z+3Wl*kbY==ADO7)q;v$#5t?rz+uf<5@z7ZIx#r81z3-@P<@|iGr{Lo>3UGUOF0{wm
zWHuM!<+<mPa%4X|S<SMkxB8e=5^1pefo-zv{Hs@g(HtX}aou61R_*a_)NJNEyJeKV
zK{#6{>i62y_k6MDN11ez75HR{eCT08KKwY&i>*t<G&kmng3d-A8ERBk?XWxaq0DvF
zd7eDHLy$dr)2O~W4|@xjjZv%J>j;=PhEIq!YeXe0*czOr8*>RCD*yPY1~@xyt*U}A
z0@E}1I_z-a@j^gn1s*0g%<1FpqbLk~X&yRbHbZKTSQm`!fp}bQL4H=v^ItS{v=8}I
zUN@5xo#`vJ2l*SGMC0MiZ#--78{VjoUolv1@1A%<F(?aw&H=J_RotIc!&^G9y>XxP
zD)@{pO=cRjf6sIVWCX+?xdMaav6(fUo`EO4kIMT9L>6dl-4P|&?3ll599U7D54b@9
zdm)+K;tQXRtIVYVn(H?qiiI{aW|xx%q>BD<-09(fWqqv($@p?o-1$BqyOdwwd^n&I
zFSq-1mggB)1TDmQT-Y>6pq!jtL*E>|%ZJ;Qdk#-dP>z)pj*g8OGn0L?Ls1+`Ukf&j
zg55Rc!=#+Pu(~iZ3+k0ludPu9mS!psV1mshQCVFse~}Cfj*5GDP!$(9z$=R0O*6&Z
zb|YJ@RYN5uX$Z514cnHlknJ@wO8nF*92y(ytH$NbX={+>D!{UC6@z6aY$F_y9Y$ap
zf8BUAKep*@Sl9d6Mg{tu)elb$m(C0?+0d@-$DT-he!dnocc!|7*JRML0#tME@_IN{
z$BQ+@62+XdR{Nom(FR7>rl~p(SaQirVW(-l7d!d=_CP0lfbD9-{)1^nhTfFEXT)#b
z^Q3cdJr<EmySkannGEUIy5zD0QDVGhS2+973hVZIpR+v4!dJtV6X{Z>qN;Df)aZ#s
z>}EZ8?Zfj?67{9TjuIIiYJ%5E563VMUaM|Ehb<&m9#0lAfj*p%u>0!hPv|kU7#Y9m
z6sYqsy5TLL92b>H{Sl(wbJl+&e?QHUiOBORbmQh@$J6(b)h{j0c0HRR_rL^uPc6%i
zaTQJ!2GV`U_iJvxw}tH}8$%Am@&;xnna!PKMIGi+Z+1LTw&WQJQH|i+VYFzox6U(=
zSoQ?`ix8q^y0`atcl+-N@;e>`CnLH5k`%}BNJVT4U`qD$VPWu|AMw5NALA8$_&D%J
zNrbCR=Sq_lk3U~t!$EButrf0jCB&!D_rQm7(E<ia0M7e5mq<2sqES<6d2e&FQl$($
z<K3q(Wwglz>=tOr$#-xc0;0%gF?U+gVG+#<v^k23{GRw{IqD1u1%I@x?3Jfwbxh6P
zKIjti{gTg!dV2Ce{eX>h*Ou4Bf#O_X6)ut{k94Q~Hv*R}=IAoh!ScJLCq{8{N-~I-
zsS(n<@~LP0us-|TIFgEKzP3UtHGo0BUOfsa`)=(H#Siv3Z_GX;JtcxN8+B%tQKFm4
zXVUWVyW^Cji@vDl4e#GnHBv7qhmcB}%|Ens<%|DlEo+DO9;<0cL#Hbl@2rTf`u4NN
z3o>qP^U(@L3*NMN*!lbhXBcTE5;znsq_6VdbT!GtnN131v*0Hmi*?MjoA(<VS5H-6
zyc&s*3D?$m!w@YxEP}0lb2r{Er{#D#&dM6<>MUa=+2RUSJKp;ykex>`wC62xWb9&s
z*l+$wZTGBqdzNw}Z;D}pVWG0!ws$Eoz$bM<{`xaj|HlQ;Zvqp%lP|0YIqr>ws)r0P
z(WUXYmb<aYIoB47OqoT$GbpVRIo;psXJB^swcBaOTfKj{c66}sj^s4oj}p(3qN|lm
z_+zZm7Dl<=DC(Y3)+(>I1{CmIRZ;RIVoN{gS&qm@(o-Kol!0o5O3v8}2?Lj0g4Dla
zinfP~^E=p~ID|sbC;@*}2x_}8E3b*NXHyi_q<>7O(?ifOcT)t^@EL}2b5Gf7v62<5
z#td;lG5X6>C>4rqD!U@Afz!*U`@!tYOd-(T;8#yikIkB=lh%xSZzrguFW)um0)ad$
zOS1hN4r{*$$R#Tx`%eTeqMBpGv@+YzKBSESe|yrpfPb~rxx5x(b!(fC2AAK}#^vIJ
zn|uk1I}J)(%Paq%w$_cy90~V}`>1VS_=;ilDdxIzR6UcM6`{)hb$d;qEgP3QxY-9K
zFwH=)rz8D+Eo9tlHJZm%X`Zm4RWOq^-qdM}Z8Wx~@dvx8q_xgFNH*-ta>J?ocEtSW
zTtn<%z%wd_R&DL4dw-oP(+W9D7gA}|<dxXkarj1YRqY5G@o>~K-#P>c?co6aR@l*w
z7fDq#L8fR(29>T3uT<#z>fbd1MD2CGdZ(iLPD}RWz#gW%i(k$^b2mDLB`X%`K^ZE;
zQpT?MvS~bGJp+TuLQ|OR7FRnzg^a^P<IA&26p5sckB`NEXljD0ObfyFvhbiv0!cob
zF2dbo)bq%jn-~*(Jp7;qEQE+&qgK7>*>XN&vY#Eb7(^8kU<nuFkxY!3#_|Z<1djfL
zsVq;)urDu<%h>N2PRLa@F*|LmEkXL|vYDu;p0Hq<i0GWgn~9<hF>lL&Kdy(;i3pp6
zc5F6dLGn9h-FBW;ZAQ#tj$whYx!Lf%83c)ppnN|=>^diY@_$qdl^J_02IRRSU-ba_
zW}s~0_&NLZ4+t^dh)|YCKvm`<WJcz{LGgLn)&g182O9PxCo=SLsOh0kEpg22hU@WP
z2-!W=D$8|>={@6xg@@a(RE;UtHK?wH&6fDWJUlD(v4ZKLO!D&78|^yG9ePkhjHZ7x
zO1-A#upK28Wc0xhd;K&^_l7>6D;NFl?A+l!8$nRf@y27*w2O6OxfIfJ)`~!#YUeM2
zk0<---?_LbCn$NpLv}fpPrj1cS2&T)rT$n}0C&4R2`r-Ni6OG(i~<cE9Ya^XY>I!K
z2bX-A8O5r2>bP28@1!+ISP~xOwOT{HAs$Krsfaj8oJfqJm0DO-^m+a7hnOF5bS|Vc
z%!0_kq@P$zIPQ(}Ya?-YZja|D5Fc^YAJIw4)hN0rt}gTD!`;HN66}68c){w(qoW`8
zt*18Ghtp>I&v({_gk#JMbRf+iKfu%fkEpi{i>mwnhmi*9?ii4e66wZ~l<w{nq@=rH
z=<e<Y>F$(nkWT56?&olS|G(>b!R0G(=IpcgUh7k9kZfD*A|MyZ65}HHGAdmjV~e_m
z&nA5-dd2}=;IdyYo4QJb(S0UODqvD%#j6toV!BIkcS(reyZa6EI*Qr-Ak^!5bHcx<
zyhknbKb*7w1B>q3#1sy?BE<}_N(wzUn6@Z^dE)o{96`HttV)JLsg*zjV*tT9!((`F
zqYV)llj$xP-1}?r7<m$DnG~NGB+X`cTBi4G(Z{4fQCeDpc_t^ctDMfgC$jm`?MUyn
z2hdX_bCd;}r*kOFiyZCmV<@=9;;89+*e)6bhh(MUt;Q=qvg0+KG%UZ-#)fkDx?v(j
zGRALfQ>Ot^V1&s;+4X!q^aF&bar%q>WQ0-!P2BXtlc>s{ytF0qNh+RjBAd|4Q~5t9
z%HoC^`YuOOW1xPJWu^?}g~5a%-nLJ4u8&a}{(HD-?A`u)VGgu`k<1i5uFX7udYV5>
zpiaQ_v$w?0T8$<1z+5|CZwX9@moAWMi>D)~ee0y-b`MWm{UE}+o1>(bSNQ?MeN#DZ
zs06jcZ4>|EC)dj*!rch|3-wYoH0DA5LrqCP%D`X&MS-3cPQe`zm(@M!4&yER2W>fS
zEABsa2>hHtJmJ@(Sym;+@&ERV-J16+42k`*ncu4V{gK}Gv=j@ed?Uey(^}i*<PsTC
zgwrgmcrZZy%;pT`&13xpi_M%Yx3X|IPORgXOXc%sV+n}K?#9Q-4b)EX`}3W)vE>h7
z^|Qzr_MVkLt%-(-1Dj8i`kS(oI-E+a(SFuL@pa_pB6*d8H0-DD)T^tTes3k<+^0sD
zb@YKpUM?*|M|@YR1<AS$VLOn&Xj%niP!42MsBHFsQD!wjaB$JNxnKjL(u#_&&Ngb&
z;4q#p=&G-Q4$mTfX!*IhpxK4FKyKHe`O=DH0n0>}^_6L&pD2aWDWFJ`CN^<kFR*ZZ
znp(zpAW?9@hK?ax-ZNBfMwlr|Lg>U633Yb5O`=YG;KvCBg|Nz6eZq>2Bu5q{Ke6oR
zTVDI}Ma*~mAv`s1dF<N=^7pXnpI_9bn<j<YNmfc(UEfs7l8Cdd=}Pa~unUw9$uk=e
zN4^hy0vHdCQ$VNY4ZCaxBic|f)i;7ICx}@y@MPm@3B)GD<>Kpk@o<k>y7Br0`lX9s
zNBK>wO`8u#{c`NUsgG%5>dMn$_s8T-7DZ{fYPWnsz~!Yq+kk~MuLw@}OQB(n-YYI_
z(ad6rpn-pTWj$awl%W0@?XNoQK8yF(ajnDsQkI=`Dg&;dBngmwlX;W+uf~ys$jI?*
zs4tZMj)B}!at$R4<{25J;I?Q#e}~c`lNHsBTx-2XNCxo<MDxq8i7hMESi4?vboI}t
zgVR$nfHTgX`KG9+7Zp2=wfSO~TinELtb=T$^u8#kx+6zVhqfchF?atZ_S_v)QX-z!
z`pnuJU0hZWRJu3~FdU7I?GYg<aq_WRSmAzg9A0#Xlbe@=GcP{x^_tYVz15xfr`3|S
zAQ$(a6rq%+kj-}CbN?>F>!@lw&#ff5cn?123ONu+nqe**dXTP4mOEbcHYOsA#}G*s
z5k4OuE++hX<X$5{Z{(|H-qg}eHPU3JTE*Z>XHWei{4B>S5cGG}1nX9)$uwW1^oJfY
z-$<@nO*qiYvQv%o9j&7sfSpVei%Zzl<@sYqlUbFt*=*KYIPx6_R!S6{Lm<S=S07(@
z3ykQ<x~1a?hU^Qc%nglZLjX-QU4u4rIL_iB+gK3nzG_<=`UZHdHO!!WJO1_zUFr1{
z;k@|zB_TXKT(Y-L2T)=UiD3Koi1zk_m6X_koIH_MAW@>YqypPY!dF5hL4(7Aj=8wr
z`sQ?%)=PhGIwt0)$3*I93%=)oa0NtC0e-LWz&v@HBhKnbB^4Ewq2`)i$KL+Yn2Ltn
zYW{IB1^9J)_;O<hXUmyp`qr-}TnUwemJY-7?}S-S9@n`*{{H3Tj5j$>=>C55PQMN0
zK@?%njB=Iq4lSi!I1}H+W}+inl^)~yB+MI=Ywk~!lx`F3+38vsBl_e#%oI7Y`m~0K
z@6fq>1*d0E4h}Po9Ge@@>wO?ntBvKJ03GbCh~e&V0_*i(K(HyUs>=9VpGZb|7Mdsv
z&ek~|ca}BI{vyJ_G(KEQocJGT(aXe@LiItxE1^`<n3q>39uGgSq#llJ8HOgt@Grcd
zz>*Yh#4+{1^OO8I*&a#2d{6rIO?=4uQ3O<FK!WV^?;QWmr{MFESc`7L?r9xp1<-I`
zM#daiNteGn#}@S2<Or7rFsIHV{KYKn_{BysD4W5?6yHrY8N7ds->-89E3!)G&0`bI
zK^Snp<_d<0+hVaX*4m9T$H>;fTVjCg_jAAowt8X}fC2t8u}r`SoqYh~i0<e64z-+O
zU+8_tV)}J4_4>Shv2mev^!b9IN#Hcwr@`~_-O6RCUB0TiGI#b+w=GA=UHf<QO7xJQ
zdCS119grKPi4QqaqyUkhy_Z3;sOe8DgT^fRUWy5gnIjx;bsa1VLKem6cg7T4!Y&pU
z=?;O$Twa|05ft$oxG0InAipI4?0>%qPXW{*OWcYXk&Y{&J4CIKG=gWpSY)%lv-kEi
z@1=*~;^KD8RMRN%%Tp{ZEP>|FuW!1JZYXFO05{Nok5{QKNi$gDLrkzu*7LC)D`Nq0
z+vY{(<A(rnd7Q~R=jF_EyNA|ZzR1fI`5lgM1kb`|GlEtkHh?)9vLraWcxM5n2HnIo
z+7cFj>wT3{==Dt!X#GT~F(r_N9Nw`SGHS;sAByGbE9j~&vB*MV%;XO-N}w^z!9_T+
zbU1cet<Go2ydnC{28eLfo1r|vrX?g1`dK%(E?oLp13qNJk%B8>%4>Jfn*soxp2PAX
zspPwmT;<f=r_dzek!CL55#UwneK>?;MT;~<+g>LiHM`-p2{<C-$)TXbxn9Q?QW1|g
zDcoQ&dy+&H)((bV=BU!cuQ9je646SRqYTZB4)9Y7T4sR-%?sv2^7H<RafRFENzp18
z!0Li{(g{e#&E(lfnXzwQxA3V}b;jt7{!ACNswyiFT`0-|2wV&=7NC!KB~c<UBxwX`
z3{>VJWqq%#J&T+%BO0@{yW0%D;N#w?;Lw1M$uDX|SawE<NSP)(`49-2r~GyW;Yn~m
zZN`5rC(xn)c`)=DI2Z6FEMkVRAYMOY{yAZloj(t{VU2_!l+!leNj&_4!n&1(I{yka
z0i}M*5&jCGxUP<c3!f4-<gzhec3%%{9Y-w*uoH?d_}e|F!uo(ZzmF;vk$M|oHnA)J
zh7;L~<rrG3a1_e1h-<dwFmAsOMXf=L34opwC+38;j%o6dBXPi(o=It|VR>RD-i>wn
z{=LK&zN80zUCp}UkYx+1+X#Qe?%?v)(Cn7^&A{V81@qy<yI`~lritCR`qXObBEdRe
z6bH%Z3xEW-1aWfO1mSy8j%(jNi2`+?H<q1WRHhSprLHq->C$s@N=6WUN}R&FGQRdL
zP~R9`5QkkDUn2~&B0|3YdQhKR3&wtN%d|xyRgj0Ot$Lp@_8heuYGi}N@G?0iayt|3
zt`)fW-<sjjV>1mf#cYmctLJVjBG55gva?bQN=fvw#$^CoK&s-xnQ>c{`p=>YP-;*B
zivKUPnB@mK0gk2sWFlvD2q5z%4%dEV)fT(M7%>oQnwC^7+(&=s@QfKy4A51)r>>9y
z-hPd5n396uJSzd+dBa~G-5-zA_HhfJI6wQscm4iqmvh2_5=z^tEI2H`CU;IWphx;}
zZ4x|@c}+CH7aj`Nb?XtZsmF-)Zo>g8SR?G`>+6D9ssULE8CFLb-u3bRKEZ$tuwva5
znoJ-%{2`A7P~=Pv4!ZG9igu>qS8LV8q3(E0`{@WUrq8k&u}8r<CJxsV#ta-bGV}}r
zmU>2ceW<oC?G<3Bb8vvSM8g#KMHKE1Kw4~Pa3i%w78&_y(;d{{M3Np8Qc#=Nm9gcw
zjNL?X-lXKG;O$W$+(+EK79al$WKb>-sSabDu`lcm?tGsE0i&$)hplU3IUqGZd;PFJ
z>4}L-4-;$x9}<!(*oqWO>`byCi|~+bw6BhPgAPC+hmo#As!G3x*x5T^pB#yG7V{J(
z^S`OH`o{Ic0g`Fjh8wSDKq7?G6xlFBN8{PU9#|NZng13B553$2YQ5d!%P%>b_6s=@
z$JrQ|QzJ(!t6wyja3ERq4|`*L=&#RrMi=&{&3juX*GFGPO0Q<wG52nUm+Fxo$M`nw
zX*UYgbdh9hnSo`S3WtmRBk52u=WU9S(I2)?u<*#`^9M?hL*#jjDnSpKjfJX*pEgnB
zIz)S(V$WguvxkGX?*#iRzsqV=$m2VH3#PkoVq!sF$EtafpL6Fxpi*oJUX`7iX-AMK
z3~$6650F1FU!vfo1Q<LT2CRBYK>aig2Ka*9tJAYMI7#2Wg(~Q+V}xteZeO@fZXdm}
zr}FiaI%Nb5pl_~7I<|^3X7L9>i{9maQd3r@Tf3Be9*>NC2Fy}mLQ)Yx=CSJKu`(5@
z>Z~QfuJ0^|PQ*vVN^gbb=H^jsQ7WdPi)3MOM1>p*vny5G8PNITxeNnBa4_JeWD%}i
zRZog6BK9rYE4p0nYFFd(JnmLCIHG_mJk0XI?(`_s=<oI@p;=&4+P=JEU@6u60fX)w
zwatBc@nFc8&oohJ1zJ(9qKML(jBhky6G1uaj5CwjxHFdRX#UTFYr#xAAhGJW6u^uT
zRVYA~nY#AEAWU-p<z~yh=rf8q%F~ud>vEO8EHyn(-1Y|A8{XE|0O9zo4FAOWd9O=x
zKjkR=RRLe~OQpSWT(n~M({^!gfJ^+n`1Ya@o>13HIMg^E9sS{Esk1C!=%GAZ$Ob5m
zq6%=C+W9;9M8gdMeyDoYx*i*kLJbb-RirI88C??y18SAQRnv6@YXr>!L^(_%HY+Vm
zpw%giR~UK(Nu)7kHBcnScN{-dFr5XCnjW^A424w0$>7^b@HegHmxI*QbGfz>G#Wlh
zAXYMfM8SQCmmJ5)&rs8utF)}le@ny8&JOZxaIvd!UbRVwd3FxYLio?1jwnw*x1opN
z+H$QNBE{g?i-Y|6!O>Cg(?SP*U(6pM+@~bstTAjC505!uA&y8@J{-YX9M_bL=h;B}
zd_@#Y*W$I>^XnQS&G5(JA}y@?iEU)+&oX6Vo>H;s6e9ZXiNHTIA;a_5(vouhiJW>8
zd=9kDirN4l@!{i|7X6Btg$y;aJWiNJs)cQDnoyTX)iYgRH!K59lq?l*JfM-x@Vevx
zS`#+@JLwDj?N;4h7w)M6X-;mZRg%3X+~4d@bHHu?p8l!WFr@voJYa-LvyFSFRswob
zP^%(8rt8Du@@hD7Whdr>o;T^YH-+br3lAwPKo$%T0mAqM^(vde=FGMPcItNrf-~g<
z6p6~WDSwHf)(FaP;QdWTlFUBFV9<xJ0yV{NaceI!eqEN&pFclbXS`NDdwS9_Fm(q-
zJf4}w<KTw4;xOK&MOKjrf2s^OgvlG>RI~Q28VnA>rZv>o&4fs8*9~rW1E!D7qZ_0o
zMJ0RemX?;7m=yYeOs`zo<ff*OI2sTi9|BIX=rc)JU0q#vHg0&66jY&fMyORp#53>b
z@BfTlR;;VfaDv?W>>S?Pji1hL+U>^q;2iMs@%EQ}JO;9`4Mq~b4Q819NMG2--RoMI
ziy+5g0zb}j?iB2WeHP@;;txJ)CtY0YATD7RE3Xf6RHEnRIQ#N<Gi6BF6yn2yq1mGO
z>FnTr^#bewVVBzQgjgvbsynZis7wxO#fgHJr}M3D`t>L@<&Ct2vq6#!)emzUvET`o
z4n8Meh&;u^@YTdtmDUSAEXdk<dkt^Y^Y{n9>kZYA<j9w_D%7OTpC6o@A4DB5TIO}O
z*v6^M<b<@^I<#M%t63sWaxOVibMnW2d`+Zvmv7hzChu5#d6hHlwk@LS$JHUg^=4rw
zA~rG@HO}h`?G8(#ByF|W0*qz5P|Rka1YDTpL8(c)h&#{1l-c@hj=VT?Q1r(q{@bnh
zTO8@OJ)G3mu({<0e6$LipDrT+L3|pXs~sTjJG5yAHtZ$!W&bdlR59eXq;x{FibElB
zCcpsjnnYO0<Z`;V0SB?9i_1)9b;lQn+U61@_nYJHQNkI2N56fXt<QmhIh}^BEuJ7~
zCJdg$=_)bQ_1hb`L-doCYU~rM6A;??-X$dyhYUn4TyeW@r31U!)iqa7_Lt6!{4d0V
zG%PJz*plCIQ<E58#Ig`Keos%<h0lupbjVyKCEo#ME+&U3^>V{CKBVbK_nn=B(emt)
z!Zb&}pdy_jsep{kBY6JOt(}e@m0wChPmO%5VT;q%KV{D!rn|7v)GmF0-|b?hNTg5O
z-6(sWWcy%W@MD&?ARVT?#u}saSh~+Zq|OjmjoqubTuv*xi~hLG-S1XseJQEewBHBJ
zLcGt<jbn7Yt->NVAvD^NjbsAnd4p$QaXEyCC)FTOXDw8r2|mS4obN*+uk`OszuaS_
zMT1i;Ru>1pDHG~@k#5?)i5?@opr%}aBHS1?!>8`E6KuUv8rna3jrJAGvkNd<@a31x
z24trIhVG)G(Cxkx+QefHDiWh=e(iAo3fddzy)38X<D3$qP`Qhl*%SeVO4mdjAOBE*
zp(k`P^K6?)w*-2L{Ws;W^E(_EZl-zMwxI4^xf(K}yS+|6GptnciSks~)(Tt`p}
zIJv~)^9);)3)^-G0dZ@B7Z&*ixivq-FpTLO;*$~r>%P8|K%}1km2xtg&I2PL&@RC>
zEw7}7kHdc)pO~0yx6C<}G+Hk0;N)~urbca`mu~pi*b{WaqU(#e71Gi{kcb?enuxb=
zX-We)VxX%A{o7?^p8J^)x=V_G(a}taW`=FNv|#Y#JK)zhIzzUJ_%jn9LqR0BS<4Dh
z*pJ%U+SHjZNzhq1nN1m)x<bQ)GUAR>!pR@Um~gKP;lEPEvf7vZt?X-EUhlrUx=6lV
z;poA<8mn&%H5vY_Q*6;_lEjB70#ks%jEU;J-}Lod<oc1af|1V@iP`nyCnn{f=2!HL
z2Q5I2j)&uO%<39Y73eDJKQHuF-Gpm$J7r)`j2~26f-mgLD&yPe{tPL>X0PxHz5ZIm
z*PQVh`GfZuSRe+ok#^Zh?-k_#-wqo(upMd<_+iu=6csB*&3l-fhn<c81~d`?m*5qw
zkNx-6TcUvkpxF}54{Q(W_Vk=wGSjVlwz_NKZJFc<VNcx&hW2gGjKYaD&Y%btD=8`A
zHlJs7NYw#{FKVigIvh`qfdW5BN<>hqb+Nv=^b-^8Y^@m#Wdc2akB(wk)6j<lsz3b@
zo|~7Xc%0s})X?Fv@oGzd)JO3tQDhMn4vziP!+|=$vIh5&FwLwzsO;S?>=B=2jBLdE
z_SGOUSLpZgJ)UneOxROiSY-C|ezAN0gJFJA;W3DP@DCz=8NgZy)T70&(iiqz=P_aP
z>v?+K4TL3ap6IlKEzn%G--$<yAc{EL^?m&I+r&uQOw5`{HS>wsW3%M)VNUR2_;LaE
zh)jX5)W!<qI^_7dsCIt>Xyy2V*?Y$UeFQ+N+sbn-Pp^e0pqJyEvD_^!L}w8GA*^e(
zB=4x^)ThaM9lK9XXZ=ul!paF4qw{Q!-DH7+b01hH02jE2s;q&{(t0Dc*8zIO8BnY^
z&)A6Td+?jBF4b2q%vYqf17b;Ks5YbVA_;5YZetmIV4bDL3r}&stZ3|8EGch-Xp90E
zS$|I;SI|{l&x9So!7MFlD!flP>+QY>)P8w`rh8880g2RU(<K`J(|K=kX(isH`@DwP
zR_pq`HT0G2n{ChE7%xOU{VKR^4`#z`KJcm?P;lJVjo0js&}!k&4{GB7IRo0g`<R0U
z;B7-;s$OZ^+rvSjdyqF*2z*bEWee4P%iO4i$c=+@ycM|`laGvoLwENcP%eSBbLl3R
z85$P$&alBjI4s6`ggFgna{ErkI^^`p@d*h?Q=vWK<x_DfF_eX0`F4(X>~aCI-u3mh
zU2_N>j@6nrp(1(+w1WJifxM>$(YJ5k0-(MYCCs?SIQk~9qwbRYBY=N2T1OZCWZ242
zcKGtINpXjPo*6mRqg0sM=y<;)_SuB5#JX`mE4;=Ytdo_i?&YQ1x2?RONDG^ItemYP
z8GmHQ%f%Whs5`G{*Q}opWvb`k3yCZk{>z-+_*ZJt3f#BC^fa4fyX*~pyLzmK{^a#-
zpTh99c?|uwVata2HRuvuQ0S18BegpqCK#IyVWBN!a{^fmaCd(q1-$;4N$Y<)sO|)5
z*@!h#Vv-ij*}rf*wKFU=;l5MYLWH^nwrr1Wgd^GfgS5S5lWYdJS>$UUk3A=&R*M6m
zD0_j{XPd}d_@wDscpqHtyvQCBO2j7qQMKGQJZ8IT^KLr|zXyHc{y@&gzFdb|VL3xX
zH*&T0R$oub9!wJGS-}MOV-R#*HUkL~=Ye9nT5?(t{)nyrvJ=0CI<hzdziGenmMF;W
zom1(;C$DDUZ4)>Gr8&pksU2EUEjqpx#GiM@9)WDK5VcOBHgHea<H5`btK~dn*?_#c
z6(zm8M81>+*gy<fYC54ut#V|d`MG7OWisM{5|+$5O2Ki3^HHDNyx7iRMb^t&R^pfx
z4jn(QqfiDLkgFq;;DjG9hv7Phx_6yo)RV=2#3;FJio(%NhXs+94Kr!=yI+&cFGcE=
z?k;UzY&z}SxU_i+2v9yUEFB(UV|qIfT&Q@F=WB&kG7nsq$mXicFF1!jE9Jc3E98Q?
zRi?urJf!LI6sj;%whqAeet7EovR(5K!nYY6Sx|Z-=x__)@Cbfp7^>;`>wrh98%?hL
z7;M9YI$H!S(nmrnBG?olUsqrv^@A9wQUS>Cb@4iCYM?^5_YSIS+3h%0eh50_$SMO_
zx6cKYN+knm5l*1y)%WDHgk->*IBfVx^e=$ezAF81ua)zPEA6=8Bgw_^_xcP#<VIaa
zPNL2Aww1y;fflhZtjf7P$d1=XNW;mRH>jgj`&ULT<M{hJH>gXGLLPn%k+&jxw3zGc
zzO+Rt9rBkzp}I9-Uvbl{v@8G=t39x=jolfcl7tGE6&`+tu@Ft#kd%=?-$&j|ngueJ
z7?ko?S(mVm1BWae;a<DIgF`<gFCt-jG8T>ZNn%&q56TW!EM`dHIDq;`N;t9!*2SF@
z`GZUE`tHkFXBY5`&t_L={B_Ib<G9fqoSK<QMGkL$m$z8ejNTnk=Mv=dhUDf{o+;~n
zt61+<8))2h^`#*{4|~0*q%o7le6+tGQ8N^Wjrd+hsJolDUu$5a<#FNAZA7U40$i^y
zf5SSvPRCcdJ*GGB4lnV;CJuNFv;H{GljOSG4IQ9w`|UA@Bu5SZpBA7;{85vY&)L*t
z`9Bmg+3X@$v=PXk;HDk0W4K!6Fo>3uIH`9m)^W{xuMZf@#pjP`#qTess0lA-iNLT5
zk}W~TS{AJUJO3RXf6XXcedr8K?VJLXh|IBq^6O@)j9d5gg~P17gDu4kujf<$aG}sH
zQ~oX^O`hZf#!~F|pbN<tdGq6wnQLy2RFN-hZAmfK{+`t|q92fj6!)v&D}6p3Ke|9x
zC~j&(WHTGb5OR1DDb=HbD`L_r6CwV|_p_7E_&i}N#xqU3vMHZBG6N|?+V1)lC2^lh
zhp#Q6*YhI*mu%b8vLPc`KNHUcS-~FctqkKC0ugZ|5?2&|u%Zv<@+}PD?m$+X43bF`
z317b<T@K6(9UMDd+N#d;$@d~Zj`2G_$`6V28|0YRlstTJb#11+baQt<PcZ8YRa)?n
zn!QMg6@U|$OQwJYwK!F(KN@+QT^}hV>p*P0|4k=1;1{sL=w1WT127b?#jwKSRD%wp
zumiG!Pp&-#loX(#=Y8Lkb-tz1InZ3YN9xk~UyE0$4oJ6kP_|WHj^_avZtmP+c)R(s
z;BNa2Tro}(P2|7rJVGdd#~yzC*i9~r4qIFH9^k4!eXgzPD<jM)EDPYeB4l+QC@Bh2
zlf^Csdt8jNvARwh^IuJV?cO~BVHITrQ2ja0$$An2dwqDLq>v`VrhMaCfiM$lA;}Z!
z*gNUCDEcRekZ!MjMDACO^@XKQfiQfg2fi>V8ZjGjxbo9R<x>f9@j-PVG9sCt>=w_S
zxsDLkDmq<dwGbSl#ha^tpJ(T|AgKb=C)ieLrDZ{9^vOeRyoX+xkDMVfi67HhUj#cb
z#Ci;zN$gUY$&mVKE9EYNU_KzT{i`ieDnh!5@Jn)dSv*3u(=8uNxjevFs&kE7ByV7I
zJdD?G<w!L8CfbrfP_?kOMj07($h-A&epj2d*@3Qje(7fabzG~QyLob-qUHSFXL$go
zvv^x+eaUC8ND$)O(C#m#hx&8?!0%f#QRxxl6HN%T(r&0#XT_xPXNK(KEPTjOs_j&X
zEt``pn{y{ewgahW8$h7K-70-2t|8vzQ^?bgO_G{H^q;BYxrvo{#;&#SByz<-TMXO>
zb#@Fs5kQ3(_2Nd2qjj$xg2hLBFYo%w!$2dyOLx>x#%TFa?QkCj7Tmg$39}A;6eb05
z<>K;Eth4*!gD9aF?k!#t@?Y>V?Pyc3m*c%*ES-#NuXlru8?jQ%m}do${1&lnZiqgT
zc+F~(u7{n7+}+hQ7CWc4xY_MX^;C)d-d502CGnJ31WmdKxTr-H&>b`r;DTa8O$HYd
zpK^A+prSOjTs}}y8QGT60SF}wNIM|?P0R@g3K5?QD6s9p(aXpk2h}s_aNyX|MNhI)
zW$+`6rQ&L59bMnUfI~b`XGIUiyk5AEmr6vso57aY)j~IEk1M(|V_rFah8P*#CLccZ
zl=KSzb`#h7$Z!$jkB$YC!(Jwap<uMM6>{PKp>E1m)VB<nY>X)Qk0b#H-@sDEadAm#
zZf-7_(@Im~Db1FpARj^t{l_W4qi<S`LvY;d*kuLKNcm6wBtJ!Mvhzr6x46)fI-e{l
z3)C_*bt2+ygopp3<Ap!&b(5MLHKk7h2dXK2as9D^kFmbCTF*yk8n`TXtv>4h0Nx>f
zM;U_l#+FZ+Z;3`}#RWMtcJCo-1-sH_)(3Orb^HV(B|3raSK_iY6xx?%spD%gCC6qA
z*!^jc7K^qwGp-*U$Abn=x^GL1sf`BaD|BcB@CgORzpMKx0vZup9WL$th$4yOcVOTB
ze++-Omn<v$1#Z{!wkLxiE>!P1h-7|dpZzf38Zb+3dj$M=Zm4+xX_XI<R?tQ2+@Ua2
zeR;N*%IN{{x)8o*Pik4Kuxo94+wKn}BGatBYO-dg@Ueo6&Y_$u_#KKQr8ee>$UGls
ziYA1=dU^^Q+rhzROE6O*1rllG(R#0NJ2<h!?r}8MfrNx?CT01e*252W!6>nj6lD0P
zYI1Tm5Mdy)K9rQt8TK{-nrLr1L$x}+(Oe#*#ImGU#)cV_!)?DVCps6I4#@3630tVb
zx+LQ_QJIXO%7Idv(T1r4B%pG4NaLQT?*2|IXt4Acl1+fnYbO9_*ljA9^u&wZsV<aA
zG(3AI5`=1#>fj(o_bVa0jzoCOM~C$0A2HRSi%}ci4?rp?e}U!m3TNB2@QLPLT8TOy
zNCTU&Bq+P%NAalvFn@I5R&ts9Nm}CCf_}i3NB9wNB>BmL)MsFJi;W`DctZ+hGL9ER
z!wKW$>0*!gUlqH)D649QmZlP2cyR+SLqQ9CSP~=V!e3cy`mqg+aUFo`jZI@^P5nnS
zz^Y1fz6@yQrF@vb1O4-_yn&wIsXWuavBxj4zmw%Qh4=7;@9_Q#z^J#nC+sI_9AtvU
zIG{GnPZe3k*JvD+q{yA29XtA>pg_6A$2X<EJ~^S0-_wv!(Q5?6LbM(KVE;I>`_~x#
z_kj}C_M-DVU5)rywcO|N%7zBjtE6YuKwe%*ZgW%a7li<h!2FH4y+9Ya91&!CZPVvU
z=jaTg<i-leG>n*et{38@_68xb?Ex+?3eY~v?z;|$B=U@x?hyWPgt=WDzxT!8kx|UU
z0Dn*b`L#NUg1nEP1O82_SGC;X#bme+9R?s#&M>lgo$&TDfW88CExDo_8WwUkSjGYZ
zj(C`4tGU@ZqlLZ=4TjH=avwe8B(=X)sZ0DNfeMC7HhUH>`FW4*GiP*>{T)>6aWF<1
zW1mijc-H>f0ZGpDhd^T%Aq;Ov2{429_URrr(UZ6v45J8@R>gn50A>PtAl_k2t1)u1
zQFt&5m1z_7-!_P=1I2LJs-0WT?*s!!W>W8GY$$$|4dWxK8A}48>y*Dy$0I;C&6v4(
zD5pb_yvJhQ_c&kzsu|w+#kR7I1XkF{(CMeGnVZkN5uk9TWseELLN|Uu6e4aTSA_vl
z{9$KpfkLb!QrZP0WlaB(8B4AP4vt!sA{H7agXiZm&|QP&FCHKJ+?$0~IH~?K0<h!Q
z!IfTyK7tLxuG5W~sMUKCj`1Il(Pp9~_Lx-ZyyS6y%fSKhM>2wlNf>#H*~qlvH}=iZ
zyYC8(<uJE~r<5Yaf|jEqb(1iH=n8jSd=kmJPHyv(rf>ir2lc@qJe03La3)MJaDiB`
zt0S}P-b5z%gAG%6YwK<3cjPafLROCQbh&AvgMI*Kpt8=lqy2Q;DHlM3E)uf)Uxl_P
zO4vHU`xbFV0}r&6S48j(fU>qMn%248mG)gz)KW7Dt@?cT8`J_oX1P@0q~`lv2g57+
z4@Jf#;?=_|l-rKpjLyi3YcG+p6}Cj}JWY@}lQ^BN&Dg9WP8^Vax@mBH0?N0kKDp0)
zA%ZqKE*Cc42RA7|NZfh|2UMg2^Sp~eV>~_p5A&aQ4!uDe#`4;>&5B@c8UZlf!P(sx
zSS}1hTf#p%#rc2>%Jxm@Geur+WnI-+guqen_?)}1ylB(#fohNLMp=t+(WBFHfmzua
zP2~CZ-o3ik9Up*-gkAqQ{8;Z0SVR|VoJyf$t2YP{b?OcSMTRGbsTR>svqI)v;iNvr
z4eB4x(o;SwMs>F9D&$|{p|y5(s}YilW193=HwJM3$cMzIZh=dQ&5wMGkG91^A;x*a
zA_{?9)7{%O7Z+=))w)%4`1Y`{pz2jYeZ>K4N(=>*mg^?1zCE+@ds$mT<S#%-(sobF
zw}8hU$4c%Wg5z|2R3a#ft75X%9;wy6Kon>Ukb&m#m=<F#q7zhY?QlX39o{NM%@rqx
z4Hv0H4l!Cm7I2aw0eB7KN`PWtLEQjX87M>DQ9}vZ{Y#u+PaD|CnGXMX%xa{@(xf^2
z;tDu7e!b|u4vcZJ^sm1>(B1sm`k|2qF-kmL%cSHHW*Z*yA1_GR-Q|4&>NInBVIWtc
zNTbR_WYZ-09v}Ly?gPjQsw;b5sob-p8^~<|O6>uI)63#6pwxaR{p+TGW_f{Y9z8@X
zPMNz|I^6<UvS%!8p#t0u#$~&cuZSbS#;DO&fTMf7arKY0ns)4Uy60I13kV1@E<XY2
z_F31L6hie=O&SX_I8Ho_GUG=XnK2*;7NR~#JI^a~<CyQyK(?pCofCXAH-~^c^V8d8
zM>oF#gekkRVaaWVqVv@Oz5f1b|E4tnqwa33zbE3d#LhK2CM^g62RuMR#6kk6s&9xo
zrqsd`sZvr93Kg77!@~!UpW2*cyw{Dnkw~{$V1grB%$6(<Pk}F+;aO7RxS}75mPAW0
zhsj2Y2({#anH<9;36G<%EWGnFd4w268kj$UE^{=)bM8SqN+%96>$?}!hJUpe340zf
zA)6Rx&~-@1(=N;VYICVw(_>33%*U5;_h+gz^|WH=Xt-SEC^c#+kL{X~K6XpYn31X`
zLkUIlxAwIzZWJ^{as+-rY7KM;Duy<O8!F3}U>2d*CnVbOl$jbJrRk3<tEB3tmbFf_
z<yNp<vuC#^!_400!Pnv`YRqYqo_#J|;b3gA)7hW*QOgl+ZQ6q-hTp{l!Gi!m7a$*>
zXiXv1=-s+f`9qFb>Rq3tk7DD*_5V+#AX5$)RmAbpBQT}$Mjt}oI5Yx|2ULyBA1(sA
zJ3BjFX@`M>8GfFRslUu7$S6q^CJkDx<z7$z*^eiUOkMYN1n|M9>ouE{?Wf(9;S_j2
z3M%ezzEpevb1^nE$qnX6*Y+9Q{?a%R3`Ajvxo(c*FJ~jtEXyZZUe@PSdnRGX-whw6
z*O7>`icGJ~_@*m&C2aXEoDC*=R_r&t5WSk&#I4YAa@22WRWT!mGE)66CxLPHI&~U#
z|ND{MsbMSlRoky;w8Y$eJ;?vk^gZj)HgZ#X$JI&#ff<zd<5t$Mtz?E@)vj!&d*dQ@
z*45i&(j!bL8j&-=k&`#2ZhQP-S7&#$B^uW83y4l}v&;lCAK6nexV^Hufoc`UQVfh?
zK#s?x%eYo{VNcK_!2^sM5!@u(atR#FscgXeP*AWW3Rv6NVfsGgImQB1<2fswU|}<m
zYS38nhS6$6v1oc{b)2$|34Odgg}3D(l@Ods&+g<#2m<u&w759ZWmQ#je^)2X**RNU
zJr{x2<c^8ga=tv7l)f5+`KOHIzC4AZj_j2DLI%Id=LVZ)vhv;D7k7T^*~VET!v|`G
zBZz_{PLOfF&acdv_W<PhRZPsS?_}6hep%l;47sR#`}#&p^(KNNFcdj#E3yQ8&uFGF
z+S8Eu7-o}Svgz`}a&&$2@d#>uCfdUH6YGAzdbRhr;pJTHngrvEO+7u}<Hh#YZnIP)
zS)Jkk#qSME`c$GMH#9h{09AnD^-*Q3*DP-cR!GJP->1Fmu9@EOt{JqV9FzYS$9I$l
zobfVN%+pP{Z;5<)KqBAfTi%8L7D|bFHJIByD6Muvuqcb0`l3akXH!W_jU^yG=Pw}3
z?p2Z#;K|(+eZxUx-XipP7=`9;N`I#Rd*N7aBb~QYTn`7HTWU@KG;9WEb>C3GKaMn=
z>bZ)i{FjRA-qbJJrj{zWhOWt9T0-eO!pXU0uhI-dkA-ah;qT#6%^isvS>|Cx(BYJD
zcm^2#-j4W)@x|`0*y`MIWd7w@<#Aw1T$_c=7x7zNW{X@6y2qJ<1wkO69!_1paG#oN
znZ$xTzaL`K7{||Qi_>pU8|Uu<F5Bu^<!>DY61iAT-eC%xcg_(qPzXkt{*u5o;d#LR
z{@*b!tu8)nNpKHn1Fwk0?CP?rz1rCQ5lBTo-W<RL%6>prc#p@3an}0<Uy0kJT>xn4
zz@}mY(dn)Oqa`ATVQ<}rA#KX`Uv-SwD3nB=(KO3;YoNm6T;`|$h$_?nA6E`1?~}~h
z&R<B->>?jX;a)*mKM=hfDx=>{Ay!*ZP$0S*s_7prSLK+h128du&u41Rd+{oi91%@u
zaG5SQYA0+4`RPi}`7Vbrfy-RTGW>YVfachhaeiq}TqGC^j)@-QB4MUPVT-<}Ic#>~
z_eE5ue*>AI>z24>qZG+{n4lfz{H$OYfWxA>bOy)y511JAu5tdaiaGjc?rR?3`+9bA
zLh4@=$J78lmbz&II2#TAtN|r_@kss@3RI-dC8=r{`~S)lKuDGK27nqu(LyDo>KhVF
zO(1#4t~*arW4y;-gH$qOtIW*-FFx*!|L_0P>i_D0&)fvnxbjoV)?F%*u^WjtodUEP
z$#<g;c&ZH-Ds`I9I^3rwn^0E;9=%mgh;uFT?8kZEtLZuxEIQfQRqqGCJ!iCZ;|}%v
z$zd}=fD$bcoWV&zYYS2sepH$7U*rwG&hNKy5sQ1PUGhcBlGrem75Y!zV!()^c=PaB
z*S>vw7puj45qccZkybTUp+q<kb>iTvh>B&u<uI<7v62bLYsXw!dFnltrs8(>?>|=N
zZ8{-DcKFuBF_G=q&mtTnDn_quTTQrN3XrGmCT-RWCt8|Aa&pHm+lbvu?Yh)HVhh#4
z-uC>&T4>Ml_om|l2EWG-aG%EZ17GEgMbc@S-^7=_)83?(MLba4PS4WmZw$xL&^AZ-
z*LQ7Jt9*7LBcvNJ$nK0&+~FV8*@-A8+s^&$N0;@xQU$R(I{znq@2V;SnxFgv!d!HZ
z=;*D{G@JGhE7Grz6*fkzOO7rt1^B^N)Vq5-Tj<_*?Tar!PDoWJ#v3jQ0l2C<Kmi40
z=3-VE{Ms<%^=Yjj5YR0b`1@;3VZLn`bH5gN%HiLE=O#%P6G<%AN%_q~lbax7V@GvM
z#(V8aP1Fa*@vhjDnStK7s))AyE^nTuCXBxt<98tL2xl&-JKwWNeIh)jxY;W|0~4_Z
z8Dh?mEsS@`1B2K3nO?}PY1QKSJJR(W@Vs2EuyU*#r~ynldYVooQ9<)x*Zu>beH8zD
zD3;(Z=uf)4{5mIWV*}M*^6Yuoya5t2j*jQ;n(h)2POV3$t^U_BO``xTf_BeJPe%Yq
z1K<7ziXb3(sK{N$>%!rctsqBY*?5^SOPwfX063#4IC4{&7xzW?@;5{$IUwaNPIrlK
zN`94%ggf6XKdhLy<6BG!*IkB)Pq%PO5&;62<-GKu;NZu90oZnicF>~BV*I(uam9bH
zTxB)U)}DryT$~@>36ccS%miwM5kHyQY_$aIBZ-L1(gEXZ{q8)Y9aE?`SCmty2LD%6
z^3gpe#8~g!41Ma_=i+q_A`<(9nxrJjR=T-(Ee?dL<qTGa^o7j3J|K3fc@+KuN;-(?
zRgc{)>Vh8miV!`{6qAUrglumPnPMmd95CKSAIPwB*(^v3+9B`+N|~!hIbfFrLGp1!
zfZqnBhqX}rR!E{BQN!gN)laZSK;vO;YC(;Ic?eYopZs0@>c3{L33_G#j{`o*r4NPn
zRyq_{A;~bTv^EUNKM=?44qptJt#ku>-%L+ECV1rL=2J~2W%&#p9<6*^%+DA^+R{Hl
zw9A@}n?t|FftX7c{Fx9EmZE=CYpdKwW6nLu!vB4D1i0Mta%oS~IK+AQ=_#93<8GqX
zn#sZI2IgSmMsSo4b`XCVm7ImI1lxRcZp=p#Njjnu<~$~?k9J$>*cY?da)?P&q>ZbQ
zfThGBiXd;hXO3yBEiZ~kQ#)-Wwc6u(aQ&51Fi*XJRP^mFZ=!a5Cz|!W1BmYfvR=@A
zdYM{ZyrG_^%u;0&YS8QdWqM-%%;K23XZ9)hSrQr9P5%IV{Q-EGk`e6hq`(tk9~>7B
zQ9T1>y&aA3)jVH1vTMK3mupZSjhM8yw2H{&Njyi=WcCx&<D=4=Mc@SQ>_yq!5d-4Z
zH`~*8o~Nw08#YGLdqQg=UN@bg%OHH8p5_wZ&JoKCNdi(`m;hsCHRb;U;j^RF1`N*B
zg@ElD>VW&~H23kxT8qllC$Gr*(EMWR?fnf3<O1$+rTmZMySi?>xaCc|c&ubG%M>-I
zF>$tU65v$dxLFY9z4y&HtYGNF>1)>N7<Jqu4H88Fv5j^8WN2uvibA=n7CG)LnMG7l
z!?mkRF{(LzMa4XsVHs%gW!Vi}Mi-2)h#pT2`$d>yM|Q|l+Nq}Xb>X)0Qp=R9HEq>6
z?wDr1vC4C7bbRzIHDT07@bcLsBJD7NlI(G8&1NcFAI>K!JhCGrKEAKY?lTHD*6Qz(
znK_=9Tr&`^47LSdKKIxAX!BL;h4z4tkj6DDFkS4R)Tn|)e3;}+9a%jemTTe@=8)To
ze6IoK_pi`X#mWqk9HOsM4aULZ&@s5ovqy<O(DF8#IByeS{lh+pirxS}lw(@m_6z2Q
zX_QzQJc|f^NKu_nyYwDs!9o~&PY1W&UCtesC%;CrdHT-{dga$=A>p&^BrMFQ&RSZj
zXY%<Nf5yfh{*}u<lDqD3P4an#oGdpOg(tIU^gs0-IJks}C>6=|7B;T^<K+_IkhxN+
zahv-X#cdpPXbKElC!ne6Ym?}^5@;z@d6$i@G2ys)mqXb5iE=uFE=&I3@z6eDjRO^k
zPzGC;?}y!~P+12!p?4EwD~d-vwvQx`-A$R2mt-8$z^4dXqf4B@N#uCLlpGruXp?o#
z<l#J0l~=m@67A$7tN!aSHp}Q?lF+a<|HYV=wYZTH`M=@KAxK^}>d?AH#TJ>B-xECk
zB+CP~qVVfeh=}*yHo3eRgr<&>gcVw4PUV^Rg&N~GB4zDr`c6ayG5?Y8@lItj8-nFR
zd4Im-o_mOJ(nN}X(tPDX4XUx@ATY*N1kbBw0{HieP!WT2ih9;OMSg;wiK&0HS6p6i
zeIums0#}ocC@28RjO%Af;EWT!c1-=s5T+TT0g~5G>Xax_KGVAM`J07Pb79S@xj&mB
zZ8>%xhg}xZqY?wYElbH55Rd!b4qUKyK3xfe?qj67)81|O+r;M4uCEJq6;Nad(O}mQ
z9F?)nk<lyMiXbZAJ`$4|J$W2Cxj$|90tSPgdpgdp2@YqTw}$sMCQXN-&4$hTo8YSF
zZLui4b*VDGw_%$<ZK*MVk*70e4zy;nbbHBp9(0C-<i5w#;14SeQaQs5Zp`w4w>=Q2
z;;vSgs=GoJW>`}zx8=ka;{U$Y!s1u0VbVGs{o9;n0|K3yQ#IuTFvtsjBRjWLF6M;S
zCl2aq-6eIBqrznp9W=uf8ynXyOEuE`5|r!mB#IH}-~M%QM=}m*&ve@!Pg@tOty+I&
zJx+__8vq9h83RvELmm$Tl9Dn@$XC_v-TyvtO<%c0e(@on0Ao*q<&p`e1=`!kpA%|f
zP7>Eg;xdy}b9EZMv5sH(?_0PL>mBigYUWf!2F!CR&+By=B$$NV_%Vc&!Zd%XWb4|0
z35GK>DKz}{VoXW$?cJ1~7syn9hM)^AkmAhSi=rE2u3`?8H24PxF(+|nEf>K?^O4U+
z6x2ApEtStJt6<K?R+2peZG7-hdR#Hz)5fDv7ggs&y_#(9i<sC2f9ubJMwJzlB!ipd
zKX+qPok;Xs2Se3H+^<iUeX-e3BIH)9p+LBXTV&273syMyD+Vyh`3!r6i=j4xxbjs0
zpt~7ZZt3J`x}Q&XNKx{Nbg{`|veXw5|1N^-q~6oOa`W=`8TaWew=5&rFUuyT<0QJY
zv%-6#ikhxo{LLu7!h|R@=-V|<{tfl(Xx4Uk$DbC@b?~_7g-)CH?fmj&V(YkEncsVN
zz5lBwu!umi{3zQ5?D5(2?JEI2{qLtRecPEJ1Q0J`nvHrO<!@L1U;Z$etplGzOJjv^
zk$sf6HTzCcNoLIC(Q;3BC_jH?`F-t-2GYy7{W>LJ8bddZ7mVQbpkyb9TgsG)1EKEk
z@7D$W_p&+nJrFZ(uB87aw(Q6OhQ1ycjXK%-GA{TTM1qm2%N=m7b|#p|Q0DA*_VFQ6
zQMU{mS9Rzum3bBQ70Y|KwA6?hKltf2ZN-)d4Uf&;%^YQlRKLr1;eCjbqr2%fD3=@k
zhB*>=CKwQ0v>GC_@X~+eyg~e-|84JJ2JKa{LUK#7p~_*R_3JXR@K51|<p#CXyNuwe
zjA8a{dE4A%ooxGxd%z@oc0FgsbPe45k&!-BoQ8dfF9N_U*DGvR2CRcrylKl0sYK=i
z!AQ*pk9$DmU66zg^`yB_ciq1J4H($vix5J8eT@pbMb~IGG0t};#QeH7{>z}3e-AeF
z<^n<VHrgf6d0&1DC!CaxSO0~~LTju<v+vIjuVq|m#O&W_A(T$7I^~b^Q>%i>%W_d}
zii%1xhKmy!ytoepH~0}J5lsfGOmvoxT>9pq^OIV9eYG)WI$Ny<$3y%0tNxWp|N1dQ
zK+|-l<^k!A5dTE!Nm7|*`&C-pfps<bg&as3*4epz8YmPiciBQB(}Cv+X<mG}PNL$(
zz#ZFd`Wta}*@gBQoYG2kIqv{vl*FXx#X6yr7t|jlyz@CX#H-j+^Io4E$JF+6aXae0
zJNIu~`@b$f_Q^cXzfe;bCtvSsEVj9_t$93JP1vCe)G6;!e9q!?nxzXT<wKKA!UjgL
z@-w%bgB$KmfW=3EZ{s?R%>}v1Y@WyF=8EH}ct*}2$WVx?R84Z6oM42$jh2>{hm{7r
z8xV*cH%Sj~MHz&^4KXVvWoq}+Ll{A>B(&(QJT)4VzYNj-%J48UYXo|Mj;RC=k&yKd
z;nrq-Pei-}V{4)M?{~j)wibc^e#H{SN>!?>#hrL@<q+=wv;bIxPonBmw~khhLvq7l
z4a<r<W&u`QsfH>sO_Bg%9GiE?tXF7EqP!(`NG>OLcu#X3a5I-?5U<*~l?^KOtDMvg
zp+-WsyKHalSgcsk{8@Y?&I>N97wg}OqW#+Z;(epcRxwmIRNs9lpEO6A?IB9j$*-#R
zrax5E%py4f0n7!Z*U#5yuzc<|{WK(idbpPTHRv8V6!{8evx4d^=YdA$3FZyb87I_G
zAL4-k&*j1QOM<Y%Lftn}6+M`9)a^0Gj;T}l&{DLVw)KG&JG3IU!lj8%(&jf-zS;J%
z5fTKNLD;xBDE8yzp^Gto;WiyP`X>d?I94LL(#}hwacQ=L{f(<XrFZUMmd^Ns>^K?F
z$a#n+4=cHHZdexU<aaDmW7GUi@C6wecM&gdtMo2m%kdb33Q_vuSagHMdg(!Uu%J13
z^B}5$fLW294}WvU(PGj{DeAjs<d$fpz<8m+XHI?rcA>)8)T!ts4GKd7E_D8LB|zV{
z*leFPBm@MzkQ<z?G=o9`PuW@Ls~`}8Za~84rL8`x<I)XxnzZ+Wv#rBAV?iR5+iT^F
z_wDp|M)?>swBb{oK6S<vqH_u-Pp!4Sw8XnO=x<<%TwDAFTC^!wggZBa8b@I`{W6$w
z)#E7M<_Uoztl@Vg^DKm9y=F)(DmNcYieFP;Y5wD1U29D#7O9(f(WNI9p2uKxqOrL6
zl$d8mS#fi_1|oIJ1qM%0rP~P^Wg@@F=2zWywR&Yk&2IMN^2J?DZ9}*z-J~#hvf$Ju
z+p7(lHm_@Jf)rLo5scqn)3FUaTyG~J_J{M8`d3$kt_Y|Blp9}tcv|-LiwJlSrTO+i
zL^9VIZ~1zi0;1?y+btQ8YXigIYzSxF`Rv61Z^)+48FrTr3(q0QXJ_OLnEEAx5hE;e
z^mb3Ivk{1b{_M4RS&K8)LU&1tJ+8}s`?ghWoJ`QeI51hmmn-rqjYbcQTQEAs5noi)
zY63T`hVxB>S~ifEncaZ*?jAV)R;J>OiSf-M-rsv%65Oyum(u1_=OZdWV6i}7ea$jJ
z5D5wpUdJPb;+D@}G>YGU!DUnf|L&5#Xi1r~3oGFBYQ&Jjc>Z{6PICfZvsVO(-}%J*
z1!aQ<Y9S^ssPemtzo&3=Hyo14G;$;?C3vYj#+<s7>X?+YX>aXOUPXuLCsxYsRJCt5
z{?npiqya&pH(|Kf$Jz)JPoYLC(;Q+Wotc}N;LQ3zi;2O&+R_gUX>+%K_!f<5)Ohr3
z`6NT6L^xV!xyjBbCMM?e7SS78&G|wNFGMLPQYvl!6ceIfCtxF?+c=;iCVX8#mtLO6
zI6trVNKLc1ldX$TxMLi$b0}qo<4P(8tzf|)Abgc)zjynck4iEfE;&|6BVJ<u=j<zR
zqXbs`C3uI_b&cwgHgqKj)Ciz&wPjJjNS97W>Cs;M1uwCMI_Gy!@WLN`)NobCA`ao%
zqysL-gpAh;i*t8J5-HcNVl&&n-a@|Bos}~D4gIhSdN|383f-sDoPNf8dqD(HN+m#|
zD!)hW4XG}68tBZifECV~A#H84rS#obwqp7?%n+a-n;Z{mm~Zge3h68+^A`C`9@|+X
zk^XOtey6H}-O0qnB=fAClVN&Banpg9)$8fbI>NFTQ@;Hj5><TgeiFVwIGPzkNA<;H
z=1|UW4UI(=t)r6q-=~q(533#Z&+}`U!aN~lS<eJN8fLa-p3TB|#HN=DH~%Csxbi3K
z?iHe<2PpUdB~~?%_>2&X!s&8KL`ql63*Ck7P$o9f_s*lR_;3bl4cazV+p5)~xbl0y
zzj0IYJBUqPG!IBZ466OD;=7vH(#`NbZj9QZMSsG1)oc|b5$}=DhEN0bWZ`RwMvR2Q
zTvv7J+N2V%+mZ{HPfDSH_vr~I(Bg<NY%cZ(hH0ErUIZjkr*~nPK84VBC<t7crU*?@
z$3m-R*Z4(*HQShqMUkO~&4uQPM<S0j34_$?z+3ZvGVcR-m+l-?L{PdRBoAw@K#g3v
zz9I3owdk7FjKj<i7Bno{_f|1@ibCww9=21VeXLXeUaoLha9blf0tUtS#`U`ZKg;N)
zC$`lty%%yU^#^PVm~ceU9~!;+N{z6iJ-w*gHFr7nam*s3LxEh3kS{NwSaoVcf}jvG
z!+TSc+5=N2<sGF8T1Of8cJI`&Pbh4^V*Jpb!?fl#gL=YynBhp*i!=4(a^&Hd2`{kA
zPb-K2G4(Q<@}PvlTG)U++7^qP_a_|%%ms7hm*()JKPk_vhZ-F8_Q%5Jj=F4I!lRa=
zPLg26hy-X*gqjQ#(n28gK}ZRFXD#m~fh?+gR%z!U)ml+wd9`R<o#GfVY3VNeg~fr}
zASS2vM1x$S^(P^G9T!T|@$^4cGKo{*ZdlUP+qa~8m&=nzzLPjZ`CtFHO4nSPJ?%Vv
z$oeUj-^<)f1c1lpPey;&=Jn69ocn^eaOdwvRh}vWt_aA=cW<JDdFAtS!X|?$Dyb#~
zD7g5?Qc)khAzw7AF7*;F(rC=rk;3)Bqk)ORx)tqenUMth%<l0}(!1>rq0P#%PK=wM
z)dT|hU2v;SD%sa+9S@*A62l%4l|bJ5`i2QSp0VTac{_hRpiZTkK)rdV|A4k`p0>5M
z<%J&t?r%K7qK6N}`u~`E3#cgHXnPnzBt%*|WN7K`9!lwwZjna1OA(kM2awJIDW#Dv
z>F)0C?&f>>-Fw&nTZ^@TSp#d{r_SE{>~kL2^UR)8mNDb9ygA|bGW(6&f>M|3JKJXm
z%ntFn+`qi6219V}{-0tT*n5KwA4m>zdWo+=X?J>YZ*;t8dAPVKK~WSrm#(l1Ce8hM
zjsfTC7t7gFir3_u-C0s^LZm%^muPQjK9<IrOzWsKMHW<aZN9Fc!T#0pB7(f7xmf2#
z%0zI?Z7j&EP%OLVIU-8PGB|%quC5|-o>~xgkHB<M)Z?G-3fK|y$+nXTC&J;%cCfz#
zKvw9TeFYKz2QY9NV2t@cQz0AO{Qm*fUBgJ%73FlRN3vqrFh&~TbJ`{xG0t*&vftJG
zDVnHYr4V&yD^i)IZ6E)Q>%MGrI+c1k1nbEZSbX&6m9X<VRs=nopa>3h+pD`f3vFku
z-CZMHr#st>0X-ZnI)QFEX#Z)-L1D1BwRl!{WNM|L&GGEk!Pws}HZ(n1?8K&Qit3q9
z(yi1owXtU{liEqSIhTF2y@RZEo!oU%ZpqmKk4f|$kKtLJXl%W}vu`=SC&q#Jpb`3C
zx~}Lpy7H$|1ct^}k~>sfbaM_G)#o5+OsYjFo=x45GXWouMGDf79xUJ#3b#gF0iLTJ
z=oNXt`VmX2tUc@9Uj)9A=C>%j$w3EX(#il#f8M$Z+&FzByyXYaKial--amgY5l2Kt
zBW{ADK$zX?S>qdjG>208YdJIlR7&_DMpcXkJ1N`merEW`_sWUbgBV1WqD{lAknsD0
z3Mpp!^LKOHxQTUP;bhk3e4f2o8)bE7*W2qM*EKFoZWH6E^;O!zo%UZkSgEjQ^bg6T
zv&Ur4p3eX0!}Mo%?0j9WjuM4>o)zPLoV5U^{S_M$XN;Vx=I*E5l|B6u8(k0O<LA|2
zm_@J3pM{$rDX!&!D|pmB!iG&rc_h$f^q3{i>#`t?E{-pc-yJ_hb_2*jt$at&&<iq%
zB;TKUWT{x^vfb04HD+Uoun;*(Ugx->koLE-7@Nm8Ga*>5rWJK8XSM;GU#}#}N*|D)
z7tzyk1J9+3hAVG?4T8dp=-2PiTG(6_^%#Y+^LM4s4Y32%>vG9S_RTthyVgV=`Jd6}
zES)o&BMm@S`<H1qrQx(66D(dNHUr7tuotOV0r&2s16>S-x(ZdrfLU@D(Rcs=^>vRn
z9Kh%=Ng;MOV;14?>cxEz{<C~b?I_lmWl=rXQJEB9xu8jr@zqfvEnQF!TToLkZ2u2)
z)V7rCs29MDER$bpYQkdEXq((gwtC%q@+rBv<KseqM*jM&ObwS?_Ow5=-X_1yG>r<s
zx3YD>|AvdkUx^Xl*Im8P)ciuQ`d6EQlBc@BdtK|sL8_G6xh`K~f06r)M11J`IqQz2
zfX??!iw`xypepOiS?ocJSDXh22h)Ix(kwFWDrDqMS9pDv{#H&UrhX#lZm^z^5slGD
zB0J!U*FnkWs`096rlaBSixzo=de1`=10V)cv@9VgaPjn;qWbKJMi?L(V+J@$Yszlx
z{vo;IL?Z6>Hs2IkYYJ`hUk=PL1*DENFJ*h|Gl#?|azI_%J-5=Xx3J5rect4FDcW9R
zhk4bi8W#$;q_5uf)kwsSBjs~6^}UKVnU!7gm!7KpexGt604<DP-&n_BVq}C|6tnjS
z1pnzN_Bap%<-O5~x2wBk6X~@Y!@Jk{Cr#po;B$S9quf4peX`a_$Iil1#P7DB_{DoX
zUsX;?3D?beJ#Rs?{{AAcY46QQA0>^7ss=1v7@Y<`DJuKd4d3znua6vmjSEXteoyT>
z0vw!w#sl|Lqg#|L_J{Ji;oWdQ6NZ^x$uMQ1PX?#X3Y3CKtO@GV1ByKv7-1c|J#FKZ
zAS?D@EXv5<<YYC2F+_R@NoLL$b8hbODfc25ndr?;0^!T)Yy$$1rw1{a4<D{52DucJ
z^Q{Zo_jz7>3kUzN7$L&?h5d>2EhE4dU$#WGC02jiDjo@h7vl4rvj*UgUHs++LAC>X
zm4WE0z^kP?@f)GRe!pIAs*(@o<oL%MF|%#bZc`;0T%1q+N#7NL%+TcXojJDiHD&5T
z#Z8+eLma26Q&UBx3{tH?WX?prdK9sshJ+8_|Lq*$Rz(i5Z!j9+IS3jm`!sq6Lg1F_
z%{s%^6p@|c<~2BY=3gB4p%8vqEGJW7kMYTU<g!vF=$`z;c1)OtXVuNgn)p{Ap1k~-
zSauGaj~L}ZnbScoO%*{YRy>7_N~{S64Fd>8nr1lnX0HJRTNm{l=uuL7d<CFxp+!dA
zxv7W-0xWxftv_PUGwOyoQZ(3#gM;nG>LU-hTxehj0pYCvp_}?s?%Nup^NYWmdu)zz
z5T)nKjB|e7G$HsU=YW5fkd3jK|DKUt-oXU$YmHSKlFtep)$Fit7@O!DaCI?44GfF|
z@KqFOp&EaOpwv4rJHQ$5|FYB-Xn`FDFZ82tecs;BH*qxr{fWQQBM?9m^Vc4q9v$>u
zR(>$BvUY~zP#@o)4S6SPe>!Tq$|GDBovoX2R?sPMe;j|v@PpO@M&0R?H6l&zP{6od
zs{S&@fmnmSG44Ngm9Wf~&mAS;q;pPy60t^Ws?y6JINq!Ng$jDTZDC#$7JQm%a<~jw
zHbjsvr!Fcg`tB=CME9OuHsX3VKFN9ZdQNN@HBPyki-MPwE|M~ew`c)7+?GljF^M|#
zoo&JVUL}nLl;UC7%V>sm_=4sISbK;k-V5!W6QwlfNX-8FHeYGBPY8DC`c+$u`#UMe
zlBcx_f0y#OyT#h8XBe^_`*<GYJ#<4tu~7Z0;6yx@oWbh7wC2a3Q$*L1IBn75?J!9N
zZqtemX&S^(2+1M)Je%mY>EZ|8ohZh_TjpmmC1iUDIiRsF*MRavc0B&uTA&YZ6IzQW
z0xJbsTYEbR_fwL9bM7@Z&0;t+^ZE3^KY9n8ljG4}M7J4F9{Zb?k$+DVt6zETx1#})
z3vko@5@$S{UT19-5J2G3+I}DSm$YLX&hEweW^v7)royBBi1rk&v|lWft*IvgDTuai
z_J{Cr5-K4MUBL(Nb<cKeAS<@Kd8Kk@K1DJStRodalmT}e2tA0KL@`}Hk2hP_tJ54q
zf-GycTJG&?siTb61$;0ZQ=RvM(jh&*<}CEtUE8y<9jQPulRajY=AlI6q39M&gw6A(
zgeq9srw%@JJBA0a$K-Iohb37obdGR1wBM}dIZ!_!3>eHwCH=RBJ9Rq43Q}5a6ENBa
z$)?-dFz_vvz^enlOCIN_&#yOGZjZlX{b<ye{`^CGFkNKM=IImb<A=PFW0p4e6hMWB
zV44cVZ2PQP*cBC9@nofD+UWj;7urOqJe_$rLy+(c#tJb~UZOW2mA;8V8N|RFxR!}H
zmu?4_cK<RymsWy*Jt0I}pA+*8AFz<pvq{MhoZJ!emPN5m(h8TSpuPPxtMcvJI3iBt
zSPD#!Q!?BM_eOU>b*{%WayB|4h7A;IZkQgz=F{!G?^?1nR-g}6faCWq3ET-?AI2Mb
zGH=_H^Wgc5T}qUk)5iEbrFxegFN8!Bpm4jT4l0Aqz-D+1A$ZT&<s6s$>YjPrlL;VI
zigK_mjYPW`lZ{3Ty!`y#1O#Se@sC0QrfnogRt95{viX?rznF)HWc;?7&4cA9KrrRb
zb((iv>_J92q0D>6sA84M_9;7s7|-}jF!VD_ZnIqgchgfg;<FoJ7l4)8F?`dUyZM&h
zpuNpPrQ#%!-9!g{tg*y~%9X+|Rg|c7ucpL2G5(5hFyd&1Y5w^Qd4K44r8li(@cmg}
z_0RAah`q|IMKXy{cF%Mg_!i<Y_i51c6_-(+QYCHAW@~%D?f##Fi*?3c)j-Zz8wT>R
z-BM#Gt$6HER4WO>*N7wQsbmU08R!$61fii~eszfc<bu?enDoHM4c?KqJoG9h?KlJE
zk^(*_sUM^`+@0FN%oxE?_^1IU1<Kb9*cSAicN9k3>LGAc`6*9TZo&OMk}PwXaO7kt
zkt?EJOUnVM-sCnzn*fp=LHakGeJHle8AJ}iujl8^uZMlyv?SBL{QQ`1+)XAE7aKas
zh}&)S)bAN-?*Cr5uk5Ui#Bo0|Ic?P=iBNX)cBfCDziG;1jgS;*4<CIkg$NT+9!Xs>
z`uDpF=#VA=tx3{iSM+JFa3XZ`J>3$ua;CYS2zOY!Z|LEtKEU93wRyfsvpnJs&LmTM
zMuuZy!}FR4K(YW3N}AZgO|PYZYu*Y^5Q^DTmA$CTH<b~ZZ2pXWtaxJgGy0^;jTL~8
z@Ec##4v0M)s1x#+p3VEKiKnO&9bmMqP;mns2<J0^T%?behU3YcD%auMX0Q*|k7~=%
z+$RImuNy58a{KL2hBYbM5saB3+P#EV$U*^;s^WL|M!Zno6o?tVJ+4_vH_KaB__xxT
zJ#92c1A@qT?1Xi*S~<#IB+e<R@_gA$WrSPHaZ@?wuu~(W*rarb4rKjXG$A|?X5s>v
z2H5i>@+SV^3%EW_46g_G<fo`c8SNcFvq#J8gD?u{NMchn$if<4ar5HKx8FcsVbUB{
zRau!_<dv0~lwy(Medi}5=l|t;;GwyoJCMM4n_0V8V_yl~Dt^GaF*iUkj)xa+?Il;y
z2N8^n9Scj*%o4SpPyulDCtYB^FA$`R4AeASrl%h%0sTb%T7tTDn^;<NY2q*T<H~lj
zjKP>6sU2a+bH*yM8@msLLmQcjv1}`?c{IgyP*$vzxrUHBM#^UcOwM-*Fj76CX-&QI
z;|6*SUvIyNUap-+0rUQ?PpstLKi{@^JGQ&npIrBQ$DEeetsEupM*;_5?5U<lU#ggS
z_e=Ma_D8X>{A8{Q^_8z~#*3t4_*?Ya`ur+WK|;JNE&uA&L95;GA2noT2lK^K)6yW6
zX-`O0_AqgmyIf?r|4RtC#ONW2$Q#NUGu2{*Nl#Y=h(AD6)Q#WW7#?!(k{d0kjJBtP
zHjA->1RaoppaRKzaX%h9?kS-<AyAk5^QATAOv4M0?izb@&1%6`UmXiwBU9Yamvy-h
z`l=Je2%Cs<c}oU~Y!lC<41)_5vO!Oyj2K-K@hGD@vBXrEym8b5z<at;Kf%H4LUQqa
zk8R*^3POc|XfNnH%Bx|pk1L|K2_yWzu(0kauLJD`4jflJ^sLJvvI3WylhD#c>VH38
z!Vkl-zUj-(imLic`t;vn@{nDg!`i8=ie$c_`6~DE|6ZMW+}FUnoSgI?ba2-J*E605
zzXYigGK19Ule0v%S$F|oa>g%lnp;El*BvAwFk<S);y{Q0Jt2@R4`)Xv8s6H&sKNjU
z_`TT5Mg4EPhjTFepQxtB#S{%e5=l}){%!=c5XkjR4O;j-0Z||sGyyCM`cZH8PK#I*
zp)ITmmiZ!Rq0E(@*z-`doeOc0$^+!O3|<%p^v!AbCFQX1Bda>`gpirSMXaVMFS_-D
z$Yy?eq%3Uv-tQ0*AUEJ0`|*D9PTNK|H|zJhy}Dl;PJ%)$ph0q<DV)n4e%lt|Oxjg0
z8VFU0Pr;h_WP4S}{Zsb*1TKnV^;EvN%MPiWW0OmbNV)VECzGfS#@+X%m&uja3Y%g}
z@VboKM>Cj{5Xw@bdWr<j`gRFP!;OBk8Zq87cZD!tqN}}z*F7k#*8++fEIyvXsa77c
zAdl0^{xJ0MljyhkoO#wcSdqoq1Qap6c27tqJ=yqdQXZj6!3zOcGbjdQa`V|K=)E<{
z4$1fJ*WJ}tk3-|yo@eJD4aT&jYw*q#_DEe7Z7$uTi>>tVHGh%9$q-zm1b>cl%B&3k
zw{q}k+shZo7Vhr{<<M_RM+kMB$FHwVITjbaRHiDIt{nP+Y&<B+VH|^gQ_K!o76(j&
z&FFXnqQsq9NOFGdN_dtSOO#ELJkQ4FRWpZZkh!-_$|!Zw6Ph*acICb#cIB;oeJGW_
zopmuKibnc}qECn<jC?~Vs5Is|bk>~RHaJb7@0ZL(p@<reta0po<fa#1(^%Aq1QHeD
zzYXiV#`tdSOP|Epr{4uK(9(LAVj*#%z6-RJbD}WXB{Bt$44~M_=jJoj;Img^TXL%M
z?ycqnLW|QH^0>IlV?K(DqIWH!Ap1<^qX}V`-^*p)F;9`V5Qf%+H`Fl@(8@NQi%3B8
zPV%KHR^KzPC(6x$+<fYHRlMT~c>UrJhkXu<pXhn;5L4XOQT}ujB8wKPYzqc>P`rp@
z|I+m0G<}upg2Zf@Dz1Z(9<QuyNNHSl)IxDsXx4~MHY_>jRY+Ul`nov`ixD@mD;4fm
zu5mT3j+y#3^Rn;6t%b4ov@x<WM~)a0t_8=M_&9go5*-QK!c=z6*mJ7w3QN>w*ChK0
zwBDOrr`2gU>V?eUwYxDp+H2L(e}i@gcBb7b`eus)5n|6F5jYfy^_dyp-@Xg-{OwZ)
z1JX31z<x5IBf>E;-?>btn)H!0g;4c|@6iT^@{d*e<%F3H+znwk>F4Jl)}V{0<Lqdm
zE>r9_!wXAZRj80L{LwD^#{48qIZL6yPmXziIU>v9u;^+kKL>2G{uZiSY6wkT?!ilO
ziWWu$3nIgR#GQwN2uC}~Zuw7oy+CmE?v}$zqRgY$ay~bH2zISDwsbn5nM^vE{xARF
zVLJmz1y1L$+q=5jcAVfve-%R|&Dmrrq3=L}3~D#gK3xqu4l>Xf+)0m~$2gu()MB(<
zW^ZLx#c8~{%PnuFn6$!DrGh+{m_9w>UD-W~0&Uzfuw}RPaYG53JU`_*g7kQ29xBS6
z*z=Z6ff~${Z0F)G%FCj+mmc1Q%Kuhe#3AbK7nvNOIT7f4(>>{DfTfFSviJe(QVh+U
zy7OEFTG$H?X6<Fmu_0*;xQ##Au{l+|`01+ZUSK66*6o-j?wh-G$eV3A+tiMaCc8Q<
z3oSGRlB)&`$nOhvan0S8=jNF1k00+Ep}!8ij~e+%Qf8TiFR$0MD?Q2>ydH8ROmAW-
zpF9FCVt`e3%=r!sf;e(LM4k|zGZcC;(H7<q4o4_Qp193Q0$UmD@&rgsql%zL1vL^t
z$XUq6bw2*pzDn560J%Q8TD}1e=FXsV#~Q!ZHzEW^=1Cm`w3g15o<Etc?0zo>F7Ql*
zRwE`Og;v)bbo$kwacp0@rIskP)>Jxv9l4)CG+}cutRk4v(P=e}?A7r;eoOX8EJ&si
zEn7`y_3*5D!4Bu2@MghDfk_lv1Y3l27*f8aqu%5QCl|-qZ^iLLfB&XQ^~=z6oOF(x
zhLce2gl+xuEbr=T@_tJH?+gi@ke49cQ8z0y#>>g^P&>iATvx^F+l8eD;1*>Ofmq9%
zxr63^Plly(;lvSXWPQfXCjY_5C;_}7Ju|466#fjlv4t3)a7X3hXh&I5?!`F8_>CUI
z(<slt$8nu<mz{E;xsyL(%Mk(%0uk1-4#JtY3L##Ht*1lSr!~tIX%K)G0^0&`A&E*f
zh}ZZnMlbf+?T<H4*f}u9-kyv5u3ZeFL_its82+u=F-a>WHqTZYdp(9zyw)gJz-7l>
z4WLu6-MkglZ*fjwmR3@W4E-h_fuiMq8L7RBvAla?^j1|W!~aphmX{Pr>_o>W(}6)c
zgI1Tur_lw8qPAX9{)7Vu1J6Tu0Z0PEs&av#fA<l5R9#HZDY!P!>4|Xp7HvN<2MQCL
zw;FiNyt7ti$^>UlOS|K`=U%>ph1Gum>h(ddaqgkI?P#qELTlQvkthQy1afS_#K*^}
z5VT->VZIq$<m@<0-D-D7Tkn9gR^T8MyAkq}SB2TgtB%A%3nvjh7O!u;q&4$)k|#eu
zQX2sNkl8O_;){O16bQdj?KM+6Dv0QRHg}*j{6%=N)li~0V85~zMKdSL9eTINbz|K4
zyZd!gR0!OowI>if!uDvpWG+fkuA42FKqn^PY3}>KejIIPts!?tNe#r#&yifaj_AJy
zpfsFYRPW_jTH62Mll^fOs#T$&cptWlMZXG;eEaj$q6X`1LQM2%5eF5H)Fi2#2HY!u
z3D69{GoFk$(>i5R>uWRKg#<-PVMIv1T~cOpchw0FXXmc^)`%`4<jH-IRuTKaAOGhC
zz<(XdS8o{|fU0wJLF+QW`8HEu$e^Hq#GGPG4m`o(E9LC`i4BaBTUHiWg3Gm?Z(l@U
zFsM~wQgISjOb?9b&bLN{TyReX-tG4Ut3UMaj(-Gp^P8$ka%^T_LkoesfDP7u1Hwj8
z_3oassE5Pr#p%+;v2P`Yo)hBx7UG9bM^BwU2=xO94Lu~tB=_PHx~;>?a&m!VGu+73
zLMAKUI!EdRK)2<n2|a*HSdr-b5RG~=KR+Q+G+Gyn1bGQx>r(azBJITkyC&}ChCJvs
zUAp7^G-FMOk5k@?e^YG=DEHmfQWt%^(F~^%Z=QbSVgBiiyDk`><))#dHNC$+p^$*?
z*7}a6Q3P5l=n7BjKg3|CfDJR_t#JZJHH{EAq~RQ48iou@xB1O9*lB4$p<h_N(Vto0
z_(vS@k?+BgrL9BjY7~Ll7ey9*B@|XK$MjI<7&3e7t_D88=>pr7zKt@wl<4&EA9(g1
zsjW=xdmE%#0@S1T6tR|$S?_2mjJS6wQvS^iK9aCftW>pD>;5qmNKTxFXJ;ctiCi)m
z@dTrNPvI4Qhq>s}_8voia4jng<Ylsmo?d^+WS)2VgfO3#zbh4&kWWeKJW7N67IB(z
zeOfZ(%q!NF4rcj*ZEC`yzs5ND8beDCWA^r>>-ua=ISeOr<Meaiv=gu~ECR4UmZTYw
z{VQh<F8NxLng71AS9qb6u?AB-UUdk~U<|#jKwWi$?L|#2+%qED2AzqG*S<Y{lIOs~
z;o`f=bDo=&#tUfQ!y=y}ulZdsF@&^8;u-tEkjYrAQka+fUDcJyQ)mYrYVpee9%)OR
zCfhZVm-Co|t=cVf;PK#y@sT9BYLeG69lw6t<Gqnn<w#`jCVH9l284`PCh1@F!GKDD
z7kTUEps)oj(wHL(>%xDmjPIG3Vvpxm|9b^AP;4hkFXVb0cMa@V#iZeU1W{O9{)=lU
zm^P{qvwjJJ4}8W!gn2eFRa%GllD$PP0^C?8b>HEIV-dBuEO~D-1;+MKhj_eQ4|`pD
z#9quJFXJNoeyiE2GgTSsM^c4^FvvC8<>L6q^C{5K7Vr8>z+<@q3@lQl1no1rP%mXN
zX-$9RWq**y<OG7@*^WDI0LI$p@&6HZkSZlWvPg~Y_{){Txba8Uh7eiD>%QyE_xR$!
z(MVKf073d)yUiHX<k6+g`{??jw}MS^zSxIU*6tN?oJ}3oW^b)#kuwFUfG@W4@I7wD
zyP}a708(WL13#ADOb_&&yt5HXT4Nx|tWxOjCjcdq9RZ0gINVNQ2oe_7_HZS;`9ew!
z-uo$txHS4NnKUMBA~l=_stMxxqNc(5u44`1FsxNTegCGGGi{h>;1@t5WVlBz;`-H8
zCybD$0c}Z>GHTi$p_*vVh|}*bu%>&fVbQ#@V84k1FBfa&|3(52LWZ!2EY5e0lAS5?
zc<d{vYU9Pksvn=YNH+I!*(}?Kvsq;7nAq@FzI9l!72B0f32`vQv;WeO2A6v#rFC$K
znq|^5Wzb=iq|tVY0UpA<>EjIOPE*L`wXZE4Wd%YHg|;XZy3+$(c-kBmsU06Tn93fv
zy;6TV)8Xqie5#F$m8)7ZBHuh3Fr#(GMXuP|v%ZItLOX7ns;tVD%4cIQh99Gv9^boZ
z?Q=SD>8Q-AfNZxJdbgpf3E7pk7@rk?fretoCIYiBy4Y=|nWvex_J1#i%;)H@({E+>
zEPh*fE5t$P%^XZeXskCwpQy$D=Bd_^<+0{US*{|-*!)(mL9dEfR>?PTZm8ic${uFx
z;b>{eDi?-4G2%pgK;r}o_E9-p{wM&CAeu02%;Uq9_5BI_4;OlXOHDNX7b)iubuvC)
zk++0z8WDaf{>--$`q2<4Jnyy1!tb$*tyGD>)^v6#)xe^3esxc^WuiP`sKAmqTU2HN
zkBy5fR%2qvkg5%N75!H~XH&tSb4Fi!E|rmBvsi(Uq44&HV^GI+s(dqj%@q%c=eKVi
zMzNa4pH#)NB-x@Ld->gp;e|bQxvn2}^YV9@5o$Xvh>`oRPF<V+RX)nIO?8}8$0eZ3
zyu?8RJN+j<`ZC8+f(TSJG)2u~-Yi<or#1p>q=eQKgAmvUn8vw|)Rzuxipgd&98`Mr
z079ch=!H&WF5A@QHXc?yQ6XTFq_~&~0_G8QgSZ>Ewo#KTAfZk@z7*4tnrcEzOABL&
zG)ZI--<;b~@}q#7Q*nGh-P$iMkO}D{Uy%Iv!C^DZkzWvuNdf(HGg3RXq8HmVw&AE%
z6O_0zC%W|5?Hky1#Hev-M@r}cPhl9{ad7?=qPaEuy~OyW>SwxW&o2jVY&(CSGaC@f
z*g*<`@fX*KrAb+Jd>n}f0Wl`hdTSdIfy{Wt=GSNRek|y^gv9?ZjT4L&^cW=NB_ksR
z^kA@ra<{SnZ^ouW{9IEBF84(Gxj#vU`0Up~XJ44fK$5~4EqE+dUUfq7L^O!>Z2qDP
zNWf#;dn!D^^9#qK(scRx<Y{oFnJWM)A++N*GJQ{JFh+6WjPc-qtZ8;u4Ja7gY8Q6}
zVfLN<t9#OQxpoI8N%!G_&$5n}T+P){S=>re>n}bchaGK<qusZ0a7on(LVV4KLU`fB
zBJjT$k~Z`W#t8P3Cu6HE<B9udn_|hcDU!8&_>F+OWU97awy~>z@Ap^#T{~6Z%9ns6
z?L9ndEs*w`KF-Mtae~qLoXkcLr<CT*kp>>Uy!`)$x5+ccQe<8bi-(NB*O6q1`5q0-
zeo8n(q#_Xb66C4RE%87V80x3&^1#4{bCCpmT!>=003L)LK%zzU%HI-_CbO*u<FJ-f
z{5Fe^WX2kPq9vJ%W#F0kK&get#kh*t3;?jIeWoO>nUMGr1a-M}&HN@etfBF9|D53E
z^b-_b=+0=pMe_?Vr~U;NI7?fw@~M|HGJs!!k+}^?tQDBxv5NvoT?H3n%M;>tf6a;#
z!q;pZFN4DTT?ZhAMUS35Z-=j$CRace&bRuApOb2oJn5jR@zWPLmUcKepA_SvrKe}(
zGl3PH>Ku#?G1G2!N?W_I4Ha_*`3L#3xa&Jinem~jQ9DyB(S)jvlT)EWt`2q;@}GjE
z&c%ya-f3o{n5+rdWP|-mZ{r=oX5QJ_d6pErSS+9C1|Eck`g|jsd2jZ@YD>CyBUiFx
zDOa@*I2<Wv{;6V8^|XhZr`kFtSQT?k^G7O?9^f<^gBPQR*X~fERXxQKy&O|JAHUJo
z!NV6|VB*y)@am^V%ys%rWU?>7^II5KnD2~|nKgmAe{g--JqR4(DKY@`;)i~NwO6wZ
zuI%Efc6GTCyFK(VZ2(LoMM&r5od9S-vqUTQ#*LcBPbaLm#CO{khjVqg$bL-uHr;eb
z#0Wp3g-w9?xCSj01;qK)=l5GxO6kf{ATNKieO8`{=x!l@Q^79k8khl+g_Noqt|T)E
z2sJaurF11f?sm-P-@!y$#JoMs#wq)`c7+RVbe*Hirs4VTFCysV&)2_NQ(nK}!r*E-
z85kOS>w=-vzfiuSwO?Y_YT&TIQk8N+P?8FM{4*9RK50Kn5D8AhvcIinNQjJeGgEhR
zmfOOC@wu>)SJXRIcaN;uFwzfYtt~b?ve@hUzmDXt47(K*ng|kFY`_~bcx1NMc7J^u
z<MdAbb&@(pkKba$;=<4z-vZ;MIyjO!157(ZVPwO<J9a?o&^4H>JLFC?YpLXJH0gU0
zb=@;5&96NdUeEi47zg#98z^z>{kJ}}cvOJ&+r+tcxwPg-R@n~mb!D+c+`+$%Z=03I
zlYM6$L}Y)2HXJ_^!^|bV!B@XZrA*tqyZUXtTST0`lMo{LY1`2^)e!E`cxG27!ofv~
zWpvHnbaCnHk!R~%qDTC!J*oD*RL|xO=K7)w?2BQb;N`RL2xCU=-hzy1QXsd(=i3nP
zh}2J42t;3SgK6UOw(o$m7CUQ_7O|%N`ue*q_0i@M+q2-;&dNcLn>4}`@e+0bEqtt|
z`ZJuu<Q(A#Q$C_OT?Ur_B86VAjjrtWuPdfevuktIh-=iAzn!;0mqTC=zdRonVWB$5
zr`k~5kmQF|SyqY&uVLi;-bpt@)XBT?likwfPVfeqS=OF}MVN}sM5uD3%|yu8b9>r<
zc1F&{N@49~*XxB%gIIdEYeY1?MG#>@y&zIvld>tVI{jP2D*4MZ@wEDkQPAK#-R)pC
z+vZ~mGUywaY$m}&uK_z$_zt+!U3fRvL7?f<kH-!h5yl+*60SsaXT_W@cHPjW7TKP$
zs-@;X`^|9e5j!;0j77;uw^2D9Rzt`0H{YsA!cItKa?&#u(COfBq1YmCg~^rVG^(}E
zc--{r<V7)7JkC-Xm*h5=^0t#N!{LQPt{e`(7;pqF3@`=^l0;>uwpnqm-WT4wBVZ9v
z-n+HUM%of)jPe|4ztok!32w5D^A`=E3Z5<`5on);Eu<D()nD`dm-;JHR0sp(8jQIh
zlnl$NYSuZ{zG*o?7^J=8MOc?kd$BA#GP24lFvg$B8)O!hI0;BiWX~=hD!UPLmoTxE
zI`OXSJ65q@EqG~Cea~>PgU#?1j0;;xbfAGXJ2S?}Icp1V4Ljmnfwz*KPltJZl?l8#
zV?r<QE<5H?pd_5KxT-gsa-b!SXaa{O7`WWPw7I~5#$Gk0jzm_Fjvgo9QiPm`hgAfx
zUTMNJ?Mq}~-RD)7>#;nxP>ymTgT&O7*EM8mgB)vYU8+XnW*>ad4=PHf6fU3i03Z#K
z`Co}Xr90Z@{e~PWZE{r48z=n|+|~xy^phK@=dtqxNH!LN{v=S5s@S)0`lQ+>W6xt`
z_Iz9uPh+6Mx&aQYi9qoy3SP*C5S@Xuaa|^@%v29<BA&u;E!$+?+~nJ3sK&W*DPi~J
zS=5e+I(Cn^a55O%yfXX*<j-Drf9toEw~#r^CKJR*TRTcj2eItl{|(#~16&(`75^if
zHYlE_XyR*aYU#`>V*1scju>7SrlM;Y!d03xRxr($33!$`-A{!WUtS6xD4Y;)E+RW3
zn-JN6??CSV`pZTtS_eTC*XUqChv`4A1mOtiq*IMPtD~?l+wIpGBjFo}dF{)MC)h#k
z!LPLs$@5Pyv6RU-?bWScxM}R;diuM?9bs?jF>5J5X&E%3mr#0=3&_t9#dDHTj}9Hj
zjZ7~F{NjSDOzMBDs}j`b*|c_KojCA5FeeSB!TNWbO?uo<%LF7|qxhoi#NE^|E%5aN
zVsdeDVA5s;Um~RJO<B|D@4hIIa=zvLt*AYy3Lg2BRUv-pq!soSaK5^II*?G*dozm1
zsggIbdUi`)bCxQkva}db82{d-hKzs04Sl0%gD8W8V6N9aQ`Mn8I)_Rw;$z3jCLU3&
zAUfI!+I8u^nSP=O(vhZvL@o<ngf5aT^bO(qU`uf9k%`OZ*Kj%F$kk2UaCuS)29LV-
z^55wcjz;6;&|;EgPK9u!ncW^6$>tO>!+VBKi}26C1ptxhpFFf+DhrI9w571=d}lX^
zZ@lQ03;ReHzgY~NcEFcHIM*^5RO4_?PfT^c#NL`^_C~<5EM)jlYs{-~Nc0q1*D@rK
z9f9Qt3OZp%D;E;$76k%7f`6^w`gdnsiFHwD!}P`OHOY;eSwGb|`?;DHvw>F`q459%
zDD+r9@?Pw1R9<txdpr=Bydo1!?=`&MG<epC<8Q)j^)im9OI7iQj7qd4*62SBkB)u=
zglT*N0`w8ND%XE=OD9m8(+I5?*J$lEq0RB!zbHn}dYH(;FUw3oMiu@{o)4J6oXRO_
z{Df;Rw&1D!uZK;_UA#}GNt72Ji^(C9u}=cZ<475Dlu#8xratn&;4NXm;*xlJf3Z6K
z6J<%Vq3=fxe@)cVMMD=fvNP1_?TrkE-FrR!<-#NWMqRxPodsRzThE1mx$YgOMBz&T
z23=Y4ERq<7g^V4O$!cq4XJ<mj7qEeJ*QLEur$Me7&$g(J>)xg3$}8rR(VtQ8_zJdO
zu*_Jy8C?pB4L68SRj&BXU^{)}DXI?o+Tb~ee^kK+a?U+m{<1i${C`u@Z%mh)QOK-Q
zH*O6^OM7<0oxRDJmk<6_JXZSF<IAKpp?!nLCG8$OvnLBw@{u9-WyPnzQy1erb`|tY
zo><xqmLq0?gN9W;0y;Opt|~DB3a(tS7+xN4O@%QWQ}tPts<Dl=DXhBAdA|KlAeS%8
z>%Crs^R%GBp?^}o2&X%Y&DHDs)-*rU@XEfAQl_8xt?kKRw)%)#*eh?meYbjsuy~~q
zq$>U^u{0OIk0|p8e`|ikqwBh7sgnop7LQmjT#y|6a#JTRXCA>yJ(Sfot=qbF=?*-p
zoHB6Fv~kaQ4E@4v$4a<AIJPC2$~s~+)8spGXUw`t1?DhoGd&@pb`}`<cI5i?YQc_{
zXqyjk*WwLCi-Te595T+Oo)VO1eq$Xtk`tQ(u~T}Suw*XRj+kE`ts=d2N<EnZo?9x;
zA{px~VKM%<zA?D8A@Z<p)UIZv78aReszm#JcO-4hWy-qTiI<FY$pt9l6`euYvzC6n
zr2I)>ENwLnd<B}i3_zmYF_L4S=JroB>Z|gg{a2x^8{hL88nY_Ei(fBj<jqIiXtJoB
z*oR|RZ<5w!AIUs`9&=W<Go4cN=PRaY(IP$bn49LIq>@^3sWkOowL_XSUuRxqTW#$Y
zXfQX_nb)Tb=gU_5K|$C2(SDA8gUdED%tjUx)-{$O99TH{L18dupK%VIiDZtH!+zh=
zg<oGKj$auc#_aIW(7aZWJQSUl81X$1bES~u(?9DnWA7U`)mJkuBQ+4S5@)?x)g~HR
zC7iXdV?m@GjP5a|DcN2jG#j2ey>AqWp=oc0+P_OcU(OoTk~;M3g#iem14$~R?;$m8
z|2lVFSz6u<vW-&oiT9I-K?a%*-tjDte`|X00?|(alH|qB3H$Q6Y{w9jkm}9f&23^>
zkK|u0(y9N<%j860+fk7UmZ!wJ67Uh2IZROGJ3^*Coba4}lI76nwpo->QB&d3Cr46C
z%S*{Sb1;{_(~+0(l1wg*VX`X(dzqpS_KBigT*|cT*mf(wZ+R_i)LPn}l?W4i^CWor
zo|RmXG=wA)^>=Mql3D3*FD$iHwJhx-q7(swH@e?zF7d6T50bUm2l}d_I*r@@Dnphy
zWpSQs?1%^cZ6f=ITNJX!RU7tqSxQ8P@#og=AyThAi6>Y8mvRI&rhr7x@^S#tj{}%{
z$!6YInY}5VZ~7;~Bj&OIP$zO>#zigtiwgH<w(F)XwXD9JjpPfB&d*08pem|@t?)Au
z9qQwsMVd|DN88a*aW_{BCZTfBT|<-Bbb8Xk==x)iK%U)fe(Kg{F#!nUCl9BYdEV(C
zO0ku)4_s1{rf&y$7;K#V<>%fx7UoV?C{{UYu=>iGC!aS6DYlfjH3!g+5g7MePNbV<
z)ZgpOH|;1zZKMe=F=NXc3ST2qJe_L}UB(>vUrtxaZ{KS=Yng6-oPv~jSazYqkRZ&5
zZc`es`}iwFl0EhN_nNf4%=z362Y1f8KPWn9s%HpAWIKu~X!)^GdP}b#!?fHim;}kF
z!)mtLFI@h;2j!ir&&PWhc6svSUp{#~u`Z77)TibDv!V6wxQ$e2BV9T?bRm<LhDm20
z9m(t7%JX{xwb~XcW462ZDh^NT&52tI1C1hI0q0+e{TY`#x#gMGfAo>a0s)%~Gy=5n
z-s$uYmnX7DIZpN%MYbL0VZl!c)LuadI72WG9$w_gy+HCE7sK6yuFKfi*c@jAdKL)-
z?sxF_eveREJG`c)9+czFwE;;q2sOSc7)R8p8Q;WQ0LXC2X*T|~*GU)ZiuC_Nh>Kh|
z%UgYN^L@zs!2jTS&#+5PXg~nwGsNfIP&;j-=j?T*m#&Nh9w5)_`c_q7WR8}A<6ei=
zv<8|Qcu1%X>&8?{PawhKo@v}WcQ^3qQr)zJ`;I&4*7fZQPnH|CwiXq=B1Ts!xoVyu
zLj#YaEG?;L+H?Xj7UM&qVj2ZrD?(ya0j86F6hPUi13tdd69RM!6NF|8<Nx1y%>3DC
z?9<LZk9q?G837k4?<HT}A~cM|muQ8Ya_@u0wk~*?5K_&2-p9to+~(^7>IAwpuBwJi
z8aa>!eQb(Nyi#44hn`r!VCiuhY{&qKF|10*#{y}Lg1@I8xLSW+j25dBOW`$E7uc6P
z<nqRJq$b(CF1Zxt^Dnp5%RC#v&kaAGrF+q#UoD)$yxDIiF*nYJ>$U58#i5aRRMYM1
z{A4BX-;r88wguFQwW?<FqL9ey<{ep5UZlip5J;oCl5C|itA5h@R_|k7-Mn^70;PeA
zUO%QI8gM;RW)(G)-SY%VB80Ql0#LUjY}x77*&+@oWPFVIUrm9Cj-;$c|KUlT0sq+>
zU~O%i)2gLkmwVyGN-nMWx@W#?S~$x_*93Jd71_A?q*JG=*ox&vWD4@K7$B$B9&IO2
z<`HVBPW+$7<*UQe#x4Ss)MPF8%CBR2HuU#UpIs$7<MPa;sLW|GIT4NUQl0G$S2@hM
zH5!2Nds4w2Z#sark@jzp@5&WWAiL%J0bCbBPZ-%<u<imqTff1S4Vls&_Oo`6ot(B}
zzAx%~RPIfoIeqi1&`K^s`C>oMM~E{4*36sIRraSALz6`_y?p(;&n*AX+8g@P5^J`>
z9uX8M%OY3tfLBF#g?Vx8sjNUMIloE;<ytxYawaUV2br)?&t^I2NuQPbFJI`qoMzEP
z_lN%WvRkw3(c>hL7_1-VXO;!jshi+<o@#5weU7s2PZn2}Hs0{;2=t?@{<qOh&N!X~
zXoKn)dam~^cJGl@CcDq|SB=T>xZPx4pVEiB_;lFFu!<4o2PK?fh*hEPdk*gb$&nuk
z#XjJ5`0E5)o_ZsvvWv@R)s?}dcxHp-dH{#KH>-dxW}vzb0#=riCf7_rTT86&oS*+`
z0)OGx`5yxlp`X9I((b!dSh?-Vy*OOv9E{L)RhX(o0{k+F?8Gg;2oz{vw?SjKVz;_L
zwPx;BR>=K1wirv6_mJ|AbiimZ?vRvQMX8D3_90sy>_)#fN-Y}JL8hwauNI09&|yQY
zlZ`T$rV?#xEBD!FT^fM#?-?A?bFtD^a2S=iQ?cP);opcQnctZ}Uq&wctZ$zoAH>AZ
zjoQ~J@y<XJ`XGucDJ8FtCd#@1y<bPfP$rFpZaPu9M5A)k`oD*ghLtsYHCuiku*O5|
z8UZl8dHXN8pWx*MwZn4ni1Wgal*ip+S4O&(?u~%9ofi7f8K^7k`>7SuDLEeMv-;=@
zGl4iei)O#;w*~iGJf-N+ehWgu(25RM5{INdNby|ND7oQkYAqRwm|*<uMu7VHFZeHg
zeWZb`l^x}bRu;Twh)tKl9x<OapTT#)T5)VL6LNhUl+xkwJ&0~s?2jYAd-siGehQy`
z`Zk-$aw^}Y?ta1S9?q3zYUCtkhiC1Uskv?ze=voQpV}kZd%gD~hn8Ic&sY`I;QM6a
zG>GrxdHpE}xI4*WgXJwzk|=IAr>y7g=SP;_6EX}*pqBJ|2#^9BE^4xGGI<NIwKum$
zy+FOiE)Wv?Zq#)7opVL4L9hCN34d0if+4=9j4|EBE1(@aW<Audi_k+b?myqfPcKwV
z-@Iw@F&N{$-ZwJseGPUPiU8EHPJb($qcLoR(}}KX=koH?-oD;jxOFhG#xT+oatZ-(
zgAPf`X35?p(&_&q{T~KXy#hnVKrBOJVk?)s+ukdA<x4u_bGXZ-1v$zW$#Y(5sfg=J
zoMRZG?-tw#1n97fIo|nTgS_2#L7lLJ>(1t{S941Q8!;Wp+@@K+aT&qPOWYaF2!M!q
z#g_1htZr2E?+)R1&I0XJxU({2SAdBg=}w8qV}cabN^C3Tcj&v6hWj`&m{jPWrPNL?
zsu@EPl-H?ggTKteau*=*gWrM?v}j<$M*g9_I#rivCF?5g?Ag{~u(cF1$IE)$+r0h}
zQQtBIIktkUENt*%Oxgc|&nQ6-_sa`GjFiS^!?Ib2%DDT1o4bfNH2g84)R4r$D1n&#
z`tt-id^;HUIGv3uc~ha0-9g%Dz<`*JWLfS%?R7^0i8n@Wsxr~zeNO&+2)%N06DR$C
z58zPyyyWbnSU)o4?!Zu3ht#5h;6Q|?(le%u{FOG~EJ8|t<Ic?p4gBgVnNH8rE+Q@r
z@S0gLcuAttYq<znJ!J*ArMExfmq|(@NqmHoBmrgtxh=JamqlL~4jY8j-hJR9H?yZs
zW20J;jHd!5qjuDCNcLF)S#O5fqO5?EE_vDy3|xKlM&*Ntk6WE6c%eR*>LP0IV&?uG
z(__l_&Tgxlg%9jFnR_Fl@zc=R3EosqdGwX#?5%q*FeLZQN16?EbqM@>2|~qE53474
zd4qU9X7uDO^^P;lpQaEPq`oJV`-s;7z4{}e<w$cH^wHJEwVFB#{!^|$Dg0(sotuzz
zQ%;rR3-qTa6P_rva8xSZ6M^nUCLz7LWcQf@;?if~`TvM7n5)c!kMMQ$ua+sqN-oB3
zQUBo-hZ14}q)#N2T<i4@z6N84+iQ+G5?Z^Rw{!CH^823u*c%aMY*f!*O5?inSv>*X
zmh0cXVGJS60~E7mt#~lW$<JBoOnZ5*OL8DwxBjlgA(X06S7;{eP8#3X(8>7!yZ~X0
zF`=iZjheTS&u<Jxl*aAxYW7`pKTa6p^`gZ70RxXO8pFV!QXPp!^YtgJn0Ew|x>Blu
zznZ1$BmSwzFqm~(52dj#mzAZ#F;I5~q^YHvdw+2mT-U27B!Ki=2Eg5C-VaIv8LIPB
zl7}C4^ypg({Rk4Onjc7RywO9kB=9-gv;zMdhxV5oW_*N9*V?o}LB&=tGGyKt3gWdG
zAT&u1^)ZM&+ro8M7!Af+3z<g0GVr&Zvo<Eh=KpS=YT9bGjOr%`*V^50&~RoPvC-rF
zln`(>7k$?&Z_b+Fv;>6M5%<khaO^3aJMhxfoHTeeV$Gp7y>KBt6&x2Lw@Cw^7l{+w
z(wHlVcEa7k6v7<$p>x7jT_nrvV>~@mF+%N415ZlCp3#}fc3>(i!7GDYl$Ah*&T?w$
z-SRYZHlhhmxQ)K|MM?q?pRUm**KCg14HYGjU`UhDjECB~9p@I!t?~{5EY{LzvnEpj
zayhkswN3+n#2w!P>J5bjAo<Nr0+2`bQ_6laUOZl#RC;<<>5o6U<-vQrVQ+32ktI{-
zjGV>83;N~9hZm<Azx*V=nIqa3fK1y*9&8fPVOyls1)B2lO>ZhF>;riw(}oM~?3CJH
z)9=|8u8EQU^L?hiumMJQJK3!P1tKq$WiyL{AWEe1y_-E8#fyQWSbY6XWx;kx8c~Wp
zOu+LIa*HR)PYmN|waMt)09!dT&ZfChQN(foM1S#5KO^sP9Fb~ton`}l#niG_Wxsfy
zB`A_Z$E;m;Qua<h&hw$0RI;*1W$Z@)N8Z=SYkU1mepMq2(_CyiH?2dk-zhe?E4P`=
zj_h^samjM?V3T6pmojaz(1v%4Tb@Cj^oA}AfyrD&qYtSCMjBOnZqa8w)djT(CYb^0
z`icc=I?XCA5EBQUUY}SnGHbgIzCv%xH`kD<mHqh&N15HP`@d&HVijp?jHiFVS#djd
zE4@av3d3H)0WM3Wwo{C0%X9}u)qv6tQttyH`-3gpQsi~b%gX5#$MT&VOHu<S+z?Aj
z5$sgJUwCe#x5G;{9IYbt4fEoxXwAX&lDOi>vE-V<19f0%J26ShdX4#b^~8&$!)-Fp
z#$sak2%@_=;DJJAUF^1#KS+cZv)GM=v{zjz*~<hcL;kcgJCf(K6_Ep%p0DCQ<03(p
z(iRIyn{*p1ZtL9LucNh;FMfC@AHHoX7irP^>8Tl$C2??WAB|k`^*@XGqQYNHO;Mq0
z^(~>O);u7n;fv$iVygikDnQ@{gjLj1x%Qw4c;R|;s!b~4*4IlJdw}88eRqIAJ<~n7
zeE#rFwL^o!bRz8Sbg@IjE<`^I<({zNt6d(M1g8V`L^V)ksSOBY@W)p3*OAq!ky<}_
zEzFZHN2V4wptrm%*q6339h72_{g;`&_Af+loaI=a1{M=I80qnp@?Lw}>^)FO0n-;?
zFB}VZ6EYx(`nm9A=Uq3~Tzgi*X<VV(XuIBU3b>Z$M)-$4ic-%o0QZbx9tW*=U6GMd
z7+a%D;}=SucJjiKi<Io1fuqLWD_MBQ4=!h?4v&<gqnxxf2TX=$fvn307q5j4$<zrD
z^O}#UKI^uO5XH<HZk7MT=YB)y>^zF{)xED?Po4i^0~2_-UihX;aQ8>RxEjO;NMgWw
zgY-tJu*U`E^zm{0di`%z)#*D<g}CLE?dDpM212qOA4Zqq@BZcdvBPy@f!~_?ROT;6
zILiGh#KCea-@u4E#rY+5xnqu!@6clxcwAo~gLI1Ry{wXE4^sB>>-P(aj^e&_+x$_0
zN8tBbrt<ogjT|Ma)pv^yHAJ2ABNK$qN7vvpXQ-w0fjU!)u{Ky^?BZ`Cefo|Hn~Xg(
zTrzy44m~u+sr!DB$u>1}{6|p}8P*i2p_Xr6Kf6!>CZ_>Z+}E|CBWw|6^ycKyP=JFW
zxaRODL80~-r=}pJG%;HGkEs$cu(=prmMawD<VVx*sfGWZ0dSj{aXBr+h^YGBUTkc-
z2-?f}T;H5B#B4YG2@kj{(C>&7N}Mf{*)Dkp!~;jJgrk2O<f_u9Xc2fDpi{ow@nSu?
z#nr!2AjmzrO>jG{UbOyqe0lJi;#3F-VxBSp1HZUij(X?HD0=57e9;A`&?&awyC;1%
z;LNL>vu=RkSTi}C7-J6Yq;HSbKK_T*`^7s}(N6<HOzOVHe&AF%mv|~xe|8_NXb%7?
zLg%H7wuy9M^6uZL{gvL7zUV(RZKh?=Q}sCf%4t2}i=NxqwOG|f4Z8TaB|2$&8R?I-
zr|Rh7bYf>;#e;~iHDo}_zO$yE!_wU)v*zTb%$M*^!V9B!^a6VO>P7EoQ|(rN46fJU
z=g`n(*R0UU{2~MyxV{GeilAkXV-&Ag1obLHl_gz2MlcW({0=pQ2f(um+S*TOBWOV-
zgNflBmZekKjoIY#A8?arQ&u=3A9%yw-t^q?>pmf$#)Q1%5q&GW8wqvQagS~6>Fz_y
zY0rZmSa{<BBHc(%wa291Iq#514@X4BOy-W;wDCpn5%?~jxmQv8vwC`7!@97>s5ade
z#K}gtP&H9CRIBk#&j(O%a_>rjdK|6O6L&(GyK#Yjq57k!dCCS)j?Kg$zWO_3eLZU~
zX$;yK+_I8Q^>77BG$!WN>u+&B(`n?E?v>dRHrQ*@0+7)O7Y@-k2N(Q{UlF%9()Y>;
zb1ugm0x~&${LQ}ql%xL<MLWvHNkX!_*Y>fglO2f+H^ibuXSyI3s8yvs9lzlDezU}2
z@fFAtQz4kWxCb&mKy3nr?+Y=r=hjAv>3D|^LSbNqsAW?A=S93Dz>Hfa*UoMMwAcR6
zzfa>XIQ&fSPHTXdF|+^|19iY1g6%lZ!TU$5bJ72^JtzUThg8PrG9PGG_iN~WW+)~!
zfI~Jh73h{l;s;df8o~rb0(*a6sL=GB+^(%kOjLfe@}vB4Q`NTOV*D_2!Oyg%FwN`;
zH#iq@gs%jJg0fHy@R6lfcs=DkpR*zAr+LuObSS~qBQzeJ-#}^H#<j-jT$90`x%2r*
z-h?b+f7bY`wc0=xd%RhuNR%Lgc~^J^<POpF1y<3?iiDu@p0OOluKa-0NIW??GOAa!
zy?Y$z*(V67tdeCNVQ#hhO3#5NHAlb9573=M19}wuJ06oCx|RBa2Tbu`LgrW0vx~Yw
zCD98eC4XfHSM^E_Ez@}ZC+!E0{gTBfCfOlrfs2lvmk*!DbPtF>mgB<f3NT5KpvA<=
zocOegv<x<^Z((AarzXPOajqj@Ja*ZEXblZb8^7kp*+8P6ioGwoo}}JOn2IzcAcl#y
z?rD=m@!O;_G&dCJ$I<rv>ZU1FH*|)A3$2Q9e6#noC-q+~m~;gu4}qV4>lQxeq5ul(
z_!xawnrB^7ni$(mf99a{ST*o$iv3Q0J2yr-B8Ab00Epi2fzDw91(inX8W*2{UvC!x
z-u8il<iBI=6)u_le-TQ`vdKF2%1ud2%g|GY7>lM2Oys5AWIQGr_g90OjS&Ut!3U;D
z(M%|5CwEwcq)qU^C3{)Us%w9H--{beRyWq^jyo@O+ng+4e(4bHrK~vE{L~l(5`3(R
zcnz*MUhkjLQ-2*2)j__|;wV5q?OMqCb>qA`pLEjWU*^1Ulii8+CnW;R@kyw#D9qFn
zgl3c^areQ~^VvXP(I{-R{({r&$7`h^QOvmd?s>nuR|LNm?|HXLivK5KQRtHA_2t}(
z)$A;3A8#6@+sIB?PNe=!i~Kb&QK{I>LGt1KCqMeGQur@|&vu7vND;LT&G~OAeIGQn
zi+JbD90RxbWSb9uI4o|=82WY;N>JlF(%{JAa)c8*)2RCE^e3&zfh<bl`MaOOi;d8h
z{a*|(%r!HXql$C<3fi}}bWxmZ70Vn3-t<4a3K-IJ<SWGEYmRClkY#cyu!)c0pO>Zv
z1(Z*cn0HFq9z8d33PtrAR_?+a;?I_ih`#9Y3Z_3g|Np3Z>!_&yFI*TzKxsrmx)GGl
zp&OA#8Ug9<ZlpV<Q@XpmyKCr@lA(LZd+__acfIdfYyN^Y^ND@-e)hBLaE1`$N|uCU
z72-n666_Ssv?W2x)WA-P2~;fpcok!!Dr4CTp{&X+r4PnKc%!=yP?pZ`f`e^dlXt6b
z#BTj+4J7DT*r&Q1y!-2ZQI})BeWJEBF24avn1@`=4BSTk?qH5;c;#fg$1!<gs&_3~
zu}<h|ycXGe0=&$?+Vc@DaE6#x8^eu1a^hvm+nm|s{*#hyqH9~-|G#zn3qNQTmQ7o&
z<XkHcP>iXi%^%N2gzs~S)Z|_^lfNr{_DeiQRH@<vcgDeWEd|dv(b!vI%REn^zX63R
z-j>WF3QT7H&q#`t3aXxYA5#DuMxM~FhPlRdTc3H5f0h8YIo@v~KOboDx?A%&+9&iL
zn09JAeu|i4uC?M1?2o?te8#y8?8HlfVE^V<Z1fg@P)<gdV{Wxr0bGV}@F-bd0yOVI
z#hF3mU;gd!vOOT@0KYy}-|cmfFRS!PHz<UQ26K-RpRjV3aCwY5yw`#q5%_bGmzcs_
zcE)^#`I2IL=<l8hynjct=kNYJPL4T4af|Sh%dYE&7AE=`5E1{Elzjx&|F3Se;2c0P
z%g}5~u@Y|MKPwIm50_4JKW@4`=DYc^H?9R|7o4G0r+9hOhpQ9nJI_?r3$!Bcv)&t`
zQ9X5hG0REX+t8+wmqIEGU@tG|2xZ3H=9O6B*-t%)MlA;+uZNn4N|{CkqIW%>!2RL`
z!mNCmP6sB|KtzuW=PDk2*@vXdUXBru^4e77y72C9;a@6BKY;2ZHO1|tXk7+e{(6J<
zf#}uc%359Rd4t;4`+@at2n776D-(Bf{`rnEH!uHfYIQnxi+}n;`LI@Z`6$vS!g933
z;-v_;GGcw!0l~VWBxZ-e!o9d^xeXgoZ$W*T%0j35XO%MqcJ&%i;nOeOZk327VHF31
zi+moQ${IJ$cJ2c@w%WOt^>U}C$}&Afv{|pvn(7H5U$L&0Db+0i%j*hXzXno4erM;s
zTMSX}Mmm=2G0rY8!x4!M%*t(goe3uDPYR!^<T-4-4VOCYpK>&rBsh%-X+SdPDb8F{
zh)#<_)v@SZY%<Pa^31K&qqt``9TJ~6pH;0k{ba>Udq+;cP>8aKj!%G(SQDUQ>sf^t
zhG}a2hNl6|I4F@%D{K9W(>wJ)M<Vqp6iZ(RupIbC>#f@NhMdr)NEk}td&HioS!Q2S
zZD2>xukWobtWh8hFU^z|2Ia<Y?m2z_y+MGKQ7fK?`@JO)Qx1X|WFHT+fAGFOhT{#*
zeulf{G>jvoq>T1YAQj6CE@Ts{oP4eHx=%YgyPeVZ?#{O&^nv{?BVJ#>xsX_MmYxtT
zjv$8o`{?|v_d?v1ls{i9CRoJY3$QleawN8FGf-~EdI(I^OdK7JkB*O94X;`o=VxUN
zb8f9VuY?qlcx26AoTD7B*5K-Rx4mX;dte#ILwH9m-AQG+tT$=dkbcvg5VZ0g?vqMs
z=WNlpk;@&Xcv&w6euqUS3u#FOtka2Oo{>NG^}CiN*a6#zN^Ips5$c{s88eB0$9IMg
zEMz5>uk3$W%22>D{!Ss8Uu(kXY4`q;=NN2(8!~HYx_@oMX_3+JVU?dCc$9bIfNb>J
z<iwfkKzRD6?<Wf#l5OG3Tw|WQmO8)W&ou+gE1q`wwZ8bEE<(NDp!#z=57cp|SPDoC
zK;aiG2Ayq>4BK^%HI%pn6rQOlqK!YLfNj&RoEaI;sdQr?+oHQSH$0l01I%d`paZ_D
z3z5dMt0Ssf3ES#4#6(o(8__=+oz(zowJH#KUUrihFU?~z#J6K-jfMH&qp)~N@PRHY
z2Btn$%J1;-@T<8fC`ctkw~#c*gT4Y(_v#fKrlM3~o~#<)QbULE6=4dyxr%Ozs&=vx
z!@xN_i6<L6JQTjsb27`!UbbEQIyYM97Z}yoZHzbWuPthsYeR1SmcJFFh?M{?aCD5f
zM8LrveWB>sNK~MNQaw+1B;vfqknCy$8l9|4J$2(Qw_kDzzGmfeD2FB2`GJZNkZkG3
zc7F)sZbZV6WXU`{f~x_`Wt6#SCk@!<+S`1m+1#pV9A&)VB7sZrsBX8@@kcluz8!(g
z|Jva?4#@$1bb3}Badti#w%l?2w`pYrncz2oPs~PUWG4Jj`1UM_y4Oi2IAHhsjrJRI
za`Iv`Jj(3qXsnp;W(-4QWMtTD!SGNhmn#dA`^{5CdW)0(&GG7ia(m+P&Qqw(z)?dE
z6kbMYNRz`_xvS_vaGc8Cw59$~OzGz~hGsQl-_Un=xzNcC`6zUwu}|&=vsl+Ggd)B1
z@%Bwb**Qp%y^qT=;_#jZKnMCP>#%mZGUYZWNBK{@Z{VSV=W+vwLR9EOoq;7A;>r~6
zw}VcxdtOTxGyS_2A`sxv-x`g0*l}@5wj<nOPat1#MLLY%1eg3s{dXKaDRT3w$5Na=
zDx0|OtLN2(1AJue??CGU^XIFpEq0o&tIMLmgDr_AJmc^3p;H&;1l9#ZP!*;Owib4B
zeZ@-+i(Dap51x#iJiFMpXDkMcq9uSP$Z(#y)9QmnO#V@T5&iy}#HSq6cRge&5lJYd
zks=tjJGMcBA%8qzJ(Ev=O~mW(v0%f(M}&GHfxFmYz3$cec)hGvV}yNuyc$Hvhgc^f
zJlQN`j?!$2LG0=9-fRwuLFG95chBh*4T=6i%Zc7XrLNd@UmwjzU*fyDJN>B0P6BIb
zhJ@$cr^mCKwYG&S{UFu78b^rr!o%{3%l7Y}m3-Tb8KWVvPKo_Wt1VWwUA67w6<B{f
z4HO*u*_C`p#|xmh*W7Uh%GLdT%__Y%#ZPH4iF}Kt-7CX+ZJ#_a#Q^@k3NP?biL_jA
zIZ#s378?IaWutc}uCexr>UP5qn4?fV(#mb^Y{|T{N~?C&^OX>klIu$@ldA30{&=T5
zn#3F&LHpEKXTSXU_FSr!$1s%c6AukynTA$4CIxXH1-d|K;AAWx<*V`6oAa91JvX1)
zCBBqlT<$NzJrh0a)SAsr8@6fGn*5GDTiBO4Tk;2=Ztv5aO|3*><b#EP^R<2HHGR3y
zjL{`_@=0PDBrPqi6y{fYQTm)3iG!;XEEh$s52!f3#g-G?WL9y&^t8G4;iI%Yll4A)
z!jV8D>7F{&Hw$J;79VfAxb64e^rM|&wme|7SZxllf0y2qCVK8D`Fz8h5jW_T-kKaY
zm@>gl6VqS-vcDUpON$$vd;GSPCTh3r`;=)T3%oAO+mqR*ZW=`esj*rf9~H>Cz1hgR
zVFqm@d4}7s^OQvSSuWXpGEy+wP_+ccq$B{5YmXEnKR2xoif)04iC@X|x??*K`{jM|
z^lBtcYID;u^oWGr?0rJTXKv1~;zrc-&&>+--wN8JfN0AfUuxX<k~1a7=vM(x?Vk6c
z@z5UhPsrd7ovP{uuSEud$4jhcN679K)Z?jd@8BvsjPd;AV-mC6=9u#uL$xI|6nlTR
ziekQO-6s(B0C$8U2-f57uQ3y}j5Ew)Vr-<RUf`)W-FfXN&_43FIk=pP_<lxCpyxCw
zu*CH@p@C0{d4rW&wFzeP&Hm;TsIho`p=x6DYRI~?DQZB+`#GW`U%KnkbcWgc$zx@o
zI*BTRkR6FcI#b&qmA6A7N+2?VfK%__{)EEu`R$6!OvD#H+wL`l0=dnR0OUDnh57}>
zQeDfeC3Le36OY>hk3U18!)AHB<4O2?x65cMCwhrWnUI6h&`)mTC)l62djiCL$Y*D2
z)h=)I+i$$8826PiK4!nYKB`<#6$wQ7iSyn+Uq~h_x5d3Po$d=Iq}Kh0*uiQ>-LiwO
zUMj``SFHKC$j%W941Vr*!%b*zGp^C@yAT_8zdI&hMySQl($@T~yvecGBEH+wZhAw$
zY-MdPvDtj#$}{}SfJa@YN32mJch@nsL@XZDwjRZK&snH<aUC=A@_4%2E<IN^;>dEn
zRr{j*F<qZKF=gUR1E-o0;u@d#z}?m7UGNT5S(Qa_*0YOATPM3%f<ItxsI>0k_LLn4
zWSd7g!#RgSSQk%%Iz}w!Wpoz>v(gQb`6oG1$FV>&1IOlBm>7}Wzu&O!=y;r;inLSv
z`*9#mWJAOvCmr`_UK(}y?$1jRxj6%sracI)4~pdY;u88NPjHPu@qXA2WNPM#>Zbki
zKFk6<k=B(jlR8TiCg|;bxYR~qN+XLSc>$>)h$({<tA;hGl-&vK1Y3T;JC(H;c#ww=
zKJGZ-ck~*FqSD>V^xirsDOcpv%D`vsH*8nc_N-$PoGp69zBM}2BjmAr?e28aUlOOS
zvTWV_*XYoh+qtCUX({KGsOopG4>+Ds>_}pMUlByKt@@vE%9CFZ<c7INH|!@Odd!V{
zOeQ-ut1eTmm7wQK^y-aoR2>h3(^gm;fze2nk5K~O6V5i%7U={pPb)WX9ar7vI0(QM
z5*-A0g^#Sp6aA{l9o=OOXi9t|^Q8~45@<x-wKr=bWm!KAcDFaCVko|r?xm^MDEaKy
zNybH|{mMQ5B!OPJ>VB;^nWY&vI&#%DTQta(9NCYGUw9)I^&w}CCt<Ijzn~tSfP)bB
z^qWu%&{nFM?`fC*y2%2lPdT=Gj#t$x8;((_P*>h+!gzre%Iql^o;wn_tXz~5GmYbu
zSFnNQ6C4hY&LZF_{ARD-$t%Oaa`e|Ws6Lde@(yvynZd!qgAGHBcuh96kau{59$eJa
z{VJH}7v~l{PRF&tc(B!+ohX54CD&GF^Id7pA2yYiK%1ha@w6f3uyRO!?)IPhxe}Gd
z<F*t2y60&r#=0`CmU&3c=;&y*<H4*#JhSCet=_%+&9P<6cC1vWg}e8GQhdAjv!=z(
z`m>KlCYz~Z3|rL09}=?`X-GtlMAV|MU%+sS?H}p}P%4LI@w9)C#DYqxnu5hMATe`Z
z>hs5|-q?!`7K`WI`r#x3YYE19edir_2PJ0DRF;fYhQ<)FYd$W6!X=tS@VZOR0y|$u
ztXT;We)0805_3<<MPhy{kJwH`WN7^(Fb|$#W>CMsZ=l4XOruWe#=oRK2C(I=oU0x+
zU$NeQwm?TAwip?wb&C|Jcx*Z))T*oagI7%?0KJXVyEh~V3Yb$c&2ZgSZ-DNA8snm)
zv!>ZBX1}bSj8D(ZSUN)HYa;O3t9!2x*B;%JWh=DXzDi4oQpI^o?X1tH+zLphpPr(u
z5i4O)8dEPWjW#LJyMMb`sJ<*!Oni9o$TUe#dEg0^8q<E|=ePG70TCDYWWGf~vpSJ<
zA9VQH)zt)ipe)`p)*{x4p%0NigM$n6Pp_YyuPMnh;}dm!cqu0S9AFP#LK%!~Z+s5R
z6k24HG=zJ4Fx`c)E)SM*bv)}ZYK%ua>n)d8&?10yOky!ay1Rk|iANCRIGGfBY)R8&
z##~i8E<42^r&@Q=C#a5#U4qxzxIFI;<>!Ej-SQRae7uC{0tP2t%$|3(<O8Xk&KVlk
z-cW4q^u6+Qp4pYJiUtO4iZkf=Y;XA7uRqs6ve*Z+sn5@^1NHQrhQmB&Er(<0KqoZ{
z(#qSy1V4n;a&1HFq@9t(CSIpF@rctCn#|7}6Tb+xk7A{k3a?~RITGB^#;G5ieV6Q>
zHxMU(>(dzQ?f2pEfjjl1`0}TRGJe>Z2pBi4Grs|Ad3#nJRRaoVy>uS=2QX`yIy>Rx
z65P>2xoVCT(S5W50jH&%CYS-N9RReBty#Dxu-cDsKW>vK8*!JrE{yFVCzcgZGH~hj
z_4T+4TdP`)=^tdBG+sb0B>7Mj*d^s*s~2;=^W^=s{${S^sOtGXF-4?HA@KY2&g6ue
z`zg&KUllGYqe<bp#d<wa2QgIn$eEBvWJ2C>=+Z6*V&`6NixB2S-#tE?u2u{AbSQZ*
z6;CxYcvTnV)xDcLMi(LNPWlLQoxY-)WMn?A!rSaX*}i;F-@a1SnIT|_-o6CQ`$`dI
z4k%2xPt5udhliMRM59LzA|6HOA6Gw>&G=|k89(c^KPBHu=z4C0h8=W{Ye{-y#4`r&
zdLT+b4S>4wfEV<j5$#;rDlY75DXmT1xFW>kiY<zJJj64=blcG>en6r93ZHw>@;pTy
za*)=qxubx9sUU+vCye=yQ1PG}v>ZI%!f(Ww{UNa2`Ye~lsM*xWI-9@MAFvwR5P<7&
zmBA906~?mMMINolVuX8?N_|(GcK)k(aH5BGg4b_6y`s0z3h3Z@(i`G(mY4XI9-<^J
zUS-MWex;Jh1$0NpMnGv4ga24>^@#d(DiU?fN9r#q|04@YiI6A3UpxX|Z}skKaI67w
zLgV}PVyiIPz&Ax?uzA2N1H0!RHL%DuyqU(Po0E!4jAta-_ENp009f|%&UlZR+h+1A
z+Wt6_erR(!cl0ogeA)hz^<czK)gY^%>D-2Qj9>S1`CEQrA68L<lYKC)Zeu%!0>q+O
z95w7HQ70?&Odw}|@9~K_gAVW8Emic2v!a|T#QUD4cOy@PxnbSn7K+{0v|QiT>owkF
z9Bp_0`l;rO7I1Qff(qTZV)QegP3G(6rmT#4dM(pJ(54ikWooF{v7ah`!+5=EWqE(g
zfoi`u){`q)LHhB_x317n{P;d1ndM??Ir1nPPn5!NKSy~BtDQ5XMVy`%n+@`hRVrxY
z^8R#TGiIq|kuVhD28_`by}5)tN3!ryzkx|bg?g<$M)S99@-^J6ka&;mIH7XXhJ}7s
zQSF&wWBb3k02cS%Pj~mY0jbpL^CJOcUhUOoMmNa;fJ3Wl$uMNjR4`FW4tCqG);(w*
zGT%Dc3zE6{+&j609-p+=WhkP2&ppW2z)s||$2tBnR>JPrNn>Dob`EpNg!F<1eB<2=
ztl4tMGHS0#BA;T?d`$l`BX!x{w|N%800nzU2dk(5(N_7?=YdK~MFiNfDc1pO>%!0C
z!;x^i)1*8+EyDN;*Mc!Z@8sakelob1xoow&AIHmi6z$9op_FRgvmUput}Rp1Qo%Q@
zabOsar_rj!9(J7_|8mBBU$~iEkl7z<IC6TE(e_0~mLUzXUqMzoL*cV`55g5DF-{=K
zsH-o)ZUwaZz5+}h!sh1Y8#J{3c-1AEOr0vb0Jq0YME9k_;O5{`)rRj&OX-3I+Pn-P
zctgUrY_p(dbYas3(zrTY<rp^N?^2pVQ~C|2t4N!U6yF&T(TRkutjZ<!j-<i%p_cZ?
zo7<5k4ia)i3OdtLG^{nV?yC0~goVEF$;K>U@t7JswKksMO|644XtPk~?ktuu)}n=)
zlSzQ5``2PRDFkGsQCEN8a?|X27TL`H?7pzs`aEw&vFqz^gF}O2g8HbsLf0F@{zi-~
zUnb#q?c<Ggvi6pvus0if6SAUFFyNKDb%+GK1;CDFD7(<%Ki0BKig<?#%ZiF&TILck
z654URscnx^BQ=Y(Cba9;Usu0>j>_?HrQc@%1y*cVHKepuSkCn<K!2ZKW-j@s?EItO
z3wjtJWG}I=OPsUI73fF)`ae&D4teo$OVM5u?QKy~P>#Vw$l&0YsF@)DS-Sb<?9YJ9
zDLxQMbUM}Qig&%?SNt0Ec?t0PG)ilgOH}`C7om8Z_3b4wdGVIF9CNdUL@2;!L~97s
zt0${d&1bvT7^yYc<!L;3c^u8<sBS)K^BqA19)7GhOnV$nmE1Df<eAC~G&70GIt^XX
zl~h(T#n<FsjE_t(es|8TKuK}pIK4c`O+<$C%cB<wM56~Uv6o}P2?*R*tT)$*yxfQl
zHTqFbHU>VSMrrtHY0b|Zm8vqOiHNwS>oxv>hlj}mNyhu@{g28CC@8n7Wa-Tjel|yS
ztFw3EAFj1OdTL&ZjYlEQy7@`rZ#gh1VNk4ACKA1qXJblCv_x-(4>CL0lY6}o6@MHY
zqTw4{O`~8%6)sHe7gT<rK|6X8KX5Wx*=A;sxX3UL*3jl0V2|OeFt|@lkg%$WJ>d~|
z{FneJ<yM??{9$w2^%55Rde!CJU0oUEnT*=pGZoigLNB^kUnRAt|7zBM|5$R4jm&K@
z%68XYsl-(%1h_A-Z6oMB<u_!2X!VGl?EDp-4cG5eZrv|ZSd)%zv$el<9fu!ppYPW-
zZ|@)#a~Osd(r(xkBt!wtjuc2Y){l>zug&rAu8tt};g%iGDqGAc?ADP1VK`(*y^^IG
zhUhSkCyO~Y-T++_!=(F4#RFikjFu`lRqNU$(4+Ws-eZ37<oNO{{adTEw5!QlE2#ct
z;pfqELlZqch_=Y4?JxY;@X<o#@8hrJ?p8w?tN;dNV7C;D5wV%W#vBxXXVtbpPcCAo
zepC0c5r*VCeTU^1IzZxG4Av0p3NF`_Lp`E$w;E*J)toi^N%27vZ|Rs}W$F&?`)Z-o
zChj$hjNfG7LQ1PLJW~4`H6M0?^1fm0Pqvr)3jTosWcDqdCcV$A&-fa(=8@vM5jKaH
z%*ug>JGGnp2=5n=CYfq6iSG}YJ)-`Y%olzzx|EhT941ZVl}#%2i|)j*;5F(n=3zew
z-yWr$t<WM_e{6eCPob&jyd?Ym-JHq}C?h_6KiY)Sm?kFy$G7K0&{Tz(4&lgG#eujm
zX^NY+=Vc|7)_t7q?cFS`N2gOilKMOgk(RqeGCE2mT_N27N4{@ae@^1=ZIq>PAKV;>
z%0vGa4=*OR<2pAqkqp$tDK9>GJq-?zh(_x~7t1N`GH)-Q&z^c7WiOL;3drY~5-#DT
z#&M9*^*&ut&J0zc;Rp2>K313JZ_3Qo{Q{p%(%yd;kvU+abt)qQ%V=XQa%g0!m=S}1
zrL8aqImewC;=4u-UOSHe`MB<edUXGAHlh`#6qOwr!HrWivdFeZ8}*+m4=3!o!DN-c
zjHm8p5&?HOsEWx1FK@He7`df3$-c<u3?6__se=y8)=Fuwj|AlPvc8~6;VQuLDH>E{
z-}<|RdmP#t`~|uEi2%P0cIbX=GPYYC|49Kj7C1qX@x0xFHTh$+1wru0`~xS;Pl7NH
zlaF^buvrYMT6uvrJh8w0qxkRMG`U_gXS+Gvw9_bzglQIoA`L#!j5fIrFSCQS!ur-d
z{PYK+^n3158_$bBg%NE=3DlRUx|Xk&H{}Z-Y|3Wxd!8=h>in4M76UlQwl9+>9v_P?
zmirx3qhoA~)rjovQ9o;&kPsgql(f7@GsPvC5^ck|PO4sX-v-rd&NxI?A-=b2yQ=T*
zTueh3Q2M#Z8u$}%bsY{I@(xgq$7uYJRXP7^sI5);UEjB8EjtAm;>bji|0P;R{6_&u
z)G@3$>$V{S<8WheOTcKTM{W8k;%xK@yQ_hHdvVQ7)2@-_HI7GCI8z+WWp0*Oa|mG0
zy5}@@R5K)1!U4OT;su7jgY8h1UuuvO>0{^GkUw*Ayye#|OS{D$K|Vg6<lxbi`qJSX
z6o7gq*U{a-V@$yyd4=~GQ3&R=iMm$>dM3lq`lj>j_t&R5=trOH1emGs$(guuFSeS9
z<bo=bX77X9RO<&lJ-yH|1<Li2BfE|Q`(79Gh3MZV(}gj^R{+a~oh(uycCkyhQ9YbS
zGf2nEicBF9L3d?Bud(1hEuF@M319i!^L|93<$96q;T1b7SX-au-Dff~p*D{2$bK>0
zrf0pS^ej@;bZ$pQcf*R9P^R0Th{~j2?{4;Klm&Yhx2U9M6EqQ46gMyEKm?E59dkGz
zE(ur+f=BMufti>siV$NeKX(28aN2s$tuU;Z6feaxfy+sezH5nOzWzu;dQVwAS2+}7
zUbV7`B>pm+g(iN^OIj3+rnvMReek%J*@@$71<ybY6~MRMZS(q8Hx!ppQSk(*CKZ~+
z@`bds`U%)Iz4Gy}ljpCioBy7#@k*3wMqu<GXQB-2C#pm>2orrMW4jqZKiRA-Rco=R
z2xK$axr@OE|3EO!@ZN;4%Sev%8;YaBJ#IPcCr}pL&Tv2OO9R>4?4RmVlE`)K>3ID7
zRB;%d&f~n<ZFKDZ`io3zoqUDFDA_{u&Bry7LWzkCv1zWg!~nu}S?)t2nBSkXbI37q
zaSQCM7HU6$ojcHg)?XEYGdwQ)q-$>X^rPuKxOHOmt#@W;t{HS11>@<bqux(Bk>^L%
zd1g;Vi#0~NC+psSX8+E7`O07J=F@h!WWK-B#EP)$rI`nwBFonm@wlDhoinNi&j4=Y
z*@Aw~>6R2@(R;1(!*HIH)*7|;@$|zs5ge~NjF#smzlT=dy7y|kzSVBqF%q*ao`~qW
z5Ds@nM%6AF<BoEjAP;-597x55*R2S^cOUGguNP$9`|2_z@^rYTfKv;RSRDAklUJ!Q
zyl*#>sMNO2V(y;}_wU96%M%~i@Bx?@6SZMuuar0VyT)KLznYKBg-P4Pmi(H8JDihR
zu77_@U&%TX001i<7h|_}o5H-8QQq&1;sqW~mrCMJA^l?uEzds7_?*@%4zd{;<-Vd|
zXXmdoIch-3rt3gD-C*Tj1t0c&=(FnqZ7!i)*LK5NZ~vI4%?$lx*kR(&I_q{Qz6@hy
zR`-E?>@(1bJ33keFY{FA@BzZe%d6+VJe#Psblp0ufDh1$0QVj?CHE*nuR_n*jG^OE
z)w$y#7RAr~j@4*f6d+ue>RW1U?i;K0F|83!OBqaO5LBHGvwYnYl|;2xJZ2*D=1*Gk
zMrO?+VT|Ib9A#26a)2%sRZ{-Y{FG1aPk7PD2?NTnBBWSkX1LS_{dnbr9az41Mh*m=
z6qLJ3F=G(7qn#OBF@iRucztKo)r*Y>Y9@GdSlX|b%m;_YyOuJ+TcgRWg{ln<%}z&w
zG#M_0tIdu*Kvdu>uiypF7&2nn{Hs^CV|8F0shOg05_1>gXfc8IhmsC|^DLDGH*VIw
zODd%p;czQjS>F0twV<9zC;I|3ocv~M+1P2l!;n+;yZv#e>daNqzZN&%@o`>*R-8{f
zlg^J@)T*}J5AWRts%vWIFal*#dbf&8nm6W`t(1Gp)8$CUll_`89hsfizB_ueo6o<$
zS#>7q4|hc(=UOlvGw3uLJ79N#?j%-uK<(1T$LX$6fT>!T77Le!8YC<Zh0S|X<a=8w
zoYv)EnX~;#34r5}0a{2)2BU>Ix$nl1$KJpzU>hL#w6o#@nlRg}t0GtOCgf|*TiZq<
zC8g2Lo}fh)<zPbmlMtGx04&oYY;Glzc$pfbvDr!AW49+Ie`x0D08p=37~AlC&7)Cg
zroN>xU#cCU?RiiBkrYExazLqqo1Cs?VL^?uAfK4ofv-NFUEf7i6c`BaMnqDGGzD^p
zf+dsM?6Fl)Q$p_Q2$WJ0%7!{}f(z}=co2Xk+3y`apQObF7a}yDthsNJY(&B9c=nCR
z4Uej`&XuI?71qMUutCQ-=O9>eyj?IaB;dbC0lRN=H`t66RPpU(tt|$FBx)HE)ecmF
zn(qC<cv?YT9>&|-`^xQd7L%<UD`#zQZ%@~L8ppw8ffyDx@0l(Uik{A$JD)mhnypQA
z!ir}2W2acz?C~1$r1R$`99={!W26D0cuuAGOQ)hpW2&5QIkF4Yj-L`d9@)Dg{+mw&
z$EuDXr@N%$RD-etKk1sY1#!^;(H@eg!G@j9iCLn?->xJNULGY&RyEi^G>17v=R<|1
z)lec0iN(G;AuWkD_~BsqFI3_kPH_1MC_c$U>#429VUZ@e*bs1MO?^H8-k;9IWS(|X
zq5qJ()AneFd<#Ej!#4$)9z5(cMgmaC1OPmmaSdCc2&x+<y9&xZ;&lXDEggwlDag%a
z_tN#;dp!kFg`O;rySXw4A*TQiQs(0%T+NDHt#qP@r)K97tCPu(32@-knp3V1`8;q&
zH}a_xyO5i!{HMhy2tWIAF3D`Og|+Lb;BzPW-!e-`BcooozJh%8$rRtF2jl;{hp|4=
z7ak;=A7+3Q@)LSKqFQVRuxx&exD36<nVx|y$IW55lZc*Ge+Y8vd42*BZ#}@rUPy}d
zm7SVOT64#5_T<1wb`>-sJp~}wS5zbvZSMD)ir?n$(G;#oy{#H3GkcSBxU2UL<*Pz8
zFuPnPo_VCTTHgaM3{&AE3~z^RAh&ov=&}41F?MaCcp7|+>(`?4RAA#BwEuAsQ+mhN
zfY10QJ@1ta5E^iOg4u`mL9P!2Z+NX-TrJu<9r;l?@G22`j8E=RhMGO^X}obe8sIEA
zw(p_X?Krf?JuStAm_9XE@id0{ah5lGAy@%dl}K+Z$-ZCoFcSqpi7Y8@4EHkm5<Vk(
zJuRfgYFw&Wsm+Yk$KUvGH!}7+<3~zr8vk;W8+7k-V;r)>`}hJ$6f^COlg56qIy{mU
zSQ;v;;Q*%@Q=u0fJe<wylRD~AhuZUlLPe-2h%XoCoDa(3{<?c&9bKYEG1}CiSoH&C
zOojXO&3gpw?UKWycLY<tN3MepL+V8VCvnxX8#j$(Tuq!rPhNsHYD{|NR^dJbsXaP1
zh%5$oxoP@u&KdDL2N^5L4u(Eq=q8WL9w+{d-DbA<SXvO17)DC3Hpf^&s5f-J-~-XN
zp4Tcsz4@yIlC2ou8MLQW)R7{UtWR%o+s3)wto<Mz2laaxtv{<dsfpbgBo8GTL6vZm
zSROW~109FKB_r}R)1;kHLEB}S3VeZpFZnB`-sv}y=wC&2PavKQx4*+);1<50=r%h5
zq!CDQp$2n;n1qDGDwK2rL1=`ZIVcyZmlNK*cI>Q)5u<a`_`!@I(9?3StdKSmdW~}v
zrX0evRrZzj1*XOkiVO_>o+K10X(F(aMfhak30nCMMd#dWWZbr~mveH-o^=;qMV7sJ
z_ae5m&-PO&af@s(24)qyXt8%Ek1}VFf)PX@D%MBH?~!x^3+T}gef|C9Aoi=^-;gTy
zGSnUlau$~9nZ>+_;($c&bpJ&jRo%I)WV%Tmt=a8$1yQA@TOv4s*lD@|C~AIT=@;|$
zzefoUa7kaTot&KW{Nf@yzrb~t*Esw$2ZaKA_FLO@R}Oa|jJfqU*;IItIfOCLv+<g}
z6&U&zJebJ*<-hOP4n)Cl)L{xA`zFgq<_mNAnh8&k5$4Lk168m(?HZ$V52hW!5)~S0
zOv;WWKQh_R1z@yqw!S04x*!X}b~wMBzRm!kFaJG--KN41Gbjk2$Q%4Hc3ae_f**lq
z$1A8CtT3tUWJ_nE^6VP}VP6cDqcfvab=*<m5KxpV&iQ~d5!=9`|8GIWx25ZpqCujG
zDNKr@)gUc%bkgn>`}duRFtIR$Ye4^(OJ)*-U$EKADHve<=BI*+^w%-C78alBN_6JS
zdKPT|C>>iT(koZUMSSQ12>ook0n-08N(`cbPodWlf<>`Cg`co>BmvL$(QRYn%zP{x
znE{3=nGf|Ylfw**Lz(a1m-klnf=&_EZ|Q3M5YcglnvJgT5daFHP?+g|$0i!D%kAS$
z(46t-Kn<X9r~ISsn=F`u{x)l1ZP|%{z5mSprpwP%UlnP!QBuMEZX*<v2GHt6%dx@1
zbg%vX@AYk>Xg$}}>f<%HD`4sI`7rEKLx)K*l^bvU{qwP~u!0&vvLrIWlZpozKdyaR
z<_TF8(AXOSRU?NF4j8nexooo!gec)eb22?*EhT~62w$OpqCif5PIXT&l<n5aq9mCD
z8*Fd;ak1nG!T#)5CTMnrpw&f5<p~zn%mswT48*sw$1>^vCsX{nF?n>RtxQTG4{11v
zdq-6C)reE*FmAy5QgNeR-YFt?K+ctg;Nju9+z#I02e5Zn%?wO7%A!0xY7IOZE{PNz
z_Wb57n&~xcet6)u;H_7X>0f1QHLYSRAGX;JK?)Ji@+Q^m(F>iW5}6Qe&M2Q=wt!zq
ztA$r(YkVca8kOCt0^=><%~rgM5q9&rKZKRWR_oC54bmIMWeU;P!gWgvnYJ(;B+sZ<
zW^+A5X%3f`Q|KKe>l$iuLsMu$awG~@qoX_#Dvp}EQs)WX&FpBF;SmE~Vu1AnOp_x@
z(;znquj8A6n%Jq|x)KH-k{4o{dkneN+e<E-!AA4-agMdx6spT#x6;<m+J#I+EM(R~
z0YBYM3GGFT&iEIiXO-julg+Xfs21tqcoIa0d?voK?q3l5$Vn9i!fw3x;R`H>w3-0-
ztE3EwiZ-i2qgX2R)dezAu^2^uZvJKusHVySYInF&1xHI;bb4_k;*!p7;=E|Mzt+Zc
z+;XD*eKL1{JEq2Xth<Loyzw)u+6;zz1u?gdD-|)}11Z4j=XD=cyxp{2(gBwZV5@#F
zpASMI3XI~1Hmaonj<WweS|1S26M9ay?iY{<7fi3LBsX2zLG>XK7BT;`m4N#vaE<dv
zZ!o6GN~3M^_((CS07(?z*=zoE0TNJiWKt5|<;gk;5KR=fkRAmkcIXetwlYIeKWjF>
zr!6WfBvK?eTGsiw*%SEZ*!|ck^<;l4N3Fq&&ujJc6NiObP(&&R#qz^<uA{aaCWHM+
z`|{Ly6buwC4itiKiWvBB40d7vW41^2e!fZ_3j`SQzTIC+SM8ok@A(gTO1^M^INay;
zhPm{(zX!<uq#0wC7@vM1G*CswUT5<>x;A4wR{)26@r4}bpsjRdlpz1lHPy|Q3ECuC
zF|pGNtKUxY>Ys(}is!8NRBvRTM4$X25?N-lwC9zs(a10oklDe_DZ$fdSOd!t2`Q{u
zRJ&GH{rdphwqbG4mDypR3R0_x|AQ|%zWf#Glq->J*=gZCXx80(K^{inp=DK&*RD@M
z$^_;GUe3#7ZB>>Cqi99Z5A3FsKG(5>y_p>z<+QnrQO;~_KLDSSvY={@e!vsFGy5-D
zEAk~-Yaf_>%Bo}8!&K}*vAegk`PVq3ujQl-Q3hFbPd0tBK#s&8{5D9(M}UQu)%J8l
zsP*Q!#f6Lk&2qW6t6K~?3_eWQe6D*ih9b;jWO!Iu(T+vsxe^H(m$AQ(USlJIlRC$G
zogdqYzf7~f*GKa!nmD_~?BLFI0lhvE?)!ykCvsGR`HBJi34y4$STtL|gNVDOLKw+D
ze)+^)5tA5HTdM?2ym}iFN4;4G))^oIG87xu-LQ0Je^Sd0{aCqRY|T<!cV2&dgND@o
z<cG*~<Y$`1?MQdr@$4n-1~v2#3Ub%F-lULF$}&CM>Z4<#u8{kagy(qVt=)oyjzZ+C
zGymudc<wEu>wr%~DEU=Us0-mL2a|xn4+!Wd{N?3~F*VpSQ9<&3ID$ok(g6&Kqc7w=
zZP(z4n4*f==KXf*tlnRPlT2QzQqCdTi0^XszsPP*B-g)DMp3Tw9FULj8gfl%3R?Z-
zJB>x3*jfB4OE(ff;mp1CTh!{+H?i@qi3ybE_lY_`{PcS<RrgL-jL;&@sXl|h*sblR
zRphc5EEhvg5Vvs_N*t#LPqPgHwLeU0zp_wOp{g{~M?)zpBQ2dqnVHALSw_t1bofJ?
zj(Z|7Ip{ea*B8)`#h(OH%YIpJ27e{Zq_)7o#<st$90@`y@sG(7-}L)CV|zp7{A20z
zFGg3FRmtPy&AL92Gp5$+BB?7!C?S`f6o^{o{p^(pWnrzc8KSbZIlQ%<MM4rcGb<hu
z9BH!Q-EmsyyV1KYO(sq>I6^gO{D8gLMJ953_5d||%nQL{?nim!8hdj#Kr-O|l2)3D
z&xaqv*uH<I&Y;zd@`#8g4_WuVAu0zP8=K6-#~JAGIF~-&0seS84;XUZ;s^;vP36e3
zQX_h((nkas5x`sL`^o>bjrCA(_-U!uRO|C|ER~eaSsy*n@uTnb;-cq3xn#0MN1(2P
z-0j`Zo-qKApmC}m_5O!&zJLTGbg29#L=6mb3tFOny-PPfw-&vDzdi|3MNKL4Isvm`
zKk|IVBA+KzDKBa=W%dyQ7>ZfBq~PM;2dDo~4|Ess;tnN6l|z9%?O8ceJ{ls-I&yi{
z*>5_-iE~i`Plc~UW$@X%gGjv8%+%Z5Z~7C;G~4)72;F3^y<9EN2?mGw`~Vtgqk4p)
zghfVxpX@i&m)eK>6J$Tp44303+W?{LI&*Ns-`kVci%2(Abksg}@b_;9nJ7!VKmy$X
zx7QOC_jWzinygw|D{5q|)in&r?35F@m#%_|p>&rUBAuEoIw{KdF%Iwq@iGXn^(ayb
zbGWrvV@y*4;*nWKP>{4d^k_@@L;1NH<Q+rzoL+*@ICJ`3ian^Ip{eJ{ma)|Nw?^p-
z(M;3#JqLi@L{d!wTbAR{l2?{5J7lDFe401k*)U7xG*nyXS&9ch2pvmM_0!rrQVrM*
zS%lz0ToYl@!6-p9QOgYx5%BuQGL00B99Ui$Qd~hs5pazQ^Goa^@vKU}_}zB2dgjqB
zieGy>7TALy@3F({`wOwwvZ8~-D3_9JchzC&GtiY^IQXT`rNnfmM(1?bU{9%tphCUI
zs2K5=!~TX)#W{!BECxV@)p_xPh9LS7<MAIrD8S6K)92@#H=l`@+xN;_X3hC6fJCnt
z)g8v_+>U!I@CDkCy~#=?S^LBJCr0ys49qdG>=f8+6{B~HsU|+ZA;t?P+Z4yJlbD@5
zM(CV!pJ8nRK+-MuAu_=7gFOJuw0$AFox6FQhfZqsNBWRESu3~sylsWyW^cQh_;@ru
zEsa&Js+-tjf1yewfU~E|&$wlTvnk63{T1kY+IP3heQk?OMjI4E<peT1%gVhXP4ERV
zKbp72GaSHpBT7~rA2fsem9?5X*V7n{;ub#*eoN)BQly1Qh-XWd-V-OCrwq9U$!193
z+z|o`3N9?X^?wagIcI@68YLpwZY>BbH?Aof4C11WdmNj*;>%DPnY>zMX@g#iYfA&L
z?A#IMU5s2DpFgoS^rKV8;OD9oyXS7cwX1WrmTQS^>;>Zt>-pO4-;LwSPLYv-Tz`q5
zKF>y;Dpbg!bNoZZZ}(cgx$Wkg^xL2BRU?V{LM`muvjHN}6^OM~0f}yIVY~F-^WUPO
zF%VK)fy>ji<2&s4GOw!{9kl92R)gKH?+$NgO3q)-vam5EB%}wHD`R%)RaV7sNWr4g
zwtl(%L@R8Ei#mPs+QjI~{rnegxSrENmO{k<JtSSmsTNdQKeu;Z4HoX(TOlm=h$Hxw
zjsRJWXLH>ZFJU|-#`pB37!|GOOQ5kFrId%(R=ulA1|S((I}xPY?caTs!RPK44^&dn
z6sx-S=ZHs~-!0HgPa}Z|*i800v>p+z=FH7BidA<C^3p%oS<u!xS-N}5rkxghyr5dT
zuUNuQ`d8siwgf%|C|btOmgi!T=x-%jOH=jnLW-{(UQqtrmp+;s(sN9f{qw0cD6y)M
z=CF4Bzzj^d;{p>|FDUbJYTLjc9AVS)+|TcMHp(iOquF4^a6{1S+_Tff`g}1lyESfn
zA<$I@XH8#6_`zbS=El19Vw@+5-#&i{hZsN9I}&e`r42<8yWH4V_oiIS<>c|^>`boS
znfd?b0z5$J1lgch;Hmbu8}w;I!%QOE++4_e#S#9KslPKB@YqD#X%fYU8?h%pJ4wer
zXQDU0c#iS$)UyWhc2<eV1~is1K<=rWzF%yVU3lHmkuwdf;pj^qs?y(EbYy60++B#C
zJw`kwJ6LX~NcGb+jojZo9j|_0c%AHT1+r!?Y&yB|82&M^+3?6XS=#;(!K4<L<TQx!
zP;AA>ds1H{9Gv&%_fXaE=%t~bQ(q)`_Blle1_x%JjK^U__8_pF#CRqeJ+9~d;8{K6
zd39*y<r=EJAn^MR!~mk@|HL<Nk4eLF`HdJc<Uqq0-wkxR6=J%8aes5dSWQ2{1TgTJ
z9!;lG;ZY<}oa?Rx0?!XvD&^lJH0tQ6<sJTKymk~-<bRNAXnQbQweR5YJ}vw)kzc>p
zg+><SZ59=X>kom~vlC9A(nQRIaLV@#_8Mgp)AiSJac~YeB0iCG;fXsD+u5xo((mlx
z8$or~(qG6s?BR(C5e;kzTWoCl7F=B+#~SS91aEQ&46ppO$k+U%z;bWZ8G^|1)O8rP
zn42<cQNaDq<L@&;D{zThL=M490q!AYLO(j#Lq(MSMx>z$i{Zc}t=_d}Hvj_AmlvA?
z@-vgIuB_o+=j)A!(@u?5-rg1~)!0~AG<w~)XV`Y5_A$C2d>K2}CrXIr<~Pq}a|?ht
z*;TOnrmMx%k#ni~Ju9)h1!cZm>apX|fmjCW9Tg4D=Cssuv$Qks;uD6MlEE+49SQOI
zzQ2c`m7OG6YN`*fic}{c_N0LP^L!vZX#atynX}uxV7IF^n#WcH3*5)vNCV8>NE9l}
zJJ^R2Ketwc{6wK_YnL71UI)R@3nXQo6b#rD3^@~Fiz)xXX@9Pww`XnCy)w$1xsXsd
z6J^fcKbKS883W%|o3((P()*~+Cpbj912&|q<5dC~{KY-`$IG~=PY*7JKv~6mhCLR_
z48quQo%MFLx;u$l%T*1_R90tVue-lA6(eRd%`%&J>-SGi$-4k0{zbA*AbzD-*}P)d
zdiCy|`=Yy6;!$*lE_zm-seOp*&cBilg?K8zUqePaBtTNX99H+v<SW1$q5sJJKpIOT
zpMN$D{8LPYzFf8$9Y2|nxB$+pMs=?aOMXw9fz5-rS&7uWGlAtDTO-CNdwW6m46OvV
z1bhVCN7I8ibWsJ{Hr8!!SKZS)BUPp|MkpUxa`oii2>wax=W;m2izMVz*#s&=>|B~Y
zuo`7cflKBJa*YT0!!o@t8RTcLr|ZoYT2@+lOEnwNqIe&^RIx#TjMbl@0t+DjqLr>g
zzW`{?0A*qSoI!|RDQOWt;(&-wC@+FE@5uU(3v%77@ruJB+Ip0)-}{z(I@ZKkf+PZN
z0bg=t6^o%U7QAJC)rosfK?eiX-R74G7K8DMr%+6RB?G}i-dwOYNYl`$5Si|~Q8R)7
z*@ki>gcMgW0BN?VAhOaFrf*o;6h6u@+VlQuKnT)E$LgmhBXMxl^Aie>+^?sOt2)$*
zv*w3_=L-^S%DX#A-|hM{wflR6uQD|^Y9XPQPg1@jL1-mljl8HfZA*v!Nn+<!Tf|*p
z)(nunjtNv5$#;R|hCqI9SBC)rm(tfA5XN-X#{2p;&g0zG--3%db0Bp5oB*rMzUrfi
z@Iy~8yDbfx`2#2>=$odbA{nU4H3;{(%w*qf2G*|3_JOomrv-L7IoJb8o<L?Ndomp_
z0zwDA$LzSq=WvD$J~RLglBSvf$ZYW|06oe82-`tan%MDm@Q-Ii=QtFM<mXm?2*kJo
z9zhrAFg-;*Z5Z_=Z1>z5W_8v404Z-l>j^^Z>!Ofgb~*jAZ#@1h>FNHYLn7$<$aVqc
z+C_gTmWnIIf~07pm>~nDL)(-0`2m^Ehe9UlRHtIX9GG!@!2ajbv{b95ez08&sF;~Q
z1erZX@>oe2v^8tT)BCBVbJwau7(001zNA}AQMd~({^${SV$N>X!hf+%<a7bQA$E3i
zC1aq_A6^+w+PM`waH)aia)zmZQc+Ou$`IZ8{35@s+b+rxJC`3}4pJosgQLPy567ld
zDm0Y$g62wgqETdE2kkBlK`>5m3xp)NiQ2-~vM}=>n%n1H2Y~q2dU8_IEuqiP6pRQI
zt((7969;ply#=0oy5OPJ1t$*zG$RsWZEc2EaDJ1u*YJ&OcZ#*e%kN-e4O`L+S|2%t
zg@u8kQQxWzhQL5z5(}s!`FUE&eTsZ4VPr&&M$8+0v|PVwqWF%%+L`u13<n*+>;^vo
zl0d8;8*mzJ>EK$NjyoA49EG50t3xn)+_UO~%H-&p*#8Ppilz6qN`Np<2<=IGganBv
ziojzvnz>3vzav=Z#Mbiu8t@ZO+Jg3UTwsCR;BZarHkq+xhm8%`&sxn&c`;q>l8haP
zEYFAPHK(V+nET#;B5uz|7{iX94{Iiv-7ZZ&Yq#Q+Y1Rp^1ARs(eHPV|0oY`0v0eq2
zBOWf8%2B?@15}|f<Irm;-9D*JFOn2}s4?0c$UGDXEL*Pa|Hj}(2go9zOsrWQJi`df
zM6iUt))F&kc7osrPT%qDL^}Y6Z_|(xaO{f%g|h8(-=4kmF<$xQ(wRycLoEv9V^}>C
zl1iBjuyQ28H`J#-R<V59Hk0u9zA^;BDR0<DY1tuio+c&vxL*VxV45ip-qrnbU;O#g
z<<AwE+38TtgUc8#fnGE4r2P>GkfOpz%*{;`a5)h%5$GKd0~FBH|30A}2Sl|cE80|Q
zEpEx!YU`0=dD)HFh@k@llaw6wuQZ39QAw_&ks3#ndC=|o(;&*xU+{rY-N6z2LDj%Z
zpl^G3H`_KR5Rc!Z5ubc~)$z=K^#1Py6sS9*S#$a0=!l-1o1g7ehVgyv32D@GSAE5E
zV%17&snw9xXyS2lr#%x~_6>OyhV&mnsWg+UQ;p7s3$@YYtC2**YhEkOZ$O7~uwt-9
zCB6X3Vq}5TvHmc1fH*e5*w;yzw%fjB-ECGSabY7{W2b4aP43FRvU9_M;e2IcOdWTB
zY{x!dMcpe7vkt^wy{(25wFsJrB{@;Go*PT)v;JS;a1Fgf9Do~6RaI4AWM%mD_T3`f
zV@Y|*$tfuK+9J`F0_eKVN(~%zn52nJ0G>Q8t}yiD#JWNrpBljQ#u&4aZSO%}U)Z;$
zA^y#R!t5gkmY*<A7|r@e_xxRby8#Q6I3lpV`!I0c68sl%Hu*y;_Xf6{QGj3E8(e-~
zUyLjCNN0lMW)Zw7Is8E?u<zpjS|-@b$mb(^XBmJYEEgZcse4^OV_~4DGd7uE5)>^t
zd%FG!6!L%ag$YI4REok%$N=yflKm_u)aT%1S~@bVD4MTsB+j`%wWwrHq>C7qu=s@l
zbI?LfbN~0JytiZ|+CV`*0P>XBnZGDSqG*v=)6?^lkpWpEdg^FQQyyA7?@1L=w7i9|
zw!VUt!=SGxW*9R1dEiKE{q<)#bmUZ}&Hq&~U}FDqskHi?wlwu1-5fP0s{E5$)S8uP
zdKzVz3z08<5k>1ndA#7d9x7i{i`Xe4n-e*B0t*xt8Z_ZM_7bFEkQtri${KNiG=_0u
zwxiueU<Y9o?Lna-)V*_o<<ETt$$<NCpc_3nU?My?#K@CWcUR=U)MkPZl6%iimcj0@
zx@7;6NHFf*InQ;xIj6%v<w)+FMp8QCzj11}eZfX%#pcJlwo<Xf-yCE(i}`mHP;Azr
zQDOG4jd+QGB`~VJ5@ktR9soQjBdsTMgz)W~cQG(VFpc%R6N!M#(FF-CFYck(D476^
zjCXTM+HZ(COm2$xPHzQKdMu^-Q}QQatn4OGt_X+E>Ri){@B8u$f5p<PTujYR{@_7t
z`qNNHEXjnH6>y;yw+ks?5>l**aw30)P>Xtk9MY*y9wT95TuVR0+$62@eE}zRu>spC
z*2AMb=R2Zj=T`c9<W*JL;Jnm;Eck74XIZS>eGF`kqBT*57V#ptz@wDGMurdTvW{}M
z50mklwBd1pH=Iyb)~jL~sWa6m<rNIp_8@&=IpQx>$24&wkma?7-HzA@2k797I-E)a
zT%lok1)IYo?%wOb@)1nwwK@y8tndL~5!n~PLgec+{faqQd9zkghN~%Z*TQxv$~WcI
zqPhAl)cz>ceCAH>-c5JxB>zdHLKB|p_zi`C06I|0A|nEhOGi$yKf`N&$_1qiFhgSq
zJD-Q4?@%gB8QC59W|Ya%#fi>l=@T4~zYglJf_1R6S6=DAYzXwSdsJZeT3?SM)#d?+
z6(<MvK;E!}@irCqD1S?~UV(^#umKVQbE<&FDT@z~kdSP+9;4mpt>>#pAY6vD12HrL
z0u^68(Q-)y12Yto#M{^1^%%WCmo;wjC+QbzV7YXVa(?sB<9TDXM-$hyyS>%<!hCbO
z)cuQKPq~6eN;^43z*0!RJ_0W^j7{nXCH$hdQ*DHDo!oAN6M4Dxrn52U{A(mIETdQ4
zFdDzkfLsAlhm6<No1U_TdBWnf@gw!6LF&%E1AW*)5#leoCtkE4AAM3`27m7upyz;5
zwRE=rhL{TzGUztg+k^&>$DmODkGY<t<;JhMANS4FfufjO(YF9>wLw+tH^n(yqpRe%
z-rgsfL^kCWG&^b+{s^y<T|X*j_SgC3HPCnajRfjTnVD5Rkil{o?A`wsad>aRQOFkg
zAgNxY=Cu4y8Bq?h`=x=M9o}{Kvb(w!<Z1K+;G225Io12KsBSit${#%l|58~m2LW&%
zk=@~@sdMs5O`EAWpMeFAza&j|Vpk0aJ@pLlfYwHWnAyoxU>adNZE#&jhJi!LP#m`5
zIS8g<BO_1ahnzSjj!inIQunSWWX1BX=^X#V)+d_{O6>&+dE(H#+2e4-wuB2=5%2}7
zR;I=N9gqkUgx5^0yYD7v{R$8QXk-R=_A4YuzbtLmiAF-nZ0S<Nk<L(*L9R%On{@N!
zjD$afSicwH=yVf+x-L-XZZ16ooM<cx#!xFuo3uV(CKfjmWW#io@b3eC6_>L1G)~ln
z%(2wmL`D>ar}`4GM}X)N_E!(lmyNIT!l^jQ&L$XDRkeSe7|#w2o&3e}6;M)G^SsG{
z<(YI;6HUYw!Vx3_+6e8l<Y3@EFZ_rQ;vpBAC!QCvLtvpkK})s@ijhNT)`X!{j&q_p
z{}2J`EhMI)u|H~45cU%pDKnd|HkA9Yh&5nK*mUxJ<g%7fBEtc$m}bgU%5(NfLog`t
zXpTeG;Qw&-o&ilYK^tg95fPCp5vfLsbm_f_L3)$k0vdYn5ITw)2n0fnbWn;^>Alzp
zz4xkgiGUD_^m};U_q+G|?*1@%IONRE&d$y}^XwjN$UCEa;C+2-dfM5aZfB17VanQl
zebczhp9fybkD|64z+qZ>H=h8Y$*Xv^zZweq7LsyyPFiaz&rSUH2QuE<=_0Ocl(5{a
zR8z(Kj$&>~Bv0THTBzsv?oe4DZ=(bn%{hVDEqz;jEl`M08zOn}V1xJDZfo@s3RxHx
z_XN{*zsow7hF<4kdCuyOTxYDT)pCHzTLmxvkp#~Y#D7(|n|QDF(fd&3jceDgEq-mS
z8a-)#DZ@CR_VjF;oN}E7Ir#|-L6oc|I9St-e;#R+9M0&MjR80sE>6!1B8}EK4*AwE
zNF7Pu&;lKtpE<ZJ>_W0wU(uCbfl3>6NTr){f<mP>-3+cI)uEZ=&D5r;m6^EK;7Pb0
zVA$4WUc_}CvR$>e8t3j+s`KT+)2MeP`EaCC?(<!+k*s<_d~-cHuaZ_!X>xyfp3%b}
zyOfyMIxaKbtob07PIbvY6d-M)mi_#017LsRM^_;WU<;wT$*x)KW_l~6ob%F~8DJOs
zBz^gPq}ttVx9RCA5qH}efwxj1Z9%TWYkb@T4G2$Oip1+l@dxJ0OckqKo5nDf@(;{N
zTQ`Vr3!va4Xi)mpVM+C9p$S(_(|l!XWrX)isc;x(_1Vm$Xvg6_z+YUp*Nj<J<>L_#
z8N0p?`Y^SFLDnf3t^lM86tL|Tf!s8P-M;b@W7@Ek?#E(8TuHyrD3u$U*pg*s`DiWE
z>mRDF?=9-1ObSNPR4jJ#yTzu+{jw{ApM!WfOU<8|*u7Xko_JFh8;@c{nOq|YjJqT&
z0IhK2E>KQT2FVfskqrwOkBSBA(;lF0EP<wl$f{}Ty~@E$$=g<6fc4Un8<#NWYT{<5
zX6iIhu*v?x%K=ynPx!)fbu^J=+`ImFHeZ8baSQN5rLs#Jn=T;D-6s<(#vNw8rWSgN
zn;oX(@<QHohKi<_gABA?#UOL@<Q)9r-9|aU!w`?B1I4R?TNWwdxlcjo<W`OJy!Sxa
z6*+hh;JRIRMN5j=_(@9qfvxzF3^fawGf?epxobz{{<P`S$Cwu6cwit9u0f>QBsei{
z$MPV{tLr(WLj!~R+a)!l5l9mEgrfu6uoNBwFR;_;--Zbo)Vuz2z9b^w_*VOn@W*lv
zs0#pIr_;Q9kCu=t@6lBj%lx{-*ivx)3U-lW&Lw%D9iZu@{O9B!0YMWG_sPH5HdPzt
z#Sf)YW=XRED6F)(^T|`17u;c?pg6pGUx7!&k@ob8?X#)Cb(S|}1%|sHyBm;;jBWsf
zj_2slzrzo`qfWC(N&rm(`I~fBAD=Q4;9=jRt4VZ7!Whqk<53@0M1j4nUcW>Mfk<GZ
zxDkCL>_bW%2ekg^EbMEu5cB1D_CL5D*`1O;xZ^9%{wDF$KoVwEJ~X4RT}Qq^WH&Mg
z2RPVJ=6*f7$9wM@S|My~A3h@r0geaI3<UX^ORBW|w|8FoK$EO1kK9)|LH$l~oy&K9
zplV2M6=O;ka2>8<Buy7M6GDp!33}8!xNp*b0$_&Y5&mTDp`gEAVF?^imM0U3sYfdV
zTgt2Mfx`oy=Zl{5a9~u5+yQfVIHe>lcmL-Bkh+k9$`auBu67c;51tmd7v<nj`js~U
z)`+XtH?9l8Ua~)HqlW)zW*M^^^a)*<+a>9%c;)w>*U7&3e4xNedTRUP#iH^ZiH%Nw
zPz%T#G)UAgCL9-dtz8F1-sdg!V{cIMU!BkZI>7+n-^c0Ptpp@aIRN|O=U9%*Kjsgl
z6L%#GwADe6<@8-70R<nf>4Ly!Z(oWh%@jTWz;#E(S?7Zfjm!#spjR33UE7bh#Jq#o
z8kh(4<=I8afn;GtpAj(jz6Te{-EAK;GbR3CmIYV$$!DqDo3+#mJRqZEwI_fvS!b)z
zyvUpyS=)$W^a1wvvS}ijg($E62mu*ky&x`tKS{M+KmN5Vl__I=#rZM~fuU{2p_M5{
z&otd49Q%h7Vka5`mVC<2xi{|hfG|HDUR;mtbV{Y%<v@uMh*!b1pQ!!~JqsHVEbR|G
zJ4H2-Xn-iPn?wN8$5KgJa}&bEKBN>5%h&D>kP$!|H08$X(%369ig{<uRTK##VxC{T
z`{`5G;h#|~Z`sh1uJNW@(bsLkm7g52)+yw_#;RWm#B$PMYoKc&V}S5OUjFFPgv5b*
znT;TEyPa+89pFe1^2w<G?#{4RG$*loHT8Zy3uNe|SaRhLUzi0*2$U#M{RRl6{q=xY
z3nKtvpv|`Hjf@P&F%%`F81RpQ+YbMI!Db1Om9TcVZGVAJMuY#Cj%^EI@H!`UvhG~E
zk;^u5nqDi_cDbGv4FmfX{%hR);Vz26cZ@MzlCm_B$DrK4Ao6F=MkpB=dfouHaX&b!
zqg6OdoeYs|gr!|5k1`V15~Yg?jUB;uoCE)!cuX!9WpiAhb5nsJ=8}qoESv`9{EyY9
zM`qO%^cSNX-}9cGlY^u6k?Zftoo4Pk>a_$>BR`Yi;6vu(Vge#lyE10wCqmwcx7KUL
zO4_IyFn^D+Gsd}!D<%~3yDG0kNW}#Obwe7?*uuJ&I?%E(Duk@6UH&g2;QN#^mwo12
z1I8$7LXU0cByfR7N8HD2HA2(W?9~8;ub~^CUpniehu~TA%MzGP>px6NYu3ZXFr_W8
z{Gu}8+94%v;o#snxb`Oo(6q&vERfz!nu{(jE#CBh-32(2eopryN!U_q_(a1oA<5~t
zUDcF~%2Rzwk^8p^V5{#5n=*h96H7~ru=1_7fYwg2>lfKZ69081b+#jErw)kM`NB3q
z{G#4xu~D6PRVm><l_wL*i1-OaV34BX-nVbxme7Y~y9RqL`t&Z5)YIJwmF6oxBZgbG
zADfeQync}#<`aSqVEYe>92aS*7khvF3B6_EcO<DwliPkC^y=i2*akHq$N60%(*M(n
z!v7uT+ztNk=|M7ix7L0+z_f?Ef*1TTz=LZ?PP`<`D~qCQp`<P=swQQ-jcVJjW~w+o
zw^nyIsRbh&$?N}mK3zsvbkNPWX~m3;u4RiJ{$<^teaje~*P;M^nZH{S_`ErprIuZ{
z?+ZXT`whqk0fu^>B0Hg>T8Kr61(>hAVtQi*q&ttoU?6^yL68Cyh#pi+6gT0YnS($5
zQXe%E5jWwfLTXn5l;*}oTQZ7-G?81NkUu|YodG0Z(&XW%d~uF{WV!r@E91&(tfp?M
zti$z-<)Bw>pw4UVSHhTW|E!B?BBjYK1PXtV8)%jgf5LtL$XNV~T#uaX3fAI_gE>Vu
zeV9RB2KWCsAV6nqOeBA9hV_BOuK3?+2i0oS?A{0iDiT?>nTZ@wGY}8x77=S;La^}&
zLs2TW0~R)xkb{@La?8iZGnG<k%WE=7`hE3vE1))U^DUw9AC-t*YUWZw{-O;rh|nZ~
zTDzm?X4C*nn-*9fO!(1y6A?c?!N;O=@NM-{51ve13`Y2qwEKg^fvSOi-kkMqPQ=Z2
za9)N=Bz)-q?K72bOq^ZjhyET-{XEHlyTJhXqUFH(1wb?YYra0B_xT^@@yV#ZzuOoR
zx2cvn7=aXAySicB3M0#=3%mNJOq`HUwM0Xm3M>uLSKXOFCa(M-gaH4QKjDDyN9b~A
zjfgVe`0=cY8fFRCM3#Zxq^yaKN7&!Cia7GE*sie}vx2RKuR_XwWUCg}@KT3^5%sqK
z9-%cz{)%zGWI;DSFms#%U__v2+sU%Ne|$Ih%AWW_{~jB~|EF-@fp;kB5atgapLPPl
z0P6Bh9hAux+c6w*1N4VJte7|L2?z>`v!s>JA>F~|;Mp~OwN^^f%Xj*0DRJB8WdKTb
z6%^280P@7nm8}-)(EtU<xWZfAc&RjbT*8%TsR{h8ACJK^=c5tz094_!W;&5-53GXX
zpfbq`J%Kg6vrpTRSIgJIeV<J0AM!l^qdmK$1_!b<5(ma#YSY@w8ugRO@B_Bp@fq7N
zmHD?lpc3((HYx?_D)G-6nn>O!6Y^_Gg(CeE^(7NVQR1PvbiC*mkri^Kh0)QKF-$z@
zLthbO4IlJn`0F<YH?shrq<twj;!mXQH>aA20o!M?j=tKy0&1OzShrd~1~%~wb+=3A
zd&-qp@qF!d5tb>vvBs9`_(1OdXKetEGOFujVbnw>081EnUbDG-MVTfyDM(0>Mp4cu
zPJpvwG~%wQ39g(}<URm;&8Jej0Trdeh|a+Xi9%9AZJJy?O}DJwT+*FS9N%!OE$R8B
z&w(3KwDlYm|2zpM9*X$edD`+ErLAIb!Kb}@F(Xakt{(LwcH)bR78R8C>J62rap}pd
z-PE*fvd^g=d^)C0V!Ipm)u3OpdnGLUWwuJ3l|O|01~vODHsUT-1<$98m1gRwlh!Y(
zHGW&)61F4{&qEi)7f44&M)YJy=0?U0#^8(u2aJYZZ6If&!cEHAw=DRRF9%dyo2vw|
zdBqtM>b*nI9(#gkqIAN|5BDD;708;c5?;Nf8bU*>X{j=F#29eRHJTC|))xmbYbSwc
z1PDn0l?CS9LcQp4y$g-d)i)eKPI#zqD;naYw6J5YOo3E2T3GAuqC=V?3go&p)H?$X
zYc9{5YJ=`Kr^w>)$$SniL&o=Y7^}?(V9WLbo+aRh8C2oRpIic-a#WfekQ#o3(c@y{
zN2${AO`l%@x2N%6_RZZyJn!@r{S_u-+aw9mt(N)3Mzr(>W2>k63w=K~1qr-_quNyq
zY9_IIwYd{hxQYIA2Xee2QI8I>)!V0VS_vFB3@%ZFW^v?dORTmutNa3Ae-w3wTn~nN
zgQ+aEp(By@5AaYw?_QR-vGe_l*fCB*!KcPrCv$Zu1)KHxs6<06nzD)Fv1I(TW=vPb
zJ9OQkZ}RmsSx_*v#n5T}jnox*IQU5Gv>Kfdjq;?gxMnnM{m0xN^MtWVi_%RA>c*aX
zRceKZ+?u9*Dd2rYr%Kh77Sn{Z0S*Qo<x9%<>eWs?;i+-QMmh2+H)k?=eVi2OSp9s6
z9`kbljYpffwR$0Mkm1eRQ9*(-Mc)IMNn63CYZQgV##*hIJ43{<<#)NS!`Z3sYX4S)
zZm4@H2?&j$oNe-}q{2NS9G*QsDXcwpf>u*gWso6!i8IXgXoe~LG3?EqTD9FOoNJ?&
zzdRg=L~oZjP1qNl`wCPhgQxmYL7w%d76U|*Th?;skKelYItHBSK^irU@*kI3{InJj
z;(K0<{_5lsIT`$S5(N8VEL=3V2Vs~|6RplzdF(o4tp7h;fKUn$C>t)Zn$(v>YhraQ
za+~b@;udc>$=htS-8*W|7~hS*B4!#juS{jN6$E1?AM~wm7``!TG9JLxMIPS5`gagl
z95lMi#)_mlSzENhGY9CP+y%*)JBd&$^B>q0L$a`+jy`4ryFuQXFlew_$hcw2MajH<
zFY2Bfg+OjXoctc-5Prwk?++v;kjf%b)($d<>OFwqut_M>@Q~-}7QO>JkcmKy(O1H_
zQ_UK3-T0(o>TimoTcP;9BK_8v$SsyKLwQHr0Z{aQ@C~Pwc($6fy{wCEVEQlv(+32S
z>SfmT{W|98nC&UxDM~f>LOz8x*F~4=>JxojC;>7ua&u0LiNdJ(Ufn8rpq*1!I(fC`
za^vr1*D+zL;7+;iCq6Nxf!7v#M!ho4nw}t4EGS@7=B3G|lN5el*?f#uVv6`QdHu7~
z>?5JzZ>U7E8oj!EtKAxe;0z=)JAj6+vkC0FcjRnHt?mSria8p50!!iqi?s)9o~HI0
zY}8C=yfVa-dcp>Xa$A;_=EoI78aIqAxSY+O^)IE~lEf=GTm+Q*dXsscH~X}rsD_>}
z#SzSY6KWoppm^!)hF!l|GuALT2AuR6T9<HKJhlt=omx5xXYQ(ZPy30jiD?`zkrFSQ
za2(Dxd)=D5u<Y!-ik*1mS3EFzb6CB*{(NHu@0eE6NFlsDh4*-7<k8!6kh}T&=h@UB
zh@H<G_f6qJU><?O1n+z6?(1k>@iwon4PKb-pvk_7+^FxY7pOqvV>$w#rfTjtd%Xd_
zqi6ns6~94Obo$W2gL0j;XIiz)z9zFTExu9kytIKPdczp+YYB-`^bDdgkm88g^m!Km
z6XOg2P9@zLvSn00rp6Z)5r3mP8ZOtY*XqPXXkI0Fm$PbpCxTWZ8y-zf2s$aWjObPB
zsOP2#Ywgr)LU`d@!`I))B@pWKNIW?bJbFBRdyA*LDQkY*!*>-A@2PYI$Zu;v&A>?&
z*UdO{`Sy#yvzf*tZ|_4we@}itL^5~#Yo+PBiHLoiaICYj_KD_F6(=JBwluX2O$s3&
z4AV%0o3s4X|8BeDO87r-b@HgQSQ>Gz9|cQfo@^U8HGJTcdRLI<Vi!A8XDPiZ+u<l9
zvi`5e;1xf<ArelsuPuCC^46&Tt5QYJe{9?(*S3j$&x*eX3RTwk8Hn5{Lhf>uqQ{=C
z(=xG0ucq;Z{CWwkW~3^$b>L^qbt5l~?zl=eB`Ra`!*qDVWa-VhdFfecW|kOP!dTz;
zbZ*Mqdl0a;_w!h)=N3sH*Io+^|F-2XzHrQBR=Wb2CsTZ{mR&72>lijW)9Q3JFMPh$
zPVh}9^GiJ@LO&O-?Ym)UzM!m%xIJ6f(QC04nXZ><DM<wXd^68Xm&3j1o1IgF_-|tZ
zE{a>U0-h(JaiuyN#<_uk#KejY{+T(V*Tr8D#+*`Syc5PFJoEz%DaxsRuSSI2duL3V
z>uFj#4T1$)7TNeJe-)+ghm?um+t%({GBhKvEy^2veBn#yryScbDkxx{&3oo+!Swvr
z_E%S#*wICak0TMAw#!AiGcI-lGj&IoI@LN31g0|li3DQZOYPBvzsy;j9}8D}J}J$s
zCzG9<rL`MVj(Loq8o`+G)?OBz@saij$h6>%sPgX587-Y+9y|{Q^;WAA?o>R~E?%2)
z&Y7K~BY1LCxj6n3HSDt%*p=1CW2tH<Y&II=F}A6}nZw;mp8-33)r!4{M7mb?E6KOH
ze)c6>!q#!*+KYAzG<R-&t{{=J4R*MwZRyPVjV=EJxrkno&%DaV(S;}sJNj$>ZCS>t
z-gGV>Ry^m{hU!MhjNFlQQvR*fU$^kJw6kH}-hSHhV-FP4IA=~y4rz(2=*H(_Yze!$
zLI)5j-TFGXEdwjQdnXqEtrCoIY+s(Gp~p?%BDZg}H}6237%gR4C8v{kFK-;beFtY<
zQw1C|=SGbwUuG;gvgym{<JBXyN$Jb7O448-lE9GD$-BWm7QmmiYVvz*a5i_+RmgPg
z%Kh#jOm0Z15*m8Vi|*JY$~<_nxmtOkxSPRWm(k+9P*$C?&WorG4(*^<{>P{*o}d3H
zaa&HbGW;VeqqS5mXI?(mL164*r5Q7Tojo&UnsBNy!q=F>bIB?PJPey5-R<834j>ds
z8Ck@#?9P6G6D{J$p-P7Pi4|0@nRcLpI73R7l3z^6KrP=Ck0mThw{+e|BRoG<{tUb3
z_n`cG>H!<wtzmLnE>C_Y9BX0~nv958mtHc~tdV_f@|&KzC$oc?fya(u)%o-^>XvoT
zOITMutk^<u*coPi022f2Z@gHySbeiDVdf(IXiu#}<QS$EzcXYzRM`SX*u*Vkm0Jv(
z)ao9Y#4eZ6^uABzascu{rmE}nl5dfQjP3YmVA5?KkNFReJF#<b=lq#x?s8&E&-uL0
zXCG(Oq#a^ktaQ27UaIO+dcgWMy_TmUPrSBX#w46GgFx+AL1F9(R=j}nr=z!V^sU>v
zOlxBTIcSo5ZeHmogmkLHwa;wM<vpW>wnm>CcX5&qT3iW7EZ&%2Kb(+~xXVZvbP)|l
zmpdXRPmA&^^y6H=U|83dOcRAWvl%>VgTMN2yUK+n^D*TqAMQG=hN$pan}_c$MNtdg
zUP)(Lik*tWx~&(Tx>p(OO5H#9!UYYkJKtZsi%B<f%du#kkPBRNtx1?zQOftnsIu3x
zy?#6%To{r~QL_+?VDWWN|FCObdv&)rcYCIv`fyG3^PvdBck%pI&zsA>DKu$Npa<|M
z@QqfA*%#BK8u~aTjwD^&D|Ks=N~>@l1z8G?=YbaanRY85oai4$wi{PHaX&`^){z^T
z0d7b|X4YkGaNEoo$D29l{dve()%Y0EbmVq=a;<cyVfnXGJ|{ajtKDFr<V6J1@Du5K
z9P7o_LJw@&h9QyMWJYl4A={Edzp#+COjd()l$%@489ps^Ia0+$Qr-=9zq_^vHEl;l
zbD%X?jOTtb9o$KWa~P3prH~fY+K{?B>`T`ukFS}20B6<yCrFrHdH3UQGqGY0*|eR|
zaQHc-s2D=lu5;NL%}j{_`m6gg*a#nk{LH3R2ap=%(Rjq?+vPo6z^Q(R{&7%QUpZW}
z?iY`&$^o*`{eGYWT3z&eI{oOQyQnnx^rHk~+4>YF0!eS;VV+gIJ4E+shnmPc$&~1p
zPKX1~2A>wX(cAPpN<+QZ=Ty(If9TA-)?)s<T-G`U8t7y3y0d{|>;+u;WUl=1XZzj)
zApJp}%+SJB5NMs!k7R0@Z_-%qZ;0PhsX<Gq&wedYo+1+k|A0wXihw)urGb?c3+<np
zYc(c^$KQw+eXdhV;@M5TmBn*61&sI#JBXbCZ46zYO)NJkgbq2CHGWKQA=4jkFaO3;
zaOMU_u221R+<!}X#wZ1W<C-(!=jrfFX%(mUImN<SHCYA5GeScTAHXZEcHdjy^5o)K
zd#HH|gW&|XH~hD<!_bIM!O*^Gs+Bj)2CB+sWYuf~Wd{omoB5kFRtnSotqc@7Z6qMj
z;A2<vN>gsm8TatbQu<#Cp;VXtBo|GRR#3Xf0It6|O0c3HH+UBiBE@OZb4!!*^Q&{4
zifn7`)kGn4;uhmC=UFqpZSdCFASu`9t^UQZ<$Tg07H^GysrwH6x`|1@2{Y)Uo`W+k
zfl7GbcP3*gh#ju^2_@UfXTakYVr+#&PFrBhe@Hoh_RsaGZ$LSyn3k|^<0s?O(d3cq
z7Etf|RKdlsR+!nkJF!G<IMq!dm4cG=_a~D4Ylo%_ekF4~+G(hUu|DUsh3QOx4CH2-
z*5F4x(fgI*Es2j7!_$2$NPE$qriGdQka*62ma>f1Sm>T@7T~DLo|TVLSNZj(Q4UFO
z796~#L6lh!A~*bran0FaRd3nXQq2Ba4|4e;?RiEC^)_087L}RX_+$ku)oUH8uU`_@
zjttVNf>>5Ued<lm*r_(Pp|R9MV)-^_!c;D?Dfs@bsTa!5R%X_V&E|g=puV~jT#|3g
zuI$Op%RZlJyt={!m12jJpTjtF+ft{>xW(!q5FbKS6?m!x)f*8_cGR?gVDe2s%-#Yp
zOU4v}i<E=g>c%}IyQ1%t@17?!CA4%N20T2}vH(967-&9k0y?W1Bj|e&R-JLirmgr-
zc9GzZY2t%bRjNK6ANz4~fxt`^B0sYh!hLp{jGpG`;;#0(*95#x3VA+C_{%m91DCA<
z7I)xEdMz1hRYGc!y!@Rk^$yU(jSsNmzOo5NWtyR(!ZBT7cwx(GkkrwP7q?}YFjZj5
z7^GI(`WKBHJ)v;}u^3mduL9Qt_=@F$WKoOCSCq!P>`HRLK-WXi3q3`Xv*kSO&M|0n
zBY{7V9&`rw=12t~eurw)l=v>&KTe_XmFZrlHZ)CgwX~&E>8A;!*Bsp#ce#dp-RUbe
zeH=M1NP5HH+!NuhTe!BPzQ?N<8PHm5HXL?2%Y^$3B84oqOt?QEMmBzFGhkht2TA&|
zbk?6FVg!ldq@DGQHY-3Zb;iH86JP-+o^I|ZLJ$EnS{-nN)KYSWwW7ZX8R#Rw9nU~a
z2F-a>yI__IG>{S#(uPI4vO&b~qoRBtO&NdxRLaCL4KYw}OcF2jj8`<CTY!CLL)k_N
zM4t~8_ks0!i^W0g{Tfzx;Os3t^C@(On3%_60La;NWP!z+ekVwp7)aHCy*kY$0$9D|
zKjDF*518(fFV8Ztu5DeV_sbFz%Kx!)U5H=VL{j<CtE~so44W_x7b~w)$OJ;N<0Rrf
z7k6qh>)+}xw7ssjwQ34=Sfmv7j0HD$y19$*C2Ea(%)Lm}P-Hq}$|$l6MlZ1DmS<<4
z#Eat0WS?A`uyo}5oZ%aT$o19e-4eQ>HX2c*o1G2f3jK2=LR)e+)+Nn7a`%@!6O+8W
z<qw<Ead@AYJN5)AvSm^raTTC1GqkLXyr|uQxS9K`B$>fH#thP^hKA|Pv^nova^$u-
zT>xnX>`TU=nUpzhZHyu2ka|d}aOpf(Z@*-qf}3(9<|)noZ0Iec;tjeom#IU&f2C4K
z+LQ6TO8s>v^Im4=`hvjYz9b^gE%G?8!|>=PZF+hho@q{Nb9ALAdTT8th2Y;(Rzc(f
zvkRBG0s<f#u`81mlc8x^?{>TM_43nSO>ePCS>X0fHw7U9&WuP~^YDPoyQ5B`K4g9#
z1^l%v8;Hyb8A*^WGed<z%aFUOqDOC9p|<kL$M6KVmw(g*Q_`8da#43W8(jVy6h)!@
zi*Cf@&U*Il?xT#=-J*pR8CbbfNxQ_yD%wVQ7aRUd^;|YAB@k(K0O7khkowvfW04#M
zi3Rt9F60G-7(|WIuSsqvc|Ka6z5ZX%Na^)h?K<n(__~kK|I@Ff5B<*;&u7{G-oCWZ
z%hn;PLmeLS!%UYRNMZ%P^(<k>Qr5%{2zMY-L^b2o3pZ~&jxxmD`J%zPhQ1V@V18u{
z_%V(Pz*hu`g<R7#=tpJ`cX>POZMr*68cNy(>~1gZH+I~NPLNXHI+GlB4jF1yytFT%
zg?S8on5wFHbi4Wu+)$hmhr12?x~SL7noG7s56q1OtyQcxYpyKPzV^tw8H2<k_r3>f
zs^_1`ttf>EV~Ti{`pGDVlnp*Wy<JbSM4pCQ6Jr%|gwRe*l_vw0EaevO!?pFy)~Q7n
z4P!}g+)knI`w!9Yec^r((_2EBPx7PR7LN%7vBEo2giV)(vv3Gr$@kZ^(*FJZr|@uK
z>0s_*)9X+4luwBx{05+oeIQ#ms$&zn9F0x{v!E8YD#b~hN<NLJ%(Z{}gZ00a)Bl*E
zkmc|r*IR+O3V}%3$aQkaaa2n>&d8_rDfC2R5WLST*(A1H(%o$8EWe!IpZXbU&YY9Z
z$v!(mW6+Nf;vaEU=;*43EJ;5bUhOhQxYV1<mW+S}LA?`S1K;JIG>}9k{^M6VR5S8Y
zf;;ihJ)4Kff1eMcxl*5YH8>nt{88&<Cvwm|R;oU>+!UAHU3ZA-94b9(+A&t4f5Yj*
zv$H|^+j3CR+{noNwSEj_X+7jz)_PV>QYh0)0Pr4WBxk0$_*_rs>XU!Hg2yvEF89!o
z>%0GTRfG12>+S!Z`rn5Y5s^CN|6U3sy^#q0zn4HXln*2Sf0sB|b*O>a^}m~jLN?of
zCiy>?v0;F{|IdH^{d{x<2z>tUB|r*xs{eJ#iq*ZBLV9)TIP%_AP}l+e%Ul2JNp~3;
z2}004owqleOTb*E?jF4;%At@0f}y}J$*dv>ge%#6SxX^+z`W;dR=5uOzk=Hgss2c1
z@{*!@42Z?5$1I=;jisbYgkA;xpFZEB6QCRt*Hi0-?#=!1Gi^kIbpY_{k;ng$JnYIB
zRI?&fnLVj~-kw)z*tuVq8iD!d4Rj<-NSh_`f64q$-}dhQA6c+r603PkgrDdmH}p{!
z`iK}8)Cxa;3HMc{_UP4PBsfC?9{vsVI}@nYRG09rgqKOLsxuOn&?bxo4P4NB__$2K
z;(lj$wvEhKNck^p=x^+{PjC5T>)8t&?<)4tz1L1+wTX#v3wN3hZ=Dxp*}yI&c(fuk
zYE!q*xEv19U2eY;y>Zj$mepLZ4y9LV7zy$v%$dHuMvg6#s|dcnufAn(y>?C8yF+z#
z$_1oW&GycV#P7LvN%)$&_l^8+V`ual5nQm@&pBdT#Dcj!g4Lx~Eu~iNm<V|E5j`&G
z^>}e|U3&G+jN(;nxks<|lx(Sue4n7UW?=HmC2`+gPN~%sMuIo)G<&=pKHu2hP#(RU
zlHjVwx)u-=+A@Tt8^8&=ZDB+4SQd}oDve~Y#Oj5rq_TIA_cnjVApg3r>F;_QM#6LS
z(e*&B=HLy2e}il}!!NHyPKv>Ml)V3%6aA#R`XX5$EAYZ)mM{_T7kn<c;c6PDz%TDI
zFcQw8-U3uiDo`hyp_#~a7icv(RR$F5R|Lt?d+pV`Ml}=z)&D2fBZU+h-^%NPepo6o
zV<LFA&?4dI*Qw^9&_L>;6Qq5j;#a|KMQAL|5Gizz6k_@sxyqUAM>+HcDxv|Es%SGC
zmTc~9i(J2V(x4bMY^~P6jkNhqC=x3}P(T`ABlo;air@Ot4CO;VL{-U4t}5MJt+_+>
zaS?9}ndUAuk`3W&X#+gFo?7DO1E{w&6%!P?U3WYfbkEe*pp&-q$&&k722$<`bbGW7
zxo*L+#6peXhL7`^=0_7Wo~_)WlFbgFgG@&t>8b)vds?Nb&d|u)+LEn{SZ&T+r^t2G
zCECrW0zJMJl#JvT$=m$0nOw8%^YIzsE~GIU>C2G|EYd+7JwPDK!bsT3kT3#@5(&sH
zj$8n;v2q!V19V^iu7-Z%ow)*7K?4Hnpsz^zW7DFt^)0rweMUjyC>=~MPGe<t`uysc
zz0yP}Y?-2P_2LKCr@fY$s+1*&pKkLIlF(6)%#L+<PJQSCTh4o2s)suj44i5nMww5M
zDE;GA*6Gnl?UUd3Eu`+dkab_AAm<%v#vSP^FyN!WxP!-#<9y9UX6K8K@hjpLyBDGd
zCrk%Nv{Z9e34aA<mI-3{4^78tX`KM<I*>JX45?CTq#x-;N_F-c2@_`@jAqT?9vp47
zrK7SazbOI<n5N|^@;9ENy;@MLlnj=X|7l({I`z4=ICl*_c~0>?1TWL~i}a{EkIl1}
zg)=vcT6ZtxF9qcI0hO#1Vp|28Ml;k!yql^9wY%okTjbq5y5S$9m+g`hbsg7S4DPB5
zdG`8qY^41ay>xT<AueUJ?GADsoG+omGeUv14?Q`$L6xBj6`&r<h3-)m^B9QydXc=}
z(%JI*wRozYCKYYYKu6Z^;{`jP#>6@49nYfE0kJ$(?lwGS#;4rnf?}#FXIzBk52R}P
zp%cAG{!}YeWNs@au!NWUkit~X(vYlul5tUdlVIh<qQCW)J^iqMy+@_0p<s^3x;wW=
zGbBNSD{sH1Le?@{{U7&jHE8_J5EDm^&0#wp0O^tRF`0B-ZNK(2)LBvCKbJ^zy((Km
zq<;P?a{Zye?rQFRTu_82%+0x*$|tbqdG!KHCjZyiw3H6@+=>#q8}m~Vx+iK-2}LLY
zk;zxUxD%Ye1S8U%<vx)3RAzE1*B(MbyHpCw_T{vpZTpasC26KAQ?t`qj;2_zUU2HZ
zj!NO^J9X7%tWgcAhmg|OUD9MFs*P^EszCgDEk7#Yea^~>YBsBs7XBORZ#FLp_KRB)
z7eC-w9mBn&`UHv<H$z)G=}9Z6CTTh4J$gfTAp_R59`bsOuA55yw|D=3<cglx@0MKE
zZ0^jW=@%8%p_F}OUK&yeE0(>ZUiG#`TEO~d!G2Ho(C!qE4&}!e9&QFDuMcavUW$=}
z-LC(pd5F}%-}oNfz$;;xHA2=)#2sMU_AX9aTw-!-Li^i>y2XY=L#TDz+diT{nz>>M
z78I%R>q$hLU)~%bK7k$bJ5BFIWQ_tRkDk%o2BHJNU3HMC7CXWe=3dfC^upwJYu{1w
zUwXK9g=ge+=!;1dBfXY$AeBL?vt2ih5Lwpj<U_>zwNKy%{xa5Q8#Sro?7a)W%LVpy
zJ~b^rL{|R>2V(RRXrZ+iejG+Z#DmGCU9N(}H;4tzAdqc-w`+8;9;wy5FHkgs-K%Z9
zsJNl#DvICYVU1fMYo!ySix#SVt@{I8t0ok~P;CKxPx69KzPEl}FO0am-VC@$W-FG9
zT%Jz8b8!q>sh8K2>_LoK@qthU4}_Ezc=+8+<O(S}>T5Cow${?Ax)W#?T0HTx-kJC5
z(Bq0F0Tff9mxgX#3Opoms<IYoYPA{nVUPg9rQ*eHfHd4RZOybd6ZK%2+8;v?Tc&kN
zQvd@=9-xeMA(PMdr8~VDmRW3goDKg#P`R#;$<}-Nbct(W4N`Z#MtzDKovK~*f!zT>
z3w!)ou<bi<(H)kw^~_gbgdkSMh#J1>O<mP|RBF@UANztvc2UU`y-D-HyRJ?(sK8<&
zZCn6Ac7=fB>Bk03fe7}?#m+)$vSj+8yol;^JDah?9+nR_i&%r71cws@kOK4IKYShA
zh&k|SO5Z)}L#3xxZXbR=tf4+|d%m@v2w?IL=+`7U{$?8Gv*lm!u9XkOSSHE9l8~sG
zLo>}TiJMi)QRd*H7aL1_i8Xgxx0NbOe7_uP6`_PQgi>C<H{T7P2kHu!W+2gyN;m&z
z2b?c{*{l?WC9xFE9VnTs`_T(C5D}pXFQ%`80<H-?+k!pVLJcR9HRG<u74L*${+Zq=
zYfZiSX3_7{wt0zP4H~LkUb5@b#;TNRW%~3U8k>L6>|-~DT6>JfH!Wjqy+}`vO~YkG
z==u)z4qSk<fKJW}hziiW&&J(*pTL$axqk6$9a%0QE`DJxAAynEYAV#+qe4B)N<d@*
zFf7`9ZYIj(%@H2S@i(#xG>Fq=-(n!@vC=uO2}$fV%+ZgYnmn^e{7GN@7KleeB)ho=
z*#Tq)^bl%AijhR>s7JE8<X+4^V!`!p42%j|7BE%AL&gLbiXMwF5nfg-iELI}#vU?y
zB%0`>rX6ptHGSbt7KzmprprL1apisZ&0_SCiwSOmAtsi1q9%H!p$sn9Cn&#;ReGf-
zwaUXt5E1s;x&b5;q`07Hcq{O$ZEVZ!XhwqN?^L-sugxs)-ZsD9SnMp!v)4~*^?l#A
z$9To-jf|V4!7>6kner|*n}h1EKf1V6Lcsf-t%#)pW2>{bFWdYmc4KO*XwY+mf?4(x
zXHjEqnFL>4P#JuFqW9nmh@CX^jO(De^xrtHW=->O6CB6^*0cP9J^~`R`WoO3;^x~}
z3ZLF>F4LQvR~ZSJ=%d*2H@%-~x9+=8FcOTuh|N!DOg)y#HC2=rGLp{U4Eb~65J`OG
zHXT#i(lBsbJyDXSV|o70)ii!u*CnlF`=n{5K<Ky#o>|UtSOlN%614n*&2sOpVkGFH
zj~FCz8f4uW4h}0=mRfI$wG{(5k?~6W1*VH2>eM8P-B+eYj|CjJ!drU;+kRt3+qcog
zWe8?G{I*}O(Nx9ACi8Hhw=#h<D<wAh<bL<BB}T$M1S0A2T$02p70<SJaUHINxy?1f
z-DCwelbK|n{-?4_XXU(AD_QjCY=Ve?E&MzqDfCU3T2E+|taz4B@1~X<VAy{Jc&;zF
z^~Ix1^=O9F4+t$8cS9vsH<<|J_w;a!rR$1yEYw4o+D&RU%MT>`2LrjJOjX?qON&J>
zHUnzf#o4DZpXo4$2kUbIXm$)7_Wi}Cj@5nI;Esh-167eZ)T~kUtRTeC=FlG_?l~-{
z0FAG7+y4<LY9k|S$KW6~P+d(hElCw&?&H#1-g3)H7qUn<%4fYkDowN74PUXkur!ls
z{o`ftdh>x{r`EJ!)cDoBQ6+lFgiqE3Nbup`S1sGl**Cv4TWm*=s6jyO#yK=XCo@Ia
zsin;geTEN8oI|Wd-8}R_AjZF{(8u})IfP%pLYBZn@=kj+*4Q1_G%7DZ6|7Eh?`_Jr
zcng>N)=w~D=$*5_LQykf_I1%s<|9q${?i9sSgUj&@0Z=zUo%h+!N)(Ija+vwpdOk-
zBT&B9W^N+fQWa9e&PCd*!7h*dvdHPSMHbouiyW%p^Y?%HXGj<bZ(2H0YLWdZuxXvU
zUSc`6jXrY3HEU9AUoAZ;p9~}1WzbpQFKn|Qe_D@MsX_o06=aiI;Ag|D7Br`aAI`w8
zUe}GqOak9q)E!sD2FomM{9}nzMnQ1^P`J`Ua|N^%KgLM-e!%Ny!}G<pa!vYHL{rm|
z_(erc4#kx4QFaaf-l;xk@c@u)@Ys?%R`ul<Y7+|h1oqj_b8B`cOaj|-h*!za0_C4?
zeor_!`5fW9tZf4ikX)Vf#$2VcaN6vPI~u(}S+q4S)$4G!;A5Ur8Qtu`e{0xB+#5XL
z!k=)MrR7BxXnSuw_3UjbgaD*=M@#?mu4k6nvEIEYKs-`!mRJU)b6Qnm(v9Z~E@Sw&
zo8K=g0WOMp%h2F!R=erDiYH&!N@AxpjBKS=oi<`RcIj=#s*Nazn8zpPoe#^dTg`VF
z`V|quvjW~@3V2f!Cvd+|Kys(4C~p)645{1??N4@_3Ow6=<Jxl1Y1oyggN1?X6^Yz0
zdpG-N6J~sT7y6w9uWpG+h8AX2=BHa4Md6p1-Wj?HyTq^D3*^rztv**uZejqYL>_4p
zSc#S`Cr`IWjR9X*CTQe{H%YtT!3$8#SidJ{_&Qe9*ecB_!OOMfSfHx+GQx(%vsZcg
z<gY77f;`p}-(>2_6ZoL~H?U4=j*GWhE=+B1&utRAemaU)hDFkG@UigS^6A~kbtX-d
zc1S8TZ-85PDIGrR-qNY?wCnbla_PCJy)ve}%q-xbCCxh`%HHke)i0TF1Lsltt&5Am
z_-4Rrm1s@trxGw5YKkrE_i|BhTwJ%!(+W;LwQLs(nSd4Mequun-9&sJKB6Nj6$(6X
zcAoqc^WLTK**hlZ9@>W59JQ71A6N%2hukxx)nx;UVXdw%^|}t7JRvvdi<?i9y1i;I
zVt|`^AG(ynvS`&0n>SDVruT0faAz?W>F1Ac>?3FnGy#6fK`p^v&WLkGw|GLHj^{7~
z_|sj|qciQei5}obqSARw;KS`zkeI5f9(|F%oV5N&4XsH(L!EIOnuBdQ{g0<S_4zJl
zrH2q7mV_)<Wv3&!!<#qUNy5$-U0iD>9Gy$m=|Zx$zijj7Vv}8g6TNK)Tx*CQBk*%l
z%HduETl)<a<l~05+bW2IELF$Y7r*KNq2$E%Kk_D4hOPm(zgamP1iGcI)>!jO?00?i
zso^E7CiBn5N2RQuW?#F)XH(*n7u`B}l}d`KKALi=tDQuZqnjSG{K>Xo0S(k-#|F7E
zY<~S(K{&H5d}bn%{Pd*D$>k!C>qw)AL76R`@RJDf%f~e_ZngMvDxPm(ZPDQ^chcD_
zZzJo(kCOTyOE8t&On?$zKbFfH=q%7+NtqtFQ_{o^+yuTEy^wsjNdLC#-%t$jEM4h?
z(+_LTz*7)~&l(nn*Q;u6r`8@a!a06RcO1wTP9%7gI^8nhW^pqx1Wv5_Dn?aj+R~~r
z9oVjux0`3Ra7eZ%DRBMQ8fue=;OYyD^L-obeR5}gK?p}fd@^i5>*E++Elpx`oPmyg
zHvHOCJNPXyn76A%nkKh76E@($GbhgFhe9qO_fBd(=5VIWp?q<MxNzYFYgzq|jFje1
zH5qFqdhjkl)R(MTy6{ddSD<5Aue?g<zKC8(NrarFS36yIDA$|}Rl@k41KLk3X^1p!
zkt<b^OP)w9X`V(r-E1`*o3AKC#Z<J7^#x>fGVn%%?k)uBT6|ju(1Ovssy`=SALztA
zfTH@fkl3|~_MnRb2LPS?o#>t5d}rffGm`n_1mGEUCz={{_+yrk>&gccBxLo5x~TNn
z01OGut~BQjL*R|M&H^!TSQgT^c1f$&>b2^GSP$xS7<I~ix*%BWw#?+Re_RV&&WV6)
zi}RErpO|tvq0%>%4p1bm;%(n7Ydu3tGH4`h$sr9N&YCAt9~yfJ5N!k|%X{1N%H4;|
z-Clh^up@?;Eo!&eMSrSbX6+iQdexJ$fP6>!l#1*!RYu40+tW|CGB67ZEh?!3`?b18
zmQZh<xB(zMx7i75>FlHXVR%x}!bq6$>Mso*zwK*|9Uy<4J%<=RZlxO>YtD*10oC34
zfXUkHV<dc;i?<<aJA_Pdo?bkty3SrjXExo<tQ}Zw0_94}pR0p?*idrSo;qwEGqXk<
zye3-d$1mAVT5j%njajk(`VKgvUtBApUjR13^NH4>zit4SI!fe#i*1=8N^5&QFg-Ez
z(2W|ie-PW!x#SR8?Fy7O^t=GDO(&V;<=(q1YfOaU@b6#W!<P5OKFChVUHnw)?^CHW
zWf@;z{&6BGCa0EbKebjLUv>2MJNzP6%`VHW<H5Z$Ll%c$+~;3W2Mtk;xIWb;#W$ix
z>sw{-cZ#O7Q*0VJ|F(vZ;lLrc;U_FAlNY9YdGq$x4H1mR4Q<aRW#%kk%gQh)e#hza
zjwN%L`dt+O3AGCHa_<#)ovxh)>|`)H7)+-T`1nt7pIG;9+K)inXqL3NAYS;z_N#3~
zM({n$sB?fiH{kU1Z|B*3AB@cjgwVH|mT8!&AR%&ekZ$f&Uk!28#Wh=-HWEby&z@$q
zT>~jZ7q_UKey=t7B8EutXC4=GhVwrVURv1;<9h`@0WydrB-iP7o!GA>NhN6<>XgsM
zlJoeT#DXRntBiT+yo=?Upm?e5^Lo=3j$bljf?m3RCM=sf3x9k|=D(+{5q0!YAR7Gc
zd3ajVKBE`F@MxCf_jH?nCU3XHnmef-5OWfNvcFmRmOZdz_HSO25&rzyNrNV4QiC12
ztH?qQ0<|cC_n^V&BrYBn_w-&}l&h;O0e!vs*52buQJOFwUZ*}I|0aUJwQ{z4GrdA!
znS%l6dXFmvsRDL3XHp0Ad7BRjx>0NFa(4LgesQn4$JcB8=>EKJ`lj%Kg$<c%B1K9h
z*R;}?*QQN={<GAa<`Jp)&RJhMy|O?EiV!lt+P&O1*@~`n%t<e$t|3>Nbzbw($i88g
zt77h{!+b522o!xsd-~gxdlxs(C#nx=RZ{3jjp3)!B54fRXTSNNv|76nJdH-(n6B@V
zCl|144f=>zHQ&biB%#H&!^;O8jonj%S>V44^Afek`3!|){Y;WaPosE}EH^e588fSl
z*&ClB$WT-sFT#JM=M7CN6B}{6q~8I}qZpHFDZ-dNjX!;P6!VVs0+)2sAhCMH=i2Kz
z&cWHz$zIc?%YeuNBM{MU=s#2PJP$0s&G~8u%#zaosyt5B+7rI-_6=6F6cI8k?a`VW
zT{%g9<K#k}EWd}Z32un5i7DB>Vd?muShW9z-CSDy_hRDXdjGSe%{30Qpa(101w2FF
zy<#Fn(G4gxG{cHxDj`uGO`6=ij)rD@+8W=H93rjxg51q2K`(9rtP=1aNV(Lim-FJ(
z)Z&}qMLQv~*{d(x;!6ylkS)vb`~sIO=#eEKwWtjHKjX7mn0vd=%>8CuHR^B;YeY)-
z+&Y)ci?NDdCb-RX(QJCgSx{A_9>ilFF^py(%J_WfCb`;^!26z&peRU&ZWyeTS`{8m
zo$1)YbtKY97A|spr7pJMuk`2^s+F+MYIo8o=<!bee)AbSp5vajS<7Ui#JLcv;UlVV
z+_zV6?5dXHl-Nz$g5zE`6QhNDL6+ZlW7ev)NMtE#?>jO`Q9W^B*2|3#?(bHmFa<ax
zpbqh#0L&~HfeW(v6p&|RZL+9i&kww`MI)s#wOz2>k!Qq+&pjCNS%Rf+iltG;w6OhX
z^@}-J=}N67Ku%P88QBr}vPIs4b(1eV-u^XRzd^Fh5vH^#*Kiq}^GXI6lxU)^S@~mM
z6c-duT+!T*@>aWlfGJ5@y24V+NB%HULEiltru9Mimw^O$NV#d*q+<`UXFO+37thRE
z0_jOWU)1L#S3dWxLW_PUx2oyx+T(ZEGR2+Vi6~A8Wk^`+!n#x@Dm-463eC;Ue!SyB
zlOU?~Ft^&d?%eQp>qX7$r6;H&ZOK*bB#DEJjAzv3M#U2`B58p$9~KSJN6cxnLrFCg
z0T>AUE)YSMdSDx9#}J=en5wv$T3Iv`rC0q@S3Ey_I#M4dqSKR5qM1toXr($WhU(QC
zwSJ#Ij<Z~R6Ab0@wzNwYdW%hRZX+&e<Ius1G7++#q->!PvOeNzKU4W4-)^>qlFf*B
z%Qk`$!-5NLOMQdHUeduke6=8ZT{Hjv2e{@?d)+PUf+K-beR)*5uOw!aX-6wxWEzK(
z0c#t(iO>Pn@-{fi2JQU*^dojx6aPy?op2Qw6c~?3;AE_~S5IhkAJ|UmEt?B7d1WBc
zRc%^6y~i*@mCjG=C5<FP69Dpk-T{gq%=MV@CCfcX3JI>vrMpn&9EZTLsn>e(W?fSC
zw18Uhu6w~;R_J=1fwu0J@av720U-jOdea{0qk`k0?{l!3R>Wq@4&T8<jKe&ILf|DV
zdY6X{-z@f-<xE49K|N(%aYS(2s+g}bsOrr!`EQ{f*H?wTLMxPI?e|3MxjWNni~-j3
z@J5Nm07)CcFx2Oz%#kkZSKqnZ^Y-f{BqnEg_5L;Ik-#O@CULZC)C_=mDBVXf<C?=Y
zg_>)uR#ja*m!JNui3o1t(7#KoyP@2TZ9b?iZ7pY0nqJY$EPrSs5cb?M&EULO4q%Cf
zwz1}Fr#v(om&oZ<cWPRS7>rTjS&HBlXG1J=`SiFayGTz77Ek6Uh?<H!V)oS`#}30E
zfF(3+TsNf_eKb?_=aU9n<j@>E6|r}59nP4Mc<%fr@4M1L$UYk2=zsy8`8>At`}n<H
zuhq@9W>v86<2?Z#zM-r~Wb*^YS?42#0_zX>Maa^}G(0?G;c_#xH{9%_m|M!Ig2@Gp
zf7grZq|U4szd*&Go0N}bx)_M;o4Vxl4gak<%Uh7&zz)<jgy)rJ&O4m#H6AP!9#2U0
zp-u3cezux>q&0>fy(yGKOtp_}pI3FxcW-@kt1~@-*aN4o2Vq}Z>JWcrDJ^3b0o%I?
zla-Ob+VOvO+F{<@8QV9!CM>q#PEEzMj=iuZ`)XK8Jf;4QdM(;G(=oWFrh6fgHAQN*
zN8@=Oe_kE=>Xw^ULIxNyDv0;@6_1&M^RsP#lA2=GdBmKvg3mv_T5W-(gE%mqcGLMP
zWm;^JF4jBNzji>ISg<ZUg$|g`wcxBNP1*;x4$g!zmk{}*GLPQv+ODi4=kIELh5>lT
z0{X^)SLX??Q&{!XkQ_7H$@@QxvSI1vc9W=RQ;H1hkeBk#uRcFK55Fzdss^?6=sgL*
z@J?ud^ry0Ttgk&&9)7;*^mUl%5p$Q9N!|d_Opopd*2ZyI_vpRF8hg=?rS(01Zte6&
zo1362<!sh~BqX{{m4-<e2}Tg)ynD&b=4RchJy~9Fhe<^Wb7$&ZXJX(PbE+``2BHWK
zo7k`tMhVb!^&8f8MV~Pdv|!7}S{XpP*`Buj=&kgsN&dqkED@WlZn0V%lb3l~2yWK+
z7bHS^nr^3|Q>nOi35{?|yPGvJTen63A;ZCNwZ+@T?1qm?SukclO+Bqyj$Zn}Kh5o7
z5!=nu@c2b4rNTSC;Fk+7O?Cz-I$X)k`1gG)KXyr@8rGXJCnw*C3t`JLxRPWcd2yTZ
zQ(t+u2C=l~J8zq!!1TVo7mu`=P|M$L1vLIpEEn+=rT{HGGAWB&6l{+@UXmwk)4%Um
z$^F8`)wTyacK2QUuHHR<*7uAjsm~mr=#2+MCJVYJ1kzqq8mU2;v_rplS7A+lzfEVZ
zZ#u7&)nlpTBejV8o2O85O-e(+h=>Tl;SshA8RpNYNquXQcpMGhtAz=BvnO<XLefH?
z4&B*oVHII<^$9z*mBC<at5hDXyDNE^GR8xI_qiJ$di{aK24eETQ?u72Sw5w_vkxm$
z`?du5*5P0sb5Muy%)3~3AutEa9t0>z2)x;Pdb7G~2xTsRjgNe$4?=fO*X$km&>`-|
zojb3@v?gcd#)#~rz><Tg*`lpO2r*_zdnv*f4t<T2|FH8%%^9)vQhiCu^#Edd2*FQo
z%)7Yc@>Ob?_ZnueHe9_9#LLtIyHC8}0TF!pL0z>d=lkD_VAlGl8I)dnC59C8)m^)z
z#J6C}{#DL@<xyR0rIw9j9G?Z&C!|D_rlt>=5oug9ql=IE*lPbk0+S;{rOiv<Btti-
z`COdu);&_IlHh&~K8my$=*s_;YD3pu6mgu(|EXGt&-p0dbUy&Y(0lK1f?*xS^kQ1*
z6!ws=3S=^iIxV7E8Tg7i?VXoKi;^Y3TU9)wH!Jk5Hre4#qPp7Di8$X>x?Q`ba;Hpu
zN|w(y*f*-dSLWf$UlSN~t#ouQH~HvFQyp_{pc8{Yr6!EgaHiZ)opbuGS_$*0kQ<Xg
zJgr`d(`3+$?P~rGP)1M-loR_@u@K`+5dhTKept9hmELLTbPGU`(~uhj<c6r-RLzc+
zqkhYx0X8YgznL{eH7d7WW?Hp6uzVWhBm*#HB{QV@UYKSPuOaKEX~mLa)VTEi^Eu8J
zG(#IiKc#R%g{AKA9BI<<we;6$L2;Us94rYxYuU;*(Xi$D5t7~8Gny{i7uW<j$CRbb
zO;(!5Gxbe@UhP1E*;KkO-8A31n@ugYtr;i8mm+?g8;!gD{QD?mvV-ZNHttcKr_^Kr
zaCq6&qzgi?eao4~1X#&EfvqJ1z${ku2-fKG(wS!S{Vb#ZlOQ^?J3R0F{{*=RM)&Ty
z()OQ5M5pECu}s$fBIJ34NICC|W-XK&kJgJ6l6$>Y`?+of&iQz^Vl<fo@QZ2t9rI#=
zk|^^%*%V~Iv8`-5a_c-uTe$8`ak8&%UB!$1!(7}Q+{$^p$hBqTA|>SUy18!qQp)}-
zv+kvOt#+Dj1<#e4m!%PNWtka+%HkEJ#w-H80hUAnfTBEO=~k#W8i^f&V|KPupr3eK
zP?oItJUjPUxx>JUYqW5DG`LtfE0JTK&sO$FTo+3+r`|1YLs~w)tozUctgXO{wz}ne
z!C&pfX#fD^6Sm1oL02R4+T^QjmFoK}ux=|vRwC^Na1CGPb}W#775j|0w}Z?xLD^Tz
zxQ;D*956SJLieM&uH9tJOZ#8`8Yt#Abj(REO|rN>h)1}Z29oY!*$ZPquTKopzpo$5
z2><{9jAK3;T-q(DXaL49P!a(EigK%^TiARDC!zj?-`!^t4JUmPCqW$ZW-0!1$C5d*
zaYmZW?OG_uFYR+($DHLtO;>uo!@l3>Med9359@vNjC^`o_o114r&pvG2{;Y_Xq~W4
zUMDFR$GnQUH%Y}-R9;}NA~RQ}Wf1xFZnqt|%;MZBYdhwtX$dW`?Mk+2;kRv`x1YE@
za4oZ)V;<$L!2bJIpOyQ?pxkR&EUYAsFhA*ifvpmaR(Wcfk4sy?lZ09Dh2xCviyW6B
zpI(u**arXr0014tG0z{1Uvn<x3&8jVN+JM2QT`C68)-IoiXY>ccZ*qM^b6Mn<sP*%
z-7#-PL}&F|q0QX(clPrOj`;&g=^S&#G4JY_C&!Lg-?wMm1nT*Yd6Y@EW7fck-wpoy
zZahs`@9G-LF~1|8px1TpMP3aiWGCC;brg>ILylR4Xl|7>VQFPo<=0`sUypYE0sxrU
zToYV<(=pGr=)f_rXF>G2fRw&%eyzXgn2&SJiEhpuvsxR)1Eo?Bm*s=)On#e1;^Dzx
zr#j|QN_EG)w)Hn}PU?G+EK90^!d!V=VP)3-W+Zm)3!8uXR$QZX4UIcQM3IX*wZ{2w
zXf<nntj1T^JuAM#7;$(I*J`VgW^=csVmRg{9rGaxyMAD?Ae4i$7L?dCbd(*_F&y(@
z_FSZ$ccqwnm_PSf9RL6T00W7R!5wqctffa`ZsyJntaz{RB6p*<0T{m^CJ_LjC}-LV
zV0~e4IOgRn_}|;jeau=Gv{7R*{BIvg^>iz@4t{hpVs#?=hXv05P~eMSC$xsgxU6y1
zH}Lyy`?F)d-rO4L#O7Npj>2YU<&)&%=*wpPqv`H)%sR-s+a_Q2{|$q*I$?Q!LLOng
zYdP+9%xCsN-D@D#&zTj}^^W;$vry1Uynk50(|IiTwQet%{O-0%wS~+gZu=m9hPvLl
zp-1KK3Xb`AS@|aIAFft+tJi8X2I@k=AakJwaMk!OOK5Br$9$3nS5;;A7WVAsw#i=P
z7RP*nJ;!@#pj`X?fGLlbBqFukS<`|Ge?L5axLSQ+q^YrGN&AJX9rHy&pnT<QFY-h|
zSs!bGZ$E17^QDgYGK;73h6M%Q*2dmXnu2S*acE$1dVpiz(&}6LoN4CHQx!$J=poC(
zwby(v@SU&s+zz_H!1}oDE$-E`Hc|40ekrtC)7c`8-`j$a6CLxm7O!ZQ#lBkK0s+5b
zLAJx=wO`alKbKg4_<&=++>~Z7TRg(0?YWF5h<?`L%h2bu>X;w4K*jp(PPcny^_Y9Y
ztE{X50001JpNhqx`i5ivv-SVa**uB%!Md0>T8uf}<`w_|;}|H3004#XBC{N`TH;(v
zMDHS^uf&df+!Ph9xJK(5YIWk6wOrAJ&$QUN{%t`gf7x$JA}{iiSrFY|6EqvymY6)@
z_-7sSq2`K72Ya7PI8G-ob#-txr3~?UU$jr^{yXL(a8a88UNhMD2XjLe%H-{V!YGOR
ziD+uQR(maF|DcfS=E_Lguallo5`mte0d0ec>80%jb??}@kGFO?w@n_dR;we;=El|*
z=(cN0(ot=swQDo&1$dK$m@}%(Dx@S*_g&W4Xi)vnnzhiDw#E6jwynL8cQ5fGS2*T_
z0;{TZVlD;Gvi9blJ}dW`j(M%1A6(tGh{r5O);42T%4EFAd1eu<L9dr0qK~H~P5e{j
zuV#sSg(<^6Vq01sfA8=mL~+5Eu`LZ%;C|BF<N*Kx0HAZ|yzms;lF+g6j$WkGV*mg^
zp8+Ki0OOCVf-91OS9e?R{pqwsUgQpSIijmgh-f`~xmUMeSKEr~R~++dK^b03$?B@u
zIn#*p$xEA=&Wl`c<yLn?>zZQZd5gDmp2p!ZbLuFcdlRyGKG>fpSINn%Fsdl3HyWFc
zG@G9_MaZJID*H7XuxT8KE4@hKq}y<{`u&k+^A;QQtxQCV5Yg+lN=R1=o$Ey&&MHq(
z@1HQst9Q3p_eJY7N;4_#9?~*f+>%{v5^`~y=+$HHvNYFPy>4g|e#=_f7qH2$7j4zD
zy1%;ITt;<^=-M-7-2tv@Aq&J|0ue>qxHu2nv1x7oTzHZD&8l-}Q_L)3mVo-~Ug$;M
zbpPd;bK8u1t#;Q)v$?b>fHyS7(_5{b|EE3UHwLcI!egHpX*O53iS;#25ni!&gksl|
z+pe49N8NuOVHQPeS(`n{+M7F6G37;WEFjDH`I9KqrlxVDo%5NTbkl?LWsMV#Fl(-z
z&00+B;t^elp-LqBPrc@y^XcGR(=*ESbdfR@N#{ilHcQK$%=&gdb3b~GjT5wwdw;?4
zO;%YV%G7fvUgS#0yu4W-Z*DRtd!`;DqMJ-1nut#N)EkWrMw-n}Tb(Ux{pm~_D?Di9
zg==+8l90!Jb}ecBTk8EeT&?ah(rj*RGJmg~yJFAVv-YfA;ze$5M@9eu000<;Dt$g^
yW7ZGZVi6q+JDdBs!xq|};zbY>5P<Ox5&b`Orj^g?S30%;0000<MNUMnLSTZ5G3s9c
literal 0
HcmV?d00001
diff --git a/docs/source/changelog/index.rst b/docs/source/changelog/index.rst
index 534b1dd3..378c73b3 100644
--- a/docs/source/changelog/index.rst
+++ b/docs/source/changelog/index.rst
@@ -83,3 +83,57 @@ a workflow using Google Colab on a GPU for free.
- A ``metrics.csv`` output file is produced for those interested in running
hundreds of samples in automated pipelines. This file can be parsed to look for
indications that a sample may need to be re-run.
+
+Human-mouse mixture benchmark
+-----------------------------
+
+The `human-mouse mixture dataset from 10x Genomics
+<https://www.10xgenomics.com/resources/datasets/12-k-1-1-mixture-of-fresh-frozen-human-hek-293-t-and-mouse-nih-3-t-3-cells-2-standard-2-1-0>`_
+is a great dataset for benchmarking noise removal methods.
+
+Figure legend:
+
+a. Log-log plot of counts of human genes versus mouse genes in each cell.
+ Each cell is a dot. Gray dots are the raw data, same in each figure. Green
+ dots are the CellBender output at FPR 0.01. Marginal histograms are shown
+ above and to the right.
+b. Quantification of cross-species contamination per cell. Blue boxes show
+ human genes in mouse cells, and orange boxes show mouse genes in human cells.
+c. Quantification of overall count removal per gene. Each dot is a gene,
+ summed over all cells. X-axis is the raw data. Y-axis is the fraction of
+ counts remaining after CellBender noise removal. This is not a comparison
+ with truth, but rather a look to see if high-count or low-count genes
+ are being treated differently with regard to noise removal. The red line
+ guides the eye for what would be a constant level of removal.
+
+This benchmark is not shown for v0.1.0, as it is deprecated.
+
+v0.2.2
+~~~~~~
+
+.. image:: /_static/remove_background/v0.2.2_hgmm.png
+ :width: 750 px
+
+Note that the very small dip in panel c for genes with raw counts > 1e5
+indicates some over-removal of highly expressed genes. (Some of the mass of
+black dots is dipping below the red line.)
+
+Nature Methods publication
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The code run in the paper is technically v0.3.0_rc, a "release candidate".
+The results are shown in the paper in Figure 5a, and reproduced here in a
+comparable format.
+
+.. image:: /_static/remove_background/v0.3.0_rc_hgmm.png
+ :width: 750 px
+
+Note that the seeming performance regression came with
+a lot of extra guarantees about the quality of the output. The dip visible
+in panel c in v0.2.2 has disappeared.
+
+v0.3.0
+~~~~~~
+
+.. image:: /_static/remove_background/v0.3.0_hgmm.png
+ :width: 750 px
diff --git a/docs/source/citation/index.rst b/docs/source/citation/index.rst
index 65b8beb1..f6f1ed3a 100644
--- a/docs/source/citation/index.rst
+++ b/docs/source/citation/index.rst
@@ -9,7 +9,6 @@ citing our paper:
Stephen J Fleming, Mark D Chaffin, Alessandro Arduini, Amer-Denis Akkad, Eric Banks,
John C Marioni, Anthony A Philippakis, Patrick T Ellinor, and Mehrtash Babadi.
Unsupervised removal of systematic background noise from droplet-based single-cell
-experiments using CellBender. *Nature Methods* (in press), 2023.
+experiments using CellBender. *Nature Methods*, 2023. https://doi.org/10.1038/s41592-023-01943-7
-Preprint `available at bioRxiv <https://www.biorxiv.org/content/10.1101/791699v2>`_.
-https://doi.org/10.1101/791699
+Preprint `available on bioRxiv <https://www.biorxiv.org/content/10.1101/791699v2>`_.
diff --git a/docs/source/contributing/index.rst b/docs/source/contributing/index.rst
index 7d89ef94..682cf7ff 100644
--- a/docs/source/contributing/index.rst
+++ b/docs/source/contributing/index.rst
@@ -9,6 +9,12 @@ with our research collaborators, your feedback is invaluable to us and allows us
to steer CellBender in the direction that you find most useful in your research. If you
have an interesting idea or suggestion, please do not hesitate to reach out to us.
+A github issue is often the right place to start a conversation about a new feature
+or a bug fix.
+From there, once we collectively agree on a sense of how to proceed, the repository
+can be forked, changes made (using the ``dev`` branch), and a pull request (PR)
+can be created. PRs are to target the ``dev`` branch.
+
If you encounter a bug, please file a detailed github
`issue <https://github.com/broadinstitute/CellBender/issues>`_
and we will get back to you as soon as possible.
diff --git a/docs/source/introduction/index.rst b/docs/source/introduction/index.rst
index 7bcf04a9..77b5525a 100644
--- a/docs/source/introduction/index.rst
+++ b/docs/source/introduction/index.rst
@@ -13,11 +13,11 @@ Despite the recent progress in improving, optimizing and standardizing droplet-b
like scRNA-seq, the complexity of these experiments leaves room for systematic biases and background noise in
the raw observations. These nuisances
can be traced back to undesirable enzymatic processes that produce spurious library fragments, contamination by
-exogeneous or endogenous ambient transcripts, impurity of barcode beads, and barcode swapping during amplification
+exogenous or endogenous ambient transcripts, impurity of barcode beads, and barcode swapping during amplification
and/or sequencing.
The main purpose of CellBender is to take raw gene-by-cell count matrices and molecule-level information produced
-by 3rd party pipelines (e.g. CellRanger, Alevin, DropSeq, etc.), to model and remove systematic biases and
+by 3rd party pipelines (e.g. CellRanger, Alevin, DropSeq, StarSolo, etc.), to model and remove systematic biases and
background noise, and to produce improved estimates of gene expression.
As such, CellBender relies on an external tool for primary processing of the raw data obtained from the
diff --git a/docs/source/troubleshooting/index.rst b/docs/source/troubleshooting/index.rst
index 893f7bb8..54c13507 100644
--- a/docs/source/troubleshooting/index.rst
+++ b/docs/source/troubleshooting/index.rst
@@ -177,7 +177,8 @@ Answers / Troubleshooting Tips
Anecdotally it seems that ATAC data is less noisy than gene expression data
to begin with, so some users opt to have CellBender ignore the ATAC features
using the input argument ``--exclude-feature-types Peaks``. There is nothing
- wrong with doing this.
+ wrong with doing this. The CellBender output file will still contain the ATAC
+ Peak features, but they will be identical to the raw input file.
.. _a8:
@@ -421,6 +422,5 @@ Answers / Troubleshooting Tips
This means that the "empty droplet plateau" could not be identified. The most likely
explanation is that the level of background RNA is extremely low, and that the value
of ``--low-count-threshold`` exceeds this level. This would result in the empty
- droplet plateau being excluded from the analysis, which is not advisable. This can be
- corrected by decreasing ``--low-count-threshold`` from its default of 15 to a value like 5.
-
+ droplet plateau being excluded from the analysis, which is not advisable. This could
+ possibly be corrected by decreasing ``--low-count-threshold`` to a value like 1.
diff --git a/docs/source/usage/index.rst b/docs/source/usage/index.rst
index bb0d5e95..ddcaebf1 100644
--- a/docs/source/usage/index.rst
+++ b/docs/source/usage/index.rst
@@ -27,7 +27,12 @@ Proposed pipeline
~~~~~~~~~~~~~~~~~
#. Run ``cellranger count`` or some other quantification tool to obtain a count matrix
-#. Run ``cellbender remove-background``
+#. Run ``cellbender remove-background`` using the command
+
+.. code-block:: console
+
+ cellbender remove-background --cuda --input input_file.h5 --output output_file.h5
+
#. Perform per-cell quality control checks, and filter out dead / dying cells,
as appropriate for your experiment
#. Perform all subsequent analyses using the CellBender count matrix. (It is useful
@@ -67,18 +72,16 @@ Run ``remove-background`` on the dataset using the following command
.. code-block:: console
(cellbender) $ cellbender remove-background \
- --input raw_feature_bc_matrix.h5 \
- --output output.h5 \
--cuda \
- --expected-cells 5000 \
- --total-droplets-included 20000 \
- --fpr 0.01 \
- --epochs 150
+ --input raw_feature_bc_matrix.h5 \
+ --output output.h5
(The output filename "output.h5" can be replaced with a filename of choice.)
-This command will produce five output files:
+This command will produce nine output files:
+* ``output_report.html``: HTML report including plots and commentary, along with any
+ warnings or suggestions for improved parameter settings.
* ``output.h5``: Full count matrix as an h5 file, with background RNA removed. This file
contains all the original droplet barcodes.
* ``output_filtered.h5``: Filtered count matrix as an h5 file, with background RNA removed.
@@ -92,15 +95,23 @@ This command will produce five output files:
* ``output.log``: Log file produced by the ``cellbender remove-background`` run.
* ``output_metrics.csv``: Metrics describing the run, potentially to be used to flag
problematic runs when using CellBender as part of a large-scale automated pipeline.
-* ``output_report.html``: HTML report including plots and commentary, along with any
- warnings or suggestions for improved parameter settings.
* ``ckpt.tar.gz``: Checkpoint file which contains the trained model and the full posterior.
+* ``output_posterior.h5``: The full posterior probability of noise counts. This is
+ not normally used downstream.
+
+If you are interested in saving space and you do not need to re-run cellbender,
+only the ``output_report.html`` and the ``output.h5`` need to be stored. The
+``ckpt.tar.gz`` in particular is a large file which can be deleted to save disk
+storage space. (However, if you keep this checkpoint file, it can be used to
+create a new output count matrix with a different ``--fpr``, without
+needing to re-run the lengthy training process. Simply run the command again
+with a different ``--fpr`` and specify ``--checkpoint ckpt.tar.gz``.)
Quality control checks
~~~~~~~~~~~~~~~~~~~~~~
-* Check the log file for any warnings.
-* Check lines 11 -18 in the log file. Ensure that the automatically-determined priors
+* Check the log file for any errors or warnings.
+* Check lines 13-21 in the log file. Ensure that the automatically-determined priors
for cell counts and empty droplet counts match your expectation from the UMI curve.
Ensure that the numbers of "probable cells", "additional barcodes", and "empty droplets"
are all nonzero and look reasonable.
From e3588886003177df498561107deed9eb86cc4fb9 Mon Sep 17 00:00:00 2001
From: kshakir <github@kshakir.org>
Date: Tue, 31 Oct 2023 10:44:51 -0400
Subject: [PATCH 03/19] Add WDL input to set number of retries. (#247)
---
.../tests/benchmarking/docker_image_check_cuda_status.wdl | 5 ++++-
wdl/cellbender_remove_background.wdl | 5 ++++-
2 files changed, 8 insertions(+), 2 deletions(-)
diff --git a/cellbender/remove_background/tests/benchmarking/docker_image_check_cuda_status.wdl b/cellbender/remove_background/tests/benchmarking/docker_image_check_cuda_status.wdl
index 8763bfe9..91ae2777 100644
--- a/cellbender/remove_background/tests/benchmarking/docker_image_check_cuda_status.wdl
+++ b/cellbender/remove_background/tests/benchmarking/docker_image_check_cuda_status.wdl
@@ -12,6 +12,8 @@ task run_check_pytorch_cuda_status {
Int? hardware_boot_disk_size_GB = 20
String? hardware_zones = "us-east1-d us-east1-c us-central1-a us-central1-c us-west1-b"
String? hardware_gpu_type = "nvidia-tesla-t4"
+ Int? hardware_premptible_tries = 2
+ Int? hardware_max_retries = 0
String? nvidia_driver_version = "470.82.01" # need >=465.19.01 for CUDA 11.3
}
command {
@@ -28,7 +30,8 @@ task run_check_pytorch_cuda_status {
gpuCount: 1
gpuType: "${hardware_gpu_type}"
nvidiaDriverVersion: "${nvidia_driver_version}"
- maxRetries: 0
+ preemptible: hardware_premptible_tries
+ maxRetries: hardware_max_retries
}
}
diff --git a/wdl/cellbender_remove_background.wdl b/wdl/cellbender_remove_background.wdl
index 4efb8faf..6f974639 100644
--- a/wdl/cellbender_remove_background.wdl
+++ b/wdl/cellbender_remove_background.wdl
@@ -58,6 +58,7 @@ task run_cellbender_remove_background_gpu {
Int? hardware_disk_size_GB = 50
Int? hardware_boot_disk_size_GB = 20
Int? hardware_preemptible_tries = 2
+ Int? hardware_max_retries = 0
Int? hardware_cpu_count = 4
Int? hardware_memory_GB = 32
String? hardware_gpu_type = "nvidia-tesla-t4"
@@ -186,7 +187,7 @@ task run_cellbender_remove_background_gpu {
nvidiaDriverVersion: "${nvidia_driver_version}"
preemptible: hardware_preemptible_tries
checkpointFile: "ckpt.tar.gz"
- maxRetries: 0 # can be used in case of a PAPI error code 2 failure to install GPU drivers
+ maxRetries: hardware_max_retries # can be used in case of a PAPI error code 2 failure to install GPU drivers
}
meta {
author: "Stephen Fleming"
@@ -214,6 +215,8 @@ task run_cellbender_remove_background_gpu {
{help: "Optional file only used by CellBender developers or those trying to benchmark CellBender remove-background on simulated data. Normally, this input would not be supplied."}
hardware_preemptible_tries :
{help: "If nonzero, CellBender will be run on a preemptible instance, at a lower cost. If preempted, your run will not start from scratch, but will start from a checkpoint that is saved by CellBender and recovered by Cromwell. For example, if hardware_preemptible_tries is 2, your run will attempt twice using preemptible instances, and if the job is preempted both times before completing, it will finish on a non-preemptible machine. The cost savings is significant. The potential drawback is that preemption wastes time."}
+ hardware_max_retries :
+ {help: "If nonzero when CellBender exits without success it will be retried. If one also sets the memory_retry_multiplier workflow option, and the exit happens to be detected as an out of memory error, then the retry will also increase the memory allocated to the next run. The potential benefit is that one can start CellBender with less memory, and memory will be increased only when needed. The potential drawback is that the job will be retried even if the error is not a memory error."}
checkpoint_mins :
{help: "Time in minutes between creation of checkpoint files. Bear in mind that Cromwell copies checkpoints to a bucket every ten minutes."}
hardware_gpu_type :
From 7a834a41aa121dce3f32ff428facfa8c7f269985 Mon Sep 17 00:00:00 2001
From: alecw <alecw@users.noreply.github.com>
Date: Tue, 31 Oct 2023 10:45:41 -0400
Subject: [PATCH 04/19] Move hash computation so that it is recomputed on
retry, and now-invalid checkpoint is not loaded. (#258)
If number of tries is exhausted, and ELBO tests are still failing, allow to complete anyway (using checkpoint) so that outputs are produced, but exit(1).
---
cellbender/remove_background/cli.py | 11 -----------
cellbender/remove_background/run.py | 25 ++++++++++++++++++++++++-
2 files changed, 24 insertions(+), 12 deletions(-)
diff --git a/cellbender/remove_background/cli.py b/cellbender/remove_background/cli.py
index ad49c018..29f7928c 100644
--- a/cellbender/remove_background/cli.py
+++ b/cellbender/remove_background/cli.py
@@ -207,17 +207,6 @@ def setup_and_logging(args):
+ ' '.join(['cellbender', 'remove-background'] + sys.argv[2:]))
logger.info("CellBender " + get_version())
- # Set up checkpointing by creating a unique workflow hash.
- hashcode = create_workflow_hashcode(
- module_path=os.path.dirname(cellbender.__file__),
- args_to_remove=(['output_file', 'fpr', 'input_checkpoint_tarball', 'debug',
- 'posterior_batch_size', 'checkpoint_min', 'truth_file',
- 'posterior_regularization', 'cdf_threshold_q', 'prq_alpha',
- 'estimator', 'use_multiprocessing_estimation', 'cpu_threads']
- + (['epochs'] if args.constant_learning_rate else [])),
- args=args)[:10]
- args.checkpoint_filename = hashcode # store this in args
- logger.info(f'(Workflow hash {hashcode})')
return args, file_handler
diff --git a/cellbender/remove_background/run.py b/cellbender/remove_background/run.py
index 7b777e4e..42cf5b4e 100644
--- a/cellbender/remove_background/run.py
+++ b/cellbender/remove_background/run.py
@@ -1,5 +1,6 @@
"""Single run of remove-background, given input arguments."""
+import cellbender
from cellbender.remove_background.model import RemoveBackgroundPyroModel
from cellbender.remove_background.data.dataset import get_dataset_obj, \
SingleCellRNACountsDataset
@@ -21,6 +22,7 @@
from cellbender.remove_background.sparse_utils import csr_set_rows_to_zero
from cellbender.remove_background.data.io import write_matrix_to_cellranger_h5
from cellbender.remove_background.report import run_notebook_make_html, plot_summary
+from cellbender.remove_background.checkpoint import create_workflow_hashcode
import pyro
from pyro.infer import SVI, JitTraceEnum_ELBO, JitTrace_ELBO, \
@@ -59,6 +61,22 @@ def run_remove_background(args: argparse.Namespace) -> Posterior:
"""
+ # Set up checkpointing by creating a unique workflow hash.
+ hashcode = create_workflow_hashcode(
+ module_path=os.path.dirname(cellbender.__file__),
+ args_to_remove=(['output_file', 'fpr', 'input_checkpoint_tarball', 'debug',
+ 'posterior_batch_size', 'checkpoint_min', 'truth_file',
+ 'posterior_regularization', 'cdf_threshold_q', 'prq_alpha',
+ 'estimator', 'use_multiprocessing_estimation', 'cpu_threads',
+ # The following settings do not affect the results, and can change when retrying,
+ # so remove them.
+ 'epoch_elbo_fail_fraction', 'final_elbo_fail_fraction',
+ 'num_failed_attempts', 'checkpoint_filename']
+ + (['epochs'] if args.constant_learning_rate else [])),
+ args=args)[:10]
+ args.checkpoint_filename = hashcode # store this in args
+ logger.info(f'(Workflow hash {hashcode})')
+
# Handle initial random state.
pyro.util.set_rng_seed(consts.RANDOM_SEED)
if torch.cuda.is_available():
@@ -771,7 +789,12 @@ def run_inference(dataset_obj: SingleCellRNACountsDataset,
sys.exit(0)
else:
logger.info(f'No more attempts are specified by --num-training-tries. '
- f'Therefore the workflow will abort here.')
+ f'Therefore the workflow will run once more without ELBO restrictions.')
+ args.epoch_elbo_fail_fraction = None
+ args.final_elbo_fail_fraction = None
+ run_remove_background(args) # start from scratch
+ # non-zero exit status in order to draw user's attention to the fact that ELBO tests
+ # were never satisfied.
sys.exit(1)
logger.info("Inference procedure complete.")
From cf71148b952e8f6ed087e56a413c44c61b3c4768 Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Tue, 31 Oct 2023 10:46:16 -0400
Subject: [PATCH 05/19] Bug fix for WDL using MTX input (#246)
---
wdl/cellbender_remove_background.wdl | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/wdl/cellbender_remove_background.wdl b/wdl/cellbender_remove_background.wdl
index 6f974639..f7507ec3 100644
--- a/wdl/cellbender_remove_background.wdl
+++ b/wdl/cellbender_remove_background.wdl
@@ -125,8 +125,15 @@ task run_cellbender_remove_background_gpu {
[ -f $dir/$name ] || mv ~{genes_file} $dir/$name
fi
+ # use the directory as the input in the case of an MTX file
+ if [[ "~{input_file_unfiltered}" == *.mtx* ]]; then
+ input=$(dirname ~{input_file_unfiltered})
+ else
+ input=~{input_file_unfiltered}
+ fi
+
cellbender remove-background \
- --input "~{input_file_unfiltered}" \
+ --input $input \
--output "~{sample_name}_out.h5" \
--cuda \
~{"--checkpoint " + checkpoint_file} \
From 322971d81eb8ae745601515aa15227cc7af6d951 Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Tue, 31 Oct 2023 10:47:02 -0400
Subject: [PATCH 06/19] Memory-efficient posterior generation (#263)
---
cellbender/remove_background/estimation.py | 45 +++-----------
cellbender/remove_background/posterior.py | 61 ++++++++++---------
cellbender/remove_background/sparse_utils.py | 6 +-
.../remove_background/tests/test_dataprep.py | 6 +-
.../tests/test_sparse_utils.py | 6 +-
5 files changed, 50 insertions(+), 74 deletions(-)
diff --git a/cellbender/remove_background/estimation.py b/cellbender/remove_background/estimation.py
index 24437248..371a7fa0 100644
--- a/cellbender/remove_background/estimation.py
+++ b/cellbender/remove_background/estimation.py
@@ -463,11 +463,9 @@ def estimate_noise(self,
if use_multiple_processes:
logger.info('Dividing dataset into chunks of genes')
- chunk_logic_list = list(
- self._gene_chunk_iterator(
- noise_log_prob_coo=noise_log_prob_coo,
- n_chunks=n_chunks,
- )
+ chunk_logic_list = self._gene_chunk_iterator(
+ noise_log_prob_coo=noise_log_prob_coo,
+ n_chunks=n_chunks,
)
logger.info('Computing the output in asynchronous chunks in parallel...')
@@ -538,10 +536,9 @@ def estimate_noise(self,
def _gene_chunk_iterator(self,
noise_log_prob_coo: sp.coo_matrix,
n_chunks: int) \
- -> Generator[np.ndarray, None, None]:
- """Yields chunks of the posterior that can be treated as independent,
- from the standpoint of MCKP count estimation. That is, they contain all
- matrix entries for any genes they include.
+ -> List[np.ndarray]:
+ """Return a list of logical (size m) arrays used to select gene chunks
+ on which to compute the MCKP estimate. These chunks are independent.
Args:
noise_log_prob_coo: Full noise log prob posterior COO
@@ -551,36 +548,14 @@ def _gene_chunk_iterator(self,
Logical array which indexes elements of coo posterior for the chunk
"""
- # TODO this generator is way too slow
-
- # approximate number of entries in a chunk
- # approx_chunk_entries = (noise_log_prob_coo.data.size - 1) // n_chunks
-
# get gene annotations
_, genes = self.index_converter.get_ng_indices(m_inds=noise_log_prob_coo.row)
genes_series = pd.Series(genes)
- # things we need to keep track of for each chunk
- # current_chunk_genes = []
- # entry_logic = np.zeros(noise_log_prob_coo.data.size, dtype=bool)
-
- # TODO eliminate for loop to speed this up
- # take the list of genes from the coo, sort it, and divide it evenly
- # somehow break ties for genes overlapping boundaries of divisions
- sorted_genes = np.sort(genes)
- gene_arrays = np.array_split(sorted_genes, n_chunks)
- last_gene_set = {}
- for gene_array in gene_arrays:
- gene_set = set(gene_array)
- gene_set = gene_set.difference(last_gene_set) # only the new stuff
- # if there is a second chunk, make sure there is a gene unique to it
- if (n_chunks > 1) and (len(gene_set) == len(set(genes))): # all genes in first set
- # this mainly exists for tests
- gene_set = gene_set - {gene_arrays[-1][-1]}
- last_gene_set = gene_set
- entry_logic = genes_series.isin(gene_set).values
- if sum(entry_logic) > 0:
- yield entry_logic
+ gene_chunk_arrays = np.array_split(np.arange(self.index_converter.total_n_genes), n_chunks)
+
+ gene_logic_arrays = [genes_series.isin(x).values for x in gene_chunk_arrays]
+ return gene_logic_arrays
def _chunk_estimate_noise(self,
noise_log_prob_coo: sp.coo_matrix,
diff --git a/cellbender/remove_background/posterior.py b/cellbender/remove_background/posterior.py
index 37587026..6aaeefa9 100644
--- a/cellbender/remove_background/posterior.py
+++ b/cellbender/remove_background/posterior.py
@@ -451,7 +451,7 @@ def _get_cell_noise_count_posterior_coo(
f'accurate for your dataset.')
raise RuntimeError('Zero cells found!')
- dataloader_index_to_analyzed_bc_index = np.where(cell_logic)[0]
+ dataloader_index_to_analyzed_bc_index = torch.where(torch.tensor(cell_logic))[0]
cell_data_loader = DataLoader(
count_matrix[cell_logic],
empty_drop_dataset=None,
@@ -468,6 +468,12 @@ def _get_cell_noise_count_posterior_coo(
log_probs = []
ind = 0
n_minibatches = len(cell_data_loader)
+ analyzed_gene_inds = torch.tensor(self.analyzed_gene_inds.copy())
+ if analyzed_bcs_only:
+ barcode_inds = torch.tensor(self.dataset_obj.analyzed_barcode_inds.copy())
+ else:
+ barcode_inds = torch.tensor(self.barcode_inds.copy())
+ nonzero_noise_offset_dict = {}
logger.info('Computing posterior noise count probabilities in mini-batches.')
@@ -505,46 +511,43 @@ def _get_cell_noise_count_posterior_coo(
)
# Get the original gene index from gene index in the trimmed dataset.
- genes_i = self.analyzed_gene_inds[genes_i_analyzed]
+ genes_i = analyzed_gene_inds[genes_i_analyzed.cpu()]
# Barcode index in the dataloader.
- bcs_i = bcs_i_chunk + ind
+ bcs_i = (bcs_i_chunk + ind).cpu()
# Obtain the real barcode index since we only use cells.
bcs_i = dataloader_index_to_analyzed_bc_index[bcs_i]
# Translate chunk barcode inds to overall inds.
- if analyzed_bcs_only:
- bcs_i = self.dataset_obj.analyzed_barcode_inds[bcs_i]
- else:
- bcs_i = self.barcode_inds[bcs_i]
+ bcs_i = barcode_inds[bcs_i]
# Add sparse matrix values to lists.
- try:
- bcs.extend(bcs_i.tolist())
- genes.extend(genes_i.tolist())
- c.extend(c_i.tolist())
- log_probs.extend(log_prob_i.tolist())
- c_offset.extend(noise_count_offset_NG[bcs_i_chunk, genes_i_analyzed]
- .detach().cpu().numpy())
- except TypeError as e:
- # edge case of a single value
- bcs.append(bcs_i)
- genes.append(genes_i)
- c.append(c_i)
- log_probs.append(log_prob_i)
- c_offset.append(noise_count_offset_NG[bcs_i_chunk, genes_i_analyzed]
- .detach().cpu().numpy())
+ bcs.append(bcs_i.detach())
+ genes.append(genes_i.detach())
+ c.append(c_i.detach().cpu())
+ log_probs.append(log_prob_i.detach().cpu())
+
+ # Update offset dict with any nonzeros.
+ nonzero_offset_inds, nonzero_noise_count_offsets = dense_to_sparse_op_torch(
+ noise_count_offset_NG[bcs_i_chunk, genes_i_analyzed].detach().flatten(),
+ )
+ m_i = self.index_converter.get_m_indices(cell_inds=bcs_i, gene_inds=genes_i)
+
+ nonzero_noise_offset_dict.update(
+ dict(zip(m_i[nonzero_offset_inds.detach().cpu()].tolist(),
+ nonzero_noise_count_offsets.detach().cpu().tolist()))
+ )
+ c_offset.append(noise_count_offset_NG[bcs_i_chunk, genes_i_analyzed].detach().cpu())
# Increment barcode index counter.
ind += data.shape[0] # Same as data_loader.batch_size
- # Convert the lists to numpy arrays.
- log_probs = np.array(log_probs, dtype=float)
- c = np.array(c, dtype=np.uint32)
- barcodes = np.array(bcs, dtype=np.uint64) # uint32 is too small!
- genes = np.array(genes, dtype=np.uint64) # use same as above for IndexConverter
- noise_count_offsets = np.array(c_offset, dtype=np.uint32)
+ # Concatenate lists.
+ log_probs = torch.cat(log_probs)
+ c = torch.cat(c)
+ barcodes = torch.cat(bcs)
+ genes = torch.cat(genes)
# Translate (barcode, gene) inds to 'm' format index.
m = self.index_converter.get_m_indices(cell_inds=barcodes, gene_inds=genes)
@@ -554,8 +557,6 @@ def _get_cell_noise_count_posterior_coo(
(log_probs, (m, c)),
shape=[np.prod(self.count_matrix_shape), n_counts_max],
)
- noise_offset_dict = dict(zip(m, noise_count_offsets))
- nonzero_noise_offset_dict = {k: v for k, v in noise_offset_dict.items() if (v > 0)}
self._noise_count_posterior_coo_offsets = nonzero_noise_offset_dict
return self._noise_count_posterior_coo
diff --git a/cellbender/remove_background/sparse_utils.py b/cellbender/remove_background/sparse_utils.py
index 4a0f26fa..ca31329e 100644
--- a/cellbender/remove_background/sparse_utils.py
+++ b/cellbender/remove_background/sparse_utils.py
@@ -10,7 +10,7 @@
@torch.no_grad()
def dense_to_sparse_op_torch(t: torch.Tensor,
tensor_for_nonzeros: Optional[torch.Tensor] = None) \
- -> Tuple[np.ndarray, ...]:
+ -> Tuple[torch.Tensor, ...]:
"""Converts dense matrix to sparse COO format tuple of numpy arrays (*indices, data)
Args:
@@ -28,9 +28,9 @@ def dense_to_sparse_op_torch(t: torch.Tensor,
tensor_for_nonzeros = t
nonzero_inds_tuple = torch.nonzero(tensor_for_nonzeros, as_tuple=True)
- nonzero_values = t[nonzero_inds_tuple].flatten()
+ nonzero_values = t[nonzero_inds_tuple].flatten().clone()
- return tuple([ten.cpu().numpy() for ten in (nonzero_inds_tuple + (nonzero_values,))])
+ return nonzero_inds_tuple + (nonzero_values,)
def log_prob_sparse_to_dense(coo: sp.coo_matrix) -> np.ndarray:
diff --git a/cellbender/remove_background/tests/test_dataprep.py b/cellbender/remove_background/tests/test_dataprep.py
index 8fcb1562..5d11380c 100644
--- a/cellbender/remove_background/tests/test_dataprep.py
+++ b/cellbender/remove_background/tests/test_dataprep.py
@@ -75,9 +75,9 @@ def test_dataloader_sorting(simulated_dataset, cuda):
bcs_i = loader.unsort_inds(bcs_i)
# Add sparse matrix values to lists.
- barcodes.append(bcs_i)
- genes.append(genes_i)
- counts.append(counts_i)
+ barcodes.append(bcs_i.detach().cpu())
+ genes.append(genes_i.detach().cpu())
+ counts.append(counts_i.detach().cpu())
# Increment barcode index counter.
ind += data.shape[0] # Same as data_loader.batch_size
diff --git a/cellbender/remove_background/tests/test_sparse_utils.py b/cellbender/remove_background/tests/test_sparse_utils.py
index 01230da8..2f2e13ea 100644
--- a/cellbender/remove_background/tests/test_sparse_utils.py
+++ b/cellbender/remove_background/tests/test_sparse_utils.py
@@ -76,9 +76,9 @@ def test_dense_to_sparse_op_torch(simulated_dataset, cuda):
bcs_i = data_loader.unsort_inds(bcs_i)
# Add sparse matrix values to lists.
- barcodes.append(bcs_i)
- genes.append(genes_i)
- counts.append(counts_i)
+ barcodes.append(bcs_i.detach().cpu())
+ genes.append(genes_i.detach().cpu())
+ counts.append(counts_i.detach().cpu())
# Increment barcode index counter.
ind += data.shape[0] # Same as data_loader.batch_size
From 12b47582d0eafc698994d6f24f7289449df44568 Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Tue, 31 Oct 2023 10:47:36 -0400
Subject: [PATCH 07/19] Fix posterior and estimator integer overflow bugs on
Windows (#259)
---
.github/workflows/run_pytest.yml | 6 ++--
cellbender/remove_background/checkpoint.py | 2 +-
cellbender/remove_background/estimation.py | 10 +++---
cellbender/remove_background/posterior.py | 4 +--
.../remove_background/tests/conftest.py | 24 +++++++++++--
.../tests/test_estimation.py | 23 ++++++++++++
.../remove_background/tests/test_posterior.py | 35 ++++++++++++++-----
7 files changed, 83 insertions(+), 21 deletions(-)
diff --git a/.github/workflows/run_pytest.yml b/.github/workflows/run_pytest.yml
index 37e652b0..7c8c6620 100644
--- a/.github/workflows/run_pytest.yml
+++ b/.github/workflows/run_pytest.yml
@@ -1,4 +1,4 @@
-# Run cellbender's tests
+# Run cellbender test suite
name: 'pytest'
@@ -7,11 +7,13 @@ on: pull_request
jobs:
build:
- runs-on: 'ubuntu-latest'
strategy:
matrix:
+ os: ['ubuntu-latest', 'windows-latest']
python-version: ['3.7']
+ runs-on: ${{ matrix.os }}
+
steps:
- name: 'Checkout repo'
uses: actions/checkout@v3
diff --git a/cellbender/remove_background/checkpoint.py b/cellbender/remove_background/checkpoint.py
index 5b0903de..206737be 100644
--- a/cellbender/remove_background/checkpoint.py
+++ b/cellbender/remove_background/checkpoint.py
@@ -297,7 +297,7 @@ def make_tarball(files: List[str], tarball_name: str) -> bool:
for file in files:
# without arcname, unpacking results in unpredictable file locations!
tar.add(file, arcname=os.path.basename(file))
- os.rename(tarball_name + '.tmp', tarball_name)
+ os.replace(tarball_name + '.tmp', tarball_name)
return True
diff --git a/cellbender/remove_background/estimation.py b/cellbender/remove_background/estimation.py
index 371a7fa0..b4866d2f 100644
--- a/cellbender/remove_background/estimation.py
+++ b/cellbender/remove_background/estimation.py
@@ -218,7 +218,7 @@ def _estimation_array_to_csr(index_converter,
data: np.ndarray,
m: np.ndarray,
noise_offsets: Optional[Dict[int, int]],
- dtype=np.int64) -> sp.csr_matrix:
+ dtype=np.int) -> sp.csr_matrix:
"""Say you have point estimates for each count matrix element (data) and
you have the 'm'-indices for each value (m). This returns a CSR matrix
that has the shape of the count matrix, where duplicate entries have
@@ -229,7 +229,7 @@ def _estimation_array_to_csr(index_converter,
a flat format, indexed by 'm'.
m: Array of the same length as data, where each entry is an m-index.
noise_offsets: Noise count offset values keyed by 'm'.
- dtype: Data type for sparse matrix. Int32 is too small for 'm' indices.
+ dtype: Data type for values of sparse matrix
Results:
noise_csr: Noise point estimate, as a CSR sparse matrix.
@@ -238,7 +238,7 @@ def _estimation_array_to_csr(index_converter,
row, col = index_converter.get_ng_indices(m_inds=m)
if noise_offsets is not None:
data = data + np.array([noise_offsets.get(i, 0) for i in m])
- coo = sp.coo_matrix((data.astype(dtype), (row.astype(dtype), col.astype(dtype))),
+ coo = sp.coo_matrix((data.astype(dtype), (row.astype(np.uint64), col.astype(np.uint8))),
shape=index_converter.matrix_shape, dtype=dtype)
coo.sum_duplicates()
return coo.tocsr()
@@ -785,7 +785,7 @@ def apply_function_dense_chunks(noise_log_prob_coo: sp.coo_matrix,
"""
array_length = len(np.unique(noise_log_prob_coo.row))
- m = np.zeros(array_length)
+ m = np.zeros(array_length, dtype=np.uint64)
out = np.zeros(array_length)
a = 0
@@ -804,7 +804,7 @@ def apply_function_dense_chunks(noise_log_prob_coo: sp.coo_matrix,
out[a:(a + len_s)] = s.detach().cpu().numpy()
a = a + len_s
- return {'m': m.astype(int), 'result': out}
+ return {'m': m, 'result': out}
def pandas_grouped_apply(coo: sp.coo_matrix,
diff --git a/cellbender/remove_background/posterior.py b/cellbender/remove_background/posterior.py
index 6aaeefa9..c4217130 100644
--- a/cellbender/remove_background/posterior.py
+++ b/cellbender/remove_background/posterior.py
@@ -555,7 +555,7 @@ def _get_cell_noise_count_posterior_coo(
# Put the counts into a sparse csr_matrix.
self._noise_count_posterior_coo = sp.coo_matrix(
(log_probs, (m, c)),
- shape=[np.prod(self.count_matrix_shape), n_counts_max],
+ shape=[np.prod(self.count_matrix_shape, dtype=np.uint64), n_counts_max],
)
self._noise_count_posterior_coo_offsets = nonzero_noise_offset_dict
return self._noise_count_posterior_coo
@@ -1528,7 +1528,7 @@ def get_m_indices(self, cell_inds: np.ndarray, gene_inds: np.ndarray) -> np.ndar
if not ((gene_inds >= 0) & (gene_inds < self.total_n_genes)).all():
raise ValueError(f'Requested gene_inds out of range: '
f'{gene_inds[(gene_inds < 0) | (gene_inds >= self.total_n_genes)]}')
- return cell_inds * self.total_n_genes + gene_inds
+ return cell_inds.astype(np.uint64) * self.total_n_genes + gene_inds.astype(np.uint64)
def get_ng_indices(self, m_inds: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
"""Given a list of 'm' index values, return two arrays: cell index values
diff --git a/cellbender/remove_background/tests/conftest.py b/cellbender/remove_background/tests/conftest.py
index 8bafa2da..6a430107 100644
--- a/cellbender/remove_background/tests/conftest.py
+++ b/cellbender/remove_background/tests/conftest.py
@@ -15,12 +15,32 @@
USE_CUDA = torch.cuda.is_available()
-def sparse_matrix_equal(mat1: sp.csc_matrix,
- mat2: sp.csc_matrix):
+def sparse_matrix_equal(mat1, mat2):
"""Fast assertion that sparse matrices are equal"""
+ if type(mat1) == sp.coo_matrix:
+ return coo_equal(mat1, mat2)
+ elif (type(mat1) == sp.csc_matrix) or (type(mat1) == sp.csr_matrix):
+ return csc_or_csr_equal(mat1, mat2)
+ else:
+ raise ValueError(f"Error with test functions: sparse_matrix_equal() was called with a {type(mat1)}")
+
+
+def csc_or_csr_equal(mat1: sp.csc_matrix,
+ mat2: sp.csc_matrix):
+ """Fast assertion that CSC sparse matrices are equal"""
return (mat1 != mat2).sum() == 0
+def coo_equal(mat1: sp.csc_matrix,
+ mat2: sp.csc_matrix):
+ """Fast assertion that COO sparse matrices are equal"""
+ return (
+ ((mat1.data != mat2.data).sum() == 0)
+ and ((mat1.row != mat2.row).sum() == 0)
+ and ((mat1.col != mat2.col).sum() == 0)
+ )
+
+
def tensors_equal(a: torch.Tensor,
b: torch.Tensor):
"""Assertion that torch tensors are equal for each element"""
diff --git a/cellbender/remove_background/tests/test_estimation.py b/cellbender/remove_background/tests/test_estimation.py
index 528ab675..c1cb931e 100644
--- a/cellbender/remove_background/tests/test_estimation.py
+++ b/cellbender/remove_background/tests/test_estimation.py
@@ -81,6 +81,29 @@ def log_prob_coo(request, log_prob_coo_base) \
raise ValueError(f'Test writing error: requested "{request.param}" log_prob_coo')
+def test_mean_massive_m(log_prob_coo):
+ """Sets up a posterior COO with massive m values that are > max(int32).
+ Will trigger github issue #252 if a bug exists, no assertion necessary.
+ """
+
+ coo = log_prob_coo['coo']
+ greater_than_max_int32 = 2200000000
+ new_row = coo.row.astype(np.int64) + greater_than_max_int32
+ new_shape = (coo.shape[0] + greater_than_max_int32, coo.shape[1])
+ new_coo = sp.coo_matrix((coo.data, (new_row, coo.col)),
+ shape=new_shape)
+ offset_dict = {k + greater_than_max_int32: v for k, v in log_prob_coo['offsets'].items()}
+
+ # this is just a shim
+ converter = IndexConverter(total_n_cells=2,
+ total_n_genes=new_coo.shape[0])
+
+ # set up and estimate
+ estimator = Mean(index_converter=converter)
+ noise_csr = estimator.estimate_noise(noise_log_prob_coo=new_coo,
+ noise_offsets=offset_dict)
+
+
@pytest.fixture(scope='module', params=['exact', 'filtered', 'unsorted'])
def mckp_log_prob_coo(request, log_prob_coo_base) \
-> Dict[str, Union[sp.coo_matrix, np.ndarray, Dict[int, int]]]:
diff --git a/cellbender/remove_background/tests/test_posterior.py b/cellbender/remove_background/tests/test_posterior.py
index e5382b3b..04b766ed 100644
--- a/cellbender/remove_background/tests/test_posterior.py
+++ b/cellbender/remove_background/tests/test_posterior.py
@@ -368,23 +368,35 @@ def test_compute_mean_target_removal_as_function(log_prob_coo, fpr, per_gene, cu
@pytest.mark.parametrize('blank_noise_offsets', [False, True], ids=['', 'no_noise_offsets'])
-def test_save_and_load(tmpdir_factory, blank_noise_offsets):
+@pytest.mark.parametrize('m', [1000, 2200000000], ids=['small', 'big'])
+def test_save_and_load(tmpdir_factory, blank_noise_offsets, m):
"""Test that a round trip through save and load gives the same thing"""
tmp_dir = tmpdir_factory.mktemp('posterior')
filename = tmp_dir.join('posterior.h5')
- m = 1000
n = 20
+ num_nonzeros = 1000
posterior = Posterior(dataset_obj=None, vi_model=None) # blank
- posterior_coo = sp.random(m, n, density=0.1, format='coo', dtype=float)
- posterior_coo2 = sp.random(m, n, density=0.08, format='coo', dtype=float)
+ # old way that cannot handle large m
+ # posterior_coo = sp.random(m, n, density=0.1, format='coo', dtype=float)
+ # posterior_coo2 = sp.random(m, n, density=0.08, format='coo', dtype=float)
+
+ m_array = np.random.randint(low=0, high=m, size=num_nonzeros - 1, dtype=np.uint64)
+ m_array = np.concatenate([m_array, np.array(m - 1, dtype=np.uint64)], axis=None)
+ n_array = np.random.randint(low=0, high=n, size=num_nonzeros)
+ val_array = np.random.rand(num_nonzeros) * -10
+ val_array2 = np.random.rand(num_nonzeros) * -5
+
+ posterior_coo = sp.coo_matrix((val_array, (m_array, n_array)), shape=(m, n))
+ posterior_coo2 = sp.coo_matrix((val_array2, (m_array, n_array)), shape=(m, n))
+
if blank_noise_offsets:
noise_offsets = {}
else:
- noise_offsets = dict(zip(np.random.randint(low=0, high=(m - 1), size=10),
+ noise_offsets = dict(zip(np.random.randint(low=0, high=(m - 1), size=10, dtype=np.uint64),
np.random.randint(low=1, high=5, size=10)))
kwargs = {'a': 'b', 'c': 1}
kwargs2 = {'a': 'method', 'c': 1}
@@ -396,7 +408,7 @@ def test_save_and_load(tmpdir_factory, blank_noise_offsets):
posterior._noise_count_regularized_posterior_coo = posterior_coo2
posterior._noise_count_regularized_posterior_kwargs = kwargs2
posterior._latents = {'p': np.random.randn(100), 'd': np.random.randn(100)}
- posterior.index_converter = IndexConverter(total_n_cells=1000, total_n_genes=1000)
+ posterior.index_converter = IndexConverter(total_n_cells=max(1000, m // 1000 + 1), total_n_genes=1000)
# save
posterior.save(file=str(filename))
@@ -406,9 +418,14 @@ def test_save_and_load(tmpdir_factory, blank_noise_offsets):
posterior2.load(file=str(filename))
# check
- for attr in ['_noise_count_posterior_coo', '_noise_count_posterior_coo_offsets',
- '_noise_count_posterior_kwargs', '_noise_count_regularized_posterior_coo',
- '_noise_count_regularized_posterior_kwargs', '_latents']:
+ for attr in [
+ '_noise_count_posterior_coo',
+ '_noise_count_posterior_coo_offsets',
+ '_noise_count_posterior_kwargs',
+ '_noise_count_regularized_posterior_coo',
+ '_noise_count_regularized_posterior_kwargs',
+ '_latents'
+ ]:
val1 = getattr(posterior, attr)
val2 = getattr(posterior2, attr)
print(f'{attr} ===================')
From be360063c9f59e5fedcaac86134fb928bd0d4be4 Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Tue, 31 Oct 2023 10:48:29 -0400
Subject: [PATCH 08/19] Move from setup.py to pyproject.toml (#240)
---
MANIFEST.in | 5 ++-
build_docker_release.sh | 2 +-
cellbender/VERSION.txt | 1 +
cellbender/__init__.py | 4 ++-
cellbender/base_cli.py | 7 +---
pyproject.toml | 45 ++++++++++++++++++++++++
requirements-dev.txt | 2 ++
setup.py | 77 -----------------------------------------
8 files changed, 57 insertions(+), 86 deletions(-)
create mode 100644 cellbender/VERSION.txt
create mode 100644 pyproject.toml
create mode 100644 requirements-dev.txt
delete mode 100644 setup.py
diff --git a/MANIFEST.in b/MANIFEST.in
index 66e89a17..679690c5 100644
--- a/MANIFEST.in
+++ b/MANIFEST.in
@@ -1,4 +1,7 @@
include README.rst
include LICENSE
include requirements.txt
-include requirements-rtd.txt
\ No newline at end of file
+include requirements-rtd.txt
+include requirements-dev.txt
+include cellbender/VERSION.txt
+include cellbender/remove_background/report.ipynb
\ No newline at end of file
diff --git a/build_docker_release.sh b/build_docker_release.sh
index ff0511fc..532bffe0 100755
--- a/build_docker_release.sh
+++ b/build_docker_release.sh
@@ -1,6 +1,6 @@
#!/bin/bash
-tag=$(cat cellbender/__init__.py | sed -e 's?__version__ = ??' | sed "s/^'\(.*\)'$/\1/")
+tag=$(cat cellbender/VERSION.txt)
release=v${tag}
docker build \
diff --git a/cellbender/VERSION.txt b/cellbender/VERSION.txt
new file mode 100644
index 00000000..9325c3cc
--- /dev/null
+++ b/cellbender/VERSION.txt
@@ -0,0 +1 @@
+0.3.0
\ No newline at end of file
diff --git a/cellbender/__init__.py b/cellbender/__init__.py
index 0404d810..9fdd33d6 100644
--- a/cellbender/__init__.py
+++ b/cellbender/__init__.py
@@ -1 +1,3 @@
-__version__ = '0.3.0'
+from .base_cli import get_version
+
+__version__ = get_version()
diff --git a/cellbender/base_cli.py b/cellbender/base_cli.py
index bf5a71b3..fca8eb7d 100644
--- a/cellbender/base_cli.py
+++ b/cellbender/base_cli.py
@@ -23,12 +23,7 @@ def read(rel_path):
def get_version() -> str:
- for line in read('__init__.py').splitlines():
- if line.startswith('__version__'):
- delim = '"' if '"' in line else "'"
- return line.split(delim)[1]
- else:
- raise RuntimeError("Unable to find version string.")
+ return read('VERSION.txt').splitlines()[0]
class AbstractCLI(ABC):
diff --git a/pyproject.toml b/pyproject.toml
new file mode 100644
index 00000000..6a0d36d2
--- /dev/null
+++ b/pyproject.toml
@@ -0,0 +1,45 @@
+[build-system]
+requires = ["setuptools>=61.2"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "cellbender"
+authors = [{name = "Stephen Fleming"}, {name = "Mehrtash Babadi"}]
+license = {text = "BSD (3-Clause)"}
+description = "A software package for eliminating technical artifacts from high-throughput single-cell RNA sequencing (scRNA-seq) data"
+keywords = ["scRNA-seq", "bioinformatics"]
+classifiers = [
+ "Development Status :: 4 - Beta",
+ "Intended Audience :: Science/Research",
+ "License :: OSI Approved :: BSD License",
+ "Programming Language :: Python :: 3.7",
+ "Topic :: Scientific/Engineering :: Bio-Informatics",
+]
+dynamic = ["version", "dependencies", "optional-dependencies"]
+
+[project.readme]
+file = "README.rst"
+content-type = "text/x-rst"
+
+[project.urls]
+Homepage = "http://github.com/broadinstitute/CellBender"
+Documentation = "http://cellbender.readthedocs.io"
+
+[project.scripts]
+cellbender = "cellbender.base_cli:main"
+
+[tool.setuptools]
+zip-safe = false
+include-package-data = true
+
+[tool.setuptools.packages]
+find = {namespaces = false}
+
+[tool.setuptools.dynamic]
+version = {attr = "cellbender.__version__"}
+dependencies = {file = ["requirements.txt"]}
+optional-dependencies.dev = {file = ["requirements-dev.txt"]}
+optional-dependencies.docs = {file = ["requirements-rtd.txt"]}
+
+[tool.setuptools.package-data]
+"*" = ["cellbender.remove_background.report.ipynb", "cellbender.VERSION.txt"]
diff --git a/requirements-dev.txt b/requirements-dev.txt
new file mode 100644
index 00000000..3257b82a
--- /dev/null
+++ b/requirements-dev.txt
@@ -0,0 +1,2 @@
+pytest
+scikit-learn
diff --git a/setup.py b/setup.py
deleted file mode 100644
index 106fa57c..00000000
--- a/setup.py
+++ /dev/null
@@ -1,77 +0,0 @@
-#!/usr/bin/env python
-
-import os
-import setuptools
-import codecs
-from typing import List
-
-
-def read(rel_path):
- here = os.path.abspath(os.path.dirname(__file__))
- with codecs.open(os.path.join(here, rel_path), 'r') as fp:
- return fp.read()
-
-
-def get_version() -> str:
- for line in read('cellbender/__init__.py').splitlines():
- if line.startswith('__version__'):
- delim = '"' if '"' in line else "'"
- return line.split(delim)[1]
- else:
- raise RuntimeError("Unable to find version string.")
-
-
-def readme() -> str:
- with open('README.rst') as f:
- return f.read()
-
-
-def _readlines(filename, filebase=''):
- with open(os.path.join(filebase, filename)) as f:
- lines = f.readlines()
- return lines
-
-
-def get_requirements() -> List[str]:
- requirements = _readlines('requirements.txt')
- if 'READTHEDOCS' in os.environ:
- requirements.extend(get_rtd_requirements())
- return requirements
-
-
-def get_rtd_requirements() -> List[str]:
- requirements = _readlines('requirements-rtd.txt')
- return requirements
-
-
-setuptools.setup(
- name='cellbender',
- version=get_version(),
- description='A software package for eliminating technical artifacts from '
- 'high-throughput single-cell RNA sequencing (scRNA-seq) data',
- long_description=readme(),
- long_description_content_type='text/x-rst',
- classifiers=[
- 'Development Status :: 4 - Beta',
- 'Intended Audience :: Science/Research',
- 'License :: OSI Approved :: BSD License',
- 'Programming Language :: Python :: 3.7',
- 'Topic :: Scientific/Engineering :: Bio-Informatics',
- ],
- keywords='scRNA-seq bioinformatics',
- url='http://github.com/broadinstitute/CellBender',
- author='Stephen Fleming, Mehrtash Babadi',
- license='BSD (3-Clause)',
- packages=setuptools.find_packages(),
- install_requires=get_requirements(),
- extras_require={
- "dev": ["pytest", "scikit-learn"],
- "docs": get_rtd_requirements(),
- },
- entry_points={
- 'console_scripts': ['cellbender=cellbender.base_cli:main'],
- },
- include_package_data=True,
- package_data={'': ['*.ipynb']}, # include the report template
- zip_safe=False
-)
From 084194b4612623976cb5a46cd1b47f4afbe6fdec Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Tue, 31 Oct 2023 11:55:37 -0400
Subject: [PATCH 09/19] Fix bugs with report generation across platforms (#302)
---
cellbender/remove_background/posterior.py | 11 ++++++++--
cellbender/remove_background/report.py | 25 ++++++++++++++---------
cellbender/remove_background/run.py | 1 +
3 files changed, 25 insertions(+), 12 deletions(-)
diff --git a/cellbender/remove_background/posterior.py b/cellbender/remove_background/posterior.py
index c4217130..d99e5338 100644
--- a/cellbender/remove_background/posterior.py
+++ b/cellbender/remove_background/posterior.py
@@ -1518,7 +1518,9 @@ def __repr__(self):
f'\n\ttotal_n_genes: {self.total_n_genes}'
f'\n\tmatrix_shape: {self.matrix_shape}')
- def get_m_indices(self, cell_inds: np.ndarray, gene_inds: np.ndarray) -> np.ndarray:
+ def get_m_indices(self,
+ cell_inds: Union[np.ndarray, torch.Tensor],
+ gene_inds: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
"""Given arrays of cell indices and gene indices, suitable for a sparse matrix,
convert them to 'm' index values.
"""
@@ -1528,7 +1530,12 @@ def get_m_indices(self, cell_inds: np.ndarray, gene_inds: np.ndarray) -> np.ndar
if not ((gene_inds >= 0) & (gene_inds < self.total_n_genes)).all():
raise ValueError(f'Requested gene_inds out of range: '
f'{gene_inds[(gene_inds < 0) | (gene_inds >= self.total_n_genes)]}')
- return cell_inds.astype(np.uint64) * self.total_n_genes + gene_inds.astype(np.uint64)
+ if type(cell_inds) == np.ndarray:
+ return cell_inds.astype(np.uint64) * self.total_n_genes + gene_inds.astype(np.uint64)
+ elif type(cell_inds) == torch.Tensor:
+ return cell_inds.type(torch.int64) * self.total_n_genes + gene_inds.type(torch.int64)
+ else:
+ raise ValueError('IndexConverter.get_m_indices received cell_inds of unkown object type')
def get_ng_indices(self, m_inds: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
"""Given a list of 'm' index values, return two arrays: cell index values
diff --git a/cellbender/remove_background/report.py b/cellbender/remove_background/report.py
index f1a34cb9..dfb5f903 100644
--- a/cellbender/remove_background/report.py
+++ b/cellbender/remove_background/report.py
@@ -16,6 +16,7 @@
import subprocess
import datetime
import os
+import shutil
import logging
from typing import Dict, Optional
@@ -42,26 +43,30 @@
def _run_notebook(file):
- subprocess.run(f'cp {file} tmp.report.ipynb', shell=True)
+ shutil.copy(file, 'tmp.report.ipynb')
subprocess.run(run_notebook_str(file='tmp.report.ipynb'), shell=True)
- subprocess.run(f'rm tmp.report.ipynb', shell=True)
+ os.remove('tmp.report.ipynb')
return 'tmp.report.nbconvert.ipynb'
def _to_html(file, output) -> str:
subprocess.run(to_html_str(file=file, output=output), shell=True)
- subprocess.run(f'mv {file.replace(".ipynb", ".html")} {output}', shell=True)
- subprocess.run(f'rm {file}', shell=True)
+ os.replace(file.replace(".ipynb", ".html"), output)
+ os.remove(file)
return output
def _postprocess_html(file: str, title: str):
- with open(file, mode='r') as f:
- html = f.read()
- html = html.replace('<title>tmp.report.nbconvert</title>',
- f'<title>{title}</title>')
- with open(file, mode='w') as f:
- f.write(html)
+ try:
+ with open(file, mode='r', encoding="utf8", errors="surrogateescape") as f:
+ html = f.read()
+ html = html.replace('<title>tmp.report.nbconvert</title>',
+ f'<title>{title}</title>')
+ with open(file, mode='w', encoding="utf8", errors="surrogateescape") as f:
+ f.write(html)
+ except:
+ logger.warning('Failed to overwrite default HTML report title. '
+ 'This is purely aesthetic and does not affect output.')
def run_notebook_make_html(file, output) -> str:
diff --git a/cellbender/remove_background/run.py b/cellbender/remove_background/run.py
index 42cf5b4e..ca2a592e 100644
--- a/cellbender/remove_background/run.py
+++ b/cellbender/remove_background/run.py
@@ -43,6 +43,7 @@
from typing import Tuple, Optional, Dict, Union
import matplotlib
+import matplotlib.backends.backend_pdf # issue #287
matplotlib.use('Agg')
import matplotlib.pyplot as plt # This needs to be after matplotlib.use('Agg')
From 97c1d408eb7bea18efded99777230cc473fce4cf Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Wed, 10 Apr 2024 22:38:53 -0400
Subject: [PATCH 10/19] Fix major bug in v0.3.1: negative counts (#347)
* issue #348 lxml_html_clean dependency
* Fix #306: noise count integer overflow
* Introduce --force-use-checkpoint for redoing v0.3.1 runs
* Add a warning to the README about v0.3.1
---
.github/workflows/miniwdl_check.yml | 7 +++-
.github/workflows/run_packaging_check.yml | 4 ++-
.github/workflows/run_pytest.yml | 4 ++-
.gitignore | 1 +
README.rst | 26 ++++++++++++++
cellbender/remove_background/argparser.py | 11 +++++-
cellbender/remove_background/checkpoint.py | 36 ++++++++++++-------
cellbender/remove_background/estimation.py | 10 ++++--
cellbender/remove_background/posterior.py | 6 ++--
cellbender/remove_background/run.py | 6 +++-
.../tests/test_checkpoint.py | 1 +
.../tests/test_estimation.py | 36 +++++++++++++++++--
.../tests/test_integration.py | 3 ++
requirements.txt | 1 +
14 files changed, 126 insertions(+), 26 deletions(-)
diff --git a/.github/workflows/miniwdl_check.yml b/.github/workflows/miniwdl_check.yml
index 8a8de2c8..34d1a4cd 100644
--- a/.github/workflows/miniwdl_check.yml
+++ b/.github/workflows/miniwdl_check.yml
@@ -1,7 +1,12 @@
name: 'validate WDL'
-on: [pull_request]
+
+on:
+ pull_request:
+ branches: [ master, dev ]
+
env:
MINIWDL_VERSION: 1.8.0
+
jobs:
miniwdl-check:
runs-on: ubuntu-latest
diff --git a/.github/workflows/run_packaging_check.yml b/.github/workflows/run_packaging_check.yml
index 8a9fc1cd..6144cbe0 100644
--- a/.github/workflows/run_packaging_check.yml
+++ b/.github/workflows/run_packaging_check.yml
@@ -2,7 +2,9 @@
name: 'packaging'
-on: pull_request
+on:
+ pull_request:
+ branches: [ master, dev ]
jobs:
build:
diff --git a/.github/workflows/run_pytest.yml b/.github/workflows/run_pytest.yml
index 7c8c6620..b6a02ae4 100644
--- a/.github/workflows/run_pytest.yml
+++ b/.github/workflows/run_pytest.yml
@@ -2,7 +2,9 @@
name: 'pytest'
-on: pull_request
+on:
+ pull_request:
+ branches: [ master, dev ]
jobs:
build:
diff --git a/.gitignore b/.gitignore
index 80b6c0c2..4d5b6868 100644
--- a/.gitignore
+++ b/.gitignore
@@ -18,3 +18,4 @@ dist/
*.csv
*.npz
*.tar.gz
+data/
\ No newline at end of file
diff --git a/README.rst b/README.rst
index d5895c76..f32bdefa 100644
--- a/README.rst
+++ b/README.rst
@@ -36,6 +36,32 @@ The current release contains the following modules. More modules will be added i
Please refer to `the documentation <https://cellbender.readthedocs.io/en/latest/>`_ for a quick start tutorial.
+ WARNING:
+
+ The release tagged v0.3.1 included a bug which caused output count matrices to be incorrect. `The bug
+ <https://github.com/broadinstitute/CellBender/blame/e2fb5977cb187cb4b12172c9f77ed556bca92cb0/cellbender/remove_background/estimation.py#L241>`_,
+ introduced in `#303 <https://github.com/broadinstitute/CellBender/pull/303>`_, compromised output denoised count matrices
+ (due to an integer overflow) and would often show up as negative entries in the output count matrices. The bug also existed on
+ the master branch until `#347 <https://github.com/broadinstitute/CellBender/pull/347>`_.
+
+ For now, we recommend using either v0.3.0 or the master branch (after #347) until v0.3.2 is released, and then using v0.3.2.
+
+ Outputs generated with v0.3.1 (or the master branch between #303 and #347) can be salvaged by making use of the
+ checkpoint file, which is not compromised. The following command will re-run the (inexpensive, CPU-only)
+ estimation of the output count matrix using the saved posterior in the checkpoint file. Note the use of the new
+ input argument ``--force-use-checkpoint`` which will allow use of a checkpoint file produced by a different CellBender version:
+
+ .. code-block:: console
+
+ (cellbender) $ cellbender remove-background \
+ --input my_raw_count_matrix_file.h5 \
+ --output my_cellbender_output_file.h5 \
+ --checkpoint path/to/ckpt.tar.gz \
+ --force-use-checkpoint
+
+ where ``path/to/ckpt.tar.gz`` is the path to the checkpoint file generated by the original run. Ensure that you pair up the right
+ ``--input`` with the right ``--checkpoint``.
+
Installation and Usage
----------------------
diff --git a/cellbender/remove_background/argparser.py b/cellbender/remove_background/argparser.py
index d53aca3c..aebe6703 100644
--- a/cellbender/remove_background/argparser.py
+++ b/cellbender/remove_background/argparser.py
@@ -48,10 +48,19 @@ def add_subparser_args(subparsers: argparse) -> argparse:
subparser.add_argument("--checkpoint", nargs=None, type=str,
dest='input_checkpoint_tarball',
required=False, default=consts.CHECKPOINT_FILE_NAME,
- help="Checkpoint tarball produced by the same version "
+ help="Checkpoint tarball produced by v0.3.0+ "
"of CellBender remove-background. If present, "
"and the workflow hashes match, training will "
"restart from this checkpoint.")
+ subparser.add_argument("--force-use-checkpoint",
+ dest='force_use_checkpoint', action="store_true",
+ help="Normally, checkpoints can only be used if the CellBender "
+ "code and certain input args match exactly. This flag allows you "
+ "to bypass this requirement. An example use would be to create a new output "
+ "using a checkpoint from a run of v0.3.1, a redacted version with "
+ "faulty output count matrices. If you use this flag, "
+ "ensure that the input file and the checkpoint match, because "
+ "CellBender will not check.")
subparser.add_argument("--expected-cells", nargs=None, type=int,
default=None,
dest="expected_cell_count",
diff --git a/cellbender/remove_background/checkpoint.py b/cellbender/remove_background/checkpoint.py
index 206737be..44ed12e5 100644
--- a/cellbender/remove_background/checkpoint.py
+++ b/cellbender/remove_background/checkpoint.py
@@ -154,7 +154,8 @@ def save_checkpoint(filebase: str,
def load_checkpoint(filebase: Optional[str],
tarball_name: str = consts.CHECKPOINT_FILE_NAME,
- force_device: Optional[str] = None)\
+ force_device: Optional[str] = None,
+ force_use_checkpoint: bool = False)\
-> Dict[str, Union['RemoveBackgroundPyroModel', pyro.optim.PyroOptim, DataLoader, bool]]:
"""Load checkpoint and prepare a RemoveBackgroundPyroModel and optimizer."""
@@ -163,6 +164,7 @@ def load_checkpoint(filebase: Optional[str],
tarball_name=tarball_name,
to_load=['model', 'optim', 'param_store', 'dataloader', 'args', 'random_state'],
force_device=force_device,
+ force_use_checkpoint=force_use_checkpoint,
)
out.update({'loaded': True})
logger.info(f'Loaded partially-trained checkpoint from {tarball_name}')
@@ -172,7 +174,8 @@ def load_checkpoint(filebase: Optional[str],
def load_from_checkpoint(filebase: Optional[str],
tarball_name: str = consts.CHECKPOINT_FILE_NAME,
to_load: List[str] = ['model'],
- force_device: Optional[str] = None) -> Dict:
+ force_device: Optional[str] = None,
+ force_use_checkpoint: bool = False) -> Dict:
"""Load specific files from a checkpoint tarball."""
load_kwargs = {}
@@ -192,19 +195,24 @@ def load_from_checkpoint(filebase: Optional[str],
else:
# no tarball loaded, so do not continue trying to load files
raise FileNotFoundError
-
- # See if files have a hash matching input filebase.
- if filebase is not None:
+
+ # If posterior is present, do not require run hash to match: will pick up
+ # after training and run estimation from existing posterior.
+ # This smoothly allows re-runs (including for problematic v0.3.1)
+ logger.debug(f'force_use_checkpoint: {force_use_checkpoint}')
+ if force_use_checkpoint or (filebase is None):
+ filebase = (glob.glob(os.path.join(tmp_dir, '*_model.torch'))[0]
+ .replace('_model.torch', ''))
+ logger.debug(f'Accepting any file hash, so loading {filebase}*')
+
+ else:
+ # See if files have a hash matching input filebase.
basename = os.path.basename(filebase)
filebase = os.path.join(tmp_dir, basename)
logger.debug(f'Looking for files with base name matching {filebase}*')
if not os.path.exists(filebase + '_model.torch'):
logger.info('Workflow hash does not match that of checkpoint.')
- raise ValueError('Workflow hash does not match that of checkpoint.')
- else:
- filebase = (glob.glob(os.path.join(tmp_dir, '*_model.torch'))[0]
- .replace('_model.torch', ''))
- logger.debug(f'Accepting any file hash, so loading {filebase}*')
+ raise ValueError('Workflow hash does not match that of checkpoint.')
out = {}
@@ -265,9 +273,10 @@ def load_from_checkpoint(filebase: Optional[str],
return out
-def attempt_load_checkpoint(filebase: str,
+def attempt_load_checkpoint(filebase: Optional[str],
tarball_name: str = consts.CHECKPOINT_FILE_NAME,
- force_device: Optional[str] = None)\
+ force_device: Optional[str] = None,
+ force_use_checkpoint: bool = False)\
-> Dict[str, Union['RemoveBackgroundPyroModel', pyro.optim.PyroOptim, DataLoader, bool]]:
"""Load checkpoint and prepare a RemoveBackgroundPyroModel and optimizer,
or return the inputs if loading fails."""
@@ -276,7 +285,8 @@ def attempt_load_checkpoint(filebase: str,
logger.debug('Attempting to load checkpoint from ' + tarball_name)
return load_checkpoint(filebase=filebase,
tarball_name=tarball_name,
- force_device=force_device)
+ force_device=force_device,
+ force_use_checkpoint=force_use_checkpoint)
except FileNotFoundError:
logger.debug('No tarball found')
diff --git a/cellbender/remove_background/estimation.py b/cellbender/remove_background/estimation.py
index b4866d2f..777bf03e 100644
--- a/cellbender/remove_background/estimation.py
+++ b/cellbender/remove_background/estimation.py
@@ -21,6 +21,10 @@
logger = logging.getLogger('cellbender')
+N_CELLS_DATATYPE = np.int32
+N_GENES_DATATYPE = np.int32
+COUNT_DATATYPE = np.int32
+
class EstimationMethod(ABC):
"""Base class for estimation of noise counts, given a posterior."""
@@ -52,7 +56,7 @@ def _estimation_array_to_csr(self,
data: np.ndarray,
m: np.ndarray,
noise_offsets: Optional[Dict[int, int]],
- dtype=np.int64) -> sp.csr_matrix:
+ dtype=COUNT_DATATYPE) -> sp.csr_matrix:
"""Say you have point estimates for each count matrix element (data) and
you have the 'm'-indices for each value (m). This returns a CSR matrix
that has the shape of the count matrix, where duplicate entries have
@@ -218,7 +222,7 @@ def _estimation_array_to_csr(index_converter,
data: np.ndarray,
m: np.ndarray,
noise_offsets: Optional[Dict[int, int]],
- dtype=np.int) -> sp.csr_matrix:
+ dtype=COUNT_DATATYPE) -> sp.csr_matrix:
"""Say you have point estimates for each count matrix element (data) and
you have the 'm'-indices for each value (m). This returns a CSR matrix
that has the shape of the count matrix, where duplicate entries have
@@ -238,7 +242,7 @@ def _estimation_array_to_csr(index_converter,
row, col = index_converter.get_ng_indices(m_inds=m)
if noise_offsets is not None:
data = data + np.array([noise_offsets.get(i, 0) for i in m])
- coo = sp.coo_matrix((data.astype(dtype), (row.astype(np.uint64), col.astype(np.uint8))),
+ coo = sp.coo_matrix((data.astype(dtype), (row.astype(N_CELLS_DATATYPE), col.astype(N_GENES_DATATYPE))),
shape=index_converter.matrix_shape, dtype=dtype)
coo.sum_duplicates()
return coo.tocsr()
diff --git a/cellbender/remove_background/posterior.py b/cellbender/remove_background/posterior.py
index d99e5338..5bdd890c 100644
--- a/cellbender/remove_background/posterior.py
+++ b/cellbender/remove_background/posterior.py
@@ -107,13 +107,15 @@ def _do_posterior_regularization(posterior: Posterior):
try:
ckpt_posterior = load_from_checkpoint(tarball_name=args.input_checkpoint_tarball,
filebase=args.checkpoint_filename,
- to_load=['posterior'])
+ to_load=['posterior'],
+ force_use_checkpoint=args.force_use_checkpoint)
except ValueError:
# input checkpoint tarball was not a match for this workflow
# but we still may have saved a new tarball
ckpt_posterior = load_from_checkpoint(tarball_name=consts.CHECKPOINT_FILE_NAME,
filebase=args.checkpoint_filename,
- to_load=['posterior'])
+ to_load=['posterior'],
+ force_use_checkpoint=args.force_use_checkpoint)
if os.path.exists(ckpt_posterior.get('posterior_file', 'does_not_exist')):
# Load posterior if it was saved in the checkpoint.
posterior.load(file=ckpt_posterior['posterior_file'])
diff --git a/cellbender/remove_background/run.py b/cellbender/remove_background/run.py
index ca2a592e..7ac466bd 100644
--- a/cellbender/remove_background/run.py
+++ b/cellbender/remove_background/run.py
@@ -307,6 +307,9 @@ def compute_output_denoised_counts_reports_metrics(posterior: Posterior,
posterior.latents_map['p'],
)
+ # Failsafe to ensure no negative counts.
+ assert np.all(denoised_counts.data >= 0), 'Negative count matrix entries in output'
+
# TODO: correct cell probabilities so that any zero-count droplet becomes "empty"
# Save denoised count matrix.
@@ -627,7 +630,8 @@ def run_inference(dataset_obj: SingleCellRNACountsDataset,
# Attempt to load from a previously-saved checkpoint.
ckpt = attempt_load_checkpoint(filebase=checkpoint_filename,
tarball_name=args.input_checkpoint_tarball,
- force_device='cuda:0' if args.use_cuda else 'cpu')
+ force_device='cuda:0' if args.use_cuda else 'cpu',
+ force_use_checkpoint=args.force_use_checkpoint)
ckpt_loaded = ckpt['loaded'] # True if a checkpoint was loaded successfully
if ckpt_loaded:
diff --git a/cellbender/remove_background/tests/test_checkpoint.py b/cellbender/remove_background/tests/test_checkpoint.py
index 59ccb9ec..0dbaa22d 100644
--- a/cellbender/remove_background/tests/test_checkpoint.py
+++ b/cellbender/remove_background/tests/test_checkpoint.py
@@ -414,6 +414,7 @@ def test_save_and_load_cellbender_checkpoint(tmpdir_factory, cuda, scheduler):
args.constant_learning_rate = not scheduler
args.debug = False
args.input_checkpoint_tarball = 'none'
+ args.force_use_checkpoint = False
create_random_state_blank_slate(0)
pyro.clear_param_store()
diff --git a/cellbender/remove_background/tests/test_estimation.py b/cellbender/remove_background/tests/test_estimation.py
index c1cb931e..2cca6418 100644
--- a/cellbender/remove_background/tests/test_estimation.py
+++ b/cellbender/remove_background/tests/test_estimation.py
@@ -6,9 +6,10 @@
import torch
from cellbender.remove_background.estimation import Mean, MAP, \
- SingleSample, ThresholdCDF, MultipleChoiceKnapsack, pandas_grouped_apply
+ SingleSample, ThresholdCDF, MultipleChoiceKnapsack, pandas_grouped_apply, _estimation_array_to_csr, COUNT_DATATYPE
from cellbender.remove_background.posterior import IndexConverter, \
dense_to_sparse_op_torch, log_prob_sparse_to_dense
+from cellbender.remove_background.tests.conftest import sparse_matrix_equal
from typing import Dict, Union
@@ -92,11 +93,15 @@ def test_mean_massive_m(log_prob_coo):
new_shape = (coo.shape[0] + greater_than_max_int32, coo.shape[1])
new_coo = sp.coo_matrix((coo.data, (new_row, coo.col)),
shape=new_shape)
+ print(f'original COO shape: {coo.shape}')
+ print(f'new COO shape: {new_coo.shape}')
+ print(f'new row minimum value: {new_coo.row.min()}')
+ print(f'new row maximum value: {new_coo.row.max()}')
offset_dict = {k + greater_than_max_int32: v for k, v in log_prob_coo['offsets'].items()}
# this is just a shim
- converter = IndexConverter(total_n_cells=2,
- total_n_genes=new_coo.shape[0])
+ converter = IndexConverter(total_n_cells=new_coo.shape[0],
+ total_n_genes=new_coo.shape[1])
# set up and estimate
estimator = Mean(index_converter=converter)
@@ -379,3 +384,28 @@ def test_parallel_pandas_grouped_apply(fun):
np.testing.assert_array_equal(reg['m'], parallel['m'])
np.testing.assert_array_equal(reg['result'], parallel['result'])
+
+
+def test_estimation_array_to_csr():
+
+ larger_than_uint16 = 2**16 + 1
+
+ converter = IndexConverter(total_n_cells=larger_than_uint16,
+ total_n_genes=larger_than_uint16)
+ m = larger_than_uint16 + np.arange(-10, 10)
+ data = np.random.rand(len(m)) * -10
+ noise_offsets = None
+
+ output_csr = _estimation_array_to_csr(index_converter=converter, data=data, m=m, noise_offsets=noise_offsets, dtype=COUNT_DATATYPE)
+
+ # reimplementation here with totally permissive datatypes
+ cell_and_gene_dtype = np.float64
+ row, col = converter.get_ng_indices(m_inds=m)
+ if noise_offsets is not None:
+ data = data + np.array([noise_offsets.get(i, 0) for i in m])
+ coo = sp.coo_matrix((data.astype(COUNT_DATATYPE), (row.astype(cell_and_gene_dtype), col.astype(cell_and_gene_dtype))),
+ shape=converter.matrix_shape, dtype=COUNT_DATATYPE)
+ coo.sum_duplicates()
+ truth_csr = coo.tocsr()
+
+ assert sparse_matrix_equal(output_csr, truth_csr)
diff --git a/cellbender/remove_background/tests/test_integration.py b/cellbender/remove_background/tests/test_integration.py
index 624c59df..078e9f23 100644
--- a/cellbender/remove_background/tests/test_integration.py
+++ b/cellbender/remove_background/tests/test_integration.py
@@ -50,3 +50,6 @@ def test_full_run(tmpdir_factory, h5_v3_file, cuda):
adata_cell_barcodes = adata.obs_names[adata.obs['cell_probability'] > consts.CELL_PROB_CUTOFF]
assert set(cell_barcodes) == set(adata_cell_barcodes), \
'Cell barcodes in h5 are different from those in CSV file'
+
+ # ensure there are no negative count matrix entries in the output
+ assert np.all(adata.X.data >= 0), 'Negative count matrix entries in output'
diff --git a/requirements.txt b/requirements.txt
index 3585b6a4..9b68a3c1 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -12,4 +12,5 @@ jupyter
jupyter_contrib_nbextensions
notebook<7.0.0
nbconvert<7.0.0
+lxml_html_clean
psutil
From 578c214e1afe9d737c4920c4296c63fcb98d4258 Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjf11@case.edu>
Date: Fri, 19 Apr 2024 20:55:35 +0000
Subject: [PATCH 11/19] Tweak readme appearance
---
README.rst | 12 +++++++-----
1 file changed, 7 insertions(+), 5 deletions(-)
diff --git a/README.rst b/README.rst
index f32bdefa..76add55e 100644
--- a/README.rst
+++ b/README.rst
@@ -1,23 +1,25 @@
CellBender
==========
-.. image:: https://img.shields.io/github/license/broadinstitute/CellBender?color=white
+|badge1| |badge2| |badge3| |badge4| |badge5|
+
+.. |badge1| image:: https://img.shields.io/github/license/broadinstitute/CellBender?color=white
:target: LICENSE
:alt: License
-.. image:: https://readthedocs.org/projects/cellbender/badge/?version=latest
+.. |badge2| image:: https://readthedocs.org/projects/cellbender/badge/?version=latest
:target: https://cellbender.readthedocs.io/en/latest/?badge=latest
:alt: Documentation Status
-.. image:: https://img.shields.io/pypi/v/CellBender.svg
+.. |badge3| image:: https://img.shields.io/pypi/v/CellBender.svg
:target: https://pypi.org/project/CellBender
:alt: PyPI
-.. image:: https://static.pepy.tech/personalized-badge/cellbender?period=total&units=international_system&left_color=grey&right_color=blue&left_text=pypi%20downloads
+.. |badge4| image:: https://static.pepy.tech/personalized-badge/cellbender?period=total&units=international_system&left_color=grey&right_color=blue&left_text=pypi%20downloads
:target: https://pepy.tech/project/CellBender
:alt: Downloads
-.. image:: https://img.shields.io/github/stars/broadinstitute/CellBender?color=yellow&logoColor=yellow)
+.. |badge5| image:: https://img.shields.io/github/stars/broadinstitute/CellBender?color=yellow&logoColor=yellow)
:target: https://github.com/broadinstitute/CellBender/stargazers
:alt: Stars
From 80514f06aea41795486ee0caea07a41b67498f90 Mon Sep 17 00:00:00 2001
From: Jura Pintar <jpintar@users.noreply.github.com>
Date: Mon, 22 Apr 2024 15:43:11 -0400
Subject: [PATCH 12/19] Add newer Cell Ranger feature types (#339)
In Cell Ranger 7+, there is a longer list of features one can specify in `cellranger multi`. These are added here to the list of allowed features.
---
cellbender/remove_background/cli.py | 12 ++++++++++--
1 file changed, 10 insertions(+), 2 deletions(-)
diff --git a/cellbender/remove_background/cli.py b/cellbender/remove_background/cli.py
index 29f7928c..fee04796 100644
--- a/cellbender/remove_background/cli.py
+++ b/cellbender/remove_background/cli.py
@@ -117,14 +117,22 @@ def validate_args(args) -> argparse.Namespace:
args.fpr = fpr_list_correct_dtypes
# Ensure that "exclude_features" specifies allowed features.
- # As of CellRanger 6.0, the possible features are:
+ # As of CellRanger 7.2, the possible features are:
# Gene Expression
# Antibody Capture
# CRISPR Guide Capture
# Custom
# Peaks
+ # Multiplexing Capture
+ # VDJ
+ # VDJ-T
+ # VDJ-T-GD
+ # VDJ-B
+ # Antigen Capture
allowed_features = ['Gene Expression', 'Antibody Capture',
- 'CRISPR Guide Capture', 'Custom', 'Peaks']
+ 'CRISPR Guide Capture', 'Custom', 'Peaks',
+ 'Multiplexing Capture', 'VDJ', 'VDJ-T',
+ 'VDJ-T-GD', 'VDJ-B', 'Antigen Capture']
for feature in args.exclude_features:
if feature not in allowed_features:
sys.stdout.write(f"Specified '{feature}' using --exclude-feature-types, "
From 72e6a3fa405c917a1e2faba971235eed037ef74f Mon Sep 17 00:00:00 2001
From: alecw <alecw@users.noreply.github.com>
Date: Mon, 22 Apr 2024 15:58:20 -0400
Subject: [PATCH 13/19] Fix test_elbo bug when retrying with more memory from
checkpoint (#345)
Keep test_elbo as part of model, and refer to it there from the retry code
---
cellbender/remove_background/train.py | 30 +++++++++++++--------------
1 file changed, 14 insertions(+), 16 deletions(-)
diff --git a/cellbender/remove_background/train.py b/cellbender/remove_background/train.py
index 8d778082..bf7bed67 100644
--- a/cellbender/remove_background/train.py
+++ b/cellbender/remove_background/train.py
@@ -152,7 +152,6 @@ def run_training(model: RemoveBackgroundPyroModel,
# Initialize train and tests ELBO with empty lists.
train_elbo = []
- test_elbo = []
lr = []
epoch_checkpoint_freq = 1000 # a large number... it will be recalculated
@@ -212,16 +211,15 @@ def run_training(model: RemoveBackgroundPyroModel,
if epoch % test_freq == 0:
model.eval()
total_epoch_loss_test = evaluate_epoch(svi, test_loader)
- test_elbo.append(-total_epoch_loss_test)
model.loss['test']['epoch'].append(epoch)
model.loss['test']['elbo'].append(-total_epoch_loss_test)
logger.info("[epoch %03d] average test loss: %.4f"
% (epoch, total_epoch_loss_test))
# Check whether test ELBO has spiked beyond specified conditions.
- if (epoch_elbo_fail_fraction is not None) and (len(test_elbo) > 2):
- current_diff = max(0., test_elbo[-2] - test_elbo[-1])
- overall_diff = np.abs(test_elbo[-2] - test_elbo[0])
+ if (epoch_elbo_fail_fraction is not None) and (len(model.loss['test']['elbo']) > 2):
+ current_diff = max(0., model.loss['test']['elbo'][-2] - model.loss['test']['elbo'][-1])
+ overall_diff = np.abs(model.loss['test']['elbo'][-2] - model.loss['test']['elbo'][0])
fractional_spike = current_diff / overall_diff
if fractional_spike > epoch_elbo_fail_fraction:
raise ElboException(
@@ -245,15 +243,15 @@ def run_training(model: RemoveBackgroundPyroModel,
# Check on the final test ELBO to see if it meets criteria.
if final_elbo_fail_fraction is not None:
- best_test_elbo = max(test_elbo)
- if test_elbo[-1] < best_test_elbo:
- final_best_diff = best_test_elbo - test_elbo[-1]
- initial_best_diff = best_test_elbo - test_elbo[0]
+ best_test_elbo = max(model.loss['test']['elbo'])
+ if model.loss['test']['elbo'][-1] < best_test_elbo:
+ final_best_diff = best_test_elbo - model.loss['test']['elbo'][-1]
+ initial_best_diff = best_test_elbo - model.loss['test']['elbo'][0]
if (final_best_diff / initial_best_diff) > final_elbo_fail_fraction:
raise ElboException(
- f'Training failed because final test loss {test_elbo[-1]:.2f} '
+ f"Training failed because final test loss {model.loss['test']['elbo'][-1]:.2f} "
f'is not sufficiently close to best test loss {best_test_elbo:.2f}, '
- f'compared to the initial test loss {test_elbo[0]:.2f}. '
+ f"compared to the initial test loss {model.loss['test']['elbo'][0]:.2f}. "
f'Fractional difference is {final_best_diff / initial_best_diff:.2f}, '
f'which is > specified final_elbo_fail_fraction {final_elbo_fail_fraction:.2f}'
)
@@ -284,14 +282,14 @@ def run_training(model: RemoveBackgroundPyroModel,
logger.info(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
# Check final ELBO meets conditions.
- if (final_elbo_fail_fraction is not None) and (len(test_elbo) > 1):
- best_test_elbo = max(test_elbo)
- if -test_elbo[-1] >= -best_test_elbo * (1 + final_elbo_fail_fraction):
- raise ElboException(f'Training failed because final test loss ({-test_elbo[-1]:.4f}) '
+ if (final_elbo_fail_fraction is not None) and (len(model.loss['test']['elbo']) > 1):
+ best_test_elbo = max(model.loss['test']['elbo'])
+ if -model.loss['test']['elbo'][-1] >= -best_test_elbo * (1 + final_elbo_fail_fraction):
+ raise ElboException(f"Training failed because final test loss ({-model.loss['test']['elbo'][-1]:.4f}) "
f'exceeds best test loss ({-best_test_elbo:.4f}) by >= '
f'{100 * final_elbo_fail_fraction:.1f}%')
# Free up all the GPU memory we can once training is complete.
torch.cuda.empty_cache()
- return train_elbo, test_elbo
+ return train_elbo, model.loss['test']['elbo']
From a71e18362be9dbe45d553c843e786edce2edeb6c Mon Sep 17 00:00:00 2001
From: alecw <alecw@users.noreply.github.com>
Date: Tue, 23 Apr 2024 09:53:21 -0400
Subject: [PATCH 14/19] Retry: prevent ZeroDivisionError if initial test ELBO
is the best test ELBO (#312)
---
cellbender/remove_background/train.py | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
diff --git a/cellbender/remove_background/train.py b/cellbender/remove_background/train.py
index bf7bed67..961e06e7 100644
--- a/cellbender/remove_background/train.py
+++ b/cellbender/remove_background/train.py
@@ -247,7 +247,12 @@ def run_training(model: RemoveBackgroundPyroModel,
if model.loss['test']['elbo'][-1] < best_test_elbo:
final_best_diff = best_test_elbo - model.loss['test']['elbo'][-1]
initial_best_diff = best_test_elbo - model.loss['test']['elbo'][0]
- if (final_best_diff / initial_best_diff) > final_elbo_fail_fraction:
+ if initial_best_diff == 0:
+ raise ElboException(
+ f"Training failed because there was no improvement from the initial test loss {model.loss['test']['elbo'][0]:.2f}. "
+ f"Final test loss was {model.loss['test']['elbo'][-1]}"
+ )
+ elif (final_best_diff / initial_best_diff) > final_elbo_fail_fraction:
raise ElboException(
f"Training failed because final test loss {model.loss['test']['elbo'][-1]:.2f} "
f'is not sufficiently close to best test loss {best_test_elbo:.2f}, '
From 2c3dbe88c660f5e05d979d38c8e99259695b5286 Mon Sep 17 00:00:00 2001
From: aawdeh <aseel.awdeh@gmail.com>
Date: Mon, 24 Jun 2024 10:50:31 -0400
Subject: [PATCH 15/19] Azurize CellBender to run on ToA (#367)
CPU-only version of WDL
---
wdl/cellbender_remove_background_azure.wdl | 250 +++++++++++++++++++++
1 file changed, 250 insertions(+)
create mode 100644 wdl/cellbender_remove_background_azure.wdl
diff --git a/wdl/cellbender_remove_background_azure.wdl b/wdl/cellbender_remove_background_azure.wdl
new file mode 100644
index 00000000..503a4659
--- /dev/null
+++ b/wdl/cellbender_remove_background_azure.wdl
@@ -0,0 +1,250 @@
+version 1.0
+
+## Copyright Broad Institute, 2022
+##
+## LICENSING :
+## This script is released under the WDL source code license (BSD-3)
+## (see LICENSE in https://github.com/openwdl/wdl).
+
+task run_cellbender_remove_background_gpu {
+
+ input {
+
+ # File-related inputs
+ String sample_name
+ File input_file_unfiltered # all barcodes, raw data
+ File? barcodes_file # for MTX and NPZ formats, the bacode information is in a separate file
+ File? genes_file # for MTX and NPZ formats, the gene information is in a separate file
+ File? checkpoint_file # start from a saved checkpoint
+ File? truth_file # only for developers using simulated data
+
+ # Outputs
+ String? output_bucket_base_directory # Google bucket path
+
+ # Docker image with CellBender
+ String? docker_image = "us.gcr.io/broad-dsde-methods/cellbender:0.3.0"
+
+ # Used by developers for testing non-dockerized versions of CellBender
+ String? dev_git_hash__ # leave blank to run CellBender normally
+
+ # Method configuration inputs
+ Int? expected_cells
+ Int? total_droplets_included
+ Float? force_cell_umi_prior
+ Float? force_empty_umi_prior
+ String? model
+ Int? low_count_threshold
+ String? fpr # in quotes: floats separated by whitespace: the output false positive rate(s)
+ Int? epochs
+ Int? z_dim
+ String? z_layers # in quotes: integers separated by whitespace
+ Float? empty_drop_training_fraction
+ Float? learning_rate
+ String? exclude_feature_types # in quotes: strings separated by whitespace
+ String? ignore_features # in quotes: integers separated by whitespace
+ Float? projected_ambient_count_threshold
+ Float? checkpoint_mins
+ Float? final_elbo_fail_fraction
+ Float? epoch_elbo_fail_fraction
+ Int? num_training_tries
+ Float? learning_rate_retry_mult
+ Int? posterior_batch_size
+ Boolean? estimator_multiple_cpu
+ Boolean? constant_learning_rate
+ Boolean? debug
+
+ # Hardware-related inputs
+ String? hardware_zones = "us-east1-d us-east1-c us-central1-a us-central1-c us-west1-b"
+ Int? hardware_disk_size_GB = 50
+ Int? hardware_boot_disk_size_GB = 20
+ Int? hardware_preemptible_tries = 0
+ Int? hardware_max_retries = 0
+ Int? hardware_cpu_count = 4
+ Int? hardware_memory_GB = 32
+ String? hardware_gpu_type = "nvidia-tesla-t4"
+ String? nvidia_driver_version = "470.82.01" # need >=465.19.01 for CUDA 11.3
+
+ }
+
+ # For development only: install a non dockerized version of CellBender
+ Boolean install_from_git = (if defined(dev_git_hash__) then true else false)
+
+ # Compute the output bucket directory for this sample: output_bucket_base_directory/sample_name/
+ String output_bucket_directory = (if defined(output_bucket_base_directory)
+ then sub(select_first([output_bucket_base_directory]), "/+$", "") + "/${sample_name}/"
+ else "")
+
+ command {
+
+ set -e # fail the workflow if there is an error
+ set -x
+
+ # install a specific commit of cellbender from github if called for (-- developers only)
+ if [[ ~{install_from_git} == true ]]; then
+ echo "Uninstalling pre-installed cellbender"
+ yes | pip uninstall cellbender
+ echo "Installing cellbender from github"
+ # this more succinct version is broken in some older versions of cellbender
+ echo "pip install --no-cache-dir -U git+https://github.com/broadinstitute/CellBender.git@~{dev_git_hash__}"
+ # yes | pip install --no-cache-dir -U git+https://github.com/broadinstitute/CellBender.git@~{dev_git_hash__}
+ # this should always work
+ git clone -q https://github.com/broadinstitute/CellBender.git /cromwell_root/CellBender
+ cd /cromwell_root/CellBender
+ git checkout -q ~{dev_git_hash__}
+ yes | pip install --no-cache-dir -U -e /cromwell_root/CellBender
+ pip list
+ cd /cromwell_root
+ fi
+
+ # put the barcodes_file in the right place, if it is provided
+ if [[ ! -z "~{barcodes_file}" ]]; then
+ dir=$(dirname ~{input_file_unfiltered})
+ if [[ "~{input_file_unfiltered}" == *.npz ]]; then
+ name="row_index.npy"
+ elif [[ "~{barcodes_file}" == *.gz ]]; then
+ name="barcodes.tsv.gz"
+ else
+ name="barcodes.tsv"
+ fi
+ echo "Moving barcodes file to "$dir"/"$name
+ echo "mv ~{barcodes_file} "$dir"/"$name
+ [ -f $dir/$name ] || mv ~{barcodes_file} $dir/$name
+ fi
+
+ # put the genes_file in the right place, if it is provided
+ if [[ ! -z "~{genes_file}" ]]; then
+ dir=$(dirname ~{input_file_unfiltered})
+ if [[ "~{input_file_unfiltered}" == *.npz ]]; then
+ name="col_index.npy"
+ elif [[ "~{genes_file}" == *.gz ]]; then
+ name="features.tsv.gz"
+ else
+ name="genes.tsv"
+ fi
+ echo "Moving genes file to "$dir"/"$name
+ echo "mv ~{genes_file} "$dir"/"$name
+ [ -f $dir/$name ] || mv ~{genes_file} $dir/$name
+ fi
+
+ # use the directory as the input in the case of an MTX file
+ if [[ "~{input_file_unfiltered}" == *.mtx* ]]; then
+ input=$(dirname ~{input_file_unfiltered})
+ else
+ input=~{input_file_unfiltered}
+ fi
+
+ cellbender remove-background \
+ --input $input \
+ --output "~{sample_name}_out.h5" \
+ ~{"--checkpoint " + checkpoint_file} \
+ ~{"--expected-cells " + expected_cells} \
+ ~{"--total-droplets-included " + total_droplets_included} \
+ ~{"--fpr " + fpr} \
+ ~{"--model " + model} \
+ ~{"--low-count-threshold " + low_count_threshold} \
+ ~{"--epochs " + epochs} \
+ ~{"--force-cell-umi-prior " + force_cell_umi_prior} \
+ ~{"--force-empty-umi-prior " + force_empty_umi_prior} \
+ ~{"--z-dim " + z_dim} \
+ ~{"--z-layers " + z_layers} \
+ ~{"--empty-drop-training-fraction " + empty_drop_training_fraction} \
+ ~{"--exclude-feature-types " + exclude_feature_types} \
+ ~{"--ignore-features " + ignore_features} \
+ ~{"--projected-ambient-count-threshold " + projected_ambient_count_threshold} \
+ ~{"--learning-rate " + learning_rate} \
+ ~{"--checkpoint-mins " + checkpoint_mins} \
+ ~{"--final-elbo-fail-fraction " + final_elbo_fail_fraction} \
+ ~{"--epoch-elbo-fail-fraction " + epoch_elbo_fail_fraction} \
+ ~{"--num-training-tries " + num_training_tries} \
+ ~{"--learning-rate-retry-mult " + learning_rate_retry_mult} \
+ ~{"--posterior-batch-size " + posterior_batch_size} \
+ ~{true="--estimator-multiple-cpu " false=" " estimator_multiple_cpu} \
+ ~{true="--constant-learning-rate " false=" " constant_learning_rate} \
+ ~{true="--debug " false=" " debug} \
+ ~{"--truth " + truth_file}
+
+ # copy outputs to google bucket if output_bucket_base_directory is supplied
+ echo ~{output_bucket_directory}
+ # commenting out b/c cant use gsutil in azure
+ if [[ ! -z "~{output_bucket_directory}" ]]; then
+ echo "Copying output data to ~{output_bucket_directory} using gsutil cp"
+ gsutil -m cp ~{sample_name}_out* ~{output_bucket_directory}
+ fi
+
+ }
+
+ output {
+ File log = "${sample_name}_out.log"
+ File pdf = "${sample_name}_out.pdf"
+ File cell_csv = "${sample_name}_out_cell_barcodes.csv"
+ Array[File] metrics_array = glob("${sample_name}_out*_metrics.csv") # a number of outputs depending on "fpr"
+ Array[File] report_array = glob("${sample_name}_out*_report.html") # a number of outputs depending on "fpr"
+ Array[File] h5_array = glob("${sample_name}_out*.h5") # a number of outputs depending on "fpr"
+ String output_dir = "${output_bucket_directory}"
+ File ckpt_file = "ckpt.tar.gz"
+ }
+
+ runtime {
+ docker: "${docker_image}"
+ bootDiskSizeGb: hardware_boot_disk_size_GB
+ disks: "local-disk ${hardware_disk_size_GB} HDD"
+ memory: "${hardware_memory_GB}G"
+ cpu: hardware_cpu_count
+ zones: "${hardware_zones}"
+ preemptible: hardware_preemptible_tries
+ checkpointFile: "ckpt.tar.gz"
+ maxRetries: hardware_max_retries # can be used in case of a PAPI error code 2 failure to install GPU drivers
+ }
+ meta {
+ author: "Stephen Fleming"
+ email: "sfleming@broadinstitute.org"
+ description: "WDL that runs CellBender remove-background on a GPU on Google Cloud hardware. See the [CellBender GitHub repo](https://github.com/broadinstitute/CellBender) and [read the documentation](https://cellbender.readthedocs.io/en/v0.3.0/reference/index.html#command-line-options) for more information."
+ }
+
+ parameter_meta {
+ sample_name :
+ {help: "Name which will be prepended to output files and which will be used to construct the output google bucket file path, if output_bucket_base_directory is supplied, as output_bucket_base_directory/sample_name/"}
+ input_file_unfiltered :
+ {help: "Input file. This must be raw data that includes all barcodes. See http://cellbender.readthedocs.io for more information on file formats, but importantly, this must be one file, and cannot be a pointer to a directory that contains MTX and TSV files."}
+ barcodes_file :
+ {help: "Supply this only if your input_file_unfiltered is a sparse NPZ matrix from Optimus lacking metadata."}
+ genes_file :
+ {help: "Supply this only if your input_file_unfiltered is a sparse NPZ matrix from Optimus lacking metadata."}
+ output_bucket_base_directory :
+ {help: "Optional google bucket gsURL. If provided, the workflow will create a subfolder called sample_name in this directory and copy outputs there. (Note that the output data would then exist in two places.) Helpful for organization."}
+ docker_image :
+ {help: "CellBender docker image. Not all CellBender docker image tags will be compatible with this WDL.",
+ suggestions: ["us.gcr.io/broad-dsde-methods/cellbender:0.3.0"]}
+ checkpoint_file :
+ {help: "Optional gsURL for a checkpoint file created using a previous run ('ckpt.tar.gz') on the same dataset (using the same CellBender docker image). This can be used to create a new output with a different --fpr without re-doing the training run."}
+ truth_file :
+ {help: "Optional file only used by CellBender developers or those trying to benchmark CellBender remove-background on simulated data. Normally, this input would not be supplied."}
+ hardware_preemptible_tries :
+ {help: "If nonzero, CellBender will be run on a preemptible instance, at a lower cost. If preempted, your run will not start from scratch, but will start from a checkpoint that is saved by CellBender and recovered by Cromwell. For example, if hardware_preemptible_tries is 2, your run will attempt twice using preemptible instances, and if the job is preempted both times before completing, it will finish on a non-preemptible machine. The cost savings is significant. The potential drawback is that preemption wastes time."}
+ hardware_max_retries :
+ {help: "If nonzero when CellBender exits without success it will be retried. If one also sets the memory_retry_multiplier workflow option, and the exit happens to be detected as an out of memory error, then the retry will also increase the memory allocated to the next run. The potential benefit is that one can start CellBender with less memory, and memory will be increased only when needed. The potential drawback is that the job will be retried even if the error is not a memory error."}
+ checkpoint_mins :
+ {help: "Time in minutes between creation of checkpoint files. Bear in mind that Cromwell copies checkpoints to a bucket every ten minutes."}
+ hardware_gpu_type :
+ {help: "Specify the type of GPU that should be used. Ensure that the selected hardware_zones have the GPU available.",
+ suggestions: ["nvidia-tesla-t4", "nvidia-tesla-k80"]}
+ }
+
+}
+
+workflow cellbender_remove_background {
+
+ call run_cellbender_remove_background_gpu
+
+ output {
+ File log = run_cellbender_remove_background_gpu.log
+ File summary_pdf = run_cellbender_remove_background_gpu.pdf
+ File cell_barcodes_csv = run_cellbender_remove_background_gpu.cell_csv
+ Array[File] metrics_csv_array = run_cellbender_remove_background_gpu.metrics_array
+ Array[File] html_report_array = run_cellbender_remove_background_gpu.report_array
+ Array[File] h5_array = run_cellbender_remove_background_gpu.h5_array
+ String output_directory = run_cellbender_remove_background_gpu.output_dir
+ File checkpoint_file = run_cellbender_remove_background_gpu.ckpt_file
+ }
+
+}
From 569b7b1bf77e622706bed2d1a6d25eb9d76e402c Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Mon, 24 Jun 2024 12:57:29 -0400
Subject: [PATCH 16/19] Updates to benchmarking scripts
---
.../remove_background/tests/benchmarking/run_benchmark.py | 2 ++
.../tests/benchmarking/run_benchmark_result_tabulation.py | 8 ++++----
2 files changed, 6 insertions(+), 4 deletions(-)
diff --git a/cellbender/remove_background/tests/benchmarking/run_benchmark.py b/cellbender/remove_background/tests/benchmarking/run_benchmark.py
index efb8520a..ab31ca3a 100644
--- a/cellbender/remove_background/tests/benchmarking/run_benchmark.py
+++ b/cellbender/remove_background/tests/benchmarking/run_benchmark.py
@@ -141,7 +141,9 @@ def cromshell_submit(wdl: str,
submit_cmd = ['cromshell', 'submit',
tmp_wdl,
inputs,
+ '--options-json',
options,
+ '--dependencies-zip',
dependencies_zip]
# submit job
diff --git a/cellbender/remove_background/tests/benchmarking/run_benchmark_result_tabulation.py b/cellbender/remove_background/tests/benchmarking/run_benchmark_result_tabulation.py
index ffc5e5a8..e5c997ca 100644
--- a/cellbender/remove_background/tests/benchmarking/run_benchmark_result_tabulation.py
+++ b/cellbender/remove_background/tests/benchmarking/run_benchmark_result_tabulation.py
@@ -69,7 +69,7 @@ def get_cromshell_output_h5(workflow: str, grep: str = '_out.h5') -> Union[str,
"""Use cromshell list-outputs to get the relevant file gsURL"""
output = grep_from_command(['cromshell', 'list-outputs', workflow], grep=grep)
- out = output[:-1].decode().split('\n')
+ out = output.decode().lstrip('run_cellbender_benchmark.h5_array: ').rstrip('\n').split('\n')
if len(out) > 1:
return out
else:
@@ -95,18 +95,18 @@ def metadata_from_workflow_id(workflow: str) -> Tuple[str, str, Optional[str]]:
# git hash
output = grep_from_command(['cromshell', 'metadata', workflow],
grep='"git_hash":')
- git_hash = output[17:-3].decode()
+ git_hash = output.decode().split('"git_hash": ')[-1].lstrip('"').split('"')[0]
# input file
output = grep_from_command(['cromshell', 'metadata', workflow],
grep='run_cellbender_benchmark.cb.input_file_unfiltered')
- input_file = output[58:-3].decode()
+ input_file = 'gs://' + output.decode().split('gs://')[-1].split('"')[0]
# truth file
output = grep_from_command(['cromshell', 'metadata', workflow],
grep='run_cellbender_benchmark.cb.truth_file')
if 'null' not in output.decode():
- truth_file = output[47:-3].decode()
+ truth_file = 'gs://' + output.decode().split('gs://')[-1].split('"')[0]
else:
truth_file = None
From 31130fd9fe5405c891d4c1efa25aeab9860a5d5e Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Mon, 24 Jun 2024 12:58:19 -0400
Subject: [PATCH 17/19] Update documentation in anticipation of v0.3.2
---
.../_static/remove_background/v0.3.2_hgmm.png | Bin 0 -> 189077 bytes
docs/source/changelog/index.rst | 82 ++++++++++++++----
docs/source/reference/index.rst | 3 +-
3 files changed, 66 insertions(+), 19 deletions(-)
create mode 100644 docs/source/_static/remove_background/v0.3.2_hgmm.png
diff --git a/docs/source/_static/remove_background/v0.3.2_hgmm.png b/docs/source/_static/remove_background/v0.3.2_hgmm.png
new file mode 100644
index 0000000000000000000000000000000000000000..a9ca57fb350324f0ccdb8c8ec1d84ac5f7c480fd
GIT binary patch
literal 189077
zcmb@uWmuF^*ET#R3WCytgn)DjB1kJDT>{b}-QA5SEg;<>B_IvL&_j1fGr-UxGBET2
zL%bXB`+48v`|<sIt^)_Au6<o=uf5iJo@;GGl@ug#vB|L^5D2ccl(;ekayJ?Rxixw3
zCU_;dp3xWlx^FL~=>&n0;$Q!}(Z_Hg3V}R@NQ=Kwbx%VgJiH0tU19Coc{;_8Ix+{}
z;a_F=z`v>x-7fjYB7TfzM#MjxDTO6uI7*W11JQeat%r}^<t-7d$<&JDaV<#<1wG*>
z{C+b}-Pi8{k}w7_Xj;WT<2%o(-z+AOe6fQ>4wzO7<~Y8hxD77kKL-sR+w_h9c>-zC
zeQt{NKhGX%{r^Ad>rc`(H8oXMj_EYK@A~wUQmHG681`SPIIMyVnoe~4r(4*NG$az4
zoBJ%bmk9Y9oHHaum3FI3t3)ll)391Aw8s`o7_C(@Y0jfn!fStn<rx>(h+)OFHE*>R
zB~s(ORA-y*3(>eR9O>H7Yh0;SQrt^VTdkG;?c4VWOF{yIfG1CCUkg8EZl=bgP=YMx
z=H@<ra5rZ}gB*r}KwiV?cun(_3*8sgizYmqMA+E}?cOsF*h0Izx_(YrrVf%IHEzDi
zQ<h%PESgBk$S9q#bP-60-NMO5BCS8Swgidtkjranr9h|j;BYuYY-q!QSkc6fygUmY
za<W#9l1cC}+uK4kG&KDDbwufJ2m)t96Oxmq#Vsr>2w=e&J+oD=#}A4t@6-P_=aI<4
z&$%!6JGEG;Xd^GM<KVr7)#iu94LmHOzNqie%sjCiiHKq~*1LN*AWv0a-A)=br5!cr
z0W*ij%Ngm>71(kX$QlkJui*rCn*MY<04mrYAp`6LyG!!CuC5N*{LEFG$t@~IuPia?
zt(qE%TUNriZ;RfEv4?VxQwI~$(gsBOZBv&Wfu&7;3?%>NAE~UYtj%G_`{q;jV*hc3
zgspaOZ?C4NCdYP(UfGm64|zmanL!-K8@o!e{ES2rf?-LIk~~!Kt^ZzdR#XIn@!7N6
zf{!2&|JmCcRYp)!i)yW7+=LO!YH%8foahkl7cXS{S2@`fJlv$>1?G3yFv_Z?zOp>Y
za-pH2YDFK)re>{qd2hDuyzlGj`SwaMbi(pO(FFK1Z)!wD#w&`*;$k-1n2h_eNlE6<
zAIHSTMo7^0N4(1$F=r!@m5~V>xOhPDx%FF246XBWa9d~>%WF?h0k1UuoDuip*|2dk
zGH!0}n3$MiwIYT5D!4U2KR;4JVdGdWe=K>>baM43;UNPn(O(aXrHNCfvg*SN`$d1T
z(=1gCvoDv@&(oYYOfH=t{@%O`VOsfe_h9<M8JvLDkDVPl6N-{0TiG6(vF2^kv*S&@
zAph`jb$uQ9RJE2354rAThs4~od1-?xxVvA#<kpdRTv{5N@Utd86xD;-+YRRH7gjxn
zTsRjZ)Q5x-Pi99qosreupYv4H_edHtLl8a4sPq2v?|hnX%7CjtH*y*$NB^y*wP5A#
z{FBbtg<bW&Ca6lz|GB)`+ug|NivzI>fvr(+oRcKQOb?qNZ40@}bhe9eYEWYOP2RBE
z22o1PM;+pu<R`ALD1gscCrLBCVH$7x^^3<f1P&?c?zi=uBFlOTH-E|zBAcUB@SoA8
zM8y6(@hz-WoOfkjPR=l$cv%**mMW7zhrCYSSb|Si?)?hsG&G1^xh?qZHIusF)>nD$
z4B&HvA*%r&O!aM#%;;VrX*d@ujB-_NZ?uFd<XQ%bAy(FD<3<DSd>`7i<q;Ir&KX%<
zU5!gN8|{z(H@5ml&b{dW-ufbO6zQ$Jlx2NC8C{3?djA`czQ%j=ZY63(2wT$43L)M<
zYDE*q+zu+!yBsic=<K#@gA)UI>ErYADQgyCX3WP54K@;-S&~$-`P0ukO=X>iOX~#$
z3t?Vg=ESlss|EP^;TNoYiT)+7jP$-;mer1|Ya{7owYsKP4xz!b__i0Ed$PE;WFk>^
zEH36cu3_>c^d@NmgG?J-c}}}zWMtwUJL`*`F%GRgE!t`~*1b(D$Z3h9l(C|BkgwD$
z())kc@y5&%%9~>arr>RIh1-o63AFdAG-`Pk)3`RI+l+gp6YAft%sa~4VV-uB;m7S1
zyfcr8p4MU27^bg@xMUrGxY@4;>)$YO_Mme5=KdNX8^s4flhBQqvbw*GL1@<A8yJ$Y
zOL2d@Jg8Jqi2jNAHkE#-^5U>YKh}}!`*`7!Z|K&!!@E+#f05MDA~;Xq&r_%_mng9@
zEuD~6gJ$i{0|>;o+Y7C$pn!K@^sG;wqaDL0ZEIWVl^G2}Nk)=OJF<|c3XYujaQ8mu
zWn>ISE020$v(#`RlgW8gs}D=c@4+ee!6M-hRudBoIH+ph(QgAmI(JT&^B3tVZz2h4
z)WMnO5VS0u;?=QcV)x#cQH-zoSBae8AL7)Cjx^YcpLLFvWBr2PmL4h#EZ?96A*!S5
zDCx6uRhoP16V#`V-2rHP*@cI+bRxN(^apF=M!9^H%C%oMhGqrY+4I;lvQ8XZkl&g1
zUpqh5*79%}re@D~Us(M%$!w1=pi(rk9Jol~&v`UATQlMpNS1^bNwZ)|{K%5MX?uUJ
zSL8cYW&yD!m)LfFTb7R5($Iq7md)7e?otbXqXnP$=Fm;JujfRWbmQaQi}6SG9WG36
z=^uGh-vvQCHY(!X9i`bg$oL!Ll3wVu<bDMX`t<op@H-8@r#$toGd}ZW>DCqRXbojy
zWD#9yX%w#S*5;S1ya!BQG<;Pmny~(1nKM!_wWTQ{rf+M*`vgR9uzJ+0277x)wvIq3
zK2@3PWZH^ooWTx%&c~N%l09NR9rRU#4nqi4nftoWaev&hx`#AqZceAvOF>q)zqgl*
zJi<CzYF=!8*#HAF+mUC}fvS!emV)^uTh?JS1-Y#g@^^W<p`>K~SVfUyls+U^?s-S0
zp@oL$Ubq)E2q+YP@87>KGx(-!Iq?-SPlN4x*tj)s>Y!<p-Z1x4Q_{iSQ%4QY9YY0F
zN7GKP;0T2tFC|q)+3n>O2kyTQ!h_xR1(_#4W2)QHEJWoAlS^+A2a^5FqkPgRqG^U-
zTI^E-d3kx#I@(>`o*X$ghx!I*4WCz}_^CVZS0~K7dh8feEA8_PvGQa*wwxlyTNA)>
z^F?Tq$;Y;D++xf_>{bZk*khQR^kdz*7BPi1&Y?x)E;sqc^Cm3ovR{x#2&5}C)_A%6
zXHo=769b4xi>~M7+etU_3(5DzdgUy=PjkMvoQHm|iG4a!oI`MMs6R%sWS%rHfUxc=
zzkbP_<zpcHOI>@3mMBFib#1_<U`PK+t^C2=E|+Y9^dX<EMpk%**UCGRx)ou2Hlg{q
z49?$u_<3}2mna<wsqxm`b$X&GiFVfSF{@h_Q>^NGlxKsPmKjtTQXmOcIj>A#6Rxu6
zeIlr_Ga+~UX4Rw$-hl&8BpjUm(N0%0MXS%fHbF*`k5PYPyQjJLmu789g!$&=sDO8o
z*QGgsU<%u2dO(!q8&}~(1n*hd^NX2Q{sVWBFhYxKZ~o%Ad+c9{?zHQ=e}2QIr@qvl
z388u0)NJjgeC5NmL)Uin4qMB1KJ+5xq|-ktKd6VkRC((D+-pV@=Z_(DT0Pd&mDPRs
z?Pr{3S0AL9Z`^{s_TP(wR}6i|Y=&Ok9z?XYV0bc5plTcVBpW5qmxnVRBW%AIL`!)W
zeDkt={fAtA)%D3vRpYFoC~+lcSR<v7Y>U@+o2PlT)?kBfa7%=#vl-d#ELp-tNtrsk
zP%(P8!+SrYf+bcPl21H4DE8jng<r_barUxRfOu5Ps=H?_XM25=9z-3JM@X~=-g)N>
zcQrV4zD26N)PI2!kNnI;5Y%%g8%ZcHF}4)l!KF{;(&fWhJ3<>pzu8~F#rTcAgwthM
zkv(52mcB`4*CAR7ex&0%`v=u_C^^I-lo7Qot|e>NAe2SXC=BAji0kbY|GtKQW~*IA
zMy1CiczQ+b31)HM)mODZGxd(FmPv0lr;*bp*2?HIx>Ni%tEjVkcig+SHzTb#bXoL`
z>egL|f6K~dBixRyQ9OeOpZbvjOow0J&t6@XR{gw@^lT2br0lAArsf_@Fl`oC(<#6k
zT|~I2&)He#ugSUfL5QQM%4tDQ+;*R_IhE9yntbvioSMA)m)WgOPI8@f-k4XHc_bNM
zMglR+i~VfwJJSzym7BK&a)rgVF1;FnLBzd@CjPn8TkNRBs>-bg7_qhqxt$<9=9l7z
zYow&ti_+<OQTEOndrUCj^RbE-`;Rv@Jimyp64dXKe)nIS7*i-Q^L;&-{G|j;a*|bL
zKdiZz9zVsy@BQl9CJPPMHo@b>yxa@FA(~Dero$*zcSs;0xtgwncYjJ@-|<*Kkm%uK
z2qYjNzjNaZQ!IGs>7%7ZzD466ysdNQMcN3Ps)!2F@2-;N*ifuc&}Fy5ql9ZaTo(7#
z)Uq<}-QJnPqOx1r@TqNCX?}Ip*c)Fb1!Hv{KX)_NAZLcoRlR>PA1;fQ%Y6`7{=lQG
zz@4x4ulIiXof!}nmB&1!ydz_x-31>v5V$C;%MK<+YaFi|^H3_^(-jtE{HAM`hV}Fd
zLrQc;>Wg%R#Pu%z1OMt@c$r%nqFM^`2y~5#eFoN;_9we17rIH8J@EAIgZClsR+~17
zjB%g8R<*C-u86JBWx47u?Y=czim6%_WNol;Jh5NQ{q-cw+3bd`DE}z1OyA;4VXJme
z7!e!VJE-T4aRsxxBY!S&;@7kbb(*zwBm}akj=vhjzFz`D7`!sMxVZT1*RKNu105aW
zRqz2slcG3aAstW(!G3mOVT5Hh10$o8;*_Jdw1NVOJ}PZzj%C=`)U*<pAI2PDLl4MK
zQAw>h(=YbpH?}-ul+cs<gH(ptUSsY$(lW0hjv`49Hm+ml8UZiB-X<(TqLC)idmQe@
zs?00i795mXmO|REY5sxD3^^K>=N26u{l@Kwr$O}ewrhgkv$vb64B214+}`L*OHWVF
z$jFF^F(lI4eNWqf$cT$GhRz~LKd;a`n%3QxbHl@?Lp=@qNQS+}fDw-!ahAG2VG9lS
z?&ubXO-M)p-%U;~95H`z_pPHNxB3&4)vk)!ZR9h@rAE3NRu-MmHM!g!=2^#>*Rg9h
z`QOsg(vp*tqoZ4W-o+{C38VA=B@RwAG;U~Sl$cP>b>?WSh%OH5U)f|E{}?pma;Mjl
z7U?QzhZ?xKvcESr)MuO06MUUPlya{#NN@CfnRMOb-L-N1t617QsmaL}(Ahe;LM|4n
zzxQ2s=bG&GVqWeh1->h0KM)B;8R)*+Vyq=@0IZ_dSVl?;#B^kTIl>mIRf6={NlB7L
zH4;ZhhMk<Td~RLbQ`ynRhlQZsX;?GrdOqB=8}9aLKYU!%k44{=qgF&nNVx4P*dRZ!
zqV~mguIQ~1bhfIhYQl}-;jwFYBM&)lk?7E-JVR{j$KRIKbETpBK>~hCXFqhy(`6K|
zj}5AYZ>_e_{M=lP6165h&b9EoPQwbV5)dQ1I%69TWbX;)08(;dQV(TJX#{yC3bSUE
zE)DR(z<{xOJ7D2mzaMfQV_*k@cR$n6(S62w3IZj{MR6}7_BQz=JHb0(L8PbK0P+<R
z<LK!4Eb@>aYw?3Hy29qr&ESmAjC6FC%uM-zq6CkQYlE(LluZ9WtIIK{$jD&Z^#Kyc
z!Rm)B=9G`+#2PYc$WnD9Sv-!VtI#KeJ>Kmu%u;Qw)Qug*V|8~7B~9ey#}tZrHcsDy
z_%GVSjv}ka#oQ21Dn1VP#XV(?-vy!>JffVqY8ON~FQh3FefBI_N9l~GC}^Vu>lB~S
z>@6wH2jd8d40gLFY*`75uz{RxLP5SKxXZ7Ha2#&%BxQBlh>5PXfUB50tqSLE4$Rgi
z<Z{;h(YwcumwhiPKckUQUJqoDo_f2lD3>7ZN~N<^XYLfvME>0C<a@4zHvX0N)hm`h
zN0ky!($<z^O3KqkqQ786<Qk;QU9BP8R!*5>>fCV;Yx)WM-aNa3g438JT)d`I(l2Gr
z{nFvZwSGw?x^iBgA?iAr_@g)R*2eC*Ez`s`Sa+t$(<iiZ^bvhI$5$fNDuZO68^)(r
z<cQurOR8!YC05o;1ftUAx@RUOCL-{Hn{Is4iiaEsDHR$T`fV4lx-M+HO0&|bqwN>M
zx)+w`{0I(DwrEk!?hVA7_?6T-;<jV<SNze4I=aiEW`wO9tTi}Xa@FFVtrGz4$!_<A
z5l@}JUwE1~*24!cT1h^)0@9RbMUL=zAbxI2Ry1K*p#cIar(%s|^$gHHN@u(R5ibn6
zlb?x5_DcDQwXTKq2!>P=_KS>w=sKE1+}qkM(7uzlydNE6Gqm=G-GM`p0AwAXu~E^T
zeBfazY{0`9XlPokdCiQCjm^xeQi*csGU>}QFr&O@)z2^Dss!yaf5l8|dCsogUt7$t
z`i%Yz0xb>KGqewo8K{bWWxA@ljy>C}qD@bCZ^)9m_N+1bPjSBT=Y}n4mAA$z<oehN
zAAF<YCxv<S6*%S}#P=nitI#ynbKi#e`#Vrp#X1u2jOIJWT1j0kKg(0MM`4;$d}WW-
zp7~C$Pzcj5Jnt-fF>MkH)hDR63qq=MyUu!>n*R9BqqrPp=%H<{eB5u$4X2l*?K0rf
zNMjguNbXe-I6h2+M#<NG6AtU}xf8#a#3)^~%Q2PHJnpTxJQ#o+>@^q2Ra(tf$sGcM
z)%5o9>Z&o2G1Akw9kl3!+#Km0I6A4y6B1UU>KV({i2C5O3ZFo*sH&{|rDY={Bjf2=
z57<(Y^x|w6g!Bh3|A8reGkc;n{lDn9e{)6Hn@Y5Y>NSo=MSpYYH-eas^FWccd;x|W
zix`5K-=Cc3OO^haszXK35VpS3>3>p4bJ)m!s#C7qJX2m7&-I$q*)h~^symIWlQp-o
z5b(m5g*Kf__B_()i<iA1t)MR^1ql^cYl-)gUy6%tRTvST1~KQw9v{mF^dl%Nt5b9x
z_l7JvW3r+|#lt^FOI;C#e|k~{g3d3k5`fIRrjz45<b)ilgyV*6p^WtO$#M+o3LD%0
z=^eAR)AD72cj&-bd>?hbI9wA~NsBvm&VBUrh4t$<Wm8VEUNQsU5Lt3=&C;G7n2T2J
z=DDM%Tz+<28I1`le|&&Lky@YGG$4k8%sWrZGH|LRE;;1<R*@M|hnC0>|C6<E{jIRq
zaKXOV!e*wqZ!=COx?PO?A>Bt3h?j;_HiKJsyY^yL3N_V1ceppcyFeTeXW9NFm6MTF
z`f+$%dew_Nz$`aBA&_0ZOD@4ESd7DhD8HUC14EvdEAStmTDh}_J=`2my%k%+Bjlpz
zxCA$DBVBKCybWFI7386+SgU9}BWj4qwCT5nnwgq9F7%~U1P~PZg8U9tv7!kU?(?M&
zK)rbRPYktBY}1R@Sc0!rAN+|~#(FRGR$qZ<zjAX|O6GaxmYjsdU2UZ(cAn(4G;2}~
zWISnuRg`X^*E47To4-e6M7J`Kw;NIY?#t761euf_cPlb*@GYL)(n;!L*kvK}Sh!wV
zz2Wlni1bQ3OPpPoz>WuN5#ksNGZQ}f)%H(LjN`T^)9H8C;?O_!2_Lrj<p)(x-ETat
z(mi&qm04Y-UFf3ibaKs74Ln3IGD9FeD@RG`nD|Gs$O&C3<$5zTW}^x5ts6%^nvwIF
z-eYD#)-M2Tfb2uP<P)ak%{C$YsmBLvneVZuw)t}6h--teuyA_Q@z2+_bQvJ=KxYAO
z#Y8*SQ4afzS))xJB=f&`{J^>_Pt@cicQOGXVTqcRnp#3fH=}O3P3S^GC&{S1l$8Da
zm&VUb9V;Jl{>umA$`begKJeNf@^~YxwocH!7{bKPE?=Vd<iTBL=u91T2LODl!}W{}
zYpcSQG9PF*8O5H?N?Z|}wg|imZ5QVLm*B`MJSG$zfuv~cR(1m{teC7uv2QU|jn~}3
z<C){)mlMZNeXK3pj2`f^Q+NF;sr%t^jkI_ABjb7?35`~&d$IkA%LK!kAyLGm-77di
zY+0>jxPxCfVVOnDI=#Yj->r<L(=-<UG!Ca`-#Xxrp){LfV)gNWY1JWa?|fLtu#C=Z
zzX$Hi?sVeALr-D57HdllUggEcdvS5@Uu1%gnmM-KRy;tA^>gsJR(Cgy#|@eqsaTia
z+_R)tQquvdDl$Vb%eWEKQ{HMa|8hOI?0F9!?tM0&p45k|kLEiYbAd!!oN}L73uHCf
zMH?IZpQp5;fs^<LCQkFbl*;A$slsY6?6w}@yll03BYb>E`79-=_&#t~DmrI<c8-ti
zCGD9mihBLb<JWXzH>NhpWpP%*kMxJAqbuW@cifbg=fqmw!5Z>o>LS9;c9HH|CkLut
z%uq_tg+y<St>L)OJms+mR%D`A!;+U9$;t89tgrks_I_W+HK8a!<NWoV9r}@@j$MZp
zv+C&1!lCcznlBaDkzE0S&}@}9XZl>>YsIIz+;3dqg?|W8a{Dv7P+k(Ef3q^Oujz%6
zI)T^|N4$Z=l90Gt?UUv};igRv_iJ~qV|$)DILC)O87)Yf((NuzP|kI8enb&{*i1wo
zzOd#4u=igHAc@I7jj0mcX-GRMzlYN+Y<(<ZS5Y3eD#Sq-C+cSIj&(VA;P6WKO1kcl
zlrbhAZf#1xqPJf8ke1Vxj8S#gXD&@$KA5O65g|q+UHk$m`(3NVRTO(Uusyut=xy0l
zct<Z@scgVC@1XvhaiygY?W8G_zo<7;9Eg7vgQR%<bhhg|(T^zp2Eg(Ato}pszq!4J
zSAX-0YN=)+k6`#`&9<fBp*DIm(Fraxc|tl=i?kh|H*)BvcVGXXDWPLGUxAn2Folop
z0Udnf{SoP*9Ske0j#3S-G0|05K(yK-gFx<mZD?k+-)`0`8}7S4Mz@YYIb*JAnLHv0
z-IjteP{|oFHZdWKkjNQHij8g0xZ_l3Y?e6QNcWT?%ZoaVFL!Dy!mWHKA<eh>z=V}b
zv>#ihfX+EYFQ7cak|zZKc%V^%ipI{?dmwvFhg^*DvqbHMFxtaVubG-S1G{ut0j5b;
z$RFFBcC&Cr-){K0f2{0JoO9Bd*1jq7swU_o5CXt_%z`B70&8BXTx({mk~QGf*8?vW
zsRBQrtITRNS<Z-Rw-A>A^_tKajeBz%b$Vrdse{(2+7=)x>WpF>A9*;CHh@o?R>9X)
zL~MtOe!Mj-Q2v>GDA1lq#KJMT@BXs!&-pk|m`z^nybws|01#=r0YmTJqNTL~W6?kx
zL(^#aWsyW0L)Qkn&#})M*U$ouc*875?c}<&=N*T*j;(bEd;XQ>v9Yn5F3T)Wo?IEC
z6+9a{A^Pek8uymZb-NpFQqm3%|F(Ec{EM^uafwr9q52PAox55xIi|PN>y{3IS`a85
zoQD?!yuP~{qpZX2cx)AT`1m;la=TyG$W_eAuVBJ*@<zCETxx2+AnxraH=GrE1!(r>
zG}Iaq_CPlIon{D%J9A0toZ5RQn$cb}WP7rYcKTufIjGx7utC(Jj~++4-$Vyj0Y%o`
z9E6;MEHFLbzS!u)-<&tx;7o|`=A3ApP&9Y4{t~%!$<3^K3QvkW^nRvrMu!?=%Mw@_
z=>FW(Z-Ohi@$jsfe^WkKQct*xK5qz=s1hXp$bNr&xFoq(;xskbIPo@r?Eb}Tf~16m
z_cc4cw)~w=38T!Q0)@8l1XZVw{G(e7ymyP9`}D#C0o8WB&e<D7-M3O=V^;_owMtA&
zZf}v!L8*W&s4V8_zMv^J07na~e;Di7J8toqwLKY4zk`fJ11Mehcu}O_q;N^;kH9B4
z{1-L4B&kY9^ed}|LL`Y$eoTXy=YwlaLTs(}1zTIWiGi+`>*k9XEjAKc+eN@`BP3QX
zH5&l4o9#R}Oae856gdXoEVfkSz?xv{KTC$*pLW7(yq|{CXW}7)9`K3kHZj$N3vzHh
zuBj|{N;y0{1TZ;ggvkdNYm>%7j)5@MJ~4tj+?&RjS@*8|KP7(oFHGqU$*|;dOdzex
z=R_xK<kZKk79#$Tom%q#peIBNX>mW^VymHZ6HB@U=uFg1sq<6TwRBx4pGhU{2!CW}
zXV$v*MCIkfY36TZOAOE;bsYc<pOXAQQBp{r|6eb>@E6D=y%v&p_jX($VTd9CeYE3}
zB)am63XMin8L||m`R!MI8feM{aaNg8vM>f`uEj=v6GdFKeC*uTZGirtKc2P<m0a){
zQ&!+WKg9kgU1-q=-v#*a(}sX4*5bp8iVE=#g!@AX<Vy{xY=h%V;2l8-6Q;7N3RGk;
zbl{aWvKfhgZ$0{iLAJn<u3n(>2_Au)HU?vXx(?N^h=1=t#Y4kCQNy>7<Kl=K5T#?A
zyLm(q|2MPu{=G|OFo5SB57dXaxDI*yey(IWqz_oDQ(K=2o<Lrgf4BKhaa&*fHW>eC
znpOqWZ%cxE)DV>#Y}PlrLCq_%a|Mic>%XHd;tZH!VLwlnyZgBn@Tn5D8g5~21yD1_
zjq<tix-Rg)_ha0iMEP%g@bKEa)!X%do<Oqh82zt2>7gnGB=0}H0RK7p|JIrRx6VdS
z``=!H*KmVsEg)Fv&Z`zpfC$O72Zo4KW(jV`M?QuS5t0$M>rn`M*S@(pqh)SD|EciU
zQlmX{JVuU>k3%}{W-)W3GNbZ+#Kdv8rZ```T<R!P5oB9W7j)BPOAv)#S-@Q5oAi@1
zspncCkncjzA3tDbX6CT7<RPzuQ^dK0v1FL%%&;J9DTB6Netkr=f9EQU;E$_P_c}?-
zBB`m7O#iv#7RIDuJji!BW)|jkz2NAdF5y@KkAc%(A3dMB1cA~J{x$S;c(pgoc%n#h
z_|$3LdcJ(=%Y^4a7jC#oFfIJLaRbcvQ{Z!p9MN1{;0NJ#_(*0qvo}^VCE^KhtzY?@
z_8@Up<>j8Cp`FPl?ZJ~2VpcWF$wv6eo}EsueYC8O&T<kTYL?e#woA3(_jiN#_c;>j
zB3^3d;gV5Q>6h@sH@&^`sk|jRbu!L!auMV&@8XijSyg*JaNL=z69<1qdTm$#kWW=B
zR;_bb=~59nSv0+5-!xe^+!)d=Yr3qL7!%>(5D{SbArsG1q6MWoZNrgGU}I+CSik65
z?hJ2m-c+tI%<GFU;<253byT+@#}KP@iA&DADl>|ht8+?s8@%{Dmlgy&-|v6>P?(5P
zCCk)%(C=bvPhY**UfOSYwYMiJsY^bk+|PEtKHf)&n?WKqZZ*vjb1Zr#{C5EiS*le{
z&U(clpW?hVZ*%^{&U&u)y}4vSo=K4!^Jgs4@vOIqdYAS#)mFU!c9*QSfCYrz;J!Hu
zqw*T4v?4_->{P6b!A{N_Q0q3+<rN+Wrlg#@zWcp5W^1l`FFlXv>dMv!G|Qiwfiv06
z);RRq%s1Le{>W=+^w>T-G65HVzzLJ{xN6!f^Gd2i22-!Yk`nZkR68y`mJc`Gu25Ia
zseE=9LAcdc6Z@7^VPRoGxYW(&y#yP42RJwnsYJYYcb7aC{H}U~shfpdix2d#L;zOx
zzj4JH5wlE~ub6qV*L^B_In7vzR>blBsalb(&L#C;NrT6*<7)ZWhOV}WPE}v6=|I8z
zOoh~S8{4wTCwxYeY#VN+x-egGiAIcyw(#R_Qho1ezk?s#4!8cf7G&-d@I!iH5+b3~
z)z|%{y1IgHzK)A6Ct5ERGev~GCXwKKGo^Yy`@NqU(8%p6^oXt(`bb)aoX7lko@}D&
zP@2d7Kyu@gsZ6r-!^;b-HdHg$-}S*%@p0rOroja@->6wd{?hV`q`1y2Ki{23JGbrm
zgUK(U6GiHN2CFz?d&PQ9rXw+(_)GY7Dk{`Fzq6DBu@{C?p{Cn2Io*+YKYo;I)r}T4
zVTL<Gt)H@eHE;31foO1zDAGJwkJyroTrbs!?L>(V(mPKHjMuQL3p)R?Kc3^K{X0@J
zTSX^9L`rINH6~;>lopVk{h{k6qTUrz>7uQg+lD)WLzk)l%UC)GLmuhlVcjW1JXNBL
zR#AxjY~EF@NlE2<l5Cb?&}%tSYNJuU*ming)D=me%@FaV>FSc`IKP(s<#af_%B<o>
zLr8gfTwI(2rQmo-I44`J%hGZuq5=JQs<_U5Gp4fjzm|T~g@nYpjF>jmB6wG-hC#nX
zE#K7zY9ustJ8D%cib`;@P9<2_3r$AC_IS605Lvdld1brMjHrlZbc&+9@_>(ix`icR
zHhmh+fpxrXHCJ2FA9VL2+1bbJ_0{ATE8Q?I>ywbA0;YL)5)RF|Kg~f!Q;(mF7j@5X
zPF+>}7);Ttb+!P0An3MZIaOM6*o2udf0o^aMnJt!9iiVCw4*3thEPFR{P8l5b&JAS
z&32`UsCq~!LpDBLhR=5Ol-Kbe-}SjV1m?Kug@d8#4?WKZ%I8V3FIa?q&JH)NPhK3s
z$8+bUhhe+4`0}S8vH46%35j1WhB7Lo3H4r*QR~MFGNqBdFvEZUcm$Qo3woK*Hi3&(
zfJ9v%N_Rhs2r{LT^I5bF%6DNF)=RW6er@8n2k}3g#(&?5&1KhgMB>$Cg=&T}%Gpyr
z`KePkI!icjT87xrv4Up@V*DxbJAW1pXQaOqpR$|RsWi`0dTt=zpKvjNJ6$m-?`cNs
z)nu5rwJxr~mf(XytN&DLt#xI$le&+7X}!=)i5~tOjK{9soEE|&{*F_(<}2>B12ZXy
zPS4%(=A^yAi@z^74Wvd2+uJh{WaUOd51*7;%qE23Q3!aiqnt*o1RPiI;a6KAQruc1
zL+V|&)@2L+b6=1~_bz0qPE)<%L`v^fddwwi<l;>DhjL*GuMK!SJMv<?Sjd5$d+zKJ
zGTx$QdTp`ym$j@G2Cvq3I~IF)*rzddTiEl^nDZF5z~}?w*us$0R*mSF(ZlUA=w|v$
zl2<di^T^KTCOzw&S&xH|KNAisJ0;rcGG~D}q!pyYQQq@D%Z(i!9g?vzAqC$H%sV;z
z^MYwQYaPn!U~tTcBs~1W3oSML(=aeh-1*SlR&QQ3tm7`2kRi>INl*p4Rtm2Vjd+>(
zq@<+iXaizB9Uvln_paZ9PcJ~NF}CxK$YM>7P&{f8adB4lOBr!w*~-Fi`J`R7Bg=`K
z#EkxKrryVz@*{Uh806eByCXWC&?m(BBR@Z|qS0I9rR{n=wsA^LS|6rhlXb$_rVZ9B
zgRUUB`<-uggMjC{x*Of-cggHJ_tIu=b1Xk_d2_BV7L$7PcfyPlsgB(7Bi^6w!WHM)
z4ZT4yc2%)`oc=ZH*f);v4XoKmoh(B9*LIX>5FRCqFhP`nn{UU(px4fWmowJqfdfT$
zux5T7(se}be4$D=--7pnew65v?@2BO<Ybg$GVS3MI$sX|%*w#G85seR@s~(ok%0j@
z3?kM7wDp8{#(*Ny*yf>UpxFSJZ>FY8K2tO3Y=iR-)1HxvXx|*nZec+4#S`8FRAvjr
zd*3`$1ov<FhapALi*afr*9yy*D9ju;&Ypt<t=}NbrGHKF^v<2($j24#o66A<dJUc*
z{wTiODb8zw@!LrRC}ar9i?g3fJo88w@Cj=&pD8A@3&5b}AnwOVP2isR)ZJB6BO_4i
zp(D$(7oz9(2Pjx3zuV4YG~F=lB>QhCDOh%+6-M10x>O!WeZ>Nw$1o6_nCwQ!#sHdn
zM*zxrfQ00XJTCv%A{mTFtykqak!cz}V%}(6<-_mgYH4X%{`n?^Nf6v-zFN`PsAS!O
zF{ANXpZ-Fh`&Kvg28QvuzcMqMkdDwNmVvi7Nfq;Ri#4jlLTKCC&W^W>M$C5#wV{T2
zVej9Mh{v5Dj)|+P>aaalOfqy*(}Ot(e^uq^p9$+}Na)s6EmThCd$ZODU4XgDFP$QZ
zob3?xUg#an;^&9WmuIZA6GeF0pMxJ#3wgT_K^KBX!huZ>$D@8s7L##uLCrFJ>upkb
z8~tQ_&o=Y#;;l}0U(b)x^q@0UY_i-Bz>q+l6Y}MDbLva0$*-EqFwvV3f9Q@UFYjJ!
zuuc~;g~!74WS8dH5B>1!aO|OQwKQj>+>zNANLHS>yT6=+Hq_eSma9BDA9wQPWTEct
zX&K*vu3g_jkqj~G8lqNDXZhPJYxO)!Vu82faf#@MC`91Zxsz8=?avQQwb~`G{MUxT
zEqe6D0SEK&L(Mm7*I3V2y3C#(ZC#4^WqtY6eFtZ<RM)CPV~!EIGEr<XW-6bGeOc|{
zgbuU<9kYF@puH97_drWTzHX^XVApOu?kn1!sTOuw?wCb2KSA1q52ZBhv`UX2Kg$N#
z=yG8d?WbFJHoZp}j7RDGiUK^&HJRyaoiU+@!`U?ME0G6AhlOBc-wtl)(0N5&{!wkg
zgIE0O)4U0YX^v4;_cfl0tHRz0)6O(V{=JEJwA0k2v7Pl9%Y%@D-#e+PzzdyCGV=bt
zxU=ly{mJk|4(SK49ZIySWfZT_P^XOyAqU5?v!lTj9-3^NL=JiUjBGqGB|7FjDS#8D
zD=>1iBt}Q$_k4R@Ln_hyExFKdJ=u&!{U;Gp1Kh~>hAt#IZ^6|{Kp?p6w8Np<i?c)5
zjG|JPA`;Civ*YdAg|!0^d+N<lm#@XKA3d@T{6yFmPG>jYNdK`l4TG^eUiO7n>xY_m
zh7sQg{}NK~wxu90n;1G+yn$T8gKl)|G?~7%3m|4yuOJ1vfBN)-+~H+NkMEky8-r*}
zevS3}+Vc-?RMf7oWODs}zU___z1+R!HK?|69p<Bev@>Ap1O#Cq_JA5_&ig?=((+Ej
zR_;phZmE;5aO{4Z(kQEXnd_gPs=I6YSSNPbH<%){ptkbjq#U{eE$S8OC0ahaLkNQ#
zN_#_`36gWXo(HCv<Au~%Yy`CABHpj?eSjn9=dqJ<ng8DIG9Kb{8cKI-;&%WbyePsd
zgo%YV`^xCb@lqw6ggvz2i#cMZnj>=t>UZ^X;+)RE@<X8`rgzS7f4}kUV@?;kiZt+@
zkWo0ex<#@4G0+gj4+EM^;Ez(Ea^JZBljxP_q4`k29WpXIdV2cKuE~(Ep`l+xEq<40
z#SQx=CBqBe=Y++xsF&$AeJE11Ydl>a$&tLE-~>xd!ru1`+-RvrwduZJe(sf*Gt4)t
zP?g2y=r2kkWif+$xSz8gl7+13wq1}D*r|5tq{rJ3-JY&+>IlLA^x#MP+lLjTC^=6?
z*!PxG#blNFDt$s7j!A0J65UVoPzUBw<rb5GwnTOG#sVm_d<h3j<3^OIf;gY;+Cz$0
zEgsTygp7(JFaJoZ)pAC76rLf9%0TXnYU**9bz51#%J?SiyWA0b3v1p*D(-u!TD7v=
z1ANL??$deShSNnwcdDoGaK4TzWL`3RWm%NH7XVz&pIa(_tPTbW9mONNhq<7c;R-+g
zCO}*xjVAqB!y=fy{4rkZnK-wV`i_SrL?jW9xeVJHP#Ho_Kgw(Wc=}$p(n*S3Agk0$
zwEZsOX1)<X?u;W?x}3~t!k7&t(fHiZ{0e`gxqp7ZDVMzeY5VU)_0|;BYRYu=V0gjF
zPIz{$e{0H4qtV#GSbWv!;_QN;AXpho=ybe^j4c1^((CHeXY@<+q)XHx<_RAM8w*b=
zyS~rS(QejTSSt1Jz*NlP=+5lbRJ4v&cmg{X9>sKt)l}`t$$2WDQ^_Rm6JBe0{ZY=t
z7qit?{dt+ff>xO0ZL{I@7pUpej+H1;3^HVSIcD4Ml7AzW9~y=oS|Hz_aOHy%PAn1H
zJMlL~FZk{DL|<Ux)mWdt4&tEWX0hx~U_XFm>gy<{osX*oy$g>+l+8zp)EmPq+m^F^
zea^l<=E^o5)N*iecqn|hSfUMWPU1fZPkxcgXER!D1x;Y53MqNNqIOyG!fcpByJo!6
zh!H<9`fmi^yb>?fLemS2m^d_ml|7~pf!MMO1(Ml(_s38setx)*G!V{quD;8B%3i)g
z3{jI&jjUcuU>@!l<NmSj`PSN>!1<7FqQEcTis0oPD7IReW)--q?3kFTdG8mr+>YZ+
z8NGd=T`#vb5-xv<BAQR#g!F~7d|=_!sV(=r>b51B6R`iY>yEhk@w!wmuMZ@hUls@J
z{t@tZ?<ClC>x<n`&p>KQMAo21Py8_|TWTey!$a7+NeI7rFVx~l4y;(?z1T1F-k5i?
z3ANAesUWIzts^v!5=v$>D0Vv{5e-u|bG5cB5*x5%X_+&N)_?-3DK#cGx%gP4Kk%vz
z`iAZC<SM{KOi+^V#0M_UQnjJjMATK7;RaNxP6aKj+<q*9J+;P~CfioGKIyAOeU%^<
z9;JwV$cC+;OZ?9RzmqJmV*9#Q1TTWAWjy@Ox5|tLKp?*P)*M860s?}$1!w9`)=Vn%
z`SZMyoLS(8r9K!EcC9oak4mluo^Fn2%ua(ze<I=gzvWq(zMhy~|7bc%auN;-ZtKVR
z5kM>Oy}<AlT){T`GpYF<0R#y*(=O8!p1ZmSc_iiLqpfg-Hs6F!-H7|3Tkm|jwfL!!
z+UJ*JXP8R9LU3^KOY3R-#aP8=AO-+bl*(&|XZ+k|=GXky)F(xeQ{&U&e#esTo}~iK
z7tJ6DgNU*!GP7m}(vB<%=X{OLY@^2k?`6&2z~&3S{R5-y>?iY6-`@sDMu}d!UtN~=
zkI$*(A5;6CSJ;naTV#4y8f^@fmzU2}TMCcn^_OS0$Hd%L54W0cgzIBH^xfYZ2NV}<
z67fRS<?CNvayA_*=Y086=d{7ntp8AGw`JZ3Ba+T*vt94f9YxvjL8E8hPN{4@`fadU
znO3#SY>jY(8?~tK@>#@|fq}u_zZ0|I0%3kv;~N=8fWI$at5*K{2o|O!K(xCeV8q14
z)piTM14&%q)?4wyGW#F+6UcjSP#zl|%5w6<GKI{4pSf%`pRcDH2Cs;sngz{<D0UR{
z6*7KhyrRh{IZ$PUKpsI@f>(|Yl|XAa=vZnD9UYZKI1M+s{GGhOAT`kqyCNCKaI3Or
z<ihiR8*jez{D1T<HsL0%k{Y<Rh(x8+Mk&qPyG3oP`lf#HIw!N^^@!BdbO9H8J&7`{
zc$Tfn5}ej28;JB;r%YXtZ4@a)hf+<)e#i^Q#(WuF)hZdJa`z80q^0AWMao><71|2t
z?24oYjVkcM!oo0Ppr&>p%*?8_Jg|_LcKaBk?#M#D5^~H&*axedq~yRfAV|72s!Z*P
zJl`rA7D>(ULQa1U4W0AXeed7^^YytBurGJ)l+m&o#^&OZEbhMKN%gX*X7xR#-)qJ4
zSup?Y;5ICLHFvP%XA}4Bcb-bVymrkhDakt9iN$oe!KVn|^o{mj2#xv9@ordFDu5Jo
z{{N9!{U3m8_#eeam0V_SQ>e`>FD*IGALgC(%tZFtMi*9m_LI&0G>UfULG$cB1_cmJ
zo#WA8qNaH?dr&}xo9UyS$!Z+C01n-Td4<ziyBlHmArRv_pG->Bvoc}NV`q<znU`$%
zfI@*|9yFaX#7Y;rqz@)Ep)&MxMoyYX`V%<_-*0PG*Sh&)w)tq4)r%6w1w*}F(9@gj
zd!Ycv8v`TVY`DPo!Y8F3PJLP`!RF<#)<;YYtv2c&0+|>;`XVengyG6<9|^AtEl@zz
zgVnKqAqBK6#6qKaRUrOENW>8r6@)a=riL`rYbRVYshV1b6}8S;o^*AnLh<j~l!8o*
zjX`-N@s9D>*qBFVguH^nu}`@AK`H)wxj^hWK>%6b;~&*~{VCUL@(TTPeu~p<$3>(A
zb4YJ0Sp2mzRC}|x!F90PW=*%j5?&pS*(iKrrkKtzEABF!e!$C$_5Qs<bkrWC*8Wj^
z+^Yb}JO18=My~w{9NSa*eCXj!M-usq5Wnc_?QGz{CkdzL;(A{;Rn@Q!cHx+sGN9yL
zL18M|%qWMJ&W@O5#s=Db+irbfrpfFfzRS*dFTdtm--nQbf_t|l^tL9Oya0()r{?!M
zF2<xH*9GTAZ=7V?BI=0^vp-)C4c(|(X*v}Ww3=LOeS6Kv6ZMBFh1~xv=40LA<@Nv0
z?|ibmk?F^~3^xNp9u+YmP?=x2?aY0|B9;RXCbpNyez66JoZ{Ub<3;Lrf65KSieT{`
z`aTyE;DjmQD^IlPfJiD4K`t#Vz2I}DTfTx#{5BXDNWcJEY1Uh7mK)uV1hK)0e(_u%
ztfSAAq>$IUfQ18b=^q832HV-@ZwS(kvhq~V38UxVij(x?9S?(e9`^`;r2U~NG7Z^B
zEO>t*UhS{yPmR!~T8{$;yMS{+lzO4o%hy~tApUGnRF$lP!cb3-{nY_TUb|z8qPKaU
z`7d76pa%n<i^doA)jrf-#bVCR!+OFwIXP~qnfNOEEQ6)C;r#2?zlW~QpLjt;&XIIV
z<+GZpbIO;AXU!DyFkJZ||Af~@tKJ!?F*u(ob-%G=PB+pV_3z0o%r_ZXjW1qmLh*(V
zg#qLQMOJ3+WJ-LOKfk%q(|e}}uwnPD5-(KK#+b!eNC8;YV4v1dFCNt^adF%y<Lbsa
zZ`5^l*->3CwDrz=v~~fqi8fFbwKA&O@)Mm#=hO33C%`0CDM)@`+k)aG&7b2WU9b^E
zA-{hh?%>assE0<<&AycX$C2da<h*^(Pl$`u?g(Z-#lgXOvU{?kB6?Qv=JQ=-`Y?bz
z)qD#Sm_Ak@wwur7O53;Nl$XKIf4a=N60>&LqmZ*Om|0YY=+%BdURB)%LjP7DF=oQb
zg0p!#TXU@M!_V_QGyW#r*cY4WZUofc&!^u5=>U}VKr>a~&QGTw6PEGyy^$e6MdH2c
zrnXz%fZB}JyQK1JUrv|Ol5;zU$W{(&`>Nwto5N)ax9_SStUoN5UMud7bU}@@3%6DX
zTAlFP&9i-vu1n@}>cXh}D4*g2y{Wi&Ay01yI+R`aIG-`?jf-+<oD94uP1n)OD051@
z`6cE55i4>&+hW@DKa#>;$2_a6_P1Qc5BH}lfUwF=!~OWdn+I%60?x=kLLVi?*|mQ?
z#`I|b(r2dlEdcvIXB3$c%=2ig3>F&FV!{V`eY)2C<BL|wP;c*^J337vA8dlamKZHk
zFJr`dVE<@M7uM|2PeS`sEw3-vje$R({{YS3ADJ#bUicR44j-ksjTY2^v*xNUov#&S
z$LVl2c5N&sh8so_QE-b2<S&i!4@Dn-sxsHM0uz3JzTsy)n2~%jk;WkJzCI@{OI__X
z+-#<tVR`u)Cs8zA#HBCl#?99SmJ6lAS$AsbgoM&UX#4jRJFbfh0f9kNdX)L&x8#Ak
zC9@X^2OTg~@ZzyeoyiCj&>%KquDXjSICOq`pzr6R(}_(S@@EN)aH>pSzsj`HDS#6;
zU>YuSjLci>|9-eR3w`KS;YdFmOIxS51ti)~CI;?GZS>Bk2Lrm^IuQx%TEDDEXlZGM
z*w_XRK-Hm-Y|6z*pCUO$$b2!O#4nBCiI<ZzD&+5E$?Vb+R^2HjL({?WbcNAPN(Kl%
zKBcgyb8sg>Cmi9O))mN)DlX$Cx|oi=6)MpNgNw7egLRv;Z76IwJu2*L0bpU7%|2se
zqhL~z=-La0+B=Sae$&wY)aCnPy6MRIC^YYl8>tlP+4gv$N*-w^5E@MfQ>4XhW}%6c
zjmtMRg@HbK+=j!+Q><^d?{`&%)MKLGW$d3?@Q%HhP_MR{^38SHm~Zk5&%F&K6lVcl
zus8KUaPCz)*sr<i4Cs($b!Ao)K;e06p=-qWRHajX->3ZS*Zx-jop4XkdAxf<C|vKN
zQmFdIPf3*-puR$tqV<Ih&NN{+TXd<e%}kBU<yN;!ekG|QE;$kL=9vlD+bI@)tM40~
zWF*_kCSy9s+oFk|6BYO)Bj9I6pOTRi+Sb-aFPqrB|F<a0+2ffAhX{*6y<wgVpflc<
zzrUkkSC?vUvMO7>4$rn$;8b<z>aaJr_8!i&Fi`+9jWRQ0Q9ZFfbKZ**-RtlV`SXY0
zX`QK>PUgPar{_RriU^C042hfvT;!%;&h7b|dm7|;pTHy0!G_<JwdLP$S#KL$x5$Yp
zDaj~>HxX{}Go%Z5YCRuNbJ5ZYc(w=OA~O;7Zc!nTdX>(Fm4fuMqwm63<Ad6lPS*Q5
zo7f}FR8({c`24<~0;>|wvjW=e7n%8rCJ!g|g&Ur1!f~6_1{xQKv=W_TWC-x?Ia%Q|
zQVzYD2D=3SIhR+gCV_T3z4V$)HD4ipsw)!Ll?bxGlpkMdbbiA3!*8!HfDG^P^N&Xe
zl{=(F#Oaq9)L4Har*^aVXs+~vA6mCtA)RP0yEc27cE`ooT|5@^WLwl-st+sp0a>H<
zsUvMzWQg}tV04POd}~R-gw-UQW=(NH?kWh1Uqil9yzFZ?j-+7<xf7$W^xXK*H-c<~
z&sb0k!qsG@SJ)&yjAJ4Qs2|Mzy)`Jjfba(2dY*mNh}i-4$dg9Gu=~bGCu>vsemdnJ
zSDgw0B2jVWj8oRp834_9JR#<11Iepvy`!bNvsi?Lncj}ySiZ86^nZaa%=qmp4O0ob
z?ZBNB(`RJ1i69X7PFtAc$>OK2sWMVx;-j50#_1&~m!-DO*0&i;mmWpm7l$yUo)%sX
zsIr$OE|;o*s|xu_*L;NS=I2oB)UUj2C9JKb9eN9^%+Nctq!K^%guUdvw@SWX$nP3r
ziQB+`S&S?PiaUcsQ$az&$KYF#N1g$~=J&!p)+W^JyaZTOi^qS+TT!E`wx$bJ#GRdi
z*cJ4U8Zi7ZQlM=DCZgwY^kE0D8Jn2Z22uGA-t;IsKK|%$t;50k;B14Pe&}i@*~VaM
zcT8tz=f%;~jSbPY#}DQ>Bh*WE@cU?tQB|?(PZLD1;*_6uE2cf_%c`yY5-;7aR#Li9
zTR6UR5(#}RG%wYu{niyVL!ZY)f!AFU&<48{_&#PL{!S8VF($ytsaS~Ei`E||-Qccs
zWiNBx7H~n;ZDeju)$zUE7HP22uIXXI?M}ZL?Nj+Y@ZCE~0oRa_NUDQTqaVZof&Zr$
zKvrZ2VD%m)Rh@lMb!JujNkm3Dc<FSxb;37q?(Mgt*W`JJRmednlm0UEclQysx4nLw
z4>2%~R`tv+b}ILaqdg$mVH=+lOcVl+moCs6d9&{+sGsLib|L%lvdaEcb|aOQGxX{l
zKZMR=K-txFL6j~#CzZ!y*=V{_`g~FsEZqd_H$ykkut@TkVHYm*X?9;6@4>kQNeOU_
z&x=-aiKHwoe=6a<4aN%(iL?b5b+EyCLRE7mMNviglSVnL4a`Ztq`2g8#$F)O|3U_S
z&{-!p?ex9h@3fQOq#y!b_f5pK8(aLix9#+FBNL1cR1qf5j}XW=0Jj^1sXUoRE!_J}
z26mtezD9OhY=QyA≺|o!>TcRhlty$*81bXSbPQ%zQvC(r6M+!a~n|?m7?kITQMu
z@j)kIx-&dcJSdPb`lur`;t3y1#Y@=qFN-m_(?&>v?e+|FFZ0fQJ!;X6rjwPZ?Hdtn
z?h2WLvIN-1OKpL70^jZ2muTK?1KF#-+Nw-<qaLYc@l5%x5J_P8Mh3OCc!L{i?`_fz
z+6!;ZT6+WO``X4)32b@<5<h{pT~Q;SF?Jgu-SQZ;%>p#@+syQ?<wK$_>|DUO_Z3LY
zK;@r4yUH(9MMw&8$HcUlMkBN>)`$~bkfJ-r27UWFT9c7Vg1S!019#^&M5k;bdV2w8
zcXgSay$m*I{96`|;D2feFwD*gyj0D1-7ZDcS59G$cW_n3^;XnsHzph1Mwn=pn$M0x
z|734FF1BI+1k>~F%gtU+wz*n^&;9yNzUL0}=!Nd?E>O+Qx@<z9@sGZKR7kV@ZmHC9
zO=gK@qNMF+p3wt}%?(*!gC)RlYSFSi*IgYOSGIF~jM?pyPJDI<&)vm0SKT|L#jgR}
z9>S=XS%z%yo6ug@%G+pwyk7u1BS1A=@7>{k5*Hk<U(?`q=Ys+&_ZikmO^Nm`3aWu5
z{+x7)F@>B8{RYpAv!nlebKD<FwBnm;eODyx>F_JOhHa0Rv#_WMWX7`IV$s$oB$9ql
z3fOMTlYN%;I>a0$GNs+soFpt*qg4&k(U-Q|&ag*iG_>!QY{A0s#FNa!&U06iI6D;+
z-xx|W;Vji_QclW5?dxwuQik@Fe|%%$dpXDNbWTwVE58|9_Wmu?r$h3}mXA-fgQ-Z^
zfEvCVjN0&16qg4{E;dS(M0&Db-ir1NAhMf8By(J*he_!q{R-jlbgQ-UWa0xT@gT2z
zTI(yGj?D{0VL80kzn|7dhrDuMgm{7d8a`qcJF-$wvsu+k?=t=PDCl-<7o+r4+<CD@
z7b@9war}icfpQjlHIge$2$Og$wfyo6lnDgnP(ns3zva-7UHhFu(aRIMqdHaEzY|5I
z@miz9q_hG)XR1#Pv02of60@j1l4$`UYI_D$i3@!%&VmCALke;`_BIbKUYHI&%bx0r
zII!!0K<uF<+BF3sprTEiQLHwYBJK>l-~aZVD?(~@I+4=5`0vEOc*#A0DB*!-nAPc8
zgZ2@l?hvN0s!f<fT-R3t$YWOIJDtYG0V{`>hM`PXYRw0!g04skVNW72pIrclQqC88
ztIUQOvIShWI5Zol6>ng*M5OXbi+y@D63?n$%a_Qhe>uWPJq1J0KNLOdq4%jW9prWv
z5EUoC_wTk9(nT%~ztIq9XA=^>tlk>0^S!8l$%h8jHSuvq=FJMmj9tI+ORx@UvbUzL
z?u*O4>BN4<!mZ47!w#3<ciDY^Oi{g1qtFUkdU|@=;I`dt3@r>1aldK;AOIt13QXX7
z+4hdM{^}Cb8Qa;R`ZPPBt?hLOkT{=eKJ7vR6E<0lb!9IvyV}fvg-r+1z76Z+{6B2H
z1z42bw>ONzgOUy%0!lYZH>d+hcc;YAAl)SbgOqfKNOyNA-Q6kO-TiHR{^vdCeD8ZZ
z*JTYeckI3PTEAGE=A(p+QIU?vWyaQ6?mwI)kO`~VWW1m65r6X<>d}D@(WtWIZi;i<
z%{820T6dIA;&$EYRU^xFEwmdT(^1rfVj_Qq9~EjX&)2(MvOtpf++FTM1#8pyfu?Db
zxD2c4b-nQ6K?{aV;+sz+?ZF_FPXlJH<pF`a)49FIAQdqG<K_JbKINI7!VI-z{Yv^x
z1)2W)!JFS=6i#Lz0p!&YL|EZ|O-jUR-rE%ADKl5=etk;J>QXN9^U(1Zs0PXSB$X5;
zg_B_k4d!){P^CP@=O1N3#V=3G3UfU8<sHV{h)+UtcYA)kJ#oFzdy>F;D<H5e6=;Zb
zF5U7E+OTMd+if2w&J@$r`EcLBJF9NjaV+X)vD2l_^%GLJ7d!A)sIbT?9_|YOH~h&B
z5%v{^VnaM8Y6c5j&zo=Dyk#OSKc&-A`FA{ioko)-{83)v;kYH%_!BVkVm>Wfq3*11
zzy8C1Jyz_qCt?rN7SimP0vYl<-Hr-oq*3*b<a+u3Ik;AEG5`+j+y4y;?Wdrg=Cfau
zEIh4tVhL<}4+V5J3RLpX%2Gzp`T$sA($e%evcR+DQh=d-3*Mfv;7rI<D%g4`xqRg1
zd!vcz<b&NWS2ss8D@?|YCIhgnq~Z<O;<3SHPR?LaWug+*-WNR?f`m~;93B&}l9Ro^
zI^4R5x87;1DKqjNPUQ7WhIbF-R{@rW>^b6lXocw*vaIB9a_tY@N^;*h*HIu!MTO{p
z2kDBk_n$d})dD|&xbYD4I@r8m)S%<Y6Z3fI{kkKS%q3kS#tZQu)bU_Mgcn)8v2#;p
zo_TN*my1v9oSgYO15}5XH)Wfkd^R4<mLu}KZV)-$f}SkP6oyW`+i*1+P7f7G7Kef}
z1u@2HB~0`SahOqGg5WQ0)lgzAaob)~$}|u<5JB|GQ>bDRqj^DQ3RFX0ddMV)gz|;o
zzSMK;z&GYyd+z$4<qKKPSJOtO){nhF@RDE8$x5b3n{z#PKc91aZ=y}ex;A~f4Lz|F
z7ym={gsiz1XiJi5SYhWf+9VKmbEk=`AKRRk6lvM;P5`qQ?JMTLU^Uo~MH)((siH5{
zdKL?$C8M4fuOV$v5fGGfC1@{+OQp6scG}Ja5%Xq2d791Frz_9O6heUw)E=f&JVI{~
zL!HoQRfONoCo)47uaj%N;k4PR-kSApYBcYh^Ijt$Wc2^?e$o+4LJYyZYYsGN@k76~
zzZ}@P-BieGyj@Ee;J-HKmAZCbaL2<Uxb(!LcG#Qct{#g1aA?i0UTHj3DjW!hAG}$7
zZU8i+jE3qWXB)Ds-BC#kv!%|g5c}<UM!F-<P#J*(AzLyY=c7on8Q))>olb#spHyDc
zx%6#~Wo&%Ov|13@X#?1l6BdAFJEXx)yO<>OMxoC(9m^HNM5g$!(v^V*YWmTc9{Fg^
z4Pa8{vy&}^ZiWc--=83%AKX>E90xajc%wx|V5;#p^y+k9tJwH(cE8a`A+zDF_0lLH
z!aRO|algJdRc&f)({nakYJG(S(7Qh(T{^6Hm+Pc^6+Xrw9?+$MB#=8RK&b;$?yjJe
zyUsI??Jk&8;LeT^q7)SJmT<zZ{@Tm0nn|m6G)E}{-k;1Hqu*&Qb8tACv(xq<5ghFN
zvYtF`|M9&yAG`7QV`En}O8F1n8V|#|N5SWxDF1*oHSZhS<$t7c!7wQ>?1{cUC&d?V
zD*o2Fc%pNhDV6CFTVr*vBNfl~+#v{r8oo$^vb+}x-f*B{X>KUHJYN8!ynyVMxrkY<
zo;nezTp9}c`wjf_X|oZ*dms^NYmJr6^xqeV&uCuhikJ$F5oxjmYs|VJqh|R=k5%LA
zw9ipnI!}=Pd0nmR1nYcp_P$<-O3H=v*>A)rR%laYu8!Oc$l-2{ktr;n5WftA`foRl
zjJz?9wZpQh)M&j>YI|azErB_)so3Z|S*5-4%YqhI1q^6MgefUz*I!+Iy1Q{)aOWfg
z@J8d+l^GebLJrRY|KZf(GBrFK+}&^HcsHgczYp4a<DiNedXy1eqmAx$f7&Br&IkNS
z-8*lOyQAo<bS8`HSxqp{?>zTWdEVp<WHM=-9W&D_PaYTY$HsCU`QnjCO4i$Nk8iP?
z;WBG=54+tAi2@qqwh0M2LiLrMBQRiI`~ifdGsAaKG+ps}v>oEYolL$9p<Oynvz<-I
zt7EUMA7_k4brpR1az5`GdY#o8;R}~fkxDw-%<?4WQjHdWf{;3!tNQw%ynlQnfoLE_
z&@%)M)R8?cj5h!tm*;+ozF_<I{?Z7DRurKa{IOhalW)olJ7%4a6``T0`%O+mWb2W)
z*XKHA4i*cdq_@sfEDKdA{j*JJ1O5F}X~}c7+{~NXN5P6v=7O`T1yAa<=T=Q~l~P1M
z3{JPlFnT}F2SiX4NjTa3c(ZWYwf&_zZFJNK%~3^|N;Bi?j}PrKfQ%&J-jpRmoduHJ
z&fXvMUuV?PJX`OX10jMCK=YB2dx%F$4rkdV`g01=^Wb43JcHcF@m=21Dok$2c6Qd+
z^R;ox$uCu#mx;;fkaOVv^fwd(g)b5{wI^6@i=RCDE*AHe7fB9~h3F*QX*Mp@LB>66
zeZcr4;lO&YY@T0Roo+D98$l)QG;_ZDg>_12$d}Dzd}k8iK@n&Gq~1&9#ymdp4dtx{
zfOgG-$MyAoLjSG101L~}yvsRIiZL_pwhi<81nj)!&78%BM-qldj&A=9&5WK;dzHxS
zcx>1k(yeI%&iw1MJ%PGpeiuLAoc<)<^-w7aYK-vYd0ka}QWfp-OnMBE>1QOd47BIM
zl?%1`H@c`9>l{iP)5yt;`fr+@0fo3Wa9^064iz5#>B=cwiLxs0b9I`5!3D%(Ao)t>
z#|fg|kbT7@`>J1-us>HR&+lRIA4?|-^l6VaoomN9QY;BP^M;VB_@I3U86Ofan8C8?
zlaYDn;{|m<cnd_@#N2M#daK<)%9$shQ7RFS{1%5wI`&1Ty#3y+TF&m{j*g;@J#<pO
zY+XxmMZmW_Qf;k&miDKg()wq6aDz*CiO`Ul$umTDJ$*DkUnVQ<S|{t#4Y@K{|F^(U
zM;X%WK@boWoZ`|!##L!Hp}bK9>PjUC&^7PjfVRSJ<PIBX5!c?(M>IDpz8|P<Z|hio
ztBUZ<T6Owg=0v;!AR?}A0QO78{{0`_ex!@-tg^7J=l}Xggs0P`loEK1+8?3L&dzdj
zat{jSS1X+_%L2az99p-;F5>+C($v{4Kt=wfI>#Ey9cwRs(@XOV;g9%QY|=@Wrt=Se
zAlC!OLoyh9arPBcRP0DW#(rlsfU|F$Q3Ev)*Vo^F`!sG5p94AnqMFbg2M(4L-Hyiy
zFTNRsOxf>EzLoT4Li_s}?;~Fu{XSt903^AFRRd$>+FfQJ(D<q?edfE+)f2Z7f7JpM
zV=t&}do5e8PP^M5{r#!b)Umwr7jmH_b5nhq^-HF$XIoRYebW_9ceRfn%n2{#o4@?W
zz2J4kfLr!$BTTdwPQ888Ne(sOxF`6Y3O*+N?|KdXAo~v>8ROxYduFea^FR4qLG89y
zHbEYI2UPB04mnFYY%U4!-Gi@b*Rw`Nct#d9_4hOY9Hy!Wg~|=RUFstHnHI32{JY@c
z$LywH64E3TZx6GFe)46$#mD`|r;3K~$BLx`NR;PY1uQR;0<heB|GgBUkEf<eyDCL@
z?q@+xrM8|alxcKPc~2>T3<Mzu+$|}OeST6aNPBEj(^!E<3|O4_iAo9nQ8fwGv8H)}
zNAtpn1`WKxvXd*fECTC6|7o(@!yJku24gFrO!GtI=X;NmA?)O~+;dKG72j2`t%S~?
zLP3InFts8`w0-a?FBTJg6F6>ijRv$m)gAKH8(ca<l6GgRmZkJL5>yJb3qYkMLG<U~
zmyF^fgfa|Nc!&sIkHNXcDHkboJcz2mfyjS8YfI{W66Uj^6MApv-zJ}@%oy_&fCbKa
zzYvgs2G@2#5HIpoPuc$nx+SBa`vGs9zQY47_kTW#UHoJ0{chF%&BFbvdO;G&=4_)y
zT{jNBpT9pI9^Sp}@>SPJmgDZM*832V=g;@0V+VmGN4G;`&-{NOXf$m{_L14&P~3aT
zd)L7{ar3v=|9j6<%?VFVMr8PJ?fjUQM>a=fl6iKICV;3~MaN^+2SebOl=i#B)?bfg
zP+tl;W`-z<4_&rw9sA?-zCd`T3gXe<>otu2M#HMGkSBVsHZ=tp@p_kC(~8`&El%UX
z0*#5R6pzfxnI~o7t*@Jz^&Q$MLuP-9nsaEpd-ralQ0uk?P_z$6fXn&jNS#cIuIM9g
zFK~LPy|PZdmdowI3RxcYmbbOdX=G>8?>q#~C)8Mt_GTOE0Qj{pnV7xf(1UNuF^)v)
zRt)sABIfx)l1x+P=^O5A-;D5eGx$B)eh6u%3MKwvOVlc#`a=BR>-gyV?-$tg?g7m%
ztUq2D8p69@Er}vW4qSm%3%T%QDX9X_{RU*q+MxgNJQuNIiTPoTLa_hehvv-IvVxC%
zrWUf4R@}TT0K@d?nRvLzS(C9^BJCQius*UhWVxZx+wGDsU%nXjXzf!1b3+zgklWj)
zKjTF`F`<O);yF?={D5~MI2l!dj+TiI6u%h*LEDEI0!SX;Yc!C6dXaiuEXL?~R!o@-
z`}zWBlk0tHpSJ90`poyj_7%DBP%*F--42aU14RL+3I=5}cGoJPg;U&kq=ZKS1+U)`
zdjH1r?qCsQR>EKX%_7R7<%d-Z*i)8jIHW2+0dfZDCOG^V28zNO{@=k7`2k>`K>1>{
zbv(eVS#Asz>`k|Kb;mluD$(!{fL9vsJ^TmP!F#_z<F0o+G@AO~r5DPduBXJP6f$r)
z3SFPYuYx$_xga21siACrJOxx1`issFPaDA*=msYM$nR;XiANYt#BVyf&I;44aTqIa
zj~(TuP5tdX41x?HDhfYUo!#LQh##=NL>xTQ{kI={-#<b4{vqyd9bnM1;mPhyN^W5B
z+nT8JhlBhpqtl^>ZvAoxs2_ZY{yKhyOn}t&rWH$OWv-e;<Ybp~0V*PjLBcaiDFud?
z2AkQ{#P1JglK5rxr(jV5$v*{ENMLA8GCsm9nn>rpBTykhM++^8LTU39bHt&XNw^V+
zj{ln^`^p~r_p0j8PE~Z0_-sXlPv#tt0ic3Od!TVau2steXMovy{SDw*z3vV`*$2hO
zAzprKs6Ee4miuCv+t3*oEWEdCSdhbe&<s|y_mULp;9SyCY;$l?(}*to<?7Doud|=#
z!dTbUqYX>?DYqMiSJf5!!}9{a3e(a>!N1#)#%IREK>ycUm>D7Q!rsL35a~i=%K7=R
zv!@C#`~IbyAZbo~KjOoQrb^d3>=HuM;uB=>1>e1o^==e^2n!1<DX4@0G;ERVlh)8k
zfIr@!!`gG8^Y-tc#0?7wcuu}NuJRIneH*l1!8|~0ZFc$B{I+_wF{F<>l-?CV>7Qd}
z{6bDd1Xe+TXs?9sUuHE?SlYpSd#XYy+?Lyegh2QhX6k?&txR-(wY_n;%(??}*oJAU
zPXb)ga=xa8ES}So<l$>VJnw@~cmCc}I-fQ@2qe&>m~tnkbpV>4mL^Zk?G3N3I41Fr
zaP1f<+Bvkv<w^y?D5_PmufM!XEux){KrN13B;?@egoR^Ion>wCw@;5sz5!r$$0d(;
z>{K_3Wx4c+hMP;z?$7=jy|R7d*{yJiv5r~z-rG#x{z8I|{?NZR6+^x&vfQm#u~AS*
zr#5O2B7jmjps4R~qaZ|efSV*P-aU1>qN*{o5{a$Asz5K%)4m{>1wfh^|BDL%{!mg-
zA;Tx30D0Fl7<571{&%rd46mtiR+Q{*3&g9mT6nG0jkLHg=vM6k=B2@>PilRe@8;rj
zWFNj#^qrFzqw%BEZ*R)17$8%VlZjj&lJSgf)JKQMb#ANQ4ul;6_1|Au*!(34RbvY%
zPTa503{4*p9B`CmN+pRv19ZYYZtZI(y}*-O55T&h(1n1W6G~<_Lg_g_ERA(_jdfo!
zGPc23&-$?uQs;G#qO7k@2tc|>%x^;x`0-$YnT*hq|D<b#dfCE);3eemM+@aA^ScHS
zu!{@XU37*4E(80>NmwdCr}06mCb8cC`}Ketv~1U?$`Sx$3Sh|51jHHX2X1U{*VWZn
z+ti#temvh4DlyP$KARn8c?^5f%Q&W3l53_GP}nZca20->G3__5IBHOKKmfWhAs~z*
z6&7ac^|}Z{h3{(&Q6SG#o`y%nFi%In7aDCzpJ}|kHl_Oj6kWq_G|RB$6~B}6jMmb*
zSiZU|BQ(<|2<&0avoO9dVNw@YALIC&Ckl<pDSr1nx#+5Qwq@3=r@7EBTMm3G8P7EY
znqsLV=Y0h2?+Sbflq!ba`}+>~_@p2qahe2m1^m5woUB*Oj#s0!Ag0z55aiHt!XOfA
za=^xbj<P=Bd9-l|0EQI<Lo*&x@b<U42Fb*V;Tw%5;$KAW^tEzbS`DO!ziQ;=x&R&x
zoHOBDeK4?CVTzgl7onu|w$7mbptt2lKwuN<9o@0vO1!(h!W@cboT1>uaXnru0Hp9o
zu@b9=)3zfA!&X?yK{4IK=Lq$g&J+^oVaBZ6)&ymML5=o=b)emk&-ILq?5HoHO8E(r
z_%k9lr@@)(>f)G1ynF7B&gzFBL_iq(AUg!~!(Pa`u;jU(L0R0P;>8w2LkfEuMoM;@
zR`{`{W!3(y)}KGWIzfK^a9~2~Q625ALjpZhI{JFn4-GOhR##RPBZ(Z1XU1VC_)k9g
zl-e!1oqSYU2h01bZ*lQjnzOtQF!eeHxvSm2$9&&zThAMj_P@~xfhO3?xcy`kZ+7<M
ztQ#lpP%Wc|M=}?{925lI_Uz;5dUHFA_vltoZf@PIoLm=Bo`Et|Sw&@AQ1=&ePc(44
zHDoI2Bcw2Yl#+;e*|NCu3`0IVqGvd(Hl8p+D-J3DQR?qcnXs6iK1wpJIhbJ3=<pC#
zT5BtXNP9t1xPp9>n-~QdNzn`D?Y|d?!)n-#eJ&9L|EMhjs}7dx2qBKb#<Ar}jNQQ3
z(OEdjmn5tGe5m*7(;|TT%y_1F+7|{KvxCtfCTwzLm6hlio||ini+hXKx(kiY$7^GD
z4pC9@(HtK276d+cTnsc4v2}I+pR?WpakHMyf~xkrI}CkfXFjr}V-$jyRd+z`G&pWC
z^R?rMg@rxg-xW;Du4QKCb{K@=#Gs&oxw-~PDPavnXg{?Klg=GIb?Elq;#<R`$WJX`
z*5ukl68cE^ZB~v9dwN~Yag(`$diH3+Lj&;PHNL_f=)|TR{0LqyUUBhB^@%Sb4Z*JG
zEgwGx<sx$gM;ZKRc|T*l4DTFn#ytK{{Qb3AjRh%z@6FdJ<m*!NS%E`W_sy?aO(DCR
z3+vIH?zV>x#KZC@y-So5g|5zsvrYSZdv_hv73@*?J{%4M=43s22GpCE!$ir8x6BBN
zPg&XP#tSr1;Dy>X^$vZ6U12IknlK0fAt4KkrF%xs52i$RMn*=Zk($5HV^l1ZR8S8B
z5TmZq-u!ZW1%;8W2&yE0&%48A(QwxULo1y{NmYjNE-Y#~H)rDEzG{tMx7VbCZrPjA
zMo1k)(&tO=lE6V64V%IMhN1ws)v|2!F$@Gk2Cx)~w6{aC&sc6wHWfy)251<2Pb1AX
zdOqV0{rT-f@Neuv5Xukgl_X=~Bz?tg?se%^wmx74GZaJTwcXP+O!mzA+{eVh!9l^r
zwc>32`Jn92pYP_<=#WiQvnJ2;J^u-wy-Dkxi^G8c9RtI3jdJ5p{B|4lHMT(Utx_V{
zO~%6$4==a3&$EQvgcUO+O$u)mue?MUr2|vnCGJ!?GGd@_xj&)l6(^_TeElV*eh3Nw
zQO}W1z(;8jSjd0oFG8lVsj-`wv~r}%_m@=6)G|=QEy2Rk$XVygAUe$*jS{`pMjs@)
z&!0RVSykTVgX!FCg@WuXSEM7_;NcFKyN=x%JZ3fdp&*Lq7#K~?+f<NX0`^+ED<lMC
z<Jg?3uI}#H26wZ*TKn@SPXI}L)=_(Ub#;|X7Yy7B&U|2&C>ubbFo#MnW1%LGg!`Ma
z0C9aVx7$}&*4?5m)l1Bx3?f0N-^Rc7LdAPdtcZjv%Z4C$?5np7M`QG#idW|KbedpH
z6{Sl!clRaY1hkvzyE7tC;h5Ftt_W(a;}Zicv)?z)RC7{C$46ERHMXnM>?A2oH&qMi
zh6cC{G$YDch8=waDSU=3DaKprVj?11vAgxnqhq79bNt;nI`&-vU&qtGZa-sVKdo^+
z_p3VJZz}S@AwE7hT-JN^j|*{?<+a&G@h8dPXG*`@&E$NGPD!`ZGSb_0S^kX)9JXtc
z`x;f!TuG`8E*0TvX_K*l^#?N}(0DvwE!@>?EKL|^<>JXz5)Y%z=M9*-`D{UgxA!|Y
zH@7;Loer0^-Z-=5WT*kBeK+8oe5rSLcjx%|sAQ7Fp-l`r0_GP7Kz)U)#nLk`u<`pz
z*Bro~n>)tPTz>tR7NA4aJi=>Z{W+G7aL0!I22jF>vCM0+>z1iSIghiG%hVp_4H>Sy
zJ<B9QASCqD5qMCzxlI<Y4}2hfI`4VE2pYsIKcu@A=0imbUWHo0vBdxCU{l$CiUt!3
zS|(m+NQ{qvuw1AuP)#!jbhI9?rMbAvboz5Fk>HnV<YM9+Ky!(#dD>BH-~7a1`2Ld5
zqrRY^DJ3PM%=U-S!MYzB6BCoL6zS6PLZc^hjfSPRwuuy|2dk~OVcXG+ZtiY4IQWwl
zKcZDhgKo7SrBW%0buCO*=apDb<Zck;!ZXt{?^R<+I|3ql#o|Ak;gntez4eOnHyxos
z_sK*<Qw^Bm!A~vGj2hF8TGt}o-8yAu`U~|nuIqcCOtC$#))xuBvR%#u>u8|Xe5vy>
zp;%q`<0ntBFtGwYs%Xq)n;mi*HWjKauIxxcNk~XKLU_Bn-CPBo&iCfJQC&`2#k4Fe
zbOi4+rP%35Kt~IBGuzxW4L9X!PYjf7=@41?oUn+;9G57x;vs1iav!eL;$74kW=InS
z5$dYuVvGMV8Sc;+(Cus2DIXoVF;y}A<a50`xf;t$!4z;`UBSg(!6+?NFFu&rsjREA
z67};pGd7ltd;61xQ=Un<Y#r8zeOzt@8U|dQ?hykITvh6?-6db3CQ!DirY9ls+@GWG
zgZcXTamXnZ7+6&V>lfK6aeGkmc6I9CpIKpv96j>#`-i=5dSh_$gGQAF7D-P}=cxI=
zL)c!9l$F&%Y43Zq&aVV6>t-}7E6gc*Wr?3AatNW!swEDLft{_RibeVzxjoR)2G{ei
zNKdZLv4{qXJnw(`V+3r>-(E-MPZg+@Ijrw#NktNgLPf@#z7RvYBo1-tZbRaF^8O90
zqdTTQebG@YI)cf`Bv5*`;y}L!aL5iky|`8}xi2%CadmdZl4G^l?NZ2urDr`S<z-~C
z%dQY7U(GL?nxE&)Er7)&C%@(!9K^nl&GoygDv)DkVWDIFo|U!UKUW9O*7A3t_tc=B
zPS47!a*4WVv$eCgmq?ZrlTc8yyg6F!>0W8c2hKJ!@c|=UjBh75R#vs-<jrHn%oe9h
zIErT}7^$!yFcl&b(`=L~&hY5y+)d5DLoKyBH8u4ZkGT}y9ztLXoSi1LIc7lNls5$#
z`t#anj#&#C)vfp@W_2t=6`fkt3B3CihtS>i<xYNnH)>>jJRP4C;mad7E@*j9Xq=h}
zdtI$z3+v9iM+ofC*?Iqdh1Zd_gY|1NFwQ>xE-WuD6cO2rVG<UBD(ah#we~+8h4m_w
z6cmwsw25vL!0*s))E2oCsn~u!%J~Mnzt_u@oYChS(x}D7^Fqv_w{IOS7hJA@P%nbo
zYH8^yR9ZAt`H#xK68_Qm=lcu#kg2{z+8==0<f-L>M#f%b)E<#h3THg5QLAff-_ZVe
z2o}v&TUQt=t!L-u)YQ3t3|aha1M7`_2NgvprQH=<(Q{{sc!|J<pFO1%091i5d2G@%
zAK)C2F)o4ZsnT-J<$k{5G08w{t3g{)6CojCbXCN_^8jdiQ5H3KewaKt+0t5+35vQ%
zHisN!d<W!aZU3aFXSxUGZ)%bMcm7peS+E55E%D%@4r#eOi!AZWq@o&pB&4J}VfXkn
z;>VBQM;Sf+)wPYt7^E-@t&z>Yr@+PqE>cBNQC`s$8y{cQj)j}Dtu=jSkzBNyoE%8_
z?zlZ0f%eG5!((+fk@>|7$Zhs;WEj}6BRDd$d&bk(ek+}s&xr_+n&Z{p?#zCeAPtvt
z#Rq>3f`nN7gxGW=!?l%_zcZ-fg{gy*Ughl$F>mGl&G7&bC+W-WUUdS27O1CRADkS~
z{lNn+-4h^*+gSOALoeVdhWP@={qW)jEMc}bJ)Ahw3JN~(rR8Nxe_vl;IogB_Wynt^
zk!BxinG`sh)6GT6L2B8j7}M??Y!qT#0St5xl9#jTCEw)PquC!JY$1>Sn~$ZExLlpk
z1{x2hz7sZLVJTzMto}vDGPr$lGgSu0VIL7ErkuJ+`=$cZ>Q+7IGx9hsVV;w+_2BZc
zqoX5knA-k4PMt3@KF`UFf&i!0Jl)jQ+U`vCg%eqLOc;T#rM87y07$|9M`gI2M2jF~
zZGRf|mV?7#Kbha>A2h(z$Arazm(^IUJwXxL!Fg4AsiLB3dTuU|rLRXd-Qp4vX+(AX
zt*wRZuM@uuK}EFgnus>hm}LZj$d=b(c5qNuTiM|>SBYP;77~^<5o}{`sn41Sf9&&R
zZ~l~)5w@HeEJjNUWdo`j<Q1K_Kom4GWGW7Q9uV;9$B#C3$J^Vg1Kq@e0*1=brwG||
zxw)_odIm{x@sRtA5;`885<Nre4=qtm_arayX*6{!f{sv}MzUl`+3bvU2V1_p*G7*a
zI=wmtQH|Ge(?<9HF8AZVh`%Y+)RYin!ahdFL#{a~t)fu}WWK`GqM?M0uUsIw%7y(|
zG6_SPqM||uW@g))n`-${5j|QU$<mprDubh*o(8TpDun`d*K-VXQXU$Y`Q>3sY9k}E
zaJ`c8q6OI`+JLCUa6%FYgaq<?8RzKoCSDo|A$9!I#m^{CtA#z_L%VJd#z+x(KCohY
z-}I&xbQ73wZ!<A7>5FI2WBGZ2?A-k1zcZx6|71uMX#z0V*;(2DGeJ_7jcffoT~ebv
zSJyD*1omfK{p)XTi00=R(r6Vj;eHk7Q{ixVd4--a?u5Qt$Nl*&r98fd`u1SbqQcC~
zwA^O0+U0#SU1NNrYZ0NxY9?`S^+k5)8>*VCzDGv#nH&uWFRZPudkUr)un*>L)E9*o
zZ74ag?r^D1fzj>2Z@Z@?83f9)rtkN!UoRRuaV4lKN8@_l+_@&CsH<?%XY!$<*4s__
zz-?@~fuol_2`HxuS`7Lu7w+7fZY$M=ni(VE*%xalI<}3Llht8TY&N6DX!aGz80iZ+
zBV<Vap8!$xl8r_{pi#>o-LPh>H;xs6BDtldb#--_e*ZwGW@hSKog|ZaqTn8Cjc$_%
ziv{YH*bS!94eo<JNa3!QLutODGC2xML?H&!@@ma&IP6pTwO6tn54UX;2hd*u-$D*g
z594?Yks&oTT9T6De@hZEHU?H!A9y5oPXp3>V7|QP_7vcz92>Ikpv9Zuv|5JK+_i|3
z$tx?y3+P`>S9ltmW~;2>+SJsXlRk~&8(2PWOJ-gwb}DIEYU8c65?r&pnHtdr{;@U2
zct)*qC!k0m?mpYgAq*m9bs3EvQfzRw1LyNPXl&<fccJR2N(p2fau^YoEW^VX_l%8l
zLy@#2xXjw}o$IbH+9@VRDvKp+{Ye_$d3UN+(oP_2eEsRSwDJFyx819`pM7w<nK|F5
zJn!x6$rBPd9RL_bH|Cn)xUhuSc8B%f;S)8O<(Q%}^cHV!Hik(n(@OWxMs);yG$9+*
zZr}WTMWL8-z?(+L@h#KuN?NodDI^5U9FR&d#Y7w{BviJ8PJbza5AuytFO5bt3*iF5
zu#9bu1sQ-qF%}xPDGhvc;Bm*pdCJX*&iGC~4YsT{719u7zG5~q`ScF@=+UE~APom^
zZ*LF<py_!9{GV>V>J>B<UPKr8`-^%!k}2N=69`eWaHZ&ce_p8v_zl1zq}CU2C<Cp3
zug$D5!jy#eQ%_G7dRDd@=!-bt)IqkTXQ0@%$YVcD#53x$Q*AAYv$MsY53-h51gyt9
zOo_boue{@Oa$;g)Y<`Mhk|CiJI(1BN(tI3-?&Zp|xmu1(zLnUluo%hI*`N2uAhkh3
zo@LUG<*_$Cg{f4rE16#zq|IQogxiFh*W)UE5cqpa(9xfw=e{VoU^RsbeW}`?hsYhn
zYJqgV&REji+#IUN7T5`T=fX<Oe*q6^x%~{sVo80FWFOTN1X+}~tOI*W2H23k%BmR6
zV)8UlIzWc}F_cPE@}GiK7K;RRV?oATsOj#uNqD|VM~Zt4j{!F?%={;*2W?lhYru<#
zQ}F7ZtTP($ZgAzX-+9GM%{k=Y_Q%Ix{M)y*NB*HNAtb0I@&y`ok0Fe9suz4*7tWtQ
z16);H+zBlGvM<4XL3TofDn?ICy#$#=m(q2i(B>vV^Aj{Fsgg&+u6m%iv!X}!_wVcI
z=nIvpK-5o_CY8Vm%oU2FD(Bs+J$lTLx&u>j^>{z^HqL)>7vn!`13G}B_%f`nQl#Es
zH^qV*I0lk<8~baHJuOEpEHWPZZhK%=*VC`Bug{cDpkkyXz9P%#t^|S~g%4WQqW<OO
zic;&35oC$pDwIXUM3AR}4bX1hJiuSj^qic-mXEOKX1>t$|CzR7#3V%#yd<P^S!>Dq
zijjh8hlNGI0MIvZyh$@rI852NAM{08FXQHxBFu|aE+!`<A|m2en>y+3MOnXBdu`cZ
z1>X8R)|sPHtqcaWyZ88mLC#;KqYsI*+wbqJt{6wKJwDn~4ts@8^)LSWkeAJYyez0V
z>?H@VJwFOdDbhJL08rTJb5?fVWy`szP?aJrccOFY_pu!r63k>|0i7<LFGSiIc`4hh
z26or}N8ZN1s>rjG3dC%nVKc~HU(wU&fnG(*jd^&?+AriVu&}y5%>@Mq(D0GfACW&s
z5c<&4TUG+Qy`t^xwmv=vjyUJ-vm_YLH08rRg-uVDz~i&CC6hRk!mN9wD+0thsQuyp
zov<k?6*QKMdlN@Kd>;aSOnAJDAPu|?s$#Oa3GQ$PCAVami2|9pn6#l@?-kX5S58yr
zB6n&C2wOPq<7G1y*H-?*fgm{pBah3l+aPYgPBM|#k?F1l$!eeIUci~|l^#eDJT#U4
z;;efL6{v5)bAfMIXeI*us>`do%Bo3pOUVqj#9*gt%xn1)qNc98RP`4q7P24*O%v7+
z!7>uw^7YYUOhpNKdsG;m&@RQRrQ^eUhBrRKigCH<`W0o|W$SL2;q@lRnX@<b5ll?Y
zy>R2qhK4sns>L%Uw?#ugMc%R-PZ8L^tzBOJ0F_A=h+{WX?(*sgB}KvDA#%tW@BMod
zQVZX*Y#lLht#q%_k7P<myl}W2HpC74;r*_HQrZ8eQqQ;L1`%p7nt@XCM&XgEJg_5+
zo2pwf`S+|M%N9S$?CJ);y6EX>K~2Xv5l9?yJa?yB*IuqN@E~LKJWMmldq>^S<apQD
zxdg#^W0!sUPy%hN^=VLQEY!Ob;gQNf0aEy+w=a*oD*PY->*1_~EMS~1dnTXZF;(Y4
zOafdDv>fe2LqpqZQn+DBNm?K<anjH|9GDQ4P~3OHm@gWfk28zEGt<-m%qaMoo=rnZ
zsWc+?j{o(GgX3cqu&Wz*m6esU46}M%<=9W@9N~M2*#80gIwW+kWktQ`{d}F{ApyZ=
zFxZ>Jcqre~(^h}Eus1}B>EE>N>QX}fy<S>E-OSfREU3nR<!%<If>j!r$w1u*j_kf2
zSU>2?q2&Two3?-0)@-6crzrcC&S-bns^xUG_4zzM0~wB?;Wf9d=^G`4?+lzV8mgT!
z;d)hfIl<WTUQb%m({a+Vamtkx6_=M6)~#!F%9`ltIq8<$l;l%}CmxQXZ^`ufz~O1T
z)<>Ntj2_6qn9s&G>buU*q-A5{Kt)E2Yj%5u#vZri7*DtbUQi0g;H|$QTd07Xp4@$S
z${DN>&~meZy%v8AjF)VKzkUfRD|1j#v`%zza)#1t5_fL4KEK=#bblhmsQ5ia6e=}3
zDp%~2`!oAhQCeDfSaMi$oLmcgU2b&J?U;4e&jT)U^5vO#F736a+rljzR?`$$1#_r%
z07L*~a+Z-%aEdPPjE94^_WUlGBRUOqsFM}8Mh|I|{gL7SnTlCiS?2(8l{|&9x?K#l
z_!HwYF<?LAa@LdqP88;%H04K-1XZ9Q&JJ6x-+o%Z?X6DMl<Ea1WT!JDKupyG7Eab-
zR-4kM=-Y79+^N;iT$hOTRCK`vn~aS77Ii=MOCl2aEK4HVOdX(hhiM5(bEp6zC*gBD
z`Lp;l?O2{+T*uQwquLlkzNthqUiP=|nyx0mnmO_{I&}RX<lG#4l|X|&<G2qmR39KC
zgjqBk4g_n}5x(Wp<o9vpmz8+-^f6g@4q}hNyIU&y>s{ZF0ldgVA7q#Z4I)$~g^x0=
za(6l>j~!xBi}#|~ZfdwQ+n8*4Xk5F=GU@g_$sjK!Kj=hoy&o)!3+2wo5uPcXot<n}
zL+071#h@m!+Zg<jSxWc(KP24fbC8Vi^A}|q!RDIq4h~!VLHq-p^!cu!SI#B!DgOMa
zP(ia^i5U-K_WC!E_U=a3w*a<|`Si*C>Rcbv@C*g3U<wfl38RvP?v?kOA75Kr5mE49
zFhiz$LuW%jv$0acM=@ZdB7kFqgFZq?DQ{tWiqbZmqMWXqhOUQBMdL3lGl&4-9K?U-
zmvL|jiR{^{inHDPr4$K63Asy;ZrNx~ueC4`gjT_F7U*>O1%AT2xoN*c<8n8FLw`Z>
zii{m6O1VJYx54Q*wR+td#>^p>I)_z%=isUQt@hNX<^m;5b-YUA(&F=BQ;EMyC1BZm
z{(R=4M+LYg0RlgHfRsWZO9s!~3^u(z%`T%P83=SJ>V9!uDb`lYs!&|q`?&fuP#~s?
zhN?pKMMLjyxDP=+>(SU~t_NxtfLDNoOoy1qt#h)qb+VO$pPrSPIxBPJuYw@k73{Ia
z!_Aw*<6i62+4=?K=rzd#jIZdfTuv?)QaqDYeVpGX{aQx>Ai(`~6qz|#IGBbayVqlG
zH#rdy4rAktV&bBNp;uQ)P?`A}dkLpv(A2^=q78Jt1H=5y01@YnyZhbEK1tE5T<yZ0
z)ST>1kgvWMVxbm0!sdGjDqf{1sBS{gXk6TH$VbA|RpwLTP;)&-?fPts1djfG>JK`8
z*O%7|4cn~7gTDxSa7-IU%gYa!E!sF3KdZ_fr30$wpi2jEWBPzVN=T4^ZkgUYi`9}|
zcs*2l!lXVGZ%F?*FSqA)(mb$9TdUpZ=C~RXVtM~UBwl`G9B1r-*+w7_HhO$>t1vvG
z3Pn2GPXJRH(pQ`?wU#w*LO=ivnWB9_K4gmHcNUEE3{C@r?S>=_25x+n8Z^S=PK~WJ
z2c|&Q-GABLS$&G*-*S1y?hk+J_NZj5i|;S8<h6Bj2r7-O$sZqv2o_l!l1_N~lKDlQ
zMC9IibPZ<^aaf)lALoTp5S7-R7K(oOFb9gv(hhnWI%O>uF!$hO+VyTY1d!7$C2>>J
zPoI7y=9>rN_RTmBZVwk|1g`a&(v?u(-z6kRk7wkSgoHRqt;&k5;o`mvC7F^*@i@6S
zjPK8QNv7@Rd2`_nmvlW>mXph=^#<$#`^ea9G7OB+I94O%fkZBAvV;i0^K&_u8XM>h
z<B`cL$r&2VUUeVtH&t2AQ1L~_gyl>bJUlvtEiy}O0m$m=`tv577tT&j#c(P~eCFD_
zpwC1N11W+NAH&hq=3FLP$mF2}=uobt&(py8x6Uk<fw)m&)#Th04c2faBjxvIcNfdr
zx~U^WKOGz##spnH6x*H`+O7{^g$)4j5m-1aU;U+mK;Gr(?jHJ5z+BJ>I2s>XKqd)~
zmVEHnyKeClz@SV628_3Wp;BU=01?jE_`>axa*7{eD{SIQjUGWsQbA$aR!>i|*KD!s
z9R|r~6l7IsCYekRrkt_^E~IPxDqimPA<&7g+8hIX%>!^4i&Y@0RMo)PTxe?g^5wm}
z&2X8~?<%Xg{+oDWfK6NkV2y+;L`D&fnKxIcc{O=TWy$7&Sg<|g=@k^oR`k8f?h=4v
zF<4<<aN%;Np)|cP$Gv$&2mz6!KDeh~N`N9}@bhPz8l6VB#kI9@JF>yll1YZub~-6V
zC1I5XBx6}11d0E=Jq`jVXukSRU3?34cLGAolB<(#YQnR<Iad&4hjQW=h~9ujgQweJ
zc9OTT5lYb#Jq?m1XOx!)hIb0;i^qsr1Buh6&OO(mq-Hdt+0>IkCu#Y!8(&A>Mt6p6
zLfH%kUoAEl3Ii}cLS7!8Ewj)V6a@GO*_m`BScB&|IwEhj_2b4@^yJI8b&i{HYV1iU
zLVftGE^H9@i<7Qdw_jj&1WlsA!ogu4Tc4kAJUK!w1OHg0kQLvBiLezZ$Z3g<yR(vR
z`1tW-T)<`J1H{vuqsP)u3u=%)ce2UQ>*6mC_-$f{R5D+N1l~_9YH84#k%o>|O#`E(
zDn*N7t$Qr5ue<{EAwGR!mifR;XY+H_%(i}gN;W(=YCN3&Jg}|TNFM=5R`&b-Y7d??
zI`!Mv@1&J+%XU*A#S6H%7i!g|X*YQsolHe!4Zy(z0o3MDtfpmhN|PXn00Y_?cI4Vs
zmUHk<(?7pJX^gsticq0uBcA(nS$gsCw-1ub+0N_Ns?y5iIjxPz0NVw`^YQVi^SOWk
zXUq&#sjaiqcm)nATkns$-tTx0wPSvzn_9sRk&0vM8f&BUZ{fB(_+w3-^;PD1U;H*Y
zf!^&*u2Nex`L*pxW||?{dT)xL(Vg4S_+)G>C>W7LKYZ~?l1_Z~{CPld@NIkqIXUCc
zjIr{VoK2vsxA4PYcIMq*UY@bv5Q4h*_JY`DTjTtI{yOYTc7&!(mz%I!FSQy<#IT#D
zQ=s;*u5qXES?{IT7U^qCh)E2@3M`+qn<*+0#_TBXkYH0<xy{dVYionfKq%sM)<0xY
zuo?<=8XN@=Rs-?&7dqBg*J){J%FU<JbQbPx*ZbL7w&ag6Ra8_KJu@oI#&2po+~ws<
zTfSTZA;KccQ_#1HkXvu`TNM4+SUg;Ncx3d%**N<^H8v~zF;mqF4NkjiUkDr(4GmPo
z^sjY+(;b+;^gg#<(fxNxQc?;u9>n3vVeg>d#J|N85`rPeW+!iB1ArknQqqC0#dISB
z8oJT5q-6yB4p^Dxe4_=ggZXD;8c3H|53sTif?e9!KwDN;K7kaV+!09CO!5~M!5jh&
z2Jj{E5lm~qa)*n31kaYrQ^uK`|2@-VR$frB;nI;qK;Ut{cR5#AWw_1Z=vdX5AsCKN
zNXP-}-3VQ9>~Y4sUyO|yp;b|_vfB5!cb{)!aPswqxnCV^-q)-gA}<NLe`~Vqx*g7V
z<bGZ2tSPU+XwN$u8;egwlmKwt#VWbCZ;3dq1YFM~g?{l{Fctt*aT=ABga}UsyUBH7
zJ`j(2*7HEf{D)L&>B0qI)=p4SziA~$BwW*hmil(RhY{4c+*~_T4I0g=jg5_1)OU%O
z?AJ3E?Twefeu;-?<azHUopjQjz)c>TKlMlf;1i1e4zj@INe1{h_(v_NhD@wS7b03x
zNf9lV@kzow*;DU6jB2W!A1-$cr1&>_;u*1jmhFoAgR_DIYMtmly;IO!RFglCL%FN7
zdla;Dt1)<#T3NTb%g7hYZKjOKv{0+>U;L3{xE*$o`VNGyh1NFEaDtA5iHVZm;!o<k
zMgajmoj?Ki8!m&+Kjpo>i{l?Uf`}u+!uFPyFc0azX~^oDnGq3PXKu9Z0ufXlA%Vqo
z{iUJP#TVye21Z8V>!IPhvkssj`5qQsN$_a|^q#`)OQ5HvH8eC7hKkhI)Bq{hyz7B(
z??%|8N9V25ue!T?LA>$PzyMHoZ--0@`i|iD@Z>Y9y`;Gyg1tQwu+Loz78WqzRg{;?
z%E@uNU-j<RISNuz^_P|%A0E<9c$TCQ>6Gif+ZS-h#1x2fScWRf{sK@=UxJ3Z8Ug->
znW=GScUM+sI-mPl?{1ogrN{j57JnoSb@Yl72(PlL@rJahC`oA5Q4HwuWABc4z6V-Z
z9Ht0lzzif~-v}`52oHUw>u9R9nlHWdkbMEMP_3FDEOz?wUF>nTdw95lwx53!D#pa9
zd2U9A(#WP_L~Bt%k~9WCet8vmdI%R7#1BVODM$g2$jNntMH=b;9Om-Sd$EAuzY0eC
z`rZIBVXF2f7{)<~v%jApnev2PrS3GDKI#@06*amHrxZlJ4v)kpxcqkhQztS}voI(z
zvHamR3p^t|2|-ygG2pLX(uRStKzQI7`SPsH2&K+qsML33e;dKmb90}Zj+XuC^pw$*
zj$s9(q!O27cdjmLQuRT~)YDZ{T^$*Xs`CmS6h!0P=+yxRa&qwu=#@u3eX@6Zon~MF
zcJ^tg{}wMmO=W6ojGR#G$Z0aGy8200Rt{_+R2d&1U(C+Ml|BgvDFBc!t@Xu`yOip2
zCWBU(m`9pBw^Wo_Cp%LL>eBu5{e`L#fq@c3s_N?MKYn}?mf}v-{8=c{oY|wo6sXPK
z(l6~K@M}S52d;P6*rax`h@k<=FH?9R7>Nk+@d@!iN=TGUp=~d~TY7GRxNQYb=U36@
zDA*1tbz}1}gFaG)=#b1$A)A*7kt4As`cI+B+S=?%&kz(NzJ8SvZj$76U^wsH+;}h|
zA&5cWDMo|?apCJUEzkXzY{sb@&a@n0^WU66Zvlw=UHHk#LWdsnLX}m~Mmje~f>JqK
z+yvVG5mgJ>)~Hz1iZkFeTl_;eOwF){0I@7l=QdX>2%V@O86Q?JHIF;sIx)Oxymjq!
zIcG9eQnpka-Eu#MIV&(;2j^vH;}bdJ%me^%7gP`MjSug?d@TX+0(I?LQrefBvnB_(
zp4qzDIMy#f>^G3iv+8*7F}1ArSDAxPK)6K^bAGNhA9$&D^jBM<bo-a!7rKo|48OV<
z`alwDG&)w@CsRN_=;PCp=7WTOh4>WULYCc;w5aPy2o*A>;-*6-rV&(_WD-}g-wzm8
zYVylZfE`JqvJ~zCTu8%_w9peSn(UZ3wGxCVk#1HaTu(w3z-VU@5&|lTxw*MJD#!q&
z1--pH!%_6WK#NKAi%h-T*t#U*_R7OGT}4UFxgE7muVC%=%HlOy8vNkEhO*^P@Ixy;
zG5x54fPf4{{0;l^(o#By;o<IYQmxd;7@lfpCHO#k%0#oVI0Xn>`{$<r3S(vMhgn!;
zc!;UVh!P-aJ^og6l@`E;?)JNifK^Z?_VZGDS-#}uJOM3Tr>l$0G-%D(Gk>>PW&~&i
zv$<RxpDxa)lg=1~AswM45IIqP{v3IaQGYa|elpA#ZH?|4r4=eg$$IysCR3%t^5Wt!
zWo1P9rO#Z_(jEJ=HAEzku=wzJkl^EfPe_T;`9*&@Jy~dx5YGF{bbI3&hRfNO5{sOL
zme%4s7ZI15@cduFq9P*6{O2;Vvbv*r**iN$h&S(tl%0hyhugR2G<wCq0PrW5j?M&Z
zmziU-+MP%nwEIr35Q*u)1(ni-&h#qOp4NdSE`n`wJeRQRFXO*ezB^O7H`UbSK2sTp
z%S_9|t$$VSIXkNc3c9P@++61OElkxl!os4i)vgaq|4!1j{T7q~O!nu%4ajYC(FsB1
zc1DvzBbdVE?R*Yb&Cl-qC}No_;Ea#&v{{M-2y7bzL!h%H_98XD9g)f3&j_7(JM<4_
zB~8>HEzOq@lkA)*+|=w>3z915`tKjd^e|54=H$?x(Q$J-lhI*Ex;_dzz+=vmOdup6
zo|6<^&q`5u+%%^lXKaC$`L@L$z&kmut)Z=btDW80%meIW*k$J-^*2@Ov6+5;UzJN7
zso&LD^6gw6`%1*zU9UV3%m(t?q&hn}dD-vb(M-8lz{|02b~Z8GZkkJZ_glaj5wW{>
zWo<Qz-Zt}%LfK5?i1lf$U@HgC5pqOCBcI2O@%`NsQUUjPK9k}0V4?uuTw?4FSjpiD
z{X;eMX0?Bgmmv-813}x_^L*gudcUdp%N%e*PZX+sfOgU5fLaj+4aIhL(+N1YOC6cq
zKG1)Q=m@^QkTZWXe*<SX?cI>m7r_eKJLl#CiKl`Tnm+l@@5OJ(jsSP9(kCsg91Tal
z3|BT8Shfj(Q#uH{W=hcLn|CyZ-v4p9yNP|PFJ2jv^5)$ICiM#qw8%u%vYcqJo%HqD
z?e$7PZLaCB23Hqi-e}ap!9j+b{qj2GP$dgc|K)%VF4@_U<I|wjZE!h_Z~FCV>1)7!
zqF_-@PR+<rSeW%TeJ|$U0h3L!xTFB6Hd>zL<!uS5y!jb(69ZoI(5USF<$3_MmL};x
zh*5dRlM_X8J=h3aem`3Lo|`!^_4ntgL@j&2NrLskY52*1mLo_raZ)k{H>#ZHj@AKv
zK%3{_{-~Jw?-2M>D}Rc{bw*|uWN5m=tQGe`UkgMhfb|$lR>^pCX1;^b)zz(SS2tO!
zwC#R<{(!?+Ei{gC*d^I+9Q-JlseB%~YAD`Yh=kx(OY7f4>-stAI^r=xYNeQz1U1=E
z4I&6CfQoH!rKh{`6eC|QY=;(FN%ST}<SPZ*@O?x<IBjY6MIm<N!aYC%GO2x+?y<sF
z7)r)lc2@GVjqL3AjY2!&4po*eC|TtdW+RyfX3o!zf}+k^&Gp_thU7$*2DfJo1Tc5k
zC_ov*diu!I&UV|dci<a-Sw=OaubGIiukAh_DnQ~C+}U}_!Xu)RKzAho7z5d$UwqP2
zpevXT0m7YV+~`<P8p;Z;eCG1ppH@1%z-hWuZ!eu|@bdB=A5Sxyga$`N@ow3(aaMs2
znH<r=NM3$oul`^`I612}dVl%@FbK2~z(P})7<`qc8TJOK&Dq%5l=C$YCqjs!<;IJ8
zx6=Ore+CUBSCdzj!C_j{{rp&Vn7pml90}E2@<0AB1s&{ue?g@50Y-qCeBu!SJ%{%+
z%mgL`WJ_*1O%4h2KSiLC(M%q^aIjH~k_Ap~H&{clhPEe$fJW^b`zqc6hvj_hA62kv
z3MPqLU)QJNnebMATCFHlG|=nhx6dzAk}69ua5&lCykUFuJvv%1&^;a);Zp6~(t>t&
zJW~W_N^G-|ljlJPRZ|{vAj3}0O%sP&&3hyW;b9)}B}9CH^1Gg3qxGzQ-ndTSM?mP}
zu$(nEw3LQA*FUzkJ^p}%v4(db?st1U!2RLLZA4+A@Nw~Y;jItcL#qT}%_?l+19+h1
z5<D^Ff@~VI?+*2pABh3P{l(D8AKj9ItAG5CSXe+8=UUHukCIGAjjCiG+svcqqtBFZ
zaLA?`Ti$Y0DxL6A$C#P%GEtqO`RL;m)E+QUeEn@Hk-N7bIV?X6J)xw{iSO{OWr`zT
zGZImdi|BhVg@v#c3%b>2=VC8-ON`FVRJG8Zd~A+BCGrb>`3|ZBx_5%6N@Zoox2dV`
zy>+{oo7k-8-2kbE(~FOgYRST~wM6vd&*I)00F$A1SG^`=dXUro1syg9z|N2j4`Y^L
z=}oyyessbi35$=OZnV&dAj#;jd*oFcB5Zfr{m^H~X}((S7qe!iUR%3Vf<HQu>3ma4
zV!fM<xcKMKodpS1!(}XiTEI(OoXqdS=X^B^n!OK^i=vb8eH0d61yU3s(rEBFj)teJ
z<{#fD8m)-&o>G}Y?e1=z(d>ig8~U(uqQ0&(GB5&(ea2*dMo;waWR30MpWhKwQY@_O
z7HPSJpo{Y5@j766+2XUG>+sR1h_G*`7417*+JJTP+Q9;Jep(RIkr5Z)0^NbKG-Wc3
z4As%`UjRJ@?SMEtrm7kmHpz~i9ffx)Soe=?7NRw2xO<9Naw}jl;o&a5T>-+Qpd-Q7
zOS~=b=T)r?2LN<L)aZLGl=L%qSoP>bQ7MjAm!hcX43M}unVXvO7<NZ)s}zOYJ7cCi
zMGwSPS5u?a3+6*G{wNZJzex1r1p&_Oak`if5*pac0$!{;ksTIptor%$8_1^pXx=1E
zVV0)jZ3F$)bu{+bh!OVKPNn%AuG78tF)pKYGM8MyHG<c)u-v=5qYMcE8O)VCOMuIC
zQC<_oW4f_jd|&d$N7@B641LJv>|lM}3+(_!dsPdxlw;cy!aie4N^^5_*KZDWfyHEc
znkR+3g?bt^|7i5NUx4s(P-QChL_^E1motH|sp(sYfaBdw&GCBRV|BKX2Di=bd3RR=
zv7hnW3|bCHC!3(R^-Ctlb5E<^^*T*~mU|CET*98gU%Kt(CRzQ7UF(OHcc(GN3XfuX
zMrm3V`<Uo?dCAM1aB4oj)1C*W8D4=#B#f7xeFORsMy=YfVb(w9M-)d-NVny83;?LB
z+Kk0}S-A0D671}P&+TFt)Bn#8<`%3b3Skiu2Veup1|hiyuzI1PRv$lhUONKQQ=_-{
z-3zmAV737j74~+PXy_jJ5vf`S(1o&O<MO08KT};QfM#$(r@Gn(ay?Zhlfs`zL!A;9
z=N+Z?C{<)~Anw6c!`#&sQ>0k~Og8{(Q#8*L)&8_N$LpvHo$C?gW@_@{v|joO)HI6S
z>ibl4wHCAW0m9{=e^SXc{$rszb}+@Rr)&##>eiN*j}L}g9If{O5TdB0go=vl5*Hsr
z%v!0zxvUq1O3J+?tO^Rb4mWrAB>;=q)PR&rD=0`Lq&SY%s*%r@w;i33Wu-eqf?6gS
z72P8`F|m31I1uq;iCQ6829=A@(b2mWSF+iZH?t=`?M#*&uQ9%uOUts9Q?|q=Jkwif
zY=~iYu6M!dqhB#GHO6<`=Ga^h05b>XR*fYs4FsQpo_9CL>pwOLHhq=K2oM41q105!
z1*R!R%U8$2D#v{gT;8xSjk*GuyIGQ_Bx3*T8$`fSd=wY2o%EJXnTJ8-hSH4QH#dXV
z+1}Ll;-Y}lf$k8MM2|PP@hSd(Ha0bBo?VyMSI;oqBu00DE*Rc9+x!x28kbFbR9ZSe
zKhHrfLP0@=2?;a?PG_K7PJ0`}&sU9z$dLBv-d(qn0X9&1eso5a2AiIa$dc^)C}V5m
zfQyFDGpe2uW-~rv4p`*usYj!N!}9W=w=t;D&8L`$wkNFz_T~&xNpvlloQIVGnrb4&
z(c(XYO3Ji2ZAnE1G*37)D_cuTTk>vOl6ldLloSe!$|Y5@i$p4l-|yW|w-!{2aA&UW
z=IX+!cfl4AJ3ACXlxK85OaPsjz#-LK5!EpXrFcisIN2J{Z@Y%JpK$U2arPEKS%qD=
z=mrE4rKF@pKtW0r327umDe00BLAs<H6p#{;mJTJPQ@TW?ySux);Vl0DoVn-Rxik09
z{f2Qy2jScM+wc3XwVv9&I&sJ_eUp=Gh!&vWpdV4Joh7yDId$!4JPaW5+_>)K>@+N$
z&_SmR{ryn>>^oguF;5m|W?$debBl)wN};Z<=U&$@+}x5EC#&2q8J-0D!o1>pH9g{C
zwKP@+wmkXfBpsV$)%<4-0Wc^dy!lc_CZ5Nrr;kCQjfJ_j-&$6->o|Z!O?39(7eDCa
zMh93#E-JiBx`5A$F8c8!E~6U|ox5a#Qj%kDp<w>p%_!Sy#|jv57$<R`RdM@eRdn@4
z4V(T7JZ4hd8)61Ud2sLcEO%L%sNQ??N9D0t_?i_?*{VbqU5OvT>X?(`Y3nSJ>3Dg+
z*xYXa1sX=4-;pTF1xj+l@X7J<ihWN2u)z1WK#K5Md0}*t$Ze8Kw2*nx`7MI{Yi@ZN
zM9QC9G~gXTtd#Y}fP{pmr+)zESXo73ibgt`fnUa0HT}Q%rVoK6yz3hb!ejKgtfR-J
zUX-KKNXWc8-+2)1>rg5D!Y=Ta!F|EUFkg<<*jW93hlFI2Ry^o-@TaUXfGJstuXHSZ
z(87W`hw_=t5*TtMbMT6xA8*jnz5e9s>T-E`OH5wg%2Z>_>w~UtvaW**9DW}ch$r>?
z!9R*17?{e?_C9s9N08z=X$uSDZqQ`khcYZ396Ln!wtLfzCE%B@;4}AVlm9<qV;<#o
zZgr@nj*jpZPM0z6yw7K|_b?&dnM+q&TN{w`+mEEAC4;_x#lgLLsmLikNTzsnXg0>%
zeMgdsiN@c7C_Fyu0pIMbwU2LfhSA>gc#Q8EDysh{S65fRCRE&is6yHK&Vw6|UTCtj
zv(L{F!Jid(Y@-C-tAm1Gd-Jw6wb9T>XQzhLTRo!y=4WW;me9Vcm4vZ_sHlVJwIG!y
z*8CR$dnW=-`u|1%k`SzhZbXQ}k(BiMb)UQZ<pE~<+uNiDCYd8Y{eI_sO@fO4aO2;2
zg~f2W(ol3rEHjy@j`<@#%%kd`aqhMJG;B+z_cX4x9G<?JT>Ab!Y>%C$phyVNFXM^P
zQhWWXcI8}60wT)LeU6ydpP)7eQvpyo)|;AwK>QpAJ6JZEzkWRz-QL~~sRt)g-<K*X
zC8#2O&fkB*c!bp0t~faLb9vIokMLDgL>Lq3P7CtGP?l9LKKFyQ<3a<js>*@OZ668R
z_vX**vWBg3D0vjw9J*&3M9fS9a=$@R$;7}FNqb^8>2{u?%1pC!UOn!Wn|oz!J=W0j
zC)PI}{kpYPndsbQRl0$WU&9)b_!)Xwd#)!pa2&OYSmGfAqFi#KcM31md^}nAUCXF-
z&|Us(qFhxL621VGY;0^W(-clrF?}G|C)Nk;-phY3{7sFG1{Y_?Z5@2j2zgKe<iZJ`
z7&|+~Gs~%R!3#{cyUD4^!lJG2n)qa7bY#m;E&Ja?3Gg`vyN7F>55YuG&p@BpvJ72g
zBfs9K2K~mh<jAI$mxr8Be{Ak9Pq-g#O)H?{7Z3KXzBi_$XQ`(L2wh0D+U-L0XHG`O
zV7_kn<Wodw`I_ht@t~H4IZWRdION2{w?jIi!+v(GVXpeJ>1)kzsA3-^XT`<>tQKBk
zO&&8XN_dCR=Ft3?g*B)7$oH@iVbR;1b%ssAb3ugI=(I__dpmrFhn`iGv+?d;$u3(m
zl>&$%J>njXz>NsjOR=KmqwQJIpZ8fW=azfXad4no@E=fbX=;0zEizqMHdI>r^QWG>
zwl+Xe#u)=YZ^X2(Ox950A0MCM2iE|RPz{?xRqPNQk$sV-p*i`tN*7-Zm&~QRT^bfq
zX&HIIOKolUW7gImKeXTAl1YgILtZ99K<WRLzw3ijtiKR(#p>3{(pRw>;r4vmxf&r{
zUEA2W7&+fs0gF66xByC8(9EYw#rdx%!}mr*^hk106e&KsfQtjjG?RB(|7hI~P6-cB
z{rn<dGG4$@S=~ZRa^(1AVt@CLo<>rVX_{1$Ns?v<AfDr$_X~rjGzEE&4c35h$~d>u
zZ*oO%T>sE4qOM-9=kcjBw(&s_oOhyt`1M=acIX#&BnYy$>n|>RXbhy9gt8kNaeO($
zzmmX7PK}N}#l4G%*DTQ*%kLcFR{^eQ7(V&Yn%za#X437H?q<4n3y$M6n~S`C{}!gE
z>k|a&)z#I$QgF*Ze_q#KU$0XKq-3xqA&U)kzc^lhJDVUzfAiQ|^zXE#Kf1e&RHajP
zpMAD|)oIhqR|LTc?jDu~s+pTi15vn->60VzgCiel6UXNl{RL<`U8k<DQ&sfo`ue=y
zd&z*{(D32}{x?E|6dEh^#L)I~>{q&Jim9my?a@7Y^!TB#7$F743<=)1MkYMGUyMwX
zV27ctU1N0!gaT|rs74H4ClE{X27Ps`0v!gdm)ti)rUy&Gw40~Lt{1VE=OY~U(6fI3
zKK}J<$;Mdrq4g~k=*~$WzZI&PsPcRGjJ?P(LlYVV--nMMZ?(s8i3P(M(MdqRb#&HK
z9`dg6<Patg<|N}@Ozdhn6!%S6CaNkzDl0t<J2&`6fBxt%f&pdb0{FGHFh1Z2_vKkW
ze{0qbH3lJi<RH+YU{cQVra)AwEPn~@MT@`0O%i5b7oz?Wi%BdSI1+@U`q$zwPFY<O
z>~hlYv2v!e!J|biDJcnW`dy3&iν;3dEo!4VE*baqwnLdCmFC|z8bz!BiP<5*ek
zdP?~$Pw?TxB`Eeke?GyzbbU8c=;M_GhrO?FXyb$A1ohvhsXxvy>2@!<)Lh;5iIGH%
z8)<_dto%C}9=&+Btg4*aSLP4`qv=f>8z~7Hr)|EylkiNE&N%0tr4Q=rD--<A>>Pbe
z1N5+9pC}bN6Of<Q8!al9MJ?`oL_`cU;!+v&z?G!ykVdI-Q+)NxJ^O^JFHZ6Jt%s#j
z#Z*@pjHU@-H<k1bodmW+MDAJpSc~yY7z35SgDj1U%yVYv8=64aBc^<#cAoaw;JuNY
z&6lf9l(Wk$RZ{dL?%kI-74CxHe4{xAv*#@1c$Tp#jrr}yp24XYz6PFV9=?B>8`M$8
zD)|`ei>x!o+N|XugIFs*o~yFT(^I5?oBL|zoI{ybpg}&hS!2|?0C*+rhQLev<FwS}
zXuSy5LOCG}=d$Wf<-g6}UxE(XqkQA<jkFaio7vCUXxZ6*hO?=H+1WmM!_2QcSAp_p
z&p&^%)6j}~H`m!d6nfZOkGbq^ZNs}>2gUOf>H(m?>)W(b!@+yA^)hnTKRyZ_Nq@e8
z)}~$K=nPS=+7T>I+1UxncR`y3>&a)|RZ|9j=ZI%iG-XVCDw`B<nnWibiTeLnp=MOp
z{hPq3@b@QsYHZUF?jQ(uFR)Aj`b2#K;3?oKK-B`S1H4l9>uDg;op4kS_zLyx2$Ino
zU2Gx}FNo`SF$R%2ARGdrpp>MzIFQ%`1uwxNv92@TTU>m6uezB8Q>xtd4Y+$XykKBv
zb~%`9o=r>~DzVswq%xQ4%2<tx-)7RzhszU%J`mVk$OV`=_I04JhW`V`=~f-u4fpw*
zPuO;2RhkIy!cJjtILS%FSpJ0s()5<2)9CT{pFAxrNeEAfbU*EgV;al>t;~GQIUNga
z;oG`TE%%kl%-t^L7(y%T){TKcd8J477<v6l?gOZdBABXILQoJx%xEqXV!}Pky>z*f
zjfoFdwziB<f>eb>MMZpWl{w8b#<XYD;@NMG6~}Nq<$jQo&PA=GtsWiS(ecO=qkq)3
z=yfdL^y(}L*~uAzB34!fFtUvSrhd_{USl3sP)UCR>H6X$&1<3|X>kPBPdoQ{ZPn+d
z9-A(8ee}d=XlyK5?Mh-+fhpSy)jS0yWjlNO!M?70G|tY>_P9aK3Td~ySz9A<$ptE*
zi%wQ)FDSu1>ArmjEb6xa1k4WaBst^LnaN!Gu2gd-AHfG(RkbiF1Q$oY%2$hzAH=2P
zDLM!ew6d}eC|)t-ICtIOB)S-yXdj0Zh~-b4{XE3M5vi$JS!+J86TA#=hlZNH7=j(&
zt}@%~fqC-5NTI=N)3q2GCrYpv2jbRlRc)<DZ7l=@XzpyKUja}nXuy1d!bj>$;MHM`
zUCg<ZZqQTZtklLw!gzeI6Stu%5QO6l_#E<zdu7fls(A!t7A7&_!B5<~lVN@YuI9L?
zRK9La11X{LYe95HOiUb%*~??U6r6D=oe0P)940EUV>rXP%+m8>&th}gv?WcHLge1%
zGw#U!$F`skLLeVeaJB*aLJ|UdS~@zQ--B37wcPkBDEJhnYcMC>j~-huFoSF;>R>H@
zr9=tsI&MT{_yZbqx9ZL~Uf#<I2sp2CNsh96>KqjfS$?s<Mt||k@Wd_k^sCA}*E%wD
z5cNv&p0{PqeQMAl74v!64w<iCyeZ$?({d*(C-tCyxGpCr$9kcYo+jZ<3C8yr6*m5K
zd{AJAg@*^E#%WjEonvXTk}4cWErw6q%=}fs@$sdTiGABweDp#FK+8=_n=b)D?qn2e
ztE)^ruCBxW;Y4@)>gwv6hhfUj0vO{Js5OSG=n*(bb0Zr}Ad13@GFV7qgGJKafTDh%
zMog>$CQ2833YqgY%m#V8{1x_#yAI6B+7XXN^K~l>LxJ)Vc=c)r-7>)N8!lRP@sM=8
zHqkz{RP6ODDr%|ZGX{oa#G~O-$6ZY-DiDjcFmbS$ogEc)G~3!AcJ;~T_P8W+zJ1ys
zY&mi~p;=MH!Qs7IAK(gA-FS^b=f25cMh%c=bXfm*UB8-qlD{=J20bI`#iw9=ceUy<
zxEN`%2lwtF(5S=9!+0QD95?hXKj=`jiQ*>r4)pPD*LzY_>ty<%$y(LdyMM2IdR~c$
zyHEHJKiAycoG}kR;&I{;<!C2H{77Rd&vLpxG0{EYtaYkww<W-7U&5hlAuZuA##vu#
z;MP=sW_k;5iGKTtrw=ZdocY@3=46s->-zGF|8OMj0~To!>Fk|jc8X~KXt<veCejP{
zrAv!5cy?hkL_DHt+Vx3xj>o0TmBbdQ1l(vq<dpUjCx-96)l1c(#@9E9?mqngDcYab
zyYGB>aUM+T<7e2{6s*L4O{j`KW%yCCn*9xwm4tXAx<=(U0<wj7a)!aRCLN-j!F^e$
za<S#*>rTA8hK9!e*j=)h%4DrgP3d#0@$vDz_M0@xc9bLB8)N@d(Mn|eLdm!N-E5M|
z)s4PKlbtFJ3g-9kGxa(mKUHej;Xrg6RhE2O=BU&ijeI%fEY-5jsgGD1-$B@GCb(J|
z6SFl6`nJB+L6G=U@i}yTI6Y{Pj8J5z$_FpdvFtbY6OA>_XGhv7-O>G-W~Qch)~FJn
z>%4lk`#$Xcb;UPqm%5wP_0bZ)5s!rXVc$IZ?tx`)Ce1^d1FRw0!3kb}8dPPCLCw{2
z6Se$}5xF{lb`&Bpp$1`s{Ez+~_k8!XgkL$kYs}~431dcfb!^T$Td>)0w4o?tF3LQY
z&dGKulpJ~>3ie;m4be`5PLlup!T1^;8Y)6aL!&JwTAkr!^fynZcz#V@!JE&8fae&8
z+6#UXwPISG$Y+PsBbP7KI;7Il<>liOFT@OfrUjTsL<K)*k1!>C6!b<dBBgon?APh(
z_t!IHrH0+Y3B2~{LTV;e{~QuOh&eW(xx2GcQ9ViklLKW^H8sE2t=Y?Be<ew{RFFhe
zpu9LsqH$gmhhIiybUK_QK}<{>Tfmk?^&AA{y&}qXGJB_o-J6rvWkC6v973V}&l~GI
zVSZDth*%B}mwl?wuCV2AkLDLgrZ%Rj%cEi{$kE>(#Fm8w_K3gk?D)43@?&<jsVOAn
zUBamS-?CM3L=mvxKs*{k8lcM1F4y}7mI^VMtAQqiz8Da}CL<w}<J2l4*{Wqr%4Ht=
ztq%E0#j8x)e-mi_|K$R7JaRoeX-^G|0g_`-o#&U{6^ZvWUOzh*57vRFYhwI6GL6Y*
zd1U{r%(|a!ZMZ@uLMAe$H}l-(9$qA~a#p%}nbonC8jow|L=}QZcDvy1^q;SBJl#Zu
zHWjNQLqk8$zDE)6E{E=`uioW%dieMbGqK6U*)eDdIDDF#+j(JH_{saG^Xk4nBV$YN
zG8tLc6FS|imRhdv*YL!<1}qvzWNI6oiJ2KF=69sTT+z<?-OuVCv4k$I_Ei>zJK9NM
zgbtk~fpW*tvw&(}VtwQlYnFN}kE`+I#kpwIBc{=!x_ke<&$rJ(cU!36Ryw~9<{x|@
zXC);I{M#`&oDJUTrY{Yl3lLvhU-$ma{b9!8rD~~6<5vi8{ANnX#Kd%CV<g|2UM?@c
zFjHO7z9SYeb#ATlzXX^EYZL2pdzWU3HFWg!_aFFe{LZSZ{4J@-DE2z{5J-7E#`Z&%
zwwc-4>nyKd-=Go4<iG&uiJqRFmHpLlj$g~mki4{F!TuI(G=AO!r;mK?l5@JF?VZ&U
z#YV|UFa_P;Tke(TbROEmrw|A#4bxVfo1e8l>=?g}dj|3wrDB3{^YNki%U4B2ge2Zx
z-acO5jjkti4%_d{ei$sy$*?7C(XBY&eav~WTa9*oY<xVPA6HI6j&7d7lu@svkcp8|
z0L(?e6y9zrvHk7i`6U~Cs{3D=aESUhC)3uln3J#ztgUKF6y<{DrHTuS^Ia)Pk9pYS
zi{scr8H^ctL!^j#pZ>TnvUFYBgMb_#)|r>O=UiBrtNZzT0k^Yp=R<v|gu8d{95HZm
zLjCkR$jLS*y{U3_b+k;kaq_oC0++j+=-<{16<Qjt4(+_DNbM5a{$-cL99EI%PVgZ)
zA24(9QzAvhMYm25bM5-wT8WdVrl;j>PjrNYgv=zp83q4x@4su1kVLlj_vg3KZ8=|H
zzG}=FugC#nzkpliY9RSEb<*L+SR9uDPGqNcks)=$7xxSx(3y=Nyb(n6lq8|3|Kxe@
z;=0%okBx92A1j*-PzbCgW{*$L`4QB?I3MR$QSomK1SbL5-v!2k4;}a>zCA}fTpO96
zUnbpwnNVX;E)Q>KtUw&<)XH>ljJE6><f*=tGYOHa;pIn?B%K^C^IMPEuXzj|bwW<2
z>5;);(aRVEk~?=w4X2aOk~d@O{4Rb=ULtDzRRY>IDX2mqb@BUU)Q;hI<A$L6L4>hh
zI)m>s{)261dcFqKtL-eEXldey`1~a01;7&@KPHc=g!SuED_QXKCnAxJXov+B6?~C0
zN*7kq1-_7!l%Bo#E+=0%-4@GVxUl|&O2A)Gxk|Ak{!Q7t_wAYEwzlHfR2R(me}1X|
zU|m(F*L>Sb96WeaQN1Nnpon23y=BEKj`@o2cy+KsON)SBIwOlz&`m?jT&DI%BmL#M
z!8yhadQ=ss^M&mC>T~inE|rTW7@KElb;dj6Qefg>Q*xLZN=beB(ssDk+Z0S2c>R{D
z(fja_h=WZ(0`knC{UM>DeSLiy2Yfa|dFVHZ|Ni^;l~QnG#JG4g$l5j|sar33vg-7g
z!E59(+lrv;iA})!R&yt(3Cj;^HRk6YC@-+juEpSSb62dz#=<z=`k0SKNgUlj+Mi8x
zC3IQz@tVQQvk*7@?RWDM)HPLCQ$JyCPdE8MIRD54t_}w;S+u}AoL?l2vW8KCcbD03
zc~og(WYo{7(Cf9fHjw`GPCTEZFyX&1cOMkhftg@}w--_yn)aGJq%h#i7Z8m{bGkas
zC(mweZ9R~T+<KnZF@Qs0hskkeou3|^M{Bb=!h+ltbldqk&0PVD&9rDzBfH0ZHnr+H
z4pN&wP0L7N=LJyY@X$!ahi9+j`Lm-7Go$lbp`LAR&rVOjr)9U&*L`|mc2`^I2iEZo
zr@bb88XfV(ZbR5U%9ocJ5KOm=Th?|{tF!gfyAP3iADm<B%>=MpNX9zl;Y@lJ+D7^H
zE{ezL;kmKadF;1u-(znfHfqC9=;-9XX4AbPALTAIK=X+DrDlC9@!OCYV-=pYq(})y
z3Abt!`n}s%w+PSJ?=Rn_G#eyb1`axk%+1depkotD8lqwEZ@LpLE-rHECFU|P?2@p^
ze+#90)}^YcNlUk&E2Tk!d=H|6IQ8n0jh=yl9JQ#pn9St9M~sXa)Z6zI2}<xg5{&$u
z_HP$N%WBEsApTf4$tP;GEUKczGg>Btu$HoV%N79C4u{<D=pW7wPFkgw)5nF2^BUAX
zGII+H1?+t$1BcyQ>B>J+bDk4o`TMV6)u_?ZLp0F(8!RDM;*CL{q7A+pfyX`Nh)L>u
zz0}mSin5CK7~b5Kep(+~0`Fn#g{EfC^D+u%Y}*+@QBqPit`;#xtm5CcxeWyk@;Dg(
zDItnAG<j|wGZk?^)QQAHhgTBzlHdyYQOU_SY7slr$&0n-P^j0uKYQn^k}0G~sg>I_
zw~-1DpwfCpe@o7D%q7GknmR?j#E^gMAFaYfPp@4VBJo3sJg#*X21%P=b8~%jhEhxs
zYHnV1r+sQecH{RYG(4skP0g?I#WB@#G=}-$DFHcjj+QPqV#jHxpQ_8@rf0y<$4|+s
ziHeYU7nlrgq~6;0$yx)SFB@da%CY@e(*HjXIG%9~kD<}}L>YlwtQ=D+thVUzzjGBZ
z5DFn66&7q__ax<hDRp|VNq^hUsmjTGq{wmUTN00^&7E5q!FZ&+=3~8!ozJ@#a7w5b
zYu?U~;$&8Z*=?}G@iAxV`%tB0siU)Xvdl!KU*COW{LOw!gVjwcs8Ke@TbUSRc`Z5^
zTb1VL=W)-@Ex^15`|Go3OLkBn)CHg&7K&UwT*8>IJeuDU=Sw1q|HoO*>?K>MQpWWl
zs4>^>S5a1FzPO<E3L&y&vvvo6J4ChQjgq6-kLz80!J&n;@|s4!CSNXR%al2mUB{)&
zP;P$2nyIYD0#ucp9CKpg)^UFR1RpN}+#A_hX>t#*<E)hwn>-qgns&4M?_1!&3u1ld
zd_t&QOT}R<b$Rhw_$rn5V7QQ8k+FSXK)Gd_<2I4fdl~}}{XQsqZYQ&G7{SQcbMq#3
z-x>o0!)@=%F^Z*u;+2%c?!bz1JtH+o88-n$D&ffw{XaNJkC$|}+0CZr;XqCf0>r`5
zF&9W1VL0S|u~}__tAi?P>KniYA(>_cQW~%A#<OOZg7AJwA;KXwvMNoW6viwTa6u6a
zU^(lgPg1=_94>5D=>SJTe}(I)^K1kQp-1V^(8x%7M&|Hk0TEJe+Z)GMc2efg5*)9U
zkUJ@la*z5+_usxEeb2vMUQ{gJrjo4ErL=g$?f&Lmn#o%`@4fQ#KlDou=4e~=C83{f
zR&?aMuQjRlL%ScU@RMCEtBe|KD7s2E0UO*nIf2Fl1PUhdCxl4>Uvbj1q-I9j+pb<O
zzxdc3joySpAc5H~VmX^=ro^4E_1AA*Yxt1;>IV<+BDcudN&~*in3<K^Z14e-<6BqP
zsJ)B2ct=NGLX!L2dq}0PL%!8~OEmWXwO40hjNxM@>Ayen4OT=1YW>_V)bg#A=8t1h
zP(B)T#2WP_k>0`xm0@LI+TE^potjDl`X<UfCSK?58L>bfC+im@XlLY)9)-5kXOHH6
z0M`JhLG?fvAIIa`(lj*GH1sStFY%}7%V<(~8Z#K4Y`5mtn3J>Vz4XBuu|Y>qO-_~}
ziGw*grGVA5lz*Q-(LaA)VB=JbTmI=2khROojzOT*G+z1_?(l-8l3@HbbR|w0pb?Y3
zfoX9g&^7ml<Yu-qC0;c*bBwn3E*l%KfD_7CsgZ)nL{ylPA_cFT>5=brOA8~+kic_I
z%)z1bZx#B{^|lzn8kow#dysBF06e<Na{H6-dRTpZz5Uvlk-0heYsk205i_j>AB&|2
zI&ag&qP~5n-q`CzprG(PTuNDz%P53fuGD=pccH%%)Edprh)`rWrS3bL_FChD|D&R7
zHXV73C%j^uag&%C@hI)Q0oGo7J3Cl7IKPXFPaRh};T@Z<bpybR+~Z}nZ%X_KGVxo0
zFyi0{934)cIJp8@dWC)PM>#pUQCfP}{G54d>Gv<BuJbGXOe(o8tt}b9G8Gv9Ec(i&
zhk140(8C@-nYvnheg0)pm+kKi{S7T2^fWUp8@W~l`uj)o+ndisvW9c;J>p`aeacch
z$y$4PxXueyYv--5$+)6oF`GXGm{vQxK<wkcg~emJRcn^OTfB}uaOL}WuO#u_(sFNN
z+G}m`>T;*`=%oV}4;s7XnDiwsxo*%K4WG2Kx}P5#xL!XeBtPBlRH(l1{y?jsw<o8p
zRKxzmF)#@Q57vIEsGuT}lC7Ca5fO7l#C+H|ZZN5FnApzAWt9~9F~agnyXGs!)70uG
zv|fy^ucS~tt}spW0nHL}#O2i)jrX^oZ~ym`r^a>nA2=#zsDMyRicr5ZigvJ1+<4}z
z5N7}VsSm#><OuAq@tMyCx~(1gMY7Vqn-vtVc6@3vEu7o~TH(=CJD;z~6C#{r7iiZ-
zgoU+goFl)-nrgoHlw+**`ymdD^^h7jvpyu6-pWlXGCVR{tGK1LHSJR?3W6o;A<V$Y
zoUxA~CF18*Qzg5qT=#2sP%5;>Q#4&J|4wu=NfnzO;Yz=b)y^Qx8_v_Q>daA1m{*T~
zfAXWkb~EQ^R@>sXKNgm*V1t*J*I&1bwfT7t!>wBERW8K}kQqeLe@(d80^58P>GDV1
zvz)SWSN{NVZmXY>K8+5$e*6C7MO#x-U*o*^7?cKObJfZ|oS(29yzyo8X?^uty2^cU
zaD242zusC1go3|Ke+{_h<!@)d(GrEexVgRM7Vga|=go_A504s33IS2^$(^0sf_*@Y
z${Eee_tj+6?@NDhb0LV6L-Mt>^@p+zm-&$k&+D|>)i!003a{kkVso%=-lV_edvGRp
zO;XeKc&4EcL~E1JD;^78e(<;?QW;dSn7Bp8-gG=87J9icP7S6OMt>Znu_H1H{G&gg
z<eV&0;2Y2ST4O|g;i0o+c=#BODir5Ub}opW)U5hPhR7uFICg9czpwI1OPfuRM1+3G
zLJ+~Y-uAf%FDkVn0V|W%112p+&wmt}(qi1;tJ;-RGd`wKbDz%0#6GXMP_x9`bw0Qt
z&?iaA{Vv5$r(mOnMh*Th1MxM~Y@GG>wtpidLu91;hgqF(N(a1R1;12{ja!sE%OWp=
zD2^Mc+;5R{Qn)X0Ub{FD&Q`q<;|c`mZnNJ*qfbe1E&uhDm28~fei3wcp9LDKSPDuX
z94f8skI!qW3l}=jJ)mcE(5cAlUEVubzZBejEuOYp3z@pOR8Kza%vn8JTtgSPvbI}c
zR_>n`)n#N<HXHk;+Xypk7L9Vz&f2$!6M`lSYap4(_`Nyy>j!}N{pvB$y{7z38836c
zYzXj$UfA!cjn?yN(8(PmJe0f`*RJ8!e6h8)1-lR(KEo|zlhx}VKf=AVP5<z%N|q~E
z`=&cS3i61{{L8Ocp21%~sy|-%bC?v3lw3wca9Y&;X>0qEVQ*Ay;eL8RnH(Dy`-Pj=
zYb7(_Ye(eV+*V#5*lG9~664>+K`6-`vHbMdo)!{6QkOk0s7#&_cU>6zpv$yMg&~mT
z6bZG8olbWHlS*u^*5w`Y%$(GJi`zA+sV~ID6+W^JwAnOT)GXVrG6udD`uWYElh48t
z5i(+UBQx|W7_*q+*5^Oh^6e)K9<R9lHPO*|&ist1WqL@?M_1f?FH2~$(D^WayubHE
z7?-H@zj8|xa^b(+*Ip?_d{zX8?eM#rlpRMirH?2m987U>v9WQnUI&x43V6}D;3%*;
zQT+n}@dE}1Mo5nixlcttXA-lWzFFkftwtvXokH({MG`^@?%Q7KTm}a8G&)=Ze}C%#
z$zhC>iL>Fcy<PYwj_R~HlB1@Cm)~jc+mGfj<*>yClJs8>1FT!&^OZNB#PFEQZGN|~
z=2B_Cje@*QNqJwKW&m{i{NlpQDJ%Z5@rJhHD#ueCq}*nOJ@ucu3Z3T@i;FXWTjpUm
zblmx=kLLC5FHo4=X69<>FV5}T)ni8F6(J2q;G(mJ(4#pt?Ip*v?XZY&0oS`4h4HP6
zH?wo~3eDrd57Zr|iyhl%t-deS)P`l^c+K<P7`!hohqRGEVvBRB1R(skks^sKR+AWB
zBbaoa+ko>(Q=%JdK@mknL{cBbf+klF)tj_o{A6yi=AWk&4fGv)<q)f#smP6siR|3O
zV%4xTZ7any_I`Axx~Eq&9RJr`!ENmh&0e0E8c*Cq_I@-e#&A@=sSn)mey5$Qc0K>$
zce*qWDhkoMpD%7p%P8xEt>^R3czBQ>r2X^_z0uy<c^&ChzZ)O({Y#nEI@yzl3u|~Z
z-60PoWn#}FS@EP2I62tarR4HYXoidS_Q08W>$wn$>Tesh&}$6pXledu;P0I~`aAdz
zE_78!GzLB2z(yCzV{g!0`6Q}qluo|13SX^FUFnLkL*DB2MDsA?`mrKjOn6r~U1qJP
zV9;U$;=rZ#!ARo%fy>;|ljM6&&(%V~?}u^L(vZ-<Jo+2UmY37z1;5<VfRf7C8#fA$
zOpC6P1*IwhSKmKAONTRi${O;w5R4O7FRlJfU0D_xFFb~mL#|kU>r`Y)L1XOfcsKdS
z4?o>#bVN>0&i(8-?PK*tTIn;a+P(yLV(Z&E&njJAI7dfwC-VTMD6`^Vf@@LKo@*;y
z9Vli(TqKPG(i{>8J~Fhr@JSv?95LIa`Zn9YEQ$_YW}N<Ni^(xo!IkT`9|KCRPzNnL
zH!Zilpfkc4&6DZbQvh_l#4E3PyqAplLd!vfiR3>?@W;x`4c_cY$qWfm_2EWpCj<{s
zJ(|BidFR#w<gh3C`AApQ8<p*#y7Qg+n@Cf06XrQ}>W|&?JArd^_dhr>G3cOqbk#%T
zb(&<P)Bb0?+mk7hGMke%L4JYQ4o8pQP6;93l6p$dE;pwiAxoeI0n3wL7M%bL$8egV
zdH_<Q*nSxI%xN!j(mmJ35d;z*KDZ5&(TUP^nw9v%FWF(K(n?ERL%7UKO^5P=!*K9J
z*s#1mt!j~@J&DTs^+PDVzony@8f$&j<kgs8?xd@B&1pa5b)3f_gzc>?+r5vQLmZkW
zovpU)a9|q8YpQ5j+ASy7_?p80gDtv;S<S#Ud&;N!z@Pxz!@aZ9sYyf#{h8aE*U<;!
z$S|SxMrlU2D8Xo=fZjLU5O(V?HSGw-r0*d$?+Bm9>u;TGqK{8aUY_d4{H#cy<<TX#
z&4-1(Wf>Wn+m)Ih4!5@4kG{Q1!+m^Lw)i3^AtAA))$8Za;im?oqLpFyj+Aljw_jjV
zcf*?7=VD`{5}W|xa3_^}AD`B>tHv!bDKXLG;X?+0-dEC6!MR%6m6qFScliUqoD5bn
z7MooejcFe*YA6vj268;%$X27GAX6_J)KcPiwh^@38}hnA@RusOeje4MvU1D%*8JR@
zA8HKCK;)-J_@~&|t(uqf?y(U~IrDpWy<|lm6E2zO<Sc@e+x1{?a4xsu!2^FOeJ~S)
z@QXsz<pLD6FZWl$P30N_)^n`8x{XrE)Q&@49kmpX>twa-c=>WXxAi|LBh3_xA|nUE
z=B@eIMIC>vp`r}9RX|I8$EmEMP`^ng92giIONlr*IH@z247pDUsT!t3VO}r(g>d&F
zmm)#OZ5Z?4|7y2h%{qD?$1|jD#`|VmT(t#G?d;9!6gOuW?m?q2{yH8~mwS5ad&gX8
zZscF!QIRw4Y21AB>mdX9Q7?l_^vt9%dR|dcmI^e8dWLKicS|ZYj$SAd5zUK!tiubT
z$?~HULG$p(n)cCGG`DZ4^=roDAj{-XRyN6cx%QZw!4LJVsi{UWTdkl+>nsHEM!gh^
ziitTq`sLhGqG*qc+cw<)w@22%C>_n?%dPB=e=o5B4F-&#R$tfN9(QKc2hW&qZf<Qr
zJ-f_do(;i7?xLgfV?oYqgg+gRQI1%x+Dy9wF>m2-NoI@yiMn24JmH?yq>J+^w(Xy3
z(}s%vGHEg}oo{Pz-{0Glkbp!a^*xSx69pD4?JBF3z(8!d)ZS&r3}dntW15k}O_Qm)
zIjJz~f<o;oM_1q%7kz2dc0c9k;8@S|^gI$_VvBP0Mt`t&J4wYayioSkYS4I2Z$9zB
zyGTrIay5K}!H<esINj*IzCNlDN!;a{0x2QkeUTLW8@x!OJOl8lE=KF9cOK*fioted
z@w2^<*3K+Zks-k2X`^U8TuG>J=)=36R)5yF;4%yh**#G2A;7kZ^XSvgr7S99j$jqE
zJut<u_`AQ*4#}b+gZBJ}#zv%f?)Xl||4LUZKEFc{O2owMgH3rm;k(1kq<cG9s=bcC
zudkw6Vb=^F_@~x()`wY+x5TV=4y{XUY|4FmV`xHAiW3IHW5X14C!pPa?)ID-d3bgh
z{NkmdCBP;b12&rzln8iYrjU{ghsdcYSlb__tB1yg7Z2#OAhF!8FFYGE22?ZZsQc((
z%%Pyi#%uhl`jjIB!@v(9q}$g;KAP(p9<8ju!3hG!iuUAuLh|zRJ{^|DMD>IGHw+4L
z=JvCnTfgaupd;j*ISGb$B?&flzE@XYtlDoPQ?53BzWB5mn&(H;XmU%siL5dfQmHhE
zXwBuly_Hkk?r=~4UUK1d?r*_KA4<nsT*Xk}Nvm=pH8OrDUmRssVTZ`p;a9U)2i8-5
zj3#FcmmjrK!6v?`@`vWWD?C{gUny;$*@cDW;2E2%-Wr%ylPT7$y1hIXwx<}qea}l4
z4(r^4{5RJT@EU!<Wzab9Ca4~)K<)CnB;rp~`em~B&D@Gomsb}n4i{O)`|2My$e!xA
zagdN8m`*!C)jM)N)nnhlM)&xtuFez3<JTJ#U4Vt~x$V7N>Q79h7uDHo3ZB{fjAx8<
zRX3ijzp>dnZHWdF(uZRgqWMeH6-`-HOvJ`9|Df8`#8MoPqTAT^Q^M&uF#Y2%bTy;8
zMe3c3gGb$$nys2I|3>mAau);L+uO_6Fe)}=Cp{uISQh-%d_T!V64$)FT33I0<ALtK
zdu+Qgje<Nj`zJLMAxXljTrvG`--Jlx5z-dLtI~#?kRPCV{0s{N7QSMI;vap8wLtU8
zkgw>zjEncgGwy&GXlCRYfMd$erxgMM^j|6RFW=upuTAW9r)v{%I{JosZ+p2Y8`#)y
zsvo$YN1Q2%h$+e8bM1*@CapX+@PSE7@mn_b=$K3<qK8PH!KPMxo>v$$)srIm;`3*@
zZn<}Mr$8o(;nMWmffGH)PANi$UVsd*RMB>@U$e-P8UgECZy)~_Q)PJtIj#3)SV%~X
z>#6SMM9T>jHQMgyM2KE`AB6ZQSF{33H^Vf$wx&_+)>C98T4z4aW4G-7b-JLqm?it+
zbq_$a{BPFd89%=ms;~!aHf4Uk@pW7@hx-NVXo>sT-Yex7|3(H<3m_~iDW-xKIn$WE
z&VY(wcqaX7?&;H$ODDv-xCR$S_LA2VH5k+0P6#teS?Vu!-!;yXA}vZBQntVS`wKG-
z0WbwP-%~gQ00GF98lfr8i@QBNy-<m|twt;kS`?eR$0tNmQJ%XBay*KRjEEq{yJLCK
zrY3-iyglwo+2=t|9tM<e(Uyb+#lxrKG>`1G_r`wXL=~8#$h0<%mAJdRn8t>O_xs>>
zyI&NumLx(#-I=w;q6}pNMU{0L2p}^%v@7ul$W8camr5po?=SyIh>v$$Xd}J!!9__-
za7(aV<?5^KuJ0`@O~*v6tgL*akB&++Zqz}MY-n6yJmHHM@~OgpLp`Tp4+IMF5jKXo
zj>|I-iW$jwd3iPG&}0!Q>Go*GmvHQ|kSwf4(~b?Jz@Y63zK^V=DMVm^vu)u*+i^(P
zoaJ0zeeeqQLnMIYQLLc*^LA&~^NrJ!agvejhJyEib<y&V&^?zbuGpBleR^E@y~Hp*
zlYeVBfoFg5XBeRhuDuHGVT)wVki}JhqJ<L6XIoJE^XOlD%(kCjgtAG7*2?5y?ciZ`
zY}PmX_b}X#Kbn{92gz&XG(IbnHpa=Zk#Te%(`_$2HA14O94?E6xrK(t2DYfS=-1C~
zo!ogA&nuZCYiY&|b>oe<9XZ@Eh>Cma%!z!Dik{}<h3lJH8ZF7Kne^;6kP*CrB_Siz
z(aH*V2N{|1=kt@8=qP55(sa)|^py(89W1PI+ls(D+%50Q#F$QLtrx#ia%pJ9^SgJ<
zZ|fPpi)+hC8{|Ey_vuZODK9Ipy4;e}l(zli*Q{=`nk3>gU~hv}``%y?+%&aH9jJjd
z&C8n>7Xn%rR}yB`JmY@jGut59)|8yo3;{jgF3DN?FOgP{S#IV$g7@D2*kF(%pS{n|
zv*^9v-;)49=VW*3Pw<SB)>nB&!HkX)hw^se_8kHO88x->b+6#5e~FDj4Hl;0<ENuT
z?RNTzhiAs?$}ZjiL|i5^bQBB)7o=kvjMmm*@4UZKVQMbWJ?4^p^rP|LU*E|)Nt`lb
zVt~qNf)KCgzIE~fIA0)LgLyR>R=hnv1ckeh?LL^RjZ(YR{o_?oQc8V4oP_tuphdZA
zHtyjhnXf#&ni(@PF#SL*%VJw><GaB5@V{Ju=zIb?S%hzV91@2lPAuQq-GxYv_ap3B
z9@jc=sq<@S!fx$SeUF|?rS0E}vGGxEDobrF%%!iY!1VB`n_d(NfzkB;`8KGeU^m1s
z%&(}pzTJ7f*!=td8uJ%B?64IOX~@_qD=S?8B>RRQY3k^}ui=*=6-np(-@c1Slc&*L
zX6xV}zo^J`MfD;#F%J)!)3Yq(?E4&6Fg(*qf)>kmpiRqzxnNIFEIhy=$MOtYA=V82
zzt?)TLgIh<sf{gqT}Qcaa&mfshS9ZblYEf?xgR?RIlFdP-d5Xx(0hLD93Ecxz!|Yk
z7z@M`M(&9dr1z{EA|BiozfrCKCoYgg-09XD(HhCBV#y8gpddZn?Qm2N`p$U2VUaj)
zUIT+g0KSw=G|VSlb`Q+jBV0ZvmZmQ$>CF{Ry~V2iRwvT%1tm#Y=wqEFob1X`|L4i>
z^h#4&+6fN#<hLC-)VKBWpNeU`f1;|0rDXhH=0frF^6_%FqpgYKKi)rcwObimj`^y6
zO1RK#ZQAZ|a45-thIspbfJQQD!_mjbFE%M08A)ZLe=?};9(-Gf3DoFdK7k0;k|5}d
z*u{4V5quL(4pnB28%&fPp!&+rMyUk{_R`l+4$sd8|G7vUNB^0&>hJG&i)UYME3;-N
zYhR2Q#2$>H5_FZ2IFnb_>d<?9>tqF&jqSqvQUJqKZkRl#U0`8qYo~qn$m=hwpC1ts
z1zGsWo7_V@-P!#Z&gJE0B}yI!?lQ1aNAk`OCjmTHVtE`u#FSG}iXdq%A4lvfE6q#G
z++<!#Nl7s?GovEhYHBHV5+eX3_YBzCRMx&TQsjgThi5BOTN|!6e<1Sxc6C^Iq{726
zNb<uWH8eFbF)-8<y-+Fi`xaa1e0nfmb$SR*?C$PPi?LuwtZ>pbn4$36n1qDKfwKz^
zqVWiqYt|Hh;rN_!W(CTU^{Dj};#*fwGq_6q_)%yAhh5}d8y;k6cvw1)Cp_$;WN4IC
ztJ3#ju9mJeN?!i&4U4`SeMz$L)byX}quCWnI)d__&BMXZrqbflEm<nEf#nj<n7L+i
z=whdTHL5p_%FeQDVB>q(*?jE<n#^F1(dyE}5`~fSb*@mkA9fo<-I^uNVA{4#&<LC}
z3bG3tTHe!<;!IK8c_Jc_&$s;i0?<63|3mAD<L}K>*#yHue)qL9%lVcRCY8Gw7+3WB
z8B9U9vu7sL!bsOZ&j`%7U%Ys+vO@MgW$mal!PC>zVtl8ivo*VI*i}-Ji;caFi8&GV
zeS`n}sQYD=Nu)V+9X^y4u{oOdz*(FauYkZ0ldZ^XwR<cEk>PP>`7VyN^**-0{}vD0
z^4Ojb!kp1e&F#64S>$)6yZirP+4v=9r--4*%F5o$tOuN1IYK2jAHxF#i6P;5rk<u|
zri*RBvd~Y}f2M$v^lumLi>vb?%n8FCxZiqo|Md$kD(0o&clr5AA>V1f1*E(xobiu!
z{&XdYz@&R|V|rt0DLql>x(C!)_%|xb3JoSp-Kq*cT+haNffjg-z!D9*JX4!-|HmpU
zf+7?_M{xIUPXEQ_uI6WKs%o3%p3pRzk3zC)M)@vkS~?$pM+|^srF?Zod4QZ8mz}KZ
z(t0McnA1d8cUHZzlMtGEzPlD}-@CUH2(H<dh}qUpYOypGd}La@^9aw0As~*y>4Gw&
z-TUU(V$G3v7C}b5eWpqmmBg4wNKxu7^dzclr~-1ehXm=^5+^$;+dTGs_Qc4SFM;G1
zzRE!<Nitc%VZn9U(U=^`7p~aA0J!!Ub$n>(XnA*$EW`Nn?9`$kcti~&BTmppQQe!~
zHE>09y{m1xGWQCRRWi1p{?m{VxR1T3qOPi8V<&TQ&<-jfC;~#nz9&RpN2;9nv!e4d
z?~9yX20<ie9snUIwcBUM8PF`r$|izAhRwe$HSLfZ0k=Jv9Z40ujV6LCevV7k;3|`d
z80M?;+O6FY)Jgs-vzVc+tjKyanu#tKq97llV0?9mv<uHG5aaj+)LfH$S~qR!hbn+B
z>Z?KA$`Z|U4a(@&<c0tf<&1|+XtiUG)+W^^>ywwKZkCQ^+z8$_O=aHCYX~umTAOUT
zTRLPRaFWEu!7fyX=4WkZykd-OUDP+idrlG&qa2xCUq?frfB)lkgCjRO+O4zK+w1eY
zj*UxU;dnHUXnzU4IVNdo<+ycx;dDAW&-{i7t)-N+Pr~@2n|Q_occd^Xyfig5vV<Zn
zR#fp&k%x~SVIUX;1`mJye#F4gJUYrW|HfEOCrQ4-E{?|-my}O$DNje5l0+Djqg;#M
zc`z$1I2;A3nh<{UE8R`7Bn?zyNOJNI-^hCJ_r15UPn%ntBCNSLlP6wLO#620A|AM(
zy6Na<8v{=rTdpFt4ng?5K+x#xOM=MCE7M8l#`9+}vT#Vx{P%vt<dYBqP0(YV>j=ps
zEr&ZtOYJd(VBrp3uo#_ntp`eL1nZk}JtGC5&$8A&t6FlBje3j)l0$GO$;eio{kp-y
z4XhA8Y6CW+7#?D;=$<lb1NT!i=O-0E)m^W%WKY6h!1^DX>Qn11U^9^s;R6FgX$ond
zgFfL2-+s&HWxD#RQh}&^mOk*s11uK-s;k1(FLJ%O2<Fh$Rr)(67LfRNTRc58bJgKs
z?K3v{@cjG#@{W*uth6Sa18wh)cVODC^joU)B}tlyT_DvF3n4c*?+p2HrV6XPlgTm&
zT!^VI)pG4~@o7IlB_aw94UOk>Cf<XH6rJGkKxqYWXj#(ne`RDWEa>ezob46a*kWUo
z<$M#9PRpv~<KHI_YE0`+&k(%(owjda>*&*Pq5k-81unu-G9e@VTDwHYbJFvA{N4DF
zY^v81GQd15-5jsb4ha=2$$H2Xn&=Mf6<Rn}^gH8cXRPW4?@cX-rop6PUZdmuI_`#`
z!{a6g*xG{pl$bSeNDx(H<BN+68Ycy%!Ef?ofnYopS5Ux)fJh;<=Z&O&_qK4&#S5iK
zNAVhW3n0LEPEuV}3Vr3>>4b4Pnt<ml038MhF}BYzQM@I(mq)@s#|$U^Zh@5yqd%mp
zuS>PS;Z<k6AOiy<F%e1QKae}Rv(j-YBU8?s6J)C#Ul%(EOLjpD!}8{LC17_tH|jFB
zwC@S0&&#W*&|;A;|5+@3ucwFV@%i6nub$${;o<<QK#l%s$nELSDR`UNSywOQo`Ar`
z#kq5uOu|L%(F2;Dex<YkWfL>RBc)m5r3520Guz|G-_8__xnZ`4P4Ns5(ekRP%~nN8
zT&TW-adNWCl?6?oUFcs=JmHtoOIB7un(gh8AAYcw`W3y{h>%fCT1i?=Nl7Lm_Rtx#
z_!&^qZK|!jQ4{~_olG9~Zk{FxZe3iQPJy6AIY%Q(TM--{h*mjFzN}hfuKD0n0UX1O
zd_;sIB*@vDN_EpZcC4)Q+;793{Cl#$E>0+1U%%L%puKJk+)wUEGz1EN73JI4H4q_l
zbMtl_K^F(@#YDh0R7=dyGx|+#?Va8SEa+sPJlQRAyB2y8_;`WA`Q^(=z{XFadyD+B
z&7~g-*w|pWd+yr!Rc2)5=iYEjmm#|bb-(`m{HG=rf8XV$6K1e`KF56ZF*Zwzgpv{k
zT-}^z(1imr7YC@}0C|8UA`c#K=@ffW>a6try0}0Q@xdnWn1c3mzh#*F*|V!@f1c{^
zwV`fZe}Ac-nHQKz3wl$Dbs*+KZx9>I)dXuT_u?PAz!Z`uq?0<ke4ZmA;V*GZAHU*0
z{1D?RaMI@FyeRxG!qhb{{)HlHZKY+}(l<KlseUYvGcf|kyw&=MQnSDhVlQ|uGyMH$
zkFA-M)n#HIU;OS|-x#k5?J+Zhm>BR}o2XiK9vm$31#-}cy80{aYF}%MuGz*8m7JsT
z8f|6e0oW>bUxf~+s_jF+AcuAl5@xf|woL7LoonS``f&8(p@~@}%*%{BO<CK486YoG
z_wxqqi@y@-pA&k3CxYMi03E$P_vVlWl^o5RPPY|vJ3xgYg>r}6rSw{WF+MdDFFh?S
zB~q#puXlO_c*PzsQieu`uNGWxUhRledw6Rk6DwFLOecpDT#<U7fifk-KI(h)*!F&)
z%Pt)?^>Y=K$n-YWhefLSl?!+lS~VLfgFZv~IgEBj!@G-}h=<IjoAN)H%gNRI2;RK$
zVc(LAZ3Qw0wbvth3U+oEJ76B4|2)CE@VQLKBQ}~(t@9B4Jv!c77*t&s#t<rgiweXW
zk7npM3*HM?zn&Q?n2#i&zYL%HhlOJwJ}*JSRniju9as#drJ?k%@A5cUj-Tv@G$_Q7
zbLE^~zXtw4VPUzD<%sGLPR!Iq`5$&B%pPNMn;DSD=2=lo`nv5>xf&G_v3tc@af|cw
zXtn-CM>yJ6)SdB;6FZPOD*v0}5z`CE0djDWU?};D(c?dVX}39|sU<(Nvb;RALPl}>
zond0<+sE7*=5woqideQFFS?F<p~;dZ;(i$y8F|QeU+?70^7+XD5#}{?bZ>lo)^@2z
zFs!ED>YmLA%+)SC*^GL6;))F{+V4`SS+@e~=ZQ#2#H*dc<6=yPN;ad52f_n^pI5aW
zseOiuRsz!HleQ=Urz#-b=@}~885zm%45p_WV%45tlfP-Sh6BQmgaUF;tw@oFhzs7>
z{(comS5AoG@Ya|(ymb2|)V(^s>)J;D!~%k!lR#y2by8I5Hrf#f?AK0Cwl?~-E?P{K
zE4@$#fV?=w`+i@5Z{(n+hMO0bGh_rkGh*{Q|FIC%@RmP*UU9Vl3O+ivZx_E&@fEx5
z`>oG5G|#V^&I}SLSifXt@sQ|1$$}tCO5skn6P<Y$M*+TnKA|?xn(U@_#y)Wcqtvc{
z>*07O(?*Q*Mj1{D;&PPX65LBumR)tzE5(`SIp)+!I*!N3ST<nQ=;)wY3J}`tvh&{3
z_E9%2t3@shcbzA8dW<J0#>US6`VqjMlb60-q}vz|=6CxJCu_%z-*K0Xmf=URM@yxB
z9RYXPBGun*zy&ha1#M2KK?i6dCOu<vxI8@*($Z2Qt)Ip5+EM7fD`=Q-gDNbWR_D6j
z{Mu9v{e%88$JZj&F0OlfYo?|ccSxnb8)_*lkNhxr{b8frX`eATPeD!|BAlY)wJZ=o
z4r~s$REzX(Ag176yc5QtX=9TyQQXtRc64mr%@2ut+>amU>7Q={aR`|Ib<=KNgKAuk
zcc|RY<iaBDr@E-02uk;1$ujgsRBse}wF|h+`7Uw3SzCW%b~Qvy(r+4aNQ5(MQ8Lo)
za{X`wNB7s1W!xjJpUAEBPJKHh4;ajOB=}&f3V1Ost*RFuO%$Y4d3%eQ0h_`mdtDy(
zP#LwTE=kEqW=+s196f1^ePX@vC90qCKVbL8(+W`UI%Ws9tCN#&J7a(j(QjdaVAWPK
zKt!g)e0D+lseZ@3luRD{#f#e^5dBs5;0g7I^Uv5wV$<7e7}xM_@ctR&w#M4s&0b$~
zK&|`ZeKSGOS<9B+La=JW%9fHM>8~Q;2PGNh7ysqs(|x5X8MB(d5b;r$E+;8w0M(OK
zL4lG9<FnY3?rsDRc=*MVf0dTeuP&eZe%l^tFAUDoQqx#fA5Yeg|5&KzC!*N+guYvo
zlBR&@5A1f5W^xQHb|g5T&G|t-0bqGe0c6xhQ-CJ@!P(2A<7LWJf6i{p!7&ys`wpfi
zVD34?W#0Ez_9<*gIV#{f<8+u)p-J$e(lY6LJB9~150M_^R-<IkXDGQ-@q~SkrNAdW
z-kPol_fsXUk<xjRf_IN8c{yqMpYrm?q$E98PQZNC4pz^Z(y2^3g?_P=+cp20$g~G~
zdpCSMKdre>Ke@lJS^D;3qUK^{UpKG=JTNb>yTJqT`Zz)YrE7hl9|mVtJZW+a@Hxim
z34B=Yc~2$q&vYzLPOOabH~~#Kkd~h7>RFm8(KewScp(oL!5pYxl2e!hu**@uNdLFr
zoWHr9Pd8ZP?(n@lYR(4lT4^a643G3BdnIeT`@kSem1iLPo+pxx9$GgfKl|lwQu|N5
z|90Ebga|m*PSd`~la#D7pFAjcCBK5g!pZf}JP;wjy1nvw&B+chVz~)}dPjlUi}O?5
zwaLjJ;`YRnp)XRf{Jezm%j0maJ}%*yrr`#cvzMSgu_YKu91ZgH5p?@vG^H9G9Q^9l
ztKQVtk-rMkIr;d;-g!3w`t~8>v9a`Hs{cOfIdCMi($avq3?vh$EeOu0AUzw-WA@4q
zvoF91cy)9aoVRAG_(3e_QF_$*;ZH{e9=j>8g%_Dp@PuMKU*X5Nt`^BKfeGpd097i@
zEv|tb3%|uU(2gX}Bjjae)3e58Ds<Ce=GYO>k(uVODE95^Gha&3X@ksyy|0E9y|c5^
zVoFt8L4JW%{pUS;LCeH^Wgnae4<6h|5N!P0fKuy3{sx~5i=UZ+mrDgL^4EVAME8Gz
zwg$jjK9{Cx$U}wyE+HvNL`cYEzrqfLdQdEL=miG~%}IKxUyJ7pWgYE;sa<;Tx2VEa
zI9GpbS+}-ks-KUhFP*x%fhI7D-+Cab&>*FzTjWNC*F$(=nh9OXX)CG@)#iV`MOvVH
zdSLK6-UVXh*qNp7;lboORVJ#yW;rA}E#<#j8?T>lTL-j%n?5khHAZ}c0%-HOu4aUE
zJyv$hq@?wqan2(m=w^;X(gw4)NWd+AIy(B#Vns13U~qPuV>fv12M+%MMY+WaG*d5L
z@LWxWw~t-eX8$Kum_s$PBQEBu%@od#{YEMO;Hk+#TUZf03ZkhzFw)Yp7@^YUgZqnt
z0hA6$V>{{<`hlNfE0L+ylD?C>KPD`t{vDQ8gHlSEJnTGO#dwez2U3hq!h-YjgY!RV
z={$!=MZpJ(-v6v4_~WO4Y^>&9MY?a33ZVz|>H=;jdz*rnQt^DhDsmnH3$P+*16;2!
zKsYp;_8RRLHp)F3ww~6Clat{WDajd9gaw5KUYt!G&F2&HVqy|c{Paw8PkGpOZ1sK`
z^S-Ae;35(FU%v{^_)U`<`6&%EA)**Ia(rDtKKp$|`B%!K4Z*F=eTAS;$uue|Tt?;_
zscD0YAQD1ezLa7ISbg~36Y6R4)(QCliE9y`c_c+B5j{P_D>cm3bQbe}@bwi?Rc>3`
z+e86T5J6f&LQ?5QWfRgRU4pcLbeBPhfV8wUNOy-)(%ndhv~<HiH=cX%cfW6pf4RpQ
zM>p*Eeb-)Vt{KmKp64nrWo4^IO628?lSFSBTe6y6k>#|RE{%?k?qa`l$EGCR_+1rW
z@b`Ok&)%!j!1)e+hY1MfY}FTPKm5O@BvcyrFV!{qdw%c?V$S}8c|>rCcJ*SduQxuO
zQg-I>J$m|2G}~;ysR2<cUlvCa-kK=4VQ7!K@cs9Z_%?VMn&!Q9EiHs}A$S;POGbS5
z#M5W*Vg+&vvN%l<+B>R54;NM-S%Q#?eCj89Z{JnjV_#q$f|AhOVMJF>+|w8aANB~9
z;iOM=DgU2Vh91mJc8GOGO56lwQ(9ZtO&|Sz`1nX9akbKol*IQxY8+ML1g_$xwOA1K
z2x@+4TMKV_)j&ao{^25m7)#01lZ&U%o0hiPprpY~4D7T2zI1f!O#g6@p<wWue28qK
zu#I6(L|9-%WbE4f8oT*g<^RBOQj3o~Ex+;7M3={4_ch`X3tJtvRF)bZk9_?Lymvy$
zpKWE%?cFW%)}Rj<9|kZkzXh~3bd3B6!a;Xm{PK8YWCWTql561}$zE_}(^h%7u+q#?
z6F?>um9$X&+A8j5-Diu7J9`NMfW~9BpU-6>fP}9B9~(~BJHMF8rzsJJHeKX;>Kj0H
zY_~&bg2*k%TORDpwad{&{x>X<VJ__!IQSXf836`dMFz#;s()uoR041z<k_1C7e$(^
zc6V!#LP8{5mi4t?*RDJ_y|p&&7kOQU%K`DTwOv+TK6vvoMW1c1cq=+OI^xNTGE#%M
zpx}bu<wR=48(Ygmo6%yaA_0-Z$jx?$oX~53VpC9(0uS*1=eqj|(Q$ESRiDFatOm2c
zwV#evlq`w+&f)R%gP2)GxRL^=;bbi&e{+>H(N0fGDyw|1iVzcnSr6r;wVa~2_`9}L
z8EHfF_5F&>O4j64Bx4@yHddq=dOj00AiMvb(C6*N7~#ZC`(q}b>&AriRPUbF3FUh~
zxOkx^;jO&PXDvulPGZb1S_>_977%eeKR|?7aq3D+g2iU;8q67{;^qcM{b4;~tSj2h
zL6kTFp;6%gJB~LFUD?bx4(+$JMNBR&>A%bygJ9xQ-EiK~nf#I=Zi~mIBh%HD2UtGV
zoZ^)X+K-h346eKJMm2Qc@}2$dMP+k?8d4mm>yLMcGnQ0o&=y2x3hqzQAh<OuT%Sv1
zrcEZLrm)*^ru7G<_Ey*~%H`nRgzly|$J1BP(x&y~iH=-Q-bU?1c`E7XNAd98Vy2K+
zByrC7rovoRSMOcE_Z1;QEPx_#OZue7WSP&*k*%erXS4{O-=!KfU?q>?wu^qnKipz$
z(;3T`Ocrnrb+2FL(n{CvoL6&fqIeyGwA^blGU%Y5V%+gJ_=jBj=H}KuuPazbRic!j
z1VF|n1nIaeyb83PeVY+=^jnr<LRbXj$(zB!z3d$GqO0|JVkbw0rR8O8uf9U5EBY#N
zh?zcW417lXV_!vyfq53qeSg1GlH?jx-r;T|Bk3h2!w~MB_-E>d?~(C+3n4?1qOHS^
zaeu#@ni?%O<pDYED@s3QnJk%^ZHsN)PIr`8xT1=<pkb={rzjeNG2F{U_q-eNma#XG
zr*QG`G&DAb^`EZ9{jCG=TkGpzMR?!Ezuz||gZT3Ry-(tE1b=XRHx>`)@y?(YSwiCN
z=@~1_{MMmlYtOgk<(_5%q=zm?Ev?Ogoa%IVn=26u3^x%p40^3o_&3f??0uhCdIoCr
zTBLve#Q)a6=q`MC>DS!6B;t41(+rDpKnwLn1`5piF)|6%4%xfsJznH1A?Y>f_`+{>
z6=68&dX4GX1KR%q=tRv<KJ*2AB4FV@wR#pzZwcNVhv^6|4xZG7yF6y<^6%Z@NY$&;
z`*y9!hVu396op76H~RU$2vW&b#@^$43Z#fS%{Uc1M2^SXw|z!CFbUW*i>vK<HKBQ0
zPA(sDS>dZh-0l`QiOq!gsJe#-SH=9be|6-fd4P~=G6KlhY?^jglkC8GH0izpFwpI-
zzH}lEvz3u#jKxH7L5wjTv<+t!7G6gDt*IeOZ4WjRQ>7bfsy2lm^riQv?w;7=iG+um
zj%dq_=0QajFpbZc9ur_bL_0ra`<|$i;v*b)miq6;G&ME!+dh99o1O4&Bt>f6s{GCO
zmJTe$n1+UdMBI4a6hKDk?oTlVWen!rE~xSDQv+7eqUFKm{!8;K$lU*`B0pfGZCQQ9
zaqitICR!~>UH8#~Tnka+7Gn$#hj0l)NKTxVycbMk)~y%QJXhE>Kc0E~F9>FH4SaK`
zNi9+Vn!~H^(4G^?{aKMQ=j$I>wt~<PvGfGKhN8eXu^L6`lV~JZrlV!=K0#y7`LB6j
z20=>{CnBJU+$Qo8gNtqDy?Twa3W`X$U;P|3pNQnX1rulr6Zj$uJ^MUuCn3!H_mcOK
z7k@zoxid~Jug7xw{RV7RCHWVom3ciZ%+CY`d*pjtT3Rb?m$_dLy~)~LB{DCapMm^O
zHtAZ4x<&x8z<6IC%$7;7C%dre(?4U%?~Q}CWN4;o*P70!L=5s^SKWsT3@V%)p<-^<
zoFKHcv_w9x!K7BK{#N$*nDn@4y%QgsDWe~<;)V<UF}JBCK7Z{L1MajOatf^~*V_aK
z%*>lFuq!U3I&_1oozu6*p}f6!x?`X`gu$f}(GkNFB^_DV*zyg~JvV$(Bah$y0KgBl
z`10y4v9=Av%r3$)4}#T6NlDvmD><sI@xh@VKYj$F*((UB%~0w6R=nQnCzQsH&&c_)
zw7g-QwD-IyORlV%b$`<e*B$to&x&0te(&M{ip_lJ&F&c2tzrW^NCCv9N*1V8?C$63
z?yQF8>a{XaOP$hv_k-|_wl$Q|%&cA<@KU$&@o(eXZO%3WA#b)?jb^^Rv1*O0hLmeH
zn7$5{C-09T$zk@+>B%)0k&YTPuXk=*m6pT3zg!^sm!Rh9=_wJx?sri}R#CX6D+?#~
zU+tuIK5oY`m)FbeR>xiD$GFhl5noAyxopQu$E%!xY#-|VYu$AskPt+?0f@7n-W;ma
z#BHy3?nGekzIgEBf&AXy<GA`V3(%ECb@<nMzpSHYP*Mz0;{CI?x4AL3c@R}%78)1|
zsfn!18wysSP1n%hZ~37Cc(w`0iHu2Xg{H$)t_5DV8fwQD7mP`c$3ieO7eeATUf-ix
z`*mJP&Z@EN)A-*9cBsW>{LZ$Ijoae~F#P<*0_0`ihv?<(^vE3D9W&H+a?({zjfY_5
zY@&Q?|7~yT>50wgIx*=U#8AHY3`?+xT)ciKg4~=gOD*Wk)FOa*5;!|Zb*wEx5{mp~
zYC_WN+FD0X7Xo@|#DUvJNIVQto6;UdEgjJ#B$B!r78(}zy)({ttr&0vAkZ4h<ood>
z=y+zO$6L()>^^kWgxjot4-kKuR(D@blx<gF0;tUE%8t%BI}|devZ~71OjRUNZXio-
zq9S}e#7O=MVxg0Vlz&BL;y&!G2iIOV1#}+vEy^Q*_&@qTP5iS84vQw1me1sZew($n
zv}RPRm$+)!YG#M0QdbKKzF1{rlnFHdJ)#_)=_Nz?Ed;3X=`SO=2*wisG5>vX4o*(8
z2=;6(rs)u$joXdpsx_L3$<<ZLm^2XC{u~*L1uNP=n!`@BD8p@aH0N6;CW4WXL-!kf
z!HsLzCPPAfPSb3cJ`6H2(aHd}dT-@d)c&=bH_;Fv%o1DWWImjqoF^cU#G?%mG}=cy
zbBdJ`ip>!7@^of>H0LElN3rT)3snn5*Hbm|mKpND4$>_ZxoQ`v8oxvHHDiDCXR@VH
zZgOr_5*Lq2RblH}<hPFp^**o_cyEnWtdCaMG5Mg;()`R()YUO2BqoF`(;`7{s`wZ~
zP_?aWtR%_BXK9K-3oFnI4Jt(}3B+fV*oeQNhpeu~{{woUB!x5O+zu|yg{3acWhlwl
zB;(@Z($bamt_>N1pk9GNdysY%*V}I&jiofrbe_Hat&hC);?9d^;q)Bn8TNT=R$}&y
z$hx<m5}iz-QIz@td>l?NuL1t{_9Aqy(w^fTmOHvIAPs3Us4C8DruO+i)qOA<=XO+7
z=ithRfdLaUhZ6&vrT-n~z6crZFHsY$bG`ZQZ8p$p+bk<$x5NL4U3kR0qO@L085F+f
zau{{hm$J~qvz$=&XJ-z-tQ2=)QIUxv(C2-lDT0L><jQNRXlBMJOVbn85^TVp%P~Ct
zTHjgw*Xq(t-H{<1ThUe9A9IspYJK@^hNhViipR4m^kb16IstX|4T6$>9-G6y>XRcM
zW(~Z(Q<IAPp@<@lo>zu)G+#Q;9|DJxk3x5-D}JJ<$M{ITp?6tv-#0h6>y4!3)bs2S
zi3sL{moi-d(1${z)ag-97yI%+PJoYZB0?gDyC~K}Jng;4GnjhZ_v~NK(wJ!JL!?C0
zB4(Od4)Slb4{zv+i(bEaU%_vdtAe06#jbl%BBaXP;g{2&+QyOc8d<k^KO=?6grEd)
zv*n!X`v4|4)Dgi)glLm!hhtBTy>oqVgt6+&dJ0V{4XRE6)2<o~6~UCG<O{Xl@q$pF
zDN<V3gj6p;g^zK!n%6zz3WBYSgboJH(}7g!8WvWzQYS!~_f|*aqT<}fyEk?+lxj|P
zNB6Jao}HUTe639QarXV)JG7g^NMShyYhRzDT9#3WbPm^b$7AJa{q*#7HaeeXJAegH
za#1N%Km&9NRy>x=_6=pir5h_qVZ{-ieF`LF^BD#^e5(=ebPTjt(djh{?YkzQKDaWF
zf`V|7N&hvFY!^J~>xz$%@!>;ba;ZtFa{g$jb(}^C&hJNWXQ2y^R!rV3bZAf}uXpmO
z5E24?0^m5emy0L#+rOMHt&V#C6osTQ6l^|DlvPGYS5zFm_qpxh?23Ffo5^KGgHTmf
zg`8uvRt-+?;bla=LS$4_RCFN@n#9!tRH^1ai3z3kxmuXZq)*XWq$aB?10%zMJ$Igo
zK9_2?-M<<i&Me0K7wdWd0Tu^`9YRLdc!6v`jrpb0izY19J+l41!*r3SvVj>HYE=-e
zJTmS}L%5xR9(;;{-UWBWJ!k~`_L`tCzAB~l2ENf6MiQhdW>@~nRUB0NM!-Sbm+<Wd
z1bw%jCCEl`BmOZe<XXzgG*@aZ`)T_}0=~!qLbS-Z2yu^|RpIS@VSnNSuKOFnok5i-
z(yOO;1#uBko|)PFyE|3-l!z?u-ke|M&i3K05AW-d@i%C1miPBAGbDd8wp-eMNTI7s
zp*vIg<82Vm*ex)t-=K?M>4!dGGwwdpu01(&C6D}7v(BMi3kL1#{5<|=W)ubwbyL?v
z1A#Q}<dDGIOqlg&TMN+nfcwE_GH`kS2vsFmm`_PfxpU{vmHnd}%_3D}aUD8>U6u#v
zOyru^OmE@g3F96TvYznIKpHa#<cdUuzZJ{tGZQ&tUPQHB0~AlQ%K2dJ#0&Z!^JW@-
z4B1_O0Me<xM7BGdBRlv%w;_}$hX8oA58TDqz&!TgX6>|$boNRzAIZp9Qn>FslK20M
ze4z(Zf+ozf+@rc0UY{`i#2W5s!@)+vB|wlkUKx*&h-DVqPJBQ>XgboDe$gH6^g`07
zzl0UHo;@?N>AL8S&qR*o@b$J=87@sGVl!6o+KoiOQ3Kft#nq+2Ixn9eD_mAd0%*6u
zy>|)R3g&*ifq}pzU@q+ldv_Jf+QP~Se<v)&0^7J-Y-r_hT<#|$Rf!<tClgJ`BqM5K
z<Mqvo=DvaU6c0Cdi)^~n-ns$v-F`yQ=O*ROJ%c_WT=)L4jHP}DIk}!}5}wQV_E+~v
zidJvn-@aYoHlc-!t~cTFE{5AS)BG4vOaaFybXxlQZUTx4z&dyOt0Hs<)G5oBG)j1R
z8L|}J_U{&0ewbU3&I68-Y0I+p@yfFGb+^$$d<2!`UAGmtm0xzw@^hzXf7Y9gBiGnH
z*23iURY;yZeUg?R>LMiT$^5Dm?JG?<89HOBW6oQek8-;~H}q((k>B~P6VlTYCWS6(
zF!De$ICz2#O=m?y99ZxcPNr?Nq+_?D7o6qGwr8Ujoa2~kUK>7=oSHnnbmTfLjzs_P
z>;aAnH!;E;uZfq7Csb7k85_HEMryo11!4{kd+nAQCLc&bM&4Q`TegoB-QW{%rlZgz
zC-+knJgdwz8hejo;TkaB2%=m#+ou4xF^A*z{_63;!PYzwo(bs~<j^yT!?6RxiQpvn
z2FC(>E?J&^bI?`EQ|OTcKVWg_dyRiPF-R>sQnK&b8b^kgtON5CA)y<n`j$SF|F(yZ
zhflPp^sT>9r0#GRMXR5w!m_tAT5>!S5bk!IH+3uSCI7*zohb9l;V#k0pS9oL@z|~*
zk+RfMQ}RH*ZZofrot4$s$0MBDkAvMhAg_?%r&i+n>sx<DS+O$c>VZH9jI6%wUrwHw
zq#<uk4*iHYG+j=1Xhn9A$i7sA0F8<aeqq%d4E$4VZ)h@=-^+z$%<0~8PUo3LcKntA
zdZ0`$==g9I7~~xPP84_3z(M~o&aVj#<`Ep`W&oUzVhJj5PBLRoC&Ipd*;%~lBBBc?
zg4%$(IPZQ!;Wp2X|7;^C4_Cxo`G9fv4<X2WP$hW3#NKagE^m)u860lz-rjh)Cx>;m
z5$mwPns!<X1y#h8463|4B4;~;fP1ew1<?Id|21=RkI7J>9OvzgCO?|BFlV|C-j2Iz
z2J$0ozsj+@ft!<%TL3764;57e2k&8@4?lk3VW;naFW=40jj!9|q=w$<?4(?Tbm$C2
zurINC!qk+w+QOJ8I)<1i<Y7TU3CB@1YQ7XDY8P@2rhOj<Q<;zE@ljKY#LIAUasu<i
zhrM+HvoU)#uS4@}<0KmSTY<jfd=>TFJUq8$jUo~(l=95du3bjRaBR)4IHIB5mi<w*
z92gj7(Dp@TG&hdpjM#0JTaZ<!{#t=Lpm$+>)ttKYqYmwTl*Kc(Jmqvn{P2)IzRaY;
z5byUFU-zh9Qn-8(kw(->U$9N~NJfW(Dj|G<X@9Na2AltI?$jtC=5Yf3WH|Tc4)n{+
zGX5N6FkNfViC1=>8L>fq6R4p-xd!VUI^t|@`uh0XPk_QLn}|eKd%_3bX&oJSsT0Ql
zr#(Fh7UGGoZY+~Cl6-39EK?=-_Z^bf>cc}KE$2M_AMbr2n_%<CS|)_n>;V-i@0)vc
z)FRULw_G+F78w0#i~9*lc&!ch#;d~o|6aDuXtK6Ru6KQ9EN`ZulRtduqX}`MLb~1B
zQ1h)1H}7_}O%<X4ovH$co>&$qDO)MVl433vo+GyjXAM1s(#B3zviL%$mUgyp@f#t^
z<gNLR8%rY%<KL%pT9x2#l|z|@%5I#@tmT#_^~rrlJ!&Bvg!&`-rfNRV9z$W_SA}gq
zr#_(gDfQ4j4L6gXw-b{h5Xu*<@D!B+5;>8ZyEm@pxTR>hllWAlB2)FGdFb)Zil>>R
zrbb~}TGA}&F__nE`6ugMFe$K@*}GWvvq$#)GPq3bv64v4y5s%2n~i)_%>3QB(taeP
z&@sI1<Aa7pO3Z5CMV%o)z-ym7^?SYM46rooTK+ySy-~Co*SUa4(U6k`Uyy`HzisHT
z?)3EZ8K)Z~1zkpAVIPoNbPgK)nQS`RUNjvyDS89ao5MwRoulTJ1OTC5R7TwZB6BRv
zbhI=^<VHG&V1N%{NC8CxV0Ctv2c^>$waT66e5UG1R#wW<Y)bO5n;IHvTUvcYr61&B
zgNQ@w)81l|Ar7DiHj~WvkbUzeZgP@e@-YkW#*OAZehio6v)oR5rXq!mU0WltTl-?U
z1|5a2{+~W^yQ*g?9_{acBIG-qbJzk&*LI?p7S9)>xvW5qS<!Vn(80iyy<>bJ)`Z@1
z?haa@pr7u_gCPA+s2%+^d9uYoY4Ncv1+y^R>gmaz)l5@BYjZ0_LeJ)A4WkBob#{2g
z(dh|Vx$R=lpR&=*D?}IwnUaB$#eJUj%A*P6hE0k1^uCUWlOF$H6*u@m*9~41&RgvZ
zaju6;COHvI!j7v&N*Et~?N&!9af++s9u>awWV*LXX=_V9JCB++($5nWO6CE>JD?5<
z1k^vuwrWlt4<Yk%htN-|Q|xtpEGelpCK>;mu6z7s6?A>Z8$9du(KYVgj(bZ&m;=KX
z5DNW=%wnoBVJ3gP7YB;A{F2nDVgdSMg?5vSik!&f&%2vp&9CxrQ%eI<My562oI#WR
z7CJutqt5xeBL1Ym@=OL-@D2EYvQIza7V3@=OY^g{3g*)ClUoOZo;OICUmLQhx3-8r
z!0|yoLt>JC0zzS*2rFCqjGC}pagImD>ajVrN=wScjW03lDi78w{3`0LI(fJDbpwDM
z-EMJQgE1V<cI|bGN~eItkYvQ8DcDUE!D2p!kP%n9je|$8YQ5MJ$}#Tjn>ttnmFo=^
zP|`S<IVQ8|1PZ-7w{Pp(ml@48a*R6|jMSX&)SS8WPVQY*6p>HQTK%l5(H$L#y6ZwV
zR3~CWY}HXIC)dDZiFg`cUth|E`+}aV@+b{wsz(MLW+0ETUVX%<WL|`&9^oUruyT0w
zuC`&TB+ndp9~ExH(x?>%#~CYsUxwW}h_+|{swhZuGiVONU+7?QjkK8<uexz#z1?m+
zE3NC*D}#W$kH@+Q@Il6ksolY7I;k0ICq}=cyTn52LQMv;i>z@`w`SURn?m>z@@RL(
z;czi^`L6B8bewLJpRPBnu}wk4RgfNLDwtat$ksR-&)2iTb(jm|iPF3>@X4PnV44sT
zSKnL<b)P&z)0)_Bld#d7TQ}&$B0HU>(DbJ~`QVG6vDZYUzB<778=D#k`(fl>{JWRt
z&l}roYF-myA2>L<?ftTQ=f!Kbpxx<`&gMM(bM)cECVx`M`IyQ+Z*E)ctvOnAa@{)h
zjef)CVY0nEn$vh;YHDgRA_8f>de|@0HXI7QJnO`eAcTv~Flk+18lI7*@=`SJh%W0W
z;tlNn&8N8@KEafe=^&XQK|b@ivQZMqxj`O-H`hFK=V8w^Z{sYiSDToKKkg2vXMdU^
zSm*;kHwm09=YXh4qgOq-Y{1vAq)(p~EVO|cvYriNi2V5R=TFfB^+Duw4s%OCcPtWP
z;;#Gm+)c(RBl;vBKUJGvnH~B)rfwR`y(wnu)gf`1UZvr*rHKVstyK*T3Bw5#b&r;{
zQY(DZ9nu`^L!+a7d0<$@%T3G3-%krMf(1unwx*=EH<e_h0_97Cw)XzHKY<h;B$uE~
zi_dgqD^b{wwKZ8%R$*IvHA;ah!@|9<J=6pje=t`p+ri$^QE|PEiT;%_*JXw{O=&$n
zbn=nhQ#C77&CaErx$^w-rftnJFN5smA{@z$#H)`!<Ks=|+(EeC;c?h}fbC#kuEzbD
zU4+Ue`1vny(R-6>+;|B0u%EnK70wP9u=#d}YH3fv3T%i+C~)h}f4VW@Mpj0d&t+vh
z*Mb&s3o3jAoKKE(wPaOCBpyd#lHsZ`dEQ4uL)+?T*1e175wTEs@%*aV-uuqHDHjl3
z=*M950bP>QQ>MGnvH$-4(mR!^YKeGnKOZi)DbTH`(Xub7zoz{`dFtKGzKQI!;^Gg?
zo5#mjF!1~M8oxwkWDrnXHcW2Yx{pjue*kK4XRgH3wKG8DuA*u4RAwT#>&Aq<we<}(
zR^_|`gN{BPDo|)6AiscG1G$O@oiS=f#-SDgv4-o_E;ez=)H>dH^j{`I;x8z(@U1^(
zGfkIh%!<i;_Yh#uy0Oo8ocRO13(ISKCAo)&Z_7%p{i3C~a`6H-)FLP72iXE+>itd9
z6uuFnrIGFOa1K<SAA?{9oE;Md`kjoADQqlGK%lUR)Tu4V;t~R(pc`vr>+kQ6RJh0V
zXxn;eY|#?hLnI_Z1Di|aeY}mvsrh3xqH$X+$%7Ua205Q-%Is?=6<BS3xPrQW#5~K(
zF?#_mvbxqXrcMKIypFgD>Qd>Q@6MzX<KOwe3PNk26PxZQ_{kJz>5h(xi2!dis*|ck
z<Fh}1#LxA!JHuP`<Q@*H#CJQ_yu2ic+m$zeq823zD%=Htk)~nba{;V)KUz|1G9MhT
zT$kXUVOW8vZMen7wOXblpZJATzKIeoEezo+MY6P^km*Gx8SK1wsRt_c$DIXvzsk`T
zeL7BD2fDV#3Oca^6;u7Zy*V9Mlr1ec0~1Biu&#A7(LYGon4$hMS5CF`D}!SYXhh`v
zE<tNcnQ$lReh}y^z|J8g=0xc>z7riPQjW_}nVDOL7X=d|EuFmbzhak)H9jXCS)HD&
zJ}-=c!`&NEtS{fj{#LZytbZQtxV^+?v^i_Sf#dVDy<w-lWSqU#Mv~QWn<m@h*j`~L
zG2O@rWmg{Hp_7qu7yj!}4WsZF)Nvu$!@oT<{ri=Hfx+S%1VWV<HmSOXI=kuE{*K#O
zMtav>ZVC#DrcZ<~l7D0bgf9TKta?mLo2AC>8>GBLi7q_kE$ywNRR(sUi6(=YQ-pAs
zU|(<F%KaO6C=Z+*9mL1qhr#A*)yd_`OA7@9^NHLa(?PJuyhW*Oma%dBC}?0qZrRTt
zyh2B4OSXk2fU3G<C@E3k?@5nK-S|wx(8OcMF*x24uXR=p9v<GGgZc8bMSv!Af#ZwZ
zZuy}EC8%FT;`rHau-u6_Ov%al>*%#k*XM6yq9z#;3EZ}$Ipd;*w|(R6H=9QqG07eu
zml<_FeDO_Vw1h#qG?%LmmOz?XFcTdH*`s^TwFnleG?@^W(}o50iqg{kr-^C%b#twL
zUM_`}$s4UIECwXJ_P>AT=?Z=9sebvA%=wRp1Qic3?l7TYy}2~tx*_xg&5fn;`uWxI
ztPlHgFuT9x=`P!U5w2Yba5`Khsj!RVbU;7%1!dcU3yx;S%1Ynco0G*s{v}fH2RAO>
z9oNHyg%{kRca~^lURc^93xOmMa&v3LCDIxy2j6SXH07EMzDV<{RQ(QBilos#zm}mZ
zki1hSoSq!@^JSVV4}q_gqbY6}_7&ch(2M<cVYJxH>F8Gj{1iv)Yf0n-+}9JutMz%h
zosuz2OSANO*mhVe`-gNnnv$aSTYel2BoIQu_#%IL;tYzR+uxj&oo?EZL;Y%k_%WE9
zFKxWbC<q9)&GOhVeL;Ua0N~}a%(fT*hG6=i{+^Sn^rp!vYZth)!n~Cg2kEC!AK3!G
zNlZfGBdXE8uRwTy8FzxRvN9Z5)x(%_FN;e+sT0Xh4!qu1A4zwtrCZG0O(5eqQTQhF
zMeo_)9~bF{Muqdpu8(h}_EhU`0wAhbr4`^^DliRZwY44@N?C_2+~wxV%IfNsD)X1m
zf6=fI`WgNeotY4AZOzBc<Y-VQU{eUQYP|v(frHJr%N%CvH5P&y{dx2z|7?!F0^c%c
z6{=u=LVjLI2%KK}f%blnjCt0=oQX+I&B8Rb?aSj{`f%_vV5FISe=D@$Dqy!ilvf59
zI9P3aZ!yEmTlE^GUza{cudLK0Fj0+{+f4Tvc^a;y1?E(pmM71cTAlK`xlS-S`G_XV
zeMFLr12IU~^cSGTtZO85{WK<l#0-r6t+TC#pLq7|2zPen{9<1hy^%j$YL4@D=2~84
zD4CCzS&eqYytw+?$NTq5viO;K!$<u9EW_ja>oe3+_sFHCrT>Z47lxI|djyVk$Ff&J
z5Ddl>76^_GJhWVmsDodeDd6-KwZdPX0?h?Nm1--qAnK9D+lqh^m6X46-lN1k-A46l
zYoqm8J_2a{OHKy6?^@0O4ozS@RoYzK3*~arfmsm-EmF&E!vcxFHeOqkl$125dN(pT
z5iV-z-ng*R5-LU>B>LdfLBjblhPoN|#=!zBTN@h-8$QeZH=(DN$@%K($Zl@bd#K^w
z_;{2Q_iEAH=HI`=e`|5)(}6e=HM*FMXp7E{mRf&{hl`1cDaT|tKE|wz=65NNKw12>
zOORYe4ecU!QxFfYScgDq`EDZhvp_SrC8M^p*YxM_KM~J`sQCZ`KP)VK+2`BY)K?r1
zIQ=AkLPP+={f?M;&DLqY9&&bJ)rIfAhxgqCt=$#bkg&#v`b^cbo`z;0Z+)q_Y{hS^
z#zi0MU;_53|M}i((HT^ZvtwyWrQ<-nhEb#Z?qS2ZZ`R<8KT#IEaayqs8`P!N{ZCSo
z8XhKy!+hoLEBJd{(E;XPg*=kYDh1kd;tQT7J}xvhZJ^^nEd82fUAH%pb8I-bYdRV!
z@Uxg+OY_aTCMmX25jW}0J6A3szWv}nr}IbTKKsag1-l3u!JQsXZUCW^rlOwmM6L|^
zNB`=f%I*tUYBh|UtgN7)lUA5i*&zuVZzM%T!@zF+w5o08b??y)=c+2Qjh^#~LKI<N
z|8OTX6e1+2MQsDEOq%xQDmA??Pww4IfXXR{owMmU<Q-BZVkMB+b>z8#5BORx(mu?f
zzAd&R<-%~%C^HL>2qV%zyhI7iz(t1(d-?JUQsKHfFuq}mSI3AUG4C7njlGiNi0Je4
zzI?G70(69-4`E!v5yW4mS|%^pML3SLGc0<oJw>JWT~OgC0wI}U(5Y#(XlF)6X~A^l
z<)!t}hs$=wzFM6zSgfoA-EaQr2GQLFcYhLk{D-g!ZotQ>oqk8t|MYjRYd?8n5?W7o
z#aq?H&1v%QtypR0sHOHBQNU4ivfV4?;nDo?;hE;?OV^k81zBB>FZ*5ve_ZCy?$`6H
zBj{N5SCq-cB+5f0!-PRY*Hbd`(Nk$_IxdIlr00AfkId?@Sw!T6oSmOOoL4hf<5Y)r
z<>uJu_;-aO7-RrP%V9c}ra;SfhcKI(nOW|njio6e;gi>ay#wT9y84HkaUR2->^nbH
z;HtS&(1yw3YD9rU0_mE=>4r4<)XlZ|kx{Pg<q5l<q~?qKyHDUMcxb3o2!^5H4>Zhc
zWRGgk(;R)5{`u$JXFwYvK)$)V5yAN~0X5kO!$<#FzMA<8P5D|`8YYYIVF<!YM~&x(
z!TrwW))pnX6PlAa8vH}o-G6VEk9YB&B0(uC&-xxvcKb4+H%!yIm0K_BEJ90h=H*zw
zzXpPCd$V=RZLH8}t5bkUCBLs}w6$kSIqTG=?&2jL0!>Fpd(}sG5eSpF!1jY0%JJY8
z0&Al2V!oa<Keom0iV6^1ii^C8nQw~)oGHDy2+*W``24TUgG*5MA4}fkg%!hX^Q|O2
zJS(fTcmtS9UU&R~4_)&7_ookUY3h4>d*3e)4%`@Ch#TZ1{d;)lq6x{T=iQ?-&PP3W
zp+%|qYbnRzvqEKnFO653KnipH1{wcYf4|?+==8$6c}rQG+pY?Ho0PV@(8Ud6$<B5g
z)1xKkQG7pG3_7b1moi$<PS&AmCDO$b;a-fnyb`Gx`Zn`@WG3n{-7D(#H@@fVg<(y}
zR&8}uoSn=s3rDB@J2TRycM-mzzgU2>@KA13(1eu}YvF)Q$w6vd_2Gz#KBgtmv3c62
z`d_H9TV3l;^deY5Ll_o(E-+E0`3Xu(2t@xBV;;frQQb`jk{^jZONE}rpP7D-rO!aV
zBPIHue8)U?f%a)@{LA=A^md<M>OC^vG`-;xJSMzrK*&IEzy0U!yQA$J7*OK)@vC>H
z6O3oo**JUv?EC3#&7I9RZrcYJ5F?!bT1kW<bfL%xfueWCMRXW{RoFP`P5t;hIXU_H
z^VhFm0dQH{5wpx^(i+Bq%C#h@F!7%CD`{E86c}`l7uc;A;%nYZ;FxeUX&cIOt*-Wz
z*No&l`T=BDd|3ejoj|tW=cMVjY9r{h^5rqGYZHSyM(9wfv|`TT{Bk4xP4EhIO&JaB
zfbw<d%a@_j{V!kW9zV`Q*W<!=D6cRWE0bR@Gw6&y;DQ*f%*t+S&FB3^GO{??M3EX^
zb4_h+ZPV0fE~nk4A``&qfsEHXQm5BYA=Y<hLQSVOW2t!I#g6Q3%<gJFr&!m+MOv?x
z(3XKOK;&YsNzrjQId!q_Lr$g6*jdg)`Hth=(MHKAYsZt1Zf8fSajwk7EiG*ltv}P8
zRBeQWeqFlk46$XEz@pX}x8v+J)EO3#Z92Eo_TuK-Qx%+P1%<w^_XUl4yGu(i#aE-u
z*Aw49q@9lqE>@j}4vUqAP1fUZ0~go*DK8MyA}3@vlv(F@`9^$V-@jCKw0kEs1mK5O
z%|Xc;75qa7vwi9E`eo%PA;jjnEBX=8zu{-1cGRetj)vXW-`8H2BI)E`ai;+L8h(3R
z&7zkd(Lh!drz<bEk;-lXlkpHcKWAyDlF6W^pDz;wV{^+-&Ofb8kZw&%N)f{J{aF@e
z;BR+sV&}Wz9wA|Nyjn7$$aTferz>ffjI)+4HjX{(P*w?lL6$-UC(wTz+8@}C`rXEf
zpCbcqIJT{@TtXB+H8qL(MKSR1jqMvwyESK^?<jK>eC{2b>e~LZjWrwX=^6c+Hl10(
zAu3scU{JjMixQ_h%DjrrbcB_fnz~M<RQx?vsPz`r;59%uG+rRNc(D{HQ=SJ{#_#U~
zk$fi9L^ka)vGX!dGtOMg%uXE19hciV$n^9Ph{?jNXw<cpOI%~76hG<GT$fEIMf%B1
zpnQX`ddB4y$QVsG!!xE{gDT}F-cs?)ET9HHPbe3W*?2*YLn&KUUHzB_zdYDtK}$ZB
zTXf}F-7}IE5=PBR8YbFT-ux;%79=|ae0+QvcJugn%*fSOtE7~qhPM=RltMzW10PYv
zz9cQ`Nz}~A=}YrTQya{#RMlru&+~t*uarsiB7Hf%?pjw+QC74D1<=XFW}r<iuAR2k
z#HRC4uG)<cmm2kg*{nHoO%zq`>?{xuXGDhhrb3z=ygjbl>o`8Yb7~m-V+tt7DBm80
z(6yF%W@VZBU8=u%bSF6(rUz7R=v8xppQr{@jy2WgR^}GD9NT^iVc-MAid~1@Q&eWc
zD_x(8i9V#fT$TGH)$II|PUG<G2U>C*pZD#yLh_H0W85<Cir@533WhF{2#_|gK^f}B
zjJyAZ3ye}yLX?y>aeO?^`@uGyipnxikhVBJ4)0J-yt}@Bi-yj5qOR@u$KmqE3?sHG
ziJL%8PW5qqyusZ3JiE{%$?v_r1$Tr@EXL5qj_VO^8K$dGyRThWx?L(<Q+;AgG_$ZS
zf&}L5EW2V2r`=n!HK#>V=f_Xh+lO{Okdi|}w9&^yw1)FtedU3zx?r+gKj)K^+==hq
zc@a7n5$?j@xhov|)apXLiAjhT<Ji<Qa9>gOy+VCPE@R3VrDek@aCl(y!KrX{Z*8Kr
zt%JR4qIPh!qWP^ce+Q!!%EB~nZJ7|?W1?|>)H7Um@B=>8Q?MhObZ6iVC3&9Tjgtb;
z42!1&;DJJG$b))nUNK2U(!i(}`0{WN=b@ux95{RASGM(du%h|IR7s^am=haT!9SZE
zV~pVjm_%5H|Gaagq?C4p&#zkia_@h9Ey#%ex}^F6SR4%IW+CWuJnVW|_wA;rw73&z
zzR6A^oZ6Q643q*KO-gaF<2n=F2HvnS{!~>^uYc4=G4QINV;aOgAEEIDN~J>bwH=w-
zJI#SrX#AQ*vo)jwnLXqt)G$$j7kR>nZLDw-$D9)8y7u{W3PIOZp|lRHC%3zkTf;RS
z7n7I%l@qCoWMhVBx;MKLiJU>tmt)Ynm>3O`$5;AGFOH~MV1)c3LQ~BN(_*Raw)4?W
z7^${To58qFca*KO{aaS+*nok+G+2k8je&uWCD^<E+0=}n7rWJKc#L~Lhx_ZKbej6#
zrDb{2VJ}+K$9vMjSM)?pr$vKYX=S3ZG3DTg@hEN9^pv9R9l~~VYH32v@o)he*;K`h
z05UF}-()9!@}9c@n`4R<5O5kFtB-mPg}VutBhxc?M3$lx8~^EkbFh~`DP=-8g{Qgy
z7&*>#owkPRK0boshO0V4^1_+KVB$DoHb}3geWgm9iezFcwPS*`n`%tF*kiaEMB(MX
z3#Ng6dL9EKQ~bY+#LpEycUGMadP4u^0L+)R7aP6yF}IqjGbWrbRm8yWU*r`>>*@iS
z(ioGZ<kUg-6a`r^NtrsEUxzl|<Ky8aluQXs0`oEOy+JkP11xd%JU0V_RQ-?FC-46)
z^adzp{@K{X-Q33^vRJ6~&#Bp5>VJRv0YsUq7_%!&-B3WuwpiRu9%+0v){J_Na_A&1
z{qm*s%UygCY91b*Es%nM5uh{u{<^yE=<X^T77abEKexqOZ{Cq2lj_{s@_ne%e(T>k
zO<&yx)s2+ol!u{nl#~q|*%QhI^S*#R<uEu#ApX#NF34zt9>JB@rap#)*}txDZ1kg{
zYHnZ@Tm%K<?5D49X{!EL4r!kWwez~(0zYIEDm%r!*F*-LQQ7>H@ATI3Lql$LcQyT8
z6A0}qpRa!t@p8w~a@nE84t*F3Gy<aKHm&==$6b!*Yc|mXzGSJDL{!bHS65C#*?iD^
ztjpMjzBSU$67_ucNa&!MCCs6BPCnF_hV$CnyYrE>bhM$zB)t1I)I`lmh6YzJ&TXS?
zO3%V#<DIbVhF5Di(V>8&{ZT^%o7KSgcV7O)Jc>y)t3-J0P94#ZM*^S=wNxz7%v>NR
zJbi_P98!Tuh1m%Z<0@I1=E)fDg2x1wPj^~_{A5cwZ@dCYu6opf&p1DxjtOs!Tole3
zRO9xX4-SkVFno5!Yl-qcYpAN(MDo(R15B&U{f+yAUXhf~bUj^kDLjS#*h^??#tt*V
zXYN2|=Rv~Qn7TMx5AjG})Rm0c#l_*$np&!HO1UDUhVpfsyr<M~30A^bKep>7>pecu
zL>t9jbT}X`8JJ7)HzYOOSgICBqA~>t!YtFQ0+Zz2r3;AKw<CQ5KTMa}-@Y*3K94%x
z!%|H93rsXHI~$FQxC_kRpq}b{FyO_OzAhWc(!wVu1R$fTf~uURrWq>x@^^fVxMwhE
z$)H+F!$?KNNY&WT-3&f-&(z;{3$LfA8__9cidKidUVM@%5B;ga!dF{C0N@E9c(-M_
zXn#YCCsxxvqN&nxhqGIqagK-Skev|{ilL#QXXZx__&qq`O2TUYr);8AmHU=^7}gG~
z{qy1NUCKR;!*jjvPY(-AF9ii;lMnmkrTNZ|w{ABEh=f|`9#|ga>$N)WUfbN*+1lBF
zKY`DaQ=-2(EJ(S!R_C~RrTE|aS|kWz@A;TC@74N3I+oFS=5pMflj`<@ru8N<w>9*t
ztZ7t8xa^jtHu|qZ$u{}YYtUNZ+3THe?*K}F9JO?)j|3Ol+3oG@tp3VS>eQ^Ntg2jH
zHKEx^(%l*6Ds!yj|N73&N(8fO1T|kTC;pa=5x&2>GZrG*Ok9svORAi9p<yC9h`9lr
z@;H~>VJLM;MY0<8%svFKM7+uHWuAovQXxSGmkt-?Ko+=#E_x^|%}=e}4=#dslI@M1
z)yZ`r%UEAs9<6q9$`ef*EZDqQ29P)*T{J_IyH%tCvn4WgkfkkhWE;f8_<;W5NB#y_
zDM=8&Lxw~}LBJR(80A=?9m>ncw}F2b(YDstEU#8<He6w+MVT?<xKcmepP|U4Uf7lI
zp)vq`vjM^+cAtl??EKD_p|Q-j`gnE_A1d;o7<+X1xw^VWl3cpPKsGe6KJ$1)yTn<y
z(4=1dI{oJbl-*KgtmCc#*a<IbaCo@EqBrc71oqHIV|jDDOdRl_wS^tYW2C4jw|u<y
zK-GN13DgQ)mR?zkeUwYKw6Y-q&_~Ea`B>-8=i(y91i@6&{7@v9xyQLRO5hXofR1H(
z?bjG6WM^z_3}$PzwpOI}0P}@V?acKn&%*T3ZW34@4_WxWztQz^a(F0A{J*MDsFmPe
z{NesIjxDl&=JEae-{A!T?8Z!rq}NT7Fnqkz{jE+89P`5V=$P1_?d_7%_4%(STwMqf
zpniu;O79!5aTDO7V}Xw33N9`?W@Z-DAC`RPg3UQghE`*?JG|u7BGlqFm)sQ*|HSkN
zcQjCkb#Uev47o+7q^M~|9Z1vCVn#^`tnYOTTdS+9F~lTQmAvcFi>Ly~ZpJYGL1tD$
zKM7h2@(78@(3sWb)%G3V_ZQ8Xq~W!eESJp4!0?cs<|o|Cp4r;3%1Ype((v)I_$*3~
z$Ngm4<DWPhQG02<>P&OHK>*gy=<u+f{;L;;bUva#XNW$6aH3%Sod`l+!_Oan6h0yf
z6n#*9pPZU{p>nLkGg{*gj#^&5;>vh6pUt`ya!yQdGiw&AJ;lYv?Sza^mf!#D3J7fM
z)+G+-v#P1&;Hq~SF}}PtPsjtUm&f~CTh$V~|Jd`o37nDgy`!b0YvSJtxsq5!A1EV=
z#1u3xYN?kG5z~{kc#Q4fSjzG86=L8bS4CKBF4_mhjG2wQs8lR!XwTu?Hot`}2%%9b
zSAiD#5-C_qg09Ch8nW>+A>7>DWt?29+>K#6i$Tc{$Ac8qjiv71dU;DL%dPbZ>V#UU
zm?&me1f55;Mr+akFz>_K=O2&Aj1^0^wXvjSQW}T+833D~JZaipTN}F+o82LWtj~9Z
zMl&$p&`VokZT4iBfsSWc(UKYa;P8M=MIABlu6ubW%>RP1zOn)SWBRvEx`2PBC&319
z2X6B@5;E*zSV>%k=^Pk}&Aq+1-8?Q#V!tqCeH_kcI$FMC>3qC8i_7Cw5zFUjy^6a(
zr2Fphz31Wnq0f8eqaIO}VJ-tVJ`$qa1xvlsltZ2uzw-YaF1omX`3ge&sp%5g6#I_;
zciQ`qfz=OTV&ZflH-?bCCYG;t$9lQP0dPbtDbhDFF$)b15nWV&=plFWMMU2yELkx3
zfn@SP&wbPkxW5Q|{wo#~lhe}D-VY6xOxX91*3Hak%}VnhWvT^`N(=BYva=8LTmbS8
z*w>1P4AQTfOvv_6?VUFRn~VKNUN0)s-(xuO&##owF(B*uKGBose#cn;>{<!xp|3B4
z?DXOo0MQrPqd*(Zl7lJWKKk|RW%~<pVOT9dbC|!j@JshDtvn67@5k5mIs#g8QZW@n
z-pDYl8(%@^Vjq#Esc)qC+O^u(TW4#FMZCb7#D$(jwxCDBICM_Wx%!EUW=eWrLgjht
zku_2-d0<(E$7go!ms4}|bDg{|P_F#d*3mJwVGk50F+4U<MAP-o)~YdE)MlW+KefcM
zT>ZQ{6qkr7H9K<_n~R&1lbgHPgc@V18!$#@m+H{ASIIh>!blK)A8wW)vSM=*rM;n`
zTVP^xorfIf`V|}=C$kt5ob%ja*EL6SPa0;+t+Fk*+4<QAR8;VBG`Jt5A@;}FyplAi
zl+T^Wa`z02gdR6@4%UId>#_ZDr}YfhJDyTuY0B|5iZsktmf7_U7!J-%=AF4+)$(yM
z+_M|;W7!qpTHCW1->D52)Bea_`lI8n;?UDxqkm6yXXC{uGjF(fq@_ELe>|IUCB+VC
z>dmLee*U_CZEdyx&K>BG=N})}6%IIE+3IA#&G}$D^k+50yt28-rK)mlY^)`8j#-5#
z{>fLl3?R3%vRcpIzSn<mcEg}O@}MD-&31?e@2kfj?2@^baQv}~a1o?xmcQ3ry-Zwj
zbS?!s%4-E*th+QoZFQu?>dhM(T3WM-Y7Ilfw4aT_G`9|qj;`3ETU%Po(S~B}?s{mH
zJ35pC0uNNKG0;r!%gRy@xE(QI+brFEWHO%#?f*N=gW=w+p(AfEI(VqZ1_f;nYJ~%1
z50rFU`<KKNODuX*1kNho{ONirprs}5b+ao{I|jD+-<k}{gG-a4p(g_1=4f9^SZvY1
z7h|`w;@R?IuSU7u*3NErmtmnQ!@!4r!|!je2pn$*_ciQ*NC+{DJ}aMkvNEu49Dp}^
zlZ-3hX)ya!8-3XJ?WTaUgN1e+HH)n#%r3F}aO4vV3`2Tla^p*7C9k;1WkeTVQ97&Q
z`Du&FfU<mheCNtjzPCZ<sf-SLzwXc}UfbVcV&}D-{W(qPD=*jUZP-+73SdpM!@CzU
zFL(A_c#`!>t^^JozPE3o+85w;CO2@efVH*fzB2WH_HEJAAKBbt{S(zOEfh<?DoP&0
za&%luPh~$TeZDj;hE!tI>dsSUzpPFB>MbI&8AFyugv9yaLs~=_R{cm)@&qjUkTBP#
z9*=V?u=7MGUATd=ft{Qv&#$zyvaMcz8woM_rK?!?5(F5dbHF_8;UXsQ8K89$Gm96%
zR@6Y^M#X8#$&O%#n=FYdhrN6mlOmTv&or}j1@?7jG0J|sizLy|iF*f3ROvQVLwQua
zIyx{=R#CB0$7o1UhbRharw{4+0;4E$w5s%59%9N0B^Lw31Tv}HP^J^B)_TYY`zzNz
z_hBY_RK$xf%QH)huV$1-bAvwiaQjClBqdlc1%}VwX;G?>iX>oOkWDd;P*#ky<Ppl~
zN5A|QvKr*^wXLISS#rLn4MTuJfkEh}w-czw&g5!f5ttAnQqRAF|Hqz$d*5z(pvoXe
z^C`eER(l#o2Rj=U1I<R%n0%1TdMzOsZkoJj3vsupB%M!+1{?YLIwm>;nsk3pO(SOv
zKB_2o&OGMG*5GlP$HW%GbozlKd#O-y-y5S#g32~3QYCV3`M3N`HxDI2E2HbJaIIV9
z!KNfdW!*{^`4%A~!+Nx#{>4~+u}PG0?#0U}q<1SQDpFBWK3`e6#Jz8_QYbova@uPX
zhqikIFz~9LA$iHu6~6TwLzkG3Xlwq%bR&2{fHNih@xu?$kRju-Z3~)G1v20o=VSZ6
z?z9<#;bHNxYP-$8MIf9+@@~AZiPPFacLx{EZhahI%h}?zQy_+f*b|Fn<oLp+Le2dU
zjIQ^eg024=`kcp<O(f*xYg5DAa!x5j+{RlJ9bG@JSpSihk{U2%MJ53_TT=2xy28yM
zyHL?_)XhG52gpOf&Juak6iUY>uUB!rJl==Qf0f(%xVy)t0%n%Zn6r13e1Z>fuBn8*
zP*QA}vlRDw{T!5<RZF$M1LTrcMWZi0ST<E^lORQ2Tl<Zu=tNgnuItI`hGGW?2S^Lg
zPt??Gc=-`Ts35nYba8@)nQOg*$hyuy_H4Aj&xc{=fkvpKZR5>aORwk;3hK{=2+;$Z
zN%?3^B!?^>a!DqKi4S21D$^^bE2#7F%*@76CZ4Cd1_S|WyE<IKF*Rbtqc)mTeMrm8
z8||%Nyz^5wk^f|c<wBg>+~(Yw^T8ZQC&fg^Kr=^#BhOrfL5eE(8*v<8q{-Qq-kGaQ
ziP>n_Ew>k0>0OIEe-V^CwyWgiYJq<55(SN(`;_Mg@ld9jgx?f6&TZ$H$L!*S!(APU
z;ZA#--l7@(7X24$KP@TmD5CZDIyXM&Udy}{@u@;5^rMZZ;CH#jZysa^JO?jox?+Mb
zvumC{R*nD!IBYyMpZcbzK@j3gmh)+t_V@5uRIka(5=)-mupiUM_3XD*b~vxt+D!5r
z#_-#SNwb17+_=lzveU8X*U~p{lGfe+a{xQvWbiaqBA#<+)coYo6Hx4fc$lRh-dI?|
zb02*F9uI;<{LcGbeo@aGMP5XrhaP`F&CQmQ0-`syno~6nC46WK==%8Bk<Yf)C|)sB
zz07#WIGX_v{SNvEy_r0!h-8&Stm+l=1o9g~erMT^=$Va`LtC=FGn*K1^~^?vO&`zE
z_V>PR>?JHw;d7c>3w;#uh_o)lb^6`1ShAY%3GYn3ub(KJ>IKPU_5(#6*o!K=>zba6
zuw?v1>sRkag@&ff(fa(%?eiDkdO1hJiz>qdfTq33`d)jyWR*&;7=%Xq7?_9~$qMN)
z>d{wt$Dw`;lJh@pM!XxMz#$qoh-1S9tR)%0bE#^UR825cFBx!OW4w@2wqDGs(b3UG
zVHW0E^zxKfW@cul79ZLKpDt_3gB%|{$-O5T@8_r)2LG!9+fn&Rsp>^%dw=QWb*bmV
z>aG{C8|Ltl3rH`g4}(sjKLZ6UA~<K32kfs4Qu*^V*n5!L36XycZp|Pzmd{v4KJ%#W
z&{Waf4bZIOr{q$h$7A7*6-&qUZ9vxZQMcwhW=W-j_0`lVIxo`oQ7z6(0*d6%pACoo
z6Gz+PC->gi*szh}5sy(n%GG`Ce7vD60$_O~k0)x6IF!P>q0NAl@1v87x}0jZ0vyh2
z?6mF!SY$eCj1hg3R5ldk9`5Fg%9%2zL~IIfDMg26Oee)XA)#iyiTVxAbFd<dy8|0f
zRL&|dV-sjf@v*N8zf8|^|Hz6vyYzc?vO(AWJ2Qt_jK&ucWs-E^jMTvbB*Xz{>Li(p
z89v^f_-XydYu8nLV)z1@yzOuHID=pHk52)tCJ|7qoo{qOlR!Q!F?;^nROng!#-n60
z5`OI`63)k3ZW_WJ*=8M5{v9+LXn%hG;ldlqUNXuF`Y30Pmb2vWn9%{BKD4oU4mnGo
z>6SbSJI3ryw7OZEa5&iJ)k(I!z9@CYjyKoWKK}YeY=#?RDZOHAi8nz*A7e4ykd?@8
zPE^$JNN}gf$@D!OadNa7)GL_Q+*$P<<E^d1FYP@sNo<2q=?-_thpnyeoiAWxZ%uBb
zL`T1l@z*<ZSO91ou!)mT>&jk1uwpoee7MSTBp-1jCiigvs_6~crl08|nn5GOLjK1O
z)dF?z?jCOtj9HeoQ*daMInC8v*P_P7z0x{HcqlkIzqmGiFGuE^_4d|nx<iNE87%mv
zMY%P~2xK|0qh7Q+>FknyVs4*(R$NUmS@^oiyPXxtN?nNSTv7=XeKo{bj2<sKJ2T_N
z!9h*QRVKE>{m&em$)&e7cW<oys=&g+QV$p!P|WBxE>b0*ma}svwW4NN)z=}%w!!?Z
z-Tdal*Y?^}6~ix$rX|LW*E0e*Ts+G~^uzR<=|UpJl7aEbzHF>7ZNVIm=7BRgIp0x0
zm3qdO-KNtYiHx@PiDt;$=>-0$Q|#*>Z~mN1k+4XuZz;5VDEiCa?`UAKyXkY(4EeB|
z)nGndRK~)l#DNdNxQ2;ybl52J?aQTMhEyOKW!PcICRiu(y>sq#Yk&OQ^~gc?e)q)k
z3MmmQGNV7{<YWxwIxveGfRBrS%q`UWaC=QRUi89O)9+}ih6s7%DmI;GXtzn(vKSD(
zg5UO3tNr=2*@GByd^X}M^D8`e+b@2oa&IYoFeUH&^!pb<jBiP2-y|wb`|tz<{IgiM
z&;HCRkwl;Pb)pgpxR6}da_yXG&JioC%KV{@kNt~=esyjBv+`F%Lq47Y1J$U=`->^`
zOc~`$ul=?i&F;45tM8O2OixFzyk5Br&;`f|T3T2fXb<Nn)kl@7sJQa;BWD+<#_xZv
z!#5dBN>0Xd+k8$M2dI^Opxs#BJ~Y$d-b2UAv;LX&?wM>#woFA?`Acab_r|9kScG5K
zRuxU;MQGbczqcwr`6!lu=X1$BC5nq_3`tA)Y^$n{I@_7C?TWMZJoXk*qN4jreX`~p
z;j=*r%naCpaw5ojN_25rv6$Dy>mje`kGASK%PPe*Klys}7sv<H=>Psya};jT(J|QV
zr4}fH#Nls0w7J6JV{t6C>#u%NLEE_+`C}zwAf+)YD|l$w$;rveNvx0U@Z|@dFg*ue
z&8FHXo>SO?s>zN1d8p^sM&}PW*z+PI+<Sz?fOgSpV4XWTf}<sYx=J|1#@2d;2yb^o
zgse>4?!)14pSuab)1@vgVGR(PrWu#RMT76(mE*2;J=&+ixY66icWX&=Dm3e-tWmRy
zXhWz1-<^gfl{HmfPlMl?)e|N=q-zNoKlG(m*!$~Eux0g)22Rjl6wB?jZ7kqm;Qmry
zHOC7>O+gfSSjDK*C$7RLZPZ2nrTw<xIfRvQ4RczCQ;<7dXqaD^c-TDon0MwNAenTg
z$2q#*>EZ3hTr)r*{>1`RoXp%XBPycI;yiIYElTbt`j&Iq`M&Y|$!5OpO!XXTgQ;Op
zmPLf3_|(9F#bWWS+H+pgoV5*)7;e4s|Hsr{2UYpK-{bHhBot{0X$0x+MnXiSyFpsI
z4jm$bbazQhH%OyMN#~(Ky1U`Pvw6M0pWi&p83+Gh=Dzp2_Lb{etFd#k(L|efb%-qD
ztqD^Mzkq;avkE2PJqAA!m0NW%g4T;!pc`Pf_+cQ4@G0^gI^Js+kg<+A)?v<9Hv#xa
znU->Ddiw9!#r~3e+LsWeC!&KAoel}G|HaiJw@QYnT&J`u6uJCKJgASGX#QAinLn$T
z*B^dcO|!hjNrTp`WvfG~WpoU}IzlB4V^(68pCzwcyCRNjFGse%eFJC3N(~I3)XfXa
z9+iz;@)H@3h`X>x(W@~o+86orjP=xmU+&Ea#lH2kvF9;7Cl>i)tL?WKo_c-yjpFw3
zWl!s`_5yMp@>XhIv8*07ee~W|GIXFfkLY^WVLk9yUD3K;i#l8Uap%wPsI%g-X!>k|
zJ43<Ry^r=#?R~yR$^Lvk)Bo!2BYHr!#%bD%EhjU|Xyw`j0}B{JOiZA)2B-}5g;Mfg
zAzJzzeI|HGS;n8?70AWf4t^2oXv~Die()qs7=3`Zzb&6{%^AbRDXMi@KX|c9MiM@*
zsG~7wam%lvD3=TCQ52$CMi<@}amPro>QgY%HA`4h?DU)Vb-t~&8{hi(vH|-tTAR+o
z*@WQ2JteIsH;%qes#64`WGpsNu`x7VPa5-<y_@!zs#w_EIyJSOE;{RrUNV&(de-)i
zBT7abVopyKTv1?BGWpWPi$-cXYBs*24JgsNe=M}nd-Gp{5z@K&{EorcG$Sr_-nkTk
z>`qlxb=QODf@(D|gTvO$ECT*$8X%3C+(!0p%g?nm1DJ*C%cafPr_$OBUl6{(?VWjp
zG|$#<KVJ@{gG3@3|HY~3p056Id(Vj>+rc<j8T2#wi0I?_9~1gN@I&efbxu`fD`h>#
zh)Gp7oe9;?id`FBP9j!j4?k8ntPV%LZr!PGnC&#@oo`VOQhs>;)Em-%*Dd$&$`c{x
zZKp<YEf%Mx^N=^Q$CXfO&j>gq1+$0lum4O}f4WX=Y^84{`RevP-`e?<rKU9{vy)iT
z`}`D*XK2eYU5Z*PWUv?W)q{_RNW+kwTE~x!hRZ)2n{!Wpe2}B!OaWdDG_SzF_0)7u
zYl=_hi6^y*D<@x+l??q1&j>+pZZeMfU*@x+;pDWq&>RvHk_Yp0_c-4k?^znIsQ(4%
zpY@8E8CgXg@2<kw+5*X5oMC18@;D}c`QkH<f^3w3ltS>bV_~ti8)|#zDH27MAL)*<
zeXg!Kh7<Fpx2>BOhk@IyG!t2PtU7NbzVTZ|$PaOK?6fSlWn~**-_OMzgHyKtFi6h%
zJOfe-3byihIl6=Vs6RgH(*tRjDou5VYr9*{9Gi-rYs%X_YFKV@*hlXMW4|e8Vj573
zR(ZZ>RmUA2%;7|c`b%EEGKPEo&gM}4$Gk1gdct4Xw%q10dpBFcUb6ky`}dAp;RC<Z
zsTl>59lw5CS%1O?zwo0LB_nPAY-~AH)KH!99!qz$)*nZVChMo}cOtLlnK-=qy?68F
z>Q<Obvkl;z0J)J&jAYT<Y}jtk-N^v_DXm=IsIpMia$MWto#DX4+Kcx^5vDNJv~-(d
zjn;G5Gwm0=TqW$Vp2-@X*LzwT6)K5b7`e=VUgR*-*doMpLi9rgtc2%f)jQJ-@y81A
z`?=8RoITq>vZEm@D6x5Yd4VWh>+PAxzw1ppu$JQE;No%y_Uv(SaiIr``+Hm5UKe{C
zSU>ZPzr{1UzIpRm%Vu?Fr@MkSWM)b=4HAVRAY=(Kl%dc?aFpx(AdAsBSI5U3;M`zs
z=+nW*a2qk_>@~>I(n8h#MoQ=6jaT%giBv6*V@bBN^~w}X(om7%14nwSO7hCax<)a>
zx<RpPs>rs?%Co08^!D<5ymhS^!N#Vao2ExDVvff8r;Sq5DsrQmBmN8}`qha4A{!Ge
z&A(8A(HUTQL=3Mw>%KKjH21;Ala-1;w~*@ClD062i0k)iZhW1DUwnU^SMXNm0J?9B
z%&Qy+b>1u>EG%nv4E$D|P#pg|{yd~dUrH@rA=?0{kph$}uSzV){serW4Y`(*x}2O{
zI!KTEuP2}))VzZ(6nPv(QWmLE$Yts(7aI0Kr0k-Fypz2lF3#507HAJ#Z2<`_ka+R{
zq3?B{#&oK&>Qaq-#^Jucfj$B6vn{Z&cd|hONc}~!dgfcx8eDf1!DiHY|9Yc`$9zbu
z=XtHo8bMNO;>CKB;o|%ExHHpePb6Hw>e#%3Ize}^u}7N0qzaVsdqkE@^;Is(5Dmn~
z?}5BC%y)xQ6L%<{1xFz-i`_12m%IEggpg$O<(!n-_gGE)VT{yk{S|b{v~M6y<;ufp
zOm!oNVi7<zs%NhP5UE4i=-)9Z@L{?~!@lzzd;P3$?TGoIHH;he559SvkfA3y$j`x)
z{hhw!O>}{K^yj(wv)!S1Lm`3Y*ZBAsF-?Pq9K-DDZ>>-?Ct0(<S9VkU=90-;5VKnf
z$79ndN+sN)-`Sk0xTz<y>e@cMZ#*;4&S_4QTsz7_-A^f<ifL%InHj;-*_Ny*z(#)p
zg6|?39mWpP09m!0gIatdSihN=P|MqY^s5~}C0!CMYktfg`St5pPmc^GHdxZiQ%b}e
zt-cgPJB0ZG+<T!yrcV7`=g%wmPHkv?eI85(v{=>=8(IKQ4~tgq7WA)e;{EZ0Kon7o
z(lS^8Uftz#ABpxj-B`gS?O=W!m!7OYxja}0t1TtgR4?$fL`d4aTHeb&G}o8>H7MtK
z8VjwcvK^6|Dg-`A<liNYwBMy$+PC7eey^g<u*vanLUB+hB;5CQFHtt>t8YE-TEo6J
zd~pMLU0^^C>rkw9SC<sNp;EMti`Q(38}h&xIQv1O`unSJmCI%P{DU_dtOf?`4|{Y$
zMFv8*zUP-_E!$G>t-A8(`-46@tMq+=$6kalfC^3>>f$5%Nl<e+sz>SL=Z_nGgp-0`
zks*^@u;E6%G^9Yq+VgkhrzORy|Jo0rf2Gby3(Hk0`dP*@fKB^&oVC7Uwl}h>Mf$Bf
zs|}03U`7P?VGQNpX+7*Wv}Xxv_U+ge7G(1K-1r&7lloV1MpwU}fPIr@&5BIbLQz<v
zX7v7a6}+rrk%5^J<RH+GRY0*trARG4sU}ZTQ)6R&b8~$Y?Ei|mZ!P;x-GlmeG!c^+
z3=%e{6RiO@0|@DD%zn(S#NpaHzv*{l3-$<1seV0#KPoU9RHl&YoP+9#RqE9e+^QL2
zn~OFBthsh`%5L={OvI1|h<x}Rdj6^_Da6UzzuB{JZ2s1}@b>+g(RjG`I((n6fn*5I
zYb+`81_x(>QRAf`-T|GZ7U$k!@fa6b*YAlF&w7f%iXT)SYIz=&cuy`>`ZYif=IG?{
z?SsY-&q~t6i44_T#J3>Yf`IH`)jTl9%xWwnRhlU%5N3~`*z}D)U#n~+o<;EV)48oN
zh1t6PGwr$_+6hjsgJd(_2TTOPjL2`XarTZP^|<K&%a1x5(~{f(K~Wq0_oDBuEp;_c
zh#(5G3kKTOOwDEo0#X=$zFZ=63?aczNR!((C!k;h&F^`@k=5CHud_JC0@xjaD#l69
z??|${C67&j_<8_a0pLeP6%5~i9)lPU4-c(?01HcN<qxpw1qXOSM*4U$vHh<E)i=oB
z6MoCgBwGLfQ4vJkCb6FUz09H|i$B^Pb3U&K;gY5;DXP*>v&Z8lYFNVbs`=~xtdTH$
zPJdpI!~pY%^c@z3A=&aZrNFepiJQD&xev(FBwrr?U0{L>-{un(IwQHfm8;V0vRjz<
zBznEKKJ{fwEy1Yx;hRc7)3xVV%0&$!F7cRGR?sk3EY$C&*LuF+!$K3afBv!DT<VLO
ztf3aQ($?{Igs+6{J}*R-%TJTBTgg17&9P&>XzsSja&_WODtwi*Gxu9#Mxx;j`MMt0
z%0*~hU7dQ#q<ISaIW*KS7?Q0-iH)WOs(tP559rp9hF*yXD>o<`(YQe0QO*}*XLsl>
zO?iOd?1I@0IR}@;2gA0zCDBE!M6%?c^2IN1!9s)`&Egu4^gZcJg<m_P3{S#rZ0t(u
zKxv64UL={A@6mGh-1fG8%YD#<JTKTd3q!^*B0I9GUyqet@I-!cj`kE6?Mm3d<X#SB
zsC3l%kvYyegSaZ0C%zH;vuC!1DkgK3wLb5_f9_(u+TIHdy!!B)jym>;Vc=Exw|ZLJ
zb(DGfes&Djsd)UNXBccRur7<(c_-Yo&K2<Rpr-MaaYseLrG>{8uBcR?XXPO9hP-o5
z^yBjqI^7Kl9?BIUfBs;}MoE41al`^K(`fg{pz><_oZzRLIIs@<K_ib!a=lruF)6&p
zz2L2^loV+HMTo+}L2JUygE)&#sYuPc&7(~W$V~J##M-@ny`wGCU;Tfxy(z!f|G5qX
z0Azoi13H?MSgsgol+-ce?6>)Ze7+B@RK_y4qTsH8DH4G423((l2F<s3H2A%FW6{ye
z0K~PwSZCeYWKZ<C0G~hsH-y^Royt5j;)$-QG2iPfl&|yDUw9?2rl+iC(#-2Kiwc|O
z<&Sa90p3=>7oO{>`@4q6<>hSY=mRcJ6_%56Ibuk*y-TpNimXtl9BPc}xzUuKf$khh
zk+N+R)2<d=HJDHY?|kPA|8?GvJ0V!2c4~sl10zn+nY!P3Lya{Bb3&q3$T!h&gCW<K
z<c2(Jhu^lqV5Y`b-<XMVl{vGjiOUgmB-A<m38-+N(m*c+GZKm37IInWM3`wOPR`8E
z?zrfHPTBv1?++#PYkunm<aR&f%+96a6EY;F0IZ?Ly!GALusp$RJy;#_P|l8P?VPSs
zOyRtru^Jvu6H1a90HEwYyRbAtVF*)DdIrxb2UfKNVvv0zI@wnRo_g!kuJvJS8xl23
zZl3y(A73PcclMdCIf4|*eW-2&bY!IRrbfT+J(J=d;0qvGHa~R><nZ8d%KiEo{iNN^
zpR*E2gN2@{{Io|yH@o?%h@=!Fi=)M@z$^vHWf>KWcW1eO7tc*ytRyUq45tjkFS<Ca
zpY5k=tg;UgaapXY)%GdEvsW_rYGx3(1UR4d&(P0~C?f%2z?RQRyq&fdN&nrXXGp&Z
z#pTF8wC}%kbJAEF-gvr6l!Hlz7=BAkE-3{R`Y3Z^Ude#?5ms40)Ei+9vqhGOdDmP0
z+Go_xn#tgLv<Dm%#wbwU`TtJgw9=<i+R8Ah%{-;DwmJ_~0)K>MBW#fp%3FXnpu6Ag
z?6Yq#4G%l*?_T@)*>~1`#X=>a3t-fa2%8=}s*XEd`g)2H|H4e&a|23psaJ?FDU<q6
z=8wtN#zT$U=CSRB#mZF{@nAT2tT)B6t+vye-WqkDxExisxFtjUSEqvKe9W4cbNP4E
z>(h2m7ZS(zFCb0%BLy8!Nhp@+N|A{F-jNDXB`NkLn)Y1UMsRbbUc--LdQ9{a$*#Y_
zTMT`UZ!9;sa-&ouQ+kFsJ#J)r%e7$sX@6#$0^Y#6pFM{tEs1&l+HF{XkGd1iMsrdz
zzN2>3b+cuXWqPgcB#h+dcQfs_)TALMAGYOj`?^meYYqiJhla0NFXH)gk%fh#R%c^n
zcfv1$`yc<Kj*T9Qc@-QiBO?$5KQ%Of@6`+o3JOVvK84`>=lr;B=YNM(2i@0+b3En^
zj_5n61T<67%k$QgO-;#crPy|R!y3=Jrya{k0unt4CeUZv&0?g_uUp`!@$;4M33rM^
zUtEMLC2J5lm-elnm_#Cc_sM(ZDV!G`JlXq`jlE4eDSY?#WnT}bK0-6A9}KnavMy@4
z=kJ@aMfTh2U$9K;JGEQOloc^n-e>fm+#iZ^7?a@pypSzIs^Q)k^#1CyT_<N5*9eO!
z%*g0KA;{0WL7}>Z!7sn4$`CW`f9|KS2;IA1U_Bwuu^zL!HIS<bE2y6*#;%A??_U^G
zGNikLa}1Lgn(W(h%X2Z8JlU(v{GZMvqW|oy4nzrmy)a7HsaxA(W<PSWMAPTCx}O{@
z0vO({v?Bnr<d(aZyFR;`Ej59RgoFghQ81a%qUY(LDJOPoS((gkBx1)-9QB+JWngSD
zNAp)3DNKuU*DE31;SaGna{8*V)5BM9PYq2Xd76bpLj+L#r~V~wITEDhlhLfc(Ql$4
zE+iiBm66q4E$>;HTRj{V?cU4rdjH0h_oEcjQy+&2t%L;Q@Q19@L}hm3<)}?jbjV~+
zG2BG@37<JD-4~g*yK#%*hw&v=Z*P4AepCm2y92DLpyu5hNhP1s)bLy{B^%~hqKemK
zGnbynC;hACHX6|#;8<`SfeU@u>dmuN$5-D+gT6j(&<?kLkq<8JG3S4rcOxHb8k$~0
zXlU>BSii0X6>PM(T+Fs&2|En8UaCMcG8TaH#3@U|{7Jtd+zH5weh!)RBB;4NY-VF(
zA%occyxuT;*wF-ACFrm3Iz3K`=z!{g3{surYLdBi1R9HYa$!?+N{T~LkgzBpEnn$&
z`D7H|94}pFmiar>quuV9_hX?^n0%Hd@FHqC<+EgtE#85b7?AHl-E-E_$3{l}-WY#Z
zz2t%=IuEjsPaIyUgk!M5qM44P@X9=lD2XQDOR!-FV2Mz8%Csz%H-%e>M}M~HNKwY@
z!A!>Zw@_3ceqzxeEME+p;rweR%}dfjftZQz_M0|yrn-}BZ8CT>C(3dIy2P~Q`8&40
z>_()RTRw$F<H2%G!Bj+Aox*Z{kzH^HsG_j>h?fi`Glzb%wJmcJ#A8PKkHC$ExBX_F
zjpBZ0c(!FR15768mf%DC$6Le2V0-G<D;Z330f|Wb7x>ITRbVi+4%)n*?%cq<A<<Wu
z;d}F_D7h?&b$)hX_b?h|;Bh{wsvxi~XLtj&bo#?W<vz49(DdifaB*iOn!^vX9_~Wu
zGNWItSIe=9Se>Uf8VE_yv3q~f@4c%QM-h>*UrCX@(V3Z1oNO(Lpv!p3UwQ6@x;WB3
ziYNG15Zu(|n#k6hV>b;SR0wCpst44Rxp;MsyibdqDLl0&|26=DQ?y}k9=GkP=D?z_
z$Mu?yr|D9Pxq#8xYK5W)f{j=?ax`4h4)48BqSQEqSvL5cp+W9yFTQFHJ#-m47@3SR
zv|q0!L9o$&y=X4?z25z&-&7y&U@*;NZ+M9H#;j?3c*$G2AC(jX(~M|5_m!hFd0JTT
zS>Sn8<R6>4PZgEdU;RrO85Y;|KWcVvS*9Yqd@IxZryOxLc6#ZzW6u4h7ygst_n3cU
zr2<^g6F>iFxdbG5Jp!bWhfA)o%fqH;kf30h)YR1C%YFFm($d2D2J6H0F4E}YDEJ&7
zpi)3O9)wSNZ5x(E>8vy94t}00giYSG9{SiDxw~H$W*yvf>q*h`G0iDku`w{L%NLlC
zk)dLdladm}zo6w~x?MVG>={VgfX<+0j5yfuNtkm7?iOfkTfeuv`q(MQ(Q_<pc<DJ%
z^*W-en-s`o@5<&y?B4FwelhjV+}GPjBZgrG?atJpt-k4SYdutH8ce>K&I&5nDr~QB
zz4)sMTQf!IkkwAv2=OtIeY^An(U9Rj6k3_QdG0*DbsO2ZHu!LP3U7z_yRgnjuD(00
zY{Sp6=$)Hi(i>mM==vcq`S~KoWVWb(>8IUpx$W_x2*nJO&GBq~(>x10A`}umye`C&
z_z#<(CDEJJP8P@Uqs0BmI#arOKL!R81SN>-O%rne|KuFP^mI~P7n<t$79Ccf4U38K
z^JAYqTb!C|Kb)<D2aDCt+1#WQE|V+z@0K?o470UYAFsuE*tN*ZDEcjJdX*RgJGRL5
zI#{?|Iy~4$|2sAtrvZ9le$Ag=5Y$L+{WMzXlbGVhU3`9nwS>AkGc0sZa`35WG&S?^
zQ_=M?aohmkkE))RDJmZ$5h&-9$wfq`y`+TRb$5<k5y){4@ogY*BZ?RNnibO+tzaB)
zEHr7s>G{0)Wi)L28RlB*sZOvks^=}jve5%Xv4|tBh2ns?UGH=BeLV8=2R_5*6`=>s
zhu^>MKjY@^JPSHBg7DAto$d2g2*2`qCGe3m&<(i(qyGMu?sESSy-@O^e_}SBlqGnG
zYi%F>M#NWa2EAlIC&FOb8!uz_?d%(xMKCMkmbNv#5l@;<7+$Aq=YU$egVrK)dse}A
z_!-8DK54&E(Eno4dlY%qCvu7S%-T*H+7|FRkxOLh(am#T$D^?J_R|Ad0AxPj+4x#B
zEKpeb`{LRn%eUn=PEibAo1Oi+rOOecGtFHY{|Sft`HQ3_w<UO<?GuSD_vMm3o2$-4
zzlZBg5r{(%nOk{!SDS#2XR0BtijyRhxHw-XzGMx1JxWkyOn`Y#1V<0VXfl7n%0$sH
z3>qDthSEJ=*Zb3%S)@$n<tVis(&sj|7wI>l%|YJ63>ICR3;Q@>+4$FEe1~HcAr0r}
zi^G13D6}m_au`zSwnQgFoMiCN5J<Ft1_CAQk1L84gnw2jolNk(xw&JyMXy)?(nH;=
zEkf69u_4IA`8sM{l-YK^8}n_UizrKmoY-Z8oR6;<1pjEhte>sbWqm4{DLoip+Sbx0
zn%3&12tZa;)mm7pj%#mQAmYmLIrKR+yb~}4NN4p>zCvo|M`>X|^=MhI2QZ>xZ0#5j
zw$`KH4XjO;<w;+#Us|m32Y3E+;Ag#`>T+&PybdTRVl1rryM1@y);Z@7xJ!3;c7=ol
zCEow|tx*LwkZWtX$s7Jq$3G>iNplo3^x7||`i-jFb{sJiKBG;<)d7L59eFS`Qemxn
zSO|Yu9I25{1fia+GXosyk?b7Z=;~m7P1dKYW?mG31-~TTi15?#r{DXZL$1AVj<(nZ
z)HCgbcm)@QlX8D<b%h5~=i1{Wn(6GIygipG6<)=IK*GPHxDtD)&!`KS*C?Vb9VOV`
zodjpuA7TWKZ$YtYkvhg)R+*N(L$Xu!!n#(mF8=_B+|P9=RChJ;ZK!^3YDpZ$!<;K*
z#`Vb`)KB;JCfg~e#d-QInM4rras2-3o8*j~V%5Psxew~<sRb{jIds94-__j`4Pq+$
z$;!5?wF*#@ZWLxzw7x|IROK#+R22bhRGEQ1n@&7vEP7hQcIM|;m<uke5ll(yKHKp`
zWJKmV_D0_@bZ&KlsZ|`<6WKG~ZyclDC!0XFo)(X;W_Gv-Musc)iD(N+bQcn4D2VD{
ztbT6H^>JOq$+~bf9KPS`GL3lA$fd#vk)km#8yw_#`C96xny9Xvt)+*H-|^XgTcpph
z4?G3Nc6F|((7GO0V#vi&02gVsG6@MvM}T~Hq<0^jgCTWr=|b<0_b3p>Up~uy8lRW+
z7J2V1Q{+L>-urZPeMUq2>$!)=DWg$?E3BSH$X-Fot)jMezzFRj)88ei^|W^wg!%Fx
zaXo%<e;i4R3}6cjWTH~ac~&;XZI8-$?-Lq*l%&wIt#Usmg1D3avU|^bimYK_N{s(%
zEl!c+L%;?ySH3)&MGD#8Ic=tZX0=P7r#2qmWx96bsEIURaQNKUW0rE$`B#Boc#2tZ
z<BbP?TAfGC_MBgH2u2>deAS&tQ3v_?f)mttRTC?-K4YSB{0#V>`E(VoVO2&*WmDa{
zE)bS@)$jOk>=|(?gDleqF;Sww#WLH{mHO4D*+lAr-Po6pCZpZpMLTgz2NdXD9N<=3
zXtbO}5W4A~wGZ3O#JF$RoM3oZH#OJrkNDCPl^DpNu)4`f`smZ+i{g$F^Q!lwAtSEs
zcH`Zsy-4c@Ox&qi8y9bI9`l@2Qt-+nB=#-Gzx&yI9t`i7`slKRulI;4Bj6{UVupUt
z#&YNIXSf`{to_|vA6@-UI)NEm6+>JRAk&>0<IrKr&Aq5LAwGKOu?Q38tb%@B@*X?K
zks-g`!Ll=OG}uq5U>z(EPq}@oeAHY44~TD#^JnBOf>iSI^<-*iXKj$!;hFZDzKJ%b
z9Feh`U{k%3Y81=341OVC6?_}ib~{x!YKU)ZU^a~)uI;&f)6(MWD0~}*ysdYTXr5G(
z-Nu@8C>5Jw9B09q0`U*r<WDr#0{mAwiGmqFIS8l%M})aIu|gjic~dtDG!JCKIB&4c
zho|oLa=&D;p~8Gx>?uV0$&)e`mN<jkU$=)3$BKS8Tb)=@t}af@syaGN>vyO2ZQBLz
z3l1nK!FE6Yq3CLFUr@m5>%^o_YRgfwJziELNfUlhdyPV{WOe;c?x%0zdajouDuh4x
zD%@?Br1uwBD3FBg_DzDK6K5H=tw*^YBKdc!?47{kRA&lE6&h-IhH-N{mhQ`Y&8wp*
zyjP4KIx@?BhBF`Nbl+Y#&CkzflEVXM4;GCbm2RgAzJEZu8u#!^KH@5Vf9o&@><HW$
zCa19}!qQ>BRA=Tp$wHAIhwvQ|JE1Y5()A)*;#fJ~4Q6-qeoR}D9GY7h84Q%y24U0N
zu0iqx+I1iL3rk%15kG&Q{3P3#0g6l(l(7B(XBhL;p)(7vFf8BeG(*Kz3FTP(pZ86h
z2Hwg-)L++{y1T(=>28A$>tT(~ZQjG;Ox<nKL*}rf;~86vf7%Hyd&$m`R!g^V)er5L
z{jOhWY(Jm~MT(XIJmA^3T#|mQ#Sj8oL*yXWCZuDu?bVt|3Hzc;jz_O6blZG?H8h<l
zLtKGL@HuA74>6}oKM3TLm#2ZX_kO_NXOFj^<xAN^+ie@XwRJY+hxG3MVFA();j6xx
z2*#y4I+VyfWIQBcys=TC#Wvmth<<3y+|0w(<lAIz8~eLi58ZJo1>xW9={!F1EG)=s
z6j>A7v|f(ierooDZLSYy$yd+*6-Wf0^nMhlrmMI(boo+7G4rj(G%~31CE`sR=DMoE
z%>h1x7M%kBu{!5oV$Nx;a<Xhe6R|eZU@dun^1GzGDONa`nE0Yz*Un?-Qt|F$REDi>
ziy8t_{aX;|`+AyRv3jgl`1u@sB`j~hNaY$9=92C=Ob&8I*XYEHD;o$G@boqx)vWts
zN{SXK^Qrfz#|YVd=`%!)0wMZ^jj5ZepP7qB4$IpWwhw#No4JY$r!)479`}gaF;=0z
z8+T-iw)auTB~r@K|1Q%GxJbZu9||viWKzuHcpHoY5kjQCZ25pDo143mmmpGOCKIc^
zQ-3gfaSpq0tZz=5cTl{|lx0itxp+!qg=*X|XYp^>y(XpQf}4+zZvxgA8(2p7cmVbr
z<oFbRcgNr4j{avleOglOGrE9mXTM>w$rtde0>Np$#<A67HNlKHgY&a-#Y1nnE2Q^>
z-0e|kR3e+<+eiSVIA?3UJ5kqdK<rRegx{Z4cxyMCy1TD+lh_Rm55lj9AH2+c<E693
z_`mN*>|<d@zoDeD#Os~@RQ*Y&mp|@#VP(qJ@nB`?=Sjjm)kz}dnz(0EeyePn1tiZQ
zMtKv<e*^+B9?oveKspUPsh{wrYz@L46+1EITrXRVmLi*g&v1hMp4C#Ey>`w+j(5L%
ze*L=kR4)A0?&%3<`1Y~yjw$EfyN2Ko1Sm9Hgr6;zQ{DXd8)l$6G(8{iC{b@wQF0Xr
z1O4f5gvnza3X2)#$CF!tkPE#m*X8%Chqebeum;9gukO49{^yY1k5PMOw{91t6*x}%
zTrMf1M2TMfxFCJ&eLE(%<agLgm+83^2Ks3|hd)mvGw<6p4L!O(@*=(e_C(z}{+GfY
zNt*G%k*wrwG6v>l_-0Q`Rl;P~o)!~?8R!rjTR2)l{sA!(C6JZTjF|Iw@R1s}+YjU6
zgw1%_R16=Yigjr1YCU_N`~Ir54<av#5DTmCuj{7>;3f#7bziY=ryr(Af~4m7IUhMz
zH>y9&n&D!aT3?@lUY7Uwd)%~;wf(DKWfLph&(Tjx^Xl2rt?;tHRV!J}PwZcixajD%
zITs%qI(uoEN^o~y5QGAMBdxsh1P@%3Y&38jL@JXo1Nf<1Ir?StZ&c|=R&QD>oc)N5
zpeWIsAFa7akkF@3S#K}QM)B^?!7j#f7##IX_Xzw+3Aa)dDD<?`cv%8H)&dmX|r
zYj|yM`V{>Hb;W*p_jD5@b?ng6Nk=7_McJ!Stxo;Z00kN!R$%{J$lt3625GCVOe(gS
zf*q_kwiYZ9ExKnL_&NfR$`J^pzuIYPeIvG|zFEvC+`>pJ0N4qQ5vs|yz0TE#=Dr%;
z%thDH2hrwrtJECT7pkwCEwg|tDkR8yIa=nL$Pik5p89R^+YTWj7g=RV<(x;2wsQQI
zs_?Kig3vq-AW5fRkM&kU_8%6bsQCE#{}}<U`FEO=jkM<8<-5~{Kk3Vrb3m1Do_ttL
z^lCZXxvhXqQ+xrwT{wD5l9~!Ww!Cu~%ol-be@^68#7TA))y?@K(A{nb<nxF-6B6hB
z&`HVC$ydKFc*eTs7gh$gFgZSAN^Hb91RLV@8-ATOvyjB%f*;vPt#E5+Cj}PKx<3QD
z@<B~)4#`DU?`k>OsU=@OyY=#Xfc{(6&&`0f+;%c!-bi1q&WKHT%8UPgjmw22KYmbj
zI3;*%nL6evFJ)TG)2zp#tC)8-XB9THS2-swd*P|`<P8K8*C#bvOCyUZb7i0QC&eGc
z;E^+$hwsx1uoAJz|M&JN-FvtKM_mY<)9!z}-JKOg&ea-voR3Bbbq4Gl94I|!!56)M
ze>R?E;QmjSP@fx5P6UXd258il>*Jmksww64Xh2(VNr_?5Ls;Yl^w7}Ip}NnAhJscd
zBhY`s=k$34Xbr(OHues;_0!T!hD)zj8x;?n_5~|eTC5ivF(LFXEkk{`1H`nnPLm9N
zDn)m0yT0Otboe8KOAn3=beJ3}+1TioaA!`rQ;udn{DG->brzM67#NS+zo9izj<pap
zLdDPD_g&X2s4Yz=ea)ndudW;A=Mq?86^s5&*9@3^{$H=dRymZ-OlP|m0o{5Vk2|;a
z-JOHc$mK88k75C_ffW@Y6J*Z`i)%F1{-miIHz0M4=DOC@7JTi(QXAlacrP|;inV=O
zSBFGf_l;u4I;PZp0D1SOhRyZma28l9BUW$S=?MJ8huf|EM#B_0*~@6ygr;0`Zb9y&
zdL~r`W$-JX)Iwe{hfl{vrgsC0WFUH8`Z_(p3~x8LhOk0%^+8X%Wc$F^va>!c$uM&u
zYk2P!Q~D>FXDrwSX0-|N@dE>hu4)N<X$X|`p|-aBI4LgN@>bTXgXXy6XV>1zpMZMP
zv$?Gw?LU6d-?#%D;@|d??Ccnf^y6lAW_=xW2vXIG3n)A!_<;uW!2-ppa8_oF_O?T*
zPTi+5IRl^bNdORXCzuKv2>|U^sT;u0qXi%n#~W)Krbrzf;-AR7K8l)MoSm<mTNJL{
zpUN{7dP=_s=Fc+d(lV}55z)7tngF{y3Coxzg6IVIxV6L{v%;2odp2S;d{%V(jdt%B
z4ItA?bGL8DcLOjW|1xTjaFd@a#cE(Qv6T1Ml*#n)_7P}3JGahdXsRy^jxQ&{z@*(m
zDk*%?a(a<}Uq?lY@<s9{)>~D6=@qeUUOgG4l9VoFWuO$l7mn~fTa=U)_Gw4jbSxxR
zPe*4;$&k<<s5DS<i{#!bb1r*`E|TI6BZ;#UrRVbi(Sw~cEKHHJ<tWfdX)MolsAhle
zaGs!ff8SSFYZz_q>3Q4K+yvd5KHj#-`kAUvR3k}Gd$hn4vqsSs6dYu3CI2k!t2~mU
zsduU5c+ggrB4PIFb1yRO1Pgq>^>(b550UB0cP}7Y&oE;F@=#^C=W|szekPTDA_G%r
zvI4aatyR~Ohc$`S^{0M%m|+f^ev)a2f}gq-G`x4uJKFd|lx8d&F=>DFEL3<H&`b?m
zc3?tilqdY8>1lS;6tHOyqxjNRUn~IK)7lO#=8cqhMQK(PEf4p4YzS0=)C&%9*NWET
zp!;u#M=Nq{e*5Va`mU3cYu}z8Z7eO87L_erCy5=3xqm}>kKnp6TjyM55J*7j?UO+(
zD;nD6{Wm<tyFol~V%o%?nU~44rG3{Q>fi@MkkQDG;Q(RYeeUmJH2vc203%#M@i9Ga
zA;taD{QV?tt|SXBu;X3P8=gBMYH4*=zttwauyrs~c=rdum=$~E_`<ZC?8fG6WuGk$
z8uOuld`ZS~mSTPJR96WsBfdn=wlyc<|J-N#6d7?=3p)^$7PYnEr851ucz3=Qlbj>_
z{~kqeBD1<|h%cfLAyxtSs^ZuXQ2cbKff#=}9;YbUXboe9e1|BA?bj;3D;LvQ5);Oc
z&-c3DUuyI2*6f#)%knyS05KiQtE5Ec$>#wz@qEZsg@FjHU&iut2j8?5*aK|x@<RCD
zP^h{4hB&*T0Sf}O4OQ>=ecSy^O@aEV^b<zTOLt?fR#R#HT?5-vF*9TN(N(h<_1QO^
zL)_+0+>j^VB?>t3AX-RXmb<&CO_<F6+V{JJhJ3`!N`{x>M(pG9$9su|b!DZu<##3a
z)0*iTEj0(U_J&?oQ001tIf_`%w=(B*`m565xo!{}L)8kvMbN+Rb`^Q!S3Y|8`-vj$
zyS6a5&b34K@dEtH-GlAzot=+Fui4nxh|pfjzs0l9Y^;+PSEkL+7p-8!#{TeHm7NGV
z6!-2v=teMcHC=(OhqH^87J<rWMV_b(*igJviZ(HgIc$7^%w9p!o&u5_Uf_-gT0i7Y
z4Yl`cg<|S2&eLo79$;UvSjvhy(*_~S(8h{Z=v-meVe4V0SE{bAwe=^WAZ^#N;`GwJ
zv7DUW0RbQ%U}a`y?7WAe<)gLrysidBFbxm)ZV-N#tC4QDE;}->t0QBGRs-w#Ogom?
zNw1{XJqK}`7Vr9H7S<gV)dmqnW1}pToNZX&nAcwn*&K;^SRVnh-OWv4#EnewaxCKC
zLLIl6R)Z-VY)X;tk_gdYnx%3AHma|{`==T_h1@`~D-6$EIDc^tM_(@PkZF^iH5=8y
z?=Va}gE&NBACIr#jRbNnLNr5$avPYZ*q!XfIZE8>tHphrCsv`HMCRK|j4gc=6VqrA
zZh7k&Xdq2!agGXiLP839N=*$ZNSMNgK#RA<b|13r)I2;}%dkP((EkLIl0sb5#2JV!
zC~Dt+ZXVX-ywk1fdlQ$H=`toed;HX^Yf{so?)Fq<xcwxw%d`o|89f%aU}0WBk5S?o
zi3sdjGHj~UKoZ64;vj@?6UZAN`NwJGi{4(=>)MwrG*m#B+Dgwb!n$HlZii79M=wA_
z=;--vPS1byto6;}D0L)pRln08h!FN-@Jr1-W}e`jKXRX@g&e2WTgLJ8u<sQXmmmZ+
zv)(MZ?`t$1%+{N24ube2=P>ujEjg>2lPgI>;GxFqHbchuUVIfQtu9IB0d3R4M#iO3
z!|$@GiY2e6c?DvTRVr@2whlT-DZ3c1QM=A+YyW3BH*4d=pcDF6J>P5{rw~Y8Blzb(
zcY`k<WQp9dqwHa)Dkobc^<Hn<U(fhGkU|!{oizv02U(NmRyiQ9pnaQvH+SmfeB$H`
z@i)eZQ^?GE|Gbz_%4&5Bov9{?Vr!`^;=FJN_2v00SN{;A|M`<$tEc!cET4h`9Mylp
zmcn`*+TgP{U{KuN$6Z`Q%dnBL>Ju}?0Uh%#b=DoQ^A!9#1d4fP%>Qo&GIMQ^x%uP`
zyKX7i83v9Qjf~iVd0dQrWaAV}tGvG2LZ^TvrDc`u$r~KVNwu@Ns~HHT-k5t8Paw?;
zc6AE9WXkw;xgcq7yN{><MZ)^pI~OeNK&4`%OnqvLbjP><V!a`jhg<H)`GmQ2cXu%|
zK;U?mM~j7tmvPAbgIVG$VDO1%EMepgQKYJBm7L#yjEFj6K<dk<i{;0oc(@3k$lVWt
z)(aj7W0IDYPTy>Lx-ZVun(4Z85!)V7+|0*1AyTMJe(Ue<s__4YwVS8cMx~=t;%Gz|
z;eIa2HTxjJAB?7<z(+7z?VhI90I%aKr~`vQ05=Xd57kHUaka3y=MYdO(lNLbRZ3dE
zH6F0T1<ZkTU*_yvTg%%F{C3tg4Lnwg3L(pR>-{Er9`j9Jo>*j}-djh;+L#QXoOp&D
z<E3gk166M$xKRAZ!xY2F<CbCHgvswdKHvLX0=@U!WfgOnRG4BPEcVN!Iy__ja4$-?
z(gvexoImh-e|W#p<oD1p=Cws^m-yR}6mleNsBG-qtgAOD&*~O7MWI2F&2?O$^lrn0
ze03Xy0&@jUtWcEDn~g5Q;a@A@g>-aU4W3M3SnHVOa|zPG#$l+Cynl0grI^`WW;k#Y
zpzwZjG)z7viiv>{hzdu#6g%>V-kn8|47V>M?yvI9mCrPu*74n6G~EBoR^_$%c^^cg
zFjGlIgwb|>G3R&RZ`hl@ThVOyn&58BBr0=Dg#K-carBR4$N2iudT<LZ$!sR0$V{}|
zN7Gcak(!HaPHT~%|4Zu(vXb1=5$a?QYmPtx>;K{C)++RG;6QRwJj4B4Wo{y(H`vHj
z-@CRU9X3@rGBN`MB^O)K2`MwMIVQA#{)K?KN8MU^#y(VO$knJ;I%cew(%tp2Z(>3S
zSq!!WUHBJNOo%^eY=iy7@N*KrStgtQ;>0;d#&`b;l<4E*QnYnc5KXiRlP!Zs2DGI(
zn%nN%vV`v(&K&*Nzl-_8-P%i{W)L2mM(%?=a|WL43H~%9H&?4Ss^D}h9&SJyl+J<o
ztZjd4j<t0S7c$*qb&8(qq-R%tgvK(<@A;dh#NK0Pt9e$-Z%1jE3O=1ZCqRO8Cz$zN
z_atG%d{dMTS%rOQ-R?LlQi&^Fi;x9Bc95VXg>QS+<>G0YAzHFW950~eAIl4%S^8Ij
z0A7_gaqSEBhwzy<xLBbL3%<7>n2Esm-Ak9{{S*>_?S7l#w&?ZX11=1K925%u6-(Q%
zS<3HlrDWhM+Cz+x>xwog>>ZUs+Qa#9xmUxd=Le^7+q=42jS`G>t!CC$(eb`dtXOgx
zk5DLd4r+5lIQwo%<YfKu+MdWsM)4|V*WcIMvu}I0N4=5BB=QxhMP;~6JuOxHDtc}H
zj_4@``lKT%T`0Buc<9UAy58_=Ax2r7AyAG4IGU>8MB;JFOVceq_H`NfG#`sBZb=lB
z8!HGz9Rb@+JAPgH&DPF#k-g>B)@qdc3;x8$i&@(j(BMIHU5$0UritwzrOXTrJn6|H
zuh9lA!nV}wTS`AXMK18U-WUXoj*IEP&zVuK^~?*oe?;Wc(HG-7V?ZKH|MCcEgUK8i
za|%?ycEN`AY($D_B^EDMbHQ?<nUq4teFkdE_@8s|rL*Zj@qU4}%#<n?y!!R?n<rE!
zP=R6s<N<-WosWG0YP;6`dVAl~-$O%~6~mj*=2mh=Yiny6nL<hw$>OIF1X)2nO~2am
zMR5|XM4pszVq)TJxs!dUjDo@tSk`L)WHulFEN~=#d2a7+a$y|MupmIHNd@0rcB90L
z4^$P<m+^1#f8NTd^b<V90x^pZ-!Az=Q1ykSZU=%;Y;Skk;OjUBaRV#0y5Dw%R+7%D
zt9zx{W82)0dS{N2W-Iv<#vT6?%+6-eu5+SAZwiSO20tT9V6Lz*GdT%Hsk1vX?d=yw
z|2h{$M5xr&<D>C0uD^V4`}L-wM^(4kFn--$-~FH712P&5Dc#t!eMu~lHK*pRJNV(I
zCsmCp`Neno?-Q5_W8HB>4Ne-^0jsO@H<X%E9sNAZq9b!y5OB`=NOoT4PyWoH-!3$i
z%gf@0;HNeQHVeqEzY4K|P+q}#b*1Ix+|!(Us!U>O24^{usl-)X2G~gpCLZkqUHD=;
zS18jrDCfs8P~a+u1z-yrJ21z^WhBI9tUCl+huwY^w~z_$8K0dQ{daz?Ba)ezNDx(6
zw*V7m7pD2k3E@v-NgG#xRr@{tUByqpd2=sA;n(8+5qb&;V^K`Dqwi+j$F3R4yh@8i
zHOeTmbi9i%s!Z0BeLsAJf^HN*0a<%_KK5BgCiBks%*A>Wir?1Ocx<j&AqYXRk&SuV
zn+A6gAd5E8>HSd)4x>k9;I8q}8Ff`%A-2rdI1+T6T=vC3D%I7dA(VMP0+yEe!1<5Q
zElf^MEl%y6L&=2zpCWsqwUp(XUk&ru*5FJXWOH>C(UZ6Zx9<+ie|kO~7(T{xkBxcY
zIq542z)-*Mkfi@aPH))m;)@;^OdN`-JMP5XnKl@(gauMCrclmpTGhA8fX2)&iRlKT
zv%0H`w#=kT;@4qMyx?R13V4V}80=zdY0>hs4ng|Z+}TliG*rE0s#LcS3r=GB_fwr#
zSz>i8u9K9qns(x*=`S=EwX4Ojg748PA|_rY&*G*@fY-(g@7m){uno-=R}KQtR%sTW
z_Iw%Z^9-au+{tN#YprbmK;9&~zF9L&x6AsTorp9QP6->}iOwjQs-+R%cx=DeeX>DA
z-J^f1_JaPeIEh5I+Fh`=x_06?^%u`-8rj%)j8g~@dtQ_OGi2hEt=y>lWmGr0_w*}2
zbAMmV^^l8MtA@h_0}2=-(+GJouyxpvCJE(#{dnI++H;zC%L2dpogcIQ&MGug00lA}
z{n$u83hv-qqcv`ht^Jt!*%;j&<-Lmt)5q>g&^UF*T}y35speWp)lgOwz#o(OQ#*zC
z<jE7}{s@cdB1Vw$cMlAdMn;-&9P&QLY!;ymm4E&mr{~LO$|5b)pQw*RoRy1QREqR&
zbqe1r<G-2#*gDYw3v$11_@M5^kqaT36<={1JV=~7_zCay^G=S7;|gYyvQ$n<YgpPA
zv32)pi?1*h6}oS$GFO6&BZ!v^dMM~|a)#U3D@}oGW|&z+y`g2&Pj8qOSJU3$W~&gH
z_#w&P{-=EOufHANFhHkMHdm92Y%SDcqs{THUt%`^)#BF$IOd%P{^~&UZJ%353|-+@
z0Ml(>O6;wj1BLyU@k-$l4U(ObU%x*2ugnRo7rUYaR6uBh>7b4K1Y((b|JSN;!MlI(
zUfHh%cKZ*N!uzKe-;<G%iMye4eN8>G+#MVV%hv;gLhPcpG^=q@qMqx+Y)Jl;R|1h~
z-DGzrb(@|)Chk0HxREG#yW{xBytKp&QuT>BjM@1oUR@2MY~kE}`IU`f(ggjRsC+*X
zWV_gQF}6p|lAbh_?~D?yzKjt)lrM>EE@tEwpn}hXuL2Q|rIA*)N?p-<I~Vg5^1b-R
zU;hfgh=}?tZFF=5<(1#Y#h4IFD~n`Kp>qN2QYY6PJN707mS1a>w~+)!%=l*Zt~`)=
z?{WF|tE>lNzpJ)~`^_xB`)(3@jCU%wwk%*=0yy3AYH~kU>#(@_9(t)Z?aWwYT3Oj<
zcrQvgj&?0IZLQcNh$}0Po*ft=@<8w*dHM9iktPw8$lX10eVpdqq}ITqt_ezj1WAq7
zIfBs2%6gKFX&<P*G|Yk<n$1%=3r@Md^IN9sXsG8ek;t9wedNtp=2uWw*N!cC@BuiZ
zPYpu+iKCwHN4UEj-)hH#3Kf9KH3tEIiU9sfEE*nAj|Y{|^Z2KqWR3sv!1R;u)DK&4
zF9AW5_lCw-a2IBUgobW*H0{AG@Ts`6T4(`8Tz!J6{AZ5mQk&uZ4GfM1cW!+fND$NL
zfojnB(L>c4kbV0-#FyV3Q9Nj=aWt<I67|=RC@5#@N751oq4e~`UBBNZHt$rFr3k5r
z&%|+72-q(HdD74%6xQWmz`(-#`aNeq=W)}Y&>YJe_1f@rrwq<eMD={DR5M}~&S9IX
z#Q1mzX33o?$EMZd;&(lGq*A~i5kx02B~l1G(O{z~Y`+nt=?9%8G}_MH<3)$g=LrgQ
z5X(8ecWZWoJCni|q{I<6#KG990%vQg<{f<`yvPB+4l)-7t}nVInfNW<j}nCH!sbbx
zNwMTdv-FQ=I|~<XW?ocC*CvmI5&NHFc%17S(q%+NX5|Dvh&ndvzTDZ>8ypPH8R32o
z5gwNLa8Y-S5wwj0SAPM)m&OT%Fr;i!kXikxu(q=cs$#_hiCE32l-A$Ce>pp!KK}7R
zqLuF>wx(7oy~P*_a&~d~?+Z7$5H2n*8Z7#sLe*1?jARjp(<-~UC2Fi9KSFn52;%39
zt<ER=VZn)Md9?Fnbh*Fqozpw&8_Jt+$9w8qdA{zGNP!u0Q_ExZvc7H<Z3pG%<pCh7
z*+i{3*!+>I_+R<bs2J5%6Z1xnUjQzl&>Z(n*ug6_MD#hFS)|W2p)}<(Iz6%Nk4n~I
z*dC!M-qCN-2K_ZElFX6=e*aFkK>F3uvHX2BdBU1ff}VEcwK*VC^o0B<Docs8Hx$M&
zLE@G6fbwMAevDkTBDN{p-$wAC&CE>vRdi^rQFPp%L;U)EHFJO7DzZIq97|O-v!-Ci
z*RRq6ONzJ$03|aSX;jFmVP)s}T;<zF!P__3;OXqB(E<$7I{tK|hwy!ssGj*q>t)Wd
z338N?9O=v)1d2Qy39jXlUtbzM8bRL$f~}$^P0SWnPG3B{sR|h@of6Qak_a1|U#lzE
z^{f8M$H&X7uw7iHX$YU>8cxrq9h6)UY0bzWgHZNoZ@Ug~2kc;7eg|8W!nupdtM2&f
zAp#0J%|R`7A|l$P6BlPj2+i6ZRT!P@%>9ncs#;oWpX?AJ(#patEIIUu(hQ*z_c!9J
z7eS-Sg<aJ?+eqBf2{Sg#ol`Kf2<>Y-6`Jf}?Ye<os~5<h3#>Gq@rrUS>b3S?@Irdu
z4OgmZ%Qif<Wh<L%e1vW22v3!X2@KjO$B)F5=kaB^CVu|)U`ZYLTDjGp!h#zG2?4_X
z!5iuP7Ct!03aQkb#+S}myk(HEEZ*1<TO{p2yS$V?FSZ7aq}K#o_98daxmfV49#fl-
zGk@YtIEm=dV)|8U@>t9^;s3V0X8C#eo_r&YWs+ml&my$LA#o&-bmua)T30zSTmpdP
z**D7gKtL~nJ8}!R8I<+o7$fY0v$W7MdcDPo5$p<eY0h^xktW0dvqXrf(-B!<g`0eO
z?Hx;_A&>#-s0s&PS}OzlW1eG)v!Yc{VgEN_ZwLs%mddA@e-uH>{+egLB<d#DpL)JK
zK9IQHpOn#J+!+)Ht1xJNl)Pl_fx5UrI!+SkViNK5N|=KVK#SvQOvrBmx@((c;Dipb
zS*-Q&Z13r1KkB4HfSi1oVbE4N6_4}vR=Jn$@f%JaG!98-`G_|l)0fOGA|G%12K!^W
z5<n82HhLF!Nu2DRG~Pw$1gRc@DZ2=lv!xN`uR(#xXQ&7ae>LoCv{+*8uu4<-VxrR%
z0XH5PmxmVAnOR4e72bS#^rervJsJv9Wa!g|X0vIrCX1yuk=yPPsubXq@n8Osp|~BC
zV`~FJd*5N2OfCahsdQJ@>s&|7q=*s3pu^TvO#?e1#f}MK4>um`<9Pnp7xEO+grqw3
zDV(7{{)IDXRGX&FKwA~gpfTMbvyj)r|6u_#({{Y^f34e5$$l~&wh9DMJ8^yQ1@EE_
z!5v5H&$zROeK`v7g`G5drxpyF`#O9#OpA`o$ly1ZL4-2*2xKVW244Vx7p)bQ01G>?
zL|WMi)b0PiCh$s1fA>9|lb((&OfHO#IjwE|101`6dcX2uYxE*A5^~)t)G^e4&vGO2
z^m}eDmk^f;6c#q}JGi3lnKna~tEmf+VqH_qUkhR$1UZ#zqQPG-%H9qxY_7IMBP8ZO
z?~CMwDQFXa5|M7dp$f^N@1qj%HPKM#IrT=U3k(+$YSxP68gZe-ULG_Ln;~UeM1`2T
zv2<3O=32MVZa+LVq2;Q*25l7ZoWQ1nlpxpF9^ga5mnjj6ELByLn_E;>1)dZM70dVT
zSSDhy{96gptUrQKpgCe2oYnM;Q#3{s(gH){zoM?f)c@t5V9F`D?Y-oXM)_q1;!Et^
zY0Hso!@<`faTnj@e!5xcQ8)3YYneZhg9xLvZw=a<S7n$hoogkCjb=%a5M3H5j%Xw>
zWY++oFLrch+h>~U)B9v(WC2UVX@f87pE@bUn<kpMXA&Y)Ir<1z^9eh~Qa)M^Ho!st
zFio10dL7Bc#soe>w|g~@L#T*|2zhz=U`#Smmo1}((O1E{J4}1^+J?*$VTj#(1l6Ru
zjV$yK02z{fZ}(2MmW#>4PqBiFiEhIHh9*5FkqY1-mQO=VzaAIJ^E|P?m->a97My>H
zi~j{gq(oZT#JoRlXi4R{^h$ItuM(xI8jFgmt8pO%Eg@1XwKdHB0((%nW*mgHNVmsP
zPx>p%N*MK22-#zw1Uz%9a^Hj!pVNGC0+VK$-gQ|RE24jN4$%4uJ@#9V!WJZt3)UYA
zu|`jlLs3uZS9R)4{51tYJdn!v!UnTS*Rg8=TrpM6&5@9o8B}XO07DEYfO(zLE%VUg
zcm&k6Lu|#Y0+UC}mD|OjOX#gt<x1T0Uy4&d6ns1VEG`~zdLjmQT(p_8cT>ktKwhr3
ztI45nN=-XY`Z>??{M;3Evo30)X?`ia1FRtb9Wh$|eE!n?Mf+WU`xzax0@<Lr>enYv
z*3HbgY8e9w1ia7ndn3c$dzWnwg<xf^8&~!}e!L%W6efAfXL-KI3o7VtzD#-s*U^oZ
zv&<<a#Mg5<aw?0|n5iFL$IS?j6z&@20N_0d(OX>UxITO|WPm@yPsPr(c5JRStOj0V
zLD_0UUdJ5lGf>OsjBjJmtwgNvq6d5*40%`de{OR65z7U9Np{MJ7=P{QISgUBfW&^_
z??C$QI%VMRqkXsgorCTDJwS`7jTz%<pxPB~YsjNgwB0IVMKQf~on?4G%L*%JxnSuV
zFf`hze0)%9`Vd&Vmq`Fi6OB0Uy$AC*0yN|`CHjxmIrO3Qk6|m1n_LOMlO11LHmBDy
zU6jc{8Na6M6L?lKKqo`(6jIWtjg6eS$;7y4*T9sMnbCGY4!M$%p-+81Du^-t-}8d`
zhHEeuf`83eOk7%t7;CTuSR_w=8$6<??P45{*P9z9(sZyE2g~qwdKg*W+S=kiQ2JvL
z^DDGv;w*v%Yzh66{Qoia7EoDj-S_Yz1O+6dBn1gULX-{xDe3O+?(Pst>6A|Ckd_vZ
zlJ4&A?)q=Mzju5e!{PPbF}%-n&fa_Nx#pT{&BQNJmCO3{Q|IFjkL^ML&WLrxO0reF
z`bl;p%O0omvctU0@se}dVu*s#gL~oYj)TJXi_vOK_>AoA>{d%eg;)xa7lNr9&k=lC
zd^TL_v!e7nOhoJ?LPH_#^ygnl07Xmk)Rt=1nhglXUr&op(wMK31TDNr5!(3|ywMY>
zZ!QiHy@#V@=glcUcd4}iN>*dbKGkJiNlo=lX$TNevb#S|ZQgI%DN(b!GJ=#rn~|cZ
zxoAVyNBy78E8k*hx1=w0in-@9_S?+ji$u7Z4SrTt=OR8zx>OO=Eb!K-RS<$jWJW=V
zR*A#&S5d7U-Lj8rHX)w=2_&E(t;v3ZUX?)pS4bwNpkTFBHI&`kx}Qi;#$avDf;)}2
z2hwI>*uul(8el}=SGw?KiH>Q@GM?6Em^7~jt<g0zf{J`m>;Fno@-V}UA9@YBN#MmB
zM6@<aytKUu$)_$=BLIW;O(OyA{}v4??O+ZW0?EFRo^wE4K^~R_(MMN!sK{iCHa;;$
z?P&&e%lElk1D2#dLq<q?WEx1tvq*iI9=77IDN~YB2BZr;xlDTd<7ArX-{n@<J85X5
zBJ3LCOUTjofW1_n?8$i6-t#V?tcZgSIlOn(nNh^=#8tK1aO_AD3v10<zMueXGw3&-
zL^&5M7B*6Mi#7u~`t(feGgmPrBu>t-$g%<|B<~^PQN`B*eB=Z4=uc@KIQU81S4^J6
ze;pn!53K$PUQEmHudW7S^6^0?_i&&{`FPevTkogwp`SB*v~AnrUQI@55L=2Qt!=<y
z^6&3+e2%XBB3^rJ@DQ-tprl~sTJXI8n0BW~@W$>@`;uCiX!%XLf&TdtJ*HmOHlT?;
zb8vpHgiw;AKeY;17z2<rbwf;4j5@elY4Vou_t>m}V=X9^^0xdmmvnS=Qp$}>$)yS=
zfj_-?lfes4(kh!w6n{Cbm@K~skqSbss8*Zgj)2dN&i_TxCk|(zflgJm;BG$+<$Pis
z)_cxy<oE!cuj1Lx6rH|@uJexWj<9I~n!xH5h09nrOno4S*`~8VXRWXxdty$NszdfG
zn$8cPosO9~a+c#0Iszi%>Dg{qO9wFUxUO4YH<TrL_eciQSg{1~eHb_?$gG<yDKO|^
z4;v}r7@yCv!V-i1LeA9*0Bt#xUgbAdGHf~-S@>Uig7vUP@*ldPJic@%Rp|Sf1NMtx
zvaK5zBrU6I`h9=?-PggR$JZLyogMk<eVTl8pa{`6`wKgFE}*gdzpB?a7o0Yi=;44y
z-PhFSAiYz-5TPds`t}6Y=P18GRnzjD3+5+u@QLyYHi~pK)YLYUd;J9l5IR!#7ZN6B
z<}jy_A5<Tv)+~qcrJsIDFOHV~MMl~j_%tg{2UK~#pXUCD0iF_t*Elskfy~B>sud=?
zF^<h{fzawH6vsh`TJO=16EL{U@~S5FZnFguX7RX(-i{*$=;|?N(z*B$_U_MPW~~+W
zgTJqs{7dvC$UYo{2T&ZDYvg`+0B)zShk6|H@+}2W@S?QHEx>^tSqAePzNUZo3A#Ce
z_w~WJ^(>DZH!q`Ci7Lfzj<$#4^k8`f3G6yQff93ev8%yKd_o5q-{xJz(2f4%>c^nC
z2MaxwXaj}`FzL<B4?l3V*ucdfU#hIL$_6>_?A5?%_iZ!#wdae#e0nM|=I;J#{(jLH
z>-RLX*`dh_+YtgLo_=DvtZiGcc>NS~j}EjRkqKY4s$+h{mkjif`P2q_Yb;w6f5~&N
z{If(?%un;|<j9cGiUrxT*%K8cELN|N$`2I?l`Jb@yazUb1|V(){u9Z<remmpS5Fkh
z6IC9hC?VX6KdDsR(;oQv_yRH^PdPhidLOyU4ERH8d{U<80Gz;k`+g3awL54WZM=xX
z!lf%zb)4SHJ-F@Bg0w|iUPKQuO#qfMvQ=r>Z+B4p*XnFii3V#))6%8YCV}Ni07<By
z&NG0ty8jGeDF8CHF~b^8qJ2K_4%AQI^V>dabqV|~#65j!sL<S;m(BFrFt?z9yt@GO
zr&+N`8SfPo@E&hby?xCr70&9%@5wd5Nr-BZa-@igCCD-+6hQPJwF**2@o0Y1g)+VA
z&)C0b#`i7Eyfzvys34@pn&!K``h%u=)3t(V#k5M>wZEV+6-e&}hyv0}#fFGAJi3(3
zrx5nm8g;GCf0;EXZ92HHiy-Vu0ukybgjAnKr3Q?5=C}unN3AZxn33Ue;?*fQ>pg#a
zH;t{3Cu-#sAZ=wu1?-lCToct-{bVLKOK6x$@&cQ?yVh)kM{%DV*Q4ZA$y9CM)>Z%4
zYr|)p^B$pX2u#C&#V)I%(O6pNda*Bc+hW~fc0Bt^cC60dAN-2fx?_Z1;yLcsF&8S9
zs8+ZF$LnFc{*Olc<6!`Yzw16xa+0z#0tri(ZD6WmQmPKSx<S$BO{uAJdDGj{Xz;e_
zx*+Ha%43lL$0B*VM(4MstZE*Op+)`XVFGt=jcz+qGrIQ6oQx(^s=vBr*Y4b)2^A0s
zn21B*su7mypaHc9Xp0$x(SLH+Hj+)-Mnpu>kP`ZS^-{#;KiriuC9m@M<1*xj3ky1U
zK}Q-;$V}B(X-V`okp+Z@Y`cxaz#D9xG>)Ym6jQ$X5oCnHm+$w!{%U8alh|XklZR3d
zvqy3T)g|XtQ3nemoWiD7i8AU^n;FqFNUzR)lYsK;xgUWuf^h1bf9rE<6}aJ%FS_r|
zX4-e8GtfE`Trfi%f?whjN-M<Xm<Y;<iuIhHxTL<4s)B>Od6OTw+*H-TD0#9bjTn78
zn|CDeB#c}t3xsedkE&;InIzH#XTm}fav-W`Rv4aDO{nvSU7Vl2;<yx)u;`m_9IIkx
zrms=MtJhyEFQ`?a;roWrL_VNZGv_+<BVHXAN@BH0)lAh$^m*f}t|uuPC$CQj?~UQ&
zp8Qf`$gR|w%;qB}lf`wqvVa9AxF|BDmr=2=Qwl(lJc_*f*Bo!uF6F+wg=6h&SjT7Z
zurRh8fmRc$n%^&<QJ7a%zl?TlepKv$Z`y}`N0W<`{`Re)92T#n<9LD*fG!C}QF3>1
za1WQAJf>!5VA{ssh}6R|AstReJFK4L|3Bv}nRrp1>k<C1C>O2CCwIzJ{U=jI@Xfdh
zyBRsWP>vs;Z7(YJw4}nkIRYfxCaIv1x}l(e4gZ*VVla6Gt4GNzyth>s@IncgD+zyL
z%JW1(@FM5x&6mfQ2nILhZQYjOl73~vf|%c%eHAz|P6VE4x8s7i9y^)e?oR@ElC#E0
zD+7BMCiL8=**vpsoiB8?NF2^~2XE0UOlZ4*lHDnbR0J-AQA8o9fo5kqY#&H&bf|;h
z(!O<MX5At)zQ(F0zjEgem?Iu~1(&JX5;ghSip7ja_=}q^+aur^UJ=jv-l^d2Rd+h<
ztQ4f|f<($hhatmdLAEzr{4_&Qhkn6l<)bNYYwmORQQ`T(7hx6lh;AUZ${|Swu?tm5
z#jI|ozl@l+(<5a`%*ldt#BfeU41Vq6h~U5Ef=R8Bi}oZ6s}`q5EP2xAHF3+q#Inb;
zajD@Q_~r$bDio+eQwsE{fMGx$cXxXh7Qb8`k^V6G<_hrvR+AdpSJ@|z71glXkSPl1
z(Lg7hhjs>vH%28~A1Vo;rV4Y=qWP;dONfWN>QaolmM*bCAL;kn#=Gu;fp0FN-*ytj
zDTjZ?eg*uGL%kaOC!ocvv+x_)B0RhC;7*GXSnc0AW7|E4$9#PZf5*prcthscW$Ft?
z*F(SiD!}%MSZ{JW2nAAHel!^<Uj0_8N0Bz|j^c4&QB|!U@3F3DAuNsvZT}MD(m(V$
z@O|$0=GDXV6S}MSoyIKf<`{)MVo|$%8Pb1j?!==wa+%G?B>n6b;7RMwlX0kewPXj^
z&&V~i<4ohVCG}yrgNKzCMw&{pvjyVFp}@_M$4*!}!v~v(yCF~?*d&F1Z5PlqU-en_
zUk5+#o)}%bD_Fu48Ms3eG6&&U7)Ux>8gvS0ItU_|(6X`NVhI9O>aT3_N3=`L;otre
z1*PY2)Sr+PS*)tEe3=)5aVvz5oNWx#z1RK-|A)Not&OGQ3~Y_!R@6(@rKtoYhnOft
z%NJk$p!$38s!EkM2v>yQ`|WK*WaN@Ii7`_`dBunBqo#WwZ4zKiJRbuDN)UOi6qYYA
z6E|gGs6MtWHxzk6oft<R`Tk`b7S>A?|J|Y@`MdXz9O}RGc|u58Osh|y5skvVHy!P4
z7=@(}$}ZW<xJoDXPNEwYhrQT@-^2tj$VD)kKf^#l;hK<DB(BR+Q3Pvey92pnhzUGa
z6jeFx8<1OZzY-h~i|6Q03Gny7%_5`i8Dbinouz&H<s4f>BPD1e|Lba)UT0;vVqq-#
zX;sZtrEgrBJ4a$pK|ujN0l37O62-HbKrg3)WbV1#>e{;GofPuF7y^Fd@Bx|>QZ*gA
zJ(N0VF?qN`&X!+Ih+QiGi|A1d&{9gRM-|sskJ&qw4)nMg_JqY0=?uMA(=~Ziz&khM
z>yFoEQK=G)Fct6RIY`iI{6{Fg2CWtUGzfCiSZ6Tx<dH&^7>FWH$br*q+qQo7YI)mV
z#jM}Z{`aQ9S^Spjj>Mrugm)Cx!a45*zQI6*k~B4hF8`NtRl?9y!pR8-5wx>(_kMmB
zk@K7!JqNpDL@F@PF1I;i!{z0-tjI4mD;8><q7$HjSd1p#R0)W;OV)D?<Hh3Mxc98?
ztg=OSXHy|b1U#HApH%{C91sRr7RuZ9^eVC>k7AHax+JVc-?F#Pa<{&p{)?}m(m*1v
zr_yy?TiLJu3X;!xv%Sb}_jgzf^wC6ean2IE-3pH&qz&X&sXfuK3ksAhTE&8#8&j>h
zQ`$(|(@^T-_EZTa4sR4FrX&WJ%mgtY=hS!fbQY$@=_ZM*<B{601&2P#jIs_6CQib7
zd%pkQrw;1+mv$m23~7`5GRdimS{j22UVOaV%v?SFyEom9WuPz4=WJItP!F$o?kyA{
zYBz1De3##J4^!+(_$*ljlc09{&6{hXESZ@Rd^$2lns+<{Bg1@r4)7uia(qb7IwEO#
z$|XxDFwx-^6kvWo#tBErdU7wzlc&e_f8-kYE`cuB{#;dmh_M2)h4o2<v-0aU9+M?g
z6jV4z>Rhx*81J|WggSid`?BY`Q2N4)!tihH-y%&^oExT4Ajg0<=CbF0hSup`&c@PO
zbzy(Dj1xiFPM{nAA!($hRu6>a8&}KAnH~NJ!8xB4M0B`|%TveI;MzbR=VN-2YYfhj
zam)Ac(oR_th6%CnxIh;LyviWnIR5FJcIrsJrD1J8hVQL@Iw+QLCO#}4e*nQt7l<=7
z3?s~l<i#dk*xst^m=_qJzvR-G3RId=Mn((_8PXxTb{LJ6T?JTzqy*ahq=GcbpJghi
zbg_M5#`Ry^dU_Z%{%K_X7li&{b+8Gdu0Vn54`4K|2gAL@J1FB)I*1y4U_P)7D#aHl
zP#4I06uI&*AtbgUP*(WBPX~*MUf&>|V$>YTv!yz*5nhW4V})dzs0sR2FEM5E3r203
zlj}*M!-uV$)@!{<r;>JP9(_gdFHz&N8piFfP@3G{-huh8Qbn6tPQ2-Aru{#F2+j$r
zR>(%_3=oqHi_yb)_a(7Ny&LG)>7L;keGcVfmq>MkD8s&Xg$$49jBu`bgy6qtn22Lc
zfQLuXZG0V}*zm=k9EnmSvx!f=HwQDwhL&HouT|M7$hguc!$dSL`@7MTPuBHnr#n~_
z>{v*Gy){acUotmOTt)gg|400QWx2HE@&cZ8AdWy#+shoMfcvr!j!jGFF|_&D%2gzA
z(9;j{uL|dtE8kPp2p$>xu*jeQ6iC%Dl`Q^=!_BX_KKG4op-Gg)8>US1&%17+IbZcD
zn3e*&^}*F{Td%&VCom60MksARX>o1Rs;kef?Pkixh~Egx<l=4#{qK~l!}Uh1c2z$A
zCK|krVMR6jn-PMz-~e}*y_!n(-p#KMlw5j~o+BVY0RBZa{H&?L?cg(<ZlV71vy2vQ
zbhK3y<PC3WRN~0`Xr7klD)1PAx`%?oza$vPRzU0x7W}8?Y4Y}lY;dF0^ZZ|=BDHS?
z3<+Nc;}Rs1IFuMgFpRa=+l2rb39(p)q?o8u1?%BqS|R!A-!%>WDWj8#>b@V@Jjd#`
zV`eq8(V`3r4kqjb`VNi0J`jK)?Fxm6m7z#BfE`39T9i3-{-Ax|eLY793aX@!1v88z
z(f{2hoA6)U^l;1VeV=GQg^(nkfW!o-|9$4UA-2Jpi4T*Hfx6p5<}_|Bkhm0Y3oowV
z$ay@QNE;Ao>|yU;zMhs7HxW_QQVm-{LBT9Fp1Y*ToE#k$e!K7tjM59N&*iKupO?Sb
zGyN7arOZF$^X<pc|6OBGQ_>Ip97*N!tfL|Yp2*tqgW8x-Cm9EXOB8HsZmO>mCN4Y+
zqsUoaX=`%)WG?d{>p)CRRZ&uQ<Hd@LCi2M<msRTId&L_a3XvjPj<W^fj*-Ap3ge>p
zKVH51-xZW-#b8(yK__LmPq2yLjZCzVxKoursf)1qnomYQtB^cc)trV<Vq`3Y>~yJl
zl}z!d{9JgSAmWl%Fm&Q`Nc(;s(+Q<5P3ne-r<TAu;gt?`;`pm%CL@M2=Vz|}UqrL#
zYXeJHUHX8Ix64+!5))GyGWglzg^+cQJ_B2fj=0p=oFpw_4_15({8#TF)-5=xYiW5l
z;f+@g^%~?JA<5t$p`p)F+9G^|Gex6>K@U*SkcM1Vc!TTd6QDGgS&vfa?OuJ0&6_mD
z186N+Uw+?c%b+JvI@PB`G*lFQ1B2<mZ+5?jZQ81-%f1#F?&=a|Kxa702{{dcH<Nr}
z>HUAWp0x6Cv3{SgLwf0pZiA0h_w)MOx$ClzJ)XB0vvp=_5>6LDvT9Vl%XVL%5(s3)
zq>RVf0_Q4*h8>$-3_Z50^569E;?_WG1L`mncG*S@I>pq^8IW^u#}_*c+F6~2F~OnK
zFIwSx)^@2EHH*fqvZTH1P+c*Y$=@zJcl7mncJ{Y1CzY23{hbv(4P!-9#-|s8i8-;F
z?Xm`;wd=+sd5({4z`wOi`E^L-Ov*U8^qnY4U}QkwoQjKm-Fcdee?D=4pB5E3>3-y&
zFFze#Y2JWEZ`5I-VH^R_$|c%Uwbvr!y_*U!c^|pnt9h~xo7+D8Z)OcrpC5*ji>jJ#
zDXpX}#lQ)#%%f1rhZlkX?(rl=FVlXxvxx*M*v5jJl+QcP&`>I`+(+hlKo{EoeI@J1
zy;R(%UhVah04mA5eweC1ILs+vfff(?B<pZvhT~ZLhH=orY~o$gt{1Fgao)kdes@dD
zNK5lx2^dgSkcjD9p`+`^enH_Jc^0!DI~3y9%p+O*wJ66Vaj2=evvcEob=GcS7J==e
z5AWT~kCxfLTz7i;R{C%RC>N;T(2!oQ>@w(lj;Ks*8An==QjDw6FyEKT9ogRKg1lk-
z8yxMB$NcEv>8KW8CJP>0D!zCMKb%JuDwGsUF_Je{^#hBdr<VbbVQPq(iD^_@Kh5tV
zQ?{UYulB$3G2k(oMM~~Z)#TM}_tPp=i>puCG-izh`?j@|UU9nK{*7I&cS2r6s?81f
z?CySBxS~nKX;?(oMpZyffygRm$vwK9;m@i}^=_2M(iylKvY@I5knJj8%juq-$W>Dm
zpW1mLuz!F_+gMt{5U;`#Bkv$P>q7ncI<hNql?aX`y1R>n^UFJQ<5AOSm_laN6(WtD
zS${rqoGsDHj-<9n@Vnpjo=A{&Pw=ru8n||gxHx<qtff>8?@h|@-_g;r+05kC^5tFg
zZB!F0kyU#*kEHV+^P8{AUckt>ud6F0)Z_uoCvbLlU0+{M=JL2M9m##TwVMdrn<2vK
zuu!lF-c#*9Q`#NU=Cf|-*6Z08?`9tOnqu<-849_Gpg?4KQ`~D&hy=;V=6A>It4xvW
z;NlRcrlpmpg)oX_DK5OyL<CBWxI$2mfupRg9nZ)B-s8J!jJ0+pFu(x{P5iOGVfuD=
zQbwP1c;$`wQzR7qp69MZr46UU+ge_2ULub?-o^5*S8dIG)7A*P-7hR=lVhyA;{~sW
zyAziaD##J0I8CY8Ha>Z-v`~@f)Ef7$y`f~`*HErg+8po#oBXM>tBe?0eX{HfVZrbs
ze?qQj`tHuE@^Cp6@p*^qNzHeLZf$+MD;e6zpKZ$_g(||TD#Y$erSZZDR#ZC~yW
z5iYK*nNEG9?$K)%_MVsiy&bVaRk4AB*WI6^lH{lr7Tq3-1%-qnbUfl1qb|RLos-x7
zQ}1)+o^UhsokjKFFWkGE@nQ(r;JDWR34c5}or|LyKO*lOj<F!Fv(@PPChqGs;M0x(
zA!uI?W*ju1SPEMo#)5G#X*>tXR6AQ`ov&DxJ(~3bP_3FzSRg`n*HIlEth}DCPo@TS
z8u%d-1Zuce?QERBbd<X<_%V4$Lhidhuq6DDr{d%12j86mB_%E~Dj}g;G0UVweONRu
z7X16`p1;*%=2!al%Ewh)f{}MLztl^rHsoTIUz3vU{?U~&G^8XTNXRQ3F%i^_yr1My
zG0c&DG%jJYKm>(rMxD;E*E5kg-@ck>*t|=Gx%AjkblMXBw+*fD2SKyn1C@7YnloSX
z;qBE4`_&cJZzC{R_xe;0P**L87SokeKW+4Yuq7GGgg^36fHYN6MNVVjuc~G(w5Cb=
z=j%#y!#rV!CVI8pmhW|51n}dT15sLoWWKa@E!PiOYV@b0rd5C1Mzmqr;+QeTiaRMW
z9XRNCE!M+yjDN_QTnT)T4r^N)WFo@BL4{zs-!U>qCwFz^@W{kB7}&mcsrc0l4;EtM
zEgfB7Pj`1;U!S;CPk+Bcz8Unpr|o3shuAh3D<j!S?-uKa-+KSk0?4si&F01RfPqq_
zrB-}=6Z?&>z#>h}ZYwxASPb_aEN{T%q-%R}vU>vTORh`2YV!B*-@oBWgBva3xqo@t
zO9EZ)R@C~Y;++oEo}A#)PBxk;HBNs9ku(C6{oz?{_AhuSGc)_z?{hd297G>*0$k>z
zqtbtS2h`hhyRD1!m&8U!+OSh^*&fGcXN!r5#1~@OLQkKe_ILM}oXj>$CM(x_-fhX{
z?qkhjAG+x1=)fU9PvkHVovF4cuC;fR>>uEHfQQ-wo||>4<>6vbkRZ-;Sp{f}Lf{og
zwS2H!?VL^CtsKP)3dmJ{E%O$O5cPAhq8hQcVMCia|IcoLySX>^9Etkly~(X6GxVyz
zH_GfZrV@VcQUlAr(7r`Fg0w)ez1iD)6JFgOt~`YF@PF~8+t?k&-!qG*w4ARSyQClv
zsMt)h%4xo<J`#g|n#e6Lk$q;ym+<tQ2mT0n-cmgGms#OFB}$VnE;SIz>*M=s?Iac}
z1n+&Rk0M|`E=i=YP$;3p`RtWkuH;yu%_CrAk3V@8B_`Sf{qEfG2`V!hNPXh{2JAeq
zW#06(hkOFB8Fx}UjWP~Ar*<X|3sF~BQP=Bmo_pnmlF0~CesiGj$sHVv@U>N_1xM-7
zWQ2l>q&6>GOqw^vcUSrQPsVN%!PxjD{$YJ27YF;Bm&J-SJ-t5-p+r9oN^@S-l@phT
zxdxFDrhvA{KtY`)i(y$g@$p=f#Ll$>W>wqF`+B|4!-9f>30(1i%oi>n<JpX{!^H#k
zik-{QUZ!PswyNE7o(!4O-V_|fNBmGVK}lgRAw7`?+EN<qCg|xTnje+|xVRreS?3wJ
zk6PQsf|6LH^-zCrq<OH|ts|a2Ptq}lq_;e(&IhKvsJwIQoF+<f&|f}<98i=0?JY5z
zd%JQQCAi*GmP^=A{B`)D6WC&bQqIfyZ$hOi#gvLj_o>9XKZ$mGotrY+EFuuH8aZ`j
z4R14hm-Hi@i3Ud|+Su42{Rk%Sd*V)<@S{*GwD;>1Z=mCcyoJd#*$98|wOAJEczozj
zN6k5WTn|lk^IbKoCcK3oV)2)ks6-sV2dFwUtmLHVIRQ(AW@c&p1TY5G)6)~a$`*;P
zwpidc?R^P3gBJYHq4)-ecTA9bv;EeESbPEVbZ<zA!_K(JQ9G_$&8;Vx>#f+*G9IwE
zv!W)em`L68nn$xKS^Y<Z92Cxd{98yzAJ>@}?ZM>{+sy`@T%;eK9X)}n3t=u?3A9IN
z2Dj{=Z<$7vOj;ExQoNz)UK^`-JSbbIy(^h)|AyQAaDP*#*YW7_yAwy3#9~SA0Ol`O
zb=x&-)yZecb{7^rNHiVK*B?B7n{aM0ogw}AdX2Q!Y2#e$_nRAkQl*&qBsn($82~UK
z60(WRna=OYFAI4xizQjiWj{G~48B^!$Dm$AhR2bV$(PsXey7qHhBY1GT5%AE{}=>f
zbHo$CV4$cny-vs7!Isn4<vG2ni~)qS4<_tMKgl%Rm;Aah-@tl@={O$Twy7h7u*Sl&
zo@CW@Z#k4nK%GC5BT=X{nZn8Q2Qpt!<RsF+^FM=lVNN%P_?uNNHHo20^QlbD(8n0v
zRhyhv!MJJaeKqgTmv1htlyy|6`!HY06kP11QRW|bSk$QTtoX(5*!rfg9^!r?$MvGJ
z8*|_(;uK7qG1Iho0xcagg&ZYg!pVn(u@CX>jFZ`|iO#||<j~y<ZY^}`#^!x;JbfeY
zoGt$%aU=f9$H%Kk=JvRD`~(Aig-2{YT^8%WHtbCnLG4d#mNqiOq?Cm9_P9CU+H`dv
z<FMb9hJGbPaaheaTp{D!-kuC(Bkyf(WoXe+Q$K@rr#rt}K6!-1^xaqNoYjR#f4d?<
z)uViH1~{qPTPwBwEtzaqDp8Hr`dLgfj_y~0WE-IxWf_l9L1P>BO%wv>o4o#YZvS-`
zJ}kEsZDfRAP{5nEL4DY7vKb8b>IuE3Dle0S)+fQlUK%-*Y!BNac`7VXY-&=tO=qsu
z->b%)!LM7A2wochxE6YZ!t6snaERS4>7U>v6a%9~$=cdr((@{_`^SNC^DitxxENKL
z1OQuB7t_HNU<&ZqdLP+BF7UMeeC<%lF&EhuK<!-;&ZX<o1Ktz5Eyo5Awp1e56<FC?
zHTl!d{uHm>*=G<6WKd8Od93yKXGYkJJ$BK4u)*4T>g`S`aNkj}47G2!M**XTUSPd6
z-cXjnu)czrp+)NBaxq^HRkP!l%+fLPqA#q$^Jd7Gt*I6-{k81anZ?sT3X#(|n_l&?
zNV}!#=_Ntbm)CH0<wzS~WuqWx>K&N)>|%AqIr9@^IS%XbP_>VFl}5q&uBo4XJfGB4
zsjZDwP;jVrT<^srYI6=0@HyL}2!{W1cXNS;#}VV+Cl2Gz;sh*3J<fJ+%RK5uvoau8
zZZ^OF7O(#btph>hIm5zCI$lSM#`B$e-HDP1HKYFi>JnvsE~4AZcjaWz;na%V?yz=b
zrtkyX-+@!l^e%ja1nP6;*Eo@TL|v)H9dGdUtBR(ITWyhgNnV)f3jj;=sXycBZN|j2
zL9mc}W1~R=@VJVa^T?D$P>L}yh`#LJ`u-?uTg8#RFVr9s#K$u!K~+)AY-spGCn75P
zx4RpU#hop@#{d}sFdolVXUN&&Sbrzpnv}y~J0>-?OlMA2K|<o;{w^!Dx*vF^;xpQk
zw=bP;&*i3C{5xF<>dRBVzzwBRUX;XIf;-npgC4BAtP2MVi?aRI|Lg5Fn2|*+zRFxp
zqf`>%8%D&DSU8fwj|d4%Hnqkjv)L?zk?!VWnFjg-1Uz2N_zHa~U%vDf6}5d1?-2M5
z20>5oDVcY09=yc;yM}~@7WDimjV3?Tv^`b}=Tu#r9~q&}&;EwW@v?Y3b3FCa9A6!C
zL(6&)jIfv(@CO6V2^uqwgkei>UenWuH56zmD{7X=76|A@rj?d{QWV3-`!o__!|9M|
zLiv$wj^st6VI>Dq{_mBeO4Sj#w<5X#DFb(_C=Mh)u7`08@af>M`F+E|#>K_KDObui
z0aU0QVVF=>$wd9nJv#6@Os7|E1%-tJZLaoaYW6&f%gb$T(O+Wemq==R^FVo~`(BuF
z@o5joc`~{_Kg}k-E-DtU+SbxCGE#gbf)LIRc)af230`9yx<qTb_01NJ%)9pZvpn9T
ztM>GfKwpg-usxZNX}_&s(uz4(&ok6mPsR`Hy}Ybzi;iymonQY7@e+<#Mu$}G1t!7G
zU>jN*w~Ot>{6`=WMQN)v#u%;IAp^_s@@V&IkJ4;|MUqWMxsZWce~Ee)31ZFa`BYI{
zW}b~k?(~oR^`MY<S<t1sk5KXY$&>YRNrY#diH4SI8+Nr^8bRZ2qs|;N<w}I8287Q7
z_}eS3q}I3B*Ve!HOhrzWzx@%VGkF)3Q1JKfak*?X#&$Eaf6f$*Et~w!=H_*O0Q1Yl
z$3&k!Rh{$=qHfuwI#HxD3Is;=n~MOHY&ekUaJ>pzxktVFQ^<(gGkg)li8yom!~Tu_
zNW|XQ*u8v>@#Lr83`;ql-xHAV4fku*k5T6C)<%IFGp#QH&dxj`%=B1HJ6a1@jo;0x
z0y(rl)ofr*aox`?V5+RekdLQNenDY|jj*8q+0W|txIx|nZrc26jQekZ9PW4G9vC^~
z5gP4!<|^kvzXZN&-qi771XD)@2dAm<=b`VhL|bhEXlQsWD3E$HStxiEQXK9b)n(G7
zJlQUo-+b$QKjdM(<7;YGTFdb-G&%g*CM}Sb!0i(KO|%aF2|4Plb}35|oFk8-Uqhn!
zc9B(1w|^gAQIS;%YTMwr7Bcil*GAbn$zOd)be#6_^Pe)a>~IG2WenQrBz$6Vd6cL`
z5VUk^h`Han3o8}<%Ii{3S&ov04y%57$6Ek^zCTX~LB)zjGrgbLHEtA46LNa#Nez@%
z$#G*LURxOvM?uN2QUvu(GbmFm97USBF>$eTO%8>m6`!9Qwaa3q4H#XdqX!9GT;!)y
zX}=6^b5KTjUK0<X#;X0e6P6&~OpW3r-jc3pw{e8dcz2A61Js;8-UNk)k|CBC(b_qb
z>wka(Lqp4M#^N6;8^rkd;=g_^|BF<sZpn`ve%&7B`+o296a~4e)grdWz$fLtWbXk?
zF>;6`PWbIx8UQOnf*tDG-rC9skyL!PF!LF}g%px#!vX_CLPGR09btVIeL@ZJ{8RF$
z)6;k0Fdw=7<euxW%+k_uU)>M<`LZzHzMRA0zoMIxIQ^Z9Rgk~Dx?w*KPI<L0j!LiI
zpk2}rKsb~x*BP+ud1hiIDJCf@Dk*w=a(;Ap_%p-LOrh#yS0nq)9zK4#V2nKe95<!t
zTLo%`(3B-vvr5%a1h7VB7PWLLijJzL!OFC8LRF#PcC0TfVO+NJEVPQ(t@j`HT9&Mf
zRJtGAA5BGz4yat+T-d278KX+oVIV{&&}Z<Jbd<N><#&JIz`(fMJ)*bSMCr}t_j`mW
zo$sc#x5sNeAn|ts7Yfk}?cz_q2%h!&?;%1q>p^`#07__B<B{bDG#kO!2WeCKZwOjE
zGQX8&5+{tvm4|+}`J9VT+fs#f>*-TO4ZjEytc&XCvwn8T)x`UI#WpK5)0{13ZuL{X
z^r3S@5$)7ntSc4LOaHc|2rzK>nZ(F&UiFKzz>h@HPT(CTHguOZo{tRhbjSXsQ>$)z
ze|FZ7#$ztMG%;3KVQUAoxFqR#W9zg50{L9^So&e-S`Vv*D3gj-;Ye(}QZk1_BJ9hn
zW14fMe`BIi7n3)(na##~(|FSEB4I9Xzr8DBw;An<48@6g2JDTk#>(4M>iuevk*Y0c
zu7|(9!;@FOopXW?2isCK82KZ!s3;Ld{JI@Y7+P~B;?tpW#3%HIMm}rVvE89^dnF*?
zSJU}Ebxq<5q<B%RXcqAkOy$b)$pq$Os>KNY7-Hn+t&(1Dv;NJ<kKoWiP?W6s!E<Xh
zzf_3^>+S}P?d)Rf`}y9KQq95S(Y2Zt3>2T)<)d-yw$gE);<X*m0ChfG7xrVm2_d+c
zTRzyuMgZ&hmGHMo0%gZ?IHiL`;H!_W_E`dY^9?R1lu{l<5Ez?Y96-EnK8URb+-jjg
zYL@Ev^mB(FsmPP37fM+dJPY$8=#2MjFf2y&B!>UOh|FJm8loJB=tl?K-ny$8B5)Q+
zl4JQ-qQ)8+My#f&7E>bX=q#C&>3b6=T<s**nm;~=vVy`{rWI}(Ph5Qb1aVC4Kmw;y
z^LqTrdRga6XXrwe*~R?bO;5qwS6kB#!)B!)hpWsaB>FPi0v;|aSSqS7fsO8NjmlJ8
z3TODx4XWG9dLI^Vcy1~f2Eka;yJXwQ^)V}Zxasb?y}f;k#vg&h6+rE*&(Tp)`X3T@
z=X;aBeudAMPmG8-T>kbBa7i0NkLhvk(_oz<k@%^plDYba+1hAAwS7}>O{<mI$gD44
zU@e;RY?OB6V`*S`Ut=}i_0O1JVTN;a7<ZxHVSeVim<3-LhWF1#(r$1HFe4AzpR+*i
zO&-|jOKR4JVaGYTQUH%^p>y0cF%Yv3`#XEHiLtKF*jczX)&<T<;tZ2F0Us4r3uE8n
zT4_8vpLrT2`}Fa-BPM_(`Xx3Tyo#q5;J?iyzuAmi{Zqr<-cwU^uZuGn)*s_r`#;*0
zX7OYc-$D}rhK!${$wWEl5j)7wp22v7Aq$+G=%0-ht#Ft{NJ&YnJ-qE~?Pi2NtI5e_
zH6vqst7ZW|l-X(`&JZ(O>g}CU+=h#NH<mwo6ivbcFuRl07Mp_^p-$LtD!;6k{>IW6
zpp``EY5$F(rco@iRu%@EDT2mMvR^W-tqxz&v7NH<`fc}GS@IaRUgy-=)?7ea1`-m|
z_6%P*9e?l`6C~hrc)?ZELoXA`&bDtWBjYYT*x|DSP~^ypV>JKNY%<_)c(v0#5kebq
z&p$^B3)rxU`hB4z#ryaHJ^WT{c?l06-_!uVe{UbVTvP6E2TLK7Vufo(o`|h`e;AwR
z?d7z?&i(z3eM?HcJ2wwg(OJ6BfOx_O_USqn3l`5-c%((p-S2U^)F0}afe_WL|E}>=
z7W||zD-|ifH~5Q%_!Vr=r?uQ0Jkp`W>`}1V8{@SkFljZ^?~kfPq6ufQ%`S<tEClF+
z$DTmwuLj#SWr~RocRo$E_ze4;n-0>A(%kSG#@5MHf6-`^%$JR9)SXJ9>2f3G-vGCy
zv9hwOYT|vZd286N00_6IcJ6UY0@IWz6+oz8GeXHVtN)>2)B{)yjpT+Ct&|M|xg(+1
z(6y+j==MUhmx@XPbL`F8&cx)=qcnBD^$M*Ub;tCV2_Oo%oU3MFFGX&5he`Sd!|%Ao
zGkgc$-P-G{&Ks6H6g{6ptP$ghoK+XGDdW>DmVe90$T(G*;yv&!N_6ZZrwKoqDo!{C
z$*+ZtlV*WCiN<`C&OW8kW&L$%McHcf2#VDc5H%9S4Lw@>2L{k^L7yS%SLI9iwL1JJ
zM8UZ72tsw*A~(t}%=EKOd4n(zmA&=Sd3&d-Q-xSXNzqlv6U;*wzQ_ndnpja^R<STZ
zR&W-SSkyaKhQ$R6n#=m%HJ{-+)V~ip8#eDu#M9TC!W-8OB1cO51F%d&0*6CE?N<%H
z$Dc20X|JG|*jS*}#`<@QMl|JI<;^mIJiPXv7>lKHP>3)lw-?5{gl>87g=_@*?^PSj
zal(!-Ww&>?e?`S&MSrv;8H|_$&V)$HiRPVDzI^$N;!lyrYt731`LXrrJ?5akj;I@6
zq*$aU5{%UzaDX%JD)(EUoozI7GBL^jdt3P;n7{i2-QI(?6Cm$6T~G6DYSYpdg0X2d
z-H!x)ZD|5l9`oP6ugD*T_NN<fe&63slxS28&RDxEwFTugvhVqAqy@OSNly;KBAN~?
zGn^;<`MY)~LCk6);YR!22TWK4?+8I8s9hJ5ej7`!^9JWztSCi%&m*60)>La)q+;ME
zp*!+xl^gO_e%<mFs3xa0OjW1Bgtdq$)%wLEzqvKS2c^Tmh^&rex3{<S8g=@%!bEs>
z|Eg83S5i@;3Vk;R;tLfIygD)%kTk@X5z)zvsch#u`?$G(OpDq^kPTFc31e_utUD3K
zE4gy^@{qUYhBD{w-dZ8&^SnMcsQ<Hw)knG^3lfwuQ(xj)*0>zJP$3DK7}<Ps3PO5T
z4vxgmmDt$$!kD*_%Jv(QKEkxGv-WyTQnL!PBQY#5I`)VeBpAUXKmq*N*tl?Tpn%gq
z5W>ETI8Z%;!Iu|AO;@@(R-q)KLJjvD=J4bUJ4m1(wE!n!-&Fm6=<};3LS^uhyYTx5
z1l-xJ=C>VosJ+8%Y!=*4B+{BZT+|8V<hyuD7I4$Cpv(rcy)nZ^St}4}E9i?)K%_-^
zp~{1?Ph?oi5a9trts3CK1AymvVOqtCb>8{VkZC9j*~d^?y@@6dx2U0`lQRW3V_+B9
z)6?TXp^A2r)t$IkZV4@FKI80ZrX~<d!e`1GH%2~~H+Y03gA|qbMjd}Ux)Ec0)>Q7h
z0PZ|Fn#&YyH~4||dsHO4Nvpc|hwBg}rX$4g7;rv63I%JaH+T+rrz!}b>8}APADtx0
z(b+hYDSsrE5D_S3(pM=ZG4X5F+2xxRmW-R@6YLY=yuz}t&7<O`dWz{-5+bfWY79&$
zNp?RtCc4#Pvk+yCn9zwon6LNL)nX(X(%$Bn3pe0DY=e^*J}_ZYg>Zwcw!qfU22=!a
z=AW+v40?W*!~o=x(EH~L%We)PINs<5(I;sxz?_WQ8&HC}t^K@h7ol_tN|C<Ff4S<v
z3NW6MHKWOxe(Lr8R=Qh7X;EXdSO+SA;*+)9j<8wpR|JCi@?USgEEC280h`qAiwKTW
z%aZiEr`(xnLpk=OR--blp_Iu)HBkV6nZqbP9W+W)I8!tp4`y|9lo6RIiB0+>Jg2Z$
z1lp+DqU71x>Y*Ws!F9E(%50pClMsFpiJcQQ01m>ya7=2L(r0pCnB0fa&_5o54#(T9
zrYW+;oam{mAV`!>z0Ph6RFw9HX>BoZWZSNTqQn!#^Q9qg$z1Fp!SDJeAA0X{iliFw
zL%v4~3F#XezKs=CQDmVeO?&T~k*FIakT0#AHr77yg70bBYzbK?XIo~@SF1{Q^0c<I
z^YiV|e7Vy3#!jj4lXVI|<zV@rGUJ_3Nmi0h3sk~*2OrxSAuij3Q%}cEP0dc-JuuVR
z*4E~aBRX2um>E*eRR0z06`@o)2O-k3?@BSdL2ru<Km$Z}*2x;Mypn%VLta1QY@O-(
zh7t>84>`2S-_+A;7h;$`ysq;IS*tBXXgd;qHf8pm?!5{-Ep0=82SDiwtAB}1qV2RL
zB)<#hbq@~KMI9eEOQzks&o@-d3_sARt?vS`v)^~cUXMX0!~o?=ra3y2(7&N#NC#ys
zB2N3oqhG6Hqt{V8rV|x3ic>}(>u1KCj1jn8cp0thuO(B?>}XEtaj~zSErR>IBxQXg
zv%`?|3<1<~q`dg+%8!EqbODJ*j<^si{zvXratBJyZfmZ)v*#x#CotYQCS?MdUL1IR
zxmsY9*;7*eq<1_I;=fl))f-jm<2gBT_$hN+{_2-!XOnj#`bw{;G?`CT)2VfN?Regt
ztpFT9Ru0B1m|WkX@o?Q-EcBzIldg6E$~1}Fqb@vmB92b2M5}p!S2+<Mt>il)d>|~i
z5vSxbRgo3<sXuMU=NA`;Crie~Y8Y)7Tg+7Jmc4|?iBs{b5>aAh-(p?E6`5<F1a&!V
zqIH7e01U6PGMie!U5E;!BO268S7FR}n++0-D5Eqt&x?H^<=05#?&C(zn3Mr`Q};2b
zT}-w_l8ol160;pnEwy!fWeJ7e+s-#so`(3`-9#&~nolmy0w&(F?~0$`3oIPq4?dVr
z9>TyOfEM0-qf*WOu)M9OFUE7WIpF_H74}{4Q~S6{|H6po3`kU81S|X;dV1MVer&4|
zq`17)&0yacN6x6-R-N6|4sxY3<s$jFXFS+(PPxp~AIVY%a3^<sebSh|N@Jd^DQUs{
z_Rd8#zxC*e^(y@cd_R0XeNsO;VO;M`-TOJ2Eo%ZE;V7>z>vOnd3Wq~`FgEHpcv3{n
zXYghd#c0ULBb${^oDMiR2|+Gz=h+{&XSzF8Q~L1RhiGM)_JW0QYRuZn<?r-X;RbTQ
z)6;G3Qp24j@3yZg^$hduqxxU?n+FFbCRQ{Jl2{Y-CaSah*Or&qQ|^uU<6;3Ec>?=M
zUmsqrrFFlOGM`zsSg}y1aL;yQZZp~+oli0GVoYflL3RRg&YmYZYuBjz!oYdW035xJ
zHNX#X0dW8MmB(nQ7Ek%E29Vp!$_@jt|2{q-(I4_^24kI;OXYI^zI}_L`IfF!xpq0e
zEdX_{!KFf#-fD)WK&$zFBv;A`lh|^$A58VRc*ViAGa|*4#>vXacoL6IgG$U65g-3X
zDL<$2r~{Nf#_>erXwA-N*?~<da6djjKMo81hR1BXY8IVq%PyZ37&!Cf$wAZ6dW~58
z%d|ZCvD<3bF@Q=-G@9ubk@?zGFZRbaLRmErCd;xu^+kK#jae=7{;HYEQ_oPoUkO*M
zu>|FR=|bj@uII|NrV7r+d9_bt3lhO-l+@XX@?)ejWh{)Lj@FOwH$auW3n#45^UC1%
z@<^#fU8B~v9aJc6ZEbPsuh&y*7Ie(EjSbtb9l)@^M&~0w8=0W+(t-3(zLb*R4@c#s
z@L0vFWj&6MIv+S}`0@*`t~S=xJQt!OBQpx}BSn7zs_MA@pkZfgOUA9!QzPXU_4|)>
z;E<v{x44@P+~B!7mD74<W%vH!sPmOt&7*zpci9XSRLd**i+w$z@coI>v`<x_k)4Rm
z{ql@juD3WsSzUejZ6z~g9~fF2<?ctPM5W+%u@C&C{~Td<R#ezsa(VW|(bWO|I!`9c
zABEWbu2E2O_4W-;b>Sx%2E|jBM5+FgykD#PbFk2$sRItnX(fp`AxZ=S6kh56r@w_X
z{_{x4^H&6SDf|+YY$3=~e**gO(06Jm9Owx%4HCd^%aEsD9fQ&p3jy;iBO_zZ?4L39
zFd_1L<H(tdTT6Ql(bt`_&*O4#?Q!Fs>E0p1R-aN57(S*uN+nFS-YMdsveJJ8N;~3~
znm!RTCPSG|pwT?DQ)H!L&9i&AgF+_!>eUV}YSw$PtA&UA8sjC_gNBom+Hn>OQFn`D
z`OuPiG4}=z&)HgJ+%6jxgZ3%Aje%6ohZ!rg^W7=4rJ_c+3&~`LvZ8{h+dTP+Y@@%!
zj0QAb8h7`c2lVQfp|zM`&^m|ipNcD|{dXRoyMKaE%gR9i=D~bpeYc!YxVy`MkMC}{
z?T_KVZn`bY?caD(n0xe{PG)QE$UcXShX4FkY0s~_?EmLa<h#bo>xqTlvlk=7!@&4G
zDmJMnE<5{&$I)6m0juRN`}bRxOLL^Yh(`JbWu=B2>vUI3H*{*AJZBdR9n?LXPo9va
zda!K$y}MPF^z`gs?*r9Ov|XOX3%X`EVe(zEjAB)o-#LH&fKub}FLg|-ZSlSGplC{Q
zedH<*4PNi&xF#<G)5^qIwHjL|kY?(A9>`6duJZ_JE)DNy{iYru##PB+7#7OjRJbJ2
z<E1LuL<K<@X;%Em?7?cQi@PZmDXRA6YWiJ7lU!NSZ{bVF{-*`tK{QHUG|nyx@GKc*
zUP7nypgiyfOAo3NL1j?7{M)H(roM<DXt79su{C9Ye^E3{1L;t=yn0og5D9wX?P2}>
zo9Coow++UEwUUYWh^j5Y99y%A^70QfJ3aaz?2wm8SA}8<5(|||VZFnZ_Ffi((Uc}5
zxoK=J9|jj5e&ix@-A}HR*$|NvM^(l=0K3sLqlf#1S03({hmoMpdU<n0P+9HJ69+8o
za8n6}YMc%xf3amqE33=zE^X<s`zI$SD|Lo)dMzyn>hRxyZBZ-Mp!}Mej{0ur&yilP
z<j1Kyk!UXlWaMx2-#H8}eBg954%ar)UUAzC+lv=T=Jxb-v)i6-X6SXs-5%a0c=eau
zSV+jvQ2E*or0^`{N*#dv#{F_sF6Gt!^46M)j7*t+_i`g5+Q*aLMGek}&6CBaA8e&S
zZ{Edrmr~dBqvLbg!m)%clj{rpn{y)FvZqDob5Rsd2eU2?r+J<pH$qk46scpmo0_0>
zX)4V$4u_uVJ*?k>I&F0ne>&gTpCLcZD=4@=3Fiq*<8E}>8K+Syy{|JJBMav?9n7H>
z8_zNtAf8KeJ=m+M|I_)6A}X^rQ^RXty~2~<Q1N{7eO7-e=R&De8rr;DDSvK&fUe9E
z@XFnvt#xecP<!$O4&k)f`92eb=$Yy8p18iN+Y?j2*F_FpjBu%*OYQN%EDs!Z&`ox~
zGN6GNP8ha>e$lo}F~EI6nEx<)GEk~oQzA>B#_87or4;vN^BRxS9w_A6*py17w&KWQ
z8FekTWV94>*l(i~5NzzN0D}=bns*PUTXIJ4)Y+~s7Ut?c);XN4(;pp;FQ~1;hg1}+
z)aPhEuybm50$1fG$GhvQoeAGBs-VO>?2nT}<jm=Kmp*ystEjGG()ru=t?k|<z9|Z`
zY2O(Ez!UX~S>sF%)HKLR4dN=}VB38#Q0Yg_9#Xl#ORPs*r9Gd;uc9O)Yx|X>VVTuZ
z_0peo;xzy~{hpV5$0`<M1qy|V#rX4KYR2G?*4BllwX=&nMxA~URfz*>imbM)T^Kt7
z{l&^<ty4#@_32WjyD*U|>}*l~Dljv*`J=rlZQ^$MK<&O7`PrE-BOt&?QZv;?j@5E@
z5!sI`<FJo4cPk!^fF-1|La#II+W!K)zLT%5pvsx#9{Z;a6UnIQCe`x_`CSwE-(K-8
zfzo~Gz}vU<N2jN&&-?~Zh<kDF9hTuug5O}3Xf{qC#$@Uq%kpFi3!_B8kzm3jb(5+#
zpO)W9GI+UO@vcd`jREhr7mb^w^uFu_ZAHm!H1Uwd!kJIUT)oKC#ZeNKeW$0i=>s?T
zeLY%9!tgD38Y>W?UM_yBK5UM`lHFWe^8iAC#h7on@1mUOjt;5lO0|31THC;*@q%(R
z{re_qcRpAts@X=jnj<mFEqf-HibvC~ZdOjqTW6XizR<oaz2_a_1P8n5>hjXsywchb
zbm&a!->k~Njp5ipMFhD9oK-8yXQ72j(7)Y!ffd1qUc+mvah4)uTnpB4PNvacojICR
zH7*y5KJo{m9n#YAv@fRCiaFvr@>GBk0;&~XPQ?w2q7%w!GaEL-#GGzyY*B2y5kv^q
zhqrhL^hk221&kzE$Z2St%`LYh_%Ryw4yI)RT4EebQ+?ENXu57uL`UmC@6!JO$g2n<
zA`a)Df<1Ao%J~tw**V2(_2-J8V5sq!4Ij%tHe^=|UNG9g$D2~g@!Pj=Vu1&?K(qu_
z5b|VlW@e+>taalrDasz(I^UZ?goarx#1+4@>ve3}te_yRA+-Z$6eeA&R1>(~Gjr20
z{PyaK$e3}uw^_b?w!)nR@Hc#8Z9-KpJVNDD6k1yP`-^RbJKOWr)2+Eq{ja+pMe6Lg
zengVdHK@RWHiE$J#fcIXxW`Ws;^J(*RDelibsgQbil>zvh*?|TAh=DW@~)Xrv~*Xz
z$W4WZNW*#h<yEtkF7~xxAYM&LRJC6oH>kD@Hp71w-O>mZ0nd`QzwodaNoHE?_%Qzw
zKv~MQl}@Qk*CjrDlJ_>g_tWl>v<6+X=be9A{KaAme~2DKP+JBuur?v$bV~hr6CRsO
zVq~(6me!|E@POXsk9t49cRMrvqh&WZc{TF|BuxLdyQC}V@ckVcWcy@%fXckuTzGRd
z;iuh2wq}Z)H^{F=;=nSV!x>uRKvdXBkoJi&h>8W|7nID(zXTobWhb*qgFca#_5+NU
zFK_M#gqKq)DqsL%?Azr25TKjjdmF1+XKw<yCH;$M$Ya^E`DAeXc}qEsch%jtsM1;<
z?4@qu#E|v%D79?Y%1JkqkRBH?7$kl$`+mPBLTG!<cG!TZ4dTsCNY-f1Ma0JiIs|f+
z!oM?}o$AS^q4*F}r6grfh{=NH!w<_jV&*I81u?NAKxQ&3`C47`(YG@-bPLhPh*Cvz
z7yBPWkpbvd(NuT>74z}+YLpZeRaPIL5g@SFSIfvotBJtZm_C6phJ~XCG+IbWSJUh2
z66v_SN?0V$PrY*`{EvFmKY@@7;|)Mg1$({CN{2_IH7Qi2Qqe(TRAVCd`g!knrOw2L
zE5UY(q!+B@l$LcaC(Zc99g7M->V!(97ca_7P||b~%gYOFp<gnp`GZsx?qwBipHK}J
zzG{q!xiEEH+e>K6E2ZpU15H;HZI($Qk&%%#j>rrK3V$K<bjT@_t)jYmvR}Q%#yh`C
z+DDf2Nsx~Qb2z#fIq{2U;R72dPhONgbxQsjnwvY(i^tey<081WJyw`ZfCki}&OlJ4
z1$e!3!6I==y@1iHt1FQo!=QQq^s!EzKd7wU9RR!I)7?2Y0KqL<TZLU59rqXtu|X78
zR_Npb6QA3QzxbYswYG_yr@5{kV${m&NWV;r;Ej@~IhduSFrV)727o6(n0(@9856R#
zE$?mt<zdG}5%aRYqM|6Gu&{6v<Ddlv)IkGj&r{t$7=^7Nn}mFji1Q1mVxiZ6=d_;Y
z)qgA3xmZ3ttm_*^!e}Ky{sP8Z!2DVzsR0SX4^5=&?HM3L+|$z)PaUJ@NF<zX@<^f!
zi4uL~b{c7iOFw`4ZVlj@6plHqVozY@12O2_I<Mr67irv_pKe{drM{%0qiah}p=s;<
zp0g`|L}7cl_d_u4<Go7TqtH9r`cnqnGu+n3HvD-KTVx3mY=hqYBon3tiML>=dS4G1
zTK<7rf#;q1eo@1$x+6O1U+@!b_j5u9(XTLK{6$vPS8!^&TgqyA?j*FkH_)L!FGT7$
zH$g24$k#R<?z3eqM3Qz%MhrisOzl%4{oL_KH|6s5^t>Zh4<IJ3iHnRIr<@g#l+aSC
zt(y4>d8-z)P+^FOXMfd9`fzpLi1({Vq51ZM=dCl8S68!EC```eL>eTZ8&Ltq4Os~%
z6QXurkL=h;8Sh+lB0%Ut;<F{k$Fe(W(uL2X)v-F4Z58A{3W>=pC`Q8FX;7cKxQ@pu
zakKrlZoPcAL9Y()%?w-tYHLD7goHZ$va=%=weM(&%l;ozUl~?K)UG|!A|Xl&NJ<Mx
zcXx+$3ew#j0!m4zbazNehte(G-BQxM`Ihf{&UYQ%fAHF}XZFlm^E~%+2jJ7vR<<id
z24fP8R~mpng4t41%0$rhO0`Mh2uQz6nbrc%eSABD{oc;UgWjb|lfRDJtZFq=Qfy)A
zObIwp$3A8&phcBVPJI%Q{t!W~L{qGyafc$0hAAPBCeI2k>$*MLt2!8nJMw2^`4xFA
z)v;5DfevQA<W+r3abnx+g%%Q}Q!3cq8)p<>x3zZd7>UMtY9;btpdteQEo^ZI556y^
z&b12pDo?~|0pewfr1ak2-ri6M;+9oYjf;!RowD@sZmUnykg+UL&s}?lP#d{z&m9w^
zQrGll@;x2)))o$!%;eCNE8vYcxkQvmP~uie3t<H3SQND6<bV$f6`w+jh6iNwyA=k&
zohku3k9~SoBALX{(r&U%1Z&*wdbQG?NaB5ede5bWO7;AbcN&t}z{YmR#g1uybbII<
zzm+;gx5W3W&HEL<-c)Vhk+xaF(?7}Y+AH*zW|L~LL+kMt;UG|4a&QAfW~u$dhmMU)
zdJ4%_!0wJvXm)t`a@_V1$=6+`SC0YlxEvTRjs%nVpg`CiiNmRqUzv4(F}-_NwRVii
z?SDqXki=l1@FaOKEiTwX;<$G^u7;kDvC*p@f@-zXZ7e>sxY!s8)G*ZK2d|wba{R^e
zB=Pq3vNd^)QEepvs2M)awJ^i{0yg{pHd-lg()Oi!N<MxGJtRm@Hn(>-vBrLWesvP1
zV?{L|Ss0UgcTT{XOUlT=K!Zn9wcm-XuaEK{Z3ZJ)`H4D>wG&j-qiAEs;7;bR66Ijz
zk|dK0SO)$%{vX0nkKv$BW<JK5$7TH>YeQg>Q-2CzwSTxX8w42Gfb9ZL-A^8rl6KGQ
zHsq2AIQ!O5Q@AW**{8b5!HSCx<`q1DfAyMbrb?vG^TUU_@Z7EKEr2zp9aulB{8C$z
z{z4~<vt7l&!sGK#QeFrJz1(MDaBP7!JLdaUUTy7bidF0Bri)esVcQCLMl)ee^Vx_;
z6^NAmP4^~S!Fes4<MY$i7SmA8^Iy+K?Ut+a%Ss#{KRt*dVIujvSk1r*?W=!b%>p%N
zwbL172#V{Tr@T<#FLb}fptDSUk$}Kq7gA>)J!y+?5`~BmDrtj>tbz|xdQ<b3;M>gq
z8KrX6DYq+kss9SnW9>5KN!jF^lh~@JRUq_D_{hzD>MGa9^kaCsToj~2$VU=<p&A_B
zbJmX0q{K!6^u_3m!~*rwexljxFd2>y$e1HFDCn3>jC2(xW$*h+!Mi8Eb-#EjI`MI!
zTqd79X7cDgeT<lNK3#s;;1>8bEWt?#{-yH(1f+B#wg}XBRx%{b4E95wo{vt}tfWSL
z9eg*?r_rX<(zP;ESY;$4uQ&R-weKg)=g!pUwivzwKlAx$J`Hi_$8Y^@#Qe^`8@@}9
z>Tg>41G@{Bh_qd^W9jkF&mNG_%!g#3K+d7Z>twit2?MNVBkCH1{KpYA&Mmf4{0cvk
z|0P{y2~)Yk&G}_rYMg1-+afVVkZ{_bCBB8e1J^*%w}ZicZB`Kq^rFf6T~YtG<$`2A
zGDZCn_Py7GXW%lDW4xyvmmMYdoaxrq7C=StVBA;tAy{Y~3wCWWHYbBYRHpN(6R}`s
z9v6Gx;={*heZ!ld(16J8&bF!C{AF9=JgHL>{bLJ1O|{(r$`$~S{KfiO&2-?SMKLXS
zePbc;FhHWu_|Drzw>h<EVPSm-_Kd-;F}Al6ma?XWD$v;27$9%N#9&pI9q4$NwjU1l
zk1#SZP3+Y|(HR-92?(&31dyO3VAZ4iajSVg-4X}2PJ$l$sS8iJCV6PQOnwZOvWf*L
zqlZ{1{{J-rEa!`T&J;Dtd6w;E&V{eVk7xtzwYJ7aM!gN1c=yP}jp5hVjKoIs@Eu^!
z)@yd-T@B-7bo;yd6C!1etGPQ`4luukyMI7JCqjYTj&NZO)>^p|pg0N#Akd%BD5!?_
z4GvJ&n8nUex$#W6#09w3;{w^+%V`_DW>%**H|c{TiIAWtQJn65N`V@8V4lmvg%5#F
zUPoQI5vWa6V1NsQ#Lj;tÚ%O1H_)9Ccd#2M2PVPte}s#rED*l(rfCVW)j&ZZN|
ztan{W_|p@B%<s--!$<X&&p2k3$;cndU8jXvdr52y=yDjZ7xcBE^Q8aqz|a~fFI&+?
z(ok@izS8KY>V|}hm7=fmRZ7lRlKe+0yc(<D<oQ&IiC|C~Vd)qoCcGUSiZ(HgwRtl%
zRy5Ud#PX9_CO-_pz8$F(e^DnwLht>1O5E^I9{@w@wAaAT)Ggf(_&21W{lI&e*_${}
zB_EJdRD58{KNH0aMfe3s(6&o0H>cg<w48SOZSD^H+rx3x3fNDBusdSyaA7{Do3VCy
z;^JWJAfBHOSZPkK2YnGfcQ!VW*5^fI3qR~QD?qiZUansUwn8txlRyGiXmzuA;_mWz
zhRF<C_T5R}^m}CYn7*%UTX*1sn9n*f{a}048>p5Mu7Ddpn8ak~9<-TpFxNb~iT=s!
zit^@k>kHthnGR<hZtNjKf*@=xt<lLvLPv)gy1OE|wuE^%>yD%W4oxO8V8~&!mtaw3
zye(~L9}`osJvw9Jll<jrHLpWuwiT7h`~)&LRKEwbD`G#$zdcIPI+!ca(t(&Kw$bzM
z@{kvBP+_tL!<p?B-}2QLQx^DK{BHZH0a0SMmeygiXlx|1+XlPXG6|_c*_(BsvfL@q
zfiJ~W*OEk0BP4q*<xnEc7$dS-_dPZVcCEuI54~d5-E{4xTBSk$fB0sUT2$at<bXN7
zNHLAHJm3ZaV$H|zgz^*C^O0|Zy-0jrZ!PWZnWd*_69#yghSyHbwUN=_vAfqCEx+RG
zk6G5$+ilIjgp{>d2)fsRbPXa6gw586vNjX0_hpO15o%c22l*;uI{y?jL1_$86ujB&
zx!7NDJ#ar>x=QqIpPXCc$^=jevx8Aq=CMJ&9pO;a{$%k)2o?qI+gsPU*zU)>o5Ce{
zWH4>I1c;c%{rv<oiBDQeB8S)XLFKmJR#{(+&i+dKN)<e^-}Q+f7&pCUHy!%=jdi}x
zW@lqde4>f2F=zE?3B+$BC5u06ubROj@wz`bOyy{O9Fqi)jjpb>b?R@W*O`?{bhf(f
zKex^V?>09w)yo*@-r}Gj-kQ09T&=~qnrC>7eD2qO7mnN8?~@GXB_td@_|nx&mFSLf
zaSC8bCGe)VsdgNEB94m7=_dkSqAh&<jk=KvrBm(Hl9<Gn*L4sn4fEgi&(91-?(X}l
zHdYw4&pR;%osSmhR%Vh0H|Veja@N=5A5|uK4uSjbp(FRR=}Mi)F=W=OgB8C$M@P;5
zX~1}WgN+(4E6b?gAO`>PCGo@k-!-T9WV6#n5Tt4bGyz)N+?hU?+~{Z!gKoobcU;ya
z(T6sR3F;cN$E1{bJ20=waQ`(kBgv=%I(N~06?uHtg4z}%K#`5Omxv*5P(|kC;4ock
zx=606(T~@_Uj%k+fq8*x#NjXPx#IUf%B84OckFRaoH+iUiCycZp~DHbSt}giez3Bz
zy!WJKjnE!mvyVp=*S18xv99r32ii1rj<;i(ZQ7M@?6`@-rr+r@Mh$L#T-1rtI97{f
zVBLU?qKU?SF{;u86c-}$R8Hq&%C?2ZKl9-k=LZwWeuXGne-A>ibO5k>e`h)b;y=pf
zYT_6TU`}^H_NndR>FLemc2pHdZ8n@yXME?15$tI^U%SuwsB^RbXSm(2CylsTD{8sb
z^ZZCHSKV&*I}9}I(n*Xsi>>t7s1)--)XsaSu;yb7LV?9ii5w0>*rF3A|Dwrs+8niF
zhTIKPckduQ2czlbffj)A2DX%*;rc#<p1@#8W-h+hYx1d_6Kpju`-wR8>O_JkNiupA
zFHl4~wZY+93|PelqY=*oE)FBs<?hlxfY7a`E66zWii$LIr$V~dC7P&X#`{*LE8u}D
z;_7&@@$6uZ6+jQA#l=t2ZHVz^K5X`B04156ZrWNhd-Kb1aiCasAn3Rx<^@Qy=l!fh
zz6`-)^%i|)3!mwEGJsyRxMoIYt|k-(zTMyMQ+}9fH10jw=uIu-cHjN|Q#6>0iVEO@
zm*9hpsv3Oc&WFCS7AW}q!?RC3`!1BxHaK2>11{`0ph&>R#^wTKFN(YFMUI+{C`7?~
z5G6LP49yB2?4V?FT!!_wUPicl4tA=ONKH%;!R?N+CD}N4k4Cq~H=F6&_51q7e{f5g
zNTzfw+W2&C`^;ktC!tMXd9qdY^z3X-k9ItQ#r9HI`QTnFMhYMkT}JN+eoCIYZDOYT
zGGEB2FWe`6=VkG&aeV0SV+37*<Ha8}h}o$J4KuWBZjChQ`HL6ha<JOz!P|>{D*0~%
z`7{YQ@M)<n99URxAP*FXNjfD;f9!kNc7^0^esr}O`4<{{VKE`*?W?AiK#?O>6r<OA
zYd2#X1=lehK`7bsHvnJ(DBrdf8sXCwbo7<i4rmsiI@zsgrewdJsm)k0VbmoP-d=a2
zj9~`?-yQuq%VuN29;48p94;Uk+gHzi{W`;%{Ojn>gWJE#+s)R&$jHc~T8@P#3e5&c
zV;^zXPi`o;W|Xd8xO#sg&2ypY{GEW40Jz_5@)K-t6QvNmUlZ)oB}y?{*uoEeCYwYc
z67a@`ynXB!R3ZO=eBj>({!^4GAEe}e0pb+Irmc&I!;ZDGLg=y}@MA1Efh@>+TOo?^
zrvBJaBUiu33)rL6+*xxNouLVD{^zi*VI9pEmHLA24^i*DI&kUY=}zxOC!9Scc(jx;
zAZMO(Dase&;s;^wK|U{BJp&KeEZM(N(<e$DRhIQKvs=0xo_cKehL5pzuHjcA0H%q&
zu#wUF^x*=HKXr!p_*_*30+V9p=|I@f`T2Da!n3I`8JQNo)>)E%6kBh6;iET9kBPxz
z2X<_CE@L_3gq%R{_glSu{p$}VA(TdpdXKMia&q@qD}KRjFwbatX?G#3z5vEgpDt7I
z|I1Mk>zVuTtD<d@(<<=Quks`oHjL%X-ehc<E^=3RRE`glMej>$YilK>L%(S+3sb(h
zLq=F00Ej`pO22{?izJv0j}<LQN^(kobkWJ8=}J!CW8w0#&7Bn<@9W?Bd4p&W;@_Uh
z8DfIe!7#-TLD%)J7z#;!?=2iB7eL8QV$}b$)QqPSn~)9<2@4BjgghtnYIJzGO`TYF
zzsc7C9oYW<zK)(=LTpN#@g0Yj0tR41S<RF=xPC<;lTFJM`~G{%AK8o0UoXPblVdx7
zS|CPG$ED6Vl}q$F*?!|i8!~PHbz=*VMZm$s&Az3Agyw&gpKxFW3FB5OtEs8SCnSCs
z{W}S+0*~8^P*}4eg+^Z~Ek^I5FR<1@fZ__|%FhU~PPcj%CJL`eez%XoJ0b(nk_Vi}
zrQl;4v&Hy^BHstyHL&8rKb!g!@S39OT#pKx(fYeyBVsyodeqzc)jNkeGzLr6f=(6?
z2r!~h$lha)_z6@Mls<vbI$lQce*s5-!)mf$CcA&oc1c8a@Vq)!hsng8(LHt(&Oaa^
z0I*c@mmRkO)y1#6q-Ih-^Mh1L<ky#M-&{Y|H&2;Q+Fm^a^uL^ub|>Dh+Q;)aMFY<p
z6G^cRjGjn>X3vX1K=^PxSEwlBodkP8mZG5G=i}d+nyM}<qmaa))>A<84+#nR$G%k^
zpr?x>5yXI6y)K<icTuKP0U=!AX0>z4%MEY&Edx`fq5sLniO+7ff+D83=fGz1_74RK
zaGxc)i&br(REWymnaHNjY@Y+P(Crwcz^j+m*ktFoj4w1?5{Zm#P0JFbpa}>HwqJ&P
zh{rq<H2?6EOOl*-0ZvHMolD$7>b}Q9(iJ6}jU}L9$@H1aM->n8{2IpLoo9>Zm|T{x
z)XT5C7iGz_X*vW+D^v@o%L9lBVs&Qg-5y2LNd}r-==U61`hV|HJW{rla5`~N^WL?(
zt-$*y84&C>&|q0IS4nXl4|<%nrR^LDJs06{^kte5(ST2JvopR?9Nb~j6p9r0ZbN??
zbQ2uk56xtV_BAfvdMuN=yG9d34Zos!o{5piD3LPX!~)BaMgLz*<@&7yT1mcnvPouN
z&>W_xuwDt;w$zB)%YweBRHF=^-BeE6qw%ElNn0DM>Cp9RAm(iB2v7_x-jE?9{o^wO
zzp$^*1RdC&4!D{HTJP++-U|zVLXTKQ;SEyi+>81#2x3{5ct=;-TWTkZ9S9neB!Gw<
zOvMxTwLxAYB_#p`^kGYXA-znmR;@;n6EJ@miFo4VT0uog*`+7Y(FeSYZtjjJ@IZm%
zvJd>P>k13a_NU2!;B!}1(WUp;@gK7dj`Q`y*d`((5&|L;1V#1MHDdiU7kJfYcs1K6
zwFhFSAfeJR9!_R>a0AjMfWO&FeRb<vgXO(~<iAA8zi}IXsj1M{SKjM7@j=P{0w{W~
z`2po)taI+uKN@U7<NC|+l7fQj9F^ktA>A+RNbF*icK>ml`WyxZ2GwK|q!>=mE?i_F
zw&8e<_mf5MMPfn<Tu?oL?oLEsPgms?{%YmaUZ4GsKXirVq&=u;sFripZ$w$`?R-nc
z7#v-d`m*^bUzi2^M?a%2Eow=~)X>w~+dEw10V2HKg*mK$bk9Cl#`u4Dk^T?sxx^K}
z(UIXPrYIClKa<Tqz~x0oMB<_60J;j4{lX7GItvJgx{Qh%IT!YQwl!3o?-AYY8+pDw
z@ekmwZ*6S_;e@(o1iXHa5UT$#0y&gyQ@fo=ViKRzr<Xa(tw1JV<|7QwPF_XYf_83i
zL?7MpU`h$zW~C}uZ{f!?yte}*!zR%A<KW2fgX!8YSDOX>$E%c@PC%^lj59GQ%rEGx
zt}ZW@`yu&zul7YS<pG+JF*pRs5I>71W*7FI{gY61`Wx0+`svWF7f(m*fQ$y3Z8PrO
zt*X@C6$kas#N4~4wDw365-xkgQgyNyp}|6wELWK|oz}W+^1QqV;5=3Eq4bdqdX&bl
ziIEXOx5!ugua0cWj&`<|1b(z86>6vf4{N_-6SV}0RHJy?hud!dhmAr6HrZxHS{=!=
zR7BOj(sCjj>WHu3eRVThjdB5$IZ;JeGCd8{Jxzi^^5yEBu2Rzg!hfq%i;eR6Kbk@J
z?-Vu_cn<@8*@Ro?|H}o4`4`9RdTnopyiYfh#?9*8<uvdrVas6youE5CCZ@`5Gy@cI
z)Nk}EfE0?9l8cL)xA}?uugqrGH&j&mE%s}+AF{<5jzG&Ye(YzCLuBm~2&%{iGq54f
zVa<&WN2Wb?sH&;)fA}SKdc3vzDG;ER>Cw?y24u9f7O((BFEB=Hw)0a{t4m5MR9zQC
zX{>yQ+^fe(7lw&C&t;$9$tZJbYmNK$Ybx-6NGabXii;SJD$l82SN9c|GZ!fzGtDa;
zV?)Q@;>rC3W!X5h1+N>H0Ev!fFa{Amza<(Zm%$8+n{yJ;sh{>^f1ED&X@)IUSzU$3
zQ3J*H!%pNOJQ9YpN**I6)8llO@MzH_JDQNal9Hx-vXuGGi8HI2@88a-#~R~6Q0Y~+
zoSs&_q@InHY4q;RY{nqss<pfrNYeN2kDLJyfK8sX$IY7i!Y#-ASh!?fvylU>ny>Vf
zl$7^?bfj^HH+1K?3~UU}*MdoS1f+tZ?*__;Kr_2Z?LR0kjzMTs*gjJNYiK<G>MUR&
z#9?UD@CA_UjGLVb{fW{3#h^O<8-D`GSv|K(widlmiRQt_Cyf|U1XIb1&GB>2r-)UB
zSAC|SlF!X{(c5Tdc)Y>e&=A$B&$!*LEw)qex)1GT(o@KI9G*$8f%g{=$EBNMKU}C=
zs4|uWLcX4cpv&bh8i6FdCB{DzGM1v+s#djunFP*cK)3AO+CruHZN5s>T1{;V^YPBW
zQ(`;0s#2ya7MQinlG%b4&e#J8lkmuw@IoTqcyUj8rw_C;E40q`+aB(i&4QnoB-tN-
zzs@3M85w!^l$T{EzH{yS*#~J;`Et;9Ye#6?$XAq)@Z4Jb{30e>LH;7#`HePEHt$BX
z!$ULDvrjq>?^~*4*8%z43_0Mpk4B|IyXS)^JR#WpkO~f?ZB$^g*^opl%eZ5ma^=?+
zXYQN!RNBbUfo?GEjfQi-Y20~5`vDnpsH(DSz1e0BM7r2t5#;2ox9Mc8H6Ig+{T_~n
zDd_PBo0outgQws)>l0c>*VC6Zj@2iXSYW++0s>`CH7FR~9U3$0w`}@HKNqR!h%{dR
zS2g!t9Q-SD^Lp0!m}6A`hbpCH_N!vVI}LqsQ)_!?pB_8xXt6B;v=kLyg;`m74Go`$
z=hIb+hlYl_Y5_qV3*24M?>|3jt9qn950uDAc5R-X+ASgZFZ<kaEWZKjdZWsMP1Nog
z13&plPFuN2gVqYkV}1DBOMdSIO)a%8w{Sc^4ztKsM&OItC(b2kW@e%R4~dauxw1dE
zq&jGWOQ+z>QnIQ&SX2oN(r*57_(ykiB*i=JCyNJqRAX7oJy=f=Bp$OL)?H^XGl8aB
z#Ar=s{xkbkHNIP5P|%lz1nju(3jik?U+xO^v0#RN-2O^{Zjt@p9_BI(nO$*NKLbZi
zP5krc#8MnQ95}$GtUyJFNT2_$lPQm?tgL~r8X)0yEvnOo+z-YiU;CV61#0~B!s+ad
zd?K{|1w9$4CbA2HHx8hVg~6B5p94!UtI`Y9Lk~)U55(C`BUWgQ33??z(esp@o--@j
z-5qG$2XDB5Y})xQ8A;*mo6?yl{xZeujvf$%1`l~Oa*Jd?hs?ixOBFaiOby>D7|kUZ
zT-awP#|-77qmyC4h4*hAk4Iro3l<xXe)RTfNq@kIUWk*)hrU}~*$z&@7V8NAnVLj~
z^l#N!5h6mTfBx`TH35Y<W^~}sK>0GwnPxYVCsD`{(2|23MvMqFKxN->f*D7>r;~5<
zbeTC?vTHf{vXxR@W_@5uI!VOki^!Aujqq=N`*9FDp!uWb%aB(X6Y{1mP*?4ZcD&i2
zJ3)kt&_;)3@>!TTL0S$r&2e7;mcFi3+#v9RDt}Y%ay32m<4Gs8+rHpOMF!kwh;UFy
zSlBNtAehWM-und4WYFq)-xWG2Eq(IzDb+qD71e8AhX6SDSDaZ8<d0}D?B_9AB>1>^
zGy77Ji#M?;oyRAP2yIue<5~@9ktYgAe|82xS)Hvi2Gm~{O~cCl;l~G)t%2JNqf>@B
zoUhLe>Q^Q=3WDKeD6u6ZBo^N?<UzNW`!l6>D|4CND*F0I`}@3hFmJ}B{R4a~k&D%P
zYLXwNG0DkXp6BLHQ!ie=oM2@UfbCL9_H66^pCZ-N8bd@rG*Ba}M&3ee?og+GsGMge
zI!OCJYAy3?*bAzd_vuoUpnf@!AOJp0fEG;tw0WKxAav$r=2(YoF;q#CUhcKeYt5y8
zVL&iO?vwVGz_Hh_&*q2gSZwenb4&(tg%6_s^BnX67p2E-#kVlug~a=O?8hN58V4{d
zsl0Yw`}IaP36x-irDxAhO>-V66f?PebL=t&-Z>EWM3EdH*fS{8sC~YgPK*eRVAm=H
zb7)Glq?m6O*~QV(&o@LG7cXrp{q)*1_<yjN75ats*T|_eGqTYL<1*aCyfBp5CKVE9
z+nmf~EdQ)`uK*HplVsEpC{4S9<$`q6&PaS`D|))5<7^ijxB+ZCe8EAJm7=Ecf2(7s
zK;ga<T|Tgreg6Cz%8J&|VWk?8`S$JG$0?2n8?(c47Gc~$cql;z5dz<ob~)~mywBFj
z&4FtbQCP~1ZAI&J4gLH0YSGJp&u@p-C}f|+8Vg{yqXU^kjW!~(-`!!3o7oW8!_^8|
zc|fzX2cVDOV}=Ink}1l|ODnXAu%f{WJz;>k#=tHShv&edLZn3|;b4*r1NnWfUS&E$
zhNQv9M*XB=Ilg+W^J`C*HI$V{UWW&X#JNyj{Rl5F>H6Fr^VAw#BQS`9Q?^OE=*Sjz
z+VL*PbvbX<XvEGp#)m@CFiALl?;HI_#?Xl83Wl!$<y8x0<EF5K$@{M$Oo98~B2m##
zI?_l*TsQC2h!>_H|Kmgk{c7jX#^s(4U!PR)oq@@WzB|kV+MQ^|jI0sun<`?TpYOsI
z?`L5>V4;eX2d<?3TYq0EbnE~f(Iwq2p~-izC5ig*k%8EyN>OFpQ>%+{k#z$#xQ~4q
zpDq6#u%+cBt|jm-OX09dk;%Vxo@OBKq|{STXi6Ju6U&(7h79~#S0gxAoYjLWOht{z
z7pfRbOZc2OLq3Cgb8c=7I0!^MlgDTi9S;67z4utue05=?ZhN!?B%NFIz7Glf8SHLH
zWsOc3qdAZqqr3O&?`%cW;R1+~94l$CJaNzh!te)`p`nsLn@UPTg7}29w6ukqi;Gz5
zU+Fk1pTBFtSxKi}S4&rIK>i98DJ$u)%dOJ>9uFZ7LGy}$&NMDp@4adgB6}CPP!3Tr
zM){<#5s38t_rkSdkDQtKE;bmUfEUzpC+KDNdPOW9c49^q$0mI|E<)QH8z0aSp!*sH
zd!U>q2evq+h9Ru?GfrJe`F%W2_5~8Qo|Y!l-%S;vk$`==upC|f*ya4+cQWJO`iG9A
zn25jV%2(#~p84LKvfqrHJ&dqg5VoE5I~gWF4R98CFC$T`q6{$%HXWb*LFM&I+<5a%
zE?=d)yEl&9R)JcCuv!cxw@_jXB>V)TX%Gk=Au;UGqEIO4wC%h<?bbP~|MCtQ!<OTt
zV(!$Rbr-5QP*0RB(E4Kuw}=KdB8HkyvEyS0)?(7q(vpUZw*G~4GeLyU5BSD^H(i+c
z_HNJswobpoL!A?D%<B!~BO>x{Y%H&<8naMpZ`?DcW=2_<Cl%=M31-2{`e~T}lI__q
zT)?Ymk+1O^t>B6RI%RC8vo&XOT6{d)hhO`Zt?GIjJCmebq|9s#o8HpJ>bW<Y8&TBq
z>EuaGkl|qpJEf!j$YP^7BBBQ(BAYn1JU$cNFZWcN=`yWeKLOu`109ecvefsnvY?#b
ze)qWpMcu>uXVTeT>tnrCaxw~vigz=}j|2U~3AS@5+L1Z03Hn*L+FEVX??!7i)GK}Y
zAJ}WUW8>L;?;?r6ijb9aTz(-uJv((qYNZzWKrx6x$ic?RX;2u>>v4*yp~w~ztdw(n
z3(Y&6Zju1H?s6M&%;7j7UGe9Ba6}V+PC|}CF0Y!$mnKmf5gK}-D;am?7mdUuZGnxL
zvrA_`YR;lHWD1cYmCI(Ba0}3%5pf@7Ceu6?#$RVCG21Zc&7iVr+kE^8kkHtpqB!TY
zU1gGA{C`Xr(vfPo-}1NOyi`?HxlgTLh~{{KgEM99_X$z>{;>WfaJa9ls{?=(=(Rzt
zScfjUWF!&ydZPmnq7pJC87M_ZC!;{hF)=qD9&gi+3f1FQ&klx$t{b={1G$}ce4AY8
zp_cDP-KQX0b8Mae<~U!}<7x>#iVqf4*(#(h>ZmeaMdG{ao}tX#5>Y=P6OT&8!7-o3
z`MWj%#T!hAhf{Xc`4k<a&E!&l0PtP8;OR)Fpr5qNz{+vcew5&&bBY!N9ACA&LyJCj
z(rk#F>t<h!56A%T$P0Z5;C0H-<e)v(0xEfY#Z7MM7Fk)vxS<!$O^DC%P&ht?WLK7X
z&J|}!wJd%AAUbw)Hoh9a`7x{@ib!C*&&ko)0*lw-?{|iA`&IuZ1^3=Yl~LgUhOi!s
z$|{{W^0HVFd?t-X$Z-w?7(bVg0!Z}`PshZ9YI1UN%7&a?oA$-#IG*nERR?q%Rav2D
ztFVX>#8YXfv?*|MPQLV!_f=LDV>HSgv+M`!>rL!sQe2$qpG6OY?-P3$`T6;cRB`vN
zJb{M6q$uNRSr(Syv?%nPqr@ra89O~*1`21|uz_vU+__|KYdfXyry4x6=M<}059n^F
z6fzF1n4(=*4dV(%xfjr!IF^%BB?AzyyxT_=#NhqQ%7%a?5I3Na*Jt^EWiY4`iCxYU
z67g{8I6r5L_3_^Qy$1Hj(G+&871#smv)2gzpfc3!pT%py6Z^#VcWJpb_e)fAbQ&5&
zKp^=h5XjvMlLS+<CSsDn&#Eda#JPV>N7AG#Xm&b&iU!o+4}ATaD^??G_X0UI6AUq<
z$<dus5qO)8^IV*pddU`TNKw|=OG--1aZ9yxG0>k{EYGul75VAFAU;9VJMe}>qbOdk
zeb?O4LL$3wXkeWiMB9cH*aZBhW;qqHpcS8>io=4$go&69AuZMXk2$yEoZ973)~RZ`
zpkGD~O?ad5b1|rIq<`BZ<3sZI@83iI+V}3G2GGpyI^Qnpz8pP^P~UGHlFQieexZX5
z{UpnV!L-A>joziwh8fLCC@za4><uyKy8B7lKZ~7Y?!yp%u7O7A$v$@AE8xq=%{_Ni
zMjAq$Vu^}Q%rm(#LI$1fjNF*-a~ODia>k_;SH!?NKsmiPXm(x&t^3{G-Luu8u#ys=
zmE$=5KO5^bDf8`*r`2YPnF=xrU_Tig8K_RB=a9Aoy+teKsC04N-kwEdKsq6)cqB14
z*H=?CX2wyqUum$Q4H|eBD6ym~IQ(&3<qaW~!u-5-y`Dq<IiO1IFU)5DQfOk*ze`Nx
zovlKzsHjMmH1&HrHKM_R8ms+3e}IoMU4%L|V#dxEoVw<2J}}$CN+}`y%aJP6AML`G
z*lNEfVs9T}{DoFtwd;#A&3hRtU_r=BA8w-c{mC`?4-60)92t5f>F5$_UNl3f^j)-x
zhwP^?O(z-3IOC%G##g2`7%J3u=X#|&m^hct@4QRu!4Vl1AuT2Kx5SL&=d|jX<sefU
z3BPa7c{NkS=wygS6>dUE&TMd-?ggv{3>;17CN^h}@W$2Oi%1CQH~*L(r%7MB{xu-j
z{!l?GXkc5JMT13Jp(tjpdk~Z98&f$!&vyOE<1FYjEcEG99QG+(dM^(npFrm&0?{uB
zGt*K|s;@nFm&`IB_6jqP3>sGF-x)8Q>Fv5>D=TN&kE<*WUc=71<5M{;H>IH2mYWsM
z74I7Q8c<UX17X2trk{kd7X^awJbXgz*e!B&j^kl$pfqlk>Mo;N^MpAXQebIlq{-eU
z1?3k73=opYUiNRE0vD)HLQFjDU|rEK)}Kbp!0<hJSmlH_opJdin>GV(Za)G@-)h?%
zC{ZcaOiXYqw^_+V3;-8WcwUNRxDTK2<9F3ma0E*rvFB^R%|ldNL(GxmP=?Qf;A_v2
zt(!;K`Prx-N&9xn*l7*$5P8LYeaC<0IrmOST6(x4ba=r<G9GHbI@-5ldM+lB3uW9Z
zAnD(XYeGWY-)NvsTwa1lcS2|p5&JqgpgkPud=4w2q#a)vIs{C@1~W=6mfJ_f|Bi`<
zC7T;~J!#WG7rZ-|;BkWQWVS`%g_60Qs0JM}eQrZ%h(w>~5sCtH0!9Ic4k1Ol-*dxn
zCWxolRBiRux4BhSrTB`$%EAqQLrvW~oDwO5L+gkvuXc98sH~}(b%EKyCZ>#A!%a<1
zO@F3a(T@O;LxU$><3=3t;xX%PTHkg;3?(OOyR8#Tef6mr1@!kxNRPD_n*RG;?IieB
zlyGvQy=IwBFhqvHRI$Z#_WS&8lNwX`d4Zi&crvp?j!L<D$rMvr%X%;N^VRlR+Lj~W
zh{<iBO_x5v<(#~^h=`0(z$krEP>&Ko8shfXhfB04h61}|DAkO^ihGku6o~3(BU|IJ
z5ca<HdHwLr7^cg72E8gz+~7(WPz(!oI$!{sdtM;|q**?V>muwjuc_V?)NQRm#QYqR
z!cTX`2~a6UfSg46pMHc_M(~$aN$-ZkeuDzXhS(fKn?s{-$aJZGQBaIADm!J)=1Z4g
z>`W#xONJ6q8xJg7;5(N5?oz#K+F=)!g7?RM&Fv4<^a?JF?=9K64O3y+@e+KW5+2U}
zaN{v0F1cLglgJoP2-e{->8G)I#C~*|sNItVv$MFV44-X+upPm>^>FLlDnF4&5I0N3
zMH9Va)2o=Kq;xLNT2StNlJ5^`2s06XCT}v>C82KdR<4ZsmT<AXVE3e>Ul2s_+{)El
zKDzbVWdHohky_qTEt7wGcRI*~I41qzi?G*^U{MF6J)vkmvOTxC%qjS@?K1t;qcw6x
z{%_SMD>~Zms&JG(JsBo+;cN2~y4{EpgbZtu<*9#l{!)OQNcMV7meUly{fZuTQiq2!
z9%}!7`T#90%;dCjGr6nsgh<k}-0}fl>?ab$LCB=|_iIG>=l~ye`D$Bpz-H8<^bK||
zkxHPx;tRukaW?I8Mt#YN_(=kFR~O-86$g1tOol-?aV+s;&F<{_XIGTBca;b2Ik_gy
z$(qzQ$IlUjqt%tP0COD6=eB_}e0PZu8W^N&SyQDp9=e!Wcxc|{8b8u3OZ;jqIE#K7
zKuU5@>88cjiSb#aYP}mzqvd*iX2tvAq&E|o3@U;vVZlC68%N!GxvleXEAb<B<!wMt
z5wLk2ZX#XrYk!pR5DMKlNPO_RqI&a&z52l53tG%=5C`uwVNWXAq_=NKUagP_3hXE9
zL=yG?wq#SxWAOM>T*&w6r1h}e6VfH-vA<v&b<EC-S^o;kQn|SpDwim|xjmTPYPWuk
zg>^D^KHPSRD9rQBwK^%?dPjdumG+x#j1_`10#rIBGr=}?+BfI>kdxOa^G4M9T}lC>
za^J{1+61~Bq}WM~y%$MzE9@>g(P{ArywwH9?yn%tp>0`oBDAO4!})(dB`0e(>PoF1
zpw5OGW%fxy4Vp>>MgC=9zJwoiOSNwcM)ONP?+&j<5s<cV&W#|K>K(Jsk6vnfKQrmN
zxQ?=p{aJ+UZ%g<15EI9{=qBV#8VnMi$pfk8mw1GOb<9GeFDp$~c)TvlYjpOp$N73A
ziBq{e9BWL)O5*e-uJ$$bG-hIa$)u<QknHCdaiM_whsx$7o)q=3jUR@<G!>YHqCqdI
z6{KWj=olHi!Iu_7kgF6;dZI4$B(EUvvS?Fw#me=yEjuLiea>@RJ$DLQJrRL!O-zeG
zVu37#Ncz$-J|c}nG+M>r$!fJBwdiw%2z7eYfGJD0(kbA+s<MC_Hzt+GFZ|rGx|$|s
zTT|T<{mH+rplcxJy?XduT7rjyCBcWnUGCwXguAk-qSVn2o3L^UU7ELu{>$^<n^2|B
zua3+AER}FhaR<;;^yz<_`pwM8ci-7hLW3idQqae{CoqpC%oFghC~t*#qIur+(mKF>
zclFcKim%t^-?g$b>&P8XNwKerGM?)vy#P0#8lNKn)d?CEC&#Zx!__rT5GeygS%LzE
zlp8fGy;a}uY+E!1aBG5Upjo)4`xO?Jg7y`;)Qr(>e6g`D{ux-maR#ehu!oxJnts|>
z;5LO)<%?K3XNOsK*A>%w9(o8XNx0Ku5)v#1MyWR*ZZGxDe~T(I>d_&#@h|25CbQ|f
z;->+J@FZ)l90{77o^eYk|04hV>Uia`EA|bA#8(s)lylWmKxl;dj`9PYuIyRdLX@GP
z4Kea7q#XD=kFzF}ao3Qt2?wsk%s7QTmy-u%O1fW+K76s$=w>5n{R4bSap_ym>~#Q*
zQ!mA#8L+I2E#h<%ZJjf!+dm^k4Uouv607}82qUkjIF^(~qZ&cjd5eyVZGzN?8WfiO
z{35DHn7Qu<(w_e$qc*p$47Ed-vAF7*1$QjCZ<TQF)%AUUI({mk^}s}L<}(6&K}H8~
z#sjvZHzT(fJEKWdiZX3mi8^erhJP@%;j_9PcpbOSj892?C8CA1UTB(Xo%>;<L+f7(
z9FfasP{wVz9XCT%%3Yae4<9adcz;IL&lQX&V-dXy2q0d6{ia;$-SB~nXh6V$gsx5R
zOXnsH^;TU29bT8y&EFlQfsD&p;q^67(a}<P_8j+yqB}Z3{{m!{R)048eTTNF1M;43
zWcvOozP-H#|B(++k?IZn0*P9x|Cj&@y}G(<Cpu?er7Q+6qi!D!b}?-+uiFt<(l=W9
z{=s2Ax|avhu3cA(G87V#49l%v3)AIHZGQv227#(Pis;=hdWAZiYC?GbzrVk4O7{cd
z0=3_x&jL`Dz`F75+HbZ4F$H#sIW1oelgSiqXBAc8ZyZE3ig}sH55PC1Q8mT>WrJuK
z7?^i2#)}b=j|7#~IrwsB&Yse1T2Zilo_QKTidbw>JMa5&yY<m*g3YJ+F_6JGqDUDP
zv3$&Yvzpk0mPrs9BI=FJn;H0%5}PB!LK7wuIH#+vJ#gmj?cHL0PQrz-B0ZilW%T)Z
zQAu&a{h9}?L^#k%9>@pim@KG;mEYeLbj7-D%zv6eXWGBe(dq{C{Mjsg;pc~*=2?cp
z*Fm439b-QIu|iT$0n@hWeH-cQVxH1=k*Jf%2fMy7?vV<#YrmL|&6g$dIjDO2lA5RI
zwBf^Ttp%^!F~|1yYaV<1^}LUT93W=y-S)OJ#JTKltkWtPebfy?i|d)%Z+aqr?g_%A
zoUh|HZGN=H#8jMqNI?BkbHPP6%4onMz(LiX6>Tm0fN#?7tE`rw_F?UVXZo@^+x8Y7
z%E;*sm3*>25a-Mz-OVy;XvF+AjvssXfRL)B;+XMqyT9%|PUp4NL#OBGS59Y}HSf`K
zFV85FQrR4gPP;;TeXl<ppmRbUW4XNuKTTP_K|%;)w&b7Slvw1P+kHbrLji-)c9Xl>
zR8G#>M0{%3O2TfXEnS%gyzhD;-3@lKlj+%-QGm126T$DeiDGe!2B}gjwt4nNv)i?h
z8F<gf7fZkF0b9Yz#3M{Y&8_iN;;~CC7PrZ3hLMFFF0Zg~>y^C2$q6`+Pwze8{XsX`
zoMQ)gBe<1B!St%~any>jJ87U3Z{NA|*>s29-yq)HoOtSm4P_x06yGAqoV!S-Mi6o`
z`kpJJ8cxYF@$WUcje|BcL$7N6;`|&4G`%KTryT(E8U46z-<eYa{(g<2dLA7h6CmYO
zWn6{xdj|11&WHb?AH^eyT!x0F?8J>3TQUfgGvwmCNa(NR=z4J`aTQizhDE3)n(kIg
zVXALrbI0(}kfQ5i>O}3Nxcu6>j5rL!N@_G~VNU?z8)pm9L#8-_(*f?1Pow?kykyM$
z6(M1$637yvK~VwifnEHVd95acPi_uCCNMR8TO`GMr2?{Me%~VRu5Y03hvikGHyIH$
ztNWq2i8oyh%QRLM)6&wEx@Z8TplO%sZ+x*TrrAfn$3-V_iyYlBByJez+^J1Xer^3Q
zRH3&b^G5!uX}HT4@Jzbw*@vDJ;cIH@mZ=xd&)MbCPq;kl&yPF;(%xcMs4W`C;fkQ^
zf#!|W|A&_lsH1*Un%M(L3(%77?96zHUigZ?r<QcCs=6fRNGm6m$3#bKQgX#g+05=G
z@i_15U)j7Jds}1bQ}0&k?P;#`cicpAKnFE1zmUK6Udyq?ez}FI{nqoGJ-7Y#Pbue$
zyD=b#Ld<4Iy3lO32P;(Ob$9qti%jHZ5h307fp?!uK?e89bcI%p<Cbh}CToHG0x&A_
z^wiFvy*hN+*X<6GE`77>e6}-+3`xHNcT1|k!|GZm^xbl600>5#{3^4g&EuBiD3CD~
z8c%r>z@8^p%Y4(?*W2p9D|aoLlD~8{-ZpcP2cA87d{vu`IP|)Q9QW}9)nQ?btyAVp
zwq<VxUCKTNJB!aV?xO^_Zx2}%zYXk9-jG3X_AK|h9D3v6%0$&>$gXpM+956b4i(X^
zhWpavw>jI`yLUkDDCR#w_~M^*xxA!;k&&(`Cq{fCVEof^G>Dipj{GQe5PGK4ku~;_
z{OOO_vm}Hr`ql&>mMjx>Psg+O9&Ofr(tPyRcy6o}SlTsb2+d59CcgFhQ^&cY$!rHq
zz}psbLV<gD+u`NHlaz;5l>g-d2yGN*@R^*h{tSeNhXc`4;6dzS`1nLNR}CQx`eC!=
zxnHJ5yy2CUS9|`ez|Rgi3cq(zOPCUGM+wWhbQ6B)c4|6WkzG(OnW|9WeycarR}CP%
z3YlZY6y~+k%65Wx>z0L2w3d6b!jChfX+i?Vl0|h@=<vhfP>b^NPK<jabR0~=7AE`S
z4)Y{lkqR(IzNgo@`SncZjnk$*(o=qV28P%{CrVW7KY7wBtI6rME}B$QSpEG9!Ek0O
z@w&{G#WGtG*0Bcjq7SkN?jKK#UKnl5)blv6uI}5l&f(h7zvVydTSF7ZHV;!sVb<3U
zUL1_+l!Rz*xW+9ku3Br4GgIEnY#WgHO8t1XSK0mp$PhG|BLF+a#o{LhB!Ncaz(tW;
z^xG?&k!F@m538_TSqDL0_4Egt!b{n$nQFp&^IKEtV-B4%D&vti2@yF$*wV{ZA3i)8
zJ%fMo5_q<VPlu}HtL!=70`VK1KPd{EeCk*3*fQK#FV=So6ZaiJ%G1fKa$JbPv__)T
z1Yd&IN=desW4$GhSwd1$XUNp(B{4C&Jh7eLQ&(se@uxz$`XGrHf^u2DCH|KmcNVU1
ze`0_tLR3LvB-Fr<{{&FGVYh>o?HQzX^||B}XJ@A=>0ThC!_RKck&R-*dk--o7HrVp
z1dMwEG5MmSqgU6a06zDGl!dM&-*&WRuK>95`<ag!Ipw<SROs=hz9>K;htMaRT<mfU
zsSC?myiX_lwY28iToFz<j(9ug*@F_8b;`)H$rLy>>Auj?)6ZB=at^2M(aerP8@*Ay
zX%4k-1-#(UC&5!b*766ydvYYCwZ4H>m~;yt0?E9Ii*bnG{Dix?vx7G)nA!IC*E0~Q
zKmky%`;%Vcg?ic3MBPgnAC`nXIRV)9OEAT2;@eYMKiHcXcaZ2An~Z;NMCwM`EzD~r
zyPTieIG#`Z;9N7$yFcAn|BZ`wPR#`l(_xEwJCi(p{b#5(JMLS3>zCr3%@NcK8=Q6F
zM?KF0lf+d^@&!yY(v1lY7YFAX;h#Hk$>1+R(O8Kap@zMVEtEHU1KLxTpsajC3eaxx
z35H(6^V_DT+xvSy7ETU<<@xw&R49OS)<u%vZyBb8m#$wqWf{TE4Ut|%uv1cgG&6Z3
zUd2#4X5RPqIj0myMf&sN*naOMhqr53BENo>(9=Kh&zXPaN{mDN688l|_wCaFA)X$L
zfv373+&M@|zdw7^NsclOFD>tLI_$P=+bx#3fydZzJYlnthc(iZIXEJal>RWndt9c~
zdK)82V&fVEoV<PNcK~+6zdx;hZNK7km4J)a{@pTRX@?8Ze&c)(lw#{s8pc<lgPK|#
zv=-ixW0MDhUihQq5B6_n8&&q9thP2Xww9y`RkBZ?c4TGEhErVbcQ)AT>P}VZzJx}E
zegh`gegd$D2!T@kf;j~|PrZ8}f<Kf2X`F~ksv1dgsm<QqsM^AXJn_sFU`Llu82m37
zo|l(bP*@m)HVwGk{go!vUebP46?yrkNSKL{{VD>J*-CJtO+@J;fU9zM)NFVC7wex;
z<nEy@a>2&=x9!0lzgJHr$;tt_&0+)q0aZ3XSyUiUc&s2zXC@U$A#V5SO2guU1k3%P
zJ>n;jZ}oZD`Y(~zpQPPe7HqG==NM?M0!U>}9FhvLacHqoJ%Umivomtca|V>LO{z5e
z^7Dm${0i_+SFp3Ojt_g9{8ruTCq_prz=hbb4nVt8rXeowadup6B_bN|M5r>nl9Mx@
z4SQ#2$Bdl~fo@e!Pmfzefm2D3@eQtU^lL<>Qnr%!%%KnHg`*~(EGI|J$MbDY)X??2
zfzQwIz7ny=0yX>5!qvs)VGAjPR#Ut(9f{{I87IA_*5y(A%3EmBX-2Q}OQV(k-WRhB
zucK^9l0h!VhNAD^3x05L(a>;kxE@^R{f@M*9~bv{sGF&dpz@2@o|-BxluO4YBozJA
zMX{86+`i(sG0aNb8}*vUdFf)N`P(e@xU|IvDEF-@T3vxaXOzj{NV9{HlKFd+G_8|r
zZyFqR1ipFJ!7Zy?GSK!cCiSNZ`nA!s%M6TB0F<|oH({;MFoSl6@pXBA^-R#_CU{aV
z=KSVB5T@iA+nD&frUB@1xg`ZZHe3mbYYDwv4pU6on_cI?^)d?fB9I#2*!*DYS`?4A
zw`EzSUBfaOV>~{K4IOL`X#Cz^M-;BrL4W|uq4(tJ!%L40>VTB{ah353KDd#0R%*Fp
z(F)ZO3|1^D)*KnNDGAQ};oXkg2L#W>%l(8vp$i5aM=y<u_wZioEKatm=c_2Q9j=M-
z;^)7B+#?5vhbQ*eo#c8ZS271I&GGT^nT1OG*>sQ)k-pG(rc+T-gj5z)6uUU#$}U~V
zA?dkHb%x_c;0t(PhhPOFV=~gvNd3(+p7n#BkyMMLJw+Sp<MNzRgYUl%Kzt_fxE|ho
z+=vWiXjX563ypz+;o1aHKELL(N;B3PB%oH%Y0rLJ|NIpZ8^BxUt3@!^3?6U&;WQhK
zy8kshRtC`+)b#ZGb+v5JH85;*Ej&v0*)PoW-L_m`3q)>*{n@v3J+2LRNcB5zC0zE=
z$X0run!e&gpWbBI;V&&M%`lpIzp<&<elLSm`I?7~m0q?@BtU_amiy*%!4^+^tr1Tj
z1r3eJP^yn@J3c;MOk8}4&1U|n)k{`J=JxXN8FX=RA-&N{!sY&5!<+`<gLQ4x;Uw_3
z-6?3Fors6qrf@rnR1bCyX|dYD`4<%xQNQ*XN{CP;h(v2j87FUYHeUdTISAaC+MAC~
z>nNl!yFs_!(ys`TYgBZJ@JL9A2?2FR%y+_e$jA0z1qizrGz{Q7i0hYcr+JUXf`VlW
zVO$=k+1=q$kEi<!mHC1V<}sx>96o=4?jN@}wdH+WTWn-%Jw@%TJ@{GYrGX$BZHkrt
zrVu3{IX!OpjIWdV;pJ^yVrAreH4UA&>U9$Gr}~J(@j5c$G}vKa9s`_;zI^!A*~z5e
zAf*fpkLt96@)#XD6;IRew*H}6jtj7@_q4#x(j(G1eqc4ruug)Q2%AB#wydlrr#fKA
zqp@MSS&;;UHPvb2pne`URi#T<TWby3Nl3uSp^2TYikwNHu2sQCgnVwrw88Qd2?!4l
z<BD{$e2P6)Z?~fV;X__O1H>$20-SuYov42@%Ly4O>DiR~<m!aML4Gzu_EJiKgcvy{
z<@+3|6?Rn?rwD7T4%_i8WzQU$?W0LHY&UGKZyg-V6$oku)jdDWVjwYH6WgJh>Pn}c
zk*>ElpuR%h_nB|qU@PgaW_WM4&tbtNa2>(NwheB}-i7p_?r+s;zvdI**bX*+JY3?<
zIn1CKUMG3padOWj_`v3@si;VB^pu}Q_abFgm+dtf7l$+i;dJWKv_5~ndORsXg2!nG
zyMLR?xwE+)ixvBk=kyPjs$CNnQ8JH5wMe!*m{&^?vvJzI+|*P5?;Q2=$GMr1cF?)%
zrD>;np6mp=my&y&)zTM!-YwugE%rS;?CbqG>)r8(FBpVBOa>D_1tR4$>To>5kkce$
z7&IH~2PH{>G7{#wE@tiQyw4{Y{Tiq!qc${qYvED~;$?7p0X94k@f!Qp3pjb6RO5`Y
z+7OxCLt-qf7cWXwSR`^|q~7z5!V6XY{R?KLsEEd;s}Jdj;Z;Niy6?pz!81KPEX+L;
z9?IB|`eE@7J0@J_)f|q}WcM>X(3SD%wai(AQrt-EM|E|8iB?KM=<n_Kz!CH&iE*=`
zpkUspYVy6toDYdpx2$qmMF}bzT5MW2ZAnFeI)H*TB<{Pr_rG#uTr|?k9PXIaX%lK$
zX_^y3Xi{?VMo%TtUo{p!?&NQeH>_$}NF?DN0B~5>;e1z_d={KB52`B^9h)M<)|}~k
zJ6CL4Z@BiW+`#)a4uQw&DjD`ci-eq9EiLCLb?an|!moN>w?^)TzG}cU4e<U50GwvG
zImLeSKMYu}uv?tx;sWF~+l)RhX89SBD|@Jy>o!_UP_fX^dQ#3Lm}uq9<#Sqyie<zN
zih`0Yj*T~SCUD<EO3;M!TQ|yh_Bq3F1&;UC(rZU%;~wWiB&1dWHm-(}g6+PwPSt|h
z88$sYrc8_y2t#=W#JJ3;uHdmO<w$RFv0AcK8cdK>ar$v9JUYz9QES+!DkO*Wc+pa0
zHgd~dxP_s}fr#FvG#W{I2BoI^56!GJ%eTgC*m<#PQmcv9c#m+oI;)JDPkJ+-K&b3Z
zBleW_b#1p%acOB>g+hjoHHTMCK!7)OoOn#$?kUj@o7JphaqKhj@g9F?ECI=25#PJr
z?O1xwfIOg~H%>2yk%OZ?E32&f*9pVX3!m<<1e{<`UXG7rwLCXBxC?M6yjpV!`q24B
z=H@$_5&#kA+kBdHzftDD)w}#?Df8zoM)|k=X>^c8sMq97%k^v}4lgwh>2Nb!y_QFO
z{@Grk^-Ku7a0`;9>#+T4RNUrj%GgXpb!d20DbG!BW+CkR#KFIZ1F+~sdxilXD$E>m
zV$teZDlDh!m<s6Hwec-uB@^P5tSW>Gxl^WQW`Q(t;@!R7F;e`+pRDbb>@D^@>KJwc
zKHx{F%UZ5UKR5myKsBFvACEn=jlUtm2ne<8*ogMU@C>={PpiLbJdIEX$-c!?mhx0D
z8=E9@*K9Z<P1NYrbiG?LWlT)w?pyfw=yqho4p+z%WPUQop0UX<G&{dRMGSoNCz>(%
zE>9ytWshAk8Y5t%A-*?l(ptvvY}mDZ`ez_=-_J=}RlPa6OaY(Ho(R0b3iHz%uii+F
zi3!Df&-FI1<C~p78eSKU%ds!^NMCzwjsdXcNYH*&fkA1?huif4-vaCI5B;O=<vHue
z-#+f?Vzx(H#|(Hw={2y{tG+6|+Y-HSBoDT<NzW-;6PD`jsXDbf_K!h%W1ngad7geP
zIU}oh<v<!T)6rXQ%H?*+!2NsQxxzXCE)ZHhm~%S)X_lb{1AuDwCt-j@$((0>c_(~3
znW(C-MerJTySNu+yfwPs)0~}>h=GUa712ilkD>H*g29subo9g=5eQOlu|N5>h0haK
zojfeITs`meVZThOC}O9kl0ncq*vUl>`%{jTbB09u(F8RM6VuEX9U~*mXs%l~PI{lj
z(CSv3tJ-em|8ez}QCU4<zwib@LKLLCr9-Kkln{{akOq<N?hr(}Q&PISTa*-OkPhkY
z&NKX<^PKgr*DttSF5P=(_BB`hLOT}a({mQoPkw2{U{pTs`n{mx@^ZI?={V-g?#qtx
zR{KZ1MWMdNKConPAwEDJU}4-*I)h+*t@n3W(D$Z1O~Ps5ZeiuJfC6Q|k1$=eqAyhE
zM*{Hk!$PC}nR!*w2wl8~`|gVb7A<2ALV%VkpBLsupaiJts{D*2>N>fy9&mb+k6HVi
zGcz+|!VRyBQcSt?jf8W`UMSVf4~*sUq%Bl6i}8jzTXudXFMYz<Y+O*(RMUeRX{+%`
zbZ-L*ng9x;-GAE>UO*_pD1h%-Tir8{?^hGVrS>issi!Y=`g6%opt|q)?wyyNV)UoH
zu`WE^)&6P?#^3#LLKq0m<LW%XM!E3AMhSaKJ_o%oZfI8_7@l5T0kAH%k#zwr)t~rE
zkpM)*H<guTp5$FOBJY0Dg(UbPxTVP2NGQ;mzAseFp75gsL+q6B3w{`ZFh(ce<&8zF
znsR%`oX`sn!f-~STJa}PczF1zgjiQD|MW_GI`{n!3f~P;Sv#pqFv$RLMg!VhpB9Dv
z1Lqva!eL<qW5w#M>M7wo4(slde4Vh~vj)5jA0V+g(Aapzt26gowV0X+wpBbFW6Isg
zUw?Vjqs2?|n9aYn;Pj_$eiPvL)VFD-A2DCEmB^u~-;vNWkUWSeM?5ztj6`u-LWjqV
zp11I)4wy)b7sjO+nHcflG-hQPf{Fp;7OZgzyy>qpt*km#;Vs#NC>7$wOY1%`0eQ;=
zS_mk-rVKRp(iyx-0a>x6TN`nIau|<uhR4wnZ+r$>un<%4P&EoX2(W(S%tnva6zc;g
zwRi%_i0J$hr#vt90rnWQc$^Kxl659Uq=L_g7mQPYNeAGc<wQT+9r$c~B2ZC6PM1}?
zu4kSYP)L7=(dlq#SN!e=*wssn?2+DC%bbZ4slI{<b1=Wt<ekE)ZdPe}0n~}3x`uU<
zm&1S+P|<JH<`AbnX$u5CKf3eq?Vi`?M(bb<s+R4{RO3PrGHO~&+@n#d=4J#^bIOUO
zv1=p*giHUHyRx@s@!pikevZ#>Yww-bpF*~T2C@~JOwbajp9JbL&@o6}r26R$H4{hp
z&1c@-%*%fR&)&vpwR_s&sId}M7)uLS&N6yp7@Zd1jVIC4ce^DZX?+ES8{H-_a{n2i
z1TDE~5ji}0nX{qKLD;9q7@W_ZEF<a#OoagpFeXXt3vIYs3b#Ia-{<hnd0INU*D#nm
z27<aD3AbAL35gV#OS)uWDp{AK#J&k!eNdawj=0?eNKyzmbO9cc)Bqy`1A`^kLfHX=
z?ARR^7EeTA9<T55QfEXnC&Cf)1m_?hXY+yH=TjjO5&u4x%9fHp@v?4Wub_gv@iH`w
zgOlCkwKiwCBg}M_D!%&@yRD?B<`YGUyQkc4QY9jpQ~yc=nE)&U>2tCUyK5CuH5V7W
zkc5Q9jjsFrwNvr^mArjPx12NbG8~gjVJdO9N({Cfebt|jT%lRR`dmOM4I=-E#Unj}
z3N^T9o{DU_$(N1(o||wMD^;VVriO&fB;U}N(Yk@pj<vjiNx+9^<Q}Zq1N6%+aNHwi
zqu<Wc+M2YV>pV1=nT07zeFqESb92?lw3a&-??him^|zNxUimAJ{p48+`?4l;1p_-2
zD9PS|Rwi|G{Hv!1CMM>t>JUc#V|N+6<rWh$udTv@faUS3WCVS&HE+}7)^ik0KYl}%
zGfkRM!*<c<J&&!fP~1hkn>{ege?yQF{<8C3P}Vr`kBOq~hKGE|<%F{*Vnh`|IS*+w
zxNejz7Ar41HTku){H>2SWe-y#Ai_h3n0lnIS^#cH2|umtBV0(#MS1U^KaS%B(jlm?
z`XaLmmVd&>2s06wauD`}IJxWa*QJBt1m|n;E-*T-DI@a16EsM8Y9EDZks#sEpLGv=
zXo9*^cYptuN-*jH2=D+#N(mFHo!v%ty+-WQirp=f<XWHeD;!m)TIlts;qL(s>-?vi
zh7nv~1_lso+#c7bs{Tmw5^~mhfiDq+=f<{3n|Noff5py!@Ya0`-5T`{pSL=Sx|RnW
z>OwzWhn(vW2sdHHc#Bqjlx!-_p?gJ5y}G(Om1xGEB3X^7h@JD}&3As)?)WgaEcs7%
z2Es^l;qMQL6XV??R)1h_aB4FDdOa$@#WDUX9y6!-hmbHvP!lcY4Wo%zPbYU;fFH@I
z&NqH5_n2QNADm=Mi+lNU0@)3_PixC_pZsuH+_N*+=lEv-3>CPSmZ-nyJNX%cW1;5}
zzT$Qq6Q7G<{h|H7K9&|F%IO~G(-F=f!3pa#C4B9Q!L!irVgrUD%|tMFwh_z9?yhc?
z$%*JhpNE@yrDy8zX{c%IRZ3%d)%lM$4Y6o5M+3<v7`H@+bO5YVQd(Njb@O}i{MR=y
znAM&ptH)ZObxBd4c&Q2)Of7+Plw(2Z<Y-0W{8Cp*5TT<x)|I!13plr+F!h9Hk#|iC
zD{JtEPHU^H4>>SFzH~(@3<)dW8-1IrP2}ol9{!StCdEd%FKs&>sZMaJ;;fC%(__m#
zQ*xn46;&h+FHm^RMT8Ef>3`*O%@g^AqRRZF%ocwl3eUX;DEu}r_l$Pi%A6X<*^Sx#
z+jhP^w9&EA-VV-5?Xk>sP!?nDz$~Ng3oCBa>(oUr^9EAreH%~KyTsqS6wo3s8~tr*
z(&SC&w@Ri(+(nMjfB&RJ^US{aA})$oy@Dw;S%-+{f11;hnm+pH1fye{Ohn366OhIU
zP0xI{p>zWO^^~FRg1Ku0QS)5{-;7scqySyie&Bya&*(_4(-;8N5r%i`zZOy!M99*W
z<T`egOIYEF05%Jz@=DKCc);~KT8TcUGyhNK9W_CTTO$pC-%Xc1cAEDY_ul|(K?5e-
zQ1R>$BD9>3I>u_aJyk-ATtww(=s^(?;%o527&#hhW!i2J{tu9EAf5of<7UO@Cr>5v
z+b)m$F9dHfAZ%<WAn!o3@k?1=pw;@<)315J#CD)+_dE?OmUZ)<^MKz?spZOk8XxFq
zWk#IjIIx+GQUj}db01|J1-50BV+{$|yO0Vjh}dGPmZZ=@Ecl(UK~8Q<a@{$g2ILg!
zzrU@($YCF&C++K1djfGVw10yQIo8f9{7Ku0EfueBenVJ*!<cS!vGWlfH77-wnkytP
zrS$9ixUUQn1VEMl<wF2sSmg)_nKu<qY$r`|f3)HFmXpLsc>*|$QaVO_IL9qIuv2*%
zCKi_HXK@jf+2Yy3!Qt%R!mR%qV53q_z1)gznGpUgj1!MT9^(GaPc8c?o**Ka+3!o~
z=-egOTI3{xd0pa*qM{6~N)hX;(ckf-M=gyfbvBEYJF_;$+ve{h32Q;sHj*vQ$;Gul
zUw=?+(?-FyWik1+&Ly8M7<6e{!JG~I$4P&pV6EpBkisB#83_}4<Mw+Ms4tQA^=^Wz
zpc`FKVIZ7b0O^13@}rE%F=&-<2m}4MtW<DWdZ4>x8Y39*&8)N8Yg~YiR7f((=-A=i
zNm9)OKA-WQ4L#3kMyBxzYr`Ku;H>5zk5#Nq*AM^sskpvAcOB2-R(r(r??O!kIrhXo
zmHYAJ^z*9M)s$N`46m|BOoNt<D?bEfVPj#XfM}EBuzh+BG-P5#Y>w>gj;S?5yOdOw
zi9iPZ8vhTjg`&ad>>06<8dswq$&-hopiB<Ql6(jR_{1A*{SaW)BPceKqg9qd-hAn*
z14uq19TEo=1Sa@TMWge~D5rStu|Wac1dCUVF9u=*#4woVUiThX2#|gU-o5)9dfsMh
z+{dqKzW8P7x~*k|l6f~!5!1BXEO<}~nd%S6&?Yv$A$s$1b#T5}K7)6{F5S^sG!Tsx
zeOB(@ee+4R<FxX_OjQ8!fz9^*_E(K+<2{Rzv>dp&Ft+yZ71J6@S;ZTfy42GN<JA^_
z@$KYQRK}<O$oyz)m5A)TNKUJn0Q4vmpwCUzzwokkwqfGX7G_+J<G=1~D6O1f2^6q?
zkYY<*yekjILvXqIgejA+@H*!Mkz!nxLsh&~ye^*2Yy2E>tWMo+^2`y_IH`CJJSYc)
zUj#&RyO+5mreORt=8T0c7P>?<F!-uU%2{c!T&fTM)IDLvEgeSSAmMB9aT${IQl=-k
zXspa}Px~t<#FBo@j}GPzpiBMLfY=B15mcxQhFP(Zg1Qa*9qllix6Zq!TUXBjv5Yrw
zF>DL=_VCADa49p=KMGJ6<v{C$8Tm?${e3~<TaVPD0Sl~o^d5MV^hEx5yn^ojZDNAM
zD}rQ+5&r&uthy~reN6NtIQ%cpvLGd8r8jgXIv*0QbATG0aKs}BtR?&)qtkhBP7;nu
zr*6Czr1_ukXUKr3mVIf|(h{P>dd)}5>2Lu3>mp-g1Ek}z;bEaMM_b!NAZp)D^XhDS
z;$Q0zjGWn}x27t&pRUE;|4$Duhf){=G35uGbk-i)(<l8$jg5J`B{JfLyBWiRZ&_Fv
z`TA$3l2R-}c4rd}c?83+fiMl+JsffWv8;d<1c^qkaED}xY6NhD8C?F~f{E?pV>Yv>
zsbN&{ioU#hD)qW}WHi*W;?n4h11y5R*Jy5JuIcJ@*an%I_D{q!$q!WD`A;wEF(ys`
zN5ee?KJh|qgy=bGU0oLFH5wL~3w>}>OQp;hDot#(K~is<B5}VI2oYP`9NdoiN2%M%
zGtQ?7TQ!|s-F0HWCFSJgtPQk4HC9l84}H}W_|vMPua5zJP$(90=g%%DVD6d@Z!@k5
ztj-Zt|DnLD_u5X0?fgi{ovF<SzuESP66)M>9MGIW5bmq#`?s{DkDlVg_X>EIQBY(u
zzo<dp75?mKL;R$nSHe^Rx%c<|Fo%(ZM3h!hTv+&qnBpoo00#lqP3|!}Y@9K^=uo#+
z0@9RJj9$UP!o-pek^ZB;twG#lN8IC3qAg05l`hry<!nz&X_cShGeSc&=ry23BDX7O
zZm)Yw!n0<pY@a=inJZE&c1sE0;Eu40(=xA_0f=`9!dr`EIa?k!hx_r7tD1>XRq7+p
zy{B>?db|`49+|wNqp){+eDgAoERaZ&$8mdMX(byDk~T#Z&`fRpIXCyr&ykt%&w9<N
zBLKFOlfd@_ar$HBW>|4?F(MS+M8J-)g$a;QX9o@&F)MlijWBiiFMVjf&5N(l1a_S~
zGFxwRVVe-Q8Iq|BM(_jxg!ix2U5Ad5PfvTU#7Nc(Sd)0QEGf~#c_@(2<M-H`m7x{@
zkER9V*!pB69$gFs%p&ok$t}kt+)&kQMd@&(S1TY|f!0g$MA3-3*>=c_>cmXCVy08y
zAPY#pRujx+hsPV^AMaK~KBeGD>=}@^YyL4#&dUweEJyPU3J!Lvj*ENCWON^crUog{
z6VPkmLmv{DSy+N2A`<>4s_<uho-v}-V#2jQ*e3yrsW2}(sddg!Di0DQU``e8Xh;nP
z(Oi~O!&yma;+a#e7V6UFLBVCnewV1=?cN1GpRM<KxIlWrf4rH!&ped+mfh-_<G^)-
z;4vb(bXA(Bjj;hE`dYVm^xsDGY+XsQhP9fK-?&C~K7TOhNH$_2`0>A30AJ_<i1@P2
zo3in^NMQGALjYe;d3401p|R)D-!kx5@J)2qX7Wrjfd7A|>kw;y*39M*A<@dc3WP(B
zPb7kQTUm6=Ls=PE3d@QRpko%vR}T`|AK=~Od+F^JuPRrZDl14vpS!vq{rXILisXoR
zk3yanHC?~Z^zZIM@bd5%7_B`+L$ftM0gg7F4O{O}m-d_Y&#OE@WCI{&#KnWRHmPq6
z0a*n2V+hzS*T1CFe#+H;v;VaydHWl?&2gSh8{*@>;=Fh3Ez}pKY=<Da-}{Oh3vX4;
zX)J0rb8Wd2tW?;-s1;vo43|wZm?>qBJkE7hrPj5LX|fgYF)?Mza$325R9ZSaa|LCI
zbU?oR=<56EN1fKG<0(`<Og)V8GKPj!kS(!GA&ahZ>h~PiqGU=3$qm$2mIceon6DKG
z^hMFwjaZ089IeH2swkFJp^oU|=7Qt%QP$nbr>8@8)_flO+P`CHx!NAuvOAhw7Bph6
zbC5sEw0jS=AGTZw?AClTGPVZBHzg${(MCK(th_h*-{7}n<ps~$Y=S(j+w8grZQ7}N
zewDY))DMD%ainWoY-DM8xX07e*3i*%IAMkz%rshD+R~CWf2w~;jPJVfFu&!6{u05=
zg?Npk=Oz7jG*Ly~S0oY%$Fj~cmc%@-CL41mayI$hUZ+v7Zk;{NA5HfFEeoLa2>KA5
zHVh$<zJ<9t1#2J@1QH=_<=EG2>vqJ_O8&vNKsA(A7F(>x37IQp!1`t5GEJm8^;6u?
zfEL3t%H4|a$BK4Cv{VV03$|M~uS5(yT*)ii*(un*CeS+@^1;Q$2L4|8IW;N>-{H7y
z7IwAN`EafR0wl3$6V%n3VA;^yqlkAB(wE!C<pMBVxY4n(93{u3<}c(DlRqyabw|~*
z0gb(Br#vMtb=SQ9jni-J#ikH)3BF~zJMIX8wo)33NR6?O7u8m#!*_E{tMN1B<>^2A
zgv${yFPSf@7?txTKE@zUs=2M01;phjeJCow(4AY6r3Q?P^ZiI(_R}TPEMWeV(Y=kx
zIW@1*BpRv)lYCusaz>7=B+$>Ytel~@j`5@qpliyXuF*gM9jd@TsXQnC;Ro{%KLpRo
z_>rO2#dbXPGHMes=vC=di_r;4CA>BaH$)JUocagW;^QZE*xl!*jx6+~^C9aDFD*;U
z68cHU!ylh#zCJ}=`W*0#ssdhrl2$-qZ)<G`9x}TP;oQPN*wS<i$!D#axa<+cweuSo
zur{2oGHN>+LDFCRfB*;icaS4fa)+MBe4&O(rXlnI7II9qy&-)_!s9CkPj2y-nBb9;
z%Hg_DPDPhY<R!KFweTkx0Vo1pt!60<pIgJayyZ_S@fPW`XV1{wTeFA7Px&zM3)G6A
zy9*GM7nf2Xi9UB!nagpJaW|Rs>E*b20h1QCGO$#T*%<`2KKcGPb`$IAfY2fSi&ULC
zPDI<WQv%j<cQ+tz+jdl<e7r?VmhslZ`lHflTdN10{+`7aWP361GVggZ0R6DP1#moF
zcQYedr!&5TmK8(2XfA*hQSlB-CIS4nP`ON?o|nqw@F?eT`;2vWs`7~o5yCD}sG~G-
zwOUHR6rZ?9UjQ}CK4vTdkKvW!@J{a5e?$s<fM*Lz%~vYS*)Jmur=j7BD6nLR4?(t7
zv!K#Qr7nocE7pAQ1)|KL2h<W48hN2y2ng(wCB)eZ67^IrKX3q2Z(8FUZnZ`TRvbXm
zO9MFQJju`rfX5dG5pJs3Ie4PakD+$RDJ}0$vtNj6OaoYGWFBANAA28MVGM+$FLA(n
zY4aZ_g5RcQWN4}CU}0lm;LNj{iLeDDy<hb_JA*;C^971>Lq&0as{1pxh^9A=XwwS{
zO15!uw{;GpoQ-cU?Wli75a0YI0M$sss0IfggwJ6*yfIq`h_Z857!@{5qjyp2CJvFx
z6n$m`&Bp2#ST|ubSSaP}059tpl10C5pAonT)Wq|8xob5^u!wf@mEa)7Y5aiYKhx~w
zA!kmth*u@!4u6eTCd@fe)~JfeAb?_dPQca~Up#<nK&`G`s1lq-J|C@%mk}FFB_H$Z
z@HTf`Jy)uW4qLP$_Sj%v`)4aUbME)L*Sr;rsjVl$m>0)qjPh~u0Yyuq8XB^ORuy_t
zX_bz+*?aDU&zTEJrg4ib%v&8OKqE~v+`T(&tL;4N*<t_D3iglZ>NL-0%a~p?>v#A)
zvRw~aT{7IR%?{iHYuM@O3nB7}>XTYNaN?M=vCNzi=By;Z7m$+z+yvHjY3;{y$KQP}
zENW_MD3H(;#zs7T5KnleL>R0cd#x!EfrEoH9Z&(<l_6PcFOuv}a=OdaidBKoT6kZT
zm-|IufB!50=f>6TYCFNRa$6whgiXe!#Unzi5r1%tsi(G1uSZ9fLZ@ZMYWkAip<UCS
z(u%ZS+_c7gZ|pD-&S$g;#zcasC@F)yO=zd!ge(=p9B~nwNfG;x%(lQA!UO1*goYar
zBzn3ps~c0k`ytTgCnoT3yB~lROglNz`2u{COt2V6PN)dX23&%pq*1Q0uXcDRh%X9D
zN*gfxkJLg{HHEj;K+D*Q>jt;jz6bj8W<#Y^O?otal~h*m{_vBA8URfg7C!s(dq|#5
zfkAbcD+k_r{i{1D@7UIR@nD;0Af`V8(_*z~{B&j)`w}P-2hz$)NX8i}LIPd_nn#Bn
zne1YV$zP8$CB69!F_5BY1|lDBWcZ|WAEk870O$fQ9>2{KXp0e6`x_ksUDDk@9E7rE
z^ppK<6cH^K3m#In$+zIgfg86pf`qiZhwIVD%6Dmi^suGRtCN~WLiwXm@L`qv5n;_?
z-o8!?-ZXgR5ewo-W-{5$k_*PTr%|tR*LqP${RqQnD*KpvtE_WV)!^w3Y2lNS6DD?+
z`^?R2dFi+>fjX%Y7FtvQ+;0!<nm1Q*dGvjZ%lz~iNNm9h#fq`Q$$4vUT^kpR`Ax0+
zYh=Zox|B9kZ}54;Dt7w*V_#NyLB>cU2|?-zTd7IRe*rHJ`cYFj=I76f7kf4I{=see
z=vD2R6ZHL45<pSJv$k!rv+@yi2n$cta$2suqOW#22=W8O+*D#d_5`bv!s3%AmmmRj
z!Skfs%MK*6(@zQGEL^(fzsUJOvCHxwp0vAmQilxBq02C3-g+gpRLf+}G1jI<6eB0C
zps+OBq?9*zLPshXR8)kuyu7?KW+^J6U0EZh#O3KJ7D^1Npt%jmYZ8;{trY)muQ0b>
zK~cV(tt{!<1?sM@%f6rO!)4mOw{(5ks9zb6VM3Js)(CbV3<t5W7&WShS&`ut;>eMY
zab^|{YI8j~ysA|aUc_^EVgb={1*vyYYA0PAT6p}2<j)n<)M%c2Kiu6TK082yFbz77
zZ@1O;I8z49EGDme2L`)*KL-omd1|AU-1HA;UWtj_v_C8imQ*x!S}f9$??sHW;Cgfc
zBwOwgz<=On*xcQ;I_$T)ypYxThc;~m=p4b1jJ%Wbyxgnbk{JqWx^~;ZPbgphq(v4Z
zmHL!ILR~bycDuU-hYhR$Na;mk#>`NqY*8GedF6-QGCB9sWTI^`MS!Ly524md7!Qsu
znuHTlRD9VOCDQSMS<N2j|I!4l)5n@^o&Jvaz!H?mQ2vKY==S9KG7jh?@_s#YQ?8);
zt^$U0%bYKW`4J%C4~C427WXufSxf4C=+{>?pueB+SDqQZ<?ehIkVcNu^3#G8v&BOp
zrOaR=8{lH_ZF+j=Z)vRM+ne2Lz!TK=xx-&<2&y3{6k|FM0pGIpipPN5jCrT%MQkMB
zKG*%tN~C~!ai#TL(?38gBIBtls>{pF?3mmqm$>=<w_r`9*tsTmmrxQB6hPuwxP;rk
ze{4;tN_~AC1Z!z)Qz9-G#)XMzQ=sH`hGB<2fvO@jv!6m>o}Oo!8Ys3Fq`6^m#lyFe
zD-8+B$duePF3?+o*Yk4?9Ltwe1R%{<l2rL}Y1b!z2vC<Yk;rb>67?Q&zqgk|3}0Wu
zaoyc4ww=^qRW7yNz~9{gj46>DkI%ww`@9_z;y4yP{Sohc?@0EFQxgwpf_8U5S{T)r
zQ9`z-OC6x<k9vq8<=^~QG9EX6(0+FwE1U92ZH{6IKX5Q4sV5|#wYpvDKw5M^k%fAH
z|Gw@YHb}X@I{<d3m%Yk1)vhDUeAewGPz)_iY{)V4=yT-fC`T?7Q?;#qX4Qq?tsf)q
zhE*Lvb|{hSDk_e=x*ODA;%=*`0L3R(t9-hpw5C!6CZzy91MOKY20};p_bo@t;C8CS
zVbj{211ptwZ#h-zgZk*ln8q7-d~6R%$n3I)@7{Qm0nJNBpa&ohZhsr_;PUm>0?E<;
zuxR;+-hD+YjB`rW$6psD^Ey4GiY)n1xqKK+)wKa|CVq1DZew|$pCDZh%C2Hgh7&l<
zC;dg>mjlB97ZdQmw1gdSHv9RBK={0t1)hg@7ch7sAk4l!Q63ZB+1dHVuxt@1a}AB`
zQvUQ1`*HyfeO0`i!9y^UPeoPoBey4(9^vSz^dqm+%y-sqDQcQAuPov;SNtMgtXisA
zEFwZoJIaxp_04-^q%%p8E=aItfCe?|zeZA{;q&Kb6U04L;n=c__%w>;v}q<J6lpNi
z<f5W7V<HV<jB!Gzf<H*{$D>Yn<J(b7`ttia7@fPRmgyS=U%}Ovp$!jJRaHaZPj9cH
zmN1%~5QJ-zi0HtOoM`6wMf1O%YOQCC(jR|j8S9-*I_nXGWXCS2r3D@mF<KlW<~x4^
zg%R?wwWud?xZXKUdNny&I6K#Ya)&|NW3LtzbT|l=z?6*qHHgwTGk^mPE(e5U9a$@8
z?p~z~uex~nhLv-^)d$_L0blPg+6CQb1;J{<%PIq$^=zvFeE(6GnAjt}ry&(<n_(Co
z5_u$}%Q&3AA8qWd!j`=_$Ur6?56d>&9>osHqLV7X*!7Iko5V6r1|gKVpA8l9DCE`$
z=r4e~rhc0E70-e(mRPjt$q`pdi1o(MgWt?gz<PY4Djbg@eoo}Cy64XJylTC_tU{Bh
zbqPqCcyOm9K@l4{(PR-M*-&?6J7vHteoA;~n<)8VcK6>f$3(Nse?k<I$UwqBA==In
zq;Vj9Nh$=jjC#*SGRan5dPX8^c&dO$SykRw_=feF5qoK24D6bNXNLglYFFd&>r1~|
zHevX9X4c+)lo3zF*_;7e*GFq`u)Wet0xhKWEMyzqj<wbOJA^*&?<O2mOV=}ZKJ5w%
z3gQL?DN0Is#O8Dv|LN3(^nJk9&dGmG{^m$qZ9Uzlz4_!}AHjF%?Ooh%mQDK}SEGg)
zNJ|i&B3yP$BOKuZaR7N%A6%$ywiyqqNP0}cNgkUIU(w!PFFx!~HLmp0W>8aQSlxvY
zhmr8<TAS5s{-Sbsd)%WYZlyh*?7mNLQb_oU>)bUc!)r}LWPxoQk&e+^O~FFycw540
zNku4<W0Okt+5JrK`!N{aIBu4HAO<c=iy$?3fd;!z>ECCA<Pt7x=^poOo?Ri4Z|@wJ
zJA-RPj(+%~D}fk|M#kwLH^;QHsR%V(ZyUU|#}aN474KHOdSecz<&I9C<<qhmZ|^q2
zaB6I*Cqui*vBM9s{r*^56TmF2qRg!^@%sC*ZfLK6%RgSrqPc9FX9Ems%rP~ZikM|}
zb$Ja1y@d=WlLmh)c}fh+wbbn#{K%U$i&hZ$z&t)kRKgZx+YWhjWbR=}5TvNJ3v(4q
zpOFCJ={uCbBw5KmTkk@K_}YGz&cxzWdoVE}{rW^j)hTzviTWN?F|AveqvDUs3zWOP
zl{RVM=$rBVPiBx}(|We9W%GGwUVMGp0a&)32Sm1I{=acIZEhj>yNLTZjI1@PwVEZ)
zO3B*FW2UuA7Mf-u<UxQVEl1*j1A9NF5}4QVSd^uG8D9F%+ID$cH*0NCq1o5#b+Uf9
zNaE6K)plboFY85u`1-{wG=dA(4;5D{ktv+-2Ffg=V`5MtB>_IYW`B^Esd+vyZ8S_k
zc+JlfHr&V64FBJ7T8+VL2hGosf<G+2n%w~_<!{7b0jIeZB2bPIeKl@ci!Fu~%nXS8
zjjdpoWklo_;jos(D9FeA{IM?DqNb*$HQ87NDqShz;eWRP)hC^QdT=RG&xZhJj3%ZS
zn=f+|9GDh$L*}Y~(nAK*`Ul_6eqgTvfuU;>W+=HteWo)1V039WEt!wwzIL9EkZ<vP
z_Y{%!rlMB(|98j?h=g@x?@+^E;*{t)I+_MGzb%QNa3Xn`vhEr`9EDM6SQ3Y3@g@b%
zx4`~*tJ62jY*Z*aH^NP$Vuzw#|8WVUf({Q=S=(2)cO8Hg#a#7}`&fqDAgJnY516&+
ze3QHtC0Amv#m`QZZ?0zO9|M`4Nd7Q@k5lsBR`!s(7!HG4_2s`-2{FO{f_qNC!%%u-
zPUZ$U_%}0&FGQ6(N^lt)5_Yv!Bd99P_V&Z(RacP`-Dk{ep`|1oMyGsEe6%!v(=zs3
zKc0}UbX1QU8_`m4M+M0Mn@K=re(cACyJl|0Edv5wA68WJXaBGCx=?#VjU%&{^`<!O
zo8hV^ptO0!*El%fpt#9vK^S5kBfgILdw=|NOfC@}u%Gk!hXp;hm_)p9_FPc)M+EQx
zS-c1TG)4{q@4-8AKk_{>ONBo{LC9Hpj)ol(O9mvceB*t-dgmy1hF8DrTe>W1B&#CM
z(*!veHrn+gvJ6b3UN*D%(dn8t3}?}6=5i_En8H5jC^M26J^{@WUi7=Htgyz$Xs?^=
z88A8rhx_UHE=JD#LqIhbc;9%aVBx@VLd{l2RUsX-7ZRmcvH8P0{WiJwraUX#bkl*L
zQhH3kIOB5S7FSloGO5n>kD;?andf1fV`_4YUloa;BSI{V?D1SuR>9rPg-a^CPC%b>
z9zM*PUr*cf?C>90xJEya@yVofg%ckvHpA%Y`(icBw#<~Wrv~o@XhAdU*-@);pw@GQ
zqZb9*u71V6+!t)I#rnKGq}coAS}H1W@or^0Y!0pRpw^?^Dj<og+Aff~pi3hJ2qI}P
z%X6hu3YtQ!4SrAReF+rT7Moob!@jY5?LD(c1he7@p5oeR`apD2Pxq^8=urU&@+C0@
zdJx8!y45G1#h<|k7}0<^LyR=vFH`oVF~NA|*pts>-JL*wRMCRaf{(>$!j2PrS2M7k
za%tEqg`b0kU9ZJqsERSqq84B`RvVh~)JUzsB>p`)mx+*@{#dqzjEtwtudl$c>+i$_
z($lAxGbA~(sRCY_yo3J@6aC2N!{m-+Hv72zewHJF(<I0r>FdBM*7$N=GtNx<Hq&gz
zcqP~?F?Nq%^8NGkXFua62GL~|y>4%6j1Xn~#W2eM(O?vUCLJ(#HnSM>zCW96Uv!!b
ze7LSS!IL;61Q;d|93yqhQ}}Eal6J}uaB#&(YLr3c-{Ub2%MsJ&bggjY_rB#ObAK#=
zUtZpS9`^cE?OA`+@%6dGMt#Z87hvv1#T@eAL*cFSTLw{F8CD<tH?cj}R@zl@jfn`M
zF9zSp0tF?%Ct&=W0OCK;Z71`*q!>B^DIA#|-$bwiL;Qr3lT9liQiBVyJM4^H0LzT^
z{9vB3++?FK4yCUTNH}Q}nM_760L6?;2wZ>r(Fa@$!UG_kRU1h_<~EfxwrBk_9sbO3
z6Mvoj;(f8=tYk=-$VdPs75~c>pj`vHKQudk%OAWSE@nXp1}uTLqi%vW=X~V~>Q`%H
zRvEAq&N-v|vpp8-SFhZZeho(aZ~5OmI?c%!e6}-*-h91{GbB)doSW#m7n(SAe+JX@
zJ_BaU6BC6h4380d&%bs=Q&pq%Zvjx_X|d%I#;A6va!WVC&+`%Sk-Qz<Sx)<HQfAA_
z&dD*Z{-I=GA&xu<ND0>XMFa(z)^PpaTf6@s9i-x}qQ3<M0e&*ke33Ol<wMrYB}~C6
zp~+qodf>Q%HAnvr0bSqb<F!QKF6lMC3+ZnRrrDGO36`VEH-Ot48MeEVe|}P{43-{w
z`3p<-G@Ov1?nAk3ihG1H%y^Fb-A5m2s`|piQ|*1zbHFJCzv@sI7OU~=xkSd|lnR$W
zs1>SA-;qeNYU%KRJ{?mJ9v)!r$W+6OtSoa(Llo(xLPAOQ*7gIwei4eje$Qz7exn(6
zFXpL5y+r6=rUGW|`~GzW9F=eSpW5HxFHZ{F!s`wFt>(9Y>FQ1XP-cS0E1f8*8uU8#
zw&WFcjxUIXc?IvV&>&mjNT|m*O7FT;S`X>3pKj^-oP?KqJ%&^a7spova@YYOA*gyD
z{{;SjNjhwtu(s6i;x+vQR<i0`R(5uW2JC}KF=Nj+`xEX6lBlFj$-Giz+v`&^jIP7O
zpx)lMqhbPXSWTv1baZ9!k00J%Xt`a1ej)Q17tRC@tT>SeXQl0Jv(pu8isYECF!ufR
z-aKe3>6{%pb;b_jOUzCudFUJN!)<IVLr~k?w5M6<yLu~g!-Hp!Nb*&Qr_gKEIVE@~
zmH!F}(YD8`{_6h*)vska)A0%MNQem6GvM#eh?+`B@QF`x7B@7)qSPZS6pNU?NsMqP
zqNVI{8vN;l1A-n4#!n(WlrfsJ^np5bBkk|s5V<G{_`&~jrWgBt%Fo}JqcyN{8o{%`
zToZ;J)8B6Bdg}U<;2(GA1~RD+3oYN^@j;GO?zM#-RsIX<Il=P(UL;Ln#);9<M?0+T
za#HD=;7?Z{rkf8c^2#pYmTYJw$xEn1c2%)!8GebCdMuf$&Jm^7&TN_b)YjG3IZw}c
z7nc&5lzK2)k82`Pe#B8AsCthn&Lt_~?V}z328BY1TybG&Wo%;|a38#EDqN4((Z)Zh
z5gV9z-%Ro{ZlXsQtDA>4DC8<_&v2Y5umL@HOJO74@TcKP;WtXXSZj;_PUbfiGQUXS
z)NE9cA&ZGFlon|MeJsm$%#}~6k<BJzTwhLHp(q;cSjq1rKP}48Z0Zu%D1g1g*5QX1
z={H9Zd5k2N%^WfmOC#^>XwO!#-sxDh_wc%{#M1L8tt>H(NO=ixQ6G1fDpXm=!%hVv
z3V@4C2w@6YYFn~@HEtJ=AZ~p1wvAj+kcofpZ{?%7`})-VyX<dePg|yDhd-)6p~%S}
zlq^j6qggr`1@nK#NjNq;umzJqJEe;4BMkrAp<iR(RqSB7)So^wCj0*Rie1~@Rd{*z
z&?&oJCq6GwZGO$?@8r0NFwg5KN`i|>s?Ly?5lkh%`VXIXmNK@t^NSiSW=gIdyziuD
z`+D!kX<n8unjvuhNRwCLSVuOhe>O)}dctjveRZEc!(|CutYfC2_*WWDWkvF6knkyD
zPY_pg6!?%c3*|?UR2F~67*#jv9*Bw4_Q(1}6S%&Wzv}qp4hU_27Vco8d6zwsOm;oF
zbau8I(&4BA+|ZH>vlj<Tv$QJl+C~Qw?ryK~XiUZ_zbNz@T<=f`lv$p~ypm2}(U~D0
zX+h(4-aKqy^wv6?cac6hzV-C7B6$IxL4|R4WA}$&*;A;YYPz`np{?(oFmkKq_MZmM
zjU-M=$@es|PV_wULV~!laL^~VzCXk5ytF56qG(k%31|V;IdY6We|5N1n?(!?nq$R_
z3$4UU3AkC%DNY%#Jg2K><ama+_N}c+fREk?d&n^+OZbeV^X}B3K6iryH_+52dGk?J
zMMd5Pb}^ICoAlO8oINlQcmQlo-9O0$yQpvTgyK{wbF83~;{)3_nNM?JNY~#8u1)$r
zEhL0Z<@<@jK@zNqOzlTv_=(b%vK9t}Wp!yR7$Tdm&eo3C>R$X?W4GAnUw#*WyEFxE
z%?L7>GJylo2L58%d!B*vxZv(abgJf*2KghVj=?;<;pR~}7OS5Y67edL#D0l4%`-GP
z1rE973g>Q{NuuhxZtn^9QL~L`rVS)dZ)MdDyaCVJ8dD>VMDx!T4xGNaGjneypqsy*
zlabC&N|zX)#+=xf#H6O6|NT{IYY|FwAqxBN9Hn$B_5fB6_EBc}xPyUn1`NFIjq#u(
zI~fL*B405$jpvXMe&ou*a=zZ@sk5IG9TV~Y;KemsW}KUnl3%|gkEaRa8Ad=42g26-
zNf~SY;2EYP)lXt6M36a&{B@|%p8lp+HE)!83fX_##UT|^r~Md46bg^X_U#+areh;d
zBcY?4!2D$c>MKj4L)A|yz5I%Ap-oKZCRBRQe-(x&TiBaRgK*5jNnFJl#2gD-B6#>^
z4kpa;s+7y$+nRfhjf~XL{Ck$8B|kB81r7I$logZ}6*=138L<&tl7y!H)OZUPSQpp%
zdh4YJ5i$jzqE?nHje12AT5Pj;%~WS9K_`AaTVc&!wViV}3LB-D_#bbhvGk3z)1DrH
zoqOs&&Fd##y!CT!4hI?TM=+{6bIxG^q|e4<+QK8w9Z4G{^~S&i@T_boU*NK8s+8Bv
zUw)A<!SBvZy8IhaS~>+(aQA@=nO^CKql5NIoZseIUxZzPLbiDTfRUiGQU$fE?jHr%
zs{**)P5<1cJmcBA%O^MJk4NnjEn4^RR!(lN8AtNpQUO=~vMU`bTP>6{oe>QqQR@H2
z0?ZDSo8<QoyeTAC`|Pbd<m#FF;thxVg;Lg2Uc{fSLp{?I-@>TiEjnW<zC?x3H+XwZ
z1--;9RPk}*YOJ~jZaQ6^oi=S(#zW}>z?Pyy!xJnQhB=1E738*ZeHI)`gA*V3Bh;c}
z`+|!_uWff{$D$bdr3iC)-y27WJeCmv_#hR=c>RvDm)+GyT9(HVQyU$hFvBJKUN-XF
zo$dZPcp;27;E?4fdRMdBlYY=PxU;&lD@zm2oUtE5q`;9*6<X^$LJ1X=U~YY#(eUe8
z)(Hu&SHt(z_~z8JL-FYWct>%W4(r$Ei9Zv9U$Seotc(4?1ngWz^P9dmH26Hg;SFZ&
z5a1ZCeAo*X&%WC@UgPVNbaO+T_-$UQcGL6qc~dpWnkMYs0w?wA4ZeW4i_-D59zQS_
z+G`Cr0{tq7NfxcAGcA|!24-d`P{9P}C$@d900oC)t03agK62-;PT$9zICm^VxWQ<T
ziHYttX~X2tmxgOWwHrZ9DJ+v*|2kzNsK!>89lygkB4yyiag$VnUWAd_xl&G!C{ALV
zbDin@h!CQQw!>r;d7h_I*b`2aKRWJOcl6=pgmvUyU}X~8a-2+=1h?1r%|W|-ql<iQ
zApxJar`%vH?oDoVCBV~@aN8rKKgT17+uYcV$l$&QnuYO+&xp=bo`cP*LY-wDD%mkN
zh2wSEZMZ0JXJBM__6(J6lpXaMD){}y{h9iQd^yN=w^K4#O2@>%Jj}_b<XLDCa%fNi
zhN@()l2)L<f;2Y%64_{7nw@v4*UdQy@%*HC(4Pn0(X>}Du<Zk^mvog8ZrMin$c0M;
zOl+kTn5F+cQ$~QlL%p7GMhjk?dFc0l=X2NA!v~`!W{km3RP7H-5kbyWAI%FVx?EeM
z_B~K93kwb&d9@4qc6|aJ)H%>tRI_A)4iWvBQ@CcB6!@xb9@?G&1`XVtj?R<2+c3t?
z?g98=Qa<;bE5mnS;~Xs?RA=9ldH-N#-LT^>_PO*v9yfbx61)Hyc7H&DHl$T?xDxqR
zyV<!^cG=Qf8jsccBJBzqZd9fv|HsYu#WK5&0tMV$&ki?7sA@H3bCvX-c@OCGmxyP3
zrCxm~Ra2(J#tar!IXOB)gr3HZ#JuOlmvK3BI=|ww7x#9HuGb>M^c63D$@oeX5f3Tx
zVy3HtK<;CFjfnQvptI4~^axT<f@;4ePsMq}RMCWaPe!ULdF)Hs($@Nw$+vnI&?dHR
z-io98m*HX?Aev3>n-on&OK_Iu^cQ$6Ybe@nZjQ*Nz=nJ-=GW1N-lFs~8g~zT<mM)x
z^U)t(5`G^^Ld56YB1{6q+wlG2D;=CjE1PT?Oa2mEn-zhn^O^4gB_T5J!_Rg<<rX{X
zmFxCq+zyvA6Jo0kD?b31HYCCQ2A85)6#o0jCPQwFcs7x|K5eX|{16*HHRiX1?vHFA
zWYAMYEoe{kJGN2Z2g(hba_T&gQphg||FNI2qCB!%4fAiz4+^JWHRo1Nr7K9FH}@@%
z;~D+HKG?m5=?p25$A$~ECdqw`AWmoe)rC6*uFe4lW&skYJB-MrZ6?WX<prsLl!|zJ
z`CY4KX)>4d&+i@ne^e}1Mrh9V=Z!;A!C7WAzDd;)GeU$A5T5?k7>iz$Pyy=ad<Nc6
zQ17al-F8sYip-<$IX>K-ty|+QCIHI^4o%xE5mY^7$l6hY)yp$*$Ye~Y1YUh+%zk+-
zEryTYGn0sfA?{Eo((S=zWs6sIr8VpA&Q0TU(gW4Z2|pE9EKR+5Rp&|+Ry{T2`KNA7
zq~AW>;D}+c{rFf#&1Pn#Bd}UjB$Z`fjbL;zNS#Upce*z=>B(tQH+J`MZ;|lmQ$#Ul
z`2aWi6C^DKDh$9$;7fAG)!zF7lVwK}E&ZQ^mtGEbwlsBN$Qws!a&{X0@7W|cAqeyA
z#Ln6BeLjo!pw8zg$pEQ|lVe*_);X#cfxt8xY!_P>Qc@02z4Z?4&k!V{)&4h3{t7il
zqO;iNe)_K>N5X_lWGgmSq=ykX-|EgrY(l!-{@fK$^+r;m^Z71Mmc2ul(IlzhZv<cG
zz1fs<@!U~hOKxZ#7yqsHc-E#&w^2!3Nh)8_%hNL}@I@NT#Ts~Ld!=m8mM_-sOb)1U
zdmb(Y{rg4qk_o(?$ETa2C|qYwh`hZ5qnQvG`WV}@&=x6M9n6+(?sYfvXR(FCL&E|6
z8aQX+?V$jHpuXOjswsrXee{>t2d^F8j>66$oo*47D0ed2$qIk6e)!uf?oMm)W!JyF
zk;9k-IimDy+aKSNwg}!wkPZM_CuS_iqvbBD_mMHs(-*H{I5lhJY<5mB5Qojp3*x@|
zs3{eVW97><rMO4B78WU*@O!5o0g0?y3!mHjn~MdX_6}w$aEb9)W>kaF4IDwH+I=2o
ziga3Bj{qt1fYq~{2LzN_+oxK|$plt`r{8=Q(R4B6(+bEH{jqTNQNM^Si$I4iM_9d=
zzX=uKb5KYTnx48=w_4-Ohgf25?d+c9FsQr4Gd|xxyy7F<O&#?{jknxi<h-TO{Mi@W
z>N=9uUBA*>)iLZ!8tOt%vk>z9F-;+>3XH-5>aiOawe1#gKoZ|^vsckSSG5)*f|cmi
zm<HqeqP=#pCee$k$0%bN1a2Klh2x%_g))Qx><K*+`(kP5Wm`9datG1{7%ACAG)nm$
z&UgQ1zn>r+=#frTGOM{d*?@hY2AI2KrURhl=Kl_OAAPF-pG8;zT7g4?<>`Djx>l{v
zF@f)h$Uy((bbei9YHyf(tIdKXD-8~{Ipli1IGxxA=lhz)<tD1s5w(D`;`Ex#M_}Qm
zK!mOsjt0kCj=OI6dPCF;7m+~~f&_DR@3!_Anqwm;66$_Y#Ntv$TgF;fqp}5uv42nu
z1!SYDPkk|D-q%~7et!Ggq6YkYEQw|Y`=Vb(|0c~d(FW>TN=&2hQ4pvM2r)4~?zrv&
zp<!Dkd;9h5z<H1J*UHQreX)aSe6U?yP;v3Q-!?pnPl$`dCY*_&OMT?bJwc+8PeX;s
zyl#(H`(kypw_9~a3!<rra52I4lNi0{;M`R9RjJhdz#l!qr>&)KLHt0swE`sD=EgJ|
z_F23d_iZ>q&w}S^(Q=Rx#A22O{C!9r+@py|z&Q|Y)P9aKjnXtvd;VO@rFQ8zI=u&+
zP<FtV-7haYsKT_s3xg|e0>-GWrlw{+Ra2<M#Or#ALG1mn1Q&u9zQKTDt==ppuLp<e
z;gVKg--C8Dmr-}v%(k|ijAu;BO^3(x)a4W4@e93nJtWup8S&bBcM9BF)-z4OY+b<l
za=2!y;cIOq$PWW5yar#KHW5PldG&9|kfuS^0C&Q7%_Qmi(#gEoc-Rj%%Q!wy-#TZT
zb}&c`3s)PPsdVCV$q%!aEJ#r<lb|ey1M(SH=-Y?3li>3!XvTtMUZmZQ&9@RVLqc@<
zBuQB#=09CivD*}2r|9f1;JvzM7Jc$#BzCy0Io?h%Q+}MT8VY<)G;^ysT`ES#+60^k
zGYxyb{Xt)F-zgo9!OY?Zjy>9RGP075NlC7kn<K!xYtUmO?P*OrAV6ltU_zS{)%i!$
zHf=giOG_OgMMcuV!H>2w6rV^XizjAyzvj!O@;91K)Jyicgvr+)E_}_;r-iTq@`o1#
zgb0(qQ>7E9up!J>nZ)Oj0_qc{oLF+LdZk`%;_zu`QUha#^W9ZUs2F{GVvjXTVzbp@
z!q{~F1^+?{L&+tVefQ=tN;mBJihX;yrsS@IV!2xt^LC<bsQU-$I4S8Hl8`5Ju=BNl
zTU*l~AK>wD6X#m)Z43-POEB3P^SQaXxf~FzO?3bIYX{`v6&3qjWIXqMO@1Ms@6Qpi
zTWdejBz^m_bNXZhZfRz%e_n=}9p-g?DsXfD94Kg<7LMW870$xA+=sufu6{Cb3wV+S
zyuptE&?cNzpb1VGQ}g(N6?jBqmn<aKh1xlPc?pE|@!t|ZRxF<~n7-+nx5=s2_AZr_
zc{eG@WDf_#e7yTY5}h&@{Z@LR&-};MG$v?mP-0+9OE=ZLxH;{1hw-i^wjbK$?BmkI
z!+jI2I}`?WXzt2HHx&&{_g}>p4sWe$VBh-22>$aCeMIq2k!mqvLd6r`u+!80J=+Ev
zMsNYc3mF-yJ|UFnc8TDQVk|Ddzd4uy;*ZR9hvoAgO^#+uqaFkF6#xEcWv7?*xZL)R
zY;IE2g(FdX+07tX+eMz56+)gS%9WDMpSYf_TBG)2n%4OFiDX_2bY)~x!9Gk7I;n}N
zMd#2>eM_fTa{gz}p()$&o;wz`X>IQlr_ig<*VDGP^FMfT5vjK~Sc5HEr*%o@(#YaM
zJ?8|Ti%=SOeLdn9#c!xZ7U``cee3RcE+aC!x0<6>UuAD!1Eu^ej|>m5b<$>*uHT<O
zIX)hurq*t>f3+bgr)rd&9UB{2?~>8x_4PT?zED1_VJ##?YWYX-0GGz+emf*ctTsz`
zfS2tXkdxUbB-mP8lgu5F3YD(2Ui5zG@bA%>R<vocJNkJ=D)7guP#Ep*HF^B+ZJc*h
zAQS4Wlu~SWfhfLck~jBHx3X2hz77{t@NERkFc|RgB0jw&%CIuiiq@>E{vu4`UvH{@
zdQPj_*k9)JcKH(iu9#F6IgliUY;fg{Z{Ieu{r#5s>`N3rLSNK$<4q|2MVs=tP(jLA
zq)o~7a`)t~hymIduRWe!#l3E#Zqb@Nm>+sQuJBn`VmznL3O{ZnK0=6Qkl@3u!+eI%
z3~xw+G(#+$ufPA3co70F%Yuo-#pvHTop9lD>>8gtKEVq)In}ONdAFOZeyX=tK^1Ga
z9gU+0`mO`6w;xoqNd<Um_-F(BY)wtaK{~$sa(y=$^}HLy9T9XbzNv_0Di|pRADQy4
zd`qtjspio*PP(Sr4O^sS3r_usN2m=VkUHVqzmp*Wo_s-pgzrUb;#)YFnJxTc<of^R
zY=k)|U6i>~T^$<gwDNHolo%KehG#x9%&Qv#B}R`uxJho7gvn9I-Kl>o=x@=U?O+lR
zG@2eQE2E&)MHCp>Il0inb=pkn8!vwF82dHujH~@(<7Yr9ur4T%2|~nK+nIbg8VMwu
z`0Ll#6)yw4UckpC_}XMr1?*5%X;Q{-O;$(XuD2-~R?Tj$3>(AB|M0gvWp5(=mmF5e
zJ08nN6&m>>dnc`i?a5CH-<1$EiJdEEEt(|Dd9i)*o~4`b{6V`wru9Z;LRBduPwjEY
ztG4+sV~|AC{`su}WK!%Ne=X<VprEzqzP9<{3-`Q@HCU9_=ibww)yF96a}MI3#h%Oo
z`+{o()B8#pREQyWt2`Cko(JOQ{$L=ZU98rEE_iL)^}8HyX>-KT*xvbk=`uR|dS`l?
z3;zp-JN*=m$>h>PxAgdgsf-O8#khW_iW;kp4NP6Vm;)al^k`9nl$i!nylGfu>5LV~
zb?8X#JcaRCHp+yMA)SWQ6{`FC)!z?BK{=lz;wZ(QN9Rc6g&4|xyKFIFGeCmS#g5HT
z*qy~~GLl}ulBL^U+HWzGxQ$p17tA3uQlc<hMxnF{6r>r~=f^!zlJUgi90Z4O>rmiS
zs>yy&^H)>6=8yYL<x<x<rMHU=^z5<P^V&hFX`)Ql^G)$~h@UHnA^aC;#dFp|?QNoi
z`yF>TYGxH%qVLf=ICsJQr)wEo-E6-{h$68!8~7CgX@scqr-c^@a1sEB9?IZ`tg51h
z9BNjWyehy{$)iWk(aQhPE*LR&>9mpMcHz2_w01U5_&`Y!D`{<QO+itPcy;5`cH0k1
zDP#AiU!pXut;=Md`uEky2`1@#kCPhvIljn07+_^Rw!0uE;r&cUumUR)J$BJ>;;fAS
zBC4_1@ja0)?v)a76{UWf21kxlSW`4(Z=bgr*_a)ZslsrzSgXgFGXV+Fj!6hfmG>e*
zF}U|GR^enlDacQKWC{#=so4fKttgTU67lh1)#?WeYLN*DIm5;W676hFj$8KYb3EP}
z%FIdxgfnm3E0CctBEkx`QFZ~TyOxlxlT+pIJ=fCoJVddNm8yiIj|Um7k0H<NPZJ@)
zk@&QxCUfMS$<Q}}kC%tW4F|c!>MjSf?}^9g#-k?vNI&t{to}R#%B=$4#1L^o&2+Vi
zn5Z~ialFZ3x`5~DX6+8_H-g_vk35IL95O{{V`Gz_)6=QU>prYC*5=mv@E17K82@m@
zR(mO1vO=tF9y^kNm8?un!I0P+-yiFO*#n=Orwbb$Iw<5+$WpNKwvTrGMu|a#LrcO4
z6eX$%r_^mT**4r>gXb2<I>U6?1)E{<_fLl4qNW{Yt0}W-N;}ovZ+t-m&!3GB>MVMs
z+^)9DHH(eP8`Y24EoWn65YodL=NjyuK&ckfs&#D$o^S1zemZmMVNrY;mhRz7i915{
zbE`HsgX4hmG39!InnYgs#zudliIE9x%L^PK7NC?<n!=S!O)H%nlid@cWj_JZ%;2&l
z#+0<UxYM2TcDWSZN8_KH^Zh%ENfwCPV_(EtS2tEX0z<?(;saNi2+AIhC72?Vy$<Uy
zRr&nH6wY@;$ZY2iwH#tqA&EB8@s^)y$kOGq;70>(>rW{XHlPX*THE{SO8n>`%QPa|
zDhge0u0$3W6_w!7>XA=X*}YSV+<!{4)#~$LRjSo;oF|)hZmcri#i*WKaF_WmJJ^49
zwy)a+=yE$8A49$ygWwpj(9mEaW$vYz2fTgk`F>2Xb9N+j+G}zm^FAHPlTE3&3S})X
z!c>WSxCFK|_ovF`1qEQk!%1Krn$o>HDSR&0Ps*-6uk4rg=gD?N-sxS|+s;i@CbH<x
z)i~{_>v<E9EvoFylY?b7=!v>S$)^#-2my14S)A58GkVEq(1Z{?<n8plT~iL*By<AG
zdO8<iT9np2Ca+kWu=v4FogB@{<@${|PJqmb%Moy$|5YEy2y8p<yL3Kdgw=fqMM`XP
zSt4+!C*|`r)wE2-$wK%(Rsh`Y9y{>H#(CcF-8C(DzlIZ~0Dn}5Z-Oqmw)#tb>?XIj
zcfWXDoE@Ap$4G+C<ex`L9D;V=nbpGpY>}7`j47zzvuMj<{~xB_GOWt3ix$2C0R;(3
zC8ebW0ZB!=k(QDM>F!jz8$`OhyOHkh4(XC^_!d6z`Oe|mmw#ll*S+?NIp-L2jHg~|
z2PgCMjpO52$RH?xB?1U4Of`QuiJROLqEUckODGYuO-|8mGObdB>)q|$-SeHk^?+y%
z*Q;h*pl0AHq}pWS==v|0+BCUhbON1_py}yh!G4L5%gv;3W58f2aSXaWIc$Eoy*aE$
z#*vC=Go9>L2HHbaT3(O^XqbGAegB^ReF+SZExiQ;am{Dz&RV|1;j@{w5@GP=P|%{|
zet3ghJF<eBEWLLXM%635g8Z0_s-6f*$#=hreY(DG56}0WGb2{~8y4b%1E-Qtpa?ao
zROZtsH8)Sn++E^qCnr`*9VkywbNs(di$C!I1!-kkgKVAQ@DJ7p35{oI`;!}lb1O3;
z1ntr#Y~<fW22r8&+q-H~%8mA1IT)Dxa{VC?=P0YIyKd~w@C%2aW4kHe$;lUgMNlG?
zX1CtIwuS!wJ)az!DZJMk=<ym`G@X5p$9;JgLRh255G}+=OFdI(I@fq}0plT*l>o$N
z!IxmdoAVeZ#O+&4=rS9QcZa6~iE%mx|D`%TFm3Q19kR79K$ZwhgG(d-4bae$N@QCI
z?<f;>$d;x;N34@MK==U*x1gp*lb6RpMTv^XQu2x)T-rB$5}}Q4o4zLw0>31bv={J2
z$I%DZ1grd6Y7*l_(8~-aXL-(LCAbfMjwH}+Qy^SkZ~R$5>hqSw@+=~AJ?#z$w`g~f
z(}-Q)SX=t5PR+xG`eFn5A?Js?>Vu(=XMNP<qxMUMDkT<LuK09s7(#%HbPYE-`?s!h
zpjP$*B@`RfkQ>Wf2c^35gS$5SVqBX{W*&2IAl(W?!77EU4?oR16T8$HUmWBs70r_8
zjXUSIf2uB8Fv?%sJX8vII&8kVcywJgk_qxH*?)=Qp8v+bYn3~<_ssri>>KN$T|)Ng
z8*#l`B*9wbg)@^uM~1$E@PnTUXBlek_7<8PpA8iz6x8Qit+1~mj4XvOP`9yP?Jvx&
z`<+A?37Qw@sNP2Tn|3Mm2wu=0I+r&Z*!j`M#C)sx1;R|)I``kPNA-?&??McePVES^
zz$j#6PBENMx56W1r+|@B2m!0npBx#xiNGq=jrXusIGn+C07))ZD!ScF&Zp?e1futI
zUB=%|OTahg<>fV=pPTG{<@hHj{Jr^NO}tfgMTPtux?)8|t}=pp`@?gP<e=fRRUA-0
zqEiWbZw#{ESuBIqkEkf`kK=kf0@}b~QnIqL2J?T32sv?DqQ6A{tV(E5^$yC&Hek88
zOG--W^!{;dRaEc1hA;e-gqAa*$$n;ke}#MTBy$(aqG->0`lQ{`Amk=J&#LyQVEVXq
zvnwx~t7oG-nPL1+#0ei7rKt9gy?BQKEuS|VC#zc+oVW|mS=HFX4<QzM9kFxynk2hX
zhP?ce9f*i&*FAXjhzBcvc^=D1TRth$sN^9}^KK>{bW4<7H$JYXf0<F{#CgiwA+U<x
zu(Lb(=7w8ViZ<Bv=sJGZ)C4azlLFgTP&+$4F6l~0lqj06wC;+bSCy5O6HTf9ltM2r
zFOPw!MpRtI`5X}ufFA|pj1AUQ>kIaqm|d^XfE3^qZS&V=Hy)_s=#<-K>oX>N5GdG?
zH8+U3Z*xKb?%o9z*DxA7^MIS)N^+13LI9Sk4m1ZSxW7YQo*)A%%1k&;s=g3_9cXsN
z2O_?b#;&140Y=_#T4QGz`+2wkuzQ*+R92j@P-#2`8u}Ull#;0;Xv5sii+6Rdii}HF
zq97t07sam099KC3Jk6okIE^JRX<y`{T`r2;MBTJx+_YXm4wFsOg*}YlCEe7)2&>lD
z4wS=Du8}*hIB&cxYi&czjR`14jkyP<+)cT_tv4RQ!CCf8TO?5wthrYbl#slV(k?Dr
zLF~81NUB;|MHfA##&6&nd&(N2{%7w=>+$~F>U-10sKk>eNv;%^Dxa4D3o}bPbiJm2
zhNTW+XRNwdEH$T|=y7QY`AYfy4PgH1Ugx|O0+u>np^pS9=I#uDP1k!Z`p~m#-_=Uc
z=qSw|C>AKL6xNaBeC>JTJU@4c;w?rZ6UbH5l*HJVKQdV}wvZpCuxL^obo@NHN<YQi
zJy&I9vXM|@aLR}k3nSB4360rbVjdpy!-ru2<X0!hFV(RkHW}_Jjdcxas}oy-p!mq=
zr@5A}us!xHzS`u2<!S~_g#cCCcd_`lmaThn@0^7MXV}<&koC#XMZ$WhF%%a{Nd_Z8
z(}V0~aH%*?#`rZWG}Zo<Oa&>!8xy#t^%H9pT>9EZ`11>|LX%a}Co*EQqs9C5CMLGa
zd_%qBxGB3L_hm8*oanl3wCa$ee~39|ej&5Y0yMRqpyMFa1k~u>ym*QpatzU|7)MS=
zM~+GvW0tuC^U&iU?a!~i7r|UtpWB`v*`9+Kk83PZ{S^b0GGYX3>>(zYqLktBH-(>`
zRzl!Q!~U0!u(Jq42!w{tyZ@dwC)^~aoU6qBL73QGt#`An{JZT7^}7C~ch1Lua_K1B
z6lGpn+gg)2>A-$yuHp>XF?x((D-Q~<{x9Zc_0B{z6uoSG-0PM7<*hlHK@e)9=I)n0
z_@kt^xtiO$(*YQeM;QUoeCvoFB}ij0eG|#fo)Dj=bkmDGwYY!=X>ku8Y$evI`W~^T
zP=Jfk2Fb=|(-glU^JcMn-AMCaej+hP;#Dvh@imS23o7&+(b&$+`sBC@<R7+uICmj7
z^W(YA_f3;vPlIa<-H3ed&mIB_k;_#tcAD?XYI_o!>uf>XJ1(Hr5UJUZjAANH#~6_q
zIaaOlG168*`vv6XQ&oOYICWQgFsG33r-KlzztH)D4<&}Gm#T`Zo6$UrB65%TD5+)?
zq>j&#-u%*wZN+MOD;<RfwTF^<Gd_Wo@ZPXk_zp>TOT8}hwPKxTXj&L3SG8^jgAoM>
z&xWo6oJW3sV5OPZY1)c%-rykD=)aLZ2P4HORwIFVzBOmRQ4puo8I(SK^tx;?ku#ac
zod`0cSOB81Sv{P!(pE6bXCa9N&b#o33_y0|IdpnBMcg$5@&}hhSHFE!mw8L~pV>Bt
zqCA~5tST=<`6UJ^%nCmaG9+{bq6XbFZ=-1^r~{N<Kj0Zj{Ayn>o{K51A}z)<F8nhc
zQN|;1n?uX}`l`EG{z^qqOW$ZXfN#7$>x6-Z+&x%nowwc4>2)X}S$H?8Omt;y_dp*q
z>NA-C<!cg?wN7V751%~y$9sAQ&8C`oJcTf0l!gr@Bj6~`gH5X>htc_BTF&+=R7_<p
zF-lbuRMJ0$J2nXAJ0?!~i)GP_2c*XdThpF=hh;UY6p>kPwTaGKkHgmcU~}GPB#QEE
z9=Kaur0~W45@?{K3V)-L5?-;O0w1NMRAKq;uH={`enTMptx@-{B%*~^HoirLWbbdQ
z*Au0z*J+fEyA-jQgQq=yTw0<dPd9E5t#(?#?-m+NrMW$;&hd5C`>%v(tV>o+xo4fm
zrA$FIcYJzU(W{|JdyWC4`BiOx_Tz(zi#)`u^Hu~R=h_lUSMsR5cmMR1M4-e;V^)86
z#t<e?fyRNYWC?@q3iaqQfEue)urySb5$yps?VX|c=ATzKC-Px5`=?74rKeY%?4Hh)
z<75dw`SO3b07;;NN7#e!F7mNAqqTVrRrr=i<z%m#BQhh}IM_{3Q~-i_#pJ~NmNwSg
z{gUv8-Ie)&;JsF-c7}y6Qjt8Kh*5iBK?e$y5Yz(r`6|b8MOxYlT&rD7cM=KroSwQE
zud>_IVmt(O=%3<^mKf_9Il3G~*X@$XErMp?(sUt!Q)lF4=#IAgHy5jSty|d`9A;(>
zYLq}wH}5ZQy#QNt^`#J&XS~MgO)TJc^up_B<EWn*GnoF<G$FR$QEp(%*!M&l_qpO5
ztZeCo;b=5MCccL}eniy4UtCo5JUaUG1?AaNAe`v``1W(5;!?fy#pz6WmsFAo?(r;W
zL@prOZ&oTM;<kaHKUS|2rK+87Kvnf6Y36gtsKTQDGRv(t&+#Mnab_8S5d%G~6)Htu
zN0+^M9T6D+4)QVyku+fRyH3%XiV+iMG7v@&qyYI+`hlLF-ujqlK<~PSOR9!7sOay8
zT>SsBGa~Rl1^%oBGv)O{LW)XC$sEq-!{KBCi5iLzSQGiNm8<g9Sd&JN;3l8lAe*(0
zso?v8-`s11VX~Pa!HK00%b(VIu`srQP=4dhd8L>a<%$G_R~RAlPkd+Q-3<J;h4}sb
z!_|unnrhytScZOd#aLrj6@S^)^}XE5yu+yZ#3{o;p9i*0g&2Vfqr+Ys%=`}Q<Gj7&
z`AIsnQu*ahT_s~(4Ix*V*MqTnx>Mq2-gVv<aSf*d6<?DY{}^3C&EFA;34czTX_r<A
zq)&v2>S~)6NXhjpo5?hLsgAdi4#z76TI<m7QT&h*9&uqjsy|e8JzWt-@k312_psJ^
z!Pi{}s@9zeiO(uszdHih&UCK&;bdqcVpSF|cw>K!mBnd#zSf$LZg++CrP~kiwc{K8
z?@zZRBqkwj?meyAp&%W9`40-%b+Vl-FMsw?W2sZ(&DIjY?e6aiU!nn05q{*_xC9Qy
zJ6Zp7a5pFP?m6E&_u@sXq&F%5!fHpQ3&+ki9w&`1kd^tQ#Kj#Ii@?FjHKJ}Ma20{K
zwSy6r-d!wD<Eq9kRy#UYC@D5VM*l1!hnFN$*f)eBI_fwdo`<KdTDuA(x4jL{7h7Ds
zVdkr2nu}gA4gXIza!nB<Wuvv~pS({=TE1%c0ZUW7X8Qa4yPIfi>;M(FQ*w;$+pCiS
zf|uawGRDfR>=<$9Jz+`4`?DzfYyM>_^=Z}dj88nCQR!QKp?hs8cgeTrN;TL<mcAjH
zi^1ga^9#ZqpB|gcm#;FY&~``YcL7)rBv$C#R0a`g5r@*>F!sy6P6S9(_EVAD?0ok7
zNlXe+Oli$oRFI0SU&8)w`3*<R<7BeS3qmIn>$e65&6EGnH8iJ*&R5*PEpt!!VJR5Z
z{?Y{$)g!$HYEx8Vr~r{J9zk2d+NHQ9g<E=hyL$rc6Sf1iz|UmSqJ0@Oxjk~+IA~+5
zTMB8PX>Lxp6XT-QH<3PFCUSsJnrlr+yi@sm@S{|x6XJbvyR%Fn<jcu@4E|0W+QT^2
z%?5QzWCm?D<6ap|VPUZ{29=Y&%|)M~EXy?y9;P*K_xG+$ti!@w0$7zbg6HQ*m6JZk
z(DE|ExPD?_ls&vm($w?;aHJ+bJ#HUL&A~ypqZ$IDSU5<Y#*Pd`#9APSS<Bp&z=3U9
ze4%Q|!v%hgG12qU?YBZSa$_LgVE4=arUcLtwm8=2)l9twTBYpSx<R`d{m(aXqgrBP
z9UF(Mh!DvEP!W;SeqluZA8PxD`akOD00A1`z%l?@$vt}He*H<N0iV$w@R+~x<+%B2
z*<%Ov+I;#&+jggabXNUC$i;O{703G%#fTZs(z&)Hb{`U#)humPFYOOH#0W+BWz^5t
zCWfyOzA3BZz{+w}S;y3feIm8Bv+9eEzG1Hs<DH+R7bucB{+aObWSb1uC+8F49)?<K
zj>-Bo62zPTg2pS+ArV#$4!PaC&h|or66Xcq1|YvBTm-1#F)|MAB~Y<&E=+!ZWS`oC
z{dL#s3bv@7-?R{0P#1M%H`2OQ+TZN@o3p$yq-a4X<8~FCl^ER~PoPXERxDmw;9)}~
zBG7JEn)}#Gl-Ym;S^aqhL18Em+96>9Jq8ZEE-QNHnnz3&Df-JplRz-Jukbl-rt3c5
zcMtXl?afa7kNLWc$9rxJ3QDb{q~Jv(0Kn|+!g(-d9NSglk}Jrz&$v{>Cp>6TWG`E>
zzVlX>7n<^e6{Y(C@<ms?Xs7#DcvZw7;LlIMbj{2#&cDYY(KWZ3>&6QL$HOD`4AxTH
zcIi`A_@;vXer2LS-<y_{SaD5{2S7NAaXS_Ti0tRjVTEnrg7q@;=HR;VzUd_O+6yzo
zS5_NBO~q-wmBNbEUbPhDKDC-D)!~9F&cBT{@prAX`8K}5?pU!fy7}R9efnwupAd}e
zl;9ski6D<CEiDbb0lA><YjCw-ROV>3((h=6J7$iKy7{i|6zH3CaDiXBMGopuXL>2%
zCWZWYQvxTGZ{!cm4C?&@c%n(YvnCecdVf|0D5o5U4e1)K8C;iFS4|_k!L`;K-oCyw
zG;fgw8(&06xfmR3xC4UTG$)R{uTkdcRF^?_$|o+n{W4oR@abh0+Ld*zS{kMXnKZc+
zuXw!nA}6<6$kpNzz_DPeHHfPGT}?Uh=+d%ttxI*S+O^c>3Y05mIMY=2=b9O*ucUnK
zOqsN%Y^+X>fAI4`%PxF;01?9<i-JqMzbeynFVHYg>{dyd)eoBFh?H;Uv(uA`t6(XU
z)D6#=9?KmFZp9ZI)rNw}21GN^muTv-2cL;!XDu?gEwBBNp`>5Q_Yn5Qz|Nt&DQSb~
zU$rcKAJBZ;hka|7w2-5V7t9y@jZnGhjeQzNSq^Z}jx^i{MUM)Qw?WH?t4a^=;Yj2^
zl4JDvAL#DBK`;B@e2_~D*7Wt`+4g_Adgrx>9i?r;*QfjW6GlWlE;g%w{1x&Qb6J03
z(<<40TXH+y8mS5yd}1p(KqdYW#-mu##dP;eSmwrn`$hTCM`;muOhTUa$<hYh|M#R;
zl{Hg)taSo@EFb}FnA>E%hJzpxJ%vn&{Pm}%JR(aPuwe`*Fmhs`mdFHEfva=j*t7h?
zrut^guCt^P3<v<}3nN^+mS|weTXBrBjMcVJ!8!~Rd>sCLqtn@PDjf7Wwr|U(4qH`G
zUa3!)Ia4-w;+v>0MxvBK>6hn-zSmW4kDM$7$jjUI8S*ak&VKa`8o?K84EnESx_`Q5
z?0y=!TqN2f1qAv&$72CD#j0ID=mfK65K#zfmK(jNitpy~u!kmB8dyL*05&&D5azqQ
ziUPfrr3IUY*>NkZ<#%4nvBb29Plbj}!fzEBoa=JXHE7<w;s&c7jQVLKb2qSco-q2e
zCBtQRFk5Pk8$V!|<IXog9xN0e9UXmjX(^+X>M?|i1P{`>oRYUEM{5NIbir9N_z?|@
zxpMOISkNDAtfx<4J#W5BNvJhAhyt&9AP*LaSw9uu=%#!Us3v1NEMlTY5d3@VOrO#7
z7zPeu_+?Osc+ATvG&~l+PECnlEVYAw4;CeZmufK~#p17~E|Wt`$(@=B9In@wM{A(3
zb;Zi@IpsSb#Co#cm+8cT3<qk5r&}{^-+$ZKS_e9j?Ev`;#W&HN6W8av!a8rY*H)qB
zor!$)dRuGFw|94UOU@6j_kre7-XTO>Y<BbP053y`+}_!l1%=^>kHu^?wh9W^!M{_O
zUdCc#8neu5Ox|J;5e@2>a}W#y{u<ER%Xt`-NaP%t8j5XOcRM}m(k^XZwySk?tjo%a
z?lo5`bZV3XR>oC<fxE5n$j@nRu5K_fbzn3#ZCET*53JI1QSgltcku`aBy`?(O8-pJ
zcsY|P{?+Y1Php8|xzVUpjUWNDlYpH?o4@rZ+;wf?Y_Fu^U)ysYu8n2?>Pgf6hmJ<-
zE~$pfmBlC>{2RA(g{hk2y1`3oqu<AWxA`09?ld+<+ZyJo)4q|ZsoZ?7d|o)|o0U0H
zDl}~Ue7KBvZt*wuy!YlIv<6oviRz>yB$E8lSi6;9WSO6ib3fwTv&yh0OGV7Xbv~S1
z{lI;8o@~%X(C+nkCO)q137@q+m6(JClUg;i<0v(Tp`IR8BqSq=n_s`OJuTGK)KB}S
z3h(aLGR4ot%_tX|;%uDTu2Bqzk}pmin~D^dnq2J;XCk9)REu)-nUf4?lK*VLwg6Ly
zQ+%1Jakct$M(C7yxOi|HNIXE$S0bYQ*{Z@qsnv=6eD^EC3uY?DL*HaKPRrZr<|%Uw
zGNdze9+3RHBBG)*W?DJQ9w%n};c9-kcW#Y)SGllSz2v5T+><RpnO>5id)FBZO8>QP
zYcWdrp`2?}BIKi`I$dfa664V#*a3g)?IRBt7fQ6+;@)k-!rh`Uybvds;c=c9m8AFz
znDOnj0G6sP(>uvBl*`@+@$Gp+?JtAv=t}on%azuv`1g^gOD>6k<5HwfHb(rx>6{^l
z&18m@uK3b;NCO8o@ATvh5GlU(#T1@wH0rpV_eRs!0^NwW@658?90jI|pc4DTCJpzy
ziL#rNY$Mf*!CP%R1Z-G!IPkLZZglx7k}JLQ_63ynyF$pv=<AXJVod}g-?<McS(fNO
zX5>8woy+Mldn4(>h-id$mTN*XTAvi6yE{5&#`@~Pg99}+z9d%|f3VYM&_l|NKhfj~
zR8s6cwW<ZXlKpStSD@epwM_N(t#tG>pg!-cVbZBGng2Vq-WL~2{IS>Jvduo<;oxAV
zwic5rXsUR|y@eMVhHkCb1ByOVD*eI53WE-^GZhs5ZPp*Y)@O~5|BDZIgXo!#Fd&=5
zkv54$U^dV>@YC`x*}Ye+?_OLokq{T4%*;e?Cj;*hSeI05!bGO&Rm(CeGERRtIwv!#
z3L(g9iT0#?hYR-2gK1fBZ|_KU;dGfUd61PO{^)W!lbsC9YiT$W(Gj@wJSk=5Te>f5
zZP-tdY}r`XbLG+Cc)d)$nK)(qsp8GL+K)Yb#2-v@j(D3dJ~}+ydm=8m-bf0ms2GZF
zZ(~xy-rZI}$hZymfK@qTy>aSRK2<xGEgleHZ#stnZKDxSu%+#%103k`@5qDeaStHf
ziHVCbCh*{6y|h^Vn+iI&`II(6%$RF`fZo-GN}QXQGk&+dP|XNc7|*e-FE-ww5xuN#
zX8k|}aXWB(X71j?!ou)=mPv|^)?0D*262*Oko0B7QYyRny@-OWDq`TUwR@53RH<Ax
z@LJrl;p5wJzexVGo_TNoV8hoMtKm8U{!28H@_m^SD8xQxCED0i3A2J}9^@>W-TXdj
z`0yRB{dF#Q)}`ffuv{Nz8-e=uD_|#nadPzR@5Rk!dT8X~yru42vn5eb$re=_o0^;J
z>cfN>cCf+mZL(0O;1J88;vpWJgjQK^bHxl*tXl075ar0{Mn*)yq&Q!g>E{W0`I^s{
zkT5fk4s_J2&sDkKUu_PP@I#t?8<N)(vT}!}A)$yK;h<N3qBr`Whyh4qYA|TrUnDg{
zDPBVwM`uWDoyY-!Yx@1{py*v+WqNmatLSnzaqD;j0m>M^e%8{?jqCmcI7zbPa;NTV
z&nIv-Y_>;d20CWC=4ct%&J|@N0x`MXQEi6W8lxPlezCb8&jn58i0}w?^>xpICj+=O
zpx60|K@P_4%bz?4lenfp{Vt?_&dVROEBF;06c5;LuXsOM9{Hf`Hh(QmGZpcx&ld-L
z%FXspn3hmrlB6cfH?B3j9U>)t@rqoLd;E?BLFv-L3AUx`(i9=M?@s{eScc=B?~g;s
zzJ<H>_Y`P^@%eXHqR7drTwld%7{M6bSpWT-IuwJ)(DNALvcBt|sl;N%E4=z-_arlS
zecFgDl2+HsTb{Ds`TbeEd2`|iS57Dh|J??VUo81^%tl1y>H8eUqkL22=h1BQ0|b=h
zj~unMwY7P9Tgs0&Ya0NE5Ec+j;C33O(tTgGpANO~TF*2}?a$#ywRd#D@;WLgxP3}3
zvN(wufOrwU)Sr7lnE^IGh-Zgr5Wo)E%o2oB?;EZ+aq&Wf;UJW6hvZaKS5KjmjaXLb
z?@<@AJ(%<5sL<*uXW|AlsNU{?nVr?H|0iIb?#2R4$mU@B1;hdYOWcv<Y!*i#DUGhE
zxbv}C3iKhAzSBMv&l>cjn4PKB@udibATVEu9>79ys-ebhb;<c&Hg5s~vJ(at&)R<f
z4)1vaMW&xT&eBKWUu&5^9mMbV0`GrM{wCFZL#baOLaH_V?XW-5`E+<CsG=h6#}Du?
z*!V-q??XlHhhYhLY-cC6K<H?+ZNXC?la`g$8DTVgZNh(az9n<b*aLB`R!>$Ddhz83
zEj3(imFn0ZE*9n#jBXW5Km675aU1yH>SVrD`;LNDfhLlp><eUZwCd>o-~k+q?@s%M
z>eP~CGn*{e5TMgtr3-C4*!wXiAJAi_pXcrE{VJ*g#j%cAUDUEPv>Mma<*Zi^+B)mR
zCn{9MA1D6kc3T^Wn`w3B{&073y<5DS9I!LvIbb)+_-y|#i{KYo6_rggT|GVQm?#Kv
zVqEpCye++@2)STe?kZSpB~haJMxxPzL)eN4um}>I9-2a94FEQcy#7gA0|ys-J4uO3
zzyGCY#AKoB)y*|`Pv67lY-Ofqc+O%YObQ<zr`=xnmQ1?sI!Fv)c@N(&zWo}32GtOR
zRHS%lC|-OSi6N7Y(%3AsL0}DLbd-%ohMERHXlJ%}{N@0CV%y^zc9zV(J=8Hcu_A#q
zSkWeR)@BMt8A2EnS(|5R_#{QenaJ<N)^AO;KE+LPip#-=Z?}w}nk`VNC%pK*zq+w}
zp0d+~+c%fEJ!rZ7q!uw!R9CGBm871JG+c!X@1D$WSbXL0N*iE;HgM0~o&a%6X=!zi
z&^{ET=LBpfTU5#S8vQxp)VaC7p#+nM`H$)Dz;kQ6^FT)Qn3x#ILq(mR^8L9dlOUtN
z4nppW_oEP<N=1$Flo^8IYY4aw5U_mv{K>Mk2U$=-;hOOV?d20l#RGJuthD)vl8=r*
z9QfLAAL}PC>b~ab?O#Rq;${HgrxMVWbYL^<YbY}~`LVFDFd%QeX@w$YGsdk~wweEz
z+c5$eN)_@vhEn)_Lxy&^nYw2jcWjZK1sUYui)Ic@HNP9|3a8dZNK8sPVkzg)oqqc+
zGzI<*ZcMdN$2)5c&{h@)EwAhSjMm{O^Mk9MZ?{<(9{#-6in`DBp{EaoR!o@DABbY-
z=t)-<t&S(@kLuhk^z}CeGUmX3!8JanVTg@a{rtK_)6E{4iB+qU@WaQ1F2K`Ri1?+Y
z#t>Eknj!8f<S?=_K?_?;nnyicWS5l$Q+@c+3W0Ve-$Kdf{6dQ4Oqm~8Wo3*|3rdq&
z(5Ctwr<U5mL23-@CSVzP8I_waxj_!JPC3=NAAnU~<q^B<MnQ7u1-Q`uRYbadEX)e-
zR4%&eO8or53rQwges)?fo>~_pJc}vxes%H1MxwOHXKE~~+|$lVW6%Z(54kyBb<2=N
z(^R^#(zk$P2wVZ%@;lk$v0vGpt8!AprlUN*@JMF5r8X(yqq%DkLp3%ahydC!GRVuj
z`vWOs^|paeJj8Ja;docq`@j^gtZeS88sv~i2)a-7{F%)+Oh$ViMU4fZ2|Ko;4m*8J
zMAGo?Wvm&A*;C~?5UMkn6(P9K`glRSfExOEBw2MJN2cUfF2(ZY@)#GhGOdpGjEo4R
zMq)yK!AVm=AS8fVW~>eRl7i#s>R5@OFQ`FBYg9iQ1ky(!XxYaTR7|LPo7*<)A2}V#
zRK3+1`T@LQt++cH5daMW*zN0lm`{~}`DjtLc4#U41oi~&v6P~`+8Mk{?m@!u^D}qt
zU<YHSayNK)DxUsd$?4_B-rm#?d)y3is%&zqdX9SbU%k=ev0SlA282Kac}YJPOH4qZ
z{Xo55J6jsjLpV+u7__Fld4sz^pWE#@gos9<c6Bz@=zOtZ$<dzmq;j1Uaf{wi*hFZg
zR{SOO{N>xvN_+|X=D)S=py8bvh%@!`mnB4&PVKi4FWn3L<OsiH98RwI?k!dy%o{_t
zx9jJo(R^APp%kM1KH)#F9Yx78AkC>wM=Gg=_LLD82q-+mfnH)47j8%*iMuB?rhw_l
z)3_u5=FmcXd3O(Ua82P2NkZ4Vk35P>?}GY3yYnd{03e^OS^62#Az2Xq)KV)FFwJ7Z
zM_#j{Pe6W*YT{&6IuRkpWX^j@OKVH&wRn6tK*oB?%F0C~H?sPFN`rZVV=B9UHdfUb
zG=I(b|LD<WCLq9@zX$4IhnJ-~t-Za(?w4A>TfNN~q$AmLFJxg_!h3>3*ls_Zd^O7J
z$pRjW^?Gx_v)3?MAq`9()MC6p3&Y#|!L{k`pdJ|F_F+$midMs<Xe>3ww0Anj?B<p@
zRQ!5%4|9(;`fgNW-`4^Zt+Q=!@9>Aql8cMQw6qI3A)rl1j>=qK2F~4Sk4}6>9v;^}
zbLRM1A4Gt04GauSOm&+syF?tPp7?DQy=pa`avAxG&5x8c0!V-0m%<!H>#-HItUY)T
zr#>=AuQnvrob{WcKHYv5fIxt0jE5ALQ3O+%OAUWrp+cWIl0Wu%A_jlwsE?s|{rW8x
z)tv5ZNRKYUlP7FhmN=26+8qJj#4fLaO6S%3X0cMi>t5^tBWQU1+w&a_&^}j%ec(s0
zT?rwg?m34Yg*`{dB6tq*ENux=L|P*w-On}#ek)4WTifH~o0aX+trUukj+-gT9s{<K
z$)Je`bUAe5K6PMZ_+I_c0s~p-GNa&>8#jaya#=0R10L?<eFr);3g>FPQHeQEx2cl1
zVxwYWczXH}J-(VZq|#=LCWS=HO*NhL>?xO5x10C%I$Et=+iG9<4_A?dhlf+rg0Dcf
zmH+~6<ilIO4z^!DY)L4MU1F${_CK<RQ+<`}A%}ql3oS9|B@s)*r{9mIqx07EU}3Yb
zZh$@LAm9~e1UMz3Y(xV%*pFq$=bN2Xbh^ycZ{BqGbtwU|0%hd(`WikK*7#*2gZ@z5
z?)nN0)O>dxFK}!_r1zN@AX?EA{;>?(Mq+wpSBHbeg^7xV4xeD|p%Sb0_t0>qF?s15
z{6ryFG8!ZZm~=dY+!hxr>4{7{q|)tnC$S;3#kz?UD(Bw?ape#%TnX@7zB>ji#KcFu
z7N~WMUDm%;Ewf>Q2)JDwE)EtbX}`ggm6iQko%G_$&E>LdC|Ubmv_A)dmLbz?8+DS{
z6PxtVG+;0BD<CLnZ@7pax~*|K0}1W#z1Q`C3;VuiYb;yZ(-RPAbXE06%VZ66kBEaO
z|EltM{TfeDF0X_DVyCLu=XRQzsTcU+_U~k221mL6dutQ0yuv<!_*B4*VGuZmr|xrk
zRhZ(=Ht~b%su__nnM`*E$aY{Mz&vGbE+RqlMIruer3tlMKEt%%n@&>tYyp&Gy8Zlu
zd_!UHj1u=In>E})0hRi@LEFJo*PAm$55BbdvQKu>jC*rvU|JfKAG(;wYpt_qEG;36
zt?qDa4L4;Oir)VI+l@JMDv6(<Le0I1+t>W<Ry_!KjtZNf?mdG_%k?D_IYA<RLXnxN
zT}u9*@$hupi$Xm1`uef}oO)R`tiS6e<2kH1U9TpTLWy{2iVNzj)_b!|kDwBeL9h`G
zHT&76jLkKk<3R<g)eQ9XK>6O7MJ2e?0-+a>PGFWVdm{0MBfhiP?%3oAnn|Go_1}5V
zA3uJr);=CVo-e3y&bhD*ZhaZscz(;2t6cX`jd&ucON$*Jodv&or{Cs-0xC$65oQ`v
z67wZm{QW<fn%vx)i#3+2OuI*MvvL26VZb_6!(#U5tGd_ln4Ycy#msdatb!}Ofi%HD
zWeK0i$Tg6dWOsEc`sL{lBnr!u!2mmHPHGMguJ=>ATzA(OOBJ@Xmh&3sBXux$*F|5h
z&c--hcbRsK0H-q_D4FMEkn`}6u~ewT5lJo`8czkZuYz)k%SjvB((PZ*qWQiyvX%Z<
zY;(Ec`|$!Kky44AprUe0))lhPQ{sO3{jSq?r}D(ywGhnddS7hu_5F>)l6x^N!8@nZ
zDGhhi&(=V&$9mC`>gr%&aAl>g_2k(z*HgDBK<E5Yrkg7!2M2+37^iEYu<B;9$rcl$
zlz3^Lxe0)^aTf==MG)SvSv|rou#^H}rh5z3A1hqz(^exesic$G9Y>exj#$@T;?oO0
zW=-46>|DBfy3E^sjLlD=Qi<z1bLlP@$prSJQEy~0^<2zxy02{m{c3DhT7TeD#o%pC
z6@RqkB?(V+YRsiQoJ`M#i$1TAdT<IV-vvtFQ+^iEbgi+HD^WKKjdz*zRFRHvsDPI(
z2^>4~OLel==9}~QAmPy2O4GT^b7N4OJwKeY2e2PdJ_QUz&d-Lrkb!WQAA>qh!^vRo
z8kO7C)n-<GeSDz!=>8liUf^I52xQB~+4$DJGq`Uw2nDti)6>&DAOUwhkp%CY)mJN@
zk<#f3Aqnx{?yPitB&U|g>vu8%o)#sb0jY9TvE8@nRF>I(%URcEdevNbA19S*1QYZy
zoI4)+@zUZ-iI#Eu#P9!b0Za%f!8>C@<TqnPyEnLbcz*AVYIgH+*_kA>+9zfE3LivL
zRfVkUMa0MA5@5gz!L8!~%h;><wKY}CT4433pNB%gBEk>}gmni3s)f3({?{y}tNI_*
z0WAZPzj4>(q|xNq80G^k88RsLNF}kmt!)gnw(^yJ>DMoBY-n(@IlMogBa8nw=+*f9
zK?6W%|KQ@<(j53WZoQoVY>iE99HI|a4yQRIRjApSa)2#)z6VoYFO5C{4ue|vWPkq_
za0a3Cp=|7|a_jX*Q3Eei&=3ax5=%cohT<fIOd9hhzI`8`CRBd`j2oibd{Kjm9&_bI
zG}Pyh)mK&FaXv$q*@FJAWS47y`P8lX+%e!kVw|rmC-=_sw#F<xd}M_G7fVLlwzHts
zQIN&hTpOX(QoP33KWhgZoHjq=aCQoGTtt&f&>^)~Aqo@BYsR*w8?m??vae<$fXJt`
z$yBYbEMXsBM)U&L!{v|AdMXH=9vzVZwII(Z4l`Jp6thC#s)Ht1fL=%TW$R9W;&Z8J
zv4WgjRQ;O(`B7Q;W}*D>-1T*_7ts<cwC&Mixw3Gq5Z>#FFPC08e=RkuwAqmrFcgAk
zJy%qUA`mznMHMjY>uCK2K>y-sj}3zNLDx~$Z7`rxgJ}Vsd(dO37@Du-UT=hL*`G3E
zg5OV@F1I<T92Zv=k|yYEa4=8LWryBn60;UgCY&&{H%lY;CW_w|?Y*stW7@4=b#?ih
zVMYhP;_}d*ax<HpZl9;cZ`P1-CYc=l4=CahR&k3b%p(P1X2!Uc#c1shK{I6kd$;=|
z??(7dnlc#34+m<AMu?`l%QBKlQ0eohyxMr(tdgWbPZffkFAnos_&m(vjIWVEeCf;M
zfWfGpSeJ>g&W(T9Bg;la%0)yjB&ZV*5<aTJO#WGICnDlG1?|0yNl8^9@PMe)>d0LL
z*QYDy$F7>{WdUfE3E>pBL5$DVto(H-vD<0dz}Mq*R8vQXANOSB`-m;u0pKO3dc7s0
z3hJ961iB`o@X9D*8m(IOTHL6Vd;blAj~?mZ(-B4s_U{))2nrLVxQLsxgB$+GyZx_}
zWxHA;^VA?vcc(_jon-qlG&KWSl$JV2)+7ewwpzZ?%%mNS)M=ig0onlkDxY*Vo4we9
zm$TQV|FJlE2La`UmZOlcSr@FOo1^Xe-xe*$wE6Ke`Ai`N2F4-WO^(km<FGk0SDF3s
z<jFt7<P^N$C<=VH$y%MFCfcM%VZ1rXrlC09w~NwLF+Mbt%tz_S?o_hPFo-$`h<lTD
zI&c;LLm^F9h_G@T`x96}yFdA8%s~8^I9;my1nskjN1oAcnUe;dwH+hO?GsR7WE=mF
zx+GWyz|nE;Z_g2q1S4fr-@YmJI>=MP_z$Z5hnfDHo-n~I$%HQk&mM!N<bSCi_}WZB
zy?M;NN9Sl~o)Dj;&vc=xa1i5hXMO<S<G-~y(+~7ePBm2Q^9sCmtvVjc?U%G9)lH$6
zC0@SPMP|J|bb_Y`-hN0ElZ}<yGP`Dkl_84FMwNpqs@TE{3dl9^f(1X86`xqX)%Dy%
z(V`$<BE}tKAqU+PtvAjNvys3j+i@s9{-?9mA<v5!_8<V#*3qD(qEu2;3h(*D(Ai0p
zl!OUl22RsDJox5!kHk!o$-f2?|5n|&<NkwbGf;3I-yTDQNBH1+pVm+p06*NVyRosU
z(d;<(@pVhb%y!RO*D)%ls08p&1wXF*o`3a4ML{|^Qugeb?Pn8@^}d+>g=*YG^icE&
zjN1Y9f6v|XtJKHxPljBAZ%3sdH`RQPgq0O@WCtw7&(jEM@)!HNhNB%BsijQk6Zz!^
zgNjq5<saP$IE+s3+z(J+vVUfHLE3}I;l%Rn^?_bNetdqIzG~I2!`U^;dXgt><ye!e
zb7&I#r>LnSW!fdc?ht#{_1`KEE65F#*L6ps(;$ECh=O0GfA)*1Nv9?oGZA!8qkX)+
zyxP>!XbR9_VcNX90kBjby_i?nDsZ@2!LsN_B_T&~Fn#R#_d?%kIDa^wTkw0j^ng13
zBIr5FfzZ6IoeiwVX+Z(-G@1d*NU~o*OtVh?+AB<V#qlu>0afdhG?`q!7$ZZ&+8TDM
zOmaA0(7+L-&4=XCV<Jl?_lFQ(_8X`Q4oD1@l$NV|Qh_$+47mRd9UWs!k`Gp<q_ir=
zoE%u5ARQ0%ORhh@gWSW@7qy2A$t#Sxjp>o%0Y2-}867+NS2zqr!;81SOLb@w#!kfT
zfKG_nCg*rMEwjd;uEM5YI>S0KQLL?_qoW}i&#XWR^Oy(RuKXnA6Z!DPWU}uzHc3B;
zhI|p<SYJ_}we)94Wyf9r%?t4{TIPT)?{<D|L3_-LnpSr0v%Ne5+Qg#v06I2g=+y9f
zB6uQZ7N>TCm?ID`9uHG|ss5#%lAJ>M==LwwyAjzr(n{{Tk*=Q{3!6E4IX+oFoOtZm
z6FAWm(Wh?@42#W&zG<t8G`#K8!x4<;Yro|_^!{G=Nm)JDH&<BJFoI3NcN*)6F8PPS
zpD8TJqqfRBF^Qh)En9tVjLcNJIGU6=PJAYxZ0U&U2TQVgnl81zJ-vHz9Tc6>><X;$
zKk?t}N+&BF75?@vwDVNGp5$F9$*7tZX{+)33flZwB*BxJXu^bv@G_4vPh`JpVsuoJ
zRhnPuMu9NIm$+WB<BtE~p>{wp*)rbA%@U)}q5h`J%6)lhAvz>Ty-w}OdbN&{xnM&`
z?B;S^N~tQWpcE}pRFMBB&%A&#{98+rZ_Ilqj7T2kfzm*JL)Y5iDg`lxy!0Z_aauW6
zZq2$!lw{krkn>(4+9*p;j5OCOB2qaH9!d}ZX~t_Cr^4D=$G&TzFU?e?xHMg}wOj0e
z9Dv<-W@9^xO%M^8v$X}<IjyHkU_uDECU*kFW9VX3K7-}(dDo&}$`U&^3I_)?`9#iS
zq>+}U>v>_p1P<)nT-0z<=0vV^1bsfxs-Ic0UJDTSXD&T;(*Bz@UEbc7JywMFlt`n?
z#H=6{NeaQovLo|ANXSI9E0(1=Ckr;1?lToXA?AE=kVpFl=(5u`#mtGEwl%?k_WaS5
zCrsz>(NgD4)og7R6%=4vifU4(%8T;za?b17i>myq8QZh2U*abdS{ofCI&w>o_kB(l
zkX?C#03CGtY<4EBn2C<NI2zpNuWDmsluWN?$lGjw5U~E}Vsn95BwYLfyEpzpniR(4
z2_n#?4W9WKb3p_p#jeBgx^RReO37c;;6Lmgt*;lTT2fFz%Vs7Hix&w`$iQL0qCfO2
zIw=|+VpCMFHd~+s8{{VeA;AaycRnX?_XY_uT7baEgdxj^LQ!&NWd+dqa~YNjx}9Ux
zqaLRdnO#4-hN!Dut;y;cw6kM0F+To>5S7^8Y^lj=-KzEG!v7F+cV`E5`()Zx?tuMn
zXsFNDYw#y(`<=Xkf`8{%G9hlKV=rf0?AN*ZV<sl1^)qv!N3UM}JzG5wMf!Z{c5zM4
zO3^XfEB2Y29FoI~{L15eYO#HPmnB(OSC=5@f}bBwy%B?xbZ}6AaZp)Ng3i<J)7hEq
zc6+4Iy4#+Sk=9Cz6KMMF+w$t_>c(EfQP%(^r3=Ft`X4I`i{VUh(ViZlW*8j&+!k*l
zRrNiHui;@}zAi4#(~pgg-dF2cgVuH#(MPK{OVMXc4DC6^xOdy+Z>l+~{ZXLdcAxnj
zHQL0Dsm!v}yq{nD429?l^NLrvCEhqKw6kV(=<+i)D)M6o7#qXftG8%-{-R+i{XVJv
z!a$H!L9^t3VY|%8%5gU8J&Q%}Gr{6u*7~<s!C6c*ymH<coIKT>v0M&(qNY6C&CYgA
zjWO92GID&``MI!>9h=mg<)gD@tFs$w*ki`D_<1(D4oS#-BCjPSQI4c?jSbI!vA*Zy
z2eY24R_h|juRd-6JfmA5xb#_XNDT6??yeT!xY^oTM{E4ICqzuDqK3A3_l49}Npcxe
zFh;sHrIJ{vZ_S=SL1$+R72Cc=it0VOy<PM3TU$BJ_nZ$qXU5C#LD$_V1vRyIPS4Pg
zxQN3Y=o@Slmw@#+R4fKesd^78EbJo+lk2ElpkZfT8+9{h*XOW<kB55jU^`c&4lU0Y
zs<y_zC;uIspGOmkiD|#T=z4MQ3U0ojemdJ<CCZyMcNLB&L-qFg5U_}|z3z|GWv1BM
z2nKQ<<L_STQ>rgEx)v7}QuNxLKUBFdT%$nCpr3DJKy`0#f2m3IOS9HppOPhnesSxq
zIz*JviI$e;;^GRElJD*K;09#omm0;z5Uzq}V7}r9PIcNP80%pF<^STusbCP!N1cHi
zQo~FPky0fxkFdOO=GxfUz?k{|-t1aysHqX+rI2fFGe>*HoN+fNfV6rVHwL|R9L0Lk
zY)Q=|+6dB@r!UAK)PDtICB^;wX@9S(qWViQMo~qTrn3$lM<jB&x-_<vdovQ0<$9j^
zQ$?@vUm57@SsM8`JBH`V(;6Em#hlL8J<hv)UQ7&*=Tg?Jm!DBmvwJ>0e8)TO_NUrp
zB(V&Yvp)y)>(BN?vt&aFI7pqsl0)%@vt*?cIXl!H4;Oh>9pB&m^?ZX*P5ro_N#$}E
z!P-7MLqQ1%2?3wKU!7mGdvbhYzJ$t{pF>VGcyd9NBz0QyVh4cfBn3bJn~EH23y$4Q
ztI6V&tQ4NP@NU%Gw$P6l4`N0}MuNh^)$X28sVQ$;K*^{l;?mx5aZELA6a)@@5kIKB
zq+VGW4%Q5>^D_*@p>tA@Q~w-kN4G=<=FVNxNl&2SRQ{odGu+J`B{R&<Q1>&=#qRE$
zi9*AiG5vuA6zJ@^kgT@0Y8@%IYLy8I(!+cst=J9GLOonddRM5Rd%$!}joG=~{r&xX
zG<dFb;gBaDEp5Jn92aMMGh=7Z5um;K+Q9mxysWDIO;ac=JhDUGT;FJCmrID)Jn6Bs
z=~<=a7#f%TAp(TwZM7+NIh@MBPWUL5<;6)&P`<0yyVINHJG6MbY9n^ld@(9UUA%C-
zdV2bIM$c9+%|sWt)(McjefwGDow&8NBnpZQ+c(n%mpi&dDcsCP`P#EBQ7kJf$3wGD
z+>x%dG}0H<$<c}G^0Lg<8wD|eSyJ&xkb;fVsV8Eg+2-{bKBQ@=KUrN&HI()YlQK6#
zWTuu_@cFaHaLMaz_dR-SdtSZ_*}?p8>Zeh%GBTyU!pZC6u}Fs2H+D6?*}}?#nf+;x
zBUfo<s+aN_Y5!v3ZYisbvXm*R$BIlAi+%x%nO4+9vI;i!bA0aU-@iYRe%CYDM28or
zU5r<d`No12yuqsNC6FY7MUyawAlNWL7LNisIXRJ<eLP<mhh7f5^Y#gPBA}8Y9UWPc
zni-%s>pWvcs+A_bXb2?@=6H;m{7TO256|zah#{7oih6r81XMioL%%S2wc8uM*7}Rq
z!%<G7)13D9KPkwYdRUlXxb0{mGq$8tAJni}G=)Xz3Z>{Asa(iHi8|5Koj{@)D<WTM
zdxp#K74iVzG>(22#*DsmgL5P=R=;38QdL$FDl|XuR52S7_A}}UWOr^yk2^Cn_r6+G
z0;4HB^a{S+yYxI+aYf76SPArpYKwsRL074CI{W_d9)cwC_LJT3x0jlm8_4)ZkB1hV
z$szlTMZVcE1qJnmx+Q<~Z7FK-j<BfZpXyXrA6RQ9*bb3ZIofE+-g&;UvA4n?Ltu+n
zP*sw?%{0*cTJPq%752#y9-c-mcD{-@^7kKZ<W1@JXA$m;i`@2w7f%^wW}6)^q@o2>
zDlAg{N~5EqAdk2>$38(_1qDf4Tj&h~wV@taGD8RJ_Nccx9400vPKojEA2zDTP<#Bl
zOTE#=?>1i;URC~+`@8O#i$vaD@)b!~$i~_VXR#nK+_E)(irN415m!U$4xzB1sEKIH
zr8cRs(H%vzIhn7DQdp;_XB|(;noC5Kk}sHLmt@ib?AvugG!*1yDU|qw!TX6i(n{|0
zXGp)^x;2DV{*;tNbh|tIs!I-iQB{$O{}7oC!?Y|CWt^9fxzs$@Xdo>nCMIHG5zEj?
z+9@<&9t!M8)$^p}vZp3SjJ*<3pFx$U8wm1v5BuM~eFLDGLZ%O=WP1zPc&USiCo%NC
z9~RQ=*Q=LwH+2m*bx%wUCnoZwiS1wc_++Hj<h+DzY;C>Yr;p^o!>28lyadk`fdF<T
z9nB7{fT7~j_?vwZ6<}*~1a`+Y+RGfF@H5mnrO&*nM<GUWaW=J<Yd(2MhptkWsO&{5
zRW`S{8Wm@bqV#NwYK*swH1u<xDf<<ckAi_=+HZ>P?4txc(^3nw==TkRLP9p8st>10
zGS)#89xxxK*mnpsVPR!`X5OM-s$8Hr`&fgO_13XV^d^$b5XJbGXVi+nh0l`wRY~S6
zgV4*3tSnGM5vcX=80_t>bGzG#8F=8b_#5hrrbC)el!MG}Ss>YdFJ`JUVY~9fTRfHl
zoXuF4(as}GCg^XiIzzLvC?6HTPk7*pnwo%iU1f!rxd6Z1uAKAzCduT)B#cLh6z#{?
zk7;UvzRKlA!ZGs=Qy+pW*}lQ;ynT>$wOLvP0gFXSh36;c<(<hp+t&EY=`R-_T)kNO
zdIu~qb`^o7f>&4uSk?wwDQ3$##UnfGr9QKT&-Z6$r!l<SqoVA5GV=J!$}5%6_Vb9$
zcy2p`1A{|>mCa36q3T@i3>(wC?TD!H`4_CLNOA8B=|0ZGg>+D4q7ez00%Fg~Ks*S)
zA4&`({f55j>OO(s;o<G7*ti41f?}<-f!?yPn1F?DrlTaU)?W?Kg6~LSe_J1GJcQo=
z@?m#(bJ`;MRicK-Jl3V+*a1uucqW7My-TmN_pN-J$W6pY+wZcM8XRxRO`2mL4<2p{
z)Bq{ns<sNnIh*${*t~Hd(01-MZJ3sxMvjgNL6?`8Uw*~Gv@E%6M2)iv3XtZ0%Ng_g
zxB_R#9TfaLTO#Rp>j=}9#&)jCbo*5tX_3gSW_--ASt}WRAp|BFpFnZi_kvY#-(^Nn
zLt(w<T;MJVH`5*OZ6O9bGlN}Tgog4LWx_`tl>10Jl5sZ}7-ZI*3e}YtfIzo0x7;1x
z0wj&|!bFgpztXKuKjx)o?D|ynCLVCj#9RA|s{C@|($Yg+T~vFM^jAJCsdEAba9X(P
zbuW$TU)CZm`T1}Tbaf#@X~aBY;^L9=CLk7p!V(i{b5VMIO2b#)5EJ9HY!JJHCDi14
z1KR?`FH{)7#&{vKwZ^a~hVsKeLmF~8{$trE*q@c9WNC>L6GQ3)W(}Mwj^7Lo4{mH3
z@uz8<t|ou~*4m#vluBTKL4e@`3|S_OC#UD;if`?(BRlZ+*V2pg$qqHa{b*|h#*a;h
z;|$OO<0;qVxbW@GO)kwAkDtD1R1&``0LzO*NKhv8cH6bRBkZ9uLH&7UXy{XUlFBA=
zmu53UnJ*2x{$cvWC&*isg+ub<-8};^5V$~J%+aC_0piK)OIeYCZ@9>+s>-kM@euCr
z?E3^sNMa#y^jQOgU~pdsGJZhxNC<T-EiL5^1mdrqx}`vMRlpy}N`dl1xiWf*=Jar?
ziO1gj{1V!IsqRb}hjp`BnaRv=eqXB(ZsbMIU&>xemx1k1Zt(alz_7@_#(rpKgo&PK
z?Ko3>Wo0Flm?tfQI-bqkjH8uQ9AO8+J@m@hxVf0%?17NbdUR@f_AtM>>sBwt>~*gY
z(1*fV2COKVkrC^Se!bFZ`Z+Hj@7jw!Wm#F$UIc!A{uAUtNOL(Q#NK$bXDk~I!q@C#
zZ1$@OFi8F&O8f!11o_k-f0CEKT&Rai9C|Qh^ecV>$M0Z17>~2v2W#9Ukk$Y|8U`{c
zC=dwx-(y~_UBEF&g_wB6xEBpdO%<sf;d3c2GJzejvuA;d)T0Bpx7Y}5jGiFix&6|c
zXlv8C^FjG&I#*{sRTQ?Y$>Ad`lPO{Oi@C#4Oj%u7kUx`1L8(2J)2#Wy5`e)}tBZ{(
z;v}rB6CVNQ6CGU{kZ0t!*yYQ_<evEPL9Jz&Jboar&^T$Q!oR1#x3|IW?mC2UAYFa8
zd9!hj#ajr@<2M|lkf2Z%tH_LnjSazg^U|8Z5THWWIG)7Byc4hZvvS~ebFknYh=YrR
zBbm$*$w%qyv#3_?L=XLpvzaMLx{#W_KiN!fc3=WY;4kBsgTuV;yH$m|ljeQ_{#Umr
zj8!HxCdLL22mA47_><ZCx=>5Q)rn&jCNA#M@NmNIs&u3hx9fG(SW2jGHXi&(Ot9u{
zg>rLqL9s`pqob>{DJa-gL0#PULF%J-R8XM!h1b+GOUolPowRZ*t303(ZDUhVScgi;
z*<gS8^*H3QSP9$0#AWF45>rlD&QyO?@9*;O`LlBUfosFL$&wOPu))|Q?O#)X<)R%l
z`C|OdpW%{{f%x}_bM=gL+}w2B3EAMVi=k6C(bpf8OV;?3kw0W$q!JLYv(m!P?~97f
zNOU!OJ*G&kp@GxF*u2N?w)Z7}U#VF2!d^tgNa%45%y-+8aqGZ&s%lkNRC+I1ht1`g
zE9vR#A`tQ5xxFU^SzC6}lfQlE`#etPe+)lrt$H>5fKBR6{rWY*YP8<|{QR8s1J~oZ
z#WGoRchK_tBPh2;z=6Q$;GnR-#Rp7Lk2n=xdA-|u?}YJGo>5*OEzJosQdxP~D@h6T
z6<m3oK-cr}p1<`kw&Rqya}6h0S!rm<U0q#;yUfnn*iJUxSpj9~;5<BB2723TfHh>d
zXJT5X^5u`q?d54tBw($GRN0pa!%kP(?p~kmnyb~`9_yL#^TR#bQ>)^~Cp9`dpYC>9
zTh|S_cCp*<cB-fIk><*bOGwtAS{=S~n0`xW`r-X7^5wBOw54xOr7bUT^9fG1xiiL3
zSXC|qhIu8}*V%D@v5Nkx4G}v+Ou?3QDa2xIa@6VO+|*{=({}_dARutYq^hau9#Hdz
zFE594#$)yyeNob7-Z<6O(c|8^Ew_|qiXRFK0OD?HYH7GodTpKE#j7;~y@^V_HA74)
zg)w8VCt2pNH(oM4o}#Ht@jnxo1*88zzTPsfs-W!~U4$s5lpr7?(jZ79Qc6j8cXvp4
zC?Vb5-60*4N_Tg6cjp<rpZhuQIUnBr<Aa2I#hx{D&GoO{$-a9v>9W@=k7^<`SO_Db
zZ$n>Kcnb)<pFcss(A&4rDGHtvi~DqT-IwsViNt>Y?C|i^(A2OV6W#Lc+BUt#(|sQm
z@A!byHlXqTdNhu4XQF7P%9eT{8{x&9TMm}<-Mk18LWsd>r7b=N`d>o|;7}x#N!z?N
zGPc;1WjszxzSJi%Qz1$$n+YsMhUVt9?ay66o-XKxaP9B!yI$AS)#)1;uz&mn3%0$}
z<TT7?Gv_Hf=^5Z79-~+VQ|a~fm1~1xn5U=b<jmtU4LnOUINa40Wp#~==$-p#OK@ly
z7@1#Mm@NO>=pw#5Iv7sc<oLZ$OWPAeYj$NOT4TM@j}B+&G(9~uL{CfG&*%F2^XJYi
zk@EFk7>@|@ht)Tac8-tBeJ6mcR|#OHS%Og31E&~QG|16j;EMHTz90<ea{1_RxVWFh
zu285tHI6DTE8iDOfA?mBCYiC^y6RxE)P13(<AH<4Y@5l?8_;_MwQSPq*^e57?hd!?
zbbdZ3Tt=Bh&lZovaD2EwiXg|FC%2j~F*1VrEQXpAfL36V4+`m@1N_1NPUIQU)JHsr
zkIVbqZCP1k%wu4SY|0X#K`lLebD?c|JJu7UL8Bab%5~He$H;whQeIhYzz~aAvSR*q
zwgDZx%4x3d_ropbYm7e^*%U5wh%YqiwE6ACgBy$YNBuClZ%hm0yd9^(+BlGWwc)m!
zSgSE#V|TZ@#rQD}XhA7)ZxwOK$jE$Z44P{5C>IDz#tYS)Z&$m#E+>Le$Tk)PRq#!q
z&i;P$DK6FxVrWuPMTGvJ{RJrEG;hxLW<aHt$_^g}K^}K`Z+-k^vVo$;<$lV5Ts&!6
z_8foEy=5XGhE~l9yypx$IuXu?-m{bs-Q$~;l@ZyZM!F0zZ%_3No7`OeuDRa4kum6f
z1>oP-2Ins^AMCB|=KnC);(}35n)>bl!kXTj=^IJAzPdU#)3h)!Fwq->-+d}2mDoMs
zFbA{P7O=q)G{VSGV6ufJS5_VnAp7Tk&_=|l<gi-H5i&J`q{ClyPPV8}C{*m_C95FU
z*4A#mP$#FMfCGsWa84^~d}u#sYr39MnW<2%)8A80KitJVX!-FLC@IJgb`ef@dT@WO
zC^k4TOMP_TPkc25Ry%23hvX!>5=?bWZ#UN)L}Nx!iUjWNvRYsY1RWC@`=Aiy7i1-A
zC<2*RMvLb|!@!(3G7hA<J2}vqpFbRkN0cX6(n*ny;}q|>KDMf|#WWzIBBh};B&n5Q
zBF5BNaQ_U_SI#C0r%ITU75u<3v^(6YPaZ{fJsF53@;go{ZcBSpT2f`xXxVY??@QD&
zpB%rhO0G)sa`Fmh)KF*I#wqL8j@C3Q-TgP>h@n<<;o?iW2|anyd0q@;ZDbwS_X5nJ
zc!pTh8M7U3c5akeLR(QvZKqT!1OZk8wV(WubcY>o1`e_p@(ozo@IK~y2%Fownak-D
zcovlu7ZnQ^X^k36Gw1IZPiB&R2tX&yA2~KY?%6sUuWJt)O#L#nYq5<R?WPiwuOTg0
zSX``?rgff8A{KyB_36BnbcnFH?Y~?ABP<3Rg=*ovfqjcvYR_cz0-YWbW<4x}F<vt`
zPrW>?Xt{)ikzXoOu6`{1F``lg+AP1y(mn+wcj=dQ8&jPaBcm8zC$H~U_A0wFb8_0=
z1^Ne;O_x<{2NJEr;VTAR%etB|7tQ8}#LEW#nEtU}M~kAFE$gC5hi4pX8=sGtk0odJ
za!=W=DB-&aX@-9*tgEdn2BM;=of0owM2m|QS`Qh0ACqXnPLG*QQ681CK)k?&bTKo1
ze1%cnT`y-i6;~kB1uVvvm=bKnP-Z$$zWx|)_}M%Txg0}&k<I&_oo?&xO%#X#BeF~=
z-_-CO3^eFb)Z(M;;UtaU&oQW94wH%;R46t^T||9hdnhH7wmP={8_(t--^{33uQ;El
zbhOF=z$GKSGTxctYEgfuS{_SiKX#|pXULkjHKibEY_&KiFAo^u)z#J^yjT2q&e*i@
zf-srUBKg(K#_jgO($<g_<xzyYfSsMCQl2k<z3mQ*6nNNFd$NcH916m^f<*Pf??7YC
znXCqoT`~;;ToIeux{g0KH8qW7tYc#E6!P;EO2orxuCrY*o33~48YY!`p^b8C1G6|*
zP?OYj3kxljsF1JX`Uwd*fIwqh+*iJ1a9$)j<9oG7qz`W8yE>+~F$cI%EcUsaxtW>;
zFYj9zdpNG-H&?R$u#g9JtEV+#x+7k)xc_RNpY!QW&buwXY3{`9)`^&pXKgQAM>%Yv
zP&DW0+@KPN_DWo7q6UOi1Sqxmd#<p=?+YWMyuy5V2+G%p^nry&qd<1C`o;I;=bx$c
zKFb?@3Kaw)ASPCjAnOSxv}iVlzn{}tLJPgUlZixvpqJ2L>f59o_Bd4j5~_pG$X#6*
z%JS;oHPKOCCULMhFi>O+O!LXG%l=_%_O_ZZ3B*BpEz)e9Szi{?7X%|b1r2h5ROPeJ
z`4%^NJY#@oD$C0Bg_ja~EfM&3yF|Z!*(fg!Dfrswh%_sa0ENhF+F!h6?Zym=yo2v`
zSlERS#EXEO_yrW15Bi3=l~J}1mjoIM^k#Eh8w&31Fn<kNCGO?M#@<=bB~yS|$T^$<
z%FPYEKDBhezI_6bkdQ0_#PAs<=Ie6I?E3m26t8gck&Sh94Z(%Id^sCsGU0~v@}(_k
zakH!76B4E<f4ZB$yLs<n#tl}2Ey}T2^za&hAMPK0q+zJ1qU|s4hVhu4{sj4cczO0y
z+siB0?8jxev}@%$0m?uX?=!zIVnGbn+p>;dzhfyoIUztFKizHUGouCt&Atc<3cPkW
zXLH^!e5GTTsBwpi8g$9>*6kR0`e9jm8X5I9J%BdZ$wF!o0lj`qY-DTocc#arq5^MZ
zn8nqj6|jCsO`QP4^_(9%c7Fy<w)noNsi{fg$Jc!VR!nl13i-LAf6V8u?=IkS#FOh^
z+6C~Wq2O{jjBSstv<JVu4|H*dU%#=Y=V8#v8mX)fBQlsQ_Ca(F4Dvg|5<#_<8Yt(@
zWuD!bTJiu8D+d;m-okgd<6v0IitTtnFC8BudlN1v8?RWX?1U=8NDSwhBaz};>e^p&
zuRJ~Nf{(_1HNom@Q(iw^=hPoqrB)YvxR5G!viNqu2!8hRjpX1~WqNBf4!!2re_*v&
zCVa<krZ@ayU$;f0(p6AB+!sMu+RLOBXYSvNrQ}C$b`IR^ONVZc&@`4mp7(IG<}bxQ
zUajeP<(Fe}geS+2@p++w-roG$rYR={-26d*s$sx0%**RZp;AzIYyT=3w7twS($c<Q
zIfxTKQq+c)Vw33N@#4oouOqjh0CP0E5i}|*Y;AOLuz8L(a|}#!43MGO6-p;2HO_ci
z3ApFZckAr=AMtG9Jo5TSYGUdVPO*<|7S3*hZ_-K#4rcobI290;T=g$-d7`U~jZ(7q
zl+<8oc56K}&P0_uh!nX`*7MxHH9y)s$SNr+XBy}Zo5X(|uVb@wMhC&H_&p<q`99)5
zsutSiq?X1wz$v2Vb+SX=**OawucV|^-{*=68R0^kTU$c6Qvo8v3ysbf=X302<$fn3
zZ^0Y~N;N@KkoyAfQD3jRtn~l`Hwgi>kd~Hmslv$6)aB^bJE@`BXwz$ObdUroPyk^8
z0X!?RFywtlgJX9W6dCHxd5)EpeK9rN$4oyWIzmpcz|<-9I(Q0kWqCM>0-W$O<A|{2
z=MbnCG2G8V=eO3`p~l%EQbBHb%f`=w@bMHt*HC{ll*rVh8nKVDX&{jv6N(iRwXit&
zrBW>|EzR@y4+83I$q7lIn1QFS8n_y35@_!I1x|#CN-XjbRQUMRl%L=Gmi5x8LJBVj
z{TYM(zN&yLXGY^23w>bIFzg8i`|0*n00oEpk*Z_@7<I6pUsNZ96Velj0cA>iTl@CT
zCYYUtU4UWC;~CkqlyCGC?CcbzW$|<D&P<#P-rx22_jgfJVv>tl?Mzt@uXb6?HxMpv
zfp#PreboW;YygGJW$2&kFQPU~P^`_?yJdu@plybgPRLiS+Ku3FUw<XpW-|G!iq9Y1
z^nWAC&8@F`u+Eh|(ZaLh^fYkSRPHV6#c}v$IX*rvSZKgkcvn6Z)ur&iQz9BwN4UP8
zC{|OccH_*^9||W$JUSW*3&UqMzcp@r=w73~|76cX^+_O!Y<R7JV|^e+0&Aat)~uF^
zhLdysNC*1G#E5C@dEbZ6VfkY~F8AVKPNb(3g76E9ftn1RSQA_~67ohWn*Qcs8UW#r
zkpqsAkbdy~$Pr6wAJJR@8JC0Q_8*x2R~eO5rIi(d>yW5^gunmsEDX-Sfs`hzFtIyd
zo^bLjJ>4<8-~ZU3kkw+{yw+?-K2E?#*RUD5{sH^{2fL47(TtShXksvazJLZxG`P`O
zKLU$1p!;TGhzkLZRrk02pk}R&xd*B!KCScD^#Z_hF%mU+s!|&Swu5FH3IsMJ2UF!S
z<q?<$Ls>mBAt@Y)5T~kY=FuqFAWS#vJeA6px|Dbcob$owISpbhO*i)EXm(8Sw;i`F
zd&LoNk2($@x<4jfDj{#AosgEFy?eK8E^WoI@b{@4#@2}+rsMcA9{uuA1UdA}Xr=wZ
z5)~WlF^X^QZ&Io)o_YZ5Lq1LiO?_kG_t&-$CFmFXFdoIx=hI^*nWF^YJ{}MIweR}-
z`y2Gfw>VA?PG06U8fFd%ir-DrHx@MA1fLALC%c=;o%xj;V9wWX%@x=;-MgZqq9}?L
zCf3)pS}f#4Mwu7Sk_k0Q_2#B4feHb}10dwEh+u;KH$Zb6<lU=-;lG`1%WVZsT@9Qx
zx!nnD7OA*$a>8N1-`v@n8XWva{%*Ou(uB{`)%kq0TCdUfY{8w-?PeTcr&~%&0B^r^
zI?|rlrv-XmHmlo@pl-kCILe8}o`1jtw_6Of3JMBJ0^BZh^*AHimWvPc4F*OuxBOp<
zC=`O=rTZ;LgEw4(j#uxZdD0Z^Dlhe~{P_d0KYjWnOsR?b?*+YG@-F>%y3#skdTqHS
zG4@!30;6NOXF;ghDdKVw;i;*M-#rn?>K&~q=t6(1NAX>+#SrvFf3O%C&owr#HeZON
zH`~5@`hqIIz<zJ)uTo~Q0H+hvw?CIh+VD!s^7@avV1ICgODan0h~2}-b2izirnmU_
zLTrCbBLhP-oZcF_{4`LMEP`^(Kq@2D6L{MKx6;2q>7QJAgK5OAF~#pLqN%YM7y#QB
ztN(VozxCc2NCAtxin4NVGN-}NL;y{1>udL2(~;Sz$S4?(u<zgN;iQyfwQ!qBdOkac
zi^59EMXi2R0mt9@1{!paQwWh4NsqIay;FPFX$4bAqpOO=+;i|U7)V5KaFx~7EifY!
zkNs!~dk}QkDzBhIohN4dtX0SxL=7B2=QGms(Ssu1!y!ej_MeRicUWy664}3aVgDlJ
zIA8cv679ja|6JPkVdK3MJ>ud%_d_T9BPllb<s)+=nnn#92Ol2?zn~<eAkXh2@;{rk
zT;M#hU8{3*zQfA9vb?<ZFZhoHMIFDO26W2-TLQ*IR6fccc#Aw<0HtTS!N4ngd|Udh
zVGy|f9qt-9Z_N(>;ZaL!uRe;==#Q{;c>K$!!GOl~?{q!|c<^grvij@8aX4P@c8wVD
z;H^BCPDE^f{=?g1rlz6Raj!d}<*{njF5ePk=FaDsDK0Az0|up0LX;$#Om34ELi0qA
zILteu(kWBN4v))~Ui_V<rIz-sM06fOapCT6K?wEi4BRAa3MM@0RrQ_5YwQ62oDpVb
zW;t2-c`60Tgy=IDlUFDRQNnBLt<|dXQZU+g8C{P4RdzCw!uvl>Sw=@?PrOi}y+92~
zl6Ncs#7dQ<_@e>4Z;I^5HhqHOOTcX)r6(t6kbngIq|tb~kvev<CLVxQ;N#3f=65z1
zBVJNulH1Y36&zm1&^$CCL5I%)&WsNKb&c~kYFcVINMPetIi|DuP4&S%Sn^ui+QD>Q
zRVmiv)X4nX{@&_*R%15cGWu%>9;qFUZ;qh}D}Uaqo1(0&sn%E&xhZ65aQovOdTC3H
z^+use)0az;<|H$s4_}?tVM|wV^j`2KVK-D^LrhFeg2%z|;2e!H>sjP2X;kV=yH~sT
zrdnGgV?)x##(^(eTPWI5;{`QoX$XH5hQ8;|+FCY!PdNJ*<_>8TV4*W~LO>w#=piBk
zkSUuzXYd+$GwE-SgXrT4FwFVSKo%KP1Rk42nQtFSg5YO+dfM7`Mk@lCm@@R!6@?or
zjDRGO%>ScX+y1))xXoW4Z?m#;yvtwn|GAD;AsJ;?#G1d@_g0-;_vhS?jY_Z3^p=|`
zO2&GY=%?!Y9jN6ka3w)&KEIz<i$|f7Zkm2Q;Q1Ot7@3YN*d=Mm)VMK%m=DMZURYku
zU332#{-AUar(h#p3*mOVI69Kb7@}aB-WthSBM2+II%!P-4K*uE%c0?67A7WOYdc!o
z)ZfKUp|`^A&gCtza%F906>hKEy#l)PJt)@qB*$3)Wj}(GJ1uE$Zsz^|{WzE>Zf~K9
zo0%Di%#C_uhpwJnz0lT9*VYCtHMgr_Qpca!rq>r|Hs3be9SzNW$_s=7{o<wo!01So
z@s5}q8w;TFiti-8>-?ig8fU{#Fxs8c+MHf~p{(;j;{nF=1x;6WyCCTF{GjRO8}MjH
z4*IRQpYQG!r_H<24$s&h)PHqYc!tUe+z5ble+@q&9KayW3kw?aQWWF{<T)cb;+7M?
z6isDhWT3;8gAZXj%_cfgQ;l#>51kpSK{HYG*;9Q_@P)tsW|j_Qig-yHl)nX;eQNAL
zts@Q_Uauo14wmLslt~zmgo2&wg2pEvow*dXZC;YU&8F&W>+`KXKLC&D{_swOBaZ)J
zV1|ZMzo+W$?|a#htf^;2*_RsnQCdbuqpiKsKOiK?^$GxQ0+=JSqfFF{<zp|wt&m9J
z#K$B^(vZv*AhO?^KD)8qSzn}7qY<S03zRVU<0Y(gbcVKby@iFuUtwW1fs5eY!H?w#
zy>Q|TK<m{xLq+jHdBfp!YA9sKTiWF&K}$(9(_sGs+MTKtk&}~q*8m#${~52L;ov!N
z>=CZ=a0+g}cFN=TYAv<fvw!;herUWB54ZOTjD{M@^Yu}>_*ekZz`s|ORabA+UK93f
zpDYz(`3rtVLP|DY?<h0qk3*aI;zdYjcmJao6uu_NL<EV+n2e{vyyHUmHx}Zx2@PuB
zB!49&sHmhg{0*}VM+7T}4vbBCH9no@=LBMGsN}!Ad`j+f_Xoj2mm%VW1wv`jUNW*A
zQ7O^{Jp6z%U^CC?=lP8)uFUjkIl}!7x7Frg>(Po*vwDNM+S`8S7XC6^1dqet_TIXe
zR{=3No}O>GT@j#%nKHeRY{Kc8@4=t-h3xDP?q(WY5iO7Sy}bcuWNB-69U8sQ`R7K*
z?Ymc9dW~*9IVQjWVIUZdC%%OC$RsTI4?_Q*L#(9hlL3Q)6u4&pm(*&`Hi0R2Dcr8t
zEJQ?{oDJDN^PQTxQWPvK@_z*!E{+aE?G4{ug@zxq5G{%E@K_27$zY1`CarI;vvdy(
zfc0TMt@6wA&BZ_p*Ij71w6c6Z!ubzxWDtfEWiOe#k~U~q#Jc>`Yh#Ks8OU8o$w<#E
z?d=anI>yGBHRtQCzkK;qVh{GLvkJ&KupY@(b~ZMF0RfB#Lm=MC!pj>FXu7((#^vQ*
z6jJE4v=uM+R=x(?g#W%2VfhP)Q+@G_Umbju-rHGO9mg>m00P+p;Bebr5-`fh2-%C2
zc5-#W>&uJ)LXdi+|BchqA5Zs+tJf95{O3abJw4@B<uWjxeSOR<4B?6rlg}*u+zwL}
z4%D~t;_K?_KpzMN6*<D5&nhA^5tzL{U3a_6DlcdB7ABR+%Mnisx0kf~YpOj5B(?w;
zn4OzrX09B~l@$@`Q7SjsSY7XCXS(`$zi|ZAV9F{g1yXqk2nbwUT);pc+!!P+KpNAf
ziW1FIZ?F@aFYWB$aCd(P$KehbUB}PG%!}2)<i-1Rn!fJ0okE7R^Vz=)(Je-Ol-Eq5
zm(`kJLvsAQ$d<CJi|>z^SRdEJasZ`F4$v9~iDpUXj)HUVWUpg+^~sdt;wG2zfj1Cs
z+2lS(Bm5sHjmWI9u*h#9G<ET9GCzm}K??vz_)m(s+-|-4Z?8uBQOc)wzj!`H9-!+u
zc=|jvZxjTB`Vp8tCG8<usIl}E&6=K^^nLu^ljP^^9c!HMo%ZeJ`PUv)F!&#|OD9F5
zX?l7)Pftz=z(djd_a`Qq7%(8Iu3rfc0kiDbLecM6SdsBz-^nus=9@NH&1V50=Hvvq
z{;QMwQ5(+o_I67QEsa{4;(S_H2e4f%QEkvRGh?<~<|8-Ac}h>OL8JK5u^Y>=gVbOG
zL^wCv2!WZ=>)w?U6kA<)td3UXbDZ7O1=uvIo7=miO(i+?Iz|bY9S`q6l=z0Yy8?k?
z(+71IKouyMRCFbn!%s^C3F#5`Hbrmqaz!tf12%6iPEJnDRh!;myiN=cKi=7&GxdmC
zaEV9^2d#7dvnd0efh4Y^%kw>h#>V5|Cq$z}z_ktZ>H&K@Z>u72#3B0n@*4e-5yH$6
zTTe}luA6D_tcwo%2p(NS+~fz4q~_)A&0KnTxWi2ns|z4nsB;H?qP<G)CIaYS;2<D7
zjxp8zsQ>f6(*7X%`1rWuG!X28fLR>3+o`W`w&`>|8+bm!{rR4du(NH^*v9+hhrh?)
zrbFM~1e=o*KYTr}p1w$DpAD-&wV6%BHBv<TwRoB*GwTizff#$q<5Cdsb==*JNiH7V
z1#aXp3s?_EB_9UPY|M*`shgA#A^V4`ui{_)x?5U^NIp6&rhb`?qLwc{#bXJ4f`)<d
zuB3U|`0iq1AeJ7uc6=XonV|<oNy1F^U%wwIgeeCn2-i+k^74@6fF_5pf4H8e@XFl`
zE8lkuADHIeRxG}rU(wOoRDP|n6x7sFhJggh@DZx=%1BT=9$t8$1Z4$x7a{WTX{BxB
z<t=g-f{%_}E|`r0Pw%hKb7rPNa$^n-j&X}g2x7P23pX)5F8$B)b|m?Iac{5h;&5QT
zzZj0-MhI=7#pUUz@R;Oab|%kEa!@~QN_TEsC^*`LIyYAe1FtWT;+S6#hb}{S!3BjT
zL=Yc*2HSN3b@F;LB#?x(GULJ6u#!-X$lxzjtDBvJ`Q!^ONDF_5RHT0ibRA-*Rb)Z>
zBrVS0DW4#cu~Cq~L)T{;O-cQQcS7R#q1S8mj#N|_Z$5#QYIIjiR8&u_vd#VAg3V<7
zJ@gFucblnkxQwb`T3l@GVku!-24Poq#4&0gMdRq`=wfOr;6nxCF?TmNhw@|qUY`j*
zzSMY-f8Z&|O_`rp|3@|{0nC(zB9)l<c&e#q)Kmr7?-exx8)>J?%mtsq4aY%OU+)RR
z8w8Id4W@@5@$vE3Gxv7{q;8O)xHtkNV{I*tx*@kYRiKbi7aAQ;7DM~iOELkhg&0rR
zRa6>LP4{+vd9<~|1GS}T*=lR~zDW9raS>=*8GK}6d9<*a?T7_A!Y3(!F98z9cXx9{
zD&DGxudHbN^zkX<QZww;E_I)K*hvFqH_fwrIcbz?9gGiMub+5(Q&Lj4a=lm5^|L@1
zQZ+W-E|{Dg9Q*|!_HAXLoln{KBL2VNnY=yifVl`QP8Q+YXC)N{Rr%=K(r{AJ5A}%=
z>0c9<9Z}J8v7GCf?iR3+l29=m1JB@hWSJSEnmqOO_2>I@Guft*Q8CtpH(dne_+Zm1
zEGmu=CIbWgw%_~Yv>WSiBhW+aSR1j%o;8++%m4Ps%T%#xzPb7HAXS)@6r#mE_;h09
z<DY{*DZrsHc9Ee%wVS!`p3A-GW1X6`&o2FA|0XvZYLfwq{pr$Q5J72wFk1`C4WTEs
zR%`yl{{E?{Z<&!&J^*Jb)i?%|=vN6*dN|PHAMX=%`v)z$SD<2-Na50fGgWPJ1DQ92
z!3s;GDekn&jfp3c`A62y04dx%oPp>pZ{gFkIoXH*0j{hM{qM^kKR=bI7zMN^#5y{S
zsDUU656qxfyjbjhMrFP=ObUfYCZa$p)n=SltIq&8BSE9Vir)#w{v*IiA3xp7%1S?2
zdmn2n>zsl3tYU$)(^J6l0^~r^>q8jG<!jJ%V?)HeIN!6Q@Wi+%Hrs0(GgPq`5--eX
z_JCqHGBPF%a+|s@Z*7XFf;kHBwLI^|wYIhffmD7^!6WQ1uD_7O$WRV7*NEUGJM|P}
z`*l&qDJl8W=}!<nm(ddf3M=FnO_=CJPAv0{t_hKmt!(EJ>_U3Hr>-mRD)$#eG?i;%
z^|M_LOnjGvf$u`wzkZe0UIwg;Cq)iR2TiqvcIDN`&yk`Rb12HU7MvHPWxImsb{9xV
z1Im^`#Z|IUj|j;!=(YTSUPy$r`?Ry8qbrnX7?dd%OcW8zPyjl&k-Jx#mKKiZ57_ma
z7l6%U;Yn7UGXAn0*=ys2dUBIP7)1eeYta*3Y0^ae#!B6rrR@c9O(2X`x_x~X6a-RL
zv|)PV*vb`3k~!2M2$``w<m&33IfxaCJ1W7{oqghkh5^?QEsh0PfmbZ4uK6X!`OC!<
zz}A1R97sqX)&|E%!(Ac?l!!gQV8QFf@<<UQ&ElTpvN4=)k6i#SmfOqH#<Z5{4uh-k
zHXxF+-WlSZhOy^cCUx|RQ5dYXrtA`GYa<2N!M(Nq@48yYt4l>?MM3t(V8MD4YuV^1
z*NeUC;b)>>LQ2q{ZroHLg{meJ@N<(G?p$oe$K!zv07H(Mq48k^kL~>K^K(87tOq)|
zc9Mip&z~=vTrb31P0@b3KRSxMzUfkPe$EB%XdY}PX+V?tNTVh9_I-rGKw8`5CuGHx
z`R~~Y3JC^PEF2u1%0B-7e!~r*Av=dulvWlM7WxJS%~e?b{hg_MfIe@8yEUyGewDs3
z*QR$hoaG9Nshd+vTid?=e!FirrnW_oOoX|ik#SHRSz20>ZNtIB^3bZqGaEdI3Y8nK
z%S&~NiV8oPxV}U?++5;aQW+l`bG^T1+t32<0el#Lzrx*{xhP1Ld;gy1Wj^wwVZ(&G
z@@)8V_1%s8OjA=>i>QbgCt!R@aNl9QKE2pGWY^EmZY?;BOGz2Xm3r4MVY4)Mb7AKH
zP{L?XwKqwmehZHsz^OBR+PaDb)F!+?;n>b**`*akM2Iml(SLutTNk5>J6>)>M*j_^
zHz4smZvB>rKv!M7xTK(9WhV7rrlBhMxqU-J!(Et8M<zishGYiQLZ_mn4{o49dUKK%
z4NVcfo14z<E%==k+J`PskVa&SN7<PQ^McRP6r^Jbe}}=RtOCZBZh%n?GM<o;0)XR4
zrz};RaAjE)cH=ns8BiT$XO9E@&g@wW5m764E|appJoZI8;A(~CjjFI(9Ge;gY+0A;
z3`{@an_|wEWk>6KD%c4$+C09dp?R{z#Kc3Gf6Q(#5y9V#=E-1e|7;axAu3r};qO97
z=rcFgMrWI<aBIlC9(J3Z%FWF+97*sN&7wtmwpff~xnO9hTkoE7d3UihmOHB6;IzKB
zA(HuF@AULYXR2NXFtb4I+Dr4^c+*CF?c0YB3{g&J)6Af7cagr{G(1b<(WHw5mC0L~
zycV1fRprD#d?$&4rXpQQ@|$E&)}#I+`ugU^dgA8t677uV9?+s|72-AyB(kL~e0;1F
zkQV4y=buP%(N3D4|G$MS4tEZ+nZNr(^4wk^*mtw(QjMi0@M>v=rLv^En_{SeE?xSj
znr6T`)iXz9@4KI`h)6U}nr@{pbAA>j(HtRSIzDhMAb_L~BE0vCHIm9wcDhCS4s-F$
ze;8e;nXnk&za~>c_+%b#*;`C(v_teIp;4klxgj1@`iO{2Mhh7RA<;k8k$Cbu9A-at
zr}5liz7uAdRQ>5bOU#UvRrEAPmx`H-HrOD{ck1o4PvHvi!_GOr^h_DgNWw!wQ65lZ
z@oQ=qpa7sOw=GZz!1-3dw<Tp`Q*p?WXJo_&2gVHxnj7zbxd7Q@^3KaZy(*V63?wO*
z4_lj?;d!I`#dCmLbysGc^GZ!nLIkG)88s;I+~!N@_BI|vRK7qK!$ZP5bbG30xkcRN
z)llhMjVJJE-}Tm3pBa%YcaRx;RleEFj~2<gz_5vQ_eJMn7~i(MG|=h5ND%>kK#Mb)
z`?Iy65#6*OArEG=0$i>d_ohRXl<O0T-e2Ehb#XUqgxVc-q9jE{Mb*E+K$HK=5&5<?
zoF&lSG!Za0tqAwoQVLz-W;QVd($q1dNz%*;jzG1R837iYm8Q)}X#Itsghd^A;P>9l
z!U6(T9hSDavD1NDX10%yY(^BzzUk@qT!KL!#nfW#e=Ba=-UbR{PTo^{ByxO(5>+{v
zVc=ocAP~5C+t%O@=OLEB;uQ3%)Ovze>Fo0#<-5&q71eM<=cv#(D=SzJyiaPwT2(E3
zGfR9e0|PM@aL(-PD%8|at^u9240pxJmSVHp=_=xdh8Zs>l$n{$>2yp&DU?`yHQ(gQ
zi^aF~$1|6Lo)-sje`TWhv|%EnqWnug9Z}WP)Euf&1zkQ8d(KOm0#Q-k5oZ*>r<rN`
z(jUCK)MXHEh+7X2us*T|r}KA60~#C~GmD_>#h3OoA$|5BOQ-aT>8LmrC5MD{ArzkR
zJ&R}=3WX^ced36I7p`^o7WR<f)nUpC*h2Jz@(Qrq-Y-JDM_re6az}^~f(1WSFt~3c
zO6>GKRoFvT*Ti>+(2P*}D1OLuf9e$#{j&PquWVQvN)4sN2q@Y6Z()o%st*hY`O+43
z)SZ(H%F2F7X$+jOIxK#Mo*^P>k4~g;*$M~t=|5J?Qgw&xW<Op@4CRAjmAP~(o+%!f
z`n)Q@EcsEImX?YT9^?-2W}qWt8XODq_XlKkqX&fVCSS`i>B1^1G+-Z_o&8NKiSARi
zQBexIAQc4);=$ctBkgI96tJ&eS0cXDe+3;QBSVa{TwrWTPLiy|CSlaMKfICpw$tJc
zO-@Y2<d8xr1fNDNPB#H}Ys@u%_2Y*R7E_=i`nC2`4Cpo}1>TD#as|I3{l;S=&LfS1
zsfQk+S{aYl`1>j3HCID#dU*)C<F3s1{lG7g#P*V0V_A>*Vx@x>5pmW)Pv2nU{gsmx
zW&)?nCYnxZ@ao#Wt;>mJ4aEVLI7=oIcy^WNSg#!%4lCe2Zf}2F9L!s$7<D1|1_vrv
zYr#Q*r?`|!n*8CXTPJ<}<MNXDGZNf%fE?cWE}NyhtB({Z-G+$1RBfdVu%s1=Rkf_d
zHejLuqw*&4t5jN+lbl%OCd3t_b%Q2rEJwhO`mECS9WgNhe*z&dRw6;yVCaKE68vwl
z5bd$rY)Udl&_baKl4!3nddp8!4rlnZjw30Ju@aM|URT*w@i3r3ltx<~xaNfa8BRwp
z<#;G5`^W=ukz=7^mFo3{cA1gknw=80-}4y?5{!&8|6eM^DcBr}Fl**5>+*J@Ba;(d
zT|Fc#A8*Ue4rlc-@{o~|!!?hasQ|GXETq4_4xYs@>E_{aW1h?XHy~Ink;x1XIz8JQ
zX*OL}B}0S&7Q4f~`{tcGi+ZEWW8G=XoF6msE!<z&Nm+Rq;EIomK>!S9y)LW3a;>9P
zkF-Y``1(`Y*mi-AnGqCGQx501Q$YA%$;`yb!3h#Kf!*Tur+Zr=M_k0km_5)T*7{NL
z7d!p`<a>O^+C0TmY8@3gOL3FN$U+7H(Vk;*;_U2Jxqe=O6elOAXkYK9&8G`?S5T5K
zn?*ak{D>J*Ns@QD(BN9VG2seKOgDcy54=gv+Pj)v-xd7FdJxP<_Bxm-c0D{KqL|=r
zeMW=}G?6=hN^szskF$D!8Oqx3YO4~K%@zif8!iY&TbPOqK}sl!MxD#mRZmRp()nMK
z9uI9Y+>YHyw1%pyXdx`-`>#)!&ek<93kvdUcI0FQzdk2WF12ba`lSLm8J&wQ1@a|t
zSnUcQXGL+A(Xl7V6U{<`YGyP!Wp()sEL_RImHIpC0Sh0}RF`~v^;s`#%<FIth!he^
zE?m%Ww)uCd_s>Df!fuU@-LhT=00X&Ar?E~JTY!9T%f<=FhEhTZKnpL?50gxoU31Gw
zm(KSG82iZ~&$C9i8^5IGqCWw%@&26iR5<BC3b*skx|t(Cn;%~p9qYwSFL(~K-n?-!
zmwzjw?f#1Dxzc`CVznb^=tbm60D3Rq{8np0YW}-Js;5G28VVD)NL9CkKIg9H0e1wv
z97?rpvk%zR9pFx;o#z%8Pc;~@ZfFS$0D?=T!flNsqkXxVO0`w@{mE^KMiaWzj=ZC-
zoQA}J#}njhCno?D%=-D(pkpIIH|M)j!PJ5zJY~3zAb3{OM+IPg9>XNyJ`cL15U{+z
zA0f3$ngvCQfNMj2h9nlyD}oom3#4M>;?1v)zFr=Eb$pthn*NS>jo`;DV0QvFYg2v3
zqC`--{D2`mgdnq+iQ6l2Vifu$<r!Jk`N6CF{61c_bek_h!e#^3LR)6y^+@lln&XT~
zUN#vlG)iBSg8+@vctscd$3fHd@Z9)80H9Nu|ADtTqAF3(IHG1dmLN1Ntlb0d539#(
zP6#WKns0N|p9q4_L$S)G_7ryi8%cr|^lFb0gRvJ5V2r{My7ehZ7{mWd@i?=!<@=)Y
z^{W7V*9&rd+Z3lc4057v0vBvqsfu{CS<V{UWYcf7<Ctqn>8Dc=`Z+2>Mi_-nr0FYr
zqsAM>FUieUq)14}em|0v8{cai69-8U4yD_)@HNj7tp+%}#KXgX!Yp}s31;+stZ!Qd
z%l0rtqlkG@ojCm|P1ZZ(y;~!{m1BW_JaaD7*QYltUqcG7?V{>4WDtHZ-1Z&61U<bd
zdmCZcM|3n+mt#O2acgS_7-)bL!Mf?zOpzW0(k#6`nzK_~yf}_;W&c<(sN_WP@yZNH
zB5ulJ&j1MWNrEM}Vx3W&&oPI>mF2USYX+#MYxSOX0Y#Aga{8asK$H@6;>)V8OD6h)
zEIoxOelmuB)iH1tZEZ>dT9r?Zs7A+-UgIkKy9_ECno|2;7E_;(z+}2TSXflp$`u0i
zHQ=REc0@tg`wK-gT<^0{cp&I`6XC?{^QBb)<(H@UnQMGR5IEKsGpt)Wxw8K~x0<iK
z@ReH}kM}m{k3{vzHpSHDSUHZA!~Jq?DhucZkRoMcub27e1(|Y*2~r8;IMp1bu&Fir
zDx_VG5b&r-|2r*V39+@Af9C{x{G(EfU=93pFftpt_<tTMQj>|V3@k>#cE^wBYnM+-
z$_J^M|AUeOrAkpee?F&+6W3Yy`|K@Pd9-R`VtQdQ367{p$jE>!@bU<5A{gt2p=xdT
z`4HI90!%I#_Hiz9OIgo`|Ctv~%0UD`(H+?-$sa>#UB+h_K&J?Ol4bFEPGGD*`~><B
z%S44`vfBQGR4j;z@<$TNZt-g$@0UNN&6UDB+9zx+kd#rG`Lgn~n1GUoy8Qogq$I}+
z2Li%@@(LL}UPJ+2l7g|2;K&M_O~a|lYa{K#HIW!{*H03m-?om|R>&n_J*fPE6djul
zo*XZ0i>D)qkjs_`FhM%@U>|W2$)BIK%cqWmVf?75*=JgY(4k&2v2q4R?+Omr^;0lK
z%XCG?lzvD`O8j^hL#vU@=FIXqM;|PZPftw?GHV)P(i8EqK%X^K$gj0f$oM3AFw~pQ
zHx~#a-{FG@1pzX$x5o)jpvHLijm;x&wzKXEi$!+tpP3l%I!ynOX-*FK&cG`3H5jyj
zQK`(<SgC7Ojzu-OEr&A)RjMK20|5VCad2X!JS}ZWoUav75q%Mw_U5P)jCS$y@PYld
zLI2<2$tU{?UrsM!Qe&JCF8qY7LJ03?ysxjXL6kiK-wx~%6$KB@KR7k=Rqzz$r_vXQ
zk;gX`7_YO1qkR$4E1a9%F|7^{mjEl}hqNwlpr?e~TaZDa$aaoO`U*AdeOIH*N5n84
zmXCjfi<aZ)3c^}WFH&%<uB>{Z^GSw*)J5@ZkE+XS60g76sQ{}A`CTIt&BptilgUz2
zYK1CO%9YDB0yRTjx~I?aVh`6*B^q`gOA-J!w*rpr(9(sGx$w7GDy6+C&Py`*@|L0%
z{eK^<$np&W6nKt{i;YJXcR4M-`-1lmaFUitVLLh068+*v<jqsMYrM2^GnLMdOu#>r
z&F9+?gMpQXtbL%6`s#@HrY<x-zRq@rr54OTVAKlw0dimZ-~T=!)8)>e1H2p@QD@2x
zEOPqVi<8(x-f(BXzVX2b^bw5KU7+OSVKG&-Ff^pt68xtu2iyzX|7Ak_y8y*+^Z)g?
zdK^ty$p8Oe{|8J5zk2-hHbBYzpMO4nPl^5df4Dy2hpWBQ?LIsiN;8v>ERWFiw>|98
zzyI`j;@V|p<pi$@X6odPh5aj|qI_q#ZLgUi(WH-El|KP237*IO>9d`q5)HU!7k78`
z1!h4GfFwK~^YK0Ws2Dk2Z-NPcd21zg@Kl=kCa~H9-7R<z^uIpfp2$A><^0cyt!Zf+
z<6&U6$|%Vy)|jpdr_WkZmPU^f^FB+f`d{=FtVo_fEv0qqklbCkYbsv5CJiVBpFci*
z^53@shxLE*U)!HY)gj>36-*Rv&Hau^Zsdx}Ha9l@q~9AG`}BoFJg2XJ`k*0KV`J-I
z4^k1p_}Bo>jw6ZD@3C*dRt>a)m>C&!@(Kz}3}fTsLI_y*H;0HpzbhPR#$c49|D?U6
zNxjY{FCrqMs*1&1fc@|WCS43Pn7=Q)0B);eU0p&VA_>A7xOjNydplPk3esq6i>BJh
z&`>aK2-r<HY(HGC*o#XrmY3DwWMyWe0ZG_^xy&sHOg+L~%P3CM($$>>DP^GG8yg>k
z^O*2OZ$-E&9&6Q30mk8=fR-PZ$HR*U6)o5>a{z*%#Y)L~U&f#zFi3%*@;pj9x^JC>
zL;?csOt!Y5q5p7o65#Le&8Mia@Jsv33Ry%XzOF7f*xY_%zbBf;9SEHg>biyoCm%b_
z0D}Vq!qCuI%4i$dK!eEE#Ty=|x#RW+;N96-*`*H~0tzh(A=hq95<9>Gl&gSb&ZZ|R
zG$B>V&+nzHKoG1)dirvoN)>oBUKJG;8Ihi`KHI-OL0NihvwL?N1380+t$n-TEhl1b
zXqxBBiW`>za0(JO=mFJc0(9pqcISRioZ>RV!ofLK`FUwQDwqEa6=51i>HyI!40nzI
zmb*i^wUKxG+-3wxv9YiAqou;aXc8>%fwg3t(bnD`O>SIN<ynD(B@J;!Sru8d|9oB;
z7gy`a{@ibcy{3WVnePjrThp^K;0-UDCQW=Ww=%UJnJ43y1v<sEXBvTo14>7eM>?|s
zuy8Lh&p@{UjT2}s?F%ZOb%8F9wAC@F&>Bmy68hYZSN(#5P#5-llh~`%!dgb${^6ga
zOX})zQ&SC27sj>Yt6dKk2mLF?!V8cG_u6=EH9zpPOc2cSdPkJSj}b1bO`=Q3s88pK
zdlNsW$P<@FrxQ~YE}JeDiS1veFkN|Yb9Z+%ey3({EixRtakCXN@;mV)d1f#1hU<d!
zfIUb9*<!XP?QKc(i6XT^F~+^9$^|x7W@>7uwFLtM!|a-?i_0bo!oZ*J-@m{12I(}l
z3tqdI4Zj8XP4E*A55KXoIqv^G+Zx%JDrb_vVoOejotG-qmPqE<nyLQKq&;K5Yj{v_
zk>Y*wUqTkz^#`Z()2Sc)$Scb${9jtTEd49J+ZKs<U!e#L-K2MR`j+E-P@b7|>b1JK
zS*gj*eFg>65yZWIpx~l^v`B`2nf*ohDHNgPEHg5yX@-yYYV-7fySvvkHTg}FxP`GX
z6)ly&Ga7zM>?;@Q-r3pP!S8U^8v{F$<&)(W8ei0tI|}1i#>c1OfE~uBl%iy6dOi|_
zAst?>sx~$@ih_zNCwn6-ECiU6^767=67F`~Wb4yGiz(3`L_`Hac*#=}3-yzAwFWl=
z$R*@8U|+Ed^>#Y}Q7JI?FgA`mm{0m>sr`KdRz3>z9tL_#OuDzb`(D(ur>BQgf8aSH
zGMtCEHRqgx8!;)3oRSp!$pClQ&=A;2qab+xmIs~2bZ_JVjd~K;MJ7G}yrW-Q?PLwB
zt5np~B?t%t0$)=az7bs=t*p4;(z~=i3$C-6Z=}a(vs(3Cvexg%GdHH1D>o3cJp<t%
zv}z!%nPOG<bMtEniSuJ`?*p6<x930$h~lr@sN0n(6yem}$;8Uq)YDr}@Wze$o%X`_
zy9*s1@#Mpa{gS((lcN0m_`5#VFKKCzrk<7-hpumJZ7rBXnCMHlySq0xH&@9eCVMv4
z*1&G*z)m;wna5vvc6(M<HO@Q#<6~zjDXEJGzdD?~?R_q@?$s4?*vw24Z^ssHs=F&|
z9GsZ0h}Z2(ikn;bgSe;R>FJ@Nq1G`U)^M<tXB5dvNzSFV;h_1Msv^bvlWEi)Pacp3
z_JX@<>2Dz0`T0k-sp0PY{Ooqk(EC-y^Nfx_DQS6gu<Z4I5R`B}zikVEnZ9GsntDLg
z_+;q&-$lh$ougwahzVav8suaCQPUf#J0qDG7zp7^VxHb#^}{4E?CMo2H^d?(qMSO`
za{AdZHAF$h#nlK$gNXcFEF%|qq^=8+6K7&!VJT4pe2UoYIby9)aBSjjtS;cI`)0mZ
zt&!|pxd&3=%x;$tANEx@wnl0nq(#Y2VqQ`UYA?R*m(m9oJdJ5#^&U>wPP<#Wi3CCE
z<3$<+Y6a=((uwINCMMRI5iLI2F6xsP{B3||&g37zxp+7fp~n#S^Tq48W|w+xNLt+k
zu}id?OoS%QD;5P4zc9EhAK&NC-@mmcbM?5wRu5dZ{*M<$yWLN)uMD2Rx3vHu|1p5&
ziq-!Z5%?_a!v`=G_IfR|+3&}gBWvi=zH%GcN!|Cm?dlX2*kv)Fe=m>&4sy<~?2Wel
zoQk^O4JjFnM_}Y>H8PGRPV6>Ekg00DO&(?l`?w1VVEi$!t*zU2KmQv1%3ND}_`5$4
z{X=Lsi$q`Rz~lx+%17d0VTDI_`hvKQSM1}UZXg2|)aYSLMM_~|x_`T;g*#WkiWp{k
zTYn3DKwmc9{wbIDXyS%WM+U#eZ?`5<(+VUoTVq4@E_Y<q><PpQ@S0m?YqOP^Sy=$3
zVkj@i$4gN31LEL`VvVN|wTMXahr)W*1CTuf*PPVJzWee@$m#YKw5hHA$V@j}TE|`o
z;0p8^OS7#zYWB%^k(vk#Z6f!@Py?cn(3h6!>3H_J?&<i}Kn6wzuZ4sP-JQ***34E3
ziy3!dSN3R&f0ytKmyXI&gM^0WtQMcRes#42z;tjPmO35bZypQvQ@sY)lXz3b8N~~<
zmrJy%+=ijgAk4QcWL0oC7KWzN09IOS?s2TWsywq5Er!h^%>rre1e1(Ut*wz;o&X22
zfn@fZ!zIs{59;#ra?Fg}h9WEqUOnB@+a=`&K>T=e%H@i+gO9i0(F%D46$RD2GV9$O
z0+s;iOXxi@G*3NzeSOE3<yrnzoV!X{io0<^ofq3hIXOLQ&7-4{v}E!Vlby+zyHn+5
z-NKNhyi~2z9^L?FOmuVwwzzBleDw$C=&ERW3nkp-N~d$+X04Z%1vWt>Yv!YB$FOIS
zVN3RfKJ5sQp{bGE$$I5}^Y?m*sIgzi4A8IA%BaXtG*t{J90L<iB_2l8H!Q4P)nXqO
z$WF_KhxMX>Gy4=U@UUudcC|?%%qmW6VN_EMB%>ylkyclilOw&IH*F6gfA|QDm?)hd
z+|GBW+S?Nc2Hp{A^*(p?7W`>fl`!KMLH-1*{59*oo9TP#c4=qlYU%dLW>ZF4neoBF
z4Wt>~(9m_bq*2F~lmW<w$Ktxu(pEe?JT9s(M*M6-7KH^jWvILldZ5;tsc-0FX@9h;
zw0-(=ttchT#@3N980=#*x-Rzb8Z$E+^GdU;#D5G0TaE*{KI81%{Ms`-Aq5Qu>f7f}
zBo$=VV#a1?IjqMEn+p`4$Qm>}f6u~FrAti8HU+fMo6Z-4u>;15%#fmz%2!}vHP<jR
z`v(iA2IuYtgfdLf$=KXVlS9LooAK9PnsM)bT4H2)N``?!Ss}lH5MShzC2V9QmSyIz
zWy@KoK|@Hg=e!z3g}lXA08Gz6<oX@b%Bt^cphpK6nC8?ko?OZpkkzz+#`D#aDtM3Y
zb#R33+^`<?vx(w<Q$ttoq@o6NsrB_~X<-^LZ~8`tl$y-%vS|`jpyxN&Z_dw{2PdrU
z>`F_k@W6_aQ+>|;1eE`24;A6V7#LqevjJotI|ndo3-7NkTk@#``-d%_{;t=DJyD^%
ziEMtq1AXc~`9wWhc4SRYkJANt3`q<R4<CpKL#Y*oqCe6|R+AYQ8p^Ax@=Z)5qeE@1
zt;fbj<U(T%=t%hZ_+j2MFdXNm5JDAnwH@=e@JanTZ3xSgVZ<bwUol@+R#&qEGLf8A
zSZE{~PL3c94r_Jxx_+>7x3KVCWrcc!>&<%_0_felcQBs?gq=Fkh`~UtQfji);gyl;
zy>pk9mBmX)fSTL&>qCpc%0_}89_~W;EHr$%7p$%5Ru51!fhGp}UQ0QNM3$c39-F*p
z=-v_+Wir)tqdCrDRb+TL8UC8;(ReUA`VxKrsZ60-`Ls=|FJR}--fR$(ecmpw7&S;>
zHJd6d&Ii3xZ2OD2B<GrvSC1oW)8wz|%`G|7I|R?uWDeJJy%|p~%-@KZOrv8>Hq?(3
zvOD$oaweF+Zo6gWs}N$Hy1J2?Hb74>{44w2!|h+oX{O-W`ozR@=2J|*#|K{Aeedw_
znI36>td5x)XOllmRaHqa0W~G{6OY3~I_OuGQPM%LZjB#FlDe7#4mz&UG|L=FNc<Hs
zRk|!THs(TDz3PS`qg+Qmb$Gd+u;PACxmx;3pw5hb92Lk})XLXK>?vqz5$eVLbCNk#
zQkGsGR_V5J>ghUgANCaGZLOT_+tRA2;H#)4lpXuK-}^N>+~3wB<EVJVMMbR8+t2T?
zTClHq-+gFirT~}ULoAbNF<W0NbFR-ezPzcG@{_)FY)lRj@#6{gzuNo-=gcy}*47pW
zGH^b685IdZ{Fo*ao&b7bh50P69M2A<mL&^~X7(!V`M*FxXlJ_6LI8rrCrZ?h3k2m!
zD5z*@sY@IZK|t)dnG_qkhb5qOp}ldtv6OxX`NDqtaP~LEct+G3_<<}dtoNDp#|X+*
zR?5rEM~6P${?+I@)(QLt^>cPZQM^ROgkEO<GM>TuIdD}^1GDIOZC#dDIDD<h#P1h@
z>b7zZLe;!YOvxjnw%EE_mR9p*+W7hL$BT&l9;d7KE)V?ECncVbtl#VDLdD8%lg0_)
ztZ&3QeS`&FQ-wv*16Ba&zJglt4X3_n3YW0WC*DAD?r(<>Wds`G!Pe5!BAsRDFD=(=
z3Mm_er2fIt;j-5TYy;NfvNPJ>INCTJ4(HPO0iYfo7M8V%;@=;-H`ep*aMABMrb3A_
ze>XflyzkH!GInKgM|Iy!0(@TAz+5gxIIY9nq)&VHOQ1`7#U>40kY4B7ThwZikpa|V
zrrLDrs;x4ph=%Q=)eFAF(Yixfu_Bp(mxotoq`!Z^I;sGG;Oysrsw(InIMz3o8Pp0i
zxWSRb!&#UZ0Gl1L-XZKao#PWh9f$p(_g@Rr^18~?gNpXbb0nJ1;w{+vrG=E@(m^6-
zR>jRl%BQiT*aInsd=`*5GW6q#{P=jj>&-AOh7}T01R2|FAoPsJej}l_H&@xy*JqX(
zHTGF;E}4O0^=x!>IG)4tct9j(ZEX$FF!NQCc)u$kA7vIVT0i6#78c5JKP;~@jZL=l
ziK{1z8~BcFNlOm{XuhJV^qku*F~bGS-yq;8DJe<J=S56Jj6nl#pqQwts_OIonVH@*
zFNNO~;0jMq03Qfn%UF1xE|b$WG7<pxMP`LDE)R1#yM#oeMTTc1WcXl}?W?DaH5y0`
zlP#Ch-`!e8SI~5>`Of>L#Tlqd$JgWVUt^7yBG`WhKA8~8;LuVWJ_g30lJv@`%4eSi
z<j0fP(N{^#OdlRjcO3BXA;9_MTMqgAGvOKGh-_$#jGQ#s2oHx>Ful>r`}u-!oaNhd
z==CSlToO+h=;ENm=JK2-vlKgeRcJ`1numRDYfYVj0YXDZr`6ni-7~1wuAR;ay#*VE
zbnQ11@Z=;Up=%$RuFv(v^Z}1kf~CwcXt}qqk90@b274&v4*6R|c)03s{k}+{2q^W*
zTQZhx%$bvye+_idm-(W9y)x6sw%vF>0E^FRLB^Qs{nMvxAXA(wpMo-r)A<~j>bwAE
zQO-y2K9E<J=7Er70R@E#NePR;(KHRZdKde%xcG!IKtBhJO?i1Uyx@5yG!_g)@2`Wv
zgv8baBjOukD}Mx(dv75jzAv&1r6wCa6Bfr-Rz2N4am`peUt}a@2S-OYbJCWU=sUtE
zKHK1{%={i2a=*NO&ceiDHM!BY!#XgM5)%~@;Ol$3m66N-ZILd6$Yi`ECfs--P(e&g
z42C@TD`@0IB_+j$DM?B`6U|PnPXtuMeE)6<ZdzOutuFP({QT`jz*G+*Z0hUnbvUW2
z%l1icwLJquDW>-VA?M1sNj0xF<OF$T7+JYB_=Ws7ek_BGoXm}Z2CHXz&!98r%BCV!
z%ailHz`(%kYXo{cF_GUX#lgY93wD@raOAImhTz_mP{VER*U<zjzEV@MQiBtq?(4E2
zF1Ph46xV~J_V&J~r8Uy3@;yrx5*EtO%iG%Cymeg{{jJXHRX5)z67%QJ-!>zrkg0MU
z2S=&M%M+XrV<7d!Ei(4-!W+2`4na{-d3$TSyQ^z#VqyUa85sqop73VzwD#odiyMc&
zi{*Km9qruo%nI!_BpNX>dORBP4;M>+YeE-8XA@r^B(cLZcek%EU!4@zIwvM2f_7KS
zv|o#6Mo3jtTB%XwRA)zC-ZE4C#%l?gktI&2!DJ30Hn&>`VdL$yMa_kkj)-u=6V(^{
z<3+zLaA?xcQXPBm^#REu97SL{0O#k6E57*#e8&rO>F$WZ1XlLLa3`mJHs_@uRnhDI
z8g5`=_YL&ruwUZ2KG{&PnEg06+w1ltrTEvcLETG01Z3>cnV+p)ULFOGz_-B!J>C7!
zpEt;<%A1?%nQ@ILiVtn{pC#^W31ysp+}^RL=2(_@;Byp1y(lNgfrCro5W+Q~2ZSp*
z5t;IwKv7;-H4sea-Sa)I+9Kupri7f5lH;e&uaw^zo`31No9eFXB-k!~{rc4<^IF9N
z5d!;JYt9Y4S193;;Z&5g!ouRupCbha`Ngrja}~IN)?F;UPX5j|q<Q%EE=wfld}J6S
zV@SQ;$#uO~16)dt(~5wSW81Y@!t@zZxe8Fc&EGY_LA;6a%qH2TewGL74PfgB=-ikX
zm=-_1^_o^x)zFDf4@hZfm}eanU3jz|<<{uA7uo`Ftg&jaUkS3PW=j}I7;w%h4H%Hg
zBY3K-{Xflp^+Qx`)a{@GA|Oh4N=tY9f^?&FD)j<Gx8nekN-8A{GAP~BlF}eU4?{O1
zJ%BXaNAG>_cmIUTZyXMDVxDvM-fOMBw&?|iHBJ|re<OKSgbym$&VM2Mm@fR?8con*
zeX@-nD9`=+b^LvM2!otzL%@zCq=1X_cx=>WEw;KACx|>Ke-rt}`Nusb&Fo!|wv7i6
zFz4h1DR;ftJJ)Pa<V*Gtf|T$CNusEzSgC32?g{*2+Pl-|!`9qDKu75N;o)z=^4TG{
zA1_Z=*mvDP-!xX>jX4TX{43paO+I+MvkK2`ipIvHz|d*P_B?=chv;arjiO#PG+y0g
z6}wtJ5d|Io(}Ug&1bB@Hs_boHB4U>M-WSXHSLeqpH0C!YyRBp-pUUsY<o-)4g8HuD
zCIC=Kz0rk*?54)B&#bMHG@{PTwz<2rOJyJqNO!*`h{*-<;59JQ3S7FJhr@6QD4%$E
zdwYwOT52N-KwU@>h6sLqKJ3qRJWi=nAAN3LC;!+*(?clk}gr+$`Y&sQ^$@Sk(KC
z^z`(AhcHr$z2^G#kZ4y4fK$U9j@<;l0+BCORCvwm_W|#D=D7jt=8~Ko6|QN`6vk85
zWgUHUXRCizenmb@EKWBCV=(&?u$)k(G%|L=F}c4PrjQmb^<Op26+#j}CL~xkcuauV
zx;4&QqM!tg{}&21TNv(jC)HB~JJ{2J)(8LV2_rHGQW4m1XqPO!7CHyYGUFS_4G_>p
z>5}$I!hSz4e(emuN-8$Iy1Z1G3_d#qJqi8is_NS8U9t2s^7n?R2;w9>U&G@pzIkm9
zrna6iOu4>rP#xLac|~}}6fso+L`&N-8$%AH#}YoVjr3bYX&+M0MmkEik^oTMiUFey
z`7B#jwE~aNMH(dgdiy6QQJ$+f*tm7(gg@Ph7(7-dCs4T9xYhX;*ZybCNFhNX)i(eU
zq&1ei!QjW_uCA>KRaH%mMMbJd$Hb`eP#7A<nyd<{@Rrh&0b^kO=lZGuJ|IT^-tTN2
z0(ZDCPOi$bTsu=%ag-33#$%k=ZRzC#&xv=OzOJTu%!(^1J-t?yjg9Ta@s_>SzeYMb
zz5SZ%>elA423PJEFY1b&92?Tzb9dbEDR?h`HdNc=iTh$zc^=x>*r;?dBtKqiq2o^-
zOiX+}=>u1=zQMZcBYb<21brbGXwsOdq&q1pN^5-SvbEa!{m0JDUIsZ^g0QNRLGs3p
zI)1_*zvCt3Ly!$rxCn1EL>kozg@@Ys`H92M*4N3abrHRaDJ#_ay1J^4rDiP|8<pET
zJF0JH>N=*zLP8XbvO^t6$rvblvbayJ`L$Q=diJh6Yy>kizKnA8|NHh~PwBV)D;Q{<
zYp2Uz2r$(wYZQ*em{bI@7#bU=D(S9T{{~=kn5$4i-|O@b?N%!95pr+ey+BS^+Y@aa
zFP)y6oL)C)WJs%8Ho4H95SE4Y5!}CzuJY!FEt?S$wP!XhefcuWPz>ND0PIR!U5t#(
z)&Q?;WVY1C$_mRMxPtY_y37wJzjtP_%Pd=muAMwE)_NAvxmuaBU0F4@TcTAA4x3s2
zXJ^~S;UT6l$9yfgPFSB!EQ4H@&A~-QJN(3k$#uXonb%r-Hqfdy2<Z)M2gJbwJ;`j|
zhkr<Gq%J)DHNhRA{B)b&E}}*LD<5?=i(EiJz_#&(NDY9-9;p{+PlAPv_T=U+Hu|R-
z2-{=7f6n_tPC?-*rr_p#?&)P<Zf5S_{6Tq_{u&e;B*et%XBlA5`WlG#QR)Lz>W*KJ
zufUMAuF7ofH~{4RT5j;%HokS9P@?S$^VzS53#Muie<95;0D3R?4JAdzgOUl0U~8((
zDo^U}?H2OwkJJHjWMoNs{9CWc7#J4bSciqpu}vT|&eB<;$lB#RPAmAEez^^n<&|VG
zF{vU7B&u7N%94(GP4)C5pWB2@<9SelPe0l~HgKPx)3espy3DX?T;NT5+m&<~?V1^E
zuuF->$y|+D`$(@m6?I1WyIU~4?unsM_CWX_3!kP+IFZNXK6=EO-XJLtbSqc^MpdED
zQ+Jn{HO^Qfh!QXtCpSao(P$1K9ED&+Lfc;L5M4Dl2S?rk9u`Yz%PnbZAAmb{psQC(
zN>sKTdMiF9uxopGcvMwaKhrb#^7ANFXl|k9$Kw2qwtDhg-Nd4_>G8tYG4n4?K<Zat
zUm;5cpl=s@du2p{teP4L`?QpVgoL=pxETL~{GVeP8K65*eXA&C4&8d493THhH_;23
z1)i8^`I3@ase+D!l#xS}b}v0fZ$g7&wL#Y{<%Faf<L~di)CqlNbuM&$g8a0DaMWqR
z#mR0=T{X$yqujqwl9`dgruB7C&MN3za7UcnYgb3fU?orog&-!=QLgJl2R{L;Ab}8A
zUw_7??c<T=Xm2kzIAtU+AFrgVHTI>_aMNfI3>9Q!XKg^Hj1>F2s&aCIP|*VhG-;-<
zi`zz)mhLPvDUmAX>yB614*2aR%!H&E*>5=Hg+)ffLBkp7Yy#&%OG|sc^|QCP*DVo9
z{Xs>hn~^)l;W++tWc7IDSH7YJ`G&(7Jv+PSb5%7B;W|q7YmRO<w%BK1#3dxa<vbhH
z@y%L)YVma_5=}-*nifl|ir{dMbaA#8g1_2oh4g`*g-6CA5tQiZX>h;Xy)s^kI+Egr
z+!qo!`;Ic~MG^i*x<WUHt8Tnr4Z6%<p3T-6y<nj+rKfm$6Eqo%2JrF{^7#|r#xbd4
zFgY%_CT7#`W5VGG22h2<vX)o^se`AdewcO@dA%yILS?qVI@|fB9UyF{+%0sULDp;-
z9U$0V8LfquAhh<e#DBr-L@UrFux>5sQ*s`zRHYE>4$AX)Uu*f8LTB|OtbIzNHa;js
z_!(SYM=<zV4ENd&kGOLoP0B4=*=p<t-6$}mDNqzuY`Mi-?=4iqtZAJ~Wq#nxK`l*%
zO#9LpFnfN^!%OS#WNCk$3`3N$FK+CfRSKPqFIMkcQGv);SL>zxYnmF-eKKJiZ@<<4
zz8I|!hwV3&diHD=6%{);Tl$KKis1P(1E2&yKtPI_M+UxGWg8ws2&^bOIn{J^N4h7I
zSH@%`K!?XvY@IEGSdV*VTtnajG%;0N5kxBL`=Xb}sD*hu8-DaIzWIY{Yr)jLsK`Lm
z=I*w(ASR|hD;3ajU;iWJAuX-E2e$8PoL|32PYDo1KL%LrD`7~NC7Qvdz65sgG;v5a
zgL%v>O3_n5#YKfzhCnDFIKW#cfGi|9WCP|8lt18rqg2t?x6ifRy=I&3Pr#T%2{j*a
zJ7r<9x)RH}vYVjm(NUQ$4mFd;{H&I&sN&+dxHx`oX}5y1LQvla&4uErsr4_DX6BB8
zpxA0&!FV_)m9@68wpwD`RC^PIwM@#&$0yhQdQjBWHFlmPB3(>@jxL{zg9GFm!NIuz
zbf@xRKmD!*v-asM1VMdgr>~*n#(qLfI6Q35qS~hvl)&NPI3Y|Rcu-OiT72^p@ZG1t
z3kM|K_*wGx5s5FeQ?3|~H{tHA`-;g%9_;*i!CL6uN$9IeBKpQA<SMU~TzddxIhDP!
zQ188lR}Uc>(%zBK$L6&qt}`!{Zh<+tsnEf6IqgDihFCLTE-qy{GdxpAla?kQ>@t`p
zqTXwF*BT8F{C6um+r;q+Sc?w*B>H8lx3S+_sXU*2-umjev8qbV#H47<<lQj4)a(5m
zW|SNS6VqAy^CS>p8RksQ%r3#STL#(JAk7_!?B^2fqEz?!jqB;Bc;gud6_?a<wnWVA
zEqxcS60rHHlZ6fL!PWKYZ3um*?BWoe%qvo)3CP=R5MN7J1`1;6Wx}OC)C`-d;>A%R
zq%<@&W9%E_bD7bU$4i4hLp^3-V83X<&{l3PhaHLq2gfHQU_1_;wzkCW-VgQl(Thf(
z;LmezWC1QnLp{Gde?ayhVmXl-Z<`5uef^~O@2vv5FqlIuMxGrsnClnjfP<@w_wHeL
zw>utQ`GQ?EjeWywv_})(?;R6PVy5v-hk{};OlrV=h!^gC=pLy<czF1Wx{f=f{4^q5
z>#}k0egj>n#JVk;-_ZJF&}}@!z{A3ant{z2>^YnIuix|2gfPnR#8@U*#m6r;c7XP%
z=;*hLDvx=}{nAtME?uu60AtGi-ubSBcDWO*YOO3%9mZX(I3&?yFggIZ5YK=0ujcPL
z=fJ~(_r)`OeEiTS_*)AXmuVdo-Dn%NACbb}?$U8bYS$}Zk2heiJI>%v5J*FT+BaPa
zAo+GqPS#FNEES(sI4G+XEO}t2GyWFRY?}7Q)po~GhUHRM^`MIU^ohy0u_|5={S%SX
z#$wn|TKC<_ImZ``20B$1_$J`@?8+wm#T&;SXHFS6`-(+E#qm&zfv1eg(5aXWcO)d}
z-S6L^e7oHS8h@n3>`)DGhl_W=DR|9`N($*k`k`z6NeTJl{ryMVZZo~pjF%pD{;=xx
zMB9_i?1*8*K5zkq-rg_Nii!*G^_Bvu#>&z|zvH4HL9(^A?ben7$L95)TU&TzLE0b2
zYA;i7WkXn<^mb;==ndsTD9oxc!e1Qs{&}PRPgai$XXJSz9HWr$*As0Nu{Bp*m(xA&
z$=ggpK_T~jZpU1!EGQ)TAoA#8rpGQCz1Q2F^AAo(BGT_7eD(|=*IUc$Z1ryLs>;h1
zq2=Y}VA`po_i;$QpP{d>ZcuwTjvYb?NLPK9pW;R@009^U@G5<>rB*XJX<5&C7w+))
z9-qxOgV1}AqXGzDkq@iEIZz<I3~nmywH>e;W;2T$!*O~FpFPvVCVac{Q>)tRrM*aT
zX_QMFH*?>PxoIFYr;HbQtAusK(-ZrcMk>eiVia_(Hr^oRB+kw@`7DP_E1uZ22fvY&
z1UTvXx;4PeHl3Rk3#1TZRJT0l%*+8+h!w=f<p^AEpwN=SLQb@(UR<YgT3R}~IMMK{
zg=gMpKeJ&;Lw#FT^aF~Gt35WJIcg<7zVgK(ASS!gAlwoJ*nKem*Dc!r%Msjx=(DoI
zU@&tTb8LVA@X)-sZ|^$!FRT;YEN7N|+o@$;Jzev-7I^9cv^8`Nz<HgLn)-_@ywUb>
z9k^^fM@`f~9_8n;<>%2KACi#l9PHHLf2-ST>jlywcq72A>s;r7-aeG&3W#?A)r|G0
zAQ$R6#Q#e0Y4;9bKD-4-=w)Q}ax&>u#0*i1<IfE&P4pEN<@JImeAm#KyZe=QzzZ8E
zh@P2Q=BeB$>=75YNfY|=+`0t2gPY6GtGX!;JT#n~g)?=oz$XkDkMUr8avnvo$!Bqz
zE$!dC;Wm6s6B#LuqlKw<{~n&Z{BqX|<Ld>|8Ubq@UtalaZKwtT-qhw1>ON7Nj=}k2
z-@&xxkn6*{sab+Hh^iutXPu5(qsZf_;kG7$q@+^RgH&zG+ZVPXA|l@C%39|+wOc-N
zJjPY_CoX_JsMU?hPLI}@F)gT4OTm=Xz3EC30M6st#l{dwPD!5EYVSja?C&6%cy-GE
zJ3s?!<M-6TR?%Ah@P*|<uha+N1o<^?ZVRwGc*)t>Vja-D-5@Mr@+)~8cXD#_n}53d
z%5+VSGYi>$hTk9J<4M1fgtoi_OS53kME3CKZJ#`;QbzWV@87qLe8d6Mh>-v2$yZf<
zX-w^dy1ckYQk$^LNhkq<BqizJ!(e8yi;7Gx9Asd%i1e$5h=6^-vaank|CoW9s)0Vw
z!K;)hT;JWN>tv)rqfbO9D$K<62G$`P&>g*WFv}ZNmX)>G{^_oFqp9=Z;`UTa$OZG^
z`Gk5_jqtdsy%FsE2)*G%^Qmb1+-mP4I2iQqds3(bUS~iH3XZO3V>T%xiP>p@S(G$>
zVn9cA#ts?vD-wv!=g+>L*xQiB<C!}1F@i}_LhWbILOIK)4P+5l2{PW%1qHotAAVPk
zrqr^dyGPhuWA(=H`r^a^Y1#QyInFs<o17_hIA103%;Prdxx29L2=L#=WN;9Ex=9PX
zy3KOeTT=G{*7a>kNKQpX&H%MS;}}9Y)&1o2n?O+P=%D`H$<3GcXV&nqX(?UH&d)!5
z{0H5s`h;1EBu}gthm0Tk2Oq<AKkzHxXaETd1+STxAR{BU5cl~IoudC)M=rCgh0rk5
zk?V)-kIJGr`zGaVj~<2EmkO+xl3ZGnh~v{s=^B(|J1J?kYiKIJk6F!ol!Yy?pwQQ?
zQvG3P+VkUw1gEQc=5!QCZD(g*cXHW<Xl~A;4CG$`TV3w5igOe=rM<8bHj3B64YJ=v
zf=PuS7j3q+_44+DFbYXsOcWbtH^a@?G%BQC9dV!oxZG<07*cgm@;}+QOcnCn7^E65
z&lZniyT{y|2aBb4XbjkyE-86jeippWs{FI{Hh3J}nYjsU?CcJ>-Lg>_?QFbFEp>iB
zq+%C{ClW@rr7+lWnvS{G*oV?B%K3XuUDZ!|f7H6v7IMNGGrrM65|fiegfV_bkJ-K?
zSib&H+Y^z>vlGP<Ei6<`Z;uBp_GCPA?KF9jOr~Vd$;$~;G<--C@EwiN6x@cO{r#2W
zXdkXWDKXwC3Wbr!+hS~#E9r$Hev~KgrNr3z<k}Xpp;A&(bxufgrG*^t)UXOu>9U9>
zpS@rG-TahZ5Wlnd9z)r`=v;9(8Y`$LLjuxv=kyq58OFIM5=f6WYH*yehmr@TnV8sw
zpu+?V0Dok1{D3S~)Oi{Jk|4I~8|qJ>Cdga(zR`I50s}Vb?2&(dw&MM{G1St2{18%d
zvd^CnC>hDAUvXy^+h3f$-}TaT=J8Dq+CAE6qt0vrkpaqN5R)R?TywY`*df;%<Z1I#
zPY+W%^8&)Cg^<ul8_ZIid;rB-v^bFgiY^@snR)H@NS5*1z)a<;6l-e=t*2o+`7{!_
z-lQ{)$6LRqvSwdfK40Y@Y55v_axMy4+-c4PIpFmn;?ys87O+k4-=}YLnwVVmb|XTS
zcN{;JXdw4MhX5p4!PORCujHd@eD<|4*2YJaU0HeZR@I!QB^ep<vWAM5Ec{#;mJ9%h
zkL6~Y{oO!@aXTpTm*d{f40;UAFCphKnN8v`27c}~1#=TyD1bY%Aom=yWG-&LqoP_b
z{&{9j2ywZCi#)+Bi%ErMms>`FM(xCJz!Ey(r2f0Ctc`+tdXLcbUs+)0{f%mcw_|y#
z_knl<QUN`yT3meayJ3cyBmo#WG*$_)C`XHXo$iO!*MlJ1STD|f1=k%}RZY0#l`oZ9
zS$Qzd`K-*M6+p03^ONi6HCwbK>b7>aNu(buCId$4#NQ{D7nXZ^q)~eu8kE;=E&<sl
z04i7z(*dW~xM-sM7+<#(Kg^K><n*{s-<s#K(ZHdtjh&XKhvR~gIMY%5CZ}q?A!=gH
z9h)4Bu{FmE*nmV)@Qf}UPch;#P!b94ZYwd3dw<dL+`ZT;2P+i_2x^3dDR@o9a{_Mw
z7XcXR3DI$6mdMu8Lf~PEI3O)iN!-5M(-XHe><vF+TPy&$I_EfNjcm22N9NwW#!rQX
z0`9po#otCAut*rAm#eBSDl0AU$u+X1Q;G8-ZR3L_sGiB#soEbGS~jaI<Imyd%?*d^
zv6Qmkl!->zwe@Z;50StU4x*#0hdGv-^e<Q2Fuj`jh0e-?@CUXBA02rin3Skog@Co)
zN6a>QY6Jnk)A^5yUYfv@#oR|qkQgEgo%UiE^cY=L?eFj3!@n1IdIq9eLsJ77M&IIO
zl3P^bwLWN^@1Vs4Y3Yf@#=&W9YN7$~e0)4FFPECCCSjwy^|zYi%9S2fe|G^B9i9E3
zx`=^dF72yhR8kvvSs5>w6S~F24pbb5I!61es%lfErog!tpMq!GD1obgpw=1C%@;l^
zObSXd^guSl%{^1xSj2TW*Q8Fto7egp4>C#|zwUe+2?8Qhs4RChnyJ%aUYgH!Z-;el
z?V;~30iC08;Joxiiz91S7YJ6^7I>tHcbaM90tLnNYPR0_#_-MVFWt&<#9Eosto7g>
zcAC8(=qP!)wk5`A*-B+&wYa)!<Lb)C#f$->J<znAGBpLj^r~{y?$#E_EI+1a-*5=1
z=C}U7JhKSFrLnQIaXFl~SQx(xQNqD7t#w%wk6cMC1gsD^-7{gHxwa=aGp-bXRZ>(`
zyt+ay{NVY)J2XVWXSq9jNbPFxyEa|(wxA#=2%^{}Yg0Lgb^9EU5-1evf>vVLbM@`~
zlgvtM9s-=rljMZr@?vD)k7fWVAUYrEJ}|C)skl+QF*ZkVFY;Z~ik2mC1HCBcHY{1K
zE8;ufdq9COZ+_)`={FOP!2nXz8ERc_mXQgEU)?dYsfEs$<mMd>^d|u)fb?27Ek+)$
z^Y)H3W=NCIPR7lSJVf08{HU*X$!_V>Cvt}LAxpDpax!n<RgP*P2tTElT8C|i>}MRj
zXEH3a*x}(mJgC0%0L;}x3WmU+`!Cqq%9?K&n4KM50uDxPSrF~!lf*LzwWgxZf4dr=
zDk>-J6#yaF(_5F?<0vsN<H7q7fItB!0obHiHE09DkBYocE7_?Vv;s>5>91;EpQp5)
za!)6yQW=Quiqk)6(^UkV(h4E5!&NdblY|75M5>gfrv4h{^Aoe4zms>(|3<I)<Yds&
z7a#IHZ#w+^>jqzd@$rqn56CN!nNk$+n}Cm9+bX`V22%uR;SLXSrRFw)T#`Zd?P1Ew
zLsH_=TW)9lWU<PPPZecT#>~om%Rve>_C=CSHo(llxWV^kZ@Ym6>Ws{k<Kyl2CnF{!
zYj9i21=*CYLtj&d*=Xe*h&wgU;_18d^Kh*=G#H$;WT(=b9vzJxMP5maDzRvGzj|eG
z!?SZ1^>rfy`Sh7^B@hFHeA>qaCOcQNg^J6T?8<DhM&XXx0XIkhPXWWpk;Q!ED3}P~
zyK-KGa-F}G-~g50u3Ipuj*DNhKk!{>7;>{b8$g0`d4RhgEV>I0uKS{bRn;}#p4ol6
zQ{Sy)lMPxCEiFV0(!E1{&PxQncbQpPsr)uaMP2s-|4xZ8F)<+}T~&@1_Vo77oHx&}
zT&Zx~n?I;sfy3dzutY&O@xzD}Fhk?3;+QWlHZV>}-;?v>Kc8=2jwi+m`BYX``)Z<G
zXx1LbEv!ve%+vD_^lfHI`eT%Ih7qTy%ggcW$#%w6R5Eb*fcr|o=`Ji5;|SLWegX&t
z0%=*luYCl-o+|F*;;w&%6|u3wQ319uEh8f}ONoykl@}Tc>^TG~iS)9+*xA}mOrl1f
z!Z^6N2&e?d-HTRAj2GM%YJt25ydWsp<Spr_6-PK6%OWBcUjl(-US4afobPT^Mn-lr
z@5IQ+3;suWd3gi`1V9jKd3r6A&^KG*KvGBvnD-O+2@&BtVo+r4&MgmrgVxsvxo_TJ
zDagmS1L|$qEnR>`Ts*Wn88BhWWI(VVkk3}1%Ma}7DEKT&<Wwx><@Es2v}E-idm#YE
z!ch>W4gA`88|oX23X4amq5-~3@3|ftHb?&YZG07`W^0m)l3GngjjOl-06N<szG|?I
z0jh1mQWJ1dHV3nxf+7L7?5xxc+|R%%cXC9Mf~gQB+yf|-Z*2YskmCa<Fdi)c+bS3v
zqvz+dRtWbw#qXcLBj+`G;w1GDOO7fvHPb`<KVg2RR%qY$(dMVWcj|dwZYKnSCG<xN
z@OJ~4+!3+$pX=kFE#u8r|8JAZr`<m)wg26(^5+x1`9Rh2Kff2i+W)6l1pIvt|F72X
zzx{t%00!d!90&Z9&PniphJAlGHd25W`+w*8|M#J9Zj+o=_d$1FD?-4Js-o7@GI^`u
F{{@Sj<Qf0~
literal 0
HcmV?d00001
diff --git a/docs/source/changelog/index.rst b/docs/source/changelog/index.rst
index 378c73b3..f16d7a6e 100644
--- a/docs/source/changelog/index.rst
+++ b/docs/source/changelog/index.rst
@@ -11,33 +11,37 @@ edge case bug fixes, speedups, and small new features might bump up the last
digit of the version number. For example, the difference between 0.2.1 and 0.2.0
represents this kind of small change.
-Version 0.1.0
+
+Version 0.3.2
-------------
-This was the initial release. The output count matrix was constructed via
-imputation, so that there were no explicit guarantees that CellBender would
-only subtract counts and never add.
+Small improvements aimed at reducing memory footprint, along with bug fixes.
-This version has been deprecated, and we do not recommend using it any longer.
+Improvements:
-- Imputes the "denoised" count matrix using a variational autoencoder
+- Make posterior generation more memory efficient
-Version 0.2.0
--------------
+New features:
-A significant overhaul of the model and the output generation procedure were
-undertaken to explicitly guarantee that CellBender only subtracts counts and
-never adds. The output is not constructed by imputation or smoothing, and
-CellBender intentionally tries to modify the raw data as little as possible in
-order to achieve denoising. A nominal false positive rate is approximately
-controlled at the level of the entire dataset, to prevent removal of too much
-signal.
+- WDL workflow updates to facilitate automatic retries on failure
+- Added to list of allowed feature types to match 2024.04 CellRanger definitions
-- Uses a variational autoencoder as a prior
+Bug fixes:
-- Computes the "denoised" count matrix using a MAP estimate and posterior regularization
+- Fix bug with MTX inputs for WDL
+- Fix Windows bug during posterior generation
+- Fix report generation bugs on Mac and Windows
+
+
+(Version 0.3.1 -- redacted)
+---------------------------
+
+WARNING: redacted
+
+If you managed to obtain a copy of v0.3.1 before it was redacted, do not use it. An integer
+overflow bug caused outputs to be incorrect in nearly all cases. For more information, see
+`github issue 347 here <https://github.com/broadinstitute/CellBender/pull/347>`_.
- - CellBender never adds counts
Version 0.3.0
-------------
@@ -84,6 +88,37 @@ a workflow using Google Colab on a GPU for free.
hundreds of samples in automated pipelines. This file can be parsed to look for
indications that a sample may need to be re-run.
+
+Version 0.2.0
+-------------
+
+A significant overhaul of the model and the output generation procedure were
+undertaken to explicitly guarantee that CellBender only subtracts counts and
+never adds. The output is not constructed by imputation or smoothing, and
+CellBender intentionally tries to modify the raw data as little as possible in
+order to achieve denoising. A nominal false positive rate is approximately
+controlled at the level of the entire dataset, to prevent removal of too much
+signal.
+
+- Uses a variational autoencoder as a prior
+
+- Computes the "denoised" count matrix using a MAP estimate and posterior regularization
+
+ - CellBender never adds counts
+
+
+Version 0.1.0
+-------------
+
+This was the initial release. The output count matrix was constructed via
+imputation, so that there were no explicit guarantees that CellBender would
+only subtract counts and never add.
+
+This version has been deprecated, and we do not recommend using it any longer.
+
+- Imputes the "denoised" count matrix using a variational autoencoder
+
+
Human-mouse mixture benchmark
-----------------------------
@@ -137,3 +172,14 @@ v0.3.0
.. image:: /_static/remove_background/v0.3.0_hgmm.png
:width: 750 px
+
+This represents a real improvement over the results published in the paper.
+
+v0.3.2
+~~~~~~
+
+.. image:: /_static/remove_background/v0.3.2_hgmm.png
+ :width: 750 px
+
+This appears identical to v0.3.0, as the changes were intended to fix bugs and
+reduce memory footprint.
diff --git a/docs/source/reference/index.rst b/docs/source/reference/index.rst
index ead0ce72..2f476294 100644
--- a/docs/source/reference/index.rst
+++ b/docs/source/reference/index.rst
@@ -185,4 +185,5 @@ The information contained in the posterior can be used to
quantitatively answer questions such as "What is the probability that the
number of viral gene counts in this cell is nonzero?" For help with these kinds
of computations, please open a
-`github issue <https://github.com/broadinstitute/CellBender/issues>`_.
+`github issue <https://github.com/broadinstitute/CellBender/issues>`_, or see
+the `semi-worked example on the github issue here <https://github.com/broadinstitute/CellBender/issues/299>`_.
From 1198d8f0b6b56bfa82707cc3a593e7ee28e7034f Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Mon, 24 Jun 2024 12:58:37 -0400
Subject: [PATCH 18/19] Update wdl for benchmarking compatibility
---
wdl/cellbender_remove_background.wdl | 1 +
1 file changed, 1 insertion(+)
diff --git a/wdl/cellbender_remove_background.wdl b/wdl/cellbender_remove_background.wdl
index f7507ec3..d3a964b9 100644
--- a/wdl/cellbender_remove_background.wdl
+++ b/wdl/cellbender_remove_background.wdl
@@ -90,6 +90,7 @@ task run_cellbender_remove_background_gpu {
git clone -q https://github.com/broadinstitute/CellBender.git /cromwell_root/CellBender
cd /cromwell_root/CellBender
git checkout -q ~{dev_git_hash__}
+ yes | pip install -U pip setuptools
yes | pip install --no-cache-dir -U -e /cromwell_root/CellBender
pip list
cd /cromwell_root
From d7d228b491858054fac7e4780f575bc366bc9165 Mon Sep 17 00:00:00 2001
From: Stephen Fleming <sjfleming@users.noreply.github.com>
Date: Mon, 24 Jun 2024 13:00:20 -0400
Subject: [PATCH 19/19] Bump version to v0.3.2
---
cellbender/VERSION.txt | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/cellbender/VERSION.txt b/cellbender/VERSION.txt
index 6833eaa5..9fc80f93 100644
--- a/cellbender/VERSION.txt
+++ b/cellbender/VERSION.txt
@@ -1 +1 @@
-0.3.1.dev0
\ No newline at end of file
+0.3.2
\ No newline at end of file