Attention.py

# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""A powerful dynamic attention wrapper object."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import collections
import functools
import math

import numpy as np

from tensorflow.contrib.framework.python.framework import tensor_util
from tensorflow.python.eager import context
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
from tensorflow.python.keras import initializers
from tensorflow.python.keras import layers
from tensorflow.python.keras.engine import base_layer_utils
from tensorflow.python.layers import base as layers_base
from tensorflow.python.layers import core as layers_core
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import check_ops
from tensorflow.python.ops import clip_ops
from tensorflow.python.ops import functional_ops
from tensorflow.python.ops import init_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import nn_ops
from tensorflow.python.ops import random_ops
from tensorflow.python.ops import rnn_cell_impl
from tensorflow.python.ops import tensor_array_ops
from tensorflow.python.ops import variable_scope
from tensorflow.python.util import nest

__all__ = [
    "AttentionMechanism",
    "AttentionWrapper",
    "AttentionWrapperState",
    "LuongAttention",
    "BahdanauAttention",
    "hardmax",
    "safe_cumprod",
    "monotonic_attention",
    "BahdanauMonotonicAttention",
    "LuongMonotonicAttention",
]

_zero_state_tensors = rnn_cell_impl._zero_state_tensors  # pylint: disable=protected-access


# 父类，其余Attention均继承此类
class AttentionMechanism(object):

  @property
  def alignments_size(self):
    raise NotImplementedError

  @property
  def state_size(self):
    raise NotImplementedError


class _BaseAttentionMechanism(AttentionMechanism):
  """A base AttentionMechanism class providing common functionality.

  Common functionality includes:
    此时query代表解码器的当前hidden值，memory代表编码器所有的隐藏值状态值
    1. Storing the query and memory layers.
    2. Preprocessing and storing the memory.
  """

  def __init__(self,
               query_layer,
               memory,
               probability_fn,
               memory_sequence_length=None,
               memory_layer=None,
               check_inner_dims_defined=True,
               score_mask_value=None,
               custom_key_value_fn=None,
               name=None):
    """Construct base AttentionMechanism class.

    参数：
        query_layer:深度与memory_layer匹配
        memory:编码器rnn的输出，形状为[batch_size,max_time,...]
        probability_fn:是一个概率转换函数、类似于softmax等函数
        memory_sequence_length:不理解，不过应该不需要指定
        memory_layer:深度与query_layer匹配
        其余不太重要
    Args:
      query_layer: Callable.  Instance of `tf.compat.v1.layers.Layer`.  The
        layer's depth must match the depth of `memory_layer`.  If `query_layer`
        is not provided, the shape of `query` must match that of `memory_layer`.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      probability_fn: A `callable`.  Converts the score and previous alignments
        to probabilities. Its signature should be: `probabilities =
          probability_fn(score, state)`.
      memory_sequence_length (optional): Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      memory_layer: Instance of `tf.compat.v1.layers.Layer` (may be None).  The
        layer's depth must match the depth of `query_layer`. If `memory_layer`
        is not provided, the shape of `memory` must match that of `query_layer`.
      check_inner_dims_defined: Python boolean.  If `True`, the `memory`
        argument's shape is checked to ensure all but the two outermost
        dimensions are fully defined.
      score_mask_value: (optional): The mask value for score before passing into
        `probability_fn`. The default is -inf. Only used if
        `memory_sequence_length` is not None.
      custom_key_value_fn: (optional): The custom function for
        computing keys and values.
      name: Name to use when creating ops.
    """
    # 检查query和memory这两个参数，非空的情况下一定要是对应的两个layer，query对应decoder、memory对应encoder
    if (query_layer is not None and
        not isinstance(query_layer, layers_base.Layer)):
      raise TypeError("query_layer is not a Layer: %s" %
                      type(query_layer).__name__)
    if (memory_layer is not None and
        not isinstance(memory_layer, layers_base.Layer)):
      raise TypeError("memory_layer is not a Layer: %s" %
                      type(memory_layer).__name__)

    
    self._query_layer = query_layer
    self._memory_layer = memory_layer
    self.dtype = memory_layer.dtype
    
    if not callable(probability_fn):
        # probability_fn对应的参数一定是可调用并且有返回值的。否则将报错
      raise TypeError("probability_fn must be callable, saw type: %s" %
                      type(probability_fn).__name__)

    if score_mask_value is None:
      score_mask_value = dtypes.as_dtype(
          self._memory_layer.dtype).as_numpy_dtype(-np.inf)
    self._probability_fn = lambda score, prev: (  # pylint:disable=g-long-lambda
        probability_fn(
            _maybe_mask_score(
                score,
                memory_sequence_length=memory_sequence_length,
                score_mask_value=score_mask_value), prev))
    with ops.name_scope(name, "BaseAttentionMechanismInit",
                        nest.flatten(memory)):
      self._values = _prepare_memory(
          memory,
          memory_sequence_length=memory_sequence_length,
          check_inner_dims_defined=check_inner_dims_defined)
      self._keys = (
          self.memory_layer(self._values) if self.memory_layer  # pylint: disable=not-callable
          else self._values)
      if custom_key_value_fn is not None:
        self._keys, self._values = custom_key_value_fn(self._keys, self._values)
      self._batch_size = (
          tensor_shape.dimension_value(self._keys.shape[0]) or
          array_ops.shape(self._keys)[0])
      self._alignments_size = (
          tensor_shape.dimension_value(self._keys.shape[1]) or
          array_ops.shape(self._keys)[1])

  @property
  def memory_layer(self):
    return self._memory_layer

  @property
  def query_layer(self):
    return self._query_layer

  @property
  def values(self):
    return self._values

  @property
  def keys(self):
    return self._keys

  @property
  def batch_size(self):
    return self._batch_size

  @property
  def alignments_size(self):
    return self._alignments_size

  @property
  def state_size(self):
    return self._alignments_size

  def initial_alignments(self, batch_size, dtype):
    """Creates the initial alignment values for the `AttentionWrapper` class.

    This is important for AttentionMechanisms that use the previous alignment
    to calculate the alignment at the next time step (e.g. monotonic attention).

    The default behavior is to return a tensor of all zeros.

    Args:
      batch_size: `int32` scalar, the batch_size.
      dtype: The `dtype`.

    Returns:
      A `dtype` tensor shaped `[batch_size, alignments_size]`
      (`alignments_size` is the values' `max_time`).
    """
    max_time = self._alignments_size
    return _zero_state_tensors(max_time, batch_size, dtype)

  def initial_state(self, batch_size, dtype):
    """Creates the initial state values for the `AttentionWrapper` class.

    This is important for AttentionMechanisms that use the previous alignment
    to calculate the alignment at the next time step (e.g. monotonic attention).

    The default behavior is to return the same output as initial_alignments.

    Args:
      batch_size: `int32` scalar, the batch_size.
      dtype: The `dtype`.

    Returns:
      A structure of all-zero tensors with shapes as described by `state_size`.
    """
    return self.initial_alignments(batch_size, dtype)


class _BaseAttentionMechanismV2(AttentionMechanism, layers.Layer):
  """A base AttentionMechanism class providing common functionality.

  Common functionality includes:
    1. Storing the query and memory layers.
    2. Preprocessing and storing the memory.

  Note that this layer takes memory as its init parameter, which is an
  anti-pattern of Keras API, we have to keep the memory as init parameter for
  performance and dependency reason. Under the hood, during `__init__()`, it
  will invoke `base_layer.__call__(memory, setup_memory=True)`. This will let
  keras to keep track of the memory tensor as the input of this layer. Once
  the `__init__()` is done, then user can query the attention by
  `score = att_obj([query, state])`, and use it as a normal keras layer.

  Special attention is needed when adding using this class as the base layer for
  new attention:
    1. Build() could be invoked at least twice. So please make sure weights are
       not duplicated.
    2. Layer.get_weights() might return different set of weights if the instance
       has `query_layer`. The query_layer weights is not initialized until the
       memory is configured.

  Also note that this layer does not work with Keras model when
  `model.compile(run_eagerly=True)` due to the fact that this layer is stateful.
  The support for that will be added in a future version.
  """

  def __init__(self,
               memory,
               probability_fn,
               query_layer=None,
               memory_layer=None,
               memory_sequence_length=None,
               **kwargs):
    """Construct base AttentionMechanism class.

    Args:
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      probability_fn: A `callable`. Converts the score and previous alignments
        to probabilities. Its signature should be: `probabilities =
          probability_fn(score, state)`.
      query_layer:  (optional): Instance of `tf.keras.Layer`.  The layer's depth
        must match the depth of `memory_layer`.  If `query_layer` is not
        provided, the shape of `query` must match that of `memory_layer`.
      memory_layer: (optional): Instance of `tf.keras.Layer`. The layer's depth
        must match the depth of `query_layer`. If `memory_layer` is not
        provided, the shape of `memory` must match that of `query_layer`.
      memory_sequence_length (optional): Sequence lengths for the batch entries
        in memory. If provided, the memory tensor rows are masked with zeros for
        values past the respective sequence lengths.
      **kwargs: Dictionary that contains other common arguments for layer
        creation.
    """
    if (query_layer is not None and not isinstance(query_layer, layers.Layer)):
      raise TypeError("query_layer is not a Layer: %s" %
                      type(query_layer).__name__)
    if (memory_layer is not None and
        not isinstance(memory_layer, layers.Layer)):
      raise TypeError("memory_layer is not a Layer: %s" %
                      type(memory_layer).__name__)
    self.query_layer = query_layer
    self.memory_layer = memory_layer
    if self.memory_layer is not None and "dtype" not in kwargs:
      kwargs["dtype"] = self.memory_layer.dtype
    super(_BaseAttentionMechanismV2, self).__init__(**kwargs)
    if not callable(probability_fn):
      raise TypeError("probability_fn must be callable, saw type: %s" %
                      type(probability_fn).__name__)
    self.probability_fn = probability_fn

    self.keys = None
    self.values = None
    self.batch_size = None
    self._memory_initialized = False
    self._check_inner_dims_defined = True
    self.supports_masking = True
    self.score_mask_value = dtypes.as_dtype(self.dtype).as_numpy_dtype(-np.inf)

    if memory is not None:
      # Setup the memory by self.__call__() with memory and memory_seq_length.
      # This will make the attention follow the keras convention which takes
      # all the tensor inputs via __call__().
      if memory_sequence_length is None:
        inputs = memory
      else:
        inputs = [memory, memory_sequence_length]

      self.values = super(_BaseAttentionMechanismV2, self).__call__(
          inputs, setup_memory=True)

  def build(self, input_shape):
    if not self._memory_initialized:
      # This is for setting up the memory, which contains memory and optional
      # memory_sequence_length. Build the memory_layer with memory shape.
      if self.memory_layer is not None and not self.memory_layer.built:
        if isinstance(input_shape, list):
          self.memory_layer.build(input_shape[0])
        else:
          self.memory_layer.build(input_shape)
    else:
      # The input_shape should be query.shape and state.shape. Use the query
      # to init the query layer.
      if self.query_layer is not None and not self.query_layer.built:
        self.query_layer.build(input_shape[0])

  def __call__(self, inputs, **kwargs):
    """Preprocess the inputs before calling `base_layer.__call__()`.

    Note that there are situation here, one for setup memory, and one with
    actual query and state.
    1. When the memory has not been configured, we just pass all the param to
    base_layer.__call__(), which will then invoke self.call() with proper
    inputs, which allows this class to setup memory.
    2. When the memory has already been setup, the input should contain query
    and state, and optionally processed memory. If the processed memory is
    not included in the input, we will have to append it to the inputs and
    give it to the base_layer.__call__(). The processed memory is the output
    of first invocation of self.__call__(). If we don't add it here, then from
    keras perspective, the graph is disconnected since the output from
    previous call is never used.

    Args:
      inputs: the inputs tensors.
      **kwargs: dict, other keyeword arguments for the `__call__()`
    """
    if self._memory_initialized:
      if len(inputs) not in (2, 3):
        raise ValueError("Expect the inputs to have 2 or 3 tensors, got %d" %
                         len(inputs))
      if len(inputs) == 2:
        # We append the calculated memory here so that the graph will be
        # connected.
        inputs.append(self.values)
    return super(_BaseAttentionMechanismV2, self).__call__(inputs, **kwargs)

  def call(self, inputs, mask=None, setup_memory=False, **kwargs):
    """Setup the memory or query the attention.

    There are two case here, one for setup memory, and the second is query the
    attention score. `setup_memory` is the flag to indicate which mode it is.
    The input list will be treated differently based on that flag.

    Args:
      inputs: a list of tensor that could either be `query` and `state`, or
        `memory` and `memory_sequence_length`. `query` is the tensor of dtype
        matching `memory` and shape `[batch_size, query_depth]`. `state` is the
        tensor of dtype matching `memory` and shape `[batch_size,
        alignments_size]`. (`alignments_size` is memory's `max_time`). `memory`
        is the memory to query; usually the output of an RNN encoder. The tensor
        should be shaped `[batch_size, max_time, ...]`. `memory_sequence_length`
        (optional) is the sequence lengths for the batch entries in memory. If
        provided, the memory tensor rows are masked with zeros for values past
        the respective sequence lengths.
      mask: optional bool tensor with shape `[batch, max_time]` for the mask of
        memory. If it is not None, the corresponding item of the memory should
        be filtered out during calculation.
      setup_memory: boolean, whether the input is for setting up memory, or
        query attention.
      **kwargs: Dict, other keyword arguments for the call method.

    Returns:
      Either processed memory or attention score, based on `setup_memory`.
    """
    if setup_memory:
      if isinstance(inputs, list):
        if len(inputs) not in (1, 2):
          raise ValueError("Expect inputs to have 1 or 2 tensors, got %d" %
                           len(inputs))
        memory = inputs[0]
        memory_sequence_length = inputs[1] if len(inputs) == 2 else None
        memory_mask = mask
      else:
        memory, memory_sequence_length = inputs, None
        memory_mask = mask
      self._setup_memory(memory, memory_sequence_length, memory_mask)
      # We force the self.built to false here since only memory is initialized,
      # but the real query/state has not been call() yet. The layer should be
      # build and call again.
      self.built = False
      # Return the processed memory in order to create the Keras connectivity
      # data for it.
      return self.values
    else:
      if not self._memory_initialized:
        raise ValueError("Cannot query the attention before the setup of "
                         "memory")
      if len(inputs) not in (2, 3):
        raise ValueError("Expect the inputs to have query, state, and optional "
                         "processed memory, got %d items" % len(inputs))
      # Ignore the rest of the inputs and only care about the query and state
      query, state = inputs[0], inputs[1]
      return self._calculate_attention(query, state)

  def _setup_memory(self, memory, memory_sequence_length=None,
                    memory_mask=None):
    """Pre-process the memory before actually query the memory.

    This should only be called once at the first invocation of call().

    Args:
      memory: The memory to query; usually the output of an RNN encoder. This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length (optional): Sequence lengths for the batch entries
        in memory. If provided, the memory tensor rows are masked with zeros for
        values past the respective sequence lengths.
      memory_mask: (Optional) The boolean tensor with shape `[batch_size,
        max_time]`. For any value equal to False, the corresponding value in
        memory should be ignored.
    """
    if self._memory_initialized:
      raise ValueError("The memory for the attention has already been setup.")
    if memory_sequence_length is not None and memory_mask is not None:
      raise ValueError("memory_sequence_length and memory_mask cannot be "
                       "used at same time for attention.")
    with ops.name_scope(self.name, "BaseAttentionMechanismInit",
                        nest.flatten(memory)):
      self.values = _prepare_memory(
          memory,
          memory_sequence_length=memory_sequence_length,
          memory_mask=memory_mask,
          check_inner_dims_defined=self._check_inner_dims_defined)
      # Mark the value as check since the memory and memory mask might not
      # passed from __call__(), which does not have proper keras metadata.
      # TODO(omalleyt): Remove this hack once the mask the has proper keras
      # history.
      base_layer_utils.mark_checked(self.values)
      if self.memory_layer is not None:
        self.keys = self.memory_layer(self.values)
      else:
        self.keys = self.values
      self.batch_size = (
          tensor_shape.dimension_value(self.keys.shape[0]) or
          array_ops.shape(self.keys)[0])
      self._alignments_size = (
          tensor_shape.dimension_value(self.keys.shape[1]) or
          array_ops.shape(self.keys)[1])
      if memory_mask is not None:
        unwrapped_probability_fn = self.probability_fn

        def _mask_probability_fn(score, prev):
          return unwrapped_probability_fn(
              _maybe_mask_score(
                  score,
                  memory_mask=memory_mask,
                  memory_sequence_length=memory_sequence_length,
                  score_mask_value=self.score_mask_value), prev)

        self.probability_fn = _mask_probability_fn
    self._memory_initialized = True

  def _calculate_attention(self, query, state):
    raise NotImplementedError(
        "_calculate_attention need to be implemented by subclasses.")

  def compute_mask(self, inputs, mask=None):
    # There real input of the attention is query and state, and the memory layer
    # mask shouldn't be pass down. Returning None for all output mask here.
    return None, None

  def get_config(self):
    config = {}
    # Since the probability_fn is likely to be a wrapped function, the child
    # class should preserve the original function and how its wrapped.

    if self.query_layer is not None:
      config["query_layer"] = {
          "class_name": self.query_layer.__class__.__name__,
          "config": self.query_layer.get_config(),
      }
    if self.memory_layer is not None:
      config["memory_layer"] = {
          "class_name": self.memory_layer.__class__.__name__,
          "config": self.memory_layer.get_config(),
      }
    # memory is a required init parameter and its a tensor. It cannot be
    # serialized to config, so we put a placeholder for it.
    config["memory"] = None
    base_config = super(_BaseAttentionMechanismV2, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  def _process_probability_fn(self, func_name):
    """Helper method to retrieve the probably function by string input."""
    valid_probability_fns = {
        "softmax": nn_ops.softmax,
        "hardmax": hardmax,
    }
    if func_name not in valid_probability_fns.keys():
      raise ValueError("Invalid probability function: %s, options are %s" %
                       (func_name, valid_probability_fns.keys()))
    return valid_probability_fns[func_name]

  @classmethod
  def deserialize_inner_layer_from_config(cls, config, custom_objects):
    """Helper method that reconstruct the query and memory from the config.

    In the get_config() method, the query and memory layer configs are
    serialized into dict for persistence, this method perform the reverse action
    to reconstruct the layer from the config.

    Args:
      config: dict, the configs that will be used to reconstruct the object.
      custom_objects: dict mapping class names (or function names) of custom
        (non-Keras) objects to class/functions.

    Returns:
      config: dict, the config with layer instance created, which is ready to be
        used as init parameters.
    """
    # Reconstruct the query and memory layer for parent class.
    from tensorflow.python.keras.layers import deserialize as deserialize_layer  # pylint: disable=g-import-not-at-top
    # Instead of updating the input, create a copy and use that.
    config = config.copy()
    query_layer_config = config.pop("query_layer", None)
    if query_layer_config:
      query_layer = deserialize_layer(
          query_layer_config, custom_objects=custom_objects)
      config["query_layer"] = query_layer
    memory_layer_config = config.pop("memory_layer", None)
    if memory_layer_config:
      memory_layer = deserialize_layer(
          memory_layer_config, custom_objects=custom_objects)
      config["memory_layer"] = memory_layer
    return config

  @property
  def alignments_size(self):
    return self._alignments_size

  @property
  def state_size(self):
    return self._alignments_size

  def initial_alignments(self, batch_size, dtype):
    """Creates the initial alignment values for the `AttentionWrapper` class.

    This is important for AttentionMechanisms that use the previous alignment
    to calculate the alignment at the next time step (e.g. monotonic attention).

    The default behavior is to return a tensor of all zeros.

    Args:
      batch_size: `int32` scalar, the batch_size.
      dtype: The `dtype`.

    Returns:
      A `dtype` tensor shaped `[batch_size, alignments_size]`
      (`alignments_size` is the values' `max_time`).
    """
    max_time = self._alignments_size
    return _zero_state_tensors(max_time, batch_size, dtype)

  def initial_state(self, batch_size, dtype):
    """Creates the initial state values for the `AttentionWrapper` class.

    This is important for AttentionMechanisms that use the previous alignment
    to calculate the alignment at the next time step (e.g. monotonic attention).

    The default behavior is to return the same output as initial_alignments.

    Args:
      batch_size: `int32` scalar, the batch_size.
      dtype: The `dtype`.

    Returns:
      A structure of all-zero tensors with shapes as described by `state_size`.
    """
    return self.initial_alignments(batch_size, dtype)


def _luong_score(query, keys, scale):
  """Implements Luong-style (multiplicative) scoring function.

  This attention has two forms.  The first is standard Luong attention,
  as described in:

  Minh-Thang Luong, Hieu Pham, Christopher D. Manning.
  "Effective Approaches to Attention-based Neural Machine Translation."
  EMNLP 2015.  https://arxiv.org/abs/1508.04025

  The second is the scaled form inspired partly by the normalized form of
  Bahdanau attention.

  To enable the second form, call this function with `scale=True`.

  Args:
    query: Tensor, shape `[batch_size, num_units]` to compare to keys.
    keys: Processed memory, shape `[batch_size, max_time, num_units]`.
    scale: the optional tensor to scale the attention score.

  Returns:
    A `[batch_size, max_time]` tensor of unnormalized score values.

  Raises:
    ValueError: If `key` and `query` depths do not match.
  """
  depth = query.get_shape()[-1]
  key_units = keys.get_shape()[-1]
  if depth != key_units:
    raise ValueError(
        "Incompatible or unknown inner dimensions between query and keys.  "
        "Query (%s) has units: %s.  Keys (%s) have units: %s.  "
        "Perhaps you need to set num_units to the keys' dimension (%s)?" %
        (query, depth, keys, key_units, key_units))

  # Reshape from [batch_size, depth] to [batch_size, 1, depth]
  # for matmul.
  query = array_ops.expand_dims(query, 1)

  # Inner product along the query units dimension.
  # matmul shapes: query is [batch_size, 1, depth] and
  #                keys is [batch_size, max_time, depth].
  # the inner product is asked to **transpose keys' inner shape** to get a
  # batched matmul on:
  #   [batch_size, 1, depth] . [batch_size, depth, max_time]
  # resulting in an output shape of:
  #   [batch_size, 1, max_time].
  # we then squeeze out the center singleton dimension.
  score = math_ops.matmul(query, keys, transpose_b=True)
  score = array_ops.squeeze(score, [1])

  if scale is not None:
    score = scale * score
  return score


class LuongAttention(_BaseAttentionMechanism):
  """Implements Luong-style (multiplicative) attention scoring.

  This attention has two forms.  The first is standard Luong attention,
  as described in:

  Minh-Thang Luong, Hieu Pham, Christopher D. Manning.
  [Effective Approaches to Attention-based Neural Machine Translation.
  EMNLP 2015.](https://arxiv.org/abs/1508.04025)

  The second is the scaled form inspired partly by the normalized form of
  Bahdanau attention.

  To enable the second form, construct the object with parameter
  `scale=True`.
  """

  def __init__(self,
               num_units,
               memory,
               memory_sequence_length=None,
               scale=False,
               probability_fn=None,
               score_mask_value=None,
               dtype=None,
               custom_key_value_fn=None,
               name="LuongAttention"):
    """Construct the AttentionMechanism mechanism.

    参数:
        num_units:Attention机制的深度
        memory：RNN的输出，形状为[batch_size,max_time,...]
        scale:False
        probability_fn:概率函数，默认用softmax
        其余参数默认即可
    Args:
    
      num_units: The depth of the attention mechanism.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length: (optional) Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      scale: Python boolean.  Whether to scale the energy term.
      probability_fn: (optional) A `callable`.  Converts the score to
        probabilities.  The default is `tf.nn.softmax`. Other options include
        `tf.contrib.seq2seq.hardmax` and `tf.contrib.sparsemax.sparsemax`.
        Its signature should be: `probabilities = probability_fn(score)`.
      score_mask_value: (optional) The mask value for score before passing into
        `probability_fn`. The default is -inf. Only used if
        `memory_sequence_length` is not None.
      dtype: The data type for the memory layer of the attention mechanism.
      custom_key_value_fn: (optional): The custom function for
        computing keys and values.
      name: Name to use when creating ops.
    """
    # For LuongAttention, we only transform the memory layer; thus
    # num_units **must** match expected the query depth.
    if probability_fn is None:
      probability_fn = nn_ops.softmax
    if dtype is None:
      dtype = dtypes.float32
    # 使用Lambda表达式计算score值
    wrapped_probability_fn = lambda score, _: probability_fn(score)
    # 调用super.init方法对父类进行初始化
    super(LuongAttention, self).__init__(
        query_layer=None,
        memory_layer=layers_core.Dense(
            num_units, name="memory_layer", use_bias=False, dtype=dtype),
        memory=memory,
        probability_fn=wrapped_probability_fn,
        memory_sequence_length=memory_sequence_length,
        score_mask_value=score_mask_value,
        custom_key_value_fn=custom_key_value_fn,
        name=name)
    self._num_units = num_units
    self._scale = scale
    self._name = name

  def __call__(self, query, state):
    #  call函数中，也就是当前类被实例化调用的时候，直接会调用这个函数
    # 个人理解是key和value相等，都代表着编码器的隐藏值状态，只不过key是用来与query也就是解码器中的当前隐藏值状态计算score的，
    # 或者说是计算Attention权重的，也就是经过softmax处理的概率分布
    # 那么value也是编码器的隐藏值状态，只不过这里面是基于上一步的运算结束后，在进行加权求和，得到的权重向量
    """Score the query based on the keys and values.

    Args:
      query: Tensor of dtype matching `self.values` and shape `[batch_size,
        query_depth]`.
      state: Tensor of dtype matching `self.values` and shape `[batch_size,
        alignments_size]` (`alignments_size` is memory's `max_time`).

    Returns:
      alignments: Tensor of dtype matching `self.values` and shape
        `[batch_size, alignments_size]` (`alignments_size` is memory's
        `max_time`).
    """
    with variable_scope.variable_scope(None, "luong_attention", [query]):
      attention_g = None
      if self._scale:
        attention_g = variable_scope.get_variable(
            "attention_g",
            dtype=query.dtype,
            initializer=init_ops.ones_initializer,
            shape=())
      score = _luong_score(query, self._keys, attention_g)
    alignments = self._probability_fn(score, state)
    next_state = alignments
    # 返回的是最终对于编码器的每个隐藏值h的权重向量
    return alignments, next_state


class LuongAttentionV2(_BaseAttentionMechanismV2):
  """Implements Luong-style (multiplicative) attention scoring.

  This attention has two forms.  The first is standard Luong attention,
  as described in:

  Minh-Thang Luong, Hieu Pham, Christopher D. Manning.
  [Effective Approaches to Attention-based Neural Machine Translation.
  EMNLP 2015.](https://arxiv.org/abs/1508.04025)

  The second is the scaled form inspired partly by the normalized form of
  Bahdanau attention.

  To enable the second form, construct the object with parameter
  `scale=True`.
  """

  def __init__(self,
               units,
               memory,
               memory_sequence_length=None,
               scale=False,
               probability_fn="softmax",
               dtype=None,
               name="LuongAttention",
               **kwargs):
    """Construct the AttentionMechanism mechanism.

    Args:
      units: The depth of the attention mechanism.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length: (optional): Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      scale: Python boolean. Whether to scale the energy term.
      probability_fn: (optional) string, the name of function to convert the
        attention score to probabilities. The default is `softmax` which is
        `tf.nn.softmax`. Other options is `hardmax`, which is hardmax() within
        this module. Any other value will result intovalidation error. Default
        to use `softmax`.
      dtype: The data type for the memory layer of the attention mechanism.
      name: Name to use when creating ops.
      **kwargs: Dictionary that contains other common arguments for layer
        creation.
    """
    # For LuongAttention, we only transform the memory layer; thus
    # num_units **must** match expected the query depth.
    self.probability_fn_name = probability_fn
    probability_fn = self._process_probability_fn(self.probability_fn_name)
    wrapped_probability_fn = lambda score, _: probability_fn(score)
    if dtype is None:
      dtype = dtypes.float32
    memory_layer = kwargs.pop("memory_layer", None)
    if not memory_layer:
      memory_layer = layers.Dense(
          units, name="memory_layer", use_bias=False, dtype=dtype)
    self.units = units
    self.scale = scale
    self.scale_weight = None
    super(LuongAttentionV2, self).__init__(
        memory=memory,
        memory_sequence_length=memory_sequence_length,
        query_layer=None,
        memory_layer=memory_layer,
        probability_fn=wrapped_probability_fn,
        name=name,
        dtype=dtype,
        **kwargs)

  def build(self, input_shape):
    super(LuongAttentionV2, self).build(input_shape)
    if self.scale and self.scale_weight is None:
      self.scale_weight = self.add_weight(
          "attention_g", initializer=init_ops.ones_initializer, shape=())
    self.built = True

  def _calculate_attention(self, query, state):
    """Score the query based on the keys and values.

    Args:
      query: Tensor of dtype matching `self.values` and shape `[batch_size,
        query_depth]`.
      state: Tensor of dtype matching `self.values` and shape `[batch_size,
        alignments_size]` (`alignments_size` is memory's `max_time`).

    Returns:
      alignments: Tensor of dtype matching `self.values` and shape
        `[batch_size, alignments_size]` (`alignments_size` is memory's
        `max_time`).
      next_state: Same as the alignments.
    """
    score = _luong_score(query, self.keys, self.scale_weight)
    alignments = self.probability_fn(score, state)
    next_state = alignments
    return alignments, next_state

  def get_config(self):
    config = {
        "units": self.units,
        "scale": self.scale,
        "probability_fn": self.probability_fn_name,
    }
    base_config = super(LuongAttentionV2, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config, custom_objects=None):
    config = _BaseAttentionMechanismV2.deserialize_inner_layer_from_config(
        config, custom_objects=custom_objects)
    return cls(**config)


def _bahdanau_score(processed_query,
                    keys,
                    attention_v,
                    attention_g=None,
                    attention_b=None):
  """Implements Bahdanau-style (additive) scoring function.

  This attention has two forms.  The first is Bhandanau attention,
  as described in:

  Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio.
  "Neural Machine Translation by Jointly Learning to Align and Translate."
  ICLR 2015. https://arxiv.org/abs/1409.0473

  The second is the normalized form.  This form is inspired by the
  weight normalization article:

  Tim Salimans, Diederik P. Kingma.
  "Weight Normalization: A Simple Reparameterization to Accelerate
   Training of Deep Neural Networks."
  https://arxiv.org/abs/1602.07868

  To enable the second form, set please pass in attention_g and attention_b.

  Args:
    processed_query: Tensor, shape `[batch_size, num_units]` to compare to keys.
    keys: Processed memory, shape `[batch_size, max_time, num_units]`.
    attention_v: Tensor, shape `[num_units]`.
    attention_g: Optional scalar tensor for normalization.
    attention_b: Optional tensor with shape `[num_units]` for normalization.

  Returns:
    A `[batch_size, max_time]` tensor of unnormalized score values.
  """
  # Reshape from [batch_size, ...] to [batch_size, 1, ...] for broadcasting.
  processed_query = array_ops.expand_dims(processed_query, 1)
  if attention_g is not None and attention_b is not None:
    normed_v = attention_g * attention_v * math_ops.rsqrt(
        math_ops.reduce_sum(math_ops.square(attention_v)))
    return math_ops.reduce_sum(
        normed_v * math_ops.tanh(keys + processed_query + attention_b), [2])
  else:
    return math_ops.reduce_sum(
        attention_v * math_ops.tanh(keys + processed_query), [2])


class BahdanauAttention(_BaseAttentionMechanism):
  """Implements Bahdanau-style (additive) attention.

  This attention has two forms.  The first is Bahdanau attention,
  as described in:

  Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio.
  "Neural Machine Translation by Jointly Learning to Align and Translate."
  ICLR 2015. https://arxiv.org/abs/1409.0473

  The second is the normalized form.  This form is inspired by the
  weight normalization article:

  Tim Salimans, Diederik P. Kingma.
  "Weight Normalization: A Simple Reparameterization to Accelerate
   Training of Deep Neural Networks."
  https://arxiv.org/abs/1602.07868

  To enable the second form, construct the object with parameter
  `normalize=True`.
  """

  def __init__(self,
               num_units,
               memory,
               memory_sequence_length=None,
               normalize=False,
               probability_fn=None,
               score_mask_value=None,
               dtype=None,
               custom_key_value_fn=None,
               name="BahdanauAttention"):
    """Construct the Attention mechanism.

    Args:
      num_units: The depth of the query mechanism.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length: (optional) Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      normalize: Python boolean.  Whether to normalize the energy term.
      probability_fn: (optional) A `callable`.  Converts the score to
        probabilities.  The default is `tf.nn.softmax`. Other options include
        `tf.contrib.seq2seq.hardmax` and `tf.contrib.sparsemax.sparsemax`.
        Its signature should be: `probabilities = probability_fn(score)`.
      score_mask_value: (optional): The mask value for score before passing into
        `probability_fn`. The default is -inf. Only used if
        `memory_sequence_length` is not None.
      dtype: The data type for the query and memory layers of the attention
        mechanism.
      custom_key_value_fn: (optional): The custom function for
        computing keys and values.
      name: Name to use when creating ops.
    """
    if probability_fn is None:
      probability_fn = nn_ops.softmax
    if dtype is None:
      dtype = dtypes.float32
    wrapped_probability_fn = lambda score, _: probability_fn(score)
    super(BahdanauAttention, self).__init__(
        query_layer=layers_core.Dense(
            num_units, name="query_layer", use_bias=False, dtype=dtype),
        memory_layer=layers_core.Dense(
            num_units, name="memory_layer", use_bias=False, dtype=dtype),
        memory=memory,
        probability_fn=wrapped_probability_fn,
        custom_key_value_fn=custom_key_value_fn,
        memory_sequence_length=memory_sequence_length,
        score_mask_value=score_mask_value,
        name=name)
    self._num_units = num_units
    self._normalize = normalize
    self._name = name

  def __call__(self, query, state):
    """Score the query based on the keys and values.

    Args:
      query: Tensor of dtype matching `self.values` and shape `[batch_size,
        query_depth]`.
      state: Tensor of dtype matching `self.values` and shape `[batch_size,
        alignments_size]` (`alignments_size` is memory's `max_time`).

    Returns:
      alignments: Tensor of dtype matching `self.values` and shape
        `[batch_size, alignments_size]` (`alignments_size` is memory's
        `max_time`).
    """
    with variable_scope.variable_scope(None, "bahdanau_attention", [query]):
      processed_query = self.query_layer(query) if self.query_layer else query
      attention_v = variable_scope.get_variable(
          "attention_v", [self._num_units], dtype=query.dtype)
      if not self._normalize:
        attention_g = None
        attention_b = None
      else:
        attention_g = variable_scope.get_variable(
            "attention_g",
            dtype=query.dtype,
            initializer=init_ops.constant_initializer(
                math.sqrt((1. / self._num_units))),
            shape=())
        attention_b = variable_scope.get_variable(
            "attention_b", [self._num_units],
            dtype=query.dtype,
            initializer=init_ops.zeros_initializer())

      score = _bahdanau_score(
          processed_query,
          self._keys,
          attention_v,
          attention_g=attention_g,
          attention_b=attention_b)
    alignments = self._probability_fn(score, state)
    next_state = alignments
    return alignments, next_state


class BahdanauAttentionV2(_BaseAttentionMechanismV2):
  """Implements Bahdanau-style (additive) attention.

  This attention has two forms.  The first is Bahdanau attention,
  as described in:

  Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio.
  "Neural Machine Translation by Jointly Learning to Align and Translate."
  ICLR 2015. https://arxiv.org/abs/1409.0473

  The second is the normalized form.  This form is inspired by the
  weight normalization article:

  Tim Salimans, Diederik P. Kingma.
  "Weight Normalization: A Simple Reparameterization to Accelerate
   Training of Deep Neural Networks."
  https://arxiv.org/abs/1602.07868

  To enable the second form, construct the object with parameter
  `normalize=True`.
  """

  def __init__(self,
               units,
               memory,
               memory_sequence_length=None,
               normalize=False,
               probability_fn="softmax",
               kernel_initializer="glorot_uniform",
               dtype=None,
               name="BahdanauAttention",
               **kwargs):
    """Construct the Attention mechanism.

    Args:
      units: The depth of the query mechanism.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length: (optional): Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      normalize: Python boolean.  Whether to normalize the energy term.
      probability_fn: (optional) string, the name of function to convert the
        attention score to probabilities. The default is `softmax` which is
        `tf.nn.softmax`. Other options is `hardmax`, which is hardmax() within
        this module. Any other value will result into validation error. Default
        to use `softmax`.
      kernel_initializer: (optional), the name of the initializer for the
        attention kernel.
      dtype: The data type for the query and memory layers of the attention
        mechanism.
      name: Name to use when creating ops.
      **kwargs: Dictionary that contains other common arguments for layer
        creation.
    """
    self.probability_fn_name = probability_fn
    probability_fn = self._process_probability_fn(self.probability_fn_name)
    wrapped_probability_fn = lambda score, _: probability_fn(score)
    if dtype is None:
      dtype = dtypes.float32
    query_layer = kwargs.pop("query_layer", None)
    if not query_layer:
      query_layer = layers.Dense(
          units, name="query_layer", use_bias=False, dtype=dtype)
    memory_layer = kwargs.pop("memory_layer", None)
    if not memory_layer:
      memory_layer = layers.Dense(
          units, name="memory_layer", use_bias=False, dtype=dtype)
    self.units = units
    self.normalize = normalize
    self.kernel_initializer = initializers.get(kernel_initializer)
    self.attention_v = None
    self.attention_g = None
    self.attention_b = None
    super(BahdanauAttentionV2, self).__init__(
        memory=memory,
        memory_sequence_length=memory_sequence_length,
        query_layer=query_layer,
        memory_layer=memory_layer,
        probability_fn=wrapped_probability_fn,
        name=name,
        dtype=dtype,
        **kwargs)

  def build(self, input_shape):
    super(BahdanauAttentionV2, self).build(input_shape)
    if self.attention_v is None:
      self.attention_v = self.add_weight(
          "attention_v", [self.units],
          dtype=self.dtype,
          initializer=self.kernel_initializer)
    if self.normalize and self.attention_g is None and self.attention_b is None:
      self.attention_g = self.add_weight(
          "attention_g",
          initializer=init_ops.constant_initializer(
              math.sqrt((1. / self.units))),
          shape=())
      self.attention_b = self.add_weight(
          "attention_b",
          shape=[self.units],
          initializer=init_ops.zeros_initializer())
    self.built = True

  def _calculate_attention(self, query, state):
    """Score the query based on the keys and values.

    Args:
      query: Tensor of dtype matching `self.values` and shape `[batch_size,
        query_depth]`.
      state: Tensor of dtype matching `self.values` and shape `[batch_size,
        alignments_size]` (`alignments_size` is memory's `max_time`).

    Returns:
      alignments: Tensor of dtype matching `self.values` and shape
        `[batch_size, alignments_size]` (`alignments_size` is memory's
        `max_time`).
      next_state: same as alignments.
    """
    processed_query = self.query_layer(query) if self.query_layer else query
    score = _bahdanau_score(
        processed_query,
        self.keys,
        self.attention_v,
        attention_g=self.attention_g,
        attention_b=self.attention_b)
    alignments = self.probability_fn(score, state)
    next_state = alignments
    return alignments, next_state

  def get_config(self):
    config = {
        "units": self.units,
        "normalize": self.normalize,
        "probability_fn": self.probability_fn_name,
        "kernel_initializer": initializers.serialize(self.kernel_initializer)
    }
    base_config = super(BahdanauAttentionV2, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config, custom_objects=None):
    config = _BaseAttentionMechanismV2.deserialize_inner_layer_from_config(
        config, custom_objects=custom_objects)
    return cls(**config)


def safe_cumprod(x, *args, **kwargs):
  """Computes cumprod of x in logspace using cumsum to avoid underflow.

  The cumprod function and its gradient can result in numerical instabilities
  when its argument has very small and/or zero values.  As long as the argument
  is all positive, we can instead compute the cumulative product as
  exp(cumsum(log(x))).  This function can be called identically to tf.cumprod.

  Args:
    x: Tensor to take the cumulative product of.
    *args: Passed on to cumsum; these are identical to those in cumprod.
    **kwargs: Passed on to cumsum; these are identical to those in cumprod.

  Returns:
    Cumulative product of x.
  """
  with ops.name_scope(None, "SafeCumprod", [x]):
    x = ops.convert_to_tensor(x, name="x")
    tiny = np.finfo(x.dtype.as_numpy_dtype).tiny
    return math_ops.exp(
        math_ops.cumsum(
            math_ops.log(clip_ops.clip_by_value(x, tiny, 1)), *args, **kwargs))


def monotonic_attention(p_choose_i, previous_attention, mode):
  """Compute monotonic attention distribution from choosing probabilities.

  Monotonic attention implies that the input sequence is processed in an
  explicitly left-to-right manner when generating the output sequence.  In
  addition, once an input sequence element is attended to at a given output
  timestep, elements occurring before it cannot be attended to at subsequent
  output timesteps.  This function generates attention distributions according
  to these assumptions.  For more information, see `Online and Linear-Time
  Attention by Enforcing Monotonic Alignments`.

  Args:
    p_choose_i: Probability of choosing input sequence/memory element i.  Should
      be of shape (batch_size, input_sequence_length), and should all be in the
      range [0, 1].
    previous_attention: The attention distribution from the previous output
      timestep.  Should be of shape (batch_size, input_sequence_length).  For
      the first output timestep, preevious_attention[n] should be [1, 0, 0, ...,
      0] for all n in [0, ... batch_size - 1].
    mode: How to compute the attention distribution.  Must be one of
      'recursive', 'parallel', or 'hard'. * 'recursive' uses tf.scan to
      recursively compute the distribution. This is slowest but is exact,
      general, and does not suffer from numerical instabilities. * 'parallel'
      uses parallelized cumulative-sum and cumulative-product operations to
      compute a closed-form solution to the recurrence relation defining the
      attention distribution.  This makes it more efficient than 'recursive',
      but it requires numerical checks which make the distribution non-exact.
      This can be a problem in particular when input_sequence_length is long
      and/or p_choose_i has entries very close to 0 or 1. * 'hard' requires that
      the probabilities in p_choose_i are all either 0 or 1, and subsequently
      uses a more efficient and exact solution.

  Returns:
    A tensor of shape (batch_size, input_sequence_length) representing the
    attention distributions for each sequence in the batch.

  Raises:
    ValueError: mode is not one of 'recursive', 'parallel', 'hard'.
  """
  # Force things to be tensors
  p_choose_i = ops.convert_to_tensor(p_choose_i, name="p_choose_i")
  previous_attention = ops.convert_to_tensor(
      previous_attention, name="previous_attention")
  if mode == "recursive":
    # Use .shape[0] when it's not None, or fall back on symbolic shape
    batch_size = tensor_shape.dimension_value(
        p_choose_i.shape[0]) or array_ops.shape(p_choose_i)[0]
    # Compute [1, 1 - p_choose_i[0], 1 - p_choose_i[1], ..., 1 - p_choose_i[-2]]
    shifted_1mp_choose_i = array_ops.concat(
        [array_ops.ones((batch_size, 1)), 1 - p_choose_i[:, :-1]], 1)
    # Compute attention distribution recursively as
    # q[i] = (1 - p_choose_i[i - 1])*q[i - 1] + previous_attention[i]
    # attention[i] = p_choose_i[i]*q[i]
    attention = p_choose_i * array_ops.transpose(
        functional_ops.scan(
            # Need to use reshape to remind TF of the shape between loop iterations
            lambda x, yz: array_ops.reshape(yz[0] * x + yz[1], (batch_size,)),
            # Loop variables yz[0] and yz[1]
            [
                array_ops.transpose(shifted_1mp_choose_i),
                array_ops.transpose(previous_attention)
            ],
            # Initial value of x is just zeros
            array_ops.zeros((batch_size,))))
  elif mode == "parallel":
    # safe_cumprod computes cumprod in logspace with numeric checks
    cumprod_1mp_choose_i = safe_cumprod(1 - p_choose_i, axis=1, exclusive=True)
    # Compute recurrence relation solution
    attention = p_choose_i * cumprod_1mp_choose_i * math_ops.cumsum(
        previous_attention /
        # Clip cumprod_1mp to avoid divide-by-zero
        clip_ops.clip_by_value(cumprod_1mp_choose_i, 1e-10, 1.),
        axis=1)
  elif mode == "hard":
    # Remove any probabilities before the index chosen last time step
    p_choose_i *= math_ops.cumsum(previous_attention, axis=1)
    # Now, use exclusive cumprod to remove probabilities after the first
    # chosen index, like so:
    # p_choose_i = [0, 0, 0, 1, 1, 0, 1, 1]
    # cumprod(1 - p_choose_i, exclusive=True) = [1, 1, 1, 1, 0, 0, 0, 0]
    # Product of above: [0, 0, 0, 1, 0, 0, 0, 0]
    attention = p_choose_i * math_ops.cumprod(
        1 - p_choose_i, axis=1, exclusive=True)
  else:
    raise ValueError("mode must be 'recursive', 'parallel', or 'hard'.")
  return attention


def _monotonic_probability_fn(score,
                              previous_alignments,
                              sigmoid_noise,
                              mode,
                              seed=None):
  """Attention probability function for monotonic attention.

  Takes in unnormalized attention scores, adds pre-sigmoid noise to encourage
  the model to make discrete attention decisions, passes them through a sigmoid
  to obtain "choosing" probabilities, and then calls monotonic_attention to
  obtain the attention distribution.  For more information, see

  Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, Douglas Eck,
  "Online and Linear-Time Attention by Enforcing Monotonic Alignments."
  ICML 2017.  https://arxiv.org/abs/1704.00784

  Args:
    score: Unnormalized attention scores, shape `[batch_size, alignments_size]`
    previous_alignments: Previous attention distribution, shape `[batch_size,
      alignments_size]`
    sigmoid_noise: Standard deviation of pre-sigmoid noise.  Setting this larger
      than 0 will encourage the model to produce large attention scores,
      effectively making the choosing probabilities discrete and the resulting
      attention distribution one-hot.  It should be set to 0 at test-time, and
      when hard attention is not desired.
    mode: How to compute the attention distribution.  Must be one of
      'recursive', 'parallel', or 'hard'.  See the docstring for
      `tf.contrib.seq2seq.monotonic_attention` for more information.
    seed: (optional) Random seed for pre-sigmoid noise.

  Returns:
    A `[batch_size, alignments_size]`-shape tensor corresponding to the
    resulting attention distribution.
  """
  # Optionally add pre-sigmoid noise to the scores
  if sigmoid_noise > 0:
    noise = random_ops.random_normal(
        array_ops.shape(score), dtype=score.dtype, seed=seed)
    score += sigmoid_noise * noise
  # Compute "choosing" probabilities from the attention scores
  if mode == "hard":
    # When mode is hard, use a hard sigmoid
    p_choose_i = math_ops.cast(score > 0, score.dtype)
  else:
    p_choose_i = math_ops.sigmoid(score)
  # Convert from choosing probabilities to attention distribution
  return monotonic_attention(p_choose_i, previous_alignments, mode)


class _BaseMonotonicAttentionMechanism(_BaseAttentionMechanism):
  """Base attention mechanism for monotonic attention.

  Simply overrides the initial_alignments function to provide a dirac
  distribution, which is needed in order for the monotonic attention
  distributions to have the correct behavior.
  """

  def initial_alignments(self, batch_size, dtype):
    """Creates the initial alignment values for the monotonic attentions.

    Initializes to dirac distributions, i.e. [1, 0, 0, ...memory length..., 0]
    for all entries in the batch.

    Args:
      batch_size: `int32` scalar, the batch_size.
      dtype: The `dtype`.

    Returns:
      A `dtype` tensor shaped `[batch_size, alignments_size]`
      (`alignments_size` is the values' `max_time`).
    """
    max_time = self._alignments_size
    return array_ops.one_hot(
        array_ops.zeros((batch_size,), dtype=dtypes.int32),
        max_time,
        dtype=dtype)


class _BaseMonotonicAttentionMechanismV2(_BaseAttentionMechanismV2):
  """Base attention mechanism for monotonic attention.

  Simply overrides the initial_alignments function to provide a dirac
  distribution, which is needed in order for the monotonic attention
  distributions to have the correct behavior.
  """

  def initial_alignments(self, batch_size, dtype):
    """Creates the initial alignment values for the monotonic attentions.

    Initializes to dirac distributions, i.e. [1, 0, 0, ...memory length..., 0]
    for all entries in the batch.

    Args:
      batch_size: `int32` scalar, the batch_size.
      dtype: The `dtype`.

    Returns:
      A `dtype` tensor shaped `[batch_size, alignments_size]`
      (`alignments_size` is the values' `max_time`).
    """
    max_time = self._alignments_size
    return array_ops.one_hot(
        array_ops.zeros((batch_size,), dtype=dtypes.int32),
        max_time,
        dtype=dtype)


class BahdanauMonotonicAttention(_BaseMonotonicAttentionMechanism):
  """Monotonic attention mechanism with Bahadanau-style energy function.

  This type of attention enforces a monotonic constraint on the attention
  distributions; that is once the model attends to a given point in the memory
  it can't attend to any prior points at subsequence output timesteps.  It
  achieves this by using the _monotonic_probability_fn instead of softmax to
  construct its attention distributions.  Since the attention scores are passed
  through a sigmoid, a learnable scalar bias parameter is applied after the
  score function and before the sigmoid.  Otherwise, it is equivalent to
  BahdanauAttention.  This approach is proposed in

  Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, Douglas Eck,
  "Online and Linear-Time Attention by Enforcing Monotonic Alignments."
  ICML 2017.  https://arxiv.org/abs/1704.00784
  """

  def __init__(self,
               num_units,
               memory,
               memory_sequence_length=None,
               normalize=False,
               score_mask_value=None,
               sigmoid_noise=0.,
               sigmoid_noise_seed=None,
               score_bias_init=0.,
               mode="parallel",
               dtype=None,
               name="BahdanauMonotonicAttention"):
    """Construct the Attention mechanism.

    Args:
      num_units: The depth of the query mechanism.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length (optional): Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      normalize: Python boolean.  Whether to normalize the energy term.
      score_mask_value: (optional): The mask value for score before passing into
        `probability_fn`. The default is -inf. Only used if
        `memory_sequence_length` is not None.
      sigmoid_noise: Standard deviation of pre-sigmoid noise.  See the docstring
        for `_monotonic_probability_fn` for more information.
      sigmoid_noise_seed: (optional) Random seed for pre-sigmoid noise.
      score_bias_init: Initial value for score bias scalar.  It's recommended to
        initialize this to a negative value when the length of the memory is
        large.
      mode: How to compute the attention distribution.  Must be one of
        'recursive', 'parallel', or 'hard'.  See the docstring for
        `tf.contrib.seq2seq.monotonic_attention` for more information.
      dtype: The data type for the query and memory layers of the attention
        mechanism.
      name: Name to use when creating ops.
    """
    # Set up the monotonic probability fn with supplied parameters
    if dtype is None:
      dtype = dtypes.float32
    wrapped_probability_fn = functools.partial(
        _monotonic_probability_fn,
        sigmoid_noise=sigmoid_noise,
        mode=mode,
        seed=sigmoid_noise_seed)
    super(BahdanauMonotonicAttention, self).__init__(
        query_layer=layers_core.Dense(
            num_units, name="query_layer", use_bias=False, dtype=dtype),
        memory_layer=layers_core.Dense(
            num_units, name="memory_layer", use_bias=False, dtype=dtype),
        memory=memory,
        probability_fn=wrapped_probability_fn,
        memory_sequence_length=memory_sequence_length,
        score_mask_value=score_mask_value,
        name=name)
    self._num_units = num_units
    self._normalize = normalize
    self._name = name
    self._score_bias_init = score_bias_init

  def __call__(self, query, state):
    """Score the query based on the keys and values.

    Args:
      query: Tensor of dtype matching `self.values` and shape `[batch_size,
        query_depth]`.
      state: Tensor of dtype matching `self.values` and shape `[batch_size,
        alignments_size]` (`alignments_size` is memory's `max_time`).

    Returns:
      alignments: Tensor of dtype matching `self.values` and shape
        `[batch_size, alignments_size]` (`alignments_size` is memory's
        `max_time`).
    """
    with variable_scope.variable_scope(None, "bahdanau_monotonic_attention",
                                       [query]):
      processed_query = self.query_layer(query) if self.query_layer else query
      attention_v = variable_scope.get_variable(
          "attention_v", [self._num_units], dtype=query.dtype)
      if not self._normalize:
        attention_g = None
        attention_b = None
      else:
        attention_g = variable_scope.get_variable(
            "attention_g",
            dtype=query.dtype,
            initializer=init_ops.constant_initializer(
                math.sqrt((1. / self._num_units))),
            shape=())
        attention_b = variable_scope.get_variable(
            "attention_b", [self._num_units],
            dtype=query.dtype,
            initializer=init_ops.zeros_initializer())
      score = _bahdanau_score(
          processed_query,
          self._keys,
          attention_v,
          attention_g=attention_g,
          attention_b=attention_b)
      score_bias = variable_scope.get_variable(
          "attention_score_bias",
          dtype=processed_query.dtype,
          initializer=self._score_bias_init)
      score += score_bias
    alignments = self._probability_fn(score, state)
    next_state = alignments
    return alignments, next_state


class BahdanauMonotonicAttentionV2(_BaseMonotonicAttentionMechanismV2):
  """Monotonic attention mechanism with Bahadanau-style energy function.

  This type of attention enforces a monotonic constraint on the attention
  distributions; that is once the model attends to a given point in the memory
  it can't attend to any prior points at subsequence output timesteps.  It
  achieves this by using the _monotonic_probability_fn instead of softmax to
  construct its attention distributions.  Since the attention scores are passed
  through a sigmoid, a learnable scalar bias parameter is applied after the
  score function and before the sigmoid.  Otherwise, it is equivalent to
  BahdanauAttention.  This approach is proposed in

  Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, Douglas Eck,
  "Online and Linear-Time Attention by Enforcing Monotonic Alignments."
  ICML 2017.  https://arxiv.org/abs/1704.00784
  """

  def __init__(self,
               units,
               memory,
               memory_sequence_length=None,
               normalize=False,
               sigmoid_noise=0.,
               sigmoid_noise_seed=None,
               score_bias_init=0.,
               mode="parallel",
               kernel_initializer="glorot_uniform",
               dtype=None,
               name="BahdanauMonotonicAttention",
               **kwargs):
    """Construct the Attention mechanism.

    Args:
      units: The depth of the query mechanism.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length: (optional): Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      normalize: Python boolean. Whether to normalize the energy term.
      sigmoid_noise: Standard deviation of pre-sigmoid noise. See the docstring
        for `_monotonic_probability_fn` for more information.
      sigmoid_noise_seed: (optional) Random seed for pre-sigmoid noise.
      score_bias_init: Initial value for score bias scalar. It's recommended to
        initialize this to a negative value when the length of the memory is
        large.
      mode: How to compute the attention distribution. Must be one of
        'recursive', 'parallel', or 'hard'. See the docstring for
        `tf.contrib.seq2seq.monotonic_attention` for more information.
      kernel_initializer: (optional), the name of the initializer for the
        attention kernel.
      dtype: The data type for the query and memory layers of the attention
        mechanism.
      name: Name to use when creating ops.
      **kwargs: Dictionary that contains other common arguments for layer
        creation.
    """
    # Set up the monotonic probability fn with supplied parameters
    if dtype is None:
      dtype = dtypes.float32
    wrapped_probability_fn = functools.partial(
        _monotonic_probability_fn,
        sigmoid_noise=sigmoid_noise,
        mode=mode,
        seed=sigmoid_noise_seed)
    query_layer = kwargs.pop("query_layer", None)
    if not query_layer:
      query_layer = layers.Dense(
          units, name="query_layer", use_bias=False, dtype=dtype)
    memory_layer = kwargs.pop("memory_layer", None)
    if not memory_layer:
      memory_layer = layers.Dense(
          units, name="memory_layer", use_bias=False, dtype=dtype)
    self.units = units
    self.normalize = normalize
    self.sigmoid_noise = sigmoid_noise
    self.sigmoid_noise_seed = sigmoid_noise_seed
    self.score_bias_init = score_bias_init
    self.mode = mode
    self.kernel_initializer = initializers.get(kernel_initializer)
    self.attention_v = None
    self.attention_score_bias = None
    self.attention_g = None
    self.attention_b = None
    super(BahdanauMonotonicAttentionV2, self).__init__(
        memory=memory,
        memory_sequence_length=memory_sequence_length,
        query_layer=query_layer,
        memory_layer=memory_layer,
        probability_fn=wrapped_probability_fn,
        name=name,
        dtype=dtype,
        **kwargs)

  def build(self, input_shape):
    super(BahdanauMonotonicAttentionV2, self).build(input_shape)
    if self.attention_v is None:
      self.attention_v = self.add_weight(
          "attention_v", [self.units],
          dtype=self.dtype,
          initializer=self.kernel_initializer)
    if self.attention_score_bias is None:
      self.attention_score_bias = self.add_weight(
          "attention_score_bias",
          shape=(),
          dtype=self.dtype,
          initializer=init_ops.constant_initializer(
              self.score_bias_init, dtype=self.dtype))
    if self.normalize and self.attention_g is None and self.attention_b is None:
      self.attention_g = self.add_weight(
          "attention_g",
          dtype=self.dtype,
          initializer=init_ops.constant_initializer(
              math.sqrt((1. / self.units))),
          shape=())
      self.attention_b = self.add_weight(
          "attention_b", [self.units],
          dtype=self.dtype,
          initializer=init_ops.zeros_initializer())
    self.built = True

  def _calculate_attention(self, query, state):
    """Score the query based on the keys and values.

    Args:
      query: Tensor of dtype matching `self.values` and shape `[batch_size,
        query_depth]`.
      state: Tensor of dtype matching `self.values` and shape `[batch_size,
        alignments_size]` (`alignments_size` is memory's `max_time`).

    Returns:
      alignments: Tensor of dtype matching `self.values` and shape
        `[batch_size, alignments_size]` (`alignments_size` is memory's
        `max_time`).
    """
    processed_query = self.query_layer(query) if self.query_layer else query
    score = _bahdanau_score(
        processed_query,
        self.keys,
        self.attention_v,
        attention_g=self.attention_g,
        attention_b=self.attention_b)
    score += self.attention_score_bias
    alignments = self.probability_fn(score, state)
    next_state = alignments
    return alignments, next_state

  def get_config(self):
    config = {
        "units": self.units,
        "normalize": self.normalize,
        "sigmoid_noise": self.sigmoid_noise,
        "sigmoid_noise_seed": self.sigmoid_noise_seed,
        "score_bias_init": self.score_bias_init,
        "mode": self.mode,
        "kernel_initializer": initializers.serialize(self.kernel_initializer),
    }
    base_config = super(BahdanauMonotonicAttentionV2, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config, custom_objects=None):
    config = _BaseAttentionMechanismV2.deserialize_inner_layer_from_config(
        config, custom_objects=custom_objects)
    return cls(**config)


class LuongMonotonicAttention(_BaseMonotonicAttentionMechanism):
  """Monotonic attention mechanism with Luong-style energy function.

  This type of attention enforces a monotonic constraint on the attention
  distributions; that is once the model attends to a given point in the memory
  it can't attend to any prior points at subsequence output timesteps.  It
  achieves this by using the _monotonic_probability_fn instead of softmax to
  construct its attention distributions.  Otherwise, it is equivalent to
  LuongAttention.  This approach is proposed in

  Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, Douglas Eck,
  "Online and Linear-Time Attention by Enforcing Monotonic Alignments."
  ICML 2017.  https://arxiv.org/abs/1704.00784
  """

  def __init__(self,
               num_units,
               memory,
               memory_sequence_length=None,
               scale=False,
               score_mask_value=None,
               sigmoid_noise=0.,
               sigmoid_noise_seed=None,
               score_bias_init=0.,
               mode="parallel",
               dtype=None,
               name="LuongMonotonicAttention"):
    """Construct the Attention mechanism.

    Args:
      num_units: The depth of the query mechanism.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length (optional): Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      scale: Python boolean.  Whether to scale the energy term.
      score_mask_value: (optional): The mask value for score before passing into
        `probability_fn`. The default is -inf. Only used if
        `memory_sequence_length` is not None.
      sigmoid_noise: Standard deviation of pre-sigmoid noise.  See the docstring
        for `_monotonic_probability_fn` for more information.
      sigmoid_noise_seed: (optional) Random seed for pre-sigmoid noise.
      score_bias_init: Initial value for score bias scalar.  It's recommended to
        initialize this to a negative value when the length of the memory is
        large.
      mode: How to compute the attention distribution.  Must be one of
        'recursive', 'parallel', or 'hard'.  See the docstring for
        `tf.contrib.seq2seq.monotonic_attention` for more information.
      dtype: The data type for the query and memory layers of the attention
        mechanism.
      name: Name to use when creating ops.
    """
    # Set up the monotonic probability fn with supplied parameters
    if dtype is None:
      dtype = dtypes.float32
    wrapped_probability_fn = functools.partial(
        _monotonic_probability_fn,
        sigmoid_noise=sigmoid_noise,
        mode=mode,
        seed=sigmoid_noise_seed)
    super(LuongMonotonicAttention, self).__init__(
        query_layer=None,
        memory_layer=layers_core.Dense(
            num_units, name="memory_layer", use_bias=False, dtype=dtype),
        memory=memory,
        probability_fn=wrapped_probability_fn,
        memory_sequence_length=memory_sequence_length,
        score_mask_value=score_mask_value,
        name=name)
    self._num_units = num_units
    self._scale = scale
    self._score_bias_init = score_bias_init
    self._name = name

  def __call__(self, query, state):
    """Score the query based on the keys and values.

    Args:
      query: Tensor of dtype matching `self.values` and shape `[batch_size,
        query_depth]`.
      state: Tensor of dtype matching `self.values` and shape `[batch_size,
        alignments_size]` (`alignments_size` is memory's `max_time`).

    Returns:
      alignments: Tensor of dtype matching `self.values` and shape
        `[batch_size, alignments_size]` (`alignments_size` is memory's
        `max_time`).
    """
    with variable_scope.variable_scope(None, "luong_monotonic_attention",
                                       [query]):
      attention_g = None
      if self._scale:
        attention_g = variable_scope.get_variable(
            "attention_g",
            dtype=query.dtype,
            initializer=init_ops.ones_initializer,
            shape=())
      score = _luong_score(query, self._keys, attention_g)
      score_bias = variable_scope.get_variable(
          "attention_score_bias",
          dtype=query.dtype,
          initializer=self._score_bias_init)
      score += score_bias
    alignments = self._probability_fn(score, state)
    next_state = alignments
    return alignments, next_state


class LuongMonotonicAttentionV2(_BaseMonotonicAttentionMechanismV2):
  """Monotonic attention mechanism with Luong-style energy function.

  This type of attention enforces a monotonic constraint on the attention
  distributions; that is once the model attends to a given point in the memory
  it can't attend to any prior points at subsequence output timesteps.  It
  achieves this by using the _monotonic_probability_fn instead of softmax to
  construct its attention distributions.  Otherwise, it is equivalent to
  LuongAttention.  This approach is proposed in

  [Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, Douglas Eck,
  "Online and Linear-Time Attention by Enforcing Monotonic Alignments."
  ICML 2017.](https://arxiv.org/abs/1704.00784)
  """

  def __init__(self,
               units,
               memory,
               memory_sequence_length=None,
               scale=False,
               sigmoid_noise=0.,
               sigmoid_noise_seed=None,
               score_bias_init=0.,
               mode="parallel",
               dtype=None,
               name="LuongMonotonicAttention",
               **kwargs):
    """Construct the Attention mechanism.

    Args:
      units: The depth of the query mechanism.
      memory: The memory to query; usually the output of an RNN encoder.  This
        tensor should be shaped `[batch_size, max_time, ...]`.
      memory_sequence_length: (optional): Sequence lengths for the batch entries
        in memory.  If provided, the memory tensor rows are masked with zeros
        for values past the respective sequence lengths.
      scale: Python boolean.  Whether to scale the energy term.
      sigmoid_noise: Standard deviation of pre-sigmoid noise.  See the docstring
        for `_monotonic_probability_fn` for more information.
      sigmoid_noise_seed: (optional) Random seed for pre-sigmoid noise.
      score_bias_init: Initial value for score bias scalar.  It's recommended to
        initialize this to a negative value when the length of the memory is
        large.
      mode: How to compute the attention distribution.  Must be one of
        'recursive', 'parallel', or 'hard'.  See the docstring for
        `tf.contrib.seq2seq.monotonic_attention` for more information.
      dtype: The data type for the query and memory layers of the attention
        mechanism.
      name: Name to use when creating ops.
      **kwargs: Dictionary that contains other common arguments for layer
        creation.
    """
    # Set up the monotonic probability fn with supplied parameters
    if dtype is None:
      dtype = dtypes.float32
    wrapped_probability_fn = functools.partial(
        _monotonic_probability_fn,
        sigmoid_noise=sigmoid_noise,
        mode=mode,
        seed=sigmoid_noise_seed)
    memory_layer = kwargs.pop("memory_layer", None)
    if not memory_layer:
      memory_layer = layers.Dense(
          units, name="memory_layer", use_bias=False, dtype=dtype)
    self.units = units
    self.scale = scale
    self.sigmoid_noise = sigmoid_noise
    self.sigmoid_noise_seed = sigmoid_noise_seed
    self.score_bias_init = score_bias_init
    self.mode = mode
    self.attention_g = None
    self.attention_score_bias = None
    super(LuongMonotonicAttentionV2, self).__init__(
        memory=memory,
        memory_sequence_length=memory_sequence_length,
        query_layer=None,
        memory_layer=memory_layer,
        probability_fn=wrapped_probability_fn,
        name=name,
        dtype=dtype,
        **kwargs)

  def build(self, input_shape):
    super(LuongMonotonicAttentionV2, self).build(input_shape)
    if self.scale and self.attention_g is None:
      self.attention_g = self.add_weight(
          "attention_g", initializer=init_ops.ones_initializer, shape=())
    if self.attention_score_bias is None:
      self.attention_score_bias = self.add_weight(
          "attention_score_bias",
          shape=(),
          initializer=init_ops.constant_initializer(
              self.score_bias_init, dtype=self.dtype))
    self.built = True

  def _calculate_attention(self, query, state):
    """Score the query based on the keys and values.

    Args:
      query: Tensor of dtype matching `self.values` and shape `[batch_size,
        query_depth]`.
      state: Tensor of dtype matching `self.values` and shape `[batch_size,
        alignments_size]` (`alignments_size` is memory's `max_time`).

    Returns:
      alignments: Tensor of dtype matching `self.values` and shape
        `[batch_size, alignments_size]` (`alignments_size` is memory's
        `max_time`).
      next_state: Same as alignments
    """
    score = _luong_score(query, self.keys, self.attention_g)
    score += self.attention_score_bias
    alignments = self.probability_fn(score, state)
    next_state = alignments
    return alignments, next_state

  def get_config(self):
    config = {
        "units": self.units,
        "scale": self.scale,
        "sigmoid_noise": self.sigmoid_noise,
        "sigmoid_noise_seed": self.sigmoid_noise_seed,
        "score_bias_init": self.score_bias_init,
        "mode": self.mode,
    }
    base_config = super(LuongMonotonicAttentionV2, self).get_config()
    return dict(list(base_config.items()) + list(config.items()))

  @classmethod
  def from_config(cls, config, custom_objects=None):
    config = _BaseAttentionMechanismV2.deserialize_inner_layer_from_config(
        config, custom_objects=custom_objects)
    return cls(**config)


class AttentionWrapperState(
    collections.namedtuple("AttentionWrapperState",
                           ("cell_state", "attention", "time", "alignments",
                            "alignment_history", "attention_state"))):
  """`namedtuple` storing the state of a `AttentionWrapper`.

  Contains:

    - `cell_state`: The state of the wrapped `RNNCell` at the previous time
      step.
    - `attention`: The attention emitted at the previous time step.
    - `time`: int32 scalar containing the current time step.
    - `alignments`: A single or tuple of `Tensor`(s) containing the alignments
       emitted at the previous time step for each attention mechanism.
    - `alignment_history`: (if enabled) a single or tuple of `TensorArray`(s)
       containing alignment matrices from all time steps for each attention
       mechanism. Call `stack()` on each to convert to a `Tensor`.
    - `attention_state`: A single or tuple of nested objects
       containing attention mechanism state for each attention mechanism.
       The objects may contain Tensors or TensorArrays.
  """

  def clone(self, **kwargs):
    """Clone this object, overriding components provided by kwargs.

    The new state fields' shape must match original state fields' shape. This
    will be validated, and original fields' shape will be propagated to new
    fields.

    Example:

    ```python
    initial_state = attention_wrapper.zero_state(dtype=..., batch_size=...)
    initial_state = initial_state.clone(cell_state=encoder_state)
    ```

    Args:
      **kwargs: Any properties of the state object to replace in the returned
        `AttentionWrapperState`.

    Returns:
      A new `AttentionWrapperState` whose properties are the same as
      this one, except any overridden properties as provided in `kwargs`.
    """

    def with_same_shape(old, new):
      """Check and set new tensor's shape."""
      if isinstance(old, ops.Tensor) and isinstance(new, ops.Tensor):
        if not context.executing_eagerly():
          return tensor_util.with_same_shape(old, new)
        else:
          if old.shape.as_list() != new.shape.as_list():
            raise ValueError("The shape of the AttentionWrapperState is "
                             "expected to be same as the one to clone. "
                             "self.shape: %s, input.shape: %s" %
                             (old.shape, new.shape))
          return new
      return new

    return nest.map_structure(
        with_same_shape, self,
        super(AttentionWrapperState, self)._replace(**kwargs))


def _prepare_memory(memory,
                    memory_sequence_length=None,
                    memory_mask=None,
                    check_inner_dims_defined=True):
  """Convert to tensor and possibly mask `memory`.

  Args:
    memory: `Tensor`, shaped `[batch_size, max_time, ...]`.
    memory_sequence_length: `int32` `Tensor`, shaped `[batch_size]`.
    memory_mask: `boolean` tensor with shape [batch_size, max_time]. The memory
      should be skipped when the corresponding mask is False.
    check_inner_dims_defined: Python boolean.  If `True`, the `memory`
      argument's shape is checked to ensure all but the two outermost dimensions
      are fully defined.

  Returns:
    A (possibly masked), checked, new `memory`.

  Raises:
    ValueError: If `check_inner_dims_defined` is `True` and not
      `memory.shape[2:].is_fully_defined()`.
  """
  memory = nest.map_structure(lambda m: ops.convert_to_tensor(m, name="memory"),
                              memory)
  if memory_sequence_length is not None and memory_mask is not None:
    raise ValueError("memory_sequence_length and memory_mask can't be provided "
                     "at same time.")
  if memory_sequence_length is not None:
    memory_sequence_length = ops.convert_to_tensor(
        memory_sequence_length, name="memory_sequence_length")
  if check_inner_dims_defined:

    def _check_dims(m):
      if not m.get_shape()[2:].is_fully_defined():
        raise ValueError("Expected memory %s to have fully defined inner dims, "
                         "but saw shape: %s" % (m.name, m.get_shape()))

    nest.map_structure(_check_dims, memory)
  if memory_sequence_length is None and memory_mask is None:
    return memory
  elif memory_sequence_length is not None:
    seq_len_mask = array_ops.sequence_mask(
        memory_sequence_length,
        maxlen=array_ops.shape(nest.flatten(memory)[0])[1],
        dtype=nest.flatten(memory)[0].dtype)
  else:
    # For memory_mask is not None
    seq_len_mask = math_ops.cast(
        memory_mask, dtype=nest.flatten(memory)[0].dtype)

  def _maybe_mask(m, seq_len_mask):
    """Mask the memory based on the memory mask."""
    rank = m.get_shape().ndims
    rank = rank if rank is not None else array_ops.rank(m)
    extra_ones = array_ops.ones(rank - 2, dtype=dtypes.int32)
    seq_len_mask = array_ops.reshape(
        seq_len_mask,
        array_ops.concat((array_ops.shape(seq_len_mask), extra_ones), 0))
    return m * seq_len_mask

  return nest.map_structure(lambda m: _maybe_mask(m, seq_len_mask), memory)


def _maybe_mask_score(score,
                      memory_sequence_length=None,
                      memory_mask=None,
                      score_mask_value=None):
  """Mask the attention score based on the masks."""
  if memory_sequence_length is None and memory_mask is None:
    return score
  if memory_sequence_length is not None and memory_mask is not None:
    raise ValueError("memory_sequence_length and memory_mask can't be provided "
                     "at same time.")
  if memory_sequence_length is not None:
    message = "All values in memory_sequence_length must be greater than zero."
    with ops.control_dependencies(
        [check_ops.assert_positive(memory_sequence_length, message=message)]):
      memory_mask = array_ops.sequence_mask(
          memory_sequence_length, maxlen=array_ops.shape(score)[1])
  score_mask_values = score_mask_value * array_ops.ones_like(score)
  return array_ops.where(memory_mask, score, score_mask_values)


def hardmax(logits, name=None):
  """Returns batched one-hot vectors.

  The depth index containing the `1` is that of the maximum logit value.

  Args:
    logits: A batch tensor of logit values.
    name: Name to use when creating ops.

  Returns:
    A batched one-hot tensor.
  """
  with ops.name_scope(name, "Hardmax", [logits]):
    logits = ops.convert_to_tensor(logits, name="logits")
    if tensor_shape.dimension_value(logits.get_shape()[-1]) is not None:
      depth = tensor_shape.dimension_value(logits.get_shape()[-1])
    else:
      depth = array_ops.shape(logits)[-1]
    return array_ops.one_hot(
        math_ops.argmax(logits, -1), depth, dtype=logits.dtype)


def _compute_attention(attention_mechanism, cell_output, attention_state,
                       attention_layer):
  """Computes the attention and alignments for a given attention_mechanism."""
  if isinstance(attention_mechanism, _BaseAttentionMechanismV2):
    alignments, next_attention_state = attention_mechanism(
        [cell_output, attention_state])
  else:
    # For other class, assume they are following _BaseAttentionMechanism, which
    # takes query and state as separate parameter.
    alignments, next_attention_state = attention_mechanism(
        cell_output, state=attention_state)

  # Reshape from [batch_size, memory_time] to [batch_size, 1, memory_time]
  expanded_alignments = array_ops.expand_dims(alignments, 1)
  # Context is the inner product of alignments and values along the
  # memory time dimension.
  # alignments shape is
  #   [batch_size, 1, memory_time]
  # attention_mechanism.values shape is
  #   [batch_size, memory_time, memory_size]
  # the batched matmul is over memory_time, so the output shape is
  #   [batch_size, 1, memory_size].
  # we then squeeze out the singleton dim.
  context_ = math_ops.matmul(expanded_alignments, attention_mechanism.values)
  context_ = array_ops.squeeze(context_, [1])

  if attention_layer is not None:
    attention = attention_layer(array_ops.concat([cell_output, context_], 1))
  else:
    attention = context_

  return attention, alignments, next_attention_state


class AttentionWrapper(rnn_cell_impl.RNNCell):
  """Wraps another `RNNCell` with attention."""

  def __init__(self,
               cell,
               attention_mechanism,
               attention_layer_size=None,
               alignment_history=False,
               cell_input_fn=None,
               output_attention=True,
               initial_cell_state=None,
               name=None,
               attention_layer=None,
               attention_fn=None):
    """Construct the `AttentionWrapper`.

    **NOTE** If you are using the `BeamSearchDecoder` with a cell wrapped in
    `AttentionWrapper`, then you must ensure that:

    - The encoder output has been tiled to `beam_width` via
      `tf.contrib.seq2seq.tile_batch` (NOT `tf.tile`).
    - The `batch_size` argument passed to the `zero_state` method of this
      wrapper is equal to `true_batch_size * beam_width`.
    - The initial state created with `zero_state` above contains a
      `cell_state` value containing properly tiled final state from the
      encoder.

    An example:

    ```
    tiled_encoder_outputs = tf.contrib.seq2seq.tile_batch(
        encoder_outputs, multiplier=beam_width)
    tiled_encoder_final_state = tf.conrib.seq2seq.tile_batch(
        encoder_final_state, multiplier=beam_width)
    tiled_sequence_length = tf.contrib.seq2seq.tile_batch(
        sequence_length, multiplier=beam_width)
    attention_mechanism = MyFavoriteAttentionMechanism(
        num_units=attention_depth,
        memory=tiled_inputs,
        memory_sequence_length=tiled_sequence_length)
    attention_cell = AttentionWrapper(cell, attention_mechanism, ...)
    decoder_initial_state = attention_cell.zero_state(
        dtype, batch_size=true_batch_size * beam_width)
    decoder_initial_state = decoder_initial_state.clone(
        cell_state=tiled_encoder_final_state)
    ```

    Args:
      cell: An instance of `RNNCell`.
      attention_mechanism: A list of `AttentionMechanism` instances or a single
        instance.
      attention_layer_size: A list of Python integers or a single Python
        integer, the depth of the attention (output) layer(s). If None
        (default), use the context as attention at each time step. Otherwise,
        feed the context and cell output into the attention layer to generate
        attention at each time step. If attention_mechanism is a list,
        attention_layer_size must be a list of the same length. If
        attention_layer is set, this must be None. If attention_fn is set, it
        must guaranteed that the outputs of attention_fn also meet the above
        requirements.
      alignment_history: Python boolean, whether to store alignment history from
        all time steps in the final output state (currently stored as a time
        major `TensorArray` on which you must call `stack()`).
      cell_input_fn: (optional) A `callable`.  The default is:
        `lambda inputs, attention: array_ops.concat([inputs, attention], -1)`.
      output_attention: Python bool.  If `True` (default), the output at each
        time step is the attention value.  This is the behavior of Luong-style
        attention mechanisms.  If `False`, the output at each time step is the
        output of `cell`.  This is the behavior of Bhadanau-style attention
        mechanisms.  In both cases, the `attention` tensor is propagated to the
        next time step via the state and is used there. This flag only controls
        whether the attention mechanism is propagated up to the next cell in an
        RNN stack or to the top RNN output.
      initial_cell_state: The initial state value to use for the cell when the
        user calls `zero_state()`.  Note that if this value is provided now, and
        the user uses a `batch_size` argument of `zero_state` which does not
        match the batch size of `initial_cell_state`, proper behavior is not
        guaranteed.
      name: Name to use when creating ops.
      attention_layer: A list of `tf.compat.v1.layers.Layer` instances or a
        single `tf.compat.v1.layers.Layer` instance taking the context and cell
        output as inputs to generate attention at each time step. If None
        (default), use the context as attention at each time step. If
        attention_mechanism is a list, attention_layer must be a list of the
        same length. If attention_layers_size is set, this must be None.
      attention_fn: An optional callable function that allows users to provide
        their own customized attention function, which takes input
        (attention_mechanism, cell_output, attention_state, attention_layer) and
        outputs (attention, alignments, next_attention_state). If provided, the
        attention_layer_size should be the size of the outputs of attention_fn.

    Raises:
      TypeError: `attention_layer_size` is not None and (`attention_mechanism`
        is a list but `attention_layer_size` is not; or vice versa).
      ValueError: if `attention_layer_size` is not None, `attention_mechanism`
        is a list, and its length does not match that of `attention_layer_size`;
        if `attention_layer_size` and `attention_layer` are set simultaneously.
    """
    super(AttentionWrapper, self).__init__(name=name)
    rnn_cell_impl.assert_like_rnncell("cell", cell)
    if isinstance(attention_mechanism, (list, tuple)):
      self._is_multi = True
      attention_mechanisms = attention_mechanism
      for attention_mechanism in attention_mechanisms:
        if not isinstance(attention_mechanism, AttentionMechanism):
          raise TypeError("attention_mechanism must contain only instances of "
                          "AttentionMechanism, saw type: %s" %
                          type(attention_mechanism).__name__)
    else:
      self._is_multi = False
      if not isinstance(attention_mechanism, AttentionMechanism):
        raise TypeError(
            "attention_mechanism must be an AttentionMechanism or list of "
            "multiple AttentionMechanism instances, saw type: %s" %
            type(attention_mechanism).__name__)
      attention_mechanisms = (attention_mechanism,)

    if cell_input_fn is None:
      cell_input_fn = (
          lambda inputs, attention: array_ops.concat([inputs, attention], -1))
    else:
      if not callable(cell_input_fn):
        raise TypeError("cell_input_fn must be callable, saw type: %s" %
                        type(cell_input_fn).__name__)

    if attention_layer_size is not None and attention_layer is not None:
      raise ValueError("Only one of attention_layer_size and attention_layer "
                       "should be set")

    if attention_layer_size is not None:
      attention_layer_sizes = tuple(
          attention_layer_size if isinstance(attention_layer_size, (
              list, tuple)) else (attention_layer_size,))
      if len(attention_layer_sizes) != len(attention_mechanisms):
        raise ValueError(
            "If provided, attention_layer_size must contain exactly one "
            "integer per attention_mechanism, saw: %d vs %d" %
            (len(attention_layer_sizes), len(attention_mechanisms)))
      self._attention_layers = tuple(
          layers_core.Dense(
              attention_layer_size,
              name="attention_layer",
              use_bias=False,
              dtype=attention_mechanisms[i].dtype)
          for i, attention_layer_size in enumerate(attention_layer_sizes))
      self._attention_layer_size = sum(attention_layer_sizes)
    elif attention_layer is not None:
      self._attention_layers = tuple(
          attention_layer if isinstance(attention_layer, (list, tuple)) else (
              attention_layer,))
      if len(self._attention_layers) != len(attention_mechanisms):
        raise ValueError(
            "If provided, attention_layer must contain exactly one "
            "layer per attention_mechanism, saw: %d vs %d" %
            (len(self._attention_layers), len(attention_mechanisms)))
      self._attention_layer_size = sum(
          tensor_shape.dimension_value(
              layer.compute_output_shape([
                  None, cell.output_size +
                  tensor_shape.dimension_value(mechanism.values.shape[-1])
              ])[-1]) for layer, mechanism in zip(self._attention_layers,
                                                  attention_mechanisms))
    else:
      self._attention_layers = None
      self._attention_layer_size = sum(
          tensor_shape.dimension_value(attention_mechanism.values.shape[-1])
          for attention_mechanism in attention_mechanisms)

    if attention_fn is None:
      attention_fn = _compute_attention
    self._attention_fn = attention_fn

    self._cell = cell
    self._attention_mechanisms = attention_mechanisms
    self._cell_input_fn = cell_input_fn
    self._output_attention = output_attention
    self._alignment_history = alignment_history
    with ops.name_scope(name, "AttentionWrapperInit"):
      if initial_cell_state is None:
        self._initial_cell_state = None
      else:
        final_state_tensor = nest.flatten(initial_cell_state)[-1]
        state_batch_size = (
            tensor_shape.dimension_value(final_state_tensor.shape[0]) or
            array_ops.shape(final_state_tensor)[0])
        error_message = (
            "When constructing AttentionWrapper %s: " % self._base_name +
            "Non-matching batch sizes between the memory "
            "(encoder output) and initial_cell_state.  Are you using "
            "the BeamSearchDecoder?  You may need to tile your initial state "
            "via the tf.contrib.seq2seq.tile_batch function with argument "
            "multiple=beam_width.")
        with ops.control_dependencies(
            self._batch_size_checks(state_batch_size, error_message)):
          self._initial_cell_state = nest.map_structure(
              lambda s: array_ops.identity(s, name="check_initial_cell_state"),
              initial_cell_state)

  def _batch_size_checks(self, batch_size, error_message):
    return [
        check_ops.assert_equal(
            batch_size, attention_mechanism.batch_size, message=error_message)
        for attention_mechanism in self._attention_mechanisms
    ]

  def _item_or_tuple(self, seq):
    """Returns `seq` as tuple or the singular element.

    Which is returned is determined by how the AttentionMechanism(s) were passed
    to the constructor.

    Args:
      seq: A non-empty sequence of items or generator.

    Returns:
       Either the values in the sequence as a tuple if AttentionMechanism(s)
       were passed to the constructor as a sequence or the singular element.
    """
    t = tuple(seq)
    if self._is_multi:
      return t
    else:
      return t[0]

  @property
  def output_size(self):
    if self._output_attention:
      return self._attention_layer_size
    else:
      return self._cell.output_size

  @property
  def state_size(self):
    """The `state_size` property of `AttentionWrapper`.

    Returns:
      An `AttentionWrapperState` tuple containing shapes used by this object.
    """
    return AttentionWrapperState(
        cell_state=self._cell.state_size,
        time=tensor_shape.TensorShape([]),
        attention=self._attention_layer_size,
        alignments=self._item_or_tuple(
            a.alignments_size for a in self._attention_mechanisms),
        attention_state=self._item_or_tuple(
            a.state_size for a in self._attention_mechanisms),
        alignment_history=self._item_or_tuple(
            a.alignments_size if self._alignment_history else ()
            for a in self._attention_mechanisms))  # sometimes a TensorArray

  def zero_state(self, batch_size, dtype):
    """Return an initial (zero) state tuple for this `AttentionWrapper`.

    **NOTE** Please see the initializer documentation for details of how
    to call `zero_state` if using an `AttentionWrapper` with a
    `BeamSearchDecoder`.

    Args:
      batch_size: `0D` integer tensor: the batch size.
      dtype: The internal state data type.

    Returns:
      An `AttentionWrapperState` tuple containing zeroed out tensors and,
      possibly, empty `TensorArray` objects.

    Raises:
      ValueError: (or, possibly at runtime, InvalidArgument), if
        `batch_size` does not match the output size of the encoder passed
        to the wrapper object at initialization time.
    """
    with ops.name_scope(type(self).__name__ + "ZeroState", values=[batch_size]):
      if self._initial_cell_state is not None:
        cell_state = self._initial_cell_state
      else:
        cell_state = self._cell.get_initial_state(
            batch_size=batch_size, dtype=dtype)
      error_message = (
          "When calling zero_state of AttentionWrapper %s: " % self._base_name +
          "Non-matching batch sizes between the memory "
          "(encoder output) and the requested batch size.  Are you using "
          "the BeamSearchDecoder?  If so, make sure your encoder output has "
          "been tiled to beam_width via tf.contrib.seq2seq.tile_batch, and "
          "the batch_size= argument passed to zero_state is "
          "batch_size * beam_width.")
      with ops.control_dependencies(
          self._batch_size_checks(batch_size, error_message)):
        cell_state = nest.map_structure(
            lambda s: array_ops.identity(s, name="checked_cell_state"),
            cell_state)
      initial_alignments = [
          attention_mechanism.initial_alignments(batch_size, dtype)
          for attention_mechanism in self._attention_mechanisms
      ]
      return AttentionWrapperState(
          cell_state=cell_state,
          time=array_ops.zeros([], dtype=dtypes.int32),
          attention=_zero_state_tensors(self._attention_layer_size, batch_size,
                                        dtype),
          alignments=self._item_or_tuple(initial_alignments),
          attention_state=self._item_or_tuple(
              attention_mechanism.initial_state(batch_size, dtype)
              for attention_mechanism in self._attention_mechanisms),
          alignment_history=self._item_or_tuple(
              tensor_array_ops.TensorArray(
                  dtype,
                  size=0,
                  dynamic_size=True,
                  element_shape=alignment.shape) if self._alignment_history else
              () for alignment in initial_alignments))

  def call(self, inputs, state):
    """Perform a step of attention-wrapped RNN.

    - Step 1: Mix the `inputs` and previous step's `attention` output via
      `cell_input_fn`.
    - Step 2: Call the wrapped `cell` with this input and its previous state.
    - Step 3: Score the cell's output with `attention_mechanism`.
    - Step 4: Calculate the alignments by passing the score through the
      `normalizer`.
    - Step 5: Calculate the context vector as the inner product between the
      alignments and the attention_mechanism's values (memory).
    - Step 6: Calculate the attention output by concatenating the cell output
      and context through the attention layer (a linear layer with
      `attention_layer_size` outputs).

    Args:
      inputs: (Possibly nested tuple of) Tensor, the input at this time step.
      state: An instance of `AttentionWrapperState` containing tensors from the
        previous time step.

    Returns:
      A tuple `(attention_or_cell_output, next_state)`, where:

      - `attention_or_cell_output` depending on `output_attention`.
      - `next_state` is an instance of `AttentionWrapperState`
         containing the state calculated at this time step.

    Raises:
      TypeError: If `state` is not an instance of `AttentionWrapperState`.
    """
    if not isinstance(state, AttentionWrapperState):
      raise TypeError("Expected state to be instance of AttentionWrapperState. "
                      "Received type %s instead." % type(state))

    # Step 1: Calculate the true inputs to the cell based on the
    # previous attention value.
    cell_inputs = self._cell_input_fn(inputs, state.attention)
    cell_state = state.cell_state
    cell_output, next_cell_state = self._cell(cell_inputs, cell_state)

    cell_batch_size = (
        tensor_shape.dimension_value(cell_output.shape[0]) or
        array_ops.shape(cell_output)[0])
    error_message = (
        "When applying AttentionWrapper %s: " % self.name +
        "Non-matching batch sizes between the memory "
        "(encoder output) and the query (decoder output).  Are you using "
        "the BeamSearchDecoder?  You may need to tile your memory input via "
        "the tf.contrib.seq2seq.tile_batch function with argument "
        "multiple=beam_width.")
    with ops.control_dependencies(
        self._batch_size_checks(cell_batch_size, error_message)):
      cell_output = array_ops.identity(cell_output, name="checked_cell_output")

    if self._is_multi:
      previous_attention_state = state.attention_state
      previous_alignment_history = state.alignment_history
    else:
      previous_attention_state = [state.attention_state]
      previous_alignment_history = [state.alignment_history]

    all_alignments = []
    all_attentions = []
    all_attention_states = []
    maybe_all_histories = []
    for i, attention_mechanism in enumerate(self._attention_mechanisms):
      attention, alignments, next_attention_state = self._attention_fn(
          attention_mechanism, cell_output, previous_attention_state[i],
          self._attention_layers[i] if self._attention_layers else None)
      alignment_history = previous_alignment_history[i].write(
          state.time, alignments) if self._alignment_history else ()

      all_attention_states.append(next_attention_state)
      all_alignments.append(alignments)
      all_attentions.append(attention)
      maybe_all_histories.append(alignment_history)

    attention = array_ops.concat(all_attentions, 1)
    next_state = AttentionWrapperState(
        time=state.time + 1,
        cell_state=next_cell_state,
        attention=attention,
        attention_state=self._item_or_tuple(all_attention_states),
        alignments=self._item_or_tuple(all_alignments),
        alignment_history=self._item_or_tuple(maybe_all_histories))

    if self._output_attention:
      return attention, next_state
    else:
      return cell_output, next_state