Bring velocity to deep-learning research.
This project hosts a collection of highly modular deep learning components that are tested to be working well together. A simple yaml-based system ties these modules together declaratively using configuration files, but everything that can be defined using config files can be coded directly in the python script as well.
This is still an early version and a hobby project so documentation is unfortunately nonexistent. I've tried to make the code as clear as possible, and provide many usage examples, but whenever there was a tradeoff to be made between simplicity and modularity I've chosen modularity first and simplicity second.
Having conducted a few research projects, I've gathered a small collection of repositories lying around with various model implementations suited to a particular usecase. Usually, starting a new project involved copying pieces of code from one or multiple of these past experiments, gluing, tweaking and debugging them until the code started working in a new setting.
After repeating that pattern multiple times, I've decided that this is the time to bite the bullet and start organising deep learning models into a structure that is designed to be reused rather than copied over.
As a goal, it should be enough to write a config file that wires existing components together and defines their hyperparameters for most common applications. If that's not the case few bits of custom glue code should do the job.
This repository is still in an early stage of that journey but it will grow as I'll be putting work into it.
Project can be installed from PyPi via pip install vel but also can be checked out from github
and installed directly by running
pip install -e .from the repository root directory.
This project requires Python at least 3.6 and PyTorch 1.0.
If you want to run YAML config examples, you'll also need a project configuration file
.velproject.yaml. An example is included in this repository.
Default project configuration writes metrics to MongoDB instance open on localhost port 27017 and Visdom instance on localhost port 8097.
If you don't want to run these services, there is included
another example file .velproject.dummy.yaml
that writes training progress to the standard output only.
To use it, just rename it to .velproject.yaml.
- Models should be runnable from the configuration files that are easy to store in version control, generate automatically and diff. Codebase should be generic and do not contain any of the model hyperparameters. Unless user intervenes, it should be obvious which model was run with which hyperparameters and what output it gave.
- The amount of "magic" in the framework should be limited and it should be easy to understand what exactly the model is doing for newcomers already comfortable with PyTorch.
- All state-of-the-art models should be implemented in the framework with accuracy matching published results.
- All common deep learning workflows should be fast to implement, while uncommon ones should be possible, at least as far as PyTorch allows.
Several models are already implemented in the framework and have example config files that are ready to run and easy to modify for other similar usecases:
- State-of-the art results on Cifar10 dataset using residual networks
- Cats vs dogs classification using transfer learning from a resnet34 model pretrained on ImageNet
- Character-level language models based on LSTM and GRU recurrent networks, with example trained on works of Shakespeare
- Sentiment analysis of IMDB movie reviews
- Continuous and discrete action spaces
- Basic support for LSTM policies for A2C and PPO
- Following published policy gradient reinforcement learning algorithms:
- Advantage Actor-Critic (A2C)
- Deep Deterministic Policy Gradient (DDPG)
- Proximal Policy Optimization (PPO)
- Trust Region Policy Optimization (TRPO)
- Actor-Critic with Experience Replay (ACER)
- Deep Q-Learning (DQN) as described by DeepMind in their Nature publication with following
improvements:
- Double DQN
- Dueling DQN
- Prioritized experience replay
- N-Step Bellman updates
- Distributional Q-Learning
- Noisy Networks for Exploration
- Rainbow (combination of the above)
Most of the examples for this framework are defined using config files in the
examples-configs directory with sane default hyperparameters already selected.
For example, to run the A2C algorithm on a Breakout atari environment, simply invoke:
python -m vel.launcher examples-configs/rl/atari/a2c/breakout_a2c.yaml train
If you install the library locally, you'll have a special wrapper created that will invoke the launcher for you. Then, above becomes:
vel examples-configs/rl/atari/a2c/breakout_a2c.yaml train
General command line interface of the launcher is:
python -m vel.launcher CONFIGFILE COMMAND --device PYTORCH_DEVICE -r RUN_NUMBER -s SEED
Where PYTORCH_DEVICE is a valid name of pytorch device, most likely cuda:0.
Run number is a sequential number you wish to record your results with.
If you prefer to use the library from inside your scripts, take a look at the
examples-scripts directory. From time to time I'll be putting some examples in there as
well. Scripts generally don't require any MongoDB or Visdom setup, so they can be run straight
away in any setup, but their output will be less rich and less informative.
Here is an example script running the same setup as a config file from above:
import torch
import torch.optim as optim
from vel.rl.metrics import EpisodeRewardMetric
from vel.storage.streaming.stdout import StdoutStreaming
from vel.util.random import set_seed
from vel.rl.env.classic_atari import ClassicAtariEnv
from vel.rl.vecenv.subproc import SubprocVecEnvWrapper
from vel.modules.input.image_to_tensor import ImageToTensorFactory
from vel.rl.models.stochastic_policy_model import StochasticPolicyModelFactory
from vel.rl.models.backbone.nature_cnn import NatureCnnFactory
from vel.rl.reinforcers.on_policy_iteration_reinforcer import (
OnPolicyIterationReinforcer, OnPolicyIterationReinforcerSettings
)
from vel.rl.algo.policy_gradient.a2c import A2CPolicyGradient
from vel.rl.env_roller.step_env_roller import StepEnvRoller
from vel.api.info import TrainingInfo, EpochInfo
def breakout_a2c():
device = torch.device('cuda:0')
seed = 1001
# Set random seed in python std lib, numpy and pytorch
set_seed(seed)
# Create 16 environments evaluated in parallel in sub processess with all usual DeepMind wrappers
# These are just helper functions for that
vec_env = SubprocVecEnvWrapper(
ClassicAtariEnv('BreakoutNoFrameskip-v4'), frame_history=4
).instantiate(parallel_envs=16, seed=seed)
# Again, use a helper to create a model
# But because model is owned by the reinforcer, model should not be accessed using this variable
# but from reinforcer.model property
model = StochasticPolicyModelFactory(
input_block=ImageToTensorFactory(),
backbone=NatureCnnFactory(input_width=84, input_height=84, input_channels=4)
).instantiate(action_space=vec_env.action_space)
# Reinforcer - an object managing the learning process
reinforcer = OnPolicyIterationReinforcer(
device=device,
settings=OnPolicyIterationReinforcerSettings(
batch_size=256,
number_of_steps=5,
),
model=model,
algo=A2CPolicyGradient(
entropy_coefficient=0.01,
value_coefficient=0.5,
max_grad_norm=0.5,
discount_factor=0.99,
),
env_roller=StepEnvRoller(
environment=vec_env,
device=device,
)
)
# Model optimizer
optimizer = optim.RMSprop(reinforcer.model.parameters(), lr=7.0e-4, eps=1e-3)
# Overall information store for training information
training_info = TrainingInfo(
metrics=[
EpisodeRewardMetric('episode_rewards'), # Calculate average reward from episode
],
callbacks=[StdoutStreaming()] # Print live metrics every epoch to standard output
)
# A bit of training initialization bookkeeping...
training_info.initialize()
reinforcer.initialize_training(training_info)
training_info.on_train_begin()
# Let's make 100 batches per epoch to average metrics nicely
num_epochs = int(1.1e7 / (5 * 16) / 100)
# Normal handrolled training loop
for i in range(1, num_epochs+1):
epoch_info = EpochInfo(
training_info=training_info,
global_epoch_idx=i,
batches_per_epoch=100,
optimizer=optimizer
)
reinforcer.train_epoch(epoch_info)
training_info.on_train_end()
if __name__ == '__main__':
breakout_a2c()Dockerized version of this library is available in from the Docker Hub as
millionintegrals/vel. Link: https://hub.docker.com/r/millionintegrals/vel/
pip install vel
or
pip install vel[gym,mongo,visdom]
- https://github.com/MillionIntegrals/baselines-experiments
- https://github.com/MillionIntegrals/vel-miniworld
For a glossary of terms used in the library please refer to Glossary. If there is anything you'd like to see there, feel free to open an issue or make a pull request.
For a more or less exhaustive bibliography please refer to Bibliography.
For each major version I'll try to keep master branch stable together with what's currently published on PyPI. At the same time I'll proceed with implementing new features on a release branch that will be merged after the testing is done and a release is ready.
Below is a hypothetical set of features I somehow speculate to include in version 0.4 of Vel:
Very likely to be included:
- Neural machine translation using RNNs and Transformer Networks
- Soft actor-critic
- Twin Delayed DDPG
Possible to be included:
- Popart reward normalization
- Parameter Space Noise for Exploration
- Hindsight experience replay
- Generative adversarial networks
Code quality:
- Rename models to policies
- Force dictionary inputs and outputs for policies
- Factor action noise back into the policy
- Use linter as a part of the build process
If you use vel in your research, you can cite it as follows:
@misc{tworek2018vel,
author = {Jerry Tworek},
title = {vel},
year = {2018},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/MillionIntegrals/vel}},
}