ilpyt
The imitation learning toolbox (ilpyt
) contains modular implementations of common deep imitation learning algorithms in PyTorch, with unified infrastructure supporting key imitation learning and reinforcement learning algorithms. You can read more about ilpyt
in our white paper.
Documentation is available here.
Table of Contents
- Main Features
- Installation
- Getting Started
- Supported Algorithms and Environments
- Benchmarks
- Citation
Main Features
- Implementation of baseline imitation learning algorithms: BC, DAgger, AppL, GCL, GAIL.
- Implementation of baseline reinforcement learning algorithms, for comparison purposes: DQN, A2C, PPO2.
- Modular, extensible framework for training, evaluating, and testing imitation learning (and reinforcement learning) algorithms.
- Simple algorithm API which exposes train and test methods, allowing for quick library setup and use (a basic usage of the library requires less than ten lines of code to have a fully functioning train and test pipeline).
- A modular infrastructure for easy modification and reuse of existing components for novel algorithm implementations.
- Parallel and serialization modes, allowing for faster, optimized operations or serial operations for debugging.
- Compatibility with the OpenAI Gym environment interface for access to many existing benchmark learning environments, as well as the flexibility to create custom environments.
Installation
Note: ilpyt
has only been tested on Ubuntu 20.04, and with Python 3.8.5.
- In order to install
ilpyt
, there are a few prerequisites required. The following commands will setup all the basics so you can runilpyt
with the OpenAI Gym environments:
# Install system-based packages
apt-get install cmake python3-pip python3-testresources freeglut3-dev xvfb
# Install Wheel
pip3 install --no-cache-dir --no-warn-script-location wheel
- Install
ilpyt
using pip:
pip3 install ilpyt
# Or to install from source:
# pip3 install -e .
- (Optional) Run the associated Python tests to confirm the package has installed successfully:
git clone https://github.com/mitre/ilpyt.git
cd ilpyt/
# To run all the tests
# If running headless, prepend the pytest command with `xvfb-run -a -s "-screen 0 1400x900x24 +extension RANDR" --`
pytest tests/
# Example: to run an individual test, like DQN
pytest tests/test_dqn.py
Getting Started
Various sample Python script(s) of how to run the toolbox can be found within the examples
directory. Documentation is available here.
Basic Usage
Various sample Python script(s) of how to run the toolbox can be found within the examples
directory. A minimal train and test snippet for an imitation learning algorithm takes less than 10 lines of code in ilpyt
. In this basic example, we are training a behavioral cloning algorithm for 10,000 epochs before testing the best policy for 100 episodes.
import ilpyt
from ilpyt.agents.imitation_agent import ImitationAgent
from ilpyt.algos.bc import BC
env = ilpyt.envs.build_env(env_id='LunarLander-v2', num_env=16)
net = ilpyt.nets.choose_net(env)
agent = ImitationAgent(net=net, lr=0.0001)
algo = BC(agent=agent, env=env)
algo.train(num_epochs=10000, expert_demos='demos/LunarLander-v2/demos.pkl')
algo.test(num_episodes=100)
Code Organization
At a high-level, the algorithm
orchestrates the training and testing of our agent
in a particular environment
. During these training or testing loops, a runner
will execute the agent
and environment
in a loop to collect (state
, action
, reward
, next state
) transitions. The individual components of a transition (e.g., state
or action
) are typically torch Tensors. The agent
can then use this batch of transitions to update its network
and move towards an optimal action policy.
Customization
To implement a new algorithm, one simply has to extend the BaseAlgorithm and BaseAgent abstract classes (for even further customization, one can even make custom networks by extending the BaseNetwork interface). Each of these components is modular (see code organization for more details), allowing components to be easily swapped out. (For example, the agent.generator used in the GAIL algorithm can be easily swapped between PPOAgent, DQNAgent, or A2Cagent. In a similar way, new algorithm implementations can utilize existing implemented classes as building blocks, or extend the class interfaces for more customization.)
Adding a custom environment is as simple as extending the OpenAI Gym Environment interface and registering it within your local gym environment registry.
See agents/base_agent.py
, algos/base_algo.py
, nets/base_net.py
for more details.
Supported Algorithms and Environments
The following imitation learning (IL) algorithms are supported:
- Behavioral Cloning (BC/ALVINN)
- Dataset Aggregation (DAgger)
- Generative Adversarial Imitation Learning (GAIL)
- Apprenticeship Learning (AppL)
- Guided Cost Learning (GCL)
The following reinforcement learning (RL) algorithms are supported:
- Advantage Actor Critic (A2C)
- Deep Q-Networks (DQN)
- Proximal Policy Optimization (PPO)
The following OpenAI Gym Environments are supported. Environments with:
- Observation space:
Box(x,)
andBox(x,y,z)
- Action space:
Discrete(x)
andBox(x,)
NOTE: To create your own custom environment, just follow the OpenAI Gym Environment interface. i.e., your environment must implement the following methods (and inherit from the OpenAI Gym Class). More detailed instructions can be found on the OpenAI GitHub repository page on creating custom Gym environments.
Benchmarks
Sample train and test results of the baseline algorithms on some environments:
CartPole-v0 | MountainCar-v0 | MountainCarContinuous-v0 | LunarLander-v2 | LunarLanderContinuous-v2 | |
---|---|---|---|---|---|
Threshold | 200 | -110 | 90 | 200 | 200 |
Expert (Mean/Std) | 200.00 / 0.00 | -98.71 / 7.83 | 93.36 / 0.05 | 268.09 / 21.18 | 283.83 / 17.70 |
BC (Mean/Std) | 200.00 / 0.00 | -100.800 / 13.797 | 93.353 / 0.113 | 244.295 / 97.765 | 285.895 / 14.584 |
DAgger (Mean/Std) | 200.00 / 0.00 | -102.36 / 15.38 | 93.20 / 0.17 | 230.15 / 122.604 | 285.85 / 14.61 |
GAIL (Mean/Std) | 200.00 / 0.00 | -104.31 / 17.21 | 79.78 / 6.23 | 201.88 / 93.82 | 282.00 / 31.73 |
GCL | 200.00 / 0.00 | - | - | 212.321 / 119.933 | 255.414 / 76.917 |
AppL(Mean/Std) | 200.00 / 0.00 | -108.60 / 22.843 | - | - | - |
DQN (Mean/Std) | - | - | - | 281.96 / 24.57 | - |
A2C (Mean/Std) | - | - | 201.26 / 62.52 | - | |
PPO (Mean/Std) | - | - | - | 249.72 / 75.05 | - |
The pre-trained weights for these models can be found in our Model Zoo.
Citation
If you use ilpyt
for your work, please cite our white paper:
@misc{ilpyt_2021,
author = {Vu, Amanda and Tapley, Alex and Bissey, Brett},
title = {ilpyt: Imitation Learning Research Code Base in PyTorch},
year = {2021},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/mitre/ilpyt}},
}