codeaudit / jaxrl

Jax (Flax) implementation of algorithms for Deep Reinforcement Learning with continuous action spaces.

Jax (Flax) RL

This repository contains Jax (Flax) implementations of Reinforcement Learning algorithms:

• Soft Actor Critic with learnable temperature
• Advantage Weighted Actor Critic
• Image Augmentation Is All You Need(only [K=1, M=1])
• Hindsight Experience Replay
• Deep Deterministic Policy Gradient with Clipped Double Q-Learning (https://arxiv.org/abs/1802.09477)
• Behavioral Cloning

The goal of this repository is to provide simple and clean implementations to build research on top of. Please do not use this repository for baseline results and use the original implementations instead (SAC, AWAC, DrQ).

Changelog

July 20th, 2021

• Added an implementation of Deep Deterministic Policy Gradient with Clipped Double Q-Learning

July 19th, 2021

• Added an implementation of Hindsight Experience Replay

May 19th, 2021

• Added an implementation of Soft Actor Critic v1

April 29th, 2021

• Added an implementation of data augmentation from Image Augmentation Is All You Need

Installation

conda install patchelf
pip install --upgrade git+https://github.com/ikostrikov/jaxrl
# For GPU support run
pip install --upgrade jaxlib==0.1.67+cuda111 -f https://storage.googleapis.com/jax-releases/jax_releases.html

If you want to run this code on GPU, please follow instructions from the official repository.

Please follow the instructions to build mujoco-py with fast headless GPU rendering.

Development

If you want to modify the code, install following the instructions below.

conda install patchelf
pip install --upgrade -e .

Examples

Troubleshooting

If you experience out-of-memory errors, especially with enabled video saving, please consider reading docs on Jax GPU memory allocation. Also, you can try running with the following environment variable:

XLA_PYTHON_CLIENT_MEM_FRACTION=0.80 python ...

If you run your code on a remote machine and want to save videos for DeepMind Control Suite, please use EGL for rendering:

MUJOCO_GL=egl python train.py --env_name=cheetah-run --save_dir=./tmp/ --save_video

Tensorboard

Launch tensorboard to see training and evaluation logs

tensorboard --logdir=./tmp/

Results

Continous control from states

Continous control from pixels

Docker

Build

Copy your MuJoCo key to ./vendor

cd remote
docker build -t ikostrikov/jaxrl . -f Dockerfile 

Test

docker run -v <examples-dir>:/jaxrl/ --gpus=all ikostrikov/jaxrl:latest python /jaxrl/train.py --env_name=HalfCheetah-v2 --save_dir=/jaxrl/tmp/

Contributing

When contributing to this repository, please first discuss the change you wish to make via issue. If you are not familiar with pull requests, please read this documentation.