JAXDL: JAX (Flax) Deep Learning Library

Clean state-of-the-art JAX/Flax deep learning algorithm implementations:

• Soft-Actor-Critic (SAC) (arXiv:1812.05905)
• Twin-Delayed DDPG (TD3) (arXiv:1802.09477)
• Transformer (arXiv:1706.03762; planned)
• Unified Graph Network Blocks (arXiv:1806.01261; planned)

If you use JAXDL in your work, please cite this repository as follows:

@misc{jaxdl,
  author = {Hart, Patrick},
  month = {10},
  doi = {10.5281/zenodo.5596512},
  title = {{JAXDL: JAX Deep Learning Algorithm Implementations.}},
  url = {https://github.com/patrickhart/jaxdl},
  year = {2021}
}

Results / Benchmark

Continous Control From States

HalfCheetah-v2	Ant-v2
Reacher-v2	Humanoid-v2

Installation

Install JAXDL using PyPi pip install jaxdl.

To use MuJoCo 2.1 you need to run pip install git+https://github.com/nimrod-gileadi/mujoco-py and place the binaries of MuJoCo in ~/.mujoco/mujoco210.

Examples / Getting Started

To get started have a look in the examples folder.

To train a reinforcement learning agent run

python examples/run_rl.py \
  --mode=train \
  --env_name=Ant-v2 \
  --save_dir=./tmp/ \
  --config=./examples/configs/sac_config.py

To visualize the trained agent use

python examples/run_rl.py \
  --mode=visualize \
  --env_name=Ant-v2 \
  --save_dir=./tmp/ \
  --config=./examples/configs/sac_config.py

Tensorboard

Monitor the training run using:

tensorboard --logdir=/save_dir/

Contributing

Contributions are welcome! This repository is meant to provide clean and simple implementations – please consider this when contributing.