Some jax Reinforcement Learning implementation.
The body is copied from https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/ppo_atari_envpool_xla_jax.py , but this implementation trying to keep the hyperparameter and training process same as SOTA MuJoCo benchmark tianshou.
In this implementation.
-
copy runningmeanstd implementation from tianshou, and modified it into jax version
-
write env wrapper in jax way.(the api is also different from traditional step and reset, the jax version wrapper should implement reset,send,recv)
-
run 5x~10x faster than tianshou(depends on environment and your gpu, good gpu and long step environment(e.g. HalfCheetah) will lead to higher speedup)
| Mujoco(Ant-v3) SPS (global_step per second) | tianshou | jax version | Speedup |
|---|---|---|---|
| 64 cores Xeon + A100 | 1243 | 13870 | 11.16x |
| 128 cores AMD 3995WX + 2080ti | 1360 | 6787 | 4.99x |
| 32 cores AMD 3950X + 4090 | 1770 | 11568 | 6.535 |
- Install required packages
you can following https://github.com/google/jax#installation to install jax and jax[cuda](CUDA 11.4 or newer is required,If you have an Ada Lovelace (e.g., RTX 4080) or Hopper (e.g., H100) GPU, you must use CUDA 11.8 or newer.)
Basically, if you has the right cuda version, you only need to run following commands.
pip install jax
pip install --upgrade "jax[cuda]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
Then you install envpool and flax
pip install envpool==0.8.1
pip install flax
you can run apt-get install libgl1 if you havn't install libgl1 yet.
If you want track your run, you should install wandb.
Windows users can use JAX on CPU and GPU via the Windows Subsystem for Linux.
If you trying to run in vastai, you can try https://hub.docker.com/r/pyhf/cuda/ mirror.