Cleanba is CleanRL-style implementation of DeepMind's Sebulba distributed training platform, but with a few different design choices to make distributed RL more reproducible and transparent to use.
Warning This repo is intended for archiving purposes. Once the codebase is stable, we will move it to CleanRL for future maintenance.
- Strong performance: Cleanba's IMPALA and PPO achieve about 165% median HNS in Atari with sticky actions, matching
monobeastIMPALA's 165% median HNS and outperformingmoolibIMPALA's 140% median HNS. - Short training time: Under the 1 GPU 10 CPU setting, Cleanba's IMPALA is 6.8x faster
monobeast's IMPALA and 1.2x faster thanmoolib's IMPALA. Under a max specification setting, Cleanba's IMPALA (8 GPU and 40 CPU) and 2x faster thanmoolib's IMPALA (8 GPU and 80 CPU) is 5x faster thanmonobeast's IMPALA (1 GPU and 80 CPU).
- Highly reproducible: Cleanba shows predictable and reproducible learning curves across 1 and 8 GPU settings given the same set of hyperparameters, whereas
moolib's learning curves can be considerably different, even if hyperparameters are controlled to be the same.
Understandable: We adopt the single-file implementation philosophy used in CleanRL, making our core codebase succinct and easy to understand. For example, our cleanba/cleanba_ppo.py is ~800 lines of code.
Prerequisites:
- Python >=3.8
- Poetry 1.3.2+
- CUDA 11.2+
- CuDNN 8.2+
poetry install
poetry run pip install --upgrade "jax[cuda11_cudnn82]==0.4.8" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
poetry run python cleanba/cleanba_ppo.py
poetry run python cleanba/cleanba_ppo.py --help
Let us use a0-l1,2,3-d1 to denote our setups, where a0 means actor on GPU 0, l1,2,3 means learner on GPUs 1,2,3, and d1 means the computation is distributed 1 time.
Here are come common setups. You can also run the commands with --track to track the experiments with Weights & Biases.
# a0-l0-d1: single GPU
python cleanba/cleanba_ppo.py --actor-device-ids 0 --learner-device-ids 0 --local-num-envs 60 --track
# a0-l0,1-d1: two GPUs
python cleanba/cleanba_ppo.py --actor-device-ids 0 --learner-device-ids 0 1 --local-num-envs 60
# a0-l1,2-d1: three GPUs
python cleanba/cleanba_ppo.py --actor-device-ids 0 --learner-device-ids 1 2 --local-num-envs 60
# a0-l1,2,3-d1: four GPUs
python cleanba/cleanba_ppo.py --actor-device-ids 0 --learner-device-ids 1 2 3
# a0-l1,2,3,4-d1: five GPUs
python cleanba/cleanba_ppo.py --actor-device-ids 0 --learner-device-ids 1 2 3 4 --local-num-envs 60
# a0-l1,2,3,4,5,6-d1: seven GPUs
python cleanba/cleanba_ppo.py --actor-device-ids 0 --learner-device-ids 1 2 3 4 5 6 --local-num-envs 60
# a0-l0-d2: 8 GPUs (distributed 2 times on 4 GPUs)
# execute them in separate terminals; here we assume all 8 GPUs are on the same machine
# however it is possible to scale to hundreds of GPUs allowed by `jax.distributed`
CUDA_VISIBLE_DEVICES="0,1,2,3" SLURM_JOB_ID=26017 SLURM_STEP_NODELIST=localhost SLURM_NTASKS=2 SLURM_PROCID=0 SLURM_LOCALID=0 SLURM_STEP_NUM_NODES=2 python cleanba/cleanba_ppo.py --distributed --actor-device-ids 0 --learner-device-ids 1 2 3 --local-num-envs 60
CUDA_VISIBLE_DEVICES="4,5,6,7" SLURM_JOB_ID=26017 SLURM_STEP_NODELIST=localhost SLURM_NTASKS=2 SLURM_PROCID=1 SLURM_LOCALID=0 SLURM_STEP_NUM_NODES=2 python cleanba/cleanba_ppo.py --distributed --actor-device-ids 0 --learner-device-ids 1 2 3 --local-num-envs 60
# if you have slurm it's possible to run the following
python -m cleanrl_utils.benchmark \
--env-ids Breakout-v5 \
--command "poetry run python cleanrl/cleanba_ppo.py --distributed --learner-device-ids 1 2 3 --local-num-envs 60 --track --save-model --upload-model" \
--num-seeds 1 \
--workers 1 \
--slurm-gpus-per-task 4 \
--slurm-ntasks 2 \
--slurm-nodes 1 \
--slurm-template-path cleanba.slurm_template
Please see benchmark.sh for the commands to reproduce all of our results.
The commands to reproduce the TPU experiments can be found in tpu.sh. Here is a video demonstrating the orchestration of TPU experiments.
Screen.Recording.2023-03-23.at.9.31.58.PM.mov
Using an earlier version of the codebase, here are some runtime numbers for different hardware settings (GPUs TPUs).
runtime (minutes) in Breakout-v5 |
|
|---|---|
| baseline (8 A100) | 30.4671 |
| a0_l0_d1 (1 A100) | 154.079 |
| a0_l0_d2 (2 A100) | 93.3155 |
| a0_l1_d1 (2 A100) | 121.107 |
| a0_l01_d1 (2 A100) | 101.63 |
| a0_l1 2_d1 (3 A100) | 70.2993 |
| a0_l1 2 3_d1 (4 A100) | 52.5321 |
| a0_l0_d4 (4 A100) | 58.4344 |
| a0_l1 2 3 4_d1 (5 A100) | 44.8671 |
| a0_l1 2 3 4 5 6_d1 (7 A100) | 38.4216 |
| a0_l1 2 3 4 5 6_d1 (7 TPUv3-8 cores) | 124.397 |
| a0_l1 2_d1 (6 TPUv4 cores ) | 44.4206 |
| a0_l1_d1 (4 TPUv4 cores) | 54.6161 |
| a0_l1_d2 (8 TPUv4 cores) | 33.1134 |
The complete learning curves can be found in the static/cleanba folder. static/cleanba/plot.sh contains the plotting script.
We thank
- Stability AI's HPC for generously providing much GPU computational resources to this project.
- Hugging Face's cluster for providing much GPU computational resources to this project.
- Google's TPU Research Cloud for providing the TPU computational resources.
@misc{huang2023cleanba,
title={Cleanba: A Reproducible and Efficient Distributed Reinforcement Learning Platform},
author={Shengyi Huang and Jiayi Weng and Rujikorn Charakorn and Min Lin and Zhongwen Xu and Santiago Ontañón},
year={2023},
eprint={2310.00036},
archivePrefix={arXiv},
primaryClass={cs.LG}
}