Replicating multi-GPU EfficientZero Atari results

[selfsim]

Hello,

I want to replicate all E0 Atari results to compare my custom learner against.

I have SSH access to 32 CPU core + 4 GPU nodes on a cluster. I managed to get the single GPU atari_efficientzero_config.py to work, but I have been unable to run the multi-GPU atari_efficientzero_multigpu_ddp_config.py successfully. Can you confirm that the latter works on the current version? And if so, can you provide the environment details (python version etc.) and the execution instructions used?

Also, do you have the hyperparameters for the E0 Atari trials? I'm assuming they are all default but it would be nice to have confirmation.

[puyuan1996]

Hello, we have confirmed that the file atari_efficientzero_multigpu_ddp_config.py on the main branch can be run normally using the following command:

python -m torch.distributed.launch --nproc_per_node=2 zoo/atari/config/atari_efficientzero_multigpu_ddp_config.py

The environment configuration is as follows:

• CentOS Linux 7
• torch==2.1.1+cu118
• python==3.9.12
• gym==0.25.1
• ale-py==0.8.0

Could you please specify the error message you are encountering? Regarding the hyperparameters for EfficientZero, you can use the default settings. If necessary, you can also adjust parameters such as model_update_ratio to optimize performance. More information on multi-gpu setting can be found here. I hope this information is helpful to you!