Question about the effect of torch_amp

[puyuan1996]

Hi,

First of all, thank you for opensourcing your nice code!

I have a question regarding the effect of torch_amp: I test the training process of EfficientZero when using and not using torch_amp in env PongNoFrameskip-v4 on k8s machine. We keep all the other setting same to compare fairly. I found that using torch.amp is a little slower than not using torch.amp. It's counterintuitive.


where the blue line is the result not using torch_amp, and the orange line is the result using torch_amp.

Could you provide some your experimental results and insights about whether to use torch_amp or not ?

Thanks a lot!

[puyuan1996]

Hi,

One more question about runtime. In the EfficientZero paper A.5 evaluation, the paper states that "To train an Atari agent for 100k steps, it only needs 4 GPUs to train 7 hours.", but in practice we found that under the same gpu and cpu configuration it takes about 14 hours to train an Atari agent for 120k steps.
What are the possible reasons? Is it because of different algorithm parameter settings or because of hardware differences?

Thanks a lot!

[rPortelas]

Hello,

I'm not 100% sure since I'm not one of EfficientMuzero authors,
but this wall clock time difference is most likely a matter of hardware differences: on my side it took me 9 hours to get 120k steps (on Breakout, with 4gpus, same hyperparams as in train.sh).