Reproducing Pong Training Curve using Monobeast about torchbeast HOT 2 CLOSED

Hey Badr,

Thanks for your interest in TorchBeast.

I have to acknowledge we never validated monobeast against the Atari suite. It cannot (easily) be run with the right batch size so it's tricky to use the known-good hyperparameters. I believe its design is sound in principle (CPU actors, one GPU learner, PyTorch shared memory tensors for inter-process communication) but there's lots of things that can go wrong with RL.

Re: The NaNs: I don't know what mean_expected_return is, I'm guessing you're seeing mean_episode_return? If so, we made the (perhaps questionable) decision to log NaN if there is no episode ending in the batch in the first place. This happens "naturally" via division by zero (the "mean" over no episodes). This implementation logs per-rollout numbers on the learner side and not all rollouts contain finished episodes. Since we also want to log speed, frame counts etc we felt we had to log something for mean_episode_return so we decided for the somewhat "natural" NaN there. Many plotting libraries ignore these values, which is what you'd want to do at that point. This approach has the obvious downside of scaring researchers, who are typically trained to associate NaNs with "something went very wrong".

from torchbeast.

I see, and yes I apologize for the typo, I indeed meant mean_episode_return.

Thank you for your prompt response!

from torchbeast.