rickstaa/stable-learning-control

Algorithm reproducibility

Closed this issue · 1 comments

Describe the bug

The algorithm results differ on different systems even with the same random seeds (i.e. non-deterministic). I know that this happens between different software versions and hardware but wanted to quickly check if we can improve this.

Reproduction steps

No response

Expected behaviour

No response

Screenshots / Live demo link

No response

System information

No response

Additional context

No response

As explained in https://pytorch.org/docs/stable/notes/randomness.html and https://discuss.pytorch.org/t/reproducibility-over-different-machines/63047/13, achieving full reproducibility between machines is challenging. I did my best to make the results similar on different devices. In the future https://www.pytorchlightning.ai/index.html could be used to improve this behaviour or docker containers.