Kai Arulkumaran

I've cleaned up my Python installation as much as possible, removed and reinstalled `conda`, and have tried again, this time with the PyTorch 1.13 stable release and Apex built at...

@vince62s this issue is about model performance (reward since this is an RL setup, but analagous to classification accuracy), not timing.

Based on https://github.com/pytorch/pytorch/issues/88258, I did new runs to see if the `Adam(fused=True)` implementation in PyTorch 1.13 caused issues, or if it was specific to APEX's `FusedAdam`. Learning happens fine, so...

I have also come across this issue when trying to `mamba uninstall pytorch`, with this cryptic message about the above `__cuda` package appearing when it fails.

Back when I did [release v1.3](https://github.com/Kaixhin/Rainbow/releases/tag/1.3), as stated, I was unable to achieve any reward on Montezuma's Revenge (the only other result I couldn't match was on H.E.R.O.). However, there...

Apart from the obvious (sort by name, stars and pulls), it would also be useful to sort by build time (see https://github.com/docker/hub-feedback/issues/141).