Adam Gleave

> I did a general code review, of the library changes, left a few comments. Thanks! > A few more notes: > > * Right after running the tutorial, this...

Thanks for flagging this! Didn't know about HuggingFace datasets, I agree we should switch to that. Support automation in the form of "there's a command we can run to do...

Hi @ernestum thanks for looking into this and laying out such a clear plan. I'm on board with step 1 and step 2(a). As you say worth thinking a bit...

> Talking to @osanseviero and @simoninithomas and some playing around with the HuggingFace `datasets` library revealed, that it is super easy to map between our current format and their datasets....

> Good point. I don't know what spaces we support right now. Do we want to clarify this? Do we want to support all of them? I think when we...

> Do we properly on-hot encode discrete spaces? We could easily flatten/unflatten spaces using these util functions: https://github.com/Farama-Foundation/Gymnasium/blob/76559ed553cffc3e5082227b0a8acea461273e84/gymnasium/spaces/utils.py#L114 I think the `stable_baselines3.common.preprocessing.preprocess_obs` function should take care of this.

At first glance this seems like a real issue -- I think we first wrote this code for SB2 before the callback API was finalized, and then it just got...

That change sounds good to me. We may want to consider deprecating our own `callback` feature, although it may still serve some purpose, so should first check if it's being...

A lot of the test failures on Mac seem to be down to `ffmpeg` not being installed. I think we removed that in https://github.com/HumanCompatibleAI/imitation/pull/539 as it was making Mac OS...

Thanks for the interest in contributing! VIPER would be a nice algorithm to have working with `imitation`. My suggestion would be to initially develop it in a separate repo but...