Vladimir Blagojevic

Sounds like a plan @lhoestq If you create a PR I'll pick it up and try it out right away!

@lhoestq I can also prepare the PR, just lmk.

> I’m not exactly sure how to read the graph but it seems that to_categorical take a lot of time here. Could you share more informations on the features/stats of...

@lhoestq the proposed branch is faster, but overall training speedup is a few percentage points. I couldn't figure out how to include the GitHub branch into setup.py, so I couldn't...

Hey @albertvillanova yes thank you, I am aware, I can easily pull it from a terminal command line but then I can't automate docker image builds as dependencies are picked...

@lhoestq @thomwolf @albertvillanova The new approach is definitely faster, dataloader now takes less than 3% cumulative time (pink rectangle two rectangles to the right of tensor.py backward invocation) ![Screen Shot...

> ```python > datasets @ git+ssh://[email protected]/huggingface/datasets.git@refs/pull/2505/head > ``` @albertvillanova how would I replace datasets dependency in https://github.com/huggingface/transformers/blob/master/setup.py as the above approach is not working.

@albertvillanova of course it works. Apologies. I needed to change datasets in all deps references , like [here](https://github.com/huggingface/transformers/blob/master/setup.py#L235) for example.

Hi all, it would be great to have this functionality - an auto_play parameter, default is false. There have been numerous requests for it, and it seems trivial to implement....

Great timing @ZanSara and @anakin87 - I need this PR for https://github.com/deepset-ai/haystack/pull/3062