✨ Multi-dataset General ood eval

[fira7s]

• Support for multi-dataset OOD eval in classification routine
• Each datamodule should define near and far OOD datasets to be used by default
• According to args defined in datamodule (OOD and shift), routine is automatically configured Not possible
• Add SOTA OOD eval techniques from OpenOOD
• fix ugly output + add id accuracy using the ood criteria Not needed for now
• fix imagent 1k eval shift
• set ood datasets names in output table when using custom datasets
• add ood datasets to huggingface
• Add imagenet 200
• hyperparam search for some ood postprocessors
• Fix Mnist perf
• Add ood tutorial
• fix docstrings
• fill coverage

[o-laurent]

@fira7s you can install the pre-commits with pre-commit install to avoid linting errors in the CI 🙂

[fira7s]

My bad i knew it would cause the error but i ignored it @o-laurent i will be also changing more stuff in the routine and datamodules so i may need your opinion after it's done i hope my changes won't break other parts of the library.

[o-laurent]

Don't worry @fira7s , it was just a piece of friendly advice! I'll have a look, thanks!

[o-laurent]

Thanks for your work @fira7s
We should put all the configuration files on HuggingFace 😉

[fira7s]

@o-laurent Yes indeed, i'm working on merging this into dev right now, i will put all config files on huggingface once i'm finished.

[o-laurent]

(NB: We'll need to squash the PR)

[o-laurent]

@fira7s can you extract your last modifications and put them in a separate PR? I can do it and put you as "commit co-author", if you prefer.