Find out where models are typically stored

[plbenveniste]

Eg: check on various nnUnet-based projects, are they hosted on GH? if so on the release artifacts (what is max size), if not where do ppl usually store these models?

Related to #38

[plbenveniste]

It seems that the best strategy is to host the model in a google drive or dropbox and add the link to the Github page.
Another solution could be to store the model in the release: but it's meant to make available versions of the code and the model weights don't rely fit in that category.

Do we have an ivadomed or neuropoly google drive section specifically designed for hosting pre-trained models ?

[jcohenadad]

It seems that the best strategy is to host the model in a google drive or dropbox and add the link to the Github page.

That could be an option indeed.

Another solution could be to store the model in the release: but it's meant to make available versions of the code and the model weights don't rely fit in that category.

I think it is totally appropriate to release models as 'release artifacts', as we did here. However, GH artifacts are size limited (I think it is 100MB but we should double check), so if models are larger, then we should host them somewhere else (eg: gdrive as you mentioned).

[plbenveniste]

I think that for a release, the maximum size is 2GB. (I was able to test uploading a 1GB file to a new release).
Should we prefer a realease rather than a google drive then ?

[jcohenadad]

Should we prefer a realease rather than a google drive then ?

Yes. Better to have it centralized in the same repos (for permission issues) and Gdrive is not accessible in some countries (eg China)