Adding mixin class for ease saving, uploading, downloading (as discussed in issue #9).

[thevasudevgupta]

This PR will add feature as discussed in issue #9.

Example

from huggingface_hub import ModelHubMixin

class MyModel(nn.Module, ModelHubMixin):
   def __init__(self, **kwargs):
      super().__init__()
      self.config = kwargs.pop("config", None)
      self.layer = ...
   def forward(self, ...):
      return ...

model = MyModel()

# saving model to local directory & pushing to hub
model.save_pretrained("mymodel", push_to_hub=True, config={"act": "gelu"})

# initiatizing model & loading it from trained-weights
model = MyModel.from_pretrained("username/mymodel@main")

Note

• In case user want to save some specific layers as per his needs, he can simple overwrite ._save_pretrained() method.
• Supporting all kind of PyTorch model.
• tf2 not supported currently

[thevasudevgupta]

Shall I integrate model uploading directly in .save_pretrained(path, model_id="username/mymodel", upload_to_hub=True) method instead of making separate method .upload_to_hub(model_id="username/mymodel")? I feel this looks better.

Please put your views on this.

[julien-c]

Hi @vasudevgupta7 thanks for the PR!

I have no more bandwidth this week but plan on actively looking at this early next week!

[julien-c]

Ooops @vasudevgupta7 I wanted to reach out to let you know that we had merged the Repository PR. But I guess you saw that :)

Will take a look at your updated PR with @philschmid and others, soon. Thanks!

[julien-c]

Shall I integrate model uploading directly in .save_pretrained(path, model_id="username/mymodel", upload_to_hub=True) method instead of making separate method .upload_to_hub(model_id="username/mymodel")? I feel this looks better.

Replying late to this but yeah I feel like it'd be nice to expose both.

[sgugger]

Thanks for the PR! I'm not entirely sure adding a complex logic in the default save_pretrained and from_pretrained is desirable as users that need such a logic will probable need to overload those methods anyway (for instance transformers will definitely need to, or the API assumes the model is initialized with a config when it exists but some library might use a class method instead). So I would limit those two methods to the base save of the model weights and document they can be overloaded to add more logic.

The push_to_hub on the other hand is absolutely awesome to have!

[LysandreJik]

This looks like a great addition!

[julien-c]

@vasudevgupta7 let us know when you feel like this is ready to squash/merge and I'll make a final review. Thanks again

[theo-m]

Cool stuff! 🚀
First time reviewing on this repo, mostly shallow review; the rest looks sound to me, but I'm not confident approving on my own.

[patrickvonplaten]

Great PR!

[thevasudevgupta]

Hoping everything works now. Sorry for the delay.

[julien-c]

Looks good!

Does anyone tagged here have any last comments? @LysandreJik and @philschmid maybe?

Otherwise I'll merge and push a release soon.

[LysandreJik]

LGTM!

[philschmid]

LGTM too! Thank you for creating a test. Looking forward to using this!🚀

[julien-c]

I've removed torch as a hard dependency, and please note that I don't think the tests can actually pass given that they are pushing to your userspace @vasudevgupta7 (not in a staging env) and the CI env of course doesn't have your git credentials.

I'll merge like this but would be nice to set up CI for this at some point

@LysandreJik By the way do you know how to activate GitHub Actions for pull requests from forks?

[julien-c]

Thanks for the contribution @vasudevgupta7 ❤️

[LysandreJik]

Yes, setting it up can be done like it was done here: https://github.com/huggingface/transformers/pull/10383/files. I'm opening a PR with that proposal now!

[thevasudevgupta]

Glad it's merged finally 🤩. Thank you @julien-c, @philschmid, @LysandreJik, @patrickvonplaten, @theo-m, @sgugger for reviewing and approving.

[thevasudevgupta]

@julien-c there is small error in the example you shared in release. It should be push_to_hub instead of upload_to_hub. Sorry, I forgot to update PR description. I changed it now.

[julien-c]

Oops, no big deal, but I will update the release notes: https://github.com/huggingface/huggingface_hub/releases/tag/v0.0.7

[julien-c]

link to 🐦