Add simple class to ease saving loading from HF hub

[thevasudevgupta]

Would like to have a class from which model class can be inherited to add support of Huggingface hub. It will look something like this :

from huggingface_hub import SavingUtils

class MyModel(nn.Module, SavingUtils):

   def __init__(self):
      self.layer1 = ....

   def forward(self, ...):
      return ...

model = MyModel()
model.save_pretrained("model_id")

# loading model
model = MyModel.from_pretrained("model_id")

I can make a PR for adding this simple feature if you approve it.

[julien-c]

Sounds like a great feature!

Not sure about calling .save_pretrained() to git push though – might be confusing vs. saving to a local directory.

Maybe .upload_to_hub() or .git_push(). or .upload_pretrained() if we wanted to maintain the symmetry with other method names?

Finally, not sure about implementing as a mixin vs. subclassing nn.Module. I think I like the mixin but I'd like to get more feedback, maybe from @sgugger @thomwolf @patrickvonplaten and others.

[patrickvonplaten]

I'm fine with adding a .upload_to_hub() method in PreTrainedModel(....) next to save_pretrained(...) and from_pretrained(...) - I don't think we really need a new Mixin class for that

[julien-c]

@patrickvonplaten this is huggingface_hub not transformers so we do not have a PreTrainedModel here. (It is meant to be used from any library)

[patrickvonplaten]

My bad! Then I think it's a good idea to have a ModelHubMixin class

[thevasudevgupta]

Okay. I will start PR then.

[sgugger]

I agree with @julien-c and @patrickvonplaten on the fact it should be a mixin. Seems like a useful feature, so thanks for tackling this @vasudevgupta7 !