Integrate with huggingface

[not-lain]

this pr will mainly add 3 methods to the uniformerv2 models

• save_pretrained
• push_to_hub
• from_pretrained : initialize and load the model weights

allowing your models to be easily integrated with huggingface using the PyTorchModelHubMixin class, I made this notebook https://colab.research.google.com/drive/1_H5LR5EclcBAhtyrarB_c49-TEceKoZU?usp=sharing explaining how to use it by the end of the notebook all users can load your model simply by

!pip install -q git+https://github.com/not-lain/UniFormerV2.git@integrate-with-huggingface # or your main branch when this is merged
from slowfast.models.uniformerv2_model import VisionTransformer
new_model = VisionTransformer.from_pretrained("not-lain/uniformerv2_b16") 
# no more renitializing the model and manually downloading the weights

I also made a huggingface space further showcasing how to use this pr.

Why you should integrate your model with huggingface ?

• easily save, load and push your model
• Keep track on how many times your model has been downloaded by the community.
• Automatic model card generation: the metadata in the card allows you to filter your searches easily to check how many UniFormerV2 models exist on the Hub .
• Recommended: after this pr is merged we can open a pull request on this file to make the "UniFormerV2" library official on the Hub meaning better discoverability + possibility to add code snippets.

do not hesitate if you have any reviews on the pr or any questions.

Kind regards, Hafedh Hichri