Add support for image classification

[regisss]

What does this PR do?

Fixes #173. This PR enables:

• to use an image-classification pipeline
• to perform quantization/optimization with computer vision models

At the moment, static post-training quantization with percentile/entropy calibration may not work with some models due to an issue with ONNX Runtime.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[philschmid]

LGTM!🔥✅ I left a few comments. Mostly regarding breaking backward compatibility.
Other than this awesome PR! 🔥

[michaelbenayoun]

Great, thanks a lot for adding this @regisss !
Left a few comments, but those are minor stuff, LGTM!

[JingyaHuang]

LGTM(G means great here). :>

Just added some small nits.

[davanstrien]

Thanks so much for doing this -- really happy to see this feature added 🤗

[regisss]

Thanks so much for doing this -- really happy to see this feature added hugs

Do not hesitate to show us how you use it :) @davanstrien

[rhkenne]

Thank you very much for the great work. The description of this PR mentioned "perform quantization/optimization with computer vision models". Does this include CLIP model (CLIPProcessor, CLIPModel) ? I am looking for an approach to serialize CLIP models (to onnx) and perform quantization/optimization. Any idea how ? @NielsRogge @JingyaHuang

[regisss]

Thank you very much for the great work. The description of this PR mentioned "perform quantization/optimization with computer vision models". Does this include CLIP model (CLIPProcessor, CLIPModel) ? I am looking for an approach to serialize CLIP models (to onnx) and perform quantization/optimization. Any idea how ? @NielsRogge @JingyaHuang

@rhkenne The first step would be to add ONNX support for CLIP in Transformers. Would you be interested in opening a PR to enable this?

Then, for optimization and quantization, additional work would be necessary because multimodal models are not supported yet.

[regisss]

@rhkenne There is an ongoing PR here to make CLIP exportable to ONNX.