Resnet

[FrancescoSaverioZuppichini]

What does this PR do?

This WIP PR adds ResNet.

Currently, the model can be used as follows

import requests
from io import BytesIO
res = requests.get('https://github.com/huggingface/transformers/blob/master/tests/fixtures/tests_samples/COCO/000000039769.png?raw=true')
image = Image.open(BytesIO(res.content))

feature_extractor = AutoFeatureExtractor.from_pretrained("Francesco/resnet50")
model = ResNetForImageClassification.from_pretrained("Francesco/resnet50").eval()

inputs = feature_extractor(image, return_tensors="pt")
outputs = model(**inputs)
print(model.config.id2label[torch.argmax(outputs.logits).item()])
# tiger cat

[HuggingFaceDocBuilder]

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

[sgugger]

Thanks for working on this model! It's very clean already, I just have one naming suggestion.
Make sure that the conversion script can work for someone who is not yourself and I think we could expand it to work for more checkpoints?

[FrancescoSaverioZuppichini]

I've removed the regression loss from ForImageClassification since it is not a loss used for image classification. This open for discussion

[sgugger]

The regression part should be added back. There are problems that can be framed as regression from images (guessing some coordinates inside the image, the age of a subject in a picture, the angle at which a picture was taken etc.).

[FrancescoSaverioZuppichini]

I've removed the D variant code, it will be added by a different PR. Three new models outputs have been added: BaseModelOutputWithNoAttention, BaseModelOutputWithNoAttentionAndWithPooling and ImageClassificationModelOutput

[NielsRogge]

This is a horrible name.. but ok i guess, since BaseModelOutputWithPooling is already taken? cc @sgugger @LysandreJik

[NielsRogge]

LGTM