Add image-feature-extraction task to transformers and timm lib->task mapping

[rwightman]

There's no pipeline snippet for models w/ the image-feature-extraction task, probably because this mapping is missing?

cc @pcuenca re slack comment

[rwightman]

See e.g.
https://huggingface.co/facebook/dinov2-large?library=transformers
https://huggingface.co/timm/vit_large_patch14_dinov2.lvd142m?library=transformers

[pcuenca]

The intent here is to enable the pipeline snippet in the "Use this model" button, and I think it'll work for that purpose. However, I'm not sure whether this will attempt to show an inference widget as well, along with the curl / js snippets. What do you think @SBrandeis?

The image-feature-extraction task is supported for inference here, but not here. The following curl command hence fails with {"error":"image-feature-extraction is not a valid pipeline"}

curl https://api-inference.huggingface.co/models/facebook/dinov2-large \ 
        -X POST \
        --data-binary '@example.jpg' \
        -H "Authorization: Bearer $TOKEN"

Also cc @merveenoyan @NielsRogge

[Vaibhavs10]

for snippets we'd need to add the task here as well AFAIR: https://github.com/huggingface/huggingface.js/tree/44e9e02cee0a3cb9d6e89b329f8c2d60b458e480/packages/tasks/src/snippets

[pcuenca]

for snippets we'd need to add the task here as well AFAIR:

That's for inference snippets / widgets IIUC. For "Use this model", would the current state of the PR work with no impact elsewhere? (i.e., no widget enabled)

[julien-c]

"Use this model" snippets are defined in https://github.com/huggingface/huggingface.js/blob/44e9e02cee0a3cb9d6e89b329f8c2d60b458e480/packages/tasks/src/model-libraries.ts AFAIK

[pcuenca]

"Use this model" snippets are defined in https://github.com/huggingface/huggingface.js/blob/44e9e02cee0a3cb9d6e89b329f8c2d60b458e480/packages/tasks/src/model-libraries.ts AFAIK

Yes! The issue is that image-feature-extraction models don't have a pipeline snippet because of this condition.

i.e, this is an image-feature-extraction model (AutoModel, but no pipeline snippet)

vs an image-classification one (AutoModel + pipeline snippets):

My concern is whether updating LIBRARY_TASK_MAPPING would have any side effects (i.e., widgets being enabled).

[julien-c]

worst case we merge and we'll check in prod:)

[merveenoyan]

I think this PR should work. I'm a bit late to party, sorry.

I think adding model-libraries.ts will not matter because it's mostly for third party libraries and already defined for transformers. There are two issues here:

1. transformers image-feature-extraction models don't have pipeline snippets. transformers models are maintained separately AFAIK
   them having pipeline_tag as image-feature-extraction won't matter either e.g. this one has the tag and also have the model class properly but doesn't have snippet mapped (this PR should solve that).

I have overlooked it, thanks for fixing. I will add other vision tasks in a separate PR.

2. timm models likewise doesn't have timmwrapper to load with transformers, which I'm surprised because they're inferred from here and TimmWrapper exists there. @rwightman opened another PR for it it seems

[merveenoyan]

enabled the same for VLMs with recently merged image-text-to-text pipeline here