The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

cc @sbucaille As you're working on adding SuperGlue - and apologies for my previous bad advise to just use AutoModel!

Thanks @amyeroberts for the quick fix - one minor remark is that I would've waited with adding an AutoModelForKeypointDetection class before we have at least 2 keypoint detection models, in order to unify their APIs. This is because Auto classes define a certain standard API which all classes that belong to that mapping should respect.

Yes, but in this case it's needed for the SuperGlue implementation

Could you clarify why it's needed? Models can be added to the "not auto mapped" list and eventually be added to an auto class

Because we need an auto model to be used within the modeling file for SuperPoint, similar to what we do in llava

@amyeroberts I was exactly about to point the problem I would reach with SuperGlue, since it uses the outputs of models like SuperPoint (keypoints data) and not images (pixel_values). But we will be able to discuss this in the appropriate PR (btw you can reopen it as I'll push new changes in the coming days). Anyway thanks for the anticipation ! 🤗