@fcakyon, why was I mentioned here? The only time I edited this file was to add a missing space in an error message.

My apologies @aphedges. Saw your name on PR.

Thanks for the issue!

Since a batch of images is a very common input I think we should support this. Pinging @NielsRogge, @sgugger

@fcakyon Also note that when resizing, the feature extractors convert the tensors to PIL images, so the most efficient way to use these is to pass a list of PIL images.

@patil-suraj I was using as:

PIL Images > Pytorch DataLoader > batch as torch.Tensor > CLIPFeatureExtractor > CLIPModel

But in this workflow CLIPFeatureExtractor treats batch as torch.Tensor as a single image tensor and as a result CLIPModel gives error.

Is the intended usage different?

We could support batch as tensors but as explained in the above comment it won't be efficient. So I would recommend passing a list of PIL images.

@patil-suraj Do you still see supporting batched tensors as necessary/beneficial? It seems pretty straightforward (e.g. checking ndim), but I thought I'd ask for confirmation before jumping into a draft PR.