Vadim Kantorov

at the very least, the actual behavior should be clearly explained in the docs...

I think for all models in torchvision, being able to serve as backbone is top-1 or top-2 usage anyway. People usually figure out how to remove the classifier, and just...

Probably this repo/detectron2 also has swin adapted as a backbone: https://github.com/facebookresearch/detectron2/tree/main/projects/ViTDet

Related discussion and pointers on generalizing fixed resolution for Swin: https://github.com/pytorch/vision/issues/6227 Also, I wonder if more relative-attention related modules can be reused from Swin

@fmassa Is general `torchvision.io.read_image` affected? In general, is `torchvision.io.read_image` ready for being used in dataset classes instead of PIL / OpenCV for reading images from disk? Is it fast?

Maybe this is not even related to transforms per se, since useful also in non-augmentation contexts as well. For transforms, I really hope, there is more native batched transforms, clear...

I think it's not a problem per se, as we almost do not encode semantics into the tensor type system in PyTorch, in the goal of being flexible and letting...

I think many transforms actually don't care about the input color space and support everything. Yes, it would be good to add notes to the docs about supported color spaces....

You are right. Of course it would be invalid to put in LAB image into a RGB-trained model. So yes, this would only be for new models or new functionality....

Well, one also does not have explicit need for drawing bounding boxes within the library (and yet it is was merged in). I think this argument is not very valid....