16bit png images for classification?

[andife]

Hi,
I wonder if anyone has patched https://github.com/ludwig-ai/ludwig/blob/master/ludwig/utils/image_utils.py to work with 16bit png images? (At the moment 16bit png images do not work)

I know there are some problems with the support in Pytorch (uint16 is not supported
pytorch/pytorch#58734),
but torchvision managed the handling with that
pytorch/vision#2218, pytorch/vision#4657

Are there any plans for adding that? Would it be enough to adjust the loading of the images, or would you then also have to adjust the model etc. accordingly. Another question with image augmentation does not arise, because that would have to be done externally by ludwig anyway?

Thank you

[ShreyaR]

Hi @andife, thanks for raising this issue. After reading through the linked PyTorch issues, I believe it should be feasible to support 16-bit PNG images. As a stopgap solution, we can use the private API _decode_16bits_png() (referenced in pytorch/vision#4731).