[Feature]: Better documentation for what W&B does when logging a PyTorch tensor as a PIL Image

[RylanSchaeffer]

Description

This is W&B's documentation for Images: https://docs.wandb.ai/ref/python/data-types/image

There are two cryptic sentences: "Note : When logging a torch.Tensor as a wandb.Image, images are normalized. If you do not want to normalize your images, please convert your tensors to a PIL Image."

Suggested Solution

Please say a little more about what this means. When you say images are normalized, what specifically is meant? Subtract the mean or divide by the maximum? Is the mean/maximum calculated per channel or over all three channels? Do we also divided by the standard deviations i.e. is this really standardization? Additionally, if we pass a numpy array instead of a torch tensor, is this normalization still applied?

More documentation about this point would be appreciated! I just wasted a lot of compute saving images that weren't the images I thought they were 😢

Alternatives

No response

Additional Context

No response

[JoanaMarieL]

Hello @RylanSchaeffer , thank you for your suggestion we appreciate this. I already created a feature request and tagged our Docs team to check it. We'll provide an update for any progress about it.

[TonyPolich]

Agreed! What is particularly unclear is if the images are normalised on a per image basis or per batch.