flash.image.SemanticSegmentationData dataloaders to convert greyscale/RGBA to RGB for train and val source images.

[kurtjcu]

🚀 Feature

Have The dataloaders convert any image type to rgb for use in segmentation.

Motivation

greyscale or RGBA image types crash a training run

Pitch

Most generic segmentation models work with RGB images [3 x length x width], some sources of images may contain either greyscale or RGBA image types. This simple modification would make the Flash segmentation more robust and easier to use.

Alternatives

Check image formats and give a clear error as to the the problem, current stack trace is not indicative of the problem.

Additional context

[ethanwharris]

Hi @kurtjcu Thanks for the issue!

I think the problem is here: https://github.com/PyTorchLightning/lightning-flash/blob/113efabc3e28e80ef03f9b3ed92b11156a4a3ca5/flash/image/segmentation/data.py#L146
We should be loading the image with torchvision.datasets.folder.default_loader and then using torchvision.transforms.ToTensor() to convert it to a tensor. Those methods in torchvision will ensure that the final tensors always have three channels even if the image was greyscale.

Would you be interested in making a PR that implements this fix?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.