models.transforms.grayscale_to_rgb does not work properly with nih_cxr14

Previously, nih-cxr14 samples were converted to RGB through PIL when used during training with multi-label densenet, which worked without any problems (afaik).

The new custom grayscale_to_rgb() does not work with nih-cxr14 due to dataset consisting of 8-bit and 32-bit .PNG samples. The following error will be thrown when these images are sampled during training:

"TypeError: Input image tensor should have 1 or 3 planes,but found 4"

Examples of such 32-bit images that raise the TypeError: images/00000004_000.png, images/00000058_000.png.

This is a bug, clearly. My tests were passing because I only have the downsized version of that database installed locally.

That figure reports the following format:

<PIL.PngImagePlugin.PngImageFile image mode=RGBA size=1024x1024 at 0x7FEBA1DDE390>

Other figures report the following:

<PIL.PngImagePlugin.PngImageFile image mode=L size=1024x1024 at 0x7FEBA11D1850>

mentioned in commit 3665101d

@ogueler, @guelero: should be fixed now, on the branch add-datamodule-gradcam. Please check and report back. Thanks!

mentioned in commit 452a7b4d

closed with commit 452a7b4d

mentioned in commit 05c24a29

closed with merge request !12 (merged)