Loading 16bit png images

[ssnl]

ported here from pytorch/pytorch#32971

Original description:

When I was trying to load 16 bit .png grayscale image with torchvision.datasets.imagefolder ,it is loading every image as white only.
I solved this issue by doing transformation operations outside Compose function.

cc @pmeier @wanifarooq @choidongyeon

[shubhamsidhwa]

What was the solution for this issue ?

[pmeier]

@shubhamsidhwa If you don't set a loader for datasets.ImageFolder, internally

vision/torchvision/datasets/folder.py

Lines 264 to 269 in 90a6206

	def default_loader(path: str) -> Any:
	from torchvision import get_image_backend
	if get_image_backend() == 'accimage':
	return accimage_loader(path)
	else:
	return pil_loader(path)

is used. Since accimage is an extra installation and activation step, I'm guessing it is falling back to Pillow in your case:

vision/torchvision/datasets/folder.py

Lines 247 to 251 in 90a6206

	def pil_loader(path: str) -> Image.Image:
	# open path as file to avoid ResourceWarning (https://github.com/python-pillow/Pillow/issues/835)
	with open(path, 'rb') as f:
	img = Image.open(f)
	return img.convert('RGB')

Their documentation states:

Pillow also provides limited support for a few special modes, including:

• [...]
• I;16 (16-bit unsigned integer pixels)
• I;16L (16-bit little endian unsigned integer pixels)
• I;16B (16-bit big endian unsigned integer pixels)
• I;16N (16-bit native endian unsigned integer pixels)
• [...]

My guess would be that you are hitting a case where Pillow does not have adequate support. Could you share an image so we can confirm hunch?

---

To overcome this you can set a custom loader that is able to handle your images. It should take a path as argument and should return the image in whatever format you need it to be.

[NicolasHug]

Pillow is kinda bad at reading 16bits grayscale pngs (e.g. python-pillow/Pillow#3011), so I'm not sure this is a torchvision issue, more probably a PIL problem.

I marked this issue to be closed once #4657 is merged: from that point you'll be able to read 16 bits pngs natively in torchvision with read_image or decode_png.