pil_to_tensor returns non-contiguous tensor

[aaronzs]

🐛 Bug torchvision.transforms.functional.pil_to_tensor returns non-contiguous tensor

To Reproduce

from PIL import Image
from torchvision.transforms.functional import pil_to_tensor

pic = Image.open("3_channel_pic.png")
pil_to_tensor(pic).is_contiguous()
# False

Expected behavior

.contiguous() is recommended to be used after .permute(), so that

1. consistent behavior with to_tensor();
2. increase computational efficiency

cc @vfdev-5

[datumbox]

I believe this would require copying the memory here:

vision/torchvision/transforms/functional.py

Line 168 in 993325d

img = img.permute((2, 0, 1))

I think the original implementation was written to avoid this. Any thoughts @vfdev-5 ?

[vfdev-5]

Yes, img.permute((2, 0, 1)) returns non-contiguous tensor and avoids data copy. Here is a discussion on why keeping channels last memory format here: #2092 (comment)

[datumbox]

@aaronzs I believe the above addresses the issue, so to keep things tidy I'm closing the ticket. If you still have concerns feel free to reopen.