transforms.functional.resize broken for (h,w) tensors

[gtangg12]

🐛 Describe the bug

torchvision version: linux condaforge '0.13.1+cu113'

import torch
import torchvision

# Breaks
"""
ValueError: Input and output must have the same number of spatial dimensions, but got input with with spatial dimensions of [640] and output size of [960, 1280]. Please provide input tensor in (N, C, d1, d2, ...,dK) format and output size in (o1, o2, ...,oK) format.
"""
a = torch.zeros((480, 640))
a = torchvision.transforms.functional.resize(a, (2 * 480, 2 * 640))
print(a.shape)

# Works
b = torch.zeros((480, 640))
b = b.unsqueeze(0)
print(b.shape)
b = torchvision.transforms.functional.resize(b, (2 * 480, 2 * 640))
print(b.shape)

Versions

wget https://raw.githubusercontent.com/pytorch/pytorch/master/torch/utils/collect_env.py
# For security purposes, please check the contents of collect_env.py before running it.
python collect_env.py

cc @vfdev-5

[pmeier]

Thanks for the report @gtangg12. I guess this could actually be considered a bug, since our documentations states

vision/torchvision/transforms/functional.py

Lines 399 to 400 in a192c95

	If the image is torch Tensor, it is expected
	to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions

IMO, "arbitrary number" could also mean "no extra leading dimensions". The documentation for the transformation is the same.

Contrast this with F.gaussian_blur:

vision/torchvision/transforms/functional.py

Lines 1328 to 1329 in a192c95

	If the image is torch Tensor, it is expected
	to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions.

However, here the transformation also doesn't feature the extra channel dimension:

vision/torchvision/transforms/transforms.py

Lines 1755 to 1756 in 7fe5136

	If the image is torch Tensor, it is expected
	to have [..., C, H, W] shape, where ... means an arbitrary number of leading dimensions.

I'm not sure if we ever officially supported 2D images cc @vfdev-5. If not, I would be ok to just require the image to be a 3D image with potentially a singleton channel dimension in the docs. If the kernel can handle 2D inputs that is fine, but not documented. Thoughts?

[pmeier]

For reference, there is #6580 that tries to add full 2D support for transforms v2.