Resize ignores `max_size` when short side equals to target size

[izikgo]

🐛 Describe the bug

When using the resize function, if the target size equals to the smaller side of the image, the image is returned as-is, even if the longer side exceeds the given max_size. This can be seen at https://github.com/pytorch/vision/blob/main/torchvision/transforms/functional_pil.py#L244

import torch
import torchvision

H, W = 32, 40

img = torch.randint(256, (3, H, W))
resized_img = torchvision.transforms.functional.resize(img, H, max_size=H)

print(resized_img.shape[2])  # This is 40 instead 32

Versions

PyTorch version: 1.10.2+cu102
Is debug build: False
CUDA used to build PyTorch: 10.2
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.3 LTS (x86_64)
GCC version: (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.31

Python version: 3.8.6 (default, Dec 12 2021, 11:32:40) [GCC 9.3.0] (64-bit runtime)
Python platform: Linux-5.10.60.1-microsoft-standard-WSL2-x86_64-with-glibc2.29
Is CUDA available: True
CUDA runtime version: Could not collect
GPU models and configuration: GPU 0: GeForce MX250
Nvidia driver version: 465.42
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A

Versions of relevant libraries:
[pip3] numpy==1.22.2
[pip3] torch==1.10.2
[pip3] torchvision==0.11.3
[conda] Could not collect

[NicolasHug]

Thanks for the report @igolan89

What exactly are you trying to do here?

It isn't natural to ask for a given smaller edge size (size) while asking for the same larger edge size (max_size). That's asking for a square image, which would be more naturally achieved with resize(img, (H, H)).

Because of this, later in the code we raise an error if max_size <= size:

vision/torchvision/transforms/functional_pil.py

Lines 249 to 253 in f2f490b

	if max_size <= size:
	raise ValueError(
	f"max_size = {max_size} must be strictly greater than the requested "
	f"size for the smaller edge size = {size}"
	)

I guess we could remove the early return that you linked to, but it would just lead to an error in this case. What do you think would make more sense?

[izikgo]

Thanks for the reply @NicolasHug.
The fact that max_size=size is just for the illustration. The same thing will happen regardless of the value of max_size (i.e., it can be H+1)

[NicolasHug]

Thanks for the details @igolan89 , I think you're right!

#5409 should fix the issue.