Could anyone explain why we need log aspect ratio for RandomResizedCrop? vision/transforms.py at 7d4154735f421b254c408c16e0980b1ca0dd9b8e · pytorch/vision · GitHub I could not find any reference for this:
log_ratio = torch.log(torch.tensor(ratio))
It seems this sampling was added in this PR to avoid the bias in the uniform sampling from the defined aspect ratios. This comment explains it in more detail.