Batch support for torchvision.transforms.functional

[Shubhammawa]

Support for a batch of images in torchvision.transforms.functional.

Currently the torchvision.transforms.functional transformations take only a single image as input. Sometimes after creating the dataloader, I need to transform the images during the training loop, but passing the batch of images to transforms.functional raises an error expecting a 2/3 dimensional input while it gets a 4-dimensional input. I am currently solving the issue by looping through the images in the batch but it would be nice if the transforms could handle a batch of images on their own.

[fmassa]

Hi,

We are progressively adding support for tensors inputs in the transforms.functional, and with this it will have support for batch of images as well.

See

vision/torchvision/transforms/functional_tensor.py

Lines 594 to 596 in 62e3fbd

	if img.ndim < 4:
	img = img.unsqueeze(dim=0)
	need_squeeze = True

which gets dispatched to from torchvision.transforms.functional.resize when the input is a tensor #2292

As such, I'm closing this as a duplicate of #1375