Public _RandomApplyTransform class

[priba]

🚀 The feature

Make _RandomApplyTransform public for custom random apply transforms.

vision/torchvision/transforms/v2/_transform.py

Line 139 in 9ebf10a

class _RandomApplyTransform(Transform):

The main change is to refactor _RandomApplyTransform to RandomApplyTransform, add it to the available classes in torchvision.transforms.v2.

vision/torchvision/transforms/v2/__init__.py

Line 5 in 9ebf10a

from ._transform import Transform # usort: skip

Motivation, pitch

This class might be beneficial for custom transformations that should have Random behaviour. We are currently playing with non-standard augmentation pipelines that needs this.

I will submit a PR with this refactoring.

Alternatives

No response

Additional context

No response

cc @vfdev-5

[pmeier]

Thanks for the request @priba. Could you clarify what exactly you are trying to do? We already have a public RandomApply transform:

vision/torchvision/transforms/v2/_container.py

Lines 102 to 110 in bf03f4e

	def forward(self, *inputs: Any) -> Any:
	sample = inputs if len(inputs) > 1 else inputs[0]
	if torch.rand(1) >= self.p:
	return sample
	for transform in self.transforms:
	sample = transform(sample)
	return sample

Imagine you want to randomly apply a rotation. You can do

from torchvision.transforms import v2 as transforms


transform = transforms.RandomApply(
    [
        transforms.RandomRotation(degrees=30)
    ],
    p=0.7,
)

The one you linked so far is just an internal helper for classes that still have the p argument on them directly. Our vision for the future is to not write such transforms again, but rather wrap into RandomApply if we need such a behavior.

[priba]

Hi @pmeier , thanks for the reply.

Indeed, I can have the same behaviour with RandomApply. For me, it felt more natural to define it as the class rather than wrapping each one of the augmentations with the RandomApply (if I want to have different probabilities for each one).

My idea was to make this internal helper class public to be able to follow the same implementation and functionalities.

[pmeier]

My idea was to make this internal helper class public to be able to follow the same implementation and functionalities.

As mentioned above, we only have _RandomApplyTransform for BC and will likely not implement any new transformations based on it. Paired with the fact that we have a public RandomApply that a potential RandomApplyTransform would compete with, I don't see a good reason to go for that.

[NicolasHug]

Not sure if technically possible due to jit etc, but ideally we should let RandomApply accept a single transform instead of a list?

emphasis mine below:

For me, it felt more natural to define it as the class rather than wrapping each one of the augmentations with the RandomApply (if I want to have different probabilities for each one).

@priba you won't be able to do that neither with RandomApply nor with _RandomApplyTransform. We don't have anything that allows that right now other than wrapping each transform.