Avoid calling torch.rand in Transformation.forward()

[datumbox]

The preferred way to structure the Transformation classes is to put the initialization of random weights/params in a static get_params() method. The method should receive any hyper parameter necessary for the sampling and it should return all the necessary random variables. This method should be called by forward() during the transformation process. This is an example of how this would look:

vision/torchvision/transforms/transforms.py

Lines 530 to 531 in 9e71fda

	@staticmethod
	def get_params(img: Tensor, output_size: Tuple[int, int]) -> Tuple[int, int, int, int]:

vision/torchvision/transforms/transforms.py

Line 589 in 9e71fda

i, j, h, w = self.get_params(img, self.size)

Unfortunately many forward() methods call directly torch.rand. Here are a few examples:

vision/torchvision/transforms/transforms.py

Lines 452 to 453 in 9e71fda

	def forward(self, img):
	if self.p < torch.rand(1):

vision/torchvision/transforms/transforms.py

Line 619 in 9e71fda

if torch.rand(1) < self.p:

vision/torchvision/transforms/transforms.py

Line 649 in 9e71fda

if torch.rand(1) < self.p:

vision/torchvision/transforms/transforms.py

Line 700 in 9e71fda

if torch.rand(1) < self.p:

vision/torchvision/transforms/transforms.py

Line 1454 in 9e71fda

if torch.rand(1) < self.p:

vision/torchvision/transforms/transforms.py

Line 1560 in 9e71fda

if torch.rand(1) < self.p:

There might be potentially others. We should refactor the codebase so that all of the above calls happen within a static get_params() method. See #3065 RandomInvert for an example on how to structure it.

cc @vfdev-5

[datumbox]

Closing this as we need to rethink the plan to cover for cases such as object detection. See #3286