class MySegmentationDataset(torch.utils.data.Dataset):
    def __init__(self, transforms=None, target_transforms=None, generators=None):
        self.transforms = transforms
        self.target_transforms = target_transforms
        self.generators = generators

    def __getitem__(self, idx):
      # load img and target
      ...
      # generate random numbers, if generator is provided
      if self.generators is not None:
            for g in self.generators:
                g.generate()

        if self.transforms is not None:
            img = self.transforms(img)

        if self.target_transforms is not None:
            target = self.transforms(target)

        return img, target

# create random generators
crop_generator = torchvision.transforms.RandomCropGenerator()
flip_generator = torchvision.transforms.RandomFlipGenerator()

generators = (crop_generator, flip_generator)

# random transforms that consumes the generators
crop = RandomCrop(img_size, generator=crop_generator)
flip = RandomHorizontalFlip(generator=flip_generator)

# composed transforms that reuses crop and flip
train_input_transform = torchvision.transforms.Compose([
    crop,
    flip,
    torchvision.transforms.ToTensor(),
    normalize,
])
train_target_transform = torchvision.transforms.Compose([
    crop,
    flip,
    torchvision.transforms.ToTensor(),
    torchvision.transforms.Lambda(lambd=lambda x: x.long())
])

# create dataset, but need to pass the generators as well
mydataset = MySegmentationDataset(train_input_transform,
                    train_target_transform,
                    generators)

img, target = mydataset[0]