Use same seed for transform function across all images in a batch

vision
1

I am training a model which requires variable sized inputs ((4, 4), (8, 8), (16, 16)) for data augmentation.

I am using the following the transform function to achieve the same

        sizes = [(4, 4), (8, 8), (16, 16)]
        
        ResizeTransformation = Transforms.RandomChoice(
            [Transforms.Resize(size) for size in sizes])

        transformList = [Transforms.RandomRotation(180, Image.BILINEAR),
                         Transforms.CenterCrop(220),
                         Transforms.RandomCrop(128),
                         Transforms.RandomHorizontalFlip(),
                         ResizeTransformation,
                         Transforms.ToTensor(),
                         Transforms.Normalize((0.1483, 0.1210, 0.1257), (0.1292, 0.1050, 0.1186))]

But this creates an issue when using a batch size > 1, As there are cases when images in the batch are resized to different dimensions.

I’m thinking that If we use the same seed value while transforming all the images in a batch, we can get a single tensor of shape (N, C, H, W) as the transform function will resize to the same dimensions.

I have looked up on the forum that we can create a list of tensors of batch size of different sizes by using a custom collate_fn, but then aren’t we training the network on a batch_size = 1 only as we’ll have to pass all the images to the networks separately since they are of different sizes?

Looking for a way to create single tensor of dimensions (N, C, H, W) with variable sized inputs.

TIA

2

Maybe it would be easier to remove the ResizeTransformation from transformList and apply F.interpolate on the complete batch inside the DataLoader loop?
This would of course not be applied from each worker in Dataset.__getitem__, but might not necessarily be slower.

Would that work for you? I’m always a bit skeptical when it comes to using the seeding mechanism to fix these random transformations, as all other “random” transformations will also be fixed (e.g. RandomRotation).