Pathology Transforms inherit `torch.nn.Module`

Is your feature request related to a problem? Please describe. While we are working on improving Digital Pathology pipelines, there is a need to chain our transforms with torchvision transforms and run them on them GPU. For some reason, implementing it using ToDevice and TorchVision transform did not show an improvement whereas using the same torchvision transforms and chaining them with the model (using nn.Sequential showed a speed up.

Describe the solution you’d like Implement #2883 for all pathology transforms.

1. Derive all transform from torch.nn.Module instead of Transform
2. Implement forward for all transforms instead of __call__
3. Make all of them work with torch.Tensors

This solution should not have any impact on the rest of MONAI and should work with full backward compatibility.

[drbeh]

@Nic-Ma @wyli, as we discussed in our meeting, the main benefit that it can provide us with is to chain the transforms with the model (using nn.Sequential) and run it on the GPU.

It seems that the main reason that we are getting more speedup in this manner is that the transforms are being applied on the batch of images instead of individual one but I will look more into this and see why using ToDevice and Compose did not give us a significant speedup. Will keep you posted.

[Nic-Ma]

Hi @drbeh ,

And I think we need to check the speedup when achieving the same target metrics.

Thanks.