Differentiable Data Augmentation With Kornia

Reviewer 1

The presented usecases and strategies can help us greatly reduce the amount of time we would need to implement difficult algorithms like STNs and to create new interesting models for new problem domains to achieve new state-of-the-art results where differentiable data augmentation can significantly help. This benefit was demonstrated in recent works like bilevel optimization (https://github.com/ElementAI/bilevel_augment).

Reviewer 2

The main contribution of this work is to present an easy-to-use implementation of differentiable data augmentation techniques in PyTorch. In addition to being differentiable, the augmentations described in the paper are also faster than existing implementations on large images. The authors make a compelling case that the contributions presented here will be useful to people doing computer vision research.

The authors do not describe any challenges they had to overcome in order to make this work possible, which makes it feel that the work amounts to replacing np.* with torch.* in some well-known image augmentation procedures. I think the paper would be stronger if it put more emphasis on explaining the difficulties related to backpropagating gradients of data augmentation procedures, and the clever ways in which the authors overcame them.