Are Kornia augs and PyTorch Unfold used in practice?

[Vadim Kantorov]

@ajabri From what I can see in the code skimage-based augmentations are used (random crop and spatial jitter) and kornia_augs.py is not used at all (and then Appendix I is slightly incorrect, since it uses Kornia). Am I right?

If correct, what were the arguments for choosing ones over the other? Thanks!

[A. Jabri]

Thanks for the question, @vadimkantorov!

As can be seen from the code, we do not use kornia augs in practice. I found doing data aug on the GPU (with kornia) to be slower than doing it in cpu threads, especially because of the image -> patches operation; it may very well be that this can be made more efficient, though. This is mentioned in a comment in kornia_augs.py. We prioritize simplicity in the pseudocode provided in the paper, including the appendix.

[Vadim Kantorov]

Thanks for addressing this :)