[FeatureRequest] Add Equivalent to Models Genesis Transform about rising HOT 2 CLOSED

Hi,
I started working on a rising implementation of the genesis transforms. Basically, there are three transforms, (1) local-shuffling, (2) in/outpainting, and (3) non-linear bezier transform.
(1) and (2) should be straightforward to implement, I should be able to do this next week.
However, (3) is difficult, because pyTorch doesn't yet support unstructured interpolation. Could be that it will be implemented in a future version, 3rd party implementations are already available: pytorch/pytorch#1552

For the moment, I'm using the interpolation function in numpy (master...weningerleon:master). Of course, it is not possible to do the augmentation on the GPU in this way.
What do you think, should we just wait until a pytorch interpolation function is available? Any other suggestions? It wouldn't be ideal if an external dependency would be required for rising....

from rising.

Theoretically you could look, whether Google Jax offers the functions you need. You could then execute it in GPU and change the results to a torch tensor via dnnpack format without memory copy. But in general I think, we can either depend on some external lib or implement it ourselves.

from rising.