Feature Request: make selectors amenable to GPU processing

[rosecers]

as requested by @Luthaf

[Luthaf]

So for more context on this: it would be nice to be able to pass a torch.Tensor (or jax array) living on GPU directly to the selectors, instead of having to move the data back to main CPU memory.

A first pass would be to make sure all the function calls are compatible with PyTorch API, but given the high usage of Python for loops in the selector code that might not give a lot of performance improvement. The second step would then be to rewrite the selector code to use more high-level operations & launch larger GPU kernels, and hopefully improve performance.

---

This is mostly unrelated to the autograd part of PyTorch, so even if we need to .detach() the tensors before passing them, that would be fine with me. I would mostly like to be able to keep the data in GPU memory.

[Luthaf]

My ideal user-facing interface for this would be to be able to do something like this:

import torch
from skmatter.feature_selection import CUR

X = torch.rand(300, 300, device="cuda")  # or device="mps" on Apple M1/M2

selector = CUR(n_to_select=4)
selector.fit(X)

Xr = selector.transform(X)
# Xr is a torch tensor, with device=X.device

A first step for this would be to add a test trying to use skmatter with a torch tensor, and check where the code starts throwing errors.

---

Depending on the number of function call (e.g. np.sum, …) that need to be updated, it might be interesting to use https://github.com/jcmgray/autoray to dispatch function calls to the right backend.

[Luthaf]

This is put on the back burner for now, if you are interested in getting skmatter to run on GPU please voice your interest here!

[Luthaf]

It looks like sklearn now has experimental support for PyTorch/CuPy (and thus GPU data) using the array API: https://scikit-learn.org/stable/modules/array_api.html. We could use the same here!

[agoscinski]

We should experiment as well how the array api works with our selection methods. FPS is probably a good candidate because we do not use very complicated mathematical operations there. So hopefully there is not so much friction in making this work.

[Luthaf]

More info on this array API in sklearn: https://labs.quansight.org/blog/array-api-support-scikit-learn.