torch.linalg.cholesky warnings

[ry-jojo]

Hi Joakim,
When training on 10shot, I am facing with warnings like the below:

WARNING batched routines are designed for small sizes. It might be better to use the
Native/Hybrid classical routines if you want good performance.

I think this warning comes from the torch.linalg.cholesky(K_ss), when K_ss's largest size() > 2048.

May I know how you deal with this warning during training? Thanks in advance!

[joakimjohnander]

Hey,

Unfortunately, we have no simple fix that removes the warning. It seems to stem from a package, magma, that is called by PyTorch.

Best regards,
Joakim

[Parskatt]

@ry-jojo You can do a for-loop instead. Not sure about the performance loss/gains.