All nan in matrix come from non negative tucker decomposition

[hahia]

Code:

from tensorly.decomposition import non_negative_tucker
tl.set_backend('pytorch')
tensor = tl.tensor(dat, device='cuda:3')
facs_overall = non_negative_tucker(tensor,rank=[8,8, 8],random_state = 2337)

dat is

The facs_overall

Could someone provide an explanation for this?

[JeanKossaifi]

What's your input tensor? Can you produce a minimal snippet of code to reproduce your error? It seems your tensor is very sparse / could it also be low-rank?

It also seems the range might be problematic - I think that's also link with another thing I want to add in tensorLy: optional normalization of the input tensor (and unnormalizing directly in factorized form) @cohenjer @aarmey

As a small test, could you try i) lower rank and ii) normalizing your tensor (e.g. remove mean and divide by std).

[cohenjer]

Just to complete @JeanKossaifi's great answer, you can easily normalize a tensor T using T = T/tl.norm(T) but centering would make some values negative in the tensor so I would not do it in the context of nonnegative decomposition.

Something bad which may also occur is if whole slices of the tensor are zero. non_negative_tucker is based on multiplicative updates, you may be dividing by zero at some point in the algorithm. There is a safeguard (a small epsilon 1e-12 is added) but maybe with your precision this epsilon is too small and considered 0?