Great feature that's been on the backlog for a while, thanks @merajhashemi and @aarmey!

Currently this is NumPy specific though and not all backends support array_function so this won't work in all cases. Shall we move it inside the backend class and replace the np. by self.?

Hi @merajhashemi -- are you still working on this?

Codecov Report

Merging #216 (01cbffe) into main (bfda610) will decrease coverage by 0.20%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##             main     #216      +/-   ##
==========================================
- Coverage   87.08%   86.88%   -0.21%     
==========================================
  Files          89       89              
  Lines        4446     4446              
==========================================
- Hits         3872     3863       -9     
- Misses        574      583       +9     

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update bfda610...01cbffe. Read the comment docs.

Thanks @merajhashemi - it would be good to have a small unit-test.

Wondering whether it would make sense to unify with cp_sign_flip @cohenjer @aarmey

I see how they address similar needs, but I don't see a way to unify this with cp_sign_flip, unless I'm missing something.

Unless you want to have a weird, general sign-flip function, I would also keep these separated? But maybe I am simply not understanding what you have in mind @JeanKossaifi

Was just a thought as SVD can be seen as a 2D CP with weights = diag(S) and factors = [U, V], so it would be possible to resolve the sign ambiguity for all cases. But indeed the way it's done is different enough that a unified function would probably be awkward. Might be best leave the two separate indeed.

If this is just missing a test, I'm happy to write one as part of #227, since that will likely require some refactoring of the tests as well.

Awesome, thanks @aarmey, in that case I'll go ahead and merge!