large sparse matrix causing crash during decomposition

[jtliso]

When I try to use Tucker decomposition of a large sparse matrix, Tensorly crashes. I have used both the MXNet and NumPy backend, and both cause the crash due to memory issues.

The dimensions of my sparse matrix are (358, 556, 2). I was hoping to use Tensorly for even larger sparse matrices. I did not know if you guys intend to release any support for sparse matrices, or if perhaps something I am doing could be incorrect.

[JeanKossaifi]

We don't yet have support for sparse tensors but it is on the roadmap!

[MaxMcGlinnPoole]

@JeanKossaifi Great to hear it is on the roadmap. I also need this functionality for the project I am working on. Keep up the great work!

[JeanKossaifi]

Thanks! We welcome contributions if you are interested in taking a crack at it! :)

[JeanKossaifi]

As a side note, you should have no problem to factorize (dense) tensors of that size. I recently ran CP and Tucker factorization (on an AWS instance, with one Tesla V100 GPU) on a tensor of size 1000x1000x1000 in about 40 seconds using pytorch, mxnet or tensorflow.

[santiago-py-chen]

Hi, I wonder if it is normal that when I was calling partial_tucker on a given tensor, it always returns core tensor and factors with the same magnitude but different signs? Even if I fixed the random_state, the sign of returned core tensor and factor were still different from time to time. (Not sure if it should be raised here, I would move it to somewhere else if inappropriate)

[JeanKossaifi]

It is the result of the sign indeterminacy of the singular value decomposition.
See for instance this article about the problem and a possible solution to it. We can add a SignFlip function and an additional parameter in the decomposition algorithms on whether to use it.

I opened #74 -- feel free to take a crack at it!