Optional order parameter in tl.reshape can't be used with PyTorch backend

[OsmanMalik]

Optional order parameter in tl.reshape can't be used with PyTorch backend

[OsmanMalik]

Thank you @JeanKossaifi for sharing that blog post. I was wondering about the ordering of unfolding in TensorLy, and this explains it very well.

I will open a PR when the implementation is done.

[JeanKossaifi]

Hi @OsmanMalik -- yes this is an issue with the way the documentation is built for the backend functions. Ideally we'd need one doc per backend.

The issue here is that PyTorch does not support order in the reshape. Why do you need it for?
If you absolutely need a reshape with the order keyword in Pytorch you'd have to use a combination of transpose and reshapes which most likely would be inefficient.

[OsmanMalik]

Hi @JeanKossaifi, I'm implementing an ALS-based function for tensor ring decomposition and was converting over code I had previously written in Matlab. Since Matlab uses Fortran ordering, I'm using order="F" to get similar behavior in Python. But I should be able to rewrite the code to avoid using the order parameter and just stick to standard ordering, which will probably be more efficient anyway.

[JeanKossaifi]

Indeed @OsmanMalik if you're planning to run things on GPU it would be much more efficient to stick with C ordering. I had written a blog post on this, and a Stack-Overflow post, if this helps. Happy to help adapt the formulas if you need!

As a side note, would love to have the ALS-TR in TensorLy if you're interested in opening a PR when you're done! :)