Spm2torch

[brudfors]

close #5

[balbasty]

A few comments:

• """ ... """ is for documentation, not for comments inside the code. Better to use # ...
• It's fine for now, but I think we'll split mathfunc differently: probably a linalg module for expm/logm, and something else for bessel. Scipy has bessel in a special module, but I don't like that too much.
• Not an issue now, but I think we'll use relative imports (rather than absolute), like in scikit-learn.
• Almost nothing in that module is automatically differentiable, nor supports batches. In the best of worlds, almost everything would be (at least) differentiable; but that's not realistic. It would be nice to find a way to make explicit what is and what is not (different module hierarchy? or a naming convention?).
• We need to reference SPM's GPL license somewhere in that file.

[brudfors]

Thanks! Will do the non-essential mods you proposed.

[brudfors]

I did not do anything about differentiable or batching.

Not sure how to acknowledge SPM in the best way. For now I just put a statement in the header. But maybe it should also go into the README of nitorch?

This link:

https://www.software.ac.uk/blog/2017-03-30-how-not-cite-software-and-how-be-cited

talks about referencing software.

They also suggest putting the license terms in each individual file? Maybe a good idea?

[balbasty]

We could make spm a package (directory with __init__.py file), and add a LICENSE file in that directory. That would maybe make things clearer.