GromovWasserstein API unclear compared to Sinkhorn

[gjhuizing]

The GromovWasserstein solver hides quite a lot of arguments in *args and **kwargs.
For a new user it is unclear that you can set an implicit_diff argument for example, since they would have to navigate first to WassersteinSolver, then to Sinkhorn.

This could be fixed either by making some arguments explicit, or by expanding the documentation to make it just as nice as the Sinkhorn solver.

[marcocuturi]

Hi Geert, thanks for your great comment.

I think you are right, this is worth refactoring a bit.

We have args, kwargs and kwargs_init. It's also true that the WassersteinSolver class is a bit obscure, and needs better context. It's essentially used to solve problems where various (e.g. barycenter) or one (GW) linear solver is repeatedly used. Will keep this issue open until we find a better solution.

[gjhuizing]

I'm not that familiar with the codebase, but would it be possible to remove default values in the parent class WassersteinSolver and just repeat the arguments explicitly in GromovWasserstein?

Actually, some arguments seem repeated in GromovWassersteinBarycenter already! So regardless of the chosen solution, this should be harmonized.