Nd total variation

[JakobAsslaender]

Hi @migrosser and @tknopp,

I developed this code a bit further. The key changes are:

• create a truly ND version, i.e., you can supply this code with 5D, 17D, or, my favorite, 42D data
• allow to freely choose any single dimension or combination of dimensions that the TV will be calculated along
• dispatch to Condat algorithm if the TV is calculated along a single (arbitrary) dimension (this is much faster)
• reduce allocations (GC reduced form 80% to 0%) and speed the code up by a few fold.

The part that bothers me the most is that you cannot dispatch on keyword arguments and I am dispatching over dims to choose the Condat or the FDG algorithm. And we do need keyword arguments with the current wrapper functions of your regularizer setup. The hack to fix this is to have a function call with keyword arguments as a wrapper and then dispatch w/o keywords. Ain't pretty--let me know if you have a better idea.

[migrosser]

To me the PR looks quite nice. Concerning the dispatch over dims, the function barrier you implemented is probably the most elegant workaround, which doesn't require any breaking changes of our regularizer interface.