[ENHANCEMENT] Call sweep_callback and region_callback from iterators?

[emstoudenmire]

We should consider having the default_sweep_callback and default_region_callback be called by iterating the SweepIterator and RegionIterator rather than having sweep_solve be responsible for calling these. (There is already a comment in the code suggesting this.) If we did this, then sweep_solve would just be some loops and little else but that's fine.

A key benefit would be that one could use the iterators outside of the sweep_solve setup and still benefit from the callbacks associated to a problem type and/or have a built-in callback system.

[mtfishman]

Do you mean the callback calls here: https://github.com/ITensor/ITensorNetworks.jl/blob/v0.15.4/src/solvers/sweep_solve.jl#L14 and here: https://github.com/ITensor/ITensorNetworks.jl/blob/v0.15.4/src/solvers/sweep_solve.jl#L16? I think that is by design, if you want to do the iteration yourself and include callbacks it is not so hard to write:

for _ in IncrementOnly(sweep_iterator)
    for _ in region_iterator(sweep_iterator)
        my_region_callback(sweep_iterator)
    end
    my_sweep_callback(sweep_iterator)
end
return sweep_iterator

I think the point is to keep the jobs of the iterators simpler and more self-contained, such that the iterators just have the job of incrementing and computing. Customizations can either be done by creating a new problem with a new compute step or modifying the way the iterators are called (with iterator adapters and callbacks). But maybe I'm misunderstanding what you are suggesting.

[emstoudenmire]

Thanks, Matt. These are good points. Between writing any code a user wants in the loop body, as you show above, and the customizability of the compute! step, I agree there are plenty of places to call custom code already. What I had in mind was what you were thinking I meant, prompted by a comment I saw in the code, but I agree it's not the best idea. So I'll close this issue.