This PR adds a class ConvergenceMonitor to monitor convergence for all the
solvers in a uniform way.
One debatable issue is that it computes the residual vector inside the object.
You could argue that this disadvantages methods (PGD, CD) that work directly with the
residual vector, but on the other hand you might say that not doing it would be
unfair to other methods (ADMM, NewtALM) that don't need to compute the residual
vector at all.
I think this is a nice addition! Thanks @jolars. I will start investigating some behavior observed on the benchopt repo with this ConvergenceMonitor Class
feat: add convergence monitor
This PR adds a class ConvergenceMonitor to monitor convergence for all the solvers in a uniform way.
One debatable issue is that it computes the residual vector inside the object. You could argue that this disadvantages methods (PGD, CD) that work directly with the residual vector, but on the other hand you might say that not doing it would be unfair to other methods (ADMM, NewtALM) that don't need to compute the residual vector at all.