ott.initializers module implement simple strategies to initialize
solvers. For convex solvers, these initializations can be used to gain
computational efficiency, but only have an impact in that respect.
When used on more advanced and non-convex problems, these initializations
play a far more important role.
Two problems and their solvers fall in the convex category, those are the
LinearProblem solved with a
Sinkhorn solver, or the fixed-support
Wasserstein barycenter problems [Cuturi and Doucet, 2014] described in
solved with a
When the problem is not convex, which describes pretty much all other pairings
of problems/solvers in
OTT, notably quadratic and neural-network based
below, initializers play a more important role: different initializations will
very likely result in different end solutions.