solving inversed problem with newton optimizer will be quiet slow

[kangbeicai]

When using newton optimizer , solving a inversed problem will be quiet slow and allocate more memory compared with a forward problem . I try to change "permc_spec" in scipy.sparse.linalg.spsolve to default ”COLAMD“, and found that it can significantly increase speed and reduce memory usage.

[pkarnakov]

Thanks for sharing this observation! I haven't tried COLAMD myself, but it's good if that makes the linear solver more efficient. The option I normally use for linear system is the multigrid solver from pyamg, called with --linsolver multigrid