Cross-Domain Learning with Normalizing Flow