Fork or Fail: Cycle-Consistent Training with Many-to-One Mappings