Rethinking Consistent Multi-Label Classification under Inexact Supervision