L EARNING D ISTRIBUTIONALLY R OBUST M ODELS AT S CALE VIA C OMPOSITE O PTIMIZATION