GANs Trained by a Two Time-Scale Update Rule Converge to a Nash Equilibrium