DRIT: Diverse image-to-image translation via disentangled representations