Learning temporal coherence via self-supervision for GAN-based video generation