Clockwork Variational Autoencoders for Video Prediction