Temporal Autoencoding Improves Generative Models of Time Series