How to Train Deep Variational Autoencoders and Probabilistic Ladder Networks