On the quantitative analysis of decoder-based generative models