VAE with Inverse Autoregressive flows

This report replicates the work of “Improving Variational Inference with Inverse Autoregressive Flow” to improve the accuracy of Generative models. VAEs introduce a parametric Inference model to approximate the true latent posterior and optimize the parameters using an Evidence Lower Bound objective function. The follow-up work improves this by increasing the flexibility of the Inference model through Nonlinear Autoregressive Transforms called Normalizing Flow. This can lead to a tighter lower bound and higher accuracy of the Generative model. The report will replicate experiments on MNIST and CIFAR-10 datasets using standard-VAEs, simple IAF-VAE, and bidirectional ResNet VAE with PixelCNN masking.