Authors: Jiarui Xu and Hao-Wen Dong
We investigate the flow-based deep generative models. We first compare different generative models, especially generative adversarial networks (GANs), variational autoencoders (VAEs) and flow-based generative models. We then survey different normalizing flow models, including non-linear independent components estimation (NICE), real-valued non-volume preserving (RealNVP) transformations, generative flow with invertible 1×1 convolutions (Glow), masked autoregressive flow (MAF) and inverse autoregressive flow (IAF). Finally, we conduct experiments on generating MNIST handwritten digits using NICE and RealNVP to examine the effectiveness of flow-based models.
- RealNVP on toy dataset:
realnvp_toy.ipynb - RealNVP on MNIST:
realnvp_mnist.ipynb
