This repo contains a Pytorch implementation of Planar Flow presented in (Rezende and Mohamed, 2015) with experiments on a 2D density and MNIST dataset.
| Input | Model | Latent Space Size | Test Lower Bound |
|---|---|---|---|
| [0,1] | VAE | 20 | -99.37 |
| [0,1] | VAE+PF (K=20) | 20 | -98.23 |
| {0,1} | VAE | 20 | -84.60 |
| {0,1} | VAE+PF (K=20) | 20 | -81.83 |
[0,1] denotes float values between 0 and 1 and {0,1} denotes binary values.
Vanilla VAE: python vae.py
VAE with Planar Flow: python vae-pf.py
Add --binary option to binarize the input dataset.
(Rezende and Mohamed, 2015) Rezende, Danilo, and Shakir Mohamed. "Variational Inference with Normalizing Flows." International Conference on Machine Learning. 2015.