CS480 Final Project: Evaluating and comparing VAE with VampPrior and standard Gaussian prior

This repository contains the code, trained models, and resulting generated images used for our CS480 final project at the University of Waterloo. Code for generating and training different models based on the VampPrior and standard Gaussian prior is based on the following paper:

Jakub M. Tomczak, Max Welling, VAE with a VampPrior, arXiv preprint, 2017

Requirements

The code is compatible with:

pytorch 0.2.0

Data

The experiments can be run on the following datasets:

static MNIST: links to the datasets can found at link;
binary MNIST: the dataset is loaded from PyTorch;
OMNIGLOT: the dataset could be downloaded from link;
Caltech 101 Silhouettes: the dataset could be downloaded from link.
Frey Faces: the dataset could be downloaded from link.
Histopathology Gray: the dataset could be downloaded from link;
CIFAR 10: the dataset is loaded from PyTorch.
Fashion MNIST: the dataset is loaded from PyTorch.

Run the experiment

Set-up your experiment in experiment.py.
Run experiment:

python experiment.py

Models

You can run a vanilla VAE, a one-layered VAE or a two-layered HVAE with the standard prior or the VampPrior by setting model_name argument to either: (i) vae or hvae_2level for MLP, (ii) convvae_2level for convnets, (iii) pixelhvae_2level for (ii) with a PixelCNN-based decoder, and specifying prior argument to either standard or vampprior.

The trained models used and evaluated in our project can be found in the trained_models directory. Specifically, models for the mnist, frey faces, cifar10, and fashion mnist datasets were trained. Each model uses either a VampPrior or standard prior built on a vanilla VAE, two-layered HVAE, 2 level convolutional network, or 2 level PixelCNN-based decoder.

Quantitative Evaluations

Scripts for generating the FID (Frechet Inception Distance) and IS (Inception Score) for the trained models can be found under the evaluations directory.

The metrics for each trained model (KL divergence, Log Likelihood, ELBO, IS, FID) can be found in metric.txt.

Results

The generated images produced from our trained models can be found under the generate directory. The script generate_fake_images.py was used to generate images from the trained models.

Citation

Code used in this project is based on the paper mentioned above and should be appropriately cited as such:

@article{TW:2017,
  title={{VAE with a VampPrior}},
  author={Tomczak, Jakub M and Welling, Max},
  journal={arXiv},
  year={2017}
}

RobertYCXu/vae_vampprior