Comparison of Deep Generative Models

In recent years, generative models have been very effective in learning the under-lying distribution of data sets and are able to reproduce a variety of results thatare indistinguishable from the original dataset. In this project we will compareand discuss the generative capabilities of three different deep generative models:GAN, Wasserstein-GAN, and Variational autoencoders (VAE) on the EMNISThand-written letters data set

Build Environment

conda env create -f environment.yml
conda activate gan

Training WGAN

To train WGAN, run all the cells in:

WGAN.ipynb

To generate a few WGAN samples from a pretrained model call:

python wgan_generate_samples.py

To evaluate the reconstruction error:

python wgan_reconstruction_error.py