KID-22 / Deep-Generative-Models

VAE, GAN, WGAN, WGAN-GP on two synthetic Gaussian datasets.

Deep Generative Models

In this project, we implemented VAE, GAN, WGAN, WGAN-GP on two synthetic datasets with Pytorch.

Folder Structure

The following shows basic folder structure.

├── fig # figures
├── data
│   ├── circle_1 # circle_1 data
│   ├── circle_1 # circle_1 data
│   └── generate_data.ipynb # generate synthetic dataset
├── code
│   ├── model # include GAN.py VAE.py WGAN.py WGAN_GP.py
│   ├── main_vae.py # gateway for VAE
│   ├── ...
│   ├── ...
│   ├── main_wgan_gp.py # gateway for WGAN_GP
│   └── utils.py # utils
├── ckpt # model checkpoints to be saved here
├── result # generation sample results and loss to be saved here
└── log # training log to be saved here

Dataset

• Circle 1: This dataset consists of 28 components, where each component is isotropic and generated from a mixture of 2-dimensional standard normal Gaussian. The 28 components are uniformly distributed on a circle (radius=70).

• Circle 2: Similar to Circle 1 dataset, Circle 2 dataset also consists of 28 isotropic standard normal Gaussian components. Differently, the 28 components are uniformly distributed over 2 circles, 8 components on the small circle (radius=25) and 20 components on the large circle (radius=50).

Reproducing Results

For a easy way to reproduce any results in this project, you can run the command as follows.

sh run.sh

To run the model on a different dataset, or with different hyper-parameters, etc, simply modify the config in the main_*.py file.

Results

Sample Quality

Circle 1

Circle 2

Visualization of Loss

Circle 1

Circle 2