ShambhaviCodes/Face-Generation

Implementations of various GANs using CelebA Dataset and PyTorch

Face-Generation

0. Overview

This repository contains implementations of various GANs using CelebA dataset. The list of papers are below,

• DCGAN
• LSGAN
• Improved Techniques for Training GANs
• WGAN-GP
• BEGAN
• SAGAN
• EBGAN
• VAE-GAN

1. Qualitative Analysis

1) Generation

DCGAN	WGAN-GP	BEGAN	SAGAN	EBGAN	VAE-GAN

2) Interpolation

GANs	Interpolation
DCGAN
WGAN-GP
BEGAN
SAGAN
EBGAN
VAE-GAN

2. Quantitative Analysis

Model	IS↑	FID↓
DCGAN	2.827 ± 0.0164	6.600
WGAN-GP	2.735 ± 0.0141	5.857
BEGAN	2.362 ± 0.0131	9.942
SAGAN	2.094 ± 0.0174	6.140
EBGAN	2.499 ± 0.0186	4.946
VAE-GAN	1.823 ± 0.0074	6.314

3. Acknowledgement

Thank you for inspiration and sharing codes for metrics of IS and FID!

• PyTorch DCGAN Tutorial
• Inception Score Pytorch
• Fréchet Inception Distance (FID score) in PyTorch

Development Environment

- Ubuntu 18.04 LTS
- NVIDIA GFORCE GTX 1080 ti
- CUDA 10.2
- torch 1.5.1
- torchvision 0.5.0
- etc