baojunshan/gan

Some gan implementation pytorch.

Generative Adversarial Networks

Some gan implementations with pytorch. Welcome to PR and issue.

Installation

$ git clone https://github.com/baojunshan/gan.git
$ cd gan/
$ pip3 install -r requirements.txt

Train Data

./*gan/utils.py contains ImageLoader class which can load train data. The input accepts incompressible pictures.

• mnist(png): Baidu Cloud (password:ynz7)
• anime face(png): Baidu Could (password:nmj2)

Usually, the data path should be like ./data/dataset_name/*.png. Some labelled data path should be like ./data/dataset_name/label_name/*.png. Otherwise, you can custom your ImageLoader to load data.

Contents

Model	Paper	Code	Results
Base
├ GAN	[2014.06] Generative Adversarial Networks	Code	Results
├ DCGAN	[2015.11] Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks	Code	Results
├ LSGAN	[2016.11]Least Squares Generative Adversarial Networks	Code	Results
├ WGAN	[2017.01] Wasserstein GAN	Code	Results
├ WGAN-gp	[2017.04] Improved Training of Wasserstein GANs	Code	Results
├ WGAN-div	[2017.12]Wasserstein Divergence for GANs	Code	Results
├ SAGAN	[Self-Attention Generative Adversarial Networks]	Code	Results
└ PGGAN			Results
Conditional
├ CGAN	[2014.11] Conditional Generative Adversarial Nets	Code	Results
├ InfoGAN			Results
└ StyleGAN			Results
Understanding
├ CCGAN			Results
├ Pix2Pix			Results
├ CycleGAN			Results
└ SRGAN			Results

Implementations

GAN

$ cd gan
$ sh run_mnist.sh
$ sh run_anime.sh

DCGAN

$ cd dcgan
$ sh run_mnist.sh
$ sh run_anime.sh

LSGAN

$ cd lsgan
$ sh run_mnist.sh
$ sh run_anime.sh

WGAN

$ cd wgan
$ sh run_mnist.sh
$ sh run_anime.sh

WGAN-gp

$ cd wgan_gp
$ sh run_mnist.sh
$ sh run_anime.sh

WGAN-div

$ cd wgan_div
$ sh run_mnist.sh
$ sh run_anime.sh

SAGAN

PGGAN

CGAN

$ cd cgan
$ sh run_mnist.sh

InfoGAN

StyleGAN

CCGAN

Pix2Pix

CycleGAN

SRGAN