GAN collections in Tensorflow

Tensorflow implemetation of various GANs on MNIST.

Generative Adversarial Networks (GANs) Papers

Name	Paper Link
GAN	Arxiv
WGAN	Arxiv
WGAN-GP	Arxiv
LSGAN	Arxiv
DRAGAN	Arxiv
CGAN	Arxiv
infoGAN	Arxiv

Access to Data

Download mnist.npz from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz .

Put it into data/mnist.

Usage

Train a wgan with gradient clipping to -0.05 to 0.05, with D and G using CNN structure, D:G=5:1.

python main.py --gan_type wgan --net_type cnn --clip 0.05 --D_iter 5 --epoch 10 

Execute run.sh to train all models and generate images.

./run.sh

Clean figs, checkpoints and logs of one model.

python main.py --todo clear --gan_type gan --net_type cnn  

Generation

CNN

Name	Epoch 0	Epoch 5	Epoch 9
GAN
WGAN
WGAN-GP
LSGAN
DRAGAN
CGAN

MLP

Name	Epoch 0	Epoch 10	Epoch 19
GAN
WGAN
WGAN-GP
LSGAN
DRAGAN
CGAN

Reference

Codes

• tensorflow-generative-model-collections
• sngan_projection
• Spectral_Normalization-Tensorflow
• WassersteinGAN
• gan_practice

CNN structure

G and D in tensorflow-generative-model-collection

MLP structure

G and D in Simple-GAN-MNIST