DCGAN_LSGAN_WGAN_WGAN-GP_SNGAN_RSGAN_BEGAN_ACGAN_PGGAN_pix2pix_BigGAN
Implementation of some different variants of GANs
Introduction
This code is mainly implement some basic GANs about 'DCGAN', 'WGAN', 'WGAN-GP', 'LSGAN', 'SNGAN', 'RSGAN'&'RaSGAN', 'BEGAN', 'ACGAN', 'PGGAN', 'pix2pix', 'BigGAN'.
More details of these GANs, please see follow papers:
-
DCGAN: Unsupervised representation learning with deep convolutional generative adversarial networks
-
WGAN: Wasserstein gan
-
SNGAN: Spectral normalization for generative adversarial networks
-
RSGAN&RaSGAN: The relativistic discriminator: a key element missing from standard GAN
-
BEGAN:BEGAN: Boundary Equilibrium Generative Adversarial Networks
-
ACGAN: Conditional Image Synthesis With Auxiliary Classifier GANs
-
PGGAN: Progressive Growing of GANs for Improved Quality, Stability, and Variation
-
pix2pix: Image-to-Image Translation with Conditional Adversarial Networks
-
BigGAN: Large Scale GAN Training for High Fidelity Natural Image Synthesis [Code]
Attention
If your computer don't have GPU to accelerate the training process, please click Google Cloud Colab to train the GANs.
How to use
Firstly, you should download the data 'facedata.mat' from Baidu Drive or Google Drive, then put the file 'facedata.mat' into the folder 'TrainingSet'.
Requirements
- python3.5
- tensorflow1.4.0
- pillow
- scipy
- numpy
Results of this code
This result is using DCGAN trained about 8000 iterations.
Compare LSGAN, WGAN, WGAN-GP, SNGAN, RSGAN of different iteration
Convergence of BEGAN
ACGAN for face generating
dataset: download address: Baidu Drive password: 5egd
Fixed label, change noise slightly | Fixed noise, change label slightly |
---|---|
PGGAN for face generating
SNGAN for cifar-10
D_loss | G_loss | results |
---|---|---|
Pix2Pix
Dataset: Google maps download address: http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/maps.tar.gz
Edges2Shoes download address: http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/edges2shoes.tar.gz