Skip to content
Implemention of "CVAE-GAN: Fine-Grained Image Generation through Asymmetric Training"
Branch: master
Clone or download
Latest commit 37fe060 Jun 21, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
cfgs trival Nov 21, 2018
classifier trival Nov 21, 2018
dataset trival Nov 21, 2018
decoder trival Nov 21, 2018
discriminator trival Nov 21, 2018
encoder trival Nov 21, 2018
generator trival Nov 21, 2018
model trival Nov 21, 2018
netutils trival Nov 21, 2018
network trival Nov 21, 2018
test trival Nov 21, 2018
tester trival Jul 10, 2018
trainer trival Sep 26, 2018
validator trival Nov 21, 2018
.gitignore trival Sep 20, 2018
LICENSE Create LICENSE May 25, 2018
LabelCheckProj.sublime-workspace trival Nov 21, 2018 trival Sep 26, 2018
VAE-GAN.sublime-project trival Sep 17, 2018
VAE-GAN.sublime-workspace trival Nov 21, 2018 trival Jul 22, 2018 trival Aug 14, 2018 trival Sep 14, 2018


This repo implements many recently emerged generative models, such as GAN, VAE, DCGAN, WGAN, WGAN-GP, and some semi-supervised model such as AAE, SemiDGM, this code is just for learning the generative models and for fast developing algorithms.

There is some problem with my code, for WGAN, I found it very likely to cause model collpase(after 30k iters on Cifar10 dataset) and the generated sample quality goes worse. For semi-supervised model AAE, I achieved 96% accuracy on MNIST dataset with 10 labels per class which is below the paper claimed accuracy 98.1%. I will keep refining this repo to support more generative and semi-supervised algorithms.

Some code is outdated and may cause bug in running, please email to me:

this code is running with python3 and tensorflow1.9.0 on both Windows and Ubuntu

Currently Implemented Models









before training, you must specify the dataset location. the currently supported dataset is

  1. MNIST
  2. Cifar10
  3. Imagenet

open the py file under dataset folder, for each dataset I write some if control flow to find the dataset location. please add the dataset location to the control flow.

the config files is under the cfgs folder, I write config file in json format, the dataset, model and train method are all specified in this file. you can train it with If you want to run the models, just run the following command:

python(3) --config=<config file>  --gpu=<your gpu number>

the "assets dir" in config file is the folder where the result stores. the tensorboard log file is under the 'log' folder


AAE Semi-supervised Classification with 100 labels

python3 --config=aae/mnist_ssl 


You can’t perform that action at this time.