tensorflow-generative-model-collections

Tensorflow implementation of various GANs and VAEs.

Pytorch version

Pytorch Version is now availabel at https://github.com/znxlwm/pytorch-generative-model-collections

Generative Adversarial Networks (GANs)

Lists

Name	Paper Link	Value Function
GAN	Arxiv
LSGAN	Arxiv
WGAN	Arxiv
WGAN-GP	Arxiv
DRAGAN	Arxiv
CGAN	Arxiv
infoGAN	Arxiv
ACGAN	Arxiv
EBGAN	Arxiv
BEGAN	Arxiv

Variants of GAN structure

Results for mnist

Network architecture of generator and discriminator is the exaclty sames as in infoGAN paper.
For fair comparison of core ideas in all gan variants, all implementations for network architecture are kept same except EBGAN and BEGAN. Small modification is made for EBGAN/BEGAN, since those adopt auto-encoder strucutre for discriminator. But I tried to keep the capacity of discirminator.

The following results can be reproduced with command:

python main.py --dataset mnist --gan_type <TYPE> --epoch 25 --batch_size 64

Random generation

All results are randomly sampled.

Name	Epoch 2	Epoch 10	Epoch 24
GAN
LSGAN
WGAN
WGAN-GP
DRAGAN
EBGAN
BEGAN

Conditional generation

Each row has the same noise vector and each column has the same label condition.

Name	Epoch 1	Epoch 10	Epoch 25
CGAN
ACGAN
infoGAN

InfoGAN : Manipulating two continous codes

Results for fashion-mnist

Comments on network architecture in mnist are also applied to here.
Fashion-mnist is a recently proposed dataset consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. (T-shirt/top, Trouser, Pullover, Dress, Coat, Sandal, Shirt, Sneaker, Bag, Ankle boot)

The following results can be reproduced with command:

python main.py --dataset fashion-mnist --gan_type <TYPE> --epoch 40 --batch_size 64

Random generation

All results are randomly sampled.

Name	Epoch 1	Epoch 20	Epoch 40
GAN
LSGAN
WGAN
WGAN-GP
DRAGAN
EBGAN
BEGAN

Conditional generation

Each row has the same noise vector and each column has the same label condition.

Name	Epoch 1	Epoch 20	Epoch 40
CGAN
ACGAN
infoGAN

Without hyper-parameter tuning from mnist-version, ACGAN/infoGAN does not work well as compared iwth CGAN.
ACGAN tends to fall into mode-collapse.
infoGAN tends to ignore noise-vector. It results in that various style within the same class can not be represented.

InfoGAN : Manipulating two continous codes

Some results for celebA

(to be added)

Variational Auto-Encoders (VAEs)

Lists

Name	Paper Link	Loss Function
VAE	Arxiv
CVAE	Arxiv
DVAE	Arxiv	(to be added)
AAE	Arxiv	(to be added)

Variants of VAE structure

Results for mnist

Network architecture of decoder(generator) and encoder(discriminator) is the exaclty sames as in infoGAN paper. The number of output nodes in encoder is different. (2x z_dim for VAE, 1 for GAN)

The following results can be reproduced with command:

python main.py --dataset mnist --gan_type <TYPE> --epoch 25 --batch_size 64

Random generation

All results are randomly sampled.

Name	Epoch 1	Epoch 10	Epoch 25
VAE
GAN

Results of GAN is also given to compare images generated from VAE and GAN. The main difference (VAE generates smooth and blurry images, otherwise GAN generates sharp and artifact images) is cleary observed from the results.

Conditional generation

Each row has the same noise vector and each column has the same label condition.

Name	Epoch 1	Epoch 10	Epoch 25
CVAE
CGAN

Results of CGAN is also given to compare images generated from CVAE and CGAN.

Results for fashion-mnist

Comments on network architecture in mnist are also applied to here.

The following results can be reproduced with command:

python main.py --dataset fashion-mnist --gan_type <TYPE> --epoch 40 --batch_size 64

Random generation

All results are randomly sampled.

Name	Epoch 1	Epoch 20	Epoch 40
VAE
GAN

Results of GAN is also given to compare images generated from VAE and GAN.

Conditional generation

Each row has the same noise vector and each column has the same label condition.

Name	Epoch 1	Epoch 20	Epoch 40
CVAE
CGAN

Results of CGAN is also given to compare images generated from CVAE and CGAN.

Results for celebA

(to be added)

Folder structure

The following shows basic folder structure.

├── main.py # gateway
├── data
│   ├── mnist # mnist data (not included in this repo)
│   |   ├── t10k-images-idx3-ubyte.gz
│   |   ├── t10k-labels-idx1-ubyte.gz
│   |   ├── train-images-idx3-ubyte.gz
│   |   └── train-labels-idx1-ubyte.gz
│   └── fashion-mnist # fashion-mnist data (not included in this repo)
│       ├── t10k-images-idx3-ubyte.gz
│       ├── t10k-labels-idx1-ubyte.gz
│       ├── train-images-idx3-ubyte.gz
│       └── train-labels-idx1-ubyte.gz
├── GAN.py # vainilla GAN
├── ops.py # some operations on layer
├── utils.py # utils
├── logs # log files for tensorboard to be saved here
└── checkpoint # model files to be saved here

Acknowledgements

This implementation has been based on this repository and tested with Tensorflow over ver1.0 on Windows 10 and Ubuntu 14.04.