Generative Adversarial Network Models

A collection of generative adversarial network models, e.g. GAN, FGAN, SoftmaxGAN, LSGAN in Tensorflow.

How to use?

• Command 1: python train.py gan_name train
• Command 2: python train.py gan_name generate

Note: Generated samples will be stored in images/{gan_model}/ directory during training.

Standard GANs

Description: The Generator is similar to a Decoder whereas the Discriminator is in the form of an Encoder.

• GAN: Generative Adversarial Networks
  • Date: 2014-06
  • Paper: Generative Adversarial Networks
  • Loss functions (Minimization):
  • 
• DCGAN: Deep Convolutional GAN
  • Date: 2016-01
  • Paper: Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
  • Loss functions (Minimization):
  • 
• FGAN: f-GAN
  • Date: 2016-06
  • Paper: f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization
  • Loss functions (Minimization):
  • 
• SoftmaxGAN: Softmax GAN
  • Date: 2017-04
  • Paper: Softmax GAN
  • Loss functions (Minimization):
  • 
• LSGAN: Least Squares GAN
  • Date: 2017-04
  • Paper: Least Squares Generative Adversarial Networks
  • Loss functions (Minimization):
  • 
• WGAN: Wasserstein GAN
  • Date: 2017-12
  • Paper: Wasserstein GAN
  • Loss functions (Minimization):
  • 
• WGAN_GP: Wasserstein GAN with Gradient Penalty
  • Date: 2017-12
  • Paper: Improved Training of Wasserstein GANs
  • Loss functions (Minimization):
  • 
• DRAGAN: Deep Regret Analytic GAN
  • Date: 2017-12
  • Paper: On Convergence and Stability of GANs
  • Loss functions (Minimization):
  • 
• BGAN: Boundary-Seeking GAN
  • Date: 2018-02
  • Paper: Boundary-Seeking Generative Adversarial Networks
  • Loss functions (Minimization):
  • 

MNIST Results

The following results can be reproduced with the command:

python train.py gan_name train

Standard GANs Results

Name	Epoch 1	Epoch 2	Epoch 3
GAN
DCGAN
FGAN
SoftmaxGAN
LSGAN
DRAGAN
WGAN
WGAN_GP
BGAN

Dependencies

1. Install miniconda https://docs.conda.io/en/latest/miniconda.html
2. Create an environment conda create --name autoencoder
3. Activate the environment source activate autoencoder
4. Install [Tensorflow] conda install -c conda-forge tensorflow
5. Install [Opencv] conda install -c conda-forge opencv
6. Install [sklearn] conda install -c anaconda scikit-learn
7. Install [matplotlib] conda install -c conda-forge matplotlib

Datasets

If you wanna try new dataset, please make sure you make it in the following way:

• Dataset_main_directory
  • train_data
    • category_1: (image1, image2, ...)
    • category_2: (image1, image2, ...)
    • ...
  • test_data
    • category_1: (image1, image2, ...)
    • category_2: (image1, image2, ...)
    • ...

The loader.py file will automatically upload all images and their labels (category_i folders)

Acknowledgements

This implementation has been based on the work of the great following repositories:

• https://github.com/hwalsuklee/tensorflow-generative-model-collections
• https://github.com/wiseodd/generative-models
• https://github.com/eriklindernoren/Keras-GAN