TensorFlow Generative Adversarial Networks

Repository for TensorFlow implementation of Generative Adversarial Networks (GANs).

Background

NOTE: Tensorflow 2.0.0-alpha is released. Update will follow.

TensorFlow implementation of GAN and its variants can be easily found from GitHub. Despite the fact that most GANs share the minimax-game philosophy with some modification in its model (model refers to the multi-layer perceptron structure in this document. i.e. layers of D and G) or graph (graph refers to the operations applied to the model in this document. e.g. optimizers, losses and regularizers of the network), there has been no implementation that can separately switch between these modifications. Say that you want to try a network that has a model of the DCGAN, combined with a graph of the GEOGAN with hinge loss. It will take a substantial amount of effort even when not writing the whole script from scratch, since the implementation of DCGAN and GEOGAN may be coded in a different style, with many customized operations.

To address this issue, I built the GANs with

1. clear demarcation of the model and the graph structure
2. using only TensorFlow module functions without custom wrapping of the functions.

This implementation approach offers appealing advantages including

1. flexibility of separately applying various model and graph structures,
2. and ease of reading/modifying the script to make a new GAN yourself.

I also considered maximizing the compatibility with the TensorBoard, so that clean summary of the network can be visualized.

Requirements

Python3 with following packages

• tensorflow >= 1.12.0
• matplotlib >= 3.0.2 for saving images

Usage

Run the main script python3 main.py. You can check the arguments by calling python3 main.py -h, and change hyperparameters like python3 main.py -g 0 -e 100 -b 128 -n 100 -tM dcgan -lD 1e-7. It will run the DCGAN model combined with the vanilla GAN graph, with batch size, noise vector length, and discriminator learning rate set to 128, 100, and 1e-7 respectively, for 100 epochs using the GPU with ID 0.

Options

Datasets

• MNIST
• Cifar-10
• Cifar-100

Models

• GAN
• DCGAN
• CGAN
• ACGAN

Graphs

• GAN
• LSGAN
• WGAN
• WGAN-GP
• GEOGAN
• ACGAN

Regularizers

• Spectral Normalization

• Mode-seek

  NOTE: an issue is ongoing when applying spectral normalization with GPUs (tensorflow/tensorflow#24660 (comment))

Other options are being updated.

Implementation

• main.py

  Main function that runs the script. Hyperparameters can be passed as arguments.

• network.py

  Defines the class NetworkGAN, which is the full network that wraps the dataset, model, graph, and session.

• dataset.py

  Defines the class DatasetGAN, which can import datasets and return image, label, and noise.

• model.py

  Defines the class ModelGAN, which can build the model by creating GeneratorComponentGAN and DiscriminatorComponentGAN objects from the model_components module.

  • model_components.py

    Defines the class GeneratorComponentGAN and the class DiscriminatorComponentGAN, which holds multi-layer structure of the generator and the discriminator.

• graph.py

  Defines the class GraphGAN, which can build the graph by defining needed graph operations such as input nodes, losses, and optimizers upon the ModelGAN model object.

• session.py

  Defines the class SessionGAN, which can run the dynamic session of the built graph.

Reference

• GAN: Generative Adversarial Nets
• DCGAN: Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• LSGAN: Least Squares Generative Adversarial Networks
• WGAN: Wasserstein GAN
• WGAN-GP: Improved Training of Wasserstein GANs
• GEOGAN: Geometric GAN
• CGAN: Conditional Generative Adversarial Nets
• ACGAN: Conditional Image Synthesis with Auxiliary Classifier GANs
• SPECTRALNORM: Spectral Normalization for Generative Adversarial Networks
• MODESEEK: Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis