Tensorflow GANs Architectures Implementation

Brief Description

GANs have proven to be very powerful generative models. So, here's a well-structured Tensorflow project containing implementations of some GANs architectures.

Utilized Frameworks

• Tensorflow 1.13.1

Repository Strucuture

1) base folder:

• contains abstract classes for both model and trainer.

2) configs folder:

• contains json files for different model configurations.

3) data folder:

• for the training data to be added.

4) data_loader folder:

• contains data generator class for data loading and preprocessing.

5) models folder:

• contains different model implementations.

6) trainers folder:

• contains trainers for models.

7) utils folder:

• contains logger for Tensorboard summary, argument parser, configuration processing and directory creation.

Implemented Architectures

• DCGAN (Deep Convolutional Generative Adverserial Networks): https://arxiv.org/abs/1511.06434
• CycleGAN (Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks): https://arxiv.org/abs/1703.10593

Usage

• Put your training images in data folder.
• Edit the configuration JSON in configs folder (optional).
• Run the main file providing config and model arguments:

    python main.py -c <config_path> -m <model_name>

• Have a nice day!

To-Do List

1. Implement more GANs architectures.
2. Add Tensorflow 2.0 compatibility.
3. Add distributed training to the trainer process.
4. Improve the current training process and fix some issues.

Acknowledgment

• https://github.com/MrGemy95/Tensorflow-Project-Template.git