Generative Models

This repository contains implementations of various generative models, including DCGAN, GAN, VAE, CycleGAN, and WGAN. Some of the implementations, specifically the GAN model, utilize the 128x128 thumbnails from the FFHQ dataset for training.

Table of Contents

• Introduction
• Installation
• Usage
• Models
  • DCGAN
  • GAN
  • VAE
  • CycleGAN
  • WGAN
  • StyleGAN
• Configuration
• License
• Contributing

Introduction

Generative models are a subset of unsupervised learning that generate new sample/data that can be considered as part of the original data. This repository provides implementations of some of the popular generative models.

Installation

1. Clone the repository:

   git clone https://github.com/kaanberke/generative-models.git
   cd generative-models

2. Install the required packages:

   pip install -r requirements.txt

3. (Optional) If you wish to develop or contribute, you can also install the development dependencies:

   python setup.py develop

Usage

To run the main script:

python main.py

Models

CycleGAN

Cycle Generative Adversarial Networks (CycleGAN) is a method for training unsupervised image-to-image translation models without paired data. The models are trained in a way that they can convert images from one domain to another and vice versa. For the code and implementation details, see models/cyclegan.py.

DCGAN

Deep Convolutional Generative Adversarial Networks (DCGAN) is a type of GAN where the generator and discriminator are deep convnets. You can find the implementation in models/dcgan.py.

GAN

Generative Adversarial Networks (GAN) consists of two networks, a generator and a discriminator, that are trained together. The generator tries to produce data that comes from some probability distribution, while the discriminator tries to tell real from fake data. The implementation in this repository uses the 128x128 thumbnails from the FFHQ dataset for training purposes. Check out the implementation in models/gan.py.

VAE

Variational Autoencoders (VAE) are a kind of autoencoder that's trained to learn the probability distribution of its input data. For more details, refer to models/vae.py.

WGAN

Wasserstein Generative Adversarial Networks (WGAN) introduces a new way of training GANs to overcome issues like mode collapse. It replaces the traditional GAN loss with a Wasserstein distance, leading to more stable training. The implementation details and code can be found in models/wgan.py.

StyleGAN

Style-Based Generator Architecture for Generative Adversarial Networks (StyleGAN) is a novel generative model proposed by Nvidia researchers. It introduces a new architecture that can produce high-resolution, photorealistic images and gives control over the synthesis process through style inputs. The implementation can be found at models/stylegan.py. To train StyleGAN with the default configuration, run:

Configuration

The repository includes a config.yaml file which allows users to configure various parameters related to the models, training, and other settings.

License

This project is licensed under the MIT License. For more details, see the LICENSE file.

Contributing

We'd love to see your contributions! Just a friendly reminder: please stick to our coding standards and project guidelines. And, before you hit that "submit" button for a pull request, do give your changes a good test run. Thanks for being awesome! 🌟