How to Develop a CycleGAN from Scratch

In this article, we implement a CycleGAN with a residual-based generator.

Here, we’ll show you how to implement a residual-based generator and train the resulting CycleGAN on a medical dataset.

‘This is a series of articles about Image-to-Image Translation with CycleGAN’.

Introduction

In this series of articles, we’ll present a Mobile Image-to-Image Translation system based on a Cycle-Consistent Adversarial Networks (CycleGAN). We’ll build a CycleGAN that can perform unpaired image-to-image translation, as well as show you some entertaining yet academically deep examples. We’ll also discuss how such a trained network, built with TensorFlow and Keras, can be converted to TensorFlow Lite and used as an app on mobile devices.

We assume that you are familiar with the concepts of Deep Learning, as well as with Jupyter Notebooks and TensorFlow. You are welcome to download the project code and files.

In the previous article of this series, we trained and evaluated a CycleGAN that used a U-Net-based generator. In this article, we’ll implement a CycleGAN with a residual-based generator.

CycleGAN from Scratch

The original CycleGan was first built using a residual-based generator. Let’s implement a CycleGAN of this type from scratch. We’ll build the network and train it to reduce artifacts in fundus images using a dataset of fundi with and without artifacts.