Cycle Consistent Generative Adversarial Networks

Tensorflow implementation of (https://arxiv.org/pdf/1703.10593.pdf).

Overview

Cycle Consistent GANs are an adaptation of Generative Adversarial Networks, in which the resulting model has the capability of performing domain adaptation between two datasets of varying domains. Again- unpaired! Images between datasets don't need to be directly matched, as the additional cycle consistency term added within the CycleGAN model allows for additional stability within training - of which pressures output domains to be consistent.

Sample mappings shown above.

Prerequisites

Dataset available at https://github.com/junyanz/CycleGAN (horse2zebra preferred, else change image dimensions). Alter path names in main for local directory for proper usage.

Packages Required in Environment:

• Tensorflow
• CV2
• Numpy
• Matplotlib

GPU training is preferred.

Execution

To execute the program, use the following command whilst in terminal:

python main.py