This model implements the CycleGAN conditional GAN architecture introduced by Zhu et al. The Horse to Zebra gif is reproduced from the authors' project. The code is a port from the authors' implementation which can be found here.
Architectures have been implemented for the following image dimensions: (256, 256).
This model references dataset attributes self.data.X and self.data.Y as the two domains: compatible datasets will process images and store them in these attributes. See the Horse to Zebra dataset for an example template.
To train with default parameters:
mantra train cyclegan --dataset my_images_data --image-dim 256 256
To import this model to your project, run:
mantra import RJT1990/models/cyclegan