Home

Awesome

Paper :

Progressive Growing of GANs for Improved Quality, Stability, and Variation

You would find some helpful comments in some key functions, which may help to find detail instructions from the paper.

ENV :

How to use :

Gen Image dataset: Download the CelebA first, then run "gen_classified_images" function in train.py file.

if __name__ == "__main__":
    gen_classified_images(r"E:\workspace\datasets\CelebA\Img\img_align_celeba", centre_crop=True, save_to_local=True)

This function just resizing the original image, if you would like to test the CelebA-HQ dataset, please follow tkarras' instructions.

Training: Open the train.py file again, modify and run the script：

if __name__ == "__main__":
    p = PGGAN(resolution=1024,            # Final Resolution.
              latent_size=512,            # Dimensionality of the latent vectors.
              criterion_type="GAN"        # "GAN" or "WGAN-GP"
              )
    p.train(r"E:\workspace\datasets\CelebA\Img\img_align_celeba_classified")

Reference and Acknowledgement