This project is a reimplementation from official NV_Lab stylegan2(link) & this repo.

I only use the first 130K images(NG: 42610, OK: 87390) from _train_map.txt for training GAN. In data preprocessing stage, I resize images to 160x160 then center crop to 128x128 and normalize the pixel value into [-1, 1] along the spatial dimension.

Stylgan2(Condtional GAN architecture). I use embedding table to lookup label and concat it with latent vector in mapping network of StyleGAN2. In Discriminator, I use label information to only output prediction for that class.

The training behavior is quite like the official Stylegan2. I traing this model for 300K iteration with initial learning rate 0.002, batch size 16, Non-saturating loss with path length regularization for Generator, logistic loss and r1 regularization for Discriminator, and Style mixing regularization. For training configuration file, please refer to StyleGAN2_Pytorch/experiments/itri-noaug.yml

The best FIDs in this training are 87.5 and 48.7 for NG and OK, respectively. For FID in whole training, please refer to fid.txt