Get the latest tech news
The Path to StyleGan2 – Implementing the Progressive Growing GAN
- Implementing the Progressive Growing GAN Author Yusuf Mohammad This post is the first in a three part series, which will see us implement the StyleGAN2 architecure at its finale. The Progressive Growing GAN architecture is the foundation of the StyleGAN models, so we start our journey here.
This is the paper we will be following in this blog post, you can find it at https://arxiv.org/pdf/1710.10196 By progressively increasing the resolution throughout the model’s layers, the PGGAN can generate high quality images which look good. We need to understand the growing scheme and the actual structure of the model itself, minibatch standard deviation (MB StdDev), equalised learning rate (Eq LR), pixel normalisation (PN) and the WGAN-GP loss function. The saturation occurs due to the loss functions trying to solve the Jenson-Shannon divergence which is not differentiable everywhere, simply put at some points the gradients provide no information for updating the models.
Or read this on Hacker News
