moono / stylegan-reproduced

StyleGAN, tensorflow, tf.estimator

stylegan-reproduced

• This repoitory is intended to understand official StyleGAN code from https://github.com/NVlabs/stylegan
• Most of the code is just copy of original source code
• In this Repository, I tried to remove tflib and some if/else statements from official code

Requirements

• Tensorflow >= 1.13
• tensorflow estimator and tf.contrib.distribute for multi-GPU
• Training time: approx. 5 days 8 hours with V100 * 4
  • note: training step on resolution 1024x1024 is much less than official one

Steps to reproduce

Dataset

• Download & setup from ffhq-dataset

To check the generator network is properly copied

1. export pretrained weight from official code
2. set variable names to current implementation
3. try to generate from official weights inference_from_official_weights.py
4. check the result

official output	current implementation

Training

• Trying to use native tensorflow training framework (tf.estimator)
• It recreates tf.estimator object every time it needs to change resolution or transition state is over to clear the optimizer states
• Run train.py
• training progress in tensorboard - 1024x1024

Generation results after training

• Generate using inference.py.