Tiled-mode for project_images.py and resumable training

rolux / stylegan2encoder

StyleGAN2 - Official TensorFlow Implementation

http://arxiv.org/abs/1912.04958

Other

379 stars 71 forks source link

Tiled-mode for project_images.py and resumable training #9

Open dimitry12 opened 4 years ago

dimitry12 commented 4 years ago

This PR contains two improvements:

new command-line argument for project_images.py: --no-tiled which allows to switch to factory-default projection into 1x512 dlatents-space (useful for evaluation of "semantic" quality of the model);
a tweak in training loop to resume from the latest network-snapshot if one is found in the --result-dir

rolux commented 4 years ago

Yes this both makes sense. I'll try and add it to the next release. There are going to be a few major changes, mostly for video encoding.

woctezuma commented 4 years ago

Would you mind explaining the idea behind tiled mode? I found your PR while looking for information about this exact change. I suspect tiled mode leads to better projection results.

That could explain why I get different projection results with Nvidia's original project and this one.

woctezuma commented 4 years ago

Alright, I discovered this thread #2 and will read about it tomorrow. I think it might contain the answer.