stylegan2encoder icon indicating copy to clipboard operation
stylegan2encoder copied to clipboard

Tiled-mode for project_images.py and resumable training

Open dimitry12 opened this issue 5 years ago • 3 comments

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

dimitry12 avatar Jan 16 '20 18:01 dimitry12

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.

rolux avatar Jan 16 '20 18:01 rolux

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 avatar Sep 07 '20 21:09 woctezuma

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

woctezuma avatar Sep 07 '20 21:09 woctezuma