mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity
5ecf7da881
DL-Art-School/recipes/glean
History
James Betker 3fd627fc62 Mods to support image classification & filtering
2020-12-26 13:49:27 -07:00
..
README.md GLEAN docs 2020-12-18 16:21:01 -07:00
train_ffhq_glean.yml Mods to support image classification & filtering 2020-12-26 13:49:27 -07:00

README.md

GLEAN

DLAS contains an attempt at implementing GLEAN, which performs image super-resolution guided by pretrained StyleGAN networks. Since this paper is currently closed-source, it was implemented entirely on what information I could glean from the paper.

Training

GLEAN requires a pre-trained StyleGAN network to operate. DLAS currently only has support for StyleGAN2 models, so you will need to use one of those. The pre-eminent StyleGAN 2 model is the one trained on FFHQ faces, so I will use that in this training example.

  1. Download the ffhq model from nVidias Drive. This repo currently only supports the "-f.pkl" files without further modifications, so choose one of those.

  2. Download and extract the FFHQ dataset.

  3. Convert the TF model to a Pytorch one supported by DLAS:

    python scripts/stylegan2/convert_weights_rosinality.py stylegan2-ffhq-config-f.pkl

  4. The above conversion script outputs a *.pth file as well as JPG preview of model outputs. Check the JPG to ensure the StyleGAN is performing as expected. If so, copy the *.pth file to your experiments/ directory within DLAS.

  5. Edit the provided trainer configuration. Find comments starting with '<--' and make changes as indicated.

  6. Train the model:

    python train.py -opt train_ffhq_glean.yml