Skip to content

Latest commit

 

History

History
108 lines (82 loc) · 1.83 KB

README.md

File metadata and controls

108 lines (82 loc) · 1.83 KB

ISF-GAN with StyleGAN v1

Configuration

  • Python: 3.6.8
  • Pytorch: 1.5.1+cu101
  • torchvision: 0.6.1+cu101
  • tensorflow-gpu: 1.14.0
  • timm: 0.4.12
  • ffmpeg
  • munch

Download pretrained models:

Download pretrained models from here, and copy them into the foloder models/pretrain:

./models/pretrain
  |__karras2018iclr-celebahq-1024x1024.pkl
  |__karras2019stylegan-celebahq-1024x1024.pkl
  |__karras2019stylegan-ffhq-1024x1024.pkl
  |__pggan_celebahq.pth
  |__stylegan_celebahq.pth
  |__stylegan_ffhq.pth

Preparing dataset

  • Generate data through stylegan or pggan:
# stylegan
$ python3 generate_data.py -m stylegan_ffhq -o data/stylegan_ffhq -n 80000

Similarly, collect 10000 testing images.

  • Download pretrained classifiers:
$ python3 preparing/download.py
  • Collect attributes of generated images:
$ python3 preparing/collect_attributes.py
$ python3 preparing/merge.py

Finally, we can organize the dataset in this format:

../datasets/stylegan-ffhq
  |__train
  |   |__xxx.jpg
  |   |__log.txt
  |   |__w.npy
  |   |__wp.npy
  |   |__z.npy
  |
  |__test
  |   |__xxx.jpg
  |   |__log.txt
  |   |__w.npy
  |   |__wp.npy
  |   |__z.npy
  |
  |__list_attr_ffhq-test.txt
  |__list_attr_ffhq-train.txt

Training

$ sh ./scripts/train.sh

Testing

  • Image Editing
$ python3 isfgan_edit.py \
  --checkpoint_dir /path/to/checkpoint_dir \
  --use_post 1 \
  --save_dir /path/to/save_dir
  • Image interpolation
$ python3 isfgan_interp.py \
  --checkpoint_dir /path/to/checkpoint_dir \
  --use_post 1 \
  --save_dir /path/to/save_dir
  • Image sampling
$ python3 isfgan_sample.py \
  --checkpoint_dir /path/to/checkpoint_dir \
  --use_post 1 \
  --save_dir /path/to/save_dir