pip install git+https://github.com/moritztng/stylegan2-pytorch.gitDownloads Flickr Faces HQ Dataset into data and creates checkpoint
stylegan2 learn --path-ffhq dataLoads checkpoint and creates images.png
stylegan2 generate
import torch
from torchvision.utils import save_image
from stylegan2.learn import StyleGAN
# Train
stylegan = StyleGAN(path_ffhq='data')
stylegan('data')
# Generate
generator = stylegan.generator_average
latent = torch.randn(16, 512, device=stylegan.device)
images = generator(latent, weight_truncation=0.7)
save_image(images * 0.5 + 0.5, 'images.png')$ stylegan2 learn --help
Usage: stylegan2 learn [OPTIONS]
Options:
--data TEXT
--n-epochs INTEGER [default: 1]
--batch-size INTEGER [default: 16]
--n-batch-log INTEGER [default: 100]
--n-log-checkpoint INTEGER [default: 1]
--path-checkpoint-save TEXT [default: checkpoint.pth]
--lr FLOAT [default: 0.001]
--scale-lr-mapper FLOAT [default: 0.01]
--equal-lr / --no-equal-lr [default: True]
--betas <FLOAT FLOAT>... [default: 0.0, 0.99]
--scale-r1 FLOAT [default: 10.0]
--resolution INTEGER [default: 128]
--n-feature-maps TEXT [default: [512, 512, 512, 512, 512, 256]]
--n-conv-d INTEGER [default: 2]
--n-layers-mapper-g INTEGER [default: 8]
--n-dim-mapper-g INTEGER [default: 512]
--n-conv-blocks-g INTEGER [default: 2]
--n-dim-const-g INTEGER [default: 4]
--decay-generator-avg FLOAT [default: 0.999]
--path-checkpoint-init TEXT
--path-ffhq TEXT
--device TEXT
--help Show this message and exit.$ stylegan2 generate --help
Usage: stylegan2 generate [OPTIONS]
Options:
--path-checkpoint TEXT [default: checkpoint.pth]
--path-images TEXT [default: images.png]
--n-images INTEGER [default: 16]
--n-layers-mapper INTEGER [default: 8]
--n-dim-mapper INTEGER [default: 512]
--n-conv-blocks INTEGER [default: 2]
--n-feature-maps TEXT [default: [512, 512, 512, 512, 512, 256]]
--n-dim-const INTEGER [default: 4]
--equal-lr / --no-equal-lr [default: True]
--weight-truncation FLOAT [default: 0.7]
--device TEXT
--help Show this message and exit.During training plots and images are visualized in view.html
This checkpoint.pth was trained on the 128x128 FFHQ Dataset for 9 epochs