GAN models (including 3D controllable models) implemented with Pytorch Lightning and Hydra configuration. This is an unofficial project and work in progress. Model correctness is not guaranteed.
Usage examples:
python run_network.py +expt=dc_gan dataset=celeb_a
python run_network.py +expt=wgan dataset=mnist calc_fid=False val.use_fid=False figure_details.fid_callback=False (Don't use FID for MNIST)
python run_network.py +expt=wgan_gp dataset=celeb_a
python run_network.py +expt=gan_stability_r1 dataset=celeb_a
python run_network.py +expt=hologan dataset=celeb_a
By default, the FID score will be monitored on the validation set during the validation step. Model checkpoints are saved when the best FID score is attained. Note that FID is not valid for MNIST dataset training due to single channel output. Generator samples and latent space interpolations are saved to the output directory. Varying view outputs saved for 3D controllable models.
cp conf/filepaths/example.yaml conf/filepaths/local.yaml
Edit conf/filepaths/local.yaml with dataset locations on your system
Conda environment yaml file coming soon.
Install pytorch and pytorch lightning.
Install required hydra version using:
pip install hydra-core==1.1.0dev3