Original pytorch implementation of preprint RPGAN: GANs Interpretability via Random Routing. (https://arxiv.org/abs/1912.10920).
RPGAN is an alternative GAN design based on random routes in generator, which can serve as a tool for generative models interpretability.
Go to ./lib.
Run train:
python train.py --out <out_directory> --model sn_resnet32 --data cifar10 --data_path <path_to_dataset>
FID computation:
python score/scoring.py --dir <train_out_directory>
To download pretrained models simply run
download_pretrained.sh
This script creates a directory ./pretrained and downloads models and configs for CIFAR-10, AnimeFaces and LSUN-Bedroom.
Once you have a RPGAN generator and .json with its parameters, run
python visualization.py --model_params <path_to_params.json> --model <path_to_model.pt> --out_dir <out_directory>
This script generates images charts with layers variations.
Code for FID computation is based on https://github.com/mseitzer/pytorch-fid
Code for Spectral Norm module is based on https://github.com/christiancosgrove/pytorch-spectral-normalization-gan