Repository for Diverse Rare Sample Generation with Pretrained GANs [Paper] [Arxiv]

The environment was builded on StyleGAN2-pytorch by rosinality. With CUDA12.1, the following command will download the corresponding docker image.

docker pull sbrblee/cuda12.1gcc6:1.0

python -u scripts/feature_extraction.py --data_path {image directory} # The default model option is VGG16.

This code generates the feature vectors from a given image directory in npz file format.

python -u scripts/nf_train.py --npz_path {path to real feature npz file}

This code trains the normalizing flow model (Glow model) from the features in a given npz file. We provide the Glow model architecture used in our paper in models folder as default. This code also saves the MinMax scaler for the feature vectors.

python -u scripts/divrare_optimization.py --zG_path {path to reference latent vectors npy file} --real_feature_path {path to real feature npz file} --nf_ckpt {path to checkpoint of normalizing flow model} --scaler_path {path to scaler} --dists_path {path to penalizing distances}

This code generates diverse rare samples for given reference latent vectors. The following options will be helpful to control the algorithm hyperparameters:

--n_sample number of rare samples to generate per reference

--rand_scale scale of noise to add to the initial latent vector for multi-start approach

--lambda1 coefficient of the similarity objective

--lambda2 coefficient of the diversity objective