README.md

Data preparation for CIFAR-10 experiments

The folder CFIAR10/prepare_data contains the code to prepare data for the CIFAR-10 experiments.

Pre-trained image classifier

First, you need a pre-trained image classifier to annotate generated images.

The image classifier that we have used is publicly available here. You can simply download the densenet-bc-L190-k40 model and unzip it to the folder CIFAR10/pretrained/classifiers/cifar10, where our code will load the checkpoint of the image classifier.

Data generation

Generate 60k images and latent variables of StyleGAN2 (including w and z):

bash scripts/run_gen_batch.sh

Use the pre-trained image classifier to annotate the generated images:

bash scripts/run_cifar10_labeling.sh

The resulting pairs of latent variables (w and z) and labels will be used to train latent classifiers.

FID reference statistics

After Data generation, you can calculate the FID statistics for real CIFAR-10 images:

bash scripts/run_calc_inception.sh

The resulting inception_cifar10.pkl will be used for computing FID scores.