Code for the image generation experiments in Latent Dirichlet Allocation in Generative Adversarial Network.

We experimented on 5 different datasets:

CIFAR10, CIFAR100, ImageNet(size 32x32), CelebA and CelebAHQ.

To train a model, use

python train.py --yaml ./config/dataset_name.yml 

To generate samples for evaluation, use

python test.py --yaml ./config/dataset_name.yml --checkpoint ./output/checkpoint/500000_G.pth --output_name outputname 

It will return a ".npy" file contains 50,000 samples by default.

./utils/InceptionScore_and_FID.py

This file contains the implementation of functions to calculate the Inception Score and the FID. It compares the ".npy" file mentioned above with pre-calculated statistic:

python ./utils/InceptionScore_and_FID.py --input npy_filename --stats pre_calculated_stats 

Precalculated statistics for datasets can be found here.