Sumching / LDAGAN

Code for image generation of Latent Dirichlet Allocation in Generative Adversarial Network

Latent Dirichlet Allocation in Generative Adversarial Network

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

Usage

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.

Experiments

dataset	IS	FID
CIFAR10	8.77	10.4
CIFAR100	8.81	15.2
ImageNet(32x32)	9.70	18.5

./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.