/GAN-Metrics

An empirical study on evaluation metrics of generative adversarial networks.

Primary LanguageJupyter Notebook

GAN Metrics

This repository provides the code for An empirical study on evaluation metrics of generative adversarial networks.

Requirement

  • Python 3.6.4
  • torch 0.4.0
  • torchvision 0.2.1
  • pot 0.4.0
  • tqdm 4.19.6
  • numpy, scipy, math

Usage

  • We create a demo for DCGAN training as well as computing all the metrics after each epoch.
    In the demo, final metrics scores of all epoches will be scored in <outf>/score_tr_ep.npy
  • If you want to compute metrics of your own images, you have to modify the codes of function compute_score_raw() in metric.py by yourself :)
python3 demo_dcgan.py \
--dataset cifar10 \
--cuda \
--dataroot <data_folder> \
--outf <output_folder> \
--sampleSize 2000

demo