An empirical study on evaluation metrics of generative adversarial networks

TL;DR

This paper critically evaluates various sample-based metrics for assessing GANs, identifying their strengths and limitations, and proposes that kernel MMD and 1-NN tests are most effective when using appropriate feature spaces.

Contribution

It systematically investigates existing evaluation metrics for GANs, establishing criteria for meaningful scores and highlighting the effectiveness of kernel MMD and 1-NN tests.

Findings

Kernel MMD and 1-NN tests perform well with suitable feature spaces.

Evaluation metrics can reveal GANs' mode dropping, overfitting, and memorization.

GAN models exhibit diverse behaviors in learning target distributions.

Abstract

Evaluating generative adversarial networks (GANs) is inherently challenging. In this paper, we revisit several representative sample-based evaluation metrics for GANs, and address the problem of how to evaluate the evaluation metrics. We start with a few necessary conditions for metrics to produce meaningful scores, such as distinguishing real from generated samples, identifying mode dropping and mode collapsing, and detecting overfitting. With a series of carefully designed experiments, we comprehensively investigate existing sample-based metrics and identify their strengths and limitations in practical settings. Based on these results, we observe that kernel Maximum Mean Discrepancy (MMD) and the 1-Nearest-Neighbor (1-NN) two-sample test seem to satisfy most of the desirable properties, provided that the distances between samples are computed in a suitable feature space. Our experiments also unveil interesting properties about the behavior of several popular GAN models, such as whether they are memorizing training samples, and how far they are from learning the target distribution.