Early prediction for mode anomaly in generative adversarial network training: An empirical study

Abstract

Mode anomaly (MA for short) significantly blocks the application of generative adversarial networks (GANs). Although diverse metrics have been proposed to measure the MA, and a lot of efforts have been made to resolve the MA, none of them gives a quantitative definition for MA detection. Moreover, very few studies concentrate on the early-stage prediction of MA. In this paper, we make the first effort to this field with a systematic empirical study. To this end, we first give a fine-grained definition where the MA is categorized into three typical sub-patterns. Afterwards, traditional MA metrics are studied with extensive experiments on numbers of representative combinations of subjects (including 13 GANs and 3 datasets) to explore their sensitivity for the MA across different training steps. We find that in most of cases, the MA can be reasonably predicted in very early training stage through our sensitivity studies. Under the insight, we propose a novel prediction strategy using conception of “anomaly sign”. The evaluation results on diverse experimental subjects demonstrate the feasibility and high accuracy for the early prediction of MA. We also discuss the prediction efficiency, as well as analyze the prediction effectiveness from human perception.