Quality Assessment Method for GAN Based on Modified Metrics Inception Score and Fréchet Inception Distance

The article examines the problem of quality assessment for generative adversarial networks (GANs). There is no unified and universal metric to compare and evaluate GAN. Well-known approaches for the GAN quality assessment are focused on images generating neural networks. This paper considers the problem of the quality determination of arbitrary GAN operating with various data sets. For problem solution, a quality assessment method of arbitrary GAN is proposed, which differs by the modification of the calculation formulas Inception Score and Frechet Inception Distance. The included changes allow the use of these metrics to assess and compare arbitrary GANs. The developed method was tested during experiments on the objects generation from marked (MNIST) and unmarked (Human Activity Recognition Using Smartphones and Epileptic Seizure Recognition) datasets. The obtained results confirm the possibility of applying the modified metrics Inception Score and Frechet Inception Distance to assess the quality of arbitrary GANs.