Quality metrics of variational autoencoders

Variational autoencoders (VAEs) are popular models for the generation of realistic images. Usually, the quality of generated images is estimated only by visual inspection, as the methods for the quantitative estimation are not fully developed. In this work, we present and test the methods that allow the possibility to evaluate the quality and the diversity of the images generated by a VAE objectively using an auxiliary classifier. The corresponding quality metrics are calculated for several implementations of VAE. The proposed metrics are especially valuable for the problems of preventing catastrophic forgetting during sequential learning. The experiments are performed on publicly available datasets.