Inception Score is a measure of the individual and overall characteristics of a generative model.
Individual characteristics: The images generated by the same GAN should belong to one category, so we use the Inception model trained by imageNet to classify the images, so the images of the same category have less uncertainty and less entropy in a certain category.
Overall characteristics: The generated images should be diverse, even if they belong to the same category, their output vectors should still be different.
So we calculate an Inception Score for its output vector:
The source code address is: https://github.com/openai/improved-gan/blob/master/inception_score/model.py
There is also a mode score that measures model diversity:
Quote: