I'm trying to understand the output of a GAN training process. I can always inspect the visual output (the generated images) to see how the GAN's evolving, but how can I evaluate training based on the losses and accuracies of the generator and discriminator?

Unless I'm mistaken, the goal is to see the GAN's accuracy go to 0.5, meaning it no longer can differentiate between real and fake data (images). Is that right? Can I learn more from the evolution of the losses?


1 Answer 1


This is in general a difficult problem.

Monitoring the discriminator's accuracy is insufficient, since it goes to 0.5 when mode collapse happens.

Some GAN variants -- for example Wasserstein GAN, have theoretical grounding which allows you to track the convergence by the loss function, however this is not true for all GANs in general, and is not true for the original GAN formulation.

Fréchet Inception Distance (FID) is the most widely accepted method of evaluating GANs (although flawed in many ways). You can try to track the convergence of your GAN by computing the FID and plotting the trend over time.

  • $\begingroup$ Sorry for the late response, but that was very informative! Thank you! $\endgroup$
    – Inkidu616
    Sep 30, 2019 at 9:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.