Challenges in training GANs

GANs are notoriously difficult to train. We'll discuss some of the main challenges in training a GAN.

Uninformative loss and metrics

When training a CNN for classification or detection tasks, we can look at the shape of the loss plots to tell whether the network has converged or is overfitting and we'll know when to stop training. Then the metrics will correlate with the loss. For example, classification accuracy is normally the highest when the loss is the lowest. However, we can't do the same with GAN loss, as it doesn't have a minimum but fluctuates around some constant values after training for a while. We also could not correlate the generated image quality with the loss. A few metrics were invented to address this in the early days of GANs and one of them is the inception score.

A classification CNN known as inception is used to predict the confidence score of an image belonging to one of 1,000 categories in the ImageNet...