Deep Learning By Example

Model losses

Now it's time to define the model losses. First off, the discriminator loss will be divided into two parts:

• One which will represent the GAN problem, which is the unsupervised loss
• The second one will compute the individual actual class probabilities, which is the supervised loss

For the discriminator's unsupervised loss, it has to discriminate between actual training images and the generated images by the generator.

As for a regular GAN, half of the time, the discriminator will get unlabeled images from the training set as an input and the other half, fake, unlabeled images from the generator.

For the second part of the discriminator loss, which is the supervised loss, we need to build upon the logits from the discriminator. ...