Generative Adversarial Network

The problem, from first principles

A Generative Adversarial Network (GAN) consists of a generator G that produces fake samples and a discriminator D that distinguishes fake from real. They are trained in an adversarial game where D tries to maximize accuracy and G tries to fool D.

In practice, use binary cross-entropy loss for both networks.

From theory to code

Implement:

gan_loss(d_real_logits, d_fake_logits, is_discriminator=True)

Computes the loss for discriminator or generator training.

Constraints

• d_real_logits shape: (N,)
• d_fake_logits shape: (N,)
• is_discriminator: bool flag for which loss to compute.
• Return scalar loss.