A Perda Adversarial

Cada participante de um GAN aprende com uma perda que avalia seu desempenho. As duas perdas puxam em direções opostas, o que torna o treinamento adversarial.

O treinamento de um GAN é um jogo minimax: o discriminador tenta maximizar sua pontuação, enquanto o gerador tenta minimizá-la. Nenhum dos dois vence completamente.

Most GANs measure each prediction with binary cross-entropy, the standard loss for a real-versus-fake yes or no decision.

First the discriminator sees real images and should call them real. Its real loss pushes those outputs toward the label 1.

Next it judges generated images and should call them fake. The fake loss pushes those outputs toward the label 0.

When updating the discriminator you call detach on the fakes so gradients do not flow back into the generator on this step.

The generator wants its fakes labeled real, so its loss compares the discriminator's verdict on fakes against the label 1.

Telling the generator to maximize real-ness, instead of just minimizing fake-ness, gives stronger early gradients. This is the non-saturating loss.

Each training round does two updates: first the discriminator on its combined loss, then the generator on its own. Order and timing matter.

Because both networks change every step, each one chases a moving target. This is why GAN losses bounce around instead of falling smoothly.

In a GAN a falling loss does not mean better images. You judge quality by looking at samples, not by the loss number alone. 👀

You saw the min-max game: cross-entropy guides both players, fakes are detached for the judge, and samples, not loss, reveal real progress. 🎯