GAN: The Non-saturating Generator Loss
Intermediate
The problem, from first principles
The original generator loss, log(1 - D(G(z))), saturates when the discriminator is confident, leaving the generator with almost no gradient. The paper's non-saturating variant maximizes log D(G(z)) instead, which gives a strong signal early in training.
From theory to code
Implement gan_generator_loss(d_fake), returning -mean(log D(G(z))).
Constraints
- Returns a Python float.
Hints
Hint 1
Take the log of the discriminator's output on fake samples, average, and negate.
Keywords
research-papersunsupervisedgangenerative-models
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.