GAN Generator Loss

Problem

Compute the non-saturating generator loss: -mean(log(D(fake))), where D(fake) are the discriminator probabilities assigned to generated samples.

The signature is gan_generator_loss(fake_probs).

Hint

This is the -log D variant that gives useful gradients when the generator is losing.

case 1

in: [[0.5,0.5]]

expected: 0.6931471806

case 2

in: [[0.9]]

expected: 0.1053605157

case 3

in: [[0.1]]

expected: 2.302585093

case 4

in: [[1]]

expected: 0

Starter codePython

import math
def gan_generator_loss(fake_probs):
    # Your code here
    pass

4 test cases · run your code against them in the editor