In the GAN example of the documentation, did we update D several times before updating G?

Hi,

In the following example of GAN for the deep learning toolbox ( https://www.mathworks.com/help/deeplearning/examples/train-generative-adversarial-network.html ), did we update Discriminator D several times before updating the generator G?

As per the GAN algorithm (say the original paper of Goodfellow 2014), we would update D k-times before updating G.

But I don't think the value of k is mentioned in the example. It looks like D and G are updated together in every loops.

The relevant codes:

        % Update the discriminator network parameters.
        [dlnetDiscriminator.Learnables,trailingAvgDiscriminator,trailingAvgSqDiscriminator] = ...
            adamupdate(dlnetDiscriminator.Learnables, gradientsDiscriminator, ...
            trailingAvgDiscriminator, trailingAvgSqDiscriminator, iteration, ...
            learnRateDiscriminator, gradientDecayFactor, squaredGradientDecayFactor);
        
        % Update the generator network parameters.
        [dlnetGenerator.Learnables,trailingAvgGenerator,trailingAvgSqGenerator] = ...
            adamupdate(dlnetGenerator.Learnables, gradientsGenerator, ...
            trailingAvgGenerator, trailingAvgSqGenerator, iteration, ...
            learnRateGenerator, gradientDecayFactor, squaredGradientDecayFactor);

Thanks!

Answer 1

[Sourav Bairagya]

GANs can be trained in many fashions. Here, discriminator, D is updated once before updating generator, G. Hence, first, discriminator, D is updated and then the generator, G is updated in every iteration.

[SC]

So, to follow the original training fashion (i.e. update (D k-times and G 1-time) n times), we need to add one more inner loop to achieve that?