Retraining Encoder/Generator in AAE about matlab-gan HOT 5 CLOSED

@NicoMigenda
From AAE.m script, I did update encoder every iteration.

The updates are listed in the while loop.
From lines 56-76,

        [GradEn,GradDe,GradDis] = ...
                dlfeval(@modelGradients,XBatch,...
                paramsEn,paramsDe,paramsDis,settings);

        % Update Discriminator network parameters
        [paramsDis,avgG.Dis,avgGS.Dis] = ...
            adamupdate(paramsDis, GradDis, ...
            avgG.Dis, avgGS.Dis, global_iter, ...
            settings.lrD, settings.beta1, settings.beta2);

        % Update Encoder network parameters
        [paramsEn,avgG.En,avgGS.En] = ...
            adamupdate(paramsEn, GradEn, ...
            avgG.En, avgGS.En, global_iter, ...
            settings.lrG, settings.beta1, settings.beta2);
        
        % Update Decoder network parameters
        [paramsDe,avgG.De,avgGS.De] = ...
            adamupdate(paramsDe, GradDe, ...
            avgG.De, avgGS.De, global_iter, ...
            settings.lrG, settings.beta1, settings.beta2);

Details are described in the modelGraidents from lines 106-125. The generator gradient (encoder+decoder structure) is updated together at line 123.
[GradEn,GradDe] = dlgradient(g_loss,paramsEn,paramsDe,'RetainData',true);

Edit:
Maybe I get what you mean. You mean the gradient of the encoder should not only depend on reconstruction loss, but also binary cross entropy (for True False output). I did include this line 120,

g_loss = .999*mean(mean(.5*(x_-x).^2,1))-.001*mean(log(d_output_fake+eps));

Hope this replies your question.

from matlab-gan.

from matlab-gan.

@NicoMigenda

Oh I see. What you mention is common for all GANs. Usually, we will not train the generator twice, but still retain that concept in mind.

For example in Keras,

# Train the discriminator 
d_loss_real = self.discriminator.train_on_batch(latent_real, valid) 
d_loss_fake = self.discriminator.train_on_batch(latent_fake, fake) 
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake) 
# --------------------- 
# Train Generator 
# --------------------- 
# Train the generator 
g_loss = self.adversarial_autoencoder.train_on_batch(imgs, [imgs, valid])

They also only train once. But they set discriminator trainable to false first before training generator.

In Matlab, they do it by setting generator dlgradient retaindata to true.

from matlab-gan.

Thank you very much for that explanation. Is there further literature about that?

from matlab-gan.

@NicoMigenda
I would suggest to read how others implement vae and gan with keras and pytorch, and compare their implementations with the matlab version.

Keras AAE:
https://github.com/eriklindernoren/Keras-GAN/blob/master/aae/aae.py
Pay attention on discriminator.trainable and train_on_batch.

PyTorch AAE:
https://github.com/eriklindernoren/PyTorch-GAN/blob/master/implementations/aae/aae.py
Pay attention on steps between zero_grad, backward and step.

Matlab dlgradient:
https://uk.mathworks.com/help/deeplearning/ref/dlarray.dlgradient.html

Also, compare the documentations in matlab deep learning toolbox, keras and pytorch.

from matlab-gan.