aae.py discriminator.trainable about keras-gan HOT 5 CLOSED

You set the discriminator to be false when training the generator is standard since you only want to update generator weights during this phase. But when you want to train the discriminator in order to provide a strong signal to generator you have to train it so it doesnt get easlily fooled by fake data.

There are TWO distinct phases of training for two distinct model types (G and D). One where discriminator gets its weights updated and another when the generator weights gets updated.

from keras-gan.

anyway cannot understand.

So you mean discriminator will be trained on
self.discriminator.train_on_batch
regardless of
self.discriminator.trainable = False
??

from keras-gan.

I understand using non trainable discriminator in
validity = self.discriminator(encoded_repr)
in
self.adversarial_autoencoder = Model(img, [reconstructed_img, validity])

but why
self.discriminator.train_on_batch

from keras-gan.

hm
or discriminator actually trainable if compiled

self.discriminator = self.build_discriminator()
self.discriminator.compile

before
self.discriminator.trainable = False
??

from keras-gan.

then all fine

from keras-gan.