CODEWITHSUNDEEP
pythonkeras
nairouz mrabah
nairouz mrabah

Reputation: 1217

Keras "trainable" scope

I have two Keras networks. Let's say for the sake of explanation that my model is similar to GAN. So, we have a Discriminator(D) and a Generator(G). Obviously, in order to train the (G), the (D) layers should be frozen. If, I freeze them using D.trainable = False then should I invert this parameter when to train the (D)?

What is the scope of model.trainable in Keras? I have seen codes that only change this argument status once: https://github.com/nairouz/Keras-GAN/blob/master/gan/gan.py

How is it possible to do so? Any explanation?

Upvotes: 1

Views: 178

Answers (1)

rvinas
rvinas

Reputation: 11895

From How can I "freeze" Keras layers?:

Additionally, you can set the trainable property of a layer to True or False after instantiation. For this to take effect, you will need to call compile() on your model after modifying the trainable property.

And the same applies to models. This means that when you set D.trainable = False, this doesn't take effect until you compile D (or any other model leveraging D), so it doesn't affect models that you have previously compiled.

Upvotes: 1

Related Questions