What does "model.trainable = False" mean in Keras?

Question

I want to freeze a pre-trained network in Keras. I found base.trainable = False in the documentation. But I didn't understand how it works. With len(model.trainable_weights) I found out that I have 30 trainable weights. How can that be? The network shows total trainable params: 16,812,353. After freezing I have 4 trainable weights. Maybe I don't understand the difference between params and weights. Unfortunately I am a beginner in Deep Learning. Maybe someone can help me.

Answer 1

A Keras Model is trainable by default - you have two means of freezing all the weights:

1. model.trainable = False before compiling the model
2. for layer in model.layers: layer.trainable = False - works before & after compiling

(1) must be done before compilation since Keras treats model.trainable as a boolean flag at compiling, and performs (2) under the hood. After doing either of the above, you should see:

print(model.trainable_weights)
# [] 

---

Regarding the docs, likely outdated - see linked source code above, up-to-date.