Should all intermediate tensor in custom layer be registered as “non-trainable-weight”?

Question

I\'m writing custom layers with Keras. But "non-trainable-weight" confused me. Let\'s say I have a simple layer as follows (It only do one thing, divi