tensorflow - Customize trainable parameters of KERAS model

Question

I just want to set all the parameters to not trainable expect the ones that are connected to dense_5 layer which are it's own parameters and the parameters of dense_6 layer that are connected to dense_5. How can I set these last parameters to trainable ? By assigning the trainable to true for the dense_6 layer, all the parameters become trainable even if they are not connected to dense_5.

question from:https://stackoverflow.com/questions/66056589/customize-trainable-parameters-of-keras-model

Answer 1

You can not change trainable for a part of layer. Change the architecture: don't concat dense_5, add dense_7 after dense_5, concat outputs of dense_6 and dense_7. Then set trainable for dense_5 and Dense_7.