How can I make kernel trainable= False and recurrent_kernel trainable = False using standard Keras LSTM

Question

I understand that using Trainable = False, I can freeze all the weights of a specific layer. But I want to work on the components (kernel, recurrent_kernel and bias) of each layer as well. I want to freeze the kernel and recurrent_kernel of a specific layer to be Trainable = False and Trainable = True. How can I do it?

I tried the following code but got an error. Can anyone suggest to me how can I make kernel and recurrent_kernel Trainable = False using standard Keras?

#transfer model layer

lstm_layer = modelTL.layers[0]

#kernel non-trainable

lstm_layer.cell.kernel.trainable = False

#Error : lstm_layer.cell.kernel.trainable = False 

AttributeError: can't set attribute

Answer 1

After building the layer, you can separate the different variables and set them to trainable or not.

import tensorflow as tf

layer = tf.keras.layers.LSTM(8)

layer.build(sequences_shape=(2, 4))

kernel, recurrent_kernel, bias, seed = layer.variables

kernel.trainable = False
recurrent_kernel.trainable = False