How to change read only properties of Matlab Deep learning layers? (using approximation of sigmoid for LSTM Layer)

I want to evaluate the accuracy if I change the activation functions of a BiLSTM-Network to different approximations.

Due this I constructed a network in Matlab.

For this, I want to change the activation functions of the BiLSTM-module of the network

For a minimal example, lets assume a network like this

numHiddenUnits = 100;

numClasses = 9;

layers = [ ...

sequenceInputLayer(inputSize)

bilstmLayer(numHiddenUnits)

fullyConnectedLayer(numClasses)

softmaxLayer

classificationLayer]

I want now to set for StateActivationFunction of the BiLSTM an approximation of the tanh function which I created by my own

if I open the struct and set another parameter it says me "its ready-only"

Has anybody an idea to avoid this?

[Kevin]

I know there is the possibility to create a custom layer de.mathworks.com/help/deeplearning/ug/… but its time intensive

Answer 1

[Damien T]

The LSTM layer in Matlab only supports tanh and softsign activations:

https://ch.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.lstmlayer.html

Moreover, you can only set this property when creating the layer. So you would have to create a new one, and replace the original layer with the new one in your layergraph or dlnetwork.

https://ch.mathworks.com/help/deeplearning/ref/nnet.cnn.layergraph.replacelayer.html