How to use setwb?

Say I have an already fully configured Neural Network net. I am trying to write the algorithm for the cascade k-fold cross validation of this network. For the first fold I initialize the weights randomly (using initnw). I would like all following folds to get the weights that resulted from the training of the previous fold using initwb and setwb. The relevant code snippet looks like the following:

for i=1:numberOfFolds
    if (i==1)
        net = init(net); % Random initialization
    else
        % bestIW{i-1} and bestLW{i-1} are the net.IW and net.LW 
        % of the best trained network of the previous fold
        net.layers{1:end}.initFcn = 'initwb';
        net = setwb(net,[bestIW{i-1},bestLW{i-1}]);
    end
end

When I run the full code I get the following error message:

Index exceeds the number of array elements (12).

I'm not sure how to format the second arguement of the setwb command in order for this to work. Also, do I need to run init(net) after setwb?

Answer 1

[Abhaya]

Hi John,

The error message Index exceeds the number of array elements (12) indicates that the size of the second input, [bestIW{i-1},bestLW{i-1}], does not align with the structure expected by the net in the MATLAB setwb function. Specifically, the setwb function requires a single vector that contains the combined weights and biases of the network, but the second input currently includes only the weights.

To solve the issue, you can follow the steps given below.

• Create the Weight Vector: Combine bestIW{i-1}, bestLW{i-1} and bias net.b into one vector:

weights = [bestIW{i-1}(:); bestLW{i-1}(:); net.b(:)]

• Set the Network Weights: Pass the combined vector to setwb():

net = setwb(net, weights);

After using MATLAB ‘setwb’ function, you do not need to run init(net) again, because the MATLAB setwb function directly sets the weights and biases.

For more information, please refer to following MATLAB documentation for ‘setwb’ function.

https://www.mathworks.com/help/deeplearning/ref/setwb.html

Hope it resolves your query.