network object custom weights initialization

I'm having trouble understanding the network object.

Question: How can I initialize the weights with a custom function?

So far, I've understood that I need to set ffnet.initFcn = 'initLay' to enable the layers init functions; then I need to set for each layer i ffnet.layers{i}.initFcn='initwb' to enable the weights init functions; but how should I continue this?

ffnet.initFcn = 'initLay';
for layer_index=1:size(ffnet.layers, 1)
    ffnet.layers{layer_index}.initFcn = 'initwb';
end

What is the signature of a well design weights initialization function?

This is the code I produced so far, but I don't think you will need, anyway I will post it:

clear all
close all


load('Xprostate_data.mat')

X=Xprostate(:,1:8);
Y=table2array(Xprostate(:,9));
ind=table2array(Xprostate(:,10));
n=size(X,1);

Xp = zscore(table2array(X));

Xtrain=Xp(ind==1,:); Ytrain=Y(ind==1);   %dataset for training
Xtest=Xp(ind==0,:); Ytest=Y(ind==0);     %dataset for test

%%% 
ffnet = feedforwardnet(netconf);
ffnet = train(ffnet, Xtrain', Ytrain');

ffnet.trainFcn = 'traingd';
ffnet.trainParam.lr = 0.001;

net.initFcn = 'initLay'; %initLay 
for layer_index=1:size(net.layers, 1)
    net.layers{layer_index}.initFcn = 'initwb';
    
end


ffnet = init(ffnet);
ffnet = train(ffnet, Xtrain', Ytrain');

function weights = randomNormalWeights(sz, mean, std)
weights = mean + randn(sz) * std;
end

Answer 1

[Neha]

Hi Fabiano,

It is my understanding that you need to initialize the weights using a custom initialization function. Please refer to the code given below:

[icRows,icCols]=size(ffnet.inputConnect);
[lcRows,lcCols]=size(ffnet.layerConnect);
for i=1:icRows
    for j=1:icCols
        if ffnet.inputConnect(i,j)==1
            ffnet.IW{i,j}= randomNormalWeights(ffnet.inputWeights{i,j}.size,1,1); %matrix of size ffnet.inputWeights{i,j}.size, here: 10x8 (number of layersxnumber of inputs)
        end
    end
end
for i=1:lcRows
    for j=1:lcCols
        if ffnet.layerConnect(i,j)==1
            ffnet.LW{i,j}=randomNormalWeights(ffnet.layerWeights{i,j}.size,1,1); %matrix of size ffnet.layerWeights{i,j}.size, here: 10x1
        end
    end
end

For detailed explanation of the functions used, please refer to the following documentation links:

Input Weights and Layer Weights, Input Connect and Layer Connect and IW and LW properties.