How to import weights in a neural network used in regression

I have weights and biases and I just want to import them to a neural network in matlab , so how can I do that ?

Answer 1

[Purvaja]

Hi @Ahmed Ramzy!

I understand that you have pre-trained weights and biases that you'd like to import into a neural network — here I'll be discussing both the Neural Network Toolbox (feedforwardnet) and Deep Learning Toolbox (dlnetwork) approaches.

1. Neural Network Toolbox (feedforwardnet)

Here’s how you can define a network, inject your weights and biases, and verify that the forward pass matches your manual calculation:

% Define layer sizes (nInput, nHidden, nOutput) 

% Example weights & biases with dummy data, ensure the sizes are similar to
% layer dimensions.
W1 = randn(nHidden, nInput);  
b1 = randn(nHidden, 1);
W2 = randn(nOutput, nHidden);  
b2 = randn(nOutput, 1);

% test input x

% Create custom network
net = network;
net.numInputs = 1;
net.numLayers = 2;
net.biasConnect = [1;1];        
net.inputConnect(1,1) = 1;     
net.layerConnect(2,1) = 1;         
net.outputConnect(2) = 1;

% Layer sizes
net.inputs{1}.size = nInput;
net.layers{1}.size = nHidden;
net.layers{2}.size = nOutput;

% Assign weights/biases
net.IW{1,1} = W1;
net.b{1}    = b1;
net.LW{2,1} = W2;
net.b{2}    = b2;

% Test forward pass
y = net(x);

2. Deep Learning Toolbox (dlnetwork)

Here’s how to use dlnetwork to load your custom weights and biases:

% Define layer sizes (nInput, nHidden, nOutput) 

% Pre-trained weights & biases (filled with dummy data for this example) 
W1 = randn(nHidden, nInput);   
b1 = randn(nHidden, 1);
W2 = randn(nOutput, nHidden);  
b2 = randn(nOutput, 1);

% Define network architecture
layers = [
    featureInputLayer(nInput,"Name","input")
    fullyConnectedLayer(nHidden,"Name","fc1")
    reluLayer("Name","relu1")
    fullyConnectedLayer(nOutput,"Name","fc2")
    ];

% Create dlnetwork
dlnet = dlnetwork(layers);

% Inject custom weights/biases
learnables = dlnet.Learnables;

% Wrap weights/biases as dlarray and assign
learnables.Value(learnables.Layer=="fc1" & learnables.Parameter=="Weights") = {dlarray(W1)};
learnables.Value(learnables.Layer=="fc1" & learnables.Parameter=="Bias")    = {dlarray(b1)};
learnables.Value(learnables.Layer=="fc2" & learnables.Parameter=="Weights") = {dlarray(W2)};
learnables.Value(learnables.Layer=="fc2" & learnables.Parameter=="Bias")    = {dlarray(b2)};

% Assign back to network
dlnet.Learnables = learnables;

% Forward pass with your imported weights 
X = randn(nInput,1);        % dummy input column vector
dlX = dlarray(X,"CB");      
dlY = predict(dlnet, dlX);
disp("Network output (dlnetwork):");
disp(extractdata(dlY));

For additional details on these APIs, refer to the MATLAB documentation:

1. Deep learning network: https://www.mathworks.com/help/deeplearning/ref/dlnetwork.html
2. Feedforward neutral network: https://www.mathworks.com/help/thingspeak/create-and-train-a-feedforward-neural-network.html

I hope this solves your doubt, I'll recommend to cross-check it after implementing.