How do I access individual weight matrices in a custom training loop?

Suppose I have a basic feedforward network of the type , minimising a loss function with respect to targets .

I would like to add a penalty term that regularises an intermediate signal, such as .

Could I define a custom loss function to do this? How can I access individual weight matrices in a custom training loop?

Answer 1

[Abhijit Bhattacharjee]

I assume by custom training loops, you are referring to custom training loops in the Deep Learning Toolbox. If so, here is a good resource to start with: Define Model Loss Function for Custom Training Loop.

There are two different methods to access the weight arrays, depending on how you defined the neural network in the custom training loop. If you used a dlnetwork, then you can access the weights of any layer that has learnable parameters.

Example for dlnetwork:

% Create feedforward neural network
layers = [
    featureInputLayer(1,"Name","featureinput")
    fullyConnectedLayer(10,"Name","fc_1")
    sigmoidLayer("Name","sigmoid_1")
    fullyConnectedLayer(10,"Name","fc_2")
    sigmoidLayer("Name","sigmoid_2")
    fullyConnectedLayer(10,"Name","fc_3")
    sigmoidLayer("Name","sigmoid_3")];

dlnet = dlnetwork(layers);

% Access weights of the first FC layer
dlnet.Layers(2).Weights

ans = 10×1

0.6758 -0.3062 -0.2549 -0.5748 0.3081 0.4561 0.4425 0.6709 -0.1209 -0.6227

If you used a model function in the custom training loop instead, then you would need to define a parameters array that contains the weights and biases of the model upfront. You can access these parameters at any time using the names you already defined.