How can we convert SeriesNetwork to Network?

Dear Guys,

I already have done some customized feedfarword networks which work on sound/speech processing with 'Network' type (processed by 'train' function). Then, I have tried to use an 'trainNetwork' function to train DNN (by ignoring lstmlayer) as well as LSTM. However, I do not have any idea how to utlize this object; like how to implement as Network object on my old project.

Is there anyway to convert SeriesNetwork to Network?

Thank you in advance,

Qilonz

Answer 1

[Akshat]

Hi @Qilonz,

As far as I know, there is no way to convert a "SeriesNetwork" to "Network" via inbuilt methods.

As a workaround, you can try extracting the weights and biases of the "SeriesNetwork" and make a custom "Network" using a script like the following:

% Assume 'trainedNet' is your SeriesNetwork object
layers = trainedNet.Layers;

% Create a new Network object
net = network;
numLayers = numel(layers);
net.numLayers = numLayers;
for i = 1:numLayers
    layer = layers(i);
    
    % Example for a fully connected layer
    if isa(layer, 'nnet.cnn.layer.FullyConnectedLayer')
        % Set the layer size
        net.layers{i}.size = layer.OutputSize;
        
        % Set weights and biases
        net.IW{i,1} = layer.Weights;
        net.b{i} = layer.Bias;
    end
    
    % Add similar logic for other layer types
end

% Configure input and output settings
net.inputs{1}.size = size(layers(1).InputSize, 1);
net.outputs{numLayers}.size = layers(end).OutputSize;

% You will need to set up connections and transfer functions manually

Hope this helps your use case.