predict function not working in custom training loop

I am building a custom training loop for a simple LSTM classification network because I need a custom loss function (specifically, 0-1 loss). I have followed a tutorial but when I call the predict function within my custom loss function, I get the error: 'Undefined function 'predict' for input arguments of type 'nnet.cnn.layer.Layer'.'

I can successfully train the network with trainNetwork, so I am wondering what trainNetwork is doing that I am missing. If I call predict after training with trainNetwork, it works, but not in the training loop.

my input is a 50x1 sequence that is either classified as 1 or 2 (depending on if its average is positive or negative).

My network is defined as follows:

numFeatures = 1; %input data value (50 time points in sequence)
numHiddenUnits = 100;
numClasses = 2; %Left/rigth decision at the end
layers = [ ...
    sequenceInputLayer(numFeatures)
    lstmLayer(numHiddenUnits,"OutputMode","last")
    fullyConnectedLayer(numClasses)
    softmaxLayer
    classificationLayer];

net = layers;

and my custom loss function is:

function [gradients,state,loss] = customGradients2Lay(net,dlX,Ylabel)
    [Y,state]=predict(net,dlX);
    loss=loss01(Y,Ylabel);
    gradients=dlgradient(loss,net.Learnables);
end


function loss = loss01(Y, T)
    if isequal(Y,T)
        loss = 0;
    else
        loss = 1;
    end
end

My training loop just calls a random dataset to test on and sends it to predict. The error again is: Undefined function 'predict' for input arguments of type 'nnet.cnn.layer.Layer'.

I also am not sure why its calling nnet.cnn. I even built a custom fully connected layer to try to get around this and it was still calling nnet.cnn class.

What am I missing?

Answer 1

[James Gross]

Hello,

To train your network in a custom training loop, you must specify your network as a dlnetwork.

net = dlnetwork(layers);

You should then be able to train and call predict on your network as desired. For examples of how to train using a custom training loop with a dlnetwork, you can refer to one of the following:

https://www.mathworks.com/help/deeplearning/ug/train-network-using-custom-training-loop.html

https://www.mathworks.com/help/deeplearning/ug/train-network-using-custom-mini-batch-datastore-for-sequence-data.html

I hope this information helps!

[Isabella]

That worked, thank you!