"size of predictions and target values must match" when using custom loop training for multi-input and multi-output

Dear all,

I build a deep learning model with 2 input and 2 output and use a custom loop training. But I have an error when call dlfeval(): "size of predictions and target values must match" . What should I fix? Here is my code:

[grad,state,loss] = dlfeval(@modelGradientsMulti, dlnet, input1, input2, output1, output2);
%support function
function [gradients,state,loss] = modelGradientsMulti(dlnet,in1,in2,out1,out2)
[pre1,pre2,state] = forward(dlnet,in1,in2,'Outputs',["fc1" "fc2"]);
loss1 = mse(pre1,output1);
loss2 = mse(pre2,output2);
loss = loss1 + loss2;
gradients = dlgradient(loss,dlnet.Learnables);
end

Thank you for reading my question. I hope someone expert will give a help.

[Katja Mogalle]

From the error message my first guess would be that pre1 and output1 don't have the same size, or that pre2 and output2 don't have the same size. Can you put a breakpoint at the line where you compute mse for the first time and check the sizes of pre1, output1, pre2, and output2? If there are inconsistencies between the sizes we need to look further if the data or the network is set up incorrectly.

To better understand the sizes of activations in the network, I always suggest using analyzeNetwork:

>> analyzeNetwork(dlnet)

Let me know how you get on and what you find.

[Van Vy]

@Katja Mogalle: Thank you for your reply, I checked my network and there was no problem. Now I'm trying to re-organize my code in another way because I stucked in this point long time already.

Answer 1

[Jaimin]

Hi @Van Vy

To resolve the error "size of predictions and target values must match" in a deep learning model with two inputs and two outputs using a custom training loop, I have created a simplified demo code.

Kindly refer to it for understanding.

% Create dummy data 
input1 = dlarray(randn(10, 32), 'CB'); % 10 features, 32 samples 
input2 = dlarray(randn(10, 32), 'CB'); % 10 features, 32 samples 
output1 = dlarray(randn(3, 32), 'CB'); % 3 outputs, 32 samples 
output2 = dlarray(randn(3, 32), 'CB'); % 3 outputs, 32 samples 

% Custom training loop 
numEpochs = 5; 
learningRate = 0.01; 

for epoch = 1:numEpochs 
    % Evaluate the model gradients and loss using dlfeval 
    [gradients, state, loss] = dlfeval(@modelGradientsMulti, dlnet, input1, input2, output1, output2); 

    % Update the network parameters using SGD 
    dlnet = dlupdate(@(w, g) w - learningRate * g, dlnet, gradients); 

    % Display the loss 
    fprintf('Epoch %d, Loss: %.4f\n', epoch, loss); 
end 

 

% Support function for computing gradients and loss 
function [gradients, state, loss] = modelGradientsMulti(dlnet, in1, in2, out1, out2) 
    % Forward pass 
    [pre1, pre2, state] = forward(dlnet, in1, in2, 'Outputs', ["output1", "output2"]); 

    % Compute losses 
    loss1 = mse(pre1, out1); 
    loss2 = mse(pre2, out2); 
    loss = loss1 + loss2; 

    % Compute gradients 
    gradients = dlgradient(loss, dlnet.Learnables); 
end 

For more information kindly refer following MathWorks documentation.

dlfeval: https://www.mathworks.com/help/deeplearning/ref/dlfeval.html

dlupdate: https://www.mathworks.com/help/deeplearning/ref/dlupdate.html

I hope this will be helpful.