Value to differentiate is not traced. It must be a traced real dlarray scalar. Use dlgradient inside a function called by dlfeval to trace the variables.

Hello, I am working on customizing the loss function to minimize dimensionality by maximizing the Bhattacharyya distance distance.

But it came up with a error shown as:

Error using dlarray/dlgradient
Value to differentiate is not traced. It must be a traced real dlarray scalar. Use dlgradient inside a function called by dlfeval to trace the variables.

Error in mutiatt_nld>modelGradients (line 54)
    gradients = dlgradient(loss, dlnet.Learnables);

Error in deep.internal.dlfeval (line 17)
[varargout{1:nargout}] = fun(x{:});

Error in deep.internal.dlfevalWithNestingCheck (line 15)
    [varargout{1:nargout}] = deep.internal.dlfeval(fun,varargin{:});

Error in dlfeval (line 31)
[varargout{1:nargout}] = deep.internal.dlfevalWithNestingCheck(fun,varargin{:});

Error in mutiatt_nld (line 24)
    [gradients, loss] = dlfeval(@modelGradients, dlnet, dlX, N);
    

The code:

    % Parameter settings
M = 10; % Dimension of input features
N = 50; % Number of samples per class
numEpochs = 100;
learnRate = 0.01;

% Generate example data
X = rand(2*N, M);
X(1:N, :) = X(1:N, :) + 1; % Data for class A
X(N+1:end, :) = X(N+1:end, :) - 1; % Data for class B

% Define the neural network
layers = [
    featureInputLayer(M, 'Normalization', 'none')
    fullyConnectedLayer(10)
    reluLayer
    fullyConnectedLayer(3)
];
dlnet = dlnetwork(layerGraph(layers));

% Custom training loop
for epoch = 1:numEpochs
    dlX = dlarray(X', 'CB'); % Transpose input data to match network's expected format
    [gradients, loss] = dlfeval(@modelGradients, dlnet, dlX, N);
    dlnet = dlupdate(@sgdmupdate, dlnet, gradients, learnRate);
    disp(['Epoch ' num2str(epoch) ', Loss: ' num2str(extractdata(loss))]);
end

% Testing phase
X_test = rand(N, M); % Assume test data is randomly generated
dlX_test = dlarray(X_test', 'CB'); % Transpose input data to match network's expected format
Y_test = predict(dlnet, dlX_test);
disp('Dimensionality reduction results during testing:');
disp(extractdata(Y_test)');

% Custom loss function
function loss = customLoss(Y, N)
    YA = extractdata(Y(:, 1:N))';
    YB = extractdata(Y(:, N+1:end))';
    muA = mean(YA);
    muB = mean(YB);
    covA = cov(YA);
    covB = cov(YB);
    covMean = (covA + covB) / 2;
    d = 0.25 * (muA - muB) / covMean * (muA - muB)' + 0.5 * log(det(covMean) / sqrt(det(covA) * det(covB)));
    loss = -d;  % Maximize Bhattacharyya distance
    loss = dlarray(loss); % Ensure loss is a tracked dlarray scalar
end

% Model gradient function
function [gradients, loss] = modelGradients(dlnet, dlX, N)
    Y = forward(dlnet, dlX);
    loss = customLoss(Y, N);
    gradients = dlgradient(loss, dlnet.Learnables);
end

% Update function
function param = sgdmupdate(param, grad, learnRate)
    param = param - learnRate * grad;
end

Answer 1

[Ganesh]

Hi @yingyu jiang,

You are getting this error as you are using "extractData" with a traced argument. This would lead to breaking of tracing. You may use the following documentation to know what norms need to be followed within your loss function:

https://www.mathworks.com/help/deeplearning/ug/include-automatic-differentiation.html

I have tried your code with a different loss function, and this worked:

% Parameter settings
M = 10; % Dimension of input features
N = 50; % Number of samples per class
numEpochs = 100;
learnRate = 0.01;

% Generate example data
X = rand(2*N, M);
X(1:N, :) = X(1:N, :) + 1; % Data for class A
X(N+1:end, :) = X(N+1:end, :) - 1; % Data for class B

% Define the neural network
layers = [
    featureInputLayer(M, 'Normalization', 'none')
    fullyConnectedLayer(10)
    reluLayer
    fullyConnectedLayer(3)
];
dlnet = dlnetwork(layerGraph(layers));

% Custom training loop
for epoch = 1:numEpochs
    dlX = dlarray(X', 'CB'); % Transpose input data to match network's expected format

    Y = forward(dlnet, dlX);
    Yact = rand(size(Y));

    [gradients, loss] = dlfeval(@modelGradients, dlnet, Y, Yact);
    dlnet = dlupdate(@sgdmupdate, dlnet, gradients, learnRate);
    disp(['Epoch ' num2str(epoch) ', Loss: ' num2str(extractdata(loss))]);
end

% Testing phase
X_test = rand(N, M); % Assume test data is randomly generated
dlX_test = dlarray(X_test', 'CB'); % Transpose input data to match network's expected format
Y_test = predict(dlnet, dlX_test);
disp('Dimensionality reduction results during testing:');
disp(extractdata(Y_test)');

% Model gradient function
function [gradients, loss] = modelGradients(dlnet, Y, Yact)
    loss = mse(Y, Yact);
    gradients = dlgradient(loss, dlnet.Learnables);
end

% Update function
function param = sgdmupdate(param, grad, learnRate)
    param = param - learnRate * grad;
end

I understand that as your loss function is unsupervised, you are running into issues. You will have to refactor the code accordingly.

Hope this helps!

[yingyu jiang]

thank you very much！

But I did not run your demonstration program successfully and encountered an error message:

Dot indexing is not supported for variables of this type.

Error in deep.internal.recording.containerfeval>iTablesToCells (line 460)
        args{input} = args{input}.Value;

Error in deep.internal.recording.containerfeval>iProcessNetwork_Nout_Nin (line 358)
iterableArgs = iTablesToCells(iterableArgs);

Error in deep.internal.recording.containerfeval>iDispatch_Nout_Nin (line 194)
    outputs = iProcessNetwork_Nout_Nin(fun, paramFun, numOut, ...

Error in deep.internal.recording.containerfeval (line 38)
    outputs = iDispatch_Nout_Nin(allowNetInput, fun, paramFun, numOut, ...

Error in deep.internal.networkContainerFixedArgsFun (line 29)
varargout = deep.internal.recording.containerfeval(...

Error in dlupdate (line 124)
    [varargout{1:nargout}] = deep.internal.networkContainerFixedArgsFun(...

Error in mutiatt_nld (line 29)
    dlnet = dlupdate(@sgdmupdate, dlnet, gradients, learnRate);