The error "Value to differentiate must be a traced dlarray scalar" occurs because:
- Missing ‘dlfeval ‘Context
- ‘dlgradient’ requires the computation to be traced (tracked for automatic differentiation).
- In the code, ‘ModelD’ was called directly instead of inside ‘dlfeval’, so MATLAB could not trace the operations.
There are some other issues with code as well:
- Non-Scalar Loss
- The loss must be a scalar for gradient computation.
- The original ‘mse(Vt, X_in)’ might return a non-scalar (e.g., vector/matrix) if ‘Vt’ and ‘X_in’ are not properly reduced.
- Improper ‘dlarray’ Handling
- Some operations (like ‘repmat’, indexing) were breaking the computation graph, preventing gradient tracing.
Please refer to this MATLAB Code with execution on dummy input:
function [gradients, loss] = ModelD(k, M, Validate_train, params, NN_layer)
params.attention.weight = dlarray(params.attention.weight);
params.attention.output_weight = dlarray(params.attention.output_weight);
params.attention.bias = dlarray(params.attention.bias);
Validate_train = dlarray(Validate_train);
validate_data_in = Validate_train(randperm(size(Validate_train,1)),:);
Validate_train_x = validate_data_in(:,1:3);
Validate_train_y = validate_data_in(:,4:end);
A_zero = dlarray(zeros(size(Validate_train_y,1),1));
Validate_train_y = [Validate_train_y, A_zero];
Validate_data_x = dlarray(zeros(k, NN_layer));
for i = 1:k
for j = 1:NN_layer
Validate_data_x(i,j) = Validate_train_x(j);
Validate_train_x(j) = Validate_train_x(j+3);
end
end
y_in = Validate_train_y(1:M,:);
Index = randi([1,M],1,1);
X_in = Validate_data_x(Index,:);
Y_in = repmat(y_in(Index,:), 1, NN_layer); % Fixed repmat dimensions
score = dlarray(zeros(NN_layer,1), 'CB');
for i = 1:NN_layer
h = X_in(i);
ht = Y_in(1,i);
A = params.attention.weight(i) .* h;
B = params.attention.output_weight * ht;
C = params.attention.bias(i);
score(i) = tanh(A + B + C);
end
a = softmax(score);
Vt = dlarray(zeros(1,NN_layer), 'CB');
for i = 1:NN_layer
Vt(i) = a(i) * X_in(i);
end
loss = sum((Vt - X_in).^2, 'all');
gradients = dlgradient(loss, params);
end
k = 5;
M = 10;
NN_layer = 3;
Validate_train = rand(100, 7);
params.attention.weight = rand(NN_layer, 1);
params.attention.output_weight = rand(1, 1);
params.attention.bias = rand(NN_layer, 1);
params.attention.weight = dlarray(params.attention.weight);
params.attention.output_weight = dlarray(params.attention.output_weight);
params.attention.bias = dlarray(params.attention.bias);
[gradients, loss] = dlfeval(@ModelD, k, M, Validate_train, params, NN_layer);
disp(loss);
disp(gradients);
Hope this helps resolve the issue!
