dlgradient throws Value to differentiate is non-scalar. It must be a traced real dlarray scalar.error

Hi everyone,

With dfleval I call a function with which I do a number of computations. I need to calculate few gradients within that function but it throws an error. Here is the code.

function [grads,loss] = schooling(data,Q,params) 
% Trains the model
%   receives the model and input parameters
l = data(2:end,1); %x coordinates
t = data(1,2:end); %time coordinates

[T,D] = NN(params,t,l); % feed forward

[Tt,Tx] = dlgradient(T,t,l); % here I have error message. % all inputs T,t,l are dlarrays. The function schooling is 
% called from within dfleval

Txdx = dlgradient(D.*Tx,l);

Tgt = data(2:end,2:end); % readings
loss1 = (T-Tgt).^2;
loss2 = (Txdx+Q-Tt).^2;
loss = sum(loss1+loss2,1);
loss = sum(loss);
grads = dlgradient(loss,params);
end

Answer 1

[Walter Roberson]

[Tt,Tx] = arrayfun(@(Tscalar) dlgradient(Tscalar,t,l), T, 'uniform', 0);

Tt and Tx will then be cell arrays the same size of T, in which the entries are the gradients for the corresponding T values.

You will also need to adjust the code further down.

The basic issue is that your T is not a scalar dlarray object.

[Alim]

Thank you Walter. Just spot on. Indeed my T was not a scalar rather a matrix. I summed up all its elements and put in dlgradient and it worked. Thank you again