The error messages you're encountering suggest that the context in which you're calling dlgradient isn't quite right. In MATLAB, dlarray should be used to encapsulate data you want to differentiate, and dlgradient should be used inside a function called by dlfeval.
Here's how you can rewrite your MATLAB function to correctly compute gradients, including addressing the nested dlfeval issue:
function get_r(model, tw, xw, a_data_n, mean_a, mean_u, kP, dim1, dim2, w1, stdt, stdx)
% Convert inputs to dlarray if they are not already
tw = dlarray(tw);
xw = dlarray(xw);
% Define a function for computing gradients
function [At, Px, ux, ut] = computeGradients(tw, xw)
% Run the model
[a, u, p] = model.net_u(tw, xw);
% Compute gradients
At = dlgradient(sum(a, 'all'), tw);
Px = dlgradient(sum(p, 'all'), xw);
ux = dlgradient(sum(u, 'all'), xw);
ut = dlgradient(sum(u, 'all'), tw);
end
% Use dlfeval to compute gradients
[At, Px, ux, ut] = dlfeval(@computeGradients, tw, xw);
end
Here's a brief explantion for the above code snippet:
- Convert Inputs to dlarray: Ensure tw and xw are of type dlarray before passing them into the function for gradient computation.
- Define a Gradient Computation Function: The computeGradients function performs the necessary operations and computes the gradients. The sum operation is used to ensure that dlgradient has a scalar to differentiate against.
- Use dlfeval: dlfeval is used to call computeGradients, which ensures that operations inside computeGradients are properly traced by the automatic differentiation engine.
