Face Difficulty when converting tensorflow model to Matlab

I have a part of tensorflow code that I need to translate to matlab, but fail to do that. I have checked deep learning toolbox and unable to resolve the issue. If someone can help me this question, it is very helpful.

My tensorflow code (Python) is the following:

def get_grad3(model, t, x,a_data,u_data,a_data_n,mean_a,mean_u,w_upd,w1,w2,w3,p_data,stdt,stdx,xdim,ydim,twf,xwf,p_dataf):
    % A tf.GradientTape is used to compute derivatives in TensorFlow
    with tf.GradientTape(persistent=True) as tape:
        % Split t and x to compute partial derivatives
        tape.watch(model.trainable_variables)
        loss, Lu, Lr, Lm, Pn,w1n,Lr1,Lr2,w_upd,kP = compute_loss3(model, t, x, a_data, u_data, a_data_n,mean_a,mean_u,w_upd,w1,w2,w3,p_data,stdt,stdx,xdim,ydim,twf,xwf,p_dataf)
    grads_u = tape.gradient(loss, model.trainable_variables) % loss function
    del tape
    return grads_u, loss, Lu, Lr, Lm,Pn,w1n,Lr1,Lr2,w_upd,kP

I followed my previous question response: https://www.mathworks.com/matlabcentral/answers/2140906-face-difficulty-when-converting-tensorflow-model-to-matlab?s_tid=mlc_ans_men_view&mentions=true#answer_1491566

My matlab code is the following:

function [grads_u, loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = get_grad3(model, t, x, a_data, u_data, a_data_n, mean_a, mean_u, w_upd, w1, w2, w3, p_data, stdt, stdx, xdim, ydim, twf, xwf, p_dataf)

    % Define a function for computing loss and other required variables
    function [loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = computeLossGradients(model, t, x)
        % Compute the loss and other outputs using the compute_loss3 function
        [loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = compute_loss3(model, t, x, a_data, u_data, a_data_n, mean_a, mean_u, w_upd, w1, w2, w3, p_data, stdt, stdx, xdim, ydim, twf, xwf, p_dataf);
        % Compute gradients of the loss with respect to the model's learnable parameters
        grads_u = dlgradient(loss, model.LearnableParameters, 'EnableHigherDerivatives', true);
    end

    % Use dlfeval to compute the loss and other variables
    [loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = dlfeval(@computeLossGradients, model, t, x);   
    
end

My error message is

Unrecognized function or variable 'w_upd'.

Error in get_grad3/computeLossGradients (line 6)
        [loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = compute_loss3(model, t, x, a_data, u_data, a_data_n, mean_a, mean_u, w_upd, w1, w2, w3, p_data, stdt, stdx, xdim, ydim, twf, xwf, p_dataf);

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

Error in deep.internal.dlfevalWithNestingCheck (line 19)
    [varargout{1:nargout}] = deep.internal.dlfeval(fun,varargin{:});

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

Error in get_grad3 (line 12)
    [loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = dlfeval(@computeLossGradients, model, t, x);

Error in PWIP_initialize_v2 (line 114)
        [grads_u, loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = get_grad3(model, tw, xw, a_data, u_data, a_data_n, mean_a, mean_u, epoch, w1, w2, w3, p_data, stdt, stdx, xdim, ydim, twf, xwf, p_dataf);

I have also tried the symbolic differentiation and it works okay.

function [grads_u, loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = get_grad3(model, t, x, a_data, u_data, a_data_n, mean_a, mean_u, w_upd, w1, w2, w3, p_data, stdt, stdx, xdim, ydim, twf, xwf, p_dataf)

    % Compute the loss and other required variables
    [loss, Lu, Lr, Lm, Pn, w1n, Lr1, Lr2, w_upd, kP] = compute_loss3(model, t, x, a_data, u_data, a_data_n, mean_a, mean_u, w_upd, w1, w2, w3, p_data, stdt, stdx, xdim, ydim, twf, xwf, p_dataf);
    
    % Define symbolic variables
    num_vars = numel(model.trainable_variables);
    syms model_sym [1 num_vars]
    
    % Convert loss to symbolic
    loss_sym = sym(loss);
    
    % Calculate gradients
    grads_u = sym('grads_u', [1 num_vars]);
    for i = 1:num_vars
        grads_u(i) = diff(loss_sym, model_sym(i));
    end
    
end 

Can anyone point out how can I improve the code by changing the function input and output?

Thanks for all suggestions.

Answer 1

[William Rose]

@Ze,

Your function compute_loss_gradients(...) calls function compute_loss3(). The call to compute_loss3(...) includes w_upd as one of the parameters passed to the function. (w_upd is also one of the outputs from the function.) The error occurs because w_upd has not been passed to compute_loss_gradients(...), so it is unknown, and cannot be passed to compute_loss3().