Fittype that calls a subfunction and integration with appdesigner

I have a subfunction that traditionally has sat in its own .m file. I'd like to integrate into an app code so everything is consolidated when I go to share it. However, the problem is the subfunction has to be called by "fittype". I think this is where I'm running into issues. The App can see the subfunction, but fittype can't see the function if its in the app.

I'm not sure if there's a way to set up the function in AppDesigner so that the "fittype" can see the subfunction. I tried making it both a function of the App, as well as as a nested type function where "fittype" is called. Neither worked.

The subfunction is very simple. It's a setup to let me do a Legendre polynomial fit:

function m0coeff = mod_legendre(n,x)
      %Legendre Polynomial return: note that MatLab legendre function returns all
      %coefficients from m=0...n, whereas the Legendre Polynomial is m=0 only.

      LegCoef = legendre(n,x);
      m0coeff = LegCoef(1,:)';
end

This then is called within appdesigner. I have to make a custom equation fit, and it needs to use the above function when it does the fit.

opts        = fitoptions( 'Method', 'NonlinearLeastSquares' );
opts.Robust  ='On';
opts.Display ='Off';
legnum       = app.NumLegModes.Value;

for ii = 0:legnum
      if ii==0
           eqn = ['p' num2str(ii) '*mod_legendre(' num2str(ii) ',x)'];
           opts.StartPoint =1;
      else
           eqn = [eqn '+p' num2str(ii) '*mod_legendre(' num2str(ii) ',x)'];
           opts.StartPoint =[opts.StartPoint 0];
      end
end
ft = fittype(eqn);
[legfit, gof, output] = fit(x,r,ft,opts);

The actual error is:

Error using fittype/testCustomModelEvaluation (line 12)

Expression

p0*mod_legendre(0,x)+p1*mod_legendre(1,x)+p2*mod_legendre(2,x)+p3*mod_legendre(3,x)+p4*mod_legendre(4,x) is

not a valid MATLAB expression, has non-scalar coefficients, or cannot be evaluated:

Error in fittype expression ==>

p0.*mod_legendre(0,x)+p1.*mod_legendre(1,x)+p2.*mod_legendre(2,x)+p3.*mod_legendre(3,x)+p4.*mod_legendre(4,x)

??? Undefined function 'mod_legendre' for input arguments of type 'double'.

Error in fittype>iCreateFittype (line 362)

testCustomModelEvaluation( obj );

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Error in fittype (line 319)

obj = iCreateFittype( obj, varargin{:} );

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Error in dblshell_Shape/leg_fit (line 483)

ft = fittype(eqn);

^^^^^^^^^^^^

Error in dblshell_Shape/LegendreLeastSquareButtonPushed (line 2162)

handles = leg_fit(app, handles);

^^^^^^^^^^^^^^^^^^^^^

Error in matlab.apps.AppBase>@(source,event)executeCallback(ams,app,callback,requiresEventData,event) (line 60)

newCallback = @(source, event)executeCallback(ams, ...

^^^^^^^^^^^^^^^^^^^^^^^^

Caused by:

Error using fittype/evaluate (line 57)

Error in fittype expression ==>

p0.*mod_legendre(0,x)+p1.*mod_legendre(1,x)+p2.*mod_legendre(2,x)+p3.*mod_legendre(3,x)+p4.*mod_legendre(4,x)

??? Undefined function 'mod_legendre' for input arguments of type 'double'.

[dpb]

I've never had any issue with an m-file not being found if it is in the working path of the app...I don't see why it wouldn't also be visible here, but without the actual app code structure it's not possible to see what might be going on.

[Matt J]

Your model is linear (in p). Using a nonlinear least squares fitter is super inefficient.

Answer 1

[Steven Lord]

Try specifying the input to fittype not as a text expression but as an anonymous function that calls the local function inside the anonymous function.

[Torsten]

@Steven Lord

So "fittype" cannot be used for text expressions containing user-defined function names ?

[Steven Lord]

It can, but I believe (though I'm not 100% certain) that like the Callback property of a uicontrol, if you specify that text as an expression "MATLAB evaluates this expression in the base workspace." So if you specify the name of a local function that won't work. A function handle to that local function or an anonymous function that calls it, though, can call the local function if it was created when that local function was in scope.

[Walter Roberson]

This is exactly the problem: quoted text is evaluated in the base workspace, outside the context that defines local functions. Use anonymous functions instead of quoted text.

[plasmageek]

Thanks! I will give this a go at some point...I did see about the anonymous functions in the matlab help, but I've never done it and I wasn't sure it would work.

Answer 2

[dpb]

As I expected, if the subfunction is an m-file in the working space, it works just fine...I built a toy app that mimics your posted code attached -- it consists of a spinner set to 4 and a button that calls your function with fake data...

 % Callbacks that handle component events
  methods (Access = private)

    % Button pushed function: Button
    function ButtonPushed(app, event)
      opts        = fitoptions( 'Method', 'NonlinearLeastSquares' );
      opts.Robust  ='On';
      opts.Display ='Off';
      legnum       = app.NumLegModesSpinner.Value;
      for ii = 0:legnum
        if ii==0
         eqn = ['p' num2str(ii) '*mod_legendre(' num2str(ii) ',x)'];
         opts.StartPoint =1;
        else
         eqn = [eqn '+p' num2str(ii) '*mod_legendre(' num2str(ii) ',x)'];
         opts.StartPoint =[opts.StartPoint 0];
        end
      end
      ft = fittype(eqn);
      x=linspace(-1,1,20); x=x(:);
      r=rand(size(x));
      [legfit, gof, output] = fit(x,r,ft,opts);
    end
  end

The fit function ran to completion although the function doesn't do anything with the outputs and they're not made part of the app properties struct so they're lost at the moment, but the code does run.

I put the mod_legendre function m-file in the same working folder; when you package the app, the dependency walker will find it is needed and include it as additional file needed.

This is a case as @Steven Lord notes, the way fittype is constructed the local function isn't in the base workspace scope in which MATLAB evaluates the passed string. The solution is to continue to use the m-file as demonstrated with the sample app or recast to use an anonymous function created in scope that refers to the local function.

Of course, this addresses only the structure of the app itself; @Matt J's observation that the model is linear in the coefficients is significant.

[plasmageek]

Thanks! Unforunatley I'm not allowed to share the actual app (work rules). You're right, it does work with the separate .m file, which how I've been fine so far. I've not done a compiled app through Appdesigner for, so I wasn't sure it would save the extra dependency. Right now that function isn't even working in my Matlab, so I can't test it either. I have to do it all myself manually. If it compiles it with the app, then I may just leave it until I have actual "free" time to update it.

I haven't done an anonymous function before; I had the thought to try it, but I wanted to post here first to see if people thought that method might work.

Answer 3

[Matt J]

Your model is linear (in p). Using a nonlinear least squares fitter is super inefficient.

So, instead, you should do,

A = cell2mat(arrayfun(@(n)mod_legendre(n,x),0:legnum,'uni',0));

p=num2cell(A\r);

ft = cfit(eqn,p{:});

gof=gofCalc(ft,x)

function gof=gofCalc(ft,x,r)

    r_fit = feval(ft, x);
    
    % Compute goodness-of-fit statistics
    SSE = sum((r - r_fit).^2);         % Sum of squared errors
    SST = sum((r - mean(r_fit)).^2); % Total sum of squares
    R_square = 1 - (SSE / SST);              % R-squared
    RMSE = sqrt(SSE / length(r));      % Root Mean Square Error
    
    % Create the gof structure
    gof = struct('sse', SSE, 'rsquare', R_square, 'rmse', RMSE);

end

function m0coeff = mod_legendre(n,x)
      %Legendre Polynomial return: note that MatLab legendre function returns all
      %coefficients from m=0...n, whereas the Legendre Polynomial is m=0 only.

      LegCoef = legendre(n,x);
      m0coeff = LegCoef(1,:)';
end