cfit object doesn't work with 'plot' or 'differentiate'

Hello,

I stored a cfit obj to hard disc. It looks like:

General model:

     ans(xd) = off+amp./(1+exp(-a.*(xd-b)))
     Coefficients (with 95% confidence bounds):
       off =      -4.047  (-4.054, -4.041)
       amp =       8.179  (8.169, 8.189)
       a =       4.053  (4.034, 4.073)
       b =       2.271  (2.27, 2.273)

In the process of creating the obj I was able to plot it and to calculate the derivatives.

Now I want to come back to these data and tried:

>>plot(cfObj)

but receive:

Error using cfit/plot (line 111)
Error evaluating CFIT function:
Error while trying to evaluate CFIT model: obj:

Error while trying to evaluate FITTYPE function obj:

Argument must contain a string or function_handle.

For:

>>differntiate(cfObj, x)

I receive:

Error using cfit/feval (line 31)
Error while trying to evaluate CFIT model: fitobj:

Error while trying to evaluate FITTYPE function obj:

Argument must contain a string or function_handle.

Error in cfit/differentiate (line 33)
    f1 = feval(fitobj,x+ms);

Can anyone please help out with that? Sure I can recreate the fit from the coefficients, but I think it is not meant to be like that.

Thanks,

Kai

[Onno Broekmans]

Your fit results are definitely not useless. You can use coeffvalues(cfObj) to get the coefficient values from the fit manually, and then plot them using:

fitCoeffs = num2cell(coeffvalues(cfObj));
fitFun = @(off,amp,a,b,xd) off+amp./(1+exp(-a.*(xd-b)));
plot(x, feval(fitFun, fitCoeffs{:}, x, '-r');

(Warning: untested code, so there might be minor syntax errors in there. I'm assuming your x data is in the variable 'x').

Note that this is what plot(cfObj,x) does internally, so the results are exactly the same :)

[Kai]

Right, that's what I will do. I was only a bit shocked not seeing it come out in the first place. The function I will use the fits in looks like:

F(t) = kE(t)+rE(t)'+uE(t)''+c

So, I see the 'expr' statement growing a bit longer than before...

Answer 1

[Onno Broekmans]

I can confirm this behavior (see code to reproduce at the bottom). Unfortunately, it also affects parallel processing: if you try to perform a large number of fits in parallel, using a 'parfor' loop, the resulting cfit objects are "broken" in the way described by Kai.

Unless I've missed something, I'd consider this a bug. I've filed a Service Request (ID: 1-NE8ATH) with TMW. I'll get back to this thread when I hear back.

As a workaround, you could try manually evaluting the fit's anonymous function, and using the coeffvalues() function to retrieve the coefficients from the cfit object (that does work fine).

- Onno

Code to reproduce the issue:

%%Generate some random data, and plot it

data_x = 1:50;
data_y = (2.* data_x) + randn(size(data_x)) + 5;

figure;
plot(data_x, data_y, '.b');
title('Original data');

%%Fit the data using a simple model
model = @(A, B, x) A.*x + B;

ftype = fittype(model);
fopt = fitoptions(ftype);
fopt.StartPoint = [1 1];

f = fit(data_x(:), data_y(:), ftype, fopt);

%%Plot the fit
figure;
hold on;
    plot(data_x, data_y, '.b');
    plot(f, '-r');
hold off;
title('Fitted data');

%%Store cfit object on disk, and then retrieve it again
save('fit.mat', 'f');
clear('f');

load('fit.mat');

figure;
hold on;
    plot(data_x, data_y, '.b');
    plot(f, '-r');      % <-- crash here: 
                        % Error using cfit/plot (line 113)
                        % Error evaluating CFIT function:
                        % Error while trying to evaluate CFIT model: obj: 
                        % Error while trying to evaluate FITTYPE function obj: 
                        % Argument must contain a string or function_handle.

hold off;
title('Fitted data (fit loaded from disk)');

[Kai]

Hi Onno,

thanks for the code. It exactly reproduces the error. Next time I will communicate with code as well ;) Adding:

          try
             d1 = differentiate(f, data_x);
          catch err
             disp(err.message)
          end

on line 36 shows the other error s. above.

Also, thanks for filing the bug report, looking forward to the response. So long, I ll use the coefficients.

Best, Kai

[Onno Broekmans]

I forgot to post back, but TMW confirmed the issue (early July), and said they had their developers working on it.

[Jon Cherrie]

We have published a patch for this issue. See the bug report 968079 for details.

Answer 2

[Andrei Bobrov]

x = (0:.01:4)';                                         %
y = -4 + 8./(1+exp(4.*(x - 2))) + .5*randn(numel(x),1); % your data

ftfun = fit(x,y,fittype(@(a,b,c,d,x)a + b./(1+exp(c.*(x - d)))),...
                                                          'Startpoint',[1 1 1 1])

dft = differentiate(ftfun,x);

plot(x,[y,ftfun(x),dft])

[Kai]

Hi Andrei,

thanks for your answer. As I see it, you describe how to create a fit object and then use plot and differentiate. This is not really an answer to my question, because I already have the fit objects and know how to use the mentioned functions. The point is that directly after creation I could plot and differentiate them. But now, after a few days, I receive the errors shown above. Weird, I think. Or did I misunderstand your point?

Kai