"parfor" skips or fail with filtfilt function

Hi, I am trying to use 'parfor' to speedy a loop that performs 3 steps to a simple time series (SignCorr): 1 - Detrend 2 - Linear regression 3 - Low Pass Filtering Apparently, 'parfor' skip the step 3. I can run the same code just changing 'parfor' to 'for' and the results are different. Then I noticed that the difference is the LowPass Filtering Step. I am attaching the code only with the LPF and the 'parfor' results in a generic error: "Struct contents reference from a non-struct array object.".

Any idea? Thanks!

Fs = 1/2; % sampling frequency (Hz)
tp = 180; % Time Points

% LowPass Filter design
FpassLOW = 0.1;   % passband frequency (Hz) 
FstopLOW = 0.1 + (Fs/tp); % stopband frequency (Hz)
ApassLOW = 1;% passband ripple (dB)
AstopLOW = 50;  % stopband attenuation (dB)
hLOW = fdesign.lowpass('Fp,Fst,Ap,Ast',FpassLOW,FstopLOW,ApassLOW,AstopLOW,Fs);
HdLOW = design(hLOW, 'equiripple');

EPIreshTMP = zeros(size(finalEPIresh));
tic
parfor tt3 = 1:size(EPIresh,2) % EPIresh = 180x176004 Matrix (Time x Spatial Dimensions)
    SignCorr = EPIresh(:,tt3);
    MMM = mean(SignCorr);

      % Low Pass Filtering
      s_filt2 = repmat(s_corr,6,1); % we use the center copy, avoinding phase distortion
      s_filtAB1LOW = filtfilt(HdLOW.Numerator,1,s_filt2);
      s_filtAB1LOW = s_filtAB1LOW((1+2*180):(3*180));
      EPIreshTMP(:,tt3) = s_filtAB1LOW(:) + MMM; %giving the mean back
  end
  toc
  newEPIresh = EPIreshTMP;

[Edric Ellis]

I'm not sure what's going on here, and unfortunately I can't run this code - it would be really helpful for us reading your question, and probably for you too, if you could simplify your code to a "Minimal, complete, and verifiable example".

[Brunno Machado de Campos]

You are right. here a total independent code, that works with 'for' but not with 'parfor'. Just tested in a fresh Matlab session:

EPIresh = rand(180,100000); % simulating my image data
Frgval = [rand(180,6),ones(180,1)]; % simulating my
% regression matrix (confounding series + column of ones

Fs = 0.5; % sampling frequency (Hz)
tp = 180; % Time Points

% LowPass Filter design
FpassLOW = 0.1;   % passband frequency (Hz) 
FstopLOW = 0.1 + (Fs/tp); % stopband frequency (Hz)
ApassLOW = 1;% passband ripple (dB)
AstopLOW = 50;  % stopband attenuation (dB)
hLOW = fdesign.lowpass('Fp,Fst,Ap,Ast',FpassLOW,FstopLOW,...
       ApassLOW,AstopLOW,Fs);
HdLOW = design(hLOW, 'equiripple');
% End of LowPass Filter design

EPIreshTMP = zeros(size(EPIresh));

tic
parfor tt3 = 1:size(EPIresh,2)
    SignCorr = EPIresh(:,tt3);
    MMM = mean(SignCorr);

      % Detrend
      SignCorr = detrend(SignCorr,'linear');% detrend
      SignCorr = SignCorr + MMM; %giving the mean back

      % Regression
      b = regress(SignCorr,Frgval); 
      WReg  = Frgval*b;
      s_corr = SignCorr - WReg;
      s_corr = s_corr + b(end); % giving the mean back

      % Low Pass Filtering
      s_filt2 = repmat(s_corr,6,1); % I put the real data in a
           % sanduiche of itself, avoinding phase distortion
      s_filtAB1LOW = filtfilt(HdLOW.Numerator,1,s_filt2);
      s_filtAB1LOW = s_filtAB1LOW((1+2*180):(3*180)); %¨taking the
           % center copy of the series
      EPIreshTMP(:,tt3) = s_filtAB1LOW(:)+ b(end);%giving the mean back
  end
  toc
  newEPIresh = EPIreshTMP;

[Jan]

Please post the complete error message. It runs successfully in Matlab Online.

[Brunno Machado de Campos]

The error was a simple: "Struct contents reference from a non-struct array object" as I mentioned on the original question. The solution suggested by OCDER worked properly. Thank you anyway for your attention.

[Jan]

@Brunno: If you post the complete error message, we could see, which line is causing the problem. It is very useful for a discussion, if the readers do not have to guess, where the problem occurs. So please do not post just a part, which seems to be relevant, but everything printed in red to the command window.

[Brunno Machado de Campos]

@Jan: As I mentioned: attached a print of the error.

And to be crystal clear, the Line 17 is: "parfor tt3 = 1:size(EPIresh,2)"

Not sure of how this is more informative than what I posted before but anyway. Thank you.

Answer 1

[OCDER]

Looks like your object HdLOW is not being passed to the parfor correctly.

The output of design seems to be a System Object. https://www.mathworks.com/help/dsp/ref/design.html

parfor processes Objects differently than a pure structure. https://www.mathworks.com/help/distcomp/objects-and-handles-in-parfor-loops.html

Try the following:

EPIresh = rand(180,100000);         % simulating my image data
Frgval = [rand(180,6),ones(180,1)]; % simulating my

% regression matrix (confounding series + column of ones
Fs = 0.5; % sampling frequency (Hz)
tp = 180; % Time Points

% LowPass Filter design
FpassLOW = 0.1;             % passband frequency (Hz)
FstopLOW = 0.1 + (Fs/tp);   % stopband frequency (Hz)
ApassLOW = 1;               % passband ripple (dB)
AstopLOW = 50;              % stopband attenuation (dB)
hLOW = fdesign.lowpass('Fp,Fst,Ap,Ast',FpassLOW,FstopLOW,ApassLOW,AstopLOW,Fs);
HdLOW = design(hLOW, 'equiripple');

EPIreshTMP = zeros(size(EPIresh));
tic

HdLOW_Numerator = HdLOW.Numerator; %<===Pull out what you need from the object
parfor tt3 = 1:size(EPIresh,2)
    SignCorr = EPIresh(:,tt3);
    MMM = mean(SignCorr);
    % Detrend
    SignCorr = detrend(SignCorr,'linear');% detrend
    SignCorr = SignCorr + MMM; %giving the mean back
    % Regression
    b = regress(SignCorr,Frgval);
    WReg  = Frgval*b;
    s_corr = SignCorr - WReg;
    s_corr = s_corr + b(end); % giving the mean back
    % Low Pass Filtering
    s_filt2 = repmat(s_corr,6,1); % I put the real data in a
    % sanduiche of itself, avoinding phase distortion
    s_filtAB1LOW = filtfilt(HdLOW_Numerator,1,s_filt2); %<===Change this to not call the object
    s_filtAB1LOW = s_filtAB1LOW((1+2*180):(3*180)); %¨taking the
    % center copy of the series
    EPIreshTMP(:,tt3) = s_filtAB1LOW(:)+ b(end);%giving the mean back
end
toc
newEPIresh = EPIreshTMP;

[Brunno Machado de Campos]

That's worked properly! I missed that "differences" that you mentioned on for and parfor. Thank you!

[Brunno Machado de Campos]

Just to add another observation: In a 'parfor' loop, we can't even check if a structure variable is empty or not (isempty(XXX)).

Thank you all!