Trouble with accessing variable in parfor loop

Hello all,

I am trying to do batch processing and I am having trouble implementing the parfor loop. I have a variable (a movie frame), that I want to subtract from every frame in a movie after I load in the frames. The basics of my loop look like this.

load('gray_mean_frame.mat');

set(handles.testMovieButton, 'String', 'Please Wait');

poolobj = parpool('local',4);
addAttachedFiles(poolobj,{'gray_mean_frame.mat'});

parfor frameNumber = 1:numFrames

      movieFrame = read(readerobj, frameNumber);

      ghostFrame = gray_mean_frame - movieFrame;
end

I always get the error

"An UndefinedFunction error was thrown on the workers for 'gray_mean_frame'.
This might be because the file containing 'gray_mean_frame' is not accessible
on the workers.  Use addAttachedFiles(pool, files) to specify the required
files to be attached.  See the documentation for
'parallel.Pool/addAttachedFiles' for more details."

But I think I am correctly passing the file gray_mean_frame to the workers. gray_mean_frame.mat is a file which contains the variable gray_mean_frame, among a couple others. I've also tried making gray_mean_frame a global variable but it did not help. It is not clear to me what I am doing wrong. Thank you for any help.

Answer 1

[Edric Ellis]

The problem here is that unfortunately the PARFOR analysis is based on the text of your program, and in this case the analysis cannot deduce that gray_mean_frame is a variable. You should be able to fix this by making a minor change:

    data = load('gray_mean_frame.mat');
    gray_mean_frame = data.gray_mean_frame;

[Andreas Holzhammer]

Thanks Edric, you helped me as well! I ran into the same problem because in a function, I used an input parser and then dynamically made variables out of every field.

varargin={};
p = inputParser;
   addParameter(p,'autosave',false,@islogical);
   addParameter(p,'initialSampleSize',[],@isnumeric);
parse(p,varargin{:});
opts=p.Results;
params=fieldnames(opts);
for i=1:numel(params)
    eval(sprintf('%s=opts.(params{i});',params{i}));
end
myZeros=cell(2,1);
parfor i=1:2
    myZeros{i}=zeros(initialSampleSize);
end

It suffices to make a dummy statement like above before the par loop, e.g.

initialSampleSize=initialSampleSize;

while just

    initialSampleSize;

won't work. Could anybody please explain why that is and whether there is a dynamic workaround? I don't want to have to add that dummy line every time somebody adds a parameter.

[Sean de Wolski]

Andreas, I would pass input parsing off to a second function (sub function probably) and have it return the parameters you care about

% First line of your main function
[x,y,z,initialSampleSize] = parse_inputs(varargin{:})

You're going to have to add them to the parse anyway so you might as well add them as outputs which is way cleaner and faster than using eval to poof them into existence.

Answer 2

[Vincenzo De Lipsis]

I found a solution to a similar issue which might be helpful to somebody.

I had a parfor loop inside a function, after loading a .mat file full of variables. For some reason the workers are not able to see variables loaded from an external .mat file; all variables need to be stated individually. But if the parfor loop part of the code is saved as an external .m Matlab script, which can be called inside the Matlab function, all variables are visible to the workers. I don't know the logic of this, and I didn't find explanation on the web.

[Sergiy Golovin]

Thank you for pointing it out! This is exactly what helped me. Simply putting the parfor loop in a separate script and then calling it fixes the error and allows the calculations to proceed. (Note: that Matlab is going to complain about the parfoor loop in the script by underlining your statements. Just ignore it.)

Answer 3

[Quentin]

Thanks Edric, that solved my problem.

Answer 4

[Kazutaka]

I have a similar issue, but not involving to load any files whose names are confusing with the variable names.

Here is the part of the code and once it reaches

"SpikeData1_1=extractdatapt...."

line, then it seems that parfor thinks I am loading a file and keep getting an error:

Error using COHshuf_SpkSpk_M1S1_allTM_y1_parfor>(parfor body) (line 87)
An UndefinedFunction error was thrown on the workers for 'SpikeData1'.  This
might be because the file containing 'SpikeData1' is not accessible on the
workers.  Use addAttachedFiles(pool, files) to specify the required files to be
attached.  See the documentation for 'parallel.Pool/addAttachedFiles' for more
details.

Error in COHshuf_SpkSpk_M1S1_allTM_y1_parfor (line 86)
    parfor i=1:length(sigpairs)

Caused by:
    Undefined function or variable 'SpikeData1'.

-----

Any ideas?

pc = parcluster('local')
pc.JobStorageLocation = strcat('/tmp/kazutaka/',getenv('SLURM_JOB_ID'))
matlabpool (pc, 12)

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% %NOW get M1S1pairs

m1=size(TMunion_M1,1);
s1=size(TMunion_S1,1);

max_units=max([m1 s1]);

covmat=[];
for r=1:m1
    d=[];
    d=r*ones(s1,1);
    d(:,2)=repmat(1:s1,1,1);
    covmat=[covmat;d];
end

frqname={'26';'613';'beta';'gamma'};
frqrng={3:11;12:26;31:63;64:107};
Tm=[];Tm=1.75:0.01:2.74;

%Create a sliced structure for M1 data  
for i=1:length(TMunion_M1);
    %tmp=[]; tmp=TMunion_M1(covmat(sigpairs(i),1),:);
    strtmp1=sprintf('elec%03d_%03d',TMunion_M1(i,1),TMunion_M1(i,2));
    eval(['lengthtemp = length(M1spk_MB_',strtmp1,');']);
    for ll = 1:lengthtemp
        eval(['SpikeData1(',num2str(TMunion_M1(i,1)),',',num2str(TMunion_M1(i,2)),',ll).times=M1spk_MB_',strtmp1,'(ll).times;']);
    end
end

%Create a sliced structure for S1 data  
for i = 1:length(TMunion_S1);
    %tmp=[]; tmp=TMunion_S1(covmat(sigpairs(i),1),:);
    strtmp1=sprintf('elec%03d_%03d',TMunion_S1(i,1),TMunion_S1(i,2));
    eval(['lengthtemp = length(S1spk_MB_',strtmp1,');']);
    for ll = 1:lengthtemp
        eval(['SpikeData2(',num2str(TMunion_S1(i,1)),',',num2str(TMunion_S1(i,2)),',ll).times=S1spk_MB_',strtmp1,'(ll).times;']);
    end
end

for frq=1:4
    %frq = 1;
    tmp=[];eval(['tmp=SIGCOHpk' frqname{frq} '(dsnum).pk;']);
    inx=[];inx=find(tmp < unique(Cconf(dsnum-1).cf));

      sigpairs=[];eval(['sigpairs=SIGCOHinx' frqname{frq} '(dsnum).cinx;']);
      sigpairs(inx)=[];%delete the nonsog (1,1(pairs

      peaktm=[];eval(['peaktm=SIGCOHpktm' frqname{frq} '(dsnum).ptm;']);
      peaktm(inx)=[];

      peak=[];eval(['peak=SIGCOHpk' frqname{frq} '(dsnum).pk;']);
      peak(inx)=[];

      %k=0;%count for sigpeak after permutation
      %tic
      sortedC = zeros(length(sigpairs),1000); 
      parfor i=1:length(sigpairs)
          size(SpikeData1)
          size(SpikeData2)
          k = 0; % Need to redefine within parfor loop

          % Need to rewrite the whole thing here and create a sliced cell or
          % structure that can be used on parfor without using eval and
          % sprintf to retrieve variables.
          %unit
           tmp=[]; tmp=TMunion_M1(covmat(sigpairs(i),1),:);
  %         strtmp1=sprintf('elec%03d_%03d',tmp(1),tmp(2));
  %         SpikeData1=sprintf('M1spk_MB_%s',strtmp1);
  %         %lfp
  %         tmp=[];tmp=TMunion_S1(covmat(sigpairs(i),2),:);
  %         strtmp2=sprintf('elec%03d_%03d',tmp(1),tmp(2));
  %         SpikeData2=sprintf('S1spk_MB_%s',strtmp2);
  %         
  %         eval(sprintf('SpikeData1=%s;',SpikeData1));
  %         eval(sprintf('SpikeData2=%s;',SpikeData2));
          SpikeData1_1=extractdatapt(squeeze(SpikeData1(tmp(1),tmp(2),1:100)),[Tm(peaktm(i))-0.25 Tm(peaktm(i))+0.25]);

Answer 5

[Joseph MacGregor]

I am receiving the same error message as the questioner(s) in R2015a for reasons that I don't understand. The answers imply to me that all variables must be explicitly assigned to the pool, but in my 2+ years experience with parfor this was never necessary, so I'm confused as to what changed.

I start the local pool and then initialize a bunch of variables. Because there are so many to initialize, I use a cell containing their names and eval in a for loop to generate them. Because of that approach, I can see why parfor might not recognize them following the above discussion, but it worked fine in R2014b. The same error occurs if I move the pool start until after those initializations. I am in no mood to explicitly call out every variable I need in the parfor loop, given that I never had to before.

Any thoughts? Thanks.

[Walter Roberson]

parfor and eval are essentially incompatible.

I suggest you use dynamic field names in a struct.

Answer 6

[BHUSHAN MUTHIYAN]

Hello,

I am facing similar issue accessing the out variable.

Can some one please help me. I have a function which returns "out" as the output%%

function out = flexible(filter, bias, rowS, input, convModel, sumModel) %#codegen

colS = rowS;

ins = [size(input,1),size(input,2),size(input,3),size(input,4)];

filter = rot90(filter, 2);

fis = [size(filter,1),size(filter,2),size(filter,3),size(filter,4)];

bis = [size(bias,1),size(bias,2),size(bias,3),size(bias,4)];

dimx = ins(1) - fis(1) + 1;

dimy = ins(2) - fis(2) + 1;

dimfox = 1+floor((dimx-1)/rowS);

dimfoy = 1+floor((dimy-1)/colS);

out = zeros(dimfox, dimfoy, fis(4), ins(4), 'like', sumModel);

frac = 32 - (convModel.WordLength - convModel.FractionLength);

H = vision.Convolver('OutputSize', 'Valid', 'CustomProductDataType', numerictype([],32,frac), 'CustomOutputDataType', numerictype([],32,frac));

parfor im = 1:ins(4)

      carve = zeros(dimfox, dimfoy, 'like', sumModel);

      for f = 1:fis(4)

          for ch = 1:fis(3)

              carve(:) = carve + step(H, input(:,:,ch,im), filter(:,:,ch,f));

          end

          carve(:) = carve + bias(1, f);

          out(:,:,f,im) = carve;

      end

end

end

It gives me below error:

The variable out in a parfor cannot be classified.

See Parallel for Loops in MATLAB, "Overview".

Use help codegen for more information on using this command.

Error using codegen

Also, can someone let me know if we can change value of global variables inside parfor loop?

[Walter Roberson]

You initialize carve in the im loop, and you accumulate into it in the f and ch loops. You write the current value of carve into out in the f loop. That means that you are essentially writing partial calculations into out. Are you sure that is what you want??

fis3 = fis(3);
fis4 = fis(4);

parfor im = 1:ins(4)

        carve = zeros(dimfox, dimfoy, fis4, 'like', sumModel);

        for f = 1:fis4
            if f ~= 1
              carve(:, :, f) = carve(:, :, f-1);
            end
            for ch = 1:fis3
                carve(:, :, f) = carve(:, :, f) + step(H, input(:,:,ch,im), filter(:,:,ch,f));
            end
            carve(:, :, f) = carve(:, :, f) + bias(1, f);
        end
        out(:,:,:,im) = carve;
  end