Parfor: Variable indexed in different ways.

I want to parallelize the following for loop but the I have the following problem: In a parfor loop variable 'perf' is indexed in different ways

The problem is in the 3rd line from the bottom because I index in a different way the matrix perf.

How can I avoid this problem?

NET = cell(5,11);
TR = cell(5,11);

ave_perfs = zeros(1,11);
perf = zeros(5,11);

profile on;
parfor j = 1:11
    for bin = 1:5
        processed_train_Ext = load([ExtensionTrainPath,'/',Features{j},'/','train',num2str(bin),'_processed_Ext'];
        processed_train_Flx = load([FlexionTrainPath,'/',Features{j},'/','train',num2str(bin),'_processed_Flx']);
        processed_test_Ext = load([ExtensionTestPath,'/',Features{j},'/','test',num2str(bin),'_processed_Ext']);
        processed_test_Flx = load([FlexionTestPath,'/',Features{j},'/','test',num2str(bin),'_processed_Flx']);
          % concatenating the training data
        train_data = [processed_train_Ext.finalMatrix(:,1:end-1); processed_train_Flx.finalMatrix(:,1:end-1)];
        true_train_labels = [processed_train_Ext.finalMatrix(:,end); processed_train_Flx.finalMatrix(:,end)];

          %concatenating the test data
        test_data = [processed_test_Ext.finalMatrix(:,1:end-1) ; processed_test_Flx.finalMatrix(:,1:end-1)];
        true_test_labels = [processed_test_Ext.finalMatrix(:,end); processed_test_Flx.finalMatrix(:,end)];

          %fit network
         net = fitnet(5);
         net.divideParam.trainRatio = 0.85;
         net.divideParam.valRatio = 0.15;
         net.divideParam.testRatio = 0.0;

          % train and validate train data
         [net, tr] = train(net,train_data',true_train_labels','useParallel','yes');
         NET{bin,j} = net;
         TR{bin,j} = tr;

          % test newly trained network with test data
         outputs = NET{bin,j}(test_data','useParallel','yes');

         % calculate accuracy on the test data
         perf(bin,j) = 1 - mse(net, true_test_labels, outputs);

         toc
         beep on; beep;
     end

     % calculate averages of the errors
     ave_perfs(1,j) = sum(perf(:,j))/5;
 end

[Greg]

Not sure it will help the parfor (but it might actually), but 2 things:

% calculate averages of the errors
ave_perfs(1,j) = sum(perf(:,j))/5;

1) Why calculate a mean manually? Use the mean function.

2) Why do it inside the loop?

Try putting this outside the parfor:

% calculate averages of the errors
ave_perfs = mean(perf,1);

Then you wouldn't even need to pre-allocate ave_perfs at the top.

[OCDER]

Greg, your comment should be in the answer section.

Ioannis, the issue is caused because you are accessing perf via perf(bin,j) AND perf(:,j) within the parfor loop. Matlab can't determine how to slice perf by column or by element, giving that error. By moving the ave_perf = mean(perf,1) outside the parfor, it's now clear perf can be sliced by element and thus no more error occurs.

[Greg]

As I stated in the first line, I wasn't sure it was actually an answer. I was hoping Ioannis would respond if it did indeed solve the problem, at which point I can move it to answer.

Thank you for clarifying that it was the source of the problem.

[OCDER]

I actually didn't know what the issue was until I tried your solution, which did the trick for me.

[Greg]

Ahh, fair enough. My MATLAB machine is not my internet machine, so I'm rarely able to test the poster's code or my own suggestions.

I'm glad we got it solved and an accepted answer. Thanks for the assist.

Answer 1

[Greg]

Try replacing this code (inside the loop):

% calculate averages of the errors
ave_perfs(1,j) = sum(perf(:,j))/5;

With this code (outside the parfor):

% calculate averages of the errors
ave_perfs = mean(perf,1);

Then you could even remove the pre-allocation of ave_perfs at the top.