private/som_subject_batch.m (916fe4f) - SVMCrossVal.git

som_subject_batch.m

%% subject loop
function decode = som_subject_batch(header,subjectdata,somOpts)

addpath 'somtoolbox2';

RANDOMIZE_DATAPOINTS = 1;

decode = struct;
decode.decodePerformance = [];
decode.rawTimeCourse     = [];

nSubjects = numel(subjectdata);
disp(sprintf('batch processing %u subjects',nSubjects));

timeline = header.timeline;
timeline.frameShiftStart = header.frameShift.frameShiftStart;
timeline.frameShiftEnd   = header.frameShift.frameShiftEnd;
timeline.decodeDuration  = header.frameShift.decodeDuration;

timeLineStart   = timeline.frameShiftStart;
timeLineEnd     = timeline.frameShiftEnd;

% TimePointMatrix
for subjectDataID = 1:nSubjects
    currentSubject = subjectdata{subjectDataID};
    timePointArgs.pst           = currentSubject.pst;
    timePointArgs.labelMap      = LabelMap(header.classDef.labelCells,header.classDef.conditionCells);
    timePointArgs.eventList     = header.classDef.eventMatrix;

timePointMatrix{subjectDataID} = buildTimePointMatrix(timeline,timePointArgs);

decode.rawTimeCourse = [decode.rawTimeCourse currentSubject.pst];

display(sprintf('%u -fold cross validation for %u timeslices.\n',somOpts.nfold,size(1:timeLineEnd-timeLineStart+1,2)));

for timeIndex = 1:timeLineEnd-timeLineStart+1
        svmstruct = calculateSVMTables(timePointMatrix{subjectDataID},timeIndex);
        nElements = length(svmstruct.svmlabel);

if RANDOMIZE_DATAPOINTS
            rndindex  = randperm(nElements);
            svmstruct.svmdata   = svmstruct.svmdata(rndindex,:);
            svmstruct.svmlabel  = svmstruct.svmlabel(rndindex);
        end

chunksize = nElements / somOpts.nfold;
        
        cross_value = 0;
        for iFold = 1:somOpts.nfold
            chunkstart = (iFold-1)*chunksize+1;
            chunkend   = min(iFold*chunksize,nElements);

svm_train_label = svmstruct.svmlabel;
            svm_train_data  = svmstruct.svmdata;

svm_validation_label = svmstruct.svmlabel(chunkstart:chunkend);
            svm_validation_data  = svmstruct.svmdata(chunkstart:chunkend,:);

svm_train_label(chunkstart:chunkend) = []; %del test set
            svm_train_data(chunkstart:chunkend,:) = [];% del test set

[sD sM] = som_train(svm_train_label, svm_train_data, somOpts);

performance = som_decode(sM, svm_validation_data,svm_validation_label);

cross_value = [cross_value, performance];
        end
        decode.decodePerformance = [decode.decodePerformance; cross_value];
    end
end
display('decode done');
end

SVMCrossVal.git

SOM Single run