function Svmclassification

% Example MultiClass SVM Classifiction

% "One against Others"

%

%

%

clc;

clear all;

close all;





%--------------------------------------------------

%n=;

sigma=1;



% Location of the data.

subjects = {'../mock-data/wesubj1.mat','../mock-data/wesubj2.mat','../mock-data/wesubj3.mat','../mock-data/wesubj4.mat','../mock-data/wesubj5.mat'};



%% Compute the Feature Vectors.

% Create the features, for each activity, for

% each sample for each subject.

for subj = 1:5

    load(subjects{subj});

    feat = [];

    label = [];

    for activity =  1: 5

        for i = 1 : actionsample{activity}.tsamples            

            temp = [];

            

                temp = [temp computestatfeatures([actionsample{activity}.sample{i}.acc1,actionsample{activity}.sample{i}.acc2])];

                      

                %temp = [temp computefirstderivativehist(actionsample{activity}.sample{i}.acc1)];                

                %temp = [temp computerawhist(actionsample{activity}.sample{i}.acc1)];

                

            feat = [feat;temp];

            label = [label; activity];

        end       

    end

    save(sprintf('feat%d.mat',subj),'feat','label');

end



% Perform a 5 fold cross validation on the feature vectors (histograms)

% using SVM.

% Variable for storing the confusion matrix.

cfmatrix = zeros(5);

%% SVM Classification.

for fold = 1:5

%     if fold == 4

%         continue;

%     end;

    % Create the train and test data set.

    trdata = [];

    trlabel = [];

    tstdata = [];

    tstlabel = [];

    trdata_norm= [];

    tstdata_norm= [];

    for j = 1 : 5

        load(sprintf('feat%d.mat',j));

        % Subject Independent

        if fold == j

            tstdata = feat;

            tstlabel = label;

          else

            trdata = [trdata;feat];

            trlabel = [trlabel;label];

        end       

    end



  %Normalize the trdata and tstdata before running SVM training

  

    for i= 1: 39   %size(trdata,2)

        trdata_norm(:,1)=(trdata(:,1)-min(trdata(:,1)))/(max(trdata(:,1))-min(trdata(:,1)));

        if (max(trdata(:,i))-min(trdata(:,i)))~= 0

        trdata_norm(:,i)= (trdata(:,i)-min(trdata(:,i)))/(max(trdata(:,i))-min(trdata(:,i)));

        else

            trdata_norm(:,i)= trdata(:,i);

        end

    end



    for i= 1: 39  %size(tstdata,2)

        if (max(tstdata(:,i))-min(tstdata(:,i)))~= 0

        tstdata_norm(:,i)= (tstdata(:,i)-min(trdata(:,i)))/(max(trdata(:,i))-min(trdata(:,i)));

        else

          tstdata_norm(:,i)= tstdata(:,i);  

        end

    end



    %xapp=trdata_norm;

    xapp=trdata_norm;

    yapp= trlabel;



    nbclass=5;

    [n1, n2]=size(xapp);



    %-----------------------------------------------------

    %   Learning and Learning Parameters

    c = 1;

    lambda = 1e-7;

    kerneloption= 2;

    kernel='gaussian';

    verbose =0;



    % % %---------------------One Against All algorithms----------------

    nbclass=5;

    [xsup,w,b,nbsv]=svmmulticlassoneagainstall(xapp,yapp,nbclass,c,lambda,kernel,kerneloption,verbose);

    [ypred_trn] = svmmultival(xapp,xsup,w,b,nbsv,kernel,kerneloption);

    fprintf( '\nRate of correct class in training data : %2.2f \n',100*sum(ypred_trn==yapp)/length(yapp)); 



    %-------------------------------------------------------

    % Testing

    %-------------------------------------------------------

    xtest=tstdata_norm;



    [ypred,maxi] = svmmultival(xtest,xsup,w,b,nbsv,kernel,kerneloption);

    accuracy(fold)=100*sum(ypred==tstlabel)/length(tstlabel);

    tcmatrix = computeconfusionmatrix(tstlabel,ypred);

    cfmatrix = cfmatrix + tcmatrix;

end

%% Display the results.

% Display Accuracy

disp('Results with SVM');

disp(sprintf('Accuracy over 5 folds'));

disp(accuracy);

disp(sprintf('Overall Accuracy over 5 folds %f ',(sum(accuracy)/5)));

% Display Confusion Matrix

disp(sprintf('Confusion Matrix'));

disp(cfmatrix);



function cfmatrix = computeconfusionmatrix(ytest,yest)



    % this function computes the confusion matrix, given the estimated and

    % actual labels.

    cfmatrix = zeros(max(ytest),max(ytest));

    for i = 1 : size(ytest,1)

        cfmatrix(ytest(i),yest(i)) = cfmatrix(ytest(i),yest(i)) + 1;

    end    

return;



function derivhist = computefirstderivativehist(data)

    % centers for the histogram

    centers = -50:50;

    l = data(2:end,:);

    d = data(1:end-1,:)-l;

    histx = hist(d(:,1),centers);

    histy = hist(d(:,2),centers);

    histz = hist(d(:,3),centers);

    derivhist = [histx/sum(histx) histy/sum(histy) histz/sum(histz)];

return;



function rawhist = computerawhist(data)

    % centers for the histogram

    % X - Axis

    centersx = 100:450;

    % Y - Axis

    centersy = 0:700;

    % Z - Axis

    centersz = 250:600;

    histx = hist(data(:,1),centersx);

    histy = hist(data(:,2),centersy);

    histz = hist(data(:,3),centersz);

    rawhist = [histx/sum(histx) histy/sum(histy) histz/sum(histz)];

return;



function feature = computestatfeatures(data)

    fnorm=true;



    % mean of the data

    m = mean(data);



    % variance of the data

    v = var(data);



    % correlation between the axis.

    c = corr(data);



    % absolute value of the fourier transform.

    fftdata = abs(fft(data));



    % remove the mean value.

    fftdata = fftdata(2:end,:);



    % energy.

    e = sum(fftdata.^2)/size(fftdata,1);



    % spectral entropy

    fftdata = 1 + fftdata/size(data,1);

    temp = [fftdata(:,1)/sum(fftdata(:,1)), fftdata(:,2)/sum(fftdata(:,2)),...

                                            fftdata(:,3)/sum(fftdata(:,3)),...

                                            fftdata(:,4)/sum(fftdata(:,4)),...

                                            fftdata(:,5)/sum(fftdata(:,5)),...

                                            fftdata(:,6)/sum(fftdata(:,6))];

    ent = -sum(temp.*log(temp),1);

    feature = [m v c(1,2:6) c(2,3:6) c(3,4:6) c(4,5:6) c(5,6) e ent];

return;