addpath('/noback/lowbow/translation/svm/helper/');

% --- Generate some toy training values ---------------------------------------
numclasses = 2;
numclassinst = 25;
Mv=100;
Nd = numclasses*numclassinst;
maxdoclength = 60;

theta = normalizeRows(rand(numclasses,Mv))';
X = zeros(Mv,Nd);
z = zeros(numclasses,Nd);
for ii=1:numclasses
    for jj=1:numclassinst
        iter = (ii-1)*numclassinst+jj;
        z(ii,iter) = 1;
        %X(:,iter) = multinomial_sample(ceil(maxdoclength*rand),Mv,theta(:,ii),randseed);
        X(:,iter) = multsamp(ceil(maxdoclength*rand),theta(:,ii));
    end
end
I = randperm(Nd);   % not really necessary to randperm
z = [-1 1]*z(:,I);  % convert cannonical labeling schema for SVM binary classification
X = X(:,I);

I=find(sum(X,2)==0);
Mv=Mv-length(I);
X(I,:)=[];




% --- Create and save Kernel --------------------------------------------------
c = 1;
t = 0.96;
gamma = 1/(2*1^2);
T = translation(X,c,t);
[ED2 A] = transD2(X,T);
K = exp(-gamma * ED2);
unit = ones(size(K))/size(K,1);
% centering in feature space!
K_n = K - unit*K - K*unit + unit*K*unit;

numPrinComp = 3;
[evecs,evals] = eig(K_n);
evals = real(diag(evals));
princomp = evecs(:,1:numPrinComp) * diag(evals(1:numPrinComp).^-0.5);


% in the real world, we would make a new kernel with numtest rows and numtrain
% cols and premultiply it with princomp
%
%     [ Krr | Kre ]
% K = [-----+-----]
%     [ Ker | Kee ]
%
%   where r denotes training and e denotes testing
% Ker would need to be normalized prior to: pX = Ker_n * princomp
pX = K_n*princomp;

%figure('vis', 'off');
	clf;
	I=find(z==1);
	plot3(pX(I,1),pX(I,2),pX(I,3),'b.');
	hold on;
	I=find(z==-1);
	plot3(pX(I,1),pX(I,2),pX(I,3),'r.');
	grid on;
%saveas(gcf, './kpca', 'png');