Dev Diary May 4th, 2011

      So, I got a job working as an intern working for a company by the name of Data In Motion as of this last Monday.  Not much to say about it as of yet except that there's a ton to the job, and I am currently... whelmed at the tasks ahead.  That is to say, it's certainly daunting, but I think if I take it one piece at a time (like one of my favorite Johnny Cash songs) I'll get to understanding it, and be able to make myself useful to my employers.  For the time being, they are doing a lot more for me than I am for them, and it is currently just a hope to be a good return on that investment.

     All that having been said, I probably won't be sharing my code from work-related projects here on my blog, unless given explicit permission to do so.  I am also starting a class for developing android applications, and may share some of the toy programs from that class (albeit if it looks even marginally profitable I'd be keeping mum).  So that's the situation if you were wondering when my next interesting tidbit of code would appear.

     And now I'm off to spend the rest of the evening getting used to some new equipment.  And to read up on marker, visitor, and singleton design patterns.  On a complete sidenote, I should remind myself to strongly suggest to my professors here at Westminster to implement a class focused on design patterns.

SVM Tester

function [score iter correct falseNeg falsePos unsures] = TestVectorMachine( weight, kernel, labels, bias )

%Test function for vector machien -- takes in a set of training data and
%its labels and tests a given set of alphas and a bias against it
[xn xm] = size(kernel);
[yn ym] = size(labels);
[an am] = size(weight);

%if(xn ~= am || ym ~= xn)
  %  display('Sorry, this is an idiot proof function.  Try again!');
 %   return;
%end
falseNeg = 0;
falsePos = 0;
unsures = 0;
iter = 0;
for i = 1:xn
    fXi = (weight .* labels) * kernel(i,:) + bias;
    if (fXi * labels(i))  <= 0
        if(fXi > 0)
            falsePos = falsePos + 1;
        end
       
        if(fXi < 0)
            falseNeg = falseNeg + 1;
        end
       
        if(fXi == 0)
            unsures = unsures + 1;
        end
       
        iter = iter + 1;
    end
end

score = (xn - iter) / xn * 100;
correct = xm - iter;

Soft Margin one-norm SVM

function [ weights bias ] = TannSchmidVectorMachineSoftMarginUno( K, H, labels, C)
%Toggle details which kernel we use

[xm xn] = size(K);
[ym yn] = size(labels);

%scale C down
C = C / xm;

%check to make sure training & labels have same dimension and toggle is
%valid

if xm ~= ym
    display('Sorry, this is an idiot proof function. Try feeding in valid parameters next time, doof!');
    return;
end

%allocate space for different parts
f = zeros(xm, 1);
A = zeros(2 * xm + 4, xm);
b = zeros(2 * xm + 4, 1);

%build constraints matrix
A(1,:) = labels';
A(2,:) = -labels';
A(3,:) = ones(1, xm);
A(4,:) = -ones(1, xm);
for i = 1:xm
    A(i+4, i) = 1;
end
for i = 1:xm
    A(i+4+xm, i) = -1;
end

b = [0; 0; 1; -1; (C - 10^(-7)) * ones(xm,1); zeros(xm, 1)];
          
[weights v] = quadprog(H, f, A, b);

%find the bias
bias = GetSoftBiasUno(weights, K, labels, C);
bias = bias / sqrt(weights' * H * weights);

save('recordedResults0', 'weights', 'bias', 'K');

Soft Margin 2-norm SVM

function [ weights bias ] = TannSchmidVectorMachineSoftMarginDos( K, H, labels, C)
%Toggle details which kernel we use

[xm xn] = size(K);
[ym yn] = size(labels);

%scale C down
C = C / xm;
H = (1/2) * (H + (1 / C) * eye(xm, xn));

%check to make sure training & labels have same dimension and toggle is
%valid

if xm ~= ym
    display('Sorry, this is an idiot proof function. Try feeding in valid parameters next time, doof!');
    return;
end

%allocate space for different parts
f = zeros(xm, 1);
A = zeros(xm + 4, xm);
b = zeros(xm + 4, 1);

%build constraints matrix
A(1,:) = labels';
A(2,:) = -labels';
A(3,:) = ones(1, xm);
A(4,:) = -ones(1, xm);
for i = 1:xm
    A(i+4, i) = -1;
end

b = [0; 0; 1; -1; zeros(xm, 1)];
          
[weights v] = quadprog(H, f, A, b);

%find the bias
bias = GetSoftBiasDos(weights, K, labels, C);
bias = bias / sqrt(weights' * H * weights);

save('recordedResults0', 'weights', 'bias', 'K');

Hard Margin one-norm SVM

function [ weights bias ] = TannSchmidVectorMachineHardMarginUno( K, H, labels)
%Toggle details which kernel we use

[xm xn] = size(K);
[ym yn] = size(labels);

%check to make sure training & labels have same dimension and toggle is
%valid

if xm ~= ym
    display('Sorry, this is an idiot proof function. Try feeding in valid parameters next time, doof!');
    return;
end

%allocate space for different parts
f = -ones(xm, 1);
A = zeros(xm +2, xm);
bias = zeros(xm +2, 1);

%build constraints matrix
A(1,:) = labels';
A(2,:) = -labels';
for i = 1:xm
    A(i+2, i) = -1;
end
          
[weights v] = quadprog(H, f, A, bias);

%find the bias
bias = getHardMarginBias(weights, K, labels);

save('recordedResults0', 'weights', 'bias', 'K');

Soft Margin one-norm Bias Calculator

function [ bias ] = GetSoftBiasUno( weights, kernel, labels, C)
[xm xn] = size(kernel);
counter = 0;
bias = 0;

for i = 1:xm
    if weights(i) > (10^-10) && weights(i) < (C - 10^(-10))  %calculate first <w xi>
        sgnLastY = labels(i) > 0;
        partialSum = 0;
        for j = 1:xm
            partialSum = partialSum + labels(j) * kernel(i,j) * weights(j);
        end
       
        %reset partial sum
        wXi = partialSum;
        partialSum = 0;
       
        for j = i:xm
            if weights(i) > (10^-10) && weights(i) < (C - 10^(-10)) && sgnLastY ~= (labels(j) > 0)
              for j = 1:xm
                 partialSum = partialSum + labels(j) * kernel(i,j) * weights(j);
              end
             
              %save second <w xj>
              wXj = partialSum;
              bias = bias + -(wXi + wXj) / 2;
              counter = counter + 1;
            end
        end
    end
end

bias = bias / counter;

Soft-Margin 2-norm Bias Calculator

function [ bias ] = GetSoftBiasDos( weights, kernel, labels, C)
[xm xn] = size(kernel);
counter = 0;
bias = 0;

for i = 1:xm
    if weights(i) > (10^-10)  %calculate first <w xi>
        sgnLastY = labels(i) > 0;
        partialSum = 0;
        for j = 1:xm
            partialSum = partialSum + labels(j) * kernel(i,j) * weights(j);
        end
       
        %reset partial sum
        wXi = partialSum;
        partialSum = 0;
       
        for j = i:xm
            if weights(i) > (10^-10) && sgnLastY ~= (labels(j) > 0)
              for j = 1:xm
                 partialSum = partialSum + labels(j) * kernel(i,j) * weights(j);
              end
             
              %save second <w xj>
              wXj = partialSum;
              bias = bias + -(wXi + wXj) / 2 - (labels(i) * weights(i) - labels(j) * weights(j)) / (2 * C);
              counter = counter + 1;
            end
        end
    end
end

bias = bias / counter;

Hard Margin Bias Calculator

function [ bias ] = getHardMarginBias(weights, kernel, labels)
%returns the bias
[xm xn] = size(kernel);
counter = 0;
bias = 0;
for i = 1:xm
    if weights(i, 1) > 0.000000000001
        partialSum = 0;
        for j = 1:xm
            partialSum = partialSum + label(j) * kernel(i,j) * weights(j);
        end
        bias = bias + labels(i) - partialSum;
        counter = counter + 1;
    end
end

bias = bias / counter;

end

Gaussian Kernel

function [ result ] = GaussKernel( x, y, sigma )
result = norm(x - y)^2;
result = result / sigma;
result = exp(-result);

end

Default Kernel

function [ result ] = defaultKernel(x, y, A)
%One of many kernel functions.  Takes vectors x and y, returns kernel
%function as a dot product using a positive definite matrix A
[xm xn] = size(x);
[ym yn] = size(y);
[R isPosDef] = chol(A);
if isPosDef ~= 0 || xm ~= ym || xn ~= yn || xn == 1
    disp('sorry, this function is idiot proof.  Please enter in a positive definite matrix A');
    result = -1;
    return;
end

result = x * (A * y');