dualgreedypcatrain.m.svn-base

a function inside machine learning

function [subspaceInfo, trainInfo] = dualGreedyPCATrain(trainData, params)
%A kernel greedy version of PCA 

if (nargin ~= 2)
    fprintf('%s\n', help(sprintf('%s', mfilename)));
    error('Incorrect number of inputs - see above usage instructions.');
end

[nameX, nameY] = getSpaceNames(params); 
[numTrainExamples, numFeatures] = getDataFieldSize(trainData, nameX);

%Store parameters
T = params.iterations; 
kernelFunction = char(params.X.kernel.name); 
kernelParams = params.X.kernel.params; 

if isfield(params.X, 'kernelCols') 
    kernelCols = params.X.kernelCols; 
else 
    kernelCols = numTrainExamples;
end 

trainKj = feval(kernelFunction, getDataFieldValue(trainData, nameX), getDataFieldValue(trainData, nameX), kernelParams);
b = sparse(numTrainExamples, T); 
tau = zeros(numTrainExamples, T); 

%Do the training 
for i=1:T 
    displayCount(i, T, 30);
    
    %Find dual projection vector
    b(:, i) = dualMaxSparseVariance2(trainKj, trainKj, 0, 0, kernelCols);
    tau(:, i) = trainKj'*b(:, i);
    
    %Now, deflate K 
    trainKj = trainKj - tau(:, i)*tau(:, i)'; 
end 

%trainKj should be symmetric - check (it is)
trainInfo.data = data; 
trainInfo.data = addDataField(trainInfo.data, 'X', tau, 'examples'); 

subspaceInfo = struct; 
subspaceInfo.b = b; 
subspaceInfo.tau = tau;