Daniel@0: function Ypredict = soft_classify(W, test_k, Xtrain, Ytrain, Xtest, Testnorm)
Daniel@0: % Ypredict = soft_classify(W, test_k, Xtrain, Ytrain, Xtest, Testnorm)
Daniel@0: %
Daniel@0: %       W       = d-by-d positive semi-definite matrix
Daniel@0: %       test_k  = k-value to use for KNN
Daniel@0: %       Xtrain  = d-by-n matrix of training data
Daniel@0: %       Ytrain  = n-by-1 vector of training labels
Daniel@0: %       Xtest   = d-by-m matrix of testing data
Daniel@0: %       Testnorm= m-by-#kernels vector of k(i,i) for each i in test
Daniel@0: %
Daniel@0: 
Daniel@0:     addpath('cuttingPlane', 'distance', 'feasible', 'initialize', 'loss', ...
Daniel@0:             'metricPsi', 'regularize', 'separationOracle', 'util');
Daniel@0: 
Daniel@0:     [d, nTrain, nKernel] = size(Xtrain);
Daniel@0:     nTest       = size(Xtest, 2);
Daniel@0:     test_k      = min(test_k, nTrain);
Daniel@0: 
Daniel@0:     if nargin < 7
Daniel@0:         Testnorm = [];
Daniel@0:     end
Daniel@0: 
Daniel@0:     % Build the distance matrix
Daniel@0:     [D, I] = mlr_test_distance(W, Xtrain, Xtest, Testnorm);
Daniel@0: 
Daniel@0:     % Compute label agreement
Daniel@0:     Ypredict  = histc(Ytrain(I(1:test_k,:)), unique(Ytrain)');
Daniel@0: 
Daniel@0: end
Daniel@0: 
Daniel@0: 
Daniel@0: function [D,I] = mlr_test_distance(W, Xtrain, Xtest, Testnorm)
Daniel@0: 
Daniel@0:     % CASES:
Daniel@0:     %   Raw:                        W = []
Daniel@0:     
Daniel@0:     %   Linear, full:               W = d-by-d
Daniel@0:     %   Single Kernel, full:        W = n-by-n
Daniel@0:     %   MKL, full:                  W = n-by-n-by-m
Daniel@0: 
Daniel@0:     %   Linear, diagonal:           W = d-by-1
Daniel@0:     %   Single Kernel, diagonal:    W = n-by-1
Daniel@0:     %   MKL, diag:                  W = n-by-m
Daniel@0:     %   MKL, diag-off-diag:         W = m-by-m-by-n
Daniel@0:     
Daniel@0:     [d, nTrain, nKernel] = size(Xtrain);
Daniel@0:     nTest = size(Xtest, 2);
Daniel@0: 
Daniel@0:     if isempty(W)
Daniel@0:         % W = []  => native euclidean distances
Daniel@0:         D = mlr_test_distance_raw(Xtrain, Xtest, Testnorm);
Daniel@0: 
Daniel@0:     elseif size(W,1) == d && size(W,2) == d
Daniel@0:         % We're in a full-projection case
Daniel@0:         D = setDistanceFullMKL([Xtrain Xtest], W, nTrain + (1:nTest), 1:nTrain);
Daniel@0: 
Daniel@0:     elseif size(W,1) == d && size(W,2) == nKernel
Daniel@0:         % We're in a simple diagonal case
Daniel@0:         D = setDistanceDiagMKL([Xtrain Xtest], W, nTrain + (1:nTest), 1:nTrain);
Daniel@0: 
Daniel@0:     elseif size(W,1) == nKernel && size(W,2) == nKernel && size(W,3) == nTrain
Daniel@0:         % We're in DOD mode
Daniel@0:         D = setDistanceDODMKL([Xtrain Xtest], W, nTrain + (1:nTest), 1:nTrain);
Daniel@0: 
Daniel@0:     else
Daniel@0:         % Error?
Daniel@0:         error('Cannot determine metric mode.');
Daniel@0: 
Daniel@0:     end
Daniel@0:     
Daniel@0:     D       = full(D(1:nTrain, nTrain + (1:nTest)));
Daniel@0:     [v,I]   = sort(D, 1);
Daniel@0: end
Daniel@0: 
Daniel@0: 
Daniel@0: 
Daniel@0: function D = mlr_test_distance_raw(Xtrain, Xtest, Testnorm)
Daniel@0: 
Daniel@0:     [d, nTrain, nKernel] = size(Xtrain);
Daniel@0:     nTest = size(Xtest, 2);
Daniel@0: 
Daniel@0:     if isempty(Testnorm)
Daniel@0:         % Not in kernel mode, compute distances directly
Daniel@0:         D = 0;
Daniel@0:         for i = 1:nKernel
Daniel@0:             D = D + setDistanceDiag([Xtrain(:,:,i) Xtest(:,:,i)], ones(d,1), ...
Daniel@0:                                     nTrain + (1:nTest), 1:nTrain);
Daniel@0:         end
Daniel@0:     else
Daniel@0:         % We are in kernel mode
Daniel@0:         D = sparse(nTrain + nTest, nTrain + nTest);
Daniel@0:         for i = 1:nKernel
Daniel@0:             Trainnorm = diag(Xtrain(:,:,i));
Daniel@0:             D(1:nTrain, nTrain + (1:nTest)) = D(1:nTrain, nTrain + (1:nTest)) ...
Daniel@0:                 +  bsxfun(@plus, Trainnorm, bsxfun(@plus, Testnorm(:,i)', -2 * Xtest(:,:,i)));
Daniel@0:         end
Daniel@0:     end
Daniel@0: end
Daniel@0: