Mercurial > hg > camir-aes2014

annotate toolboxes/distance_learning/mlr/util/soft_classify.m @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 function Ypredict = soft_classify(W, test_k, Xtrain, Ytrain, Xtest, Testnorm)
wolffd@0 2 % Ypredict = soft_classify(W, test_k, Xtrain, Ytrain, Xtest, Testnorm)
wolffd@0 3 %
wolffd@0 4 % W = d-by-d positive semi-definite matrix
wolffd@0 5 % test_k = k-value to use for KNN
wolffd@0 6 % Xtrain = d-by-n matrix of training data
wolffd@0 7 % Ytrain = n-by-1 vector of training labels
wolffd@0 8 % Xtest = d-by-m matrix of testing data
wolffd@0 9 % Testnorm= m-by-#kernels vector of k(i,i) for each i in test
wolffd@0 10 %
wolffd@0 11
wolffd@0 12 addpath('cuttingPlane', 'distance', 'feasible', 'initialize', 'loss', ...
wolffd@0 13 'metricPsi', 'regularize', 'separationOracle', 'util');
wolffd@0 14
wolffd@0 15 [d, nTrain, nKernel] = size(Xtrain);
wolffd@0 16 nTest = size(Xtest, 2);
wolffd@0 17 test_k = min(test_k, nTrain);
wolffd@0 18
wolffd@0 19 if nargin < 7
wolffd@0 20 Testnorm = [];
wolffd@0 21 end
wolffd@0 22
wolffd@0 23 % Build the distance matrix
wolffd@0 24 [D, I] = mlr_test_distance(W, Xtrain, Xtest, Testnorm);
wolffd@0 25
wolffd@0 26 % Compute label agreement
wolffd@0 27 Ypredict = histc(Ytrain(I(1:test_k,:)), unique(Ytrain)');
wolffd@0 28
wolffd@0 29 end
wolffd@0 30
wolffd@0 31
wolffd@0 32 function [D,I] = mlr_test_distance(W, Xtrain, Xtest, Testnorm)
wolffd@0 33
wolffd@0 34 % CASES:
wolffd@0 35 % Raw: W = []
wolffd@0 36
wolffd@0 37 % Linear, full: W = d-by-d
wolffd@0 38 % Single Kernel, full: W = n-by-n
wolffd@0 39 % MKL, full: W = n-by-n-by-m
wolffd@0 40
wolffd@0 41 % Linear, diagonal: W = d-by-1
wolffd@0 42 % Single Kernel, diagonal: W = n-by-1
wolffd@0 43 % MKL, diag: W = n-by-m
wolffd@0 44 % MKL, diag-off-diag: W = m-by-m-by-n
wolffd@0 45
wolffd@0 46 [d, nTrain, nKernel] = size(Xtrain);
wolffd@0 47 nTest = size(Xtest, 2);
wolffd@0 48
wolffd@0 49 if isempty(W)
wolffd@0 50 % W = [] => native euclidean distances
wolffd@0 51 D = mlr_test_distance_raw(Xtrain, Xtest, Testnorm);
wolffd@0 52
wolffd@0 53 elseif size(W,1) == d && size(W,2) == d
wolffd@0 54 % We're in a full-projection case
wolffd@0 55 D = setDistanceFullMKL([Xtrain Xtest], W, nTrain + (1:nTest), 1:nTrain);
wolffd@0 56
wolffd@0 57 elseif size(W,1) == d && size(W,2) == nKernel
wolffd@0 58 % We're in a simple diagonal case
wolffd@0 59 D = setDistanceDiagMKL([Xtrain Xtest], W, nTrain + (1:nTest), 1:nTrain);
wolffd@0 60
wolffd@0 61 elseif size(W,1) == nKernel && size(W,2) == nKernel && size(W,3) == nTrain
wolffd@0 62 % We're in DOD mode
wolffd@0 63 D = setDistanceDODMKL([Xtrain Xtest], W, nTrain + (1:nTest), 1:nTrain);
wolffd@0 64
wolffd@0 65 else
wolffd@0 66 % Error?
wolffd@0 67 error('Cannot determine metric mode.');
wolffd@0 68
wolffd@0 69 end
wolffd@0 70
wolffd@0 71 D = full(D(1:nTrain, nTrain + (1:nTest)));
wolffd@0 72 [v,I] = sort(D, 1);
wolffd@0 73 end
wolffd@0 74
wolffd@0 75
wolffd@0 76
wolffd@0 77 function D = mlr_test_distance_raw(Xtrain, Xtest, Testnorm)
wolffd@0 78
wolffd@0 79 [d, nTrain, nKernel] = size(Xtrain);
wolffd@0 80 nTest = size(Xtest, 2);
wolffd@0 81
wolffd@0 82 if isempty(Testnorm)
wolffd@0 83 % Not in kernel mode, compute distances directly
wolffd@0 84 D = 0;
wolffd@0 85 for i = 1:nKernel
wolffd@0 86 D = D + setDistanceDiag([Xtrain(:,:,i) Xtest(:,:,i)], ones(d,1), ...
wolffd@0 87 nTrain + (1:nTest), 1:nTrain);
wolffd@0 88 end
wolffd@0 89 else
wolffd@0 90 % We are in kernel mode
wolffd@0 91 D = sparse(nTrain + nTest, nTrain + nTest);
wolffd@0 92 for i = 1:nKernel
wolffd@0 93 Trainnorm = diag(Xtrain(:,:,i));
wolffd@0 94 D(1:nTrain, nTrain + (1:nTest)) = D(1:nTrain, nTrain + (1:nTest)) ...
wolffd@0 95 + bsxfun(@plus, Trainnorm, bsxfun(@plus, Testnorm(:,i)', -2 * Xtest(:,:,i)));
wolffd@0 96 end
wolffd@0 97 end
wolffd@0 98 end
wolffd@0 99