Mercurial > hg > camir-aes2014

annotate toolboxes/distance_learning/mlr/cuttingPlane/cuttingPlaneRandom.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 [dPsi, M, SO_time] = cuttingPlaneRandom(k, X, W, Ypos, Yneg, batchSize, SAMPLES, ClassScores)
wolffd@0 2 %
wolffd@0 3 % [dPsi, M, SO_time] = cuttingPlaneRandom(k, X, W, Yp, Yn, batchSize, SAMPLES, ClassScores)
wolffd@0 4 %
wolffd@0 5 % k = k parameter for the SO
wolffd@0 6 % X = d*n data matrix
wolffd@0 7 % W = d*d PSD metric
wolffd@0 8 % Yp = cell-array of relevant results for each point
wolffd@0 9 % Yn = cell-array of irrelevant results for each point
wolffd@0 10 % batchSize = number of points to use in the constraint batch
wolffd@0 11 % SAMPLES = indices of valid points to include in the batch
wolffd@0 12 % ClassScores = structure for synthetic constraints
wolffd@0 13 %
wolffd@0 14 % dPsi = dPsi vector for this batch
wolffd@0 15 % M = mean loss on this batch
wolffd@0 16 % SO_time = time spent in separation oracle
wolffd@0 17
wolffd@0 18 global SO PSI SETDISTANCE CPGRADIENT;
wolffd@0 19
wolffd@0 20 [d,n] = size(X);
wolffd@0 21
wolffd@0 22
wolffd@0 23 if length(SAMPLES) == n
wolffd@0 24 % All samples are fair game (full data)
wolffd@0 25 Batch = randperm(n);
wolffd@0 26 Batch = Batch(1:batchSize);
wolffd@0 27 D = SETDISTANCE(X, W, Batch);
wolffd@0 28
wolffd@0 29 else
wolffd@0 30 Batch = randperm(length(SAMPLES));
wolffd@0 31 Batch = SAMPLES(Batch(1:batchSize));
wolffd@0 32
wolffd@0 33 Ito = sparse(n,1);
wolffd@0 34
wolffd@0 35 if isempty(ClassScores)
wolffd@0 36 for i = Batch
wolffd@0 37 Ito(Ypos{i}) = 1;
wolffd@0 38 Ito(Yneg{i}) = 1;
wolffd@0 39 end
wolffd@0 40 D = SETDISTANCE(X, W, Batch, find(Ito));
wolffd@0 41 else
wolffd@0 42 D = SETDISTANCE(X, W, Batch, 1:n);
wolffd@0 43 end
wolffd@0 44 end
wolffd@0 45
wolffd@0 46
wolffd@0 47 M = 0;
wolffd@0 48 S = zeros(n);
wolffd@0 49 dIndex = sub2ind([n n], 1:n, 1:n);
wolffd@0 50
wolffd@0 51 SO_time = 0;
wolffd@0 52
wolffd@0 53
wolffd@0 54 if isempty(ClassScores)
wolffd@0 55 TS = zeros(batchSize, n);
wolffd@0 56 parfor j = 1:batchSize
wolffd@0 57 i = Batch(j);
wolffd@0 58 if isempty(Yneg)
wolffd@0 59 Ynegative = setdiff((1:n)', [i ; Ypos{i}]);
wolffd@0 60 else
wolffd@0 61 Ynegative = Yneg{i};
wolffd@0 62 end
wolffd@0 63 SO_start = tic();
wolffd@0 64 [yi, li] = SO(i, D, Ypos{i}, Ynegative, k);
wolffd@0 65 SO_time = SO_time + toc(SO_start);
wolffd@0 66
wolffd@0 67 M = M + li /batchSize;
wolffd@0 68 TS(j,:) = PSI(i, yi', n, Ypos{i}, Ynegative);
wolffd@0 69 end
wolffd@0 70 S(Batch,:) = TS;
wolffd@0 71 S(:,Batch) = S(:,Batch) + TS';
wolffd@0 72 S(dIndex) = S(dIndex) - sum(TS, 1);
wolffd@0 73 else
wolffd@0 74 for j = 1:length(ClassScores.classes)
wolffd@0 75 c = ClassScores.classes(j);
wolffd@0 76 points = find(ClassScores.Y(Batch) == c);
wolffd@0 77 if ~any(points)
wolffd@0 78 continue;
wolffd@0 79 end
wolffd@0 80
wolffd@0 81 Yneg = find(ClassScores.Yneg{j});
wolffd@0 82 yp = ClassScores.Ypos{j};
wolffd@0 83
wolffd@0 84 TS = zeros(length(points), n);
wolffd@0 85 parfor x = 1:length(points)
wolffd@0 86 i = Batch(points(x));
wolffd@0 87 yl = yp;
wolffd@0 88 yl(i) = 0;
wolffd@0 89 Ypos = find(yl);
wolffd@0 90 SO_start = tic();
wolffd@0 91 [yi, li] = SO(i, D, Ypos, Yneg, k);
wolffd@0 92 SO_time = SO_time + toc(SO_start);
wolffd@0 93
wolffd@0 94 M = M + li /batchSize;
wolffd@0 95 TS(x,:) = PSI(i, yi', n, Ypos, Yneg);
wolffd@0 96 end
wolffd@0 97 S(Batch(points),:) = S(Batch(points),:) + TS;
wolffd@0 98 S(:,Batch(points)) = S(:,Batch(points)) + TS';
wolffd@0 99 S(dIndex) = S(dIndex) - sum(TS, 1);
wolffd@0 100 end
wolffd@0 101 end
wolffd@0 102
wolffd@0 103 dPsi = CPGRADIENT(X, S, batchSize);
wolffd@0 104
wolffd@0 105 end