Mercurial > hg > camir-aes2014

annotate core/magnatagatune/DistMeasureNNet.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 % ---
wolffd@0 2 % The DistMeasureMahal class states a wrapper for
wolffd@0 3 % special Mahalanobis similarity distance, and is compatible with the
wolffd@0 4 % DistMeasure class
wolffd@0 5 % ---
wolffd@0 6 classdef DistMeasureNNet < handle
wolffd@0 7
wolffd@0 8 properties (SetAccess = private)
wolffd@0 9
wolffd@0 10 net;
wolffd@0 11
wolffd@0 12 featX;
wolffd@0 13
wolffd@0 14 ids;
wolffd@0 15 ret_ids;
wolffd@0 16 end
wolffd@0 17
wolffd@0 18 methods
wolffd@0 19
wolffd@0 20 % ---
wolffd@0 21 % constructor
wolffd@0 22 % ---
wolffd@0 23 function m = DistMeasureNNet(clips, net, featX)
wolffd@0 24
wolffd@0 25 if size(featX, 2) ~= numel(clips)
wolffd@0 26 error 'wrong input format'
wolffd@0 27 end
wolffd@0 28
wolffd@0 29 % fill index and generate matrix;
wolffd@0 30 m.ids = [clips.id];
wolffd@0 31
wolffd@0 32 % reverse index
wolffd@0 33 m.ret_ids = sparse(numel(m.ids),1);
wolffd@0 34 m.ret_ids(m.ids) = 1:numel(m.ids);
wolffd@0 35
wolffd@0 36 % ---
wolffd@0 37 % save neural net and lazy-copy features
wolffd@0 38 % ---
wolffd@0 39 m.net = net;
wolffd@0 40
wolffd@0 41 m.featX = featX;
wolffd@0 42
wolffd@0 43 end
wolffd@0 44
wolffd@0 45
wolffd@0 46 % ---
wolffd@0 47 % this function returns the
wolffd@0 48 % similarity of two clip indices
wolffd@0 49 % ---
wolffd@0 50 function out = mat(m, idxa, idxb)
wolffd@0 51
wolffd@0 52 if nargin == 1
wolffd@0 53 idxa = 1:numel(m.ids);
wolffd@0 54 idxb = 1:numel(m.ids);
wolffd@0 55 end
wolffd@0 56
wolffd@0 57 % cycle through all index combinations
wolffd@0 58 out = zeros(numel(idxa), numel(idxb));
wolffd@0 59 for i = 1:numel(idxa)
wolffd@0 60 for j = 1:numel(idxb)
wolffd@0 61
wolffd@0 62 % calculate distance vector
wolffd@0 63 deltas = m.featX(:,idxa(i)) - m.featX(:,idxb(j));
wolffd@0 64
wolffd@0 65 % return distance from net
wolffd@0 66 out(i,j) = m.net.calcValue(deltas);
wolffd@0 67 end
wolffd@0 68 end
wolffd@0 69
wolffd@0 70 end
wolffd@0 71
wolffd@0 72 % ---
wolffd@0 73 % returns the distance for the two input clips
wolffd@0 74 % ---
wolffd@0 75 function out = distance(m, clipa, clipb)
wolffd@0 76 posa = m.get_clip_pos(clipa);
wolffd@0 77 posb = m.get_clip_pos(clipb);
wolffd@0 78
wolffd@0 79 out = m.mat(posa, posb);
wolffd@0 80 end
wolffd@0 81
wolffd@0 82 % ---
wolffd@0 83 % returns a list of n (default = 10) clips most
wolffd@0 84 % similar to the input
wolffd@0 85 % ---
wolffd@0 86 function [clips, dist] = get_nearest(m, clip, n)
wolffd@0 87 % list = get_nearest(m, clip, n)
wolffd@0 88 %
wolffd@0 89 % returns a list of n (default = 10) clips most
wolffd@0 90 % similar to the input
wolffd@0 91
wolffd@0 92 % default number of results
wolffd@0 93 if nargin == 2
wolffd@0 94
wolffd@0 95 n = 10;
wolffd@0 96 end
wolffd@0 97
wolffd@0 98 % return all clips in case n = 0
wolffd@0 99 if n == 0; n = numel(m.ids); end
wolffd@0 100
wolffd@0 101 % get clip positions
wolffd@0 102 pos = m.get_clip_pos(clip);
wolffd@0 103
wolffd@0 104 % sort according to distance
wolffd@0 105 [sc, idx] = sort( m.mat(pos, 1:numel(m.ids)), 'ascend');
wolffd@0 106
wolffd@0 107 % we only output relevant data
wolffd@0 108 idx = idx(sc < inf);
wolffd@0 109
wolffd@0 110 if numel(idx) > 0
wolffd@0 111 % create clips form best ids
wolffd@0 112 clips = MTTClip( m.ids( idx(1:min(n, end))));
wolffd@0 113 dist = m.mat(pos, idx(1:min(n, end)));
wolffd@0 114
wolffd@0 115 else
wolffd@0 116 clips = [];
wolffd@0 117 dist = [];
wolffd@0 118 end
wolffd@0 119 end
wolffd@0 120
wolffd@0 121
wolffd@0 122
wolffd@0 123 function [clips, dist] = present_nearest(m, clip, n)
wolffd@0 124 % plays and shows the n best hits for a given clip
wolffd@0 125
wolffd@0 126 % default number of results
wolffd@0 127 if nargin == 2
wolffd@0 128
wolffd@0 129 n = 3;
wolffd@0 130 end
wolffd@0 131
wolffd@0 132 % get best list
wolffd@0 133 [clips, dist] = get_nearest(m, clip, n);
wolffd@0 134
wolffd@0 135 clip.audio_features_basicsm.visualise();
wolffd@0 136 for i = 1:numel(clips)
wolffd@0 137 fprintf('\n\n\n- Rank %d, distance: %1.4f \n\n',i, dist(i));
wolffd@0 138
wolffd@0 139 clips(i).audio_features_basicsm.visualise();
wolffd@0 140 h = gcf();
wolffd@0 141 t = clips(i).play(20);
wolffd@0 142 pause(t);
wolffd@0 143 close(h);
wolffd@0 144 end
wolffd@0 145 end
wolffd@0 146
wolffd@0 147 function a = visualise(m)
wolffd@0 148
wolffd@0 149 figure;
wolffd@0 150
wolffd@0 151 % plot data
wolffd@0 152
wolffd@0 153 imagesc(m.mat);
wolffd@0 154
wolffd@0 155 a = gca;
wolffd@0 156 set(a,'YTick',[1:numel(m.ids)], 'YTickLabel',m.ids);
wolffd@0 157 set(a,'XTick',[1:numel(m.ids)], 'XTickLabel', m.ids);
wolffd@0 158
wolffd@0 159 axis xy;
wolffd@0 160 colormap(hot);
wolffd@0 161 end
wolffd@0 162
wolffd@0 163 % end methods
wolffd@0 164 end
wolffd@0 165
wolffd@0 166 % ---
wolffd@0 167 % private methods
wolffd@0 168 % ---
wolffd@0 169 methods(Access = private)
wolffd@0 170
wolffd@0 171 function out = get_clip_pos(m, clip)
wolffd@0 172 % returns position in mat for given clip
wolffd@0 173
wolffd@0 174 out = m.ret_ids(clip.id);
wolffd@0 175 end
wolffd@0 176
wolffd@0 177 end
wolffd@0 178
wolffd@0 179 end