Mercurial > hg > camir-aes2014

annotate toolboxes/MIRtoolbox1.3.2/MIRToolbox/mircluster.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 [a d] = mircluster(a,varargin)
wolffd@0 2 % c = mircluster(a,f) clusters the segments in the audio sequence(s)
wolffd@0 3 % contained in the audio object a, along the analytic feature(s)
wolffd@0 4 % f, using the k-means strategy. Multiple analytic features have to
wolffd@0 5 % be grouped into one array of cells.
wolffd@0 6 % Example:
wolffd@0 7 % sg = mirsegment(a);
wolffd@0 8 % mircluster(sg, mirmfcc(sg))
wolffd@0 9 % mircluster(sg, {mirmfcc(sg), mircentroid(sg)})
wolffd@0 10 % c = mircluster(d) clusters the frame-decomposed data d into groups
wolffd@0 11 % using K-means clustering.
wolffd@0 12 % Example:
wolffd@0 13 % cc = mirmfcc(a,'Frame');
wolffd@0 14 % mircluster(cc)
wolffd@0 15 % Optional argument:
wolffd@0 16 % mircluster(...,n) indicates the maximal number of clusters.
wolffd@0 17 % Default value: n = 2.
wolffd@0 18 % mircluster(...,'Runs',r) indicates the maximal number of runs.
wolffd@0 19 % Default value: r = 5.
wolffd@0 20 %
wolffd@0 21 % Requires SOM Toolbox (included in the MIRtoolbox distribution).
wolffd@0 22
wolffd@0 23
wolffd@0 24 nruns.key = 'Runs';
wolffd@0 25 nruns.type = 'Integer';
wolffd@0 26 nruns.default = 5;
wolffd@0 27 option.nruns = nruns;
wolffd@0 28
wolffd@0 29 nclust.position = 2;
wolffd@0 30 nclust.type = 'Integer';
wolffd@0 31 nclust.default = 2;
wolffd@0 32 option.nclust = nclust;
wolffd@0 33
wolffd@0 34 specif.option = option;
wolffd@0 35
wolffd@0 36 specif.nochunk = 1;
wolffd@0 37
wolffd@0 38 d = a;
wolffd@0 39 if isa(a,'mirdesign')
wolffd@0 40 if not(get(a,'Eval'))
wolffd@0 41 % During bottom-up construction of the general design
wolffd@0 42
wolffd@0 43 [unused option] = miroptions(@mircluster,a,specif,varargin);
wolffd@0 44 type = get(a,'Type');
wolffd@0 45 a = mirdesign(@mircluster,a,option,{},struct,type);
wolffd@0 46 a = set(a,'NoChunk',1);
wolffd@0 47 else
wolffd@0 48 % During top-down evaluation initiation
wolffd@0 49
wolffd@0 50 e = evaleach(a);
wolffd@0 51 if iscell(e)
wolffd@0 52 e = e{1};
wolffd@0 53 end
wolffd@0 54 a = mircluster(e,varargin{:});
wolffd@0 55 end
wolffd@0 56 else
wolffd@0 57 if not(isa(a,'mirdata'))
wolffd@0 58 mirerror('mircluster','The input should be either frame- or segment-decomposed.');
wolffd@0 59 end
wolffd@0 60
wolffd@0 61 if isempty(varargin) || (not(isa(varargin{1},'mirdata') || ...
wolffd@0 62 (iscell(varargin{1}) && ...
wolffd@0 63 isa(varargin{1}{1},'mirdata'))))
wolffd@0 64 % mircluster version for frame-decomposed data:
wolffd@0 65 % frames are clustered into groups using K-means clustering.
wolffd@0 66 [unused option] = miroptions(@mircluster,a,specif,varargin);
wolffd@0 67 da = get(a,'Data');
wolffd@0 68 lva = length(da); % Number of audio files in the audio object.
wolffd@0 69 c = cell(1,lva);
wolffd@0 70 display('Clustering frames...');
wolffd@0 71 if mirwaitbar
wolffd@0 72 handle = waitbar(0,'Clustering frames...');
wolffd@0 73 else
wolffd@0 74 handle = 0;
wolffd@0 75 end
wolffd@0 76 for j = 1:lva % For each audio file,...
wolffd@0 77 va = []; % Data transmitted to the kmeans_cluster function.
wolffd@0 78 v = da{j};
wolffd@0 79 if iscell(v)
wolffd@0 80 v = uncell(v,-Inf); %v{1};
wolffd@0 81 end
wolffd@0 82 if size(v,4)>1
wolffd@0 83 v(end+1:2*end,:,:,1) = v(:,:,:,2);
wolffd@0 84 v(:,:,:,2) = [];
wolffd@0 85 end
wolffd@0 86 % Standardization
wolffd@0 87 %stv = std(v,0,2);
wolffd@0 88 %stv(find(stv == 0)) = 1;
wolffd@0 89 va(end+1:end+size(v,1),:,:) = v;%...
wolffd@0 90 %(v - repmat(mean(v,2),[1 size(v,2) ])) ...
wolffd@0 91 %./ repmat(stv,[1 size(v,2) ]);
wolffd@0 92 if isa(a,'mirscalar')
wolffd@0 93 m = get(a,'Mode');
wolffd@0 94 if not(isempty(m))
wolffd@0 95 m = m{j};
wolffd@0 96 val = [];
wolffd@0 97 for l = 1:nseg
wolffd@0 98 vl = m{l};
wolffd@0 99 if iscell(vl)
wolffd@0 100 vl = vl{1};
wolffd@0 101 end
wolffd@0 102 val(:,l) = vl;
wolffd@0 103 end
wolffd@0 104 stv = std(val,0,2);
wolffd@0 105 stv(find(stv == 0)) = 1;
wolffd@0 106 va(end+1:end+size(val,1),:) = ...
wolffd@0 107 (val - repmat(mean(val,2),[1 size(val,2) ])) ...
wolffd@0 108 ./ repmat(stv,[1 size(val,2) ]);
wolffd@0 109 end
wolffd@0 110 end
wolffd@0 111 if size(va,3)>1
wolffd@0 112 mel = 1;
wolffd@0 113 va = reshape(va,size(va,2),size(va,3))';
wolffd@0 114 else
wolffd@0 115 mel = 0;
wolffd@0 116 end
wolffd@0 117 [cc, p, err, ind] = kmeans_clusters(va',option.nclust,option.nruns);
wolffd@0 118 [minind select] = min(ind);
wolffd@0 119 c{j}.centr = cc{select}';
wolffd@0 120 c{j}.index = p{select};
wolffd@0 121 c{j}.weight = zeros(1,size(cc{select},1));
wolffd@0 122 c{j}.covar = zeros(size(cc{select}'));
wolffd@0 123 ii = 1;
wolffd@0 124 for i = 1:size(c{j}.centr,2)
wolffd@0 125 clus = va(:,c{j}.index == ii);
wolffd@0 126 if isempty(clus)
wolffd@0 127 higher = find(c{j}.index > ii);
wolffd@0 128 c{j}.index(higher) = c{j}.index(higher)-1;
wolffd@0 129 c{j}.centr(:,ii) = [];
wolffd@0 130 c{j}.weight(ii) = [];
wolffd@0 131 c{j}.covar(:,ii) = [];
wolffd@0 132 else
wolffd@0 133 c{j}.weight(ii) = size(clus,2)/size(va,2);
wolffd@0 134 if c{j}.weight(ii) == 0
wolffd@0 135 pause
wolffd@0 136 end
wolffd@0 137 c{j}.covar(:,ii) = mean((clus'-ones(1,size(clus,1))*c{j}.centr(:,ii)).^2);
wolffd@0 138 ii = ii+1;
wolffd@0 139 end
wolffd@0 140 end
wolffd@0 141 if handle
wolffd@0 142 waitbar(j/lva,handle);
wolffd@0 143 end
wolffd@0 144 end
wolffd@0 145 if handle
wolffd@0 146 delete(handle)
wolffd@0 147 end
wolffd@0 148 a = set(a,'Clusters',c);
wolffd@0 149 else
wolffd@0 150 % mircluster version for segmented audio:
wolffd@0 151 % segments are clustered into groups using K-means clustering.
wolffd@0 152 da = varargin{1};
wolffd@0 153 varargin(1) = [];
wolffd@0 154 [unused option] = miroptions(@mircluster,a,specif,varargin);
wolffd@0 155 display('Clustering segments...');
wolffd@0 156 if isa(da,'mirdata') || (iscell(da) && isa(da{1},'mirdata'))
wolffd@0 157 if not(iscell(da))
wolffd@0 158 da = {da};
wolffd@0 159 end
wolffd@0 160 vala = get(a,'Data'); % Data contained in the audio object a.
wolffd@0 161 lva = length(vala); % Number of audio files in the audio object.
wolffd@0 162 clus = cell(1,lva);
wolffd@0 163 for j = 1:lva % For each audio file,...
wolffd@0 164 va = []; % Data transmitted to the kmeans_cluster function.
wolffd@0 165 nseg = length(vala{j}); % Number of segments in the audio file.
wolffd@0 166 for i = 1:length(da) % For each analytic feature,...
wolffd@0 167 v = get(da{i},'Data');
wolffd@0 168 v = v{j};
wolffd@0 169 if iscell(v)
wolffd@0 170 v = uncell(v,-Inf); %v{1};
wolffd@0 171 end
wolffd@0 172 val = [];
wolffd@0 173 if size(v,4)>1
wolffd@0 174 v(end+1:2*end,:,:,1) = v(:,:,:,2);
wolffd@0 175 v(:,:,:,2) = [];
wolffd@0 176 end
wolffd@0 177
wolffd@0 178 % Standardization
wolffd@0 179 stv = std(v,0,2);
wolffd@0 180 stv(find(stv == 0)) = 1;
wolffd@0 181 va(end+1:end+size(v,1),:) = ...
wolffd@0 182 (v - repmat(mean(v,2),[1 size(v,2) ])) ...
wolffd@0 183 ./ repmat(stv,[1 size(v,2) ]);
wolffd@0 184 if isa(da{i},'mirscalar')
wolffd@0 185 m = get(da{i},'Mode');
wolffd@0 186 if not(isempty(m))
wolffd@0 187 m = m{j};
wolffd@0 188 val = [];
wolffd@0 189 for l = 1:nseg
wolffd@0 190 vl = m{l};
wolffd@0 191 if iscell(vl)
wolffd@0 192 vl = vl{1};
wolffd@0 193 end
wolffd@0 194 val(:,l) = vl;
wolffd@0 195 end
wolffd@0 196 stv = std(val,0,2);
wolffd@0 197 stv(find(stv == 0)) = 1;
wolffd@0 198 va(end+1:end+size(val,1),:) = ...
wolffd@0 199 (val - repmat(mean(val,2),[1 size(val,2) ])) ...
wolffd@0 200 ./ repmat(stv,[1 size(val,2) ]);
wolffd@0 201 end
wolffd@0 202 end
wolffd@0 203
wolffd@0 204 end
wolffd@0 205 [cc, p, err, ind] = kmeans_clusters(va',min(option.nclust,nseg),option.nruns);
wolffd@0 206 clus{j} = p{end};
wolffd@0 207 end
wolffd@0 208 a = set(a,'Clusters',clus);
wolffd@0 209 t = get(a,'Time');
wolffd@0 210 fp = get(a,'FramePos');
wolffd@0 211 for j = 1:lva % For each audio file,...
wolffd@0 212 aj = vala{j};
wolffd@0 213 tj = t{j};
wolffd@0 214 fpj = fp{j};
wolffd@0 215 clj = clus{j};
wolffd@0 216 k = 2;
wolffd@0 217 while k <= length(aj)
wolffd@0 218 if clj(k) == clj(k-1)
wolffd@0 219 aj{k-1} = [aj{k-1};aj{k}];
wolffd@0 220 aj(k) = [];
wolffd@0 221 tj{k-1} = [tj{k-1};tj{k}];
wolffd@0 222 tj(k) = [];
wolffd@0 223 fpj{k-1} = [fpj{k-1}(1);fpj{k}(2)];
wolffd@0 224 fpj(k) = [];
wolffd@0 225 clj(k) = [];
wolffd@0 226 k = k-1;
wolffd@0 227 end
wolffd@0 228 k = k+1;
wolffd@0 229 end
wolffd@0 230 vala{j} = aj;
wolffd@0 231 t{j} = tj;
wolffd@0 232 fp{j} = fpj;
wolffd@0 233 cl{j} = clj;
wolffd@0 234 end
wolffd@0 235 a = set(a,'Data',vala,'Time',t,'FramePos',fp,'Clusters',cl);
wolffd@0 236 end
wolffd@0 237 end
wolffd@0 238 end