Mercurial > hg > camir-aes2014

annotate toolboxes/MIRtoolbox1.3.2/somtoolbox/som_hits.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 [hits] = som_hits(sMap, sData, mode)
wolffd@0 2
wolffd@0 3 %SOM_HITS Calculate the response of the given data on the map.
wolffd@0 4 %
wolffd@0 5 % hits = som_hits(sMap, sData, [mode])
wolffd@0 6 %
wolffd@0 7 % h = som_hits(sMap,sData);
wolffd@0 8 % h = som_hits(sMap,sData,'fuzzy');
wolffd@0 9 %
wolffd@0 10 % Input and output arguments ([]'s are optional):
wolffd@0 11 % sMap (struct) map struct
wolffd@0 12 % (matrix) codebook matrix, size munits x dim
wolffd@0 13 % sData (struct) data struct
wolffd@0 14 % (matrix) data matrix, size dlen x dim
wolffd@0 15 % [mode] (string) 'crisp' (default), 'kernel', 'fuzzy'
wolffd@0 16 %
wolffd@0 17 % hits (vector) the number of hits in each map unit, length = munits
wolffd@0 18 %
wolffd@0 19 % The response of the data on the map can be calculated e.g. in
wolffd@0 20 % three ways, selected with the mode argument:
wolffd@0 21 % 'crisp' traditional hit histogram
wolffd@0 22 % 'kernel' a sum of dlen neighborhood kernels, where kernel
wolffd@0 23 % is positioned on the BMU of each data sample. The
wolffd@0 24 % neighborhood function is sMap.neigh and the
wolffd@0 25 % neighborhood width is sMap.trainhist(end).radius_fin
wolffd@0 26 % or 1 if this is empty or NaN
wolffd@0 27 % 'fuzzy' fuzzy response calculated by summing 1./(1+(q/a)^2)
wolffd@0 28 % for each data sample, where q is a vector containing
wolffd@0 29 % distance from the data sample to each map unit and
wolffd@0 30 % a is average quantization error
wolffd@0 31 %
wolffd@0 32 % For more help, try 'type som_hits' or check out online documentation.
wolffd@0 33 % See also SOM_AUTOLABEL, SOM_BMUS.
wolffd@0 34
wolffd@0 35 %%%%%%%%%%%%% DETAILED DESCRIPTION %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 36 %
wolffd@0 37 % som_hits
wolffd@0 38 %
wolffd@0 39 % PURPOSE
wolffd@0 40 %
wolffd@0 41 % Calculate the response of the given data on the map.
wolffd@0 42 %
wolffd@0 43 % SYNTAX
wolffd@0 44 %
wolffd@0 45 % hits = som_hits(sMap, sData)
wolffd@0 46 % hits = som_hits(M, D)
wolffd@0 47 % hits = som_hits(..., mode)
wolffd@0 48 %
wolffd@0 49 % DESCRIPTION
wolffd@0 50 %
wolffd@0 51 % Returns a vector indicating the response of the map to the data.
wolffd@0 52 % The response of the data on the map can be calculated e.g. in
wolffd@0 53 % three ways, selected with the mode argument:
wolffd@0 54 % 'crisp' traditional hit histogram: how many times each map unit
wolffd@0 55 % was the BMU for the data set
wolffd@0 56 % 'kernel' a sum of neighborhood kernels, where a kernel
wolffd@0 57 % is positioned on the BMU of each data sample. The
wolffd@0 58 % neighborhood function is sMap.neigh and the
wolffd@0 59 % neighborhood width is sMap.trainhist(end).radius_fin
wolffd@0 60 % or 1 if this is not available
wolffd@0 61 % 'fuzzy' fuzzy response calculated by summing
wolffd@0 62 %
wolffd@0 63 % 1
wolffd@0 64 % ------------
wolffd@0 65 % 1 + (q/a)^2
wolffd@0 66 %
wolffd@0 67 % for each data sample, where q is a vector containing
wolffd@0 68 % distance from the data sample to each map unit and
wolffd@0 69 % a is average quantization error
wolffd@0 70 %
wolffd@0 71 % REQUIRED INPUT ARGUMENTS
wolffd@0 72 %
wolffd@0 73 % sMap The vectors from among which the BMUs are searched
wolffd@0 74 % for. These must not have any unknown components (NaNs).
wolffd@0 75 % (struct) map struct
wolffd@0 76 % (matrix) codebook matrix, size munits x dim
wolffd@0 77 %
wolffd@0 78 % sData The data vector(s) for which the BMUs are searched.
wolffd@0 79 % (struct) data struct
wolffd@0 80 % (matrix) data matrix, size dlen x dim
wolffd@0 81 %
wolffd@0 82 % OPTIONAL INPUT ARGUMENTS
wolffd@0 83 %
wolffd@0 84 % mode (string) The respond mode: 'crisp' (default), 'kernel'
wolffd@0 85 % or 'fuzzy'. 'kernel' can only be used if
wolffd@0 86 % the first argument (sMap) is a map struct.
wolffd@0 87 %
wolffd@0 88 % OUTPUT ARGUMENTS
wolffd@0 89 %
wolffd@0 90 % hits (vector) The number of hits in each map unit.
wolffd@0 91 %
wolffd@0 92 % EXAMPLES
wolffd@0 93 %
wolffd@0 94 % hits = som_hits(sM,D);
wolffd@0 95 % hits = som_hits(sM,D,'kernel');
wolffd@0 96 % hits = som_hits(sM,D,'fuzzy');
wolffd@0 97 %
wolffd@0 98 % SEE ALSO
wolffd@0 99 %
wolffd@0 100 % som_bmus Find BMUs and quantization errors for a given data set.
wolffd@0 101
wolffd@0 102 % Copyright (c) 1997-2000 by the SOM toolbox programming team.
wolffd@0 103 % http://www.cis.hut.fi/projects/somtoolbox/
wolffd@0 104
wolffd@0 105 % Version 1.0beta juuso 220997
wolffd@0 106 % Version 2.0beta juuso 161199
wolffd@0 107
wolffd@0 108 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 109 %% check arguments
wolffd@0 110
wolffd@0 111 error(nargchk(2, 3, nargin)); % check no. of input args is correct
wolffd@0 112
wolffd@0 113 if isstruct(sMap),
wolffd@0 114 switch sMap.type,
wolffd@0 115 case 'som_map', munits = prod(sMap.topol.msize);
wolffd@0 116 case 'som_data', munits = size(sMap.data,1);
wolffd@0 117 otherwise,
wolffd@0 118 error('Illegal struct for 1st argument.')
wolffd@0 119 end
wolffd@0 120 else
wolffd@0 121 munits = size(sMap,1);
wolffd@0 122 end
wolffd@0 123 hits = zeros(munits,1);
wolffd@0 124
wolffd@0 125 if nargin<3, mode = 'crisp'; end
wolffd@0 126
wolffd@0 127 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 128 %% action
wolffd@0 129
wolffd@0 130 % calculate BMUs
wolffd@0 131 [bmus,qerrs] = som_bmus(sMap,sData,1);
wolffd@0 132
wolffd@0 133 switch mode,
wolffd@0 134 case 'crisp',
wolffd@0 135
wolffd@0 136 % for each unit, check how many hits it got
wolffd@0 137 for i=1:munits, hits(i) = sum(bmus == i); end
wolffd@0 138
wolffd@0 139 case 'kernel',
wolffd@0 140
wolffd@0 141 % check that sMap really is a map
wolffd@0 142 if ~isstruct(sMap) & ~strcmp(sMap.type,'som_map'),
wolffd@0 143 error('Kernel mode can only be used for maps.');
wolffd@0 144 end
wolffd@0 145
wolffd@0 146 % calculate neighborhood kernel
wolffd@0 147 Ud = som_unit_dists(sMap.topol).^2;
wolffd@0 148 sTrain = sMap.trainhist(end);
wolffd@0 149 if ~isempty(sTrain),
wolffd@0 150 rad = sTrain.radius_fin;
wolffd@0 151 if isempty(rad) | isnan(rad), rad = 1; end
wolffd@0 152 else
wolffd@0 153 rad = 1;
wolffd@0 154 end
wolffd@0 155 rad = rad^2;
wolffd@0 156 if rad==0, rad = eps; end % to avoid divide-by-0 errors
wolffd@0 157 switch sTrain.neigh,
wolffd@0 158 case 'bubble', H = (Ud<=rad);
wolffd@0 159 case 'gaussian', H = exp(-Ud/(2*rad));
wolffd@0 160 case 'cutgauss', H = exp(-Ud/(2*rad)) .* (Ud<=rad);
wolffd@0 161 case 'ep', H = (1-Ud/rad) .* (Ud<=rad);
wolffd@0 162 end
wolffd@0 163
wolffd@0 164 % weight hits with neighborhood kernel
wolffd@0 165 hits = sum(H(bmus,:),1)';
wolffd@0 166
wolffd@0 167 case 'fuzzy',
wolffd@0 168
wolffd@0 169 % extract the two matrices (M, D) and the mask
wolffd@0 170 mask = [];
wolffd@0 171 if isstruct(sMap),
wolffd@0 172 if strcmp(sMap.type,'som_data'), M = sMap.data;
wolffd@0 173 else M = sMap.codebook; mask = sMap.mask;
wolffd@0 174 end
wolffd@0 175 else M = sMap;
wolffd@0 176 end
wolffd@0 177 if any(isnan(M(:))),
wolffd@0 178 error('Data in first argument must not have any NaNs.');
wolffd@0 179 end
wolffd@0 180
wolffd@0 181 if isstruct(sData),
wolffd@0 182 switch sData.type,
wolffd@0 183 case 'som_map',
wolffd@0 184 D = sData.codebook;
wolffd@0 185 if isempty(mask), mask = sData.mask; end
wolffd@0 186 case 'som_data', D = sData.data;
wolffd@0 187 otherwise, error('Illegal 2nd argument.');
wolffd@0 188 end
wolffd@0 189 else D = sData;
wolffd@0 190 end
wolffd@0 191 [dlen dim] = size(D);
wolffd@0 192 if isempty(mask), mask = ones(dim,1); end
wolffd@0 193
wolffd@0 194 % scaling factor
wolffd@0 195 a = mean(qerrs).^2;
wolffd@0 196
wolffd@0 197 % calculate distances & bmus
wolffd@0 198 % (this is better explained in som_batchtrain and som_bmus)
wolffd@0 199 Known = ~isnan(D); D(find(~Known)) = 0; % unknown components
wolffd@0 200 blen = min(munits,dlen); % block size
wolffd@0 201 W1 = mask*ones(1,blen); W2 = ones(munits,1)*mask'; D = D'; Known = Known';
wolffd@0 202 i0 = 0;
wolffd@0 203 while i0+1<=dlen,
wolffd@0 204 inds = [(i0+1):min(dlen,i0+blen)]; i0 = i0+blen; % indeces
wolffd@0 205 Dist = (M.^2)*(W1(:,1:length(inds)).*Known(:,inds)) ...
wolffd@0 206 + W2*(D(:,inds).^2) ...
wolffd@0 207 - 2*M*diag(mask)*D(:,inds); % squared distances
wolffd@0 208 hits = hits + sum(1./(1+Dist/a),2);
wolffd@0 209 end
wolffd@0 210
wolffd@0 211 otherwise,
wolffd@0 212 error(['Unknown mode: ' mode]);
wolffd@0 213
wolffd@0 214 end
wolffd@0 215
wolffd@0 216 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 217