Mercurial > hg > camir-aes2014

annotate toolboxes/MIRtoolbox1.3.2/somtoolbox/som_prototrain.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 [sM,sTrain] = som_prototrain(sM, D)
wolffd@0 2
wolffd@0 3 %SOM_PROTOTRAIN Use sequential algorithm to train the Self-Organizing Map.
wolffd@0 4 %
wolffd@0 5 % [sM,sT] = som_prototrain(sM, D)
wolffd@0 6 %
wolffd@0 7 % sM = som_prototrain(sM,D);
wolffd@0 8 %
wolffd@0 9 % Input and output arguments:
wolffd@0 10 % sM (struct) map struct, the trained and updated map is returned
wolffd@0 11 % (matrix) codebook matrix of a self-organizing map
wolffd@0 12 % size munits x dim or msize(1) x ... x msize(k) x dim
wolffd@0 13 % The trained map codebook is returned.
wolffd@0 14 % D (struct) training data; data struct
wolffd@0 15 % (matrix) training data, size dlen x dim
wolffd@0 16 %
wolffd@0 17 % This function is otherwise just like SOM_SEQTRAIN except that
wolffd@0 18 % the implementation of the sequential training algorithm is very
wolffd@0 19 % straightforward (and slower). This should make it easy for you
wolffd@0 20 % to modify the algorithm, if you want to.
wolffd@0 21 %
wolffd@0 22 % For help on input and output parameters, try
wolffd@0 23 % 'type som_prototrain' or check out the help for SOM_SEQTRAIN.
wolffd@0 24 % See also SOM_SEQTRAIN, SOM_BATCHTRAIN.
wolffd@0 25
wolffd@0 26 % Contributed to SOM Toolbox vs2, February 2nd, 2000 by Juha Vesanto
wolffd@0 27 % http://www.cis.hut.fi/projects/somtoolbox/
wolffd@0 28
wolffd@0 29 % Version 2.0beta juuso 080200 130300
wolffd@0 30
wolffd@0 31 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 32 %% Check input arguments
wolffd@0 33
wolffd@0 34 % map
wolffd@0 35 struct_mode = isstruct(sM);
wolffd@0 36 if struct_mode,
wolffd@0 37 M = sM.codebook;
wolffd@0 38 sTopol = sM.topol;
wolffd@0 39 mask = sM.mask;
wolffd@0 40 msize = sTopol.msize;
wolffd@0 41 neigh = sM.neigh;
wolffd@0 42 else
wolffd@0 43 M = sM; orig_size = size(M);
wolffd@0 44 if ndims(sM) > 2,
wolffd@0 45 si = size(sM); dim = si(end); msize = si(1:end-1);
wolffd@0 46 M = reshape(sM,[prod(msize) dim]);
wolffd@0 47 else
wolffd@0 48 msize = [orig_size(1) 1]; dim = orig_size(2);
wolffd@0 49 end
wolffd@0 50 sM = som_map_struct(dim,'msize',msize); sTopol = sM.topol;
wolffd@0 51 mask = ones(dim,1);
wolffd@0 52 neigh = 'gaussian';
wolffd@0 53 end
wolffd@0 54 [munits dim] = size(M);
wolffd@0 55
wolffd@0 56 % data
wolffd@0 57 if isstruct(D), data_name = D.name; D = D.data;
wolffd@0 58 else data_name = inputname(2);
wolffd@0 59 end
wolffd@0 60 D = D(find(sum(isnan(D),2) < dim),:); % remove empty vectors from the data
wolffd@0 61 [dlen ddim] = size(D); % check input dimension
wolffd@0 62 if dim ~= ddim, error('Map and data input space dimensions disagree.'); end
wolffd@0 63
wolffd@0 64 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 65 %% initialize (these are default values, change as you will)
wolffd@0 66
wolffd@0 67 % training length
wolffd@0 68 trainlen = 20*dlen; % 20 epochs by default
wolffd@0 69
wolffd@0 70 % neighborhood radius
wolffd@0 71 radius_type = 'linear';
wolffd@0 72 rini = max(msize)/2;
wolffd@0 73 rfin = 1;
wolffd@0 74
wolffd@0 75 % learning rate
wolffd@0 76 alpha_type = 'inv';
wolffd@0 77 alpha_ini = 0.2;
wolffd@0 78
wolffd@0 79 % initialize random number generator
wolffd@0 80 rand('state',sum(100*clock));
wolffd@0 81
wolffd@0 82 % tracking
wolffd@0 83 start = clock; trackstep = 100;
wolffd@0 84
wolffd@0 85 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 86 %% Action
wolffd@0 87
wolffd@0 88 Ud = som_unit_dists(sTopol); % distance between map units on the grid
wolffd@0 89 mu_x_1 = ones(munits,1); % this is used pretty often
wolffd@0 90
wolffd@0 91 for t = 1:trainlen,
wolffd@0 92
wolffd@0 93 %% find BMU
wolffd@0 94 ind = ceil(dlen*rand(1)+eps); % select one vector
wolffd@0 95 x = D(ind,:); % pick it up
wolffd@0 96 known = ~isnan(x); % its known components
wolffd@0 97 Dx = M(:,known) - x(mu_x_1,known); % each map unit minus the vector
wolffd@0 98 dist2 = (Dx.^2)*mask(known); % squared distances
wolffd@0 99 [qerr bmu] = min(dist2); % find BMU
wolffd@0 100
wolffd@0 101 %% neighborhood
wolffd@0 102 switch radius_type, % radius
wolffd@0 103 case 'linear', r = rini+(rfin-rini)*(t-1)/(trainlen-1);
wolffd@0 104 end
wolffd@0 105 if ~r, r=eps; end % zero neighborhood radius may cause div-by-zero error
wolffd@0 106 switch neigh, % neighborhood function
wolffd@0 107 case 'bubble', h = (Ud(:,bmu) <= r);
wolffd@0 108 case 'gaussian', h = exp(-(Ud(:,bmu).^2)/(2*r*r));
wolffd@0 109 case 'cutgauss', h = exp(-(Ud(:,bmu).^2)/(2*r*r)) .* (Ud(:,bmu) <= r);
wolffd@0 110 case 'ep', h = (1 - (Ud(:,bmu).^2)/(r*r)) .* (Ud(:,bmu) <= r);
wolffd@0 111 end
wolffd@0 112
wolffd@0 113 %% learning rate
wolffd@0 114 switch alpha_type,
wolffd@0 115 case 'linear', a = (1-t/trainlen)*alpha_ini;
wolffd@0 116 case 'inv', a = alpha_ini / (1 + 99*(t-1)/(trainlen-1));
wolffd@0 117 case 'power', a = alpha_ini * (0.005/alpha_ini)^((t-1)/trainlen);
wolffd@0 118 end
wolffd@0 119
wolffd@0 120 %% update
wolffd@0 121 M(:,known) = M(:,known) - a*h(:,ones(sum(known),1)).*Dx;
wolffd@0 122
wolffd@0 123 %% tracking
wolffd@0 124 if t==1 | ~rem(t,trackstep),
wolffd@0 125 elap_t = etime(clock,start); tot_t = elap_t*trainlen/t;
wolffd@0 126 fprintf(1,'\rTraining: %3.0f/ %3.0f s',elap_t,tot_t)
wolffd@0 127 end
wolffd@0 128
wolffd@0 129 end; % for t = 1:trainlen
wolffd@0 130 fprintf(1,'\n');
wolffd@0 131
wolffd@0 132 % outputs
wolffd@0 133 sTrain = som_set('som_train','algorithm','proto',...
wolffd@0 134 'data_name',data_name,...
wolffd@0 135 'neigh',neigh,...
wolffd@0 136 'mask',mask,...
wolffd@0 137 'radius_ini',rini,...
wolffd@0 138 'radius_fin',rfin,...
wolffd@0 139 'alpha_ini',alpha_ini,...
wolffd@0 140 'alpha_type',alpha_type,...
wolffd@0 141 'trainlen',trainlen,...
wolffd@0 142 'time',datestr(now,0));
wolffd@0 143
wolffd@0 144 if struct_mode,
wolffd@0 145 sM = som_set(sM,'codebook',M,'mask',mask,'neigh',neigh);
wolffd@0 146 sM.trainhist(end+1) = sTrain;
wolffd@0 147 else
wolffd@0 148 sM = reshape(M,orig_size);
wolffd@0 149 end
wolffd@0 150
wolffd@0 151 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 152
wolffd@0 153
wolffd@0 154
wolffd@0 155