annotate DL/RLS-DLA/SMALL_rlsdla.m @ 40:6416fc12f2b8

(none)
author idamnjanovic
date Mon, 14 Mar 2011 15:35:24 +0000
parents
children 55faa9b5d1ac
rev   line source
idamnjanovic@40 1 function Dictionary = SMALL_rlsdla(X, params)
idamnjanovic@40 2
idamnjanovic@40 3
idamnjanovic@40 4
idamnjanovic@40 5
idamnjanovic@40 6
idamnjanovic@40 7
idamnjanovic@40 8 CODE_SPARSITY = 1;
idamnjanovic@40 9 CODE_ERROR = 2;
idamnjanovic@40 10
idamnjanovic@40 11
idamnjanovic@40 12 % Determine which method will be used for sparse representation step -
idamnjanovic@40 13 % Sparsity or Error mode
idamnjanovic@40 14
idamnjanovic@40 15 if (isfield(params,'codemode'))
idamnjanovic@40 16 switch lower(params.codemode)
idamnjanovic@40 17 case 'sparsity'
idamnjanovic@40 18 codemode = CODE_SPARSITY;
idamnjanovic@40 19 thresh = params.Tdata;
idamnjanovic@40 20 case 'error'
idamnjanovic@40 21 codemode = CODE_ERROR;
idamnjanovic@40 22 thresh = params.Edata;
idamnjanovic@40 23
idamnjanovic@40 24 otherwise
idamnjanovic@40 25 error('Invalid coding mode specified');
idamnjanovic@40 26 end
idamnjanovic@40 27 elseif (isfield(params,'Tdata'))
idamnjanovic@40 28 codemode = CODE_SPARSITY;
idamnjanovic@40 29 thresh = params.Tdata;
idamnjanovic@40 30 elseif (isfield(params,'Edata'))
idamnjanovic@40 31 codemode = CODE_ERROR;
idamnjanovic@40 32 thresh = params.Edata;
idamnjanovic@40 33
idamnjanovic@40 34 else
idamnjanovic@40 35 error('Data sparse-coding target not specified');
idamnjanovic@40 36 end
idamnjanovic@40 37
idamnjanovic@40 38
idamnjanovic@40 39 % max number of atoms %
idamnjanovic@40 40
idamnjanovic@40 41 if (codemode==CODE_ERROR && isfield(params,'maxatoms'))
idamnjanovic@40 42 maxatoms = params.maxatoms;
idamnjanovic@40 43 else
idamnjanovic@40 44 maxatoms = -1;
idamnjanovic@40 45 end
idamnjanovic@40 46
idamnjanovic@40 47
idamnjanovic@40 48 % Forgetting factor
idamnjanovic@40 49
idamnjanovic@40 50 if (isfield(params,'forgettingMode'))
idamnjanovic@40 51 switch lower(params.forgettingMode)
idamnjanovic@40 52 case 'fix'
idamnjanovic@40 53 if (isfield(params,'forgettingFactor'))
idamnjanovic@40 54 lambda=params.forgettingFactor;
idamnjanovic@40 55 else
idamnjanovic@40 56 lambda=1;
idamnjanovic@40 57 end
idamnjanovic@40 58 otherwise
idamnjanovic@40 59 error('This mode is still not implemented');
idamnjanovic@40 60 end
idamnjanovic@40 61 elseif (isfield(params,'forgettingFactor'))
idamnjanovic@40 62 lambda=params.forgettingFactor;
idamnjanovic@40 63 else
idamnjanovic@40 64 lambda=1;
idamnjanovic@40 65 end
idamnjanovic@40 66
idamnjanovic@40 67 % determine dictionary size %
idamnjanovic@40 68
idamnjanovic@40 69 if (isfield(params,'initdict'))
idamnjanovic@40 70 if (any(size(params.initdict)==1) && all(iswhole(params.initdict(:))))
idamnjanovic@40 71 dictsize = length(params.initdict);
idamnjanovic@40 72 else
idamnjanovic@40 73 dictsize = size(params.initdict,2);
idamnjanovic@40 74 end
idamnjanovic@40 75 end
idamnjanovic@40 76 if (isfield(params,'dictsize')) % this superceedes the size determined by initdict
idamnjanovic@40 77 dictsize = params.dictsize;
idamnjanovic@40 78 end
idamnjanovic@40 79
idamnjanovic@40 80 if (size(X,2) < dictsize)
idamnjanovic@40 81 error('Number of training signals is smaller than number of atoms to train');
idamnjanovic@40 82 end
idamnjanovic@40 83
idamnjanovic@40 84
idamnjanovic@40 85 % initialize the dictionary %
idamnjanovic@40 86
idamnjanovic@40 87 if (isfield(params,'initdict'))
idamnjanovic@40 88 if (any(size(params.initdict)==1) && all(iswhole(params.initdict(:))))
idamnjanovic@40 89 D = X(:,params.initdict(1:dictsize));
idamnjanovic@40 90 else
idamnjanovic@40 91 if (size(params.initdict,1)~=size(X,1) || size(params.initdict,2)<dictsize)
idamnjanovic@40 92 error('Invalid initial dictionary');
idamnjanovic@40 93 end
idamnjanovic@40 94 D = params.initdict(:,1:dictsize);
idamnjanovic@40 95 end
idamnjanovic@40 96 else
idamnjanovic@40 97 data_ids = find(colnorms_squared(X) > 1e-6); % ensure no zero data elements are chosen
idamnjanovic@40 98 perm = randperm(length(data_ids));
idamnjanovic@40 99 D = X(:,data_ids(perm(1:dictsize)));
idamnjanovic@40 100 end
idamnjanovic@40 101
idamnjanovic@40 102
idamnjanovic@40 103 % normalize the dictionary %
idamnjanovic@40 104
idamnjanovic@40 105 D = normcols(D);
idamnjanovic@40 106
idamnjanovic@40 107 % Training data
idamnjanovic@40 108
idamnjanovic@40 109 data=X;
idamnjanovic@40 110
idamnjanovic@40 111 %
idamnjanovic@40 112
idamnjanovic@40 113 C=(100000*thresh)*eye(dictsize);
idamnjanovic@40 114 w=zeros(dictsize,1);
idamnjanovic@40 115 u=zeros(dictsize,1);
idamnjanovic@40 116
idamnjanovic@40 117
idamnjanovic@40 118 for i = 1:size(data,2)
idamnjanovic@40 119
idamnjanovic@40 120 if (codemode == CODE_SPARSITY)
idamnjanovic@40 121 w = ompmex(D,data(:,i),[],[],thresh,1,-1,0);
idamnjanovic@40 122 else
idamnjanovic@40 123 w = omp2mex(D,data(:,i),[],[],[],thresh,0,-1,maxatoms,0);
idamnjanovic@40 124 end
idamnjanovic@40 125
idamnjanovic@40 126 spind=find(w);
idamnjanovic@40 127
idamnjanovic@40 128 residual = data(:,i) - D * w;
idamnjanovic@40 129
idamnjanovic@40 130 if (lambda~=1)
idamnjanovic@40 131 C = C *(1/ lambda);
idamnjanovic@40 132 end
idamnjanovic@40 133
idamnjanovic@40 134 u = C(:,spind) * w(spind);
idamnjanovic@40 135
idamnjanovic@40 136
idamnjanovic@40 137 alfa = 1/(1 + w' * u);
idamnjanovic@40 138
idamnjanovic@40 139 D = D + (alfa * residual) * u';
idamnjanovic@40 140
idamnjanovic@40 141
idamnjanovic@40 142 C = C - (alfa * u)* u';
idamnjanovic@40 143
idamnjanovic@40 144 end
idamnjanovic@40 145
idamnjanovic@40 146 Dictionary = D;
idamnjanovic@40 147
idamnjanovic@40 148 end