Mercurial > hg > camir-aes2014

annotate core/magnatagatune/MTTMixedFeatureSlaney08GenreBasicSmRBM.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 % This class loads and hanles the aufdio features included with the MTT
wolffd@0 3 % Library
wolffd@0 4 %
wolffd@0 5 % Features are extracted and then porcessed by son's RBM toolbox
wolffd@0 6 % ---
wolffd@0 7 classdef MTTMixedFeatureSlaney08GenreBasicSmRBM < MTTAudioFeature & handle
wolffd@0 8
wolffd@0 9 properties(Constant = true)
wolffd@0 10
wolffd@0 11 my_revision = str2double(substr('$Rev: 457 $', 5, -1));
wolffd@0 12 end
wolffd@0 13
wolffd@0 14 properties
wolffd@0 15
wolffd@0 16 % ---
wolffd@0 17 % Set default parameters
wolffd@0 18 % ---
wolffd@0 19 my_basetype = 'MTTMixedFeatureSlaney08GenreBasicSm';
wolffd@0 20
wolffd@0 21 my_params = MTTAudioFeature.inherited_params(...
wolffd@0 22 'MTTMixedFeatureSlaney08GenreBasicSm', ...
wolffd@0 23 ... % ---
wolffd@0 24 ... % Following: RBM params
wolffd@0 25 ... % ---
wolffd@0 26 'norm_pre_rbm', 0, ... % norm before RBM?
wolffd@0 27 'norm_post_rbm', 0, ... % norm before RBM?
wolffd@0 28 'rbm_hidNum',200, ... % number of hidden units
wolffd@0 29 'rbm_eNum', 100, ...
wolffd@0 30 'rbm_bNum', 1, ...
wolffd@0 31 'rbm_gNum', 1, ...
wolffd@0 32 'rbm_lrate1' , 0.05, ... % initial learning rate
wolffd@0 33 'rbm_lrate2', 0.1, ... % learning rate
wolffd@0 34 'rbm_momentum', 0.5, ...
wolffd@0 35 'rbm_cost', 0.00002, ... % cost function
wolffd@0 36 'rbm_N', 50, ...
wolffd@0 37 'rbm_MAX_INC', 10 ...
wolffd@0 38 );
wolffd@0 39 end
wolffd@0 40 % ---
wolffd@0 41 % member functions
wolffd@0 42 % ---
wolffd@0 43 methods
wolffd@0 44
wolffd@0 45 % ---
wolffd@0 46 % constructor: pointer to feature in database
wolffd@0 47 % ---
wolffd@0 48 function feature = MTTMixedFeatureSlaney08GenreBasicSmRBM(varargin)
wolffd@0 49
wolffd@0 50 feature = feature@MTTAudioFeature(varargin{:});
wolffd@0 51 end
wolffd@0 52
wolffd@0 53 % ---
wolffd@0 54 % load feature data from xml file
wolffd@0 55 % ---
wolffd@0 56 function data = extract(feature, clip)
wolffd@0 57 % load feature data by parsing xml
wolffd@0 58
wolffd@0 59 global globalvars;
wolffd@0 60
wolffd@0 61 % ---
wolffd@0 62 % we extract the base features, and save
wolffd@0 63 % the pointers to these.
wolffd@0 64 % the main work is then done in the define_global_transf
wolffd@0 65 % and finalise functions.
wolffd@0 66 % ---
wolffd@0 67 data.basefeat = clip.features(feature.my_basetype,...
wolffd@0 68 feature.my_params);
wolffd@0 69
wolffd@0 70 % save info data
wolffd@0 71 data.info.type = class(feature);
wolffd@0 72 data.info.owner = clip;
wolffd@0 73 data.info.owner_id = clip.id;
wolffd@0 74 data.info.creatorrev = feature.my_revision;
wolffd@0 75
wolffd@0 76 % save param data
wolffd@0 77 data.info.params = feature.my_params;
wolffd@0 78
wolffd@0 79 % prepare field for final features
wolffd@0 80 data.final.vector = [];
wolffd@0 81 data.final.dim = 0;
wolffd@0 82 data.final.vector_info.labels = {};
wolffd@0 83 end
wolffd@0 84
wolffd@0 85 function define_global_transform(features)
wolffd@0 86
wolffd@0 87 if numel(features) == 1
wolffd@0 88 error ('Insert feature array for this method');
wolffd@0 89 end
wolffd@0 90
wolffd@0 91 % ---
wolffd@0 92 % We collect all the relevant genretag
wolffd@0 93 % features and get the transform on this basis.
wolffd@0 94 % ---
wolffd@0 95 for i = 1:numel(features)
wolffd@0 96 basef(i) = features(i).data.basefeat;
wolffd@0 97 end
wolffd@0 98
wolffd@0 99 % call the features own transsform function
wolffd@0 100 basef.define_global_transform();
wolffd@0 101
wolffd@0 102 % ---
wolffd@0 103 % finalise the basic features, and
wolffd@0 104 % get the feature vectors;
wolffd@0 105 % ---
wolffd@0 106 X = basef.vector();
wolffd@0 107
wolffd@0 108 % check dataset dimension
wolffd@0 109 if numel(features) < basef.dim;
wolffd@0 110
wolffd@0 111 error ('Not enough feature vectors for RBM calculation. need %d samples', ...
wolffd@0 112 basef.dim);
wolffd@0 113 end
wolffd@0 114
wolffd@0 115 % ---
wolffd@0 116 % NOTE: should the data be normalised and scaled to -1:1
wolffd@0 117 % instead of being in a range of 0-1 AND max-min = 1
wolffd@0 118 % ---
wolffd@0 119 if features(1).my_params.norm_pre_rbm == 1
wolffd@0 120
wolffd@0 121 [X, pstd] = mapminmax(X,-1,1);
wolffd@0 122 common.rbm.pre_norm = pstd;
wolffd@0 123 elseif features(1).my_params.norm_pre_rbm == 2
wolffd@0 124
wolffd@0 125 [X, pstd] = mapstd(X,0,1);
wolffd@0 126 common.rbm.pre_norm = pstd;
wolffd@0 127 end
wolffd@0 128
wolffd@0 129
wolffd@0 130 % ---
wolffd@0 131 % TRAIN RBM
wolffd@0 132 % ---
wolffd@0 133 conf.sNum = size(X,2);
wolffd@0 134 conf.eNum = features(1).my_params.rbm_eNum;
wolffd@0 135 conf.bNum = features(1).my_params.rbm_bNum;
wolffd@0 136 conf.gNum = features(1).my_params.rbm_gNum;
wolffd@0 137 conf.hidNum = features(1).my_params.rbm_hidNum;
wolffd@0 138 conf.MAX_INC = features(1).my_params.rbm_MAX_INC;
wolffd@0 139 conf.N = features(1).my_params.rbm_N;
wolffd@0 140 conf.params = [ features(1).my_params.rbm_lrate1 features(1).my_params.rbm_lrate2 ...
wolffd@0 141 features(1).my_params.rbm_momentum features(1).my_params.rbm_cost];
wolffd@0 142 W1 = zeros(0,0);
wolffd@0 143 [common.rbm.W1 common.rbm.vB1 common.rbm.hB1] = training_rbm_(conf,W1,X');
wolffd@0 144
wolffd@0 145 % ---
wolffd@0 146 % This transforms the features using the learnt RBM
wolffd@0 147 % ---
wolffd@0 148 Y = logistic(X' * common.rbm.W1 + repmat(common.rbm.hB1,conf.sNum,1))';
wolffd@0 149
wolffd@0 150
wolffd@0 151 % normalise values after processing
wolffd@0 152 if features(1).my_params.norm_post_rbm
wolffd@0 153
wolffd@0 154 [Y,pmm] = mapminmax(Y,0,1);
wolffd@0 155 common.rbm.post_norm = pmm;
wolffd@0 156 end
wolffd@0 157
wolffd@0 158 % ---
wolffd@0 159 % set common feature values for mixed features
wolffd@0 160 % ---
wolffd@0 161 features(1).my_db.set_common(common);
wolffd@0 162
wolffd@0 163 % save the transformed features straight away!
wolffd@0 164 features.finalise(Y);
wolffd@0 165 end
wolffd@0 166
wolffd@0 167 function finalise(feature, final)
wolffd@0 168 % applies a final transformation and
wolffd@0 169 % collects the information of this feature within a single vector
wolffd@0 170 % see info for types in specific dimensions
wolffd@0 171
wolffd@0 172 max_size = feature(1).my_params.rbm_hidNum;
wolffd@0 173
wolffd@0 174 % prepare information
wolffd@0 175 info = {'RBM'};
wolffd@0 176 info(2:max_size) = num2cell(2:max_size);
wolffd@0 177
wolffd@0 178 % check if features have been finalised already
wolffd@0 179 if nargin == 2 && isempty(final)
wolffd@0 180
wolffd@0 181 % the final vector etc already are set to zero;
wolffd@0 182 return;
wolffd@0 183
wolffd@0 184 elseif nargin == 2 && (numel(feature) == size(final, 2))
wolffd@0 185
wolffd@0 186 for i = 1:numel(feature)
wolffd@0 187
wolffd@0 188 % save final vector and description
wolffd@0 189 feature(i).data.final.vector = final(:,i);
wolffd@0 190 feature(i).data.final.dim = max_size;
wolffd@0 191 feature(i).data.final.vector_info.labels = info;
wolffd@0 192 end
wolffd@0 193
wolffd@0 194 else
wolffd@0 195 % features have to be transformed first
wolffd@0 196 % ---
wolffd@0 197 % TODO: this code remains untested
wolffd@0 198 % ---
wolffd@0 199
wolffd@0 200 % check for neccesary parameters
wolffd@0 201 if isempty(feature(1).my_db.commondb)
wolffd@0 202
wolffd@0 203 error('Define the global transformation first')
wolffd@0 204 return;
wolffd@0 205 end
wolffd@0 206
wolffd@0 207
wolffd@0 208 for i = 1:numel(feature)
wolffd@0 209
wolffd@0 210 % check for neccesary parameters
wolffd@0 211 if isempty(feature(i).my_db.commondb)
wolffd@0 212
wolffd@0 213 error('Define the global transformation first')
wolffd@0 214 end
wolffd@0 215
wolffd@0 216 % ---
wolffd@0 217 % get feature vector and apply transformation
wolffd@0 218 % ---
wolffd@0 219 X = feature(i).data.basefeat.vector();
wolffd@0 220
wolffd@0 221 % ---
wolffd@0 222 % apply normalisation used for removing mean
wolffd@0 223 % in training data
wolffd@0 224 % ---
wolffd@0 225 if feature(1).my_params.norm_pre_rbm == 1
wolffd@0 226
wolffd@0 227 X = mapminmax('apply', X, feature(1).common.rbm.pre_norm);
wolffd@0 228 elseif feature(1).my_params.norm_pre_rbm == 2
wolffd@0 229
wolffd@0 230 X = mapstd('apply', X, feature(1).common.rbm.pre_norm);
wolffd@0 231 end
wolffd@0 232
wolffd@0 233 % ---
wolffd@0 234 % RBM: This transforms the features using the learnt RBM
wolffd@0 235 % ---
wolffd@0 236 conf.sNum = size(X,1);
wolffd@0 237 vec = logistic(X * common.rbm.W1 + repmat(common.rbm.hB1,conf.sNum,1));
wolffd@0 238
wolffd@0 239
wolffd@0 240 % normalise pca values after transformation
wolffd@0 241 if feature(1).my_params.norm_post_rbm
wolffd@0 242
wolffd@0 243 vec = mapminmax('apply', vec,...
wolffd@0 244 feature(1).common.rbm.post_norm);
wolffd@0 245 end
wolffd@0 246
wolffd@0 247 % ---
wolffd@0 248 % cut vector to final size.
wolffd@0 249 % NOTE: this should be done before
wolffd@0 250 % transformation to reduce computation time
wolffd@0 251 % ---
wolffd@0 252 vec = vec(1:max_size);
wolffd@0 253
wolffd@0 254 % save final vector and description
wolffd@0 255 feature(i).data.final.vector = vec;
wolffd@0 256 feature(i).data.final.dim = numel(vec);
wolffd@0 257 feature(i).data.final.vector_info.labels = info;
wolffd@0 258 end
wolffd@0 259 end
wolffd@0 260 end
wolffd@0 261
wolffd@0 262 end
wolffd@0 263 end