gerard@1: % Copyright 2013 MUSIC TECHNOLOGY GROUP, UNIVERSITAT POMPEU FABRA
gerard@1: %
gerard@1: % Written by Gerard Roma <gerard.roma@upf.edu>
gerard@1: %
gerard@1: % This program is free software: you can redistribute it and/or modify
gerard@1: % it under the terms of the GNU Affero General Public License as published by
gerard@1: % the Free Software Foundation, either version 3 of the License, or
gerard@1: % (at your option) any later version.
gerard@1: %
gerard@1: % This program is distributed in the hope that it will be useful,
gerard@1: % but WITHOUT ANY WARRANTY; without even the implied warranty of
gerard@1: % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
gerard@1: % GNU Affero General Public License for more details.
gerard@1: %
gerard@1: % You should have received a copy of the GNU Affero General Public License
gerard@1: % along with this program.  If not, see <http://www.gnu.org/licenses/>.
gerard@1: 
gerard@1: function [ predict_label ] = classify_scenes(tmp_path, train_file,test_file, output_file, grid_search)
gerard@1:     addpath('rastamat/');
gerard@1:     addpath('libsvm/')
gerard@1:     
gerard@1:     classes = {'bus' 'busystreet' 'office' 'openairmarket' 'park' 'quietstreet' 'restaurant' 'supermarket' 'tube' 'tubestation'};
gerard@1:     
gerard@1:     [tr_names,tr_labels] = loadClassificationOutput(train_file);
gerard@1:     [te_names,te_labels] = loadClassificationOutput(test_file);
gerard@1:     
gerard@1:     
gerard@1:     %% extract features
gerard@1:     disp('analyzing training files');
gerard@1:     train_features = analyze_files(tr_names,tmp_path);
gerard@1:     disp('analyzing test files');
gerard@1:     test_features = analyze_files(te_names, tmp_path);
gerard@1:     train_labels = [];
gerard@1:     test_labels = ones(length(te_names),1);
gerard@1:     
gerard@1:     train_labels = get_class_indices(tr_labels);
gerard@1:     
gerard@1:     [train_z, mu, sigma] = zscore(train_features);
gerard@1:     MU = repmat(mu,size(test_features,1),1);
gerard@1:     SIGMA = repmat(sigma,size(test_features,1),1);
gerard@1:     test_z = (test_features -MU)./SIGMA;
gerard@1:     
gerard@1:     
gerard@1:     %% SVM grid Search inspired in http://labrosa.ee.columbia.edu/projects/consumervideo/
gerard@1:     
gerard@1:     if grid_search
gerard@1:         disp('performing grid search');
gerard@1:         tuning_data_fraction = 0.9;
gerard@1:         train_size = round(tuning_data_fraction*size(train_z,1));
gerard@1:         p = randperm(size(train_z,1));
gerard@1:         trainX = train_z(p(1:train_size),:);
gerard@1:         validateX = train_z(p(train_size+1:end),:);
gerard@1:         trainY = get_class_indices(tr_labels(p(1:train_size)));
gerard@1:         validateY = get_class_indices(tr_labels(p(train_size+1:end)));
gerard@1:                 
gerard@1:         
gerard@1:         gamma = 2.^[-10:1:10];
gerard@1:         C = 2.^[0:13];
gerard@1:         
gerard@1:         params = zeros(length(gamma),length(C));
gerard@1:         best_a = 0;
gerard@1:         best_g = 0;
gerard@1:         best_C = 0;
gerard@1:      
gerard@1:         for gi= 1:length(gamma)
gerard@1:             for Ci = 1:length(C)
gerard@1:                 m = svmtrain(trainY', trainX, sprintf('-c %d -g %2.5f -q',C(Ci),gamma(gi)));
gerard@1:                 [p,a] = svmpredict(validateY', validateX, m, '-q');
gerard@1:                 if a(1) > best_a
gerard@1:                     best_C = C(Ci);
gerard@1:                     best_g = gamma(gi);
gerard@1:                     best_a = a(1);
gerard@1:                 end
gerard@1:             end
gerard@1:         end
gerard@1:         disp('grid search done');
gerard@1:     else
gerard@1:         best_C=70;
gerard@1:         best_g = 0.003;
gerard@1:     end
gerard@1:  
gerard@1:     
gerard@1:     model = svmtrain(train_labels', train_z, sprintf('-c %d -g %2.5f -q',best_C,best_g));
gerard@1:     [predict_indices, accuracy_obj, prob_values] = svmpredict(test_labels, test_z, model,'-q');
gerard@1:     predict_label = classes(predict_indices);
gerard@1:     
gerard@1:     
gerard@1:     outfd = fopen(output_file,'w+');
gerard@1:     
gerard@1:     for i = 1:length(te_names)
gerard@1:          fprintf(outfd,'%s\t',[char(te_names(i))]);
gerard@1:          fprintf(outfd,'%s\n',[char(predict_label(i))]);
gerard@1:     end
gerard@1:     fclose(outfd);    
gerard@1:     
gerard@1:     function idx = get_class_indices(labels)
gerard@1:         idx = [];            
gerard@1:         for i=1:length(labels)
gerard@1:         class =  find(strcmp(classes,labels(i)));
gerard@1:         idx(i) = class;
gerard@1:     end
gerard@1:         
gerard@1: end
gerard@1: 
gerard@1:     
gerard@1: end
gerard@1: