wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: % Experiment with gradient ascent                                         %
wolffd@0: % Project: sub-euclidean distance for music similarity                    %
wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: %% Load features
wolffd@0: %feature_file = 'rel_music_raw_features.mat';
wolffd@0: feature_file = 'rel_music_raw_features+simdata_ISMIR12.mat';
wolffd@0: 
wolffd@0: vars = whos('-file', feature_file);
wolffd@0: A = load(feature_file,vars(1).name,vars(2).name,vars(3).name,vars(4).name);
wolffd@0: raw_features = A.(vars(1).name);
wolffd@0: indices      = A.(vars(2).name);
wolffd@0: tst_inx      = A.(vars(3).name);
wolffd@0: trn_inx      = A.(vars(4).name);
wolffd@0: %% Params setting
wolffd@0: dmr    = [0 5 10 20 30 50];    % dimension reduction by PCA
wolffd@0: ws     = [0 5 10 20 30 50 70]; % window size
wolffd@0: % parameters of rbm (if it is used for extraction)
wolffd@0: hidNum = 0;
wolffd@0: lr_1   = 0;
wolffd@0: lr_2   = 0;
wolffd@0: mmt    = 0;
wolffd@0: cost   = 0;
wolffd@0: %% Select parameters (if grid-search is not applied)
wolffd@0:  di  = 1;
wolffd@0: 
wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: % If grid search is define
wolffd@0: % Define directory to save parameters & results
wolffd@0: if ~isempty(findstr('WIN',computer()))
wolffd@0:         dir    = 'C:\Pros\Experiments\ISMIR_2013\grad\'; % In windows platform
wolffd@0:         dlm    = '\';
wolffd@0: elseif ~isempty(findstr('linux',computer())) || ~isempty(findstr('LNX',computer()))
wolffd@0:         dir    = '/home/funzi/Documents/Experiments/ISMIR_2013/grad/'; % In lunix platform
wolffd@0:         dlm    = '/';
wolffd@0: end
wolffd@0: 
wolffd@0: EXT_TYPE = 2;
wolffd@0: switch (EXT_TYPE)
wolffd@0:     case 1
wolffd@0:         dir = strcat(dir,'pca',dlm);
wolffd@0:     case 2
wolffd@0:         dir = strcat(dir,'rbm',dlm);
wolffd@0:         
wolffd@0:         hidNum = [100 500 1000 1200];
wolffd@0:         lr_1   = [0.5 0.7];
wolffd@0:         lr_2   = [0.7];
wolffd@0:         mmt    = [0.1];
wolffd@0:         cost   = [0.00002];
wolffd@0:     otherwise            
wolffd@0:         dir = strcat(dir,'none',dlm);
wolffd@0: end
wolffd@0: 
wolffd@0: w_num = size(ws,2);
wolffd@0: 
wolffd@0: for iiii = 1:200 % set the higher range to search for better features in case of ext using rbm
wolffd@0:     log_file = strcat(dir,'exp',num2str(iiii),'.mat')
wolffd@0:     inx = resume_from_grid(log_file,8 + w_num);
wolffd@0:     if inx(end-w_num+1:end)==ones(1,w_num)
wolffd@0:         max_= zeros(1,w_num); 
wolffd@0:     else
wolffd@0:         max_ = inx(end-w_num+1:end);
wolffd@0:     end
wolffd@0: 
wolffd@0:     results = zeros(1,w_num);
wolffd@0:     W_max = cell(1,w_num);
wolffd@0:     vB_max = cell(1,w_num);
wolffd@0:     hB_max = cell(1,w_num);
wolffd@0:     Ws_max = cell(1,w_num);
wolffd@0: 
wolffd@0: for hi = inx(1):size(hidNum,2)
wolffd@0: for l1i = inx(2):size(lr_1,2)
wolffd@0: % for l1i = inx(3):size(lr_2,2)
wolffd@0: for mi  = inx(4):size(mmt,2)
wolffd@0: for ci  = inx(5):size(cost,2)    
wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: %% Feature extraction
wolffd@0: features = raw_features;
wolffd@0: switch (EXT_TYPE)   
wolffd@0:     case 1  % Using PCA        
wolffd@0:         assert(~exist('OCTAVE_VERSION'),'This script cannot run in octave');
wolffd@0:         coeff = princomp(raw_features);
wolffd@0:         coeff = coeff(:,1:6);  % best = 6
wolffd@0:         features = raw_features*coeff;
wolffd@0:           % normalizing
wolffd@0:         mm = minmax(features')';
wolffd@0:         inn= (find(mm(1,:)~=mm(2,:)));
wolffd@0:         mm = mm(:,inn);
wolffd@0:         features = features(:,inn);
wolffd@0:         features = (features-repmat(mm(1,:),size(features,1),1))./(repmat(mm(2,:),size(features,1),1)-repmat(mm(1,:),size(features,1),1));
wolffd@0:     case 2  % Using rbm      
wolffd@0:         conf.hidNum = hidNum(hi);
wolffd@0:         conf.eNum   = 100;
wolffd@0:         conf.sNum   = size(raw_features,1);
wolffd@0:         conf.bNum   = 1;
wolffd@0:         conf.gNum   = 1;
wolffd@0:         conf.params = [lr_1(l1i) lr_1(l1i) mmt(mi) cost(ci)];
wolffd@0:         conf.N    = 50;
wolffd@0:         conf.MAX_INC = 10;
wolffd@0:         W1 = zeros(0,0);
wolffd@0:         [W1 vB1 hB1] = training_rbm_(conf,W1,raw_features);
wolffd@0:         features = logistic(raw_features*W1 + repmat(hB1,conf.sNum,1));
wolffd@0:     otherwise
wolffd@0:         % normalizing
wolffd@0: %         mm = minmax(features')';
wolffd@0: %         inn= (find(mm(1,:)~=mm(2,:)));
wolffd@0: %         mm = mm(:,inn);
wolffd@0: %         features = features(:,inn);
wolffd@0: %         features = (features-repmat(mm(1,:),size(features,1),1))./(repmat(mm(2,:),size(features,1),1)-repmat(mm(1,:),size(features,1),1));
wolffd@0: end
wolffd@0: 
wolffd@0: for wi = inx(6):w_num
wolffd@0: %% Sub-euclidean computation
wolffd@0: w = ws(wi);   % w = subspace window size
wolffd@0: num_case = size(trn_inx,1); 
wolffd@0: [trnd_12 trnd_13] = subspace_distances(trn_inx,features,indices,w,0);
wolffd@0: [tstd_12 tstd_13] = subspace_distances(tst_inx,features,indices,w,0);
wolffd@0: cr_ = 0;   % correct rate for training
wolffd@0: cr  = 0;   % correct rate for testing
wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: %% CODE HERE                               %%
wolffd@0: [Ws cr_] = gradient_ascent(trnd_12,trnd_13,0.1,0.1,0.00002);
wolffd@0: 
wolffd@0: for i = 1:num_case       
wolffd@0:  cr = cr + sum((tstd_13{i}-tstd_12{i})*Ws{i}' > 0, 1)/size(tstd_12{i},1);
wolffd@0: end
wolffd@0: cr = cr/num_case;
wolffd@0: if cr_>max_(wi)
wolffd@0:     max_(wi) = cr_;
wolffd@0:     results(wi)  = cr;
wolffd@0:     if EXT_TYPE==2
wolffd@0:         W_max{wi} = W1;
wolffd@0:         vB_max{wi} = vB1;
wolffd@0:         hB_max{wi} = hB1;
wolffd@0:         Ws_max{wi} = Ws;
wolffd@0:     end
wolffd@0: end
wolffd@0: 
wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: fprintf('[window|train|test]= %2d |%f |%f\n',w,cr_,cr);
wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: % Using the logging function to save paramters
wolffd@0: % and the result for plotting or in grid search
wolffd@0: switch EXT_TYPE
wolffd@0:     case 1
wolffd@0:        % logging(log_file,[100 100 100 100 100 wi cr_ cr max_]);  
wolffd@0:     case 2
wolffd@0:         logging(log_file,[hi l1i l1i mi ci wi cr_ cr max_ conf.hidNum conf.eNum conf.params]);  
wolffd@0:     otherwise
wolffd@0:         logging(log_file,[100 100 100 100 100 wi cr_ cr max_]);  
wolffd@0: end
wolffd@0: end
wolffd@0: inx(6)=1;
wolffd@0: end
wolffd@0: inx(5) = 1;
wolffd@0: end
wolffd@0: inx(4) = 1;
wolffd@0: end
wolffd@0: inx(2) = 1;
wolffd@0: end
wolffd@0: inx(1) = 1;
wolffd@0: %% Test on best features
wolffd@0: 
wolffd@0: save(strcat(dir,'res_',num2str(iiii),'.mat'),'max_','results','W_max','vB_max','hB_max','Ws_max','ws');
wolffd@0: [dummy pos] = max(max_);
wolffd@0: fprintf('Accuracy (RBM best fts): w = %d train = %f test = %f\n',ws(pos),max_(pos),results(pos));
wolffd@0: clc;
wolffd@0: end
wolffd@0: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0: clear;