Mercurial > hg > camir-aes2014

annotate core/magnatagatune/tests_evals/rbm_subspace/Exp_normalise_deltas.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 % Experiment code templat %
wolffd@0 3 % Project: sub-euclidean distance for music similarity,
wolffd@0 4 % in the last part all the
wolffd@0 5 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 6 %% Load features
wolffd@0 7 feature_file = 'rel_music_raw_features.mat';
wolffd@0 8 vars = whos('-file', feature_file);
wolffd@0 9 A = load(feature_file,vars(1).name,vars(2).name,vars(3).name,vars(4).name);
wolffd@0 10 raw_features = A.(vars(1).name);
wolffd@0 11 indices = A.(vars(2).name);
wolffd@0 12 tst_inx = A.(vars(3).name);
wolffd@0 13 trn_inx = A.(vars(4).name);
wolffd@0 14 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 15 % Define directory to save parameters & results
wolffd@0 16 % dir = '/home/funzi/Documents/';
wolffd@0 17 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 18 dmr = [0 5 10 20 30 50]; % dimension reduction by PCA
wolffd@0 19 ws = [0 5 10 20 30 50 70]; % window size
wolffd@0 20 % parameters of rbm (if it is used for extraction)
wolffd@0 21 hidNum = [30 50 100 500];
wolffd@0 22 lr_1 = [0.05 0.1 0.5];
wolffd@0 23 lr_2 = [0.1 0.5 0.7];
wolffd@0 24 mmt = [0.02 0.05 0.1];
wolffd@0 25 cost = [0.00002 0.01 0.1];
wolffd@0 26
wolffd@0 27 %% Select parameters (if grid-search is not applied)
wolffd@0 28 di = 1;
wolffd@0 29 wi = 1;
wolffd@0 30 hi = 1;
wolffd@0 31 l1i = 1;
wolffd@0 32 l2i = 1;
wolffd@0 33 mi = 1;
wolffd@0 34 ci = 1;
wolffd@0 35 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 36 % If grid search is define
wolffd@0 37 % log_file = strcat(dir,'exp_.mat');
wolffd@0 38 % inx = resume_from_grid(log_file,8);
wolffd@0 39 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 40 %% Feature extraction
wolffd@0 41 EXT_TYPE = 2;
wolffd@0 42 switch (EXT_TYPE)
wolffd@0 43 case 1 % Using PCA
wolffd@0 44 assert(~exist('OCTAVE_VERSION'),'This script cannot run in octave');
wolffd@0 45 coeff = princomp(raw_features);
wolffd@0 46 coeff = coeff(:,1:end-dmr(di)); % Change value of dmr(di) to reduce the dimensionality
wolffd@0 47 features = raw_features*coeff;
wolffd@0 48 % normalizing
wolffd@0 49 mm = minmax(features')';
wolffd@0 50 inn= (find(mm(1,:)~=mm(2,:)));
wolffd@0 51 mm = mm(:,inn);
wolffd@0 52 features = features(:,inn);
wolffd@0 53 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 54 case 2 % Using rbm
wolffd@0 55 conf.hidNum = hidNum(hi);
wolffd@0 56 conf.eNum = 100;
wolffd@0 57 conf.sNum = size(raw_features,1);
wolffd@0 58 conf.bNum = 1;
wolffd@0 59 conf.gNum = 1;
wolffd@0 60 conf.params = [lr_1(l1i) lr_2(l2i) mmt(mi) cost(ci)];
wolffd@0 61 conf.N = 50;
wolffd@0 62 conf.MAX_INC = 10;
wolffd@0 63 W1 = zeros(0,0);
wolffd@0 64 [W1 vB1 hB1] = training_rbm_(conf,W1,raw_features);
wolffd@0 65 features = raw_features*W1 + repmat(hB1,conf.sNum,1);
wolffd@0 66 end
wolffd@0 67
wolffd@0 68 %% Sub-euclidean computation
wolffd@0 69 num_case = size(trn_inx,1);
wolffd@0 70 trnd_12 = cell(1,num_case);
wolffd@0 71 trnd_13 = cell(1,num_case);
wolffd@0 72 tstd_12 = cell(1,num_case);
wolffd@0 73 tstd_13 = cell(1,num_case);
wolffd@0 74
wolffd@0 75 w = ws(wi);
wolffd@0 76
wolffd@0 77 % w = subspace window size
wolffd@0 78 if w == 0 % trnd_12 = d(a,b) , trnd_13= d(a,c)
wolffd@0 79 for i = 1:num_case % over all cross-validation folds (num_case)
wolffd@0 80 [trnd_12{i} trnd_13{i}] = simple_dist(trn_inx{i},features,indices);
wolffd@0 81 [tstd_12{i} tstd_13{i}] = simple_dist(tst_inx{i},features,indices);
wolffd@0 82 end
wolffd@0 83 else
wolffd@0 84 for i = 1:num_case % for w > 1
wolffd@0 85 [trnd_12{i} trnd_13{i}] = conv_euclidean_dist(trn_inx{i},features,indices,w,1); %% normalize is better than no normalize
wolffd@0 86 [tstd_12{i} tstd_13{i}] = conv_euclidean_dist(tst_inx{i},features,indices,w,1);
wolffd@0 87 end
wolffd@0 88 end
wolffd@0 89 %% Data preparation
wolffd@0 90 trn_dat1 = cell(1,num_case);
wolffd@0 91 trn_dat2 = cell(1,num_case);
wolffd@0 92 tst_dat1 = cell(1,num_case);
wolffd@0 93 tst_dat2 = cell(1,num_case);
wolffd@0 94
wolffd@0 95 for i=1:num_case
wolffd@0 96 %=> Compute hypothesis
wolffd@0 97 trn_dat1{i} = trnd_13{i} - trnd_12{i};
wolffd@0 98 trn_dat2{i} = trnd_12{i} - trnd_13{i};
wolffd@0 99 tst_dat1{i} = tstd_13{i} - tstd_12{i};
wolffd@0 100 tst_dat2{i} = tstd_12{i} - tstd_13{i};
wolffd@0 101
wolffd@0 102
wolffd@0 103 % ---
wolffd@0 104 % Cheat: Normalize over all training and test delta values using min-max
wolffd@0 105 % Son reports this can give about 95% accuracy
wolffd@0 106 % ---
wolffd@0 107
wolffd@0 108 mm = minmax([trn_dat1{i};tst_dat1{i}]')';
wolffd@0 109 inn= find(mm(1,:)~=mm(2,:));
wolffd@0 110 mm = mm(:,inn);
wolffd@0 111 trn_dat1{i} =
wolffd@0 112 (trn_dat1{i}(:,inn)-repmat(mm(1,:),size(trn_dat1{i},1),1))./repmat(mm(2,:)-mm(1,:),size(trn_dat1{i},1),1);
wolffd@0 113 tst_dat1{i} = (tst_dat1{i}(:,inn)-repmat(mm(1,:),size(tst_dat1{i},1),1))./repmat(mm(2,:)-mm(1,:),size(tst_dat1{i},1),1);
wolffd@0 114
wolffd@0 115 mm = minmax([trn_dat2{i};tst_dat2{i}]');
wolffd@0 116 inn= find(mm(1,:)~=mm(2,:));
wolffd@0 117 mm = mm(:,inn);
wolffd@0 118 trn_dat2{i} =
wolffd@0 119 (trn_dat2{i}(:,inn)-repmat(mm(1,:),size(trn_dat2{i},1),1))./repmat(mm(2,:)-mm(1,:),size(trn_dat2{i},1),1);
wolffd@0 120 tst_dat2{i} = (tst_dat2{i}(:,inn)-repmat(mm(1,:),size(tst_dat2{i},1),1))./repmat(mm(2,:)-mm(1,:),size(tst_dat2{i},1),1);
wolffd@0 121
wolffd@0 122
wolffd@0 123
wolffd@0 124
wolffd@0 125 end