crossvalidation.m

   1 % Crossvalidation to estimate the best parameters
   2
   3 clear all;
   4
   5 %% common part
   6
   7 FEATURE_EXTRACTION_METHOD = 2;
   8
   9 disp('Reading the data...');
  10 read_data; % has to be configured first - see the code
  11
  12 % Set the range of potential boundary height values
  13 min_height = 1;
  14 max_height = size(Visuals{1}, 1);
  15
  16 disp('Extracting features...');
  17 Features = cell(size(Visuals));
  18 for i=1:length(Features)
  19         Features{i} = extract_features(Visuals{i}, FEATURE_EXTRACTION_METHOD);
  20 end
  21
  22 disp('Generating the hashtable for all possible feature vectors... (this may take a while)');
  23 X_ht = get_all_possible_x(FEATURE_EXTRACTION_METHOD, min_height:max_height);
  24
  25
  26
  27 %% <--
  28
  29 %% the crossvalidation part
  30 kernel_types = {'gaussian', 'flat'};
  31 kernel_widths = [0.02 0.1 0.5 1 2 4 7];
  32 hmm_fb_methods = {'mean', 'map'};
  33
  34 for kt = 1:length(kernel_types)
  35         for kw = 1:length(kernel_widths)
  36                 for fbm = 1:length(hmm_fb_methods)
  37                         fprintf(1, '====== Iteration for parameters (%s, %i, %s) ======\n', kernel_types{kt}, kernel_widths(kw), hmm_fb_methods{fbm});
  38                         disp('Computing model parameters...');
  39                         clear s S X;
  40                         [s, S, X] = get_probabilities(Features, Boundary1, X_ht, min_height, max_height, FEATURE_EXTRACTION_METHOD, kernel_types{kt}, kernel_widths(kw));
  41
  42                         % Inference:
  43                         fim = Features{1};
  44                         true_sequence = Boundary1{1};
  45
  46
  47                         % create emission probabilities matrix as requested by
  48                         % hmm_viterbi()
  49                         X_viterbi = zeros(size(fim, 2), length(s));
  50                         for i=1:size(X_viterbi, 1)
  51                                 fid = num2str(fim(:, i)'); % convert the feature to string -> feature id
  52                                 X_viterbi(i, :) = X(X_ht.get(fid), :); % get this features' P(feature|state) for all states
  53                         end
  54                         % <--
  55                         disp('Inferring boundary 1...');
  56                         inferred_sequence = hmm_forward_backward(s, S', X_viterbi, hmm_fb_methods{fbm});
  57
  58                         true_sequence_cleaned = true_sequence(true_sequence > 1); % remove missing data
  59                         inferred_sequence_cleaned = inferred_sequence(true_sequence > 1);
  60
  61                         dist = sum(norm(inferred_sequence_cleaned - true_sequence_cleaned'));
  62                         fprintf(1, '====== Total euclidean distance between the true and the predicted sequence is: %f\n', dist);
  63                 end
  64         end
  65 end
  66
  67