view reeval/classification/perform_classification.m @ 4:a1f6a08f624c tip

Completed version 0.0.2

function [results] = perform_classification(experiment, db, condition)

  db = svm_calc_kernel(db,'gaussian','square',1:8:size(db.features,2));
  

  optt.kernel_type = 'gaussian';
  optt.C = 2.^[0:4:8];
  optt.gamma = 2.^[-16:4:-8];
  optt.search_depth = 3;

  % This causes the accuracy to be lower than it could!
  switch(experiment)
    case('time_scat_l3')
      optt.full_test_kernel = 0;
    otherwise 
      optt.full_test_kernel = 1;
  end  

  %

  if nargin < 3
    condition = cellstr(['none '; 'fault']);
  else
    condition = cellstr(condition);
  end

  for ii=1:length(condition)

    [train_set,test_set] = createFolds(condition{ii});
    train_set = find(train_set)';
    test_set = find(test_set)';

    % Only consider excerpts that are in src struct

    elems = length(db.src.objects);
    ids = zeros(elems, 1);
    for jj=1:elems
      ids(jj) = db.src.objects(jj).ind;
    end

    train_set = intersect(train_set, ids);
    test_set = intersect(test_set, ids);

    %

    [dev_err_grid,C_grid,gamma_grid] = ...
      svm_adaptive_param_search(db,train_set,[],optt);

    [dev_err,ind] = min(mean(dev_err_grid{end},2));
    C = C_grid{end}(ind);
    gamma = gamma_grid{end}(ind);

    optt1 = optt;
    optt1.C = C;
    optt1.gamma = gamma;

    model = svm_train(db,train_set,optt1);
    labels = svm_test(db,model,test_set);
    err = classif_err(labels,test_set,db.src);

    % dummy renaming of variables
    id = test_set;
    true_label = kron((1:10), ones(1, 100));
    true_label = true_label(test_set)';
    pred_label = labels';

    % saving results in struct
    results.(condition{ii}).err = err;
    results.(condition{ii}).labels = labels;
    results.(condition{ii}).dev_err = dev_err;
    results.(condition{ii}).C = C;
    results.(condition{ii}).gamma = gamma;
    results.(condition{ii}).tab = table(id, true_label, pred_label);

  end

end