run classify svm skl

%% Classification using svm
%
% #   For CoSMoMVPA's copyright information and license terms,   #
% #   see the COPYING file distributed with CoSMoMVPA.           #

if ~cosmo_check_external('svm', false)
    fprintf('No SVM implementation is available, skipping\n');
    return
end

%% Define data
config = cosmo_config();
data_path = fullfile(config.tutorial_data_path, 'ak6', 's01');

data_fn = fullfile(data_path, 'glm_T_stats_perrun.nii');
mask_fn = fullfile(data_path, 'vt_mask.nii');
ds = cosmo_fmri_dataset(data_fn, 'mask', mask_fn, ...
                        'targets', repmat(1:6, 1, 10)', ...
                        'chunks', floor(((1:60) - 1) / 6) + 1)';

% remove constant features
ds = cosmo_remove_useless_data(ds);

%% Two class classification
ds_2class = cosmo_slice(ds, ds.sa.targets == 2 | ds.sa.targets == 5);
ds_2class_train = cosmo_slice(ds_2class, ds_2class.sa.chunks <= 5);
ds_2class_test = cosmo_slice(ds_2class, ds_2class.sa.chunks > 5);

pred2 = cosmo_classify_svm(ds_2class_train.samples, ...
                           ds_2class_train.sa.targets, ...
                           ds_2class_test.samples);

fprintf('2-class accuracy %.3f\n', sum(pred2 == ds_2class_test.sa.targets) / numel(pred2));
mx = cosmo_confusion_matrix(ds_2class_test.sa.targets, pred2);
imagesc(mx);
colorbar();

%% Four class classification
ds_4class = cosmo_slice(ds, ds.sa.targets >= 2 & ds.sa.targets <= 5);
ds_4class_train = cosmo_slice(ds_4class, ds_4class.sa.chunks <= 5);
ds_4class_test = cosmo_slice(ds_4class, ds_4class.sa.chunks > 5);

pred4 = cosmo_classify_svm(ds_4class_train.samples, ...
                           ds_4class_train.sa.targets, ...
                           ds_4class_test.samples);

fprintf('4-class: accuracy %.3f\n', sum(pred4 == ds_4class_test.sa.targets) / numel(pred4));

mx = cosmo_confusion_matrix(ds_4class_test.sa.targets, pred4);
imagesc(mx);
colorbar();