| 202 | // ---------------------------------------------------------------------------------------- |
| 203 | |
| 204 | int main() |
| 205 | { |
| 206 | // We need a dataset to test the machine learning algorithms. So we are going to |
| 207 | // define a HMM based on the following two matrices and then randomly sample a |
| 208 | // set of data from it. Then we will see if the machine learning method can |
| 209 | // recover the HMM model from the training data. |
| 210 | |
| 211 | |
| 212 | matrix<double> transition_probabilities(num_label_states, num_label_states); |
| 213 | transition_probabilities = 0.05, 0.90, 0.05, |
| 214 | 0.05, 0.05, 0.90, |
| 215 | 0.90, 0.05, 0.05; |
| 216 | |
| 217 | matrix<double> emission_probabilities(num_label_states,num_sample_states); |
| 218 | emission_probabilities = 0.5, 0.5, 0.0, |
| 219 | 0.0, 0.5, 0.5, |
| 220 | 0.5, 0.0, 0.5; |
| 221 | |
| 222 | std::vector<std::vector<unsigned long> > samples; |
| 223 | std::vector<std::vector<unsigned long> > labels; |
| 224 | // sample 1000 labeled sequences from the HMM. |
| 225 | make_dataset(transition_probabilities,emission_probabilities, |
| 226 | samples, labels, 1000); |
| 227 | |
| 228 | // print out some of the randomly sampled sequences |
| 229 | for (int i = 0; i < 10; ++i) |
| 230 | { |
| 231 | cout << "hidden states: " << trans(mat(labels[i])); |
| 232 | cout << "observed states: " << trans(mat(samples[i])); |
| 233 | cout << "******************************" << endl; |
| 234 | } |
| 235 | |
| 236 | // Next we use the structural_sequence_labeling_trainer to learn our |
| 237 | // prediction model based on just the samples and labels. |
| 238 | structural_sequence_labeling_trainer<feature_extractor> trainer; |
| 239 | // This is the common SVM C parameter. Larger values encourage the |
| 240 | // trainer to attempt to fit the data exactly but might overfit. |
| 241 | // In general, you determine this parameter by cross-validation. |
| 242 | trainer.set_c(4); |
| 243 | // This trainer can use multiple CPU cores to speed up the training. |
| 244 | // So set this to the number of available CPU cores. |
| 245 | trainer.set_num_threads(4); |
| 246 | |
| 247 | |
| 248 | // Learn to do sequence labeling from the dataset |
| 249 | sequence_labeler<feature_extractor> labeler = trainer.train(samples, labels); |
| 250 | |
| 251 | // Test the learned labeler on one of the training samples. In this |
| 252 | // case it will give the correct sequence of labels. |
| 253 | std::vector<unsigned long> predicted_labels = labeler(samples[0]); |
| 254 | cout << "true hidden states: "<< trans(mat(labels[0])); |
| 255 | cout << "predicted hidden states: "<< trans(mat(predicted_labels)); |
| 256 | |
| 257 | |
| 258 | |
| 259 | // We can also do cross-validation. The confusion_matrix is defined as: |
| 260 | // - confusion_matrix(T,P) == the number of times a sequence element with label T |
| 261 | // was predicted to have a label of P. |
nothing calls this directly
no test coverage detected