| 145 | // ---------------------------------------------------------------------------------------- |
| 146 | |
| 147 | int main() |
| 148 | { |
| 149 | std::vector<sample_type> samples; |
| 150 | std::vector<string> labels; |
| 151 | |
| 152 | // First, get our labeled set of training data |
| 153 | generate_data(samples, labels); |
| 154 | |
| 155 | cout << "samples.size(): "<< samples.size() << endl; |
| 156 | |
| 157 | // Define the trainer we will use. The second template argument specifies the type |
| 158 | // of label used, which is string in this case. |
| 159 | typedef one_vs_one_trainer<any_trainer<sample_type>, string> ovo_trainer; |
| 160 | |
| 161 | |
| 162 | ovo_trainer trainer; |
| 163 | |
| 164 | // Now tell the one_vs_one_trainer that, by default, it should use the simple_custom_trainer |
| 165 | // to solve the individual binary classification subproblems. |
| 166 | trainer.set_trainer(simple_custom_trainer()); |
| 167 | |
| 168 | // Next, to make things a little more interesting, we will setup the one_vs_one_trainer |
| 169 | // to use kernel ridge regression to solve the upper_left vs lower_right binary classification |
| 170 | // subproblem. |
| 171 | typedef radial_basis_kernel<sample_type> rbf_kernel; |
| 172 | krr_trainer<rbf_kernel> rbf_trainer; |
| 173 | rbf_trainer.set_kernel(rbf_kernel(0.1)); |
| 174 | trainer.set_trainer(rbf_trainer, "upper_left", "lower_right"); |
| 175 | |
| 176 | |
| 177 | // Now let's do 5-fold cross-validation using the one_vs_one_trainer we just setup. |
| 178 | // As an aside, always shuffle the order of the samples before doing cross validation. |
| 179 | // For a discussion of why this is a good idea see the svm_ex.cpp example. |
| 180 | randomize_samples(samples, labels); |
| 181 | cout << "cross validation: \n" << cross_validate_multiclass_trainer(trainer, samples, labels, 5) << endl; |
| 182 | // This dataset is very easy and everything is correctly classified. Therefore, the output of |
| 183 | // cross validation is the following confusion matrix. |
| 184 | /* |
| 185 | 50 0 0 0 |
| 186 | 0 50 0 0 |
| 187 | 0 0 50 0 |
| 188 | 0 0 0 50 |
| 189 | */ |
| 190 | |
| 191 | |
| 192 | // We can also obtain the decision rule as always. |
| 193 | one_vs_one_decision_function<ovo_trainer> df = trainer.train(samples, labels); |
| 194 | |
| 195 | cout << "predicted label: "<< df(samples[0]) << ", true label: "<< labels[0] << endl; |
| 196 | cout << "predicted label: "<< df(samples[90]) << ", true label: "<< labels[90] << endl; |
| 197 | // The output is: |
| 198 | /* |
| 199 | predicted label: upper_right, true label: upper_right |
| 200 | predicted label: lower_left, true label: lower_left |
| 201 | */ |
| 202 | |
| 203 | |
| 204 | // Finally, let's save our multiclass decision rule to disk. Remember that we have |
nothing calls this directly
no test coverage detected