| 172 | // ---------------------------------------------------------------------------------------- |
| 173 | |
| 174 | int main() |
| 175 | { |
| 176 | try |
| 177 | { |
| 178 | // Get the training samples we defined above. |
| 179 | dlib::array<graph_type> samples; |
| 180 | std::vector<std::vector<bool> > labels; |
| 181 | make_training_examples(samples, labels); |
| 182 | |
| 183 | |
| 184 | // Create a structural SVM trainer for graph labeling problems. The vector_type |
| 185 | // needs to be set to a type capable of holding node or edge vectors. |
| 186 | typedef matrix<double,0,1> vector_type; |
| 187 | structural_graph_labeling_trainer<vector_type> trainer; |
| 188 | // This is the usual SVM C parameter. Larger values make the trainer try |
| 189 | // harder to fit the training data but might result in overfitting. You |
| 190 | // should set this value to whatever gives the best cross-validation results. |
| 191 | trainer.set_c(10); |
| 192 | |
| 193 | // Do 3-fold cross-validation and print the results. In this case it will |
| 194 | // indicate that all nodes were correctly classified. |
| 195 | cout << "3-fold cross-validation: " << cross_validate_graph_labeling_trainer(trainer, samples, labels, 3) << endl; |
| 196 | |
| 197 | // Since the trainer is working well. Let's have it make a graph_labeler |
| 198 | // based on the training data. |
| 199 | graph_labeler<vector_type> labeler = trainer.train(samples, labels); |
| 200 | |
| 201 | |
| 202 | /* |
| 203 | Let's try the graph_labeler on a new test graph. In particular, let's |
| 204 | use one with 5 nodes as shown below: |
| 205 | |
| 206 | (0 F)-----(1 T) |
| 207 | | | |
| 208 | | | |
| 209 | | | |
| 210 | (3 T)-----(2 T)------(4 T) |
| 211 | |
| 212 | I have annotated each node with either T or F to indicate the correct |
| 213 | output (true or false). |
| 214 | */ |
| 215 | graph_type g; |
| 216 | g.set_number_of_nodes(5); |
| 217 | g.node(0).data = 1, 0; // Node data indicates a false node. |
| 218 | g.node(1).data = 0, 1; // Node data indicates a true node. |
| 219 | g.node(2).data = 0, 0; // Node data is ambiguous. |
| 220 | g.node(3).data = 0, 0; // Node data is ambiguous. |
| 221 | g.node(4).data = 0.1, 0; // Node data slightly indicates a false node. |
| 222 | |
| 223 | g.add_edge(0,1); |
| 224 | g.add_edge(1,2); |
| 225 | g.add_edge(2,3); |
| 226 | g.add_edge(3,0); |
| 227 | g.add_edge(2,4); |
| 228 | |
| 229 | // Set the edges up so nodes 1, 2, 3, and 4 are all strongly connected. |
| 230 | edge(g,0,1) = 0; |
| 231 | edge(g,1,2) = 1; |
nothing calls this directly
no test coverage detected