| 73 | |
| 74 | template <typename psi_type> |
| 75 | matrix<double,0,1> solve_structural_svm_problem_impl( |
| 76 | py::object problem |
| 77 | ) |
| 78 | { |
| 79 | const double C = problem.attr("C").cast<double>(); |
| 80 | const bool be_verbose = py::hasattr(problem,"be_verbose") && problem.attr("be_verbose").cast<bool>(); |
| 81 | const bool use_sparse_feature_vectors = py::hasattr(problem,"use_sparse_feature_vectors") && |
| 82 | problem.attr("use_sparse_feature_vectors").cast<bool>(); |
| 83 | const bool learns_nonnegative_weights = py::hasattr(problem,"learns_nonnegative_weights") && |
| 84 | problem.attr("learns_nonnegative_weights").cast<bool>(); |
| 85 | |
| 86 | double eps = 0.001; |
| 87 | unsigned long max_cache_size = 10; |
| 88 | if (py::hasattr(problem, "epsilon")) |
| 89 | eps = problem.attr("epsilon").cast<double>(); |
| 90 | if (py::hasattr(problem, "max_cache_size")) |
| 91 | max_cache_size = problem.attr("max_cache_size").cast<double>(); |
| 92 | |
| 93 | const long num_samples = problem.attr("num_samples").cast<long>(); |
| 94 | const long num_dimensions = problem.attr("num_dimensions").cast<long>(); |
| 95 | |
| 96 | pyassert(num_samples > 0, "You can't train a Structural-SVM if you don't have any training samples."); |
| 97 | |
| 98 | if (be_verbose) |
| 99 | { |
| 100 | cout << "C: " << C << endl; |
| 101 | cout << "epsilon: " << eps << endl; |
| 102 | cout << "max_cache_size: " << max_cache_size << endl; |
| 103 | cout << "num_samples: " << num_samples << endl; |
| 104 | cout << "num_dimensions: " << num_dimensions << endl; |
| 105 | cout << "use_sparse_feature_vectors: " << std::boolalpha << use_sparse_feature_vectors << endl; |
| 106 | cout << "learns_nonnegative_weights: " << std::boolalpha << learns_nonnegative_weights << endl; |
| 107 | cout << endl; |
| 108 | } |
| 109 | |
| 110 | svm_struct_prob<psi_type> prob(problem, num_dimensions, num_samples); |
| 111 | prob.set_c(C); |
| 112 | prob.set_epsilon(eps); |
| 113 | prob.set_max_cache_size(max_cache_size); |
| 114 | if (be_verbose) |
| 115 | prob.be_verbose(); |
| 116 | |
| 117 | oca solver; |
| 118 | matrix<double,0,1> w; |
| 119 | if (learns_nonnegative_weights) |
| 120 | solver(prob, w, prob.get_num_dimensions()); |
| 121 | else |
| 122 | solver(prob, w); |
| 123 | return w; |
| 124 | } |
| 125 | |
| 126 | // ---------------------------------------------------------------------------------------- |
| 127 |
nothing calls this directly
no test coverage detected