| 89 | thundersvm.model_free(c_void_p(self.model)) |
| 90 | |
| 91 | def fit(self, X, y): |
| 92 | if self.model is not None: |
| 93 | thundersvm.model_free(c_void_p(self.model)) |
| 94 | self.model = None |
| 95 | sparse = sp.isspmatrix(X) |
| 96 | self._sparse = sparse and not callable(self.kernel) |
| 97 | X, y = check_X_y(X, y, dtype=np.float64, order='C', accept_sparse='csr') |
| 98 | y = column_or_1d(y, warn=True).astype(np.float64) |
| 99 | |
| 100 | solver_type = SVM_TYPE.index(self._impl) |
| 101 | |
| 102 | if self.gamma == 'auto': |
| 103 | self._gamma = 1.0 / X.shape[1] |
| 104 | else: |
| 105 | self._gamma = self.gamma |
| 106 | if self.kernel not in KERNEL_TYPE: |
| 107 | print("The kernel parameter not recognized, please refer to the document.") |
| 108 | exit() |
| 109 | else: |
| 110 | kernel = KERNEL_TYPE.index(self.kernel) |
| 111 | |
| 112 | fit = self._sparse_fit if self._sparse else self._dense_fit |
| 113 | thundersvm.model_new.restype = c_void_p |
| 114 | self.model = thundersvm.model_new(solver_type) |
| 115 | if self.max_mem_size != -1: |
| 116 | thundersvm.set_memory_size(c_void_p(self.model), self.max_mem_size) |
| 117 | fit(X, y, solver_type, kernel) |
| 118 | if self._train_succeed[0] == -1: |
| 119 | print("Training failed!") |
| 120 | return |
| 121 | self.n_sv = thundersvm.n_sv(c_void_p(self.model)) |
| 122 | csr_row = (c_int * (self.n_sv + 1))() |
| 123 | csr_col = (c_int * (self.n_sv * self.n_features))() |
| 124 | csr_data = (c_float * (self.n_sv * self.n_features))() |
| 125 | data_size = (c_int * 1)() |
| 126 | sv_indices = (c_int * self.n_sv)() |
| 127 | thundersvm.get_sv(csr_row, csr_col, csr_data, data_size, sv_indices, c_void_p(self.model)) |
| 128 | self.row = np.frombuffer(csr_row, dtype=np.int32) |
| 129 | self.col = np.frombuffer(csr_col, dtype=np.int32)[:data_size[0]] |
| 130 | self.data = np.frombuffer(csr_data, dtype=np.float32)[:data_size[0]] |
| 131 | |
| 132 | self.support_vectors_ = sp.csr_matrix((self.data, self.col, self.row)) |
| 133 | if not self._sparse: |
| 134 | self.support_vectors_ = self.support_vectors_.toarray(order='C') |
| 135 | self.support_ = np.frombuffer(sv_indices, dtype=np.int32).astype(int) |
| 136 | |
| 137 | dual_coef = (c_float * ((self.n_classes - 1) * self.n_sv))() |
| 138 | thundersvm.get_coef(dual_coef, self.n_classes, self.n_sv, c_void_p(self.model)) |
| 139 | |
| 140 | self.dual_coef_ = np.frombuffer(dual_coef, dtype=np.float32)\ |
| 141 | .astype(float)\ |
| 142 | .reshape((self.n_classes - 1, self.n_sv)) |
| 143 | |
| 144 | rho_size = int(self.n_classes * (self.n_classes - 1) / 2) |
| 145 | self.n_binary_model = rho_size |
| 146 | rho = (c_float * rho_size)() |
| 147 | thundersvm.get_rho(rho, rho_size, c_void_p(self.model)) |
| 148 | self.intercept_ = np.frombuffer(rho, dtype=np.float32).astype(float) |