(self, X, Y)
| 592 | self.loss = CrossEntropyLoss() |
| 593 | |
| 594 | def fit(self, X, Y): |
| 595 | # 分类问题将 Y 转化为 one-hot 编码 |
| 596 | if not self.is_regression: |
| 597 | Y = to_categorical(Y.flatten()) |
| 598 | else: |
| 599 | Y = Y.reshape(-1, 1) if len(Y.shape) == 1 else Y |
| 600 | self.out_dims = Y.shape[1] |
| 601 | self.trees = np.empty((self.n_estimators, self.out_dims), dtype=object) |
| 602 | Y_pred = np.full(np.shape(Y), np.mean(Y, axis=0)) |
| 603 | self.weights = np.ones((self.n_estimators, self.out_dims)) |
| 604 | self.weights[1:, :] *= self.learning_rate |
| 605 | # 迭代过程 |
| 606 | for i in self.progressbar(range(self.n_estimators)): |
| 607 | for c in range(self.out_dims): |
| 608 | tree = RegressionTree( |
| 609 | min_samples_split=self.min_samples_split, |
| 610 | min_impurity=self.min_impurity, |
| 611 | max_depth=self.max_depth) |
| 612 | # 计算损失的梯度，并用梯度进行训练 |
| 613 | if not self.is_regression: |
| 614 | Y_hat = softmax(Y_pred) |
| 615 | y, y_pred = Y[:, c], Y_hat[:, c] |
| 616 | else: |
| 617 | y, y_pred = Y[:, c], Y_pred[:, c] |
| 618 | neg_grad = -1 * self.loss.grad(y, y_pred) |
| 619 | tree.fit(X, neg_grad) |
| 620 | # 用新的基学习器进行预测 |
| 621 | h_pred = tree.predict(X) |
| 622 | # line search |
| 623 | if self.line_search == True: |
| 624 | self.weights[i, c] *= line_search(y, y_pred, h_pred) |
| 625 | # 加法模型中添加基学习器的预测，得到最新迭代下的加法模型预测 |
| 626 | Y_pred[:, c] += np.multiply(self.weights[i, c], h_pred) |
| 627 | self.trees[i, c] = tree |
| 628 | |
| 629 | def predict(self, X): |
| 630 | Y_pred = np.zeros((X.shape[0], self.out_dims)) |
nothing calls this directly
no test coverage detected