更新 function.py

function.py

@@ -1,21 +1,14 @@
 import numpy as np
 
-"""
-实验说明:
-- 自行编写实现数据加载与预测病马死亡的对率回归算法。
-注意事项:
-- 在 function.py 中实现对率回归算法。
-- 可以自行导入其他库，需要在 requirements.txt 添加依赖。
-自动评测脚本仅会加载 train_logreg、predict_logreg 和 load_data 函数，请勿修改函数名。
-"""
 
 id = "202210423213"
 name = "熊浩燃"
 
-# 对率回归训练函数，包含调参和优化
-def train_logreg(X, y, learning_rate=0.01, iterations=1000, regularization=0.01, early_stopping_rounds=100):
+
+# 对率回归训练函数，包含加权损失、正则化、学习率调度和提前停止
+def train_logreg(X, y, learning_rate=0.01, iterations=1000, regularization=0.01, early_stopping_rounds=100, class_weights={0: 1, 1: 1}):
     """
-    对率回归训练函数，带有学习率调度和正则化。
+    对率回归训练函数，带有加权损失和正则化。
     Args:
     - X (numpy array): 特征数据。
     - y (numpy array): 标签数据。
@@ -23,6 +16,7 @@ def train_logreg(X, y, learning_rate=0.01, iterations=1000, regularization=0.01,
     - iterations (int): 最大迭代次数。
     - regularization (float): 正则化参数。
     - early_stopping_rounds (int): 提前停止轮数。
+    - class_weights (dict): 类别权重字典，控制类别不平衡。
     Returns:
     - numpy array: 训练得到的权重。
     """
@@ -31,13 +25,14 @@ def train_logreg(X, y, learning_rate=0.01, iterations=1000, regularization=0.01,
     stopping_step = 0
 
     for i in range(iterations):
-        # 预测值：通过对率回归模型得到
+        # 预测值
         z = np.dot(X, weights)
         predictions = 1 / (1 + np.exp(-z))
-        
-        # 计算损失（含正则化项）
-        loss = -np.mean(y * np.log(predictions) + (1 - y) * np.log(1 - predictions)) + regularization * np.sum(weights**2)
-        
+
+        # 加权损失
+        loss = -np.mean(class_weights[1] * y * np.log(predictions) + 
+                        class_weights[0] * (1 - y) * np.log(1 - predictions)) + regularization * np.sum(weights**2)
+
         # 提前停止判断
         if loss < min_loss:
             min_loss = loss
@@ -48,7 +43,7 @@ def train_logreg(X, y, learning_rate=0.01, iterations=1000, regularization=0.01,
             print("提前停止于迭代:", i)
             break
 
-        # 计算梯度并更新权重（含正则化项）
+        # 计算梯度并更新权重
         gradient = (np.dot(X.T, (predictions - y)) + regularization * weights) / y.size
         weights -= learning_rate * gradient
 
@@ -59,18 +54,19 @@ def train_logreg(X, y, learning_rate=0.01, iterations=1000, regularization=0.01,
 
     return weights
 
-# 对率回归预测函数
-def predict_logreg(inX, weights):
+# 对率回归预测函数，带有自定义阈值
+def predict_logreg(inX, weights, threshold=0.6):
     """
     对率回归预测函数。
     Args:
     - inX (numpy array): 输入待分类的样本。
     - weights (numpy array): 训练得到的权重。
+    - threshold (float): 分类决策阈值。
     Returns:
     - int: 预测的分类标签。
     """
     probability = 1 / (1 + np.exp(-np.dot(inX, weights)))
-    return 1 if probability >= 0.5 else 0
+    return 1 if probability >= threshold else 0
 
 # 数据加载函数，包含缺失值填充
 def load_data(filepath):
@@ -107,7 +103,7 @@ def load_data(filepath):
     return data, labels
 
 # 网格搜索函数
-def grid_search(X, y, learning_rates, iterations_list, regularizations):
+def grid_search(X, y, learning_rates, iterations_list, regularizations, class_weights_list):
     """
     网格搜索来优化对率回归模型。
     Args:
@@ -116,6 +112,7 @@ def grid_search(X, y, learning_rates, iterations_list, regularizations):
     - learning_rates (list of float): 学习率列表。
     - iterations_list (list of int): 迭代次数列表。
     - regularizations (list of float): 正则化参数列表。
+    - class_weights_list (list of dict): 类别权重列表。
     Returns:
     - dict: 包含最佳超参数及其对应的权重。
     """
@@ -126,17 +123,18 @@ def grid_search(X, y, learning_rates, iterations_list, regularizations):
     for lr in learning_rates:
         for iters in iterations_list:
             for reg in regularizations:
-                print(f"尝试参数: 学习率={lr}, 迭代次数={iters}, 正则化={reg}")
-                weights = train_logreg(X, y, learning_rate=lr, iterations=iters, regularization=reg)
-                z = np.dot(X, weights)
-                predictions = 1 / (1 + np.exp(-z))
-                loss = -np.mean(y * np.log(predictions) + (1 - y) * np.log(1 - predictions))
-                print(f"损失: {loss}")
+                for cw in class_weights_list:
+                    print(f"尝试参数: 学习率={lr}, 迭代次数={iters}, 正则化={reg}, 类别权重={cw}")
+                    weights = train_logreg(X, y, learning_rate=lr, iterations=iters, regularization=reg, class_weights=cw)
+                    z = np.dot(X, weights)
+                    predictions = 1 / (1 + np.exp(-z))
+                    loss = -np.mean(cw[1] * y * np.log(predictions) + cw[0] * (1 - y) * np.log(1 - predictions))
+                    print(f"损失: {loss}")
 
-                if loss < best_loss:
-                    best_loss = loss
-                    best_weights = weights
-                    best_params = {'learning_rate': lr, 'iterations': iters, 'regularization': reg}
+                    if loss < best_loss:
+                        best_loss = loss
+                        best_weights = weights
+                        best_params = {'learning_rate': lr, 'iterations': iters, 'regularization': reg, 'class_weights': cw}
 
     print("最佳参数:", best_params)
     return best_weights, best_params
@@ -149,13 +147,14 @@ if __name__ == "__main__":
     learning_rates = [0.001, 0.01, 0.1]
     iterations_list = [1000, 2000, 3000]
     regularizations = [0.01, 0.1, 1.0]
+    class_weights_list = [{0: 1, 1: 1}, {0: 1, 1: 2}, {0: 1, 1: 3}]
 
     # 进行网格搜索以找到最佳超参数组合
-    best_weights, best_params = grid_search(X_train, y_train, learning_rates, iterations_list, regularizations)
+    best_weights, best_params = grid_search(X_train, y_train, learning_rates, iterations_list, regularizations, class_weights_list)
     print("最佳权重:", best_weights)
 
     # 使用最佳权重进行预测
     X_test, y_test = load_data("./data/horse-colic.test")
-    y_pred = [predict_logreg(x, best_weights) for x in X_test]
+    y_pred = [predict_logreg(x, best_weights, threshold=0.6) for x in X_test]
     accuracy = np.mean(y_pred == y_test)
     print(f"测试集准确率: {accuracy}")