generated from LufsX/Experiment-LogReg
109 lines
4.3 KiB
Python
109 lines
4.3 KiB
Python
import numpy as np
|
|
|
|
id = "202210423213"
|
|
name = "熊浩燃"
|
|
|
|
# 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}):
|
|
weights = np.zeros(X.shape[1])
|
|
min_loss = float('inf')
|
|
stopping_step = 0
|
|
|
|
for i in range(iterations):
|
|
z = np.dot(X, weights)
|
|
predictions = 1 / (1 + np.exp(-z))
|
|
|
|
loss = -np.mean(class_weights[1] * y * np.log(predictions) +
|
|
class_weights[0] * (1 - y) * np.log(1 - predictions)) + regularization * np.sum(weights**2)
|
|
|
|
if loss < min_loss:
|
|
min_loss = loss
|
|
stopping_step = 0
|
|
else:
|
|
stopping_step += 1
|
|
if stopping_step >= early_stopping_rounds:
|
|
print("提前停止于迭代:", i)
|
|
break
|
|
|
|
gradient = (np.dot(X.T, (predictions - y)) + regularization * weights) / y.size
|
|
weights -= learning_rate * gradient
|
|
|
|
if i % 100 == 0 and i != 0:
|
|
learning_rate *= 0.9 # 每100轮衰减学习率
|
|
print(f"迭代 {i}，损失: {loss}，学习率: {learning_rate}")
|
|
|
|
return weights
|
|
|
|
# 对率回归预测函数，带有自定义阈值
|
|
def predict_logreg(inX, weights, threshold=0.6):
|
|
probability = 1 / (1 + np.exp(-np.dot(inX, weights)))
|
|
return 1 if probability >= threshold else 0
|
|
|
|
# 数据加载函数，包含缺失值填充
|
|
def load_data(filepath):
|
|
data = []
|
|
labels = []
|
|
with open(filepath, 'r') as file:
|
|
for line in file:
|
|
fields = line.strip().split()
|
|
features = [float(x) if x != '?' else None for x in fields[:-1]]
|
|
label = float(fields[-1])
|
|
data.append(features)
|
|
labels.append(label)
|
|
|
|
data = np.array(data, dtype=float)
|
|
labels = np.array(labels, dtype=float)
|
|
|
|
col_means = np.nanmean(data, axis=0)
|
|
indices = np.where(np.isnan(data))
|
|
data[indices] = np.take(col_means, indices[1])
|
|
|
|
return data, labels
|
|
|
|
# 暴力搜索函数
|
|
def brute_force_search(X, y, X_test, y_test, learning_rates, iterations_list, regularizations, class_weights_list, thresholds):
|
|
best_params = {}
|
|
best_weights = None
|
|
best_accuracy = 0
|
|
|
|
for lr in learning_rates:
|
|
for iters in iterations_list:
|
|
for reg in regularizations:
|
|
for cw in class_weights_list:
|
|
for threshold in thresholds:
|
|
print(f"参数组合: 学习率={lr}, 迭代次数={iters}, 正则化={reg}, 类别权重={cw}, 阈值={threshold}")
|
|
weights = train_logreg(X, y, learning_rate=lr, iterations=iters, regularization=reg, class_weights=cw)
|
|
y_pred = [predict_logreg(x, weights, threshold=threshold) for x in X_test]
|
|
accuracy = np.mean(y_pred == y_test)
|
|
print(f"当前组合的测试集准确率: {accuracy}")
|
|
|
|
if accuracy > best_accuracy:
|
|
best_accuracy = accuracy
|
|
best_weights = weights
|
|
best_params = {'learning_rate': lr, 'iterations': iters, 'regularization': reg, 'class_weights': cw, 'threshold': threshold}
|
|
|
|
print("最佳参数组合:", best_params)
|
|
print("最佳准确率:", best_accuracy)
|
|
return best_weights, best_params
|
|
|
|
if __name__ == "__main__":
|
|
# 加载训练集
|
|
X_train, y_train = load_data("./data/horse-colic.data")
|
|
X_test, y_test = load_data("./data/horse-colic.test")
|
|
|
|
# 调整后的超参数组合
|
|
learning_rates = [0.0001, 0.001, 0.01, 0.05, 0.1]
|
|
iterations_list = [1000, 2000, 5000, 10000]
|
|
regularizations = [0.01, 0.1, 1.0, 10.0]
|
|
class_weights_list = [{0: 1, 1: 2}, {0: 1, 1: 3}, {0: 1, 1: 4}, {0: 1, 1: 5}]
|
|
thresholds = [0.5, 0.55, 0.6, 0.65, 0.7]
|
|
|
|
# 进行暴力搜索找到最佳参数组合
|
|
best_weights, best_params = brute_force_search(X_train, y_train, X_test, y_test, learning_rates, iterations_list, regularizations, class_weights_list, thresholds)
|
|
print("最佳权重:", best_weights)
|
|
|
|
# 使用最佳权重和阈值在测试集上评估
|
|
y_pred = [predict_logreg(x, best_weights, threshold=best_params['threshold']) for x in X_test]
|
|
accuracy = np.mean(y_pred == y_test)
|
|
print(f"最终测试集准确率: {accuracy}")
|