【自用学习笔记】损失函数——求模型权重W

最新推荐文章于 2024-02-06 00:24:08 发布

coyude

最新推荐文章于 2024-02-06 00:24:08 发布

阅读量184

点赞数

文章标签：学习笔记深度学习机器学习

本文链接：https://blog.csdn.net/qq_26105303/article/details/134471059

版权

损失函数——求模型权重W

随机搜索

import numpy as np
from sklearn import datasets
from sklearn.model_selection import train_test_split

# 加载数据集
digits = datasets.load_digits()
X = digits.data   # 特征
y = digits.target # 标签

# 将标签转换为二分类问题（例如：数字0 vs 非数字0）
y_binary = (y == 0).astype(int)

# 划分数据集
X_train, X_test, y_train, y_test = train_test_split(X, y_binary, test_size=0.2, random_state=42)

# 定义铰链损失函数
def hinge_loss(W, X, y):
    # 计算预测值
    scores = np.dot(X, W)
    # 计算损失
    loss = np.maximum(0, 1 - scores * y)
    return np.mean(loss)

# 随机搜索优化权重W
def random_search(train_data, train_labels, num_iterations=100):
    best_loss = float('inf')
    best_W = None
    for i in range(num_iterations):
        W = np.random.randn(train_data.shape[1])
        loss = hinge_loss(W, train_data, train_labels)
        if loss < best_loss:
            best_loss = loss
            best_W = W
        print(f"第{i}次，当前损失值为{loss}，最小损失值为{best_loss}")
    return best_W

# 执行随机搜索
W_best = random_search(X_train, y_train)

# 评估模型
train_loss = hinge_loss(W_best, X_train, y_train)
test_loss = hinge_loss(W_best, X_test, y_test)
print("最优权重的训练损失：", train_loss)
print("最优权重的测试损失：", test_loss)

随机本地搜索

import numpy as np
from sklearn import datasets
from sklearn.model_selection import train_test_split

# 加载数据集
digits = datasets.load_digits()
X = digits.data  # 特征
y = digits.target  # 标签

# 将标签转换为二分类问题（例如：数字0 vs 非数字0）
y_binary = (y == 0).astype(int) * 2 - 1  # 将标签转换为 +1/-1


# 划分数据集
X_train, X_test, y_train, y_test = train_test_split(
    X, y_binary, test_size=0.2, random_state=42
)


# 定义铰链损失函数
def hinge_loss(W, X, y):
    # 计算预测值
    scores = np.dot(X, W)
    # 计算损失
    loss = np.maximum(0, 1 - scores * y)
    return np.mean(loss)


# 随机本地搜索优化权重W
def random_local_search(train_data, train_labels, num_iterations=1000, lr=0.1):
    best_W = np.random.randn(train_data.shape[1])
    best_loss = hinge_loss(best_W, train_data, train_labels)
    for i in range(num_iterations):
        new_W = best_W + np.random.randn(train_data.shape[1]) * lr
        loss_new = hinge_loss(new_W, train_data, train_labels)
        if loss_new < best_loss:
            best_loss = loss_new
            best_W = new_W
        print(f"第{i}次，当前损失值为{loss_new}，最小损失值为{best_loss}")
    return best_W


# 执行随机搜索
W_best = random_local_search(X_train, y_train)

# 评估模型
train_loss = hinge_loss(W_best, X_train, y_train)
test_loss = hinge_loss(W_best, X_test, y_test)
print("最优权重的训练损失：", train_loss)
print("最优权重的测试损失：", test_loss)

梯度下降法（解析梯度）

import numpy as np
from sklearn import datasets
from sklearn.model_selection import train_test_split

# 加载数据集
digits = datasets.load_digits()
X = digits.data  # 特征
y = digits.target  # 标签

# 将标签转换为二分类问题（例如：数字0 vs 非数字0）
y_binary = (y == 0).astype(int) * 2 - 1  # 将标签转换为 +1/-1


# 划分数据集
X_train, X_test, y_train, y_test = train_test_split(
    X, y_binary, test_size=0.2, random_state=42
)


# 定义铰链损失函数
def hinge_loss(W, X, y):
    # 计算预测值
    scores = np.dot(X, W)
    # 计算损失
    loss = np.maximum(0, 1 - scores * y)
    return np.mean(loss)


# 随机本地搜索优化权重W
def random_local_search(train_data, train_labels, num_iterations=1000, lr=0.1):
    best_W = np.random.randn(train_data.shape[1])
    best_loss = hinge_loss(best_W, train_data, train_labels)
    for i in range(num_iterations):
        new_W = best_W + np.random.randn(train_data.shape[1]) * lr
        loss_new = hinge_loss(new_W, train_data, train_labels)
        if loss_new < best_loss:
            best_loss = loss_new
            best_W = new_W
        print(f"第{i}次，当前损失值为{loss_new}，最小损失值为{best_loss}")
    return best_W


# 执行随机搜索
W_best = random_local_search(X_train, y_train)

# 评估模型
train_loss = hinge_loss(W_best, X_train, y_train)
test_loss = hinge_loss(W_best, X_test, y_test)
print("最优权重的训练损失：", train_loss)
print("最优权重的测试损失：", test_loss)