【机器学习】Logistic回归代码实现_在右侧编辑器补充代码,计算并得出logistic回归回归线的x,y坐标点数组。

# 改进的随机梯度上升算法
def stoc_grad_ascent2(dataset, labels, times=300):
    dataset = np.array(dataset)
    m, n = np.shape(dataset)
    weight = np.ones(n)
    weight_change = []
    for i in range(times):
        data_index = list(range(m))
        # 在一次数据集上训练时，不重复抽取数据集中一条样本进行权重训练
        for j in range(m):
            # 在 0至len(data_index) 上随机抽取一个整数rand_index
            rand_index = int(np.random.uniform(0, len(data_index)))
            # j增大时，减小步长的大小，减小权重调整的高频变动
            alpha = 4.0/(1.0 + i + j) + 0.001
            error = labels[rand_index] - sigmoid(sum(dataset[rand_index] * weight))
            weight = weight + alpha * error * dataset[rand_index]
            weight_change.append(weight)
            del(data_index[rand_index])
    return weight, weight_change

 
def load_dataset(filename):
    # 特征数据
    dataset = []
    # 标签数据
    labels = []
    with open(filename) as file:
        for eachline in file.readlines():
            # 将每一行的换行符去除并以空格拆分整行
            data_list = eachline.strip('\n').split('\t')
            # 初始化Xo为1，X1，X2为读取数据中的第一位、第二位值
            # 注意加上数值类型转化，不加容易出错
            dataset.append([1.0, float(data_list[0]), float(data_list[1])])
            labels.append(int(data_list[2]))
    return dataset, labels
 
def sigmoid(x):
    # 跃阶函数表达式
    return 1.0/(1 + np.exp(-x))
 
import numpy as np
# 梯度上升优化算法
def grad_ascent(dataset, labels):
    # 将列表转换成矩阵，方便进行矩阵运算
    data_matrix_mn = np.mat(dataset)
    label_matrix_m1 = np.mat(labels).transpose()
    # m：样本个数 n：特征个数
    m, n = np.shape(data_matrix_mn)
    # 初始化权重矩阵
    weight_matrix_n1 = np.ones((n, 1))
    # 步长
    alpha = 0.001
    # 训练次数
    times = 500
    for i in range(times):
        h_matrix_m1 = sigmoid(data_matrix_mn * weight_matrix_n1)
        error_matrix_m1 = label_matrix_m1 - h_matrix_m1
        # 根据计算出来error矩阵和步长调整权重矩阵
        weight_matrix_n1 = weight_matrix_n1 + alpha * data_matrix_mn.transpose() * error_matrix_m1
    return weight_matrix_n1
 
def stoc_grad_ascent(dataset, labels, times=300):
    dataset = np.array(dataset)
    m, n = np.shape(dataset)
    weight = np.ones(n)
    alaph = 0.001
    # 记录权重变化情况
    weight_change = []
    # 在数据集中循环训练次数
    for j in range(times):
        for i in range(m):
            # 利用每一条样本调整权重系数
            error = labels[i] - sigmoid(sum(dataset[i] * weight))
            weight = weight + alaph * error * dataset[i]
            # 记录权重
            weight_change.append(weight)
    return weight, weight_change
 
# 改进的随机梯度上升算法
def stoc_grad_ascent2(dataset, labels, times=300):
    dataset = np.array(dataset)
    m, n = np.shape(dataset)
    weight = np.ones(n)
    weight_change = []
    for i in range(times):
        data_index = list(range(m))
        # 在一次数据集上训练时，不重复抽取数据集中一条样本进行权重训练
        for j in range(m):
            # 在 0至len(data_index) 上随机抽取一个整数rand_index
            rand_index = int(np.random.uniform(0, len(data_index)))
            # j增大时，减小步长的大小，减小权重调整的高频变动
            alpha = 4.0/(1.0 + i + j) + 0.001
            error = labels[rand_index] - sigmoid(sum(dataset[rand_index] * weight))
            weight = weight + alpha * error * dataset[rand_index]
            weight_change.append(weight)
            del(data_index[rand_index])
    return weight, weight_change
 
def draw_weight_change(weight_change):
    import matplotlib.pyplot as plt
 
    times = np.arange(0, len(weight_change), 1)
 
    W0 = [weight[0] for weight in weight_change]
    W1 = [weight[1] for weight in weight_change]
    W3 = [weight[2] for weight in weight_change]
 
    fig = plt.figure('权重变化')
    ax1 = fig.add_subplot(311)
    ax2 = fig.add_subplot(312)
    ax3 = fig.add_subplot(313)
 
    ax1.plot(times, W0)
    ax2.plot(times, W1)
    ax3.plot(times, W3)
 
    ax1.set_ylabel('W0')
    ax2.set_ylabel('W1')
    ax3.set_ylabel('W2')
    plt.show()
 
 
 
 
 
 
def draw_divide_line(dataset, labels, weights):
    import matplotlib.pyplot as plt
    # 类别1和类别0的坐标
    x1 = []
    y1 = []
    x0 = []
    y0 = []
    # 将数据集的数据分类并存入坐标
    for i in range(len(dataset)):
        if labels[i] == 1:
            x1.append(dataset[i][1])
            y1.append(dataset[i][2])
        else:
            x0.append(dataset[i][1])
            y0.append(dataset[i][2])
 
    # 计算决策边界坐标
    x = np.arange(-3, 3, 0.1)
    # weight[0] + weight[1]*x + weight[2]*y = 0 . 反算出y值
    y = (-weights[0] - weights[1] * x) / (weights[2])
 
    # 绘制数据集散点
    fig = plt.figure('数据集及决策边界')
    ax = fig.add_subplot(111)
    type1 = ax.scatter(x1, y1, c='r', marker='s')
    type0 = ax.scatter(x0, y0, c='g')
 
    # 绘制决策边界
    ax.plot(x, y.reshape((60, 1)))
 
    # 添加标签
    plt.xlabel('x1')
    plt.ylabel('x2')
    # 添加图例
    ax.legend((type0, type1), ('0', '1'))
    plt.savefig('divide_line.png')
 
 
 
 
if __name__ == '__main__':
    dataset, labels = load_dataset(r'machine learning\Ch05\testSet.txt')
    weights, weight_change = stoc_grad_ascent2(dataset, labels)
    draw_divide_line(dataset, labels, weights)
    draw_weight_change(weight_change)