机器学习之正则化实现

正则化实现

//正则化
import numpy as np
import matplotlib.pyplot as plt


# 1.	完成数据集的读取
data_s1 = np.loadtxt('s1.txt', delimiter=',')
data_s2 = np.loadtxt('s2.txt', delimiter=',')

# x_s1, y_s1 = data_s1[:, :-1], data_s1[:, -1:]
# x_s2, y_s2 = data_s2[:, :-1], data_s2[:, -1:]

x_s1, y_s1 = np.hsplit(data_s1, [-1, ])
x_s2, y_s2 = np.hsplit(data_s2, [-1, ])


X_s1 = np.insert(x_s1, 0, 1, axis=1)
X_s2 = np.insert(x_s2, 0, 1, axis=1)
m, n = X_s1.shape


def funcZ(X, theta):
    z = X.dot(theta)
    return z


# 2.	实现Sigmoid函数,并画出该函数
def Sigmoid(z):
    h = 1/(1+np.exp(-z))
    return h


# 3.	实现逻辑回归的代价函数，实现正则化逻辑回归

def computeCost(h, y, lamda, theta):
    cost = -np.mean(y*np.log(h)+(1-y)*np.log(1-h))
    R = lamda/(2*m)*np.sum(theta[1:]**2)
    return cost+R


# 4.	实现梯度下降函数

def grad(X, h, y, lamda, theta):
    d_theta = (1/m)*np.dot(X.T, h-y)
    R = lamda/m*np.insert(theta[1:], 0, 0, axis=0)
    return d_theta+R


# 5.	通过梯度下降计算回归模型，用所得模型对测试集的数据进行预测，并计算准确率

def train(alpha, iters, theta, X, lamda, y):
    Jcost = []
    for step in range(iters):
        z = funcZ(X, theta)
        h = Sigmoid(z)
        jcost = computeCost(h, y, lamda, theta)
        d_theta = grad(X, h, y, lamda, theta)
        theta -= alpha*d_theta
        if step % 200 == 0:
            print('step:', step, 'cost:', jcost)
            Jcost.append(jcost)
    return theta, Jcost


if __name__ == '__main__':
    alpha = 0.1
    iters = 10000
    lamda = 4
    theta = np.random.uniform(-0.5, 0.5, [n, 1])
    theta, Jcost = train(alpha, iters, theta, X_s1, lamda, y_s1)

    z = funcZ(X_s2, theta)
    h = Sigmoid(z)
    print('predict:', h)
    print('label_y:', y_s2)
    print('精确度：', np.mean(y_s2 == [h > 0.5]))

    # 6.	使用X2，X3两组特征画出逻辑回归0-1分布图
  
    # numpy中的ravel()、flatten()、squeeze()都有将多维数组转换为一维数组的功能，区别：
    # ravel()：如果没有必要，不会产生源数据的副本，返回的是视图，修改时会影响原始矩阵
    # flatten()：返回源数据的副本，返回一份拷贝，对拷贝所做修改不会影响原始矩阵
    # squeeze()：只能对维数为1的维度降维
    plt.scatter(X_s2[:, 2], X_s2[:, 3], c=y_s2.ravel())
    plt.show()

    plt.scatter(X_s2[:, 2], X_s2[:, 3], c=h.ravel())
    plt.show()

    plt.scatter(y_s2, y_s2)
    plt.scatter(y_s2, h)
    plt.show()

    # plt.scatter(X_s2[:, 2], X_s2[:, 3])
    # plt.show()

    # plt.plot(X_s2[:, 2], y_s2, 'rx')
    # plt.plot(X_s2[:, 2], h, 'gx')
    # plt.show()
    # plt.plot(X_s2[:, 3], y_s2, 'rx')
    # plt.plot(X_s2[:, 3], h, 'gx')
    # plt.show()

    # pos = np.array([x_s2[i] for i in range(m) if y_s2[i] == 0])
    # neg = np.array([x_s2[i] for i in range(len(x_s2)) if y_s2[i] == 1])
    # plt.scatter(pos[:, 1], pos[:, 2])
    # plt.scatter(neg[:, 1], neg[:, 2])
    # plt.show()

    # y_s2 = np.squeeze(y_s2)
    # plt.scatter(X_s2[y_s2 == 0, 2], X_s2[y_s2 == 0, 3])
    # plt.scatter(X_s2[y_s2 == 1, 2], X_s2[y_s2 == 1, 3])
    # plt.show()

    # # 实现Sigmoid函数,并画出该函数
    # x = np.linspace(-10, 10, 200)
    # y = Sigmoid(x)
    # plt.plot(x, y)
    # plt.show()

本文中数据集下载地址:
链接：https://pan.baidu.com/s/1io3aZH-QRCqnGardXOH5yg 提取码：83mu

学习讨论群请加QQ群:521364338,扫码进群领取人工智能学习资料,转载标明出处,侵权必究!