随机梯度下降简易版

import numpy as np
import matplotlib.pyplot as plt


a=0.01
LEN = 100
X = np.arange(0, LEN)
# rand从-5到+5
np.random.seed(1)
rand = (np.random.random(LEN) * 2 - 1) * 5
Y = X * 3 + rand

# X、Y分布如下图所示


# X、Y 连接
X = X.reshape(LEN, 1)
Y = Y.reshape(LEN, 1)

allData = np.concatenate((X, Y), axis = 1)
np.random.shuffle(allData)
# 训练集：测试机 = 4:1
ratio = 0.8
index = (int)(allData.shape[0] * ratio)
trainData = allData[:index]
testData = allData[index:]

# 学习率
lr = 0.00005
# 训练集大小（每个batch随机梯度下降迭代次数）
N = trainData.shape[0]
# 待估及参数（theta）
theta = np.random.rand()
grad=0
# 迭代次数标识
iter = 1
epsilon=350
while True:
    for i in range(N):
        # 随机样本
        x = trainData[i, 0]
        y = trainData[i, 1]
        # 计算梯度
        grad = (theta * x - y) * x
        # 更新参数
        theta = theta - lr * grad
    i=0
    loss = np.sum(0.5 * (trainData[:, 0] * theta - trainData[:, 1]) ** 2)
    print("No.%d:\t grad = %f\t theta: %f\tloss: %f" %(iter, grad, theta, loss))
    iter += 1
    # 达到允许的误差，结束训练
    if loss < epsilon:
        print("Traing Completed!")
        break
result=testData[:,0]*theta

plt.plot(testData[:,0], result,'bo',label='predict')
plt.plot(testData[:,0],testData[:,1],'ro',label='standard')
plt.legend()
plt.show()