“择栖工作室招新”

import numpy as np
import tensorflow as tf
#题目自带创建随机点的函数
def data_create_3d(w1, w2, b, amount, size, turb):
    X = np.random.normal(size=(amount, 2), scale=size)
    w = np.array((w1, w2))
    delta = np.random.normal(size=amount, scale=size) * turb
    y = np.dot(X, w) + b + delta
    return X[:, 0], X[:, 1], y

print("目标值->","w1:", 2, "w2:", 3, "b:", 20)

#得到各点的值
x, y, z = data_create_3d(2, 3, 20, 1000, 10, 3)

#构造损失函数f
def loss(w1, w2, b):
    f = 0
    for i in range(0, 1000):
        f += (w1*x[i]+w2*y[i]+b-z[i])**2
    f / 1000  
    
    return f

if __name__ == "__main__":
#构造变量 原函数为y=w1x+w2z+b

    w1=tf.Variable(1.0)
    w2=tf.Variable(2.0)
    b=tf.Variable(25.0)

    
    cnt = 0#循环计数器
    foot1 = 0.000001 #w1学习率
    foot2 = 0.000001 #w2学习率
    foot3 = 0.0001 #b学习率
    yuzhi = 100 #迭代阈值



    while True:
        cnt += 1
        with tf.GradientTape(persistent=True) as tape:
            losss = loss(w1, w2, b)
#计算三个变量的偏导数↓        
        dw1 = tape.gradient(losss, w1)
        dw2 = tape.gradient(losss, w2)
        db = tape.gradient(losss, b)
#储存上次损失函数,用作下面作差使用
        pre = loss(w1, w2, b)
#根据导数更新参量的值
        w1.assign_sub(foot1*dw1)
        w2.assign_sub(foot2*dw2)
        b.assign_sub(foot3*db)
#作两次迭代函数的差
        derta = abs(loss(w1, w2, b)-pre)
#每轮各数据打印
       	print("=========\n迭代次数：",cnt)
        tf.print("loss:",loss(w1,w2,b))  
        a=tf.print("w1:"w1)
        d=tf.print("w2:",w2)
        c=tf.print("b:",b)
#结束条件
        if derta < yuzhi:
            print("迭代结束！")
            break