83 lines
2.4 KiB
Python
83 lines
2.4 KiB
Python
|
||
"""
|
||
线性回归测试
|
||
"""
|
||
|
||
import numpy as np
|
||
import torch
|
||
import torch.nn as nn
|
||
import torch.nn.functional as F
|
||
import torch.optim as optim
|
||
|
||
import matplotlib.pyplot
|
||
|
||
def sigmoid(z):
|
||
return 1.0/(1.0+np.exp(-z)) #定义sigmoid函数
|
||
|
||
"""
|
||
定义学习次数 epochs为100
|
||
定义学习率为 0.01
|
||
对x在-2到2区间内采样1000个值
|
||
定义回归目标函数 y = x^完整例子
|
||
"""
|
||
epochs = 100
|
||
learning_rate=0.01
|
||
lr = learning_rate
|
||
x = torch.linspace(-2., 2., 1000)
|
||
y = x**2
|
||
|
||
"""
|
||
定义三个权重，分别为 wih, whh, who
|
||
wih: weight of input layer to hidden layer 1 矩阵长度【 1X4 】
|
||
whh: weight of hidden layer 1 to 完整例子 矩阵长度【 4X4 】
|
||
who: weight of hidden layer 完整例子 to output layer 矩阵长度【 4X1 】
|
||
"""
|
||
|
||
wih = np.random.rand(1, 4)
|
||
whh = np.random.rand(4, 4)
|
||
who = np.random.rand(4, 1)
|
||
|
||
def forward(input):
|
||
|
||
result1 = input@wih #初始值传入第一个隐藏层时进行加权计算
|
||
input1 = sigmoid(result1) #第一次加权运算后对结果进行sigmoid压缩
|
||
result2 = input1@whh #压缩后的结果传入第二层隐藏层时进行加权运算
|
||
input2 = sigmoid(result2) #对第二次加权运算结果进行sigmoid压缩
|
||
result3 = input2@who #压缩后的结果传入输出层时进行加权运算，得到的结果result3即为输出值
|
||
|
||
final_result = result3
|
||
return final_result, result2, result1, input
|
||
|
||
"""
|
||
定义损失函数，用来计算三个层面的权值改变，分别为eoh， ehh， ehi
|
||
eoh：输出层与第二隐藏层之间的损失值
|
||
ehh：第二隐藏层与第一隐藏层之间的损失值
|
||
ehi ：第一隐藏层与输入层之间的损失值
|
||
"""
|
||
def loss(result):
|
||
eho = target - result
|
||
ehh = np.dot(who, [eho])
|
||
ehi = np.dot(whh.T, ehh)
|
||
return eho, ehh, ehi
|
||
|
||
def backward(eho,ehh,ehi,final_result,result2,result1,input):
|
||
global who,whh,wih
|
||
who += lr * np.dot((eho * final_result * (1.0 - final_result)), np.transpose(result2))
|
||
whh += lr * np.dot((ehh * result2 * (1.0 - result2)), np.transpose(result1))
|
||
wih += lr * np.dot((ehh * result1 * (1.0 - result1)), np.transpose(input))
|
||
|
||
|
||
for z in range(4):
|
||
target = z*z
|
||
final_result,result2,result1,input=forward([z])
|
||
|
||
eho,ehh,ehi = loss(final_result)
|
||
|
||
backward(eho,ehh,ehi,final_result,result2,result1,input)
|
||
|
||
print("输出结果误差",eho)
|
||
|
||
|
||
|
||
|