【三层神经网络 python实现】

三层神经网络 python实现

一、 输入层隐藏层输出层图示

#图中x1为x[0]
#图中x2为x[1]
#w1,w2,w3,w4,w5,w6为权值
#b1,b2,b3为偏置量
h1 = sigmoid(x[0]*self.w1+x[1]*self.w2+self.b1)
h2 = sigmoid(x[0]*self.w3+x[1]*self.w4+self.b3)
o1 = sigmoid(h1*self.w5+h2*self.w6+self.b3)

二、 初始化

def __init__(self):
    self.w1 = np.random.normal()
    self.w2 = np.random.normal()
    self.w3 = np.random.normal()
    self.w4 = np.random.normal()
    self.w5 = np.random.normal()
    self.w6 = np.random.normal()
    self.b1 = np.random.normal()
    self.b2 = np.random.normal()
    self.b3 = np.random.normal()

三、 前向传播

#计算出预测组
def feedForward(self, x):
    h1 = sigmoid(x[0]*self.w1+x[1]*self.w2+self.b1)
    h2 = sigmoid(x[0]*self.w3+x[1]*self.w4+self.b3)
    o1 = sigmoid(h1*self.w5+h2*self.w6+self.b3)
    ypred = o1
    return ypred

四、 后向传播

def train(self, data, all_y_trues):
        ...
            for x,ytrue in zip(data, all_y_trues):
                
                h1 = sigmoid(x[0]*self.w1+x[1]*self.w2+self.b1)
                h2 = sigmoid(x[0]*self.w3+x[1]*self.w4+self.b3)
                o1 = sigmoid(h1*self.w5+h2*self.w6+self.b3)
                
                d_L_d_ypred = -2 * (ytrue - o1)
                
                d_ypred_d_h1 = self.w5 * der_sigmoid(o1)
                d_ypred_d_h2 = self.w6 * der_sigmoid(o1)
                d_ypred_d_w5 = h1 * der_sigmoid(o1)
                d_ypred_d_w6 = h2 * der_sigmoid(o1)
                d_ypred_d_b3 = der_sigmoid(o1)
                
                d_h1_d_w1 = x[0] * der_sigmoid(h1)
                d_h1_d_w2 = x[1] * der_sigmoid(h1)
                d_h1_d_b1 = der_sigmoid(h1)
                
                d_h2_d_w3 = x[0] * der_sigmoid(h2)
                d_h2_d_w4 = x[1] * der_sigmoid(h2)
                d_h2_d_b2 = der_sigmoid(h2)
                
                #计算出梯度，即损失函数对权值和偏置的偏导，即权值和偏置量对损失函数的影响大小
                #利用链式法则求偏导
                d_L_d_w1 = d_L_d_ypred * d_ypred_d_h1 * d_h1_d_w1
                d_L_d_w2 = d_L_d_ypred * d_ypred_d_h1 * d_h1_d_w2
                d_L_d_w3 = d_L_d_ypred * d_ypred_d_h2 * d_h2_d_w3
                d_L_d_w4 = d_L_d_ypred * d_ypred_d_h2 * d_h2_d_w4
                d_L_d_w5 = d_L_d_ypred * d_ypred_d_w5
                d_L_d_w6 = d_L_d_ypred * d_ypred_d_w6
                d_L_d_b1 = d_L_d_ypred * d_ypred_d_h1 * d_h1_d_b1
                d_L_d_b2 = d_L_d_ypred * d_ypred_d_h2 * d_h2_d_b2
                d_L_d_b3 = d_L_d_ypred * d_ypred_d_b3
                
            if(i%100 == 0):
                ypreds = np.apply_along_axis(self.feedForward, 1, data)
                #根据损失函数，计算出损失值
                loss = Loss_MSE(ypreds, all_y_trues)
                print(ypreds)
                print("第 %d 轮，Loss：%.6f\n"%(i, loss))

五、 更新权值和偏置

learnrate = 0.01
self.w1 -= learnrate * self.w1
self.w2 -= learnrate * self.w2
self.w3 -= learnrate * self.w3
self.w4 -= learnrate * self.w4
self.w5 -= learnrate * self.w5
self.w6 -= learnrate * self.w6
self.b1 -= learnrate * self.b1
self.b2 -= learnrate * self.b2
self.b3 -= learnrate * self.b3

六、 数据集

# Define dataset
data = np.array([
  [-2, -1],  # Alice	处理后的身高，体重数据
  [25, 6],   # Bob
  [19, 4],   # Charlie
  [-15, -6], # Diana
])
all_y_trues = np.array([
  1, # Alice	女性
  0, # Bob		男性
  0, # Charlie
  1, # Diana
])

七、 总结

通过学习Victor Zhou的博客，https://victorzhou.com/blog/intro-to-neural-networks/
使用python实现了三层神经网络。
接下来将继续学习pytorch文档中的图片/视频、音频、文本三个基础任务，并更新。

八、 效果图

九、 源代码

import numpy as np

def sigmoid(x):
    return 1/(1+np.exp(-x))

def der_sigmoid(x):
    return x*(1-x)

def Loss_MSE(ytrue, ypred):
    return ((ytrue-ypred)*2).mean()

class myNN:
    
    def __init__(self):
        self.w1 = np.random.normal()
        self.w2 = np.random.normal()
        self.w3 = np.random.normal()
        self.w4 = np.random.normal()
        self.w5 = np.random.normal()
        self.w6 = np.random.normal()
        self.b1 = np.random.normal()
        self.b2 = np.random.normal()
        self.b3 = np.random.normal()
    
    def feedForward(self, x):
        
        h1 = sigmoid(x[0]*self.w1+x[1]*self.w2+self.b1)
        h2 = sigmoid(x[0]*self.w3+x[1]*self.w4+self.b3)
        o1 = sigmoid(h1*self.w5+h2*self.w6+self.b3)
        ypred = o1
        return ypred
    
    def train(self, data, all_y_trues):
        learnrate = 0.01
        i = 1
        for i in range(10001):
            for x,ytrue in zip(data, all_y_trues):
                
                h1 = sigmoid(x[0]*self.w1+x[1]*self.w2+self.b1)
                h2 = sigmoid(x[0]*self.w3+x[1]*self.w4+self.b3)
                o1 = sigmoid(h1*self.w5+h2*self.w6+self.b3)
                
                d_L_d_ypred = -2 * (ytrue - o1)
                
                d_ypred_d_h1 = self.w5 * der_sigmoid(o1)
                d_ypred_d_h2 = self.w6 * der_sigmoid(o1)
                d_ypred_d_w5 = h1 * der_sigmoid(o1)
                d_ypred_d_w6 = h2 * der_sigmoid(o1)
                d_ypred_d_b3 = der_sigmoid(o1)
                
                d_h1_d_w1 = x[0] * der_sigmoid(h1)
                d_h1_d_w2 = x[1] * der_sigmoid(h1)
                d_h1_d_b1 = der_sigmoid(h1)
                
                d_h2_d_w3 = x[0] * der_sigmoid(h2)
                d_h2_d_w4 = x[1] * der_sigmoid(h2)
                d_h2_d_b2 = der_sigmoid(h2)
                
                self.w1 -= learnrate * d_L_d_ypred * d_ypred_d_h1 * d_h1_d_w1
                self.w2 -= learnrate * d_L_d_ypred * d_ypred_d_h1 * d_h1_d_w2
                self.w3 -= learnrate * d_L_d_ypred * d_ypred_d_h2 * d_h2_d_w3
                self.w4 -= learnrate * d_L_d_ypred * d_ypred_d_h2 * d_h2_d_w4
                self.w5 -= learnrate * d_L_d_ypred * d_ypred_d_w5
                self.w6 -= learnrate * d_L_d_ypred * d_ypred_d_w6
                self.b1 -= learnrate * d_L_d_ypred * d_ypred_d_h1 * d_h1_d_b1
                self.b2 -= learnrate * d_L_d_ypred * d_ypred_d_h2 * d_h2_d_b2
                self.b3 -= learnrate * d_L_d_ypred * d_ypred_d_b3
                
            if(i%100 == 0):
                ypreds = np.apply_along_axis(self.feedForward, 1, data)
                loss = Loss_MSE(ypreds, all_y_trues)
                print(ypreds)
                print("第 %d 轮，Loss：%.6f\n"%(i, loss))

# Define dataset
data = np.array([
  [-2, -1],  # Alice
  [25, 6],   # Bob
  [19, 4],   # Charlie
  [-15, -6], # Diana
])
all_y_trues = np.array([
  1, # Alice
  0, # Bob
  0, # Charlie
  1, # Diana
])    

nn = myNN()
nn.train(data, all_y_trues)