R语言——实验4-人工神经网络

带包实现：

rm(list=ls())
setwd("C:/Users/Administrator/Desktop/R语言与数据挖掘作业/实验4-人工神经网络")

Data=read.csv("sales_data.csv")[,2:5]

library(nnet)
colnames(Data)<-c("x1","x2","x3","y")

model1=nnet(y~.,data=Data,size=6,decay=5e-4,maxit=1000)

pred=predict(model1,Data[,1:3],type="class")
(p=sum(as.numeric(pred==Data$y))/nrow(Data))

table(Data$y,pred)

prop.table(table(Data$y,pred),1)

2. 深入理解BP人工神经网络算法，并用R语言实现该算法

自己打出一个简单的神经网络

rm(list=ls())
#install.packages("sampling")
library(sampling)
setwd("C:/Users/Administrator/Desktop/R语言与数据挖掘作业/实验4-人工神经网络")

data("iris")
#as.numeric(data[,5])
#这里我们按照每种“Species”抽取3/5个样本进行抽样。
n=round(3/5*nrow(iris)/3)

sub_train=strata(iris,stratanames=("Species"),size=rep(n,3),method="srswor")

head(sub_train)

data_train=iris[sub_train$ID_unit,]
data_test=iris[-sub_train$ID_unit,]
dim(data_train)
dim(data_test)

#保存数据集
write.csv(data_train,"./iris_data_train.csv")
write.csv(data_test,"./iris_data_test.csv")

#对data_train归一化处理
#colnames()
data_train$Sepal.Length=(data_train$Sepal.Length-min(data_train$Sepal.Length))*1.0/
  (max(data_train$Sepal.Length)-min(data_train$Sepal.Length))

data_train$Sepal.Width=(data_train$Sepal.Width-min(data_train$Sepal.Width))*1.0/
  (max(data_train$Sepal.Width)-min(data_train$Sepal.Width))

data_train$Petal.Length=(data_train$Petal.Length-min(data_train$Petal.Length))*1.0/
  (max(data_train$Petal.Length)-min(data_train$Petal.Length))

data_train$Petal.Width=(data_train$Petal.Width-min(data_train$Petal.Width))*1.0/
  (max(data_train$Petal.Width)-min(data_train$Petal.Width))

#4个输入，5个的隐藏层，3个输出层
#第一块连接的地方需要4*5个w，5个a,第二块需要5*3个w,5个b

f<-function(x)
{
  x=1*1.0/(1+exp(-1*x))
  return(x)
}

#定义训练次数
global_time=100

#定义学习率
learning_rate=1.8

#随机生成第一块的w1
w1=matrix(sample((0:100)*1.0/100,size=20),4,5)
a1=matrix(sample((0:100)*1.0/100,size=5),1,5)

w2=matrix(sample((0:100)*1.0/100,size=15),5,3)
a2=matrix(sample((0:100)*1.0/100,size=3),1,3)

#遍历每一条数据，每扔进一条数据就调参数
for(time in 1:global_time)
{
  for(i in 1:length(data_train$Species))
  {
    #1*4
    x1=matrix(c(data_train$Sepal.Length[i],data_train$Sepal.Width[i],data_train$Petal.Length[i],data_train$Petal.Width[i]),1,4)

    #矩阵相乘 %*%
    #得到经过第一个隐藏层的输出，也就是最后输出层的输入
    x2=f(x1%*%w1-a1)

    #得到最后的输出层，是1*3的矩阵
    x3=f(x2%*%w2-a2)

    #把标签变成1*3的矩阵
    y=matrix(0.1,1,3)
    if(data_train$Species[i]=="setosa" ){y[1]=0.9}
    if(data_train$Species[i]=="versicolor"){y[2]=0.9}
    if(data_train$Species[i]=="virginica"){y[3]=0.9}
    #与标签比较调参,输出层的误差项为output*(1-output)*(y-output)
    #print(y)
    #print(data_train$Species[i])
    cha=x3*(1-x3)*(y-x3)
    #cat("loss",mean(cha))
    #print("")
    #更新隐藏层和输出层之间的w2,dw2=w2+learning_rate * cha * xi
    #通过x2(竖着),t-o(横着)相乘得到5*3的矩阵和w2相加来更新
    tx2=t(x2)
    dw2=learning_rate * (tx2 %*% cha)
    #把之前的w2存下来,之后更新要用
    before_w2=w2
    w2=w2+dw2

    #更新输入层和隐藏层之间的w1,dw1=w1+learning_rate * cha * xi
    #隐藏层的误差项不是直接得到的，需要通过后一层的误差项计算,为(和(cha1*wi))
    #隐藏层的误差项为 w2(5*3) %*% cha(3*1),的cha2(5*1),注:用的是每更新前的w2
    cha2 = before_w2 %*% t(cha)
    tx1=t(x1)
    dw1=learning_rate * (tx1 %*% t(cha2))
    w1=w1+dw1

  }
}

#看看拟合度
SUM=length(data_train$Species)
right=0
for(i in 1:length(data_train$Species))
{
  #1*4
  x1=matrix(c(data_train$Sepal.Length[i],data_train$Sepal.Width[i],data_train$Petal.Length[i],data_train$Petal.Width[i]),1,4)

  #矩阵相乘 %*%
  #得到经过第一个隐藏层的输出，也就是最后输出层的输入
  x2=f(x1%*%w1-a1)

  #得到最后的输出层，是1*3的矩阵
  x3=f(x2%*%w2-a2)
  print(x3)

  y1=matrix(c(0.9,0.1,0.1),1,3)
  y2=matrix(c(0.1,0.9,0.1),1,3)
  y3=matrix(c(0.1,0.1,0.9),1,3)

  # cha11=x3*(1-x3)*(y1-x3)
  # cha22=x3*(1-x3)*(y2-x3)
  # cha33=x3*(1-x3)*(y3-x3)
  cha11=(y1-x3)
  cha22=(y2-x3)
  cha33=(y3-x3)

  cha1=0
  cha2=0
  cha3=0

  for(j in 1:3)
  {
    cha1=cha1+abs(cha11[j])
    cha2=cha2+abs(cha22[j])
    cha3=cha3+abs(cha33[j])
  }

  micha=min(cha1,cha2,cha3)
  #cat("micha",micha,"\n")
  #cat("cha1",cha1,"\n")
  #cat("cha2",cha2,"\n")
  #cat("cha3",cha3,"\n")

  if(micha==cha1 & data_train$Species[i]=="setosa") {print(1)
    right=right+1}
  if(micha==cha2 & data_train$Species[i]=="versicolor"){print(2)
    right=right+1}
  if(micha==cha3 & data_train$Species[i]=="virginica") {print(3)
    right=right+1}
}

print("拟合度为：")
print((right*1.0/SUM))
print("sum")
print(SUM)
print("right")
print(right)

#训练结束,看看参数
print("w1")
print(w1)
print("w2")
print(w2)

data_test$Sepal.Length=(data_test$Sepal.Length-min(data_test$Sepal.Length))*1.0/
  (max(data_test$Sepal.Length)-min(data_test$Sepal.Length))

data_test$Sepal.Width=(data_test$Sepal.Width-min(data_test$Sepal.Width))*1.0/
  (max(data_test$Sepal.Width)-min(data_test$Sepal.Width))

data_test$Petal.Length=(data_test$Petal.Length-min(data_test$Petal.Length))*1.0/
  (max(data_test$Petal.Length)-min(data_test$Petal.Length))

data_test$Petal.Width=(data_test$Petal.Width-min(data_test$Petal.Width))*1.0/
  (max(data_test$Petal.Width)-min(data_test$Petal.Width))

#用测试数据测试一下准确率如何

SUM=length(data_test$Species)
right=0

for(i in 1:length(data_test$Species))
{
  #1*4
  x1=matrix(c(data_test$Sepal.Length[i],data_test$Sepal.Width[i],data_test$Petal.Length[i],data_test$Petal.Width[i]),1,4)

  #矩阵相乘 %*%
  #得到经过第一个隐藏层的输出，也就是最后输出层的输入
  x2=f(x1%*%w1-a1)

  #得到最后的输出层，是1*3的矩阵
  x3=f(x2%*%w2-a2)

  cha11=(y1-x3)
  cha22=(y2-x3)
  cha33=(y3-x3)

  cha1=0
  cha2=0
  cha3=0

  for(j in 1:3)
  {
    cha1=cha1+abs(cha11[j])
    cha2=cha2+abs(cha22[j])
    cha3=cha3+abs(cha33[j])
  }

  micha=min(cha1,cha2,cha3)
  #cat("micha",micha,"\n")
  #cat("cha1",cha1,"\n")
  #cat("cha2",cha2,"\n")
  #cat("cha3",cha3,"\n")

  if(micha==cha1 & data_test$Species[i]=="setosa") {print(1)
    right=right+1}
  if(micha==cha2 & data_test$Species[i]=="versicolor"){print(2)
    right=right+1}
  if(micha==cha3 & data_test$Species[i]=="virginica") {print(3)
    right=right+1}
}

print("accuracy:")
print((right*1.0/SUM))
cat("right",right)
print("")
cat("SUM",SUM)

2. 带包实现BP人工神经完成iris

rm(list=ls())
#install.packages("sampling")
library(nnet)
library(sampling)

setwd("C:/Users/Administrator/Desktop/R???????????ฺพ???าต/สต??4-?หน?????????")

data("iris")

iris$Sepal.Length=(iris$Sepal.Length-min(iris$Sepal.Length))*1.0/
  (max(iris$Sepal.Length)-min(iris$Sepal.Length))

iris$Sepal.Width=(iris$Sepal.Width-min(iris$Sepal.Width))*1.0/
  (max(iris$Sepal.Width)-min(iris$Sepal.Width))

iris$Petal.Length=(iris$Petal.Length-min(iris$Petal.Length))*1.0/
  (max(iris$Petal.Length)-min(iris$Petal.Length))

iris$Petal.Width=(iris$Petal.Width-min(iris$Petal.Width))*1.0/
  (max(iris$Petal.Width)-min(iris$Petal.Width))

n=round(3/5*nrow(iris)/3)

sub_train=strata(iris,stratanames=("Species"),size=rep(n,3),method="srswor")
head(sub_train)
colnames(iris)<-c("x1","x2","x3","x4","y")

data_train=iris[sub_train$ID_unit,]
data_test=iris[-sub_train$ID_unit,]
dim(data_train)
dim(data_test)

model1=nnet(y~.,data=data_train,size=6,decay=5e-5,maxit=1000)
pred=predict(model1,data_test[,1:4],type="class")
P=sum(as.numeric(pred==data_test$y))/nrow(data_test)
cat("accuracy",P*100,"%\n")
table(data_test$y,pred)