我有一個理解R中的gradDescent package的問題。讓我們說我有一個獨立變量的數據集,我想對這些數據運行一個簡單的線性迴歸並估計一個模型,並且其參數使用分批梯度下降(GD)算法。graddescent包和lm函數不同
例如,我正在使用here中給出的數據集。第一列是自變量(X),第二列是因變量(Y)。
我寫了我自己的R代碼批量梯度下降算法。我使用的學習率爲0.01,迭代次數爲1500.估計的模型是y = -3.630291 + 1.166362 x。參數的初始值都選擇爲1.
我也想檢查我的代碼是否正常工作。我用R內置函數lm包來比較。這些參數非常接近R中線性迴歸函數給出的結果。因此,在這種情況下,通過我們的梯度下降算法獲得的線性迴歸模型是y = -3.896 + 1.193 x。
但是,最近我發現了一個R包,gradDescent,我想看看它是如何工作的。使用相同的學習率和最大迭代次數,我得到了模型y = -1.229567 + 0.9257195x的結果(這些值隨着我每次運行,因爲我設置了seed = NULL)。
GD <- function(dataTrain, alpha=0.1, maxIter=10, seed=NULL){
#convert data.frame dataSet in matrix
dataTrain <- matrix(unlist(dataTrain), ncol=ncol(dataTrain), byrow=FALSE)
#shuffle data train
set.seed(seed)
dataTrain <- dataTrain[sample(nrow(dataTrain)), ]
set.seed(NULL)
#initialize theta
theta <- getTheta(ncol(dataTrain), seed=seed)
#bind 1 column to dataTrain
dataTrain <- cbind(1, dataTrain)
#parse dataTrain into input and output
inputData <- dataTrain[,1:ncol(dataTrain)-1]
outputData <- dataTrain[,ncol(dataTrain)]
#temporary variables
temporaryTheta <- matrix(ncol=length(theta), nrow=1)
updateRule <- matrix(0, ncol=length(theta), nrow=1)
#constant variables
rowLength <- nrow(dataTrain)
#loop the gradient descent
for(iteration in 1:maxIter){
error <- (inputData %*% t(theta)) - outputData
for(column in 1:length(theta)){
term <- error * inputData[,column]
#calculate gradient
gradient <- sum(term)/rowLength
updateRule[1,column] <- updateRule[1,column] + (alpha*gradient)
temporaryTheta[1,column] = theta[1,column] - updateRule[1,column]
}
#update all theta in the current iteration
theta <- temporaryTheta
}
result <- theta
return(result)
}
在這裏,給出getTheta功能,
getTheta <- function(columnLength, minTheta=0, maxTheta=1, seed=NULL){
#create static random
set.seed(seed)
#random a value
thetaList <- runif(columnLength, min=minTheta, max=maxTheta)
#clear static random
set.seed(seed)
#transform into matrix
result <- matrix(unlist(thetaList), ncol=columnLength, nrow=1, byrow=FALSE)
return(result)
}
的包裝選擇的初始值隨機。此外,它在運行GD算法之前將數據洗牌。我玩了一下。我將參數的初始值分配爲1,並停止對數據進行混洗。但是我不能認真地理解哪裏出了問題(或者是正確的),因爲我不能和我自己的GD代碼和R的lm函數得到相同的結果。請問有人能解釋一下嗎?
install.packages("gradDescent")
library(gradDescent)
URL_subs <-"https://raw.githubusercontent.com/ahawker/machine-learning-coursera/master/ex1/ex1data1.txt"
data <- read.table(URL_subs, header=FALSE, sep=",")
########## gradDescent Function ##########
GD(data, alpha = 0.01, maxIter = 1500, seed = NULL)
# [,1] [,2]
#[1,] -1.312882 0.9281769
########## R bulit-in function ##########
model <- lm(data$V2~ ., data = data)
model
#Call:
# lm(formula = data$V2 ~ ., data = data)
#
#Coefficients:
# (Intercept) V1
# -3.896 1.193
注意:我可以提供我寫的,但基本上,我試圖理解爲什麼這個包提供了比LM包更多不同的參數估計。
編輯: 是因爲代碼中的那一行嗎?
updateRule[1,column] <- updateRule[1,column] + (alpha*gradient)
當(在1(柱:長度(THETA)))第二循環結束時,代碼不重置updateRule矩陣但保持增加(阿爾法*梯度)的兩列矩陣在每次迭代中。我錯了嗎?
當我在迭代中找到參數更新後將此updateRule矩陣重置爲零時,我得到的模型y = -3.570819 +1.160388 x非常接近我擁有的和lm包提供的值。
EDIT 2 與gradDescent包在我原來的帖子中提到的什麼是錯的。 updateRule矩陣未被重置。我只是在循環中添加一行代碼,並沒有改變其他任何東西。 getTheta和GD函數與發佈包的作者相同。
我舉兩個例子來糾正它。我使用的第一個數據集有一個獨立變量,第二個數據集有兩個獨立變量。對於這兩個示例,我使用隨機生成的首字母,這是包中的想法。對於第二個例子,我規範化了數據,因爲輸入變量的大小順序是不同的。房屋面積(尺寸)比臥室面積大1000倍左右。
例1
URL_subs <-"https://raw.githubusercontent.com/ahawker/machine-learning-coursera/master/ex1/ex1data1.txt"
data <- read.table(URL_subs, header=FALSE, sep=",")
getTheta <- function(columnLength, minTheta=0, maxTheta=1, seed=NULL){
#create static random
set.seed(seed)
#random a value
thetaList <- runif(columnLength, min=minTheta, max=maxTheta)
#clear static random
set.seed(seed)
#transform into matrix
result <- matrix(unlist(thetaList), ncol=columnLength, nrow=1, byrow=FALSE)
return(result)
}
GD <- function(dataTrain, alpha=0.1, maxIter=10, seed=NULL){
#convert data.frame dataSet in matrix
dataTrain <- matrix(unlist(dataTrain), ncol=ncol(dataTrain), byrow=FALSE)
#shuffle data train
set.seed(seed)
dataTrain <- dataTrain[sample(nrow(dataTrain)), ]
set.seed(NULL)
#initialize theta
theta <- getTheta(ncol(dataTrain), seed=seed)
#bind 1 column to dataTrain
dataTrain <- cbind(1, dataTrain)
#parse dataTrain into input and output
inputData <- dataTrain[,1:ncol(dataTrain)-1]
outputData <- dataTrain[,ncol(dataTrain)]
#temporary variables
temporaryTheta <- matrix(ncol=length(theta), nrow=1)
updateRule <- matrix(0, ncol=length(theta), nrow=1)
#constant variables
rowLength <- nrow(dataTrain)
#loop the gradient descent
for(iteration in 1:maxIter){
error <- (inputData %*% t(theta)) - outputData
for(column in 1:length(theta)){
term <- error * inputData[,column]
#calculate gradient
gradient <- sum(term)/rowLength
updateRule[1,column] <- updateRule[1,column] + (alpha*gradient)
temporaryTheta[1,column] = theta[1,column] - updateRule[1,column]
}
updateRule <- matrix(0, ncol=length(theta), nrow=1)
#update all theta in the current iteration
theta <- temporaryTheta
}
result <- theta
return(result)
}
GD(data, alpha = 0.01, maxIter = 1500, seed = NULL)
# [,1] [,2]
#[1,] -3.602297 1.16355
########## R built-in lm function ##########
model <- lm(data$V2~ ., data = data)
model
#Call:
# lm(formula = data$V2 ~ ., data = data)
#
#Coefficients:
# (Intercept) V1
# -3.896 1.193
例2
data <- read.csv("https://raw.githubusercontent.com/ethen8181/machine-learning/master/linear_regression/housing.txt",
header = TRUE,
sep = ",")
getTheta <- function(columnLength, minTheta=0, maxTheta=1, seed=NULL){
#create static random
set.seed(seed)
#random a value
thetaList <- runif(columnLength, min=minTheta, max=maxTheta)
#clear static random
set.seed(seed)
#transform into matrix
result <- matrix(unlist(thetaList), ncol=columnLength, nrow=1, byrow=FALSE)
return(result)
}
GD <- function(dataTrain, alpha=0.1, maxIter=10, seed=NULL){
#convert data.frame dataSet in matrix
dataTrain <- matrix(unlist(dataTrain), ncol=ncol(dataTrain), byrow=FALSE)
#shuffle data train
set.seed(seed)
dataTrain <- dataTrain[sample(nrow(dataTrain)), ]
set.seed(NULL)
#initialize theta
theta <- getTheta(ncol(dataTrain), seed=seed)
#bind 1 column to dataTrain
dataTrain <- cbind(1, dataTrain)
#parse dataTrain into input and output
inputData <- dataTrain[,1:ncol(dataTrain)-1]
outputData <- dataTrain[,ncol(dataTrain)]
#temporary variables
temporaryTheta <- matrix(ncol=length(theta), nrow=1)
updateRule <- matrix(0, ncol=length(theta), nrow=1)
#constant variables
rowLength <- nrow(dataTrain)
#loop the gradient descent
for(iteration in 1:maxIter){
error <- (inputData %*% t(theta)) - outputData
for(column in 1:length(theta)){
term <- error * inputData[,column]
#calculate gradient
gradient <- sum(term)/rowLength
updateRule[1,column] <- updateRule[1,column] + (alpha*gradient)
temporaryTheta[1,column] = theta[1,column] - updateRule[1,column]
}
updateRule <- matrix(0, ncol=length(theta), nrow=1)
#update all theta in the current iteration
theta <- temporaryTheta
}
result <- theta
return(result)
}
GD(data, alpha = 0.05, maxIter = 500, seed = NULL)
# [,1] [,2] [,3]
#[1,] 340412.7 110630 -6648.375
########## R built-in lm function ##########
housing <- read.csv("https://raw.githubusercontent.com/ethen8181/machine-learning/master/linear_regression/housing.txt",
header = TRUE,
sep = ",")
normalized <- apply(housing[ , -3 ], 2, scale)
normalized_data <- data.frame(cbind(normalized, price = housing$price))
model <- lm(price ~ ., data = normalized_data)
model
#Call:
# lm(formula = price ~ ., data = normalized_data)
#
#Coefficients:
# (Intercept) area bedrooms
# 340413 110631 -6649
是的,謝謝!這不是我的包或功能。我還做了一些小改動,比如重置更新規則,現在我得到了與lm包類似的結果。我將編輯我的文章並顯示結果。 –