Question

我正在使用RTextTools中的DocumentTermMatrix创建一个create_matrix()，并在此基础上创建container和model。它适用于非常大的数据集。

我为每个类别（因子水平）执行此操作。因此，对于每个类别，它都必须运行矩阵，容器和模型。当我在下面的代码中运行时（例如16 core / 64 gb）-它仅在一个内核中运行，并且使用的内存不到10％。

是否可以加快此过程？也许使用doparallel和foreach？任何信息肯定会有所帮助。

#import the required libraries
library("RTextTools")
library("hash")
library(tm)

for ( n in 1:length(folderaddress)){
    #Initialize the variables
    traindata = list()
    matrix = list()
    container = list()
    models = list()
    trainingdata = list()
    results = list()
    classifiermodeldiv = 0.80`

    #Create the directory to place the models and the output files
    pradd = paste(combinedmodelsaveaddress[n],"SelftestClassifierModels",sep="")
    if (!file.exists(pradd)){
        dir.create(file.path(pradd))
    }  
    Data$CATEGORY <- as.factor(Data$CATEGORY)

    #Read the training files
    X <- split(Data, Data$CATEGORY)
    data <- lapply(seq_along(X), function(x) as.data.frame(X[[x]])[,5])
    names(data) <- levels(Data$CATEGORY)
    list2env(data, envir = .GlobalEnv)
    files=as.matrix(names(data))
    fileno=length(files)
    fileno=as.integer(fileno)
    print(fileno)

    #For all the files in the training folder(the number of files in the training folder = Number of categories in Taxonomy)
    for(i in 1:fileno){
        filename = as.character(files[i,1])
        data1 = as.data.frame(data[i])
        data1 = as.matrix(data1)
        filenamechanged = gsub ("\\.[[:alnum:]]+","",filename)
        type = matrix(data = as.character(filenamechanged),nrow = length(data1[,1]),ncol=1 )
        data1 = cbind(data1,type)
        traindata[[i]] = data1
        print(i)
    }

    for(i in 1:fileno){
        #Obtain the unique classified data from the train files for one category
        trainingdata1 = as.data.frame(traindata[[i]][,1])
        uniquetraintweet = hash()
        typetrain1 = matrix(data=as.character(traindata[[i]][1,2]), ncol =1, nrow = length(trainingdata1[,1]))

        #If the training data is less than 10 records for a category, do not create a model
        #The model created based on a smaller set of data will not be accurate
        if (length(trainingdata1[,1])<200){
            matrix[[i]] = NULL
            next
        }

        #Obtain the unique classified data from the train files of all the other category except that is considered as training category
        trainingdata2=matrix(data="",nrow=0,ncol=1)

        for (j in 1:fileno){
            if ( j==i) next
            trainingdata2dummy = as.data.frame(traindata[[j]][,1])
            length(trainingdata1[,1])
            colnames(trainingdata2)="feedbacks"
            colnames(trainingdata2dummy)="feedbacks"
            trainingdata2 = rbind(trainingdata2,trainingdata2dummy)

        }

        #Consider one category as training set and make the remaining categories as Others
        typetrain2 = matrix(data="ZZOther",nrow=length(trainingdata2[,1]),ncol=1)
        colnames(trainingdata1)="feedbacks"
        trainingdata[[i]]=rbind(trainingdata1,trainingdata2)
        colnames(typetrain1)="type"
        colnames(typetrain2)="type"
        type=rbind(typetrain1,typetrain2)
        trainingdata[[i]] = cbind(trainingdata[[i]],type)
        trainingdata[[i]]=trainingdata[[i]][sample(nrow(trainingdata[[i]])),]

        #Input the training set and other set to the classifier
        mindoc = max(1,floor(min(0.001*length(trainingdata[[i]][,1]),3)))

        #Create Matrix        
        matrix[[i]] <- create_matrix(trainingdata[[i]][,1], language="english",
                                     removeNumbers=FALSE, stemWords=FALSE,weighting=weightTf,minWordLength=3, minDocFreq=mindoc, maxDocFreq=floor(0.5*(length(trainingdata[[i]][,1]))))
        #rowTotals <- apply(matrix[[i]] , 1, sum) #Find the sum of words in each Document
        #matrix[[i]]   <- matrix[[i]][rowTotals> 0,] 
        print(i)

        #Create Container             
        container[[i]] <- create_container(matrix[[i]],trainingdata[[i]][,2],trainSize=1:length(trainingdata[[i]][,1]),virgin=FALSE)
        print(i)

        #Create Models  
        models[[i]] <- train_models(container[[i]], algorithms=c("SVM"))
        print(i)
    }

    save(matrix, file = paste(pradd,"/Matrix",sep=""))
    save(models, file = paste(pradd,"/Models",sep=""))   
}

Answer 1

以下是并行处理RTextTools的示例。我使用要查找的信息here创建了虚拟函数。

函数myFun遵循上述链接中的介绍-最后，它写入一个包含分析/摘要的csv文件（未指定目录）。然后，它是base R软件包parallel的直接应用，以便并行运行myFun。

library(parallel)
library(RTextTools)
# I. A dummy function
# Uses RTextTools
myFun <- function (trainMethod) {
  library(RTextTools)
  data(USCongress)
  # Create the document-term matrix
  doc_matrix <- create_matrix(USCongress$text, language="english", removeNumbers=TRUE,
                              stemWords=TRUE, removeSparseTerms=.998)
  container <- create_container(doc_matrix, USCongress$major, trainSize=1:4000,
                                testSize=4001:4449, virgin=FALSE)
  # Train
  model <- train_model(container,trainMethod)
  classify <- classify_model(container, model)
  # Analytics
  analytics <- create_analytics(container,
                                cbind(classify))
  summary(analytics)
  # Saving
  nameToSave <- paste(trainMethod, 'DocumentSummary.csv', sep = '_')
  write.csv(analytics@document_summary, nameToSave)
}

# II. Parallel Processing
# 
# 1. Vector for parallelization & number of cores available
trainMethods <- c('SVM','GLMNET','MAXENT','SLDA','BOOSTING')
num_cores <- detectCores() - 1L
# 2. Start a cluster
cl <- makeCluster(num_cores)
# 3. Export Variables needed to the cluster
# specifying exactly which variables should be exported
clusterExport(cl, varlist = c('myFun', 'trainMethods'))
# 4. do in parallel
parLapply(cl, seq_along(trainMethods), function (n) myFun(trainMethod = trainMethods[n]))
# stop the cluster
stopCluster(cl)

在您的情况下，您必须将代码转换为函数myFun (n, ...)，其中n是seq_along(folderaddress)的元素，并且当然用seq_along(trainMethods)代替{{ 1}}在seq_along(folderaddress)中。

当然有机会通过并行化来增强代码。问题在于没有样本数据，任何建议的改进只是猜测。

Create_Matrix'RTextTools'包的并行计算

1 个答案: