Question

How to save my trained Random Forest model and apply it to test data files one by one?

我已经问过这个问题并阅读了上面提到的帖子。现在我在这里发布我的尝试和最少数据，试图解决我的问题。

示例数据的链接在这里：

https://drive.google.com/drive/folders/11GttkeW2A1peRIE2DmiB9WJxTp8wPlEo?usp=sharing

在这个文件夹中，有 16 个控制训练数据文件，子文件夹“testset”包含完全不同的新测试数据，我想将我训练的随机森林 (RF) 分类器应用于这些数据。

我想根据少数感兴趣的药物治疗选择一些文件（不同生物样本上的数百个文件）作为训练集来构建模型（RF 分类器），保存它，并在以下情况下将其用于新数据我需要它。

代码：

 library(data.table)
 library(caret)
 library(caTools)
 library(randomForest)
 library(e1071)
 library(FNN)
 library(tidyverse)


 #####read data files in the dir and random sample a percentage of rows in each file

 txt_files_ls = list.files(pattern=c("*.txt") )
 txt_files_df <- lapply(txt_files_ls, fread)
 sampled_txt_files_df <- lapply(txt_files_df,function(x){
   x[sample(1:nrow(x), ceiling(nrow(x) * 0.4)), 1:131]
 })
 combined_df <- rbindlist(sampled_txt_files_df)
 fwrite(combined_df,"Sampled40percent.csv",row.names = FALSE)


 ##### I closed off R after the first piece to save memory space

 #read in the merged file adn tidy it up for classifier buiding
 #The idea later is to use "treatmentsum" as class labels to differentiate different treatment conditions
 #Pid_treatmentsum contains finer details in each individual patient cell responses

 df.raw <-fread("Sampled40percent.csv", header =TRUE,data.table = FALSE) %>% 
   select(c(18:131)) %>% 
   select(-c(3:4)) %>% 
   mutate(treatmentsum = factor(treatmentsum)) %>% 
   mutate(Pid_treatmentsum=factor(Pid_treatmentsum))->df

 str(df)

 ###################################feature cleaning
 #find nearZerovarance features
 nzv <- nearZeroVar(df[-c(1,2)], saveMetrics= TRUE)
 nzv.flagged <- nzv[nzv$nzv=="TRUE",]


 #remove nzv features
 df %>% 
   select(-row.names(nzv.flagged)) -> df.sanitized

 #identify correlated features
 df.Cor <- cor(df.sanitized[-c(1:2)])
 highlyCor <- findCorrelation(df.Cor, cutoff = .95,verbose = TRUE)

 corF <- names(df[highlyCor])
 

 df.sanitized<- df.sanitized[ ,-highlyCor]

 

#define control groups for kmeans based on my own interest
unique(df.sanitized$treatmentsum)

predead <- subset(df.sanitized,df.sanitized$treatmentsum== "3S_DMSO_00_predead")
live <- subset(df.sanitized,df.sanitized$treatmentsum== "3S_DMSO_00_live")
lena <- subset(df.sanitized,df.sanitized$treatmentsum== "3S_Lenalidomide_1uM_none")
STS <- subset(df.sanitized,df.sanitized$treatmentsum== "3S_STSVEN_100nM+10nM_none")

#bind to list
mylist <- list(predead,live,lena,STS)

# use kmeans to subsample data points by over-clustering
Kmeans.list <- map(.x = mylist,
                   .f = ~kmeans(scale(.x[,-c(1:2)]),
                                centers =100,
                                nstart=25,
                               iter.max = 100)) %>% 
  purrr::set_names(c("predead", "live", "Lena", "STS"))

#bind all the centers into a list for mapping
Kmeans_centers <- map(Kmeans.list, ~.x$centers)

#use map2 to select 10 cells that are close to the centers in each cluster
y <- map2(.x = mylist,.y=Kmeans_centers,
          .f=~get.knnx(scale(.x[,-c(1:2)]),.y, 10)) %>% 
  purrr::set_names(c("predead", "live", "Lena", "STS"))


#get the nn.index and select each datapoint out to form a lits
#bind all the centers into a list for mapping
y.nnindex <- map(y, ~.x$nn.index)%>% 
  purrr::set_names(c("predead", "live", "Lena", "STS"))

#subset by knn index

cl.list <- list()
for (i in 1:4) {
  idx <- sort(y.nnindex[[i]])
  cl.list[[i]] <- as.data.frame(cbind(idx, mylist[[i]][idx,]))
}

df.sampled.by.KNN <- do.call("rbind", cl.list)

write.csv(df.sampled.by.KNN, "4k representative samples each from 100 clusters .csv", row.names = F)



###### from here is the model training
df.sampled.by.KNN <- read.csv("70k representative samples each from 1000 clusters each phenotype by Kmeans then KNN.csv", header = T)



#going to train random forest
#splt datainto training and test
set.seed(123)
split <- sample.split(df.sampled.by.KNN$treatmentsum, SplitRatio = 0.8)
rftraining_set <- subset(df.sampled.by.KNN, split==TRUE)
rftest_set <- subset(df.sampled.by.KNN, split==FALSE)

#get feature importance

unique(as.factor(rftraining_set[,2]))
control <- trainControl(method="cv", number=10,verboseIter = TRUE, search = 'grid')
mtry <- sqrt(ncol(training_set))
tunegrid <- expand.grid(.mtry = mtry)
model <- train(x=scale(rftraining_set[,-c(1:3)]), y=as.factor(rftraining_set[,2]), data=rftraining_set, method = "rf", trControl = control,metric= "Accuracy", maximize = TRUE ,importance = TRUE, type="classification", ntree =500,tuneGrid=tunegrid,na.action = na.omit)


#upto here, the model works just fine
print(model)
prediction2 <- predict(model, rftest_set)
cm<-confusionMatrix(prediction2, as.factor(rftest_set[,2]),  positive = "1") 
cm <- as.table(cm)
cm<-as.data.frame(cm)
write.csv(cm, "Confusion matrix.csv", row.names = TRUE)

从这里开始，我认为我的模型经过训练并且工作正常，我应该将其保存/复制到测试集数据所在的位置，然后加载它，然后只需调用预测函数即可。

#problem starts here

saveRDS(model, "myrf.RData")
getwd()
setwd("C:/Users/mli/Desktop/Stack project/testset")
myrf <-load("myrf.RData")

我在加载模型时出错

Error in load("myrf.RData") : 
  bad restore file magic number (file may be corrupted) -- no data loaded
In addition: Warning messages:
1: In readChar(con, 5L, useBytes = TRUE) :
  truncating string with embedded nuls
2: file ‘myrf.RData’ has magic number 'X'
  Use of save versions prior to 2 is deprecated > )

加载保存的训练好的机器学习模型时出错

0 个答案: