“tfread”存在吗？

Question

在R中是否有一种有效的方式来阅读转置的.csv文件？

例如，请考虑以下文本文件：

Name,Peter,Paul,Marry
Age,40,9,38

这可以使用以下内容读入包含有用列类的data.table：

library(data.table)    
file <-  tempfile("tmp.txt")
writeLines("Name,Peter,Paul,Mary\nAge,40,5,38\n", file)    

lines <- readLines(file)
lines <- lapply(lines, function(x) gsub(pattern=",", replacement="\n", x, fixed=TRUE))
lines <- lapply(lines[-3], fread)
do.call(cbind,lines)
#>     Name Age
#> 1: Peter  40
#> 2:  Paul   5
#> 3:  Mary  38

有没有更简单的方法来实现这一目标？是否有更高效的版本（我的文件是1 GB）？

请注意，对于列式存储，此类列主存储应该比data.table更容易阅读。

Answer 1

这是评论中@Dirk Eddelbuettel建议方法的实现。

String name = "yyyymmddword";
int year = Integer.parseInt(name.substring(0, 4));
int day = Integer.parseInt(name.substring(4, 6));
int month = Integer.parseInt(name.substring(6, 8));
String year = Integer.parseInt(name.substring(8));

Answer 2

DT=setDT(read.table(text=do.call(paste,transpose(fread(file,h=F))),h=T,stringsAsFactors = F))
DT
    Name Age
1: Peter  40
2:  Paul   5
3:  Mary  38



sapply(DT,class)
       Name         Age 
"character"   "integer" 

Answer 3

不幸的是，tfread似乎不存在。

200 observations of 20000 character/integer variables / 20000 observations of 200 character/integer variables的建议解决方案的时间安排：

1. readLines-fread (@jan-glx)：7 s / 1.2 s
2. fread-transpose-paste-read.table (@Onymambu)：8 s / 36 s
3. a ll-within-j: fread-transpose-write.csv-paste-fread (@Clayton Stanley)：5分钟/ 12秒
4. command line transpose-fread (@jan-glx)：2.4 s / 1.6 s
5. iotools-paste-paste-fread (@jan-glx)：1.4 s / 1.2 s
6. fread - transpose - type.convert（@ Frank）：4.2 s / 3.6 s

代码：

library(data.table)    
file <-  tempfile("tmp.txt")
p <- 100 # = 200 lines/columns
n <- 10000 # = 20000 values per line / rows
writeLines(rep(c(paste("Name",paste0(rep(c("Peter","Paul"), n), collapse = ","), sep=","),
                 paste("Age",paste0(rep(c("40","5"), n), collapse = ","), sep=",")
               ), p), file(file,"wb"))    

system.time({ # 1
lines <- readLines(file)
lines <- lapply(lines, function(x) gsub(pattern=",", replacement="\n", x, fixed=TRUE))
lines <- lapply(lines, fread)
dt <- do.call(cbind,lines)
dim(dt)
})

system.time({ # 2
DT=setDT(read.table(text=do.call(paste,transpose(fread(file,h=F))),h=T,stringsAsFactors = F))
dim(DT)
})

system.time({ # 3
aTbl = fread(file, colClasses="character", header=F)
invisible(
  aTbl[, .SD
       ][, transpose(.SD)
         ][, setnames(.SD, .SD[1, t(.SD)])                                                                                                                   
           ][2:.N                                                                                                                  
             ][, fread(paste0(capture.output(write.csv(.SD, stdout(), row.names=F, quote=F)), collapse='\n'))                        
               ][, {bTbl <<- copy(.SD); .SD}                                                                                           
                 ]  
)
dim(bTbl)
})

system.time({ # 4 wide
  dt <- fread(paste0("transpose -t -l 20005x205 --fsep , \"", file, "\""))
  dim(dt)
})

system.time({ # 4 long
dt <- fread(paste0("transpose -t -l 205x20005 --fsep , \"", file, "\""))
dim(dt)
})

system.time({ # 5
  infile <- file(file, "rb")
  df <- iotools::chunk.tapply(infile, function(x) {
    fread(paste0(apply(iotools::mstrsplit(x, sep=","), 2, paste0, collapse = ","), collapse = "\n"))
  }, CH.MERGE = cbind)
  dim(df)
})

system.time({ # 6
d <-  fread(file, header=FALSE); 
d <- d[, lapply(transpose(.SD[,-1]), type.convert)][, setnames(.SD, d[[1]])]
dim(d)
})

Answer 4

将@lmo的评论扩展为使用iotools：

  dt <- iotools::chunk.tapply(file(file, "rb"),  function(x) {
    fread(paste0(apply(iotools::mstrsplit(x, sep=","), 2, paste0, collapse = ","), collapse = "\n"))
  }, CH.MERGE = cbind)

这比其他现有解决方案更有效。

Answer 5

扩展@ngm的注释以使用命令行工具：

1. 下载：transpose.c
2. 编译：{{1​​}}
3. 使用：gcc transpose.c -o transpose

这需要了解表的粗略大小，并要求单个字符行结尾，即fread(paste0("transpose -t -l 205x20005 --fsep , \"", file, "\""))。它非常不方便，但比其他解决方案更快。