data.table (version 1.9.2)

fread: Fast and friendly file finagler

Description

Similar to read.table but faster and more convenient. All controls such as sep, colClasses and nrows are automatically detected. bit64::integer64 types are also detected and read directly without needing to read as character before converting. Dates are read as character currently. They can be converted afterwards using the excellent fasttime package or standard base functions.

`fread` is for regular delimited files; i.e., where every row has the same number of columns. In future, secondary separator (sep2) may be specified within each column. Such columns will be read as type list where each cell is itself a vector.

Usage

fread(input, sep="auto", sep2="auto", nrows=-1L, header="auto", na.strings="NA",
stringsAsFactors=FALSE, verbose=FALSE, autostart=30L, skip=-1L, select=NULL,
drop=NULL, colClasses=NULL,
integer64=getOption("datatable.integer64"),         # default: "integer64"
showProgress=getOption("datatable.showProgress")    # default: TRUE
)

Arguments

input

Either the file name to read (containing no \n character), a shell command that preprocesses the file (e.g. fread("grep blah filename")) or the input itself as a string (containing at least one \n), see examples. In both cases, a length 1 cha

sep

The separator between columns. Defaults to the first character in the set [,\t |;:] that exists on line autostart outside quoted ("") regions, and separates the rows above autostart into a consistent num

sep2

The separator within columns. A list column will be returned where each cell is a vector of values. This is much faster using less working memory than strsplit afterwards or similar techniques. For each column sep2

nrows

The number of rows to read, by default -1 means all. Unlike read.table, it doesn't help speed to set this to the number of rows in the file (or an estimate), since the number of rows is automatically determined and is already fast. Only set <

header

Does the first data line contain column names? Defaults according to whether every non-empty field on the first data line is type character. If so, or TRUE is supplied, any empty column names are given a default name.

na.strings

A character vector of strings to convert to NA_character_. By default for columns read as type character ",," is read as a blank string ("") and ",NA," is read as NA_character_. Typical alte

stringsAsFactors

Convert all character columns to factors?

verbose

Be chatty and report timings?

autostart

Any line number within the region of machine readable delimited text, by default 30. If the file is shorter or this line is empty (e.g. short files with trailing blank lines) then the last non empty line (with a non empty line above that) is used. This li

skip

If -1 (default) use the procedure described below starting on line autostart to find the first data row. skip>=0 means ignore autostart and take line skip+1 as the first data row (or column names accordi

select

Vector of column names or numbers to keep, drop the rest.

drop

Vector of column names or numbers to drop, keep the rest.

colClasses

A character vector of classes (named or unnamed), as read.csv. Or a named list of vectors of column names or numbers, see examples. colClasses in fread is intended for rare overrides, not for routine use. fread will only promote a column to a higher type

integer64

"integer64" (default) reads columns detected as containing integers larger than 2^31 as type bit64::integer64. Alternatively, "double"|"numeric" reads as base::read.csv does; i.e., possibly with loss of precision and

showProgress

TRUE displays progress on the console using \r. It is produced in fread's C code where the very nice (but R level) txtProgressBar and tkProgressBar are not easily available.

Value

• A data.table.

Details

Once the separator is found on line autostart, the number of columns is determined. Then the file is searched backwards from autostart until a row is found that doesn't have that number of columns. Thus, the first data row is found and any human readable banners are automatically skipped. This feature can be particularly useful for loading a set of files which may not all have consistently sized banners. Setting skip>0 overrides this feature by setting autostart=skip+1 and turning off the search upwards step.

The first 5 rows, middle 5 rows and last 5 rows are then read to determine column types. The lowest type for each column is chosen from the ordered list integer, integer64, double, character. This enables fread to allocate exactly the right number of rows, with columns of the right type, up front once. The file may of course still contain data of a different type in rows other than first, middle and last 5. In that case, the column types are bumped mid read and the data read on previous rows is coerced. Setting verbose=TRUE reports the line and field number of each mid read type bump, and how long this type bumping took (if any).

There is no line length limit, not even a very large one. Since we are encouraging list columns (i.e. sep2) this has the potential to encourage longer line lengths. So the approach of scanning each line into a buffer first and then rescanning that buffer is not used. There are no buffers used in fread's C code at all. The field width limit is limited by R itself: the maximum width of a character string (currenly 2^31-1 bytes, 2GB).

character columns can be quoted (...,2,"Joe Bloggs",3.14,...) or not quoted (...,2,Joe Bloggs,3.14,...). Spaces and other whitepace (other than sep and \n) may appear in an unquoted character field, provided the field doesn't contain sep itself. Therefore quoting character values is only required if sep itself appears in the string value. Quoting may also be used to signify that numeric data should be read as text (or that can be achieved by specifying the column type via colClasses). Field quoting is automatically detected and no arguments are needed to control it.

The filename extension (such as .csv) is irrelevant for "auto" sep and sep2. Separator detection is entirely driven by the file contents. This can be useful when loading a set of different files which may not be named consistently, or may not have the extension .csv despite being csv. Some datasets have been collected over many years, one file per day for example. Sometimes the file name format has changed at some point in the past or even the format of the file itself. So the idea is that you can loop fread through a set of files and as long as each file is regular and delimited, fread can read them all. Whether they all stack is another matter but at least each one is read quickly without you needing to vary colClasses in read.table or read.csv.

All known line endings are detected automatically: \n (*NIX including Mac), \r\n (Windows CRLF), \r (old Mac) and \n\r (just in case). There is no need to convert input files first. fread running on any architecture will read a file from any architecture. Both \r and \n may be embedded in character strings (including column names) provided the field is quoted.

If an empty line is encountered then reading stops there, with warning if any text exists after the empty line such as a footer. The first line of any text discarded is included in the warning message.

References

Background : http://cran.r-project.org/doc/manuals/R-data.html http://stackoverflow.com/questions/1727772/quickly-reading-very-large-tables-as-dataframes-in-r www.biostat.jhsph.edu/~rpeng/docs/R-large-tables.html https://stat.ethz.ch/pipermail/r-help/2007-August/138315.html http://www.cerebralmastication.com/2009/11/loading-big-data-into-r/ http://stackoverflow.com/questions/9061736/faster-than-scan-with-rcpp http://stackoverflow.com/questions/415515/how-can-i-read-and-manipulate-csv-file-data-in-c http://stackoverflow.com/questions/9352887/strategies-for-reading-in-csv-files-in-pieces http://stackoverflow.com/questions/11782084/reading-in-large-text-files-in-r http://stackoverflow.com/questions/45972/mmap-vs-reading-blocks http://stackoverflow.com/questions/258091/when-should-i-use-mmap-for-file-access http://stackoverflow.com/a/9818473/403310 http://stackoverflow.com/questions/9608950/reading-huge-files-using-memory-mapped-files

finagler = "to get or achieve by guile or manipulation" http://dictionary.reference.com/browse/finagler

See Also

read.csv, url html{}

Examples

# Demo speedup
n=1e6
DT = data.table( a=sample(1:1000,n,replace=TRUE),
                 b=sample(1:1000,n,replace=TRUE),
                 c=rnorm(n),
                 d=sample(c("foo","bar","baz","qux","quux"),n,replace=TRUE),
                 e=rnorm(n),
                 f=sample(1:1000,n,replace=TRUE) )
DT[2,b:=NA_integer_]
DT[4,c:=NA_real_]
DT[3,d:=NA_character_]
DT[5,d:=""]
DT[2,e:=+Inf]
DT[3,e:=-Inf]

write.table(DT,"test.csv",sep=",",row.names=FALSE,quote=FALSE)
cat("File size (MB):", round(file.info("test.csv")$size/1024^2),"\n")
# 50 MB (1e6 rows x 6 columns)

system.time(DF1 <-read.csv("test.csv",stringsAsFactors=FALSE))
# 60 sec (first time in fresh R session)

system.time(DF1 <- read.csv("test.csv",stringsAsFactors=FALSE))
# 30 sec (immediate repeat is faster, varies)

system.time(DF2 <- read.table("test.csv",header=TRUE,sep=",",quote="",
    stringsAsFactors=FALSE,comment.char="",nrows=n,
    colClasses=c("integer","integer","numeric",
                 "character","numeric","integer")))
# 10 sec (consistently). All known tricks and known nrows, see references.

require(data.table)
system.time(DT <- fread("test.csv"))
#  3 sec (faster and friendlier)

require(sqldf)
system.time(SQLDF <- read.csv.sql("test.csv",dbname=NULL))
# 20 sec (friendly too, good defaults)

require(ff)
system.time(FFDF <- read.csv.ffdf(file="test.csv",nrows=n))
# 20 sec (friendly too, good defaults)

identical(DF1,DF2)
all.equal(as.data.table(DF1), DT)
identical(DF1,within(SQLDF,{b<-as.integer(b);c<-as.numeric(c)}))
identical(DF1,within(as.data.frame(FFDF),d<-as.character(d)))

# Scaling up ...
l = vector("list",10)
for (i in 1:10) l[[i]] = DT
DTbig = rbindlist(l)
tables()
write.table(DTbig,"testbig.csv",sep=",",row.names=FALSE,quote=FALSE)
# 500MB (10 million rows x 6 columns)

system.time(DF <- read.table("testbig.csv",header=TRUE,sep=",",         
    quote="",stringsAsFactors=FALSE,comment.char="",nrows=1e7,                     
    colClasses=c("integer","integer","numeric",
                 "character","numeric","integer")))
# 100-200 sec (varies)

system.time(DT <- fread("testbig.csv"))
# 30-40 sec

all(mapply(all.equal, DF, DT))


# Real data example (Airline data)
# http://stat-computing.org/dataexpo/2009/the-data.html

download.file("http://stat-computing.org/dataexpo/2009/2008.csv.bz2",
              destfile="2008.csv.bz2")
# 109MB (compressed)

system("bunzip2 2008.csv.bz2")                                          
# 658MB (7,009,728 rows x 29 columns)

colClasses = sapply(read.csv("2008.csv",nrows=100),class)
# 4 character, 24 integer, 1 logical. Incorrect.

colClasses = sapply(read.csv("2008.csv",nrows=200),class)
# 5 character, 24 integer. Correct. Might have missed data only using 100 rows
# since read.table assumes colClasses is correct.

system.time(DF <- read.table("2008.csv", header=TRUE, sep=",",          
    quote="",stringsAsFactors=FALSE,comment.char="",nrows=7009730,      
    colClasses=colClasses)
# 360 secs

system.time(DT <- fread("2008.csv"))
#  40 secs

table(sapply(DT,class))
# 5 character and 24 integer columns. Correct without needing to worry about colClasses
# issue above.


# Reads URLs directly :
fread("http://www.stats.ox.ac.uk/pub/datasets/csb/ch11b.dat")

# Reads text input directly :
fread("A,B
1,2
3,4")

# Reads pasted input directly :
fread("A,B
1,2
3,4
")

# Finds the first data line automatically :
fread("This is perhaps a banner line or two or ten.
A,B
1,2
3,4
")

# Detects whether column names are present automatically :
fread("1,2
3,4
")

# Numerical precision :

DT = fread("A
1.010203040506070809010203040506
")   # silent loss of precision
DT[,sprintf("%.15E",A)]   # stored accurately as far as double precision allows

DT = fread("A
1.46761e-313
")   # detailed warning about ERANGE; read as 'numeric'
DT[,sprintf("%.15E",A)]   # beyond what double precision can store accurately to 15 digits

# For greater accuracy use colClasses to read as character, then package Rmpfr.

# colClasses
data = "A,B,C,D
1,3,5,7
2,4,6,8
"
fread(data, colClasses=c(B="character",C="character",D="character"))  # as read.csv
fread(data, colClasses=list(character=c("B","C","D")))    # saves typing
fread(data, colClasses=list(character=2:4))     # same using column numbers

# drop
fread(data, colClasses=c("B"="NULL","C"="NULL"))   # as read.csv
fread(data, colClasses=list(NULL=c("B","C")))      # 
fread(data, drop=c("B","C"))      # same but less typing, easier to read
fread(data, drop=2:3)             # same using column numbers

# select
# (in read.csv you need to work out which to drop)
fread(data, select=c("A","D"))    # less typing, easier to read
fread(data, select=c(1,4))        # same using column numbers