Faster way to read fixed-width files
I work with a lot of fixed width files (i.e., no separating character) that I need to read into R. So, there is usually a definition of the column width to parse the string into variables. I can use read.fwf to read in the data without a problem. However, for large files, this can take a long time. For a recent dataset, this took 800 seconds to read in a dataset with ~500,000 rows and 143 variables.
seer9 <- read.fwf("~/data/rawdata.txt",
widths = cols,
header = FALSE,
buffersize = 250000,
colClasses = "character",
stringsAsFactors = FALSE))
fread in the data.table package in R is awesome for solving most data read problems, except it doesn't parse fixed width files. However, I can read each line in as a single character string (~500,000 rows, 1 column). This takes 3-5 seconds. (I love data.table.)
seer9 <- fread("~/data/rawdata.txt", colClasses = "character",
sep = "\n", header = FALSE, verbose = TRUE)
There are a number of good posts on SO on how to parse text files. See JHoward's suggestion here, to create a matrix of start and end columns, and substr to parse the data. See GSee's suggestion here to use strsplit. I couldn't figure out how to make that work with this data. (Also, Michael Smith made some suggestions on the data.table mailing list involving sed that were beyond my ability to implement.) Now, using fread and substr() I can do the whole thing in about 25-30 seconds. Note that coercing to a data.table at end takes a chunk of time (5 sec?).
end_col <- cumsum(cols)
start_col <- end_col - cols + 1
start_end <- cbind(start_col, end_col) # matrix of start and end positions
text <- lapply(seer9, function(x) {
apply(start_end, 1, function(y) substr(x, y[1], y[2]))
})
dt <- data.table(text$V1)
setnames(dt, old = 1:ncol(dt), new = seervars)
What I am wondering is whether this can be improved any further? I know I am not the only one who has to read fixed width files, so if this could be made faster, it would make loading even larger files (with millions of rows) more tolerable. I tried using parallel with mclapply and data.tableinstead of lapply, but those didn't change anything. (Likely due to my inexperience in R.) I imagine that an Rcpp function could be written to do this really fast, but that is beyond my skill set. Also, I may not be using lapply and apply appropriately.
My data.table implementation (with magrittr chaining) takes the same time:
text <- seer9[ , apply(start_end, 1, function(y) substr(V1, y[1], y[2]))] %>%
data.table(.)
Can anyone make suggestions to improve the speed of this? Or is this about as good as it gets?
Here is code to create a similar data.table within R (rather than linking to actual data). It should have 331 characters, and 500,000 rows. There are spaces to simulate missing fields in the data, but this is NOT space delimited data. (I am reading raw SEER data, in case anyone is interested.) Also including column widths (cols) and variable names (seervars) in case this helps someone else. These are the actual column and variable definitions for SEER data.
seer9 <-
data.table(rep((paste0(paste0(letters, 1000:1054, " ", collapse = ""), " ")),
500000))
cols = c(8,10,1,2,1,1,1,3,4,3,2,2,4,4,1,4,1,4,1,1,1,1,3,2,2,1,2,2,13,2,4,1,1,1,1,3,3,3,2,3,3,3,3,3,3,3,2,2,2,2,1,1,1,1,1,6,6,6,2,1,1,2,1,1,1,1,1,2,2,1,1,2,1,1,1,1,1,1,1,1,1,1,1,1,1,1,7,5,4,10,3,3,2,2,2,3,1,1,1,1,2,2,1,1,2,1,9,5,5,1,1,1,2,2,1,1,1,1,1,1,1,1,2,3,3,3,3,3,3,1,4,1,4,1,1,3,3,3,3,2,2,2,2)
seervars <- c("CASENUM", "REG", "MAR_STAT", "RACE", "ORIGIN", "NHIA", "SEX", "AGE_DX", "YR_BRTH", "PLC_BRTH", "SEQ_NUM", "DATE_mo", "DATE_yr", "SITEO2V", "LATERAL", "HISTO2V", "BEHO2V", "HISTO3V", "BEHO3V", "GRADE", "DX_CONF", "REPT_SRC", "EOD10_SZ", "EOD10_EX", "EOD10_PE", "EOD10_ND", "EOD10_PN", "EOD10_NE", "EOD13", "EOD2", "EOD4", "EODCODE", "TUMOR_1V", "TUMOR_2V", "TUMOR_3V", "CS_SIZE", "CS_EXT", "CS_NODE", "CS_METS", "CS_SSF1", "CS_SSF2", "CS_SSF3", "CS_SSF4", "CS_SSF5", "CS_SSF6", "CS_SSF25", "D_AJCC_T", "D_AJCC_N", "D_AJCC_M", "D_AJCC_S", "D_SSG77", "D_SSG00", "D_AJCC_F", "D_SSG77F", "D_SSG00F", "CSV_ORG", "CSV_DER", "CSV_CUR", "SURGPRIM", "SCOPE", "SURGOTH", "SURGNODE", "RECONST", "NO_SURG", "RADIATN", "RAD_BRN", "RAD_SURG", "SS_SURG", "SRPRIM02", "SCOPE02", "SRGOTH02", "REC_NO", "O_SITAGE", "O_SEQCON", "O_SEQLAT", "O_SURCON", "O_SITTYP", "H_BENIGN", "O_RPTSRC", "O_DFSITE", "O_LEUKDX", "O_SITBEH", "O_EODDT", "O_SITEOD", "O_SITMOR", "TYPEFUP", "AGE_REC", "SITERWHO", "ICDOTO9V", "ICDOT10V", "ICCC3WHO", "ICCC3XWHO", "BEHANAL", "HISTREC", "BRAINREC", "CS0204SCHEMA", "RAC_RECA", "RAC_RECY", "NHIAREC", "HST_STGA", "AJCC_STG", "AJ_3SEER", "SSG77", "SSG2000", "NUMPRIMS", "FIRSTPRM", "STCOUNTY", "ICD_5DIG", "CODKM", "STAT_REC", "IHS", "HIST_SSG_2000", "AYA_RECODE", "LYMPHOMA_RECODE", "DTH_CLASS", "O_DTH_CLASS", "EXTEVAL", "NODEEVAL", "METSEVAL", "INTPRIM", "ERSTATUS", "PRSTATUS", "CSSCHEMA", "CS_SSF8", "CS_SSF10", "CS_SSF11", "CS_SSF13", "CS_SSF15", "CS_SSF16", "VASINV", "SRV_TIME_MON", "SRV_TIME_MON_FLAG", "SRV_TIME_MON_PA", "SRV_TIME_MON_FLAG_PA", "INSREC_PUB", "DAJCC7T", "DAJCC7N", "DAJCC7M", "DAJCC7STG", "ADJTM_6VALUE", "ADJNM_6VALUE", "ADJM_6VALUE", "ADJAJCCSTG")
UPDATE: LaF did the entire read in just under 7 seconds from the raw .txt file. Maybe there is an even faster way, but I doubt anything could do appreciably better. Amazing package.
27 July 2015 Update Just wanted to provide a small update to this. I used the new readr package, and I was able to read in the entire file in 5 seconds using readr::read_fwf.
seer9_readr <- read_fwf("path_to_data/COLRECT.TXT",
col_positions = fwf_widths(cols))
Also, the updated stringi::stri_sub function is at least twice as fast as base::substr(). So, in the code above that uses fread to read the file (about 4 seconds), followed by apply to parse each line, the extraction of 143 variables took about 8 seconds with stringi::stri_sub compared to 19 for base::substr. So, fread plus stri_sub is still only about 12 seconds to run. Not bad.
seer9 <- fread("path_to_data/COLRECT.TXT",
colClasses = "character",
sep = "\n",
header = FALSE)
text <- seer9[ , apply(start_end, 1, function(y) substr(V1, y[1], y[2]))] %>%
data.table(.)
10 Dec 2015 update:
Please also see the answer below by @MichaelChirico who has added some great benchmarks and the iotools package.
Answer
Now that there are (between this and the other major question about effective reading of fixed-width files) a fair amount of options on the offer for reading in such files, I think some benchmarking is appropriate.
I'll use the following on-the-large-side (400 MB) file for comparison. It's just a bunch of random characters with randomly defined fields and widths:
set.seed(21394)
wwidth = 400L
rrows = 1000000
#creating the contents at random
contents = write.table(
replicate(
rrows,
paste0(sample(letters, wwidth, replace = TRUE), collapse = "")
),
file = "testfwf.txt",
quote = FALSE, row.names = FALSE, col.names = FALSE
)
#defining the fields & writing a dictionary
n_fields = 40L
endpoints = unique(
c(1L, sort(sample(wwidth, n_fields - 1L)), wwidth + 1L)
)
cols = list(
beg = endpoints[-(n_fields + 1L)],
end = endpoints[-1L] - 1L
)
dict = data.frame(
column = paste0("V", seq_len(length(endpoints)) - 1L)),
start = endpoints[-length(endpoints)] - 1,
length = diff(endpoints)
)
write.csv(dict, file = "testdic.csv", quote = FALSE, row.names = FALSE)
I'll compare five methods mentioned between these two threads (I'll add some others if the authors would like): the base version (read.fwf), piping the result of in2csv to fread (@AnandaMahto's suggestion), Hadley's new readr (read_fwf), that using LaF/ffbase (@jwijffls' suggestion), and an improved (streamlined) version of that suggested by the question author (@MarkDanese) combining fread with stri_sub from stringi.
Here is the benchmarking code:
library(data.table)
library(stringi)
library(readr)
library(LaF)
library(ffbase)
library(microbenchmark)
microbenchmark(
times = 5L,
utils = read.fwf("testfwf.txt", diff(endpoints), header = FALSE),
in2csv = fread(cmd = sprintf(
"in2csv -f fixed -s %s %s",
"testdic.csv", "testfwf.txt"
)),
readr = read_fwf("testfwf.txt", fwf_widths(diff(endpoints))),
LaF = {
my.data.laf = laf_open_fwf(
'testfwf.txt',
column_widths = diff(endpoints),
column_types = rep("character", length(endpoints) - 1L)
)
my.data = laf_to_ffdf(my.data.laf, nrows = rrows)
as.data.frame(my.data)
},
fread = {
DT = fread("testfwf.txt", header = FALSE, sep = "\n")
DT[ , lapply(seq_len(length(cols$beg)), function(ii) {
stri_sub(V1, cols$beg[ii], cols$end[ii])
})]
}
)
And the output:
# Unit: seconds
# expr min lq mean median uq max neval cld
# utils 423.76786 465.39212 499.00109 501.87568 543.12382 560.84598 5 c
# in2csv 67.74065 68.56549 69.60069 70.11774 70.18746 71.39210 5 a
# readr 10.57945 11.32205 15.70224 14.89057 19.54617 22.17298 5 a
# LaF 207.56267 236.39389 239.45985 237.96155 238.28316 277.09798 5 b
# fread 14.42617 15.44693 26.09877 15.76016 20.45481 64.40581 5 a
So it seems readr and fread + stri_sub are pretty competitive as the fastest; built-in read.fwf is the clear loser.
Note that the real advantage of readr here is that you can pre-specify column types; with fread you'll have to type convert afterwards.
EDIT: Adding some alternatives
At @AnandaMahto's suggestion I am including some more options, including one that appears to be a new winner! To save time I excluded the slowest options above in the new comparison. Here's the new code:
library(iotools)
microbenchmark(
times = 5L,
readr = read_fwf("testfwf.txt", fwf_widths(diff(endpoints))),
fread = {
DT = fread("testfwf.txt", header = FALSE, sep = "\n")
DT[ , lapply(seq_len(length(cols$beg)), function(ii) {
stri_sub(V1, cols$beg[ii], cols$end[ii])
})]
},
iotools = input.file(
"testfwf.txt", formatter = dstrfw,
col_types = rep("character", length(endpoints) - 1L),
widths = diff(endpoints)
),
awk = fread(header = FALSE, cmd = sprintf(
"awk -v FIELDWIDTHS='%s' -v OFS=', ' '{$1=$1 \"\"; print}' < testfwf.txt",
paste(diff(endpoints), collapse = " ")
))
)
And the new output:
# Unit: seconds
# expr min lq mean median uq max neval cld
# readr 7.892527 8.016857 10.293371 9.527409 9.807145 16.222916 5 a
# fread 9.652377 9.696135 9.796438 9.712686 9.807830 10.113160 5 a
# iotools 5.900362 7.591847 7.438049 7.799729 7.845727 8.052579 5 a
# awk 14.440489 14.457329 14.637879 14.472836 14.666587 15.152156 5 b
So it appears iotools is both very fast and very consistent.