Increase performance by moving away from a for loop

Question

The gist of the argument is the following:

A function that I wrote, takes into consideration one argument, an alphanumeric string, and should output a string where the values of each element of this alphanumeric string are switched for some 'mapping'. MRE as follows:

#This is the original and switches value map
map = data.table(mapped = c(0:35), original = c(0:9,LETTERS))
#the function that I'm using:
as_numbers <- function(string) {
  #split string unlisted
  vector_unlisted <- unlist(strsplit(string,""))
  #match the string in vector
  for (i in 1:length(vector_unlisted)) {

    vector_unlisted[i] <- subset(map, map$original==vector_unlisted[i])[[1]][1]

  }
  vector_unlisted <- paste0(vector_unlisted, collapse = "")

  return(vector_unlisted)
}

I am trying to move away from the for loop for something that increases performance, as the function works, but it is pretty slow for the amount of elements I have supplied in this form:

unlist(lapply(dat$alphanum, function(x) as_numbers(x)))

An example of the input strings would be:549300JV8KEETQJYUG13. This should result in a string like 5493001931820141429261934301613

Supplying just one string in this case:

> as_numbers("549300JV8KEETQJYUG13")
[1] "5493001931820141429261934301613"

Relevant post: [Convert binary string to binary or decimal value](http://stackoverflow.com/questions/12892348) — zx8754, Feb 23 '16 at 19:37

zx8754 · Answer 1 · 2016-02-24T08:35:52.800

We can use base conversion:

#input and expected output
x <- "549300JV8KEETQJYUG13"
# "5493001931820141429261934301613"

#output
res <- paste0(strtoi(unlist(strsplit(x, "")), base = 36), collapse = "")

#test output
as_numbers(x) == res
# [1] TRUE

Performance

Since this post is about performance, here is benchmarking* for 3 solutions:

#input set up
map = data.table(mapped = c(0:35), original = c(0:9,LETTERS))
x <- rep(c("549300JV8KEETQJYUG13", "5493V8KE300J"), 1000)

#define functions
base_f <- function(string) {
  sapply(string, function(x) {
    paste0(strtoi(unlist(strsplit(x, "")), base = 36), collapse = "")
    })
  }

match_f <- function(string) {
  mapped <- map$mapped
  original <- map$original
  sapply(strsplit(string, ""), function(y) {
    paste0(mapped[match(y, original)], collapse= "")})
  }

reduce_f <- function(string) {
  Reduce(function(string,r) 
    gsub(map$original[r],
         map$mapped[r], string, fixed = TRUE),
    seq_len(nrow(map)), string)
  }

#test if all return same output
all(base_f(x) == match_f(x))
# [1] TRUE
all(base_f(x) == reduce_f(x))
# [1] TRUE

library(rbenchmark)
benchmark(replications = 1000,
          base_f(x),
          match_f(x),
          reduce_f(x))
#          test replications elapsed relative user.self sys.self user.child sys.child
# 1   base_f(x)         1000   22.15    4.683     22.12        0         NA        NA
# 2  match_f(x)         1000   19.18    4.055     19.11        0         NA        NA
# 3 reduce_f(x)         1000    4.73    1.000      4.72        0         NA        NA

_{*Note: microbenchmark() keeps throwing warnings, hence used rbenchmark() instead. Feel free to test with other libraries and update this post.}

score 6 · Accepted Answer · answered Feb 23 '16 at 14:16

Using Reduce and gsub, you could define the following function

replacer <- function(x) Reduce(function(x,r) gsub(map$original[r],
             map$mapped[r], x, fixed=T), seq_len(nrow(map)),x)


# Let's test it
replacer("549300JV8KEETQJYUG13")
#[1] "5493001931820141429261934301613"

MichaelChirico · Answer 3 · 2016-02-23T14:19:16.030

4

Seems like a merge:

map[as.data.table(unlist(strsplit(string, ""))),
    .(mapped), on = c(original = "V1")][ , paste0(mapped, collapse = "")]

Note that both "D1" and "1V" will be mapped to "131"...

On your example output is: "5493001931820141429261934301613"

You can use sep = "." if you actually want this to be a reversible mapping...

edited Feb 23 '16 at 14:19

answered Feb 23 '16 at 14:10

MichaelChirico

33,841
14
113
198

Yeah, collapsing is a must. It's part of a standard in this case – erasmortg Feb 23 '16 at 14:14
@erasmortg why not use "00" through "35"? – MichaelChirico Feb 23 '16 at 14:14
it's part of the ISO 7064 standard. It starts on 0 = 0 until z = 36 – erasmortg Feb 23 '16 at 14:16
@erasmortg how about a separator... e.g. `"D1"` -> `"13.1"`, `"1V"` -> `"1.31"`... – MichaelChirico Feb 23 '16 at 14:17
I would have to get rid of the separator at one point, the output string should be an integer – erasmortg Feb 23 '16 at 14:20

talat · Answer 4 · 2016-02-23T14:59:21.407

I would use match:

as_numbers <- function(string) {
  lapply(strsplit(string, ""), function(y) {
    paste0(map$mapped[match(y, map$original)], collapse= "")})
}

as_numbers(c("549300JV8KEETQJYUG13", "5493V8KE300J"))
#[[1]]
#[1] "5493001931820141429261934301613"
#
#[[2]]
#[1] "5493318201430019"

Added an lapply call to handle length > 1 input correctly.

If you need further speed up, you can store map$mapped and map$original in separate vectors and use them in the match call instead of map$... so you don't need to subset the data.frame/data.table so many times (which is quite costly).

Since the Q was about performance, here's a benchmark of two of the solutions:

map = data.table(mapped = c(0:35), original = c(0:9,LETTERS))
x <- rep(c("549300JV8KEETQJYUG13", "5493V8KE300J"), 1000)

ascii_func <- function(string) {
  lapply(string, function(x) {
    x_ascii <- strtoi(charToRaw(x), 16)
    paste(ifelse(x_ascii >= 65 & x_ascii <= 90,
                  x_ascii - 55, x_ascii - 48),
                  collapse = "")
  })
}

match_func <- function(string) {
  mapped <- map$mapped
  original <- map$original
    lapply(strsplit(string, ""), function(y) {
      paste0(mapped[match(y, original)], collapse= "")})
}

library(microbenchmark)
microbenchmark(ascii_func(x), match_func(x), times = 25L)
#Unit: milliseconds
#          expr   min    lq  mean median     uq    max neval
# ascii_func(x) 83.47 92.55 96.91  96.82 103.06 112.07    25
# match_func(x) 24.30 24.74 26.86  26.11  28.67  31.55    25

identical(ascii_func(x), match_func(x))
#[1] TRUE

You might want to include @mtoto's solution as well in the benchmark. On my PC that was the fastest one. — Jaap, Feb 23 '16 at 15:06

Increase performance by moving away from a for loop

4 Answers4

Performance