[R] Possible Improvement to sapply

[Martin Morgan]

On 03/13/2018 09:23 AM, Doran, Harold wrote:
> While working with sapply, the documentation states that the simplify argument will yield a vector, matrix etc "when possible". I was curious how the code actually defined "as possible" and see this within the function
> 
> if (!identical(simplify, FALSE) && length(answer))
> 
> This seems superfluous to me, in particular this part:
> 
> !identical(simplify, FALSE)
> 
> The preceding code could be reduced to
> 
> if (simplify && length(answer))
> 
> and it would not need to execute the call to identical in order to trigger the conditional execution, which is known from the user's simplify = TRUE or FALSE inputs. I *think* the extra call to identical is just unnecessary overhead in this instance.
> 
> Take for example, the following toy example code and benchmark results and a small modification to sapply:
> 
> myList <- list(a = rnorm(100), b = rnorm(100))
> 
> answer <- lapply(X = myList, FUN = length)
> simplify = TRUE
> 
> library(microbenchmark)
> 
> mySapply <- function (X, FUN, ..., simplify = TRUE, USE.NAMES = TRUE){
> 	FUN <- match.fun(FUN)
>      answer <- lapply(X = X, FUN = FUN, ...)
>      if (USE.NAMES && is.character(X) && is.null(names(answer)))
>          names(answer) <- X
>      if (simplify && length(answer))
>          simplify2array(answer, higher = (simplify == "array"))
>      else answer
> }
> 
> 
>> microbenchmark(sapply(myList, length), times = 10000L)
> Unit: microseconds
>                     expr    min     lq     mean median     uq    max neval
>   sapply(myList, length) 14.156 15.572 16.67603 15.926 16.634 650.46 10000
>> microbenchmark(mySapply(myList, length), times = 10000L)
> Unit: microseconds
>                       expr    min     lq     mean median     uq      max neval
>   mySapply(myList, length) 13.095 14.864 16.02964 15.218 15.573 1671.804 10000
> 
> My benchmark timings show a timing improvement with only that small change made and it is seemingly nominal. In my actual work, the sapply function is called millions of times and this additional overhead propagates to some overall additional computing time.
> 
> I have done some limited testing on various real data to verify that the objects produced under both variants of the sapply (base R and my modified) yield identical objects when simply is both TRUE or FALSE.
> 
> Perhaps someone else sees a counterexample where my proposed fix does not cause for sapply to behave as expected.
> 

Check out ?sapply for possible values of `simplify=` to see why your 
proposal is not adequate.

For your example, lengths() is an order of magnitude faster than 
sapply(., length). This is a example of the advantages of vectorization 
(single call to an R function implemented in C) versus iteration (`for` 
loops but also the *apply family calling an R function many times). 
vapply() might also be relevant.

Often performance improvements come from looking one layer up from where 
the problem occurs and re-thinking the algorithm. Why would one need to 
call sapply() millions of times, in a situation where this becomes 
rate-limiting? Can the algorithm be re-implemented to avoid this step?

Martin Morgan

> Harold
> 
> ______________________________________________
> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 


This email message may contain legally privileged and/or...{{dropped:2}}