Thanks very much for this package!
I am running roommate(utils=utils) 1000s of times with changing utils but small population size (N < 1000). So, any overhead adds up.
For small N, the overhead is actually substantial, and I traced it to roommate.checkPreferences. I wonder if this check could be skipped using a flag to roommate. In fact, when pref is generated from utils by sortIndexOneSided(as.matrix(utils)), I wonder if checkPreferences can always be skipped? Since it seems sortIndexOneSided returns prefs as needed in cpp.
I thought the bottleneck in checkPreferences was the loop to create comp, but it seems not (I guess it is the sorting). Still, I think comp can be generated more easily by:
comp <- array(seq_len(NROW(pref)), dim = dim(pref)) - upper.tri(pref)
MWE & benchmarking
library('matchingR')
library('microbenchmark')
irving = utils::getFromNamespace(ns='matchingR','cpp_wrapper_irving')
validate = utils::getFromNamespace(ns='matchingR','roommate.validate')
checkPref = utils::getFromNamespace(ns='matchingR','roommate.checkPreferences')
gen.utils = function(n,seed=NULL){
set.seed(seed)
x = rnorm(n)
u = dnorm(outer(x,x,`-`))
}
my.validate = function(u){
n = ncol(u)
sortIndexOneSided(matrix(u[-seq(1,n^2,n+1)],n-1,n))
}
my.checkpref <- function(pref) {
# check if pref is using R instead of C++ indexing
if (all(apply(rbind(pref, seq_len(NCOL(pref))), 2, sort) ==
matrix(seq_len(NCOL(pref)), nrow = NCOL(pref), ncol = NCOL(pref)))) {
return(pref - 1)
}
comp <- array(seq_len(NROW(pref)), dim = dim(pref)) - upper.tri(pref)
# check if pref has a complete listing otherwise given an error
if (all(apply(pref, 2, sort) == comp)) {
return(pref)
}
return(NULL)
}
microbenchmark(
# validate(u),my.validate(u),my.checkpref(my.validate(u)), # VALIDATION
roommate(u=u),irving(validate(u))+1,irving(my.validate(u))+1, # FULL MODEL
setup={u=gen.utils(100,seed=0)},
check='equal',
times=10)output
Loading required package: Rcpp
Unit: microseconds
expr min lq mean median uq max neval
roommate(u = u) 5937.956 6142.166 6575.959 6349.9525 6501.501 8439.542 10
irving(validate(u)) + 1 6113.019 6245.212 6938.835 6516.9050 7663.952 8583.290 10
irving(my.validate(u)) + 1 855.716 866.679 1161.646 884.0885 958.542 3303.109 10
Thanks very much for this package!
I am running
roommate(utils=utils)1000s of times with changingutilsbut small population size (N < 1000). So, any overhead adds up.For small N, the overhead is actually substantial, and I traced it to
roommate.checkPreferences. I wonder if this check could be skipped using a flag toroommate. In fact, whenprefis generated fromutilsbysortIndexOneSided(as.matrix(utils)), I wonder ifcheckPreferencescan always be skipped? Since it seemssortIndexOneSidedreturnsprefsas needed in cpp.I thought the bottleneck in
checkPreferenceswas the loop to createcomp, but it seems not (I guess it is the sorting). Still, I thinkcompcan be generated more easily by:comp <- array(seq_len(NROW(pref)), dim = dim(pref)) - upper.tri(pref)MWE & benchmarking
output