CODEWITHSUNDEEP
rperformance
mcsim
mcsim

Reputation: 1798

Fast apply on a matrix

I have a question on the way how to speedup certain things in R. The code example is following:

n=1000
A=matrix(rnorm(n), 1, n)
T=diag(n)
Rprof()
for (i in 1:100)
    A = apply(T, 1, function(row) max(A*row)) + matrix(rnorm(n), 1,n)
Rprof(NULL)
summaryRprof()

Profile is following:

> $by.self
                self.time self.pct total.time total.pct
"apply"              2.16    37.50       5.74     99.65
"*"                  2.06    35.76       2.06     35.76
"aperm.default"      1.08    18.75       1.08     18.75
"max"                0.42     7.29       0.42      7.29
"FUN"                0.02     0.35       2.50     43.40
"rnorm"              0.02     0.35       0.02      0.35

As you can see the slowest thing is function apply. Could you suggest a way to get rid of this function to speedup the whole computation?

Upvotes: 0

Views: 580

Answers (1)

jlhoward
jlhoward

Reputation: 59385

Generally you're better off with lists than matrices. In your specific case:

set.seed(1)   # for reproducibility
n=1000
A=matrix(rnorm(n), 1, n)
T=diag(n)
set.seed(1)   # for reproducibility
system.time({
  for (i in 1:100)
    X = apply(T, 1, function(row) max(A*row)) + matrix(rnorm(n), 1,n)  
})
#    user  system elapsed 
#    5.83    0.13    5.98 


df <- as.data.frame(t(T))
set.seed(1)   # for reproducibility
system.time({
  for (i in 1:100)
    Y = sapply(df, function(row) max(A*row)) + matrix(rnorm(n), 1,n)
})
#    user  system elapsed 
#    0.97    0.00    0.96 

identical(X,Y)
# [1] TRUE

Upvotes: 2

Related Questions