ichunk

[dselivanov]

I eveluated both itertools::ichunk and itertools2::ichunk, and the first (at least for me) looks more logical while trying to handle objects which length not divisible by chunk_size. For example I have vector v=1:5 and want to obtain list(c(1,2), c(3,4), c(5)), is it possible with itertools2::ichunk?

[ramhiser]

itertools2::ichunk has a similar behavior, but could be improved. Here's a similar example to the one you provided:

> it <- itertools2::ichunk(1:5, chunk_size=2)
> nextElem(it)
[[1]]
[1] 1

[[2]]
[1] 2

> nextElem(it)
[[1]]
[1] 3

[[2]]
[1] 4

> nextElem(it)
[[1]]
[1] 5

[[2]]
[1] NA

[dselivanov]

I understand how to use itertools2::ichunk, but I think last chunk with value c(5) will be much more logical. I'm ok with fill=NA option, but fill=NULL doesn't work.

Also there 2 notes:

1. How to stop iterate over ichunk iterator? after the end it silently fills all values with NA and continue iteration:

it <- ichunk(1:10, 3)
while(TRUE) {
  val = try(nextElem(it), silent = T)
  if (class(val) == "try-error") break
}

Of course I can check it with anyNA() call, NA can be in underlying data... Believe, user should not care about such sort of things.
2. this note is about performance. I want to heavily use iterators in my text2vec package, but there is big issue with performance. Consider following examples:

devtools::install_github('dselivanov/text2vec@redesign')
# reviews of 25000 movies ~ 30mb
data("movie_review")
it <- ichunk(movie_review[['review']], chunk_size = 1000)
limit <- length(movie_review[['review']]) / 1000
i <- 1
system.time(
  # i <= limit because of behavior from #1
  while(TRUE && i <= limit) {
    val = try(nextElem(it), silent = T)
    if (class(val) == "try-error") break
    i <- i + 1
  }
)
#user  system elapsed 
#1.141   0.002   1.142 

#Now lets check without iterators

chunk <- function(x, n) suppressWarnings(split(x, rep(seq_len(n), each = ceiling(length(x) / n)) ))
ich <- chunk(seq_len(nrow(movie_review)), 25)
system.time(
  for(i in ich ) {
    tmp<-movie_review[['review']][i]
  }
)
#user  system elapsed 
#0.002   0.000   0.003

As you can see, there is very very big overhead of using iterators here. Can we do something with that?

[ramhiser]

I'll take a look at the ichunk behavior this weekend and see what needs to improve.

As for the iterators package, I'm aware of its unfortunate overhead -- I have seen it crop up in several cases. Contributing to or forking iterators has been on my agenda for some time.

[dselivanov]

iterators::idiv() solved my problems,. I write custom iterator based on
top of idiv and now everything works fine.

2015-11-05 16:52 GMT+03:00 John Ramey [email protected]:

I'll take a look at the ichunk behavior this weekend and see what needs
to improve.

As for the iterators package, I'm aware of its unfortunate overhead -- I
have seen it crop up in several cases. Contributing to or forking
iterators has been on my agenda for some time.

—
Reply to this email directly or view it on GitHub
#46 (comment).

Regards
Dmitriy Selivanov

[ramhiser]

That is good to hear.

After you created the issue, I realized the ichunk is not ideal. I still intend to investigate it this weekend.