map and lazy evaluation

[mplourde]

library(purrr)
packageVersion('purrr')
# [1] '0.2.1.9000'

This is an issue I encountered with newer releases of purrr when using map to wrap functions. map in prior versions of purrr didn't have this issue, but I'm not sure when the change occurred. Consider the following example where I create a list of functions and then wrap them with lapply:

funs <- list(
    function() message('fun1'),
    function() message('fun2'),
    function() message('fun3')
)

wrapper <- function(fun) {
    function() fun()
}

. <- funs %>%
    lapply(wrap_fun) %>% 
    invoke_map()
# fun1
# fun2
# fun3

When I perform this task with map:

. <- funs %>% 
    map(wrapper) %>%
    invoke_map()
# fun3
# fun3
# fun3

forceing the fun argument of wrapper fixes the issue:

forced_wrapper <- function(fun) {
    force(fun)
    function() fun()
}

. <- funs %>% 
    map(forced_wrapper) %>%
    invoke_map()
# fun1
# fun2
# fun3

Since map basically supplants lapply, it would be helpful if they behaved the same in this respect.

[zhiyuanshi]

First of all, thanks @hadley for writing such wonderful package. I second this issue and can provide another (simpler) example.

Very often I want to make a list of functions that depends on some varying parameters. For example, making a list of functions that print out 1..5.

> f <- map(1:5, function(i) function() print(i))

However, if I try to call the first one, it gives 5 instead of 1.

> f[[1]]()
[1] 5

This contradicts to my intuition with functional programming and lexical scoping, and for the unaware, can lead to serious bugs in code.

With a bit of searching, I found one workaround is to make another factory function and explicitly reference i in its body, such as:

> make_f <- function(i) { i; function() print(i) }
> f <- map(1:5, make_f)
> f[[1]]()
[1] 1

But I think we don't have the issue with lapply:

> f <- lapply(1:5, function(i) function() { print(i) })
> f[[1]]()
[1] 1

[hadley]

This is a consequence of R's delayed evaluation - however lapply() somehow works around it so we should too.

[jimhester]

For reference the R-devel thread where this was changed in lapply is https://stat.ethz.ch/pipermail/r-devel/2015-February/070686.html

[hadley]

Implemented in wch/r-source@14f904c