uniqueN could supports list input

[jangorecki]

And not just lists but any type which is not supported by uniqueN could simply redirects to length(unique(.)).

library(data.table)
d1 <- data.table(a = 1:4, l = list(list(letters[1:2]),list(Sys.time()),list(1:10),list(letters[1:2])))
d1[,length(unique(l))]
# [1] 3
d1[,uniqueN(l)]
#Error in uniqueN(l) : 
#  x must be an atomic vector or data.frames/data.tables

Looks pretty easy to do, I will try to make PR for that.

[arunsrinivasan]

@jangorecki this seems to do something different.

dt = data.table(x=c(1,1,2,3), y=c(2,2,3,4))
nrow(unique(dt)) # [1] 3

length(unique(as.list(dt))) # [1] 2

It seems to not check row-wise, rather if any element of list is identical to another.

[jangorecki]

I used uniqueN / length(unique( on single column while you did it on both columns. To make it more clear I simplified the example data:

library(data.table)
d1 <- data.table(l = list(list(letters[1:2]),list(Sys.time()),list(1:10),list(letters[1:2])))
d1[,length(unique(l))]
d1[,uniqueN(l)]

[arunsrinivasan]

Okay. I guess the idea here is that a list is a vector as well.. that's fine.
Could you make sure it'd work on multiple columns as well? For example:

d1 <- data.table(a = c(1:3,1), l = list(list(letters[1:2]),list(Sys.time()),list(1:10),list(letters[1:2])))
nrow(unique(as.data.frame(d1))) # [1] 3

We'd want that to work, if someone did:

d1[, uniqueN(.SD)] 

This wouldn't work at the moment. Adding a check as to whether any column is of list type might be quite expensive when one performs a grouping operation (checking type for each group). And this is quite a rare case.. Hm.

[arunsrinivasan]

Maybe for single column case we can already use the fix you've provided, and take care of .SD / data.table cases later. Will merge asap. Thanks @jangorecki.

[jangorecki]

This should not affect .SD as the check is if (!is.atomic(x) && !is.data.frame(x)) return(length(unique(x))).

data.table(a = list(1:2,4))[,is.data.frame(.SD)]
# [1] TRUE
data.table(a = list(1:2,4))[,uniqueN(a)]
# [1] 2
data.table(a = list(1:2,4))[,uniqueN(.SD)]
#Error in forderv(x, by = by, retGrp = TRUE) : 
#  First column being ordered is type 'list', not yet supported
data.table(aa = 1:2, a = list(1:2,4))[,uniqueN(.SD)]
# [1] 2

It still can throw error in a specific case of list column as first column, but this is related to #1229.
I will update doc, readme, and squash soon.

[arunsrinivasan]

This fix returns wrong result for list(1:2, 3:4). length(unique()) should only be dispatched for list of lists.. not list of atomics.

[jangorecki]

@arunsrinivasan I'm not sure if I follow. You are saying that list should be handled same as data.table? I don't think it is a proper behavior for list, while it is for data.table.
I can create uniqueN generic so it can be flexible.

[arunsrinivasan]

Nope. I'm comfortable with uniqueN not supporting list-of-lists. It was not the original purpose. But it needs to do what forder() does for list of atomics (using mget() in j argument would not work as intended otherwise).

[jangorecki]

I will quote from source code, it is quite on point.

# simple straightforward helper function to get the number 
# of groups in a vector or data.table. Here by data.table, 
# we really mean `.SD` - used in a grouping operation