HenrikBengtsson
commented
1 year ago Thanks again.
First observation
There's nothing that changed in the behavior of useNames = TRUE since matrixStats 0.63.0;
library(matrixStats)
stopifnot(packageVersion("matrixStats") == "0.63.0")
X <- matrix(1:5, ncol = 1L)
Xs <- list(
none = X,
null = structure(X, dimnames = list(NULL, NULL))
)
ys <- lapply(Xs, FUN = colCumsums, useNames = TRUE)
str(ys)gives
List of 2
$ none: int [1:5, 1] 1 3 6 10 15
$ null: int [1:5, 1] 1 3 6 10 15
..- attr(*, "dimnames")=List of 2
.. ..$ : NULL
.. ..$ : NULLWhat changed is that we're moving from useNames = NA to useNames = TRUE, by default (#232).
Second observation
Base R treats empty dimnames the same as no dimnames, e.g.
> ys <- lapply(Xs, FUN = colSums)
> str(ys)
List of 2
$ none: num 15
$ null: num 15
ys <- lapply(Xs, FUN = colMeans)
> str(ys)
List of 2
$ none: num 3
$ null: num 3
ys <- lapply(Xs, FUN = apply, MARGIN = 2L, sum)
> str(ys)
List of 2
$ none: int 15
$ null: int 15
ys <- lapply(Xs, FUN = apply, MARGIN = 2L, mean)
> str(ys)
List of 2
$ none: num 3
$ null: num 3Thus, this is how matrixStats should also do it.
Third observation
matrixStats behaves like base R, except for the cumulative functions, e.g. all of the following returns unnamed results:
ys <- lapply(Xs, FUN = colMeans2, useNames = TRUE)
ys <- lapply(Xs, FUN = colMedians, useNames = TRUE)
ys <- lapply(Xs, FUN = colMads, useNames = TRUE)
ys <- lapply(Xs, FUN = colVars, useNames = TRUE)
ys <- lapply(Xs, FUN = colIQRs, useNames = TRUE)
ys <- lapply(Xs, FUN = colLogSumExps, useNames = TRUE)
ys <- lapply(Xs, FUN = colWeightedMeans, useNames = TRUE)
ys <- lapply(Xs, FUN = colWeightedMedians, useNames = TRUE)However, the following, cumulative functions treats the two cases differently:
ys <- lapply(Xs, FUN = colCumsums, useNames = TRUE)
ys <- lapply(Xs, FUN = colCumprods, useNames = TRUE)
ys <- lapply(Xs, FUN = colCummins, useNames = TRUE)
ys <- lapply(Xs, FUN = colCummaxs, useNames = TRUE)Fourth observation
BTW, setting dimnames to character(0L) is the same as setting them to NULL in R, e.g.
> str(X)
int [1:5, 1] 1 2 3 4 5
> rownames(X) <- character(0)
> str(X)
int [1:5, 1] 1 2 3 4 5
- attr(*, "dimnames")=List of 2
..$ : NULL
..$ : NULL
> rownames(X) <- NULL
> str(X)
int [1:5, 1] 1 2 3 4 5
- attr(*, "dimnames")=List of 2
..$ : NULL
..$ : NULLIn other words, we don't have to treat character(0) specially; it's taken care of by R itself.
Action
- [ ] Fix all cumulative functions to treat NULL dimnames the same as non-existing dimnames.
yaccos
const-ae
Continuing the discussion from #232.
With the new implementation of
useNames = TRUE,matrixStats::colCumsumsand other functions that return a matrix, treat a matrix without a dimnames attribute anddimnames = list(NULL, NULL)differently (v0.63.0-9008).Created on 2023-05-30 with reprex v2.0.2
Previously, they were treated the same (v0.63.0):
Created on 2023-05-30 with reprex v2.0.2
From the sparse matrix point-of-view, the old behavior was preferable because dgCMatrix treats matrices without a dimnames attribute and
dimnames = list(NULL, NULL)the same.