I have been asked (by Roger; thank you for the good question, and I hope it's fine to answer to the public) : > with Pi a sparse matrix and x,y, and ones > compatible n-vectors ? when I do: >> c(t(x) %*% Pi %*% ones, t(ones) %*% Pi %*% y ) > [[1]] 1 x 1 Matrix of class "dgeMatrix" > [,1] [1,] > 0.1338527 > [[2]] 1 x 1 Matrix of class "dgeMatrix" [,1] [1,] > 0.7810341 > I get a list whereas if Pi is an ordinary matrix I get a > vector. Is this intentional? Well, no. But it has been "unavoidable" in the sense that it had not been possible to provide S4 methods for '...' in the "remote" past, when Matrix was created. Later ... also quite a few years ago, John Chambers had added that possibility, with still some limitation (all '...' must be of the same class), and also plans to remove some of the limitations, see ?dotsMethods in R. I honestly have forgotten the history of my trying to provide 'c' methods for our "Matrix" objects after the 'dotsMethods' possibility had emerged, but I know I tried and had not seen a way to succeed "satisfactorily", but maybe I now think I maybe should try again. I currently think this needs changes to R before it can be done satisfactorily, and this is the main reason why this is a public answer to R-devel at ..., but I'm happy if I'am wrong. The real challenge here is that I think that if it should "work", it should work so in all cases, e.g., also for c(NA, 3:2, Matrix(2:1), matrix(10:11)) and that's not so easy, e.g., the following class and method definitions do *not* achieve the desired result: ## "mMatrix" is already hidden in Matrix pkg: setClassUnion("mMatrix", members = c("matrix", "Matrix")) setClassUnion("numMatrixLike", members c("logical", "integer","numeric", "mMatrix")) c.Matrix <- function(...) unlist(lapply(list(...), as.vector)) ## NB: Must use signature '(x, ..., recursive = FALSE)' : setMethod("c", "Matrix", function(x, ..., recursive) c.Matrix(x, ...)) ## The above is not sufficient for ## c(NA, 3:2, <Matrix>, <matrix>) : setMethod("c", "numMatrixLike", function(x, ..., recursive) c.Matrix(x, ...)) ## but the above does not really help:> c(Diagonal(3), NA, Matrix(10:11)) ## works fine,[1] 1 0 0 0 1 0 0 0 1 NA 10 11> c(NA, Diagonal(3)) ## R's lowlevel c() already decided to use list():[[1]] [1] NA [[2]] [,1] [,2] [,3] [1,] 1 . . [2,] . 1 . [3,] . . 1>---------------------------------------------- BTW, I (and the package users) suffer from exactly the same problem with the "MPFR" (multi precision numbers) provided by my package Rmpfr:> require(Rmpfr) > c(mpfr(3,100), 1/mpfr(7, 80)) ## works fine2 'mpfr' numbers of precision 80 .. 100 bits [1] 3 0.14285714285714285714285708> c(pi, 1/mpfr(7, 80)) ## "fails" even worse than in 'Matrix' case[[1]] [1] 3.141593 [[2]] 'mpfr1' 0.14285714285714285714285708>Yes, it would be very nice if c(.) could be used to concatenate quite arbitrary R objects into one long atomic vector, but I don't see how to achieve this easily. The fact, that c() just builds a list of its arguments if it ("thinks" it) cannot dispatch to a method, is a good strategy, but I'd hope it should be possible to have c() try to do better (and hence work for this case, and without a noticable performance penalty. Suggestions are very welcome. Martin
(Brief reply, I'm traveling but as per below, this is on my radar right now so wanted to comment.) Two points regarding "dotsMethods". 1. To clarify the limitation. It's not that all the arguments have to be of the same class, but there must be one class that they belong to or subclass. (So, as in the example in the documentation, the method could be defined for a class union or other virtual class that all the actual arguments inherit from.) 2. The current documentation is a mess. In common with lots of other very old documentation. I'm in the process of rewriting a large chunk of the documentation including that for dotsMethods. Sometime in the next few weeks, I hope to have it coherent enough to commit. So far, I'm not trying to change any significant aspects of the code, including for "..." methods, which seem to do roughly what was intended. John On Sep 10, 2016, at 8:27 AM, Martin Maechler <maechler at stat.math.ethz.ch> wrote:> I have been asked (by Roger; thank you for the good question, > and I hope it's fine to answer to the public) : > >> with Pi a sparse matrix and x,y, and ones >> compatible n-vectors ? when I do: > >>> c(t(x) %*% Pi %*% ones, t(ones) %*% Pi %*% y ) >> [[1]] 1 x 1 Matrix of class "dgeMatrix" >> [,1] [1,] >> 0.1338527 >> [[2]] 1 x 1 Matrix of class "dgeMatrix" > [,1] [1,] >> 0.7810341 > >> I get a list whereas if Pi is an ordinary matrix I get a >> vector. Is this intentional? > > Well, no. But it has been "unavoidable" in the sense that it had not > been possible to provide S4 methods for '...' in the "remote" > past, when Matrix was created. > > Later ... also quite a few years ago, John Chambers had added > that possibility, with still some limitation (all '...' must be > of the same class), and also plans to remove some of the > limitations, see ?dotsMethods in R. > > I honestly have forgotten the history of my trying to provide 'c' > methods for our "Matrix" objects after the 'dotsMethods' > possibility had emerged, but I know I tried and had not seen a > way to succeed "satisfactorily", > but maybe I now think I maybe should try again. > I currently think this needs changes to R before it can be done > satisfactorily, and this is the main reason why this is a public > answer to R-devel at ..., but I'm happy if I'am wrong. > > The real challenge here is that I think that if it should "work", > it should work so in all cases, e.g., also for > > c(NA, 3:2, Matrix(2:1), matrix(10:11)) > > and that's not so easy, e.g., the following class and method > definitions do *not* achieve the desired result: > > ## "mMatrix" is already hidden in Matrix pkg: > setClassUnion("mMatrix", members = c("matrix", "Matrix")) > setClassUnion("numMatrixLike", members > c("logical", "integer","numeric", "mMatrix")) > > c.Matrix <- function(...) unlist(lapply(list(...), as.vector)) > ## NB: Must use signature '(x, ..., recursive = FALSE)' : > setMethod("c", "Matrix", function(x, ..., recursive) c.Matrix(x, > ...)) > ## The above is not sufficient for > ## c(NA, 3:2, <Matrix>, <matrix>) : > setMethod("c", "numMatrixLike", function(x, ..., recursive) > c.Matrix(x, ...)) > > ## but the above does not really help: > >> c(Diagonal(3), NA, Matrix(10:11)) ## works fine, > [1] 1 0 0 0 1 0 0 0 1 NA 10 11 > >> c(NA, Diagonal(3)) ## R's lowlevel c() already decided to use list(): > [[1]] > [1] NA > > [[2]] > [,1] [,2] [,3] > [1,] 1 . . > [2,] . 1 . > [3,] . . 1 > >> > ---------------------------------------------- > > BTW, I (and the package users) suffer from exactly the same > problem with the "MPFR" (multi precision numbers) provided by my > package Rmpfr: > >> require(Rmpfr) >> c(mpfr(3,100), 1/mpfr(7, 80)) ## works fine > 2 'mpfr' numbers of precision 80 .. 100 bits > [1] 3 0.14285714285714285714285708 > >> c(pi, 1/mpfr(7, 80)) ## "fails" even worse than in 'Matrix' case > [[1]] > [1] 3.141593 > > [[2]] > 'mpfr1' 0.14285714285714285714285708 > >> > > > Yes, it would be very nice if c(.) could be used to > concatenate quite arbitrary R objects into one long atomic > vector, but I don't see how to achieve this easily. > > The fact, that c() just builds a list of its arguments if it > ("thinks" it) cannot dispatch to a method, is a good strategy, > but I'd hope it should be possible to have c() try to do better > (and hence work for this case, and > without a noticable performance penalty. > > Suggestions are very welcome. > Martin > > ______________________________________________ > R-devel at r-project.org mailing list > stat.ethz.ch/mailman/listinfo/r-devel
One option would be to use the same strategy that we use for cbind() and rbind(), i.e., if dispatch fails, call a binary generic, c2(), recursively. Could do the same for pmin() and pmax(). Michael On Sat, Sep 10, 2016 at 8:27 AM, Martin Maechler <maechler at stat.math.ethz.ch> wrote:> I have been asked (by Roger; thank you for the good question, > and I hope it's fine to answer to the public) : > > > with Pi a sparse matrix and x,y, and ones > > compatible n-vectors ? when I do: > > >> c(t(x) %*% Pi %*% ones, t(ones) %*% Pi %*% y ) > > [[1]] 1 x 1 Matrix of class "dgeMatrix" > > [,1] [1,] > > 0.1338527 > > [[2]] 1 x 1 Matrix of class "dgeMatrix" > [,1] [1,] > > 0.7810341 > > > I get a list whereas if Pi is an ordinary matrix I get a > > vector. Is this intentional? > > Well, no. But it has been "unavoidable" in the sense that it had not > been possible to provide S4 methods for '...' in the "remote" > past, when Matrix was created. > > Later ... also quite a few years ago, John Chambers had added > that possibility, with still some limitation (all '...' must be > of the same class), and also plans to remove some of the > limitations, see ?dotsMethods in R. > > I honestly have forgotten the history of my trying to provide 'c' > methods for our "Matrix" objects after the 'dotsMethods' > possibility had emerged, but I know I tried and had not seen a > way to succeed "satisfactorily", > but maybe I now think I maybe should try again. > I currently think this needs changes to R before it can be done > satisfactorily, and this is the main reason why this is a public > answer to R-devel at ..., but I'm happy if I'am wrong. > > The real challenge here is that I think that if it should "work", > it should work so in all cases, e.g., also for > > c(NA, 3:2, Matrix(2:1), matrix(10:11)) > > and that's not so easy, e.g., the following class and method > definitions do *not* achieve the desired result: > > ## "mMatrix" is already hidden in Matrix pkg: > setClassUnion("mMatrix", members = c("matrix", "Matrix")) > setClassUnion("numMatrixLike", members > c("logical", "integer","numeric", "mMatrix")) > > c.Matrix <- function(...) unlist(lapply(list(...), as.vector)) > ## NB: Must use signature '(x, ..., recursive = FALSE)' : > setMethod("c", "Matrix", function(x, ..., recursive) c.Matrix(x, > ...)) > ## The above is not sufficient for > ## c(NA, 3:2, <Matrix>, <matrix>) : > setMethod("c", "numMatrixLike", function(x, ..., recursive) > c.Matrix(x, ...)) > > ## but the above does not really help: > >> c(Diagonal(3), NA, Matrix(10:11)) ## works fine, > [1] 1 0 0 0 1 0 0 0 1 NA 10 11 > >> c(NA, Diagonal(3)) ## R's lowlevel c() already decided to use list(): > [[1]] > [1] NA > > [[2]] > [,1] [,2] [,3] > [1,] 1 . . > [2,] . 1 . > [3,] . . 1 > >> > ---------------------------------------------- > > BTW, I (and the package users) suffer from exactly the same > problem with the "MPFR" (multi precision numbers) provided by my > package Rmpfr: > >> require(Rmpfr) >> c(mpfr(3,100), 1/mpfr(7, 80)) ## works fine > 2 'mpfr' numbers of precision 80 .. 100 bits > [1] 3 0.14285714285714285714285708 > >> c(pi, 1/mpfr(7, 80)) ## "fails" even worse than in 'Matrix' case > [[1]] > [1] 3.141593 > > [[2]] > 'mpfr1' 0.14285714285714285714285708 > >> > > > Yes, it would be very nice if c(.) could be used to > concatenate quite arbitrary R objects into one long atomic > vector, but I don't see how to achieve this easily. > > The fact, that c() just builds a list of its arguments if it > ("thinks" it) cannot dispatch to a method, is a good strategy, > but I'd hope it should be possible to have c() try to do better > (and hence work for this case, and > without a noticable performance penalty. > > Suggestions are very welcome. > Martin > > ______________________________________________ > R-devel at r-project.org mailing list > stat.ethz.ch/mailman/listinfo/r-devel
>>>>> John Chambers <jmc at r-project.org> >>>>> on Sat, 10 Sep 2016 09:16:38 -0700 writes:> (Brief reply, I'm traveling but as per below, this is on my radar right now so wanted to comment.) > Two points regarding "dotsMethods". > 1. To clarify the limitation. It's not that all the arguments have to be of the same class, but there must be one class that they belong to or subclass. (So, as in the example in the documentation, the method could be defined for a class union or other virtual class that all the actual arguments inherit from.) Thank you for the clarification. I knew that the limitation "the same class" has not been a big one, that's why I did use a class union in my example (below). I thought there were other limitations.. never mind. > 2. The current documentation is a mess. In common with lots of other very old documentation. I'm in the process of rewriting a large chunk of the documentation including that for dotsMethods. Sometime in the next few weeks, I hope to have it coherent enough to commit. That's great! > So far, I'm not trying to change any significant aspects of the code, including for "..." methods, which seem to do roughly what was intended. Yes, I'm sorry if I sounded like saying something different. That I think this [getting c() to work for a collection objects, some S4] needs changes in R is because it seems that do_c() fails to dispatch here, and hence the problem was with our C function that has carried the comment | * To call this an ugly hack would be to insult all existing ugly hacks | * at large in the world. and I don't think I would be able to correctly patch that infamous function (in src/main/eval.c) ... Martin > John > On Sep 10, 2016, at 8:27 AM, Martin Maechler <maechler at stat.math.ethz.ch> wrote: >> I have been asked (by Roger; thank you for the good question, >> and I hope it's fine to answer to the public) : >> >>> with Pi a sparse matrix and x,y, and ones >>> compatible n-vectors ? when I do: >> >>>> c(t(x) %*% Pi %*% ones, t(ones) %*% Pi %*% y ) >>> [[1]] 1 x 1 Matrix of class "dgeMatrix" >>> [,1] [1,] >>> 0.1338527 >>> [[2]] 1 x 1 Matrix of class "dgeMatrix" >> [,1] [1,] >>> 0.7810341 >> >>> I get a list whereas if Pi is an ordinary matrix I get a >>> vector. Is this intentional? >> >> Well, no. But it has been "unavoidable" in the sense that it had not >> been possible to provide S4 methods for '...' in the "remote" >> past, when Matrix was created. >> >> Later ... also quite a few years ago, John Chambers had added >> that possibility, with still some limitation (all '...' must be >> of the same class), and also plans to remove some of the >> limitations, see ?dotsMethods in R. >> >> I honestly have forgotten the history of my trying to provide 'c' >> methods for our "Matrix" objects after the 'dotsMethods' >> possibility had emerged, but I know I tried and had not seen a >> way to succeed "satisfactorily", >> but maybe I now think I maybe should try again. >> I currently think this needs changes to R before it can be done >> satisfactorily, and this is the main reason why this is a public >> answer to R-devel at ..., but I'm happy if I'am wrong. >> >> The real challenge here is that I think that if it should "work", >> it should work so in all cases, e.g., also for >> >> c(NA, 3:2, Matrix(2:1), matrix(10:11)) >> >> and that's not so easy, e.g., the following class and method >> definitions do *not* achieve the desired result: >> >> ## "mMatrix" is already hidden in Matrix pkg: >> setClassUnion("mMatrix", members = c("matrix", "Matrix")) >> setClassUnion("numMatrixLike", members >> c("logical", "integer","numeric", "mMatrix")) >> >> c.Matrix <- function(...) unlist(lapply(list(...), as.vector)) >> ## NB: Must use signature '(x, ..., recursive = FALSE)' : >> setMethod("c", "Matrix", function(x, ..., recursive) c.Matrix(x, >> ...)) >> ## The above is not sufficient for >> ## c(NA, 3:2, <Matrix>, <matrix>) : >> setMethod("c", "numMatrixLike", function(x, ..., recursive) >> c.Matrix(x, ...)) >> >> ## but the above does not really help: >> >>> c(Diagonal(3), NA, Matrix(10:11)) ## works fine, >> [1] 1 0 0 0 1 0 0 0 1 NA 10 11 >> >>> c(NA, Diagonal(3)) ## R's lowlevel c() already decided to use list(): >> [[1]] >> [1] NA >> >> [[2]] >> [,1] [,2] [,3] >> [1,] 1 . . >> [2,] . 1 . >> [3,] . . 1 >> >>> >> ---------------------------------------------- >> >> BTW, I (and the package users) suffer from exactly the same >> problem with the "MPFR" (multi precision numbers) provided by my >> package Rmpfr: >> >>> require(Rmpfr) >>> c(mpfr(3,100), 1/mpfr(7, 80)) ## works fine >> 2 'mpfr' numbers of precision 80 .. 100 bits >> [1] 3 0.14285714285714285714285708 >> >>> c(pi, 1/mpfr(7, 80)) ## "fails" even worse than in 'Matrix' case >> [[1]] >> [1] 3.141593 >> >> [[2]] >> 'mpfr1' 0.14285714285714285714285708 >> >>> >> >> >> Yes, it would be very nice if c(.) could be used to >> concatenate quite arbitrary R objects into one long atomic >> vector, but I don't see how to achieve this easily. >> >> The fact, that c() just builds a list of its arguments if it >> ("thinks" it) cannot dispatch to a method, is a good strategy, >> but I'd hope it should be possible to have c() try to do better >> (and hence work for this case, and >> without a noticable performance penalty. >> >> Suggestions are very welcome. >> Martin >> >> ______________________________________________ >> R-devel at r-project.org mailing list >> stat.ethz.ch/mailman/listinfo/r-devel
Apparently Analagous Threads
- c(<Matrix>, <Matrix>) / help(dotsMethods) etc
- c(<Matrix>, <Matrix>) / help(dotsMethods) etc
- Problems with S4 methods dispatching on `...` (aka dotsMethods)
- Problems with S4 methods dispatching on `...` (aka dotsMethods)
- Problems with S4 methods dispatching on `...` (aka dotsMethods)