Overloading methods in R
-----Original Message----- From: r-devel-bounces@stat.math.ethz.ch [mailto:r-devel-bounces@stat.math.ethz.ch] On Behalf Of Ali - Sent: Thursday, April 21, 2005 1:28 PM To: hb@maths.lth.se Cc: r-devel@stat.math.ethz.ch Subject: RE: [Rd] Overloading methods in R Henrik, Thanks for a reply as the author of the package. I understand that you people are interested in discussing this COOP-FOOP fight, however, the original question of this thread is all forgotton. So, I declare the question again: -> How to overload methods in classes created by R.oo package? <- I even sent you this question by the email provided in the package, but still no answer!
[That one got stuck in my spam filter; I've sent you a reply just before
seing this one. I'll paste th?t answer here too, if someone else is interest
or can add to the topic.]
Unfortunately, there is no immediate solution to this in the R language;
neither in S3 nor S4. The problem as is basically that you can only have one
generic function and that all methods dispatch by this generic function are
required to have the same arguments. I've been trying to get a discussion on
this problem, because I see the problem of two people creating two
independent package containing generic functions of the same name but
different arguments - that won't work today (at least not well). [namespaces
help a bit, though]
However, in S3 you can create a "generic" generic function by not specifying
arguments but only '...' - this way any methods can take any arguments (and
you don't force your argument names onto other developer's). Example:
foo <- function(...) UseMethodS3("foo")
foo.ClassA <- function(object, ...) { <code> }
foo.ClassB <- function(x, y, ...) { <code> }
and so on. This will *not* solve your problem of have overloaded methods in
the same class [using COOP thinking]. That is not possible to do, what I
understand. The only way I can think of is to have an ad hoc method which in
turn checks the arguments and dispatch on them.
foo.ClassA <- function(...) {
args <- list(...);
# Investigate names(args) and lapply(args, FUN=class) for further
dispatching
# to "private" methods .foo_x.ClassA(...), .foo_x_y.ClassA(...).
}
(This is inline with what Gabor Grothendieck outlined.) However, I am not
sure if the further methods dispatching with NextMethod() will work. You
probably want to define generic functions for .foo_x() and .foo_x_y() so
they in turn can be overloaded by subclasses.
If you figure a good schema for the above, maybe it can be made automatic so
that one can have a
setOverloadMethodS3("foo", "ClassA", function(object, ...) { <code> })
setOverloadMethodS3("foo", "ClassA", function(x, y, ...) { <code> })
to do the above. That would require some inspection of the arguments, but
that is not hard using formals().
So, the above is just a sketch that might or might not work. I think you
best shot is indeed to use S3, because it is a bit more flexible; S4 is
probably too rigid for this purpose.
BTW, I think it would be nice if you can develope an easy way to define
wrappers for C++ and other similar language! [Just don't reinvent the
wheel.]
I am trying to wrap some C++ class into R, so, in the way that you described, I am trying to wrap some COOP 'things' by some FOOP 'things'. Although there are some packages in R that already wrapped some C++ classes into R, none of them kept the originality of the wrapped classes. That is, the original C++ classes are customised and dissolved to R functions. There is nothing wrong with doing this, but, I am wrapping a few hundered C++ classes automatically and apparently it is not possible to customise each of them. The good thing about R.oo is that it is a trouble-less interface between COOP and FOOP. In fact, I have already used the package successfully to generate S3 classes from C++ *automatically*. The only big issue that I have with it is that R.oo overwrites overloaded methods and, to keep the originality of the C++ class and bein nice to the end user, I don't want to change the name of the overloaded functions. Also there is this point that sometime in feature S3 classes will become completely obsolete and new packages shouldn't be based on it. So it would be nice to have a 'S4-R.oo'.
I certainly hope S3 will not be obsolete on day and if it is every planned I certainly would to see a real and sensible open discussion on this long before being done! [You already know my standpoint here; S3 and S4 complement, not fight, each other]
Finally, I remind the package author about the original question: - How to overload methods in classes created by R.oo package?
So answered aboved.
with one more question that I add it right now: - Why R.00 is not upgraded to S4?
Indeed, I know that Nathan Whitehouse worked on this, see http://maths.newcastle.edu.au/~rking/R/devel/03b/0584.html. I do not know the current status of it, but I think he wrote some "white papers" on the topic. I don't know where and how many they are. There might be something in the http://www.rho-project.org project. Best wishes Henrik
and I hope we have some 'answer'. -Ali
Hi. Some clarification on R.oo:
-----Original Message----- From: r-devel-bounces@stat.math.ethz.ch [mailto:r-devel-bounces@stat.math.ethz.ch] On Behalf Of
A.J. Rossini
Sent: Wednesday, April 20, 2005 6:21 PM To: Ali - Cc: r-devel@stat.math.ethz.ch Subject: Re: [Rd] Overloading methods in R R.oo tries to implement an old-fashioned OO system as
found in Java,
Python, C++, etc. R's S4 methods implement a nice modern system based on the generic function approach , dispatch on argument signatures, which is different.
I would call them different, rather than old and modern -
each kind has
its own use. To the details: It is the Object class that is related to Java & co; the
setMethodS3()
and setConstructorS3() methods are orthogonal and just
userfriendly wrappers to
creating functions manually. The main purpose of Object is provide *reference variables*,
with the
main purpose to save memory! This is done by utilizing environment, which is standard R code, no ugly hacks are used. The Object class defines operators "$" and "$<-" (and a few others) to access
variables
within the environment, which is unique to each instance of class Object. Indeed, there was a similar feature added in R
v1.9.0 (or was
it v2.0.0) to the environment variable; "$" and "$<-" wraps up get() and assign() methods for easy use. For an object Object, the environment is in a list structure, contrary to being an environment directly. The reason for this is that attr(), save() and load() on environments does (did?) not work as you would expect, cf. https://stat.ethz.ch/pipermail/r-devel/2002-October/025197.html. The the Object class defines some other methods to simplify
life, and
yes, to imitate Java in the sense that it is convenient to
inherit from
one single root Class. It does not allow multiple
inheritance (although
you can update the class attributes yourself if you wish too). To differ between OOP in Java and S4/Dylan, I prefer to refer to the former as class-object-oriented programming (COOP) and the latter as function-object-oriented programming (FOOP). Then, comparing
COOP style
with FOOP style is a bit like comparing peas to apples. I would say that choosing COOP or FOOP is a design issue that has to do what you are trying to implement and not a once-in-a-lifetime/I-want-to-belong-to-this-group-of-people
decision. For
what I am working on, I found that higher level
implementation, where it is
clear that a method "belongs" to a class, is easier using
COOP. Classes for
statistical and mathematical modelling, where functions does
not belong to
a specific object, is probably better i FOOP. So, please do
not rule out one
for the other!
The setMethodS3() is a wrapper to automatically test for generic and
default
functions and create generic functions when needed etc. So
setMethodS3("foo", "MyClass", function(object, ...) {
#something
})
replaces things like
if (exists("foo.MyClass", mode="function"))
warning/stop("Replacing foo.MyClass")
if (exists("foo", mode="function") && !"not a generic function")
try to rename foo() to foo.default(), but only if foo.default()
does not already exists.
... and so on until you can safely write
foo.MyClass <- function(object, ...) {
# something
}
setConstructorS3() is basically like the above, but it does
not create
a generic function nor a class specific method, but a
"plain" function
setConstructorS3("MyClass", function(args, ...) {
# Something
}
to get
MyClass <- function(args, ...) {
# Something
}
with check for naming conflicts etc.
Cheers
Henrik Bengtsson
(author of R.oo)
While the R documentation for S4 classes is quite useful
(spanning
the green book, the BioC developer help pages, V&R's book on programming, and some other papers), I've found that for a nice background, Paul Graham's ANSI Lisp book, and in particular the nicely written chapter on CLOS, provides a nice
introduction to the
thought process. With respect to the R.oo package, the author might be the best source for that. Another package which you might take a look at is the
proto package,
which provides prototype object-orientation similar to
that found in
XLispStat, and also might help with what you are trying to do. However, I suspect that learning about the S4 system will provide more benefit in the future. best, -tony On 4/20/05, Ali - <saveez@hotmail.com> wrote:
Sean, Thanks, but, I am actually talking about overloading
'methods' and not
'functions', or you may want to answer this question: How
to overload
methods in classes created by R.oo package?
On Apr 20, 2005, at 8:16 AM, Ali - wrote:
(1) It seems to me that, generally, in R it is not
possible to
overload functions. Is that right? (2) Assuming that the above is true, or partially true,
is there any
extra packages to handle overloading in R? (3) Assuming (1) is TRUE and (2) is FALSE, can anyone
provide some
advice on developing some function that understand what the
arguments are and then calls the right overloaded function?
It would be something like this:
overloadedFunction1 <- function(x) {};
overloadedFunction2 <- function(x, y) {};
theFunction <- function(...)
{
# How to identify ... and call the right
overloaded function?
}
Ali, You are probably interested in "methods". Functions can
have different
"methods" depending on what the arguments and their types
are. A first
place to look is:
-
and-methods Sean
______________________________________________ R-devel@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
-- best, -tony "Commit early,commit often, and commit in a repository
from which we
can easily roll-back your mistakes" (AJR, 4Jan05). A.J. Rossini blindglobe@gmail.com
______________________________________________ R-devel@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
______________________________________________ R-devel@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-devel