[Rd] Problem with _new_ if class "lm" in object representatio n.
John Chambers
jmc at research.bell-labs.com
Fri Oct 1 22:29:51 CEST 2004
Eric Lecoutre wrote:
>
> Hi,
>
> For your lm problem, you could begin to create a virtual class that accepts
> either NULL or a lm object:
>
> > setClassUnion("mylm", c("NULL","lm"))
> > setClass("foo", representation(x="numeric", y="mylm"))
> > new("foo",x=1,y=NULL)
>
> Then, you have to check the content of slot y in any function that will
> manipulate objects of class "foo":
>
> setMethod("show","foo",function(object)
> {
> cat("\n object of class foo")
> if (!is.null(object at y)) cat(" does have a lm object in slot y!")
> })
This is OK, but it illustrates why it would _not_ be a good idea to
allow NULL values automatically for an arbitrary class. The
implementation here says that the "y" slot can be either a NULL or an
"lm" object. That is a different statement from saying that "y" must be
an "lm" object.
Either approach could make sense. But what one really wants (in any
definiton of a new class) is to match the actual definition to the
intended meaning.
With your suggestion, the designer of class "foo" is basically passing
on to the implenters of methods the responsibility to check for a NULL
"y" slot. If it makes sense for legitimate "foo" objects to exist that
have a NULL "y" slot, then that extra work is part of the design of the
"foo" class. Your definition of the class essentially says that the
"lm" model is an optional part of the object.
But if in fact it _only_ makes sense for the object to have a true
model, then one should not define a class that conflicts with the
intended meaning. There was no indication in the initial question that
NULL made sense. If we're just introducing the NULL possibility as a
programming hack, that's not really a good idea.
The other approach is to insist that the computations creating a "foo"
object generate a valid "lm" object, if the user didn't supply one.
Then methods for "foo" objects are not required to check. One way to
make the stricter approach work would be to have a method for
initialize() for class "foo" that generated a model from the other slots
in the object.
For example, the class might represent "x" and "y" data, and a model
that by default was the linear model for y~x. Something like the
following would implement that idea:
setClass("withLm", representation(x="numeric", y="numeric", model =
"lm"))
setMethod("initialize", "withLm",
function(.Object, x = numeric(), y = numeric(), model = NULL)
{
if(is.null(model))
model <- lm(y ~ x)
callNextMethod(.Object, x=x, y=y, model=model)
}
)
(unfortunately, lm() does not allow fitting vectors of length 0, so if
this were a real example there are a few boundary cases to consider ;-)
>
> Eric
>
> At 14:35 30/09/2004, Witold Eryk Wolski wrote:
> >Hi,
> >
> >But "any" rises some other problems well known from S3. One has "any" for
> >"free" in S3. You don't need S4. . But I know how "if" polluted
> >functions look like in S3. They are hard to understand and to maintain.
> >Hence I am quite happy to use S4.
> >Type-checking is usefull if you program with data. Also if you are on the
> >"C side" of the object. A further problem which I have with "any" is, that
> >lm="any" in setClass tells you exactly ANYthing (and you are back in S3).
> >And thats not what I like.*
> >Hence, I do not want to use any in this case.
> >
> >I think that allowing to assing NULL to a slot will solve some problems
> >which I have with S4 so far.
> >
> >a) How to express that an a slot content was not assigned so far? J.C.
> >suggest to use character(0), lm(1~1) etc. The problem I have with
> >insterting lm(1~1) in slot mod in my class Ctest are that I have 100000
> >instances of object Ctest. At initialization everyone contains an object
> >lm(1~1). But please note the following.
> >
> > > x<-lm(1~1)
> > > object.size(x)
> >[1] 6388
> > > object.size(NULL)
> >[1] 0
> >
> >Unfortunately S4 do not have references. Hence, I can not store a
> >reference to a single instance of lm(1~1) .
> >
> >Exactly the same problems I have if I want to "delete" an object. For
> >example you like to have a slot which stores some data which is important
> >if the object is in one state. But if the state of the object changes you
> >do not need that data anymore. Hence, you would like to delete it to get
> >storage. You can only do it if you have like you suggest "any", but the
> >consequence is "if" polluted code, or using slot(obj,"mod",check=FALSE)<- NULL.
> >
> >But the documentation for slot states.
> >check: " You should never set this to 'FALSE' in normal
> > use, since the result can create invalid objects. "
> >
> >
> >b) The Problem which I have writing intitialization functions like
> >suggested by J.C is: Imagine you have an object with twenty slots. You
> >have to write and initialize function with approximately 20 "if"'s. One of
> >the contribution of S4, as mentioned above, is to help you to avoid
> >writing such "if" polluted functions. But as I sometimes forget any role
> >has it's exceptions. So I can live with one function looking ugly in S4
> >instead of having all functions "if" polluted like in case of S3. The
> >only consequence so far is that I try to avoid to write initialize
> >functions. I do it only in some rare very special cases.
> >
> >The NULL exception in slot assignement will allow me object initialization
> >using just "new" without providing a "if" polluted function initialize.
> >
> > > setClass("bar")
> >[1] "bar"
> > > setClass("foo", representation(x="numeric", y="bar"))
> >[1] "foo"
> > > new("foo", x=1)
> >Error in validObject(.Object) : Invalid "foo" object: Invalid object for
> >slot "y" in class "foo": got class "NULL", should be or extend class "bar"
> > >
> >
> >
> >Are there S4 inherent solutions to this questions?
> >
> >
> >Yours.
> >/E
> >
> >*(I would rather compare "any" to the Object class then to null in java.
> >It has a different function.)
> >
> >Pfaff, Bernhard wrote:
> >
> >>>Hi,
> >>>
> >>>Thanks a lot for pointing me to setOldClass().
> >>>
> >>>What is exactly the point against it, to allow NULL to be a special case
> >>>which can be assigned to any slot? I think in java one can assing null
> >>>to any slot (I do not like java but not because of that).
> >>>I am asking because for some objects it may be hard to define something
> >>>equivalent to numeric(0). And even lm(1~1) is not that what I am looking
> >>>for. I would like to have the case
> >>>of an unassigned slot (Just think about storage?) even to the expense
> >>>checking every time if the slot is not null.
> >>>
> >>
> >>
> >>Hello Eryk,
> >>
> >>how about "ANY" and checking where suitable/needed if the passed object for
> >>"mod" is of class "lm" at a later stage in your code, like in:
> >>
> >>setClass("Ctest"
> >> ,representation(
> >> test="character"
> >> ,bla="character"
> >> ,mod="ANY" )
> >> )
> >>
> >>test1 <- new("Ctest",mod=lm(1~1))
> >>is(test1 at mod, "lm")
> >>
> >>test2 <- new("Ctest",test="bla")
> >>is(test2 at mod, "lm")
> >>
> >>test3 <- new("Ctest",bla="brum")
> >>is(test3 at mod, "lm")
> >>
> >>test4 <- new("Ctest",test="bla",bla="brum")
> >>is(test4 at mod, "lm")
> >>
> >> From there on you could use stop() or warning() or whatever you want to
> >>return to the user as information which class mod should belong to.
> >>
> >>HTH,
> >>Bernhard
> >>
> >
> >______________________________________________
> >R-devel at stat.math.ethz.ch mailing list
> >https://stat.ethz.ch/mailman/listinfo/r-devel
>
> Eric Lecoutre
> UCL / Institut de Statistique
> Voie du Roman Pays, 20
> 1348 Louvain-la-Neuve
> Belgium
>
> tel: (+32)(0)10473050
> lecoutre at stat.ucl.ac.be
> http://www.stat.ucl.ac.be/ISpersonnel/lecoutre
>
> If the statistics are boring, then you've got the wrong numbers. -Edward
> Tufte
>
> ______________________________________________
> R-devel at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
--
John M. Chambers jmc at bell-labs.com
Bell Labs, Lucent Technologies office: (908)582-2681
700 Mountain Avenue, Room 2C-282 fax: (908)582-3340
Murray Hill, NJ 07974 web: http://www.cs.bell-labs.com/~jmc
More information about the R-devel
mailing list