On 05/16/2018 10:22 AM, Michael Lawrence wrote:
Factors and data.frames are not structures, because they must have a
class attribute. Just call them "objects". They are higher level than
structures, which in practice just shape data without adding a lot of
semantics. Compare getClass("matrix") and getClass("factor").
I agree that inheritance through explicit coercion is confusing. As
far as I know, there are only 2 places where it is used:
1) Objects with attributes but no class, basically "structure" and its
subclasses "array" <- "matrix"
2) Classes that extend a reference type ("environment", "name" and
"externalptr") via hidden delegation (@.xData)
I'm not sure if anyone should be doing #2. For #1, a simple "fix"
would be just to drop inheritance of "structure" from "vector". I
think the intent was to mimic base R behavior, where it will happily
strip (or at least ignore) attributes when passing an array or matrix
to an internal function that expects a vector.
A related problem, which explains why factor and data.frame inherit
from "vector" even though they are objects, is that any S4 object
derived from those needs to be (for pragmatic compatibility reasons)
an integer vector or list, respectively, internally (the virtual
@.Data slot). Separating that from inheritance would probably be
difficult.
Yes, we can consider these to be problems, to some extent stemming
from the behavior and design of R itself, but I'm not sure it's worth
doing anything about them at this point.
Thanks for the informative discussion. It still doesn't explain
why 'm' gets its attributes stripped and 'x' does not though:
m <- matrix(1:12, ncol=3)
x <- structure(1:3, titi="A")
setGeneric("foo", function(x) standardGeneric("foo"))
setMethod("foo", "vector", identity)
foo(m)
# [1] 1 2 3 4 5 6 7 8 9 10 11 12
foo(x)
# [1] 1 2 3
# attr(,"titi")
# [1] "A"
If I understand correctly, both are "structures", not "objects".
Why aren't these problems worth fixing? More generally speaking
the erratic behavior of the S4 system with respect to S3 objects
has been a plague since the beginning of the methods package.
And many people have complained about this in many occasions in
one way or another. For the record, here are some of the most
notorious problems:
class(as.numeric(1:4))
# [1] "numeric"
class(as(1:4, "numeric"))
# [1] "integer"
is.vector(matrix())
# [1] FALSE
is(matrix(), "vector")
# [1] TRUE
is.list(data.frame())
# [1] TRUE
is(data.frame(), "list")
# [1] FALSE
extends("data.frame", "list")
# [1] TRUE
setClassUnion("vector_OR_factor", c("vector", "factor"))
is(data.frame(), "vector")
# [1] FALSE
is(data.frame(), "factor")
# [1] FALSE
is(data.frame(), "vector_OR_factor")
# [1] TRUE
etc...
Many people stay away from S4 because of these incomprehensible
behaviors.
Finally note that even pure S3 operations can produce output that
doesn't make sense:
is.list(data.frame())
# [1] TRUE
is.vector(list())
# [1] TRUE
is.vector(data.frame())
# [1] FALSE
(that is: a data frame is a list and a list is a vector but
a data frame is not a vector!)
Why aren't these problems taken more seriously?
Thanks,
H.
Michael
On Wed, May 16, 2018 at 8:33 AM, Hervé Pagès <hpa...@fredhutch.org> wrote:
On 05/15/2018 09:13 PM, Michael Lawrence wrote:
My understanding is that array (or any other structure) does not
"simply" inherit from vector, because structures are not vectors in
the strictest sense. Basically, once a vector gains attributes, it is
a structure, not a vector. The methods package accommodates this by
defining an "is" relationship between "structure" and "vector" via an
"explicit coerce", such that any "structure" passed to a "vector"
method is first passed to as.vector(), which strips attributes. This
is very much by design.
It seems that the problem is really with matrices and arrays, not
with "structures" in general:
f <- factor(c("z", "x", "z"), levels=letters)
m <- matrix(1:12, ncol=3)
df <- data.frame(f=f)
x <- structure(1:3, titi="A")
Only the matrix looses its attributes when passed to a "vector"
method:
setGeneric("foo", function(x) standardGeneric("foo"))
setMethod("foo", "vector", identity)
foo(f) # attributes are preserved
# [1] z x z
# Levels: a b c d e f g h i j k l m n o p q r s t u v w x y z
foo(m) # attributes are stripped
# [1] 1 2 3 4 5 6 7 8 9 10 11 12
foo(df) # attributes are preserved
# f
# 1 z
# 2 x
# 3 z
foo(x) # attributes are preserved
# [1] 1 2 3
# attr(,"titi")
# [1] "A"
Also if structures are passed to as.vector() before being passed to
a "vector" method, shouldn't as.vector() and foo() be equivalent on
them? For 'f' and 'x' they're not:
as.vector(f)
# [1] "z" "x" "z"
as.vector(x)
# [1] 1 2 3
Finally note that for factors and data frames the "vector" method gets
selected despite the fact that is( , "vector") is FALSE:
is(f, "vector")
# [1] FALSE
is(m, "vector")
# [1] TRUE
is(df, "vector")
# [1] FALSE
is(x, "vector")
# [1] TRUE
Couldn't we recognize these problems as real, even if they are by
design? Hopefully we can all agree that:
- the dispatch mechanism should only dispatch, not alter objects;
- is() and selectMethod() should not contradict each other.
Thanks,
H.
Michael
On Tue, May 15, 2018 at 5:25 PM, Hervé Pagès <hpa...@fredhutch.org> wrote:
Hi,
This was quite unexpected:
setGeneric("foo", function(x) standardGeneric("foo"))
setMethod("foo", "vector", identity)
foo(matrix(1:12, ncol=3))
# [1] 1 2 3 4 5 6 7 8 9 10 11 12
foo(array(1:24, 4:2))
# [1] 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
22 23
24
If I define a method for array objects, things work as expected though:
setMethod("foo", "array", identity)
foo(matrix(1:12, ncol=3))
# [,1] [,2] [,3]
# [1,] 1 5 9
# [2,] 2 6 10
# [3,] 3 7 11
# [4,] 4 8 12
So, luckily, I have a workaround.
But shouldn't the dispatch mechanism stay away from the business of
altering objects before passed to it?
Thanks,
H.
--
Hervé Pagès
Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024
E-mail: hpa...@fredhutch.org
Phone: (206) 667-5791
Fax: (206) 667-1319
______________________________________________
R-devel@r-project.org mailing list
https://urldefense.proofpoint.com/v2/url?u=https-3A__stat.ethz.ch_mailman_listinfo_r-2Ddevel&d=DwIFaQ&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=gynT4YhbmVKZhnX4srXlCWZZRyVBMXG211CKgftdEs0&s=_I0aFHQVnXdBfB5kTLg9TxK_2LHdSuaB6gqZwSx1orQ&e=
--
Hervé Pagès
Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024
E-mail: hpa...@fredhutch.org
Phone: (206) 667-5791
Fax: (206) 667-1319
--
Hervé Pagès
Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024
E-mail: hpa...@fredhutch.org
Phone: (206) 667-5791
Fax: (206) 667-1319
______________________________________________
R-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel
