[Rd] Dispatch mechanism seems to alter object before calling method on it

Michael Lawrence l@wrence@mich@el @ending from gene@com
Wed May 16 22:24:26 CEST 2018


On Wed, May 16, 2018 at 12:23 PM, Hervé Pagès <hpages at fredhutch.org> wrote:
> 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".
>

The structure 'x' has no class, so nothing special is going to happen.
As you know, S4 has a well-defined class hierarchy. Just look at
getClass("structure") to see its subclasses. There was at some point
an attempt to create a sort of dynamic inheritance, where a 'test'
function would be called and could figure this out. However, that was
never implemented. For one thing, it would be even more confusing.

> 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"
>

This is not really a problem with the methods package. is.numeric(1L)
is TRUE, thus integer extends numeric, so coercing an integer to
numeric is a no-op. as.numeric() should really be called as.double()
or something. But that's not going to change, of course.

>   is.vector(matrix())
>   # [1] FALSE
>   is(matrix(), "vector")
>   # [1] TRUE
>

We already discussed this in the context of "structure" inheriting
from "vector" and explicit coercion.

>   is.list(data.frame())
>   # [1] TRUE
>   is(data.frame(), "list")
>   # [1] FALSE
>   extends("data.frame", "list")
>   # [1] TRUE
>

This is a compromise for compatibility with inherits(), since the
result of data.frame() is an S3 object.

>
>   is(data.frame(), "vector")
>   # [1] FALSE
>   is(data.frame(), "factor")
>   # [1] FALSE
>   is(data.frame(), "vector_OR_factor")
>   # [1] TRUE
>

The question is: which inheritance to follow, S3 or S4? Since "vector"
is a basic class, inheritance follows S3 rules. But the class union is
an S4 class, so it follows S4 rules.

>   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!)
>

R has no notion of inheritance here. These are just different
functions checking different things. Bringing this up again after so
many discussions borders on trolling.

> Why aren't these problems taken more seriously?
>

They are taken seriously. But there are serious semantic differences
between S3, S4 and base type checking functions. The S3/S4 integration
should be viewed as a tool that is useful in practice, despite forced
compromises.

There are changes that would resolve some of these issues, like those
suggested earlier in this thread, but it's likely too disruptive to
make them now. Energy is better spent thinking about how we will do it
"right" the next time around.

> Thanks,
> H.
>
>>
>> Michael
>>
>> On Wed, May 16, 2018 at 8:33 AM, Hervé Pagès <hpages at 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 <hpages at 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: hpages at fredhutch.org
>>>>> Phone:  (206) 667-5791
>>>>> Fax:    (206) 667-1319
>>>>>
>>>>> ______________________________________________
>>>>> R-devel at 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: hpages at 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: hpages at fredhutch.org
> Phone:  (206) 667-5791
> Fax:    (206) 667-1319
>



More information about the R-devel mailing list