Dispatch mechanism seems to alter object before calling method on it

On Wed, May 16, 2018 at 12:23 PM, Herv? Pag?s <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, Herv? Pag?s <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, Herv? Pag?s <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.

--
Herv? Pag?s

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

--
Herv? Pag?s

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

--
Herv? Pag?s

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