as_tibble(matrix w/ attributes) copies too many attributes from matrix to its columns

[BillDunlap]

The as_tibble method for matrices copies all the attributes of the input matrix to
each of its columns, then removes the 'names' and 'dim' attributes (see matrixToDataFrame.cpp:
copy_most_attributes). This can make illegal columns because, e.g., certain classes require certain attributes to be present. E.g., "mts" and "matrix" objects are expected to have a "dim" attribute so you can
get problems like the following

mts <- cbind(A=ts(c(1,1,2,2), start=2016, freq=4), B=ts(c(11,11,12,13), start=2016, freq=4))
tblMts <- tibble::as_tibble(mts)
str(attributes(tblMts[["A"]]))
   # List of 2
   # $ tsp  : num [1:3] 2016 2017 4
   # $ class: chr [1:3] "mts" "ts" "matrix"
duplicated(tblMts[["A"]])
   # Error in duplicated.matrix(tblMts[["A"]]) : character(0)

With TIBCO's TERR there are further problems with the same code because its SET_ATTRIB checks that the dim attribute is compatible with the object's length and R's SET_ATTRIB does not.

Calling out to the R-level "[" to extract the columns would solve the problem.

(tibble-1.2, R-3.3.0)

[krlmlr]

Tibbles currently don't support matrix columns, so the above should give an error.

[BillDunlap]

Note that I am not trying insert a matrix column into a tibble,
I am making a 2-column tibble from a 2-column time-series matrix.

I would expect each column to be a time-series vector or perhaps
a pure vector. as_tibble tries to produces such a thing but it
can put illegal combinations of attributes on the columns (since
it does not go through the "[" method for the time-series matrix).
The illegal combinations of attributes are not immediately apparent,
but will show up when a method assumes the input object is legal.

On Thu, Nov 17, 2016 at 7:07 AM, Kirill Müller notifications@github.com
wrote:

Tibbles currently don't support matrix columns, so the above should give
an error.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#184 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/ALHWBZROuqOdVxl9jQ7-IkdT46IvfF_1ks5q_G2wgaJpZM4KdlI8
.

[cfhammill]

@krlmlr I think this problem is more general, any time a matrix is given an extra class it breaks matrixToDataFrame.

e.g.

mat <- matrix(1:100, nrow = 10)
class(mat) <- c("special_matrix", "matrix")
as_tibble(mat) %>% mutate(colnum = 1:10) ## error

[krlmlr]

I currently don't have a better solution except to add special handling for each matrix class that breaks, as I did with ts here.

[BillDunlap]

One might be able to fix it by doing the subscripting in C++ if the object has dims but no class (or exactly the class "matrix") and calling out to the interpreter to evaluate 'mat[,column]' or 'mat[,column,drop=TRUE]' for those objects with dims but unrecognized class. We can expect that their [ methods should return a column with appropriate attributes.

…

On Tue, Jan 10, 2017 at 11:55 AM, Chris Hammill ***@***.***> wrote: @krlmlr <https://github.com/krlmlr> This problem is more general, any time a matrix is given an extra class it breaks matrixToDataFrame. e.g. mat <- matrix(1:100, nrow = 10) attr(mat, "special") <- "a little extra" class(mat) <- c("special_matrix", "matrix") as_tibble(mat) %>% mutate(colnum = 1:10) ## error — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#184 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ALHWBQTHdBTOmwuWTOwkfWuXKpoMnYR1ks5rQ-JDgaJpZM4KdlI8> .

[krlmlr]

@BillDunlap: Your strategy fails for an existing test for a matrix-like object of class Date, this is supposed to be converted to a tibble with date columns. @hadley: Do you remember the rationale behind this behavior?

[BillDunlap]

I find it surprising that you want your function to act differently than '[' when splitting a matrix into its columns.

…

On Tue, Jan 10, 2017 at 12:21 PM, Kirill Müller ***@***.***> wrote: @BillDunlap <https://github.com/BillDunlap>: Your strategy fails for an existing test <https://github.com/tidyverse/tibble/blob/3d6f8b417f5ed994c59b60528075a6825013c5da/tests/testthat/test-matrix.R#L25-L35> for a matrix-like object of class Date, this is supposed to be converted to a tibble with date columns. @hadley <https://github.com/hadley>: Do you remember the rationale behind this behavior? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#184 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ALHWBRiYG0JccL0piCPSvK0q94wHX5uOks5rQ-hTgaJpZM4KdlI8> .

[cfhammill]

Would it make more sense to special case the matrix-like objects that copy their attributes to the
columns? That way the general rule would be to either discard extra class/attribute information and
users would be free to arbitrarily subclass their matrices.

[hadley]

The original motivation is probably that as.data.frame(date) doesn't do the right thing.

[hadley]

And I have a vague memory that's important for tidyr.

[BillDunlap]

I think that all of the as.data.frame methods should worry about their argument being a matrix, since dates, factors, etc. can have a dims attribute. But, if some class's [ method chooses to drop or change attributes when extracting a column, the matrix->list-of-columns code should respect that choice.

…

On Tue, Jan 10, 2017 at 12:35 PM, Hadley Wickham ***@***.***> wrote: The original motivation is probably that as.data.frame(date) doesn't do the right thing. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#184 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ALHWBTt5IwfklGn6IUeqR00La4hKaipgks5rQ-uegaJpZM4KdlI8> .