Tim: As the developer of DelayedMatrixStats (and enthusiastic 'canary down
the coal mine' user-dev of DelayedArray) I'm obviously invested in reducing
the confusion around these packages
I'm going to write some blog posts-cum-vignettes-cum-F1000 around these
issues over the coming weeks, with the ultimate goal of improving the
packages themselves.
Pete
On Mon., 30 Apr. 2018, 12:11 pm Tim Triche, Jr., <tim.triche at gmail.com>
wrote:
But if you merge methods like that, the error method can be that much more
difficult to identify. It took a couple of weeks to chase that bug down
properly, and it ended up down to rowMeans2 vs rowMeans.
I suppose the merged/abstracted method allows to centralize any such
dispatch into one place and swap out ill-behaved methods once identified,
so as long as DelayedArray/DelayedMatrixStats quirks are
documented/understood, maybe it is better to create this union class?
The Matrix/matrixStats/DelayedMatrix/DelayedMatrixStats situation has
been
"interesting" in practical terms, as seemingly simple abstractions appear
to require more thought. That was my only point.
--t
On Mon, Apr 30, 2018 at 11:28 AM, Martin Morgan <
martin.morgan at roswellpark.org> wrote:
But that issue will be fixed, so Tim's advice is inappropriate.
On 04/30/2018 10:42 AM, Tim Triche, Jr. wrote:
I am trying to extend my package to handle `HDF5Matrix` class ( or
generally `DelayedArray`). I currently have S4 functions for `matrix`
class. Usually I have a method for `SummarizedExperiment`, which will
call
call the method on `assay(x)` and I want the method to be able to deal
with
if `assay(x)` is a `DelayedArray`.
Most of my functions, however, do not require separate code depending
whether `x` is a `matrix` or `DelayedArray`. They are making use of
existing functions that will make that choice for me, e.g. rowMeans or
subsetting. My goal right now is compatibility, not cleverness, and
not
creating HDF5 methods to handle other cases. (If something doesn't
currently exist, then I just enclose `x` with `data.matrix` or
`as.matrix`
and call the matrix into memory ? for cleanliness and ease in updating
with
appropriate methods in future, I could make separate S4 functions for
these
specific tasks to dispatch, but that's outside of the scope of my
question). So for simplicity assume I don't really need to dispatch
code* -- that the methods I'm going to use do that.
The natural solution for me seem to use `setClassUnion` and I was
wondering if such a virtual class already exists? Or is there a better
way
to handle this?
Here's a simple example, using `rowMeans` as my example:
```
setGeneric("myNewRowMeans", function(x,...) { standardGeneric("
myNewRowMeans")})
setClassUnion("matrixOrDelayed",members=c("matrix", "DelayedArray"))
#' @importFrom DelayedArray rowMeans
setMethod("myNewRowMeans",
signature = "matrixOrDelayed",
definition = function(x,...){
# a lot of code independent of x
print("This is a lot of code shared
of
class of x\n")
# a lot of code that depends on x, but is
dispatched by the functions called
out<-rowMeans(x)
#a lot of code based on output of out
out<-out+1
return(out)
}
)
```