[R] Reshaping an array - how does it work in R

[Dénes Tóth]

Hi Martin,


On 03/22/2016 10:20 AM, Martin Maechler wrote:
>>>>>> >>>>>Dénes Tóth<toth.denes at ttk.mta.hu>
>>>>>> >>>>>     on Fri, 18 Mar 2016 22:56:23 +0100 writes:
>      > Hi Roy,
>      > R (usually) makes a copy if the dimensionality of an array is modified,
>      > even if you use this syntax:
>
>      > x <- array(1:24, c(2, 3, 4))
>      > dim(x) <- c(6, 4)
>
>      > See also ?tracemem, ?data.table::address, ?pryr::address and other tools
>      > to trace if an internal copy is done.
>
> Well, without using strange (;-) packages,  indeed standard R's
> tracemem(), notably the help page is a good pointer.
>
> According to the help page memory tracing is enabled in the
> default R binaries for Windows and OS X.
> For Linux (where I, as R developer, compile R myself anyway),
> one needs to configure with --enable-memory-profiling .
>
> Now, let's try:
>
>     > x <- array(rnorm(47), dim = c(1000,50, 40))
>     > tracemem(x)
>     [1] "<0x7f79a498a010>"
>     > dim(x) <- c(1000* 50, 40)
>     > x[5] <- pi
>     > tracemem(x)
>     [1] "<0x7f79a498a010>"
>     >
>
> So,*BOTH*   the re-dimensioning*AND*   the  sub-assignment did
> *NOT*  make a copy.

This is interesting. First I wanted to demonstrate to Roy that recent R 
versions are smart enough not to make any copy during reshaping an 
array. Then I put together an example (similar to yours) and realized 
that after several reshapes, R starts to copy the array. So I had to 
modify my suggestion... And now, I realized that this was an 
RStudio-issue. At least on Linux, a standard R terminal behaves as you 
described, however, RStudio (version 0.99.862, which is not the very 
latest) tends to create copies (quite randomly, at least to me). If I 
have time I will test this more thoroughly and file a report to RStudio 
if it turns out to be a bug.

Denes

>
> Indeed, R has become much smarter  in these things in recent
> years ... not thanks to me, but very much thanks to
> Luke Tierney (from R-core), and also thanks to contributions from "outside",
> notably Tomas Kalibera.
>
> And hence:*NO*  such strange workarounds are needed in this specific case:
>
>      > Workaround: use data.table::setattr or bit::setattr to modify the
>      > dimensions in place (i.e., without making a copy). Risk: if you modify
>      > an object by reference, all other objects which point to the same memory
>      > address will be modified silently, too.
>
> Martin Maechler, ETH Zurich  (and R-core)
>
>      > HTH,
>      > Denes
>
> (generally, your contributions help indeed, Denes, thank you!)
>
>
>      > On 03/18/2016 10:28 PM, Roy Mendelssohn - NOAA Federal wrote:
>      >> Hi All:
>      >>
>      >> I am working with a very large array.  if noLat is the number of latitudes, noLon the number of longitudes and noTime the number of  time periods, the array is of the form:
>      >>
>      >> myData[noLat, no Lon, noTime].
>      >>
>      >> It is read in this way because that is how it is stored in a (series) of netcdf files.  For the analysis I need to do, I need instead the array:
>      >>
>      >> myData[noLat*noLon, noTime].  Normally this would be easy:
>      >>
>      >> myData<- array(myData,dim=c(noLat*noLon,noTime))
>      >>
>      >> My question is how does this command work in R - does it make a copy of the existing array, with different indices for the dimensions, or does it just redo the indices and leave the given array as is?  The reason for this question is my array is 30GB in memory, and I don’t have enough space to have a copy of the array in memory.  If the latter I will have to figure out a work around to bring in only part of the data at a time and put it into the proper locations.
>      >>
>      >> Thanks,
>      >>
>      >> -Roy
>