Skip to content

Reshaping an array - how does it work in R

14 messages · Bert Gunter, Jeff Newmiller, Henrik Bengtsson +3 more

#
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



**********************
"The contents of this message do not reflect any position of the U.S. Government or NOAA."
**********************
Roy Mendelssohn
Supervisory Operations Research Analyst
NOAA/NMFS
Environmental Research Division
Southwest Fisheries Science Center
***Note new address and phone***
110 Shaffer Road
Santa Cruz, CA 95060
Phone: (831)-420-3666
Fax: (831) 420-3980
e-mail: Roy.Mendelssohn at noaa.gov www: http://www.pfeg.noaa.gov/

"Old age and treachery will overcome youth and skill."
"From those who have been given much, much will be expected" 
"the arc of the moral universe is long, but it bends toward justice" -MLK Jr.
#
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.

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.

HTH,
   Denes
On 03/18/2016 10:28 PM, Roy Mendelssohn - NOAA Federal wrote:
#
arrays are vectors stored in column major order.  So the answer is: reindexing.

Does this make it clear:
[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
, , 1

     [,1] [,2] [,3]
[1,]    1    3    5
[2,]    2    4    6

, , 2

     [,1] [,2] [,3]
[1,]    7    9   11
[2,]    8   10   12

, , 3

     [,1] [,2] [,3]
[1,]   13   15   17
[2,]   14   16   18

, , 4

     [,1] [,2] [,3]
[1,]   19   21   23
[2,]   20   22   24
[,1] [,2] [,3] [,4]
[1,]    1    7   13   19
[2,]    2    8   14   20
[3,]    3    9   15   21
[4,]    4   10   16   22
[5,]    5   11   17   23
[6,]    6   12   18   24
[1] TRUE


However copying may occur anyway as part of R's semantics. Others will
have to help you on that, as the details here are beyond me.

Cheers,
Bert



Cheers,
Bert


Bert Gunter

"The trouble with having an open mind is that people keep coming along
and sticking things into it."
-- Opus (aka Berkeley Breathed in his "Bloom County" comic strip )


On Fri, Mar 18, 2016 at 2:28 PM, Roy Mendelssohn - NOAA Federal
<roy.mendelssohn at noaa.gov> wrote:
#
Hi Bert:

Thanks for your response.  The only part I was concerned with is whether a copy was made, that is what my memory usage would be.  Sorry if that wasn?t clear in the original.

-Roy


**********************
"The contents of this message do not reflect any position of the U.S. Government or NOAA."
**********************
Roy Mendelssohn
Supervisory Operations Research Analyst
NOAA/NMFS
Environmental Research Division
Southwest Fisheries Science Center
***Note new address and phone***
110 Shaffer Road
Santa Cruz, CA 95060
Phone: (831)-420-3666
Fax: (831) 420-3980
e-mail: Roy.Mendelssohn at noaa.gov www: http://www.pfeg.noaa.gov/

"Old age and treachery will overcome youth and skill."
"From those who have been given much, much will be expected" 
"the arc of the moral universe is long, but it bends toward justice" -MLK Jr.
#
R always makes a copy for this kind of operation. There are some operations that don't make copies, but I don't think this one qualifies.
#
Thanks.  That is what I needed to know.  I don?t want to play around with some of the other suggestions, as I don?t totally understand what they do, and don?t want to risk messing up something and not be aware of it.

There is a way to read in the data chunks at a time and reshape it and put, it into the (reshaped) larger array, harder to program but probably worth the pain to be certain of what I am doing.

I had a feeling a copy was made, just wanted to make certain of it.

Thanks again,

-Roy
**********************
"The contents of this message do not reflect any position of the U.S. Government or NOAA."
**********************
Roy Mendelssohn
Supervisory Operations Research Analyst
NOAA/NMFS
Environmental Research Division
Southwest Fisheries Science Center
***Note new address and phone***
110 Shaffer Road
Santa Cruz, CA 95060
Phone: (831)-420-3666
Fax: (831) 420-3980
e-mail: Roy.Mendelssohn at noaa.gov www: http://www.pfeg.noaa.gov/

"Old age and treachery will overcome youth and skill."
"From those who have been given much, much will be expected" 
"the arc of the moral universe is long, but it bends toward justice" -MLK Jr.
#
On Fri, Mar 18, 2016 at 3:15 PM, Roy Mendelssohn - NOAA Federal
<roy.mendelssohn at noaa.gov> wrote:
I recommend this approach; whenever I work with reasonable large data
(that may become even larger in the future), I try to implement a
constant-memory version of the algorithm, which often comes down to
processing data in chunks.  The simplest version of this is to read
all data into memory and the subset, but if you can read data in in
chunks that is even better.

Though, I'm curious to what matrix operations you wish to perform.
Because if you wish to do regular summation, then base::.rowSums() and
base::.colSums() allow you to override the default dimensions on the
fly without inducing extra copies, e.g.
[1] 40 44 48 52 56 60
[1] 40 44 48 52 56 60

For other types of calculations, you might want to look at
matrixStats.  It has partial(*) support for overriding the default
dimensions in a similar fashion.  For instance,
The above effectively calculates rowVars(matrix(X, nrow=6, ncol=4))
without making copies.

(*) By partial I mean that this is a feature that hasn't been pushed
through to all of matrixStats functions, cf.
https://github.com/HenrikBengtsson/matrixStats/issues/83.

Cheers,

Henrik
(author of matrixStats)
#
Hi Henrik:

I want to do want in oceanography is called an EOF, which is just a PCA analysis. Unless I am missing something, in R I need to flatten my 3-D matrix into a 2-D data matrix. I can fit the entire 30GB matrix into memory, and I believe I have enough memory to do the PCA by constraining the number of components returned .  What I don?t think I have enough memory for is an operation that makes a copy of the matrix.

As I said, in theory I know how to do the flattening, it a simple command, but in practice I don?t have enough memory.  So I spent the afternoon rewriting my code to read in parts of the data at a time and then putting those in the appropriate places of a matrix already flattened of appropriate size.  In case someone is wondering, on the 3D grid the matrix is [1001,1001,3650].  So I create an empty matrix size [1001*1001,3650], and read in a slice of the lat-lon grid, and map those into the appropriate places in the flattened matrix.  By reading in appropriately sized chunks  my memory usage is not pushed too far.

-Roy
**********************
"The contents of this message do not reflect any position of the U.S. Government or NOAA."
**********************
Roy Mendelssohn
Supervisory Operations Research Analyst
NOAA/NMFS
Environmental Research Division
Southwest Fisheries Science Center
***Note new address and phone***
110 Shaffer Road
Santa Cruz, CA 95060
Phone: (831)-420-3666
Fax: (831) 420-3980
e-mail: Roy.Mendelssohn at noaa.gov www: http://www.pfeg.noaa.gov/

"Old age and treachery will overcome youth and skill."
"From those who have been given much, much will be expected" 
"the arc of the moral universe is long, but it bends toward justice" -MLK Jr.
#
On Fri, Mar 18, 2016 at 8:28 PM, Roy Mendelssohn - NOAA Federal
<roy.mendelssohn at noaa.gov> wrote:
Sounds good.  There's another small caveat. Make sure to specify the
'data' argument for matrix() we allocating an "empty" matrix, e.g.

    X <- matrix(NA_real_, nrow=1001*1001, ncol=3650)

This will give you a double matrix with all missing value.  If you use
the default

    X <- matrix(nrow=1001*1001, ncol=3650)

you'll get a logical matrix, which will introduce a copy as soon as
you assign a double value (e.g. X[1,1] <- 3.14). The latter is a
complete waste of memory/time. See
http://www.jottr.org/2014/06/matrixNA-wrong-way.html for details.

/Henrik
#
Thanks.  Yes one time for some reason I can?t remember I did ?NA  where that is documented but it is not something you would think of offhand.

-Roy


**********************
"The contents of this message do not reflect any position of the U.S. Government or NOAA."
**********************
Roy Mendelssohn
Supervisory Operations Research Analyst
NOAA/NMFS
Environmental Research Division
Southwest Fisheries Science Center
***Note new address and phone***
110 Shaffer Road
Santa Cruz, CA 95060
Phone: (831)-420-3666
Fax: (831) 420-3980
e-mail: Roy.Mendelssohn at noaa.gov www: http://www.pfeg.noaa.gov/

"Old age and treachery will overcome youth and skill."
"From those who have been given much, much will be expected" 
"the arc of the moral universe is long, but it bends toward justice" -MLK Jr.
2 days later
#
> 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.

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
#
Hi Martin,
On 03/22/2016 10:20 AM, Martin Maechler wrote:
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
#
Thanks all.  This is interesting, and for what I am doing worthwhile and helpful.  I have to be careful in each operation whether a copy is made or not,  and knowing this allows me to test on small examples what any command will do before I use,

Thanks again, I appreciate all the help.  I will have a related question, but will put it under a different heading.

-Roy
**********************
"The contents of this message do not reflect any position of the U.S. Government or NOAA."
**********************
Roy Mendelssohn
Supervisory Operations Research Analyst
NOAA/NMFS
Environmental Research Division
Southwest Fisheries Science Center
***Note new address and phone***
110 Shaffer Road
Santa Cruz, CA 95060
Phone: (831)-420-3666
Fax: (831) 420-3980
e-mail: Roy.Mendelssohn at noaa.gov www: http://www.pfeg.noaa.gov/

"Old age and treachery will overcome youth and skill."
"From those who have been given much, much will be expected" 
"the arc of the moral universe is long, but it bends toward justice" -MLK Jr.
1 day later
#
> Hi Martin,
> On 03/22/2016 10:20 AM, Martin Maechler wrote:
>>>>>>> >>>>>D?nes T?th<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.

Interesting, indeed.

I can confirm the bugous  Rstudio behavior
using the latest version of Rstudio (64 bit Linux, Fedora 22)
      RStudio Version 0.99.891 ? ? 2009-2016 RStudio, Inc.

The attached small R script is very transparent in demonstrating
the problem.
If you have a tracemem-enabled version of R, the output is even
more revealing, inside Rstudio it gives
+     if(capabilities("profmem")) {
+         tracemem(x)
+     } else {
+         cat("R version not configured for memory tracing\n")
+         .Internal(inspect(x))# also works w/o tracemem
+     }
+ }
[1] "<0x7fad78b37010>"
tracemem[0x7fad78b37010 -> 0x7fad77bf4010]:
[1] "<0x7fad77bf4010>"
tracemem[0x7fad77bf4010 -> 0x1ad05f50]:
[1] "<0x1ad05f50>"
Martin


    > 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