DelayedArray objects
Wrapping an array-like object (typically an on-disk object) in a DelayedArray object allows one to perform common array operations on it without loading the object in memory. In order to reduce memory usage and optimize performance, operations on the object are either delayed or executed using a block processing mechanism.
DelayedArray(seed) # constructor function type(x)
seed |
An array-like object. |
x |
Typically a DelayedArray object. More generally |
To realize a DelayedArray object (i.e. to trigger execution of the
delayed operations carried by the object and return the result as an
ordinary array), call as.array
on it. However this realizes the
full object at once in memory which could require too much memory
if the object is big. A big DelayedArray object is preferrably realized
on disk e.g. by calling writeHDF5Array
on
it (this function is defined in the HDF5Array package) or coercing it
to an HDF5Array object with as(x, "HDF5Array")
.
Other on-disk backends can be supported. This uses a block processing
strategy so that the full object is not realized at once in memory. Instead
the object is processed block by block i.e. the blocks are realized in
memory and written to disk one at a time.
See ?writeHDF5Array
in the HDF5Array package
for more information about this.
type()
is the DelayedArray equivalent of typeof()
(or
storage.mode()
) for ordinary arrays and vectors. Note that, for
convenience and consistency, type()
also supports ordinary arrays
and vectors. It should also support any array-like object, that is, any
object x
for which dim(x)
is not NULL.
A DelayedArray object can be subsetted with [
like an ordinary array,
but with the following differences:
N-dimensional single bracket subsetting (i.e. subsetting
of the form x[i_1, i_2, ..., i_n]
with one (possibly missing)
subscript per dimension) returns a DelayedArray object where the
subsetting is actually delayed. So it's a very light
operation. One notable exception is when drop=TRUE
and the
result has only one dimension, in which case it is realized
as an ordinary vector (atomic or list).
Note that NAs in the subscripts are not supported.
1D-style single bracket subsetting (i.e. subsetting of the
form x[i]
) only works if the subscript i
is a numeric
or logical vector, or a logical array-like object with the same
dimensions as x
, or a numeric matrix with one column per
dimension in x
. When i
is a numeric vector, all the
indices in it must be >= 1 and <= length(x)
. NAs in the
subscripts are not supported.
This is NOT a delayed operation (block processing is triggered)
i.e. the result is realized as an ordinary vector (atomic
or list). One exception is when x
has only one dimension
and drop
is set to FALSE
, in which case the subsetting
is delayed.
Subsetting with [[
is supported but only the 1D-style form of it
at the moment, that is, subsetting of the form x[[i]]
where i
is a single numeric value >= 1 and <= length(x)
. It is
equivalent to x[i][[1]]
.
Subassignment to a DelayedArray object with [<-
is also supported
like with an ordinary array, but with the following restrictions:
N-dimensional subassignment (i.e. subassignment of the
form x[i_1, i_2, ..., i_n] <- value
with one (possibly
missing) subscript per dimension) only accepts a replacement
value (a.k.a. right value) that is an array-like object (e.g.
ordinary array, dgCMatrix object, DelayedArray object, etc...)
or an ordinary vector (atomic or list) of length 1.
1D-style subassignment (a.k.a. 1D-style subassignment, that
is, subassignment of the form x[i] <- value
) only works if
the subscript i
is a logical DelayedArray object of the same
dimensions as x
and if the replacement value is an ordinary
vector (atomic or list) of length 1.
Filling with a vector, that is, subassignment of the form
x[] <- v
where v
is an ordinary vector (atomic or
list), is only supported if the length of the vector is a divisor
of nrow(x)
.
These 3 forms of subassignment are implemented as delayed operations so are very light.
Single value replacement (x[[...]] <- value
) is not supported yet.
showtree
for DelayedArray accessors
nseed
, seed
, and path
.
realize
for realizing a DelayedArray object in memory
or on disk.
blockApply
and family for convenient block
processing of an array-like object.
DelayedArray-utils for common operations on DelayedArray objects.
DelayedMatrix-utils for common operations on DelayedMatrix objects.
DelayedArray-stats for statistical functions on DelayedArray objects.
DelayedMatrix-stats for DelayedMatrix row/col summarization.
RleArray objects.
HDF5Array objects in the HDF5Array package.
DataFrame objects in the S4Vectors package.
array objects in base R.
## --------------------------------------------------------------------- ## A. WRAP AN ORDINARY ARRAY IN A DelayedArray OBJECT ## --------------------------------------------------------------------- a <- array(runif(1500000), dim=c(10000, 30, 5)) A <- DelayedArray(a) A ## The seed of a DelayedArray object is **always** treated as a ## "read-only" object so will never be modified by the operations ## we perform on A: stopifnot(identical(a, seed(A))) type(A) ## N-dimensional single bracket subsetting: m <- a[11:20 , 5, -3] # an ordinary matrix M <- A[11:20 , 5, -3] # a DelayedMatrix object stopifnot(identical(m, as.array(M))) ## 1D-style single bracket subsetting: A[11:20] A[A <= 1e-5] stopifnot(identical(a[a <= 1e-5], A[A <= 1e-5])) ## Subassignment: A[A < 0.2] <- NA a[a < 0.2] <- NA stopifnot(identical(a, as.array(A))) A[2:5, 1:2, ] <- array(1:40, c(4, 2, 5)) a[2:5, 1:2, ] <- array(1:40, c(4, 2, 5)) stopifnot(identical(a, as.array(A))) ## Other operations: crazy <- function(x) (5 * x[ , , 1] ^ 3 + 1L) * log(x[, , 2]) b <- crazy(a) head(b) B <- crazy(A) # very fast! (all operations are delayed) B cs <- colSums(b) CS <- colSums(B) stopifnot(identical(cs, CS)) ## --------------------------------------------------------------------- ## B. WRAP A DataFrame OBJECT IN A DelayedArray OBJECT ## --------------------------------------------------------------------- ## Generate random coverage and score along an imaginary chromosome: cov <- Rle(sample(20, 5000, replace=TRUE), sample(6, 5000, replace=TRUE)) score <- Rle(sample(100, nrun(cov), replace=TRUE), runLength(cov)) DF <- DataFrame(cov, score) A2 <- DelayedArray(DF) A2 seed(A2) # 'DF' ## Coercion of a DelayedMatrix object to DataFrame produces a DataFrame ## object with Rle columns: as(A2, "DataFrame") stopifnot(identical(DF, as(A2, "DataFrame"))) t(A2) # transposition is delayed so is very fast and memory-efficient colSums(A2) ## --------------------------------------------------------------------- ## C. AN HDF5Array OBJECT IS A (PARTICULAR KIND OF) DelayedArray OBJECT ## --------------------------------------------------------------------- library(HDF5Array) A3 <- as(a, "HDF5Array") # write 'a' to an HDF5 file A3 is(A3, "DelayedArray") # TRUE seed(A3) # an HDF5ArraySeed object B3 <- crazy(A3) # very fast! (all operations are delayed) B3 # not an HDF5Array object anymore because # now it carries delayed operations CS3 <- colSums(B3) stopifnot(identical(cs, CS3)) ## --------------------------------------------------------------------- ## D. PERFORM THE DELAYED OPERATIONS ## --------------------------------------------------------------------- as(B3, "HDF5Array") # "realize" 'B3' on disk ## If this is just an intermediate result, you can either keep going ## with B3 or replace it with its "realized" version: B3 <- as(B3, "HDF5Array") # no more delayed operations on new 'B3' seed(B3) path(B3) ## For convenience, realize() can be used instead of explicit coercion. ## The current "automatic realization backend" controls where ## realization happens e.g. in memory if set to NULL or in an HDF5 ## file if set to "HDF5Array": D <- cbind(B3, exp(B3)) D setAutoRealizationBackend("HDF5Array") D <- realize(D) D ## See '?setAutoRealizationBackend' for more information about ## "realization backends". ## --------------------------------------------------------------------- ## E. MODIFY THE PATH OF A DelayedArray OBJECT ## --------------------------------------------------------------------- ## This can be useful if the file containing the array data is on a ## shared partition but the exact path to the partition depends on the ## machine from which the data is being accessed. ## For example: ## Not run: library(HDF5Array) A <- HDF5Array("/path/to/lab_data/my_precious_data.h5") path(A) ## Operate on A... ## Now A carries delayed operations. ## Make sure path(A) still works: path(A) ## Save A: save(A, file="A.rda") ## A.rda should be small (it doesn't contain the array data). ## Send it to a co-worker that has access to my_precious_data.h5. ## Co-worker loads it: load("A.rda") path(A) ## A is broken because path(A) is incorrect for co-worker: A # error! ## Co-worker fixes the path (in this case this is better done using the ## dirname() setter rather than the path() setter): dirname(A) <- "E:/other/path/to/lab_data" ## A "works" again: A ## End(Not run) ## --------------------------------------------------------------------- ## F. WRAP A SPARSE MATRIX IN A DelayedArray OBJECT ## --------------------------------------------------------------------- ## Not run: M <- 75000L N <- 1800L p <- sparseMatrix(sample(M, 9000000, replace=TRUE), sample(N, 9000000, replace=TRUE), x=runif(9000000), dims=c(M, N)) P <- DelayedArray(p) P p2 <- as(P, "sparseMatrix") stopifnot(identical(p, p2)) ## The following is based on the following post by Murat Tasan on the ## R-help mailing list: ## https://stat.ethz.ch/pipermail/r-help/2017-May/446702.html ## As pointed out by Murat, the straight-forward row normalization ## directly on sparse matrix 'p' would consume too much memory: row_normalized_p <- p / rowSums(p^2) # consumes too much memory ## because the rowSums() result is being recycled (appropriately) into a ## *dense* matrix with dimensions equal to dim(p). ## Murat came up with the following solution that is very fast and ## memory-efficient: row_normalized_p1 <- Diagonal(x=1/sqrt(Matrix::rowSums(p^2))) ## With a DelayedArray object, the straight-forward approach uses a ## block processing strategy behind the scene so it doesn't consume ## too much memory. ## First, let's see block processing in action: DelayedArray:::set_verbose_block_processing(TRUE) ## and check the automatic block size: getAutoBlockSize() row_normalized_P <- P / sqrt(DelayedArray::rowSums(P^2)) ## Increasing the block size increases the speed but also memory usage: setAutoBlockSize(2e8) row_normalized_P2 <- P / sqrt(DelayedArray::rowSums(P^2)) stopifnot(all.equal(row_normalized_P, row_normalized_P2)) ## Back to sparse representation: DelayedArray:::set_verbose_block_processing(FALSE) row_normalized_p2 <- as(row_normalized_P, "sparseMatrix") stopifnot(all.equal(row_normalized_p1, row_normalized_p2)) setAutoBlockSize() # reset automatic block size to factory settings ## End(Not run)
Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.