ffapply: Apply for ff objects

In ff: Memory-Efficient Storage of Large Data on Disk and Fast Access Functions

Apply for ff objects

Description

The ffapply functions support convenient batched processing of ff objects such that each single batch or chunk will not exhaust RAM and such that batchs have sizes as similar as possible, see bbatch. Differing from R's standard apply which applies a FUNction, the ffapply functions do apply an EXPRession and provide two indices FROM="i1" and TO="i2", which mark beginning and end of the batch and can be used in the applied expression. Note that the ffapply functions change the two indices in their parent frame, to avoid conflicts you can use different names through FROM="i1" and TO="i2". For support of creating return values see details.

Usage

ffvecapply(EXPR, X = NULL, N = NULL, VMODE = NULL, VBYTES = NULL, RETURN = FALSE
, CFUN = NULL, USE.NAMES = TRUE, FF_RETURN = TRUE, BREAK = ".break"
, FROM = "i1", TO = "i2"
, BATCHSIZE = .Machine$integer.max, BATCHBYTES = getOption("ffbatchbytes")
, VERBOSE = FALSE)
ffrowapply(EXPR, X = NULL, N = NULL, NCOL = NULL, VMODE = NULL, VBYTES = NULL
, RETURN = FALSE, RETCOL = NCOL, CFUN = NULL, USE.NAMES = TRUE, FF_RETURN = TRUE
, FROM = "i1", TO = "i2"
, BATCHSIZE = .Machine$integer.max, BATCHBYTES = getOption("ffbatchbytes")
, VERBOSE = FALSE)
ffcolapply(EXPR, X = NULL, N = NULL, NROW = NULL, VMODE = NULL, VBYTES = NULL
, RETURN = FALSE, RETROW = NROW, CFUN = NULL, USE.NAMES = TRUE, FF_RETURN = TRUE
, FROM = "i1", TO = "i2"
, BATCHSIZE = .Machine$integer.max, BATCHBYTES = getOption("ffbatchbytes")
, VERBOSE = FALSE)
ffapply(EXPR = NULL, AFUN = NULL, MARGIN = NULL, X = NULL, N = NULL, DIM = NULL
, VMODE = NULL, VBYTES = NULL, RETURN = FALSE, CFUN = NULL, USE.NAMES = TRUE
, FF_RETURN = TRUE, IDIM = "idim"
, FROM = "i1", TO = "i2", BREAK = ".break"
, BATCHSIZE = .Machine$integer.max, BATCHBYTES = getOption("ffbatchbytes")
, VERBOSE = FALSE)

Arguments

EXPR	the expression to be applied
AFUN	ffapply only: alternatively to EXPR the name of a function to be applied, automatically converted to EXPR
MARGIN	ffapply only: the margins along which to loop in ffapply
X	an ff object from which several parameters can be derived, if they are not given directly: N, NCOL, NROW, DIM, VMODE, VBYTES, FF_RETURN
N	the total number of elements in the loop, e.g. number of elements in ffvecapply or number of rows in ffrowapply
NCOL	ffrowapply only: the number of columns needed to calculate batch sizes
NROW	ffcolapply only: the number of rows needed to calculate batch sizes
DIM	ffapply only: the dimension of the array needed to calculate batch sizes
VMODE	the vmode needed to prepare the RETURN object and to derive VBYTES if they are not given directly
VBYTES	the bytes per cell – see .rambytes – to calculate the RAM requirements per cell
BATCHBYTES	the max number of bytes per batch, default getOption("ffbatchbytes")
BATCHSIZE	an additional restriction on the number of loop elements, default=.Machine$integer.max
FROM	the name of the index that marks the beginning of the batch, default 'i1', change if needed to avoid naming-conflicts in the calling frame
TO	the name of the index that marks the end of the batch, default 'i2', change if needed to avoid naming-conflicts in the calling frame
IDIM	ffapply only: the name of an R variable used for loop-switching, change if needed to avoid naming-conflicts in the calling frame
BREAK	ffapply only: the name of an R object in the calling frame that triggers break out of the batch loop, if 1) it exists 2) is.logical and 3) is TRUE
RETURN	TRUE to prepare a return value (default FALSE)
CFUN	name of a collapsing function, see CFUN
RETCOL	NULL gives return vector[1:N], RETCOL gives return matrix[1:N, 1:RETCOL]
RETROW	NULL gives return vector[1:N], RETROW gives return matrix[1:RETROW, 1:N]
FF_RETURN	FALSE to return a ram object, TRUE to return an ff object, or an ff object that is ffsuitable to absorb the return data
USE.NAMES	FALSE to suppress attaching names or dimnames to the result
VERBOSE	TRUE to verbose the batches

Details

ffvecapply is the simplest ffapply method for ff_vectors. ffrowapply and ffcolapply is for ff_matrix, and ffapply is the most general method for ff_arrays and ff_vectors.
There are many ways to change the return value of the ffapply functions. In its simplest usage – batched looping over an expression – they don't return anything, see invisible. If you switch RETURN=TRUE in ffvecapply then it is assumed that all looped expressions together return one vector of length N, and via parameter FF_RETURN, you can decide whether this vector is in ram or is an ff object (or even which ff object to use). ffrowapply and ffcolapply additionally have parameter RETCOL resp. RETROW which defaults to returning a matrix of the original size; in order to just return a vector of length N set this to NULL, or specify a number of columns/rows for the return matrix. It is assumed that the expression will return appropriate pieces for this return structure (see examples). If you specify RETURN=TRUE and a collapsing function name CFUN, then it is assumed that the batched expressions return aggregated information, which is first collected in a list, and finally the collapsing function is called on this list: do.call(CFUN, list). If you want to return the unmodified list, you have to specify CFUN="list" for obvious reasons.
ffapply allows usages not completly unlike apply: you can specify the name of a function AFUN to be applied over MARGIN. However note that you must specify RETURN=TRUE in order to get a return value. Also note that currently ffapply assumes that your expression returns exactly one value per cell in DIM[MARGINS]. If you want to return something more complicated, you MUST specify a CFUN="list" and your return value will be a list with dim attribute DIM[MARGINS]. This means that for a function AFUN returning a scalar, ffapply behaves very similar to apply, see examples. Note also that ffapply might create a object named '.ffapply.dimexhausted' in its parent frame, and it uses a variable in the parent frame for loop-switching between dimensions, the default name 'idim' can be changed using the IDIM parameter. Finally you can break out of the implied loops by assigning TRUE to a variable with the name in BREAK.

Value

see details

Note

xx The complete generation of the return value is preliminary and the arguments related to defining the return value might still change, especially ffapply is work in progress

Author(s)

Jens Oehlschlägel

See Also

apply, expression, bbatch, repfromto, ffsuitable

Examples

   message("ffvecapply examples")
   x <- ff(vmode="integer", length=100)
   message("loop evaluate expression without returning anything")
   ffvecapply(x[i1:i2] <- i1:i2, X=x, VERBOSE=TRUE)
   ffvecapply(x[i1:i2] <- i1:i2, X=x, BATCHSIZE=20, VERBOSE=TRUE)
   ffvecapply(x[i1:i2] <- i1:i2, X=x, BATCHSIZE=19, VERBOSE=TRUE)
   message("lets return the combined expressions as a new ff object")
   ffvecapply(i1:i2, N=length(x), VMODE="integer", RETURN=TRUE, BATCHSIZE=20)
   message("lets return the combined expressions as a new ram object")
   ffvecapply(i1:i2, N=length(x), VMODE="integer", RETURN=TRUE, FF_RETURN=FALSE, BATCHSIZE=20)
   message("lets return the combined expressions in existing ff object x")
   x[] <- 0L
   ffvecapply(i1:i2, N=length(x), VMODE="integer", RETURN=TRUE, FF_RETURN=x, BATCHSIZE=20)
   x
   message("aggregate and collapse")
   ffvecapply(summary(x[i1:i2]), X=x, RETURN=TRUE, CFUN="list", BATCHSIZE=20)
   ffvecapply(summary(x[i1:i2]), X=x, RETURN=TRUE, CFUN="crbind", BATCHSIZE=20)
   ffvecapply(summary(x[i1:i2]), X=x, RETURN=TRUE, CFUN="cmean", BATCHSIZE=20)

   message("how to do colSums with ffrowapply")
   x <- ff(1:1000, vmode="integer", dim=c(100, 10))
   ffrowapply(colSums(x[i1:i2,,drop=FALSE]), X=x, RETURN=TRUE, CFUN="list", BATCHSIZE=20)
   ffrowapply(colSums(x[i1:i2,,drop=FALSE]), X=x, RETURN=TRUE, CFUN="crbind", BATCHSIZE=20)
   ffrowapply(colSums(x[i1:i2,,drop=FALSE]), X=x, RETURN=TRUE, CFUN="csum", BATCHSIZE=20)

   message("further ffrowapply examples")
   x <- ff(1:1000, vmode="integer", dim=c(100, 10))
   message("loop evaluate expression without returning anything")
   ffrowapply(x[i1:i2, ] <- i1:i2, X=x, BATCHSIZE=20)
   message("lets return the combined expressions as a new ff object (x unchanged)")
   ffrowapply(2*x[i1:i2, ], X=x, RETURN=TRUE, BATCHSIZE=20)
   message("lets return a single row aggregate")
   ffrowapply(t(apply(x[i1:i2,,drop=FALSE], 1, mean)), X=x, RETURN=TRUE, RETCOL=NULL, BATCHSIZE=20)
   message("lets return a 6 column aggregates")
   y <- ffrowapply( t(apply(x[i1:i2,,drop=FALSE], 1, summary)), X=x
   , RETURN=TRUE, RETCOL=length(summary(0)), BATCHSIZE=20)
   colnames(y) <- names(summary(0))
   y
   message("determine column minima if a complete column does not fit into RAM")
   ffrowapply(apply(x[i1:i2,], 2, min), X=x, RETURN=TRUE, CFUN="pmin", BATCHSIZE=20)

   message("ffapply examples")
   x <- ff(1:720, dim=c(8,9,10))
   dimnames(x) <- dummy.dimnames(x)
   message("apply function with scalar return value")
   apply(X=x[], MARGIN=3:2, FUN=sum)
   apply(X=x[], MARGIN=2:3, FUN=sum)
   ffapply(X=x, MARGIN=3:2, AFUN="sum", RETURN=TRUE, BATCHSIZE=8)
   message("this is what CFUN is based on")
   ffapply(X=x, MARGIN=2:3, AFUN="sum", RETURN=TRUE, CFUN="list", BATCHSIZE=8)

   message("apply functions with vector or array return value currently have limited support")
   apply(X=x[], MARGIN=3:2, FUN=summary)
   message("you must use CFUN, the rest is up to you")
   y <- ffapply(X=x, MARGIN=3:2, AFUN="summary", RETURN=TRUE, CFUN="list", BATCHSIZE=8)
   y
   y[[1]]

   rm(x); gc()

ff documentation built on April 4, 2025, 6:05 a.m.