Authors: Koki Tsuyuzaki [aut, cre]
Last modified: 2023-10-24 14:41:21.735014
Compiled: Tue Oct 24 16:56:13 2023

1 Setting

suppressPackageStartupMessages(library("DelayedTensor"))
suppressPackageStartupMessages(library("DelayedArray"))
suppressPackageStartupMessages(library("HDF5Array"))
suppressPackageStartupMessages(library("DelayedRandomArray"))

darr1 <- RandomUnifArray(c(2,3,4))
darr2 <- RandomUnifArray(c(2,3,4))

There are several settings in DelayedTensor.

First, the sparsity of the intermediate DelayedArray objects calculated inside DelayedTensor is set by setSparse.

Note that the sparse mode is experimental.

Whether it contributes to higher speed and lower memory is quite dependent on the sparsity of the DelayedArray, and the current implementation does not recognize the block size, which may cause out-of-memory errors, when the data is extremely huge.

Here, we specify as.sparse as FALSE (this is also the default value for now).

DelayedTensor::setSparse(as.sparse=FALSE)

Next, the verbose message is suppressed by setVerbose. This is useful when we want to monitor the calculation process.

Here we specify as.verbose as FALSE (this is also the default value for now).

DelayedTensor::setVerbose(as.verbose=FALSE)

The block size of block processing is specified by setAutoBlockSize. When the sparse mode is off, all the functions of DelayedTensor are performed as block processing, in which each block vector/matrix/tensor is expanded to memory space from on-disk file incrementally so as not to exceed the specified size.

Here, we specify the block size as 1E+8.

setAutoBlockSize(size=1E+8)
## automatic block size set to 1e+08 bytes (was 1e+08)

Finally, the temporal directory to store the intermediate HDF5 files during running DelayedTensor is specified by setHDF5DumpDir.

Note that in many systems the /var directory has the storage limitation, so if there is no enough space, user should specify the other directory.

# tmpdir <- paste(sample(c(letters,1:9), 10), collapse="")
# dir.create(tmpdir, recursive=TRUE))
tmpdir <- tempdir()
setHDF5DumpDir(tmpdir)

These specified values are also extracted by each getter function.

DelayedTensor::getSparse()
## $delayedtensor.sparse
## [1] FALSE
DelayedTensor::getVerbose()
## $delayedtensor.verbose
## [1] FALSE
getAutoBlockSize()
## [1] 1e+08
getHDF5DumpDir()
## [1] "/tmp/RtmpNtbkDZ"

2 Tensor Arithmetic Operations

2.1 Unfold/Fold Operations

Unfold (a.k.a. matricizing) operations are used to reshape a tensor into a matrix.

Figure 1: Unfold/Fold Operasions

In unfold, row_idx and col_idx are specified to set which modes are used as the row/column.

dmat1 <- DelayedTensor::unfold(darr1, row_idx=c(1,2), col_idx=3)
dmat1
## <6 x 4> HDF5Matrix object of type "double":
##            [,1]       [,2]       [,3]       [,4]
## [1,] 0.14201871 0.74171836 0.89457361 0.22288631
## [2,] 0.36894753 0.61222268 0.50908460 0.56925740
## [3,] 0.03697118 0.65646913 0.83733904 0.47793673
## [4,] 0.56545517 0.43655634 0.70188635 0.59687095
## [5,] 0.92319873 0.35864518 0.28317491 0.38453918
## [6,] 0.30437558 0.53211447 0.01556021 0.36664361

fold is the inverse operation of unfold, which is used to reshape a matrix into a tensor.

In fold, row_idx/col_idx are specified to set which modes correspond the row/column of the output tensor and modes is specified to set the mode of the output tensor.

dmat1_to_darr1 <- DelayedTensor::fold(dmat1,
    row_idx=c(1,2), col_idx=3, modes=dim(darr1))
dmat1_to_darr1
## <2 x 3 x 4> DelayedArray object of type "double":
## ,,1
##            [,1]       [,2]       [,3]
## [1,] 0.14201871 0.03697118 0.92319873
## [2,] 0.36894753 0.56545517 0.30437558
## 
## ,,2
##           [,1]      [,2]      [,3]
## [1,] 0.7417184 0.6564691 0.3586452
## [2,] 0.6122227 0.4365563 0.5321145
## 
## ,,3
##            [,1]       [,2]       [,3]
## [1,] 0.89457361 0.83733904 0.28317491
## [2,] 0.50908460 0.70188635 0.01556021
## 
## ,,4
##           [,1]      [,2]      [,3]
## [1,] 0.2228863 0.4779367 0.3845392
## [2,] 0.5692574 0.5968710 0.3666436
identical(as.array(darr1), as.array(dmat1_to_darr1))
## [1] TRUE

There are some wrapper functions of unfold and fold.

For example, in k_unfold, mode m is used as the row, and the other modes are is used as the column.

k_fold is the inverse operation of k_unfold.

dmat2 <- DelayedTensor::k_unfold(darr1, m=1)
dmat2_to_darr1 <- k_fold(dmat2, m=1, modes=dim(darr1))
identical(as.array(darr1), as.array(dmat2_to_darr1))
## [1] TRUE
dmat3 <- DelayedTensor::k_unfold(darr1, m=2)
dmat3_to_darr1 <- k_fold(dmat3, m=2, modes=dim(darr1))
identical(as.array(darr1), as.array(dmat3_to_darr1))
## [1] TRUE
dmat4 <- DelayedTensor::k_unfold(darr1, m=3)
dmat4_to_darr1 <- k_fold(dmat4, m=3, modes=dim(darr1))
identical(as.array(darr1), as.array(dmat4_to_darr1))
## [1] TRUE

In rs_unfold, mode m is used as the row, and the other modes are is used as the column.

rs_fold and rs_unfold also perform the same operations.

On the other hand, cs_unfold specifies the mode m as the column and the other modes are specified as the column.

cs_fold is the inverse operation of cs_unfold.

dmat8 <- DelayedTensor::cs_unfold(darr1, m=1)
dmat8_to_darr1 <- DelayedTensor::cs_fold(dmat8, m=1, modes=dim(darr1))
identical(as.array(darr1), as.array(dmat8_to_darr1))
## [1] TRUE
dmat9 <- DelayedTensor::cs_unfold(darr1, m=2)
dmat9_to_darr1 <- DelayedTensor::cs_fold(dmat9, m=2, modes=dim(darr1))
identical(as.array(darr1), as.array(dmat9_to_darr1))
## [1] TRUE
dmat10 <- DelayedTensor::cs_unfold(darr1, m=3)
dmat10_to_darr1 <- DelayedTensor::cs_fold(dmat10, m=3, modes=dim(darr1))
identical(as.array(darr1), as.array(dmat10_to_darr1))
## [1] TRUE

In matvec, m=2 is specified as unfold.

unmatvec is the inverse operation of matvec.

dmat11 <- DelayedTensor::matvec(darr1)
dmat11_darr1 <- DelayedTensor::unmatvec(dmat11, modes=dim(darr1))
identical(as.array(darr1), as.array(dmat11_darr1))
## [1] TRUE

ttm multiplies a tensor by a matrix.

m specifies in which mode the matrix will be multiplied.

dmatZ <- RandomUnifArray(c(10,4))
DelayedTensor::ttm(darr1, dmatZ, m=3)
## <2 x 3 x 10> DelayedArray object of type "double":
## ,,1
##          [,1]     [,2]     [,3]
## [1,] 1.762417 1.762341 1.783421
## [2,] 1.852990 2.043180 1.143807
## 
## ,,2
##           [,1]      [,2]      [,3]
## [1,] 1.1389921 1.0942043 1.3704663
## [2,] 1.2973165 1.3599547 0.9420316
## 
## ,,3
##           [,1]      [,2]      [,3]
## [1,] 1.4108541 1.3425973 1.2340361
## [2,] 1.2877555 1.4934812 0.6834915
## 
## ...
## 
## ,,8
##           [,1]      [,2]      [,3]
## [1,] 0.9122185 1.1025346 0.7308660
## [2,] 1.0006646 1.1934765 0.4297805
## 
## ,,9
##           [,1]      [,2]      [,3]
## [1,] 1.0394879 0.9587450 1.2648265
## [2,] 1.1345717 1.2275962 0.8009346
## 
## ,,10
##           [,1]      [,2]      [,3]
## [1,] 0.3717713 0.4356210 0.6050505
## [2,] 0.5576262 0.6335131 0.3729372

ttl multiplies a tensor by multiple matrices.

ms specifies in which mode these matrices will be multiplied.

dmatX <- RandomUnifArray(c(10,2))
dmatY <- RandomUnifArray(c(10,3))
dlizt <- list(dmatX = dmatX, dmatY = dmatY)
DelayedTensor::ttl(darr1, dlizt, ms=c(1,2))
## <10 x 10 x 4> DelayedArray object of type "double":
## ,,1
##            [,1]      [,2]      [,3] ...      [,9]     [,10]
##  [1,] 0.3030929 0.9363495 0.7489267   . 0.3584228 0.7491084
##  [2,] 0.4764815 1.3444037 1.0762268   . 0.3804979 1.0374256
##   ...         .         .         .   .         .         .
##  [9,] 0.4486242 1.3743412 1.0993325   . 0.5138868 1.0960479
## [10,] 0.2032797 0.6180891 0.4944421   . 0.2261841 0.4915294
## 
## ...
## 
## ,,4
##            [,1]      [,2]      [,3] ...      [,9]     [,10]
##  [1,] 0.2835593 0.9837275 0.8079973   . 0.4845977 0.8058227
##  [2,] 0.3566083 1.2210540 0.9633387   . 0.5958302 1.0201492
##   ...         .         .         .   .         .         .
##  [9,] 0.4116051 1.4264823 1.1680603   . 0.7021886 1.1703174
## [10,] 0.1832565 0.6345062 0.5180863   . 0.3121257 0.5213038

2.2 Vectorization

vec collapses a DelayedArray into a 1D DelayedArray (vector).

Figure 2: Vectorization

DelayedTensor::vec(darr1)
## <24> HDF5Array object of type "double":
##        [1]        [2]        [3]          .       [23]       [24] 
## 0.14201871 0.36894753 0.03697118          .  0.3845392  0.3666436

2.3 Norm Operations

fnorm calculates the Frobenius norm of a DelayedArray.

Figure 3: Norm Operations

DelayedTensor::fnorm(darr1)
## [1] 2.637225

innerProd calculates the inner product value of two DelayedArray.

DelayedTensor::innerProd(darr1, darr2)
## [1] 6.776278

2.4 Outer Product

Inner product multiplies two tensors and collapses to 0D tensor (norm). On the other hand, the outer product is an operation that leaves all subscripts intact.