CODEWITHSUNDEEP
r
RSK
RSK

Reputation: 755

Rowwise matrix multiplication in R

I have a matrix with the dimension of 100 million records and 100 columns.

Now I want to multiply that matrix by rowwise.

My sample code for matrix multiplication is

df<-as.matrix(mtcars)
result<-apply(df,1,prod)

The above syntax is very slow in my case.

I tried rowprods function in Rfast package.

result<-rowprods(mtcars)

But the above function giving me space issues.

NOTE: I have 8 GB ram in my system.

Upvotes: 5

Views: 1862

Answers (4)

Michail
Michail

Reputation: 187

The Rfast command "rowprods" accepts a matrix, not a data.frame. Secondly, any row or colprods command will have numerical overflow errors. So ti best to use Rfast::colprods(x, method = "expsumlog").

Upvotes: 1

chinsoon12
chinsoon12

Reputation: 25223

Some timings for reference

library(matrixStats)
library(inline)
library(data.table)
#devtools::install_github("privefl/bigstatsr")
library(bigstatsr)
library(RcppArmadillo)
library(microbenchmark)
set.seed(20L)
N <- 1e6
dat <- matrix(rnorm(N*100),ncol=100)

fbm <- FBM(N, 100)
big_apply(fbm, a.FUN = function(X, ind) {
    print(min(ind))
    X[, ind] <- rnorm(nrow(X) * length(ind))
    NULL
}, a.combine = 'c')   

bigstatsrMtd <- function() {
    prods <- big_apply(fbm, a.FUN = function(X, ind) {
        print(min(ind))
        matrixStats::rowProds(X[ind, ])
    }, a.combine = 'c', ind = rows_along(fbm),
        block.size = 100e3, ncores = nb_cores())  
}

df <- data.table(as.data.frame(dat), keep.rownames=TRUE)
data.tableMtd <- function() {
    df[, rowprods:= Reduce("*", .SD), .SDcols = -1]
    df[, .(rn, rowprods)]    
}

code <- '
  arma::mat prodDat = Rcpp::as<arma::mat>(dat);
  int m = prodDat.n_rows;
  int n = prodDat.n_cols;
  arma::vec res(m);
  for (int row=0; row < m; row++) {
    res(row) = 1.0;
    for (int col=0; col < n; col++) {
      res(row) *= prodDat(row, col);
    }
  }
  return Rcpp::wrap(res);
'
rcppProd <- cxxfunction(signature(dat="numeric"), code, plugin="RcppArmadillo")

rcppMtd <- function() {
    rcppData <- rcppProd(dat)                # generated by C++ code
}

baseMtd <- function() {
    apply(dat, 1, prod)   
}

microbenchmark(bigstatsrMtd(),
    data.tableMtd(),
    rcppMtd(),
    baseMtd(),
    times=5L
)

Note: Compiling the function in cxxfunction seems to take some time

Here are the timing results:

# Unit: milliseconds
#            expr       min        lq      mean    median        uq       max
#  bigstatsrMtd() 4519.1861 4993.0879 5296.7000 5126.2282 5504.3981 6340.5995
# data.tableMtd()  443.1946  444.9686  690.3703  493.2399  513.4787 1556.9695
#       rcppMtd()  787.9488  799.1575  828.3647  809.0645  871.0347  874.6178
#       baseMtd() 5658.1424 6208.5123 6232.0040 6331.7431 6458.6806 6502.9417

Upvotes: 3

F. Priv&#233;
F. Priv&#233;

Reputation: 11738

If you have a matrix that is too large to fit in memory, you can use package bigstatsr (disclaimer: I'm the author) to use data stored on your disk (instead of the RAM). Using function big_apply enables you to apply standard R functions on data blocks (and to combine them).

library(bigstatsr)
fbm <- FBM(10e6, 100)
# inialize with random numbers
system.time(
  big_apply(fbm, a.FUN = function(X, ind) {
    print(min(ind))
    X[, ind] <- rnorm(nrow(X) * length(ind))
    NULL
  }, a.combine = 'c')
) # 78 sec

# compute row prods, possibly in parallel
system.time(
  prods <- big_apply(fbm, a.FUN = function(X, ind) {
    print(min(ind))
    matrixStats::rowProds(X[ind, ])
  }, a.combine = 'c', ind = rows_along(fbm),
  block.size = 100e3, ncores = nb_cores())  
) # 22 sec with 1 core and 18 sec with 6 cores

Upvotes: 6

Roland
Roland

Reputation: 132969

Try package data.table with Reduce. That might avoid internal copies of a 1e10 length vector.

library(data.table)
df <- data.table(df, keep.rownames=TRUE)
df[, rowprods:= Reduce("*", .SD), .SDcols = -1]
df[, .(rn, rowprods)]
#                     rn   rowprods
# 1:           Mazda RX4          0
# 2:       Mazda RX4 Wag          0
# 3:          Datsun 710  609055152
# 4:      Hornet 4 Drive          0
# 5:   Hornet Sportabout          0
# 6:             Valiant          0
# 7:          Duster 360          0
# 8:           Merc 240D          0
# 9:            Merc 230          0
#10:            Merc 280          0
#11:           Merc 280C          0
#12:          Merc 450SE          0
#13:          Merc 450SL          0
#14:         Merc 450SLC          0
#15:  Cadillac Fleetwood          0
#16: Lincoln Continental          0
#17:   Chrysler Imperial          0
#18:            Fiat 128  470578906
#19:         Honda Civic  564655046
#20:      Toyota Corolla  386281789
#21:       Toyota Corona          0
#22:    Dodge Challenger          0
#23:         AMC Javelin          0
#24:          Camaro Z28          0
#25:    Pontiac Firebird          0
#26:           Fiat X1-9  339825992
#27:       Porsche 914-2          0
#28:        Lotus Europa 1259677924
#29:      Ford Pantera L          0
#30:        Ferrari Dino          0
#31:       Maserati Bora          0
#32:          Volvo 142E 1919442833
#                     rn    rowsums

However, 8 GB RAM (minus what your OS and other software needs) is not much if you want to work with data of this size. R sometimes needs to make internal copies to use your data.

Upvotes: 3

Related Questions