Chapter 3 Working with an FBM

3.1 Similar accessor as R matrices

library(bigstatsr)
X <- FBM(2, 5, init = 1:10, backingfile = "test")$save()

X$backingfile                ## the file where the data is actually stored

#> [1] "C:\\Users\\au639593\\OneDrive - Aarhus universitet\\Desktop\\bigsnpr-extdoc\\test.bk"

X <- big_attach("test.rds")  ## can get the FBM from any R session

You can access the whole FBM as an R matrix in memory using X[]. However, if the matrix is too large to fit in memory, you should always access only a subset of columns. Note that the elements of the FBM are stored column-wise (as for a standard R matrix). Therefore, be careful not to access a subset of rows, since it would read non-contiguous elements from the whole matrix from disk.

X[, 1]  ## ok (must read first column only)

#> [1] 1 2

X[1, ]  ## bad (must read all data from disk)

#> [1] 1 3 5 7 9

X[]     ## super bad (standard R matrix in memory)

#>      [,1] [,2] [,3] [,4] [,5]
#> [1,]    1    3    5    7    9
#> [2,]    2    4    6    8   10

3.2 Split-(par)Apply-Combine Strategy

colSums(X[])  ## super bad

#> [1]  3  7 11 15 19

---

How to apply standard R functions to big matrices (in parallel); implemented in big_apply().

Learn more with this tutorial on big_apply().

Compute the sum of each column of X <- big_attachExtdata() using big_apply().

3.3 Similar accessor as Rcpp matrices

// [[Rcpp::plugins(cpp11)]]
// [[Rcpp::depends(bigstatsr, rmio)]]
#include <bigstatsr/BMAcc.h>

// [[Rcpp::export]]
NumericVector bigcolsums(Environment BM) {
  
  XPtr<FBM> xpBM = BM["address"];   // get the external pointer
  BMAcc<double> macc(xpBM);         // create an accessor to the data
  
  size_t n = macc.nrow();           // similar code as for an Rcpp::NumericMatrix
  size_t m = macc.ncol();           // similar code as for an Rcpp::NumericMatrix
  
  NumericVector res(m);
  
  for (size_t j = 0; j < m; j++) 
    for (size_t i = 0; i < n; i++)
      res[j] += macc(i, j);         // similar code as for an Rcpp::NumericMatrix
  
  return res;
}

For a subset of the data:

// [[Rcpp::plugins(cpp11)]]
// [[Rcpp::depends(bigstatsr, rmio)]]
#include <bigstatsr/BMAcc.h>

// [[Rcpp::export]]
NumericVector bigcolsums2(Environment BM,
                          const IntegerVector& rowInd,
                          const IntegerVector& colInd) {
  
  XPtr<FBM> xpBM = BM["address"];   
  // accessor to a sub-view of the data -> the only line of code that should change
  SubBMAcc<double> macc(xpBM, rowInd, colInd, 1);  
  
  size_t n = macc.nrow();
  size_t m = macc.ncol();
  
  NumericVector res(m);
  
  for (size_t j = 0; j < m; j++) 
    for (size_t i = 0; i < n; i++)
      res[j] += macc(i, j); 
  
  return res;
}

3.4 Some summary functions are already implemented

big_colstats(X)  # sum and var (for each column)

#>   sum var
#> 1   3 0.5
#> 2   7 0.5
#> 3  11 0.5
#> 4  15 0.5
#> 5  19 0.5

big_scale()(X)   # mean and sd (for each column)

#>   center     scale
#> 1    1.5 0.7071068
#> 2    3.5 0.7071068
#> 3    5.5 0.7071068
#> 4    7.5 0.7071068
#> 5    9.5 0.7071068

To only use a subset of the data stored as an FBM, you should almost never make a copy of the data; instead, use parameters ind.row (or ind.train) and ind.col to apply functions to a subset of the data.