Generic way to apply function to a dimension of a matrix

New to Julia
1

Hello!
The mean function from statistics has the “dims” option, which lets me average a matrix into a vector by averaging the rows or columns.
Is there a generic way to do this without for loops? Say I want to do the same thing with the dot function, where I have a 1x3 array and a 10x3 matrix and I want to do the dot product of the array by each line of the matrix. Using dot.(a,b) wouldn’t really work as Julia doesn’t know which dims to use.
Thanks!

1 Like
2

This particular computation is exactly a matrix product, so just doing M * v' will be the best approach.

More generally, you can use the eachslice, eachrow, or eachcol iterators.

Is there a generic way to do this without for loops?

Note also that there’s nothing wrong with a loop in Julia. Loops are fast, and are sometimes the simplest and clearest way to code something.

6 Likes
3

Thanks, Steven.
You’re right, bad example. Say I want the norm of the 10x3 matrix as a 10 element array (i.e., norm of each line).
I’d like to know how to do it without for loops to be able to easily use CUDA.
Thanks!

4

Yep, combining eachrow with the . (say norm.(eachrow(matrix))) worked! Thanks!

5

It is a bit more efficient to do map(norm, eachrow(matrix)) or [norm(v) for v in eachrow(matrix)] instead of norm.(eachrow(matrix)) here — with the dot syntax, if you apply it to an iterator like eachrow(matrix) then it first runs collect (to form an explicit array of the row views).

6 Likes
6

Just be aware that eachrow iterate over Vectors not Arrays{T, 2} of dimensions 1xN, this confused a colleague of mine when he started using Julia.

7

Thanks, guys. This is all very helpful.
I tested norm.(eachrow(matrix)) and map(norm, eachrow(matrix)) and the latter is indeed much more efficient in terms of number of allocations (measured with @time).
Thanks again!

2 Likes