Julia: Generic reducedim

Created on 21 Mar 2019 · 6Comments · Source: JuliaLang/julia

julia> add(a, b) = a+b
add (generic function with 1 method)

julia> reduce(add, rand(5,5), dims = (1, 2))
ERROR: MethodError: no method matching reducedim_init(::typeof(identity), ::typeof(add), ::Array{Float64,2}, ::Tuple{Int64,Int64})

It seems like Julia should just do something sensible and generic here, like map does already – is there a specific reason not too, or is it just unimplemented?

arrays

Source

MikeInnes

Most helpful comment

Yes, it should behave identically to how it does with +:

julia> reduce(+, rand(5,5), dims = (1, 2))
1×1 Array{Float64,2}:
 12.549021410347184

MikeInnes on 21 Mar 2019

👍7

All 6 comments

What are you expecting the generic reducedim to return? A n-dimensional array, with reduce operation on each of the dims ?

eulerkochy on 21 Mar 2019

Yes, it should behave identically to how it does with +:

julia> reduce(+, rand(5,5), dims = (1, 2))
1×1 Array{Float64,2}:
 12.549021410347184

MikeInnes on 21 Mar 2019

👍7

FWIW I just came across this as well, trying to "flatten" a 2D array of 2D arrays into a single larger 2D array.

As = [rand(5, 4) for _ in 1:4, _ in 1:5]
reduce(hcat, reduce(vcat, As, dims=1))

ssfrr on 21 Mar 2019

Bump, we were hit by this.

AzamatB on 24 Oct 2019

I think I'm being hit by this, https://github.com/FluxML/Zygote.jl/issues/594

cossio on 9 Apr 2020

A "sensible and generic" default might be to pass the first two elements along a dimension to the operator and then iterate over the remaining items. But what do you do if a dimension has size 1?

Conversely, adding the suggested reducedim_init method makes it easy to support dimensions of size 1. So my question is, how big of a problem is this really?