Very slow loss aggregation

[MilesCranmer]

The mean and sum implementations in this package are extremely slow as they rely on generators. Could this be changed to a faster implementation? For example, here is how the current implementation would compute sum-squared-error:

square(x) = x ^ 2

# example of generator-based approach:
function sse(outputs, targets)
    sum(square(ŷ - y) for (ŷ, y) in zip(outputs, targets))
end

(i.e., like this code)

which gives us the following time:

julia> @btime sse(outputs, targets) setup=(outputs=randn(100_000); targets=randn(100_000))
  92.833 μs (0 allocations: 0 bytes)

but if we change this to an approach using sum(<function>, <indices>), it's much faster:

function sse2(outputs, targets)
    sum(i -> square(outputs[i] - targets[i]), eachindex(outputs, targets))
end

julia> @btime sse2(outputs, targets) setup=(outputs=randn(100_000); targets=randn(100_000))
  26.708 μs (0 allocations: 0 bytes)

which is a 3.5x speedup.

Could this be implemented as the default loss calculation? I thought this was the method that used to be used. Perhaps it got changed in the recent refactoring?

[juliohm]

I don't remember using this feature anywhere else in downstream packages. If you can submit a PR dropping support for generators, we can review and merge it.

Em qui., 24 de ago. de 2023 04:31, Miles Cranmer @.***> escreveu:

The mean and sum implementations in this package are extremely slow as they rely on generators. Could this be changed to a faster implementation? For example, here is how the current implementation would compute sum-squared-error:

square(x) = x ^ 2

example of generator-based approach:function sse(outputs, targets)

sum(square(ŷ - y) for (ŷ, y) in zip(outputs, targets))end

(i.e., like this code https://github.com/JuliaML/LossFunctions.jl/blob/7318c582874988fcc325464831aa2ce94eb36ff2/src/losses.jl#L37-L39 )

which gives us the following time:

julia> @btime sse(outputs, targets) setup=(outputs=randn(100_000); targets=randn(100_000)) 92.833 μs (0 allocations: 0 bytes)

but if we change this to an approach using sum(, ), it's much faster:

function sse2(outputs, targets) sum(i -> square(outputs[i] - targets[i]), eachindex(outputs, targets))end

julia> @btime sse2(outputs, targets) setup=(outputs=randn(100_000); targets=randn(100_000)) 26.708 μs (0 allocations: 0 bytes)

which is a 3.5x speedup.

Could this be implemented as the default loss calculation? I thought this was the method that used to be used. Perhaps it got changed in the recent refactoring?

— Reply to this email directly, view it on GitHub https://github.com/JuliaML/LossFunctions.jl/issues/172, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAZQW3KI56D2NUGNSS7FNQLXW37NRANCNFSM6AAAAAA34SKW4E . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[MilesCranmer]

I don't remember using this feature anywhere else in downstream packages.

Sorry I don't think I was clear. This is the main sum function in LossFunctions.jl which is the main interface to the package. It seems to be using a generator (e.g., sum(f(x) for x in X)) for summing, which is very slow:

https://github.com/JuliaML/LossFunctions.jl/blob/7318c582874988fcc325464831aa2ce94eb36ff2/src/losses.jl#L37-L39

If you can submit a PR dropping support for generators, we can review and merge it.

Sure!

[MilesCranmer]

Implemented in https://github.com/JuliaML/LossFunctions.jl/pull/173

[juliohm]

Yes, i meant using the version with generators in downstream packages that depend on LossFunctions.jl

Em qui., 24 de ago. de 2023 05:49, Miles Cranmer @.***> escreveu:

Implemented in #173 https://github.com/JuliaML/LossFunctions.jl/pull/173

— Reply to this email directly, view it on GitHub https://github.com/JuliaML/LossFunctions.jl/issues/172#issuecomment-1691274625, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAZQW3OTKEZIWOCK55VHWWTXW4IRZANCNFSM6AAAAAA34SKW4E . You are receiving this because you commented.Message ID: @.***>

[MilesCranmer]

Got it, thanks!