Comparing split convolutions

split_convolutions_3x3.jl

using CuArrays
using Flux: Chain, Conv, gpu

batch = rand(256, 256, 1, 16) |> gpu

m1 = Chain(Conv((1,3), 1=>1), Conv((3,1), 1=>1)) |> gpu
m2 = Conv((3,3), 1=>1) |> gpu

using BenchmarkTools: @benchmark

@benchmark m1(batch) # ~ 333.1 us
@benchmark m2(batch) # ~ 271.8 us

split_convolutions_7x7.jl

using CuArrays
using Flux: Chain, Conv, gpu

batch = rand(256, 256, 1, 16) |> gpu

m1 = Chain(Conv((1,7), 1=>1), Conv((7,1), 1=>1)) |> gpu
m2 = Conv((7,7), 1=>1) |> gpu

using BenchmarkTools: @benchmark

@benchmark m1(batch) # ~ 383.5us
@benchmark m2(batch) # ~ 810.718us