Benchmark

[mmikhasenko]

The repository misses check of performance. Some tests should be added.

[mmikhasenko]

Benchmark 1: gauss on randn()

using AlgebraPDF
#
gauss = FGauss((μ=2.4, σ=1.3)) |> Normalized((-10.0, 10.0))
data = randn(10_000)
nll = NegativeLogLikelihood(gauss, data)

Now, calling the function with four different signatures

@btime nll(1.1)
@btime nll(1.1, [2.4, 1.3])
@btime nll(1.1; p=(μ=2.4, σ=1.3))
@btime nll(1.1; p=(σ=1.3, μ=2.4))

gives the same result and takes about 200us

  195.200 μs (135 allocations: 159.16 KiB)
  197.300 μs (144 allocations: 159.47 KiB)
  196.900 μs (135 allocations: 159.16 KiB)
  197.700 μs (135 allocations: 159.16 KiB)

[mmikhasenko]

Roughly same time when using FlaggedNamedTuple for parameters

gauss = FGauss(Ext(μ=2.4, σ=1.3)) |> Normalized((-10.0, 10.0))

  213.200 μs (165 allocations: 160.39 KiB)
  206.500 μs (175 allocations: 160.77 KiB)
  201.800 μs (156 allocations: 159.98 KiB)
  202.100 μs (156 allocations: 159.98 KiB)

interestingly, code_warntype indicated some problems of input arguments

MethodInstance for (::NegativeLogLikelihood{Normalized{FGauss{FlaggedNamedTuple{(:μ, :σ)}}, Tuple{Float64, Float64}}, Vector{Float64}})(::Float64)
  from (d::AbstractFunctionWithParameters)(x; p) in AlgebraPDF at Documents\AlgebraPDF.jl\src\functionwithparameters.jl:15
Arguments
  d::NegativeLogLikelihood{Normalized{FGauss{FlaggedNamedTuple{(:μ, :σ)}}, Tuple{Float64, Float64}}, Vector{Float64}}
  x::Float64
Body::Any
1 ─ %1 = AlgebraPDF.freepars(d)::Any
│   %2 = AlgebraPDF.:(var"#_#4")(%1, d, x)::Any
└──      return %2