Rosenbrock example: Type instability

[renatobellotti]

Hi,

I'm profiling one of my optimisation codes and noticed that a lot of time is spent in typeinf(), so I wondered whether there might be some type instability. Running the Rosenbrock example displays indeed some warnings:

using Optim

rosenbrock(x) =  (1.0 - x[1])^2 + 100.0 * (x[2] - x[1]^2)^2

@code_warntype optimize(rosenbrock, zeros(2), BFGS())

MethodInstance for Optim.optimize(::typeof(rosenbrock), ::Vector{Float64}, ::BFGS{LineSearches.InitialStatic{Float64}, LineSearches.HagerZhang{Float64, Base.RefValue{Bool}}, Nothing, Nothing, Flat})
  from optimize(f, initial_x::AbstractArray, method::Optim.AbstractOptimizer) in Optim at /data/user/bellotti_r/julia_packages/packages/Optim/Zq1jM/src/multivariate/optimize/interface.jl:138
Arguments
  #self#::Core.Const(Optim.optimize)
  f::Core.Const(rosenbrock)
  initial_x::Vector{Float64}
  method::BFGS{LineSearches.InitialStatic{Float64}, LineSearches.HagerZhang{Float64, Base.RefValue{Bool}}, Nothing, Nothing, Flat}
Locals
  @_5::Optim.Options
Body::Optim.MultivariateOptimizationResults{BFGS{LineSearches.InitialStatic{Float64}, LineSearches.HagerZhang{Float64, Base.RefValue{Bool}}, Nothing, Nothing, Flat}, Vector{Float64}, Float64, Float64, Vector{OptimizationState{Float64, BFGS{LineSearches.InitialStatic{Float64}, LineSearches.HagerZhang{Float64, Base.RefValue{Bool}}, Nothing, Nothing, Flat}}}, Bool}
1 ─ %1  = Base.NamedTuple()::Core.Const(NamedTuple())
│   %2  = Optim.default_options(method)::Dict{Symbol, Any}
│   %3  = Base.merge(%1, %2)::NamedTuple
│   %4  = Base.isempty(%3)::Bool
└──       goto #3 if not %4
2 ─       (@_5 = Optim.Options())
└──       goto #4
3 ─ %8  = Core.kwfunc(Optim.Options)::Core.Const(Core.var"#Type##kw"())
└──       (@_5 = (%8)(%3, Optim.Options))
4 ┄ %10 = @_5::Optim.Options
│   %11 = (#self#)(f, initial_x, method, %10)::Optim.MultivariateOptimizationResults{BFGS{LineSearches.InitialStatic{Float64}, LineSearches.HagerZhang{Float64, Base.RefValue{Bool}}, Nothing, Nothing, Flat}, Vector{Float64}, Float64, Float64, Vector{OptimizationState{Float64, BFGS{LineSearches.InitialStatic{Float64}, LineSearches.HagerZhang{Float64, Base.RefValue{Bool}}, Nothing, Nothing, Flat}}}, Bool}
└──       return %11

Is this relevant for performance?

[pkofod]

Seems to go away if you manually input the options

[renatobellotti]

I think something went wrong with the image upload...

[pkofod]

I think something went wrong with the image upload...

Lol, I see. It was something like optimize(f, x, method, Optim.Options()) so it has to do with the promotion of the tolerances or something like that I'd guess.

[pkofod]

[pkofod]

That wasn't it. Seems to be a splatting of a dict with Any typed keys..