Dispatch on solve_model (#108)

* Dispatch on `solve_model`

docs/src/doityourself.md

@@ -14,8 +14,8 @@ AdaptiveRegularization.ALL_solvers
 
 To make your own variant we need to implement:
 - A new data structure `<: PData{T}` for some real number type `T`.
-- A `preprocess(PData::TPData, H, g, gNorm2, α)` function called before each trust-region iteration.
-- A `solve_model(PData::PDataST, H, g, gNorm2, n1, n2, δ::T)` function used to solve the algorithm subproblem.
+- A `preprocess!(PData::TPData, H, g, gNorm2, α)` function called before each trust-region iteration.
+- A `solve_model!(PData::TPData, H, g, gNorm2, n1, n2, δ::T)` function used to solve the algorithm subproblem.
 
 In the rest of this tutorial, we implement a Steihaug-Toint trust-region method using `cg_lanczos` from [`Krylov.jl`](https://github.com/JuliaSmoothOptimizers/Krylov.jl) to solve the linear subproblem with trust-region constraint.
 
@@ -57,13 +57,13 @@ end
 ```
 For our Steihaug-Toint implementation, we do not run any preprocess operation, so we use the default one.
 ```@example 1
-function AdaptiveRegularization.preprocess(PData::AdaptiveRegularization.TPData, H, g, gNorm2, n1, n2, α)
+function AdaptiveRegularization.preprocess!(PData::AdaptiveRegularization.TPData, H, g, gNorm2, n1, n2, α)
     return PData
 end
 ```
 We now solve the subproblem.
 ```@example 1
-function solve_modelST_TR(PData::PDataST, H, g, gNorm2, calls, max_calls, δ::T) where {T}
+function AdaptiveRegularization.solve_model!(PData::PDataST, H, g, gNorm2, calls, max_calls, δ::T) where {T}
     ζ, ξ, maxtol, mintol = PData.ζ, PData.ξ, PData.maxtol, PData.mintol
     n = length(g)
     # precision = max(1e-12, min(0.5, (gNorm2^ζ)))
@@ -92,7 +92,7 @@ end
 
 We can now proceed with the main solver call specifying the used `pdata_type` and `solve_model`. Since, `Krylov.cg_lanczos` only uses matrix-vector products, it is sufficient to evaluate the Hessian matrix as an operator, so we provide `hess_type = HessOp`.
 ```@example 1
-ST_TROp(nlp; kwargs...) = TRARC(nlp, pdata_type = PDataST, solve_model = solve_modelST_TR, hess_type = HessOp; kwargs...)
+ST_TROp(nlp; kwargs...) = TRARC(nlp, pdata_type = PDataST, hess_type = HessOp; kwargs...)
 ```
 Finally, we can apply our new method to any [`NLPModels`](https://github.com/JuliaSmoothOptimizers/NLPModels.jl).
 ```@example 1

src/AdaptiveRegularization.jl

@@ -46,7 +46,6 @@ The keyword arguments include
 - `TR::TrustRegion`: structure with trust-region/ARC parameters, see [`TrustRegion`](@ref). Default: `TrustRegion(T(10.0))`.
 - `hess_type::Type{Hess}`: Structure used to handle the hessian. The possible values are: `HessDense`, `HessSparse`, `HessSparseCOO`, `HessOp`. Default: `HessOp`.
 - `pdata_type::Type{ParamData}` Structure used for the preprocessing step. Default: `PDataKARC`.
-- `solve_model::Function` Function used to solve the subproblem. Default: `solve_modelKARC`.
 - `robust::Bool`: `true` implements a robust evaluation of the model. Default: `true`.
 - `verbose::Bool`: `true` prints iteration information. Default: `false`.
 Additional `kwargs` are used for stopping criterion, see `Stopping.jl`.
@@ -127,7 +126,7 @@ function TRARC(nlp_stop::NLPStopping; kwargs...)
 end
 
 for fun in union(keys(solvers_const), keys(solvers_nls_const))
-  ht, pt, sm, ka = merge(solvers_const, solvers_nls_const)[fun]
+  ht, pt, ka = merge(solvers_const, solvers_nls_const)[fun]
   @eval begin
     function $fun(nlpstop::NLPStopping; kwargs...)
       kw_list = Dict{Symbol, Any}()
@@ -137,7 +136,7 @@ for fun in union(keys(solvers_const), keys(solvers_nls_const))
         end
       end
       merge!(kw_list, Dict(kwargs))
-      TRARC(nlpstop; hess_type = $ht, pdata_type = $pt, solve_model = $sm, kw_list...)
+      TRARC(nlpstop; hess_type = $ht, pdata_type = $pt, kw_list...)
     end
   end
   @eval begin

src/SolveModel/SolveModelKARC.jl

@@ -1,4 +1,4 @@
-function solve_modelKARC(X::PDataKARC, H, g, gNorm2, calls, max_calls, α::T) where {T}
+function solve_model!(X::PDataKARC{S, T, Fatol, Frtol}, H, g, gNorm2, calls, max_calls, α::T) where {S, T, Fatol, Frtol}
   # target value should be close to satisfy αλ=||d||
   start = findfirst(X.positives)
   if isnothing(start)

src/SolveModel/SolveModelNLSST_TR.jl

@@ -1,4 +1,4 @@
-function solve_modelNLSST_TR(PData::PDataNLSST, Jx, Fx, norm_∇f, calls, max_calls, δ::T) where {T}
+function solve_model!(PData::PDataNLSST{S, T, Fatol, Frtol}, Jx, Fx, norm_∇f, calls, max_calls, δ::T) where {S, T, Fatol, Frtol}
   # cas particulier Steihaug-Toint
   # ϵ = sqrt(eps(T)) # * 100.0 # old
   # cgtol = max(ϵ, min(cgtol, 9 * cgtol / 10, 0.01 * norm(g)^(1.0 + PData.ζ))) # old

src/SolveModel/SolveModelST_TR.jl

@@ -1,4 +1,4 @@
-function solve_modelST_TR(PData::PDataST, H, g, gNorm2, calls, max_calls, δ::T) where {T}
+function solve_model!(PData::PDataST{S, T, Fatol, Frtol}, H, g, gNorm2, calls, max_calls, δ::T) where {S, T, Fatol, Frtol}
   # cas particulier Steihaug-Toint
   # ϵ = sqrt(eps(T)) # * 100.0 # old
   # cgtol = max(ϵ, min(cgtol, 9 * cgtol / 10, 0.01 * norm(g)^(1.0 + PData.ζ))) # old

src/SolveModel/SolveModelTRK.jl

@@ -1,4 +1,4 @@
-function solve_modelTRK(X::PDataTRK, H, g, gNorm2, calls, max_calls, α::T) where {T}
+function solve_model!(X::PDataTRK{S, T, Fatol, Frtol}, H, g, gNorm2, calls, max_calls, α::T) where {S, T, Fatol, Frtol}
   # target value should be close to satisfy α=||d||
   start = findfirst(X.positives)
   if isnothing(start)

src/main.jl

@@ -19,11 +19,11 @@ function compute_direction(
   ∇f,
   norm_∇f,
   α,
-  solve_model,
 ) where {T}
   max_hprod = stp.meta.max_cntrs[:neval_hprod]
   Hx = workspace.Hstruct.H
-  return solve_model(PData, Hx, ∇f, norm_∇f, neval_hprod(stp.pb), max_hprod, α)
+  solve_model!(PData, Hx, ∇f, norm_∇f, neval_hprod(stp.pb), max_hprod, α)
+  return PData.d, PData.λ
 end
 
 function compute_direction(
@@ -33,12 +33,12 @@ function compute_direction(
   ∇f,
   norm_∇f,
   α,
-  solve_model,
 )
   max_prod = stp.meta.max_cntrs[:neval_jprod_residual]
   Jx = jac_op_residual(stp.pb, workspace.xt)
   Fx = workspace.Fx
-  return solve_model(PData, Jx, Fx, norm_∇f, neval_jprod_residual(stp.pb), max_prod, α)
+  solve_model!(PData, Jx, Fx, norm_∇f, neval_jprod_residual(stp.pb), max_prod, α)
+  return PData.d, PData.λ
 end
 
 function compute_direction(
@@ -48,12 +48,12 @@ function compute_direction(
   ∇f,
   norm_∇f,
   α,
-  solve_model,
 ) where {T, S, Hess <: HessGaussNewtonOp}
   max_prod = stp.meta.max_cntrs[:neval_jprod_residual]
   Jx = jac_op_residual!(stp.pb, workspace.xt, workspace.Hstruct.Jv, workspace.Hstruct.Jtv)
   Fx = workspace.Fx
-  return solve_model(PData, Jx, Fx, norm_∇f, neval_jprod_residual(stp.pb), max_prod, α)
+  solve_model!(PData, Jx, Fx, norm_∇f, neval_jprod_residual(stp.pb), max_prod, α)
+  return PData.d, PData.λ
 end
 
 function hessian!(workspace::TRARCWorkspace, nlp, x)
@@ -100,7 +100,6 @@ function SolverCore.solve!(
   solver::TRARCSolver{T, S},
   nlp_stop::NLPStopping{Pb, M, SRC, NLPAtX{Score, T, S}, MStp, LoS},
   stats::GenericExecutionStats{T, S};
-  solve_model::Function = solve_modelKARC,
   robust::Bool = true,
   verbose::Integer = false,
   kwargs...,
@@ -156,7 +155,7 @@ function SolverCore.solve!(
 
     success = false
     while !success & (unsuccinarow < max_unsuccinarow)
-      d, λ = compute_direction(nlp_stop, PData, workspace, ∇f, norm_∇f, α, solve_model)
+      d, λ = compute_direction(nlp_stop, PData, workspace, ∇f, norm_∇f, α)
 
       slope = ∇f ⋅ d
       Δq = compute_Δq(workspace, Hx, d, ∇f)

src/solvers.jl

@@ -1,31 +1,30 @@
 const solvers_const = Dict(
   :ARCqKdense =>
-    (HessDense, PDataKARC, solve_modelKARC, [(:shifts => 10.0 .^ (collect(-20.0:1.0:20.0)))]),
+    (HessDense, PDataKARC, [(:shifts => 10.0 .^ (collect(-20.0:1.0:20.0)))]),
   :ARCqKOp =>
-    (HessOp, PDataKARC, solve_modelKARC, [:shifts => 10.0 .^ (collect(-20.0:1.0:20.0))]),
+    (HessOp, PDataKARC, [:shifts => 10.0 .^ (collect(-20.0:1.0:20.0))]),
   :ARCqKsparse =>
-    (HessSparse, PDataKARC, solve_modelKARC, [:shifts => 10.0 .^ (collect(-20.0:1.0:20.0))]),
+    (HessSparse, PDataKARC, [:shifts => 10.0 .^ (collect(-20.0:1.0:20.0))]),
   :ARCqKCOO =>
-    (HessSparseCOO, PDataKARC, solve_modelKARC, [:shifts => 10.0 .^ (collect(-20.0:1.0:20.0))]),
-  :ST_TRdense => (HessDense, PDataST, solve_modelST_TR, ()),
-  :ST_TROp => (HessOp, PDataST, solve_modelST_TR, ()),
-  :ST_TRsparse => (HessSparse, PDataST, solve_modelST_TR, ()),
-  :TRKdense => (HessDense, PDataTRK, solve_modelTRK, ()),
-  :TRKOp => (HessOp, PDataTRK, solve_modelTRK, ()),
-  :TRKsparse => (HessSparse, PDataTRK, solve_modelTRK, ()),
+    (HessSparseCOO, PDataKARC, [:shifts => 10.0 .^ (collect(-20.0:1.0:20.0))]),
+  :ST_TRdense => (HessDense, PDataST, ()),
+  :ST_TROp => (HessOp, PDataST, ()),
+  :ST_TRsparse => (HessSparse, PDataST, ()),
+  :TRKdense => (HessDense, PDataTRK, ()),
+  :TRKOp => (HessOp, PDataTRK, ()),
+  :TRKsparse => (HessSparse, PDataTRK, ()),
 )
 
 const solvers_nls_const = Dict(
   :ARCqKOpGN => (
     HessGaussNewtonOp,
     PDataKARC,
-    solve_modelKARC,
     [:shifts => 10.0 .^ (collect(-10.0:0.5:20.0))],
   ),
-  :ST_TROpGN => (HessGaussNewtonOp, PDataST, solve_modelST_TR, ()),
+  :ST_TROpGN => (HessGaussNewtonOp, PDataST, ()),
   :ST_TROpGNLSCgls =>
-    (HessGaussNewtonOp, PDataNLSST, solve_modelNLSST_TR, [:solver_method => :cgls]),
+    (HessGaussNewtonOp, PDataNLSST, [:solver_method => :cgls]),
   :ST_TROpGNLSLsqr =>
-    (HessGaussNewtonOp, PDataNLSST, solve_modelNLSST_TR, [:solver_method => :lsqr]),
-  :ST_TROpLS => (HessOp, PDataNLSST, solve_modelNLSST_TR, ()),
+    (HessGaussNewtonOp, PDataNLSST, [:solver_method => :lsqr]),
+  :ST_TROpLS => (HessOp, PDataNLSST, ()),
 )

src/utils/pdata_struct.jl

@@ -28,9 +28,10 @@ function preprocess!(PData::TPData{T}, H, g, gNorm2, n1, n2, α) where {T}
 end
 
 """
-    solve_model(PData::TPData, H, g, gNorm2, n1, n2, α)
+    solve_model!(PData::TPData, H, g, gNorm2, n1, n2, α)
 
 Function called in the `TRARC` algorithm to solve the subproblem.
+
 # Arguments
 - `PData::TPData`: data structure used for preprocessing.
 - `H`: current Hessian matrix.
@@ -40,9 +41,9 @@ Function called in the `TRARC` algorithm to solve the subproblem.
 - `n2`: Maximum number of Hessian-vector products accepted.
 - `α`: current value of the TR/ARC parameter.
 
-It returns a couple `(PData.d, PData.λ)`. Current implementations include: `solve_modelKARC`, `solve_modelTRK`, `solve_modelST_TR`.
+It returns a couple `(PData.d, PData.λ)`.
 """
-function solve_model(X::TPData{T}, H, g, gNorm2, n1, n2, α) where {T} end
+function solve_model!(X::TPData{T}, H, g, gNorm2, n1, n2, α) where {T} end
 
 """
     PDataKARC(::Type{S}, ::Type{T}, n)

test/allocation_test.jl

@@ -15,17 +15,13 @@ function alloc_preprocess(XData, H, g, ng, calls, max_calls, α)
   return nothing
 end
 
-function alloc_solve_model(solve, XData, H, g, ng, calls, max_calls, α)
-  solve(XData, H, g, ng, calls, max_calls, α)
+function alloc_solve_model(XData, H, g, ng, calls, max_calls, α)
+  solve_model!(XData, H, g, ng, calls, max_calls, α)
   return nothing
 end
 
 @testset "Allocation test in preprocess and solvemodel" begin
-  for (Data, solve) in (
-    (PDataKARC, AdaptiveRegularization.solve_modelKARC),
-    (PDataTRK, AdaptiveRegularization.solve_modelTRK),
-    (PDataST, AdaptiveRegularization.solve_modelST_TR),
-  )
+  for Data in (PDataKARC, PDataTRK, PDataST)
     XData = Data(S, T, n)
     @testset "Allocation test in preprocess with $(Data)" begin
       alloc_preprocess(XData, H, g, ng, calls, max_calls, α)
@@ -34,8 +30,8 @@ end
     end
 
     @testset "Allocation test in $solve with $(Data)" begin
-      alloc_solve_model(solve, XData, H, g, ng, calls, max_calls, α)
-      al = @allocated alloc_solve_model(solve, XData, H, g, ng, calls, max_calls, α)
+      alloc_solve_model(XData, H, g, ng, calls, max_calls, α)
+      al = @allocated alloc_solve_model(XData, H, g, ng, calls, max_calls, α)
       @test al == 0
     end
   end

test/allocation_test_main.jl

@@ -9,7 +9,7 @@ S, T = typeof(x0), eltype(x0)
 PData = PDataKARC(S, T, n)
 
 function alloc_AdaptiveRegularization(stp, solver, stats)
-  solve!(solver, stp, stats, solve_model = AdaptiveRegularization.solve_modelKARC)
+  solve!(solver, stp, stats)
   return nothing
 end