Add interface for custom scaling

lib/MadNLPGPU/src/MadNLPGPU.jl

@@ -12,7 +12,7 @@ import MadNLP
 
 import MadNLP:
     @kwdef, Logger, @debug, @warn, @error,
-    AbstractOptions, AbstractLinearSolver, set_options!, MadNLPLapackCPU,
+    AbstractOptions, AbstractLinearSolver, AbstractNLPModel, set_options!, MadNLPLapackCPU,
     SymbolicException,FactorizationException,SolveException,InertiaException,
     introduce, factorize!, solve!, improve!, is_inertia, inertia, tril_to_full!
 

lib/MadNLPGPU/src/kernels.jl

@@ -31,11 +31,16 @@ end
 MadNLP.is_valid(src::CuArray) = true
 
 # Constraint scaling
-function MadNLP.set_con_scale!(con_scale::AbstractVector, jac::CuMatrix, nlp_scaling_max_gradient)
+function MadNLP.scale_constraints!(
+    nlp::AbstractNLPModel,
+    con_scale::AbstractVector,
+    jac::CuMatrix;
+    max_gradient=1e-8,
+)
     # Compute reduction on the GPU with built-in CUDA.jl function
     d_con_scale = maximum(abs, jac, dims=2)
     copyto!(con_scale, d_con_scale)
-    con_scale .= min.(1.0, nlp_scaling_max_gradient ./ con_scale)
+    con_scale .= min.(1.0, max_gradient ./ con_scale)
 end
 
 @kernel function _treat_fixed_variable_kernell!(dest, ind_fixed)

src/MadNLP.jl

@@ -31,6 +31,7 @@ include("enums.jl")
 include("utils.jl")
 include("options.jl")
 include("matrixtools.jl")
+include("scaling.jl")
 include(joinpath("LinearSolvers","linearsolvers.jl"))
 include(joinpath("KKT", "KKTsystem.jl"))
 include("interiorpointsolver.jl")

src/interiorpointsolver.jl

@@ -537,7 +537,7 @@ function initialize!(ips::AbstractInteriorPointSolver)
     compress_jacobian!(ips.kkt)
     if (ips.m > 0) && ips.opt.nlp_scaling
         jac = get_raw_jacobian(ips.kkt)
-        set_con_scale!(ips.con_scale, jac, ips.opt.nlp_scaling_max_gradient)
+        scale_constraints!(ips.nlp, ips.con_scale, jac; max_gradient=ips.opt.nlp_scaling_max_gradient)
         set_jacobian_scaling!(ips.kkt, ips.con_scale)
         ips.l./=ips.con_scale
     end
@@ -547,7 +547,7 @@ function initialize!(ips::AbstractInteriorPointSolver)
     eval_grad_f_wrapper!(ips, ips.f,ips.x)
     @trace(ips.logger,"Computing objective scaling.")
     if ips.opt.nlp_scaling
-        ips.obj_scale[] = min(1,ips.opt.nlp_scaling_max_gradient/norm(ips.f,Inf))
+        ips.obj_scale[] = scale_objective(ips.nlp, ips.f; max_gradient=ips.opt.nlp_scaling_max_gradient)
         ips.f.*=ips.obj_scale[]
     end
 
@@ -1505,22 +1505,6 @@ function initialize_variables!(x,xl,xu,bound_push,bound_fac)
     end
 end
 
-function set_con_scale!(con_scale::AbstractVector, jac::SparseMatrixCOO, nlp_scaling_max_gradient)
-    @simd for i in 1:nnz(jac)
-        row = @inbounds jac.I[i]
-        @inbounds con_scale[row] = max(con_scale[row], abs(jac.V[i]))
-    end
-    con_scale .= min.(1.0, nlp_scaling_max_gradient ./ con_scale)
-end
-function set_con_scale!(con_scale::AbstractVector, jac::Matrix, nlp_scaling_max_gradient)
-    for row in 1:size(jac, 1)
-        for col in 1:size(jac, 2)
-            @inbounds con_scale[row] = max(con_scale[row], abs(jac[row, col]))
-        end
-    end
-    con_scale .= min.(1.0, nlp_scaling_max_gradient ./ con_scale)
-end
-
 function adjust_boundary!(x_lr,xl_r,x_ur,xu_r,mu)
     adjusted = 0
     c1 = eps(Float64)*mu

src/scaling.jl

@@ -0,0 +1,71 @@
+
+function _set_scaling!(con_scale::AbstractVector, jac::SparseMatrixCOO)
+    @simd for i in 1:nnz(jac)
+        row = @inbounds jac.I[i]
+        @inbounds con_scale[row] = max(con_scale[row], abs(jac.V[i]))
+    end
+end
+function _set_scaling!(con_scale::AbstractVector, jac::Matrix)
+    for row in 1:size(jac, 1)
+        for col in 1:size(jac, 2)
+            @inbounds con_scale[row] = max(con_scale[row], abs(jac[row, col]))
+        end
+    end
+end
+
+"""
+    scale_constraints!(
+        nlp::AbstractNLPModel,
+        con_scale::AbstractVector,
+        jac::AbstractMatrix;
+        max_gradient=1e-8,
+    )
+
+Compute the scaling of the constraints associated
+to the nonlinear model `nlp`. By default, Ipopt's scaling
+is applied. The user can write its own function to scale
+appropriately any custom `AbstractNLPModel`.
+
+### Notes
+
+This function assumes that the Jacobian `jac` has been evaluated
+before calling this function.
+
+"""
+function scale_constraints!(
+    nlp::AbstractNLPModel,
+    con_scale::AbstractVector,
+    jac::AbstractMatrix;
+    max_gradient=1e-8,
+)
+    fill!(con_scale, 0.0)
+    _set_scaling!(con_scale, jac)
+    con_scale .= min.(1.0, max_gradient ./ con_scale)
+end
+
+"""
+    scale_objective(
+        nlp::AbstractNLPModel,
+        grad::AbstractVector;
+        max_gradient=1e-8,
+    )
+
+Compute the scaling of the objective associated to the
+nonlinear model `nlp`. By default, Ipopt's scaling
+is applied. The user can write its own function to scale
+appropriately the objective of any custom `AbstractNLPModel`.
+
+### Notes
+
+This function assumes that the gradient `gradient` has been evaluated
+before calling this function.
+
+"""
+function scale_objective(
+    nlp::AbstractNLPModel,
+    grad::AbstractVector;
+    max_gradient=1e-8,
+)
+    return min(1, max_gradient / norm(grad, Inf))
+end
+