Remove a side effect where a stopping criterion accidentally updated …

…a gradient.
JuliaManifolds · Oct 30, 2023 · 7002e01 · 7002e01
1 parent b4b5603
commit 7002e01
Show file tree

Hide file tree

Showing 5 changed files with 15 additions and 11 deletions.
diff --git a/src/plans/adabtive_regularization_with_cubics_plan.jl b/src/plans/adabtive_regularization_with_cubics_plan.jl
@@ -71,10 +71,13 @@ function get_cost(
 )
     M = base_manifold(TpM)
     p = TpM.point
-    c = get_objective_cocst(M, arcmo, p)
+    c = get_objective_cost(M, arcmo, p)
     G = get_objective_gradient(M, arcmo, p)
     Y = get_objective_hessian(M, arcmo, p, X)
-    return c + inner(M, p, G, X) + 1 / 2 * inner(M, p, Y, X) + σ / 3 * norm(M, p, X)^3
+    return c + inner(M, p, G, X) + 1 / 2 * inner(M, p, Y, X) + arcmo.σ / 3 * norm(M, p, X)^3
+end
+function get_cost_function(arcmo::AdaptiveRagularizationWithCubicsModelObjective)
+    return (TpM, X) -> get_cost(TpM, arcmo, X)
 end
 @doc raw"""
     get_gradient(TpM, trmo::AdaptiveRagularizationWithCubicsModelObjective, X)
@@ -105,4 +108,8 @@ function get_gradient!(
     Y .= Y + get_objective_gradient(M, arcmo, p) + arcmo.σ * norm(M, p, X) * X
     return Y
 end
+function get_gradient_function(arcmo::AdaptiveRagularizationWithCubicsModelObjective)
+    return (TpM, X) -> get_gradient(TpM, arcmo, X)
+end
+
 # Also Implement the Hessian for Newton subsubsolver?
diff --git a/src/plans/subsolver_plan.jl b/src/plans/subsolver_plan.jl
@@ -32,8 +32,8 @@ get_objective(amso::AbstractManifoldSubObjective)
 Evaluate the cost of the (original) objective stored within the subobjective.
 """
 function get_objective_cost(
-    M::AbstractManifold, amso::AbstractManifoldSubObjective{O}, p
-) where {O<:AbstractManifoldCostObjective}
+    M::AbstractManifold, amso::AbstractManifoldSubObjective{E,O}, p
+) where {E,O<:AbstractManifoldCostObjective}
     return get_cost(M, get_objective(amso), p)
 end
 

diff --git a/src/solvers/Lanczos.jl b/src/solvers/Lanczos.jl
@@ -2,7 +2,6 @@
 #
 # Lanczos sub solver
 #
-
 @doc raw"""
     LanczosState{P,T,SC,B,I,R,TM,V,Y} <: AbstractManoptSolverState
 
@@ -263,8 +262,7 @@ function (c::StopWhenFirstOrderProgress)(
     TpM = get_manifold(dmp)
     p = TpM.point
     M = base_manifold(TpM)
-    get_gradient!(dmp, ls.X, p)
-    nX = norm(M, p, ls.X)
+    nX = norm(M, p, get_gradient(dmp, p))
     y = @view(ls.coefficients[1:(i - 1)])
     Ty = @view(ls.tridig_matrix[1:i, 1:(i - 1)]) * y
     ny = norm(y)

diff --git a/src/solvers/adaptive_regularization_with_cubics.jl b/src/solvers/adaptive_regularization_with_cubics.jl
@@ -456,7 +456,6 @@ function step_solver!(dmp::AbstractManoptProblem, arcs::AdaptiveRegularizationSt
     set_manopt_parameter!(arcs.sub_problem, :Objective, :σ, arcs.σ)
     set_iterate!(arcs.sub_state, M, copy(M, arcs.p, arcs.X))
     set_manopt_parameter!(arcs.sub_state, :σ, arcs.σ)
-    set_manopt_parameter!(arcs.sub_state, :p, copy(M, arcs.p))
     #Solve the sub_problem – via dispatch depending on type
     solve_arc_subproblem!(M, arcs.S, arcs.sub_problem, arcs.sub_state, arcs.p)
     # Compute ρ

diff --git a/test/solvers/test_adaptive_regularization_with_cubics.jl b/test/solvers/test_adaptive_regularization_with_cubics.jl
@@ -159,12 +159,12 @@ include("../utils/example_tasks.jl")
 
         # test both inplace and allocating variants of grad_g
         X0 = grad_f(M, p0)
-        X1 = grad_g(M2, X0)
+        X1 = get_gradient(M2, arcmo, X0)
         X2 = zero_vector(M, p0)
-        grad_g(M2, X2, X0)
+        get_gradient!(M2, X2, arcmo, X0)
         @test isapprox(M, p0, X1, X2)
 
-        sub_problem = DefaultManoptProblem(M2, ManifoldGradientObjective(g, grad_g))
+        sub_problem = DefaultManoptProblem(M2, arcmo)
         sub_state = GradientDescentState(
             M2,
             zero_vector(M, p0);