Extend code coverage also to reducing threshold and radius, fix a nasty bug where I accidentially overwrote the iterate.

Author: kellertuer

src/solvers/gradient_sampling.jl

@@ -86,7 +86,7 @@ mutable struct GradientSamplingState{
         Pr <: Union{F, AbstractManoptProblem} where {F}, St <: AbstractManoptSolverState,
         SP <: AbstractVector{<:P}, ST <: AbstractVector{<:T}, A <: AbstractVector{<:R},
         SC <: StoppingCriterion, S <: Stepsize, RTM <: AbstractRetractionMethod, VTM <: AbstractVectorTransportMethod,
-    } <: AbstractManoptSolverState
+    } <: AbstractGradientSolverState
     convex_hull_coeffs::A
     p::P
     sampled_points::SP
@@ -177,7 +177,6 @@ function GradientSamplingState(
         vector_transport_method = vector_transport_method, X = X, Y = copy(M, p, X),
     )
 end
-
 function default_stepsize(
         M::AbstractManifold, ::Type{GradientSamplingState}; retraction_method = default_retraction_method(M),
     )
@@ -387,13 +386,13 @@ function step_solver!(
     # solve sub problem in convex_hull_coeffs
     _gradient_sampling_subsolver(M, gss)
     # reconstruct tangent vector from the coefficients (w_l in HU17) in Y
-    zero_vector!(M, gss.p, gss.Y)
+    zero_vector!(M, gss.Y, gss.p)
     for (λj, Xj) in zip(gss.convex_hull_coeffs, gss.sampled_vectors)
         gss.Y .+= λj * Xj
     end
     # Decide whether to accept the step or update radius
     if norm(M, gss.p, gss.Y) < gss.subgradient_norm_tolerance
-        # do not accept
+        # do not accept this step but decrease radius and tolerance
         gss.sampling_radius *= gss.sampling_radius_reduction
         gss.subgradient_norm_tolerance *= gss.subgradient_norm_reduction
     else
@@ -420,12 +419,3 @@ function _gradient_sampling_subsolver(
     return gss
 end
 # (c) (not yet needed / implemented) an actual sub solver call
-
-
-function (d::DebugStepsize)(
-        dmp::P, gss::GradientSamplingState, k::Int
-    ) where {P <: AbstractManoptProblem}
-    (k < 1) && return nothing
-    Printf.format(d.io, Printf.Format(d.format), get_last_stepsize(dmp, gss, k))
-    return nothing
-end

test/solvers/test_gradient_sampling.jl

@@ -26,10 +26,11 @@ _debug_gradient_sampling = false
         record = [:Iteration, :Cost, RecordGradientNorm()]
     )
 
+    Random.seed!(23)
     m2 = gradient_sampling(
         M, f, grad_f, p0;
         return_state = true,
-        debug = _debug_gradient_sampling ? [:Iteration, :Cost, " ", :subgradient_norm_tolerance, " ", :sampling_radius, " | ", :GradientNorm, " ", :Change, "\n", :Stop, 100] : [],
+        debug = _debug_gradient_sampling ? [:Iteration, :Cost, " ", :subgradient_norm_tolerance, " ", :sampling_radius, " | ", :GradientNorm, " ", :Change, "\n", :Stop, 10] : [],
         record = [:Iteration, :Cost, RecordGradientNorm()]
     )
 
@@ -40,6 +41,19 @@ _debug_gradient_sampling = false
     p2 = get_solver_result(s2)
     @test f(M, p2) < f(M, p0)
 
+    p3 = copy(M, p0)
+    Random.seed!(23)
+    gradient_sampling!(
+        M, f, grad_f, p3;
+        sampling_radius = 0.1,
+        subgradient_norm_tolerance = 0.02,
+        sub_problem = gradient_sampling_subsolver,
+        sub_state = AllocatingEvaluation(),
+    )
+    # The parameters of this run are chosen so that reduction is necessary,
+    # they hence to not work that well and we end up a bit further away.
+    @test isapprox(M, p2, p3; atol = 2.0e-3)
+
     if _debug_gradient_sampling
         p1 = get_solver_result(m1)
         p2 = get_solver_result(m2)