OptunaSharp.Interop.pythonnet 1.1.0

OptunaSharp.Interop.pythonnet

OptunaSharp's bridge to the live Python Optuna ecosystem via Python.NET (pythonnet) — most importantly OptunaHub, Optuna's community registry of black-box-optimization methods and benchmarks.

It is the .NET analog of NumSharp.Interop.pythonnet (which bridges NDArray ↔ NumPy): here, OptunaHub.LoadModule("category/name") mirrors optunahub.load_module, and the loaded contributor packages are wrapped so an OptunaSharp Study can drive them directly.

Core stays pure-managed. This package takes a foreign-runtime dependency (an embedded CPython), so it ships separately — the NumSharp.Interop.OpenBLAS optional-package pattern.

Prerequisites

A discoverable CPython with Optuna and OptunaHub:

pip install optuna optunahub

git on PATH (OptunaHub's default download path). Set PYTHONNET_PYDLL to your libpython if it can't be found automatically (e.g. C:\...\python312.dll, /usr/lib/.../libpython3.12.so).

Benchmarks — optimize a real hub problem with any OptunaSharp sampler

using OptunaSharp;
using OptunaSharp.Interop.PythonNet;

using var bbob = OptunaHub.LoadModule("benchmarks/bbob");
var problem = bbob.CreateProblem("Problem", new() { ["function_id"] = 1, ["dimension"] = 2 });

var study = Optuna.CreateStudy(directions: problem.Directions);   // Sphere is single-objective
study.Optimize(problem.SingleObjective, nTrials: 50);
Console.WriteLine(study.BestValue);

problem.Directions and problem.SearchSpace are exposed as OptunaSharp types; problem.Objective (multi-objective) / problem.SingleObjective suggest each parameter on the C# trial, then call the Python evaluate. With OptunaHub ≥ 0.5.0 they first record the problem's evaluate_constraints(params) on the trial through trial.SetConstraint, as BaseProblem.__call__ does, so constraint-aware samplers see feasibility without a constraints_func.

Samplers — drive an OptunaSharp study with a hub sampler

using var nm = OptunaHub.LoadModule("samplers/nelder_mead");
using var sampler = nm.CreateSampler("NelderMeadSampler", new() { ["seed"] = 0 });

var study = Optuna.CreateStudy(sampler: sampler);
study.Optimize(t => {
    var x = t.SuggestFloat("x", -10, 10);
    return (x - 2) * (x - 2);
}, nTrials: 50);

The real Python sampler decides each suggestion. The bridge keeps a persistent Python shadow study in lockstep with the OptunaSharp study: every trial with its state, values, parameters, attributes and intermediate values. It forwards before_trial, infer_relative_search_space, sample_relative, sample_independent and after_trial with the trial objects Optuna passes, and writes the attributes the sampler sets back to the OptunaSharp trial. The pruner bridge does the same for prune. Registry samplers and pruners give the same trials as when run natively (test/OptunaSharp.Tests.Interop/OptunaHubNativeParityTests.cs).

Loading

• OptunaHub.LoadModule(package, repoOwner, repoName, gitRef, forceReload, baseUrl) — mirrors optunahub.load_module; baseUrl/repoOwner/repoName reach third-party registries.
• OptunaHub.LoadLocalModule(package, registryRoot) — mirrors optunahub.load_local_module against a local registry checkout (e.g. the pinned refs/optunahub-registry submodule; no network).

Part of the SciSharp STACK.