Python ecosystem

OptunaSharp can run Python optimization libraries inside the .NET process while keeping the study, trial API, objective and storage in C#. Each bridge is an optional package; installing OptunaSharp alone does not install or start Python.

Integration	.NET package	Python dependency	What runs in Python
Python.NET runtime	OptunaSharp.Interop.pythonnet	A compatible CPython	Interpreter, imports and object conversion
OptunaHub	OptunaSharp.Interop.pythonnet	optuna, optunahub, selected package requirements	Registry samplers, pruners and benchmark problems
BoTorch	OptunaSharp.Interop.BoTorch	botorch and its dependencies	GP fitting and acquisition optimization
pycma	OptunaSharp.Interop.PyCma	cma and its dependencies	CMA-ES covariance updates and candidate generation
pymoo	OptunaSharp.Interop.Pymoo	pymoo and its dependencies	Eleven evolutionary algorithms, operators and survival

All four packages target .NET 8 and .NET 10 and share one embedded CPython runtime. BoTorch, pycma and pymoo bring the Python.NET bridge transitively, but NuGet does not install their Python dependencies. Start with the runtime setup, then follow the integration's installation and example.

Choose a bridge

Use OptunaHub to load a contributor package, run a benchmark with an OptunaSharp sampler, or try a Python sampler or pruner through its supported Optuna protocol. Its shadow-study adapter has a narrower state contract than a Python study: read the compatibility limits before choosing a registry method.

Use BoTorch for acquisition functions implemented by the actual PyTorch/GPyTorch stack, including multi-objective and constrained Bayesian optimization. Use pycma when reproducing the Python PyCmaSampler integration; its numerical backend differs from the managed CmaEsSampler. Use pymoo for the exposed multi-objective evolutionary algorithms and their reference directions, aspiration points and genetic operators.

The native scikit-learn backend also reproduces part of a Python library, but its regression and cross-validation implementation is managed C# in the core OptunaSharp package. It does not use this runtime.

What reproducibility means

A seed is one input to reproducibility. Match the Python package versions, numerical dependencies, device and execution order as well. BoTorch and pycma have tests against their Python Optuna integrations; pymoo has behavioral tests for all eleven exposed algorithms and direct replay comparisons for specific configurations. Those scopes are described on each page. They do not establish identical results across arbitrary package upgrades, CPU/GPU builds, or concurrent execution schedules.

The bridges run their Python on the runtime's single Python thread and acquire the GIL there (the Python thread). That protects access to Python objects and makes any calling thread and any nJobs safe. It does not turn a stateful sampler into a shared service: create one sampler per study, and start with nJobs: 1 when reproducing a reference run.