OptunaSharp.Interop.BoTorch 1.0.0

OptunaSharp.Interop.BoTorch

BoTorchSampler for OptunaSharp — a native OptunaSharp BaseSampler that drives the real BoTorch (PyTorch + GPyTorch) library in-process via Python.NET.

It is a faithful C# port of the glue in optuna_integration.BoTorchSampler — the sampler protocol (search-space intersection, one-hot transform, startup trials, running-trial handling, direction flipping, seed handling) and the candidates_func selection — while every numerical step (SingleTaskGP, fit_gpytorch_mll, the acquisition functions, optimize_acqf) is a direct call into botorch. So you get the genuine article: byte-exact with BoTorch and its acquisition suite — with the Optuna side staying native OptunaSharp (your study, storage, and RNG). The bridge keeps pythonnet off its public surface: you pick an acquisition with the BoTorchCandidatesFunc enum, never a torch.Tensor.

Unlike the Python integration, this package does not depend on optuna_integration — we port it. You only need botorch installed in a discoverable CPython:

pip install botorch          # pulls torch + gpytorch

Usage

using OptunaSharp;
using OptunaSharp.Interop.BoTorch;

var study = Optuna.CreateStudy(sampler: new BoTorchSampler(seed: 0));
study.Optimize(t => Math.Pow(t.SuggestFloat("x", -10, 10) - 2, 2), nTrials: 30);

Console.WriteLine($"{study.BestValue} at {study.BestParams["x"]}");

By default the acquisition function is chosen automatically (BoTorchCandidatesFunc.Auto) from the number of objectives and whether running trials are considered (single-objective → log-Expected Improvement). Pick one explicitly:

var study = Optuna.CreateStudy(
    sampler: new BoTorchSampler(candidatesFunc: BoTorchCandidatesFunc.QNei, seed: 0));

How it works

BoTorchSampler runs the first nStartupTrials (default 10) through an independent RandomSampler (bit-exact with Optuna). From then on, on each ask it assembles the completed (and, if considerRunningTrials, running) trials into the train_x / train_obj / bounds tensors exactly as optuna_integration.botorch does, wraps the call in BoTorch's manual_seed, and hands them to the selected candidates_func, which builds and fits a SingleTaskGP and optimizes the acquisition function — all in real BoTorch. The returned candidate is un-transformed back to your parameters.

Requirements

• A discoverable CPython (set PYTHONNET_PYDLL to pin the interpreter) with pip install botorch.
• Not AOT-compatible (it hosts CPython). This is a satellite package; OptunaSharp Core stays pure managed.

Part of the SciSharp STACK.