botorch

Introduction to BoTorch

BoTorch is a library designed for Bayesian Optimization, built on top of PyTorch, a popular machine learning library. As an evolving project currently in beta, BoTorch targets researchers and advanced practitioners involved in Bayesian Optimization and AI. For end-users who are not focused on research, BoTorch recommends using Ax, a platform that integrates with BoTorch for user-friendly optimization tasks.

Key Features

BoTorch offers several compelling features:

• Modular and Extensible: BoTorch provides a flexible interface for Bayesian optimization primitives. Users can easily compose probabilistic models, acquisition functions, and optimizers.
• Powered by PyTorch: Utilizing PyTorch’s capabilities, including auto-differentiation and GPU support, BoTorch enables efficient computations.
• Monte Carlo Acquisition Functions: BoTorch supports these functions through techniques like the reparameterization trick, which simplify the implementation of novel ideas without heavy assumptions.
• Integration with PyTorch Architectures: It allows for smooth integration with deep and convolutional network architectures.
• State-of-the-Art Models: Deep integration with GPyTorch enables advanced probabilistic models, such as Gaussian Processes, multi-task GPs, and deep kernel learning.

Audience

BoTorch’s primary users are researchers and sophisticated AI practitioners. For those not deeply engaged in optimization research, Ax serves as a more accessible tool, offering flexible management of Bayesian Optimization tasks.

Installation

Requirements:

• Python 3.10 or higher
• PyTorch 2.0.1 or higher
• Other necessary packages include gpytorch 1.13, linear_operator 0.5.3, pyro-ppl 1.8.4 or newer, scipy, and multiple-dispatch.

MacOS Users with Intel Processors: Before installing BoTorch, PyTorch should be manually installed to ensure it is linked to MKL, optimizing performance on Intel processors.

Installation Options:

1. Via Anaconda (Recommended):

   conda install botorch -c pytorch -c gpytorch -c conda-forge
   

2. Using pip:

   pip install botorch
   

3. Latest Development Version: To access cutting-edge features, install directly from GitHub.

4. Editable Installation: Suitable for contributors, allowing for local changes to be reflected immediately.

Getting Started

A simple Bayesian optimization loop in BoTorch consists of:

1. Fitting a Gaussian Process (GP) Model:

   import torch
   from botorch.models import SingleTaskGP
   from botorch.fit import fit_gpytorch_mll
   from gpytorch.mlls import ExactMarginalLogLikelihood
   
   train_X = torch.rand(10, 2, dtype=torch.double) * 2
   Y = 1 - (train_X - 0.5).norm(dim=-1, keepdim=True)
   Y += 0.1 * torch.rand_like(Y)
   
   gp = SingleTaskGP(train_X, train_Y=Y)
   mll = ExactMarginalLogLikelihood(gp.likelihood, gp)
   fit_gpytorch_mll(mll)
   

2. Constructing an Acquisition Function:

   from botorch.acquisition import LogExpectedImprovement
   
   logEI = LogExpectedImprovement(model=gp, best_f=Y.max())
   

3. Optimizing the Acquisition Function:

   from botorch.optim import optimize_acqf
   
   bounds = torch.stack([torch.zeros(2), torch.ones(2)]).to(torch.double)
   candidate, acq_value = optimize_acqf(logEI, bounds, q=1, num_restarts=5, raw_samples=20)
   

Contribution and Citation

BoTorch is MIT licensed, and contributions are welcome. For those using BoTorch in research, citation details are provided:

M. Balandat et al., "BoTorch: A Framework for Efficient Monte-Carlo Bayesian Optimization," Advances in Neural Information Processing Systems 33, 2020.

In summary, BoTorch is a robust tool in the sphere of Bayesian Optimization, offering comprehensive support for advanced users while remaining flexible and extensible for continuous development and research integration.