Repository files navigation

Getting Started |Documentation |Discord Server

sbi is a Python package for simulation-based inference, designed to meet the needs ofboth researchers and practitioners. Whether you need fine-grained control or aneasy-to-use interface,sbi has you covered.

Withsbi, you can perform parameter inference using Bayesian inference: Given asimulator that models a real-world process, SBI estimates the full posteriordistribution over the simulator’s parameters based on observed data. This distributionindicates the most likely parameter values while additionally quantifying uncertaintyand revealing potential interactions between parameters.

sbi offers a blend of flexibility and ease of use:

• Low-Level Interfaces: For those who require maximum control over the inferenceprocess,sbi provides low-level interfaces that allow you to fine-tune many aspectsof your workflow.
• High-Level Interfaces: If you prefer simplicity and efficiency,sbi also offershigh-level interfaces that enable quick and easy implementation of complex inferencetasks.

In addition,sbi supports a wide range of state-of-the-art inference algorithms (seebelow for a list of implemented methods):

• Amortized Methods: These methods enable the reuse of posterior estimators acrossmultiple observations without the need to retrain.
• Sequential Methods: These methods focus on individual observations, optimizing thenumber of simulations required.

Beyond inference,sbi also provides:

• Validation Tools: Built-in methods to validate and verify the accuracy of yourinferred posteriors.
• Plotting and Analysis Tools: Comprehensive functions for visualizing and analyzingresults, helping you interpret the posterior distributions with ease.

Getting started withsbi is straightforward, requiring only a few lines of code:

fromsbi.inferenceimportNPE# Given: parameters theta and corresponding simulations xinference=NPE(prior=prior)inference.append_simulations(theta,x).train()posterior=inference.build_posterior()

sbi requires Python 3.10 or higher. While a GPU isn't necessary, it can improveperformance in some cases. We recommend using a virtual environment withconda for an easy setup.

Ifconda is installed on the system, an environment for installingsbi can be created as follows:

conda create -n sbi_env python=3.10&& conda activate sbi_env

To installsbi from PyPI run

python -m pip install sbi

To install and addsbi to a project withpixi, from the project directory run

pixi add sbi

and to install into a particular conda environment withconda, in the activated environment run

conda install --channel conda-forge sbi

Open a Python prompt and run

fromsbi.examples.minimalimportsimpleposterior=simple()print(posterior)

If you're new tosbi, we recommend starting with ourGettingStarted tutorial.

You can also access and run these tutorials directly in your browser by openingCodespace. To do so, click the green“Code” button on the GitHub repository and select “Open with Codespaces.” This providesa fully functional environment where you can exploresbi through Jupyter notebooks.

The following inference algorithms are currently available. You can find instructions onhow to run each of these methodshere.

• (S)NPE_A(including amortized single-roundNPE) from Papamakarios G and Murray IFastε-free Inference of Simulation Models with Bayesian Conditional DensityEstimation(NeurIPS 2016).

• (S)NPE_CorAPT from Greenberg D, Nonnenmacher M, and Macke JAutomatic PosteriorTransformation for likelihood-free inference (ICML2019).

• TSNPE from Deistler M, Goncalves P, and Macke JTruncated proposals for scalableand hassle-free simulation-based inference(NeurIPS 2022).

• FMPEfrom Wildberger, J., Dax, M., Buchholz, S., Green, S., Macke, J. H., & Schölkopf, B.Flow matching for scalable simulation-basedinference.(NeurIPS 2023).

• NPSE fromGeffner, T., Papamakarios, G., & Mnih, A.Compositional score modeling forsimulation-based inference.(ICML 2023)

• (S)NLEor justSNL from Papamakarios G, Sterrat DC and Murray ISequential NeuralLikelihood (AISTATS 2019).

• (S)NRE_AorAALR from Hermans J, Begy V, and Louppe G.Likelihood-free Inference withAmortized Approximate Likelihood Ratios (ICML2020).

• (S)NRE_BorSRE from Durkan C, Murray I, and Papamakarios G.On Contrastive Learning forLikelihood-free Inference (ICML 2020).

• (S)NRE_CorNRE-C from Miller BK, Weniger C, Forré P.Contrastive Neural RatioEstimation (NeurIPS 2022).

• BNRE fromDelaunoy A, Hermans J, Rozet F, Wehenkel A, and Louppe G.Towards ReliableSimulation-Based Inference with Balanced Neural RatioEstimation (NeurIPS 2022).

• SNVIfrom Glöckler M, Deistler M, Macke J,Variational methods for simulation-basedinference (ICLR 2022).

• MNLE fromBoelts J, Lueckmann JM, Gao R, Macke J,Flexible and efficient simulation-basedinference for models of decision-making(eLife 2022).

We welcome any feedback on howsbi is working for your inference problems (seeDiscussions) and are happy to receive bugreports, pull requests, and other feedback (seecontribute). We wish to maintain a positivecommunity; please read ourCode of Conduct.

sbi is the successor (using PyTorch) of thedelfi package. It started as a fork of Conor M.Durkan'slfi.sbi runs as a community project. See alsocredits.

sbi has been supported by the German Federal Ministry of Education and Research (BMBF)through project ADIMEM (FKZ 01IS18052 A-D), project SiMaLeSAM (FKZ 01IS21055A) and theTübingen AI Center (FKZ 01IS18039A). Since 2024,sbi is supported by the appliedAIInstitute for Europe, and by NumFOCUS.

Apache License Version 2.0 (Apache-2.0)

If you usesbi consider citing thesbi softwarepaper, in addition to the original researcharticles describing the specific sbi-algorithm(s) you are using.

@article{tejero-cantero2020sbi,  doi = {10.21105/joss.02505},  url = {https://doi.org/10.21105/joss.02505},  year = {2020},  publisher = {The Open Journal},  volume = {5},  number = {52},  pages = {2505},  author = {Alvaro Tejero-Cantero and Jan Boelts and Michael Deistler and Jan-Matthis Lueckmann and Conor Durkan and Pedro J. Gonçalves and David S. Greenberg and Jakob H. Macke},  title = {sbi: A toolkit for simulation-based inference},  journal = {Journal of Open Source Software}}

The above citation refers to the original version of thesbi project and has apersistent DOI. Additionally, new releases ofsbi are citable viaZenodo, where we create a new DOI for everyrelease.