Swyft is a system for scientific simulation-based inference at scale.
Swyft is the official implementation of Truncated Marginal Neural Ratio Estimation (TMNRE), a hyper-efficient, simulation-based inference technique for complex data and expensive simulators.
Swyft in action#
Swyft makes it convenient to perform Bayesian or Frequentist inference of hundreds, thousands or millions of parameter posteriors by constructing optimal data summaries.
To this end, Swyft estimates likelihood-to-evidence ratios for arbitrary marginal posteriors; they typically require fewer simulations than the corresponding joint.
Swyft performs targeted inference by prior truncation, combining simulation efficiency with empirical testability.
Swyft is based on stochastic simulators, which map parameters stochastically to observational data. Swyft makes it convenient to define such simulators as graphical models.
In scientific settings, a cost-benefit analysis often favors approximating the posterior marginality; swyft provides this functionality.
The package additionally implements our prior truncation technique, routines to empirically test results by estimating the expected coverage, and a simulator manager with zarr storage to simplify use with complex simulators.
Papers using Swyft/TMNRE#
“Fast and Credible Likelihood-Free Cosmology with Truncated Marginal Neural Ratio Estimation“ Cole+ https://arxiv.org/abs/2111.08030
“Estimating the warm dark matter mass from strong lensing images with truncated marginal neural ratio estimation” Anau Montel+, https://arxiv.org/abs/2205.09126
“SICRET: Supernova Ia Cosmology with truncated marginal neural Ratio EsTimation” Karchev+ https://arxiv.org/abs/2209.06733
“One never walks alone: the effect of the perturber population on subhalo measurements in strong gravitational lenses” Coogan+ https://arxiv.org/abs/2209.09918
“Detection is truncation: studying source populations with truncated marginal neural ratio estimation” Anau Montel+ https://arxiv.org/abs/2211.04291
“Debiasing Standard Siren Inference of the Hubble Constant with Marginal Neural Ratio Estimation” Gagnon-Hartman+ https://arxiv.org/abs/2301.05241
“Constraining the X-ray heating and reionization using 21-cm power spectra with Marginal Neural Ratio Estimation” Saxena+ https://arxiv.org/abs/2303.07339
“Peregrine: Sequential simulation-based inference for gravitational wave signals”, Bhardwaj+ https://arxiv.org/abs/2304.02035
“Albatross: A scalable simulation-based inference pipeline for analysing stellar streams in the Milky Way”, Alvey+ https://arxiv.org/abs/2304.02032
Documentation & installation: https://swyft.readthedocs.io/
Source code: https://github.com/undark-lab/swyft
Support & discussion: https://github.com/undark-lab/swyft/discussions
Bug reports: https://github.com/undark-lab/swyft/issues
As of v0.4.0, Swyft is based on pytorch-lightning, with a completely updated
v0.1.2 is the implementation of the paper Simulation-efficient marginal posterior estimation with swyft: stop wasting your precious time.
sbi is a collection of simulation-based inference methods. Unlike Swyft, the repository does not include our truncation scheme nor marginal estimation of posteriors.
lampe is an implementation of amoritzed simulation-based inference methods aimed at simulation-based inference researchers due to its flexibility.
- A - Building inference networks with SwyftModule
- B - Multi-dimensional posteriors and corner plots
- C - Linear regression with Swyft
- D - Informative data summaries
- E- Coverage tests
- F - Training hyper parameters
- G - Simulator and Graphical models
- H - Truncation and bounds
- I - Model comparison
- J - ZarrStore and Parallel Simulations