sbi: simulation-based inference Versions Save

API change

• We have moved sbi to an new github organization: https://github.com/sbi-dev/sbi
• We have changed the website of the sbi docs: https://sbi-dev.github.io/sbi/.
• sbi.analysis.pairplot: upper was replaced by offdiag and will be deprecated in a future release.

Features and enhancements

• size-invariant embedding nets for amortized inference with iid-data (@janfb, #808)
• option for new using MAF with rational quadratic splines (thanks to @ImahnShekhzadeh, #819)
• improved docstring for process_prior (thanks to @musoke, #813)
• extended tutorial for SBI with iid data (@janfb, #857)
• new tutorial for SBI with experimental conditions and mixed data (@janfb, #829)
• New options for pairplot:
  • upper is now called offdiag to match other kwargs.
  • alternating colors for samples and points
  • option to add a legend and pass kwargs for the legend.

Bug fixes

• fixed memory leak in in append_simulations (thanks to @VictorSven, #803)
• bug fix for CNRE (thanks to @bkmi, #815)
• bug fix for iid-inference with posterior ensembles (@janfb, #826)
• bug fix for simulation-based calibration with VI posteriors (@janfb, #834, #838)
• bug fix for BoxUniform device handling (@janfb, #854, #856)
• bug fix for MAP estimates with independent priors (@janfb, #867)
• bug fix for tutorial on SBC (@michaeldeistler, #891)
• fix spurious seeding for simulate_for_sbi (@jan-matthis, #876)
• bump python version of github action tests to 3.9.13 (@michaeldeistler, #888, #900)

• implementation of "Contrastive Neural Ratio Estimation" (thanks to @bkmi, #787)
• implementation of "Balanced Neural Ratio Estimation" (thanks to @ADelau, #779)
• bugfixes for SBC, device handling and handling of iid-observations (#793, #789, #780)

Major changes and bug fixes

• implementation of "Truncated proposals for scalable and hassle-free sbi" (#754)
• sample-based expected coverage tests (#754)
• permutation invariant embedding to allow iid data in SNPE (thanks @coschroeder, #751)
• convolutional neural network embedding (thanks @coschroeder, #745, #751, #769)
• disallow invalid simulations when using SNLE, SNRE, or atomic SNPE-C (#768)

Enhancements

• add tutorial on all available methods (#754)
• allow seeding of simulate_for_sbi on multiple workers (#762)
• expose enable_transforms in sampler interface (#756)
• bugfix for building the transformation of transformed distributions (#756)

• Rely on new version of pyknos with bugfix for APT with MDNs (#734)
• bugfix: atomic SNPE-C now allows any kind of proposal (#732)
• bugfix for SNPE with implicit prior on GPU (#730)
• SNPE-A has force_first_round_loss=True as default (#729)

• bug fix for ArviZ integration (#727)

Major changes and bug fixes

• new option to sample posterior using importance sampling (#692)
• new option to use arviz for posterior plotting and MCMC diagnostics (#546, #607, thanks to @sethaxen)
• fixes for using the VIPosterior with MultipleIndependent prior, a51e93b
• bug fix for sir (sequential importance reweighting) for MCMC initialization (#692)
• bug fix for SNPE-A 565082c
• bug fix for validation loader batch size (#674, thanks to @bkmi)
• small bug fixes for pairplot and MCMC kwargs

Enhancements

• improved and new tutorials:
  • Tutorial for simulation-based calibration (SBC) (#629, thanks to @psteinb)
  • Tutorial for sampling the conditional posterior (#667)
• new option to use first-round loss in all rounds
• simulated data is now stored as Dataset to reduce memory load and add flexibility with large data sets (#685, thanks to @tbmiller-astro)
• refactoring of summary write for better training logs with tensorboard (#704)
• new option to find peaks of 1D posterior marginals without gradients (#707, #708, thanks to @Ziaeemehr)
• new option to not use parameter transforms in DirectPosterior for more flexibility with custom priors (#714)

Breaking changes

• Posteriors saved under sbi v0.17.2 or older can not be loaded under sbi v0.18.0 or newer.
• sample_with can no longer be passed to .sample(). Instead, the user has to rerun .build_posterior(sample_with=...). (#573)
• the posterior no longer has the the method .sample_conditional(). Using this feature now requires using the sampler interface (see tutorial here) (#573)
• retrain_from_scratch_each_round is now called retrain_from_scratch (#598, thanks to @jnsbck)
• API changes that had been introduced in sbi v0.14.0 and v0.15.0 are not enforced. Using the interface prior to those changes leads to an error (#645)
• prior passed to SNPE / SNLE / SNRE must be a PyTorch distribution (#655), see FAQ-7 for how to pass use custom prior.

Major changes and bug fixes

• new sampler interface (#573)
• posterior quality assurance with simulation-based calibration (SBC) (#501)
• added Sequential Neural Variational Inference (SNVI) (Glöckler et al. 2022) (#609, thanks to @manuelgloeckler)
• bugfix for SNPE-C with mixture density networks (#573)
• bugfix for sampling-importance resampling (SIR) as init_strategy for MCMC (#646)
• new density estimator for neural likelihood estimation with mixed data types (MNLE, #638)
• MCMC can now be parallelized across CPUs (#648)
• improved device check to remove several GPU issues (#610, thanks to @LouisRouillard)

Enhancements

• pairplot takes ax and fig (#557)
• bugfix for rejection sampling (#561)
• remove warninig when using multiple transforms with NSF in single dimension (#537)
• Sampling-importance-resampling (SIR) is now the default init_strategy for MCMC (#605)
• change mp_context to allow for multi-chain pyro samplers (#608, thanks to @sethaxen)
• tutorial on posterior predictive checks (#592, thanks to @LouisRouillard)
• add FAQ entry for using a custom prior (#595, thanks to @jnsbck)
• add methods to plot tensorboard data (#593, thanks to @lappalainenj)
• add option to pass the support for custom priors (#602)
• plotting method for 1D marginals (#600, thanks to @guymoss)
• fix GPU issues for conditional_pairplot and ActiveSubspace (#613)
• MCMC can be performed in unconstrained space also when using a MultipleIndependent distribution as prior (#619)
• added z-scoring option for structured data (#597, thanks to @rdgao)
• refactor c2st; change its default classifier to random forest (#503, thanks to @psteinb)
• MCMC init_strategy is now called proposal instead of prior (#602)
• inference objects can be serialized with pickle (#617)
• preconfigured fully connected embedding net (#644, thanks to @JuliaLinhart #624)
• posterior ensembles (#612, thanks to @jnsbck)
• remove gradients before returning the posterior (#631, thanks to @tomMoral)
• reduce batchsize of rejection sampling if few samples are left (#631, thanks to @tomMoral)
• tutorial for how to use SBC (#629, thanks to @psteinb)
• tutorial for how to use SBI with trial-based data and mixed data types (#638)
• allow to use a RestrictedPrior as prior for SNPE (#642)

Minor changes

• bug fix for transforms in KDE (#552)

Minor changes

• improve kwarg handling for rejection abc and smcabc
• typo and link fixes (#549, thanks to @pitmonticone)
• tutorial notebook on crafting summary statistics with sbi (#511, thanks to @ybernaerts)
• small fixes and improved documenentation for device handling (#544, thanks to @milagorecki)

Major changes

• New API for specifying sampling methods (#487). Old syntax:

posterior = inference.build_posterior(sample_with_mcmc=True)

New syntax:

posterior = inference.build_posterior(sample_with="mcmc")  # or "rejection"

• Rejection sampling for likelihood(-ratio)-based posteriors (#487)
• MCMC in unconstrained and z-scored space (#510)
• Prior is now allowed to lie on GPU. The prior has to be on the same device as the one passed for training (#519).
• Rejection-ABC and SMC-ABC now return the accepted particles / parameters by default, or a KDE fit on those particles (kde=True) (#525).
• Fast analytical sampling, evaluation and conditioning for DirectPosterior trained with MDNs (thanks @jnsbck #458).

Minor changes

• scatter allowed for diagonal entries in pairplot (#510)
• Changes to default hyperparameters for SNPE_A (thanks @famura, #496, #497)
• bugfix for within_prior checks (#506)