Swyft Versions Save

A system for scientific simulation-based inference at scale.

v0.4.5

8 months ago

What's Changed

Updates:

  • Modularisation of SwyftModule
  • Restructuring of configure_optimizers and configure_callbacks
  • New default is AdamW with early stopping
  • Alternatives with AdamW and OneCycleLR and ReduceLROnPlateau exist as well
  • Automatic reloading of best model at the end of training is default
  • Extended plot options
  • Add smooth_prior flag for interpolating likelihoods
  • Renaming of plotting routines (plot_posterior, plot_corner, plot_pair)
  • Update example notebook
  • Bug fix for labels in corner plot by @anchal-009 in https://github.com/undark-lab/swyft/pull/136

New Contributors

Full Changelog: https://github.com/undark-lab/swyft/compare/v0.4.4...v0.4.5

v0.4.4

9 months ago

Overview

  • Included updated author list (AUTHOR file) and update LICENSE information
  • Add torch conjugate gradient to dependencies (for forthcoming applications)
  • Add list of Swyft publications and events to docs
  • pytorch-lighthing 2.0 compatibility not fully established (pending until 2.0 has feature parity with 1.9)
  • Add swyft.get_class_probs for extracting classification results from ratio estimators
  • Add LogRatioEstimator_Gaussian, a Gaussian ratio estimator for arbitrary dimensions
  • Deprecate LogRatioEstimator_1dim_Gaussian (worked only with 1dim features and 1dim parameters)
  • Rewritten logratios aggregator in swyft.Module to be more resilient
  • Add AuxLoss, which just enables to add arbitrary extra losses (e.g. related to regularisation) to the training object
  • Enable Spectral embedding for parameters in LogRatioEstimator_Ndim for high fidelity posteriors
  • Add LogRatioEstimator_Autoregressive, an autoregressive neural ratio estimator module

What's Changed

Full Changelog: https://github.com/undark-lab/swyft/compare/v0.4.3...v0.4.4

v0.4.3

1 year ago

What's Changed

Full Changelog: https://github.com/undark-lab/swyft/compare/v0.4.2...v0.4.3

v0.4.2

1 year ago

What's Changed

New Contributors

Full Changelog: https://github.com/undark-lab/swyft/compare/v0.4.1...v0.4.2

v0.4.2-pre1

1 year ago

Overview

Fixes:

  • Math in docstrings
  • Functionality of LogRatioEstimator_1dim_Gaussian, added minstd argument for stability
  • Fix tqdm progress bars in notebooks

Enhancements:

  • Automatic plot axes assignment
  • LogRatioEstimators can take different summaries for different ratios
  • Coverage estimation uses noisy logratios in order to make coverage plots useful for classification problems

What's Changed

Full Changelog: https://github.com/undark-lab/swyft/compare/v0.4.1.post1...v0.4.2-pre1

v0.4.1.post1

1 year ago

Fix image link in readme.rst.

v0.4.1

1 year ago

Fixing installation and documentation bugs

Full Changelog: https://github.com/undark-lab/swyft/compare/v0.4.0...v0.4.1

v0.4.0

1 year ago

Release of Swyft with pytorch-lightning based API.

What's changed

  • Training is handled by pytorch-lightning
  • More flexible inference network architectures (enabling training of conditional ratios, ratios of different variables, etc)
  • New simulation model based on definition of the full graphical model
  • New ZarrStore for parallel simulations
  • More efficient handling of coverage tests

See tutorial notebooks for example use-cases.

What's next

  • Incorporation of all Swyft v0.3.x functionality (violin plots, etc)

Full Changelog: https://github.com/undark-lab/swyft/compare/v0.3.2...v0.4.0

v0.4.0-pre1

1 year ago

First prerelease version of swyft v0.4.0 (aka swyft-lightning) with complete interface overhaul.

v0.3.2

2 years ago
  • Introduced a version using import swyft; swyft.__version__
  • Added labels for violin marginals.
  • Followed the requested changes from the JOSS review thread.

Full Changelog: https://github.com/undark-lab/swyft/compare/v0.3.1...v0.3.2