v0.4.1

Release of glasflow v0.4.1.

This is a bug fix release and is not believed to introduce backwards incompatible changes.

Fixed

• Fix automatic mixed precision with spline transforms by @mj-will in #66

Full Changelog: v0.4.0...v0.4.1

v0.4.0

Release of glasflow v0.4.0.

Added

• Add various autoregressive flows using the existing transforms in nflows (#62)
• Add scale_activation keyword argument to nflows.transforms.autoregressive.MaskedAffineAutoregressiveTransform (uofgravity/nflows#11)

Changed

• Drop support for Python 3.7 (#61)

Full Changelog: v0.3.1...v0.4.0

v0.3.1

Release of version 0.3.1 of glasflow. This is a bug-fix release and should be backwards compatible.

Fixed

• Addressed a deprecation warning in the nflows submodule when using LU decomposition (uofgravity/nflows#10, #57)

Full Changelog: v0.3.0...v0.3.1

v0.3.0

Release v0.3.0 of glasflow

This release includes new features and bug fixes. See below for details and for the complete set of changes see: v0.2.0...v0.3.0

What's changed

Added

• Keyword arguments passed to glasflow.transform.coupling.AffineCouplingTransform are now propogated to the parent class. (#51)
• Add support scale_activation to glasflow.transform.coupling.AffineCouplingTransform and set the default to nflows_general. (#52, #54)

Changed

• Default scale activation for glasflow.transform.coupling.AffineCouplingTransform is changed from DEFAULT_SCALE_ACTIVATION in nflows to nflows_general from glasflow. This changes the default behaviour, the previous behaviour can be recovered by setting scale_activation='nflows'. (#52, #54)

Fixed

• fix a bug in glasflow.nflows/utils/torchutils.searchsorted, see uofgravity/nflows#9 for details. (#53)

Deprecated

• The scaling_method argument in glasflow.transform.coupling.AffineCouplingTransform is now deprecated in favour of scale_activation and will be removed in a future release. (#52)

v0.2.0

This is release 0.2.0 of glasflow.

This release includes a couple of new features and a change to the logging when importing the package. It also bumps the minimum PyTorch version to 1.11.0.

What's Changed

Added

• Add a multi-layer perceptron (glasflow.nets.mlp.MLP). (#40)
• Add a resampled Gaussian distribution that uses Learnt Accept/Reject Sampling (glasflow.distributions.resampled.ResampledGaussian). (#40)
• Add a multivariate uniform distribution for Neural Spline Flows (glasflow.distributions.uniform.MultivariateUniform). (#47)

Changed

• Change logging statements on import to, by default, only appear when an external version of nflows is being used. (#44)
• Bump minimum version of PyTorch to 1.11.0. (#38)

Full Changelog: v0.1.2...v0.2.0

v0.1.2

Another patch to fix CI not uploading releases to PyPI.

Includes an update to the fork of nflows.

v0.1.1

Patch to fix CI not uploading release to PyPI

v0.1.0

First release of glasflow

Full Changelog: https://github.com/igr-ml/glasflow/commits/v0.1.0