v12.3.0

What's Changed

The main thing of note is a bug fix to ContrastiveLoss (see #164). Apologies for any inconvenience!

• Fix bug in ContrastiveLoss normalization and add improve numpyro transformed wrapper by @danielward27 in #164

Full Changelog: v12.2.0...v12.3.0

v12.2.0

What's Changed

• More consistent parameter type conversion #158
• Add logistic distribution #159
• Int type dummy array in wrappers #160
• Fix bug in masked autoregressive flow masking #162
• Add a mixture distribution #163

Full Changelog: v12.1.2...v12.2.0

v12.1.2

What's Changed

• Avoid doctest for now by @danielward27 in #156
• Update github workflows and version by @danielward27 in #157

Full Changelog: v12.1.1...v12.1.2

v12.1.1

What's Changed

• Fix param, static order in MaximumLikelihoodLoss by @danielward27 in #148
• Jaxtyping by @danielward27 in #154
• Update faq by @danielward27 in #153
• Add clipping for numerical stability by @danielward27 in #152

Full Changelog: v12.0.1...v12.1.1

v12.0.1

Change jax.tree -> jax.tree_util to maintain compatibility with earlier jax versions. Thanks @mdmould!

What's Changed

• Replace jax.tree API with jax.tree_util for backwards compatibility by @mdmould in #143
• Test for weight normalization by @danielward27 in #144
• Version bump by @danielward27 in #145

Full Changelog: v12.0.0...v12.0.1

v12.0.0

All the breaking changes in this major release were introduced in #142, so if any issues arise, check #142 to see if it gives a hint on how to update. If anyone has any issues, please open an issue :)

What's Changed

• Reshape by @danielward27 in #138
• Wrap bijections for converting transformed distributions to numpyro by @danielward27 in #139
• min scale for affine transformer by @danielward27 in #140
• Negative scales by @danielward27 in #141
• Wrap by @danielward27 in #142

Full Changelog: v11.3.0...v12.0.0

v11.3.0

A mix of additions/improvements/bug fixes, listed below. Thanks @mdmould for your help!

What's Changed

• Add guide for citing by @danielward27 in #126
• numpyro sample wrapper by @danielward27 in #127
• Add distribution_to_numpyro by @danielward27 in #128
• Exponential by @danielward27 in #129
• bugfix for concatenate array splitting by @mdmould in #131
• bugfix for concatenate bijection to allow jax tracing by @mdmould in #133
• end knots transformed by spline instead of identity by @mdmould in #134
• Update README.md typo by @danielward27 in #135
• Numpyro conditional bug fix by @danielward27 in #136
• Option to control returning best parameters vs last parameters in training by @danielward27 in #137

New Contributors

• @mdmould made their first contribution in #131

Full Changelog: v11.2.0...v11.3.0

v11.2.0

Release adds bisection search to invert the bijection in block neural autoregressive flows (BlockAutoregressiveNetwork).

Possible but unlikely breaking change:

• BlockAutoregressiveNetwork inverse methods will no longer raise a NotImplementedError, which could be a breaking change e.g. if a user relies on exception handling.

What's Changed

• Add bisection inverter by @danielward27 in #122
• Adapt bisection interval bug fix by @danielward27 in #123
• Handle exact solution in bisection search by @danielward27 in #125

Full Changelog: v11.1.0...v11.2.0

v11.1.0

Should be no breaking changes from v11.0.0. Adds LogNormal and Laplace.

What's Changed

• Update readme by @danielward27 in #119
• Add LogNormal by @danielward27 in #120
• add Laplace by @danielward27 in #121

Full Changelog: v11.0.0...v11.1.0

v11.0.0

The most important breaking changes (in importance order) are:

1. Renaming of flows

• All the flows have moved from being classes (with standard PascalCase naming) to functions with the standard snake_case naming.
• Reason for change: None of the flows implemented any methods or extended the base class AbstractTransformed other than forwarding basic information about the bijection, so the use of classes was not needed. Instead, flows now are functions that return a Transformed distribution.

2. Key word only arguments

• A lot of functions and classes now make use of key word only arguments to ensure more readable code (e.g. as TriangularAffine(loc, arr, lower=True) is much clearer than TriangularAffine(loc, arr, True).
• This may introduce errors along the lines of TypeError: function takes at most ... arguments (... given), but this should be straightforward to fix by converting to using key word arguments where required.

3. Abstract{Distribution/Bijection}

• Bijection has been renamed to AbstractBijection
• Distribution has been renamed to AbstractDistribution

4. Custom bijections

• Custom bijection implementations no longer need to manually validate input shapes, as the methods are automatically wrapped during an __init_subclass__ method in AbstractBijection. The bijection method _argcheck_and_cast has been removed as it is no longer needed.

What's Changed

• Bnaf Issue by @danielward27 in #103
• Tidy flow definitions and add merge_chains by @danielward27 in #104
• Numpyro by @danielward27 in #107
• Abstract var by @danielward27 in #108
• Make losses consistent by @danielward27 in #109
• Ensure int masks and ranks by @danielward27 in #110
• Pass params in step by @danielward27 in #111
• Flows to functions by @danielward27 in #112
• Add identity by @danielward27 in #113
• Add default shape to identity by @danielward27 in #114
• pypi trusted publishing by @danielward27 in #115
• Bijection checks by @danielward27 in #116
• Multivariate normal by @danielward27 in #117
• Docs by @danielward27 in #118

Full Changelog: v10.1.0...v11.0.0