Release notes

This is the 0.12.1 release of TensorFlow Probability. It is tested and stable against TensorFlow version 2.4.0.

Change notes

NOTE: Links point to examples in the TFP 0.12.1 release Colab.

Bijectors:

• Add implementation of GLOW at tfp.bijectors.Glow.
• Add RayleighCDF bijector.
• Add Ascending bijector and deprecate Ordered.
• Add optional low parameter to the Softplus bijector.
• Enable ScaleMatvecLinearOperator bijector to wrap blockwise LinearOperators to form a multipart bijectors.
• Allow passing kwargs to Blockwise.
• Bijectors now share a global cache, keyed by the bijector parameters and the value being transformed.

Distributions:

• BREAKING: Remove deprecated HiddenMarkovModel.num_states property.
• BREAKING: Change the naming scheme of un-named variables in JointDistributions.
• BREAKING: Remove deprecated batch_shape and event_shape arguments of TransformedDistribution.
• Add Skellam distribution.
• JointDistributionCoroutine{AutoBatched} now uses namedtuples as the sample dtype.
• von-Mises Fisher distribution now works for dimensions > 5 and implements VonMisesFisher.entropy.
• Add ExpGamma and ExpInverseGamma distributions.
• JointDistribution*AutoBatched now support (reproducible) tensor seeds.
• Add KL(VonMisesFisher || SphericalUniform).
• Added Distribution.parameter_properties method.
• experimental_default_event_space_bijector now accepts additional arguments to pin some distribution parts.
• Add JointDistribution.experimental_pin and JointDistributionPinned.
• Add NegativeBinomial.experimental_from_mean_dispersion method.
• Add tfp.experimental.distribute, with DistributionStrategy-aware distributions that support cross-device likelihood computations.
• HiddenMarkovModel can now accept time varying observation distributions if time_varying_observation_distribution is set.
• Beta, Binomial, and NegativeBinomial CDF no longer returns nan outside the support.
• Remove the "dynamic graph" code path from the Mixture sampler. (Mixture now ignores the use_static_graph parameter.)
• Mixture now computes standard deviations more accurately and robustly.
• Fix incorrect nan samples generated by several distributions.
• Fix KL divergence between Categorical distributions when logits contain -inf.
• Implement Bernoulli.cdf.
• Add a log_rate parameter to tfd.Gamma.
• Add option for parallel filtering and sampling to LinearGaussianStateSpaceModel.

MCMC:

• Add tfp.experimental.mcmc.ProgressBarReducer.
• Update experimental.mcmc.sample_sequential_monte_carlo to use new MCMC stateless kernel API.
• Add an experimental streaming MCMC framework that supports computing statistics over a (batch of) Markov chain(s) without materializing the samples. Statistics supported (mostly on arbitrary functions of the model variables): mean, (co)variance, central moments of arbitrary rank, and the potential scale reduction factor (R-hat). Also support selectively tracing history of some but not all statistics or model variables. Add algorithms for running mean, variance, covariance, arbitrary higher central moments, and potential scale reduction factor (R-hat) totfp.experimental.stats.
• untempered_log_prob_fn added as init kwarg to ReplicaExchangeMC Kernel.
• Add experimental support for mass matrix preconditioning in Hamiltonian Monte Carlo.
• Add ability to temper part of the log prob in ReplicaExchangeMC.
• tfp.experimental.mcmc.{sample_fold,sample_chain} support warm restart.
• even_odd_swap exchange function added to replica_exchange_mc.
• Samples from ReplicaExchangeMC can now have a per-replica initial state.
• Add omitted n/(n-1) term to tfp.mcmc.potential_scale_reduction_factor.
• Add KernelBuilder and KernelOutputs to experimental.
• Allow tfp.mcmc.SimpleStepSizeAdaptation and DualAveragingStepSizeAdaptation to take a custom reduction function.
• Replace make_innermost_getter et al. with tfp.experimental.unnest utilities.

VI:

• Add automated construction of ASVI surrogate posteriors.

Math + Stats:

• Add tfp.math.bessel_ive, tfp.math.bessel_kve, tfp.math.log_bessel_ive.
• Add optional weights to tfp.stats.histogram.
• Add tfp.math.erfcinv.
• Add tfp.math.reduce_log_harmonic_mean_exp.

Other:

• Add tfp.math.psd_kernels.GeneralizedMaternKernel (generalizes MaternOneHalf, MaternThreeHalves and MaternFiveHalves).
• Add tfp.math.psd_kernels.Parabolic.
• Add tfp.experimental.unnest utilities for accessing nested attributes.
• Enable pytree flattening for TFP distributions in JAX
• More careful handling of nan and +-inf in {L-,}BFGS.
• Remove Edward2 from TFP. Edward2 is now in its own repo at https://github.com/google/edward2 .
• Support vector-valued offsets in sts.Sum.
• Make DeferredTensor actually defer computation under JAX/NumPy backends.

Huge thanks to all the contributors to this release!

• Adrian Buzea
• Alexey Radul
• Ben Lee
• Ben Poole
• Brian Patton
• Christopher Suter
• Colin Carroll
• Cyril Chimisov
• Dave Moore
• Du Phan
• Emily Fertig
• Eugene Brevdo
• Federico Tomasi
• François Chollet
• George Karpenkov
• Giovanni Palla
• Ian Langmore
• Jacob Burnim
• Jacob Valdez
• Jake VanderPlas
• Jason Zavaglia
• Jean-Baptiste Lespiau
• Jeff Pollock
• Joan Puigcerver
• Jonas Eschle
• Josh Darrieulat
• Joshua V. Dillon
• Junpeng Lao
• Kapil Sachdeva
• Kate Lin
• Kibeom Kim
• Luke Metz
• Mark Daoust
• Matteo Hessel
• Michal Brys
• Oren Bochman
• Padarn Wilson
• Pavel Sountsov
• Peter Hawkins
• Rif A. Saurous
• Ru Pei
• ST John
• Sharad Vikram
• Simeon Carstens
• Srinivas Vasudevan
• Tom O'Malley
• Tomer Kaftan
• Urs Köster
• Yash Katariya
• Yilei Yang