Release notes

This is the 0.8 release of TensorFlow Probability. It is tested and stable against TensorFlow version 2.0.0 and 1.15.0rc1.

Change notes

• GPU-friendly "unrolled" NUTS: tfp.mcmc.NoUTurnSampler

  • Open-source the unrolled implementation of the No U-Turn Sampler.
  • Switch back to original U turn criteria in Hoffman & Gelman 2014.
  • Bug fix in Unrolled NUTS to make sure it does not lose shape for event_shape=1.
  • Bug fix of U turn check in Unrolled NUTS at the tree extension.
  • Refactor U turn check in Unrolled NUTS.
  • Fix dynamic shape bug in Unrolled NUTS.
  • Move NUTS unrolled into mcmc, with additional clean up.
  • Make sure the unrolled NUTS sampler handle scalar target_log_probs correctly.
  • Change implementation of check U turn to using a tf.while_loop in unrolled NUTS.
  • Implement multinomial sampling across tree (instead of Slice sampling) in unrolled NUTS.
  • Expose additional diagnostics in previous_kernel_results in unrolled NUTS so that it works with *_step_size_adaptation.

• MCMC

  • Modify the shape handling in DualAveragingStepSizeAdaptation so that it works with non-scalar event_shape.
  • support structured samples in tfp.monte_carlo.expectation.
  • Minor fix for docstring example in leapfrog_integrator

• VI

  • Add utilities for fitting variational distributions.
  • Improve Csiszar divergence support for joint variational distributions.
  • ensure that joint distributions are correctly recognized as reparameterizable by monte_carlo_csiszar_f_divergence.
  • Rename monte_carlo_csiszar_f_divergence to monte_carlo_variational_loss.
  • Refactor tfp.vi.csiszar_vimco_helper to expose useful leave-one-out statistical tools.

• Distributions

  • Added tfp.distributions.GeneralizedPareto
  • Multinomial and DirichletMultinomial samplers are now reproducible.
  • HMM samples are now reproducible.
  • Cleaning up unneeded conversion to tensor in quantile().
  • Added support for dynamic num_steps in HiddenMarkovModel
  • Added implementation of quantile() for exponential distributions.
  • Fix entropy of Categorical distribution when logits contains -inf.
  • Annotate float-valued Deterministic distributions as reparameterized.
  • Establish patterns which ensure that TFP objects are "GradientTape Safe."
  • "GradientTape-safe" distributions: FiniteDiscrete, VonMises, Binomial, Dirichlet, Multinomial, DirichletMultinomial, Categorical, Deterministic
  • Add tfp.util.DeferredTensor to delay Tensor operations on tf.Variables (also works for tf.Tensors).
  • Add probs_parameter, logits_parameter member functions to Categorical-like distributions. In the future users should use these new functions rather than probs/logits properties because the properties might be None if that's how the distribution was parameterized.

• Bijectors

  • Add log_scale parameter to AffineScalar bijector.
  • Added tfp.bijectors.RationalQuadraticSpline.
  • Add SoftFloor bijector. (Note: Known inverse bug WIP.)
  • Allow using an arbitrary bijector in RealNVP for the coupling.
  • Allow using an arbitrary bijector in MaskedAutoregressiveFlow for the coupling.

• Experimental auto-batching system: tfp.experimental.auto_batching

  • Open-source the program-counter-based auto-batching system.
  • Added tfp.experimental.auto_batching, an experimental system to recover batch parallelism across recursive function invocations.
  • Autobatched NUTS supports batching across consecutive trajectories.
  • Add support for field references to autobatching.
  • Increase the amount of Python syntax that "just works" in autobatched functions.
  • pop-push fusion optimization in the autobatching system (also recently did tail-call optimization but forgot to add a relnote).
  • Open-source the auto-batched implementation of the No U-Turn Sampler.

• STS

  • Support TF2/Eager-mode fitting of STS models, and deprecate build_factored_variational_loss.
  • Use dual averaging step size adaptation for STS HMC fitting.
  • Add support for imputing missing values in structural time series models.
  • Standardize parameter scales during STS inference.

• Layers

  • Add WeightNorm layer wrapper.
  • Fix gradients flowing through variables in the old style variational layers.
  • tf.keras.model.save_model and model.save now defaults to saving a TensorFlow SavedModel.

• Stats/Math

  • Add calibration metrics to tfp.stats.
  • Add output_gradients argument to value_and_gradient.
  • Add Geyer initial positive sequence truncation criterion to tfp.mcmc.effective_sample_size.
  • Resolve shape inconsistencies in PSDKernels API.
  • Support dynamic-shaped results in tfp.math.minimize.
  • ODE: Implement the Adjoint Method for gradients with respect to the initial state.

Huge thanks to all the contributors to this release!

• Alexey Radul
• Anudhyan Boral
• Arthur Lui
• Brian Patton
• Christopher Suter
• Colin Carroll
• Dan Moldovan
• Dave Moore
• Edward Loper
• Emily Fertig
• Gaurav Jain
• Ian Langmore
• Igor Ganichev
• Jacob Burnim
• Jeff Pollock
• Joshua V. Dillon
• Junpeng Lao
• Katherine Wu
• Mark Daoust
• Matthieu Coquet
• Parsiad Azimzadeh
• Pavel Sountsov
• Pavithra Vijay
• PJ Trainor
• prabhu prakash kagitha
• prakashkagitha
• Reed Wanderman-Milne
• refraction-ray
• Rif A. Saurous
• RJ Skerry-Ryan
• Saurabh Saxena
• Sharad Vikram
• Sigrid Keydana
• skeydan
• Srinivas Vasudevan
• Yash Katariya
• Zachary Nado