Release Notes

This is the 0.4.0 release of TensorFlow Probability. It's tested and stable against TensorFlow 1.11.

Distributions & Bijectors

• All TF core Distributions now live in TFP (old ones are deprecated but still there)
• All TF core Bijectors now live in TFP (old ones are deprecated but still there)
• CDF support for Independent
• Seasonal effects and Sum components for Structural Time Series
• Add Half-Cauchy Distribution
• DCT (discrete cosine transform) Bijector
• Support dynamic shapes in BatchNormalization Bijector
• Add InverseGaussian Distribution
• Scaffolding for Bayesian Structural Time Series models
• Add VonMises Distribution
• Add Pareto distrubiton
• Implicit reparameterization for TruncatedNormal

Sampling & Inference

• Various improvements to BFGS and line search methods
• Add tape interceptor in Edward2
• GLM fitting with proximal gradient and L1/L2-regularization
• Support nested interceptors in Edward2

Documentation & Examples

• Update docs to point to tfp instead of tf/tf.contrib
• Credit card interest notebook
• Bayesian Neural Net (VGG, ResNet18) examples on CIFAR-10 data
• Disentangled Sequential Autoencoder example
• Eager-mode, single-chain No-U-Turn Sampler
• Grammar VAE example using Edward2
• Deep exponential family example using Edward2
• Latent Dirichlet Allocation example using Edward2
• Factorial Mixtures notebook

Huge thanks to all the contributors to this release!

• Asim Shankar
• Ben Zinberg
• Billy Lamberta
• Brian Patton
• Christopher Suter
• Copybara-Service
• cyrilchimisov
• Dave Moore
• davmre
• Dustin Tran
• Ian Langmore
• jiamingz
• Jiaming Zeng
• Jonathan J Hunt
• Joshua V. Dillon
• Keyon Vafa
• Mark Daoust
• Matthew McAteer
• mgorinova
• mhoffman
• Michael Figurnov
• Mike Dusenberry
• rif
• saxeas
• srvasude
• William D. Irons
• yunhao