0.8.0

Breaking changes

Switch to softplus transforms for autoguide scales (thanks to experiments performed by @vitkl).

New Features

• New autoguide: AutoDAIS leverages HMC and annealed importance sampling within a variational inference framework
• New distributions: MixtureSameFamily, and directional distributions SineBivariateVonMises, SineSkewed
• New constraints: l1_ball for vectors with L1 norm less than 1
• New transforms: L1BallTransform, SimplexToOrderedTransform, ScaledUnitLowerCholeskyTransform
• #1116 New format_shapes utility to interpret the shapes of random variables/plates in a model.
• #1109 Allow direct use of TFP distributions in numpyro.sample
• New tutorials and examples:
  • Principled prior with Dirichlet distribution for Ordinal Regression case study
  • Horseshoe regression
  • Bad posterior geometry and how to deal with it

Enhancements and Bug Fixes

• #1108 Avoid numerical problems when using BernoulliProbs
• #1118 Recommend AutoNormal guide when hessian in AutoLaplace is singular
• #1126 Smarter warning about discrete inference in SVI models
• #1136 Support to use SA sampler with arviz
• #1139 Document Poisson is_sparse argument
• #1140 Make Sigmoid and StickBreakingTransform more stable
• #1149 Raise value error if num_steps bad in svi.run
• #1162 Use black[jupyter] in notebooks

This release is composed of great contributions and feedback from the Pyro community: @MarcoGorelli @OlaRonning @d-diaz @quattro @svilupp @peterroelants @prashjet @freddyaboulton @tcbegley @julianstastny @alexlyttle and many others. Thank you!