1.6.0

Breaking changes

• Update to PyTorch 1.8 release (required).
• Enable validation by default #2701. To disable globally call pyro.enable_validation(False); or disable locally to one distribution via e.g. Normal(loc, scale, validate_args=False).
• Switch from LKJCorrCholesky distribution to upstream LKJCholesky distribution #2771.

New Tutorials

• Product LDA for probabilistic topic modeling #2729.

New features

• Support masked conditioning via pyro.sample(..., obs_mask=...) #2772.
• Distributions:
  • Added a Distribution.infer_shapes() method for static shape analysis #2739.
  • Added ProjectedNormal for circular and spherical data (with more functionality than VonMises) #2736.
  • Added AffineBeta distribution over arbitrary intervals #2735 by @ordabayevy.
  • Added SpanningTree properties .mode, .edge_mean #2727
  • Added OneOneMatching and OneTwoMatching distributions to approximately marginalize over bipartite matchings #2707, #2697.
• Reparametrizers:
  • Added ProjectedNormalReparam for inference with ProjectedNormal distributions #2736.
  • Support more batch distributions in HaarReparam #2731.
• Constraints:
  • Added a Constraint.event_dim property #2753.
  • Added positive_ordered_vector, corr_matrix #2762
  • Added sphere #2736.
  • Added softplus_positive and softplus_lower_cholesky constraints with numerically stable SoftplusTransform and SoftplusLowerCholeskyTransform #2767.
• Added a TraceMarkovEnum_ELBO and VectorizedMarkovMessenger for parallel scan enumeration #2703, #2703 by @ordabayevy.
• Added an experimental autoname handler #2745 by @ordabayevy.
• Added goodness of fit tests for testing new distribution implementations #2738.
• Support MaskedDistribution in contrib.forecast #2709.

Bugfixes

• Many fixes to distribution shapes #2759, #2746, #2739.
• Detach AutoGuide.median() and .quantiles() methods #2743.