Release Notes

This is the 0.3.0 release of TensorFlow Probability. It's tested and stable against TensorFlow 1.10.

Distributions & Bijectors

Add the LKJ distribution on correlation matrices.
Add GammaGamma distribution.
Adds the VonMisesFisher distribution over points on the unit hypersphere.
Add CholeskyToInvCholesky bijector.
Added reparametrizable TruncatedNormal
Add tfp.bijectors.Transpose.
Add tanh bijection.
Introduce GaussianProcessRegressionModel
Introduce GaussianProcess distribution
Gamma distribution and the derived distributions (Beta, Dirichlet, Student's t, inverse Gamma) are fully reparameterized.
Add low and high as properties to quantized distribution.
Collapse WishartCholesky and WishartFull into a single Wishart distribution that takes either a scale or a scale_cholesky argument.
Add adjoint arg to tfp.bijectors.Affine.

Enable nested interceptors in Edward2.
Provide interface for controlling the number of HMC iterations during which to adapt the step size.
Added support for dynamic shapes in the slice sampler.
Make HMC more efficient and usable for MCEM.
- Allow stop_gradient to be applied as new state is built (thus enabling recycling kernel_results.accepted.target_log_prob).
  - Add hook for user defined adaptive step size code and provide default implementation.
Added implementation of the Nelder Mead derivative free optimization method.
Add tfp.math.random_rayleigh.

Add Edward2 README.md.
Add migration guide from Edward to TFP.
Add documentation matching tfp-0.2 release.
Add colab example which compares fitting HLM's between TF distributions, Stan, and R. Colab was written in collaboration with safyan@.
Added a preliminary version of a Probabilistic PCA Edward 2 example, and changed the BUILD file accordingly.
Latent Dirichlet Allocation for 20 newsgroups dataset.
A detailed case study in using TensorFlow Probability for estimating a covariance matrix.