ChangeLog.txt

- 0.3.0:
    - Added: poisson_batchify_data for Poisson subsampling of data sets
    - Added: DPSVI.update now accepts a mask indicating padded samples in the batch.
- 0.2.0:
    - ! Renamed main package 'dppp' to 'd3p' !
    - Upgraded stable numpyro dependency to 0.8.0, removed supported for earlier versions
    - Removed support for jax versions prior to v0.2.20.
    - Removed support for Python 3.6
    - Change: DPSVI now errors if model has mutable state
    - Fix: DPSVI perturbation was too large (scaled by minibatch size)
    - Fix: dputil.approximate_sigma could enter an infinite loop for some parameters
    - Feature: Fast minibatch sampling using Feistel-based CUDA-Shuffle
    - Feature: Cryptographically-secure random number generator for DP noise.
    - Removed: num_obs_total argument to DPSVI.
    - Removed: TunableSVI class.
    - Removed: minibatch context handler - superseded by numpyro plate primitive.
    - Moved: Routines for sampling from posterior/prior in d3p.svi.
- 0.1.0: Initial not-quite-a-release