* Change `pyprob.sample` arguments ordering to `distribution, name` followed by other optional arguments * Disable progress information printing for `verbosity==0` * Raise error if MCMC posterior is run with an empty trace (model with no `pyprob.sample` statement) * Density estimation of Empiricals with Gaussian mixtures * String representation of distributions are executable as Python code, making it easy to copy and instantiate distributions
1.2.2
* Support indexing traces by variable name, e.g., `trace['mean']` for a variable with prior `pyprob.sample(..., name='mean') * Issue warnings using Python `warnings` module * Add von Mises distribution * Improve autocorrelation diagnostic * `Empirical.from_distribution` as a shorthand to generate empirical distributions from other distributions * `Empirical.median` to compute median values * `model.posterior` raises error if a `None` observation is given
1.2.1
* Support slicing of on-disk Empiricals * Diagnostics: per-variable trace plots * Diagnostics: improve autocorrelation plots * Skewness, kurtosis for Empiricals * Support PyTorch version >= 1.5.1 (skip 1.5.0 due to a pickle-related bug in PyTorch)
1.2.0
* Improve diagnostic address graphs to show distribution type * Improve Gelman-Rubin diagnostic address selection * Added `min_address_count` and `exclude_addresses_regex` options to diagnostic graphs * Support `pyprob.sample` to sample from distributions even when not currently recording a trace (e.g., via `model.prior` or `model.posterior`) * Support weighted resampling of Empirical with `min_index`, `max_index` * Rename `_log_weights`, `_values` to `log_weights`, `values` in `Empirical` * Fix bug in Binomial string representation * Fix bug in diagnostics for variables without value * Fix bug in Empirical `resample` not correctly handling optional `min_index`, `max_index` arguments * Fix bug in Empirical destructor
1.1.0
* Added support for Bernoulli, Beta, Exponential, Gamma, LogNormal, Binomial, Weibull distributions * Fixed minor bug where concatenated Empirical distributions on disk were not checked correctly for uniform weights