CHANGELOG.md

Changelog

0.16.5 - 2024-03-14

• Moved repository to GitHub.

0.16.4 - 2024-01-31

• Bump min rust version to 1.72
• Fix bug in dist::Beta that sometimes resulted in invalid cache setting alphaor beta.
• Add set_a and set_b methods to dist::Uniform

0.16.3 - 2024-01-23

• Implement Variance<f64>for mixture models whose component distribution type also implements Variance<f64> and Mean<f64>.
• Implement Variance<f32>for mixture models whose component distribution type also implements Variance<f32> and Mean<f32>.

0.16.2 - 2024-01-02

• Fix edge case in misc::logsumexp that would return NaN if the first value in the slice was -inf.

0.16.1 - 2023-12-14

• Performance improvements in misc::logsumexp

0.16.0 - 2023-06-07

• Fix bug in InvGaussianSuffStat
• Add ln_f_stat method to HasSuffStat trait. This method computes the log likelihood of the data represented by the statistic.
• Un-implement HasSuffStat<f64> for InverseGamma
• Add InvGammaSuffStat
• Add BetaSuffStat
• Cache ln(weights) for Mixture to improve ln_f computation performance

0.15.1 - 2023-05-22

• Fix bug in Empirical::draw that causes an index out of bounds error when the last last bin is drawn.

0.15.0 - 2023-03-07

• All structs and enums use snake_case naming.
• Dependency updates (thanks @bsull).
• Entropy impl for Bernoulli Mixuture.
• Corrected NormalInverseGamma Posterior.
• Corrected NormalInverseWishart Posterior.
• Added Kernel::reparameterize.
• Improve CI.

0.14.4 - 2022-02-16

• re-export nalgebra (flag arraydist)
• add reparameterize method to process::gaussian::Kernel, which will replace from_parmeters in 0.15.0
• Implement Error for NormalInvGammaError and NormalInvChiSquaredError

0.14.3 - 2022-02-16

• Remove dependency on peroxide

0.14.2 - 2022-01-31

• Fix some incorrect documentation
• Pass --all-features to docs.rs builds

0.14.1 - 2022-01-21

• Fix bug in computation of entropies for continuous mixture models

0.14.0 - 2022-01-05

• Fix typo: CategoricalError::EmptyWights to CategoricalError::EmptyWeights
• Allow Categorical::from_ln_weights to accept zero (-Inf) weights
• Update lru crate to 0.7 due to security vulnerability in 0.6.6

0.13.1 - 2021-08-14

• When using dist::Gamma to draw a dist::Poisson, the Poisson rate parameter is shifted to f64::EPSILON in the event of underflow.

0.13.0 - 2021-07-15

• Clippy lints
• Use &'a [T] instead of &'a Vec<T> in DataOrSuffStat
• Use &[T] instead of &Vec<T> in GewekeTester
• Renamed BesselIvError variants
• Implement Rv<usize> for Poisson
• Additional caching of expensive computations
• Estimate Gaussian mixture entropy with Gauss–Legendre quadrature for a 10x speedup.
• Place ndarray dependent distributions behind arraydist flag.
• Fix bug the caused weights to be computed incorrectly when using Mixture::combine where one or more of the inputs was an empty mixture

0.12.0 - 2021-05-05

• Added Inverse X^2 distribution
• Added Scaled Inverse X^2 distribution (InvChiSquared)
• Added Inverse Gaussian distribution
• Added Inverse Gaussian sufficient statistic
• Added Normal Inverse X^2 (NormalInvChiSquared) distribution as prior for Gaussian
• Added Inverse X^2 distribution (InvChiSquared)
• Implemented From instead of Into for sufficient statistic converters to allow the more explicit/ergonomic From conversions.
• Improved some error messages in NormalGamma
• Caching of a normalizing constant in BetaBinomial
• Lots of testing verifying proper conjugate prior behavior
• Added Gaussian prior geweke tests and example
• More robust dpgmm example -- can use different priors
• Version updates
• Misc styling and warning fixes

0.11.1 - 2021-02-12

• Added Normal Inverse Gamma (NormalInvGamma) distribution as prior for Gaussian

0.11.0 - 2020-11-20

• Added _with_cache variants to ConjugatePrior ln_m and ln_pp methods for use cases where these methods are called many times in different data without changing the underlying distribution. This also adds two more associated types to the trait (LnMCache and LnPpCache), and one method each to compute the caches.
• Remove all references to serde_derive crate
• Gaussian process improvements including new kernels

0.10.4 - 2020-10-09

• Added Gaussian processes

0.10.3 - 2020-10-07

• Updated math in NormalGamma and GaussianSuffStat to reduce rounding errors

0.10.2 - 2020-09-07

• Categorical datum converters work like standard usize-to-boolean casting where 0 is false, and anything greater than 0 is true.

0.10.1 - 2020-08-10

• Add from_parts_unchecked method for constructing sufficient statistic since there was no way of manipulating statistic's fields or creating them manually.

0.10.0 - 2020-06-02

• Implement Poisson Mode
• Store GaussianSuffStat in a more numerically stable way. The extra numerical stability comes at the cost of slower observe and forget. Serialization on GaussianSuffStat has changed.

0.9.3 - 2020-05-30

• Entropy for Gaussian mixtures more robust to models with dispersed low-variance components
• Categorical::new_unchecked is now public

0.9.2 - 2020-05-23

• Benchmarks
• implement Entropy and KlDivergence for Poisson
• Implement QuadBounds and Entropy for Mixture<Poisson>
• Implement Mean for mixtures of distributions that implement Mean<f64> or Mean<f32>
• Rv trait has sample_stream method that produces a never-ending iterator of random numbers.

0.9.1 - 2020-05-16

• Remove printlns

0.9.0 - 2020-05-16

• Refactor errors to provide more information
• Errors implement the std::error::Error trait
• Fixed a bug with rv::dist::Mixture that would compute means incorrectly
• MvGaussian::dims renamed to MvGaussian::ndims
• More aggressive caching of log likelihood terms
• Cached terms no longer unnecessarily computed on Clone
• Remove dependency on getset
• Setters do input validation unless prefixed by _unchecked.
• feature serde_support renamed to serde1

0.8.3 - 2020-05-05

• Added Negative Binomial distribution
• Added Skellam Distribution
• Add try_quad function for numerical integration of functions that return Result<f64, E>.

0.8.2 - 2020-02-12

• Remove dependency on quadrature crate in favor of hand-rolled adaptive Simpson's rule, which handles multimodal distributions better.