v0.6.1

HiddenMarkovModels v0.6.1

Diff since v0.6.0

Merged pull requests:

• Clarify discrete aspect in README (#117) (@adannenberg)
• Functional API for applying transition matrix (#121) (@gdalle)
• Restrict single-threading to single sequence (#122) (@gdalle)

Closed issues:

• Linear memory Baum-Welch? (#7)

v0.6.0

HiddenMarkovModels v0.6.0

Breaking changes

In-place single-sequence routines were hidden from the public API. Most user code will not be affected.

Performance improvements

• Multithreading is now deactivated if seq_ends is passed as an NTuple. This helps remove overhead for very few sequences (especially for a single one).

Diff since v0.5.4

Merged pull requests:

• Decompose forward function into initialize, predict, update (#105) (@THargreaves)
• Update inference routines (#116) (@gdalle)

Closed issues:

• Stop testing against HMMBase (#109)
• Avoid @threads for a single sequence (#110)

v0.5.4

HiddenMarkovModels v0.5.4

Diff since v0.5.3

Merged pull requests:

• Minor optimizations and simd (#108) (@gdalle)
• Make HMMBase optional in the test suite (#111) (@gdalle)
• Fix loglikelihood increase check in Baum-Welch (#112) (@gdalle)
• Disable multithreading when seq_ends is passed as a tuple (#113) (@gdalle)
• Bump version to 0.5.4 (#114) (@gdalle)
• Improve code coverage (#115) (@gdalle)

Closed issues:

• Avoid repeated transposition when using time-homogenous transition matrix (#106)

v0.5.3

HiddenMarkovModels v0.5.3

Diff since v0.5.2

Merged pull requests:

• Allow heterogeneous distributions (#102) (@gdalle)

Closed issues:

• Allow dist to be a Vector{Distribution{Univariate, Continuous}} in Baum Welch (#101)

v0.5.2

HiddenMarkovModels v0.5.2

Diff since v0.5.1

Merged pull requests:

• Add JOSS citation (#98) (@gdalle)
• Allow different types for elementwise log (#100) (@gdalle)

Closed issues:

• Allow different types for transition matrix and elementwise log (#99)

v0.5.1

HiddenMarkovModels v0.5.1

Diff since v0.5.0

Merged pull requests:

• Final benchmarks (#94) (@gdalle)
• Fix benchmark reproducibility (#95) (@gdalle)
• Remove precompilation (#97) (@gdalle)

Closed issues:

• Benchmarks (#2)
• Porting issues from HMMBase.jl (#92)

v0.5.0

HiddenMarkovModels v0.5.0

Diff since v0.4.1

See #93 for the user-facing changes

Merged pull requests:

• Fix benchmarks (#90) (@gdalle)
• Speed up viterbi (#93) (@gdalle)

v0.4.1

HiddenMarkovModels v0.4.1

Diff since v0.4.0

Merged pull requests:

• Document fit! (#88) (@gdalle)

v0.4.0

HiddenMarkovModels v0.4.0

Diff since v0.3.1

Breaking changes

See the API reference for more details

Multiple sequences:

• Revamped multiple sequence formatting: multiple sequences are concatenated for the input and for the output

Removals:

• Removed alias HMMs for the package name
• Removed everything related to (Abstract)MarkovChain
• Removed PermutedHMM

Renamings:

• initial_distribution(hmm) became initialization(hmm)
• obs_distribution(hmm, i) became obs_distributions(hmm)
• in baum_welch, check_loglikelihood_increasing became loglikelihood_increasing

Storage:

• in ForwardBackwardStorage, ξ is now a vector of abstract matrices instead of a 3-dimensional array

Method signatures:

• Revamped multiple sequence formatting: algorithm(hmm, obs_seqs, nb_seqs) became algorithm(hmm, obs_seq, [control_seq]; seq_ends).
• Modified the arguments for in-place fitting: fit!(hmm, init_count, trans_count, obs_seq, state_marginals) became fit!(hmm, fb_storage, obs_seq, control_seq; seq_ends)

Return types;

• Revamped multiple sequence formatting: results are now concatenated, and loglikelihoods are computed sequencewise

Merged pull requests:

• Stdlib compat + Documenter v1 + bump to 0.3.1 (#54) (@gdalle)
• Add TagBot workflow (#55) (@gdalle)
• Move Python deps from pip to conda (#57) (@gdalle)
• Fix docs benchmark (#58) (@gdalle)
• Complete revamp (#59) (@gdalle)
• Minor fixes (#60) (@gdalle)
• Reactivate multithreaded BW (#62) (@gdalle)
• Put benchmarks inside HMMBenchmark subpackage (#63) (@gdalle)
• Move DependaBot (#64) (@gdalle)
• Split (#66) (@gdalle)
• Simplify docstrings (#67) (@gdalle)
• Typo (#68) (@gdalle)
• Benchmarks with Python updated (#69) (@gdalle)
• More benchmarks (#70) (@gdalle)
• Don't tune forward (#71) (@gdalle)
• Threads (#72) (@gdalle)
• Remove deps and fix Enzyme (#73) (@gdalle)
• Remove pre 1.9 stuff (#74) (@gdalle)
• Switch to StableRNGs (#75) (@gdalle)
• Clean up benchmarks (#77) (@gdalle)
• Add multithreading (#78) (@gdalle)
• 1 OpenBLAS thread (#79) (@gdalle)
• Parallelize fitting (#80) (@gdalle)
• Fairer Python benchmarks (#82) (@gdalle)
• Format (#83) (@gdalle)
• Correct benchmarks (#84) (@gdalle)
• Add codecov token and documenter key (#85) (@gdalle)
• Better docs and numerical stability (#87) (@gdalle)

Closed issues:

• Chain rule (#3)
• Benchmarks on different distributions (#8)
• Add SIMD (#9)
• Add links to (and comparisons with) other related packages (#14)
• Make modification to allow support for Heterogeneous (Periodic) Hidden Markov Models (#52)
• Clarify docs on the arguments of fit! (#53)
• Fix benchmarks (#61)
• MethodError Using baum_welch with MvLogNormal Emission Distributions (#76)
• Test numerical stability (#86)

v0.3.1

What's Changed

• Local benchmarks leveraging parallelism by @gdalle in #50
• Stdlib compat + Documenter v1 + bump to 0.3.1 by @gdalle in #54

Full Changelog: v0.3.0...v0.3.1