hmm-craft 0.4.0__tar.gz

hmm_craft-0.4.0/.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,43 @@
+name: Bug report
+description: Something gives a wrong result, raises, or crashes.
+body:
+  - type: markdown
+    attributes:
+      value: |
+        This fork has one maintainer and no response time. A report with a
+        short example that runs as posted is looked at first. Questions about
+        using hidden Markov models are not answered here.
+  - type: textarea
+    id: example
+    attributes:
+      label: Example
+      description: >
+        A short, self-contained example that fails when pasted into a new empty
+        file. Generate the data with a seeded random number generator inside
+        the example; do not attach data files.
+      render: python
+    validations:
+      required: true
+  - type: textarea
+    id: observed
+    attributes:
+      label: What happens, and what you expected
+    validations:
+      required: true
+  - type: textarea
+    id: versions
+    attributes:
+      label: Versions
+      description: hmm-craft, Python, NumPy, SciPy, scikit-learn, and the operating system.
+    validations:
+      required: true
+  - type: dropdown
+    id: upstream
+    attributes:
+      label: Does the same example also fail on hmmlearn 0.3.3?
+      options:
+        - "Yes"
+        - "No"
+        - Not tried
+    validations:
+      required: true

hmm_craft-0.4.0/.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1 @@
+blank_issues_enabled: false

hmm_craft-0.4.0/.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,22 @@
+name: Feature request
+description: Ask for something hmm-craft does not do.
+body:
+  - type: markdown
+    attributes:
+      value: |
+        A feature request without a pull request may be closed.
+  - type: textarea
+    id: request
+    attributes:
+      label: What you want, and why
+    validations:
+      required: true
+  - type: dropdown
+    id: pr
+    attributes:
+      label: Will you send a pull request?
+      options:
+        - "Yes"
+        - "No"
+    validations:
+      required: true

hmm_craft-0.4.0/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,19 @@
+## What changes
+
+## Does it change the result of a fit?
+
+- [ ] No.
+- [ ] Yes. It is behind an option whose default is hmmlearn 0.3.3's behaviour.
+
+## Use of AI tools
+
+- [ ] No AI tool helped write this change.
+- [ ] An AI tool helped: name it here. Each commit it helped write ends with
+      `Assisted-by: <tool>` (see `CONTRIBUTING.md`).
+
+## Checklist
+
+- [ ] A test that fails without the change.
+- [ ] `python -m pytest --pyargs hmmcraft.tests doc` passes.
+- [ ] No Python line is 80 characters or longer.
+- [ ] Every commit is signed off (`git commit -s`).

hmm_craft-0.4.0/.github/dependabot.yml

@@ -0,0 +1,14 @@
+# Proposes newer releases of the pinned GitHub Actions once a month, in one
+# pull request. Build requirements (ci/*.txt) are updated by hand.
+version: 2
+updates:
+  - package-ecosystem: github-actions
+    directory: /
+    schedule:
+      interval: monthly
+    groups:
+      actions:
+        patterns: ["*"]
+    commit-message:
+      prefix: chore
+    labels: []

hmm_craft-0.4.0/.github/workflows/ci.yml

@@ -0,0 +1,80 @@
+name: ci
+
+# The suite and the bit-equality check against hmmlearn 0.3.3, in two cells:
+# the declared minimums ("floor") and the newest wheels ("latest"). Wheels and
+# the sdist are built and tested by wheels.yml. The weekly run shows within a
+# week a new numpy, scipy or scikit-learn release that breaks the suite, or
+# under which this package's fits differ from hmmlearn 0.3.3's (both are
+# installed at the same versions). On a pull request, every commit must be
+# signed off.
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+  workflow_dispatch:
+  schedule:
+    - cron: "47 3 * * 1"
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ci-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  style:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 5
+    steps:
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      with:
+        persist-credentials: false
+    - name: No Python line of 80 characters or more
+      run: |
+        shopt -s globstar &&
+        ! grep -E '.{80}' -- **/*.py
+
+  sign-off:
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-24.04
+    timeout-minutes: 5
+    steps:
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      with:
+        fetch-depth: 0
+        persist-credentials: false
+    - name: Every commit signed off (DCO)
+      env:
+        BASE: ${{ github.event.pull_request.base.sha }}
+        HEAD: ${{ github.event.pull_request.head.sha }}
+      # The base branch's copy of the script, so a pull request that edits
+      # the script is checked by the copy it does not change.
+      run: |
+        git show "$BASE:scripts/check_signoff.py" > "$RUNNER_TEMP/check_signoff.py"
+        python3 -I "$RUNNER_TEMP/check_signoff.py" "$BASE" "$HEAD"
+
+  bit-equality:
+    name: bit equality (${{ matrix.cell }})
+    runs-on: ubuntu-24.04
+    timeout-minutes: 40
+    strategy:
+      fail-fast: false
+      matrix:
+        cell: [floor, latest]
+    env:
+      HMMCRAFT_SEED: "0"
+      OMP_NUM_THREADS: "1"
+      OPENBLAS_NUM_THREADS: "1"
+      MKL_NUM_THREADS: "1"
+    steps:
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      with:
+        fetch-depth: 0
+        persist-credentials: false
+    - uses: actions/setup-python@6d4fd22b20465ccc0f7354476588e8df7c9ca0e0  # v7.1.0
+      with:
+        python-version: "3.12"
+    - name: Suite and bit equality against hmmlearn 0.3.3
+      run: scripts/ci_bit_equality.sh "${{ matrix.cell }}" "$RUNNER_TEMP/check"

hmm_craft-0.4.0/.github/workflows/release.yml

@@ -0,0 +1,55 @@
+name: release
+
+# Publishes to PyPI by trusted publishing: no stored credential exists in this
+# repository. A tag "v<version>" builds the 20 wheels and the sdist through
+# wheels.yml; the publish job then waits for the maintainer's approval of the
+# "pypi" environment before it uploads. setuptools-scm takes the version from
+# the tag, and the publish job checks that every file carries it.
+
+on:
+  push:
+    tags:
+      - "v[0-9]*"
+
+permissions: {}
+
+jobs:
+  build:
+    permissions:
+      contents: read
+    uses: ./.github/workflows/wheels.yml
+    with:
+      full: true
+
+  publish:
+    needs: build
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    environment:
+      name: pypi
+      url: https://pypi.org/project/hmm-craft/
+    permissions:
+      id-token: write
+    steps:
+    - uses: actions/download-artifact@9000827ccba6bdab643e8b6fd33ac0654aef8333  # v8.0.2
+      with:
+        pattern: dist-*
+        path: dist
+        merge-multiple: true
+    - name: Check what is about to be uploaded
+      env:
+        TAG: ${{ github.ref_name }}
+      run: |
+        version="${TAG#v}"
+        find dist -type f -print
+        test "$(find dist -type f -name '*.whl' | wc -l)" -eq 20
+        test "$(find dist -type f -name '*.tar.gz' | wc -l)" -eq 1
+        for f in dist/*; do
+          case "$(basename "$f")" in
+            "hmm_craft-$version-"*.whl | "hmm_craft-$version.tar.gz") ;;
+            *) echo "$f does not carry version $version" >&2; exit 1 ;;
+          esac
+        done
+    - uses: pypa/gh-action-pypi-publish@7245906fb61fa33a13554f9d5fb856fe61758086  # v1.14.3
+      with:
+        packages-dir: dist/

hmm_craft-0.4.0/.github/workflows/wheels.yml

@@ -0,0 +1,129 @@
+name: wheels
+
+# 20 wheels and the sdist. A pull request builds and tests the Linux x86_64
+# cells only; a manual run, the weekly schedule and a release build everything.
+# Every wheel is tested after it is built, from the installed wheel, so a cell
+# that cannot run the tests ships no wheel. Build requirements are pinned with
+# hashes in ci/build_requirements.txt and installed before each build; nothing
+# fetches them at whatever version is current.
+
+on:
+  pull_request:
+  workflow_dispatch:
+  schedule:
+    - cron: "17 3 * * 1"
+  workflow_call:
+    inputs:
+      full:
+        description: Build every platform, not only Linux x86_64.
+        type: boolean
+        default: true
+
+permissions:
+  contents: read
+
+concurrency:
+  group: wheels-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  plan:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 5
+    outputs:
+      matrix: ${{ steps.matrix.outputs.matrix }}
+    steps:
+    - id: matrix
+      env:
+        FULL: ${{ inputs.full || github.event_name == 'workflow_dispatch' || github.event_name == 'schedule' }}
+      run: |
+        cells='{"name":"linux-x86_64","os":"ubuntu-24.04","archs":"x86_64"}'
+        if [ "$FULL" = true ]; then
+          cells="$cells"',{"name":"linux-aarch64","os":"ubuntu-24.04-arm","archs":"aarch64"}'
+          cells="$cells"',{"name":"macos-arm64","os":"macos-15","archs":"arm64"}'
+          cells="$cells"',{"name":"macos-x86_64","os":"macos-15-intel","archs":"x86_64"}'
+          cells="$cells"',{"name":"windows-amd64","os":"windows-2025","archs":"AMD64"}'
+        fi
+        echo "matrix={\"include\":[$cells]}" >> "$GITHUB_OUTPUT"
+
+  wheels:
+    name: wheels (${{ matrix.name }})
+    needs: plan
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJSON(needs.plan.outputs.matrix) }}
+    steps:
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      with:
+        fetch-depth: 0
+        persist-credentials: false
+    - uses: pypa/cibuildwheel@f9b966cfadf4109556cadc5bb39ddb72d0ce664e  # v4.3.5
+      env:
+        # CPython 3.12 to 3.15 with the GIL; no free-threaded, musllinux or
+        # 32-bit builds. Linux wheels are manylinux_2_28 (cibuildwheel's
+        # default image).
+        CIBW_BUILD: "cp312-* cp313-* cp314-* cp315-*"
+        CIBW_SKIP: "*-musllinux_*"
+        CIBW_ARCHS: ${{ matrix.archs }}
+        CIBW_BUILD_FRONTEND: "pip; args: --no-build-isolation"
+        CIBW_BEFORE_BUILD: >-
+          python -m pip install --only-binary :all: --require-hashes
+          -r {project}/ci/build_requirements.txt
+        CIBW_TEST_REQUIRES: pytest
+        CIBW_TEST_ENVIRONMENT: HMMCRAFT_SEED=0
+        # The project's pytest configuration turns a DeprecationWarning from
+        # the package into an error. The installed wheel is tested, not the
+        # tree.
+        CIBW_TEST_COMMAND: >-
+          python -I -m pytest -c {project}/pyproject.toml -p no:cacheprovider
+          --pyargs hmmcraft.tests {project}/doc
+    - uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9  # v7.0.2
+      with:
+        name: dist-wheels-${{ matrix.name }}
+        path: wheelhouse/*.whl
+        if-no-files-found: error
+
+  sdist:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 30
+    env:
+      HMMCRAFT_SEED: "0"
+    steps:
+    - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+      with:
+        fetch-depth: 0
+        persist-credentials: false
+    - uses: actions/setup-python@6d4fd22b20465ccc0f7354476588e8df7c9ca0e0  # v7.1.0
+      with:
+        python-version: "3.12"
+    - name: Build the sdist
+      run: |
+        python -m pip install --only-binary :all: --require-hashes \
+          -r ci/tool_requirements.txt
+        python -m pip install --only-binary :all: --require-hashes \
+          -r ci/build_requirements.txt
+        python -m build --sdist --no-isolation
+        ls -l dist
+        tar tzf dist/*.tar.gz > sdist_files.txt
+        for f in ext/_hmmc.cpp pyproject.toml LICENSE.txt AUTHORS.rst; do
+          grep -q "/$f\$" sdist_files.txt || { echo "missing $f" >&2; exit 1; }
+        done
+    - name: Install the sdist and run the suite
+      run: |
+        python -m venv "$RUNNER_TEMP/sdist-env"
+        env_python="$RUNNER_TEMP/sdist-env/bin/python"
+        "$env_python" -m pip install --only-binary :all: --require-hashes \
+          -r ci/build_requirements.txt
+        "$env_python" -m pip install --only-binary :all: \
+          numpy scipy scikit-learn pytest
+        "$env_python" -m pip install --no-build-isolation --no-deps \
+          dist/*.tar.gz
+        cd "$RUNNER_TEMP"
+        "$env_python" -I -m pytest -p no:cacheprovider --pyargs hmmcraft.tests
+    - uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9  # v7.0.2
+      with:
+        name: dist-sdist
+        path: dist/*.tar.gz
+        if-no-files-found: error

hmm_craft-0.4.0/.gitignore

@@ -0,0 +1,20 @@
+*.egg-info/
+.*.swp
+*.o
+*.pyc
+*.pyd
+*.so
+.cache/
+.eggs/
+.ipynb_checkpoints/
+.pytest_cache/
+build/
+wheelhouse/
+dist/
+htmlcov/
+oprofile_data/
+.coverage
+.coverage.*
+.gdb_history
+# generated by doctest:
+filename.pkl

hmm_craft-0.4.0/ARCHITECTURE.md

@@ -0,0 +1,170 @@
+# How hmm-craft is built
+
+This file describes the package from the Python API down to the C++ kernel,
+what the package promises about its fits, and how a change to a fit is made.
+It is written from the code in this repository; the code wins where the two
+disagree.
+
+## Layout
+
+| Path | Holds |
+|---|---|
+| `src/hmmcraft/base.py` | `ConvergenceMonitor`, `_AbstractHMM` (the public methods and the fit loop), `BaseHMM` (expectation-maximization), `VariationalBaseHMM` (variational inference) |
+| `src/hmmcraft/_emissions.py` | one base class per emission family: how to compute the emission probabilities, collect the statistics an M-step needs, and draw a sample |
+| `src/hmmcraft/hmm.py` | the five EM models: `CategoricalHMM`, `GaussianHMM`, `GMMHMM`, `MultinomialHMM`, `PoissonHMM` |
+| `src/hmmcraft/vhmm.py` | the two variational models: `VariationalCategoricalHMM`, `VariationalGaussianHMM` |
+| `src/hmmcraft/utils.py` | public helpers: `normalize`, `log_normalize`, `fill_covars` |
+| `src/hmmcraft/stats.py` | `log_multivariate_normal_density` for the four covariance types |
+| `src/hmmcraft/_utils.py`, `_kl_divergence.py` | private helpers; the KL divergences of the variational lower bound |
+| `ext/_hmmc.cpp` | the C++ kernel, built by `setup.py` with pybind11 as `hmmcraft._hmmc` |
+| `src/hmmcraft/tests/` | the suite, installed with the package |
+| `scripts/` | the bit-equality check and the CI helpers; not installed |
+
+## Classes
+
+```mermaid
+classDiagram
+    BaseEstimator <|-- _AbstractHMM
+    _AbstractHMM <|-- BaseHMM
+    _AbstractHMM <|-- VariationalBaseHMM
+    _AbstractHMM <|-- BaseCategoricalHMM
+    _AbstractHMM <|-- BaseGaussianHMM
+    BaseHMM <|-- BaseGMMHMM
+    BaseHMM <|-- BaseMultinomialHMM
+    BaseHMM <|-- BasePoissonHMM
+    BaseCategoricalHMM <|-- CategoricalHMM
+    BaseHMM <|-- CategoricalHMM
+    BaseGaussianHMM <|-- GaussianHMM
+    BaseHMM <|-- GaussianHMM
+    BaseGMMHMM <|-- GMMHMM
+    BaseMultinomialHMM <|-- MultinomialHMM
+    BasePoissonHMM <|-- PoissonHMM
+    BaseCategoricalHMM <|-- VariationalCategoricalHMM
+    VariationalBaseHMM <|-- VariationalCategoricalHMM
+    BaseGaussianHMM <|-- VariationalGaussianHMM
+    VariationalBaseHMM <|-- VariationalGaussianHMM
+```
+
+`BaseEstimator` is scikit-learn's: it gives `get_params`, `set_params` and
+cloning. A public model combines two halves. The emission half (a class of
+`_emissions.py`) knows the distribution of one observation in one state. The
+learning half (`BaseHMM` or `VariationalBaseHMM`) knows how the start
+probabilities and the transition matrix are estimated. Categorical and
+Gaussian emissions exist in both halves, so their emission classes derive
+from `_AbstractHMM` and are combined with a learning class; the other three
+emission classes derive from `BaseHMM` directly.
+
+## The fit loop
+
+`_AbstractHMM.fit(X, lengths)` is the only fit loop. `X` holds every
+sequence, one after another; `lengths` gives the length of each, a positive
+integer. `fit` first passes both to `_utils.split_X_lengths`, which raises
+`ValueError` for any other entry and for a total that is not the number of
+samples, so no start is computed from invalid `lengths`.
+
+```mermaid
+flowchart TD
+    A["fit(X, lengths)"] --> B["_init: starting parameters<br/>(named in init_params, or not yet set)"]
+    B --> C["_check: shapes and sums"]
+    C --> D["_do_estep"]
+    D --> E["_compute_lower_bound"]
+    E --> F["_do_mstep: new parameters<br/>(those named in params)"]
+    F --> G{"ConvergenceMonitor:<br/>n_iter reached, or gain below tol?"}
+    G -- no --> D
+    G -- yes --> H["return self"]
+```
+
+`_do_estep` splits `X` by `lengths` and, for each sequence:
+
+1. computes the emission probability of every sample in every state
+   (`_compute_likelihood`, or `_compute_log_likelihood` when
+   `implementation="log"`, the default);
+2. calls the kernel's forward and backward passes on that matrix with
+   `startprob_` and `transmat_`, and turns the two lattices into the
+   posterior probability of each state at each sample;
+3. adds what the M-step needs to a dictionary of statistics
+   (`_accumulate_sufficient_statistics`); for the transitions, the kernel
+   computes the expected transition counts (`compute_log_xi_sum` or
+   `compute_scaling_xi_sum`).
+
+It also sums the posteriors into `state_occupancy_`, the expected number of
+samples in each state. `BaseHMM` returns the summed log-likelihood as the
+lower bound; `VariationalBaseHMM` subtracts the KL divergences of its
+posteriors from their priors, and runs the passes on subnormalized
+probabilities (`_estep_begin`).
+
+A model class supplies these hooks; the loop never changes per class:
+
+| Hook | Does |
+|---|---|
+| `_init` | sets the starting parameters named in `init_params`, and any that is not yet set (k-means for Gaussian means, for instance) |
+| `_check` | validates shapes and that probabilities sum to one |
+| `_compute_likelihood` / `_compute_log_likelihood` | at least one implemented; the base class derives a missing one from the other |
+| `_initialize_sufficient_statistics` | the empty dictionary of statistics |
+| `_accumulate_sufficient_statistics` | adds one sequence's statistics |
+| `_do_mstep` | sets the parameters named in `params` from the statistics |
+| `_generate_sample_from_state` | draws one observation, for `sample` |
+| `_get_n_fit_scalars_per_param` | the free parameters, for the warning on too little data and for `aic` and `bic` |
+
+`score` runs the forward pass alone; `score_samples`, `predict_proba` and
+`decode(algorithm="map")` run the forward and backward passes; none of them
+fits. `decode` with
+`algorithm="viterbi"` calls the kernel's `viterbi`. `sample` is a Python loop
+over the transition matrix and `_generate_sample_from_state`.
+
+## The C++ kernel
+
+`ext/_hmmc.cpp` holds the recursions that loop over every sample and every
+pair of states. It knows nothing about emissions: it receives the emission
+probabilities (or log-probabilities) of one sequence as a matrix of shape
+`(n_samples, n_components)`, already computed in Python. Every argument is a
+float64 array.
+
+| Function | Takes | Returns |
+|---|---|---|
+| `forward_log` | start probabilities, transition matrix, emission log-probabilities | the log-likelihood and the forward lattice, in log space |
+| `backward_log` | the same | the backward lattice, in log space |
+| `forward_scaling` | start probabilities, transition matrix, emission probabilities | the log-likelihood, the forward lattice and the scaling factor of each sample |
+| `backward_scaling` | the same, and the scaling factors | the backward lattice, scaled with those factors |
+| `compute_log_xi_sum` | forward lattice, transition matrix, backward lattice, emission log-probabilities | the expected transition counts, in log space |
+| `compute_scaling_xi_sum` | forward lattice, transition matrix, backward lattice, emission probabilities | the expected transition counts |
+| `viterbi` | start probabilities, transition matrix, emission log-probabilities | the log-probability of the best path and the path |
+
+A shape that does not match raises `ValueError`, and so does a matrix with no
+rows (a sequence without samples) or no columns (a model without states): the
+recursions use unchecked accessors, and hmmlearn 0.3.3's kernel indexed
+outside its arrays there. Each function releases the GIL while it loops. The
+kernel's source is hmmlearn 0.3.3's apart from that check and the module's
+package name.
+
+## What the package promises
+
+**Fits are checked to equal hmmlearn 0.3.3's, bit for bit, apart from the
+differences listed in `DIFFERENCES.md`.** `scripts/bit_equality.py` fits every
+model class, covariance type and implementation with both packages, in two
+environments with the same versions, on synthetic data generated inside the
+script from a fixed seed, with one thread, and compares every fitted array,
+the log-likelihood, both decodes and a drawn sample as bytes, not within a
+tolerance. `DIFFERENCES.md` names the data, the versions, the platform and the
+result; the claim covers that check and nothing beyond it. Both CI cells
+(`ci.yml`) run it on every pull request and every week, each with the
+reference installed at the same versions as the package.
+
+### Step 1 and step 2
+
+The fork changes fits in two steps.
+
+**Step 1** is the fork as first published: hmmlearn 0.3.3 under a new name,
+with newer minimum versions, wheels, CI, and the changes listed in
+`DIFFERENCES.md`. Only one of them changes a fit with every option at its
+default, and it replaces a result that hmmlearn 0.3.3 computed wrongly or
+failed to compute (the Gaussian fit of covariances without means).
+`GaussianHMM(covars_floor=...)` changes fits only when it is set. In the check
+above, every other fit equals 0.3.3's.
+
+**Step 2** holds the later changes that move a fit hmmlearn 0.3.3 computes
+correctly. Each one comes behind an option whose default keeps 0.3.3's result,
+is listed in `DIFFERENCES.md` with an exception of its own in the check, is
+named in `CHANGES.rst`, and ships in a minor release (`RELEASING.md`).
+Equality is always checked against hmmlearn 0.3.3, with every option at its
+default.

hmm_craft-0.4.0/AUTHORS.rst

@@ -0,0 +1,32 @@
+Authors and contributors, in no particular order:
+
+* Ron Weiss
+* Shiqiao Du
+* Jaques Grobler
+* David Cournapeau
+* Fabian Pedregosa
+* Gael Varoquaux
+* Andreas Mueller
+* Bertrand Thirion
+* Daniel Nouri
+* Gilles Louppe
+* Jake Vanderplas
+* John Benediktsson
+* Lars Buitinck
+* Mikhail Korobov
+* Robert McGibbon
+* Stefano Lattarini
+* Vlad Niculae
+* csytracy
+* Alexandre Gramfort
+* Sergei Lebedev
+* Daniela Huppenkothen
+* Christopher Farrow
+* Alexandr Yanenko
+* Antony Lee
+* Matthew Danielson
+* Alex Rockhill
+
+Authors and contributors of the hmm-craft fork, from 2026:
+
+* rudolfmleziva (maintainer)