v0.12.0

Major changes

• Rework preprocessing module (see #177). (#179)
  • Add Custom transformation.
  • Rename Preprocess to Compose.
  • Don't validate observation sequences after each transformation in Compose.
  • Remove progress bars and verbose parameter.
  • Stop unnecessarily copying each observation sequence before transformations.
  • Change transform() function on Transform objects to accept a single observation sequence.
  • Remove _apply() function on Transform objects.
  • Make _is_fitted() public on Transform objects (change to is_fitted()).
  • Use __str__ instead of _describe() for transformation descriptions.
• Remove need to send DeepGRU to device explicitly, so we can now do DeepGRU(..., device=device) instead of DeepGRU(..., device=device).to(device). (#178)
• Add dev, test, docs and notebooks extras. (#174)
• Remove Equalize transform as it goes against the point of variable-length sequence classification. (#172)
• Change TrimZeros transform to TrimConstants, allowing any constant-valued observation to be trimmed. (#172)
• Add DeepGRU classifier implementation. (#169)
• Add sequentia[torch] extra for optional torch CPU installation. (#169)

Minor changes

• Keep batch lengths on CPU (pytorch/pytorch#43227). (#178)
• Remove docs/requirements.txt and specify docs extra in .readthedocs.yml. (#176)
• Move Sphinx extensions from docs/conf.py to requirements.py. (#176)
• Bump development status classifier to beta. (#175)
• Move package dependency specifications to requirements.py. (#174)
• Add docs/README.md, notebooks/README.md and lib/test/README.md. (#174)
• Update HMM classifier diagram. (#173)
• Add build status to README.md. (#171)
• Fix patch description in CONTRIBUTING.md. (#170)
• Fix wording in README.md. (#167, #168)

v0.11.1

Major changes

• Fix validation for univariate sequences. (#164)

Minor changes

• Clean up README.md and add examples. (#165)
• Clean up validation logical expressions. (#164)

v0.11.0

Major changes

• Add trailing underscore to variables containing trainable parameters (see #154). (#158)
• Add properties for GMM emission distribution parameters (see #153). (#156)
• Add selective GMMHMM parameter freezing/unfreezing (see #150). (#155)
• Fix random transition matrix initialization for _LeftRightTopology (see #149). (#151)

Minor changes

• Add access to Baum-Welch algorithm convergence monitor (see #139). (#162)
• Prefix _Validator functions with is_ (see #159). (#161)
• Add validation for checking fitted parameters (see #157). (#160)
• Clean up __repr__ for GMMHMM, HMMClassifier and KNNClassifier. (#160)
• Add classifier documentation links to README.md. (#152)
• Simplify random transition matrix initialization for _LinearTopology and _LeftRightTopology. (#151)

v0.10.3

Major changes

• Fix setup.py encoding problem. (#145)
• Add docs/robots.txt and sphinx-version-warning package to prevent search engines from indexing old package versions (see #143). (#147)

Minor changes

• Add @Prhmma as a contributor for #145. (#146)

v0.10.2

Major changes

• Add support for dependent feature warping (addresses #124). (#135)
• Add multi-processed predictions for HMMClassifier (addresses #121). (#136)
• Re-order predict() and evaluate() arguments. (#138)

Minor changes

• Add original_labels documentation to KNNClassifier. (#133)
• Simplify GMMHMM documentation. (#134)
• Fix posterior comment in classifier.svg. (#137)

v0.10.1

Minor changes

• Remove references to sigment. (#130)
• Fix type specifiers in documentation (see #129). (#131)

v0.10.0

Major changes

• Switch out pomegranate HMM backend to hmmlearn. (#105)
• Remove separate HMM and GMM-HMM implementations – only keep a single GMM-HMM implementation (in the GMMHMM class) and treat multivariate Gaussian emission HMM as a special case of GMM-HMM. (#105)
• Support string and numeric labels by using label encodings (from sklearn.preprocessing.LabelEncoder). (#105)
• Add support for Python v3.6, v3.7, v3.8, v3.9 and remove support for v3.5. (#105)
• Switch from approximate DTW algorithm (fastdtw) to exact implementation (dtaidistance) for KNNClassifier. (#106)

Minor changes

• Switch to use duck-typing for iterables instead of requiring lists. (#105)
• Rename 'strict left-right' HMM topology to 'linear'. (#105)
• Switch m2r to m2r2, as m2r is no longer maintained. (#105)
• Change covariance to covariance_type, to match hmmlearn. (#105)
• Use numpy.random.RandomState(seed=None) as default instead of numpy.random.RandomState(seed=0). (#105)
• Switch KNNClassifier serialization from HDF5 to pickling. (#106)
• Use intersphinx for external documentation links, e.g. to numpy. (#108)
• Change MinMaxScale bounds to floats. (#112)
• Add __repr__ function to GMMHMM, HMMClassifier and KNNClassifier. (#120)
• Use feature-independent warping (DTWI). (#121)
• Ensure minimum Sakoe-Chiba band width is 1. (#126)
• Add .readthedocs.yml configuration file. (#127)

v0.7.2

Major changes

• Stop referring to sequences as temporal, as non-temporal sequences can also be used. (#103)

v0.7.1

Major changes

• Fix deserialization for KNNClassifier. (#93)
  • Sort HDF5 keys before loading as numpy.ndarrays.
  • Pass weighting function into deserialization constructor.

v0.7.0

Major changes

• Fix pomegranate version to v0.12.0. (#79)
• Add serialization and deserialization support for all classifiers. (#80)
  • HMM, HMMClassifier: Serialized in JSON format.
  • KNNClassifier: Serialized in HDF5 format.
• Finish preprocessing documentation and tests. (#81)
• (Internal) Remove nested helper functions in KNNClassifier.predict(). (#84)
• Add strict left-right HMM topology. (#85)
  Note: This is the more traditional left-right HMM topology.
• Implement GMM-HMMs in the GMMHMM class. (#87)
• Implement custom, uniform and frequency-based HMM priors. (#88)
• Implement distance-weighted DTW-kNN predictions. (#90)
• Rename DTWKNN to KNNClassifer. (#91)

Minor changes

• (Internal) Simplify package imports. (#82)
• (Internal) Add Validator.func() for validating callables. (#90)