v0.9.0

v0.9.0

Added

• Allow parallel computation of DTW barycenters and plug it in TimeSeriesKMeans.
• NonMyopicEarlyClassifier can now be used with yet incomplete series or streamed inputs
  to retrieve optimal classification timing.
• Add mae, mse and mase performance metrics for datasets
• TimeSeriesFeatureSynchronizer is a new preprocessing to deal with acquisition of features
  at different sampling rates or desynchronized timestamps.
• Add root_dir parameter to UCR_UEA_datasets to allow users to specify
  where downloaded datasets should be cached. If None, a default directory is
  used, either in $XDG_DATA_HOME/tslearn/UCR_UEA if the environment variable
  is set or ~/.tslearn/datasets otherwise.
• Add VARIMA and AutoVARIMA forcasting estimator.

Changed

• Fix masking issue in LearningShapelets with Keras 3.14
• Fix imputer crash (#659)
• TimeSeriesScalerMinMax and TimeSeriesScalerMeanVariance now scales based on fitted data characteristics
  when per_timeseries is False. (#280)
• Fix LearningShapelets.shapelets_ attribute computation.
• Fix LearningShapelets with MPS due to unsupported float64 precision
• Rework dtw and frechet computation for torch backend to speed things up.