tinkoff-ai/etna 2.2.0

etna 2.2.0

on GitHub

Highlights

• Add extension with models from statsforecast
• Speed up metrics computation
• Speed up DeepARModel and TFTModel
• Add DeseasonalityTransform
• Add PatchTSModel
• Add new category mode into HolidayTransform
• Add documentation warning about using dill during loading
• Add inverse transformation into predict method of pipelines
• Fix CLI to work with pipeline ensembles

Full changelog

Added

• DeseasonalityTransform (#1307)
• Add extension with models from statsforecast: StatsForecastARIMAModel, StatsForecastAutoARIMAModel, StatsForecastAutoCESModel, StatsForecastAutoETSModel, StatsForecastAutoThetaModel (#1295)
• Notebook feature_selection (#875)
• Implementation of PatchTS model (#1277)

Changed

• Add modes binary and category to HolidayTransform (#763)
• Add sorting by timestamp before the fit in CatBoostPerSegmentModel and CatBoostMultiSegmentModel (#1337)
• Speed up metrics computation by optimizing segment validation, forbid NaNs during metrics computation (#1338)
• Unify errors, warnings and checks in models (#1312)
• Remove upper limitation on version of numba (#1321)
• Optimize TSDataset.describe and TSDataset.info by vectorization (#1344)
• Add documentation warning about using dill during loading (#1346)
• Vectorize metric computation (#1347)

Fixed

• Pipeline ensembles fail in etna forecast CLI (#1331)
• Fix performance of DeepARModel and TFTModel (#1322)
• mrmr feature selection working with categoricals (#1311)
• Fix version of statsforecast to 1.4 to avoid dependency conflicts during installation (#1313)
• Add inverse transformation into predict method of pipelines (#1314)
• Allow saving large pipelines (#1335)
• Fix link for dataset in classification notebook (#1351)

Removed

• Building docker images with cuda 10.2 (#1306)