tinkoff-ai/etna 2.1.0

etna 2.1.0

on GitHub

Highlights

• Add class etna.auto.Tune for tuning hyperparameters
• Extend functionality of class etna.auto.Auto to include a tuning stage
• Add notebook about AutoML
• Add utilities for estimating number of folds for backtesting and forecasting and integrate them into CLI
• Add parameter for setting the start of prediction into CLI
• Add etna.transforms.ExogShiftTransform to shift all exogenous variables
• Add etna.models.DeepStateModel
• Update requirements for holidays, scipy, ruptures, sqlalchemy, tsfresh
• Optimize make_samples of etna.models.RNNNet and etna.models.MLPNet
• Add parameter fast_redundancy in etna.analysis.feature_selection.mrmm and etna.transforms.MRMRFeatureSelectionTransform to speed it up

Full changelog

Added

• Notebook forecast_interpretation.ipynb with forecast decomposition (#1220)
• Exogenous variables shift transform ExogShiftTransform(#1254)
• Parameter start_timestamp to forecast CLI command (#1265)
• DeepStateModel (#1253)
• Function estimate_max_n_folds for folds number estimation (#1279)
• Parameters estimate_n_folds and context_size to forecast and backtest CLI commands (#1284)
• Class Tune for hyperparameter optimization within existing pipeline (#1200)
• Add etna.distributions for using it instead of using optuna.distributions (#1292)

Changed

• Set the default value of final_model to LinearRegression(positive=True) in the constructor of StackingEnsemble (#1238)
• Add microseconds to FileLogger's directory name (#1264)
• Inherit SaveMixin from AbstractSaveable for mypy checker (#1261)
• Update requirements for holidays and scipy, change saving library from pickle to dill in SaveMixin (#1268)
• Update requirement for ruptures, add requirement for sqlalchemy (#1276)
• Optimize make_samples of RNNNet and MLPNet (#1281)
• Remove to_be_fixed from inference tests on SpecialDaysTransform (#1283)
• Rewrite TimeSeriesImputerTransform to work without per-segment wrapper (#1293)
• Add default params_to_tune for catboost models (#1185)
• Add default params_to_tune for ProphetModel (#1203)
• Add default params_to_tune for SARIMAXModel, change default parameters for the model (#1206)
• Add default params_to_tune for linear models (#1204)
• Add default params_to_tune for SeasonalMovingAverageModel, MovingAverageModel, NaiveModel and DeadlineMovingAverageModel (#1208)
• Add default params_to_tune for DeepARModel and TFTModel (#1210)
• Add default params_to_tune for HoltWintersModel, HoltModel and SimpleExpSmoothingModel (#1209)
• Add default params_to_tune for RNNModel and MLPModel (#1218)
• Add default params_to_tune for DateFlagsTransform, TimeFlagsTransform, SpecialDaysTransform and FourierTransform (#1228)
• Add default params_to_tune for MedianOutliersTransform, DensityOutliersTransform and PredictionIntervalOutliersTransform (#1231)
• Add default params_to_tune for TimeSeriesImputerTransform (#1232)
• Add default params_to_tune for DifferencingTransform, MedianTransform, MaxTransform, MinTransform, QuantileTransform, StdTransform, MeanTransform, MADTransform, MinMaxDifferenceTransform, SumTransform, BoxCoxTransform, YeoJohnsonTransform, MaxAbsScalerTransform, MinMaxScalerTransform, RobustScalerTransform and StandardScalerTransform (#1233)
• Add default params_to_tune for LabelEncoderTransform (#1242)
• Add default params_to_tune for ChangePointsSegmentationTransform, ChangePointsTrendTransform, ChangePointsLevelTransform, TrendTransform, LinearTrendTransform, TheilSenTrendTransform and STLTransform (#1243)
• Add default params_to_tune for TreeFeatureSelectionTransform, MRMRFeatureSelectionTransform and GaleShapleyFeatureSelectionTransform (#1250)
• Add tuning stage into Auto.fit (#1272)
• Add params_to_tune into Tune init (#1282)
• Skip duplicates during Tune.fit, skip duplicates in top_k, add AutoML notebook (#1285)
• Add parameter fast_redundancy in mrmm, fix relevance calculation in get_model_relevance_table (#1294)

Fixed

• Fix plot_backtest and plot_backtest_interactive on one-step forecast (1260)
• Fix BaseReconciliator to work on pandas==1.1.5 (#1229)
• Fix TSDataset.make_future to handle hierarchy, quantiles, target components (#1248)
• Fix warning during creation of ResampleWithDistributionTransform (#1230)
• Add deep copy for copying attributes of TSDataset (#1241)
• Add tsfresh into optional dependencies, remove instruction about pip install tsfresh (#1246)
• Fix DeepARModel and TFTModel to work with changed prediction_size (#1251)
• Fix problems with flake8 B023 (#1252)
• Fix problem with swapped forecast methods in HierarchicalPipeline (#1259)
• Fix problem with segment name "target" in StackingEnsemble (#1262)
• Fix BasePipeline.forecast when prediction intervals are estimated on history data with presence of NaNs (#1291)
• Teach BaseMixin.set_params to work with nested list and tuple (#1201)
• Fix get_anomalies_prediction_interval to work when segments have different start date (#1296)
• Fix classification notebook to download FordA dataset without error (#1299)
• Fix signature of Auto.fit, Tune.fit to not have a breaking change (#1300)