gluonts

Backporting fixes:

• Fix type annotation for device in PyTorchPredictor #3094 by @lostella
• extend exception for invalid offset_name #3115 by @Borda

Backporting:

• Fix Rotbaum serialization and deserialization #3068 by @pantanurag555
• Fix Rotbaum to handle short series #3073 by @leica2023

Backporting:

• Fix Rotbaum serialization and deserialization #3068 by @pantanurag555
• Fix Rotbaum to handle short series #3073 by @leica2023

Backporting fixes:

• Fix iterable.Cached. #3060 by @jaheba
• Torch: Remove double caching of dataset. #3061 by @jaheba

Backporting fixes:

• Fix edge cases in metric computation #3037 (@lostella)
• Refactor tests for ev.aggregations #3038 (@lostella)
• Add plt.show() to README #3043 (@lostella)
• Rotbaum: Add item-id to forecast. #3049 (@jaheba)
• Fix mypy checks #3052 (@lostella)

Backporting fixes:

• Fix edge cases in metric computation #3037 (@lostella)
• Refactor tests for ev.aggregations #3038 (@lostella)
• Rotbaum: Add item-id to forecast. #3049 (@jaheba)
• Fix mypy checks #3052 (@lostella)

Release notes coming soon

Fixing issues with updated requirements, which would prevent the package from installing alongside old versions of PyTorch Lightning:

• Fix lightning import (#3018 by @shchur)
• Update requirements-pytorch.txt (#3019 by @shchur)

Backporting fixes:

• [torch] Return a model even if callback has no best model path #2952 by @kashif
• Move from pytorch_lightning to lightning #3013 by @shchur

Backporting fixes:

• Raise warning in QuantileForecast.mean when mean is not there #2843 by @lostella
• [R] Bug fix for running R methods via ParallelizedPredictor #2983 by @rshyamsundar
• Fix Forecast plot method #3006 by @lostella

Backporting fixes:

• Turn type comparison into isinstance #2958 by @lostella
• Fix Cython version in XGBoost tests #2966 by @lostella
• Clean up RepresentablePredictor #2967 by @lostella
• Fix: use isinstance instead of type comparison #2973 by @melopeo
• Fix ArrowDecoder.decode to return instead of yield #2976 by @lostella
• Unpin pyarrow version #2977 by @lostella

Backporting fixes:

• Zebras: Fix index handling of SplitFrame.resize. #2938 by @jaheba
• Docs: fix missing values use-case in tutorial for PandasDataset #2941 by @cneely33
• Ignore F403 errors in preludes. #2948 by @jaheba
• Fix: prevent accumulation of SelectFields in PyTorchPredictor #2951 by @Cameronwood611
• [Docs] fix link to NPTS implementation #2953 by @lostella

Backporting fixes:

• Fix NaNs in seasonal error #2894 by @gorold
• Filter unused fields during inference #2905 by @abdulfatir
• Define repr instead of str for PandasDataset #2906 by @lostella
• Fix typo in background.md #2907 by @tnixon
• Fix another typo in background.md #2908 by @tnixon

Backporting fixes:

• Speedup is_uniform for PandasDataset. #2878 by @jaheba
• Fix docstrings in torch lightning modules #2880 by @lostella
• Fix default scale in torch DeepAR #2885 by @lostella

Overview

We're happy to release gluonts version 0.13! This release contains a few new features and breaking changes compared to 0.12, especially around PyTorch models, data handling and model evaluation:

• Added Support for PyTorch 2.0 , Lightning 2.0, Pandas 2.0.
• New, more ergonomic model evaluation routines were added in #2673 (see the PR description for details on how to use those).
• New PyTorch-based models, including PatchTST (#2748), and other torch-models-related breaking changes and improvements like #2603, #2614, #2628, #2688, and #2618.
• Faster PandasDataset (#2663, #2665, #2860)

There are several more improvements and fixes compared to 0.12, which you can find in the changelog below. This release was possible thanks to the great work of several contributors: @jaheba, @MarcelK1102, @lostella, @gorold, @kashif, @dcmaddix, @abdulfatir, @melopeo, @huibinshen, @shchur, @pablovicente, @Gandor26, @Linbo-Liu. Thanks everyone, and thanks to users and authors of issue reports for the precious feedback!

Changelog

Breaking changes

• Add transform.Valmap, improve transform.Chain. (#2629)
• make Pytorch scaler's forward API consistent (#2627)
• Remove torch specific Dataloader, remove num_workers from torch models. (#2628)
• Pass prediction inputs as dict. (#2646)
• Simplify scalers, move to gluonts.torch.scaler (#2632)
• Remove TimeSeriesSlice (#2680)
• Add model.Input. (#2684)
• Remove FallbackPredictor. (#2686)
• Move model init to lightning module. (#2688)
• torch.SimpleFeedForward: Rename context to past_target. (#2704)
• Fix style and type issues (#2711)
• Fix off-by-one in torch DeepAR (#2618)
• Remove dataset.Schema. (#2798)
• Simplify univariate R wrapper (#2830)
• Simplify plotting of forecasts. (#2864)

Major improvements or new features

• Expose weight_decay in torch TFT estimator class (#2603)
• Allow ReduceLROnPlateau to track val_loss when validation set is available (#2614)
• Add wrapper for Nixtla/hierarchicalforecast (#2591)
• Add zebras freq/period. (#2651)
• Faster index building in PandasDataset (#2663)
• Speed up PandasDataset.from_long_dataframe (#2665)
• Add zebras.TimeFrame. (#2672)
• Allow PyTorch 2.0 (#2724)
• Fix Pandas 2.0 compatibility issues (#2710)
• Allow PyTorch Lightning 2.0 (#2728)
• Add itertools.PickleCache. (#2756)
• Add Monash repository datasets (#2771)
• Merge zebras from proof of concept branch. (#2776)
• Add interface to gluonts.ev (#2673)
• Zebras: Add from_pandas classmethods to TimeFrame and Periods. (#2807)
• Zebras: Improve TimeFrame.split. (#2808)
• Zebras: Add TimeFrame.rename. (#2810)
• Zebras: Add TimeFrame.rolsplit. (#2809)
• Add fourier.arima for long seasonal time series (#2789)
• Add patch-TST, D-Linear and a new lag-TST model (#2748)

Minor improvements or new features

• Rework torch MeanScaler. (#2600)
• Add dataset.loader.as_stacked_batches. (#2638)
• Add Cyclic.stream. (#2639)
• Add Wiki 2000 dataset (#2642)
• Make hierarchicalforecast a single module. (#2666)
• Add itertools.pluck_attr. (#2668)
• Add itertools.power_set (#2682)
• Allow axis to be a tuple in ev.aggregations (#2681)
• Export torch models in torch module directly. (#2685)
• Add zebras resize. (#2705)
• improve comprehension list style (#2715)
• Improve check for validation loop in lightning modules (#2726)
• Add warning for "object" features in PandasDataset (#2731)
• Remove usage of glide in tsf-reader. (#2737)
• Add cdf and icdf to torch NegativeBinomial distribution (#2749)
• Reduce depth of get_dataset import (#2796)
• Reduce depth of gluonts.dataset.repository imports in code and docs (#2803)
• Fix: remove multiprocessing from TSFReader (#2806)
• Add maybe stub file. (#2811)
• Add SizedIterableSlice, an IterableSlice that supports len() (#2815)
• add validated to DeepNPTS (#2823)
• Refactor base metrics computation in Evaluator class (#2825)
• Remove second call to create_lightning_module on torch estimator (#2834)
• Ingore hidden files in FileDataset by default. (#2847)
• Add --compression argument to arrow writer cli. (#2848)
• Zebras: handling of weekday offsets. (#2849)
• Zebras: Add unix_epoch method to Period and Periods. (#2851)
• Zebras: Improve equality for Periods. (#2857)
• Add join_items to itertools. (#2859)
• Zebras: Add eq_to to TimeFrame. (#2858)
• Zebras: Add eq_shape to TimeFrame. (#2863)
• Zebras: Allow to slice TimeFrame using plain strings for time info. (#2865)
• Zebras: Add TimeSeries.to_numpy. (#2866)
• Cache groupby result in PandasDataset (#2860)
• Add feat_static_cat for TSF datasets (#2871)

Bug fixes

• Ensure dtype on feat_time in torch DeepAR. (#2596)
• Add assertion to split function ensuring valid windows (#2587)
• Fix bug with static cardinalities in PandasDataset (#2599)
• Add gluonts.util.safe_extract (#2606)
• Expose aggregation method in ensemble NBEATS, fix forecast shape (#2598)
• Fix incorrect import in tsbench, apply latest black (#2613)
• Fix: torch PoissonOutput scaling (#2619)
• Remove dataclasses requirement (#2623)
• Implement equals for init_passed_kwargs. (#2630)
• Fix bugs in MeanScaler (#2633)
• Fix validation_data usage in torch. (#2643)
• Fix norm-freq to consider freq starts. (#2645)
• Fix call to extractall (#2648)
• Delay instantiation of ScipyStudentT object (#2660)
• Fix DateSplitter when split date is before start (#2670)
• Remove creation of ragged sequences in MultivariateGrouper (#2671)
• Fix ev.seasonal_error (#2696)
• Fix zebras period time features. (#2700)
• Update hierarchicalforecast for new release (#2709)
• Fix DistributionForecast failure on GPU (#2714)
• Fix version location for sdist. (#2729)
• Fix validation loop check for Lightning modules (#2733)
• Fix version cmdclass handling. (#2735)
• Fix: use non-strict inequality in definition of coverage (#2738)
• Fix MXNet NOPScaler (#2744)
• Fix dataset file discovery. (#2777)
• Fix: Loading of nested paths in FileDataset. (#2779)
• Prophet: Pass 'item_id' and 'info' to forecast. (#2780)
• Avoid zero scale in StudentTOutput (#2791)
• Zebras: time length fixes. (#2799)
• Remove .to_timestamp() to fix interval plotting (#2800)
• Fix pd.Period serialization (#2827)
• Fix torch DeepAREstimator in case context_length=1 (#2841)

Documentation

• Fix installation docs, fix typos in docstrings (#2625)
• Fix r-forecast doc strings (#2669)
• Add doctests. (#2683)
• Fix MultivariateEvaluator docstrings (#2693)
• Docs: minor spelling fix (#2701)
• Docs: Add extra requirements. (#2732)
• Update Available Models (#2740)
• Update REFERENCES.md (#2824)
• Update plotting in readme example. (#2867)
• Docs: Fix and simplify tutorials. (#2869)
• Docs: Fix black formatting of % instructions. (#2870)
• Docs: Add download link to notebooks. (#2872)
• Docs: Use torch DeepAR in README. (#2874)

Test / setup changes

• Fix version in requirements to comply with stricter setuptools. (#2604)
• Test: Increase timeout for xgboost tests. (#2601)
• Ignore warnings in tests. (#2620)
• Test: Remove -v option from pytest. (#2631)
• Add hierarchicalforecast to github workflows. (#2659)
• Make nursery tests opt-in. (#2667)
• Add test for torch models tracing (#2658)
• Relax pandas requirement to include pandas 2.x. (#2713)
• Update CP-Flow fork as extra dependency for MQF2 (#2727)
• Add scipy requirement (#2745)
• Fix pandas removed deprecations in tests (#2778)
• Test: Set caplog level for shell tests. (#2786)
• Bump numpy from 1.19.2 to 1.22.0 in /src/gluonts/nursery/daf (#2787)
• Update setuptools and wheel in test workflow (#2802)
• Bump torch from 1.6.0 to 1.13.1 in /src/gluonts/nursery/daf (#2788)
• Add test workflow for R based models (#2814)
• Add tests for hierarchical model to R workflow (#2819)
• Move notebook compilation logic to docs workflow (#2831)
• Fix bug in docs workflow (#2836)
• Fix docs workflow further (#2837)
• Fix string literal in docs workflow (#2839)
• Update action to configure AWS credentials (#2873)

Others

• Move NPTS back to gluonts.model (#2597)
• Remove mxnet from default dataset path (#2635)
• Roll back MQF2 import (#2687)
• add DAF source code (#2769)
• Add code for multivariate attack paper (#2697)
• Update wiki2k tarball path (#2805)
• [Nursery] CoP-DeepAR: Model for temporal hierarchical forecasting (#2812)
• Cop deepar: Reset the no. of epochs to the default value (#2813)
• Guard scripts execution in nursery (#2832)

Backporting fixes:

• Remove .to_timestamp() to fix interval plotting #2800 by @abdulfatir
• Fix pd.Period serialization #2827 by @abdulfatir
• Remove second call to create_lightning_module on torch estimator #2834 by @pablovicente
• Fix torch DeepAREstimator in case context_length=1 #2841 by @lostella
• Ignore hidden files in FileDataset by default. #2847 by @jaheba

Backporting fixes:

• Test: Set caplog level for shell tests. #2786 by @jaheba
• Avoid zero scale in StudentTOutput #2791 by @shchur
• Update setuptools and wheel in test workflow #2802 by @lostella
• Fix: remove multiprocessing from TSFReader #2806 by @lostella