Backporting fixes:

• Fix: Hard threshold positive distribution parameters (#1950 by @lostella)
• Fix forecast keys (quantiles) output by TemporalFusionTransformer (#1952 by @lostella)

Backporting fixes:

• Fix: use broadcast_lesser in place of comparisons in ISQF (#1920 by @vincentqb)
• Fix dummy estimator (#1931 by @canerturkmen)
• Fix Pytorch Lightning tutorial (#1933 by @jaheba)
• Fixed autograd inplace operations error in Transformed Distribution (#1938 by @shubhamkapoor)

Backporting fixes:

• Fix AddTimeFeatures transformation for multiples of base frequencies (https://github.com/awslabs/gluon-ts/pull/1916)
• Update docs requirements (https://github.com/awslabs/gluon-ts/pull/1919)

Backporting fixes:

• Added QuarterlyBegin time feature (#1903)
• Fix the use of the scaling parameter in Transformer model (#1909)

Changelog

New Features

• Add ckpt_path argument to PyTorchLightningEstimator. (#1872)
• Add TSBench (#1865)
• add SCott code to nursery (#1827)
• Add dynamic code for shell. (#1821)
• Adding torch.isqf (#1815)
• Add tsbench readme placeholder (#1808)
• Adding ISQF distribution class (#1746)
• Adding IQF to remove quantile crossing and required retraining for ne… (#1693)
• Hierarchical Forecaster: End-to-End model based on DeepVAR (#1665)
• Adding glouonts.torch.piecewise_linear (#1663)
• Add quantitle regression mode to AutoGluon-based TabularEstimator (#1611)
• add dummy estimator to trivial models (#1602)

Bug Fixes

• Add file path argument to m5 dataset generation (#1896)
• Fix negative binomial parameter map (#1893)
• Fix negative binomial sampling (#1884)
• Fixes for Monash Forecasting Repository datasets (#1879)
• Fix serde.flat type handling. (#1851)
• Fix datesplitter. (#1850)
• changed metadata creation function (#1847)
• Check equality of transformations. (#1844)
• Fix samples scaling in PyTorch DeepAR (#1836)
• Fix _version for cases when git is not installed. (#1825)
• Fixed data leakage bug in implementation of dynamic real and categorical features (#1809)
• fix for #1725, reverse breaking changes to data loader and handle all zero batches (#1779)
• Upgrade pytorch and pytorch-lightning requirements and some fixes. (#1765)
• Fix torch NOPScaler shape. (#1752)
• Convert batchify list to np array (#1732)
• Fix gluonts.json; added bdump/bdumps. (#1721)
• Fix scaling for pytorch negative binomial output (#1702)
• Fix frequency string conversion from ts format, add test (#1652)
• Fix NegativeBinomial constructor args in NegativeBinomialOutput (torch) (#1651)
• Add batch_size attribute to MQCNNEstimator and MQRNNEstimator (#1645)
• Add additional datasets from the Monash Time Series Forecasting Repository (#1632)

Breaking Changes

• Extend default quantiles for MQ* Estimators to match MSIS quantiles. (#1866)
• changed metadata creation function (#1847)
• Remove support module. (#1792)
• Set minimum Python version to 3.7. (#1791)
• Exceptions cleanup. (#1615)

Other Changes & Improvements

• Update mypy to 0.910. (#1875)
• Bump ujson from 4.3.0 to 5.1.0 in /src/gluonts/nursery/tsbench (#1869)
• Update black to v22. (#1867)
• Fix docstring typo in feature.py (#1863)
• Fix scott checks. (#1845)
• Remove requirement for @validated in from_hyperparameters. (#1826)
• Fix test collect ignore. (#1817)
• Split tests into one workflow for each framework. (#1805)
• Mark transformer as flaky. (#1801)
• Mark empirical_distribution test as flaky. (#1798)
• Use of int/float/object over np.int/float/object for dtype. (#1795)
• Rework tests. (#1786)
• Update typing_extension version. (#1785)
• Use of independent random seed. (#1767)
• Upgrade pytorch and pytorch-lightning requirements and some fixes. (#1765)
• Remove sphinx-autobuild sphinx-autorun, update sphinx version. (#1745)
• Exlude bin folders from apidoc. (#1744)
• Don't run doctest on nursery. (#1743)
• Hierarchical: Compute relative reconciliation error and add tests (#1722)
• Fixing doc build from mqcnn-iqf commit (#1699)
• Replace miniver with custom versioning code. (#1662)
• Cap numba<0.54, ipykernel<6.2.0 (#1661)
• Removed assert for cardinality and static feats (#1659)

Backporting fixes:

• loosen RTOL in test/distribution/test_flows.py to make test_flow_invertibility pass (#1604)
• Add batch_size attribute to MQCNNEstimator and MQRNNEstimator (#1645)
• Fix NegativeBinomial constructor args in NegativeBinomialOutput (torch) (#1651)
• Fix frequency string conversion from ts format, add test (adapted from #1652)

Backporting fixes:

• Fix frequency metadata bug for lstnet datasets (#1593)
• Add batch_size attribute to MQCNNEstimator and MQRNNEstimator (#1645)
• Fix NegativeBinomial constructor args in NegativeBinomialOutput (torch) (#1651)

New Features

• add dummy estimator for seasonal_naive (#1598)
• Add STL-AR as one more R baseline model (#1568)
• Allow validation data for TabularEstimator. (#1562)
• QRX fixes and added functionality (#1544)
• Extend FileDataset's Parameters to load_datasets (#1538)
• Serde: Allow encoding of functions and methods. (#1519)
• Settings: Enable partial assignment (#1504)
• Settings: Support for nested args in _inject. (#1503)
• Transform.apply (#1494)
• PyTorch implementation of DeepAR (#1460)
• support Min freq for get_seasonality() method (#1459)
• add deep renewal processes for intermittent demand forecasting (#1458)
• Add transform objects for dealing with sparse time series. (#1421)
• spliced binned pareto (#1410)
• Add callbacks mechanism to Trainer class (#1168)

Bug Fixes

• Fix frequency metadata bug for lstnet datasets (#1593)
• Fix single dispatch register for py36 (#1591)
• R fixes for methods that produce point forecasts or prediction intervals directly (#1564)
• Fix computation of OWA (#1557)
• Fixed QRX bug: ".values()" to ".values" (#1552)
• QRX fixes and added functionality (#1544)
• Fix serde issue with some distribution output types, add test (#1543)
• Add item_id to r forecast predictors (#1537)
• fix ProphetPredictor serialization issue (#1535)
• Add constant dummy time features to TFT for yearly data (#1518)
• Settings: Fix partial assignment. (#1516)
• Fix anomaly detection example (#1515)
• Fix Settings._inject to check if it can provide the value. (#1501)
• Change miniver fallback version from unknown to 0.0.0. (#1457)
• Fix get_lags_for_frequency for minute data in DeepVAR (#1455)
• Fix missing import in gluonts.mx.model.GluonEstimator (#1450)
• Fix train-test split data leakage for m4_yearly and wiki-rolling_nips. (#1445)
• fix compatibility for pandas < 1.1 in time_feature/_base.py (#1437)
• fix edge case in iteration based model averaging (#1345)

Breaking Changes

• QRX fixes and added functionality (#1544)
• Transform.apply (#1494)

Other Changes & Improvements

• shallow import for gluonts.mx module (#1592)
• Mark torch distribution inference tests as flaky (#1586)
• Update REFERENCES.md (#1583)
• Delete pytorch_predictor_example.ipynb (#1574)
• Improve tests for R methods (#1567)
• Rename flake8 action step. (#1555)
• Set max_idle_transforms to the length of the dataset (#1546)
• Add datasets from forecastingdata.org (#1542)
• Train invoke with (#1530)
• Consolidate ZeroFeature from DeepState (#1522)
• Fix indentation (#1500)
• Simplify loader.py (#1495)
• adjustments to variable length functionality in batchify (#1442)
• Use miniver for version resolution. (#1434)
• Add docstrings for metrics. (#1422)
• Fixes for MXNet 1.8 (#1403)

Backporting fixes:

• Fix serde issue with some distribution output types, add test (#1543)

Backporting fixes:

• Train invoke with (#1530)
• fix ProphetPredictor serialization issue (#1535)
• Add item_id to r forecast predictors (#1537)
• Serde: Allow encoding of functions and methods. (#1519)
• Disable tests on Windows for PRs, fix other workflows (#1525)

Backporting fixes:

• Fix Settings._inject to check if it can provide the value. (#1501)
• Fix indentation (#1500)
• Fix anomaly detection example (#1515)
• Add constant dummy time features to TFT for yearly data (#1518)

Backporting fixes:

• Fix get_lags_for_frequency for minute data in DeepVAR (#1455)

Backporting fixes:

• Fixes for MXNet 1.8 (#1403)
• Fix train-test split data leakage for m4_yearly and wiki-rolling_nips. (#1445)
• Lock the version for mxnet theme to 0.3.15 (#1451)
• Fix missing import in gluonts.mx.model.GluonEstimator (#1450)

Backporting fixes:

• Fix train-test split data leakage for m4_yearly and wiki-rolling_nips. (#1445)
• Lock the version for mxnet theme to 0.3.15 (#1451)

Backporting fixes:

• fix s3fs ImportError for fsspec by updating the requirement depending on the python version (#1391)
• fix compatibility for pandas < 1.1 in time_feature/_base.py (#1437)

Backporting fixes:

• fix compatibility for pandas < 1.1 in time_feature/_base.py (#1437)

GluonTS adds improved support for PyTorch-based models, new options for existing models, and general improvements to components and tooling.

Breaking changes

This release comes with a few breaking changes (but for good reasons). In particular, models trained and serialized prior to 0.7.0 may not be de-serializable using 0.7.0.

• Changes in model components and abstractions:
  • #1256 and #1206 contain significant changes to the GluonEstimator abstract class, as well as InstanceSplitter and InstanceSampler implementations. You are affected by this change only if you implemented custom models based on GluonEstimator. The change makes it easier to define (and understand, in case you're reading the code) how fixed-length instances are to be sampled from the original dataset for training or validation purposes. Furthermore, this PR breaks data transformation into more explicit "pre-processing" steps (deterministic ones, e.g. feature engineering) vs "iteration" steps (possibly random, e.g. random training instance sampling), so that a cache_data option is now available in the train method to have the pre-processed data cached to memory, and be iterated quicker, whenever it fits.
  • #1233 splits normalized/unnormalized time features from gluonts.time_features into distinct types.
  • #1223 updates the interface of ISSM types, making it easier to define custom ones e.g. by having a custom set of seasonality patterns. Related changes to DeepStateEstimator enable these customizations when defining a DeepState model.
• Changes in Trainer:
  • #1178 removes the input_names argument from the __call__ method. Now the provided data loaders are expected to produce batches containing only the fields that the network being trained consumes. This can be easily obtained by transforming the dataset with SelectFields.
• Package structure reorg:
  • #1183 puts all MXNet-dependant modules under gluonts.mx, with some exceptions (gluonts.model and gluonts.nursery). With the new structure, one is not forced to install MXNet unless they specifically require modules that depend on it.
  • #1402 makes the Evaluator class lighter, by moving the evaluation metrics to gluonts.evaluation.metrics instead of having them as static methods of the class.

New features

PyTorch support:

• PyTorchPredictor serde (#1086)
• Add equality operator for PytorchPredictor (#1190)
• Allow Pytorch predictor to be trained and loaded on different devices (#1244)
• Add distribution-based forecast types for torch, output layers, tests (#1266)
• Add more distribution output classes for PyTorch, add tests (#1272)
• Add pytorch tutorial notebook (#1289)

Distributions:

• Zero Inflated Poisson Distribution (#1130)
• GenPareto cdf and quantile functions (#1142)
• Added quantile function based on cdf bisection (#1145)
• Add AffineTransformedDistribution (#1161)

Models:

• add estimator/predictor types for autogluon tabular (#1105)
• Added thetaf method to the R predictor (#1281)
• Adding neural ode code for lotka volterra and corresponding notebook (#1023)
• Added lightgbm support for QRX/Rotbaum (#1365)
• Deepar imputation model (#1380)
• Initial commit for GMM-TPP (#1397)

Datasets & tooling:

• Implemented generate_rolling_datasets (#844)
• Add a MinMax scaler (#1134)
• introduce functional api for data generation recipes (#1153)
• include m3 dataset (#1169)
• Improvements for data generation (#1195)
• Add most forecasters as entry points. (#1351)

Backporting fixes:

• Added lead_time argument to PyTorchPredictor (#1316)
• Fix serialization of lead_time (#1328)
• Fix serialization of lead_time in torch predictor (#1329)

Backporting fixes:

• Use broadcast_logical_or in inflated_beta (#1226)
• Fix type error for using quantile_weights and add a proper Pytest (#1231)
• Fixed MASE in N-BEATS: removed redundant factor (#1288)
• Fixing bug where dropout was not used, also remove unused halt option (#1315)
• Fixes for Python 3.8 (#1318)

Backporting fixes:

• Fix serde for np.dtype. (#1299)