Pytorch Forecasting Versions Save

Breaking Changes

• Upgraded to pytorch 2.0 and lightning 2.0. This brings a couple of changes, such as configuration of trainers. See the lightning upgrade guide. For PyTorch Forecasting, this particularly means if you are developing own models, the class method epoch_end has been renamed to on_epoch_end and replacing model.summarize() with ModelSummary(model, max_depth=-1) and Tuner(trainer) is its own class, so trainer.tuner needs replacing. (#1280)
• Changed the predict() interface returning named tuple - see tutorials.

Changes

• The predict method is now using the lightning predict functionality and allows writing results to disk (#1280).

Fixed

• Fixed robust scaler when quantiles are 0.0, and 1.0, i.e. minimum and maximum (#1142)

Fixed

• Removed pandoc from dependencies as issue with poetry install (#1126)
• Added metric attributes for torchmetric resulting in better multi-GPU performance (#1126)

Added

• "robust" encoder method can be customized by setting "center", "lower" and "upper" quantiles (#1126)

Added

• DeepVar network (#923)
• Enable quantile loss for N-HiTS (#926)
• MQF2 loss (multivariate quantile loss) (#949)
• Non-causal attention for TFT (#949)
• Tweedie loss (#949)
• ImplicitQuantileNetworkDistributionLoss (#995)

Fixed

• Fix learning scale schedule (#912)
• Fix TFT list/tuple issue at interpretation (#924)
• Allowed encoder length down to zero for EncoderNormalizer if transformation is not needed (#949)
• Fix Aggregation and CompositeMetric resets (#949)

Changed

• Dropping Python 3.6 suppport, adding 3.10 support (#479)
• Refactored dataloader sampling - moved samplers to pytorch_forecasting.data.samplers module (#479)
• Changed transformation format for Encoders to dict from tuple (#949)

Contributors

• jdb78

Fixed

• Fix with creating tensors on correct devices (#908)
• Fix with MultiLoss when calculating gradient (#908)

Contributors

• jdb78

Added

• Added new N-HiTS network that has consistently beaten N-BEATS (#890)
• Allow using torchmetrics as loss metrics (#776)
• Enable fitting EncoderNormalizer() with limited data history using max_length argument (#782)
• More flexible MultiEmbedding() with convenience output_size and input_size properties (#829)
• Fix concatentation of attention (#902)

Fixed

• Fix pip install via github (#798)

Contributors

• jdb78
• christy
• lukemerrick
• Seon82

Added

• Added support for running pytorch_lightning.trainer.test (#759)

Fixed

• Fix inattention mutation to x_cont (#732).
• Compatability with pytorch-lightning 1.5 (#758)

Contributors

• eavae
• danielgafni
• jdb78

Added

• Use target name instead of target number for logging metrics (#588)
• Optimizer can be initialized by passing string, class or function (#602)
• Add support for multiple outputs in Baseline model (#603)
• Added Optuna pruner as optional parameter in TemporalFusionTransformer.optimize_hyperparameters (#619)
• Dropping support for Python 3.6 and starting support for Python 3.9 (#639)

Fixed

• Initialization of TemporalFusionTransformer with multiple targets but loss for only one target (#550)
• Added missing transformation of prediction for MLP (#602)
• Fixed logging hyperparameters (#688)
• Ensure MultiNormalizer fit state is detected (#681)
• Fix infinite loop in TimeDistributedEmbeddingBag (#672)

Contributors

• jdb78
• TKlerx
• chefPony
• eavae
• L0Z1K

Breaking changes

• Removed dropout_categoricals parameter from TimeSeriesDataSet. Use categorical_encoders=dict(<variable_name>=NaNLabelEncoder(add_nan=True)) instead (#518)

• Rename parameter allow_missings for TimeSeriesDataSet to allow_missing_timesteps (#518)

• Transparent handling of transformations. Forward methods should now call two new methods (#518):

  • transform_output to explicitly rescale the network outputs into the de-normalized space
  • to_network_output to create a dict-like named tuple. This allows tracing the modules with PyTorch's JIT. Only prediction is still required which is the main network output.

  Example:

  def forward(self, x):
      normalized_prediction = self.module(x)
      prediction = self.transform_output(prediction=normalized_prediction, target_scale=x["target_scale"])
      return self.to_network_output(prediction=prediction)
  

Added

• Improved validation of input parameters of TimeSeriesDataSet (#518)

Fixed

• Fix quantile prediction for tensors on GPUs for distribution losses (#491)
• Fix hyperparameter update for RecurrentNetwork.from_dataset method (#497)

Added

• Allow lists for multiple losses and normalizers (#405)
• Warn if normalization is with scale < 1e-7 (#429)
• Allow usage of distribution losses in all settings (#434)

Fixed

• Fix issue when predicting and data is on different devices (#402)
• Fix non-iterable output (#404)
• Fix problem with moving data to CPU for multiple targets (#434)

Contributors

• jdb78
• domplexity

Added

• Adding a filter functionality to the timeseries datasset (#329)
• Add simple models such as LSTM, GRU and a MLP on the decoder (#380)
• Allow usage of any torch optimizer such as SGD (#380)

Fixed

• Moving predictions to CPU to avoid running out of memory (#329)
• Correct determination of output_size for multi-target forecasting with the TemporalFusionTransformer (#328)
• Tqdm autonotebook fix to work outside of Jupyter (#338)
• Fix issue with yaml serialization for TensorboardLogger (#379)

Contributors

• jdb78
• JakeForsey
• vakker