Feature/tft darts (#513)

* added tff.py for TFFModel

* savepoint switching to master

* TFT with MixedCovariates

* savepoint software update

* checked tft_submodels

* original model performs well

* all submodels checked and ready

* new tft submodels work as original TFT

* github test

* adapted forward for fitting with darts

* predict returns output but issue with target scaling

* fixed bug in _produce_train_output(), now predictions look much better

* train and prediction work

* TFT now supports variable prediction lengths

* added multivariate TFT forecast support

* added model docstrings

* general clean up of unnecessary variables and code

* removed files used for building the model

* added TFTModel to README.md

* added probabilistic forecasting support for TFT with likelihood models

* minor docstring fixes

* fixed error from testing with MixedCovariatesInferenceDataset for regression models

* removed unused changes in darts/utils/data/inference_dataset.py

* Revert "removed unused changes in darts/utils/data/inference_dataset.py"

This reverts commit a8521cc.

* remove unused changes in darts/utils/data/inference_dataset.py

* added TFT unit tests

* cleaned up unused features in TFT unit test

* fixed *args to predict() method for model.historical_forecastings()

* solve an issue with Cuda

* made compatible with pandas==1.3.0 and statsmodels==0.13.0

* added package pytorch-forecasting>=0.9.1

* added submodels from pytorch-forecasting

* removed pytorch-forecasting dependency and applied part of PR review suggestions

* Feat/quantile loss as likelihood (#526)

* applied changes from PR review

Co-authored-by: Julien Herzen <[email protected]>
Co-authored-by: Julien Herzen <[email protected]>

Author: dennisbader

README.md

@@ -138,6 +138,7 @@ Model | Univariate | Multivariate | Probabilistic | Multiple-series training | P
 `NBEATSModel` | ✅ | ✅ | | ✅ | ✅ | | [N-BEATS paper](https://arxiv.org/abs/1905.10437)
 `TCNModel` | ✅ | ✅ | ✅ | ✅ | ✅ | | [TCN paper](https://arxiv.org/abs/1803.01271), [DeepTCN paper](https://arxiv.org/abs/1906.04397), [blog post](https://medium.com/unit8-machine-learning-publication/temporal-convolutional-networks-and-forecasting-5ce1b6e97ce4)
 `TransformerModel` | ✅ | ✅ | ✅ | ✅ | ✅ | | 
+`TFTModel` | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | [TFT paper](https://arxiv.org/pdf/1912.09363.pdf), [PyTorch Forecasting](https://pytorch-forecasting.readthedocs.io/en/latest/models.html)
 Naive Baselines | ✅ | | | | | |
 
 

darts/models/__init__.py

@@ -32,6 +32,7 @@
     from darts.models.forecasting.tcn_model import TCNModel
     from darts.models.forecasting.nbeats import NBEATSModel
     from darts.models.forecasting.transformer_model import TransformerModel
+    from darts.models.forecasting.tft_model import TFTModel
 
 except ModuleNotFoundError:
     logger.warning("Support Torch based models not available. To enable it, install u8darts[torch] or u8darts[all].")

darts/models/forecasting/fft.py

@@ -171,7 +171,7 @@ def _crop_to_match_seasons(series: TimeSeries, required_matches: Optional[set])
         return series
 
     first_ts = series.time_index[0]
-    freq = first_ts.freq
+    freq = series.freq
     pred_ts = series.time_index[-1] + freq
 
     # start at first timestamp of given series and move forward until a matching timestamp is found