Feat/optimized hfc torch

[dennisbader]

Fixes #1543, fixes #2025.

Summary

• adds optimized historical forecasting for torch models
• available with all configurations (overlap, last_points_only, autoregression, int/dt index, single/multi series, ...)
• fixes a bug when using future covariates with historical forecasting, the predictions didn't extend to the expended end time
• refactored _get_historical_forecastable_time_index() to account for forecast_horizon and overlap_end directly

Other Information

• last potential performance boost would be to avoid generating the time series in PLForecastingModule.predict_step(). this can be left for the another PR IMO

Benchmarking

Tested different configuration for using NLinearModel using the optimized and on non-optimized historical forecasting (hist fc)

Some parameters explained:

• multi_series:
  • False computes hist fcs on a single series with a total of roughly 10k iterations.
  • True computes hist fcs on 100 series with a total of roughly 10k iterations.

last_points_only	is_autoregressive	is_int_idx	multi_series	time_non_optimized	time_optimized	boost up [times]
True	False	False	False	105.118	8.034	13.08
True	False	False	True	96.251	7.667	12.55
True	False	True	False	96.185	5.592	17.2
True	False	True	True	86.841	4.933	17.6
True	True	False	False	104.748	7.727	13.56
True	True	False	True	92.924	7.496	12.4
True	True	True	False	91.359	5.302	17.23
True	True	True	True	82.67	4.791	17.26
False	False	False	False	99.843	7.599	13.14
False	False	False	True	92.049	7.093	12.98
False	False	True	False	89.718	5.43	16.52
False	False	True	True	83.32	4.562	18.26
False	True	False	False	99.991	7.927	12.61
False	True	False	True	92.429	7.126	12.97
False	True	True	False	91.653	5.29	17.33
False	True	True	True	81.988	4.874	16.82

[codecov-commenter]

Codecov Report

Attention: 9 lines in your changes are missing coverage. Please review.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Files	Coverage Δ
darts/models/forecasting/ensemble_model.py	94.73% <100.00%> (+0.09%)	⬆️
darts/models/forecasting/forecasting_model.py	94.96% <100.00%> (+0.02%)	⬆️
darts/models/forecasting/pl_forecasting_module.py	93.97% <100.00%> (+0.02%)	⬆️
darts/models/forecasting/regression_model.py	96.76% <100.00%> (+0.26%)	⬆️
darts/utils/historical_forecasts/__init__.py	100.00% <100.00%> (ø)
...casts/optimized_historical_forecasts_regression.py	97.56% <100.00%> (ø)
darts/utils/timeseries_generation.py	96.18% <100.00%> (+0.02%)	⬆️
...arts/models/forecasting/torch_forecasting_model.py	90.71% <86.66%> (-0.17%)	⬇️
..._forecasts/optimized_historical_forecasts_torch.py	94.59% <94.59%> (ø)
darts/utils/historical_forecasts/utils.py	92.71% <95.00%> (+0.32%)	⬆️
... and 1 more

... and 5 files with indirect coverage changes

📢 Thoughts on this report? Let us know!.

[madtoinou]

Really nice job @dennisbader, we should maybe include the "trick" used to optimize each historical forecasts routine in the docstring so that if other contributors want to improve it, they don't have to figure it out themself.