Forecaster

PytorchForecastingNHiTS

pytorch-forecasting NHiTS model.

Quickstart

python

from sktime.forecasting.pytorchforecasting import PytorchForecastingNHiTS

estimator = PytorchForecastingNHiTS(model_params: dict [str, Any ] | None=None, dataset_params: dict [str, Any ] | None=None, train_to_dataloader_params: dict [str, Any ] | None=None, validation_to_dataloader_params: dict [str, Any ] | None=None, trainer_params: dict [str, Any ] | None=None, model_path: str | None=None, random_log_path: bool=False, broadcasting: bool=False)

Parameters(6)

model_paramsdict[str, Any] (default=None): parameters to be passed to initialize the pytorch-forecasting NBeats model [1] for example: {“interpolation_mode”: “nearest”, “activation”: “Tanh”}
dataset_paramsdict[str, Any] (default=None): parameters to initialize TimeSeriesDataSet [2] from pandas.DataFrame max_prediction_length will be overwrite according to fh time_idx, target, group_ids, time_varying_known_reals, time_varying_unknown_reals will be inferred from data, so you do not have to pass them
train_to_dataloader_paramsdict[str, Any] (default=None): parameters to be passed for TimeSeriesDataSet.to_dataloader() by default {“train”: True}
validation_to_dataloader_paramsdict[str, Any] (default=None): parameters to be passed for TimeSeriesDataSet.to_dataloader() by default {“train”: False}
model_path: string (default=None): try to load a existing model without fitting. Calling the fit function is still needed, but no real fitting will be performed.
random_log_path: bool (default=False): use random root directory for logging. This parameter is for CI test in Github action, not designed for end users.

Examples

>>> # import packages
>>> from sktime.forecasting.base import ForecastingHorizon
>>> from sktime.forecasting.pytorchforecasting import PytorchForecastingNHiTS
>>> from sktime.utils._testing.hierarchical import _make_hierarchical
>>> from sklearn.model_selection import train_test_split
>>> # generate random data
>>> data = _make_hierarchical (
... hierarchy_levels = (5, 200), max_timepoints = 50, min_timepoints = 50, n_columns = 3
... )
>>> # define forecast horizon
>>> max_prediction_length = 5
>>> fh = ForecastingHorizon (range (1, max_prediction_length + 1), is_relative = True)
>>> # split X, y data for train and test
>>> x = data ["c0", "c1" ]
>>> y = data ["c2" ]. to_frame ()
>>> X_train, X_test, y_train, y_test = train_test_split (
... x, y, test_size = 0.2, train_size = 0.8, shuffle = False
... )
>>> len_levels = len (y_test. index. names)
>>> y_test = y_test. groupby (level = list (range (len_levels - 1))). apply (
... lambda x: x. droplevel (list (range (len_levels - 1))). iloc [: - max_prediction_length ]
... )
>>> # define the model
>>> model = PytorchForecastingNHiTS (
... trainer_params = {
... "max_epochs": 5, # for quick test
... "limit_train_batches": 10, # for quick test
... },
... )
>>> # fit and predict
>>> model. fit (y = y_train, X = X_train, fh = fh) # doctest skip PytorchForecastingNHiTS(trainer_params={'limit_train_batches': 10, 'max_epochs': 5})
>>> y_pred = model. predict (fh, X = X_test, y = y_test)
>>> print (y_test) c2 h0 h1 time h0_0 h1_180 2000-01-01 8.184178 2000-01-02 5.444128 2000-01-03 5.992600 2000-01-04 5.223143 2000-01-05 6.191883
... 
... h0_4 h1_199 2000-02-10 7.498591 2000-02-11 5.910466 2000-02-12 7.409602 2000-02-13 4.670040 2000-02-14 5.454403 [4500 rows x 1 columns]
>>> print(y_pred) c2 h0 h1 time h0_0 h1_180 2000-02-15 5.764410 2000-02-16 5.826406 2000-02-17 5.925301 2000-02-18 5.792100 2000-02-19 5.760923 … … h0_4 h1_199 2000-02-15 5.376267 2000-02-16 5.227071 2000-02-17 5.070744 2000-02-18 5.249713 2000-02-19 5.047630 [500 rows x 1 columns]

References

[1 ] https://pytorch-forecasting.readthedocs.io/en/stable/api/pytorch_forecasting.models.nbeats._nbeats.NBeats.html # noqa: E501 [2 ] https://pytorch-forecasting.readthedocs.io/en/stable/api/pytorch_forecasting.data.timeseries._timeseries.TimeSeriesDataSet.html # noqa: E501