autogluon.timeseries 1.4.1b20250907__py3-none-any.whl → 1.5.1b20260122__py3-none-any.whl

autogluon/timeseries/models/ensemble/per_item_greedy.py

@@ -0,0 +1,172 @@
+import logging
+import pprint
+import time
+from typing import Any
+
+import pandas as pd
+from joblib import Parallel, delayed
+
+from autogluon.timeseries import TimeSeriesDataFrame
+from autogluon.timeseries.utils.constants import AG_DEFAULT_N_JOBS
+
+from .abstract import AbstractTimeSeriesEnsembleModel
+from .ensemble_selection import fit_time_series_ensemble_selection
+
+logger = logging.getLogger(__name__)
+
+
+class PerItemGreedyEnsemble(AbstractTimeSeriesEnsembleModel):
+    """Per-item greedy ensemble that fits separate weighted ensembles for each individual time series.
+
+    This ensemble applies the greedy Ensemble Selection algorithm by Caruana et al. [Car2004]_ independently
+    to each time series in the dataset, allowing for customized model combinations that adapt to the
+    specific characteristics of individual series. Each time series gets its own optimal ensemble weights
+    based on predictions for that particular series. If items not seen during training are provided at prediction
+    time, average model weight across the training items will be used for their predictions.
+
+    The per-item approach is particularly effective for datasets with heterogeneous time series that
+    exhibit different patterns, seasonalities, or noise characteristics.
+
+    The algorithm uses parallel processing to efficiently fit ensembles across all time series.
+
+    Other Parameters
+    ----------------
+    ensemble_size : int, default = 100
+        Number of models (with replacement) to include in the ensemble.
+    n_jobs : int or float, default = joblib.cpu_count(only_physical_cores=True)
+        Number of CPU cores used to fit the ensembles in parallel.
+
+    References
+    ----------
+    .. [Car2004] Caruana, Rich, et al. "Ensemble selection from libraries of models."
+        Proceedings of the twenty-first international conference on Machine learning. 2004.
+    """
+
+    def __init__(self, name: str | None = None, **kwargs):
+        if name is None:
+            name = "PerItemWeightedEnsemble"
+        super().__init__(name=name, **kwargs)
+        self.weights_df: pd.DataFrame
+        self.average_weight: pd.Series
+
+    @property
+    def model_names(self) -> list[str]:
+        return list(self.weights_df.columns)
+
+    def _get_default_hyperparameters(self) -> dict[str, Any]:
+        return {"ensemble_size": 100, "n_jobs": AG_DEFAULT_N_JOBS}
+
+    def _fit(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        data_per_window: list[TimeSeriesDataFrame],
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
+    ) -> None:
+        model_names = list(predictions_per_window.keys())
+        item_ids = data_per_window[0].item_ids
+        n_jobs = min(self.get_hyperparameter("n_jobs"), len(item_ids))
+
+        predictions_per_item = self._split_predictions_per_item(predictions_per_window)
+        data_per_item = self._split_data_per_item(data_per_window)
+
+        ensemble_selection_kwargs = dict(
+            ensemble_size=self.get_hyperparameter("ensemble_size"),
+            eval_metric=self.eval_metric,
+            prediction_length=self.prediction_length,
+            target=self.target,
+        )
+
+        time_limit_per_item = None if time_limit is None else time_limit * n_jobs / len(item_ids)
+        end_time = None if time_limit is None else time.time() + time_limit
+
+        # Fit ensemble for each item in parallel
+        executor = Parallel(n_jobs=n_jobs)
+        weights_per_item = executor(
+            delayed(self._fit_item_ensemble)(
+                data_per_item[item_id],
+                predictions_per_item[item_id],
+                time_limit_per_item=time_limit_per_item,
+                end_time=end_time,
+                **ensemble_selection_kwargs,
+            )
+            for item_id in item_ids
+        )
+        self.weights_df = pd.DataFrame(weights_per_item, index=item_ids, columns=model_names)  # type: ignore
+        self.average_weight = self.weights_df.mean(axis=0)
+
+        # Drop models with zero average weight
+        if (self.average_weight == 0).any():
+            models_to_keep = self.average_weight[self.average_weight > 0].index
+            self.weights_df = self.weights_df[models_to_keep]
+            self.average_weight = self.average_weight[models_to_keep]
+
+        weights_for_printing = {model: round(float(weight), 2) for model, weight in self.average_weight.items()}
+        logger.info(f"\tAverage ensemble weights: {pprint.pformat(weights_for_printing, width=1000)}")
+
+    def _split_predictions_per_item(
+        self, predictions_per_window: dict[str, list[TimeSeriesDataFrame]]
+    ) -> dict[str, dict[str, list[TimeSeriesDataFrame]]]:
+        """Build a dictionary mapping item_id -> dict[model_name, list[TimeSeriesDataFrame]]."""
+        item_ids = list(predictions_per_window.values())[0][0].item_ids
+
+        predictions_per_item = {}
+        for i, item_id in enumerate(item_ids):
+            item_predictions = {}
+            for model_name, preds_per_window in predictions_per_window.items():
+                item_preds_per_window = [
+                    pred.iloc[i * self.prediction_length : (i + 1) * self.prediction_length]
+                    for pred in preds_per_window
+                ]
+                item_predictions[model_name] = item_preds_per_window
+            predictions_per_item[item_id] = item_predictions
+        return predictions_per_item
+
+    def _split_data_per_item(self, data_per_window: list[TimeSeriesDataFrame]) -> dict[str, list[TimeSeriesDataFrame]]:
+        """Build a dictionary mapping item_id -> ground truth values across all windows."""
+        item_ids = data_per_window[0].item_ids
+        data_per_item = {item_id: [] for item_id in item_ids}
+
+        for data in data_per_window:
+            indptr = data.get_indptr()
+            for item_idx, item_id in enumerate(item_ids):
+                new_slice = data.iloc[indptr[item_idx] : indptr[item_idx + 1]]
+                data_per_item[item_id].append(new_slice)
+        return data_per_item
+
+    @staticmethod
+    def _fit_item_ensemble(
+        data_per_window: list[TimeSeriesDataFrame],
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        time_limit_per_item: float | None = None,
+        end_time: float | None = None,
+        **ensemble_selection_kwargs,
+    ) -> dict[str, float]:
+        """Fit ensemble for a single item."""
+        if end_time is not None:
+            assert time_limit_per_item is not None
+            time_left = end_time - time.time()
+            time_limit_per_item = min(time_limit_per_item, time_left)
+        return fit_time_series_ensemble_selection(
+            data_per_window, predictions_per_window, time_limit=time_limit_per_item, **ensemble_selection_kwargs
+        )
+
+    def _predict(self, data: dict[str, TimeSeriesDataFrame], **kwargs) -> TimeSeriesDataFrame:
+        assert all(model in data for model in self.weights_df.columns)
+        item_ids = list(data.values())[0].item_ids
+        unseen_item_ids = set(item_ids) - set(self.weights_df.index)
+        if unseen_item_ids:
+            logger.debug(f"Using average weights for {len(unseen_item_ids)} unseen items")
+        weights = self.weights_df.reindex(item_ids).fillna(self.average_weight)
+
+        result = None
+        for model_name in self.weights_df.columns:
+            model_pred = data[model_name]
+            model_weights = weights[model_name].to_numpy().repeat(self.prediction_length)
+            weighted_pred = model_pred.to_data_frame().multiply(model_weights, axis=0)
+            result = weighted_pred if result is None else result + weighted_pred
+
+        return TimeSeriesDataFrame(result)  # type: ignore
+
+    def remap_base_models(self, model_refit_map: dict[str, str]) -> None:
+        self.weights_df.rename(columns=model_refit_map, inplace=True)

autogluon/timeseries/models/ensemble/weighted/__init__.py

@@ -0,0 +1,8 @@
+from .basic import PerformanceWeightedEnsemble, SimpleAverageEnsemble
+from .greedy import GreedyEnsemble
+
+__all__ = [
+    "SimpleAverageEnsemble",
+    "PerformanceWeightedEnsemble",
+    "GreedyEnsemble",
+]

autogluon/timeseries/models/ensemble/weighted/abstract.py

@@ -0,0 +1,45 @@
+import functools
+from abc import ABC
+
+import numpy as np
+
+from autogluon.timeseries.dataset import TimeSeriesDataFrame
+
+from ..abstract import AbstractTimeSeriesEnsembleModel
+
+
+class AbstractWeightedTimeSeriesEnsembleModel(AbstractTimeSeriesEnsembleModel, ABC):
+    """Abstract base class for weighted ensemble models that assign global weights to base models.
+
+    Weighted ensembles combine predictions from multiple base models using learned or computed weights,
+    where each base model receives a single global weight applied across all time series and forecast
+    horizons. The final prediction is computed as a weighted linear combination of base model forecasts.
+    """
+
+    def __init__(self, name: str | None = None, **kwargs):
+        super().__init__(name=name, **kwargs)
+        self.model_to_weight: dict[str, float] = {}
+
+    @property
+    def model_names(self) -> list[str]:
+        return list(self.model_to_weight.keys())
+
+    @property
+    def model_weights(self) -> np.ndarray:
+        return np.array(list(self.model_to_weight.values()), dtype=np.float64)
+
+    def _predict(self, data: dict[str, TimeSeriesDataFrame], **kwargs) -> TimeSeriesDataFrame:
+        weighted_predictions = [data[model_name] * weight for model_name, weight in self.model_to_weight.items()]
+        return functools.reduce(lambda x, y: x + y, weighted_predictions)
+
+    def get_info(self) -> dict:
+        info = super().get_info()
+        info["model_weights"] = self.model_to_weight.copy()
+        return info
+
+    def remap_base_models(self, model_refit_map: dict[str, str]) -> None:
+        updated_weights = {}
+        for model, weight in self.model_to_weight.items():
+            model_full_name = model_refit_map.get(model, model)
+            updated_weights[model_full_name] = weight
+        self.model_to_weight = updated_weights

autogluon/timeseries/models/ensemble/{basic.py → weighted/basic.py}

@@ -1,4 +1,4 @@
-from typing import Any, Optional
+from typing import Any
 
 import numpy as np
 
@@ -8,19 +8,20 @@ from .abstract import AbstractWeightedTimeSeriesEnsembleModel
 
 
 class SimpleAverageEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
-    """Constructs a weighted ensemble using a simple average of the constituent models' predictions."""
+    """Simple ensemble that assigns equal weights to all base models for uniform averaging.
 
-    def __init__(self, name: Optional[str] = None, **kwargs):
-        if name is None:
-            name = "SimpleAverageEnsemble"
-        super().__init__(name=name, **kwargs)
+    This ensemble computes predictions as the arithmetic mean of all base model forecasts,
+    giving each model equal influence. Simple averaging is robust and often performs well when base
+    models have similar accuracy levels or when validation data is insufficient to reliably
+    estimate performance differences.
+    """
 
     def _fit(
         self,
         predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
         data_per_window: list[TimeSeriesDataFrame],
-        model_scores: Optional[dict[str, float]] = None,
-        time_limit: Optional[float] = None,
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
     ):
         self.model_to_weight = {}
         num_models = len(predictions_per_window)
@@ -29,16 +30,23 @@ class SimpleAverageEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
 
 
 class PerformanceWeightedEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
-    """Constructs a weighted ensemble, where the weights are assigned in proportion to the
-    (inverse) validation scores.
+    """Performance-based weighted ensemble that assigns weights proportional to validation scores.
+
+    This ensemble computes model weights based on their validation performance, giving higher
+    weights to better-performing models. The weighting scheme transforms validation scores
+    (higher is better) into ensemble weights using configurable transformation functions.
+
+    .. warning::
+        This ensemble method is deprecated and may be removed in a future version.
 
     Other Parameters
     ----------------
-    weight_scheme: Literal["sq", "inv", "loginv"], default = "loginv"
+    weight_scheme : Literal["sq", "inv", "sqrt"], default = "sqrt"
         Method used to compute the weights as a function of the validation scores.
-        - "sqrt" computes weights in proportion to `sqrt(1 / S)`. This is the default.
-        - "inv" computes weights in proportion to `(1 / S)`.
-        - "sq" computes the weights in proportion to `(1 / S)^2` as outlined in [PC2020]_.
+
+        - "sqrt" computes weights in proportion to ``sqrt(1 / S)``. This is the default.
+        - "inv" computes weights in proportion to ``(1 / S)``.
+        - "sq" computes the weights in proportion to ``(1 / S)^2`` as outlined in [PC2020]_.
 
     References
     ----------
@@ -47,11 +55,6 @@ class PerformanceWeightedEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
         36.1 (2020): 93-97.
     """
 
-    def __init__(self, name: Optional[str] = None, **kwargs):
-        if name is None:
-            name = "PerformanceWeightedEnsemble"
-        super().__init__(name=name, **kwargs)
-
     def _get_default_hyperparameters(self) -> dict[str, Any]:
         return {"weight_scheme": "sqrt"}
 
@@ -59,12 +62,12 @@ class PerformanceWeightedEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
         self,
         predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
         data_per_window: list[TimeSeriesDataFrame],
-        model_scores: Optional[dict[str, float]] = None,
-        time_limit: Optional[float] = None,
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
     ):
         assert model_scores is not None
 
-        weight_scheme = self.get_hyperparameters()["weight_scheme"]
+        weight_scheme = self.get_hyperparameter("weight_scheme")
 
         # drop NaNs
         model_scores = {k: v for k, v in model_scores.items() if np.isfinite(v)}

autogluon/timeseries/models/ensemble/weighted/greedy.py

@@ -0,0 +1,64 @@
+import logging
+import pprint
+from typing import Any
+
+from autogluon.timeseries import TimeSeriesDataFrame
+
+from ..ensemble_selection import fit_time_series_ensemble_selection
+from .abstract import AbstractWeightedTimeSeriesEnsembleModel
+
+logger = logging.getLogger(__name__)
+
+
+class GreedyEnsemble(AbstractWeightedTimeSeriesEnsembleModel):
+    """Greedy ensemble selection algorithm that iteratively builds an ensemble by selecting models with
+    replacement.
+
+    Also known as ``WeightedEnsemble`` for backward compatibility.
+
+    This class implements the Ensemble Selection algorithm by Caruana et al. [Car2004]_, which starts
+    with an empty ensemble and repeatedly adds the model that most improves the ensemble's validation
+    performance. Models can be selected multiple times, allowing the algorithm to assign higher effective
+    weights to better-performing models.
+
+    Other Parameters
+    ----------------
+    ensemble_size : int, default = 100
+        Number of models (with replacement) to include in the ensemble.
+
+    References
+    ----------
+    .. [Car2004] Caruana, Rich, et al. "Ensemble selection from libraries of models."
+        Proceedings of the twenty-first international conference on Machine learning. 2004.
+    """
+
+    def __init__(self, name: str | None = None, **kwargs):
+        if name is None:
+            # FIXME: the name here is kept for backward compatibility. it will be called
+            # GreedyEnsemble in v1.4 once ensemble choices are exposed
+            name = "WeightedEnsemble"
+        super().__init__(name=name, **kwargs)
+
+    def _get_default_hyperparameters(self) -> dict[str, Any]:
+        return {"ensemble_size": 100}
+
+    def _fit(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        data_per_window: list[TimeSeriesDataFrame],
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
+    ):
+        model_to_weight = fit_time_series_ensemble_selection(
+            data_per_window=data_per_window,
+            predictions_per_window=predictions_per_window,
+            ensemble_size=self.get_hyperparameter("ensemble_size"),
+            eval_metric=self.eval_metric,
+            prediction_length=self.prediction_length,
+            target=self.target,
+            time_limit=time_limit,
+        )
+        self.model_to_weight = {model: weight for model, weight in model_to_weight.items() if weight > 0}
+
+        weights_for_printing = {model: round(float(weight), 2) for model, weight in self.model_to_weight.items()}
+        logger.info(f"\tEnsemble weights: {pprint.pformat(weights_for_printing, width=200)}")

autogluon/timeseries/models/gluonts/abstract.py

@@ -3,7 +3,7 @@ import os
 import shutil
 from datetime import timedelta
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, Optional, Type, Union, cast, overload
+from typing import TYPE_CHECKING, Any, Callable, Type, cast, overload
 
 import gluonts
 import gluonts.core.settings
@@ -21,7 +21,7 @@ from autogluon.core.hpo.constants import RAY_BACKEND
 from autogluon.tabular.models.tabular_nn.utils.categorical_encoders import (
     OneHotMergeRaresHandleUnknownEncoder as OneHotEncoder,
 )
-from autogluon.timeseries.dataset.ts_dataframe import ITEMID, TimeSeriesDataFrame
+from autogluon.timeseries.dataset import TimeSeriesDataFrame
 from autogluon.timeseries.models.abstract import AbstractTimeSeriesModel
 from autogluon.timeseries.utils.warning_filters import disable_root_logger, warning_filter
 
@@ -72,12 +72,12 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
 
     def __init__(
         self,
-        freq: Optional[str] = None,
+        freq: str | None = None,
         prediction_length: int = 1,
-        path: Optional[str] = None,
-        name: Optional[str] = None,
-        eval_metric: Optional[str] = None,
-        hyperparameters: Optional[dict[str, Any]] = None,
+        path: str | None = None,
+        name: str | None = None,
+        eval_metric: str | None = None,
+        hyperparameters: dict[str, Any] | None = None,
         **kwargs,  # noqa
     ):
         super().__init__(
@@ -89,9 +89,9 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
             hyperparameters=hyperparameters,
             **kwargs,
         )
-        self.gts_predictor: Optional[GluonTSPredictor] = None
-        self._ohe_generator_known: Optional[OneHotEncoder] = None
-        self._ohe_generator_past: Optional[OneHotEncoder] = None
+        self.gts_predictor: GluonTSPredictor | None = None
+        self._ohe_generator_known: OneHotEncoder | None = None
+        self._ohe_generator_past: OneHotEncoder | None = None
         self.callbacks = []
         # Following attributes may be overridden during fit() based on train_data & model parameters
         self.num_feat_static_cat = 0
@@ -105,7 +105,7 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
         self.past_feat_dynamic_cat_cardinality: list[int] = []
         self.negative_data = True
 
-    def save(self, path: Optional[str] = None, verbose: bool = True) -> str:
+    def save(self, path: str | None = None, verbose: bool = True) -> str:
         # we flush callbacks instance variable if it has been set. it can keep weak references which breaks training
         self.callbacks = []
         # The GluonTS predictor is serialized using custom logic
@@ -153,18 +153,17 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
                 assert dataset.static_features is not None, (
                     "Static features must be provided if num_feat_static_cat > 0"
                 )
-                feat_static_cat = dataset.static_features[self.covariate_metadata.static_features_cat]
-                self.feat_static_cat_cardinality = feat_static_cat.nunique().tolist()
+                self.feat_static_cat_cardinality = list(self.covariate_metadata.static_cat_cardinality.values())
 
         disable_known_covariates = model_params.get("disable_known_covariates", False)
         if not disable_known_covariates and self.supports_known_covariates:
             self.num_feat_dynamic_cat = len(self.covariate_metadata.known_covariates_cat)
             self.num_feat_dynamic_real = len(self.covariate_metadata.known_covariates_real)
             if self.num_feat_dynamic_cat > 0:
-                feat_dynamic_cat = dataset[self.covariate_metadata.known_covariates_cat]
                 if self.supports_cat_covariates:
-                    self.feat_dynamic_cat_cardinality = feat_dynamic_cat.nunique().tolist()
+                    self.feat_dynamic_cat_cardinality = list(self.covariate_metadata.known_cat_cardinality.values())
                 else:
+                    feat_dynamic_cat = dataset[self.covariate_metadata.known_covariates_cat]
                     # If model doesn't support categorical covariates, convert them to real via one hot encoding
                     self._ohe_generator_known = OneHotEncoder(
                         max_levels=model_params.get("max_cat_cardinality", 100),
@@ -180,10 +179,12 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
             self.num_past_feat_dynamic_cat = len(self.covariate_metadata.past_covariates_cat)
             self.num_past_feat_dynamic_real = len(self.covariate_metadata.past_covariates_real)
             if self.num_past_feat_dynamic_cat > 0:
-                past_feat_dynamic_cat = dataset[self.covariate_metadata.past_covariates_cat]
                 if self.supports_cat_covariates:
-                    self.past_feat_dynamic_cat_cardinality = past_feat_dynamic_cat.nunique().tolist()
+                    self.past_feat_dynamic_cat_cardinality = list(
+                        self.covariate_metadata.past_cat_cardinality.values()
+                    )
                 else:
+                    past_feat_dynamic_cat = dataset[self.covariate_metadata.past_covariates_cat]
                     # If model doesn't support categorical covariates, convert them to real via one hot encoding
                     self._ohe_generator_past = OneHotEncoder(
                         max_levels=model_params.get("max_cat_cardinality", 100),
@@ -277,8 +278,8 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
 
         return torch.cuda.is_available()
 
-    def get_minimum_resources(self, is_gpu_available: bool = False) -> dict[str, Union[int, float]]:
-        minimum_resources: dict[str, Union[int, float]] = {"num_cpus": 1}
+    def get_minimum_resources(self, is_gpu_available: bool = False) -> dict[str, int | float]:
+        minimum_resources: dict[str, int | float] = {"num_cpus": 1}
         # if GPU is available, we train with 1 GPU per trial
         if is_gpu_available:
             minimum_resources["num_gpus"] = 1
@@ -289,8 +290,8 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
     @overload
     def _to_gluonts_dataset(self, time_series_df: TimeSeriesDataFrame, known_covariates=None) -> GluonTSDataset: ...
     def _to_gluonts_dataset(
-        self, time_series_df: Optional[TimeSeriesDataFrame], known_covariates: Optional[TimeSeriesDataFrame] = None
-    ) -> Optional[GluonTSDataset]:
+        self, time_series_df: TimeSeriesDataFrame | None, known_covariates: TimeSeriesDataFrame | None = None
+    ) -> GluonTSDataset | None:
         if time_series_df is not None:
             # TODO: Preprocess real-valued features with StdScaler?
             if self.num_feat_static_cat > 0:
@@ -388,10 +389,10 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
     def _fit(
         self,
         train_data: TimeSeriesDataFrame,
-        val_data: Optional[TimeSeriesDataFrame] = None,
-        time_limit: Optional[float] = None,
-        num_cpus: Optional[int] = None,
-        num_gpus: Optional[int] = None,
+        val_data: TimeSeriesDataFrame | None = None,
+        time_limit: float | None = None,
+        num_cpus: int | None = None,
+        num_gpus: int | None = None,
         verbosity: int = 2,
         **kwargs,
     ) -> None:
@@ -438,8 +439,8 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
 
     def _get_callbacks(
         self,
-        time_limit: Optional[float],
-        early_stopping_patience: Optional[int] = None,
+        time_limit: float | None,
+        early_stopping_patience: int | None = None,
     ) -> list[Callable]:
         """Retrieve a list of callback objects for the GluonTS trainer"""
         from lightning.pytorch.callbacks import EarlyStopping, Timer
@@ -454,7 +455,7 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
     def _predict(
         self,
         data: TimeSeriesDataFrame,
-        known_covariates: Optional[TimeSeriesDataFrame] = None,
+        known_covariates: TimeSeriesDataFrame | None = None,
         **kwargs,
     ) -> TimeSeriesDataFrame:
         if self.gts_predictor is None:
@@ -471,8 +472,8 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
     def _predict_gluonts_forecasts(
         self,
         data: TimeSeriesDataFrame,
-        known_covariates: Optional[TimeSeriesDataFrame] = None,
-        num_samples: Optional[int] = None,
+        known_covariates: TimeSeriesDataFrame | None = None,
+        num_samples: int | None = None,
     ) -> list[Forecast]:
         assert self.gts_predictor is not None, "GluonTS models must be fit before predicting."
         gts_data = self._to_gluonts_dataset(data, known_covariates=known_covariates)
@@ -566,7 +567,7 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
     ) -> TimeSeriesDataFrame:
         from gluonts.torch.model.forecast import DistributionForecast
 
-        item_ids = forecast_index.unique(level=ITEMID)
+        item_ids = forecast_index.unique(level=TimeSeriesDataFrame.ITEMID)
         if isinstance(forecasts[0], SampleForecast):
             forecast_df = self._stack_sample_forecasts(cast(list[SampleForecast], forecasts), item_ids)
         elif isinstance(forecasts[0], QuantileForecast):

autogluon/timeseries/models/gluonts/dataset.py

@@ -1,11 +1,11 @@
-from typing import Any, Iterator, Optional, Type
+from typing import Any, Iterator, Type
 
 import numpy as np
 import pandas as pd
 from gluonts.dataset.common import Dataset as GluonTSDataset
 from gluonts.dataset.field_names import FieldName
 
-from autogluon.timeseries.dataset.ts_dataframe import TIMESTAMP, TimeSeriesDataFrame
+from autogluon.timeseries.dataset import TimeSeriesDataFrame
 from autogluon.timeseries.utils.datetime import norm_freq_str
 
 
@@ -17,14 +17,14 @@ class SimpleGluonTSDataset(GluonTSDataset):
         target_df: TimeSeriesDataFrame,
         freq: str,
         target_column: str = "target",
-        feat_static_cat: Optional[np.ndarray] = None,
-        feat_static_real: Optional[np.ndarray] = None,
-        feat_dynamic_cat: Optional[np.ndarray] = None,
-        feat_dynamic_real: Optional[np.ndarray] = None,
-        past_feat_dynamic_cat: Optional[np.ndarray] = None,
-        past_feat_dynamic_real: Optional[np.ndarray] = None,
+        feat_static_cat: np.ndarray | None = None,
+        feat_static_real: np.ndarray | None = None,
+        feat_dynamic_cat: np.ndarray | None = None,
+        feat_dynamic_real: np.ndarray | None = None,
+        past_feat_dynamic_cat: np.ndarray | None = None,
+        past_feat_dynamic_real: np.ndarray | None = None,
         includes_future: bool = False,
-        prediction_length: Optional[int] = None,
+        prediction_length: int | None = None,
     ):
         assert target_df is not None
         # Convert TimeSeriesDataFrame to pd.Series for faster processing
@@ -44,11 +44,11 @@ class SimpleGluonTSDataset(GluonTSDataset):
         # Replace inefficient groupby ITEMID with indptr that stores start:end of each time series
         self.item_ids = target_df.item_ids
         self.indptr = target_df.get_indptr()
-        self.start_timestamps = target_df.index[self.indptr[:-1]].to_frame(index=False)[TIMESTAMP]
+        self.start_timestamps = target_df.index[self.indptr[:-1]].to_frame(index=False)[TimeSeriesDataFrame.TIMESTAMP]
         assert len(self.item_ids) == len(self.start_timestamps)
 
     @staticmethod
-    def _astype(array: Optional[np.ndarray], dtype: Type[np.generic]) -> Optional[np.ndarray]:
+    def _astype(array: np.ndarray | None, dtype: Type[np.generic]) -> np.ndarray | None:
         if array is None:
             return None
         else:

autogluon/timeseries/models/gluonts/models.py

@@ -41,10 +41,8 @@ class DeepARModel(AbstractGluonTSModel):
         Number of steps to unroll the RNN for before computing predictions
     disable_static_features : bool, default = False
         If True, static features won't be used by the model even if they are present in the dataset.
-        If False, static features will be used by the model if they are present in the dataset.
     disable_known_covariates : bool, default = False
         If True, known covariates won't be used by the model even if they are present in the dataset.
-        If False, known covariates will be used by the model if they are present in the dataset.
     num_layers : int, default = 2
         Number of RNN layers
     hidden_size : int, default = 40
@@ -170,13 +168,10 @@ class TemporalFusionTransformerModel(AbstractGluonTSModel):
         Distribution output object that defines how the model output is converted to a forecast, and how the loss is computed.
     disable_static_features : bool, default = False
         If True, static features won't be used by the model even if they are present in the dataset.
-        If False, static features will be used by the model if they are present in the dataset.
     disable_known_covariates : bool, default = False
         If True, known covariates won't be used by the model even if they are present in the dataset.
-        If False, known covariates will be used by the model if they are present in the dataset.
     disable_past_covariates : bool, default = False
         If True, past covariates won't be used by the model even if they are present in the dataset.
-        If False, past covariates will be used by the model if they are present in the dataset.
     hidden_dim : int, default = 32
         Size of the LSTM & transformer hidden states.
     variable_dim : int, default = 32
@@ -470,10 +465,8 @@ class TiDEModel(AbstractGluonTSModel):
         Number of past values used for prediction.
     disable_static_features : bool, default = False
         If True, static features won't be used by the model even if they are present in the dataset.
-        If False, static features will be used by the model if they are present in the dataset.
     disable_known_covariates : bool, default = False
         If True, known covariates won't be used by the model even if they are present in the dataset.
-        If False, known covariates will be used by the model if they are present in the dataset.
     feat_proj_hidden_dim : int, default = 4
         Size of the feature projection layer.
     encoder_hidden_dim : int, default = 64

autogluon/timeseries/models/local/__init__.py

@@ -1,5 +1,3 @@
-import joblib.externals.loky
-
 from .naive import AverageModel, NaiveModel, SeasonalAverageModel, SeasonalNaiveModel
 from .npts import NPTSModel
 from .statsforecast import (
@@ -15,8 +13,3 @@ from .statsforecast import (
     ThetaModel,
     ZeroModel,
 )
-
-# By default, joblib w/ loky backend kills processes that take >300MB of RAM assuming that this is caused by a memory
-# leak. This leads to problems for some memory-hungry models like AutoARIMA/Theta.
-# This monkey patch removes this undesired behavior
-joblib.externals.loky.process_executor._MAX_MEMORY_LEAK_SIZE = int(3e10)