autogluon.timeseries 1.0.1b20240407__tar.gz → 1.1.0__tar.gz

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.1
 Name: autogluon.timeseries
-Version: 1.0.1b20240407
-Summary: AutoML for Image, Text, and Tabular Data
+Version: 1.1.0
+Summary: Fast and Accurate ML in 3 Lines of Code
 Home-page: https://github.com/autogluon/autogluon
 Author: AutoGluon Community
 License: Apache-2.0
@@ -14,7 +14,7 @@ Description:
         <div align="center">
         <img src="https://user-images.githubusercontent.com/16392542/77208906-224aa500-6aba-11ea-96bd-e81806074030.png" width="350">
         
-        ## AutoML for Image, Text, Time Series, and Tabular Data
+        ## Fast and Accurate ML in 3 Lines of Code
         
         [![Latest Release](https://img.shields.io/github/v/release/autogluon/autogluon)](https://github.com/autogluon/autogluon/releases)
         [![Conda Forge](https://img.shields.io/conda/vn/conda-forge/autogluon.svg)](https://anaconda.org/conda-forge/autogluon)
@@ -107,7 +107,7 @@ Description:
         This library is licensed under the Apache 2.0 License.
         
 Platform: UNKNOWN
-Classifier: Development Status :: 4 - Beta
+Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Education
 Classifier: Intended Audience :: Developers
 Classifier: Intended Audience :: Science/Research

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/setup.py

@@ -30,8 +30,7 @@ install_requires = [
     "pytorch_lightning",  # version range defined in `core/_setup_utils.py`
     "transformers[sentencepiece]",  # version range defined in `core/_setup_utils.py`
     "accelerate",  # version range defined in `core/_setup_utils.py`
-    "statsmodels>=0.13.0,<0.15",
-    "gluonts>=0.14.0,<0.15",
+    "gluonts>=0.14.0,<0.14.4",  # 0.14.4 caps pandas<2.2
     "networkx",  # version range defined in `core/_setup_utils.py`
     # TODO: update statsforecast to v1.5.0 - resolve antlr4-python3-runtime dependency clash with multimodal
     "statsforecast>=1.4.0,<1.5",
@@ -56,16 +55,16 @@ extras_require = {
         "black~=23.0",
     ],
     "chronos-openvino": [  # for faster CPU inference in pretrained models with OpenVINO
-        "optimum[openvino,nncf]>=1.17,<1.18",
+        "optimum-intel[openvino,nncf]>=1.15,<1.17",
+        "optimum[openvino,nncf]>=1.17,<1.19",
     ],
     "chronos-onnx": [  # for faster CPU inference in pretrained models with ONNX
-        "optimum[onnxruntime]>=1.17,<1.18",
+        "optimum[onnxruntime]>=1.17,<1.19",
     ],
 }
 
-extras_require["all"] = list(
-    set.union(*(set(extras_require[extra]) for extra in ["chronos-onnx", "chronos-openvino"]))
-)
+# TODO: add openvino back to "all" after dependency versions are relaxed
+extras_require["all"] = list(set.union(*(set(extras_require[extra]) for extra in ["chronos-onnx"])))
 
 install_requires = ag.get_dependency_version_ranges(install_requires)
 

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/configs/presets_configs.py

@@ -33,13 +33,13 @@ TIMESERIES_PRESETS_CONFIGS = dict(
     chronos_ensemble={
         "hyperparameters": {
             "Chronos": {"model_path": "small"},
-            **get_default_hps("default"),
+            **get_default_hps("light_inference"),
         }
     },
     chronos_large_ensemble={
         "hyperparameters": {
             "Chronos": {"model_path": "large", "batch_size": 8},
-            **get_default_hps("default"),
+            **get_default_hps("light_inference"),
         }
     },
 )

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/dataset/ts_dataframe.py

@@ -134,7 +134,7 @@ class TimeSeriesDataFrame(pd.DataFrame, TimeSeriesDataFrameDeprecatedMixin):
     ----------
     freq : str
         A pandas-compatible string describing the frequency of the time series. For example ``"D"`` for daily data,
-        ``"H"`` for hourly data, etc. This attribute is determined automatically based on the timestamps. For the full
+        ``"h"`` for hourly data, etc. This attribute is determined automatically based on the timestamps. For the full
         list of possible values, see `pandas documentation <https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_.
     num_items : int
         Number of items (time series) in the data set.
@@ -759,12 +759,6 @@ class TimeSeriesDataFrame(pd.DataFrame, TimeSeriesDataFrameDeprecatedMixin):
                 2019-02-07     4.0
 
         """
-        if self.freq is None:
-            raise ValueError(
-                "Please make sure that all time series have a regular index before calling `fill_missing_values`"
-                "(for example, using the `convert_frequency` method)."
-            )
-
         # Convert to pd.DataFrame for faster processing
         df = pd.DataFrame(self)
 
@@ -772,6 +766,12 @@ class TimeSeriesDataFrame(pd.DataFrame, TimeSeriesDataFrameDeprecatedMixin):
         if not df.isna().any(axis=None):
             return self
 
+        if not self.index.is_monotonic_increasing:
+            logger.warning(
+                "Trying to fill missing values in an unsorted dataframe. "
+                "It is highly recommended to call `ts_df.sort_index()` before calling `ts_df.fill_missing_values()`"
+            )
+
         grouped_df = df.groupby(level=ITEMID, sort=False, group_keys=False)
         if method == "auto":
             filled_df = grouped_df.ffill()
@@ -961,12 +961,12 @@ class TimeSeriesDataFrame(pd.DataFrame, TimeSeriesDataFrameDeprecatedMixin):
                 2021-06-30     6.0
                 2021-09-30     7.0
                 2021-12-31     8.0
-        >>> ts_df.convert_frequency("Y")
+        >>> ts_df.convert_frequency("YE")
                             target
         item_id timestamp
         0       2020-12-31     2.5
                 2021-12-31     6.5
-        >>> ts_df.convert_frequency("Y", agg_numeric="sum")
+        >>> ts_df.convert_frequency("YE", agg_numeric="sum")
                             target
         item_id timestamp
         0       2020-12-31    10.0

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/learner.py

@@ -43,6 +43,7 @@ class TimeSeriesLearner(AbstractLearner):
         self.prediction_length = prediction_length
         self.quantile_levels = kwargs.get("quantile_levels", [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9])
         self.cache_predictions = cache_predictions
+        self.freq: Optional[str] = None
 
         self.feature_generator = TimeSeriesFeatureGenerator(
             target=self.target, known_covariates_names=self.known_covariates_names
@@ -87,6 +88,8 @@ class TimeSeriesLearner(AbstractLearner):
         if val_data is not None:
             val_data = self.feature_generator.transform(val_data, data_frame_name="tuning_data")
 
+        self.freq = train_data.freq
+
         trainer_init_kwargs = kwargs.copy()
         trainer_init_kwargs.update(
             dict(
@@ -155,7 +158,9 @@ class TimeSeriesLearner(AbstractLearner):
                 f"known_covariates are missing information for the following item_ids: {reprlib.repr(missing_item_ids.to_list())}."
             )
 
-        forecast_index = get_forecast_horizon_index_ts_dataframe(data, prediction_length=self.prediction_length)
+        forecast_index = get_forecast_horizon_index_ts_dataframe(
+            data, prediction_length=self.prediction_length, freq=self.freq
+        )
         try:
             known_covariates = known_covariates.loc[forecast_index]
         except KeyError:
@@ -245,8 +250,8 @@ class TimeSeriesLearner(AbstractLearner):
                     raise ValueError(f"Feature {fn} not found in covariate metadata or the dataset.")
 
         if len(set(features)) < len(features):
-            logger.warning(
-                "Duplicate feature names provided to compute feature importance. This will lead to unexpected behavior. "
+            raise ValueError(
+                "Duplicate feature names provided to compute feature importance. "
                 "Please provide unique feature names across both static features and covariates."
             )
 

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/models/abstract/abstract_timeseries_model.py

@@ -31,7 +31,7 @@ class AbstractTimeSeriesModel(AbstractModel):
         If None, a new unique time-stamped directory is chosen.
     freq: str
         Frequency string (cf. gluonts frequency strings) describing the frequency
-        of the time series data. For example, "H" for hourly or "D" for daily data.
+        of the time series data. For example, "h" for hourly or "D" for daily data.
     prediction_length: int
         Length of the prediction horizon, i.e., the number of time steps the model
         is fit to forecast.
@@ -373,13 +373,14 @@ class AbstractTimeSeriesModel(AbstractModel):
         val_data: TimeSeriesDataFrame,
         store_val_score: bool = False,
         store_predict_time: bool = False,
+        **predict_kwargs,
     ) -> None:
         """Compute val_score, predict_time and cache out-of-fold (OOF) predictions."""
         past_data, known_covariates = val_data.get_model_inputs_for_scoring(
             prediction_length=self.prediction_length, known_covariates_names=self.metadata.known_covariates
         )
         predict_start_time = time.time()
-        oof_predictions = self.predict(past_data, known_covariates=known_covariates)
+        oof_predictions = self.predict(past_data, known_covariates=known_covariates, **predict_kwargs)
         self._oof_predictions = [oof_predictions]
         if store_predict_time:
             self.predict_time = time.time() - predict_start_time

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/models/autogluon_tabular/mlforecast.py

@@ -252,6 +252,15 @@ class AbstractMLForecastModel(AbstractTimeSeriesModel):
         if static_features is not None:
             df = pd.merge(df, static_features, how="left", on=ITEMID, suffixes=(None, "_static_feat"))
 
+        for col in self.metadata.known_covariates_real:
+            # Normalize non-boolean features using mean_abs scaling
+            if not df[col].isin([0, 1]).all():
+                df[f"__scaled_{col}"] = df[col] / df[col].abs().groupby(df[ITEMID]).mean().reindex(df[ITEMID]).values
+
+        # Convert float64 to float32 to reduce memory usage
+        float64_cols = list(df.select_dtypes(include="float64"))
+        df[float64_cols] = df[float64_cols].astype("float32")
+
         # We assume that df is sorted by 'unique_id' inside `TimeSeriesPredictor._check_and_prepare_data_frame`
         return df.rename(columns=column_name_mapping)
 
@@ -332,7 +341,7 @@ class AbstractMLForecastModel(AbstractTimeSeriesModel):
             Seasonal naive forecast for short series, if there are any in the dataset.
         """
         ts_lengths = data.num_timesteps_per_item()
-        short_series = ts_lengths.index[ts_lengths <= self._sum_of_differences]
+        short_series = ts_lengths.index[ts_lengths <= self._sum_of_differences + 1]
         if len(short_series) > 0:
             logger.warning(
                 f"Warning: {len(short_series)} time series ({len(short_series) / len(ts_lengths):.1%}) are shorter "
@@ -474,7 +483,7 @@ class DirectTabularModel(AbstractMLForecastModel):
         if known_covariates is not None:
             data_future = known_covariates.copy()
         else:
-            future_index = get_forecast_horizon_index_ts_dataframe(data, self.prediction_length)
+            future_index = get_forecast_horizon_index_ts_dataframe(data, self.prediction_length, freq=self.freq)
             data_future = pd.DataFrame(columns=[self.target], index=future_index, dtype="float32")
         # MLForecast raises exception of target contains NaN. We use inf as placeholder, replace them by NaN afterwards
         data_future[self.target] = float("inf")
@@ -606,7 +615,7 @@ class RecursiveTabularModel(AbstractMLForecastModel):
         if self._max_ts_length is not None:
             new_df = self._shorten_all_series(new_df, self._max_ts_length)
         if known_covariates is None:
-            future_index = get_forecast_horizon_index_ts_dataframe(data, self.prediction_length)
+            future_index = get_forecast_horizon_index_ts_dataframe(data, self.prediction_length, freq=self.freq)
             known_covariates = pd.DataFrame(columns=[self.target], index=future_index, dtype="float32")
         X_df = self._to_mlforecast_df(known_covariates, data.static_features, include_target=False)
         # If both covariates & static features are missing, set X_df = None to avoid exception from MLForecast

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/models/chronos/model.py

@@ -53,42 +53,6 @@ MODEL_ALIASES = {
 }
 
 
-class ChronosInferenceDataset:
-    """A container for time series datasets that implements the ``torch.utils.data.Dataset`` interface"""
-
-    def __init__(
-        self,
-        target_df: TimeSeriesDataFrame,
-        context_length: int,
-        target_column: str = "target",
-    ):
-        assert context_length > 0
-        self.context_length = context_length
-        self.target_array = target_df[target_column].to_numpy(dtype=np.float32)
-        self.freq = target_df.freq
-
-        # store pointer to start:end of each time series
-        cum_sizes = target_df.num_timesteps_per_item().values.cumsum()
-        self.indptr = np.append(0, cum_sizes).astype(np.int32)
-
-    def __len__(self):
-        return len(self.indptr) - 1  # noqa
-
-    def _get_context(self, a: np.ndarray, pad_value=np.nan):
-        a = a[-self.context_length :]
-        pad_size = self.context_length - len(a)
-        if pad_size > 0:
-            pad = np.full(shape=(pad_size,), fill_value=pad_value)
-            a = np.concatenate((pad, a))
-        return a
-
-    def __getitem__(self, idx) -> np.ndarray:
-        start_idx = self.indptr[idx]
-        end_idx = self.indptr[idx + 1]
-
-        return self._get_context(self.target_array[start_idx:end_idx])
-
-
 class ChronosModel(AbstractTimeSeriesModel):
     """Chronos pretrained time series forecasting models, based on the original
     `ChronosModel <https://github.com/amazon-science/chronos-forecasting>`_ implementation.
@@ -196,6 +160,7 @@ class ChronosModel(AbstractTimeSeriesModel):
         )
 
         self.model_pipeline: Optional[Any] = None  # of type OptimizedChronosPipeline
+        self.time_limit: Optional[float] = None
 
     def save(self, path: str = None, verbose: bool = True) -> str:
         pipeline = self.model_pipeline
@@ -288,14 +253,16 @@ class ChronosModel(AbstractTimeSeriesModel):
         **kwargs,
     ) -> None:
         self._check_fit_params()
+        self.time_limit = time_limit
 
     def _get_inference_data_loader(
         self,
         data: TimeSeriesDataFrame,
         context_length: int,
         num_workers: int = 0,
+        time_limit: Optional[float] = None,
     ):
-        import torch
+        from .utils import ChronosInferenceDataLoader, ChronosInferenceDataset, timeout_callback
 
         chronos_dataset = ChronosInferenceDataset(
             target_df=data,
@@ -303,11 +270,12 @@ class ChronosModel(AbstractTimeSeriesModel):
             context_length=context_length,
         )
 
-        return torch.utils.data.DataLoader(
+        return ChronosInferenceDataLoader(
             chronos_dataset,
             batch_size=self.batch_size,
             shuffle=False,
             num_workers=num_workers,
+            on_batch=timeout_callback(seconds=time_limit),
         )
 
     def _predict(
@@ -333,6 +301,12 @@ class ChronosModel(AbstractTimeSeriesModel):
                 # load model pipeline to device memory
                 self.load_model_pipeline(context_length=context_length)
 
+            inference_data_loader = self._get_inference_data_loader(
+                data=data,
+                num_workers=self.data_loader_num_workers,
+                context_length=context_length,
+                time_limit=kwargs.get("time_limit"),
+            )
             self.model_pipeline.model.eval()
             with torch.inference_mode():
                 prediction_samples = [
@@ -345,11 +319,7 @@ class ChronosModel(AbstractTimeSeriesModel):
                     .detach()
                     .cpu()
                     .numpy()
-                    for batch in self._get_inference_data_loader(
-                        data=data,
-                        num_workers=self.data_loader_num_workers,
-                        context_length=context_length,
-                    )
+                    for batch in inference_data_loader
                 ]
 
         samples = np.concatenate(prediction_samples, axis=0).swapaxes(1, 2).reshape(-1, self.num_samples)
@@ -360,10 +330,23 @@ class ChronosModel(AbstractTimeSeriesModel):
         df = pd.DataFrame(
             np.concatenate([mean, quantiles], axis=1),
             columns=["mean"] + [str(q) for q in self.quantile_levels],
-            index=get_forecast_horizon_index_ts_dataframe(data, self.prediction_length),
+            index=get_forecast_horizon_index_ts_dataframe(data, self.prediction_length, freq=self.freq),
         )
 
         return TimeSeriesDataFrame(df)
 
     def _more_tags(self) -> Dict:
         return {"allow_nan": True}
+
+    def score_and_cache_oof(
+        self,
+        val_data: TimeSeriesDataFrame,
+        store_val_score: bool = False,
+        store_predict_time: bool = False,
+        **predict_kwargs,
+    ) -> None:
+        # All computation happens during inference, so we provide the time_limit at prediction time
+        # TODO: Once custom predict_kwargs is allowed, make sure that `time_limit` is not among the keys
+        super().score_and_cache_oof(
+            val_data, store_val_score, store_predict_time, time_limit=self.time_limit, **predict_kwargs
+        )

autogluon.timeseries-1.1.0/src/autogluon/timeseries/models/chronos/utils.py

@@ -0,0 +1,66 @@
+import time
+from typing import Callable, Optional
+
+import numpy as np
+import torch
+
+from autogluon.core.utils.exceptions import TimeLimitExceeded
+from autogluon.timeseries.dataset.ts_dataframe import TimeSeriesDataFrame
+
+
+class ChronosInferenceDataset:
+    """A container for time series datasets that implements the ``torch.utils.data.Dataset`` interface"""
+
+    def __init__(
+        self,
+        target_df: TimeSeriesDataFrame,
+        context_length: int,
+        target_column: str = "target",
+    ):
+        assert context_length > 0
+        self.context_length = context_length
+        self.target_array = target_df[target_column].to_numpy(dtype=np.float32)
+        self.freq = target_df.freq
+
+        # store pointer to start:end of each time series
+        cum_sizes = target_df.num_timesteps_per_item().values.cumsum()
+        self.indptr = np.append(0, cum_sizes).astype(np.int32)
+
+    def __len__(self):
+        return len(self.indptr) - 1  # noqa
+
+    def _get_context(self, a: np.ndarray, pad_value=np.nan):
+        a = a[-self.context_length :]
+        pad_size = self.context_length - len(a)
+        if pad_size > 0:
+            pad = np.full(shape=(pad_size,), fill_value=pad_value)
+            a = np.concatenate((pad, a))
+        return a
+
+    def __getitem__(self, idx) -> np.ndarray:
+        start_idx = self.indptr[idx]
+        end_idx = self.indptr[idx + 1]
+
+        return self._get_context(self.target_array[start_idx:end_idx])
+
+
+class ChronosInferenceDataLoader(torch.utils.data.DataLoader):
+    def __init__(self, *args, **kwargs):
+        self.callback: Callable = kwargs.pop("on_batch", lambda: None)
+        super().__init__(*args, **kwargs)
+
+    def __iter__(self):
+        for item in super().__iter__():
+            yield item
+            self.callback()
+
+
+def timeout_callback(seconds: Optional[float]) -> Callable:
+    """Return a callback object that raises an exception if time limit is exceeded."""
+    start_time = time.time()
+
+    def callback() -> None:
+        if seconds is not None and time.time() - start_time > seconds:
+            raise TimeLimitExceeded
+
+    return callback

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/models/gluonts/abstract_gluonts.py

@@ -15,7 +15,6 @@ from gluonts.dataset.field_names import FieldName
 from gluonts.model.estimator import Estimator as GluonTSEstimator
 from gluonts.model.forecast import Forecast, QuantileForecast, SampleForecast
 from gluonts.model.predictor import Predictor as GluonTSPredictor
-from pandas.tseries.frequencies import to_offset
 from sklearn.compose import ColumnTransformer
 from sklearn.preprocessing import QuantileTransformer, StandardScaler
 
@@ -37,15 +36,13 @@ logger = logging.getLogger(__name__)
 gts_logger = logging.getLogger(gluonts.__name__)
 
 
-GLUONTS_SUPPORTED_OFFSETS = ["Y", "Q", "M", "W", "D", "B", "H", "T", "min", "S"]
-
-
 class SimpleGluonTSDataset(GluonTSDataset):
     """Wrapper for TimeSeriesDataFrame that is compatible with the GluonTS Dataset API."""
 
     def __init__(
         self,
         target_df: TimeSeriesDataFrame,
+        freq: str,
         target_column: str = "target",
         feat_static_cat: Optional[np.ndarray] = None,
         feat_static_real: Optional[np.ndarray] = None,
@@ -57,7 +54,6 @@ class SimpleGluonTSDataset(GluonTSDataset):
         prediction_length: int = None,
     ):
         assert target_df is not None
-        assert target_df.freq, "Initializing GluonTS data sets without freq is not allowed"
         # Convert TimeSeriesDataFrame to pd.Series for faster processing
         self.target_array = target_df[target_column].to_numpy(np.float32)
         self.feat_static_cat = self._astype(feat_static_cat, dtype=np.int64)
@@ -66,7 +62,7 @@ class SimpleGluonTSDataset(GluonTSDataset):
         self.feat_dynamic_real = self._astype(feat_dynamic_real, dtype=np.float32)
         self.past_feat_dynamic_cat = self._astype(past_feat_dynamic_cat, dtype=np.int64)
         self.past_feat_dynamic_real = self._astype(past_feat_dynamic_real, dtype=np.float32)
-        self.freq = self._to_gluonts_freq(target_df.freq)
+        self.freq = self._get_freq_for_period(freq)
 
         # Necessary to compute indptr for known_covariates at prediction time
         self.includes_future = includes_future
@@ -89,19 +85,22 @@ class SimpleGluonTSDataset(GluonTSDataset):
             return array.astype(dtype)
 
     @staticmethod
-    def _to_gluonts_freq(freq: str) -> str:
-        # FIXME: GluonTS expects a frequency string, but only supports a limited number of such strings
-        # for feature generation. If the frequency string doesn't match or is not provided, it raises an exception.
-        # Here we bypass this by issuing a default "yearly" frequency, tricking it into not producing
-        # any lags or features.
-        pd_offset = to_offset(freq)
-
-        # normalize freq str to handle peculiarities such as W-SUN
-        offset_base_alias = norm_freq_str(pd_offset)
-        if offset_base_alias not in GLUONTS_SUPPORTED_OFFSETS:
-            return "A"
+    def _get_freq_for_period(freq: str) -> str:
+        """Convert freq to format compatible with pd.Period.
+
+        For example, ME freq must be converted to M when creating a pd.Period.
+        """
+        offset = pd.tseries.frequencies.to_offset(freq)
+        freq_name = norm_freq_str(offset)
+        if freq_name == "SME":
+            # Replace unsupported frequency "SME" with "2W"
+            return "2W"
+        elif freq_name == "bh":
+            # Replace unsupported frequency "bh" with dummy value "Y"
+            return "Y"
         else:
-            return f"{pd_offset.n}{offset_base_alias}"
+            freq_name_for_period = {"YE": "Y", "QE": "Q", "ME": "M"}.get(freq_name, freq_name)
+            return f"{offset.n}{freq_name_for_period}"
 
     def __len__(self):
         return len(self.indptr) - 1  # noqa
@@ -161,6 +160,8 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
     """
 
     gluonts_model_path = "gluon_ts"
+    # we pass dummy freq compatible with pandas 2.1 & 2.2 to GluonTS models
+    _dummy_gluonts_freq = "D"
     # default number of samples for prediction
     default_num_samples: int = 250
     supports_cat_covariates: bool = False
@@ -234,13 +235,6 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
 
     def _deferred_init_params_aux(self, dataset: TimeSeriesDataFrame) -> None:
         """Update GluonTS specific parameters with information available only at training time."""
-        self.freq = dataset.freq or self.freq
-        if not self.freq:
-            raise ValueError(
-                "Dataset frequency not provided in the dataset, fit arguments or "
-                "during initialization. Please provide a `freq` string to `fit`."
-            )
-
         model_params = self._get_model_params()
         disable_static_features = model_params.get("disable_static_features", False)
         if not disable_static_features:
@@ -371,7 +365,7 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
         init_args.setdefault("early_stopping_patience", 20)
         init_args.update(
             dict(
-                freq=self.freq,
+                freq=self._dummy_gluonts_freq,
                 prediction_length=self.prediction_length,
                 quantiles=self.quantile_levels,
                 callbacks=self.callbacks,
@@ -502,6 +496,7 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
 
             return SimpleGluonTSDataset(
                 target_df=time_series_df[[self.target]],
+                freq=self.freq,
                 target_column=self.target,
                 feat_static_cat=feat_static_cat,
                 feat_static_real=feat_static_real,
@@ -592,7 +587,7 @@ class AbstractGluonTSModel(AbstractTimeSeriesModel):
             predicted_targets = self._predict_gluonts_forecasts(data, known_covariates=known_covariates, **kwargs)
             df = self._gluonts_forecasts_to_data_frame(
                 predicted_targets,
-                forecast_index=get_forecast_horizon_index_ts_dataframe(data, self.prediction_length),
+                forecast_index=get_forecast_horizon_index_ts_dataframe(data, self.prediction_length, freq=self.freq),
             )
         return df
 

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/models/gluonts/torch/models.py

@@ -423,6 +423,4 @@ class WaveNetModel(AbstractGluonTSModel):
         init_kwargs.setdefault("seasonality", get_seasonality(self.freq))
         init_kwargs.setdefault("time_features", get_time_features_for_frequency(self.freq))
         init_kwargs.setdefault("num_parallel_samples", self.default_num_samples)
-        # WaveNet model fails if an unsupported frequency such as "SM" is provided. We provide a dummy freq instead
-        init_kwargs["freq"] = "H"
         return init_kwargs

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/models/local/abstract_local_model.py

@@ -144,9 +144,10 @@ class AbstractLocalModel(AbstractTimeSeriesModel):
 
         # timeout ensures that no individual job takes longer than time_limit
         # TODO: a job started late may still exceed time_limit - how to prevent that?
-        timeout = None if self.n_jobs == 1 else self.time_limit
+        time_limit = kwargs.get("time_limit")
+        timeout = None if self.n_jobs == 1 else time_limit
         # end_time ensures that no new jobs are started after time_limit is exceeded
-        end_time = None if self.time_limit is None else time.time() + self.time_limit
+        end_time = None if time_limit is None else time.time() + time_limit
         executor = Parallel(self.n_jobs, timeout=timeout)
 
         try:
@@ -165,23 +166,28 @@ class AbstractLocalModel(AbstractTimeSeriesModel):
                 f"({fraction_failed_models:.1%}). Fallback model SeasonalNaive was used for these time series."
             )
         predictions_df = pd.concat([pred for pred, _ in predictions_with_flags])
-        predictions_df.index = get_forecast_horizon_index_ts_dataframe(data, self.prediction_length)
+        predictions_df.index = get_forecast_horizon_index_ts_dataframe(data, self.prediction_length, freq=self.freq)
         return TimeSeriesDataFrame(predictions_df)
 
     def score_and_cache_oof(
-        self, val_data: TimeSeriesDataFrame, store_val_score: bool = False, store_predict_time: bool = False
+        self,
+        val_data: TimeSeriesDataFrame,
+        store_val_score: bool = False,
+        store_predict_time: bool = False,
+        **predict_kwargs,
     ) -> None:
-        super().score_and_cache_oof(val_data, store_val_score, store_predict_time)
-        # Remove time_limit for future predictions
-        self.time_limit = None
+        # All computation happens during inference, so we provide the time_limit at prediction time
+        super().score_and_cache_oof(
+            val_data, store_val_score, store_predict_time, time_limit=self.time_limit, **predict_kwargs
+        )
 
     def _predict_wrapper(self, time_series: pd.Series, end_time: Optional[float] = None) -> Tuple[pd.DataFrame, bool]:
         if end_time is not None and time.time() >= end_time:
             raise TimeLimitExceeded
 
+        model_failed = False
         if time_series.isna().all():
             result = self._dummy_forecast.copy()
-            model_failed = True
         else:
             try:
                 result = self._predict_with_local_model(
@@ -190,7 +196,6 @@ class AbstractLocalModel(AbstractTimeSeriesModel):
                 )
                 if not np.isfinite(result.values).all():
                     raise RuntimeError("Forecast contains NaN or Inf values.")
-                model_failed = False
             except Exception:
                 if self.use_fallback_model:
                     result = seasonal_naive_forecast(
@@ -225,8 +230,6 @@ def seasonal_naive_forecast(
         return arr[np.maximum.accumulate(idx)]
 
     forecast = {}
-    # Convert to float64 since std computation can be unstable in float32
-    target = target.astype(np.float64)
     # At least seasonal_period + 2 values are required to compute sigma for seasonal naive
     if len(target) > seasonal_period + 1 and seasonal_period > 1:
         if np.isnan(target[-(seasonal_period + 2) :]).any():

{autogluon.timeseries-1.0.1b20240407 → autogluon.timeseries-1.1.0}/src/autogluon/timeseries/models/multi_window/multi_window_model.py

@@ -189,6 +189,7 @@ class MultiWindowBacktestingModel(AbstractTimeSeriesModel):
         val_data: TimeSeriesDataFrame,
         store_val_score: bool = False,
         store_predict_time: bool = False,
+        **predict_kwargs,
     ) -> None:
         # self.val_score, self.predict_time, self._oof_predictions already saved during _fit()
         assert self._oof_predictions is not None