def predict(self):
"""
Use the gluon dataset of training to predict future values and
concat all forecasts timeseries of different identifiers and quantiles together
"""
forecasts = self.predictor.predict(self.gluon_dataset)
forecasts_list = list(forecasts)
forecasts_timeseries = self._compute_forecasts_timeseries(
forecasts_list)
multiple_df = concat_timeseries_per_identifiers(forecasts_timeseries)
self.forecasts_df = concat_all_timeseries(multiple_df)
self.time_column_name = self.gluon_dataset.list_data[0][
TIMESERIES_KEYS.TIME_COLUMN_NAME]
self.identifiers_columns = (list(self.gluon_dataset.list_data[0][
TIMESERIES_KEYS.IDENTIFIERS].keys()) if TIMESERIES_KEYS.IDENTIFIERS
in self.gluon_dataset.list_data[0] else [])
if self.include_history:
self.forecasts_df = self._include_history(
self.frequency, history_length_limit=self.history_length_limit)
self.forecasts_df = add_row_origin(self.forecasts_df,
both=ROW_ORIGIN.FORECAST,
left_only=ROW_ORIGIN.HISTORY)
self.forecasts_df = self.forecasts_df.rename(
columns={"index": self.time_column_name})
def _include_history(self, frequency, history_length_limit=None):
"""Include the historical data on which the model was trained to the forecasts dataframe.
Args:
frequency (str): Used to reconstruct the date range (because a gluon ListDataset only store the start date).
history_length_limit (int): Maximum number of values to retrieve from historical data per timeseries. Default to None which means all.
Returns:
DataFrame containing both the historical data and the forecasted values.
"""
history_timeseries = self._retrieve_history_timeseries(
frequency, history_length_limit)
multiple_df = concat_timeseries_per_identifiers(history_timeseries)
history_df = concat_all_timeseries(multiple_df)
return history_df.merge(self.forecasts_df,
on=["index"] + self.identifiers_columns,
how="left",
indicator=True)
def train_evaluate(self,
train_list_dataset,
test_list_dataset,
make_forecasts=False,
retrain=False):
"""Train Model on train_list_dataset and evaluate it on test_list_dataset. Then retrain on test_list_dataset if retrain=True.
Args:
train_list_dataset (gluonts.dataset.common.ListDataset): ListDataset created with the GluonDataset class.
test_list_dataset (gluonts.dataset.common.ListDataset): ListDataset created with the GluonDataset class.
make_forecasts (bool, optional): Whether to make the evaluation forecasts and return them. Defaults to False.
retrain (bool, optional): Whether to retrain model on test_list_dataset after the evaluation. Defaults to False.
Returns:
Evaluation metrics DataFrame for each target and aggregated.
List of timeseries identifiers column names. Empty list if none found in train_list_dataset.
DataFrame of predictions for the last prediction_length timesteps of the test_list_dataset timeseries if make_forecasts is True.
"""
logger.info(f"Evaluating {self.get_label()} model performance...")
start = perf_counter()
evaluation_predictor = self._train_estimator(train_list_dataset)
agg_metrics, item_metrics, forecasts = self._make_evaluation_predictions(
evaluation_predictor, test_list_dataset)
self.evaluation_time = perf_counter() - start
logger.info(
f"Evaluating {self.get_label()} model performance: Done in {self.evaluation_time:.2f} seconds"
)
if retrain:
self.train(test_list_dataset)
metrics, identifiers_columns = self._format_metrics(
agg_metrics, item_metrics, train_list_dataset)
if make_forecasts:
median_forecasts_timeseries = self._compute_median_forecasts_timeseries(
forecasts, train_list_dataset)
multiple_df = concat_timeseries_per_identifiers(
median_forecasts_timeseries)
forecasts_df = concat_all_timeseries(multiple_df)
return metrics, identifiers_columns, forecasts_df
return metrics, identifiers_columns
def predict(self):
"""
Use the gluon dataset of training to predict future values and
concat all forecasts timeseries of different identifiers and quantiles together
"""
model_handler = ModelHandler(self.model_name)
if self.model_name and not model_handler.can_use_external_feature(
) and TIMESERIES_KEYS.FEAT_DYNAMIC_REAL in self.gluon_dataset.list_data[
0]:
# remove external features from the ListDataset used for predictions if the model cannot use them
gluon_dataset_without_external_features = remove_unused_external_features(
self.gluon_dataset, self.frequency)
forecasts = self.predictor.predict(
gluon_dataset_without_external_features)
else:
forecasts = self.predictor.predict(self.gluon_dataset)
forecasts_list = list(forecasts)
forecasts_timeseries = self._compute_forecasts_timeseries(
forecasts_list)
multiple_df = concat_timeseries_per_identifiers(forecasts_timeseries)
self.forecasts_df = concat_all_timeseries(multiple_df)
self.time_column_name = self.gluon_dataset.list_data[0][
TIMESERIES_KEYS.TIME_COLUMN_NAME]
self.identifiers_columns = (list(self.gluon_dataset.list_data[0][
TIMESERIES_KEYS.IDENTIFIERS].keys()) if TIMESERIES_KEYS.IDENTIFIERS
in self.gluon_dataset.list_data[0] else [])
if self.include_history:
self.forecasts_df = self._include_history(
self.frequency, history_length_limit=self.history_length_limit)
self.forecasts_df = add_row_origin(self.forecasts_df,
both=ROW_ORIGIN.FORECAST,
left_only=ROW_ORIGIN.HISTORY)
self.forecasts_df = self.forecasts_df.rename(
columns={"index": self.time_column_name})