def __init__(self,
train_dataset: DeepARDataset,
predictor_filepath: str,
mean: bool = True,
num_samples: int = 100,
quantiles: List[float] = [],
nan_padding: bool = True):
""" constructs DeepAR forecast object
if mean False, will return median point estimates
"""
self.train_dataset = train_dataset
self.train_frame = train_dataset.get_frame()
self.predictor = GluonPredictor.deserialize(Path(predictor_filepath))
self.mean = mean
self.prediction_length = train_dataset.get_pred_length()
self.context_length = train_dataset.get_context_length()
self.num_samples = num_samples
self.quantiles = quantiles
self.nan_padding = nan_padding
self.data = []
self.series_idxs = []
self.max_intervals = []
self.pre_pad_lens = []
self.total_in_samples = []
def __init__(
self,
train_dataset: NBEATSDataset,
predictor_filepath: str,
interpretable: bool = True,
mean: bool = True,
nan_padding: bool = True,
):
"""constructs NBEATS forecast object
if mean False, will return median point estimates
"""
self.train_dataset = train_dataset
self.train_frame = train_dataset.get_frame()
self.predictor = GluonPredictor.deserialize(Path(predictor_filepath))
self.interpretable = interpretable
if interpretable:
self.mean = True
else:
self.mean = mean
self.prediction_length = train_dataset.get_pred_length()
self.context_length = train_dataset.get_context_length()
self.nan_padding = nan_padding
self.data = []
self.series_idxs = []
self.max_intervals = []
self.pre_pad_lens = []
self.total_in_samples = []
def unserialize_all(
base_path: str,
) -> List[GluonPredictor]:
""" unserialize all predictors from directory
Arguments:
base_path {str} -- base dir name from which to load models
Returns:
List[GluonPredictor] -- list of unserialized GluonPredictor objects
"""
return [
GluonPredictor.deserialize(Path(base_dir))
for base_dir in glob(base_path + "-fold*")
]
def evaluate_deepar(
predictor: GluonPredictor,
train_data: ListDataset,
test_data: ListDataset,
hierarchy_dict: Dict[int, List[int]],
output_file: str = None,
output_mean: bool = True,
output_residuals: bool = True,
) -> Dict[Union[int, str], Dict[str, float]]:
""" aggregates error metrics for each level of the hierarchy, optionally writes predictions/in-sample residuals
to output file
Arguments:
predictor {GluonPredictor} -- predictor
train_data {ListDataset} -- train dataset
test_data {ListDataset} -- test dataset
hierarchy_dict {Dict[int, List[int]]} -- mapping from hierachy level to series prediction idxs included
in that level of hierarchy
Keyword Arguments:
output_file {str} -- output_file to save predictions (default: {None})
output_mean {bool} -- whether to output the mean (or median) predictions (default: {False})
output_residuals {bool} -- whether to output the residuals of in-sample predictions. If True,
the in-sample residuals will be prepended to the out-of-sample predictions. Thus,
if the in-sample data contains 24 timeteps, and the out-of-sample data contains 6 timesteps,
the output data frame will contain 30 rows (timesteps) (default: {True})
Returns:
Dict[Union[int, str], Dict[str, float]] -- mapping of hierarchy level (0-indexed) to dictionaries of aggregated metrics
for that level of the hierarchy
"""
eval_forecasts = []
output_forecasts = []
with tqdm(predictor.predict(train_data),
total=len(train_data),
desc="Making Predictions") as it, np.errstate(invalid='ignore'):
for forecast in it:
output_forecasts.append(
forecast.mean if output_mean else forecast.quantile(0.5))
eval_forecasts.append(forecast)
preds = np.array(output_forecasts)
if output_file:
if output_residuals:
preds = np.concatenate(
(predictor.prediction_net.residuals.asnumpy(), preds), axis=1)
out_df = pd.DataFrame(preds).T
out_df.to_csv(output_file, index=False)
eval_forecasts = np.array(eval_forecasts)
evaluator = Evaluator(quantiles=[0.5])
evaluations = {
level: evaluator([
to_pandas(series)
for series in np.array(list(test_data))[np.array(idxs)]
], eval_forecasts[np.array(idxs)])[0]
for level, idxs in hierarchy_dict.items()
}
evaluations['all'] = evaluator(
[to_pandas(series) for series in np.array(list(test_data))],
eval_forecasts)[0]
return evaluations