def get_gluon_train_dataset(self):
""" Retrieve the GluonDataset object with training data that was saved in the model folder during training """
gluon_train_dataset_path = f"{self.session_path}/gluon_train_dataset.pk.gz"
gluon_train_dataset = read_from_folder(self.folder,
gluon_train_dataset_path,
ObjectType.PICKLE_GZ)
return gluon_train_dataset
def get_model_predictor(self):
""" Retrieve the GluonTS Predictor object obtained during training and saved into the model folder """
if not self.manual_selection:
self.session_name = self._get_last_session()
self.session_path = os.path.join(self.partition_root,
self.session_name)
self.model_label = self._get_best_model()
model_path = os.path.join(self.session_path, self.model_label,
"model.pk.gz")
try:
model = read_from_folder(self.folder, model_path,
ObjectType.PICKLE_GZ)
except ValueError as e:
raise ModelSelectionError(
f"Unable to retrieve model '{self.model_label}' from session '{self.session_name}'. "
+
f"Please make sure that it exists in the Trained model folder. Full error: {e}"
)
return model
def _get_best_model(self):
"""Find the best model according to self.performance_metric based on the aggregated metric rows
Returns:
Label of the best model.
"""
available_models_labels = list_available_models_labels()
df = read_from_folder(self.folder, f"{self.session_path}/metrics.csv",
ObjectType.CSV)
try:
if (df[METRICS_DATASET.TARGET_COLUMN] ==
METRICS_DATASET.AGGREGATED_ROW).any():
df = df[df[METRICS_DATASET.TARGET_COLUMN] ==
METRICS_DATASET.AGGREGATED_ROW]
assert df[METRICS_DATASET.MODEL_COLUMN].nunique() == len(
df.index), "More than one row per model"
model_label = df.loc[df[self.performance_metric].idxmin()][
METRICS_DATASET.MODEL_COLUMN] # or idxmax() if maximize metric
assert model_label in available_models_labels, "Best model retrieved is not an available models"
except Exception as e:
raise ModelSelectionError(
f"Unable to find the best model of session '{self.session_name}' with the performance metric '{self.performance_metric}'. Full error: {e}"
)
return model_label