Expose NBEATSEnsemble aggregation_method and wrong dimensions

[gorold]

Description

(A clear and concise description of what the bug is.)

1. NBEATSEnsemblePredictor has the option of selecting aggregation_method, but it is not exposed to the estimator class. This means that there is no natural way of selecting this parameter. It is possible to interpret the forecasts from the ensemble as samples, and calculate the quantiles from an ensemble. This parameter should be exposed to the estimator class.
2. Manually selecting the aggregation_method="none" still leads to a bug, wrong dimension size for SampleForecast. The bug fails silently when evaluated with the Evaluator class, and leads to erroneous metrics.

gluonts/src/gluonts/mx/model/n_beats/_ensemble.py

Lines 171 to 196 in 784a882

	for item in dataset:
	output = []
	start_date = None
	for iterator in iterators:
	prediction = next(iterator)
	# on order to avoid mypys complaints
	assert isinstance(prediction, SampleForecast)
	output.append(prediction.samples)
	# get the forecast start date
	if start_date is None:
	start_date = prediction.start_date
	output = np.stack(output, axis=0)
	# aggregating output of different models
	# default according to paper is median,
	# but we can also make use of not aggregating
	if self.aggregation_method == "median":
	output = np.median(output, axis=0)
	elif self.aggregation_method == "mean":
	output = np.mean(output, axis=0)
	else: # "none": do not aggregate
	pass

Issue should be from here, where output is shape (N, 1, T)

To Reproduce

(Please provide minimal example of code snippet that reproduces the error. For existing examples, please provide link.)

from gluonts.mx import NBEATSEnsembleEstimator, NBEATSEnsemblePredictor
from gluonts.dataset.repository.datasets import get_dataset

dataset = get_dataset("airpassengers")

training_data, test_gen = split(dataset.train, offset=-36)
model = NBEATSEnsembleEstimator(
    prediction_length=12, freq="M", 
    meta_context_length=[2], meta_bagging_size=1,
    trainer=Trainer(epochs=1)
).train(training_data)

predictor = NBEATSEnsemblePredictor(12, model.predictors, "none")
test_data = test_gen.generate_instances(12)
forecast_it = predictor.predict(test_data.input)
forecast = next(forecast_it)

print(forecast.samples.shape)

Error message or code output

(Paste the complete error message, including stack trace, or the undesired output that the above snippet produces.)

(3, 1, 12)

Desired dimensionality should be (3, 12)

Environment

• Operating system:
• Python version:
• GluonTS version: 0.11.6
• MXNet version:

(Add as much information about your environment as possible, e.g. dependencies versions.)

[lostella]

@gorold both points make perfect sense, thank you. A fix is ready in #2598, do you want to take a look?