[DistributionLoss] Add a "return samples" option

[Antoine-Schwartz]

Description

As already proposed by DeepAR on Sagemaker, it would be interesting to be able to retrieve the output samples when using a DistributionLoss.

Use case

Retrieving these trajectories directly in outputs allows for greater independence of use cases that come after forecasting. Typically, being able to recalculate different quantiles without having to relaunch a forecast job is a must for many supply chain optimization modules.

[elephaint]

Hi, apologies for the late reply. You can achieve what you want by specifying a large number of quantiles, for example:

loss = DistributionLoss(distribution='StudentT', quantiles=np.linspace(0, 1, 100))

This will effectively allow you to obtain the sampled values for the distribution per timestep in the forecast.

Does that answer your question?