[FEAT] DistributionLoss

[kdgutier]

This PR has minimal implications to NeuralForecast models.

• Added a new domain_map method for all the PyTorch train losses.
• Added call to domain_map at the end of all the models' forward.
• Deprecated if condition on the dimension of the univariate losses in favor of a domain_map squeezing automatically the output last dimension.
• Changed the BaseRecurrent normalize and invert_normalize methods to operate with new squeezed univariate outputs.
• Added DistributionLoss, with Poisson, Normal and StudentT options.
• Validated minimal effects on all the models.

PENDING:

• Use the AffineTransformation on the loss function to anchor the predictions scale and have the loss function operate in the data's "original" scale.
• Add a circle-ci test for quantilic predictions.
• Change PMM and GMM for DistributionLoss operation compatibility.
• Change Normal, and StudentT loss to initialize or wrangle the scale parameter so that it initializes overestimating it, that will smooth the optimization process (maybe clip gradients also).

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB