Robustified Distribution Outputs

[kdgutier]

This pull request contains partial work towards robustifying the Distribution outputs of NeuralForecast models.

Before this PR, Base classes performed self._normalization and self._inv_normalization before sending the outsample_y signal to the distribution losses. This processing had bad implications for the Poisson distribution because the self._inv_normalization methods induce in edge cases negative values due to round errors, in addition to changing the signal type from integer to real.

The PR improves Poisson with the following:

• Relaxes the support constraints for the Poisson distribution to accept real-valued positive signals (almost integers).
• Changes BaseWindows.train_step and BaseWindows.validation_step to use original_outsample_y in the DistributionLoss rather than the processed outsample_y.
• Changes TFT.train_step and BaseWindows.validation_step to use original_outsample_y in the DistributionLoss rather than the processed outsample_y.

The PR improves GMM/PMM with the following:

• Added torch.logsumexp into GMM/PMM.
• Improved documentation of GMM/PMM.
• Added possibility to model batch and horizon correlations on GMM/PMM.
• Added mean ratesPMM regularization.

I document improvements in the HierarchicalNetworks.ipynb notebook that reaches SoTA CRPS with GMM/PMM, and performs reasonably well with Poisson.

There is missing work on the original_outsample_y signal for the BaseRecurrent and BaseMultivariate classes.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB