Feat/likelihood parameters prediction

[madtoinou]

Fixes #1735, fixes #1445.

Summary

Edit: Final version:
After fitting a regression or torch-based model with a likelihood, it's possible to pass the flag predict_likelihood_parameters=True to predict() and directly predict the distribution parameters instead of sampling (which remains the best way to simulate/visualize uncertainty).

Draft version:
The distribution parameters are returned by the Likelihood.sample() method (in addition of the sampled values), and processed in the same way as the target forecast values in predict()/get_batch_prediction() which remain necessary if the number of forecasted values is great (auto-regression). If the flag is set, the output of predict() contains only the distribution parameters (no target forecast)

Other Information

Based on @hrzn comment in #1445, a warning should probably be raised if the predict TimeSeries contain many component to avoid dimension explosion. Another solution would eventually be to restrict this feature only to a few Likelihoods.

Verifying that the model converge to the good parameters is too time consuming for the unittests but I did check manually for some distributions (Gaussian, Poisson, Quantile) and it looked rather good.

[codecov-commenter]

Codecov Report

Patch coverage: 89.61% and project coverage change: -0.17% ⚠️

Comparison is base (933316b) 94.00% compared to head (822d54c) 93.84%.

❗ Your organization is not using the GitHub App Integration. As a result you may experience degraded service beginning May 15th. Please install the Github App Integration for your organization. Read more.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1811      +/-   ##
==========================================
- Coverage   94.00%   93.84%   -0.17%     
==========================================
  Files         126      126              
  Lines       11905    12162     +257     
==========================================
+ Hits        11191    11413     +222     
- Misses        714      749      +35     

Files Changed	Coverage Δ
darts/models/forecasting/transformer_model.py	98.95% <ø> (ø)
darts/utils/timeseries_generation.py	96.15% <ø> (ø)
darts/utils/likelihood_models.py	95.55% <82.35%> (-2.31%)	⬇️
darts/models/forecasting/catboost_model.py	95.23% <83.33%> (-4.77%)	⬇️
darts/models/forecasting/regression_model.py	95.33% <84.72%> (-2.23%)	⬇️
darts/models/forecasting/forecasting_model.py	95.14% <90.00%> (-0.43%)	⬇️
darts/models/forecasting/ensemble_model.py	95.42% <94.59%> (-0.41%)	⬇️
darts/models/forecasting/baselines.py	96.15% <96.77%> (-0.15%)	⬇️
darts/ad/anomaly_model/forecasting_am.py	93.51% <100.00%> (ø)
darts/explainability/explainability.py	96.42% <100.00%> (ø)
... and 21 more

... and 4 files with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[dennisbader]

Really nice @madtoinou, this will be a great feature 🚀

From some points from our discussions:

• enforce n <= output_chunk_length when likelihood_params=True
• enforce num_samples == 1 when likelihood_params=True (I tried with DLinear, and it didn't raise an error with num_samples>1
• avoid sampling when likelihood_params=True

[dennisbader]

let's update the self.output_chunk_length of all TFMs to self._output_chunk_length as well, and then update the property accordingly

[madtoinou]

Updated, I tried to keep self.output_chunk_length when possible (available after the model/pl module is instantiated)

[dennisbader]

Congrats @madtoinou for this great PR 💯
Ready to be merged 🚀