Automatic marginalization of finite discrete variables in the logp

[ricardoV94]

Added a Model transformation, to automatically marginalize discrete variables from the logp graph.

I first attempted to modify the original model in-place so that the FiniteDiscreteMarginalRVs are introduced in the model RV graph, but this led to many issues, in the initial_point, step sampler assignment and so on, which don't really need to know about the modified logp graph.

My latest approach is to clone the underlying model, modify that in place and request its logp, everytime the logp of the user model is requested.

TODO

• Marginalize finite discrete RVs and replace by SymbolicRV subgraph
• Suppress annoying RVs missing in graph
  • Fixed by making the marginalized RV part of the subgraph
• Test nested marginalized RVs
• Test multiple dependent RVs
  • Aeppl is too restrictive in that it expects one logp-term per value variable, can we group them in a tuple?
• Test transformed dependent RV
  • Currently not supported due to Aeppl limitations
• Dispatch default_transform and moment
  • No longer needed because we don't modify the original model
• Add sampling procedure to retrieve marginalized RVs from prior and posterior
  • Will do in follow up Implement sampling routine to recover marginalized RVs from posterior trace #94
• Allow Deterministics that don't depend on marginalized RVs
• Allow Potentials that don't depend on marginalized RVs
• Fix graphviz for marginalized RVs
  • No longer a problem because we don't modify the original model
• Fix string representation for marginalized RVs
  • No longer a problem because we don't modify the original model
• Finish tests
• Finish docstrings
• Update PyMC dependency

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[ricardoV94]

This is ready for review, I decided to leave the posterior recovery of marginalized RVs for a subsequent PR as this one is already big as is

[ricardoV94]

Need an approving review :)

[zaxtax]

Looks good to me.