A features table will also be compiled for the various PPLs. These may include:
support / workarounds for missing data, with MWE.
support / workarounds for ragged arrays, with MWE.
support / workarounds for inference of discrete parameters, with MWE
Nimble and Pyro supports direct inference on discrete parameters. The recommended workaround in other PPL's is marginalizing discrete parameters, but this is not always possible.
support for automatic differentiation
Customizability
e.g. For MCMC, using a custom (user-provided) implementation to update a subset of model parameters, and use default update mechanisms (Metropolis-within-Gibbs or HMC) for the other parameters.
support for HMC, Metropolis-within-Gibbs, ADVI / BBVI, and auto-tuning for each PPL.
[ ] The table below is an example of what the Feature Comparisons table could look like.
A features table will also be compiled for the various PPLs. These may include:
support / workarounds for missing data, with MWE.
support / workarounds for ragged arrays, with MWE.
support / workarounds for inference of discrete parameters, with MWE
support for automatic differentiation
Customizability
support for HMC, Metropolis-within-Gibbs, ADVI / BBVI, and auto-tuning for each PPL.
[ ] The table below is an example of what the Feature Comparisons table could look like.