Facilities for manipulating models as sets of particles

[axch]

• [x] Fork a model (copying all the particles)
• [ ] Split a model (separating the particles into different models according to some pattern)
• [ ] Concatenate models by unioning the particle sets (must they have come from the same model program?)
• [ ] Expand a model by running more particles from the prior and likelihood weighting them
• [ ] Option: Clean a model by eliminating particles with -inf log weight (without otherwise resampling, which would do this automatically)
  • Problem 1: What if there are no particles left after this? Technically OK, but can't sample or resample any more
  • Problem 2: To conserve average weight (e.g. for estimating the evidence) would want to retain a count of how many particles were removed

What's the universal interface that lets the user smoothly code the above?

[axch]

See also #369.

[axch]

The central locations of the relevant code are python/lib/engine/trace_set.py and python/lib/engine/engine.py. To expose these facilities to the inference programming language, it will also be necessary to touch backend/lite/inference_sps.py and probably python/lib/engine/inference.py.