mcmc inference

[andymiller]

added support for MCMC inference

• implemented annealed importance sampling to approximate the marginal likelihood of a star vs galaxy
• code for running slice-sampling within AIS on both star and galaxy conditional distributions
• code for consolidating samples and assembling output dataframes (mcmc/mcmc_infer.jl)
• added a process_source_mcmc function in ParallelRun.jl; added a boolean do_vi flag to one_node_single_infer (defaults to VI)

The following call should return a list of res objects (dictionaries)

 res_list = ParallelRun.one_node_single_infer(
     catalog_entries, target_sources, neighbor_map, images;
     config=Config(25.0), do_vi=false)

each object is a dictionary, with the following entries

• star_samples and gal_samples : dataframe of parameter samples
• star_lls and gal_lls : log joint probability for the star or galaxy samples
• ave_pstar : average log probability that the source is a star
• star_bootstrap, gal_bootstrap: bootstrap samples of the marginal likelihood of the data under the star model and the gal model, respectively (these are used to compute the posterior prob(star | data)).

The list of dictionaries above can be consolidated into a one-source-per-row dataframe with approximate posterior means and stderr values with the following call

star_summary, gal_summary, joint_summary = MCMC.consolidate_samples(res_list)

Here, star_summary assumes all sources are stars, gal_summary assumes all sources are galaxies, and joint_summary presents star posterior inferences if exp(ave_pstar) > .5, and galaxy posterior inferences otherwise.

Closes #387 . Closes #714 . Closes #365 .

[jeff-regier]

Hi @andy Will you rebase on master, or merge master into mcmc-inference again? Looks like a lot of stuff ended up in this PR that was in previous merges. (The first commit here is from 11 months ago).

[jeff-regier]

process_source_mcmc looks good. It's hard to parse the rest because there's a lot in the diff that I think was already committed. Everything through removed stale NPZ reference was part of PR #681