Targeting a non-Julian model
Suppose you have some code implementing vanilla MCMC, written in an arbitrary "foreign" language such as C++, Python, R, Java, etc. You would like to turn this vanilla MCMC code into a Parallel Tempering algorithm able to harness large numbers of cores, including distributing this algorithm over MPI. However, you do not wish to learn anything about MPI/multi-threading/Parallel Tempering.
Surprisingly, it is very simple to bridge such code with Pigeons. The only requirement on the "foreign" language is that it supports reading the standard in and writing to the standard out, hence virtually any languages can be interfaced in this fashion. Based on this minimalist "standard stream bridge" with worker processes running foreign code (one such process per replica; not necessarily running on the same machine), Pigeons will coordinate the execution of an adaptive non-reversible parallel tempering algorithm.
This behaviour is implemented in StreamTarget, see its documentation for details. In a nutshell, there will be one child process for each PT chain. These processes will not necessarily be on the same machine: indeed distributed sampling is the key use case of this bridge. Pigeons will do some lightweight coordination with these child processes to orchestrate non-reversible parallel tempering. Interprocess communication only involves pigeons telling each child process to perform exploration at a pigeons-provided annealing parameter.
StreamTarget implements log_potential and explorer by invoking worker processes via standard stream communication. The standard stream is less efficient than alternatives such as protobuff, but it has the advantage of being supported by nearly all programming languages in existence. Also in many practical cases, since the worker process is invoked only three times per chain per iteration, it is unlikely to be the bottleneck (overhead is in the order of 0.1ms per interprocess call).
Usage example
To demonstrate this capability, we show here how it enables running Blang models in pigeons. Blang is a Bayesian modelling language designed for sampling combinatorial spaces such as phylogenetic trees.
We first setup Blang as follows (assuming Java 11 is accessible in the PATH variable):
using Pigeons
redirect_stdout(devnull) do
Pigeons.setup_blang("blangDemos")
end┌ Warning: Using a precompiled build for blangDemos: double check it is up to date
└ @ Pigeons ~/work/Pigeons.jl/Pigeons.jl/src/targets/BlangTarget.jl:141
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0
0 0 0 0 0 0 0 0 --:--:-- 0:00:01 --:--:-- 0
0 0 0 0 0 0 0 0 --:--:-- 0:00:02 --:--:-- 0
0 0 0 0 0 0 0 0 --:--:-- 0:00:03 --:--:-- 0
0 0 0 0 0 0 0 0 --:--:-- 0:00:04 --:--:-- 0
0 152M 0 609k 0 0 124k 0 0:20:53 0:00:04 0:20:49 124k
2 152M 2 3538k 0 0 603k 0 0:04:19 0:00:05 0:04:14 729k
6 152M 6 9953k 0 0 1449k 0 0:01:48 0:00:06 0:01:42 2048k
12 152M 12 19.7M 0 0 2566k 0 0:01:01 0:00:07 0:00:54 4150k
19 152M 19 29.1M 0 0 3362k 0 0:00:46 0:00:08 0:00:38 6140k
24 152M 24 37.0M 0 0 3841k 0 0:00:40 0:00:09 0:00:31 7464k
28 152M 28 44.3M 0 0 4179k 0 0:00:37 0:00:10 0:00:27 8366k
33 152M 33 50.8M 0 0 4378k 0 0:00:35 0:00:11 0:00:24 8386k
37 152M 37 57.3M 0 0 4564k 0 0:00:34 0:00:12 0:00:22 7732k
42 152M 42 64.6M 0 0 4767k 0 0:00:32 0:00:13 0:00:19 7258k
46 152M 46 71.5M 0 0 4929k 0 0:00:31 0:00:14 0:00:17 7087k
51 152M 51 78.5M 0 0 5062k 0 0:00:30 0:00:15 0:00:15 6978k
55 152M 55 85.5M 0 0 5184k 0 0:00:30 0:00:16 0:00:14 7100k
60 152M 60 92.4M 0 0 5299k 0 0:00:29 0:00:17 0:00:12 7189k
65 152M 65 99.9M 0 0 5411k 0 0:00:28 0:00:18 0:00:10 7187k
70 152M 70 107M 0 0 5517k 0 0:00:28 0:00:19 0:00:09 7252k
74 152M 74 113M 0 0 5556k 0 0:00:28 0:00:20 0:00:08 7123k
78 152M 78 119M 0 0 5601k 0 0:00:27 0:00:21 0:00:06 7015k
82 152M 82 126M 0 0 5664k 0 0:00:27 0:00:22 0:00:05 6964k
87 152M 87 133M 0 0 5736k 0 0:00:27 0:00:23 0:00:04 6978k
92 152M 92 141M 0 0 5826k 0 0:00:26 0:00:24 0:00:02 7060k
97 152M 97 149M 0 0 5900k 0 0:00:26 0:00:25 0:00:01 7340k
100 152M 100 152M 0 0 5937k 0 0:00:26 0:00:26 --:--:-- 7563kNext, we run a Blang implementation of our usual unidentifiable toy example:
using Pigeons
blang_unidentifiable_example(n_trials, n_successes) =
Pigeons.BlangTarget(
`$(Pigeons.blang_executable("blangDemos", "demos.UnidentifiableProduct")) --model.nTrials $n_trials --model.nFails $n_successes`
)
pt = pigeons(target = blang_unidentifiable_example(100, 50))┌ Info: Neither traces, disk, nor online recorders included.
│ You may not have access to your samples (unless you are using a custom recorder, or maybe you just want log(Z)).
└ To add recorders, use e.g. pigeons(target = ..., record = [traces; record_default()])
Preprocess {
2 samplers constructed with following prototypes:
RealScalar sampled via: [RealSliceSampler]
} [ endingBlock=Preprocess blockTime=409.1ms blockNErrors=0 ]
Inference {
────────────────────────────────────────────────────────────────────────────
scans Λ time(s) allc(B) log(Z₁/Z₀) min(α) mean(α)
────────── ────────── ────────── ────────── ────────── ────────── ──────────
2 0.992 5.85 1.61e+07 -4.05 0.621 0.89
4 1.42 0.0375 9.33e+06 -4.43 0.192 0.842
8 1.44 0.113 1.77e+07 -5.35 0.581 0.84
16 1.7 0.135 3.63e+07 -4.72 0.577 0.811
32 1.51 0.231 6.76e+07 -4.8 0.602 0.833
64 1.52 0.407 1.34e+08 -4.95 0.727 0.831
128 1.56 0.595 2.67e+08 -4.93 0.713 0.827
256 1.52 1.11 5.35e+08 -5.06 0.78 0.831
512 1.52 1.69 1.05e+09 -4.97 0.793 0.831
1.02e+03 1.54 3.04 2.1e+09 -4.97 0.789 0.829
────────────────────────────────────────────────────────────────────────────As shown above, create a StreamTarget amounts to specifying which command will be used to create a child process.
To terminate the child processes associated with a stream target, use:
Pigeons.kill_child_processes(pt)