bayesian-cpp

A fast Bayesian network inference engine written in C++20, exposed to Python via pybind11.

bayesian-cpp performs exact probabilistic inference on discrete Bayesian networks: marginals, full joint distributions, conditionals with evidence, and MAP queries. All probability tables are kept in log-space for numerical stability, and inference runs in native C++ — orders of magnitude faster than pure-Python implementations such as pgmpy.

Features

• Variable elimination — exact marginal / conditional inference with a min-fill elimination order and per-query order caching.
• Junction tree propagation — build the clique tree once and answer many repeated queries (or new evidence) cheaply from calibrated cliques.
• Queries: full joint P(X, Y, Z, …), marginals P(X, Z) (all other variables summed out), conditionals P(X | Y, Z) and P(X | Y=0, Z=1), and maximum a posteriori (MAP) assignments.
• Log-space tables with log-sum-exp marginalization — robust against underflow on large networks.
• Self-contained — pure C++20, no third-party C++ dependencies; the Python extension builds from source with only pybind11 (fetched automatically by pip).

Install

pip install bayesian-cpp

Requires Python ≥ 3.8 and a C++20 compiler. On Linux a prebuilt manylinux wheel is installed; on other platforms pip builds from the source distribution.

Quick start

import bayesian
from bayesian import Variable, Potential, Inference, JunctionTree

x = Variable(0, "X", 2)
y = Variable(1, "Y", 2)
z = Variable(2, "Z", 2)

factors = [
    Potential([x], [0.5, 0.5]),                     # P(X)
    Potential([x, y], [0.9, 0.1, 0.2, 0.8]),       # P(Y | X)
    Potential([x, z], [0.8, 0.2, 0.1, 0.9]),       # P(Z | X)
]

engine = Inference(factors)

p_y          = engine.marginal([y])                          # P(Y)
p_xz         = engine.marginal([x, z])                       # P(X, Z)
p_x_given    = engine.conditional_given([x], {y: 0, z: 1})   # P(X | Y=0, Z=1)
posterior    = engine.full_joint()                           # P(X, Y, Z)
mode         = engine.map_query([x], {y: 1})                 # MAP assignment

# Repeated queries are cheap after the junction tree is calibrated once.
jt = JunctionTree(factors)
jt.set_evidence({z: 0})
p_x = jt.marginal([x])

Every query returns a Potential; use .probabilities() for a flat list in row-major order over .variables, or .value({var: state, ...}) for a single cell.

Performance

Because inference runs in native C++ with log-space arithmetic, it is far faster than pure-Python engines. On a 16-variable tree network (dataset fitted with MLE, same queries):

Outputs agree with pgmpy to within ~1e-7 (float32 vs float64 rounding).

Exceptions

• ValueError — invalid arguments (negative probabilities, unknown variables, overlapping query/evidence, out-of-range evidence states).
• RuntimeError — evidence with zero probability.

Repository

Source: https://github.com/Kemsekov/BayesianNetworkCpp

The repository contains the full C++ engine (src/), a pytest/gtest test suite, a benchmark suite, and a example-python/ folder comparing bayesian-cpp head-to-head against pgmpy (outputs and wall-clock performance).