Module general_graph 

General-graph CRF training via loopy belief propagation.

Supports arbitrary pairwise factor graphs (not just linear chains). Inference uses sum-product message passing; MAP uses max-product.

Reference: Koller & Friedman 2009, “Probabilistic Graphical Models”, Chapter 11 (Belief Propagation).

Graph structure:

• n_nodes variable nodes, each with n_labels possible states.
• edges: Vec<(usize, usize)> — undirected pairwise edges.
• Node potentials φ_i(y_i): [n_nodes × n_labels].
• Edge potentials ψ_{ij}(y_i, y_j): [n_edges × n_labels × n_labels].

Loopy BP is approximate for graphs with cycles. Guaranteed exact for trees and linear chains.

Structs

Edge

One undirected pairwise edge in the factor graph.

GeneralGraphCrf

General-graph pairwise CRF with loopy belief propagation inference.

GraphCrfConfig

Configuration for the general-graph CRF.