Migrate to pgmpy>=1.1.2,<1.2: torch becomes optional (closes #131)

[dbolser]

This implements the migration proposed in #131, following the decisions you settled there. It's a small diff in the end (~80 lines over 9 files) — your recent master commits (vendored CITests.py, the pypickle and lingam pins) did a lot of the heavy lifting, and this PR builds directly on them.

Why

pgmpy 0.1.25 imports torch unconditionally, so every bnlearn install pulls the full CUDA stack. In pgmpy 1.x torch is an optional extra, and numpy is the default backend: a fresh bnlearn venv drops from ~1.4 GB to ~519 MB with no loss of functionality.

What changed

• Pin: pgmpy>=1.1.2,<1.2 in pyproject.toml/requirements.txt, and the import-time version check now asks for that range — the middle ground you preferred: patch fixes flow in, pgmpy's announced 1.2/1.3 removals can't break installs.
• Renamed classes: BayesianNetwork → DiscreteBayesianNetwork, MarkovNetwork → DiscreteMarkovNetwork (the old names are tombstones in 1.x whose __init__ raises). Score classes BicScore/K2Score/… → BIC/K2/BDeu/BDs/AIC.
• Parameter learning: model.fit() now takes an initialized estimator instance (DiscreteBayesianEstimator(prior_type=..., ...)) instead of a class plus kwargs.
• Hill-climb search: black_list/white_list/fixed_edges kwargs were removed from HillClimbSearch.estimate(); they map onto the new ExpertKnowledge object (forbidden_edges/search_space/required_edges). The bnlearn API is unchanged.
• Constraint search (PC): skeleton_to_pdag() is gone; estimate(return_type='pdag', variant='stable') runs the whole pipeline (variant='stable' matches the pre-1.x default). The skeleton is derived from the PDAG, so the CI tests run once instead of twice.
• Structure scores: structure_score() → pgmpy.metrics.StructureScore(...).evaluate(); the historic 'bic'/'aic' names are mapped to pgmpy's new 'bic-d'/'aic-d' so user code keeps working.
• print_CPD on MarkovNetworks: pgmpy 1.x removed to_bayesian_model(); there is no conversion left to print CPDs from, so this now warns and returns empty instead of raising.
• Old saved models: as agreed, no load-time shim — but the failure mode needed handling: 0.x pickles reference pgmpy.models.BayesianNetwork.BayesianNetwork, which 1.x still exposes as a bare tombstone, so pickle.load succeeds silently and returns a half-broken object that fails later in confusing ways. bn.load() now detects the tombstone, prints that the file needs re-saving (or pip install "bnlearn<0.14"), and returns None. Regression test included.

What's deliberately not in here

• No switch to the new sklearn-style pgmpy.causal_discovery estimators — the estimator-style classes remain supported until pgmpy 1.3, the <1.2 pin covers us, and that interface is still settling. Noted in a comment for whenever it's worth revisiting.
• No requires-python cap mirroring pgmpy's <3.15 — resolvers enforce pgmpy's bound transitively, and a local cap goes stale the moment pgmpy widens theirs.

Testing

• Full suite (90 tests, including the Add characterization tests for the pgmpy API surface (safety net for #131) #132 characterization tests with the locked inference value) passes under pgmpy 1.1.2.
• Green on the fork's 3-OS × Python 3.10/3.12/3.13 CI matrix: Migrate from pgmpy 0.1.25 to pgmpy >=1.1.2 (torch becomes optional) dbolser/bnlearn#10
• CodeRabbit review on the fork PR addressed (its remaining findings are fixed or answered on-thread).

🤖 Generated with Claude Code

https://claude.ai/code/session_017dh1KPAuSYmLpERF2T3Drm

[erdogant]

Nice work!