Session 23: God-class drain: `predict_model`

Baseline: session 22 drained the 4 persistence verbs. Session 23 drains the 5th OOP verb — predict_model. The heart of the rewrite is a task-aware dispatch that handles classification / regression / clustering / anomaly without ever touching self._legacy.

CHANGED — engine#

• CHANGED, BREAKING — packages/engine/pycaret/core/experiment.py — Experiment.predict_model. No longer delegates to self._legacy.predict_model. Rewritten as ~170 LoC of native dispatch. Signature slimmed from (estimator, *args, **kwargs) to (estimator, data=None, *, raw_score=False, round=4, verbose=False). All 3.x-era params are gone:
  • probability_threshold — removed. Callers thresholding on positive-class probability can do out["prediction_label"] = out["prediction_score"] >= t.
  • encoded_labels — removed. Label encoding happens inside the preprocessor; for integer labels, out["prediction_label"].map(class_to_int) is a one-liner.
  • preprocess — removed. In 4.0 the pipeline either preprocesses itself (Pipeline case) or we apply self.preprocess_pipeline automatically (bare estimator case, transitional).
  • ml_usecase — removed. Comes from self.task directly now.
• CHANGED — Transitional bare-estimator accommodation. CreateResult.pipeline today is a bare sklearn estimator (e.g. LogisticRegression), not a Pipeline. Once create_model's drain lands (session 24), it becomes a proper Pipeline with preprocessing baked in, and the transitional code-path in predict_model collapses to a one-line estimator.predict(X). Flagged in the docstring + guarded by isinstance(estimator, sklearn.pipeline.Pipeline).
• CHANGED — Metric computation inlined. Native predict_model uses the existing metric registry (pycaret.containers.metrics.{classification,regression}.get_all_metric_containers + pycaret.utils.generic.calculate_metrics) directly — no longer depends on self._legacy.pull() to surface the holdout metrics DataFrame. Wraps the whole metric block in a broad try/except → returns None on any registry hiccup. Rationale: metrics are advisory; a predict must never fail because a metric choked.
• CHANGED — Per-task output columns:
  • Classification binary → prediction_label + prediction_score (positive-class probability).
  • Classification multiclass, raw_score=False → prediction_label + prediction_score (winning-class probability).
  • Classification multiclass, raw_score=True → prediction_label + prediction_score_<class> per class.
  • Regression → prediction_label only.
  • Clustering → Cluster column with "Cluster {i}" labels.
  • Anomaly → Anomaly + Anomaly_Score (when the detector exposes decision_function).

ADDED — tests#

• ADDED — packages/engine/tests/test_session23_predict.py — 12 new tests, split by speed:
  • Fast (7 tests) — fabricate a tiny StandardScaler + {LogReg,LinReg} pipeline + a fit-sentinel Experiment; exercise raw predict paths. Confirms: non-estimator rejection (TypeError on dict), NotFittedError without fit, metrics absent when data lacks target, metrics present + model name set when data has target, regression has no score column, multiclass prediction_score is winning-class prob, multiclass raw_score sums to ~1 per row.
  • Slow (5 tests, @slow marker) — full engine E2E on juice / boston. Covers binary classification output columns, classification raw_score, regression output + metrics, event stream captures MODEL_PREDICTED with n_rows + duration_ms, and the drain-lock test (test_predict_model_does_not_call_legacy_predict_model).
• ADDED — Drain-lock test pattern. Monkeypatches exp._legacy.predict_model with a raise-on-call function, then calls exp.predict_model(pipeline) and asserts it succeeds. Any future refactor that accidentally re-delegates will fail on the Ubuntu + Windows matrix. Same shape as session 22's test_save_model_does_not_touch_legacy.

INTERNAL#

• INTERNAL — Why bare estimators are still accepted (temporarily). The pyramid of the 10-verb drain is save_model → predict_model → create_model → .... Sessions 22–23 drain verbs that consume the output of create_model. create_model still returns a bare estimator today, so strictly rejecting that in predict_model would red-light the slow E2E suite (and would also break every notebook in the wild). Session 24 drains create_model, replacing the returned bare estimator with a Pipeline. At that point the transitional branch in predict_model (+ the self.preprocess_pipeline transform call, + the estimator_is_pipeline check) can be deleted in a follow-up cleanup commit.
• INTERNAL — Task dispatch via self.task not isinstance. The legacy code used self._ml_usecase (a MLUsecase enum on the god-class). The native code reads self.task (a pycaret.core.tasks.TaskType enum on the Experiment). Cleaner because TaskType already lives on the 4.0 surface; no import needed from the internal module.
• INTERNAL — Metric registry called with empty globals_dict. get_all_metric_containers(globals_dict, raise_errors=False) — we pass {} for globals_dict because the default-behavior metrics don't need the legacy experiment's variables. If any metric registers via globals_dict reads, the raise_errors=False means it gets silently skipped rather than crashing the predict. This is consistent with the legacy behavior (which also had try/except around calculate_metrics calls for robustness).

Session 23 delta summary#