| """Calibrated classifier tests (D-CAL-01..03; OQ-2 single-wrap).""" |
|
|
| from pathlib import Path |
|
|
| import numpy as np |
| import pytest |
| from lightgbm import LGBMClassifier |
| from sklearn.calibration import CalibratedClassifierCV |
|
|
| from model.features import CLASSES, load_split |
| from model.train_classifier import LGBM_PARAMS, train_calibrated_classifier |
|
|
|
|
| def _smoke_subset( |
| X: np.ndarray, y: np.ndarray, per_class: int = 100 |
| ) -> tuple[np.ndarray, np.ndarray]: |
| """Take first `per_class` rows of each class — fast smoke fixture. |
| |
| With cv=5 stratified k-fold we need >=5 samples per class; 100 is plenty |
| and keeps test wall time <30s. |
| """ |
| idx_list: list[int] = [] |
| for c in range(len(CLASSES)): |
| class_idx = np.where(y == c)[0][:per_class] |
| idx_list.extend(class_idx.tolist()) |
| idx = np.array(idx_list) |
| return X[idx], y[idx] |
|
|
|
|
| def test_lgbm_params_pin_determinism_trio() -> None: |
| """Pitfall 1 — random_state + deterministic + force_col_wise required for 1e-4.""" |
| |
| assert LGBM_PARAMS["deterministic"] is True |
| assert LGBM_PARAMS["force_col_wise"] is True |
| assert LGBM_PARAMS["class_weight"] == "balanced" |
| assert LGBM_PARAMS["n_estimators"] == 500 |
| assert LGBM_PARAMS["learning_rate"] == 0.05 |
| assert LGBM_PARAMS["num_leaves"] == 63 |
|
|
|
|
| @pytest.mark.skipif( |
| not Path("data/train.parquet").exists(), |
| reason="data/train.parquet not generated (run `make synth` first)", |
| ) |
| def test_classifier_predicts_all_10_classes() -> None: |
| """CLASS-01: argmax over calibrated probs returns one of the 10 CLASSES.""" |
| X, y, _ = load_split(Path("data/train.parquet")) |
| Xs, ys = _smoke_subset(X, y, per_class=100) |
| clf = train_calibrated_classifier(Xs, ys, classifier_seed=42, cv_seed=43) |
| preds = clf.predict(Xs) |
| assert preds.min() >= 0 |
| assert preds.max() < len(CLASSES) |
| assert len(set(preds.tolist())) >= 5 |
|
|
|
|
| @pytest.mark.skipif( |
| not Path("data/train.parquet").exists(), |
| reason="data/train.parquet not generated (run `make synth` first)", |
| ) |
| def test_predict_proba_rows_sum_to_one() -> None: |
| """CLASS-02: calibrated probability simplex.""" |
| X, y, _ = load_split(Path("data/train.parquet")) |
| Xs, ys = _smoke_subset(X, y, per_class=100) |
| clf = train_calibrated_classifier(Xs, ys, classifier_seed=42, cv_seed=43) |
| proba = clf.predict_proba(Xs) |
| assert proba.shape == (len(Xs), len(CLASSES)) |
| np.testing.assert_allclose(proba.sum(axis=1), 1.0, atol=1e-9) |
|
|
|
|
| @pytest.mark.skipif( |
| not Path("data/train.parquet").exists(), |
| reason="data/train.parquet not generated (run `make synth` first)", |
| ) |
| def test_no_double_ovr_pitfall_4() -> None: |
| """Pitfall 4 / OQ-2: single-wrap. Inner estimator is LGBMClassifier, not OvR.""" |
| X, y, _ = load_split(Path("data/train.parquet")) |
| Xs, ys = _smoke_subset(X, y, per_class=100) |
| clf = train_calibrated_classifier(Xs, ys, classifier_seed=42, cv_seed=43) |
| assert isinstance(clf, CalibratedClassifierCV) |
| assert len(clf.calibrated_classifiers_) == 5 |
| inner = clf.calibrated_classifiers_[0].estimator |
| assert isinstance(inner, LGBMClassifier), ( |
| f"Pitfall 4: inner estimator should be LGBMClassifier (single-wrap), " |
| f"got {type(inner).__name__}" |
| ) |
|
|
