"""配布用の参照実装（`deploy/color_model.py`）のテスト. 仕様（実装サマリーの設計を仕様として扱う）: - NumPy と標準ライブラリだけに依存する（本パッケージの他モジュール・sklearn・torch・pydantic・ Pillow 等を import しない．TIASjudge にこのファイルだけをコピーして使うため） - `downscale_nearest` は Pillow の `Image.resize(..., NEAREST)` による縮小と画素単位で完全一致する - `extract_features` は学習側の `extract_color_features(downscale_nearest(...))` と完全一致する - `predict_proba` は学習側の sklearn パイプラインの `predict_proba` と差 1e-12 以内 - JSON に書いて読み直しても予測は変わらない - 形状・dtype の不正，舌領域が 0 画素，未対応の版・種類・キー欠落・形状不一致は ValueError """ from __future__ import annotations import ast import json from pathlib import Path import numpy as np import pytest from PIL import Image from tias_judge_model.data.classes import CLASS_DISPLAY_NAMES, CLASS_SLUGS from tias_judge_model.data.color_features import ( NUM_COLOR_FEATURES, color_feature_names, extract_color_features, ) from tias_judge_model.data.color_features import downscale_nearest as train_downscale from tias_judge_model.deploy import color_model as deploy from tias_judge_model.schema import ExperimentConfig from tias_judge_model.training.color_cross_validation import fit_classifier, make_classifier from tias_judge_model.training.color_export import ( ColorExportResult, CSelection, build_color_model_document, ) DEPLOY_PATH = Path(deploy.__file__) ALLOWED_TOP_LEVEL_IMPORTS = { "__future__", "json", "os", "collections", "dataclasses", "pathlib", "typing", "numpy", "math", } MAX_SIDE = 48 SEED = 7 # --- 依存の制約 --- def _imported_modules(path: Path) -> set[str]: tree = ast.parse(path.read_text(encoding="utf-8")) modules: set[str] = set() for node in ast.walk(tree): if isinstance(node, ast.Import): modules.update(alias.name for alias in node.names) elif isinstance(node, ast.ImportFrom): assert node.level == 0, "相対 import をしてはならない（単一ファイルで使うため）" assert node.module is not None modules.add(node.module) return modules def test_deploy_module_imports_only_numpy_and_stdlib() -> None: modules = _imported_modules(DEPLOY_PATH) top = {m.split(".")[0] for m in modules} forbidden = {"tias_judge_model", "sklearn", "torch", "pydantic", "PIL", "cv2", "scipy"} assert not (top & forbidden), top & forbidden assert top <= ALLOWED_TOP_LEVEL_IMPORTS, top - ALLOWED_TOP_LEVEL_IMPORTS def test_deploy_module_has_no_dynamic_imports() -> None: src = DEPLOY_PATH.read_text(encoding="utf-8") assert "importlib" not in src assert "__import__" not in src # --- テスト用のモデル --- def _config(max_side: int | None = MAX_SIDE, class_weighting: str = "none") -> ExperimentConfig: return ExperimentConfig.model_validate( { "data": {"root": "unused", "severities": ["strong", "normal"]}, "split": {"seed": SEED, "n_folds": 5}, "train": {"class_weighting": class_weighting}, "color": { "classifier": "logistic", "max_side": max_side, "logistic": {"c_values": [0.5, 2.0], "max_iter": 2000}, }, } ) def _random_tongue(rng: np.random.Generator, height: int, width: int) -> np.ndarray: """楕円の舌領域（ランダムな赤み）＋背景の黒の RGB 画像.""" yy, xx = np.mgrid[0:height, 0:width] inside = ((yy - height / 2) / (height * 0.45)) ** 2 + ((xx - width / 2) / (width * 0.45)) ** 2 mask = inside <= 1.0 base = rng.integers(60, 230, size=3) img = np.clip(base + rng.normal(0, 25, size=(height, width, 3)), 1, 255).astype(np.uint8) img[~mask] = 0 return img @pytest.fixture(scope="module") def trained() -> dict: rng = np.random.default_rng(SEED) images = [ _random_tongue(rng, int(rng.integers(30, 90)), int(rng.integers(30, 90))) for _ in range(40) ] config = _config() features = np.stack([extract_color_features(train_downscale(img, MAX_SIDE)) for img in images]) labels = np.arange(len(images)) % 5 assert config.color is not None pipeline = make_classifier(config.color, {"C": 1.0}, SEED) fit_classifier(pipeline, features, labels, None) result = ColorExportResult( model=pipeline, selection=CSelection( candidates=[ {"C": 0.5, "fold_macro_f1": [0.1], "mean_macro_f1": 0.1}, {"C": 1.0, "fold_macro_f1": [0.2], "mean_macro_f1": 0.2}, ], selected_index=1, fold_counts=[{"train": 32, "held_out": 8}], ), class_counts=(8, 8, 8, 8, 8), class_weights=None, seed=SEED, ) doc = build_color_model_document(config, result, run_id="RUN", git_commit="abc1234") return {"doc": doc, "pipeline": pipeline, "images": images, "rng": rng} # --- downscale_nearest の一致 --- @pytest.mark.parametrize( ("height", "width", "max_side"), [ (100, 60, 48), (60, 100, 48), (1000, 999, 512), (513, 512, 512), (37, 1001, 50), (1001, 37, 50), (7, 3, 2), (3, 7, 1), (997, 991, 131), (1500, 2000, 512), (255, 255, 254), ], ) def test_downscale_matches_pillow_nearest(height: int, width: int, max_side: int) -> None: rng = np.random.default_rng(height * 31 + width) img = rng.integers(0, 256, size=(height, width, 3), dtype=np.uint8) got = deploy.downscale_nearest(img, max_side) scale = max_side / max(height, width) size = (max(1, round(width * scale)), max(1, round(height * scale))) expected = np.asarray(Image.fromarray(img).resize(size, Image.Resampling.NEAREST)) assert got.shape == expected.shape assert max(got.shape[:2]) <= max_side np.testing.assert_array_equal(got, expected) np.testing.assert_array_equal(got, train_downscale(img, max_side)) def test_downscale_returns_input_when_small_enough() -> None: img = np.arange(4 * 5 * 3, dtype=np.uint8).reshape(4, 5, 3) np.testing.assert_array_equal(deploy.downscale_nearest(img, 5), img) np.testing.assert_array_equal(deploy.downscale_nearest(img, 100), img) def test_downscale_rejects_non_positive_max_side() -> None: img = np.zeros((4, 4, 3), dtype=np.uint8) with pytest.raises(ValueError): deploy.downscale_nearest(img, 0) def test_downscale_matches_pillow_over_many_sizes() -> None: rng = np.random.default_rng(123) for _ in range(200): height, width = int(rng.integers(2, 700)), int(rng.integers(2, 700)) max_side = int(rng.integers(1, max(height, width) + 1)) img = rng.integers(0, 256, size=(height, width, 3), dtype=np.uint8) np.testing.assert_array_equal( deploy.downscale_nearest(img, max_side), train_downscale(img, max_side), err_msg=f"{height}x{width} -> {max_side}", ) # --- 特徴量の一致 --- def test_extract_features_matches_training_side(trained: dict) -> None: model = deploy.load_color_model(trained["doc"]) rng = np.random.default_rng(99) for size in [(20, 30), (48, 48), (120, 70), (70, 300), (1, 1)]: img = _random_tongue(rng, *size) if min(size) > 1 else np.full((1, 1, 3), 9, np.uint8) got = deploy.extract_features(img, model) expected = extract_color_features(train_downscale(img, MAX_SIDE)) np.testing.assert_array_equal(got, expected) assert got.dtype == np.float64 assert got.shape == (NUM_COLOR_FEATURES,) def test_extract_features_without_downscale_uses_full_image(trained: dict) -> None: model = deploy.load_color_model(trained["doc"]) img = _random_tongue(np.random.default_rng(5), 200, 150) np.testing.assert_array_equal( deploy.extract_features(img, model, downscale=False), extract_color_features(img) ) def test_extract_features_with_null_max_side_does_not_downscale(trained: dict) -> None: doc = json.loads(json.dumps(trained["doc"])) doc["features"]["max_side"] = None model = deploy.load_color_model(doc) img = _random_tongue(np.random.default_rng(6), 200, 150) np.testing.assert_array_equal(deploy.extract_features(img, model), extract_color_features(img)) def test_extract_features_covers_all_rgb_values(trained: dict) -> None: # 256 階調すべてを含む画像（線形化の LUT の全要素）でも学習側と一致する model = deploy.load_color_model(trained["doc"]) rng = np.random.default_rng(11) img = np.stack([rng.permutation(256), rng.permutation(256), np.arange(256)], axis=-1).astype( np.uint8 )[None] # (1, 256, 3)．(0,0,0) の画素を含みうる img = np.repeat(img, 3, axis=0) np.testing.assert_array_equal( deploy.extract_features(img, model, downscale=False), extract_color_features(img) ) @pytest.mark.parametrize( "img", [ np.zeros((4, 4), dtype=np.uint8), np.zeros((4, 4, 4), dtype=np.uint8), np.ones((4, 4, 3), dtype=np.float32), np.ones((4, 4, 3), dtype=np.int64), ], ids=["2d", "4ch", "float32", "int64"], ) def test_extract_features_rejects_bad_input(trained: dict, img: np.ndarray) -> None: model = deploy.load_color_model(trained["doc"]) with pytest.raises(ValueError): deploy.extract_features(img, model) def test_extract_features_rejects_all_background(trained: dict) -> None: model = deploy.load_color_model(trained["doc"]) with pytest.raises(ValueError): deploy.extract_features(np.zeros((10, 10, 3), dtype=np.uint8), model) # --- 予測確率の一致 --- def test_predict_proba_matches_sklearn(trained: dict) -> None: model = deploy.load_color_model(trained["doc"]) rng = np.random.default_rng(1234) for i in range(30): img = _random_tongue(rng, int(rng.integers(10, 150)), int(rng.integers(10, 150))) got = deploy.predict_proba(img, model) feats = extract_color_features(train_downscale(img, MAX_SIDE)) expected = trained["pipeline"].predict_proba(feats[None])[0] assert got.shape == (5,) assert np.max(np.abs(got - expected)) <= 1e-12, i assert got.sum() == pytest.approx(1.0, abs=1e-12) def test_predict_proba_from_features_batch_matches_sklearn(trained: dict) -> None: model = deploy.load_color_model(trained["doc"]) rng = np.random.default_rng(3) feats = np.stack([extract_color_features(_random_tongue(rng, 40, 40)) for _ in range(10)]) got = deploy.predict_proba_from_features(feats, model) expected = trained["pipeline"].predict_proba(feats) assert got.shape == (10, 5) assert np.max(np.abs(got - expected)) <= 1e-12 def test_predict_proba_from_features_is_stable_for_extreme_values(trained: dict) -> None: model = deploy.load_color_model(trained["doc"]) feats = np.full(NUM_COLOR_FEATURES, 1e6) probs = deploy.predict_proba_from_features(feats, model) assert np.all(np.isfinite(probs)) assert probs.sum() == pytest.approx(1.0) @pytest.mark.parametrize("shape", [(NUM_COLOR_FEATURES - 1,), (2, 3, NUM_COLOR_FEATURES), (3,)]) def test_predict_proba_from_features_rejects_bad_shape(trained: dict, shape: tuple) -> None: model = deploy.load_color_model(trained["doc"]) with pytest.raises(ValueError): deploy.predict_proba_from_features(np.zeros(shape), model) def test_json_round_trip_keeps_predictions(trained: dict, tmp_path: Path) -> None: path = tmp_path / "color_model.json" path.write_text(json.dumps(trained["doc"], ensure_ascii=False, indent=2), encoding="utf-8") from_file = deploy.load_color_model(path) from_str_path = deploy.load_color_model(str(path)) from_dict = deploy.load_color_model(trained["doc"]) rng = np.random.default_rng(77) for _ in range(10): img = _random_tongue(rng, 64, 50) p = deploy.predict_proba(img, from_dict) np.testing.assert_array_equal(deploy.predict_proba(img, from_file), p) np.testing.assert_array_equal(deploy.predict_proba(img, from_str_path), p) # --- モデルの読み込み --- def test_loaded_model_exposes_classes_and_features(trained: dict) -> None: model = deploy.load_color_model(trained["doc"]) assert model.num_features == NUM_COLOR_FEATURES assert model.num_classes == 5 assert model.feature_names == color_feature_names() assert [c["slug"] for c in model.classes] == list(CLASS_SLUGS) assert [c["display_name"] for c in model.classes] == [ CLASS_DISPLAY_NAMES[s] for s in CLASS_SLUGS ] assert model.max_side == MAX_SIDE def _broken(doc: dict, mutate) -> dict: d = json.loads(json.dumps(doc)) mutate(d) return d @pytest.mark.parametrize( "mutate", [ lambda d: d.update(format_version=2), lambda d: d.update(format_version="1"), lambda d: d.update(model_type="color_hist_gb"), lambda d: d.pop("scaler"), lambda d: d.pop("classes"), lambda d: d["features"].pop("lab"), lambda d: d["classifier"].update(coef=d["classifier"]["coef"][:4]), lambda d: d["classifier"].update(intercept=d["classifier"]["intercept"][:4]), lambda d: d["scaler"].update(mean=d["scaler"]["mean"][:-1]), lambda d: d["scaler"].update(scale=[float("nan")] * NUM_COLOR_FEATURES), lambda d: d["features"].update(names=d["features"]["names"][:-1]), lambda d: d["features"].update(num_features=106), lambda d: d["features"].update(max_side=0), lambda d: d["features"]["ab_hist"].update(a_range=[50.0, 0.0]), lambda d: d["features"]["l_hist"].update(range=[20.0]), lambda d: d["features"]["lab"].update(srgb_to_xyz=[[1.0, 0.0], [0.0, 1.0]]), lambda d: d.update(classes=list(reversed(d["classes"]))), ], ids=[ "future_version", "string_version", "model_type", "no_scaler", "no_classes", "no_lab", "coef_rows", "intercept_len", "mean_len", "nan_scale", "names_len", "num_features", "max_side_0", "a_range_reversed", "l_range_len", "matrix_shape", "class_order", ], ) def test_load_rejects_invalid_documents(trained: dict, mutate) -> None: with pytest.raises(ValueError): deploy.load_color_model(_broken(trained["doc"], mutate)) def test_load_rejects_non_object_json(tmp_path: Path) -> None: path = tmp_path / "color_model.json" path.write_text("[1, 2, 3]", encoding="utf-8") with pytest.raises(ValueError): deploy.load_color_model(path)