defect-analysis/defect_analysis/ml/model_bundle.py

"""可持久化的 ML 模型包。"""

import warnings
from datetime import datetime, timezone

import joblib
import pandas as pd
from sklearn.model_selection import train_test_split

from defect_analysis.ml.datasets import build_supervised_dataset
from defect_analysis.ml.features import build_feature_frame
from defect_analysis.ml.model_registry import detect_optional_model_backends
from defect_analysis.ml.tabular_models import classification_metrics, extract_feature_importance, train_tabular_model
from defect_analysis.schemas import normalize_defect_schema


MODEL_BUNDLE_VERSION = 1


def _target_config(target_defect_type=None, target_severity=None):
    return {
        "defect_type": target_defect_type,
        "severity": target_severity,
        "default": target_defect_type is None and target_severity is None,
    }


def _align_features(features, feature_columns):
    """按训练时特征签名对齐新数据，避免 one-hot 列漂移导致推理失败。"""
    aligned = features.reindex(columns=feature_columns, fill_value=0.0)
    return aligned.astype(float)


def create_model_bundle(
    df,
    *,
    model_name="random_forest",
    target_defect_type=None,
    target_severity=None,
    random_state=42,
    test_size=0.25,
):
    """训练并创建可保存的模型包。"""
    normalized = normalize_defect_schema(df)
    X, y = build_supervised_dataset(
        normalized,
        target_defect_type=target_defect_type,
        target_severity=target_severity,
    )
    if y.nunique() < 2:
        raise ValueError("目标标签只有一个类别，无法训练监督模型")

    min_count = int(y.value_counts().min())
    if min_count < 2:
        warnings.warn(
            f"最小类别仅 {min_count} 个样本，已关闭分层抽样。验证集可能不包含少数类别。",
            UserWarning,
        )
    stratify = y if min_count >= 2 else None
    X_train, X_valid, y_train, y_valid = train_test_split(
        X,
        y,
        test_size=test_size,
        random_state=random_state,
        stratify=stratify,
    )
    validation_model = train_tabular_model(model_name, X_train, y_train, random_state=random_state)["model"]
    validation_metrics = classification_metrics(validation_model, X_valid, y_valid, prefix="validation")

    trained = train_tabular_model(model_name, X, y, random_state=random_state)
    feature_importance = extract_feature_importance(trained["model"], X.columns)
    return {
        "bundle_version": MODEL_BUNDLE_VERSION,
        "created_at": datetime.now(timezone.utc).isoformat(),
        "model_name": model_name,
        "target": _target_config(target_defect_type, target_severity),
        "feature_columns": list(X.columns),
        "metrics": trained.get("metrics", {}),
        "validation_metrics": validation_metrics,
        "feature_importance": feature_importance,
        "optional_backends": detect_optional_model_backends(),
        "model": trained["model"],
    }


def save_model_bundle(bundle, path):
    """保存模型包。"""
    joblib.dump(bundle, path)
    return path


def load_model_bundle(path):
    """加载模型包。"""
    bundle = joblib.load(path)
    if bundle.get("bundle_version") != MODEL_BUNDLE_VERSION:
        raise ValueError("模型包版本不兼容")
    return bundle


def predict_with_bundle(bundle, df):
    """使用模型包对新数据打分。"""
    normalized = normalize_defect_schema(df).reset_index(drop=True)
    features = build_feature_frame(normalized)
    X = _align_features(features, bundle["feature_columns"])
    model = bundle["model"]

    scored = normalized.copy()
    scored["ml_prediction"] = model.predict(X)
    if hasattr(model, "predict_proba"):
        scored["ml_probability"] = model.predict_proba(X)[:, 1]
    else:
        scored["ml_probability"] = pd.NA
    scored["model_name"] = bundle["model_name"]
    return scored