model_training.py

# ...existing code...
import joblib, json
from pathlib import Path
from xgboost import XGBClassifier
from lightgbm import LGBMClassifier
# ...existing code...
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.calibration import CalibratedClassifierCV

try:
    from ingest import metrics_report  # type: ignore
except Exception:
    try:
        from preprocess import metrics_report  # type: ignore
    except Exception:
        # Minimal fallback implementation returning a dict compatible with existing summary DataFrame usage
        from sklearn.metrics import log_loss, accuracy_score, brier_score_loss, roc_auc_score

        def metrics_report(y_true, y_proba, name="model"):
            """

            Minimal metrics report compatible with the rest of the script.

            Returns a dict with at least the key "model" so it can be placed into the summary DataFrame.

            """
            y_true = np.asarray(y_true)
            proba = np.asarray(y_proba)

            # handle binary vs multiclass probabilities
            try:
                if proba.ndim == 1 or proba.shape[1] == 1:
                    # binary
                    proba_flat = proba.ravel()
                    y_pred = (proba_flat >= 0.5).astype(int)
                    ll = float(log_loss(y_true, proba_flat)) if y_true.size and proba_flat.size else float("nan")
                    try:
                        auc = float(roc_auc_score(y_true, proba_flat))
                    except Exception:
                        auc = float("nan")
                    try:
                        brier = float(brier_score_loss(y_true, proba_flat))
                    except Exception:
                        brier = float("nan")
                else:
                    # multiclass
                    y_pred = proba.argmax(axis=1)
                    try:
                        ll = float(log_loss(y_true, proba))
                    except Exception:
                        ll = float("nan")
                    try:
                        auc = float(roc_auc_score(y_true, proba, multi_class="ovr"))
                    except Exception:
                        auc = float("nan")
                    # multiclass Brier as mean squared error against one-hot
                    try:
                        n_classes = proba.shape[1]
                        one_hot = np.eye(n_classes)[y_true]
                        brier = float(np.mean(np.sum((proba - one_hot) ** 2, axis=1)))
                    except Exception:
                        brier = float("nan")
            except Exception:
                y_pred = np.zeros_like(y_true)
                ll = auc = brier = float("nan")

            acc = float(accuracy_score(y_true, y_pred)) if y_true.size else float("nan")

            return {
                "model": name,
                "accuracy": acc,
                "log_loss": ll,
                "roc_auc": auc,
                "brier": brier
            }

def load_processed_data(target_col="target", data_dir=Path("data/processed")):
    """

    Try multiple ways to obtain a processed dataframe:

    1) call a load_processed_data / load_data function from local ingest or preprocess modules

    2) look for common filenames under data/processed (parquet/csv)

    Returns: df (pd.DataFrame)

    """
    # 1) try local modules
    try:
        from ingest import load_processed_data as _lp  # type: ignore
        df = _lp()
        if isinstance(df, pd.DataFrame):
            return df
    except Exception:
        pass

    try:
        from preprocess import load_processed_data as _lp2  # type: ignore
        df = _lp2()
        if isinstance(df, pd.DataFrame):
            return df
    except Exception:
        pass

    # 2) look for common files
    candidates = [
        data_dir / "processed.parquet",
        data_dir / "dataset.parquet",
        data_dir / "processed.csv",
        data_dir / "dataset.csv",
        data_dir / "train.parquet",
        data_dir / "train.csv",
    ]
    for fp in candidates:
        if fp.exists():
            if fp.suffix == ".parquet":
                return pd.read_parquet(fp)
            else:
                return pd.read_csv(fp)

    raise FileNotFoundError(
        f"No processed data found. Checked modules ingest/preprocess and files under {data_dir}. "
        "Add a processed dataset or expose load_processed_data() in ingest/preprocess."
    )

# Load data and build train/valid/test splits
df = load_processed_data()

# infer target column
TARGET = None
for candidate in ("target", "label", "y", "outcome"):
    if candidate in df.columns:
        TARGET = candidate
        break
if TARGET is None:
    raise KeyError("No target column found. Expected one of: target,label,y,outcome")

# If dataset already includes a 'split' column with values 'train'/'valid'/'test', use it
if "split" in df.columns:
    train_df = df[df["split"] == "train"].drop(columns=["split"])
    valid_df = df[df["split"] == "valid"].drop(columns=["split"])
    test_df  = df[df["split"] == "test"].drop(columns=["split"])
else:
    # create splits: train/valid/test = 64%/16%/20% (approx)
    train_val, test_df = train_test_split(df, test_size=0.20, stratify=df[TARGET], random_state=42)
    train_df, valid_df = train_test_split(train_val, test_size=0.20, stratify=train_val[TARGET], random_state=42)

X_cols = [c for c in df.columns if c != TARGET]
WINDOW = int(df.attrs.get("WINDOW", 1)) if hasattr(df, "attrs") else 1

X_train = train_df[X_cols]
y_train = train_df[TARGET]
X_valid = valid_df[X_cols]
y_valid = valid_df[TARGET]
X_test  = test_df[X_cols]
y_test  = test_df[TARGET]

# ...existing code...
xgb = XGBClassifier(
    n_estimators=6000,
    max_depth=50,
    learning_rate=0.05,
    subsample=0.9,
    colsample_bytree=0.9,
    objective="multi:softprob",
    num_class=3,
    reg_lambda=1.0,
    random_state=42,
    tree_method="hist"
)


xgb = XGBClassifier(
    n_estimators=6000,
    max_depth=50,
    learning_rate=0.05,
    subsample=0.9,
    colsample_bytree=0.9,
    objective="multi:softprob",
    num_class=3,
    reg_lambda=1.0,
    random_state=42,
    tree_method="hist"
)
xgb.fit(X_train, y_train)
proba_xgb_valid = xgb.predict_proba(X_valid)
proba_xgb_test  = xgb.predict_proba(X_test)

m_xgb_valid = metrics_report(y_valid, proba_xgb_valid, "xgb_valid")
m_xgb_test  = metrics_report(y_test,  proba_xgb_test,  "xgb_test")

# Isotonic calibration
cal_xgb = CalibratedClassifierCV(xgb, method="isotonic", cv="prefit")
cal_xgb.fit(X_valid, y_valid)
proba_xgb_cal_valid = cal_xgb.predict_proba(X_valid)
proba_xgb_cal_test  = cal_xgb.predict_proba(X_test)

m_xgb_cal_valid = metrics_report(y_valid, proba_xgb_cal_valid, "xgb_cal_valid")
m_xgb_cal_test  = metrics_report(y_test,  proba_xgb_cal_test,  "xgb_cal_test")

# Platt (optional)
cal_xgb_platt = CalibratedClassifierCV(xgb, method="sigmoid", cv="prefit")
cal_xgb_platt.fit(X_valid, y_valid)
proba_xgb_platt_test = cal_xgb_platt.predict_proba(X_test)
m_xgb_platt_test = metrics_report(y_test, proba_xgb_platt_test, "xgb_platt_test")

# -----------------------------
# 8) LightGBM + Calibration
# -----------------------------
lgbm = LGBMClassifier(
    n_estimators=12000,
    learning_rate=0.03,
    num_leaves=63,
    subsample=0.9,
    colsample_bytree=0.9,
    objective="multiclass",
    class_weight=None,
    random_state=42
)
lgbm.fit(X_train, y_train)
proba_lgb_valid = lgbm.predict_proba(X_valid)
proba_lgb_test  = lgbm.predict_proba(X_test)

m_lgb_valid = metrics_report(y_valid, proba_lgb_valid, "lgb_valid")
m_lgb_test  = metrics_report(y_test,  proba_lgb_test,  "lgb_test")

cal_lgb = CalibratedClassifierCV(lgbm, method="isotonic", cv="prefit")
cal_lgb.fit(X_valid, y_valid)
proba_lgb_cal_valid = cal_lgb.predict_proba(X_valid)
proba_lgb_cal_test  = cal_lgb.predict_proba(X_test)

m_lgb_cal_valid = metrics_report(y_valid, proba_lgb_cal_valid, "lgb_cal_valid")
m_lgb_cal_test  = metrics_report(y_test,  proba_lgb_cal_test,  "lgb_cal_test")

from sklearn.base import BaseEstimator, ClassifierMixin

class PriorProbaPredictor(BaseEstimator, ClassifierMixin):
    """Predict class probabilities equal to the class distribution in training data."""
    def fit(self, X, y):
        y = np.asarray(y)
        classes, counts = np.unique(y, return_counts=True)
        self.classes_ = classes
        self.class_proba_ = counts / counts.sum()
        return self

    def predict_proba(self, X):
        n = len(X)
        # return array shape (n_samples, n_classes) following classes_ order
        return np.tile(self.class_proba_, (n, 1))

    def predict(self, X):
        proba = self.predict_proba(X)
        return proba.argmax(axis=1)

odds = PriorProbaPredictor()
odds.fit(X_train, y_train)
proba_odds_valid = odds.predict_proba(X_valid)
proba_odds_test  = odds.predict_proba(X_test)

m_odds_valid = metrics_report(y_valid, proba_odds_valid, "odds_valid")
m_odds_test  = metrics_report(y_test,  proba_odds_test,  "odds_test")
# ...existing code...
# -----------------------------
# 9) Summary table
# -----------------------------
summary = pd.DataFrame([
    m_odds_valid, m_odds_test,
    m_xgb_valid, m_xgb_test, m_xgb_cal_valid, m_xgb_cal_test, m_xgb_platt_test,
    m_lgb_valid, m_lgb_test, m_lgb_cal_valid, m_lgb_cal_test
]).sort_values("model").reset_index(drop=True)

try:
    from IPython.display import display as _display
    _display(summary)
except Exception:
    print(summary.to_string(index=False))


# Optional: save metrics
summary.to_csv("./evaluation/baseline_metrics.csv", index=False)
print("Saved: baseline_metrics.csv")

Path(".").mkdir(exist_ok=True)

# เลือกโมเดลที่ต้องการใช้ inference (แนะนำตัวที่ calibrated แล้ว)
joblib.dump(cal_xgb, "./model/model_xgb_isotonic.joblib")
joblib.dump(cal_lgb, "./model/model_lgb_isotonic.joblib")

# เก็บคอลัมน์ฟีเจอร์ และพารามิเตอร์สำคัญ
with open("feature_columns.json", "w", encoding="utf-8") as f:
    json.dump({"X_cols": X_cols, "WINDOW": int(WINDOW)}, f, ensure_ascii=False, indent=2)

print("Saved: model_xgb_isotonic.joblib, model_lgb_isotonic.joblib, feature_columns.json")