scripts/train_sota_v2.py

import os
import sys
import yaml
import argparse
import pandas as pd
import numpy as np
from sklearn.model_selection import StratifiedKFold
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import accuracy_score, f1_score, roc_auc_score, precision_score, recall_score, classification_report
import xgboost as xgb
import joblib
import shap

sys.path.append(os.path.dirname(os.path.abspath(__file__)))
from models_v2 import SOTAHybridDetector

# Les 30 features stylométriques SOTA
STYLOMETRIC_COLS_V2 = [
    'num_chars', 'num_words', 'num_sentences', 'avg_sentence_len', 'std_sentence_len',
    'slv_normalized', 'avg_word_len', 'ratio_long_words',
    'vocabulary_diversity', 'hapax_ratio', 'yules_k', 'maas_index',
    'information_entropy', 'brunet_w',
    'ratio_punctuation', 'freq_uppercase', 'freq_digits',
    'connector_ratio', 'connector_diversity', 'repetition_ratio',
    'stopword_ratio', 'mean_polarity_diff',
    'syntactic_complexity_score', 'ratio_interrogative', 'ratio_exclamative',
    'ratio_declarative', 'imparfait_ratio', 'futur_ratio',
    'conditional_ratio', 'passive_voice_ratio'
]

def load_config(config_path):
    with open(config_path, "r", encoding="utf-8") as f:
        return yaml.safe_load(f)

def main():
    parser = argparse.ArgumentParser(description="Entraînement du détecteur SOTA hybride (Stylométrie + CamemBERT + XGBoost).")
    parser.add_argument("--config", default="configs/config.yaml", help="Chemin vers le fichier de config")
    args = parser.parse_args()
    
    config = load_config(args.config)
    processed_dir = config["paths"]["processed_dir"]
    models_dir = config["paths"]["models_dir"]
    reports_dir = config["paths"]["reports_dir"]
    os.makedirs(models_dir, exist_ok=True)
    os.makedirs(reports_dir, exist_ok=True)
    
    # ========== 1. Chargement des données ==========
    print("=" * 60)
    print("ENTRAÎNEMENT DU DÉTECTEUR SOTA HYBRIDE v2")
    print("Architecture: Stylométrie (30) + CamemBERT (768) → XGBoost")
    print("=" * 60)
    
    # Stylometric features
    train_sty_path = os.path.join(processed_dir, "train_features_v2.csv")
    if not os.path.exists(train_sty_path):
        print(f"ERREUR: {train_sty_path} introuvable. Lancez build_features_v2.py d'abord.")
        sys.exit(1)
    df_sty = pd.read_csv(train_sty_path)
    
    # CamemBERT embeddings
    train_emb_path = os.path.join(processed_dir, "train_embeddings_camembert.csv")
    if not os.path.exists(train_emb_path):
        print(f"ERREUR: {train_emb_path} introuvable. Lancez camembert_encoder.py d'abord.")
        sys.exit(1)
    df_emb = pd.read_csv(train_emb_path)
    
    print(f"Features stylométriques: {df_sty.shape}")
    print(f"Embeddings CamemBERT: {df_emb.shape}")
    
    # Extract arrays
    X_sty = df_sty[STYLOMETRIC_COLS_V2].values
    emb_cols = [c for c in df_emb.columns if c.startswith("camembert_")]
    X_emb = df_emb[emb_cols].values
    y = df_sty["label_human_ai"].values
    
    print(f"\nDimensions: Stylométrie={X_sty.shape[1]}, CamemBERT={X_emb.shape[1]}, Total={X_sty.shape[1]+X_emb.shape[1]}")
    print(f"Classes: Humain={np.sum(y==0)}, IA={np.sum(y==1)}")
    
    # ========== 2. Scaling ==========
    print("\nNormalisation des features...")
    scaler_sty = StandardScaler()
    X_sty_scaled = scaler_sty.fit_transform(X_sty)
    
    scaler_emb = StandardScaler()
    X_emb_scaled = scaler_emb.fit_transform(X_emb)
    
    # Concatenate
    X_combined = np.hstack([X_sty_scaled, X_emb_scaled])
    print(f"Vecteur combiné final: {X_combined.shape}")
    
    # ========== 3. Cross-Validation 5-Fold ==========
    print("\n" + "=" * 60)
    print("VALIDATION CROISÉE STRATIFIÉE (5-Fold)")
    print("=" * 60)
    
    cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
    oof_preds = np.zeros(len(y))
    oof_probs = np.zeros(len(y))
    fold_metrics = []
    
    for fold, (train_idx, val_idx) in enumerate(cv.split(X_combined, y)):
        print(f"\n--- Fold {fold+1}/5 ---")
        X_train, X_val = X_combined[train_idx], X_combined[val_idx]
        y_train, y_val = y[train_idx], y[val_idx]
        
        xgb_model = xgb.XGBClassifier(
            n_estimators=300,
            learning_rate=0.05,
            max_depth=8,
            min_child_weight=3,
            subsample=0.8,
            colsample_bytree=0.8,
            gamma=0.1,
            reg_alpha=0.1,
            reg_lambda=1.0,
            random_state=42,
            eval_metric="logloss",
            tree_method="hist"
        )
        xgb_model.fit(
            X_train, y_train,
            eval_set=[(X_val, y_val)],
            verbose=False
        )
        
        fold_probs = xgb_model.predict_proba(X_val)[:, 1]
        fold_preds = (fold_probs >= 0.5).astype(int)
        oof_probs[val_idx] = fold_probs
        oof_preds[val_idx] = fold_preds
        
        fold_acc = accuracy_score(y_val, fold_preds)
        fold_f1 = f1_score(y_val, fold_preds)
        fold_auc = roc_auc_score(y_val, fold_probs)
        fold_metrics.append({"fold": fold+1, "accuracy": fold_acc, "f1": fold_f1, "auc": fold_auc})
        print(f"  Accuracy: {fold_acc:.4f} | F1: {fold_f1:.4f} | AUC: {fold_auc:.4f}")
    
    # OOF metrics
    oof_acc = accuracy_score(y, oof_preds)
    oof_f1 = f1_score(y, oof_preds)
    oof_auc = roc_auc_score(y, oof_probs)
    oof_prec = precision_score(y, oof_preds)
    oof_rec = recall_score(y, oof_preds)
    
    print(f"\n{'=' * 60}")
    print(f"RÉSULTATS OUT-OF-FOLD (Validation Croisée Complète)")
    print(f"{'=' * 60}")
    print(f"Accuracy: {oof_acc:.4f}")
    print(f"Precision: {oof_prec:.4f}")
    print(f"Recall: {oof_rec:.4f}")
    print(f"F1-Score: {oof_f1:.4f}")
    print(f"ROC-AUC: {oof_auc:.4f}")
    
    # ========== 4. Train Final Model on Full Data ==========
    print(f"\nEntraînement du modèle final sur l'intégralité des données...")
    xgb_final = xgb.XGBClassifier(
        n_estimators=300,
        learning_rate=0.05,
        max_depth=8,
        min_child_weight=3,
        subsample=0.8,
        colsample_bytree=0.8,
        gamma=0.1,
        reg_alpha=0.1,
        reg_lambda=1.0,
        random_state=42,
        eval_metric="logloss",
        tree_method="hist"
    )
    xgb_final.fit(X_combined, y, verbose=False)
    
    # ========== 5. SHAP Explainability ==========
    print("\nCalcul des valeurs SHAP (TreeExplainer)...")
    all_feature_names = STYLOMETRIC_COLS_V2 + emb_cols
    
    explainer = shap.TreeExplainer(xgb_final)
    # Compute SHAP on a subsample for speed
    sample_size = min(500, len(X_combined))
    np.random.seed(42)
    sample_idx = np.random.choice(len(X_combined), sample_size, replace=False)
    shap_values = explainer.shap_values(X_combined[sample_idx])
    
    # Get top 10 most important features by mean |SHAP|
    mean_shap = np.abs(shap_values).mean(axis=0)
    top_indices = np.argsort(mean_shap)[::-1][:20]
    print("\nTop 20 features par importance SHAP moyenne:")
    for rank, idx in enumerate(top_indices):
        fname = all_feature_names[idx] if idx < len(all_feature_names) else f"feature_{idx}"
        print(f"  {rank+1:2d}. {fname:40s} SHAP moyen: {mean_shap[idx]:.4f}")
    
    # ========== 6. Save Model Package ==========
    # Get n-gram column info
    ngram_cols = [c for c in df_sty.columns if c.startswith("ngram_word_") or c.startswith("ngram_char_")]
    
    # Friendly feature names for stylometric cols
    sty_friendly = {
        'num_chars': 'Nombre de caractères',
        'num_words': 'Nombre de mots',
        'num_sentences': 'Nombre de phrases',
        'avg_sentence_len': 'Longueur moyenne des phrases',
        'std_sentence_len': 'Écart-type longueur des phrases',
        'slv_normalized': 'Variance normalisée des phrases (SLV)',
        'avg_word_len': 'Longueur moyenne des mots',
        'ratio_long_words': 'Ratio de mots longs (>6 chars)',
        'vocabulary_diversity': 'Diversité lexicale (TTR)',
        'hapax_ratio': "Ratio d'Hapax (mots uniques)",
        'yules_k': 'K de Yule (richesse vocabulaire)',
        'maas_index': 'Indice de Maas (diversité log)',
        'information_entropy': 'Entropie informationnelle (burstiness)',
        'brunet_w': 'W de Brunet (richesse)',
        'ratio_punctuation': 'Ratio de ponctuation',
        'freq_uppercase': 'Fréquence des majuscules',
        'freq_digits': 'Fréquence des chiffres',
        'connector_ratio': 'Ratio de connecteurs logiques',
        'connector_diversity': 'Diversité des connecteurs',
        'repetition_ratio': 'Ratio de répétitions lexicales',
        'stopword_ratio': 'Ratio de mots vides',
        'mean_polarity_diff': 'Variation de polarité inter-phrases',
        'syntactic_complexity_score': 'Complexité syntaxique',
        'ratio_interrogative': 'Ratio phrases interrogatives',
        'ratio_exclamative': 'Ratio phrases exclamatives',
        'ratio_declarative': 'Ratio phrases déclaratives',
        'imparfait_ratio': "Ratio verbes à l'imparfait",
        'futur_ratio': 'Ratio verbes au futur',
        'conditional_ratio': 'Ratio verbes au conditionnel',
        'passive_voice_ratio': 'Ratio de voix passive'
    }
    
    friendly_names_sty = [sty_friendly.get(c, c) for c in STYLOMETRIC_COLS_V2]
    friendly_names_emb = [f"CamemBERT dim {i}" for i in range(len(emb_cols))]
    
    detector = SOTAHybridDetector(
        xgb_meta=xgb_final,
        scaler_sty=scaler_sty,
        scaler_emb=scaler_emb,
        num_sty_features=len(STYLOMETRIC_COLS_V2),
        num_emb_features=len(emb_cols),
        feature_names_sty=friendly_names_sty,
        feature_names_emb=friendly_names_emb
    )
    
    package = {
        "model_name": "SOTA Hybrid Detector v2 (Stylométrie + CamemBERT + XGBoost)",
        "model_key": "hybrid_v2",
        "model": detector,
        "xgb_raw": xgb_final,
        "stylometric_cols": STYLOMETRIC_COLS_V2,
        "emb_cols": emb_cols,
        "ngram_cols": ngram_cols,
        "scalers": {"sty": scaler_sty, "emb": scaler_emb},
        "shap_explainer": explainer,
        "shap_mean_abs": mean_shap,
        "feature_names_all": all_feature_names,
        "friendly_names_sty": friendly_names_sty,
        "friendly_names_emb": friendly_names_emb,
        "vectorizer_words_path": os.path.join(models_dir, "word_vectorizer_v2.pkl"),
        "vectorizer_chars_path": os.path.join(models_dir, "char_vectorizer_v2.pkl"),
        "oof_metrics": {
            "accuracy": oof_acc, "precision": oof_prec, "recall": oof_rec,
            "f1": oof_f1, "auc": oof_auc
        },
        "fold_metrics": fold_metrics,
        "top_shap_features": [(all_feature_names[i], float(mean_shap[i])) for i in top_indices]
    }
    
    model_path = os.path.join(models_dir, "best_detector_v2.pkl")
    joblib.dump(package, model_path)
    print(f"\n🎉 Modèle SOTA v2 sauvegardé dans {model_path}")
    
    # ========== 7. Write Evaluation Report ==========
    report_path = os.path.join(reports_dir, "evaluation_report_v2.md")
    with open(report_path, "w", encoding="utf-8") as f:
        f.write(f"""# Rapport d'Évaluation — Détecteur SOTA Hybride v2\n\n""")
        f.write(f"**Date d'entraînement** : Juin 2026\n")
        f.write(f"**Architecture** : Stylométrie (30 features) + CamemBERT gelé (768 dims) → XGBoost\n\n")
        f.write(f"## Résultats de Validation Croisée (5-Fold OOF)\n\n")
        f.write(f"| Métrique | Score |\n|---|---|\n")
        f.write(f"| Accuracy | {oof_acc:.4f} |\n")
        f.write(f"| Precision | {oof_prec:.4f} |\n")
        f.write(f"| Recall | {oof_rec:.4f} |\n")
        f.write(f"| F1-Score | {oof_f1:.4f} |\n")
        f.write(f"| ROC-AUC | {oof_auc:.4f} |\n\n")
        f.write(f"## Résultats par Fold\n\n")
        f.write(f"| Fold | Accuracy | F1 | AUC |\n|---|---|---|---|\n")
        for m in fold_metrics:
            f.write(f"| {m['fold']} | {m['accuracy']:.4f} | {m['f1']:.4f} | {m['auc']:.4f} |\n")
        f.write(f"\n## Top 20 Features SHAP\n\n")
        f.write(f"| Rang | Feature | SHAP moyen |\n|---|---|---|\n")
        for rank, idx in enumerate(top_indices):
            fname = all_feature_names[idx] if idx < len(all_feature_names) else f"feature_{idx}"
            f.write(f"| {rank+1} | {fname} | {mean_shap[idx]:.4f} |\n")
        f.write(f"\n## Description de l'Architecture\n\n")
        f.write(f"""### Phase 1 : Module Stylométrique (30 features invariantes)
Extraction de 30 caractéristiques linguistiques prouvées résistantes aux attaques de paraphrase :\n- **Structure** : longueur/variance des phrases, SLV normalisée\n- **Richesse lexicale** : K de Yule, Indice de Maas, W de Brunet, Entropie\n- **Discours** : connecteurs, polarité, répétitions\n- **Syntaxe** : complexité, temps verbaux, voix passive\n\n### Phase 2 : Module Neural (CamemBERT gelé)
Embeddings denses de 768 dimensions extraits du token [CLS] de `almanach/camembert-base`.\nPoids entièrement gelés — aucun fine-tuning pour préserver la généralisation.\n\n### Phase 3 : Méta-Classifieur XGBoost
XGBoost Classifier entraîné sur le vecteur concaténé de 798 dimensions.\nHyperparamètres optimisés : 300 estimateurs, lr=0.05, depth=8.\n\n### Phase 4 : Explainabilité SHAP (TreeExplainer)
Valeurs SHAP calculées par TreeExplainer pour décomposer chaque prédiction\nen contributions individuelles de chaque feature.\n""")
    
    print(f"📊 Rapport d'évaluation sauvegardé dans {report_path}")
    print(f"\n✅ Pipeline d'entraînement SOTA v2 terminé avec succès.")

if __name__ == "__main__":
    main()