kmeans train and validation

README.md

@@ -77,8 +77,15 @@ python3 TopMotsParBin.py
 ```
 
 # Clustering methods
+## 1. K-Means clustering
+### • Script (inspired by [the github of the paper : SD-HuBERT: Sentence-Level Self-Distillation Induces Syllabic Organization in HuBERT](https://github.com/cheoljun95/sdhubert))
+**km.py :** Here is the training and validation with K-Means on différents k. 
 
-## 1. A help to chose the k number for K-Means clustering
+```bash
+python3 km.py   --train_file files/filesTrainWAV.txt   --val_file files/filesValWAV.txt   --train_dir /path/to/your/SegmentfeaturesTrain/   --val_dir /path/to/your/SegmentfeaturesVal/   --k_list 100,150,200   --save_models saved_models/
+```
+
+## 2. A help to chose the k number for K-Means clustering
 ### • Method with taking in considar of the distance of the last fusion
 **hierarchiqueSansDerniereDistance.py :** This method uses the average and the complete linkage. 
 

km.py

+import argparse
+import numpy as np
+from pathlib import Path
+from sklearn.cluster import MiniBatchKMeans
+from sklearn.metrics import silhouette_score, davies_bouldin_score
+import joblib
+from tqdm import tqdm
+
+def load_features(file_list, segment_dir):
+    features = []
+    for file in tqdm(file_list, desc=f"Chargement depuis {segment_dir}"):
+        segment_path = Path(segment_dir) / f"{Path(file).stem}_segmentfeature.npy"
+        if segment_path.exists():
+            data = np.load(segment_path)
+            if data.ndim == 1:
+                features.append(data)
+            else:
+                features.extend(data)
+        else:
+            print(f" Fichier manquant : {segment_path}")
+    return np.stack(features)
+
+def main(args):
+    with open(args.train_file, 'r') as f:
+        train_files = [line.strip() for line in f]
+    with open(args.val_file, 'r') as f:
+        val_files = [line.strip() for line in f]
+
+    print(" Chargement des features d'entraînement...")
+    X_train = load_features(train_files, args.train_dir)
+
+    print(" Chargement des features de validation...")
+    X_val = load_features(val_files, args.val_dir)
+
+    results = []
+    k_range = [int(k) for k in args.k_list.split(",")]
+
+    if args.save_models:
+        model_dir = Path(args.save_models)
+        model_dir.mkdir(parents=True, exist_ok=True)
+
+    for k in k_range:
+        print(f"\n Entraînement KMeans (k={k})...")
+        km = MiniBatchKMeans(n_clusters=k, batch_size=10000, max_iter=100, n_init=5, random_state=42)
+        km.fit(X_train)
+
+        print(" Prédiction sur Val...")
+        y_pred = km.predict(X_val)
+
+        print(" Évaluation...")
+        wcss = km.inertia_
+        silhouette = silhouette_score(X_val, y_pred)
+        db_index = davies_bouldin_score(X_val, y_pred)
+
+        print(f" k={k} | WCSS={wcss:.2f} | Silhouette={silhouette:.4f} | Davies-Bouldin={db_index:.4f}")
+        results.append((k, wcss, silhouette, db_index))
+
+        if args.save_models:
+            model_path = model_dir / f"kmeans_k{k}.pt"
+            joblib.dump(km, model_path)
+            print(f" Modèle sauvegardé dans {model_path}")
+
+    print("\n Résumé des résultats :")
+    print(f"{'k':>5} | {'WCSS':>12} | {'Silhouette':>10} | {'Davies-Bouldin':>15}")
+    for k, wcss, sil, db in results:
+        print(f"{k:5d} | {wcss:12.2f} | {sil:10.4f} | {db:15.4f}")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--train_file", type=str, required=True, help="Fichier texte avec la liste des fichiers d'entraînement (une ligne par fichier .wav)")
+    parser.add_argument("--val_file", type=str, required=True, help="Fichier texte avec la liste des fichiers de validation (une ligne par fichier .wav)")
+    parser.add_argument("--train_dir", type=str, required=True, help="Répertoire contenant les .npy d'entraînement")
+    parser.add_argument("--val_dir", type=str, required=True, help="Répertoire contenant les .npy de validation")
+    parser.add_argument("--k_list", type=str, required=True, help="Liste des k à tester, ex: 5,10,20")
+    parser.add_argument("--save_models", type=str, default=None, help="Répertoire pour sauvegarder les modèles KMeans")
+    args = parser.parse_args()
+    main(args)