ref: update somes structures blabla and adjust training script with result saving

Readme.md

@@ -32,7 +32,7 @@ Total de arquivos em FASDD_UAV: 25097
 
 # Detalhes da estrutura pra treino CV e UAV
 ```
-data/FASDD_MERGED/
+data/FASDD/
 ├── images/
 │   ├── train/
 │   ├── val/

scripts/train.py

@@ -1,48 +1,52 @@
+
 from ultralytics import YOLO
+from datetime import datetime
+import os
+import shutil
 
-def train_yolo_model(data_path, model, epochs=10, batch_size=32, imgsz=320):
+def train_yolo_model(data_path, model='yolov9s.pt', epochs=10, batch_size=8, imgsz=320):
     """
-    Train YOLO model with optimized parameters for fastest training (testing purposes)
-    
-    Args:
-        data_path: Path to dataset YAML file
-        model: Model size ('yolov9n.pt' is fastest, 'yolov9s.pt' for better accuracy)
-        epochs: Number of training epochs (10-20 for quick tests)
-        batch_size: Batch size (16-64 depending on GPU memory)
-        imgsz: Image size (320 fastest, 640 standard, 1280 highest quality)
+    Train YOLOv9s model on Raspberry Pi 5 with light settings and CSV export.
     """
-    # Load a YOLO model
     model = YOLO(model)
 
-    # Train the model with optimized parameters for speed
     results = model.train(
-        data=data_path, 
+        data=data_path,
         epochs=epochs,
         batch=batch_size,
         imgsz=imgsz,
-        device='cpu',  # Change to 'cuda' or '0' if you have a GPU
-        workers=4,     # Number of dataloader workers
-        cache=True,    # Cache images for faster training
-        amp=True,      # Automatic Mixed Precision (faster on modern GPUs)
-        patience=5,    # Early stopping patience
-        save_period=5, # Save checkpoint every 5 epochs
-        plots=False,   # Disable plots to save time
+        device='cpu',
+        workers=2,
+        cache=True,     # PS: como True consome RAM, mas acelera. Desligar se travar
+        amp=False,
+        patience=5,
+        save_period=5,
+        plots=False,
         verbose=True
     )
-    
+
+    save_results_csv()
     return results
 
-if __name__ == "__main__":
-    data_path = "fasdd.yaml"  # Path to the dataset configuration file
+def save_results_csv():
+    now = datetime.now().strftime("%Y%m%d_%H%M%S")
+    source = "runs/detect/train/results.csv"
+    target_dir = "training_logs"
+    target = f"{target_dir}/yolov9s_{now}.csv"
+    os.makedirs(target_dir, exist_ok=True)
     
-    # For fastest training (testing purposes):
-    train_yolo_model(data_path, model='yolov9s.pt', epochs=5, batch_size=64, imgsz=320)
+    if os.path.exists(source):
+        shutil.copy(source, target)
+        print(f"✅ Training metrics saved to: {target}")
+    else:
+        print("⚠️ results.csv not found. Was training successful?")
 
-    # For balanced speed/accuracy (uncomment to use):
-    # train_yolo_model(data_path, epochs=20, batch_size=32, imgsz=480)
-
-    # For full training (uncomment to use):
-    # train_yolo_model(data_path, model='yolov9s.pt', epochs=100, batch_size=16, imgsz=640)
-    
-    # For balanced speed/accuracy (uncomment to use):
-    # train_yolo_model(data_path, epochs=20, batch_size=32, imgsz=480)
+if __name__ == "__main__":
+    data_path = "fasdd.yaml"
+    train_yolo_model(
+        data_path=data_path,
+        model='yolov9s.pt',
+        epochs=10,
+        batch_size=8, # Adjust if it crashes
+        imgsz=320    # Test 416 later
+    )