Ultralytics MCP Server

""" Convert existing YOLO training CSV results to TensorBoard format This script creates TensorBoard event files from results.csv for visualization """ import pandas as pd from pathlib import Path from torch.utils.tensorboard import SummaryWriter import yaml import sys def get_model_name(training_dir): """Extract a short, meaningful model name from args.yaml""" args_file = training_dir / "args.yaml" if args_file.exists(): try: with open(args_file, 'r') as f: args = yaml.safe_load(f) # Get model variant (yolo11n, yolo11s, etc.) model_variant = "yolo11" if 'model' in args: model_path = str(args['model']) # Extract just the model name without .pt model_name = Path(model_path).stem if model_name.startswith('yolo'): model_variant = model_name # Get dataset name (simplified) dataset_name = "unknown" if 'data' in args: data_path = Path(str(args['data'])) # Get parent directory name as dataset parent_dir = data_path.parent.name if parent_dir and parent_dir not in ['custom_datasets', 'YOLO_MultiLevel_Datasets']: # Shorten long dataset names if 'kaggle' in parent_dir.lower(): dataset_name = "kaggle" elif 'panoramic' in parent_dir.lower(): dataset_name = "panoramic" elif 'dental' in parent_dir.lower(): dataset_name = "dental" else: # Take first word or first 10 chars words = parent_dir.split() dataset_name = words[0] if words else parent_dir[:10] elif parent_dir == 'YOLO_MultiLevel_Datasets': # For built-in datasets, use the folder name before data.yaml folder_name = data_path.parent.parent.name if data_path.parent.parent.exists() else "builtin" dataset_name = folder_name.replace('YOLO_', '').replace('_', '') # Get epochs epochs = args.get('epochs', '?') # Create compact name: model_dataset_epochs return f"{model_variant}_{dataset_name}_{epochs}ep" except Exception as e: print(f" Warning: Could not parse args.yaml: {e}") pass # Fallback to training directory name return training_dir.name def convert_csv_to_tensorboard(training_dir, tensorboard_root): """Convert a single training's CSV results to TensorBoard format""" results_csv = training_dir / "results.csv" if not results_csv.exists(): print(f"⚠️ No results.csv found in {training_dir.name}") return False try: # Get model name model_name = get_model_name(training_dir) # Read CSV results df = pd.read_csv(results_csv) df.columns = df.columns.str.strip() # Remove whitespace from column names # Create TensorBoard writer with model name as the run name # This will show as "model_name" in TensorBoard UI log_dir = tensorboard_root / model_name writer = SummaryWriter(log_dir=str(log_dir)) print(f"📊 Converting {training_dir.name} → {model_name}...") # Add metadata writer.add_text("Training/Directory", training_dir.name, 0) writer.add_text("Training/Epochs", str(len(df)), 0) # Map CSV columns to TensorBoard metrics metric_mapping = { 'train/box_loss': 'train/box_loss', 'train/cls_loss': 'train/cls_loss', 'train/dfl_loss': 'train/dfl_loss', 'metrics/precision(B)': 'metrics/precision', 'metrics/recall(B)': 'metrics/recall', 'metrics/mAP50(B)': 'metrics/mAP50', 'metrics/mAP50-95(B)': 'metrics/mAP50-95', 'val/box_loss': 'val/box_loss', 'val/cls_loss': 'val/cls_loss', 'val/dfl_loss': 'val/dfl_loss', 'lr/pg0': 'learning_rate/pg0', 'lr/pg1': 'learning_rate/pg1', 'lr/pg2': 'learning_rate/pg2' } # Write metrics to TensorBoard for idx, row in df.iterrows(): epoch = int(row['epoch']) if 'epoch' in df.columns else idx for csv_col, tb_name in metric_mapping.items(): if csv_col in df.columns: value = row[csv_col] if pd.notna(value): # Only write non-NaN values writer.add_scalar(tb_name, float(value), epoch) writer.close() print(f"✅ Successfully converted {training_dir.name} → {model_name} ({len(df)} epochs)") return True except Exception as e: print(f"❌ Error converting {training_dir.name}: {str(e)}") return False def main(): # Find all training directories runs_dir = Path("ultralytics/runs/detect") if not runs_dir.exists(): print(f"❌ Directory not found: {runs_dir}") return # Create central TensorBoard logs directory tensorboard_root = runs_dir / "tensorboard_logs" tensorboard_root.mkdir(exist_ok=True) training_dirs = [d for d in runs_dir.iterdir() if d.is_dir() and d.name != "tensorboard_logs"] if not training_dirs: print("❌ No training directories found!") return print(f"\n🔄 Found {len(training_dirs)} training directories") print("=" * 60) success_count = 0 total_count = len(training_dirs) for training_dir in sorted(training_dirs): if convert_csv_to_tensorboard(training_dir, tensorboard_root): success_count += 1 print("=" * 60) print(f"\n✨ Conversion complete: {success_count}/{total_count} successful") print(f"\n🚀 Now run: tensorboard --logdir=ultralytics/runs/detect/tensorboard_logs") print(f"📊 Then open: http://localhost:6006") print(f"\n💡 Each training will appear with its model name in TensorBoard!") if __name__ == "__main__": main()