Skip to main content

Model Training and Conversion

After completing dataset annotation, you can train a custom model directly on the Jetson and convert it for optimized deployment.

note

This tutorial focuses on CLI-based training and conversion. For Python-based workflows, refer to the official Ultralytics documentation.

1. Model Training​

Use the Ultralytics CLI to train a model.

1.1 Prepare Training Directory​

Copy the pretrained model (yolo11n.pt) into the directory containing your dataset configuration file, then open a terminal in that directory:

cd /home/jetson/ultralytics/ultralytics/data/yahboom_data/orange_data

1.2 Start Training​

yolo detect train data=orange.yaml model=yolo11n.pt epochs=100 imgsz=640

Parameter explanation:

  • data: Dataset configuration file\
  • model: Pretrained YOLO model\
  • epochs: Number of training epochs\
  • imgsz: Input image size

1.3 Training Output​

During training, logs and checkpoints are saved automatically.

Training Progress

Loss Curves

Metrics

2. Model Conversion​

After training, the best-performing model is saved in the runs directory.

2.1 Locate Trained Model​

/home/jetson/ultralytics/ultralytics/data/yahboom_data/orange_data/
└── runs/detect/train/weights/
    ├── best.pt
    └── last.pt

Use best.pt for deployment.

2.2 Convert PyTorch Model to TensorRT​

Navigate to the weights directory:

cd /home/jetson/ultralytics/ultralytics/data/yahboom_data/orange_data/runs/detect/train/weights

Run model export:

yolo export model=best.pt format=engine

The TensorRT engine file (.engine) will be generated in the same directory.

TensorRT Export

Engine Generation Complete

References​

Maintained by HemiHex for Jetson-based advanced vision workflows.