Skip to main content

Model Conversion on Jetson

This section explains how to convert YOLO models on NVIDIA Jetson devices for optimal inference performance using TensorRT.

The standard conversion pipeline is:

PyTorch (.pt) → ONNX (.onnx) → TensorRT (.engine)

1. Enable Optimal Jetson Performance​

Before model conversion or inference, configure Jetson for maximum performance.

Enable MAX Power Mode​

sudo nvpmodel -m 2

Enable Jetson Clocks​

sudo jetson_clocks

2. Model Conversion Overview​

TensorRT provides the fastest inference on Jetson. Ultralytics YOLO supports direct export to TensorRT, automatically generating an intermediate ONNX model.

3. CLI Model Conversion​

Navigate to the Ultralytics directory:

cd ~/ultralytics/ultralytics

Export YOLO models to TensorRT:

yolo export model=yolo11n.pt format=engine
# yolo export model=yolo11n-seg.pt format=engine
# yolo export model=yolo11n-pose.pt format=engine
# yolo export model=yolo11n-cls.pt format=engine
# yolo export model=yolo11n-obb.pt format=engine

The generated .engine file will be saved alongside the original model.

4. Python Model Conversion​

Navigate to the demo directory:

cd ~/ultralytics/ultralytics/yahboom_demo

Run the conversion script:

python3 model_pt_onnx_engine.py

Example Python Code​

from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.export(format="engine")

5. Model Inference​

USB Camera Inference (CLI)​

yolo predict model=yolo11n.engine source=0 show save=False

ONNX Inference​

yolo predict model=yolo11n.onnx source=0 show save=False
note

CLI inference supports USB cameras. For CSI cameras, use Python-based inference.

6. Common Issues​

onnxslim Error​

If you encounter an onnxslim error:

sudo pip3 install onnxslim

Then re-run the export command.

7. Summary​

  • Always enable MAX power mode and Jetson clocks
  • TensorRT delivers the best inference performance
  • Ultralytics simplifies conversion using CLI and Python APIs
  • Use .engine models for production deployment

Maintained by HemiHex for Jetson-based advanced vision workflows.