Model Conversion

1. Jetson Orin YOLO11 (Benchmark)

YOLO11 benchmark data is sourced from the Ultralytics team, testing various model formats (data for reference only)

2. Enable Maximum Board Performance

2.1. Enable MAX Power Mode

Enabling MAX Power Mode on Jetson will ensure that all CPU and GPU cores are turned on:

#Orin Nano
sudo nvpmodel -m 2
#Orin Nx
sudo nvpmodel -m 0

2.2. Enable Jetson Clocks

Enabling Jetson Clocks will ensure that all CPU and GPU cores run at maximum frequency:

xxxxxxxxxx
sudo jetson_clocks

3. Model Conversion

According to the test parameters provided by the Ultralytics team for different format models, we can find that using TensorRT provides the best inference performance!

xxxxxxxxxx
First time using YOLO11 export mode will automatically install some dependencies, please wait for the automatic completion!

3.1. CLI: pt → onnx → engine

Convert PyTorch format model to TensorRT: The conversion process will automatically generate ONNX model

xxxxxxxxxx
cd /home/jetson/ultralytics/ultralytics

xxxxxxxxxx
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

3.2. Python: pt → onnx → engine

Convert PyTorch format model to TensorRT: The conversion process will automatically generate ONNX model

xxxxxxxxxx
cd /home/jetson/ultralytics/ultralytics/yahboom_demo

xxxxxxxxxx
python3 model_pt_onnx_engine.py

​x
from ultralytics import YOLO
​
# Load a YOLO11n PyTorch model
# model = YOLO("/home/jetson/ultralytics/ultralytics/yolo11n.pt")
model = YOLO("/home/jetson/ultralytics/ultralytics/yolo11n-seg.pt")
# model = YOLO("/home/jetson/ultralytics/ultralytics/yolo11n-pose.pt")
# model = YOLO("/home/jetson/ultralytics/ultralytics/yolo11n-cls.pt")
# model = YOLO("/home/jetson/ultralytics/ultralytics/yolo11n-obb.pt")
​
# Export the model to TensorRT
model.export(format="engine")

Note: The converted model files are located at the same location as the original model files

4. Model Prediction

CLI Usage

CLI currently only supports USB cameras. For CSI cameras, users can directly modify the previous Python code to call ONNX and ENGINE models!

xxxxxxxxxx
cd /home/jetson/ultralytics/ultralytics

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

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

Common Issues

ERROR: onnxslim

Solution: Enter onnxslim installation command in terminal

xxxxxxxxxx
sudo pip3 install onnxslim

References

https://docs.ultralytics.com/guides/nvidia-jetson/

https://docs.ultralytics.com/integrations/tensorrt/