Model Conversion

Note: Using the Docker container in the factory image does not require re-setting up the environment. The environment is already set up. Simply enter Docker and run the corresponding function commands according to the previous tutorial.

1. Raspberry Pi 5 YOLO11 (Benchmark)

YOLO11 benchmark data comes from the Ultralytics team, and is tested on models in various formats (data for reference only).

Officially, only YOLO11n and YOLO11s models were benchmarked. Other models are too large to run on Raspberry Pis and may not provide good performance.

2. Model Conversion

**Conversion cannot be performed on the Jetson Nano board in Docker because the PyTorch/Ultralytics dependencies are too new and the CPU/GPU instruction sets are incompatible. You can first export to TensorRT/ONNX on a Raspberry Pi or PC, then run it on the Jetson Nano:

Based on the test parameters for different model formats provided by the Ultralytics team, we found that using TensorRT for inference performance is the best!

xxxxxxxxxx
The first time you use the export mode for YOLO11, some dependencies will automatically be installed. Wait for it to complete!

2.1. CLI: pt → onnx, pt → ncnn

Convert the PyTorch model to onnx and ncnn

xxxxxxxxxx
cd ~/ultralytics/ultralytics/

Run the following command in the terminal:

xxxxxxxxxx
yolo export model=yolo11n.pt format=onnx
# yolo export model=yolo11n-seg.pt format=onnx
# yolo export model=yolo11n-pose.pt format=onnx
# yolo export model=yolo11n-cls.pt format=onnx
# yolo export model=yolo11n-obb.pt format=onnx
​
yolo export model=yolo11n.pt format=ncnn
# yolo export model=yolo11n-seg.pt format=ncnn
# yolo export model=yolo11n-pose.pt format=ncnn
# yolo export model=yolo11n-cls.pt format=ncnn
# yolo export model=yolo11n-obb.pt format=ncnn

2.2. Python: pt → onnx → ncnn

Converting a PyTorch model to TensorRT: The conversion process automatically generates an ONNX model.

xxxxxxxxxx
cd ~/ultralytics/ultralytics/yahboom_demo

xxxxxxxxxx
python3 model_pt_onnx_ncnn.py

Sample code:

xxxxxxxxxx
from ultralytics import YOLO
​
# Load a YOLO11n PyTorch model
model = YOLO("/root/ultralytics/ultralytics/yolo11n.pt")
# model = YOLO("/root/ultralytics/ultralytics/yolo11n-seg.pt")
# model = YOLO("/root/ultralytics/ultralytics/yolo11n-pose.pt")
# model = YOLO("/root/ultralytics/ultralytics/yolo11n-cls.pt")
# model = YOLO("/root/ultralytics/ultralytics/yolo11n-obb.pt")
​
# Export the model to ONNX format
model.export(format="onnx") # This will create 'yolo11n.onnx' in the same directory
​
# Export the model to NCNN format
model.export(format="ncnn") # Creates 'yolo11n_ncnn_model'

3. Model Prediction

CLI Usage

Currently, the CLI only supports USB camera calls!

xxxxxxxxxx
cd ~/ultralytics/ultralytics/

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

xxxxxxxxxx
yolo predict model=yolo11n_ncnn_model source=0 save=False show

References

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

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