Model Conversion

1. Raspberry Pi 5 YOLO26 (Benchmark)

The Ultralytics team benchmarked YOLO26 across ten different model formats, measuring speed and accuracy. These formats include: PyTorch, TorchScript, ONNX, OpenVINO, TF SavedModel, TF GraphDef, TF Lite, MNN, NCNN, ExecuTorch. The benchmarks were run on a Raspberry Pi 5 at FP32 precision with a default input image size of 640.

Comparison Chart

2. Model Conversion

The first time you use YOLO26's export mode, it will automatically install some dependencies. Please wait for it to complete!

2.1. CLI: pt → onnx, pt → ncnn

Convert PyTorch format models to onnx and ncnn.

xxxxxxxxxx
cd /home/sunrise/ultralytics/ultralytics

​x
yolo export model=yolo26n.pt format=onnx
# yolo export model=yolo26n-seg.pt format=onnx
# yolo export model=yolo26n-pose.pt format=onnx
# yolo export model=yolo26n-cls.pt format=onnx
# yolo export model=yolo26n-obb.pt format=onnx
​
yolo export model=yolo26n.pt format=ncnn
# yolo export model=yolo26n-seg.pt format=ncnn
# yolo export model=yolo26n-pose.pt format=ncnn
# yolo export model=yolo26n-cls.pt format=ncnn
# yolo export model=yolo26n-obb.pt format=ncnn

2.2. Python: pt → onnx → ncnn

Convert PyTorch format models: The conversion process will automatically generate an ONNX model.

xxxxxxxxxx
cd /home/sunrise/ultralytics/ultralytics/yahboom_demo_26

xxxxxxxxxx
python3 model_pt_onnx_ncnn.py

xxxxxxxxxx
from ultralytics import YOLO
​
# Load a YOLo PyTorch model
model = YOLO("/home/sunrise/ultralytics/ultralytics/yolo26n.pt")
# model = YOLO("/home/sunrise/ultralytics/ultralytics/yolo26n-seg.pt")
# model = YOLO("/home/sunrise/ultralytics/ultralytics/yolo26n-pose.pt")
# model = YOLO("/home/sunrise/ultralytics/ultralytics/yolo26n-cls.pt")
# model = YOLO("/home/sunrise/ultralytics/ultralytics/yolo26n-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'
​

Note: The converted model files are located in the same directory as the original model file.

4. Model Prediction

CLI Usage

Currently, CLI only supports USB cameras. For CSI cameras, users can directly modify the Python code above to use onnx and ncnn models.

xxxxxxxxxx
cd /home/sunrise/ultralytics/ultralytics

xxxxxxxxxx
yolo predict model=yolo26n.onnx source=0 save=False show

xxxxxxxxxx
yolo predict model=yolo26n_ncnn_model source=0 save=False show

References

Quick Start Guide: Raspberry Pi with Ultralytics YOLO26