Jeston Orin Nano平台部署Ultralytics

默认Jeston Orin Nano平台已经安装JetPack6.2.1。

step1 : 更新pip

sudo apt update
sudo apt install python3-pip -y
sudo -H pip install -U pip (这一步更新系统pip，可能会有warning)

step2 : 安装ultralytics

这里使用清华源安装

sudo -H pip install ultralytics -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple

如果安装中报错无法卸载sympy1.9，使用如下命令卸载后再安装

sudo apt purge python3-sympy

step3 :（重新）安装torch/torchvision

稳妥起见，先把之前torch/torchvision卸载

sudo -H pip uninstall torch torchvision

然后安装ultralytics特供版，JetPack6.2需要用torch-2.8.0-cp310-cp310-linux_aarch64.whl和相对应的torchvision-0.23.0-cp310-cp310-linux_aarch64.whl，从github上下载后安装

sudo -H pip install torch-2.8.0-cp310-cp310-linux_aarch64.whl
sudo -H pip install torchvision-0.23.0-cp310-cp310-linux_aarch64.whl

step 4：安装cuSPARSELt修复依赖关系

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/arm64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
sudo apt-get -y install libcusparselt0 libcusparselt-dev

step 5：安装onnx/onnxslim/onnxruntime-gpu

使用清华源安装onnx和onnxslim

sudo -H pip install onnx onnxslim -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple

onnxruntime-gpu还是ultralytics的特供版，版本onnxruntime_gpu-1.20.0-cp310-cp310-linux_aarch64.whl，从github上下载后安装

sudo -H pip install onnxruntime_gpu-1.20.0-cp310-cp310-linux_aarch64.whl

step 6：安装numpy1.23.5（替换更新中安装的num2.x）

sudo -H pip install numpy==1.23.5 -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple

部署完成，可以用python脚本中测试

from ultralytics import YOLO
model = YOLO("yolo11n.pt")     # 可以从github上预先下载
model.export(format="engine")  # creates 'yolo11n.engine'

trt_model = YOLO("yolo11n.engine")
results = trt_model("https://ultralytics.com/images/bus.jpg") # FP32精度，640x640的图片推理时间大约20ms