编译部署 - Yolov5

自己部署一次 RK 官方的 rknn_model_zoo yolov5

交叉编译工具链:

RV1126 链接: https://pan.baidu.com/s/1RZC4oGlDzX6iyWIBd9zkoQ?pwd=btgt 提取码: btgt 复制这段内容后打开百度网盘手机App，操作更方便哦

1. 准备编译环境

1.1 电脑硬件配置

编译 Ubuntu 开发环境电脑硬件配置参考:

硬件	信息
CPU	64位 4核 ⬆️
系统	Ubuntu Server 24.04.3 LTS \| 建议最低 Ubunu 22 LTS ⬆️ 不要使用 Windows 子系统 (WSL)
内存	8GB ⬆️
存储	64GB ⬆️ 空闲
用户	普通用户, 不要使用 root 用户编译

1.2 电脑软件配置

项目	配置
使用系统	Ubuntu Server 24.04.3 LTS
主板芯片	`RV1126B-p`
SDK	`rv1126-repo-v20230717-dev.tar.gz` (已包含 `prebuilts` 目录)
交叉编译环境	`rv1126-repo-v20230717-dev_prebuilts.zip` 下载了 SDK 无需单独下载 `prebuilts`

sudo apt update
sudo apt install git unzip vim
sudo apt install -y git cmake build-essential

## Ubuntu 24 server (minimized) 缺这个
sudo apt update
sudo apt install -y libgl1
sudo apt install -y libsm6 libxext6 libxrender1
sudo apt install -y libglib2.0-0 libgomp1

1.3 安装 Python 3.8

推荐使用 PPA（Personal Package Archives）仓库来安装:

添加 PPA 仓库:

sudo add-apt-repository ppa:deadsnakes/ppa

按 Enter 键继续

更新软件包列表:
```
sudo apt update
```

安装 Python 3.8:

sudo apt install python3.8
sudo apt install python3.8-venv
sudo apt install python3.8-dev build-essential

验证安装:
```
python3.8 --version
```
显示 Python 3.8.x 表示安装成功

1.3 创建虚拟环境

# 创建虚拟环境 venv_1126
python3.8 -m venv venv_1126
source venv_1126/bin/activate

## 虚拟环境里升级 pip
pip install --upgrade pip

2. 准备软件环境

2.0 激活虚拟环境再操作

激活虚拟环境:
```
cd ~
source venv_1126/bin/activate
```

2.1 准备环境转换软件

这里下载 https://github.com/rockchip-linux/rknn-toolkit/releases

环境安装 rknn-toolkit v1.75:

cd ~
mkdir ~/rknn-toolkit-v1.75
cd ~/rknn-toolkit-v1.75
wget https://github.com/rockchip-linux/rknn-toolkit/releases/download/v1.7.5/rknn-toolkit-v1.7.5-packages.tar.gz

# 解压
tar -zxvf rknn-toolkit-v1.7.5-packages.tar.gz

复制一个 ubuntu22 改为 ubuntu24 用

cp packages/requirements-cpu-ubuntu20.04_py38.txt packages/requirements-cpu-ubuntu24.04_py38.txt

内容版本如下: vim packages/requirements-cpu-ubuntu24.04_py38.txt

tensorflow==2.2.0
torch==1.6.0
torchvision==0.7.0
mxnet==1.5.0

# 卸载可能的旧环境
# pip uninstall protobuf onnx

# 安装 rknn_toolkit-v1.7.5
pip install -r packages/requirements-cpu-ubuntu24.04_py38.txt
pip install packages/rknn_toolkit-1.7.5-cp38-cp38-linux_x86_64.whl

准备环境 rknn_model_zoo

cd ~
git clone https://github.com/airockchip/rknn_model_zoo.git

3. 转换模型

3.1 获取并转换模型

下载模型:

cd ~/rknn_model_zoo/examples/yolov5/model/
wget https://ftrg.zbox.filez.com/v2/delivery/data/95f00b0fc900458ba134f8b180b3f7a1/examples/yolov5/yolov5n.onnx

转换模型:

cd ~/rknn_model_zoo/examples/yolov5/python/
python3 convert.py ../model/yolov5n.onnx rv1126 u8 ../model/yolov5n_rv1126.rknn

模型下载:

rknn_model_zoo 提供的范例模型列表见首页 readme: https://github.com/airockchip/rknn_model_zoo

当前测试模型 yolov5n: https://ftrg.zbox.filez.com/v2/delivery/data/95f00b0fc900458ba134f8b180b3f7a1/examples/yolov5/yolov5n.onnx

3.2 转换结果

转换过程:

(venv_1126) roy@Debian-i7-7700:~/rknn_model_zoo/examples/yolov5/python$ python3 convert.py ../model/yolov5n.onnx rv1126 u8 ../model/yolov5n_rv1126.rknn 
W rknn-toolkit version: 1.7.5
--> Config model
done
--> Loading model
W Catch exception when optimize onnx model, skip it! Exception: Unresolved value references: /model.11/Constant_1_output_0,/model.11/Constant_output_0,/model.15/Constant_1_output_0,/model.15/Constant_output_0,
/home/roy/venv_1126/lib/python3.8/site-packages/rknn/api/rknn.py:160: FutureWarning: Possible nested set at position 7
  ret = self.rknn_base.load_onnx(model, inputs, input_size_list, outputs)
done
--> Building model
W found outlier value, this may affect quantization accuracy
const name                                   abs_mean    abs_std     outlier value
Conv_/model.0/conv/conv/Conv_194:weight      0.68        0.89        -11.603     
W:tensorflow:From /home/roy/venv_1126/lib/python3.8/site-packages/tensorflow/python/framework/func_graph.py:981: py_func (from tensorflow.python.ops.script_ops) is deprecated and will be removed in a future version.
Instructions for updating:
tf.py_func is deprecated in TF V2. Instead, there are two
    options available in V2.
    - tf.py_function takes a python function which manipulates tf eager
    tensors instead of numpy arrays. It's easy to convert a tf eager tensor to
    an ndarray (just call tensor.numpy()) but having access to eager tensors
    means `tf.py_function`s can use accelerators such as GPUs as well as
    being differentiable using a gradient tape.
    - tf.numpy_function maintains the semantics of the deprecated tf.py_func
    (it is not differentiable, and manipulates numpy arrays). It drops the
    stateful argument making all functions stateful.
    
W:tensorflow:From /home/roy/venv_1126/lib/python3.8/site-packages/tensorflow/python/ops/resource_variable_ops.py:1659: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
Instructions for updating:
If using Keras pass *_constraint arguments to layers.


done
--> Export rknn model
done



(venv_1126) roy@Debian-i7-7700:~/rknn_model_zoo/examples/yolov5/python$ ls ../model/
anchors_yolov5.txt  bus.jpg  coco_80_labels_list.txt  download_model.sh  yolov5n.onnx  yolov5n_rv1126.rknn

转换后的文件位于: ~/rknn_model_zoo/examples/yolov5/model/

包含文件:

yolov5n.onnx - 原始 ONNX 模型
yolov5n_rv1126.rknn - 转换后的 RKNN 模型
anchors_yolov5.txt - 锚点文件
coco_80_labels_list.txt - 标签文件
bus.jpg - 测试图片
download_model.sh - 范例提供了一个 yolov5s_relu.onnx 可以自行尝试

4. 交叉编译 YOLOV5

4.0 激活虚拟环境再操作

激活虚拟环境:
```
cd ~
source venv_1126b/bin/activate
```

4.1 准备交叉编译工具链

如果已经下载 SDK rv1126b_linux6.1_sdk_v1.1.0.tar.gz无需再次下载rv1126-repo-v20230717-dev_prebuilts.zip

直接使用 prebuilts 目录即可

下载交叉编译工具链:

下载地址见 1.2 电脑软件配置

下载并解压交叉编译环境:

cd ~
unzip rv1126-repo-v20230717-dev_prebuilts.zip

获取交叉编译工具路径:

RV1126 是 arm 获取下对应环境的路径

cd prebuilts/gcc/linux-x86/aarch64/gcc-arm-10.3-2021.07-x86_64-aarch64-none-linux-gnu
(venv_1126b) roy@Debian-i7-7700:~/prebuilts/gcc/linux-x86/aarch64/gcc-arm-10.3-2021.07-x86_64-aarch64-none-linux-gnu$ pwd 

/home/roy/prebuilts/gcc/linux-x86/aarch64/gcc-arm-10.3-2021.07-x86_64-aarch64-none-linux-gnu

输出路径: /home/roy/prebuilts/gcc/linux-x86/arm/gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf/bin

4.2 交叉编译 yolov5

设置环境变量:

export PATH=~/prebuilts/gcc/linux-x86/arm/gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf/bin:$PATH
export GCC_COMPILER=~/prebuilts/gcc/linux-x86/arm/gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf/bin/arm-linux-gnueabihf

开始编译:

cd ~/rknn_model_zoo
chmod +x build-linux.sh
./build-linux.sh -t rv1126 -a armhf -d yolov5

编译输出内容: 程序位于 ~/rknn_model_zoo/install/rv1126b_linux_aarch64/rknn_yolov5_demo

4.3 编译结果

(venv_1126) roy@Debian-i7-7700:~/rknn_model_zoo$ ./build-linux.sh -t rv1126 -a armhf -d yolov5

./build-linux.sh -t rv1126 -a armhf -d yolov5
/home/roy/prebuilts/gcc/linux-x86/arm/gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf/bin/arm-linux-gnueabihf
===================================
BUILD_DEMO_NAME=yolov5
BUILD_DEMO_PATH=examples/yolov5/cpp
TARGET_SOC=rv1126
TARGET_ARCH=armhf
BUILD_TYPE=Release
ENABLE_ASAN=OFF
DISABLE_RGA=OFF
DISABLE_LIBJPEG=OFF
INSTALL_DIR=/home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo
BUILD_DIR=/home/roy/rknn_model_zoo/build/build_rknn_yolov5_demo_rv1126_linux_armhf_Release
CC=/home/roy/prebuilts/gcc/linux-x86/arm/gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf/bin/arm-linux-gnueabihf-gcc
CXX=/home/roy/prebuilts/gcc/linux-x86/arm/gcc-arm-8.3-2019.03-x86_64-arm-linux-gnueabihf/bin/arm-linux-gnueabihf-g++
===================================
-- Configuring done (0.0s)
-- Generating done (0.0s)
-- Build files have been written to: /home/roy/rknn_model_zoo/build/build_rknn_yolov5_demo_rv1126_linux_armhf_Release
[ 50%] Built target imageutils
[ 50%] Built target imagedrawing
[ 50%] Built target fileutils
[ 66%] Built target audioutils
[100%] Built target rknn_yolov5_demo
[ 16%] Built target imageutils
[ 33%] Built target fileutils
[ 50%] Built target imagedrawing
[ 83%] Built target rknn_yolov5_demo
[100%] Built target audioutils
Install the project...
-- Install configuration: "Release"
-- Installing: /home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/./rknn_yolov5_demo
-- Set non-toolchain portion of runtime path of "/home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/./rknn_yolov5_demo" to "$ORIGIN/lib"
-- Installing: /home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/./model/bus.jpg
-- Installing: /home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/./model/coco_80_labels_list.txt
-- Installing: /home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/model/yolov5n_rv1126.rknn
-- Installing: /home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/model/yolov5n_rv1126b.rknn
-- Installing: /home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/lib/librknn_api.so
-- Installing: /home/roy/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/lib/librga.so

5. RV1126 设备验证

5.1 部署文件

复制模型到程序目录:

准备一个 RV1126 主板，并且刷好固件开机
复制模型到程序目录:

如果 rknn_yolov5_demo 没有 yolov5n_rv1126.rknn 可以手动复制一个

cp ~/rknn_model_zoo/examples/yolov5/model/yolov5n_rv1126.rknn ~/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/model/

安装 ADB 工具:

sudo apt install google-android-platform-tools-installer

连接设备并上传程序:

有线连接: 直接接入 usb 数据线

roy@Debian-i7-7700:~$ adb devices
List of devices attached
fec3d3812d367ee	device

adb push ~/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/ /root/

无线连接: 先获取设备 IP 地址，例如: 10.0.1.226

roy@Debian-i7-7700:~$ adb connect 10.0.1.226
connected to 10.0.1.226:5555

roy@Debian-i7-7700:~$ adb devices 
List of devices attached
10.0.1.226:5555	device

adb push ~/rknn_model_zoo/install/rv1126_linux_armhf/rknn_yolov5_demo/ /root/

5.2 执行验证

先确保连接上设备:

(venv_1126) roy@Debian-i7-7700:~$ adb devices 
List of devices attached
10.0.1.226:5555	device

(venv_1126) roy@Debian-i7-7700:~$ adb shell 
root@rv1126_rv1109:/# 

root@rv1126_rv1109:~# ls /root/
rknn_yolov5_demo

root@rv1126_rv1109:~# cd /root/rknn_yolov5_demo/
root@rv1126_rv1109:/root/rknn_yolov5_demo# ls
lib  model  rknn_yolov5_demo

运行测试程序:

root@rv1126_rv1109:/root/rknn_yolov5_demo# ./rknn_yolov5_demo ./model/yolov5n_rv1126.rknn ./model/bus.jpg

5.3 运行结果

root@rv1126_rv1109:/root/rknn_yolov5_demo# ./rknn_yolov5_demo ./model/yolov5n_rv1126.rknn ./model/bus.jpg
load lable ./model/coco_80_labels_list.txt
librknn_runtime version 1.7.3 (2e55827 build: 2022-08-25 10:45:32 base: 1131)
model input num: 1, output num: 3
input tensors:
  index=0, name=images_209, n_dims=4, dims=[1, 3, 640, 640], n_elems=1228800, size=1228800, fmt=NCHW, type=UINT8, qnt_type=AFFINE, zp=0, scale=0.003922
output tensors:
  index=0, name=Sigmoid_/model.24/Sigmoid/out0_0, n_dims=4, dims=[1, 255, 80, 80], n_elems=1632000, size=1632000, fmt=NCHW, type=UINT8, qnt_type=AFFINE, zp=0, scale=0.003917
  index=1, name=Sigmoid_/model.24/Sigmoid_1/out0_1, n_dims=4, dims=[1, 255, 40, 40], n_elems=408000, size=408000, fmt=NCHW, type=UINT8, qnt_type=AFFINE, zp=0, scale=0.003918
  index=2, name=Sigmoid_/model.24/Sigmoid_2/out0_2, n_dims=4, dims=[1, 255, 20, 20], n_elems=102000, size=102000, fmt=NCHW, type=UINT8, qnt_type=AFFINE, zp=0, scale=0.003920
model is NCHW input fmt
model input height=640, width=640, channel=3
origin size=640x640 crop size=640x640
input image: 640 x 640, subsampling: 4:2:0, colorspace: YCbCr, orientation: 1
scale=1.000000 dst_box=(0 0 639 639) allow_slight_change=1 _left_offset=0 _top_offset=0 padding_w=0 padding_h=0
librga fail to get driver version! Compatibility mode will be enabled.

rga_api version 1.10.1_[0]
rknn_run
person @ (108 235 203 523) 0.855
person @ (209 246 288 506) 0.830
person @ (483 223 568 510) 0.510
bus @ (122 158 557 451) 0.476
write_image path: out.png width=640 height=640 channel=3 data=0xa5347008

最终输出: 文件信息:

root@rv1126-buildroot:~/rknn_yolov5_demo# ls 
lib               model             out.png           rknn_yolov5_demo

效果图:

bus

总结

通过以上步骤，我们成功完成了:

✅ 搭建交叉编译环境
✅ 转换 Yolov5 模型为 RKNN 格式
✅ 编译 RKNN Yolov5 演示程序
✅ 在 RV1126 设备上部署并验证

注意事项

项目	RV1126B-p	RV1126	说明
RKNN Toolkit 版本	rknn-toolkit2 v2.x	rknn-toolkit v1.7.5	两者不兼容，不可混用
系统架构	64位 ARM (aarch64)	32位 ARM (armhf)	需要对应的交叉编译工具链
编译选项	`-t rv1126b -a aarch64`	`-t rv1126 -a armhf`	编译脚本参数不同
部署方式	文件直接部署	文件直接部署	运行方式差异
模型转换警告	可能出现，通常不影响使用	可能出现，通常不影响使用	可忽略的量化精度警告
Python 环境	Python 3.8 + 虚拟环境	Python 3.8 + 虚拟环境	确保环境一致性

1. 准备编译环境​

1.1 电脑硬件配置​

1.2 电脑软件配置​

1.3 安装 Python 3.8​

1.3 创建虚拟环境​

2. 准备软件环境​

2.0 激活虚拟环境再操作​

2.1 准备环境转换软件​

3. 转换模型​

3.1 获取并转换模型​

3.2 转换结果​

4. 交叉编译 YOLOV5​

4.0 激活虚拟环境再操作​

4.1 准备交叉编译工具链​

4.2 交叉编译 yolov5​

4.3 编译结果​

5. RV1126 设备验证​

5.1 部署文件​

5.2 执行验证​

5.3 运行结果​

总结​

注意事项​