Paddle-Lite——在树莓派端基于python API部署Paddle模型

Paddle Lite环境准备

硬件准备

• 树莓派4B
• usb摄像头
• 装好Buster的镜像源的SD卡

基本软件环境准备

摄像头准备

参考文章：树莓派摄像头的安装、配置与验证

编译库准备

完成gcc、g++、opencv、cmake的安装：

sudo apt-get update

sudo apt-get install gcc g++ make wget unzip libopencv-dev pkg-config

#下载cmake
wget https://www.cmake.org/files/v3.10/cmake-3.10.3.tar.gz

在这一步如果下载很慢，这里我也提供了cmake-3.10.3.tar.gz的包，需要的可以自行下载。

#解压
tar -zxvf cmake-3.10.3.tar.gz

#进入文件夹
cd cmake-3.10.3

#环境配置
sudo ./configure

#make
sudo make

sudo make install

到这里就完成了所有的环境准备。

下载Paddle-Lite

git clone很慢的同学参考博客：git clone速度加快方法

# 1. 下载Paddle-Lite源码 并切换到release分支
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite && git checkout release/v2.6

# 删除此目录，编译脚本会自动从国内CDN下载第三方库文件
rm -rf third-party

编译

cd Paddle-Lite
./lite/tools/build_linux.sh --arch=armv7hf --with_python=ON --python_version=3.7 --with_extra=ON --with_cv=ON

编译的可选项

./lite/tools/build_linux.sh help

• 具体选项

--------------------------------------------------------------------------------------------------------------------------------------------------------
| Methods of compiling Padddle-Lite Linux library:                                                                                                     |
--------------------------------------------------------------------------------------------------------------------------------------------------------
|  compile linux library: (armv8, gcc)                                                                                                                 |
|     ./lite/tools/build_linux.sh                                                                                                                      |
|  print help information:                                                                                                                             |
|     ./lite/tools/build_linux.sh help                                                                                                                 |
|                                                                                                                                                      |
|  optional argument:                                                                                                                                  |
|     --arch: (armv8|armv7hf|armv7), default is armv8                                                                                                  |
|     --toolchain: (gcc|clang), defalut is gcc                                                                                                         |
|     --with_extra: (OFF|ON); controls whether to publish extra operators and kernels for (sequence-related model such as OCR or NLP), default is OFF  |
|     --with_python: (OFF|ON); controls whether to build python lib or whl, default is OFF                                                             |
|     --python_version: (2.7|3.5|3.7); controls python version to compile whl, default is None                                                         |
|     --with_cv: (OFF|ON); controls whether to compile cv functions into lib, default is OFF                                                           |
|     --with_log: (OFF|ON); controls whether to print log information, default is ON                                                                   |
|     --with_exception: (OFF|ON); controls whether to throw the exception when error occurs, default is OFF                                            |
|                                                                                                                                                      |
|  arguments of striping lib according to input model:                                                                                                 |
|     ./lite/tools/build_linux.sh --with_strip=ON --opt_model_dir=YourOptimizedModelDir                                                                |
|     --with_strip: (OFF|ON); controls whether to strip lib accrding to input model, default is OFF                                                    |
|     --opt_model_dir: (absolute path to optimized model dir) required when compiling striped library                                                  |
|  detailed information about striping lib:  https://paddle-lite.readthedocs.io/zh/latest/user_guides/library_tailoring.html                           |
|                                                                                                                                                      |
|  arguments of opencl library compiling:                                                                                                              |
|     ./lite/tools/build_linux.sh --with_opencl=ON                                                                                                     |
|     --with_opencl: (OFF|ON); controls whether to compile lib for opencl, default is OFF                                                              |
|                                                                                                                                                      |
|  arguments of rockchip npu library compiling:                                                                                                        |
|     ./lite/tools/build_linux.sh --with_rockchip_npu=ON --rockchip_npu_sdk_root=YourRockchipNpuSdkPath                                                |
|     --with_rockchip_npu: (OFF|ON); controls whether to compile lib for rockchip_npu, default is OFF                                                  |
|     --rockchip_npu_sdk_root: (path to rockchip_npu DDK file) required when compiling rockchip_npu library                                            |
|                                                                                                                                                      |
|  arguments of baidu xpu library compiling:                                                                                                           |
|     ./lite/tools/build_linux.sh --with_baidu_xpu=ON --baidu_xpu_sdk_root=YourBaiduXpuSdkPath                                                         |
|     --with_baidu_xpu: (OFF|ON); controls whether to compile lib for baidu_xpu, default is OFF                                                        |
|     --baidu_xpu_sdk_root: (path to baidu_xpu DDK file) required when compiling baidu_xpu library                                                     |
--------------------------------------------------------------------------------------------------------------------------------------------------------

编译完成

安装编译生成的python paddle-lite包

进入dist目录下
cd /Paddle-Lite/build.lite.linux.armv7hf.gcc/inference_lite_lib.armlinux.armv7hf/python/install/dist

pip3 install paddlelite-2708c2fe-cp37-cp37m-linux_armv7l.whl

运行基于python API的demo程序

准备模型文件

• 下载模型

wget http://paddle-inference-dist.bj.bcebos.com/mobilenet_v1.tar.gz
tar zxf mobilenet_v1.tar.gz

• 采用opt工具转换模型

paddle_lite_opt --model_dir=./mobilenet_v1  \
                --optimize_out=mobilenet_v1_opt \
                --optimize_out_type=naive_buffer \
                --valid_targets=arm

转化成功

运行模型

需要注意的是，在这里有两个demo文件，他们的区别在于

• full_api.py中需要的模型文件为__model__和__param__文件，API详解：CxxPredictor
• light_api.py中需要的模型文件为opt转换后的model.nb文件，API详解：LightPredictor

# light api的输入为优化后模型文件mobilenet_v1_opt.nb
python3 mobilenetv1_light_api.py --model_dir=mobilenet_v1_opt.nb

部署自己的模型

Paddle用于推理的模型是通过save_inference_model这个API保存下来的，保存的格式有两种，在这里将在AI studio上运行生成的模型参数文件下载并挂载在树莓派上：

两种模型格式

• non-combined形式：参数保存到独立的文件，如设置 model_filename 为 None , params_filename 为 None
  

• combined形式：参数保存到同一个文件，如设置 model_filename 为 model , params_filename 为 params

编译opt工具

在树莓派端编译：

cd Paddle-Lite
./lite/tools/build.sh build_optimize_tool

使用opt转换模型

paddle_lite_opt --model_dir=./mobilenet_v1 \
      --valid_targets=arm \
      --optimize_out_type=naive_buffer \
      --optimize_out=mobilenet_v1_opt

具体的API详情参考：

• Python API — Paddle-Lite 文档

编译过程参考：

• 源码编译 (ARMLinux)
• 模型优化工具 opt