瑞芯微RK3588s部署百度飞桨fastdeploy_fastdeploy python安装 rk3588

0. 环境

- 虚拟机ubuntu20
- ROC-RK3588S-PC_Ubuntu20.04-Xfce-r3115_v1.3.0c_240131

1. 编译C++

1.1 [DEV]编译

#安装依赖
$ sudo apt install git cmake g++ python3-pip
#获取源码
$ cd ~/Downloads
$ git clone https://github.com/PaddlePaddle/FastDeploy.git
#压缩备份：
$ 7z a FastDeploy_git_src_20240311.7z FastDeploy
 
# 编译安装
 
$ cd FastDeploy
$ git checkout develop
$ mkdir build && cd build
$ cmake .. -DENABLE_ORT_BACKEND=ON \
         -DENABLE_VISION=ON \
         -DENABLE_RKNPU2_BACKEND=ON \
         -DRKNN2_TARGET_SOC=RK3588 \
         -DCMAKE_INSTALL_PREFIX=${PWD}/fastdeploy-dev
$ make -j8
$ make install

1.2 [PC]交叉编译

 
# 准备交叉编译工具链
sudo apt install cmake build-essential
wget https://bj.bcebos.com/fastdeploy/third_libs/gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu.tar.gz
tar -xzvf gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu.tar.gz -C /path/to/save
 
# 编译
cd FastDeploy
mkdir build && cd build
cmake ..  -DCMAKE_C_COMPILER=/home/xxjianvm/gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu/bin/aarch64-linux-gnu-gcc \
          -DCMAKE_CXX_COMPILER=/home/xxjianvm/gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu/bin/aarch64-linux-gnu-g++ \
          -DCMAKE_TOOLCHAIN_FILE=./../cmake/toolchain.cmake \
          -DTARGET_ABI=arm64 \
          -DENABLE_ORT_BACKEND=OFF \
          -DENABLE_RKNPU2_BACKEND=ON \
          -DENABLE_VISION=ON \
          -DRKNN2_TARGET_SOC=RK3588 \
          -DCMAKE_INSTALL_PREFIX=/home/xxjianvm/Downloads/FastDeploy/install
make -j8
make install

RKNN2_TARGET_SOC可以选择RK3588/RK356X

2. [DEV]编译Python安装包

# 依赖
sudo apt install git cmake g++ python3-pip
# 获取FastDeploy源码
git clone https://github.com/PaddlePaddle/FastDeploy.git
cd FastDeploy/python
git checkout develop
# 编译
#     Python通过export环境变量设置编译选项
export ENABLE_ORT_BACKEND=ON
export ENABLE_RKNPU2_BACKEND=ON
export ENABLE_VISION=ON
#     请根据你的开发版的不同，选择RK3588和RK356X
export RKNN2_TARGET_SOC=RK3588
#     如果你的核心板的运行内存大于等于8G，我们建议您执行以下命令进行编译。
python3 setup.py build
#     值得注意的是，如果你的核心板的运行内存小于8G，我们建议您执行以下命令进行编译。
python3 setup.py build -j1

# 安装
 
python3 setup.py bdist_wheel
cd dist
pip3 install fastdeploy_python-0.0.0-cp38-cp38-linux_aarch64.whl

3. 模型转换

# 这里测试直接下载Paddle静态图模型并解压
cd ~/Downloads
wget https://paddledet.bj.bcebos.com/deploy/Inference/picodet_s_416_coco_lcnet.tar
tar xvf picodet_s_416_coco_lcnet.tar

3.1 [PC]模型转换 paddle2onnx

#设置环境
source /home/xxjianvm/Downloads/FastDeploy/install/fastdeploy_init.sh 
export PATH="$PATH:/home/xxjianvm/.local/bin"
 
# 静态图模型转换onnx模型，
paddle2onnx --model_dir picodet_s_416_coco_lcnet \
        --model_filename model.pdmodel \
        --params_filename model.pdiparams \
        --save_file picodet_s_416_coco_lcnet/picodet_s_416_coco_lcnet.onnx \
        --enable_dev_version True

问题：paddle2onnx: command not found
解决办法：通过pip3 install 得到的 paddle2onnx，默认路径在 ~/.local/bin

export PATH="$PATH:/home/xxjianvm/.local/bin"

# 固定shape

python3 -m paddle2onnx.optimize --input_model picodet_s_416_coco_lcnet/picodet_s_416_coco_lcnet.onnx \
                            --output_model picodet_s_416_coco_lcnet/picodet_s_416_coco_lcnet.onnx \
                            --input_shape_dict "{'image':[1,3,416,416]}"

# 最后模型保存在当前目录下picodet_s_416_coco_lcnet/picodet_s_416_coco_lcnet.onnx

3.2 [PC]模型转换 onnx2rknn

切换目录

cd FastDeploy/tools/rknpu2

把

~/Downloads/picodet_s_416_coco_lcnet/picodet_s_416_coco_lcnet

拷贝到

FastDeploy/tools/rknpu2/picodet_s_416_coco_lcnet

打开 netron
 打开文件 ->

FastDeploy/tools/rknpu2/picodet_s_416_coco_lcnet/picodet_s_416_coco_lcnet.onnx

导出rknn模型

python3 export.py --config_path config/picodet_s_416_coco_lcnet_unquantized.yaml --target_platform rk3588

4. 推理

4.1 [DEV]推理测试 python

# 依赖 
pip3 install opencv-python
# 切换目录
cd /home/firefly/Downloads/FastDeploy/examples/vision/detection/paddledetection/rknpu2/python
# 获取模型文件
把带有 picodet_s_416_coco_lcnet_rk3588_unquantized.rknn 的文件夹 picodet_s_416_coco_lcnet 拷贝到当前目录
 
# 获取测试图片
wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg
 
# 运行推理程序，--model_file 指定模型文件，--config_file 指定配置文件， --image 指定需要推理的图片
$ export LD_LIBRARY_PATH=/home/firefly/Downloads/FastDeploy/build/fastdeploy-dev/third_libs/install/opencv/lib
$ python3 infer.py --model_file picodet_s_416_coco_lcnet/picodet_s_416_coco_lcnet_rk3588_unquantized.rknn --config_file picodet_s_416_coco_lcnet/infer_cfg.yml --image 000000014439.jpg

4.2 [DEV]推理测试 c++

 
# 切换目录
cd /home/firefly/Downloads/FastDeploy/examples/vision/detection/paddledetection/rknpu2/cpp
# 修改源码中的模型名
  auto model_file = model_dir + "/picodet_s_416_coco_lcnet_rk3588_unquantized.rknn";
# 配置
mkdir build && cd build
cmake .. -DFASTDEPLOY_INSTALL_DIR=/home/firefly/Downloads/FastDeploy/build/fastdeploy-dev
# 编译
make -j8
# 准备模型
把带有 picodet_s_416_coco_lcnet_rk3588_unquantized.rknn 的文件夹 picodet_s_416_coco_lcnet 拷贝到/home/firefly/Downloads/FastDeploy/examples/vision/detection/paddledetection/rknpu2/cpp
# 准备测试图片
cd /home/firefly/Downloads/FastDeploy/examples/vision/detection/paddledetection/rknpu2/cpp
wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg
# 设置环境变量
# export LD_LIBRARY_PATH=/home/firefly/Downloads/FastDeploy/build/fastdeploy-dev/third_libs/install/opencv/lib:/home/firefly/Downloads/FastDeploy/build/fastdeploy-dev/third_libs/install/onnxruntime/lib:/home/firefly/Downloads/FastDeploy/build/fastdeploy-dev/third_libs/install/paddle2onnx/lib
source /home/firefly/Downloads/FastDeploy/build/fastdeploy-dev/fastdeploy_init.sh 
# 推理
cd /home/firefly/Downloads/FastDeploy/examples/vision/detection/paddledetection/rknpu2/cpp/build
./infer_picodet_demo /home/firefly/Downloads/FastDeploy/examples/vision/detection/paddledetection/rknpu2/cpp/picodet_s_416_coco_lcnet ../000000014439.jpg 0
./infer_picodet_demo /home/firefly/Downloads/FastDeploy/examples/vision/detection/paddledetection/rknpu2/cpp/picodet_s_416_coco_lcnet ../000000014439.jpg 1

参考：

[1]FastDeploy RKNPU2 导航文档，https://github.com/PaddlePaddle/FastDeploy/blob/develop/docs/cn/build_and_install/rknpu2.md
[2]3. 飞桨（PaddlePaddle） FastDeploy，https://doc.embedfire.com/linux/rk356x/Ai/zh/latest/lubancat_ai/example/paddlepaddle_fastdeploy.html