1. 使用yolov5训练出自己的模型pen.pt
2. 将pen.pt模型转换成pen.onnx格式
克隆下面的开源代码,执行得到pen.onnx
git clone https://github.com/shaoshengsong/yolov5_62_export_ncnn.git
cd yolov5_62_export_ncnn
python export.py --weights pen.pt --img 640 --batch 1 --include onnx torchscript --simplify
3. 克隆下面的ncnn开源库并进行编译安装
git clone https://github.com/Tencent/ncnn
cd ncnn
mkdir build
cd build
cmake ..
make
make install
4. 将pen.onnx转换成pen.param和pen.bin
cd ncnn/build/tools/onnx
./onnx2ncnn pen.onnx pen.param pen.bin
5. 将.param文件中的Reshape行中的6400、1600、400都改为-1
6. 将 pen.param和pen.bin转化成C++可用的pen.id.h和pen.mem.h
cd ncnn/build/tools
./ncnn2mem pen.param pen.bin pen.id.h pen.mem.h