onnx2ncnn.ipynb

{"nbformat":4,"nbformat_minor":0,"metadata":{"accelerator":"GPU","colab":{"name":"onnx2ncnn.ipynb","provenance":[],"collapsed_sections":["OuIWCvAbdchP","9Fq3tyNWbRDj","MzcrEjJAb3Qe","uiXGXSy8ccux","HMyG8SbKcsF4","Mou49jI5dNir","NgblgO1BdnKZ","pd7D9-xWdvm4","FchcofWNd-1w","0au5V2FceQci","iMGa8mIXecjR"],"toc_visible":true,"mount_file_id":"1e2tDu9o1H6lK7ZDXcRVSPWARu2xGfsQI","authorship_tag":"ABX9TyM7A8oakanYTmzDCOMSzPbH"},"kernelspec":{"display_name":"Python 3","name":"python3"},"language_info":{"name":"python"}},"cells":[{"cell_type":"markdown","metadata":{"id":"XokSpzsWbDjw"},"source":["# python->onnx->ncnn（YOLOv5为例）"]},{"cell_type":"markdown","metadata":{"id":"OuIWCvAbdchP"},"source":["## python->onnx"]},{"cell_type":"markdown","metadata":{"id":"9Fq3tyNWbRDj"},"source":["### 获取YOLOv5并安装所需环境"]},{"cell_type":"code","metadata":{"id":"_6kkkltzWBX3"},"source":["# 克隆yolov5\n","!git clone https://github.com/ultralytics/yolov5  \n","# cd到yolov5路径下\n","%cd yolov5\n","# 安装所需环境\n","%pip install -qr requirements.txt"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"MzcrEjJAb3Qe"},"source":["### 导入torch并检测是否为GPU环境"]},{"cell_type":"code","metadata":{"id":"oYepV859b3lC"},"source":["# 导入torch\n","import torch\n","# 查看是否为GPU环境\n","print(f\"Setup complete. Using torch {torch.__version__} ({torch.cuda.get_device_properties(0).name if torch.cuda.is_available() else 'CPU'})\")"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"uiXGXSy8ccux"},"source":["### 安装转onnx格式的导出环境"]},{"cell_type":"code","metadata":{"id":"cdQ1w_1nXS-p"},"source":["!pip install onnx>=1.7.0  # for ONNX export\n","!pip install coremltools==4.0  # for CoreML export\n","!pip install onnx-simplifier"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"HMyG8SbKcsF4"},"source":["### 使用yolov5自带的export.py进行转换"]},{"cell_type":"code","metadata":{"id":"Fea0mezfh2Ym"},"source":["!python /content/yolov5/models/export.py --weights /content/drive/MyDrive/yolov5s-bug.pt --img 640 --batch 1"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"gj7jPiejdFH7"},"source":["### 对生成的onnx进行优化"]},{"cell_type":"code","metadata":{"id":"PsDjBL2siOTq"},"source":["!python -m onnxsim /content/drive/MyDrive/yolov5s-bug.onnx /content/drive/MyDrive/yolov5s-bug-sim.onnx"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"Mou49jI5dNir"},"source":["## onnx->ncnn"]},{"cell_type":"markdown","metadata":{"id":"NgblgO1BdnKZ"},"source":["### 方案1：使用编译环境"]},{"cell_type":"markdown","metadata":{"id":"pd7D9-xWdvm4"},"source":["#### 安装g++编译环境"]},{"cell_type":"code","metadata":{"id":"Kz_JrTrTmFZP"},"source":["!sudo apt-get install autoconf automake libtool curl make g++ unzip"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"FchcofWNd-1w"},"source":["#### 克隆protobuf并配置make环境"]},{"cell_type":"code","metadata":{"id":"YjdJy_iFmimg"},"source":["cd /content/\n","!git clone https://github.com/protocolbuffers/protobuf.git\n","%cd /content/protobuf\n","!git submodule update --init --recursive\n","!./autogen.sh\n","!./configure\n","!make\n","!make check\n","!sudo make install\n","!sudo ldconfig"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"0au5V2FceQci"},"source":["#### 配置腾讯ncnn编译环境"]},{"cell_type":"code","metadata":{"id":"rdKeb5PumkWM"},"source":["%cd /content\n","!git clone https://github.com/Tencent/ncnn.git\n","%cd /content/ncnn\n","!mkdir -p build\n","%cd /content/ncnn/build\n","!cmake -DNCNN_VULKAN=OFF ..  #vulkan是针对gpu的，如果想要ncnn能调用gpu做推理，那么选项需要打开，设置为ON。\n","!make -j4  #开始编译"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"iMGa8mIXecjR"},"source":["#### cd到onnx2ncnn目录下，使用其进行转化ncnn"]},{"cell_type":"code","metadata":{"id":"rgF5g80vwd5Y"},"source":["%cd /content/ncnn/build/tools/onnx/\n","!./onnx2ncnn /content/drive/MyDrive/yolov5s-bug-sim.onnx /content/drive/MyDrive/model.param /content/drive/MyDrive/model.bin"],"execution_count":null,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"g8r3vhAsrM1-"},"source":["\n","出现以下错误为正常现象，因为YOLOv5的Focus模块需要修改，并进行定义\n","```\n","Unsupported slice step !\n","Unsupported slice step !\n","Unsupported slice step !\n","Unsupported slice step !\n","Unsupported slice step !\n","Unsupported slice step !\n","Unsupported slice step !\n","Unsupported slice step !\n","```\n","\n"]},{"cell_type":"markdown","metadata":{"id":"PLL7E-GpfPTe"},"source":["#### 注意：YOLOv5在此到下一步之前需要对.param文件进行更改\n","\n","```\n","7767517\n","176 200\n","Input                    images                   0 1 images\n","YoloV5Focus              focus                    1 1 images 167\n","Convolution              Conv_41                  1 1 167 168 0=32 1=3 4=1 5=1 6=3456\n","Swish                    Mul_43                   1 1 168 170\n","Convolution              Conv_44                  1 1 170 171 0=64 1=3 3=2 4=1 5=1 6=18432\n","Swish                    Mul_46                   1 1 171 173\n","Split                    splitncnn_0              1 2 173 173_splitncnn_0 173_splitncnn_1\n","Convolution              Conv_47                  1 1 173_splitncnn_1 174 0=32 1=1 5=1 6=2048\n","Swish                    Mul_49                   1 1 174 176\n","Split                    splitncnn_1              1 2 176 176_splitncnn_0 176_splitncnn_1\n","Convolution              Conv_50                  1 1 176_splitncnn_1 177 0=32 1=1 5=1 6=1024\n","```\n","\n"]},{"cell_type":"markdown","metadata":{"id":"aXJgx2KNeqi2"},"source":["#### cd到ncnnoptimize目录下，使用其进行优化"]},{"cell_type":"code","metadata":{"colab":{"base_uri":"https://localhost:8080/"},"id":"mKe7AZTsrV1_","executionInfo":{"status":"ok","timestamp":1618375952893,"user_tz":-480,"elapsed":2939,"user":{"displayName":"Hou Hou","photoUrl":"","userId":"12076772147629762505"}},"outputId":"ecb71c48-5236-4e4a-a161-e0070be0d401"},"source":["%cd /content/ncnn/build/tools/\n","!./ncnnoptimize /content/drive/MyDrive/yolov5s.param  /content/drive/MyDrive/yolov5s.bin /content/drive/MyDrive/new.param /content/drive/MyDrive/new.bin 65536"],"execution_count":null,"outputs":[{"output_type":"stream","text":["create_custom_layer YoloV5Focus\n","model has custom layer, shape_inference skipped\n","model has custom layer, estimate_memory_footprint skipped\n"],"name":"stdout"}]},{"cell_type":"markdown","metadata":{"id":"YNhmEmReuYjY"},"source":["### 方案2：使用转换网站\n","[转换网站](https://convertmodel.com/)"]},{"cell_type":"markdown","metadata":{"id":"6_Sv8uEptRmN"},"source":["## 参考文献\n","[YOLO5实现安卓APP目标检测](https://blog.csdn.net/qq_29750461/article/details/114818608)\n","\n","[yolov5 pt转onnx_yolov5＞onnx&gt;ncnn＞安卓apk](https://blog.csdn.net/weixin_39808893/article/details/111136990)\n","\n","[详细记录u版YOLOv5目标检测ncnn实现](https://zhuanlan.zhihu.com/p/275989233)\n","\n","[指尖上的YOLOv5](https://my.oschina.net/u/4883419/blog/4942998)\n","\n","[YOLO5实现安卓APP目标检测](https://blog.csdn.net/qq_29750461/article/details/114818608)\n","\n","[模型部署之 ONNX ONNXRuntime](https://www.jianshu.com/p/476478c17b8e)\n","\n","[pytorch 转换onnx_Pytorch转NCNN的流程记录](https://blog.csdn.net/weixin_39752434/article/details/112174118?utm_medium=distribute.pc_relevant.none-task-blog-baidujs_title-1&spm=1001.2101.3001.4242)\n","\n","[onnxruntime模型部署流程](https://zhuanlan.zhihu.com/p/261307813)"]}]}