算法产品化---在ArmNN上运行ONNX

主要参考文档：https://developer.arm.com/technologies/machine-learning-on-arm/developer-material/how-to-guides/configuring-the-arm-nn-sdk-build-environment-for-onnx/setup-and-download-libraries

为了确保arm网站上面不做大的修改，我这边将文档保存在如下的路径，可以自行取阅：

https://download.youkuaiyun.com/download/putiancaijunyu/10892970

注意几点：

Setup and download libraries

这个阶段下载protobuf.git的时候，会找不到v3.5.0这边branch，建议切换到3.6.x上。

git clone https://github.com/google/protobuf.git

执行完成这个命令之后,切换branch

git checkout 3.6.x

 

Build the Compute Library

如果提示没有安装aarch64-linux-gnu-g++命令，那就是电脑没有安装toolchain，可以考虑执行：

sudo apt-get install g++-aarch64-linux-gnu

 

Build the Google protobuf library

请注意将branch切换到3.6.x，不然后边编译armnn的时候会出错。

如果出现如下错误：

root@JunyuSS3:/home/admin1/armnn-onnx/protobuf# ./autogen.sh 
+ mkdir -p third_party/googletest/m4
+ autoreconf -f -i -Wall,no-obsolete
./autogen.sh: 37: ./autogen.sh: autoreconf: not found

这是因为没有安装automake工具。可以考虑执行如下命令：

 sudo apt-get install autoconf automake libtool

 

Generate the ONNX protobuf source files

在这边步骤中，主要需要切换onnx.git的commit Id，不然在编译armnn的时候会出现如下的问题：

如果出现如下的错误：

[ 74%] Building CXX object CMakeFiles/armnnOnnxParser.dir/src/armnnOnnxParser/OnnxParser.cpp.o
/home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp: In function ‘armnn::TensorInfo armnnOnnxParser::{anonymous}::ToTensorInfo(const onnx::ValueInfoProto&)’:
/home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp:217:101: error: invalid conversion from ‘google::protobuf::int32 {aka int}’ to ‘onnx::TensorProto::DataType {aka onnx::TensorProto_DataType}’ [-fpermissive]
                                 onnx::TensorProto::DataType_Name(info.type().tensor_type().elem_type()) %
                                                                                                     ^
In file included from /home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.hpp:11:0,
                 from /home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp:5:
/home/admin1/armnn-onnx/onnx/onnx/onnx.pb.h:1787:38: note:   initializing argument 1 of ‘static const string& onnx::TensorProto::DataType_Name(onnx::TensorProto::DataType)’
   static inline const ::std::string& DataType_Name(DataType value) {
                                      ^
/home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp: In member function ‘void armnnOnnxParser::OnnxParser::SetupInfo(const google::protobuf::RepeatedPtrField<onnx::ValueInfoProto>*)’:
/home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp:640:85: error: invalid conversion from ‘google::protobuf::int32 {aka int}’ to ‘onnx::TensorProto::DataType {aka onnx::TensorProto_DataType}’ [-fpermissive]
         m_TensorsInfo[tensor.name()].m_dtype = tensor.type().tensor_type().elem_type();
                                                                                     ^
/home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp: In member function ‘void armnnOnnxParser::OnnxParser::ParseConstant(const onnx::NodeProto&)’:
/home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp:853:78: error: invalid conversion from ‘google::protobuf::int32 {aka int}’ to ‘onnx::TensorProto::DataType {aka onnx::TensorProto_DataType}’ [-fpermissive]
     CHECK_VALID_DATATYPE(node.name(), onnxTensor.name(), onnxTensor.data_type(), onnx::TensorProto::FLOAT);
                                                                              ^
/home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp:48:35: note: in definition of macro ‘CHECK_VALID_DATATYPE’
 CheckValidDataType({__VA_ARGS__}, ACTUAL, #__VA_ARGS__, NODE, TENSOR, CHECK_LOCATION())
                                   ^
/home/admin1/armnn-onnx/armnn/src/armnnOnnxParser/OnnxParser.cpp:24:6: note:   initializing argument 2 of ‘void armnnOnnxParser::{anonymous}::CheckValidDataType(std::initializer_list<onnx::TensorProto_DataType>, onnx::TensorProto::DataType, const char*, std::__cxx11::string, std::__cxx11::string, const armnn::CheckLocation&)’
 void CheckValidDataType(std::initializer_list<onnx::TensorProto::DataType> validInputTypes,
      ^
CMakeFiles/armnnOnnxParser.dir/build.make:62: recipe for target 'CMakeFiles/armnnOnnxParser.dir/src/armnnOnnxParser/OnnxParser.cpp.o' failed
make[2]: *** [CMakeFiles/armnnOnnxParser.dir/src/armnnOnnxParser/OnnxParser.cpp.o] Error 1
CMakeFiles/Makefile2:68: recipe for target 'CMakeFiles/armnnOnnxParser.dir/all' failed
make[1]: *** [CMakeFiles/armnnOnnxParser.dir/all] Error 2
Makefile:127: recipe for target 'all' failed
make: *** [all] Error 2

原因是onnx后面将elem_type的类型从DateType变成int32。

https://github.com/ARM-software/armnn/issues/113，这个给出了一个修改意见，建设将onnx的版本更新到0a4d5abdf4939ab0842a5eadcc16a3bf0738f901

使用的命令可以是：

cd onnx; 
git checkout 0a4d5abdf4939ab0842a5eadcc16a3bf0738f901

然后在重新编译onnx