Mxnet C++接口安装以及收集到的资料整理

Mxnet C++接口

---

install

C++接口需要从源码编译安装 跟着教程来也很简单（无脑复制粘贴）

ubuntu安装地址

官方的STEP1里安装了OpenBLAS数学库（后面STEP3是三个命令选一个，我选For a CPU-only build with OpenBLAS math library run:）

跟随官网教程做到STEP3时候需要注意的点：

• 运行cmake时候要添加参数-USE_CPP_PACKAGE=1，这样才会开启C++接口，比如这样:

rm -rf build
mkdir -p build && cd build
cmake -GNinja \
    -DUSE_CUDA=OFF \
    -DUSE_MKL_IF_AVAILABLE=OFF \
    -DCMAKE_BUILD_TYPE=Release \
    -USE_CPP_PACKAGE=1 \
..
ninja

• 如果是笔记本推荐直接用CPU的跑就行了，如果选GPU，我没折腾过不知道会不会有什么坑。。。我选择的是For a CPU-only build with OpenBLAS math library run:因为STEP1里安装了
• 完成编译之后，就会生成.so和.a两个文件，文件位置可能在lib里（我跟着mac的教程是生成在lib里）也有可能在build文件夹里（官网说ubuntu的是在这里）。然后可以继续跟着教程安装python或者其他语言接口，要不要安装python接口看你自己了（我自己是不安装了python语言接口，要是安装：pip卸载之前安装的，然后跟着教程安装，这样能使用tensorrt，当然这种需要在前面STEP3那儿开启CUDA支持）

---

关于mxnet的C/C++接口的资料

先熟悉下cmake，寻找头文件以及导入第三方库或者自己写的静态动态库都需要cmake，推荐IDE：Clion

• 官方C++接口的Tutorials
• 官方C++多分类的demo
• 官方C多分类demo
• 官方C++的YOLO3
• 民间C的SSD

C和C++的区别： 在官方C++接口的Tutorials中提到了C++的前向传播接口是没有测试过线程安全的，如果需要多线程请使用C接口（但是这个貌似是很久之前的文档，不知道现在的版本C++接口是否线程安全，我正在纠结这个）

---

我的cmakelist.txt文件

一共两个，第一个自己写的，比较简单，第二个是YOLO3里的（可能有错）
自行根据自己的项目、环境进行修改

cmake_minimum_required(VERSION 3.15)
project(HellowMxnet)

set(CMAKE_CXX_STANDARD 14)

find_package(OpenCV REQUIRED)

file(GLOB SOURCES "${PROJECT_SOURCE_DIR}/src/*.cpp")

include_directories(
        ${OpenCV_INCLUDE_DIRS}
        ${PROJECT_SOURCE_DIR}/include
        /Users/rainweic/DeepLearning/mxnet/cpp-package/include
        /Users/rainweic/DeepLearning/mxnet/include
)

link_directories(
        ${OpenCV_LIBRARY_DIRS}
        /Users/rainweic/DeepLearning/mxnet/lib
)

#add_library(HellowMxnet SHARED ${SOURCES})
add_executable(HellowMxnet ${SOURCES})

target_link_libraries(HellowMxnet
        ${OpenCV_LIBS}
        mxnet)

cmake_minimum_required(VERSION 3.1)

project(YOLOV3)

set(CMAKE_CXX_STANDARD 11)

set(MXNET_ROOT “/Users/rainweic/DeepLearning/mxnet”)

# — MXNet
include_directories(
  ${MXNET_ROOT}/include
  ${MXNET_ROOT}/cpp-package/include
  ${MXNET_ROOT}/3rdparty/dlpack/include
  ${MXNET_ROOT}/3rdparty/dmlc-core/include
  ${MXNET_ROOT}/3rdparty/mkldnn/include
  ${MXNET_ROOT}/3rdparty/mshadow
  ${MXNET_ROOT}/3rdparty/tvm/nnvm/include
  ${CMAKE_CURRENT_LIST_DIR}/src
)

SET(CMAKE_FIND_LIBRARY_PREFIXES “lib”)
find_library(MXNET_LIBS NAMES mxnet HINTS “${MXNET_ROOT}/lib”)
message(STATUS MXNET_LIBS: ${MXNET_LIBS})
message(STATUS CURRENT_DIR: ${CMAKE_CURRENT_LIST_DIR})
set(LINKER_LIBS ${MXNET_LIBS})

# — OpenCV
find_package(OpenCV COMPONENTS core highgui imgproc imgcodecs)
if(NOT OpenCV_FOUND) # if not OpenCV 3.x, then imgcodecs are not found
  message(STATUS “OpenCV imgcodecs missing”)
  find_package(OpenCV REQUIRED COMPONENTS core highgui imgproc)
endif()
include_directories(SYSTEM ${OpenCV_INCLUDE_DIRS})
list(APPEND LINKER_LIBS ${OpenCV_LIBS})
message(STATUS “ OpenCV_LIBS=${OpenCV_LIBS}”)
message(STATUS "OpenCV ${OpenCV_VERSION} found (${OpenCV_CONFIG_PATH})")

# — Targets and Installation
set(CMAKE_INSTALL_PREFIX ${CMAKE_CURRENT_LIST_DIR})

add_executable(gluoncv-detect src/detect.cpp)
target_link_libraries(gluoncv-detect ${LINKER_LIBS})

set(EXECS gluoncv-detect)
install(TARGETS ${EXECS} DESTINATION install)
if(APPLE)
  install(FILES ${MXNET_LIBS} DESTINATION install/lib)
elseif(MSVC)
  install(FILES ${MXNET_ROOT}/lib/libmxnet.dll DESTINATION install)
  file(GLOB OPENBLAS_FILES $ENV{OpenBLAS_HOME}/*.dll $ENV{OpenBLAS_HOME}/bin/*.dll)
  file(GLOB OPENCV_FILES ${_OpenCV_LIB_PATH}/*core*.dll ${_OpenCV_LIB_PATH}/*highgui*.dll ${_OpenCV_LIB_PATH}/*imgproc*.dll ${_OpenCV_LIB_PATH}/*imgcodecs*.dll ${_OpenCV_LIB_PATH}/*videoio*.dll)
  install(FILES ${OPENBLAS_FILES} DESTINATION install)
  install(FILES ${OPENCV_FILES} DESTINATION install)
else()
 install(FILES ${MXNET_LIBS} DESTINATION install)
endif()