Pytorch_2: libtorch导致OPENCV错误:对‘cv::imread(std::string const&, int)’未定义的引用

#1.错误描述 - opencv 4.0.0 - libtorch1.0.0 (官网下载,非源码安装) - gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6ubuntu1~16.04.11) - cmake version 3.5.1 ``` Scanning dependencies of target frameworks [ 50%] Building CXX object CMakeFiles/frameworks.dir/main.cpp.o [100%] Linking CXX executable frameworks CMakeFiles/frameworks.dir/main.cpp.o：在函数‘main’中： /media/xx/data/cpps/frameworks/main.cpp:26：对‘cv::imread(std::string const&, int)’未定义的引用 collect2: error: ld returned 1 exit status CMakeFiles/frameworks.dir/build.make:159: recipe for target 'frameworks' failed make[3]: *** [frameworks] Error 1 CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/frameworks.dir/all' failed make[2]: *** [CMakeFiles/frameworks.dir/all] Error 2 CMakeFiles/Makefile2:79: recipe for target 'CMakeFiles/frameworks.dir/rule' failed make[1]: *** [CMakeFiles/frameworks.dir/rule] Error 2 Makefile:118: recipe for target 'frameworks' failed make: *** [frameworks] Error 2 ``` #2.错误原因 类似的参考资料 - [github上相同的issue及解答](https://github.com/pytorch/pytorch/issues/14684) - [由于C++类库版本不同导致的OpenCV编译链接错误](https://www.cnblogs.com/zzdyyy/p/8128032.html) - [利用Pytorch的C++前端(libtorch)读取预训练权重并进行预测](https://oldpan.me/archives/pytorch-c-libtorch-inference) 原因如下: ``` # TorchConfig.cmake line 71 # When we build libtorch with the old GCC ABI, dependent libraries must too. if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU") set(TORCH_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=0") endif() ``` (经过诸位大神的分析)TorchConfig.cmake文件中,让ABI=0,具体意思我也不大懂,貌似就关闭了CXX11_ABI,而opencv的东东就在这个里面呀,所以冲突,导致_"对‘cv::imread(std::string const&, int)’未定义的引用"_,上面的第三个参考网址同样提到了这个问题. #3.解决办法 目前还在测试,主要考虑以下两个方面(opencv降版本,libtorch源码安装),测试结果我也会稍后告知. - 使用opencv3.4版本,(不想测试了,我同事说OK!),还是要装这个版本; ``` wget -c https://codeload.github.com/opencv/opencv/zip/3.4.3 wget -c https://codeload.github.com/opencv/opencv_contrib/tar.gz/3.4.3 cmake -D CMAKE_BUILD_TYPE=Releas \ -D OPENCV_EXTRA_MODULES_PATH=/home/download/opencv_contrib-master/modules \ -D WITH_TBB=ON \ -D WITH_CUDA=OFF \ -D WITH_CUBLAS=OFF \ -D BUILD_NEW_PYTHON_SUPPORT=ON \ -D WITH_V4L=ON \ -D INSTALL_C_EXAMPLES=ON \ -D INSTALL_PYTHON_EXAMPLES=ON \ -D PYTHON2_EXECUTABLE=/usr/bin/python2 \ -D PYTHON2_LIBRARY=/usr/lib/python2.7 \ -D PYTHON2_INCLUDE_DIR=/usr/include/python2.7 \ -D PYTHON2_NUMPY_INCLUDE_DIRS=/usr/lib/python2.7/dist-packages/numpy/core/include/ \ -D PYTHON3_EXECUTABLE=/usr/bin/python3 \ -D PYTHON3_LIBRARY=/usr/lib/python3.5 \ -D PYTHON3_INCLUDE_DIR=/usr/include/python3.5 \ -D PYTHON3_NUMPY_INCLUDE_DIRS=/usr/lib/python3.5/dist-packages/numpy/core/include/ \ -D BUILD_EXAMPLES=ON .. ``` - 有人说添加这样的头文件,来源[OpenCV4.0+VS2017下运行程序总是出现未定义标识符](https://ask.csdn.net/questions/718187), 我发现这个方法的时候已经,转战到opencv3.4版本了,所以没有测试. **这种办法后面实测了，行不通呀！** ``` #include "opencv2/imgproc/imgproc_c.h" ``` - 使用源码安装libtorch（**推荐**）,测试```cv::imread```和```cv::imwrite```成功.~~(划掉的部分应该是我搞错了)但是还是有未定义的引用问题发生在其他函数上面,```cv::imshow```,```cv::namewindow```,```cv::waitkey```等.**所以我在重新源码安装libtorch的同时,opencv源码安装3.4.3版本.**~~ 官方源码安装指南 [https://github.com/pytorch/pytorch#from-source](https://github.com/pytorch/pytorch#from-source) 法1.适合1.0rc版本，目前版本见法2 ``` # libtorch的源码安装方法 PYTORCH_COMMIT_ID="8619230" git clone https://github.com/pytorch/pytorch.git cd pytorch && git checkout ${PYTORCH_COMMIT_ID} mkdir build && cd ./build python3 ../tools/build_libtorch.py # 输出目录在./pytorch/pytorch/torch/lib/tmp_install ``` 法2.源码安装pytorch1.2,@2019-06-10 ``` conda activate torch # your conda env conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing #python的依赖项，不行就把conda换pip # Add LAPACK support for the GPU if needed conda install -c pytorch magma-cuda100 # or [magma-cuda92 | magma-cuda100 ] depending on your cuda version git clone --recursive https://github.com/pytorch/pytorch # 下载源码 cd pytorch # if you are updating an existing checkout # 一直执行下面两行，直到没有错误发生，我的渣渣网，搞了几个小时。 git submodule sync git submodule update --init --recursive #正常等待半小时 # 编译安装 python setup.py install #等待半小时 ``` torchvision的源码安装 [Torchvision 源码安装[Ubuntu]](https://www.cnblogs.com/wwwjjjnnn/p/10410547.html) ``` git clone https://github.com/pytorch/vision.git cd vision python setup.py install ``` ## 参考网 1.[Building PyTorch with LibTorch From Source with CUDA Support](https://michhar.github.io/how-i-built-pytorch-gpu/) 2.[github的一些说明](https://github.com/pytorch/pytorch#from-source) 3.[利用Pytorch的C++前端(libtorch)读取预训练权重并进行预测](https://oldpan.me/archives/pytorch-c-libtorch-inference) 4.[https://github.com/pytorch/pytorch#from-source](https://github.com/pytorch/pytorch#from-source)