libtorch c++调用 （五）Linux下的调用

1、Linux libtorch-cpu

libtorch下载地址：https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-1.5.1%2Bcpu.zip

后来使用中发现libtorch1.5.1版本有内存泄漏问题，具体是openmp的内存泄漏造成的，所以后来实际使用的是libtorch1.6.0.

在linux系统下新建一个文件夹如：pytorch_test

文件夹下新建一个文件：main.cpp，文件内容如下：

#include 
#include 
 
using namespace std;
 
int main()
{
    torch::Tensor tensor = torch::eye(3);
    std::cout << tensor << std::endl;
    cout << "Hello World!" << endl;
    return 0;
}

文件夹下新建一个文件：CMakeLists.txt文件内容如下：（因为我把下载的libtorch1.5.1解压到共享文件夹下的vmsharefolders文件夹下，所以下面配置Torch_DIR是：/mnt/hgfs/vmsharefolders/libtorch/share/cmake/Torch）

cmake_minimum_required(VERSION 3.5)
 
project(libtorch_demo LANGUAGES CXX)
 
set(CMAKE_CXX_STANDARD 11)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
set(Torch_DIR /mnt/hgfs/vmsharefolders/libtorch/share/cmake/Torch)
find_package(Torch REQUIRED)
 
add_executable(libtorch_demo main.cpp)
target_link_libraries(libtorch_demo "${TORCH_LIBRARIES}")
set_property(TARGET libtorch_demo PROPERTY CXX_STANDARD 17)

在文件夹下开启终端，执行如下命令：

mkdir build
cd build
cmake ..
make 

分别执行上述四条命令，就会编译完程序，执行效果如下：

wmz@ubuntu:~/Desktop/mypractice/pytorch_test/build$ cmake ..
-- The CXX compiler identification is GNU 5.4.0
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Looking for C++ include pthread.h
-- Looking for C++ include pthread.h - found
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Failed
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE  
-- Found torch: /mnt/hgfs/vmsharefolders/libtorch/lib/libtorch.so  
-- Configuring done
-- Generating done
-- Build files have been written to: /home/wmz/Desktop/mypractice/pytorch_test/build
wmz@ubuntu:~/Desktop/mypractice/pytorch_test/build$ ls
CMakeCache.txt  CMakeFiles  cmake_install.cmake  Makefile
wmz@ubuntu:~/Desktop/mypractice/pytorch_test/build$ make 
Scanning dependencies of target libtorch_demo
[ 50%] Building CXX object CMakeFiles/libtorch_demo.dir/main.cpp.o
[100%] Linking CXX executable libtorch_demo
[100%] Built target libtorch_demo
wmz@ubuntu:~/Desktop/mypractice/pytorch_test/build$ ls
CMakeCache.txt  CMakeFiles  cmake_install.cmake  libtorch_demo  Makefile
wmz@ubuntu:~/Desktop/mypractice/pytorch_test/build$ ./libtorch_demo 
 1  0  0
 0  1  0
 0  0  1
[ CPUFloatType{3,3} ]
Hello World!

参考：UBUNTU 18.04 INSTALL QT 5.12.9和LIBTORCH 1.5.1

参考：CMakeLists.txt 语法介绍与实例演练

问题：RuntimeError: [enforce fail at CPUAllocator.cpp:64] . DefaultCPUAllocator: can’t allocate memory: you tried to allocate 

参考：我与导师的聊天记录

首先感谢参考博主的热心贡献，我遇到的问题与他的不同，但是同样是内存不足的问题，我是用的是libtorch1.5.1-cpu的c++ inference部分。因为要模拟部署，所以在一台电脑上编译，在另一台电脑上执行。编译的虚拟机配置2核4G，测试的虚拟机1核2G，然后就报错内存不足。

当我把测试虚拟机的内存调到2核4G时程序就可以正常运行了。

2、Linux libtorch-gpu

libtorch GPU版本编译时报错：

/usr/bin/ld: cannot find -lCUDA_cublas_device_LIBRARY-NOTFOUND
collect2: error: ld returned 1 exit status

查找原因，说是cmake版本太低，我用的是cmake3.10，然后就卸载了cmake3.10，安装了cmake3.18.2.

参考：NOTFOUND CUDA_cublas_device_LIBRARY，Ubuntu16.04安装torch遇到的错误

参考：ubuntu卸载/更新Cmake

通过升级cmake版本上面的问题解决了,但是还有报错：

In function `main':
test_db.cpp:(.text+0x39ef): undefined reference to `cv::imread(std::string const&, int)'
test_db.cpp:(.text+0x4aec): undefined reference to `cv::imwrite(std::string const&, cv::_InputArray const&, std::vector > const&)'
CMakeFiles/dbnet_demo.dir/test_db.cpp.o: In function `cv::Mat::Mat(cv::Size_, int, void*, unsigned long)':
test_db.cpp:(.text._ZN2cv3MatC2ENS_5Size_IiEEiPvm[_ZN2cv3MatC5ENS_5Size_IiEEiPvm]+0x14b): undefined reference to `cv::error(int, std::string const&, char const*, char const*, int)'
test_db.cpp:(.text._ZN2cv3MatC2ENS_5Size_IiEEiPvm[_ZN2cv3MatC5ENS_5Size_IiEEiPvm]+0x219): undefined reference to `cv::error(int, std::string const&, char const*, char const*, int)'

原因及分析参考：Opencv4 和 libtorch不兼容的问题

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

解决办法：

方法一、降低opencv版本，比如opencv3.4就可以通过。

方法二、源码编译libtorch（推荐）。

已经测试通过：使用官方libtorch1.6.0-cu102+opencv3.4.2静态库编译测试程序通过（系统ubuntu18.04,gcc 7.4）。

看执行时间确认使用了GPU。

$ ./dbnet_demo 
ok
 file_names.size():5
0001004.jpg
./data/0001004.jpg
[W TensorIterator.cpp:918] Warning: Mixed memory format inputs detected while calling the operator. The operator will output contiguous tensor even if some of the inputs are in channels_last format. (function operator())
The run time is: 0.440735s
CLOCKS_PER_SEC is: 1000000
0001001.jpg
./data/0001001.jpg
The run time is: 0.103798s
CLOCKS_PER_SEC is: 1000000
0001002.jpg
./data/0001002.jpg
The run time is: 0.101153s
CLOCKS_PER_SEC is: 1000000
0001003.jpg
./data/0001003.jpg
The run time is: 0.100296s
CLOCKS_PER_SEC is: 1000000
0000000.jpg
./data/0000000.jpg
The run time is: 0.098885s
CLOCKS_PER_SEC is: 1000000

在第二台GPU电脑上编译报错(系统centos7.8,gcc4.8.5)：

原因是gcc版本太低。

-- Caffe2: CUDA detected: 10.1
-- Caffe2: CUDA nvcc is: /usr/local/cuda-10.1/bin/nvcc
-- Caffe2: CUDA toolkit directory: /usr/local/cuda-10.1
-- Caffe2: Header version is: 10.1
-- Found cuDNN: v7.6.4  (include: /usr/local/cuda-10.1/include, library: /usr/local/cuda-10.1/lib64/libcudnn.so)
-- Autodetected CUDA architecture(s):  7.5
-- Added CUDA NVCC flags for: -gencode;arch=compute_75,code=sm_75
-- Configuring done
-- Generating done
-- Build files have been written to: /home/wmz/Documents/dbnet/dbnet/build2
[wintone@localhost build2]$ make
[ 25%] Building CXX object CMakeFiles/dbnet.dir/DBnetAPI.cpp.o
c++: error: unrecognized command line option ‘-std=c++14’
make[2]: *** [CMakeFiles/dbnet.dir/DBnetAPI.cpp.o] Error 1
make[1]: *** [CMakeFiles/dbnet.dir/all] Error 2
make: *** [all] Error 2

在第三台GPU电脑上测试通过（系统centos7.2, gcc6.3）

中间遇到过的问题是，这台电脑上没有安装cudnn，报错：

Could NOT find CUDNN(missing: CUDNN_LIBRARY_PATH CUDNN_INCLUDE_PATH)

原因是官方Caffe2链接了cudnn,所以编译时需要本地有cudnn安装，或者编译器可以找到cudnn。

我的解决办法是直接在命令行解决(因为我下载的linux cudnn库直接解压到test目录下，默认文件夹名为cuda)：

cmake .. -DCUDNN_INCLUDE_DIR=/home/wmz/test/cuda/include -DCUDNN_LIBRARY=/home/wmz/test/cuda/lib64