Ubuntu 16.04上源码编译和安装pytorch教程，并编写C Demo CMakeLists.txt

本文首发于个人博客https://kezunlin.me/post/54e7a3d8/，欢迎阅读最新内容！

tutorial to compile and use pytorch on ubuntu 16.04

PyTorch for Python

install pytorch from anaconda

    conda info --envs
    conda activate py35

    # newest version
    # 1.1.0 pytorch/0.3.0 torchvision
    conda install pytorch torchvision cudatoolkit=9.0 -c pytorch

    # old version [NOT]
    # 0.4.1 pytorch/0.2.1 torchvision
    conda install pytorch=0.4.1 cuda90 -c pytorch

output

    The following NEW packages will be INSTALLED:

      pytorch            pytorch/linux-64::pytorch-1.1.0-py3.5_cuda9.0.176_cudnn7.5.1_0
      torchvision        pytorch/linux-64::torchvision-0.3.0-py35_cu9.0.176_1

download from channel pytorch will cost much time!

下载pytorch/linux-64::pytorch-1.1.0-py3.5_cuda9.0.176_cudnn7.5.1_0速度非常慢！

install pytorch from tsinghua

add tsinghua pytorch channels

    conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
    # for legacy win-64
    conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/peterjc123/
    conda config --set show_channel_urls yes

使用anaconda官方pytorch源非常慢，用清华源代替。

see tsinghua anaconda

cat ～/.condarc

channels:
  - https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
  - defaults

install pytorch from tsinghua

    conda create --name torch python==3.7
    conda activate torch

    conda install -y pytorch torchvision
    conda install -y scikit-learn scikit-image pandas matplotlib pillow opencv

The following NEW packages will be INSTALLED:

      pytorch            anaconda/cloud/pytorch/linux-64::pytorch-1.1.0-py3.5_cuda9.0.176_cudnn7.5.1_0
      torchvision        anaconda/cloud/pytorch/linux-64::torchvision-0.3.0-py35_cu9.0.176_1

test pytorch

    import torch
    print(torch.__version__)
    '1.1.0'

or

    python -c 'import torch; print(torch.cuda.is_available())'
    True

pre-trained models

pre-trained model saved to /home/kezunlin/.cache/torch/checkpoints/

    Downloading: "https://download.pytorch.org/models/shufflenetv2_x0.5-f707e7126e.pth" to /home/kezunlin/.cache/torch/checkpoints/shufflenetv2_x0.5-f707e7126e.pth

PyTorch for C

download LibTorch

download from LibTorch

compile from source

compile pytorch

    # method 1
    git clone --recursive https://github.com/pytorch/pytorch
    cd pytorch

    # method 2, if you are updating an existing checkout
    git clone https://github.com/pytorch/pytorch
    cd pytorch 
    git submodule sync
    git submodule update --init --recursive

check tags

    git tag -l 

    v0.4.0
    v0.4.1
    v1.0.0
    v1.0.1
    v1.0rc0
    v1.0rc1
    v1.1.0

now compile

    git checkout v1.1.0

    # method 1: offical build will generate lots of errors
    #python setup.py install 

     # method 2: normal make
    mkdir build && cd build && cmake-gui ..

with configs

    BUILD_PYTHON OFF

be sure to use stable version 1.1.0 from here instead of latest version 20190724 (unstable version 1.2.0)

because error will occurs when load models.

• for 1.1.0:

      std::shared_ptr module = torch::jit::load("./model.pt");

• for latest 1.2.0

      torch::jit::script::Module module = torch::jit::load("./model.pt");

configure output

    ******** Summary ********
    General:
      CMake version         : 3.5.1
      CMake command         : /usr/bin/cmake
      System                : Linux
      C   compiler          : /usr/bin/c  
      C   compiler id       : GNU
      C   compiler version  : 5.4.0
      BLAS                  : MKL
      CXX flags             :  -fvisibility-inlines-hidden -fopenmp -O2 -fPIC -Wno-narrowing -Wall -Wextra -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math
      Build type            : Release
      Compile definitions   : ONNX_ML=1;ONNX_NAMESPACE=onnx_torch;USE_GCC_ATOMICS=1;HAVE_MMAP=1;_FILE_OFFSET_BITS=64;HAVE_SHM_OPEN=1;HAVE_SHM_UNLINK=1;HAVE_MALLOC_USABLE_SIZE=1
      CMAKE_PREFIX_PATH     : 
      CMAKE_INSTALL_PREFIX  : /usr/local

      TORCH_VERSION         : 1.1.0
      CAFFE2_VERSION        : 1.1.0
      BUILD_CAFFE2_MOBILE   : ON
      BUILD_ATEN_ONLY       : OFF
      BUILD_BINARY          : OFF
      BUILD_CUSTOM_PROTOBUF : ON
        Link local protobuf : ON
      BUILD_DOCS            : OFF
      BUILD_PYTHON          : OFF
      BUILD_CAFFE2_OPS      : ON
      BUILD_SHARED_LIBS     : ON
      BUILD_TEST            : OFF
      INTERN_BUILD_MOBILE   : 
      USE_ASAN              : OFF
      USE_CUDA              : ON
        CUDA static link    : OFF
        USE_CUDNN           : ON
        CUDA version        : 9.2
        cuDNN version       : 7.1.4
        CUDA root directory : /usr/local/cuda
        CUDA library        : /usr/local/cuda/lib64/stubs/libcuda.so
        cudart library      : /usr/local/cuda/lib64/libcudart.so
        cublas library      : /usr/local/cuda/lib64/libcublas.so
        cufft library       : /usr/local/cuda/lib64/libcufft.so
        curand library      : /usr/local/cuda/lib64/libcurand.so
        cuDNN library       : /usr/local/cuda/lib64/libcudnn.so
        nvrtc               : /usr/local/cuda/lib64/libnvrtc.so
        CUDA include path   : /usr/local/cuda/include
        NVCC executable     : /usr/local/cuda/bin/nvcc
        CUDA host compiler  : /usr/bin/cc
        USE_TENSORRT        : OFF
      USE_ROCM              : OFF
      USE_EIGEN_FOR_BLAS    : ON
      USE_FBGEMM            : OFF
      USE_FFMPEG            : OFF
      USE_GFLAGS            : OFF
      USE_GLOG              : OFF
      USE_LEVELDB           : OFF
      USE_LITE_PROTO        : OFF
      USE_LMDB              : OFF
      USE_METAL             : OFF
      USE_MKL               : OFF
      USE_MKLDNN            : OFF
      USE_NCCL              : ON
        USE_SYSTEM_NCCL     : OFF
      USE_NNPACK            : ON
      USE_NUMPY             : ON
      USE_OBSERVERS         : ON
      USE_OPENCL            : OFF
      USE_OPENCV            : OFF
      USE_OPENMP            : ON
      USE_TBB               : OFF
      USE_PROF              : OFF
      USE_QNNPACK           : ON
      USE_REDIS             : OFF
      USE_ROCKSDB           : OFF
      USE_ZMQ               : OFF
      USE_DISTRIBUTED       : ON
        USE_MPI             : ON
        USE_GLOO            : ON
        USE_GLOO_IBVERBS    : OFF
      NAMEDTENSOR_ENABLED   : OFF
      Public Dependencies  : Threads::Threads
      Private Dependencies : qnnpack;nnpack;cpuinfo;/usr/lib/x86_64-linux-gnu/libnuma.so;fp16;/usr/lib/openmpi/lib/libmpi_cxx.so;/usr/lib/openmpi/lib/libmpi.so;gloo;aten_op_header_gen;foxi_loader;rt;gcc_s;gcc;dl
    Configuring done

install pytorch

now compile and install

    make -j8
    sudo make install

output

    Install the project...
    -- Install configuration: "Release"
    -- Old export file "/usr/local/share/cmake/Caffe2/Caffe2Targets.cmake" will be replaced.  Removing files [/usr/local/share/cmake/Caffe2/Caffe2Targets-release.cmake].
    -- Set runtime path of "/usr/local/bin/protoc" to "$ORIGIN"
    -- Old export file "/usr/local/share/cmake/Gloo/GlooTargets.cmake" will be replaced.  Removing files [/usr/local/share/cmake/Gloo/GlooTargets-release.cmake].
    -- Set runtime path of "/usr/local/lib/libonnxifi_dummy.so" to "$ORIGIN"
    -- Set runtime path of "/usr/local/lib/libonnxifi.so" to "$ORIGIN"
    -- Set runtime path of "/usr/local/lib/libfoxi_dummy.so" to "$ORIGIN"
    -- Set runtime path of "/usr/local/lib/libfoxi.so" to "$ORIGIN"
    -- Set runtime path of "/usr/local/lib/libc10.so" to "$ORIGIN"
    -- Set runtime path of "/usr/local/lib/libc10_cuda.so" to "$ORIGIN:/usr/local/cuda/lib64"
    -- Set runtime path of "/usr/local/lib/libthnvrtc.so" to "$ORIGIN:/usr/local/cuda/lib64/stubs:/usr/local/cuda/lib64"
    -- Set runtime path of "/usr/local/lib/libtorch.so" to "$ORIGIN:/usr/local/cuda/lib64:/usr/lib/openmpi/lib"
    -- Set runtime path of "/usr/local/lib/libcaffe2_detectron_ops_gpu.so" to "$ORIGIN:/usr/local/cuda/lib64"
    -- Set runtime path of "/usr/local/lib/libcaffe2_observers.so" to "$ORIGIN:/usr/local/cuda/lib64"

pytorch 1.1.0

compile and install will cost more than 2 hours

lib install to /usr/local/lib/libtorch.so

cmake install to /usr/local/share/cmake/Torch

C example

load pytorch model in c see load pytorch model in c

cpp

#include  // One-stop header.

#include 
#include 

int main(int argc, const char* argv[]) {
  if (argc != 2) {
    std::cerr << "usage: example-app \n";
    return -1;
  }

  // Deserialize the ScriptModule from a file using torch::jit::load().
  std::shared_ptr module = torch::jit::load(argv[1]);

  assert(module != nullptr);
  std::cout << "ok\n";
  
  // Create a vector of inputs.
std::vector inputs;
inputs.push_back(torch::ones({1, 3, 224, 224}));

// Execute the model and turn its output into a tensor.
at::Tensor output = module->forward(inputs).toTensor();

std::cout << output.slice(/*dim=*/1, /*start=*/0, /*end=*/5) << '\n';
}

CMakeLists.txt

cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
project(custom_ops)

# /usr/local/share/cmake/Torch
find_package(Torch REQUIRED)
MESSAGE( [Main] " TORCH_INCLUDE_DIRS = ${TORCH_INCLUDE_DIRS}") 
MESSAGE( [Main] " TORCH_LIBRARIES = ${TORCH_LIBRARIES}")  
include_directories(${TORCH_INCLUDE_DIRS})

add_executable(example-app example-app.cpp)
target_link_libraries(example-app "${TORCH_LIBRARIES}")
set_property(TARGET example-app PROPERTY CXX_STANDARD 11)

output

    Found torch: /usr/local/lib/libtorch.so  
    [Main] TORCH_INCLUDE_DIRS = /usr/local/include;/usr/local/include/torch/csrc/api/include
    [Main] TORCH_LIBRARIES = torch;torch_library;/usr/local/lib/libc10.so;/usr/local/cuda/lib64/stubs/libcuda.so;/usr/local/cuda/lib64/libnvrtc.so;/usr/local/cuda/lib64/libnvToolsExt.so;/usr/local/cuda/lib64/libcudart.so;/usr/local/lib/libc10_cuda.so
    [TOLOWER] ALGORITHM_TARGET = algorithm

make

    mkdir build 
    cd build && cmake-gui ..
    make -j8
> set `Torch_DIR` to `/home/kezunlin/program/libtorch/share/cmake/Torch`
> auto-set `Torch_DIR` to `/usr/local/share/cmake/Torch`

run

    ./example-app model.pt
    -0.2698 -0.0381  0.4023 -0.3010 -0.0448

errors and solutions

compile errors with libtorch

• Build simple c example-cpp using Libtorch fails on arm with undefined reference to c10::Error::Error

@soumith

You might be building libtorch with a compiler that is incompatible with the compiler building your final app.

For example, you built libtorch with gcc 4.9.2 and your final app with gcc 5.1, and the C ABI between both of them is not the same, so you are seeing linker errors like these

• issues-linking-with-libtorch-c-11-abi

@christianperone

    if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
      set(TORCH_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=0")
    endif()
> Which forces GCC to use the old C  11 ABI.

@ smth

we have that flag set because we build with gcc 4.9.x, which only has the old ABI.

In GCC 5.1, the ABI for std::string was changed, and binaries compiling with gcc >= 5.1 are not ABI-compatible with binaries build with gcc < 5.1 (like pytorch) unless you set that flag.

resons and solutions

• Reasons: LibTorch compiled with GCC-4.9.X (only has the old ABI), and binaries compiling with gcc >= 5.1 are not ABI-compatible
• Solution: compile pytorch from source instead of using LibTroch downloaded from the website.

runtime errors with pytorch

errors

    /usr/local/lib/libopencv_imgcodecs.so.3.1.0: undefined reference to `TIFFReadRGBAStrip@LIBTIFF_4.0'

which means opencv link against libtiff 4.0.6

ldd check

    ldd /usr/local/lib/libopencv_imgcodecs.so.3.1.0
        linux-vdso.so.1 =>  (0x00007ffc92ffc000)
        libopencv_imgproc.so.3.1 => /usr/local/lib/libopencv_imgproc.so.3.1 (0x00007f32afbca000)
        libjpeg.so.8 => /usr/local/lib/libjpeg.so.8 (0x00007f32af948000)
        libpng12.so.0 => /lib/x86_64-linux-gnu/libpng12.so.0 (0x00007f32af723000)
        libtiff.so.5 => /usr/lib/x86_64-linux-gnu/libtiff.so.5 (0x00007f32af4ae000)

when compile opencv-3.1.0, cmake find /usr/lib/x86_64-linux-gnu/libtiff.so.5

locate libtiff

    locate libtiff.so

    /home/kezunlin/anaconda3/envs/py35/lib/libtiff.so
    /home/kezunlin/anaconda3/envs/py35/lib/libtiff.so.5
    /home/kezunlin/anaconda3/envs/py35/lib/libtiff.so.5.4.0
    /home/kezunlin/anaconda3/lib/libtiff.so
    /home/kezunlin/anaconda3/lib/libtiff.so.5
    /home/kezunlin/anaconda3/lib/libtiff.so.5.4.0
    /home/kezunlin/anaconda3/pkgs/libtiff-4.0.10-h2733197_2/lib/libtiff.so
    /home/kezunlin/anaconda3/pkgs/libtiff-4.0.10-h2733197_2/lib/libtiff.so.5
    /home/kezunlin/anaconda3/pkgs/libtiff-4.0.10-h2733197_2/lib/libtiff.so.5.4.0
    /opt/MATLAB/R2016b/bin/glnxa64/libtiff.so.5
    /opt/MATLAB/R2016b/bin/glnxa64/libtiff.so.5.0.5
    /usr/lib/x86_64-linux-gnu/libtiff.so
    /usr/lib/x86_64-linux-gnu/libtiff.so.5
    /usr/lib/x86_64-linux-gnu/libtiff.so.5.2.4

It seems that my OpenCV was compiled against libtiff 4, but I have libtiff 5, how to solve this problem?

re-compile opencv-3.1.0 again, new errors occursee here

    CMake Error: The following variables are used in this project, but they are set to NOTFOUND.
    Please set them or make sure they are set and tested correctly in the CMake files:
    CUDA_nppi_LIBRARY (ADVANCED)
        linked by target "opencv_cudev" in directory /home/kezunlin/program/opencv-3.1.0/modules/cudev
        linked by target "opencv_cudev" in directory /home/kezunlin/program/opencv-3.1.0/modules/cudev
        linked by target "opencv_test_cudev" in directory /home/kezunlin/program/opencv-3.1.0/modules/cudev/test

solutions:

    WITH_CUDA OFF
    WITH_VTK OFF
    WITH_TIFF OFF
    BUILD_PERF_TESTS OFF 

for python2, use default /usr/bin/python2.7

for python3, NOT USE anaconda version

编译的过程中，尽量避免使用anaconda目录下的lib

install libwebp

    sudo apt-get -y install libwebp-dev

Reference

• pytorch
• pytorch github
• deep_learning_60min_blitz
• pytorch-tutorial
• pytorch notebooks
• pytorch-beginner
• pytorch cppdocs

History

• 20190626: created.

Copyright

• Post author: kezunlin
• Post link: https://kezunlin.me/post/54e7a3d8/
• Copyright Notice: All articles in this blog are licensed under CC BY-NC-SA 3.0 unless stating additionally.