c++ - Raspberry 上的 Libtorch 无法加载 pt 文件但在 ubuntu 上工作

Question

我正在尝试在 Raspberry PI 上使用 libtorch 构建一个 C++ 程序。该程序在 Ubuntu 上运行，但在 Raspberry 上构建时出现以下错误:

error: use of deleted function ‘void torch::jit::script::Module::operator=(const torch::jit::script::Module&)’
In file included from /usr/include/torch/csrc/jit/ir.h:18,
                 from /usr/include/torch/csrc/jit/tracer.h:9,
                 from /usr/include/torch/csrc/autograd/generated/variable_factories.h:8,
                 from /usr/include/torch/csrc/api/include/torch/types.h:7,
                 from /usr/include/torch/script.h:3,
                 from /tmp/tmp.k6618dczxt/src/../include/suvoNet.h:26,
                 from /tmp/tmp.k6618dczxt/src/../include/classifier.h:17,
                 from /tmp/tmp.k6618dczxt/src/classifier.cpp:11:
/usr/include/torch/csrc/jit/script/module.h:319:3: note: declared here
   TH_DISALLOW_COPY_AND_ASSIGN(Module);

这是崩溃的代码:

MyClass::MyClass() {
    try {
        // Deserialize the ScriptModule from a file using torch::jit::load().
        network = torch::jit::load(MODEL_FILE);
    }
    catch (const c10::Error& e) {
        std::cerr << "Error loading the model\n";
        exit(-1);
    }
}

network 声明为私有(private) torch::jit::script::Module network

我使用来自 github 的版本为“1.0.0a0+8322165”的 pyTorch 为 Raspberry (ARM) 构建了 libtorch

Answer 1

TLDR:在 1.6.0 中编译 libtorch 并且工作正常。

如何为 Raspberry 编译 Libtorch 并在我的 C++ 项目中使用？

准备构建

增加RBPi SWAP

首先，如果您有 Raspberry PI 3 或更低版本，则需要增加 SWAP，因为构建会消耗 RAM。

If you have a RBPi 4 or higher with more than 3GB of RAM, skip this step.

修改文件/etc/dphys-swapfile:

CONF_SWAPFILE=2048M

然后调用以下命令来更新更改。

sudo dphys-swapfile setup

安装基础包

安装以下软件包:

sudo apt install build-essential make cmake git python3-pip libatlas-base-dev

Libtorch 需要 CMake>=3.15 才能正确构建，使用 检查 cmake 版本 cmake --version``

如果低于 3.15，请按照以下命令构建更新的版本并删除以前的版本:

wget https://github.com/Kitware/CMake/releases/download/v3.18.0-rc1/cmake-3.18.0-rc1.tar.gz
tar -xzf cmake-3.18.0-rc1.tar.gz
cd cmake<version>
mkdir build
cd build
cmake ..
make
sudo make install

sudo apt remove cmake
sudo ln -s /usr/local/bin/cmake /usr/bin/cmake
sudo ldconfig

从源代码构建 PyTorch 以获得 ARM 的 libtorch 后端

Don't forget to increase the SWAP to 2048M if you don't have 3GB or RAM.

获取所有需要的库:

sudo apt-get update
sudo apt-get install build-essential tk-dev libncurses5-dev libncursesw5-dev libreadline6-dev libdb5.3-dev libgdbm-dev libsqlite3-dev libssl-dev libbz2-dev libexpat1-dev liblzma-dev zlib1g-dev

获取 PyTorch 源代码:

git clone --recursive https://github.com/pytorch/pytorch --branch=release/1.6
cd pytorch

初始化所有子模块:

git submodule update --init --recursive
git submodule update --remote third_party/protobuf # To prevent a bug I had

获取所有需要的库:

sudo apt-get update
sudo apt-get install build-essential tk-dev libncurses5-dev libncursesw5-dev libreadline6-dev libdb5.3-dev libgdbm-dev libsqlite3-dev libssl-dev libbz2-dev libexpat1-dev liblzma-dev zlib1g-dev

获取 PyTorch 源代码:

git clone --recursive https://github.com/pytorch/pytorch --branch=release/1.6
cd pytorch

为build设置环境变量。

将以下行添加到 ~/.bashrc 文件。

export NO_CUDA=1
export NO_DISTRIBUTED=1
export NO_MKLDNN=1 
export NO_NNPACK=1
export NO_QNNPACK=1

以root身份登录，使用.bashrc文件设置环境变量

sudo su
source /home/<user>/.bashrc

安装python依赖

pip3 install setuptools pyyaml numpy

构建并安装 PyTorch，该喝杯咖啡了，这需要一段时间。

Don't forget the -E that forces the environment variables to be used.

sudo -E python3 setup.py install

检查安装是否成功:

cd 
python3
import torch
torch.__version__

用 Torch 构建你的程序

在你的 CMakeLists.txt 中:

cmake_minimum_required(VERSION 2.6)
project(projectName)

set(CMAKE_PREFIX_PATH "/home/pi/pytorch/torch") # Adding the directory where torch as been installed
set(CMAKE_CXX_STANDARD 14) # C14 required to compile Torch
set(CMAKE_CXX_STANDARD_REQUIRED TRUE)
add_compile_definitions(_GLIBCXX_USE_CXX11_ABI=0) # Torch is compiled with CXX11_ABI, so your program needs to be also, or you may have conflicts in some libraries (such as GTest for example)

# Specifying we are using pthread for UNIX systems.
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS} -pthread -Wall")

find_package(Torch REQUIRED)

if(NOT Torch_FOUND)
    message(FATAL_ERROR "Pytorch Not Found!")
endif(NOT Torch_FOUND)

message(STATUS "Pytorch status :")
message(STATUS "    libraries: ${TORCH_LIBRARIES}")
message(STATUS "    Torch Flags: ${TORCH_CXX_FLAGS}")

# Program executable
add_executable(projectName <sources>)

target_link_libraries(projectName PRIVATE pthread dl util ${TORCH_LIBRARIES})