libTorch的安装

在官网https://pytorch.org/选择与cuda版本对应的libtorch下载，然后解压即可

TorchVision编译

升级CMake

wget https://cmake.org/files/v3.20/cmake-3.20.0-linux-x86_64.tar.gz
tar -xzvf cmake-3.20.0-linux-x86_64.tar.gz
#解压出来的包，将其放在 /opt 目录下，其他目录也可以，主要别以后不小心删了
sudo mv cmake-3.20.0-linux-x86_64 /opt/cmake-3.20.0
#建立软链接
sudo ln -sf /opt/cmake-3.20.0/bin/*  /usr/bin/
#查看 cmake 版本
cmake --version

编译安装TorchVision

下载：https://github.com/pytorch/vision

注意：因为Ubuntu默认是python3.5，故最好下载Vision0.6.1及以下版本 如果python的版本不对，是python2.7的话，需要通过下面设置优先级来设置默认的Python为python3.5:

sudo update-alternatives --install /usr/bin/python python /usr/bin/python2.7 1 
sudo update-alternatives --install /usr/bin/python python /usr/bin/python3.5 2

编译安装

mkdir build
cd build
cmake -DCMAKE_PREFIX_PATH=/home/alex/github/libtorch  -DWITH_CUDA=ON -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_STANDARD=14 ..
make -j4
sudo make install

报错

cmake的时候：

links to target "Python3::Python" but the target was not found. Perhaps a find_package() call is missing for an IMPORTED

解决：这个问题很奇怪，明明已经是默认python3了，还是说找不到。再次安装python3就好了 sudo apt install python3-dev

make的时候

shufflenetv2.cpp:83:33: error: call of overloaded ‘channel_shuffle(at::Tensor&, int)’ is ambiguous

解决：这是多个代码冲突了，将报错地方改为：vision::models::channel_shuffle()

代码实例

CMakeLists.txt

cmake_minimum_required(VERSION 3.19)
project(LibTorchHelloWorld)

set(CMAKE_CXX_STANDARD 14)

set(CMAKE_PREFIX_PATH "/usr/local/lib;/home/alex/github/libtorch")

find_package(TorchVision REQUIRED)
find_package(Torch REQUIRED)
include_directories(${TORCH_INCLUDE_DIRS})

add_executable(LibTorchHelloWorld main.cpp)
target_compile_features(LibTorchHelloWorld PUBLIC cxx_range_for)
target_link_libraries(LibTorchHelloWorld  ${TORCH_LIBRARIES} TorchVision::TorchVision)
set_property(TARGET LibTorchHelloWorld PROPERTY CXX_STANDARD 14)

main.cpp

#include <iostream>
#include <torch/torch.h>
#include <torchvision/vision.h>
#include <torchvision/models/resnet.h>


// Define a new Module.
struct Net : torch::nn::Module {
    Net() {
        // Construct and register two Linear submodules.
        fc1 = register_module("fc1", torch::nn::Linear(784, 64));
        fc2 = register_module("fc2", torch::nn::Linear(64, 32));
        fc3 = register_module("fc3", torch::nn::Linear(32, 10));
    }

    // Implement the Net's algorithm.
    torch::Tensor forward(torch::Tensor x) {
        // Use one of many tensor manipulation functions.
        x = torch::relu(fc1->forward(x.reshape({x.size(0), 784})));
        x = torch::dropout(x, /*p=*/0.5, /*train=*/is_training());
        x = torch::relu(fc2->forward(x));
        x = torch::log_softmax(fc3->forward(x), /*dim=*/1);
        return x;
    }

    // Use one of many "standard library" modules.
    torch::nn::Linear fc1{nullptr}, fc2{nullptr}, fc3{nullptr};
};


int main()
{
    auto model = vision::models::ResNet18();
    model->eval();

    // Create a random input tensor and run it through the model.
    auto in = torch::rand({1, 3, 10, 10});
    auto out = model->forward(in);

    std::cout << out.sizes();

    if (torch::cuda::is_available()) {
        // Move model and inputs to GPU
        model->to(torch::kCUDA);
        auto gpu_in = in.to(torch::kCUDA);
        auto gpu_out = model->forward(gpu_in);

        std::cout << gpu_out.sizes();
    }
}