C++ LibTorch安装与模型部署

进入PyTorch官网，获取对应的LibTorch链接。根据个人设备配置获取对应的LibTorch下载地址，这里我选择 Stable > Linux > LibTorch > C++/Java CUDA 10.2，获得下载地址：

Download here (Pre-cxx11 ABI):
https://download.pytorch.org/libtorch/cu102/libtorch-shared-with-deps-1.11.0%2Bcu102.zip
Download here (cxx11 ABI):
https://download.pytorch.org/libtorch/cu102/libtorch-cxx11-abi-shared-with-deps-1.11.0%2Bcu102.zip

获得下载地址之后，在安装目录(例如${HOME})下执行

wget https://download.pytorch.org/libtorch/cu102/libtorch-cxx11-abi-shared-with-deps-1.11.0%2Bcu102.zip
uzip libtorch-cxx11-abi-shared-with-deps-1.11.0+cu102.zip

(base) cilinyan@amax:~$ cd libtorch
(base) cilinyan@amax:~/libtorch$ tree -d -L 1
.
├── bin      
├── include		# 包含程序需要include的头文件
├── lib			# 包含含链接的共享库
└── share		# 包含必要的CMake配置文件使得find_package(Torch)

4 directories

Next, let’s write a tiny C++ file called dcgan.cpp that includes torch/torch.h and for now simply prints out a three by three identity matrix:

''' main.cpp '''
#include 
#include 

int main() {
  torch::Tensor tensor = torch::eye(3);
  std::cout << tensor << std::endl;
  return 0;
}

To build this tiny application as well as our full-fledged training script later on we’ll use this CMakeLists.txt file:

# CMakeLists.txt
cmake_minimum_required(VERSION 3.16) # 指定 CMake 最低版本

project(myTest CXX)

set(CMAKE_CXX_STANDARD 14) # C++11 会报错
set(CMAKE_CXX_STANDARD_REQUIRED ON)

set(Torch_DIR /data2/cilinyan/libtorch/share/cmake/Torch) # 指定 LibTorch 路径
find_package(Torch REQUIRED) # 指示 CMake 查找 LibTorch 库的构建配置

add_executable(myTest main.cpp)
target_link_libraries(myTest "${TORCH_LIBRARIES}")
set_property(TARGET myTest PROPERTY CXX_STANDARD 14)

简单的编译运行一下：

(base) cilinyan@amax:~/clyan/xshell$ tree 
.
├── CMakeLists.txt
└── main.cpp

0 directories, 2 files

(base) cilinyan@amax:~/clyan/xshell$ cmake .
-- The CXX compiler identification is GNU 8.4.0
-- Check for working CXX compiler: /usr/bin/c++
...
-- Found Torch: /data2/cilinyan/libtorch/lib/libtorch.so  
-- Configuring done
-- Generating done
-- Build files have been written to: /data2/cilinyan/clyan/xshell

(base) cilinyan@amax:~/clyan/xshell$ make
Scanning dependencies of target myTest
[ 50%] Building CXX object CMakeFiles/myTest.dir/main.cpp.o
[100%] Linking CXX executable myTest
[100%] Built target myTest

(base) cilinyan@amax:~/clyan/xshell$ tree -L 1
.
├── CMakeCache.txt
├── CMakeFiles
├── cmake_install.cmake
├── CMakeLists.txt
├── detect_cuda_compute_capabilities.cu
├── detect_cuda_version.cc
├── main.cpp
├── Makefile
└── myTest

1 directory, 8 files

(base) cilinyan@amax:~/clyan/xshell$ ./myTest
 1  0  0
 0  1  0
 0  0  1
[ CPUFloatType{3,3} ]

与 Python 接口一致，基于 C++ 前端的神经网络由称为模块的可重用构建块组成。有一个基本模块类，所有其他模块都从该类派生。在 Python 中，这个类是 torch.nn.Module，而在 C++ 中，它是torch::nn::Module. 除了 forward()实现模块封装的算法的方法外，模块通常包含三种子对象中的任何一种：参数、缓冲区和子模块。

参数和缓冲区以张量的形式存储状态。参数记录梯度，而缓冲区不记录。参数通常是神经网络的可训练权重。缓冲区的示例包括用于批量标准化的均值和方差。为了重用特定的逻辑和状态块，PyTorch API 允许嵌套模块。嵌套模块称为 子模块。

Parameters, buffers and submodules must be explicitly registered. Once registered, methods like parameters() or buffers() can be used to retrieve a container of all parameters in the entire (nested) module hierarchy.类似地，类似的方法to(...)，例如to(torch::kCUDA)将所有参数和缓冲区从 CPU 移动到 CUDA 内存，适用于整个模块层次结构。

Python 模型保存以及测试

import torch
import torch.nn as nn

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.fc = nn.Linear(in_features=5, out_features=3)

    def forward(self, x):
        return self.fc(x)

if __name__ == '__main__':
    torch.manual_seed(5)
    x = torch.tensor([[-9., 2., 3., 4., 5.], [2., 3., 4., 5., 6.]])
    model = Net()
    y = model(x)
    cls = torch.argmax(y, dim=1)
    scripted_module = torch.jit.script(model.eval())
    scripted_module.save("model.pt")
    print(y)
    print(cls)

输出：

tensor([[ 0.8604,  4.0944, -0.2227],
        [ 4.8016,  0.2598,  0.8346]], grad_fn=)
tensor([1, 0])

C++ 模型部署及测试

#include 
#include 
#include 

using namespace std;

int main() {
    // torch::Device device(torch::cuda::is_available() ? torch::kCUDA : torch::kCPU);
    float nums[2][5] = {{-9, 2, 3, 4, 5}, {2,  3, 4, 5, 6}};
    torch::Tensor x = torch::from_blob(nums, {2, 5}, dtype(torch::kFloat));
    auto model = torch::jit::load("model.pt");
    torch::Tensor y = model.forward({x}).toTensor();
    auto cls = torch::argmax(y, 1);
    cout << "x  :" << endl << x   << endl;
    cout << "y  :" << endl << y   << endl;
    cout << "cls:" << endl << cls << endl;
    return 0;
}

输出：

x  :
-9  2  3  4  5
 2  3  4  5  6
[ CPUFloatType{2,5} ]
y  :
 0.8604  4.0944 -0.2227
 4.8016  0.2598  0.8346
[ CPUFloatType{2,3} ]
cls:
 1
 0
[ CPULongType{2} ]