1. LibTorch安装

下载cuda版本为11.3的LibTorch安装包并解压即完成安装：

# If you need cpu support, please replace "cu113" with "cpu" in the URL below.
wget https://download.pytorch.org/libtorch/nightly/cu113/libtorch-shared-with-deps-latest.zip
unzip libtorch-shared-with-deps-latest.zip
1
2
3

如果需要cpu版本的安装包按照上面注释方法替换即可。

2. C++调用PyTorch模型

2.1 Python中保存tensor数据

保存tensor的函数：

def save_tensor(data_tensor: Tensor, name: str):
    print("[python] %s: "%name, data_tensor)
    f = io.BytesIO()
    torch.save(data_tensor, f, _use_new_zipfile_serialization=True)
    with open('/home/chenxin/peanut/DenseTNT/test_cpp/dense_tnt_infer/data/%s.pt'%name, "wb") as out_f:
        # Copy the BytesIO stream to the output file
        out_f.write(f.getbuffer())
1
2
3
4
5
6
7

2.2 C++加载tensor并调用模型

首先需要将PyTorch模型转换为C++支持的TorchScript模型，具体步骤以后找个时间再写一篇笔记，这边先略过。

C++调用TorchScript模型代码：

#include <torch/torch.h>
#include "torch/script.h"
#include <iostream>
#include <string>
#include <chrono>

std::vector<char> get_the_bytes(std::string filename)
{
  std::ifstream input(filename, std::ios::binary);
  std::vector<char> bytes(
      (std::istreambuf_iterator<char>(input)),
      (std::istreambuf_iterator<char>()));

  input.close();
  return bytes;
}

// 加载tensor数据
torch::Tensor GetTensor(const std::string &path)
{
  std::vector<char> f = get_the_bytes(path);
  torch::IValue x = torch::pickle_load(f);
  torch::Tensor my_tensor = x.toTensor();
  return my_tensor;
}
int main()
{
  torch::Device device(torch::kCPU);
  if (torch::cuda::is_available())
  {
    device = torch::Device(torch::kCUDA, 0);
  }

  // 读取推理用例数据
  torch::Tensor vector_data = GetTensor("test_cpp/dense_tnt_infer/data/vector_data.pt");
  torch::Tensor vector_mask = GetTensor("test_cpp/dense_tnt_infer/data/vector_mask.pt");
  torch::Tensor traj_polyline_mask = GetTensor("test_cpp/dense_tnt_infer/data/traj_polyline_mask.pt");
  torch::Tensor map_polyline_mask = GetTensor("test_cpp/dense_tnt_infer/data/map_polyline_mask.pt");
  torch::Tensor cent_point = GetTensor("test_cpp/dense_tnt_infer/data/cent_point.pt");

  torch::set_num_threads(1);
  std::vector<torch::jit::IValue> torch_inputs;
  torch_inputs.push_back(std::move(vector_data.to(device)));
  torch_inputs.push_back(std::move(vector_mask.to(device)));
  torch_inputs.push_back(std::move(traj_polyline_mask.to(device)));
  torch_inputs.push_back(std::move(map_polyline_mask.to(device)));
  torch_inputs.push_back(std::move(cent_point.to(device)));
  
  // 加载torch script模型
  torch::jit::script::Module torch_script_model = torch::jit::load("models.densetnt.1/model_save/model.16_script.bin", device);

  for (int i = 0; i < 100; ++i)
  {
    auto t1 = std::chrono::high_resolution_clock::now();
    auto torch_output_tuple = torch_script_model.forward(torch_inputs);
    auto t2 = std::chrono::high_resolution_clock::now();
    std::chrono::duration<double, std::milli> ms_double = t2 - t1;
    std::cout << ms_double.count() << "ms\n";
    // std::cout << torch_output_tuple << std::endl;
  }
  return 0;
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62

3. 编译执行C++推理用例

3.1 编写CMakeLists

需要安装cuda和cudnn才能进行cmake；

在C++推理代码同级目录下，创建文件 CMakeLists.txt，写入：

cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
project(dense_tnt_infer)  # 调用模型的c++文件名称
set(CMAKE_PREFIX_PATH "/home/chenxin/libtorch")  # 这里填解压libtorch时的路径

find_package(Torch REQUIRED)

add_executable(${PROJECT_NAME} "dense_tnt_infer.cc")  # 调用模型的c++文件名称
target_link_libraries(${PROJECT_NAME} ${TORCH_LIBRARIES})
set_property(TARGET ${PROJECT_NAME} PROPERTY CXX_STANDARD 14)
1
2
3
4
5
6
7
8
9

3.2 编译并执行用例

在CMakeLists.txt同级目录下执行命令：

$ mkdir build
$ cd build
$ cmake ..
$ make 
$ ./dense_tnt_infer.cc
1
2
3
4
5

输出：

...
...
12.7836ms
12.7527ms
13.0666ms
14.3305ms
13.804ms
14.1567ms
13.3143ms
13.0827ms
13.0853ms
13.5594ms
1
2
3
4
5
6
7
8
9
10
11
12

推理时长还需要通过修改代码再优化。