libtorch 常用api

Posted 2021-02-19 无左无右

其实pytorch的函数libtorch都有，只是写法上有些出入。libtorch的官方文档链接：
https://pytorch.org/cppdocs/api/library_root.html
只是官方文档只是类似与函数申明，没有告诉干嘛的，只能通过函数名字猜了。比如我要一个一个函数和已知的一个torch::Tensor变量形状一样，只是填充指定的数值，我记得在哪里看到过的有个full开头的函数，然后我就搜素full，然后找到一个函数full_like好像是我需要的。（见0）

目录

• 调试技巧：
• CMakeLists.txt
• 0.torch::full_like
• 1.创建tensor torch::rand torch::empty torch::ones
• 2.拼接tensor torch::cat

调试技巧：

torch::Tensor box_1 = torch::rand({5,4});
std::cout<<box_1<<std::endl; //可以打印出数值
box_1.print();//可以打印形状

CMakeLists.txt

cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
project(main)
SET(CMAKE_BUILD_TYPE "Debug")

set(CMAKE_PREFIX_PATH "/data_2/everyday/0429/pytorch/torch")
find_package(Torch REQUIRED)

set(CMAKE_PREFIX_PATH "/home/yhl/software_install/opencv3.2")
find_package(OpenCV REQUIRED)

add_executable(main main.cpp)
target_link_libraries(main "${TORCH_LIBRARIES}")

target_link_libraries(main ${OpenCV_LIBS})
set_property(TARGET main PROPERTY CXX_STANDARD 11)

0.torch::full_like

static Tensor at::full_like(const Tensor &self, Scalar fill_value, const TensorOptions &options = {}, c10::optional memory_format = c10::nullopt)
然后就自己试：

#include <iostream>
#include "torch/script.h"
#include "torch/torch.h"
using namespace std;

int main() {   
    torch::Tensor tmp_1 = torch::rand({2,3});
    torch::Tensor tmp_2 = torch::full_like(tmp_1,1);
    
    cout<<tmp_1<<endl;
    cout<<tmp_2<<endl;
}

打印的结果如下：
0.8465 0.5771 0.4404
0.9805 0.8665 0.7807
[ Variable[CPUFloatType]{2,3} ]
1 1 1
1 1 1
[ Variable[CPUFloatType]{2,3} ]

1.创建tensor torch::rand torch::empty torch::ones

1.1 torch::rand

torch::Tensor input = torch::rand({ 1,3,2,3 });

(1,1,.,.) =
0.5943 0.4822 0.6663
0.7099 0.0374 0.9833

(1,2,.,.) =
0.4384 0.4567 0.2143
0.3967 0.4999 0.9196

(1,3,.,.) =
0.2467 0.5066 0.8654
0.7873 0.4758 0.3718
[ Variable[CPUFloatType]{1,3,2,3} ]

1.2 torch::empty

   torch::Tensor a = torch::empty({2, 4});
    std::cout << a << std::endl;

7.0374e+22 5.7886e+22 6.7120e+22 6.7331e+22
6.7120e+22 1.8515e+28 7.3867e+20 9.2358e-01
[ Variable[CPUFloatType]{2,4} ]

1.3 torch::ones

    torch::Tensor a = torch::ones({2, 4});
    std::cout << a<< std::endl;

1 1 1 1
1 1 1 1
[ Variable[CPUFloatType]{2,4} ]

2.拼接tensor torch::cat

2.1 按列拼接

    torch::Tensor a = torch::rand({2,3});
    torch::Tensor b = torch::rand({2,1});
    torch::Tensor cat_1 = torch::cat({a,b},1);//按列拼接--》》前提是行数需要一样

    std::cout<<a<<std::endl;
    std::cout<<b<<std::endl;
    std::cout<<cat_1<<std::endl;

0.3551 0.7215 0.3603
0.1188 0.4577 0.2201
[ Variable[CPUFloatType]{2,3} ]
0.5876
0.3040
[ Variable[CPUFloatType]{2,1} ]
0.3551 0.7215 0.3603 0.5876
0.1188 0.4577 0.2201 0.3040
[ Variable[CPUFloatType]{2,4} ]
注意：如果行数不一样会报如下错误
terminate called after throwing an instance of ‘std::runtime_error‘
what(): invalid argument 0: Sizes of tensors must match except in dimension 1. Got 2 and 4 in dimension 0 at /data_2/everyday/0429/pytorch/aten/src/TH/generic/THTensor.cpp:689

2.2 按行拼接

    torch::Tensor a = torch::rand({2,3});
    torch::Tensor b = torch::rand({1,3});
    torch::Tensor cat_1 = torch::cat({a,b},0);

    std::cout<<a<<std::endl;
    std::cout<<b<<std::endl;
    std::cout<<cat_1<<std::endl;

0.0004 0.7852 0.4586
0.1612 0.6524 0.7655
[ Variable[CPUFloatType]{2,3} ]
0.5999 0.5445 0.2152
[ Variable[CPUFloatType]{1,3} ]
0.0004 0.7852 0.4586
0.1612 0.6524 0.7655
0.5999 0.5445 0.2152
[ Variable[CPUFloatType]{3,3} ]

2.3 其他例子

    torch::Tensor box_1 = torch::rand({5,4});
    torch::Tensor score_1 = torch::rand({5,1});
    torch::Tensor label_1 = torch::rand({5,1});
    torch::Tensor result_1 = torch::cat({box_1,score_1,label_1},1);
    result_1.print();

[Variable[CPUFloatType] [5, 6]]