libtorch 常用api


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
  • 3.torch的切片操作 【select（浅拷贝）】【index_select 深拷贝）】【index 深拷贝】【slice 浅拷贝】

调试技巧：

1
2
3
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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
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：

1
2
3
4
5
6
7
8
9
10
11
12
#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

1
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

1
2
   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

1
2
    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

1
2
3
4
5
6
7
    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

1
2
3
4
5
6
7
    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

1
2
3
4
5
    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]]

3.torch的切片操作 【select（浅拷贝）】【index_select 深拷贝）】【index 深拷贝】【slice 浅拷贝】

select【浅拷贝】只能指定取某一行或某一列
index【深拷贝】只能指定取某一行
index_select【深拷贝】可以按行或按列，指定多行或多列
slice【浅拷贝】 连续的行或列

3.1 inline Tensor Tensor::select(int64_t dim, int64_t index) ；好像只能整2维的。第一个参数是维度，0是取行，1是取 列，第二个参数是索引的序号
3.1.1 select//按行取

1
2
3
4
    torch::Tensor a = torch::rand({2,3});
    std::cout<<a<<std::endl;
    torch::Tensor b = a.select(0,1);//按行取
    std::cout<<b<<std::endl;

0.6201 0.7021 0.1975
0.3080 0.6304 0.1558
[ Variable[CPUFloatType]{2,3} ]
0.3080
0.6304
0.1558
[ Variable[CPUFloatType]{3} ]
3.1.2 select//按列取

1
2
3
4
5
    torch::Tensor a = torch::rand({2,3});
    std::cout<<a<<std::endl;

    torch::Tensor b = a.select(1,1);
    std::cout<<b<<std::endl;

0.8295 0.9871 0.1287
0.8466 0.7719 0.2354
[ Variable[CPUFloatType]{2,3} ]
0.9871
0.7719
[ Variable[CPUFloatType]{2} ]
注意：这里是浅拷贝，就是改变b，同时a的值也会同样的改变
3.1.3 select浅拷贝

1
2
3
4
5
6
7
8
9
    torch::Tensor a = torch::rand({2,3});
    std::cout<<a<<std::endl;
   
    torch::Tensor b = a.select(1,1);
    std::cout<<b<<std::endl;
   
    b[0] = 0.0;
    std::cout<<a<<std::endl;
    std::cout<<b<<std::endl;

0.0938 0.2861 0.0089
0.3481 0.5806 0.3711
[ Variable[CPUFloatType]{2,3} ]
0.2861
0.5806
[ Variable[CPUFloatType]{2} ]
0.0938 0.0000 0.0089
0.3481 0.5806 0.3711
[ Variable[CPUFloatType]{2,3} ]
0.0000
0.5806
[ Variable[CPUFloatType]{2} ]
可以看到，b[0] = 0.0;然后a，b的对应位置都为0了。浅拷贝！！

3.2 inline Tensor Tensor::index_select(Dimname dim, const Tensor & index) //同样的，dim0表示按行，1表示按列 index表示取的行号或者列号，这里
比较奇怪，index一定要是toType(torch::kLong)这种类型的。还有一个奇怪的地方是我准备用数组导入tensor的，发现idx全是0，原因未知

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
 torch::Tensor a = torch::rand({2,6});
    std::cout<<a<<std::endl;
slice

     torch::Tensor idx = torch::empty({4}).toType(torch::kLong);
     idx[0]=0;
     idx[1]=2;
     idx[2]=4;
     idx[3]=1;

//    int idx_data[4] = {1,3,2,4};
//    torch::Tensor idx = torch::from_blob(idx_data,{4}).toType(torch::kLong);//idx全是0  ?????????????????

    std::cout<<idx<<std::endl;

    torch::Tensor b = a.index_select(1,idx);
    std::cout<<b<<std::endl;

0.4956 0.5028 0.0863 0.9464 0.6714 0.5348
0.3523 0.2245 0.0924 0.7088 0.6913 0.2237
[ Variable[CPUFloatType]{2,6} ]
0
2
4
1
[ Variable[CPULongType]{4} ]
0.4956 0.0863 0.6714 0.5028
0.3523 0.0924 0.6913 0.2245
[ Variable[CPUFloatType]{2,4} ]

3.2.2 index_select深拷贝

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
    torch::Tensor a = torch::rand({2,6});
    std::cout<<a<<std::endl;


     torch::Tensor idx = torch::empty({4}).toType(torch::kLong);
     idx[0]=0;
     idx[1]=2;
     idx[2]=4;
     idx[3]=1;

//    int idx_data[4] = {1,3,2,4};
//    torch::Tensor idx = torch::from_blob(idx_data,{4}).toType(torch::kLong);

    std::cout<<idx<<std::endl;

    torch::Tensor b = a.index_select(1,idx);
    std::cout<<b<<std::endl;

    b[0][0]=0.0;
    std::cout<<a<<std::endl;
    std::cout<<b<<std::endl;

0.6118 0.6078 0.5052 0.9489 0.6201 0.8975
0.0901 0.2040 0.1452 0.6452 0.9593 0.7454
[ Variable[CPUFloatType]{2,6} ]
0
2
4
1
[ Variable[CPULongType]{4} ]
0.6118 0.5052 0.6201 0.6078
0.0901 0.1452 0.9593 0.2040
[ Variable[CPUFloatType]{2,4} ]
0.6118 0.6078 0.5052 0.9489 0.6201 0.8975
0.0901 0.2040 0.1452 0.6452 0.9593 0.7454
[ Variable[CPUFloatType]{2,6} ]
0.0000 0.5052 0.6201 0.6078
0.0901 0.1452 0.9593 0.2040
[ Variable[CPUFloatType]{2,4} ]

3.3 index inline Tensor Tensor::index(TensorList indices)
这个函数实验下来，只能按行取，且是深拷贝

1
2
3
4
5
6
7
8
9
10
11
12
13
14
    torch::Tensor a = torch::rand({2,6});
    std::cout<<a<<std::endl;


    torch::Tensor idx_1 = torch::empty({2}).toType(torch::kLong);
    idx_1[0]=0;
    idx_1[1]=1;


    torch::Tensor bb = a.index(idx_1);
    bb[0][0]=0;

    std::cout<<bb<<std::endl;
    std::cout<<a<<std::endl;

0.1349 0.8087 0.2659 0.3364 0.0202 0.4498
0.4785 0.4274 0.9348 0.0437 0.6732 0.3174
[ Variable[CPUFloatType]{2,6} ]
0.0000 0.8087 0.2659 0.3364 0.0202 0.4498
0.4785 0.4274 0.9348 0.0437 0.6732 0.3174
[ Variable[CPUFloatType]{2,6} ]
0.1349 0.8087 0.2659 0.3364 0.0202 0.4498
0.4785 0.4274 0.9348 0.0437 0.6732 0.3174
[ Variable[CPUFloatType]{2,6} ]
3.4 slice inline Tensor Tensor::slice(int64_t dim, int64_t start, int64_t end, int64_t step) //dim0表示按行取，1表示按列取，从start开始，到end（不含）结束
可以看到结果，是浅拷贝！！！

1
2
3
4
5
6
7
8
9
10
11
12
 torch::Tensor a = torch::rand({2,6});
    std::cout<<a<<std::endl;

    torch::Tensor b = a.slice(0,0,1);
    torch::Tensor c = a.slice(1,0,3);

    b[0][0]=0.0;

    std::cout<<b<<std::endl;
    std::cout<<c<<std::endl;

    std::cout<<a<<std::endl;

0.8270 0.7952 0.3743 0.7992 0.9093 0.5945
0.3764 0.8419 0.7977 0.4150 0.8531 0.9207
[ Variable[CPUFloatType]{2,6} ]
0.0000 0.7952 0.3743 0.7992 0.9093 0.5945
[ Variable[CPUFloatType]{1,6} ]
0.0000 0.7952 0.3743
0.3764 0.8419 0.7977
[ Variable[CPUFloatType]{2,3} ]
0.0000 0.7952 0.3743 0.7992 0.9093 0.5945
0.3764 0.8419 0.7977 0.4150 0.8531 0.9207
[ Variable[CPUFloatType]{2,6} ]