目录

一、通过索引获取值

 二、通过索引设置值

三、掩码操作

---

在PyTorch C++ API（libtorch）中对张量进行索引的方式与Python API的方式很相似。诸如None / ... / integer / boolean / slice / tensor的索引类型在C++ API里同样有效，这样就可以很方便的实现Python代码与C++代码的转换。主要的不同是将Python API里对张量的“[ ]”操作符转换成了以下形式：

torch::Tensor::index ( )    // 获取值

torch::Tensor::index_put_ ( )   // 设置值

有关官方文档请看这里。但是官方文档只简单地给了转换方式，没有具体例子有时还不太好理解。下面通过举例说明libtorch与pytorch中的向量索引/切片的方式，左边为Python方式，右边为C++方式：

一、通过索引获取值

1、tensor[Ellipsis, ...]  -->  tensor.index({Ellipsis, "..."})

import torch

a = torch.linspace(1,27,27).reshape(3, 3, 3)
print(a)
c = a[..., 2]
print(c)

#===================运行结果===============#
tensor([[[ 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.]]])
tensor([[ 3.,  6.,  9.],
        [12., 15., 18.],
        [21., 24., 27.]])

#include "iostream"
#include "torch/script.h"
int main()
{
    torch::Tensor a = torch::linspace(1, 27, 27).reshape({3, 3, 3});
    std::cout << a << std::endl;
    at::Tensor b = a.index({"...", 2});
    std::cout << b << std::endl;

    return 0;
}

/****************输出结果******************/
(1,.,.) = 
  1  2  3
  4  5  6
  7  8  9

(2,.,.) = 
  10  11  12
  13  14  15
  16  17  18

(3,.,.) = 
  19  20  21
  22  23  24
  25  26  27
[ CPUFloatType{3,3,3} ]
  3   6   9
 12  15  18
 21  24  27
[ CPUFloatType{3,3} ]

 2、tensor[1, 2] --> tensor.index({1, 2})

import torch

a = torch.linspace(1,27,27).reshape(3, 3, 3)
print(a)
c = a[1, 2]
print(c)

#===================运行结果=================#
tensor([[[ 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.]]])
tensor([16., 17., 18.])

#include "iostream"
#include "torch/script.h"

int main()
{
    torch::Tensor a = torch::linspace(1, 27, 27).reshape({3, 3, 3});
    std::cout << a << std::endl;
    at::Tensor b = a.index({1, 2});
    std::cout << b << std::endl;

    return 0;
}
/*****************运行结果***************/
(1,.,.) = 
  1  2  3
  4  5  6
  7  8  9

(2,.,.) = 
  10  11  12
  13  14  15
  16  17  18

(3,.,.) = 
  19  20  21
  22  23  24
  25  26  27
[ CPUFloatType{3,3,3} ]
 16
 17
 18
[ CPUFloatType{3} ]

3、tensor[1::2] --> tensor.index({Slice(1, None, 2)})

import torch

a = torch.linspace(1, 6, 6)
print(a)
c = a[1::2]
print(c)
#==================运行结果==================#
tensor([1., 2., 3., 4., 5., 6.])
tensor([2., 4., 6.])

#include "iostream"
#include "torch/script.h"
using namespace torch::indexing;
int main()
{
    torch::Tensor a = torch::linspace(1, 6, 6);
    std::cout << a << std::endl;
    at::Tensor b = a.index({Slice(1, None, 2)});
    std::cout << b << std::endl;

    return 0;
}
/*******************运行结果*********************/
 1
 2
 3
 4
 5
 6
[ CPUFloatType{6} ]
 2
 4
 6
[ CPUFloatType{3} ]

3.33、tensor[..., 1:] --> tensor.index({"...", Slice(1)})

import torch

a = torch.linspace(1,27,27).reshape(3, 3, 3)
b = a[..., 1:]
print(b)
#===============运行结果===================#
tensor([[[ 2.,  3.],
         [ 5.,  6.],
         [ 8.,  9.]],

        [[11., 12.],
         [14., 15.],
         [17., 18.]],

        [[20., 21.],
         [23., 24.],
         [26., 27.]]])

#include "iostream"
#include "torch/script.h"
using namespace torch::indexing;
int main()
{
    torch::Tensor a = torch::linspace(1, 27, 27).reshape({3, 3, 3});
    torch::Tensor b = a.index({"...", Slice(1)});
    std::cout << b << std::endl;

    return 0;
}
/******************运行结果**********************/
(1,.,.) = 
  2  3
  5  6
  8  9

(2,.,.) = 
  11  12
  14  15
  17  18

(3,.,.) = 
  20  21
  23  24
  26  27
[ CPUFloatType{3,3,2} ]

3.66、tensor[..., :2] --> tensor.index({"...", Slice({None, 2})})

#include "iostream"
#include "torch/script.h"
using namespace torch::indexing;
int main()
{
    float a[2][3][3] = {
   {
   {1,2,2},{3,4,4},{5,6,6}},{
   {1,2,2},{3,4,4},{5,6,6}}};
    at::Tensor b = at::from_blob(a, {2, 3, 3}, at::kFloat);
    at::Tensor c = b.index({"...", Slice({None, 2})});  // 留下最后一维的前两列，相当于Python中的b[..., :2]
    auto d = c.sizes();
    std::cout << b << std::endl;
    std::cout << c << std::endl;
    std::cout << d << std::endl;

    return 0;
}
/********************运行结果********************/
(1,.,.) = 
  1  2  2
  3  4  4
  5  6  6

(2,.,.) = 
  1  2  2
  3  4  4
  5  6  6
[ CPUFloatType{2,3,3} ]
(1,.,.) = 
  1  2
  3  4
  5  6

(2,.,.) = 
  1  2
  3  4
  5  6
[ CPUFloatType{2,3,2} ]
[2, 3, 2]

 4、tensor[torch.tensor([1, 2])] --> tensor.index({torch::tensor({1, 2})})

import torch

a = torch.linspace(1,4,4)
b = torch.tensor([0, 1, 3, 2])
c = a[b]
print(a)
print(c)
#===============运行结果===============#
tensor([1., 2., 3., 4.])
tensor([1., 2., 4., 3.])

#include "iostream"
#include "torch/script.h"
using namespace torch::indexing;
int main()
{
    torch::Tensor a = torch::linspace(1, 4, 4);
    torch::Tensor b = torch::tensor({0, 1, 3, 2});
    torch::Tensor c = a.index({b});
    std::cout << a << std::endl;
    std::cout << b << std::endl;

    return 0;
}
/*******************运行结果********************/
 1
 2
 3
 4
[ CPUFloatType{4} ]
 0
 1
 3
 2
[ CPULongType{4} ]

 二、通过索引设置值

1、tensor[1, 2] = 1 --> tensor.index_put_({1, 2}, 1)

import torch

a = torch.linspace(1,4,4).reshape(2, 2)
print(a)
a[1, 1] = 100
print(a)
#==================运行结果=====================#
tensor([[1., 2.],
        [3., 4.]])
tensor([[  1.,   2.],
        [  3., 100.]])

#include "iostream"
#include "torch/script.h"
using namespace torch::indexing;
int main()
{
    torch::Tensor a = torch::linspace(1, 4, 4).reshape({2, 2});
    std::cout << a << std::endl;
    a.index_put_({1, 1}, 100);
    std::cout << a << std::endl;

    return 0;
}
/***************运行结果****************/
 1  2
 3  4
[ CPUFloatType{2,2} ]
   1    2
   3  100
[ CPUFloatType{2,2} ]

 2、tensor[Ellipsis, ...] = 1 --> tensor.index_put_({Ellipsis, "..."}, 1)

import torch

a = torch.linspace(1,4,4).reshape(2, 2)
print(a)
a[..., 1] = 100
print(a)
#====================运行结果=====================#
tensor([[1., 2.],
        [3., 4.]])
tensor([[  1., 100.],
        [  3., 100.]])

#include "iostream"
#include "torch/script.h"
using namespace torch::indexing;
int main()
{
    torch::Tensor a = torch::linspace(1, 4, 4).reshape({2, 2});
    std::cout << a << std::endl;
    a.index_put_({"...", 1}, 100);
    std::cout << a << std::endl;

    return 0;
}
/***************运行结果****************/
 1  2
 3  4
[ CPUFloatType{2,2} ]
   1  100
   3  100
[ CPUFloatType{2,2} ]

 3、tensor[torch.tensor([1, 2])] = 1 --> tensor.index_put_({torch::tensor({1, 2})}, 1)

import torch

a = torch.linspace(1,4,4)
b = torch.tensor([0, 2])
print(a)
a[b] = 100
print(a)
#===============运行结果==================#
tensor([1., 2., 3., 4.])
tensor([100.,   2., 100.,   4.])

#include "iostream"
#include "torch/script.h"
using namespace torch::indexing;
int main()
{
    torch::Tensor a = torch::linspace(1, 4, 4);
    torch::Tensor b = torch::tensor({0, 2});
    std::cout << a << std::endl;
    a.index_put_({b}, 100);
    std::cout << a << std::endl;

    return 0;
}
/*****************运行结果*****************/
 1
 2
 3
 4
[ CPUFloatType{4} ]
 100
   2
 100
   4
[ CPUFloatType{4} ]

三、掩码操作

按理说这部分也算索引的一部分，为什么要拿出来单独说呐？因为他比较特殊。首先看下例子吧。

在PyTorch中，我们使用Python的切片方式可以方便地通过掩码向量或矩阵来获取想要的tensor区域，比如以下操作：

import torch

a = torch.randint(0, 9,(4,4))          # 创建shape为4*4，值为[0,9]的随机整数的tensor
b = torch.tensor([0, 1, 0, 1]).bool()  # 创建bool值向量，最终的结果是对应行向量的取舍
c = torch.tensor([0, 1, 0, 1])         # 注意这里不是bool值,最终的结果只是按行索引
a1 = a[b]  # 掩码操作
a2 = a[c]  # 索引操作
print('a:', a)
print('b:', b)
print('c:', c)
print('a1:', a1)
print('a2:', a2)

#=================运行结果===================#
a: tensor([[8, 0, 4, 8],
           [5, 7, 7, 3],
           [0, 5, 8, 1],
           [3, 3, 2, 1]])
b: tensor([False,  True, False,  True])
c: tensor([0, 1, 0, 1])
a1: tensor([[5, 7, 7, 3],
            [3, 3, 2, 1]])
a2: tensor([[8, 0, 4, 8],
            [5, 7, 7, 3],
            [8, 0, 4, 8],
            [5, 7, 7, 3]])

在PyThon的C++API即LibTorch中，也有类似的操作，举例如下：

#include "iostream"
#include "ATen/Parallel.h"
#include "torch/script.h"
#include <c10/core/ScalarType.h>
using namespace torch::indexing;
int main()
{
    torch::Tensor a = torch::randint(0, 9, {4,4});   // 初始化
    torch::Tensor b = at::tensor({0,1,0,1}).toType(at::kBool);  // bool型掩码向量，注意与非bool型的结果区别
    torch::Tensor c = at::tensor({0,1,0,1}).toType(at::kLong);  // 非bool型按行索引(这里的索引用tensor必须是long, byte or bool类型的)
    torch::Tensor a1 = a.index({b});  // 掩码操作
    torch::Tensor a2 = a.index({c});  // 按行索引
    std::cout << a << std::endl;
    std::cout << b << std::endl;
    std::cout << c << std::endl;
    std::cout << a1 << std::endl;
    std::cout << a2 << std::endl;

    return 0;
}

//==================运行结果====================//
 7  5  8  0
 1  0  1  2
 0  6  6  3
 2  1  5  3
[ CPUFloatType{4,4} ]
 0
 1
 0
 1
[ CPUBoolType{4} ]
 0
 1
 0
 1
[ CPULongType{4} ]
 1  0  1  2
 2  1  5  3
[ CPUFloatType{2,4} ]
 7  5  8  0
 1  0  1  2
 7  5  8  0
 1  0  1  2
[ CPUFloatType{4,4} ]

所以一定要注意0和1在bool型与非bool型情况下对索引结果的影响！

另外，LibTorch还有一个masked_select()方法，该方法是按照对应位置上的掩码值(bool值)来抽取值，先看定义：

inline Tensor Tensor::masked_select(const Tensor & mask)

其中mask必须是与要操作的tensor同shape或拥有相同数量元素的tensor，不满足条件就会自动复制补全，直至与目标tensor具有相同数量的元素。

举例如下：

#include "iostream"
#include "torch/script.h"
#include <c10/core/ScalarType.h>

using namespace torch::indexing;
int main()
{
    torch::Tensor a = torch::randint(0, 9, {4,4});   // 初始化
    torch::Tensor b = a > 5;   // 获取满足条件的bool矩阵
    torch::Tensor c = a.masked_select({b});  // 按照bool矩阵来取舍对应的值
    std::cout << a << std::endl;
    std::cout << b << std::endl;
    std::cout << c << std::endl;

    return 0;
}
//====================运行结果=======================//
 5  2  8  8
 1  0  3  0
 2  2  4  4
 8  4  7  5
[ CPUFloatType{4,4} ]
 0  0  1  1
 0  0  0  0
 0  0  0  0
 1  0  1  0
[ CPUBoolType{4,4} ]
 8
 8
 8
 7
[ CPUFloatType{4} ]

注：这里当时测试用的具体libtorch版本记不清了，不同版本间的操作可能会有差异