libtorch c++ 调用（四）api函数示例

仅记录我用到过的一些API，后面会逐渐补充。

1、大于> 小于< 运算 （这是一个很好的 “二值化“ 操作）

torch::Tensor a = torch::randn({3,4});
cout << ( a > 0.1) << endl;
结果：
tensor a 的值为：
-0.4388 -0.8234  0.3935  0.0000
 0.0121  1.0354  0.0000  1.5286
 0.1590  2.7148 -0.0737 -0.5168
[ CPUFloatType{3,4} ]

 0  0  1  0
 0  1  0  1
 1  1  0  0
[ CPUBoolType{3,4} ]

2、Tensor::slice，取tensor数据块中roi区域的数据，不复制，内存共享

切片操作方便获取数据中的某些维度，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！！！

 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;

输出：

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} ]

3、narrow narrow_copy

inline Tensor Tensor::narrow(int64_t dim, int64_t start, int64_t length) const
inline Tensor Tensor::narrow_copy(int64_t dim, int64_t start, int64_t length) const

    torch::Tensor a = torch::rand({4,6});
    torch::Tensor b = a.narrow(0,1,2);
    torch::Tensor c = a.narrow_copy(0,1,2);

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

结果：

0.9812 0.4205 0.4169 0.2412 0.8769 0.9873
0.8052 0.0312 0.9901 0.5065 0.6344 0.3408
0.0182 0.6933 0.9375 0.8675 0.5201 0.9521
0.5119 0.3880 0.1117 0.5413 0.8203 0.4163
[ Variable[CPUFloatType]{4,6} ]
0.8052 0.0312 0.9901 0.5065 0.6344 0.3408
0.0182 0.6933 0.9375 0.8675 0.5201 0.9521
[ Variable[CPUFloatType]{2,6} ]
0.8052 0.0312 0.9901 0.5065 0.6344 0.3408
0.0182 0.6933 0.9375 0.8675 0.5201 0.9521
[ Variable[CPUFloatType]{2,6} ]

4.squeeze() unsqueeze()

inline Tensor Tensor::squeeze() const//不加参数的，把所有为1的维度都压缩
inline Tensor Tensor::squeeze(int64_t dim)const//加参数的，指定哪个维度压缩
inline Tensor & Tensor::squeeze_() const //暂时不知道啥区别
inline Tensor & Tensor::squeeze_(int64_t dim) const //暂时不知道啥区别
4.1 squeeze()

4.2 squeeze(int64_t dim) 指定压缩哪个维度

    torch::Tensor a = torch::rand({1,1,3});
    std::cout<<a<<std::endl;
    
    torch::Tensor b = a.squeeze();
    std::cout<<b<<std::endl;
    
    torch::Tensor c = a.squeeze(0);
    std::cout<<c<<std::endl;
    
    torch::Tensor d = a.squeeze(1);
    std::cout<<d<<std::endl;
    
    torch::Tensor e = a.squeeze(2);
    std::cout<<e<<std::endl;

结果：

(1,.,.) =
0.8065 0.1287 0.8073
[ Variable[CPUFloatType]{1,1,3} ]
0.8065
0.1287
0.8073
[ Variable[CPUFloatType]{3} ]
0.8065 0.1287 0.8073
[ Variable[CPUFloatType]{1,3} ]
0.8065 0.1287 0.8073
[ Variable[CPUFloatType]{1,3} ]
(1,.,.) =
0.8065 0.1287 0.8073
[ Variable[CPUFloatType]{1,1,3} ]

4.3. unsqueeze

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

    torch::Tensor b = a.unsqueeze(0);
    std::cout<<b<<std::endl;

    torch::Tensor bb = a.unsqueeze(1);
    std::cout<<bb<<std::endl;

    torch::Tensor bbb = a.unsqueeze(2);
    std::cout<<bbb<<std::endl;

结果：

0.7945 0.0331 0.1666
0.7821 0.3359 0.0663
[ Variable[CPUFloatType]{2,3} ]
(1,.,.) =
0.7945 0.0331 0.1666
0.7821 0.3359 0.0663
[ Variable[CPUFloatType]{1,2,3} ]
(1,.,.) =
0.7945 0.0331 0.1666

(2,.,.) =
0.7821 0.3359 0.0663
[ Variable[CPUFloatType]{2,1,3} ]
(1,.,.) =
0.7945
0.0331
0.1666

(2,.,.) =
0.7821
0.3359
0.0663
[ Variable[CPUFloatType]{2,3,1} ]

5.torch::nonzero 输出非0的坐标

    torch::Tensor a = torch::rand({2,3});
    a[0][1] = 0;
    a[1][2] = 0;
    std::cout<<a<<std::endl;
     torch::Tensor b = torch::nonzero(a);
     std::cout<<b<<std::endl;

结果：

0.4671 0.0000 0.3360
0.9320 0.9246 0.0000
[ Variable[CPUFloatType]{2,3} ]
0 0
0 2
1 0
1 1
[ Variable[CPULongType]{4,2} ]

6.访问tensor值 a.item()就把1*1 的 tensor的a转为float

取出tensor的某个值 为int或者float ===》》》auto bbb = a[1][1].item().toFloat();
一般情况下取出tensor某个值可以直接下标索引即可。比如a[0][1],但是这个值还是tensor类型的，要想为c++的int或者float的，如下：

    torch::Tensor a = torch::rand({2,3});
    std::cout<<a<<std::endl;
    auto bbb = a[1][1].item().toFloat();
    std::cout<<bbb<<std::endl;

结果：

0.7303 0.6608 0.0024
0.5917 0.0145 0.6472
[ Variable[CPUFloatType]{2,3} ]
0.014509
[ Variable[CPUFloatType]{} ]
0.014509

另外的例子：

    torch::Tensor scores = torch::rand({10});
    std::tuple<torch::Tensor,torch::Tensor> sort_ret = torch::sort(scores.unsqueeze(1), 0, 1);
    torch::Tensor v = std::get<0>(sort_ret).squeeze(1).to(scores.device());
    torch::Tensor idx = std::get<1>(sort_ret).squeeze(1).to(scores.device());
    std::cout<<scores<<std::endl;
    std::cout<<v<<std::endl;
    std::cout<<idx<<std::endl;

    for(int i=0;i<10;i++)
    {
         int idx_1 = idx[i].item<int>();
         float s = v[i].item<float>();

          std::cout<<idx_1<<"  "<<s<<std::endl;
    }

结果：

0.1125
0.9524
0.7033
0.3204
0.7907
0.8486
0.7783
0.3215
0.0378
0.7512
[ Variable[CPUFloatType]{10} ]
0.9524
0.8486
0.7907
0.7783
0.7512
0.7033
0.3215
0.3204
0.1125
0.0378
[ Variable[CPUFloatType]{10} ]
1
5
4
6
9
2
7
3
0
8
[ Variable[CPULongType]{10} ]
1 0.952351
5 0.848641
4 0.790685
6 0.778329
9 0.751163
2 0.703278
7 0.32146
3 0.320435
0 0.112517
8 0.0378203

 

参考博客：libtorch 常用api函数示例（史上最全、最详细）