[人工智能]C++ Eigen3库Tensor操作整理(持续更新)

项目引入

VS2019环境下

1. 点击<项目>
2. 管理NuGet程序包
3. 搜索Eigen3，安装即可
   
   在项目源代码添加头文件

#include <unsupported/Eigen/CXX11/Tensor>

类型转换

C++ 数组转Eigen::Tensor(包含reshape操作)

例如从float数组转到储存类型为float的Tensor

1. float数组转Eigen::TensorMap
2. Eigen::TensorMap类型转换回Eigen::Tensor

#include <iostream>
#include <unsupported/Eigen/CXX11/Tensor>

int main()
{
	// 定义一个一维数组，在下面将其转换为2x2的tensor
    float arr[] = { 0.1, 0.2, 0.3, 0.4 };
    // 定义转换的Eigen::TensorMap，同时做了一个reshape操作
    auto mapped_t = Eigen::TensorMap<Eigen::Tensor<float, 2>>(arr, 2, 2);
    std::cout << typeid(mapped_t).name() << std::endl;
    // 强制转换为Tensor
    auto result = Eigen::Tensor<float, 2>(mapped_t);
    std::cout << typeid(result).name() << std::endl;
    std::cout << result << std::endl;
}

输出

class Eigen::TensorMap<class Eigen::Tensor<float,2,0,__int64>,0,struct Eigen::MakePointer>
class Eigen::Tensor<float,2,0,__int64>
0.1 0.3
0.2 0.4

上面的操作与python numpy的narray.reshape()是一致的。下面代码是对narray.reshape()的实现。已验证与上面结果一致。

Eigen::Tensor<float, 3> reshape(float* oneDimArray, int channel,int height, int width)
{
    const int frameSize = height * width;
    const auto length = channel * height * width;
    int index = 0;
    Eigen::Tensor<float, 3> result(channel, height, width);
    for (int c = 0; c < channel; c++)
    {
        for (int w = 0; w < width; w++)
        {
            for (int h = 0; h < height; h++)
            {
                index = c * frameSize + w * height + h;
                float value = oneDimArray[index];
                result(c, h, w) = value;
            }
        }
    }
    return result;
}

Tensor操作

concatenate

#include <iostream>
#include <unsupported/Eigen/CXX11/Tensor>

int main()
{
	// 创建一个2维，2x2的tensor
    Eigen::Tensor<float, 2, 1> a(2, 2);
    Eigen::Tensor<float, 2, 1> b(2, 2);
    // 给tensor设置默认值
    a.setConstant(1.0);
    b.setConstant(2.0);
    // 从维度0进行拼接
    auto res = a.concatenate(b, 0);
    std::cout << res << std::endl;
}

输出结果:

1 1
1 1
2 2
2 2

transpose

#include <iostream>
#include <unsupported/Eigen/CXX11/Tensor>
using namespace std;
int main()
{
    Eigen::Tensor<int, 3> m(2, 3, 3);
    m.setValues(
        {
            {{1, 2, 3},
            {4, 5, 6},
            {7, 8, 9}},
            {{10, 11, 12},
            {13, 14, 15},
            {16, 17, 18}}
        });

    Eigen::array<int, 3> shuffling({ 2, 1, 0});

    Eigen::Tensor<int, 3> transposed = m.shuffle(shuffling);

    cout << transposed(0, 0, 0) << ',' << transposed(0, 0, 1) << endl;
    cout << transposed(0, 1, 0) << ',' << transposed(0, 1, 1) << endl;
    cout << transposed(0, 2, 0) << ',' << transposed(0, 2, 1) << endl;
    cout << transposed(1, 0, 0) << ',' << transposed(1, 0, 1) << endl;
    cout << transposed(1, 1, 0) << ',' << transposed(1, 1, 1) << endl;
    cout << transposed(1, 2, 0) << ',' << transposed(1, 2, 1) << endl;
    cout << transposed(2, 0, 0) << ',' << transposed(2, 0, 1) << endl;
    cout << transposed(2, 1, 0) << ',' << transposed(2, 1, 1) << endl;
    cout << transposed(2, 2, 0) << ',' << transposed(2, 2, 1) << endl;
}

输出

1,10
4,13
7,16
2,11
5,14
8,17
3,12
6,15
9,18

该操作与python的numpy库的ndarray.transpose一致

import numpy as np

mat = [
            [[1, 2, 3],
            [4, 5, 6],
            [7, 8, 9]],
            [[10, 11, 12],
            [13, 14, 15],
            [16, 17, 18]]
        ]
mat = np.array(mat)
mat = mat.transpose(2, 1, 0)
print(mat[0][0][0], mat[0][0][1])
print(mat[0][1][0], mat[0][1][1])
print(mat[0][2][0], mat[0][2][1])
print(mat[1][0][0], mat[1][0][1])
print(mat[1][1][0], mat[1][1][1])
print(mat[1][2][0], mat[1][2][1])
print(mat[2][0][0], mat[2][0][1])
print(mat[2][1][0], mat[2][1][1])
print(mat[2][2][0], mat[2][2][1])

输出

1 10
4 13
7 16
2 11
5 14
8 17
3 12
6 15
9 18

reshape