用 2d 数组索引 3d numpy 数组

Question

我想根据 numpy 3d-array 中的值创建一个 numpy 2d-array，使用另一个 numpy 2d-array 来确定在轴 3 中使用哪个元素。

import numpy as np
#--------------------------------------------------------------------
arr_3d = np.arange(2*3*4).reshape(2,3,4)
print('arr_3d shape=', arr_3d.shape, '\n', arr_3d)
arr_2d = np.array(([3,2,0], [2,3,2]))
print('\n', 'arr_2d shape=', arr_2d.shape, '\n', arr_2d)
res_2d = arr_3d[:, :, 2]
print('\n','res_2d example using element 2 of each 3rd axis...\n', res_2d)
res_2d = arr_3d[:, :, 3]
print('\n','res_2d example using element 3 of each 3rd axis...\n', res_2d)

结果...

arr_3d shape= (2, 3, 4) 
 [[[ 0  1  2  3]
  [ 4  5  6  7]
  [ 8  9 10 11]]

 [[12 13 14 15]
  [16 17 18 19]
  [20 21 22 23]]]

 arr_2d shape= (2, 3) 
 [[3 2 0]
 [2 3 2]]

 res_2d example using element 2 of each 3rd axis...
 [[ 2  6 10]
 [14 18 22]]

 res_2d example using element 3 of each 3rd axis...
 [[ 3  7 11]
 [15 19 23]]

2 个示例结果显示了我使用轴 3 的第二个和第三个元素时得到的结果。但我想从 arr_3d 中获取由 arr_2d 指定的元素。所以...

- res_2d[0,0] would use the element 3 of arr_3d axis 3
- res_2d[0,1] would use the element 2 of arr_3d axis 3
- res_2d[0,2] would use the element 0 of arr_3d axis 3
etc

所以 res_2d 应该是这个样子...

[[3 6 8]
[14 19 22]]

我尝试使用这一行来获取 arr_2d 条目，但它会生成一个 4-dim 数组，我想要一个 2-dim 数组。

res_2d = arr_3d[:, :, arr_2d[:,:]]

Answer 1

花式索引和广播结果的形状是索引数组的形状。您需要为arr_3d 的每个轴传递二维数组

ax_0 = np.arange(arr_3d.shape[0])[:,None]
ax_1 = np.arange(arr_3d.shape[1])[None,:]

arr_3d[ax_0, ax_1, arr_2d]

Out[1127]:
array([[ 3,  6,  8],
       [14, 19, 22]])

Answer 2

In [107]: arr_3d = np.arange(2*3*4).reshape(2,3,4)                                                           
In [108]: arr_2d = np.array(([3,2,0], [2,3,2]))                                                              
In [109]: arr_2d.shape                                                                                       
Out[109]: (2, 3)
In [110]: arr_3d[[[0],[1]],[0,1,2],arr_2d]                                                                   
Out[110]: 
array([[ 3,  6,  8],
       [14, 19, 22]])

[[0],[1]],[0,1,2]相互广播索引一个（2,3）块，大小与`arr_2d相同。

ix_ 可用于构造这两个索引：

In [114]: I,J = np.ix_(range(2), range(3))                                                                   
In [115]: I,J                                                                                                
Out[115]: 
(array([[0],
        [1]]), array([[0, 1, 2]]))
In [116]: arr_3d[I, J, arr_2d]                                                                               
Out[116]: 
array([[ 3,  6,  8],
       [14, 19, 22]])