python、numpy 、pytorch高级切片索引，[]放数组。

python、numpy 、pytorch高级切片索引，[]里面放数组。

只实验了二维数组，表示对同一个维度进行多次切片，二维数组每行的值为每次的切片索引，有n行就切片n次，同时把该维度切成n个维度。
numpy和pytorch结果一样，只放了numpy代码。

直接上结果：

import numpy as np 
a=np.random.randn(2,3,4)
b=[[1,2],[0,1]]
c=a[:,b,:]
print(a)
print("c--------")
print(c)

结果：

[[[-1.49800039  0.58922711  0.93803902  0.87814673]
  [ 0.31813503  0.61408442 -0.43424169  0.01723555]
  [ 0.88257481  0.43878342  0.25124952  0.28165291]]

 [[ 0.50707757 -1.04502387  1.02290272 -1.34753527]
  [-2.84167515  1.777284   -0.49271416  0.23517798]
  [ 1.75211101 -0.49689842 -1.84401291 -1.82799111]]]
c--------
[[[[ 0.31813503  0.61408442 -0.43424169  0.01723555]
   [ 0.88257481  0.43878342  0.25124952  0.28165291]]

  [[-1.49800039  0.58922711  0.93803902  0.87814673]
   [ 0.31813503  0.61408442 -0.43424169  0.01723555]]]


 [[[-2.84167515  1.777284   -0.49271416  0.23517798]
   [ 1.75211101 -0.49689842 -1.84401291 -1.82799111]]

  [[ 0.50707757 -1.04502387  1.02290272 -1.34753527]
   [-2.84167515  1.777284   -0.49271416  0.23517798]]]]

打印形状

print(a.shape)
print("c--------")
print(c.shape)
(2, 3, 4)
c--------
(2, 2, 2, 4)

---

补充：后面看代码发现不对劲，又加个新的实验：

import numpy as np 
a=np.random.randn(2,3,4)
b=[[1,2],[0,1]]
d=[[0],[1]]
c=a[d,b,:]
print(a)
print("c--------")
print(c)

[[[-1.35228819  1.63435636 -0.34565672  0.84414161]
  [-0.14872702 -1.03662345 -0.40969845 -0.90026686]
  [-0.52720655  0.82482226  0.07821771  1.49102265]]

 [[-1.71620832 -1.10852628 -0.94867257 -2.42966257]
  [ 1.57493612  1.64461755 -0.33925698 -0.63227565]
  [-0.9511253   1.36616056 -0.37967324  1.91686236]]]
c--------
[[[-0.14872702 -1.03662345 -0.40969845 -0.90026686]
  [-0.52720655  0.82482226  0.07821771  1.49102265]]

 [[-1.71620832 -1.10852628 -0.94867257 -2.42966257]
  [ 1.57493612  1.64461755 -0.33925698 -0.63227565]]]

print(a.shape)
print("c--------")
print(c.shape)

(2, 3, 4)
c--------
(2, 2, 4)

两个维度里两个数组，结果最后维度数量不变。可以理解为后面二维数组的每行只去切前面二维数组的那行，如b[1,2]去切第一维的索引1，如d的[0]。然后b的[0,1]去切第一维的索引[1],如d的[1]。所以当b的行数等于d的行数，就会出现这种效果。