np.take
与axis
关键字参数一起使用:
>>> a = np.arange(2*3*4).reshape(2, 3, 4)
>>> a
array([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]]])
>>> b = np.arange(3)
>>> np.random.shuffle(b)
>>> b
array([1, 0, 2])
>>> np.take(a, b, axis=1)
array([[[ 4, 5, 6, 7],
[ 0, 1, 2, 3],
[ 8, 9, 10, 11]],
[[16, 17, 18, 19],
[12, 13, 14, 15],
[20, 21, 22, 23]]])
如果你想使用花哨的索引,你只需要用足够的空切片填充索引元组:
>>> a[:, b]
array([[[ 4, 5, 6, 7],
[ 0, 1, 2, 3],
[ 8, 9, 10, 11]],
[[16, 17, 18, 19],
[12, 13, 14, 15],
[20, 21, 22, 23]]])
或者在更一般的环境中:
>>> axis = 1
>>> idx = (slice(None),) * axis + (b,)
>>> a[idx]
array([[[ 4, 5, 6, 7],
[ 0, 1, 2, 3],
[ 8, 9, 10, 11]],
[[16, 17, 18, 19],
[12, 13, 14, 15],
[20, 21, 22, 23]]])
但np.take
真的应该是你的第一选择。