我有一个2D numpy数组

weights = np.array(
    [
        [1, 2, 3],
        [4, 5, 6],
        [7, 8, 9],
    ]
)

我有两个一维numpy数组，我想用它们来索引权重

positions1 = np.array([2, 1, 0])
positions2 = np.array([1, 1, 0])

如果我想得到遍历数组的结果，并索引到矩阵中，我可以这样做

print(weights[positions1[...], positions2[...]])

然后得到[8 5 1]
但是现在我想索引到矩阵中所有可能的位置1和位置2的组合，这样我就得到了一个像这样的矩阵

[
[weights[2, 1], weights[2, 1], weights[2, 0]],
[weights[1, 1], weights[1, 1], weights[1, 0]],
[weights[0, 1], weights[0, 1], weights[0, 0]],
]

所以呢

[
[weights[pos1[0], pos2[0]], weights[pos1[0], pos2[1]], weights[pos1[0], pos2[2]]],
[weights[pos1[1], pos2[0]], weights[pos1[1], pos2[1]], weights[pos1[1], pos2[2]]],
[weights[pos1[2], pos2[0]], weights[pos1[2], pos2[1]], weights[pos1[2], pos2[2]]],
]

在numpy中做这件事的规范方法是什么？我知道这有点像外积，但我实际上并不想将数组中的值相乘，而是将它们的值作为元组索引到矩阵中。