[Python] Multidimensional matrix permutation Julia vs. Python disagreement

I have noticed a difference in behavior between python's numpy.permute_dims and julia's Base.permutedims.

On an input 3x3x3 matrix containing elements 0:26, inclusive in both languages, they agree for the axes argument (1,2,0) but disagree for (0,2,1).

As far as I can tell from the docs, these functions should be equivalent. There's a note about permutedims being non-recursive, but I don't see why that should have behavioral consequences.

Julia is also column-major order, but again I don't see why that should effect the overall behavior.

Python code:

arr = np.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],
[24, 25, 26]]])

arr_perm = np.permute_dims(arr, axes=[0,2,1])


Output:

array([[[ 0, 3, 6],
[ 1, 4, 7],
[ 2, 5, 8]],

[[ 9, 12, 15],
[10, 13, 16],
[11, 14, 17]],

[[18, 21, 24],
[19, 22, 25],
[20, 23, 26]]])


Julia code:

arr = [
0 1 2
3 4 5
6 7 8;;;
9 10 11
12 13 14
15 16 17;;;
18 19 20
21 22 23
24 25 26
]

arr_perm = permutedims(arr, [1,3,2])


Output:

3×3×3 Array{Int64, 3}:
[:, :, 1] =
0 9 18
3 12 21
6 15 24

[:, :, 2] =
1 10 19
4 13 22
7 16 25

[:, :, 3] =
2 11 20
5 14 23
8 17 26

Continue reading...