awkward.argsort is wrong for arrays with dimensions > 2

Results for awkard.argsort is wrong for arrays with dimensions greater than 2. Example output from awkward1 version 0.3.1

Code

import awkward1
x = awkward1.Array([[[0.01539855, 0.45536142, 0.9510973 , 0.17593855, 0.30485241],
                     [0.57157337, 0.2281758 , 0.1437675 , 0.80287155, 0.11580015],],
                    [[0.77897828, 0.75505657, 0.61302232, 0.56694878, 0.99761142],
                     [0.38639971, 0.69037858, 0.61298759, 0.6602239 , 0.93297311],],]) # 2x2x5 array generated from numpy.random
print(awkward1.to_numpy(awkward1.argsort(x,axis=-1)))

Resulting output:

[[[0 3 4 1 2]
  [4 2 1 0 3]]
 [[3 2 1 0 4]
  [0 2 3 1 4]]]

Expected output:

[[[0 3 4 1 2]
  [3 2 1 4 0]]
 [[3 2 1 0 4]
  [0 3 1 2 4]]]

[jpivarski]

Aha! You thought it was the inverse, a natural assumption to make.

That is, if selecting an item from an array is a function, f: int → dtype, then slicing by an array, g: int → int, is composition: f ⊗ g: int → int ⊗ int → dtype or f ⊗ g: int → dtype. That g: int → int function can have an inverse, g⁻¹, (depending on how general we take it to be, but from argsort, it’s always a permutation), and it’s natural to guess that it’s supposed to be g⁻¹ when it’s supposed to be g. They have the same int → int signature, after all.

That’s why we check to see what NumPy does, to be sure we pick the same conventions.

[yimuchen]

@jpivarski Ah, thanks. It looks like I have misunderstood the meaning of the output: So basically, the argsort function outputs the indices to retrieve the objects to place the numbers in sorted order. While the output I was expecting was the position at which the numbers will end up. Thanks for the clarification! (I guess it was unfortunate that the first row of the example outputs has two answers coincided, so I as confused as to why the output is ‘sometimes right’ when it as I that was wrong all along)

So yes this issue is closed.