When using torch.sort(), how do I assign the minimal index to those same elements?

1

Like if I have a Tensor

a = [1, 2, 2, 3]

If I perform torch.sort(a), then the index it returns will be

[0, 1, 2, 3]

What if I want to get the result like

[0, 1, 1, 3]

I know there’s a function exists in scipy “scipy.data.rankdata”, in the function, I can specify the method it uses when ranking the data. But how can I achieve the similar goal with pytorch?

2 
torch.unique(x, sorted=True, return_inverse=True)
(tensor([1, 2, 3]), tensor([0, 1, 1, 2]))
3

Thanks!
But the result I expect to get is

[0, 1, 1, 3]

That is, when same elements appear in an array, the returned index will not be “continous”. Are there any other ways to do it?

4

I dont think there’s a direct way to do [0, 1, 1, 3]. Unique gives you what you want from an ordering perspective, but does not account for the right magnitude,

5

It’s also useful. Thanks anyway!

6

For anyone who stumbles over this thread again, a small modification to @smth 's answer works:

import torch
x = torch.tensor([1, 2, 2, 3])
uq_vals,inverse_ixs,counts = torch.unique(x, sorted=True, return_inverse=True,return_counts=True)
cumsum_counts = counts.cumsum(dim=0)
possible_ranks = torch.zeros_like(counts)
possible_ranks[1:] = cumsum_counts[:-1]
ranks = possible_ranks[inverse_ixs]
print(x,
      ranks)