Question about indexing a tensor - X[1,2,3] vs X[1][2][3]

[pbsepto82]

Hello!

I was trying out indexing a tensor and to my surprise, I found indexing returns different results depending on whether I use a single square bracket (with comma separated indices) OR multiple square brackets (with each index having it's own bracket). Here's my code:-

`import torch

tensorA = torch.arange(1,19).reshape(2,3,3)
print(tensorA)
print(tensorA[0][:][2])

print("\n")

x = torch.arange(1,19).reshape(2,3,3)
print(x)
print(x[0,:,2])`

This is what the output is:-

tensor([[[ 1, 2, 3],
[ 4, 5, 6],
[ 7, 8, 9]],

    [[10, 11, 12],
     [13, 14, 15],
     [16, 17, 18]]])

tensor([7, 8, 9])

tensor([[[ 1, 2, 3],
[ 4, 5, 6],
[ 7, 8, 9]],

    [[10, 11, 12],
     [13, 14, 15],
     [16, 17, 18]]])

tensor([3, 6, 9])

My intention is to get the second outcome but I'm not able to understand why the first one is returning tensor([7, 8, 9])

I'm very new to this, would really appreciate if anyone can guide. Thanks in advance!

[Perian-Yan]

Hi,
I think the problem comes from slicing [:], which is different when using a single bracket and multiple brackets. You can print tensorA step by step and see what's going on:

For indexing, I think they are the same.

[pbsepto82]

Thank you!
I see, it's little misleading though I'm specifying index 2 for 3rd dimension to get the 2nd values obviously, but the result is not in accordance with what it looks like with [0][:][2]. Would I be right to say that we should always use single bracket when slicing to avoid any confusion?

[Perian-Yan]

Yes, to get, e.g., tensor([3,6,9]), we can only use single bracket slicing.