Difference between ":" and "..." in tensor slicing

I am new to PyTorch. Can someone explain to me the difference between “:” and “…” in tensor slicing.
for instance

tensor = torch.rand(3,4)
print(f"Last column: {tensor[:,-1]}")
print(f"Last column: {tensor[...,-1]}")

outputs the same results
Thankyou in advance

It seems that when using : and ... in this context they yield the same result. However, it seems when using :, and ... in terms of adding a new dim to a Tensor they give different results.

>>> M=torch.randn(3,4)
>>> M
tensor([[-1.1915,  1.3102,  2.4116, -0.6022],
        [-1.4025, -0.0796, -0.6342, -1.5389],
        [-0.3399, -0.1049,  0.0192,  0.0186]])

#grab last column
>>> M[..., -1]
tensor([-0.6022, -1.5389,  0.0186])
>>> M[:, -1]
tensor([-0.6022, -1.5389,  0.0186])

#adding new dim
#shape torch.Size([3, 1, 4])
>>> M[:, None]
tensor([[[-1.1915,  1.3102,  2.4116, -0.6022]],

        [[-1.4025, -0.0796, -0.6342, -1.5389]],

        [[-0.3399, -0.1049,  0.0192,  0.0186]]])
#shape torch.Size([3, 4, 1])
>>> M[..., None]
tensor([[[-1.1915],
         [ 1.3102],
         [ 2.4116],
         [-0.6022]],

        [[-1.4025],
         [-0.0796],
         [-0.6342],
         [-1.5389]],

        [[-0.3399],
         [-0.1049],
         [ 0.0192],
         [ 0.0186]]])
>>> 

So, in your cases they are the same? but in other cases they’ll give different behaviour!

It’s the same as numpy slicing, use colons ( : ) when you have multiple dimensions you need to slice differently, ie: tensor[:-1. 2:-1, :] and semicolons (…) when all following (or previous) dimensions should be kept the same. For example, a tensor of shape: (8, 3, 256) can be sliced as:

  • tensor[0:1, …] or equivalently tensor[0:1, :, :]
  • tensor[…, -1] or equivalently tensor [:, :, -1]
  • tensor[0:1, 0, …] or tensor[0:1, 0, :]
    and so on and so on