Advanced Indexing Not Update Values as Expected?

1

I’m encountering a confusing behavior when using advanced indexing to update values in a tensor. Here’s an example:

import torch  
A = torch.zeros(2, 2)  
Inds = torch.tensor([[0,0,1],[1,1,1]])
A[Inds.unbind()] += torch.tensor([1, 0, 1])

I expected A[0, 1] to increment by 1, but its value remains 0 while the value of A[1,1] is incremented by 1. At first I thought that this was related to the in-place operation. I tried to use an explicit addition as shown bellow:

A[Inds.unbind()] = A[Inds.unbind()] + torch.tensor([1, 0, 1])

After this operation, A[0, 1] is still 0 whereas A[1,1] equals 2 (which the corredct value since it was equal to 1 before).

My current fix is done using a loop over the values of Inds but I would like to know why does this happen and if there is a way to vecotrize this instead?

2

The behavior this usage of advanced indexing is documented here Tensor Indexing API — PyTorch main documentation to be equivalent to doing index_put_. And index_put_ eis undefined when indices repeat and accumulate=False. To get the behavior you want, it sounds like you want to set accumulate=True?

>>> A.index_put_(inds.unbind(), torch.tensor([1., 0., 1.]), accumulate=True)
tensor([[0., 1.],
        [0., 1.]])