How to implement in pytorch - numpy's .unique WITH(!) return_index = True?

Question

In numpy.unique there is an option return_index=True - which returns positions of unique elements (first occurrence - if several).

Unfortunately, there is no such option in torch.unique !

Question: What are the fast and torch-style ways to get indexes of the unique elements ?

=====================

More generally my issue is the following: I have two vectors v1, v2, and I want to get positions of those elements in v2, which are not in v1 and also for repeated elements I need only one position. Numpy's unique with return_index = True immediately gives the solution. How to do it in torch ? If we know that vector v1 is sorted, can it be used to speed up the process ?

Answer 1

You can achieve this in PyTorch with the following approach:

def get_unique_elements_first_idx(tensor):
    # sort tensor
    sorted_tensor, indices = torch.sort(tensor)
    # find position of jumps
    unique_mask = torch.cat((torch.tensor([True]), sorted_tensor[1:] != sorted_tensor[:-1]))
    return indices[unique_mask]

Example usage:

v1 = torch.tensor([2, 3, 3])
v2 = torch.tensor([1, 2, 6, 2, 3, 10, 4, 6, 4])

# Mask to find elements in v2 that are not in v1
mask = ~torch.isin(v2, v1)
v2_without_v1 = v2[mask]

# Get unique elements and their first indices
unique_indices = get_unique_elements_first_idx(v2_without_v1)

print(unique_indices)           #[0, 3, 1, 2]
print(v2[mask][unique_indices]) #[1, 4, 6, 10]

P.S. On my computer, the function processes a vector of size 10 million in about (1.1 ± 0.1)s.