How Pytorch Tensor Get The Index Of Specific Value

In python list, we can use list.index(somevalue). How can pytorch do this? For example: a=[1,2,3] print(a.index(2)) Then, 1 will be output. How can a pytorch tensor do th

Solution 1:

I think there is no direct translation from list.index() to a pytorch function. However, you can achieve similar results using tensor==number and then the nonzero() function. For example:

t = torch.Tensor([1, 2, 3])
print ((t == 2).nonzero(as_tuple=True)[0])

This piece of code returns

1

[torch.LongTensor of size 1x1]

Solution 2:

For multidimensional tensors you can do:

(tensor == target_value).nonzero(as_tuple=True)

The resulting tensor will be of shape number_of_matches x tensor_dimension. For example, say tensor is a 3 x 4 tensor (that means the dimension is 2), the result will be a 2D-tensor with the indexes for the matches in the rows.

tensor = torch.Tensor([[1, 2, 2, 7], [3, 1, 2, 4], [3, 1, 9, 4]])
(tensor == 2).nonzero(as_tuple=False)
>>> tensor([[0, 1],
        [0, 2],
        [1, 2]])

Solution 3:

Based on others' answers:

t = torch.Tensor([1, 2, 3])
print((t==1).nonzero().item())

Solution 4:

Can be done by converting to numpy as follows

import torch
x = torch.range(1,4)
print(x)
===> tensor([ 1.,  2.,  3.,  4.]) 
nx = x.numpy()
np.where(nx == 3)[0][0]
===> 2

Solution 5:

The answers already given are great but they don't handle when I tried it when there is no match. For that see this:

def index(tensor: Tensor, value, ith_match:int =0) -> Tensor:
    """
    Returns generalized index (i.e. location/coordinate) of the first occurence of value
    in Tensor. For flat tensors (i.e. arrays/lists) it returns the indices of the occurrences
    of the value you are looking for. Otherwise, it returns the "index" as a coordinate.
    If there are multiple occurences then you need to choose which one you want with ith_index.
    e.g. ith_index=0 gives first occurence.

    Reference: https://stackoverflow.com/a/67175757/1601580
    :return:
    """# bool tensor of where value occurred
    places_where_value_occurs = (tensor == value)
    # get matches as a "coordinate list" where occurence happened
    matches = (tensor == value).nonzero()  # [number_of_matches, tensor_dimension]if matches.size(0) == 0:  # no matchesreturn -1else:
        # get index/coordinate of the occurence you want (e.g. 1st occurence ith_match=0)index = matches[ith_match]
        returnindex

credit to this great answer: https://stackoverflow.com/a/67175757/1601580