Pytorch: How can I find indices of first nonzero element in each row of a 2D tensor?

Question

I have a 2D tensor with some nonzero element in each row like this:

import torch
tmp = torch.tensor([[0, 0, 1, 0, 1, 0, 0],
                    [0, 0, 0, 1, 1, 0, 0]], dtype=torch.float)

I want a tensor containing the index of first nonzero element in each row:

indices = tensor([2],
                 [3])

How can I calculate it in Pytorch?

Asked By: Iman Aliabdi

||

Source

Answer 1

I could find a tricky answer for my question:

  tmp = torch.tensor([[0, 0, 1, 0, 1, 0, 0],
                     [0, 0, 0, 1, 1, 0, 0]], dtype=torch.float)
  idx = reversed(torch.Tensor(range(1,8)))
  print(idx)

  tmp2= torch.einsum("ab,b->ab", (tmp, idx))

  print(tmp2)

  indices = torch.argmax(tmp2, 1, keepdim=True)
  print(indeces)

The result is:

tensor([7., 6., 5., 4., 3., 2., 1.])
tensor([[0., 0., 5., 0., 3., 0., 0.],
       [0., 0., 0., 4., 3., 0., 0.]])
tensor([[2],
        [3]])

Answered By: Iman Aliabdi

Answer 2

I have simplified Iman’s approach to do the following:

idx = torch.arange(tmp.shape[1], 0, -1)
tmp2= tmp * idx
indices = torch.argmax(tmp2, 1, keepdim=True)

Answered By: Shay

Answer 3

Assuming that all the nonzero values are equal, argmax returns the first index.

tmp = torch.tensor([[0, 0, 1, 0, 1, 0, 0],
                    [0, 0, 0, 1, 1, 0, 0]])
indices = tmp.argmax(1)

Answered By: Seppo Enarvi

Answer 4

Building on @Seppo answer, we can eliminate the assumption that "all the nonzero values are equal" by simply creating a mask from the original tensor and then using pytorch functions

# tmp = some tensor of whatever shape and values
indices = torch.argmax((tmp != 0).to(dtype=torch.int), dim=-1)

However, if a row of tensor is all zeros, then the information returned isn’t the first nonzero element index. I suppose the nature of the question makes it that this case doesn’t occur.

Answered By: cisimon7

Pytorch: How can I find indices of first nonzero element in each row of a 2D tensor?

Question:

Answers: