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

问题

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?

回答1:

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)

回答2:

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]])

来源：https://stackoverflow.com/questions/56088189/pytorch-how-can-i-find-indices-of-first-nonzero-element-in-each-row-of-a-2d-ten