Count nonzeros element along an axis

vision
1

Hi!
I have a (32x750) tensor

tensor([[ 0.0000,  0.0000,  0.0000,  ...,  0.0000,  0.0000,  0.0043],
        [ 0.0000,  0.0000,  0.0000,  ...,  0.0000,  0.0000,  0.0043],
        [ 0.0000,  0.0044,  0.0000,  ...,  0.0044,  0.0000,  0.0000],
        ...,
        [ 0.0059,  0.0000,  0.0059,  ...,  0.0059,  0.0000,  0.0000],
        [ 0.0059,  0.0000,  0.0059,  ...,  0.0059,  0.0000,  0.0000],
        [ 0.0000,  0.0000,  0.0000,  ...,  0.0000,  0.0056,  0.0000]], device='cuda:0')

And I want to get the number of nonzero elements along each rows. Something like that
[12 47 0 5 .... 8 7 50]
This discussion didn’t solve my problem and concerned the number of nonzero elements for 1-D tensor.

Thanks

1 Like
2

Maybe something like this:

750 - (tensor == 0).sum(dim=1)

3 Likes
3

Oh, you want differentiable counting

4

Thanks,
It worked
I’m using smooth label regularization and the idea is to normalize the output with target values which will contribute to the grad

1 Like
5

(tensor != 0).sum(dim=1) works as well and you don’t have to mention dimentionality (750).