How to add an element to the diagonal

[ldv1]

Hi,

I want to add a constant to the diagonal of a sparse matrix. Looking at GCNConv, I assume this is what SparseTensor.fill_diag does (it would be great if methods were documented).
With

import torch
from torch_sparse import SparseTensor
 
col = torch.tensor([1,1,5,0]).type(torch.long)
row = torch.tensor([2,5,3,0]).type(torch.long)
adj = SparseTensor(col=col, row=row)
A = adj.to_dense()
print(A)

I got as expected

tensor([[1., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 1., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 1.],
        [0., 0., 0., 0., 0., 0.],
        [0., 1., 0., 0., 0., 0.]])

Further

adj = SparseTensor.fill_diag(adj, 4.)
print(adj.to_dense())

returns

tensor([[1., 0., 0., 0., 0., 0.],
        [0., 1., 0., 0., 0., 0.],
        [0., 1., 1., 0., 0., 0.],
        [0., 0., 0., 1., 0., 1.],
        [0., 0., 0., 0., 1., 0.],
        [0., 1., 0., 0., 0., 1.]])

So, the diagonal is filled with 1s, instead of 4s. What is wrong with my code?
Thanks for any help.

[rusty1s]

It will only fill the diagonal with values in case value is present in adj:

col = torch.tensor([1,1,5,0]).type(torch.long)
row = torch.tensor([2,5,3,0]).type(torch.long)
value = torch.tensor([1, 1, 1, 1].type(torch.float)
adj = SparseTensor(col=col, row=row, value=value)

[ldv1]

Thanks for the prompt answer.
My code will then work with:

if not adj_t.has_value():
    adj = adj.fill_value(0.)
adj = SparseTensor.fill_diag(adj, 1234.)
print( adj.to_dense() )

So, I understand that value is sort of a placeholder. We first define the placeholder and then fill it.
May I ask why you opted for this strange construct, i.e. why fill_diag(adj,my_float), as the name implies, would not simply fill the diagonal elements with my_float?

What would be the simplest way to add to the diagonal instead of filling?

[rusty1s]

I'm sorry for the lack of documentation :)

In general, SparseTensor supports two formats: One where we only care about connections (value=None), and one which is similar to any other sparse matrix library (value!=None). In case there exists no values, we simply want to add diagonal connections to the matrix.

There's no easy way yet to add diagonal entries to existing ones, since sum is currently not implemented.

[ldv1]

There's no easy way yet to add diagonal entries to existing ones, since sum is currently not implemented

That's bad news. I wanted to extent SIGN to more general propagation matrices. For that I need to add to the diagonal of an adjacency matrix.

[rusty1s]

But for that, adding diagonal entries is sufficient, isn't it? No need to sum them up in case there already exists some.

[ldv1]

The idea is to define a continuum of propagation matrices. To make the problem more easily understandable, I will push a PR.

[ldv1]

We have a set_diag method. Is there a get_diag?

[rusty1s]

Getting diagonal entries is quite straightforward:

row, col, value = A.coo()
diag_values = value[row == col]

[ldv1]

Thanks.
Why not a new function set_diag in diag.py ? That might be useful for others as well.

[rusty1s]

Just pushed mat.get_diag() to master, see here.

[ldv1]

Just what I needed. Thanks.

[ldv1]

Currently, we can only set the diagonal elements of a sparse matrix by specifying a scalar, right ? Would it be possible to specify a tensor instead ?

[rusty1s]

You can either fill with a scaler via fill_diag or with a tensor via set_diag.

[github-actions]

This issue had no activity for 6 months. It will be closed in 2 weeks unless there is some new activity. Is this issue already resolved?