Add_self_loops doesn't change a result

[kenenbek]

I have such model with self-loops

model = GCNConv(in_channels=1, out_channels=1, add_self_loops=True, normalize=False)
But I don't specify any edgees here:

simple_edge_index = torch.tensor([[],
                           []], dtype=torch.long)

And I have simple input data:

x = torch.tensor([[1], [1], [1], [1], [1]], dtype=torch.float)

I run the model:
h = model(x, simple_edge_index)

And as a result I get:

tensor([[0.],
        [0.],
        [0.],
        [0.],
        [0.]], grad_fn=<AddBackward0>)

Could you explain why it happened? Because I have self_loops and non-zero weights of the model

OrderedDict([('bias', tensor([0.])), ('lin.weight', tensor([[0.1189]]))])

I expected to get something like (because of self-loops):

tensor([[0.1189],
        [0.1189],
        [0.1189],
        [0.1189],
        [0.1189]], grad_fn=<AddBackward0>)

[rusty1s]

Mh, I cannot reproduce this. This works for me

import torch

from torch_geometric.nn import GCNConv

conv = GCNConv(1, 1)

edge_index = torch.empty(2, 0, dtype=torch.long)
x = torch.ones(4, 1)

out = conv(x, edge_index)
print(out)

Can you confirm?

[rusty1s]

The issue seems to be with normalize=False, since it currently also disables adding self-loops :(