Output propagate function in Message passing network

[maryamastero]

I have implemented a simple message passing network.

from torch_geometric.nn.conv import MessagePassing

class GConv(MessagePassing):
    def __init__(self, in_channels , out_channels,  **kwarg):
        super(GConv, self).__init__(aggr= 'add', **kwarg)
        self.in_channels = in_channels
        self.out_channels = out_channels
    def forward(self, x, edge_index, edge_attr):
        out = self.propagate(edge_index, x=x, edge_attr=edge_attr, size=None)
        return out
    def message(self, x_i, x_j, edge_attr):
        m = torch.cat([x_i, x_j, edge_attr], dim=-1)
        return m

mygc = GConv(1 , 3)

node_features = torch.tensor([[1.],
        [10.],
        [20.],
        [30.],
        [40.]])

edge_features = torch.tensor([[1.],
        [1.],
        [1.],
        [1.],
        [1.],
        [1.],
        [1.],
        [1.]])

edge_index = torch.tensor([[0, 0, 0, 1, 1, 2, 2, 3],
        [1, 2, 3, 3, 4, 3, 4, 4]])

I do not understand how propagate function works.
Why I got this results? Why first node has zero feature(if I understand right!)

mygc.forward(node_features, edge_index, edge_features)
tensor([[  0.,   0.,   0.],
    [ 10.,   1.,   1.],
    [ 20.,   1.,   1.],
    [ 90.,  31.,   3.],
    [120.,  60.,   3.]])
```
`

[EdisonLeeeee]

Hi @maryamkiashemshaki

The propagate function works by gathering neighboring messages and aggregating them for each node. The results are assigned to the destination nodes of edges in the graphs. In your example, the first node 0 did not appear as a destination node of any edges, i.e., in-degree is zero, so the result is initialized as zero by default. This can be solved by converting your graph into an undirected one https://pytorch-geometric.readthedocs.io/en/latest/modules/utils.html#torch_geometric.utils.to_undirected