How to use edge weighted?

[wryyy1999]

Currently I am using the GAT model that comes with DGL. Now I hope that the parameter ‘g’ of the input model is a weighted graph. During the forward propagation process, the node will take the weight of the edges into account when aggregating neighbor information.

class GAT(nn.Module):

def __init__(self, in_feats, hidden_feats, n_classes, activation,dropout):
    super(GAT, self).__init__()
    self.dropout = nn.Dropout(dropout)
    self.layers = nn.ModuleList()
    self.layers.append(GATConv(in_feats, hidden_feats, num_heads=4, activation=activation))
    self.layers.append(GATConv(hidden_feats * 4, n_classes, num_heads=1, activation=None))  

def forward(self, g, features):
    x = features
    for layer in self.layers[:-1]:
        x = layer(g, x).flatten(1)
        x = self.dropout(x) 
    x = self.layers[-1](g, x).flatten(1)
    return x

[wryyy1999]

In my understanding, I use the parameter ‘g’, the adjacency matrix of the dgl graph, to obtain the graph structure to facilitate subsequent message dissemination. But I currently don’t know how to modify this adjacency matrix into a weighted matrix.

[peizhou001]

Hi @wryyy1999, you can do this by setting edge attributes:
`g.edata[‘x’] = xxxxx
1.3 Node and Edge Features — DGL 0.8.2post1 documentation.

[wryyy1999]

@peizhou001 Yes, I already know that there is an attribute ‘edata’ in which the edge weights can be stored, but what I am curious about is how to use the weights in ‘endata’ for the message propagation of the model.

[peizhou001]

Just fetch the attribute and do anything you want, which depends on your specific model.