RelGraphConv 的输入是一个包含三个部分的元组 (g, feat, rel),其中 g 是一个 DGLGraph 对象,表示输入的异质图;feat 是一个张量,表示节点特征;rel 是一个表示边类型的张量。 RelGraphConv 的主要工作原理是将不同类型的节点和边的特征进行聚合,并将聚合后的结果传递给下一层。具体来说,对于每一层,RelGrap...
I think Da is asking why the nn.GraphConv module does not support weighted graph by default. My concern is what is the right formulation then? Collaborator Author zheng-da commented Feb 24, 2020 • edited I don't mean supporting weighted graph by default. I think we need such an opti...
🐛 Bug In our experiment, we split the training/validation/test set for a graph data set (taking Cora as an example) many times. Each split was trained in a run and the classification accuracy was calculated. We found that when we use dgl...
(hidden_dim * num_heads, hidden_dim, num_heads) # GATConv(in_dim, hidden_dim, num_heads), # GATConv(hidden_dim * num_heads, hidden_dim, num_heads), # GATConv(hidden_dim * num_heads, hidden_dim, num_heads) ]) # for the read out from the hidden features of the graph; ...