VirtualNode

gtrick.dgl.VirtualNode(in_feats, out_feats, dropout=0.5, residual=False)

Bases: nn.Module

Virtual Node from OGB Graph Property Prediction Examples.

It adds an virtual node to all nodes in the graph. This trick is helpful for Graph Level Task.

Note

To use this trick, call update_node_emb at first, then call update_vn_emb.

Examples:

VirtualNode (DGL)

Parameters:

Name	Type	Description	Default
in_feats	int	Feature size before conv layer.	required
out_feats	int	Feature size after conv layer.	required
dropout	float	Dropout rate on virtual node embedding. Default: 0.5.	0.5
residual	bool	If True, use residual connection. Default: False.	False

update_node_emb(graph, x, vx=None)

Add message from virtual nodes to graph nodes.

Parameters:

Name	Type	Description	Default
graph	dgl.DGLGraph	The graph.	required
x	torch.Tensor	The input node feature.	required
vx	torch.Tensor	Optional virtual node embedding. Default: None.	None

Returns:

Type	Description
torch.Tensor	The output node feature.
torch.Tensor	The output virtual node embedding.

update_vn_emb(graph, x, vx)

Add message from graph nodes to virtual node.

Parameters:

Name	Type	Description	Default
graph	dgl.DGLGraph	The graph.	required
x	torch.Tensor	The input node feature.	required
vx	torch.Tensor	Optional virtual node embedding.	required

Returns:

Type	Description
torch.Tensor	The output virtual node embedding.