Question about g.prop_nodes

Questions
#1

Hi, Thank you very much for your time to read my question . I was reading about prop_nodes just now and havent yet totally understood .

g.prop_nodes([[2,3],[4]] , fn.copy_src('x', 'm'),   fn.sum('m', 'x'))

g.prop_nodes([[2,3]] , fn.copy_src('x', 'm'),   fn.sum('m', 'x'))
g.prop_nodes([4] , fn.copy_src('x', 'm'),   fn.sum('m', 'x'))

^^ are these two the same ?

Question on dgl
Question on dgl1362Ă—926 19 KB

Also given I have a directed graph . what would be the easiest method to access earlier node features (read some posts about k-hop but unsure of its implementation…), as simple as accessing last node feature . like at node 4, to get node 3 and node 2 and node 1’s node feature, can I send message from one node to another node that is not direct successor, (given there are edges in between them.)

def f_edge(edges):
    return {'h': edges.src['x']}

Thank you very much for your time !

#2

Yes, it’s going to iterate one set at a time.

predecessors function for nodes and then use ndata to get those features

This is in effect what message passing is doing across multiple layers.
If you think you want a higher receptive fields than just the layers capable in the model, this is a good related paper Multi-hop Attention Graph Neural Network

1 Like
#3

Thank you very much for your reply !

#4

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.