About batch training

Questions
#1

When I use the sum of all node features as GNN model forward function’s return, the batch doesn’t work. Is there any alternative to the sum function ？
my code looks like

def forward(self, g, h):
    g.edata['h'] = h
    h = self.nn_layers(h) # (V, edge_out_feats) V is the number of edges
    return h.sum
#2

In my opinion, the problem is that the graph after batch in DGL is a large graph, so the return value after sum cannot be distinguished

#3

I assume you want to sum over all the edge features for each graph individually? If so, try dgl.sum_edges.

1 Like
#4

Wow! It’s worked! Thank you a lot!
Do I have to read the entire DGL document to find the functions I need, or do I have to be imaginative and observant, observe how functions are named, and believe that a graph-processing library should come with such functions :thinking:

#5

What you are trying to do is quite similar to graph classification, which I think is similar to what GIN does. I would say that examples and tutorials (e.g. Training a GNN for Graph Classification — DGL 0.7.1 documentation) should be a good starting point.

#6

You’re right. Actually，I have also practiced GIN on my own data sets.

#7

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