Connecting a linear layer to a homogenous graph

I have a linear layer which has an output size of 128 dimensions
I have a graph which I have passed through gcn and increased the node size to 128 dimensions as well.

The problem is, I want to connect the outputs of the linear layer to the zero in-degree nodes of my graph. Can I do that?

This is the sample model

class gnn(nn.Module):
    def __init__(self):
        self.fc1 = nn.Linear(4,128)
        self.gcn = GraphConv(128,128, allow_zero_in_degree=True)

   def forward(self, x, dgl_graph, feat, weight):
        x = torch.flatten(x,1)
        x = self.fc1(x)
        x = F.relu(x)
        graph_out = self.gcn(dgl_graph, feat=feat, edge_weight=weight)
        return out

Are you asking how to filter out the zero degree nodes in the graph?

Yes. So that it won’t be a problem when I use prop_nodes. When I use prop_nodes function I am getting a warning of the graph contains invalid edges. Is this behaviour expected?

Why are you using prop_nodes?

For filtering zero degree nodes, you can try zero_degree_node_ids = th.where(g.in_degrees() == 0)

I need to use prop_nodes to pull sequentially the nodes and apply some user defined message passing function of my own. I want to concatenate the embeddings generated by GCN with my edge features. But I am not able to do that. Can I get some help regarding that?

def edge_udf(self, edges):
nodesToPull = edges.src
# Here I want to concatenate the gcn generated embeddings(of the pulled nodes) with the edge #features of the graph and return that as the message function. But I am not able to understand how to #do it for the same tensors pulled by prop_nodes function.
return {‘m’: finalvalue}


Could you provide a code sample snippet? I don’t think I understand your problem currently

Sure. Suppose I have one node feature ‘h’ and 2 edge features ‘e’. Now what I want in my message function is that, I want the 20 dimension embedding(generated by gcn) of the node to be multiplied with my edge feature ‘e’. How can I do that?

class gnn(nn.Module):
def init(self):
super().init() = 0
self.embed = None
self.fc1 = nn.Linear(4,7)
self.fcfeat1 = nn.Linear(4,20)
self.gcn = GraphConv(20, 20, allow_zero_in_degree=True)

def edge_udf(self, edges):
    # multiply the particular node dimension with its edge feature 'e'
    m = dimension(edges.src) * 'e'
    return {'m': m}

def forward(self, x, dgl_graph, feat, edges):
    x = torch.flatten(x,1)
    x = self.fc1(x)
    x = F.relu(x) = x
    feat = self.fcfeat1(feat)
    x = self.gcn(dgl_graph, feat=feat)
    x = F.relu(x)
    self.embed = x
    reduce_func = dgl.function.sum('m', 'h')
    dgl.prop_nodes_topo(dgl_graph, self.edge_udf, reduce_func)
    out = dgl_graph.ndata['h']
    return out

You can use

import dgl.function as fn
g.update_all(fn.u_mul_e('h', 'e'), fn.sum('e', 'out'))

to set data on the node or

def efunc(edges):
  return {'m': edges.src['h']*['e']}

to put data on edges

But this operation updates the whole graph in parallel manner. I want it to happen in a sequential and levelized manner. That’s the reason why I opted for prop_nodes function. Is this the correct approach?

Also, suppose I have 3 edge features ‘e1’, ‘e2’, ‘e3’ and a node feature ‘h’. How can I perform the following operation:

[e1]*[[h][e2][e3]][‘e’] is getting the data regarding all the nodes at a time and throwing this error when I tried to concatenate h, e2, e3
Error: RuntimeError: Sizes of tensors must match except in dimension 1. Got 15 and 7 in dimension 0 (The offending index is 1)

Did you try[node.mailbox['h'], node.mailbox['e2'], node.mailbox['e3']], dim=-1) to indicate it to concat them at the last dim?

But node.mailbox comes into picture when I am writing a reduce function and not a message function. Right? Please correct me if I’m wrong.

Yes. Do you want the concat result on edge or nodes?

Concatenation on edges

You can use g.apply_edges(lambda edges:[edges.src['h'], edges.dst['h']], dim=1)) to concat them to the edges