How to combine DGL with other neural network models?

Models & Apps
#1

Hi, DGL is a graceful tool for graph learning.
However, it seems that there is no tutorial or examples in DGL to combine the existing models easily?
For example,

G.ndata['feature'] = features

note the node features are derived from other modules (e.g., a CNN/RNN feature extractor), so in this way, would the whole pytorch computation graph be complete for backpropagation?

Or, if we have obtained the results of GCN Layers, such as:

x = self.gcn1(g, features)
x = self.gcn2(g, x)

If I continue to conduct other operations on x or part columns of x, can the computation graph be complete?
Thanks.

#2

Hi,

Everything in DGL will maintain the backpropagation trace. You can use this with other model in any way as you want.

#3

Thanks. I’ll try it to combine with others.

#4

@htliu Hi, this is a very interesting problem and I wonder if you have solved it. Could you show some examples in code?

Thanks a lot.

1 Like
#5

Hi, I have one question. I am also new to DGL and I am wondering when you feed features to a graph like the following:

G.ndata['feature'] = features

Does the first row in feature corresponds to the node with node id 0 in the graph?

Thanks~

#6

Yeah, hence the dimensions should be consistent.

1 Like
#7

Yes, the i-th row of features always corresponds to the node i in the graph.