Neighbor sampling and aggregation with graphbolt

Questions
#1

Hi there.

I’m migrating code that was originally written using DGL1.X when neighbor sampling was done with dgl.contrib.sampling.NeighborSampler.

The output of the sampled graph was a generator that generated objects of the NodeFlow type. With NodeFlow, it was possible to access graph data simply with layers[i].data. For example, below is my old code for message passing on the subgraph after neighbor sampling (which is from a documentation that apparently no longer exists; see: What's the best practice to train node embeddings for a big graph):

    def encode(self, data_flows, training=True):
        # print(data_flows)
        x = self.embeddings
        nf = next(iter(data_flows))
        nf.copy_from_parent()
        nf.layers[0].data['activation'] = x[nf.layers[0].data['feature']]
        for i, layer in enumerate(self.layers):

            h = nf.layers[i].data.pop('activation')
            h = F.dropout(h, p=self.dropout, training=training)
            nf.layers[i].data['h'] = h
            nf.block_compute(i,
                             fn.copy_src(src='h', out='m'),
                             lambda node : {'h': node.mailbox['m'].mean(dim=1)},
                             layer)

        h = nf.layers[-1].data.pop('activation')

        return h

But, now, with graphbolt, it seems obtaining the graph data and aggregating it is not so simple. next(iter(subgraphs)).sampled_subgraphs[0].sampled_csc results in a sparse matrix of csc format, but I still don’t know how to proceed with aggregation because the copy_from_parent() attribute no longer exists and 'CSCFormatBase' object has no attribute 'layers'

May I request some guidance on how to handle the output from graphbolt’s neighbor sampler for message passing on the subgraph?

Thanks in advance.

#2

Here is an example of how to use graphbolt for node classification. Node Classification — DGL 2.3 documentation. The MiniBatch contains a blocks object, which includes the message flow graphs (MFGs) for message passing on the subgraph. The example also contains how to utilize the MFGs in model forward (Defining Model section).

1 Like