hierarchical pooling methods for different size graph inputs

[mfarazi1991]

Hi,

I have this very big issue on graph classification of varying size graphs (primarily meshes with 10k+ nodes) using the Cheb_conv method. In all examples, there is just one graph fixed (they assume fixed size graphs) and then use pooling once since the topology is fixed. However, when the inputs are not of fixed size, I cannot figure out how to do pooling on each instance separately. Mainly I have issues with how to batch the input graphs (they have different laplacian and dimensions).

[rusty1s]

Can you give a small example where exactly you have issues? In general, you can simply use the torch_geometric.data.DataLoader to generate mini-batches out of our examples, and these can be well inputted into ChebConv as well. For pooling, all pooling operators support the batch vector argument, which is necessary to take care of correctly pooling graphs in mini-batch mode.

[rusty1s]

This should be possible. All pooling operators should work just fine in mini-batch mode. Your example looks correct to me. What is the error message you encounter?

Note that you do not need to (re-)compute Laplacian in case you use normalization="sym" within ChebConv.

[mfarazi1991]

This is the error I get either with default code (your example on Graclus clustering) with MNIST and my dataset.
86
87 if edge_weight is not None:
---> 88 assert edge_weight.numel() == edge_index.size(1)
89 loop_weight = edge_weight.new_full((N, ), fill_value)
90 edge_weight = torch.cat([edge_weight, loop_weight], dim=0)

Also, in other codes available for ChebNet paper, there is a line of code similar in all of them:

L, perm = coarsen(A, coarsening_levels)  --> A is a fixed graph shared among all, in my case A is varying in size
g_arr = [dgl.from_scipy(csr) for csr in L]

Therefore, the batching is based on the fixed Ls and perm but in my case these are varying.

[rusty1s]

Are you calling add_self_loops anywhere in your code? You might wanna check why the shapes of edge_index and edge_weight do not align in your case. The MNIST Graclus examples provided by PyG runs fine for me.

[mfarazi1991]

I found the error. It seems data.x = F.elu(self.conv1(data.x, data.edge_index)) is true not adding edge_attr to the argument for the latest PyG. I removed it and it worked. But I am not sure why the last Graclus calling does not change the number of nodes (I used some print-out functions to track the number of nodes) and after the last one which comes with max_pool_x instead of max_pool function, does not change the number of nodes at all. Any thoughts? Btw, huge thanks for your swift and helpful response.

[rusty1s]

The number of nodes depend on the number of unique values in cluster, i.e., cluster.unique().numel(). In case this value aligns with the number of output nodes, this indicates that graclus cannot downscale the graph any further.