Cluster vectors for very high dimensional graphs

[hkim716]

Hi Matt, thanks for very helpful comments before.
I'm trying to use different dataset, which has tens of thousands nodes and edges.
This dataset also has fixed geometry with 3-dimensional pos values, x, y, z.
Each example looks like Data(edge_attr=[1459862, 3], edge_index=[2, 1459862], pos=[366883, 3], x=[366883, 1]), and I have 100 examples as entire dataset. (edge_attr, edge_index, and pos are the same for all examples but x(node feature), as my previous discussion)
I tried to obtain cluster vectors for a single example using

def pre_process(train_dataset, depth):
    train_loader_pre = DataLoader(train_dataset, batch_size=1, shuffle=False)
    clusters = []
    for data in train_loader_pre:
        for i in range(depth):
            x, edge_index, edge_attr, pos = data.x, data.edge_index, data.edge_attr, data.pos
            cluster = graclus(edge_index)
            cluster, _ = consecutive_cluster(cluster)
            clusters.append(cluster)
            data = max_pool(clusters[i], data, transform=transform)
        break
 
    return clusters

However, when I tried to run pre_process function, there was an error caused with depth > 5.
Sometimes I can obtain cluster vectors with depth=6 (succeeded once when I tried ten times), but most of the time, it's hard to get them with large value of depth.
I think it's because my dataset has too many nodes and edges to cluster, but is it common issue? or is there better way to do it?

[rusty1s]

I think the graclus method currently has problems dealing with isolated nodes. Can you check if there exists isolated nodes in your graph after a certain amount of coarsening?

[hkim716]

Thanks for your help Matt. I found some self-connected nodes on my original graph. There were 2432 self-connected nodes. However, even after I deleted them in my original dataset, the smallest size of cluster vector I can get is still [181596].
Here's my edge_index, which is the version without self-connected nodes, such as [285, 285].

edge_index.zip

[rusty1s]

The reason why graclus cannot further reduce the number of nodes is because there exists exactly 181590 isolated nodes in graph:

from torch_geometric.utils import degree
deg = degree(edge_index[0])
print((deg == 0).sum())

You can probably those in a pre-processing step:

edge_index, _ = subgraph(deg > 0, edge_index, relabel_nodes=True)

[hkim716]

Thank you so much Matt, that solved the problem!
In that case, is there any difference between when I'm doing it in pre-process function and doing this when I create original edge_index using MyDataset.py to get rid of all of isolated nodes?

class MyData(InMemoryDataset):
    ...

    def process(self):
        raw_data = np.load(self.raw_paths[0])

        x = torch.from_numpy(raw_data['x'])
        z = torch.from_numpy(raw_data['z'])

        pos = torch.stack([x, z], dim=-1)
        value = torch.from_numpy(raw_data['value'])
        y = torch.from_numpy(raw_data['label'])

        edge_index = torch.from_numpy(raw_data['edge_index'])
        edge_index -= 1
        edge_index = edge_index.t().contiguous()
        edge_index = to_undirected(edge_index)

        data_list = []
        for i in range(cp.size(0)):
            data = Data(x=value[i].view(-1,1),pos=pos.view(-1,2), edge_index=edge_index, y=y[i,:].view(-1))
            data_list.append(data)

        torch.save(self.collate(data_list), self.processed_paths[0])

[rusty1s]

There shouldn't be a difference. I suggest to do these kinds of transformations as a pre-processing step though to disentangle data representation and model.

[hkim716]

Thank you so much Matt. I really appreciate your help! I'm getting interested in this GNN world and hope I can also contribute on this awesome library later!