the principle of ClusterData

[Chrisable]

As I known,ClusterData uses the method of metis.But how to get the result of partition?
What does cluster_data.partptr stand for？
I understand, all_ node_index[partptr[0]:partptr[1]] is partition1, all_ node_index[partptr[1]:partptr[2]] is partition2.
If so,What is the difference between this and random node partitioning，If not,how to get the result of partition?

Here is my code:

import torch
from torch_geometric.datasets import Planetoid
from torch_geometric.transforms import NormalizeFeatures

dataset = Planetoid(root='data/Planetoid', name='PubMed', transform=NormalizeFeatures())
data = dataset[0]
from torch_geometric.loader import ClusterData, ClusterLoader
torch.manual_seed(12345)
cluster_data = ClusterData(data, num_parts=2)  # 1. Create subgraphs.
print("cluster_data.partptr:",cluster_data.partptr)
train_loader = ClusterLoader(cluster_data, batch_size=1, shuffle=True)
total_num_nodes = 0
for step, sub_data in enumerate(train_loader):
    print(f'Step {step + 1}:')
    print('=======')
    print(f'Number of nodes in the current batch: {sub_data.num_nodes}')
    print(sub_data.edge_index)
    total_num_nodes += sub_data.num_nodes

print(f'Iterated over {total_num_nodes} of {data.num_nodes} nodes!')

[rusty1s]

The easiest way to get the final permutation and partition is to add a node ID attribute to your data:

data.n_id = torch.arange(data.num_nodes)
for subdata in ClusterData(data, …):
    print(subdata.n_id)

[Chrisable]

Is the result of each experiment of ClusterData the same？when i use (edgecuts, parts) = metis.part_graph(G, 2),i found that the partition results of Metis and clusterdata are different

[rusty1s]

I think the METIS algorithm should be deterministic but not 100% sure TBH.

[Chrisable]

Thank you for your reply!By the way,What does cluster_data.partptr stand for？In my example,cluster_data.partptr: tensor([ 0, 9870, 19717])

[Chrisable]

right?

[rusty1s]

Yes :)

[karuna-bhaila]

Hi, just wanted to ask for some more clarification on this.

dataset = Planetoid(root='data/Planetoid', name='PubMed', transform=NormalizeFeatures())
data = dataset[0]
cluster_data = ClusterData(data, num_parts=2)

data.y contains the node labels in the original order that it was read in and cluster_data.data.y contains node labels in the new permuted order so that cluster_data.data.y[partptr[0]:partptr[1]] contains node labels for all nodes in the 1st cluster? Is this correct?

[rusty1s]

Yes, this is correct.

[karuna-bhaila]

Thank you for your reply