Convert (homogeneous) data with edge type "to_heterogeneous"?

[isratnisa]

Hi,
I want to use the new heterogeneous graph APIs to implement models like RGCN and use MUTAG and other common RGCN datasets. Looks like these dataset are loaded as (homogeneous) Data object with a list of edge type, not as a HeteroData object where each edge types is stored in disjunct storage objects. You can convert heteroData to homogeneous data by using to_homogeneous, is the opposite direction also possible? We already have the edge type for each edge! I want to use to_hetero on my model which require the metadata() information from heteroData. Please let me know if there is a easy way around.

[wsad1]

For datasets like Mutag you could directly use RGCNConv, it has a different parameters for each edge_type.

But if you want to convert something like Mutag to HeteroData, one way to do it would be

from torch_geometric.data import HeteroData
data = HeteroData()
data['node_type1'].x = mutag.x # assuming there's just one node type.
for i in mutag.edge_type.unique():
    data['node_type1', str(int(i)), 'node_type1'] = mutag.edge_index[:,mutage.edge_type ==i]

[rusty1s]

Note that combining RGCNConv and to_hetero does not make a lot of sense, since RGCNConv and to_hetero are similar in that they allow to operate on graphs with different relations. As an alternative, you can try the following:

   class GNN(torch.nn.Module):
       def __init__(self, hidden_channels, out_channels):
           super().__init__()
           self.conv1 = SAGEConv((hidden_channels, hidden_channels), hidden_channels)
           self.conv2 = SAGEConv((hidden_channels, hidden_channels), out_channels)

       def forward(self, x, edge_index):
           x = self.conv1(x, edge_index).relu()
           x = self.conv2(x, edge_index)
           return x

model = GNN(hidden_channels=32, out_channels=16)
model = to_hetero(model, data.metadata(), aggr='sum')

emb = torch.nn.Embedding(num_nodes, 32)

x_dict = data.x_dict
x_dict['node_type'] = emb.weight
model(x_dict, data.edge_index_dict)

[isratnisa]

@rusty1s: I agree! Actually, the goal is to compare runtime between these two versions. Do you loose any performance because of using hetero data structure. We have similar implementation and observed difference in runtime. Thanks for the quick responses.

@wsad1 : Can't use None as edge_type without changing the nn module. Will try that. Thanks!

[rusty1s]

I see. There shouldn't be any major differences in runtime, but I'm interested to hear your findings. Please keep me updated :)

[rusty1s]

See #3148

[isratnisa]

Ohh nice! Thank you!!

[faizan1234567]

@wsad1 how about when we have a multiple nodes type? I want to convert MUTAG datasets to a heterogenous graph structures with node and edge information like it is given in heterData object. I am not sure how to do that. Actually we want to convert TU datasets such MUTAG, COLLAB, IMDB-BINARY, etc etc to heterogenous graph for creating connection matrix.

[rusty1s]

This should work via to_heterogeneous() if node_type and edge_type are present attributes in your Data object. Can you confirm?