Heterogeneous graph classification task

[symd]

I have around 4000 heterodata objects in my dataset. But some nodes can be part of multiple graphs and all types of edges may not be present in all graphs. I have given a unique id for each node in each heterodata object. Do the node ids for these nodes needs to be unique across dataset?

While training I get the following error:
File "<eval_with_key>.5", line 110, in forward
pool_3 = torch.add(pool__nodetype1, pool__nodetype2); pool__nodetype1 = pool__nodetype2 = None
RuntimeError: The size of tensor a (64) must match the size of tensor b (62) at non-singleton dimension 0

I would like to know the reason and how to solve it

[rusty1s]

If there exists some graphs who miss node types or edge types, you need to create dummy tensors for them, e.g.:

data[node_type].x = torch.empty(0, 64)  # A feature matrix with 64 features but zero nodes
data[edge_type].edge_index = torch.empty(2, 0, dtype=torch.long)  # A matrix with zero edges

[symd]

Thank you. I have tried the same. But still the issue is there. I have given the following code
data[node_type].node_id=torch.empty(0).long()
data[node_type].x = torch.empty(0, 64)
data[edge_type].edge_index = torch.empty(2, 0, dtype=torch.long)
I have noticed that a number is repeating in ptr of one node type. Is there any problem in giving node_id like that?
I tried giving node_id as a tensor([0])

[rusty1s]

Do you have a minimal example to reproduce?

[symd]

edges.csv
graph.csv
nodes.csv

[rusty1s]

How do I need to load this data in?

[symd]

it worked when I replace the line
data[node_type].node_id=torch.empty(0).long()

with
data[node_type].node_id=torch.tensor([0])

Thank you