Expected behavior of homogeneous to heterogeneous conversion of empty graphs

[ekosman]

Hi,

I was facing an issue with the conversion of homogeneous data to heterogeneous data that represent an empty graph.

When trying to convert an empty homogeneous graph into a heterogeneous graph, is it expected that the NodeStorage contains an edge_index attribute? As far as I checked, if the homogeneous graph is initialized with an empty tensor, i.e. edge_index=torch.tensor([[],[]]), the object that I get after converting it into a HeteroData object using the to_heterogeneous() method contains edge_index in the NodeStorage referenced by the sole node type.

Example:

import torch
from torch_geometric.data import Data

d = Data(x=torch.randn(5,10), edge_index=torch.tensor([[],[]]))
hd = d.to_heterogeneous()
sole_node_type = hd.node_types[0]
has_edge_index = "edge_index" in hd[sole_node_type]
print(f"Does node {sole_node_type} has the edge_index attribute: {has_edge_index}")
print(f"Still, there are {len(hd.edge_types)} edge types in the heterogeneous graph")

For reference, if we add an edge to the graph the behavior changes:

import torch
from torch_geometric.data import Data

d = Data(x=torch.randn(5,10), edge_index=torch.tensor([[1],[1]]))
hd = d.to_heterogeneous()
sole_node_type = hd.node_types[0]
has_edge_index = "edge_index" in hd[sole_node_type]
print(f"Does node {sole_node_type} has the edge_index attribute: {has_edge_index}. (should be False)")
print(f"Now, there is {len(hd.edge_types)} edge types in the heterogeneous graph (should be 1)")

In the first example, hd[sole_node_type] has an attribute called edge_index. In the second example, it doesn't. I expected the second behavior.

[rusty1s]

Yeah, this is indeed a bug, sorry about that. I fixed it in #6304. If you don't want to fix it by installing from source, you can pass the edge_type_names argument to prevent this from happening.