RandomLinkSplit expects reverse edge types for heterogeneous undirected graphs

[saiden89]

🐛 Describe the bug

When splitting heterogeneous undirected graphs RandomLinkSplit expects rev_edge_types to be passed, which in this scenario is ambiguous as no reverse edges technically exists. If None is passed then an assert isinstance(rev_edge_types, list) is triggered. If, on the other hand, the very same edge_types are passed, then a different AttributeError is raised.

To reproduce:

import torch
from torch_geometric.data import HeteroData
from torch_geometric.transforms import RandomLinkSplit


edge_index = torch.tensor([[0, 1, 2, 1, 2, 3], [1, 2, 3, 0, 1, 2]])
x = torch.rand(4, 2)

hetero_data = HeteroData()
hetero_data['n'] = x
hetero_data['n', 'link', 'n'].edge_index = edge_index
data = HeteroData(x=x, edge_index=edge_index)
assert data.is_undirected()
RandomLinkSplit(is_undirected=True, edge_types=hetero_data.edge_types)(hetero_data)

Traceback

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
issue.ipynb Cell 1’ in <cell line: 14>()
     12 data = HeteroData(x=x, edge_index=edge_index)
     13 assert data.is_undirected()
---> 14 RandomLinkSplit(is_undirected=True, edge_types=hetero_data.edge_types)(hetero_data)

File torch_geometric/transforms/random_link_split.py:107, in RandomLinkSplit.__init__(self, num_val, num_test, is_undirected, key, split_labels, add_negative_train_samples, neg_sampling_ratio, disjoint_train_ratio, edge_types, rev_edge_types)
    104 self.rev_edge_types = rev_edge_types
    106 if isinstance(edge_types, list):
--> 107     assert isinstance(rev_edge_types, list)
    108     assert len(edge_types) == len(rev_edge_types)

AssertionError:

Explicitly passing rev_edge_types:

RandomLinkSplit(is_undirected=True, edge_types=hetero_data.edge_types, rev_edge_types=hetero_data.edge_types)(hetero_data)

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
issue.ipynb Cell 1' in <cell line: 14>()
     12 data = HeteroData(x=x, edge_index=edge_index_1)
     13 assert data.is_undirected()
---> 14 RandomLinkSplit(is_undirected=True, edge_types=hetero_data.edge_types, rev_edge_types=hetero_data.edge_types)(hetero_data)

File torch_geometric/transforms/random_link_split.py:195, in RandomLinkSplit.__call__(self, data)
    191 num_neg_test = int(num_test * self.neg_sampling_ratio)
    193 num_neg = num_neg_train + num_neg_val + num_neg_test
--> 195 size = store.size()
    196 if store._key is None or store._key[0] == store._key[-1]:
    197     size = size[0]

File torch_geometric/data/storage.py:365, in EdgeStorage.size(self, dim)
    361 if self._key is None:
    362     raise NameError("Unable to infer 'size' without explicit "
    363                     "'_key' assignment")
--> 365 size = (self._parent()[self._key[0]].num_nodes,
    366         self._parent()[self._key[-1]].num_nodes)
    368 return size if dim is None else size[dim]

AttributeError: 'Tensor' object has no attribute 'num_nodes'

Environment

• PyG version: 2.0.4+master
• PyTorch version: 1.12+nightly
• OS: Linux
• Python version: 3.10.4
• CUDA/cuDNN version: 10.2
• How you installed PyTorch and PyG (conda, pip, source): conda for torch, pip for pyg and dependencies.

[rusty1s]

It looks like hetero_data is somewhat wrongly created. This fixes the issue:

import torch
from torch_geometric.data import HeteroData, Data
from torch_geometric.transforms import RandomLinkSplit

edge_index = torch.tensor([[0, 1, 2, 1, 2, 3], [1, 2, 3, 0, 1, 2]])
x = torch.rand(4, 2)

hetero_data = HeteroData()
hetero_data['n'].x = x
hetero_data['n', 'link', 'n'].edge_index = edge_index

RandomLinkSplit(
    is_undirected=True,
    edge_types=hetero_data.edge_types,
)(hetero_data)

[saiden89]

Yup, my bad for wrongly creating the object.
The proposed solution still triggers the assert, but if the user manually assigns rev_edge_types to be hetero_data.edge_types it passes. I think it is still a bit misleading to force the user to manually specify the reverse edges in an undirected context.

[rusty1s]

Which PyG version are you operating on? This does not throw an error for me.

[saiden89]

I was on 2.0.4 stable, indeed upgrading to master gets rid of the assert.

[saiden89]

Thank you for the help, but I'm still confused by the underlying reason why if the user sets rev_edge_types=hetero_data.edge_types the splitter returns directed objects (even though is_undirected=True is specified), whereas if rev_edge_types=None the returned splits are undirected.
Could somebody please elaborate on this behavior?

train, val, test = RandomLinkSplit(
    is_undirected=True,
    edge_types=hetero_data.edge_types,
    rev_edge_types=hetero_data.edge_types
)(hetero_data)
print(train['n', 'link', 'n'].is_undirected())
>>> False

train, val, test = RandomLinkSplit(
    is_undirected=True,
    edge_types=hetero_data.edge_types,
    rev_edge_types=None
)(hetero_data)
print(train['n', 'link', 'n'].is_undirected())
>>> True

[rusty1s]

Sorry for the late reply. This is indeed a bug. I fixed it in #4757.