Changed edge_label_index after LinkNeighborLoader

[songsong0425]

Before starting, sorry for the frequent question.
While I running the LinkNeighborLoader code, I found that the edge_label_index was changed.
For example, I have a graph dataset and split it into the train, validation, and test dataset. (For simplicity, I only consider the training dataset)

train_data
# Data(edge_index=[2, 2159112], pos_edge_label=[343], pos_edge_label_index=[2, 343], neg_edge_label=[343], neg_edge_label_index=[2, 343], edge_label=[686], edge_label_index=[2, 686], x=[47822, 259])
train_data.edge_label_index
# tensor([[ 5296,   613,   322,  ..., 19970, 24669, 27129],
#         [18473,  9092,  9092,  ...,  9153, 12671, 19081]])

When I run the LinkNeighborLoader, the edge_label_index is changed to the below format:

train_loader = LinkNeighborLoader(train_data, edge_label_index=train_data.edge_label_index, batch_size=3, shuffle=True, neg_sampling_ratio=1.0, num_neighbors=[3,3])

for data in train_loader:
    print(data.edge_label_index)
# tensor([[ 1,  3,  5,  4,  7,  6],
#         [ 2,  0, 10,  8, 11,  9]])
# tensor([[ 0,  2,  6, 10,  1, 11],
#         [ 9,  3,  5,  7,  8,  4]])
# ...

Although I confirmed that the model learned the features and read the PyG documentation (https://pytorch-geometric.readthedocs.io/en/latest/advanced/batching.html), I cannot understand why the range of the edge index is changed.
If I performed the wrong process, please tell me where should I fix it.

[akihironitta]

Hey @songsong0425! I reported this issue via PyG Slack last night, and #7791 has just fixed the issue!

[songsong0425]

Thank you for announcing the Slack channel. I'll explore it for more information. Once again, thank you very much for your help!

[EdisonLeeeee]

You're welcome. If you have any questions or need further assistance, don't hesitate to reach out. We're always here to help :)

[pstrybol]

Hi @EdisonLeeeee , I also noticed this recently but I fail to see how the new labels are mapped? The node_id and n_id are still the same as in the original HeterData object?

[EdisonLeeeee]

Hi @pstrybol , the node index in a batch is remapped to a new index and currently, we do not provide the mapping from the original index to the new one. For n_id, it represents the global node index of each node in the batch, i.e., the original node index.

[pstrybol]

Thanks for the reply! The reason I would need the node mapping is because I want to calculate the correlation between the predicted probabilities and original scores but I guess I could just input custom edge_index after each epoch for example, in eval() mode so it does not update weights? I also had some more question around LinkNeighborLoader but I asked them in a separate discussion topic #7797.
Cheers