Issue with adding negative sampling with "RandomLinkSplit" for a heterogeneous graph

[HamidHadipour]

Hi community,
I am new to Pyg, I followed the example code in https://github.com/pyg-team/pytorch_geometric/blob/master/examples/hetero/hetero_link_pred.py and I need to do link prediction on a heterogeneous graph which is technically a bi-partite graph.
My task is link prediction and I have the dataset of connections or (label = 1) so, I need to do negative sampling. But as time as I add a ratio except "0.0" for the "negative sampling ration" property I get "Assertion error"

torch-geometric==2.1.0

Here is the data definition:
`
data = HeteroData()

data['chem'].num_nodes = len(cp_mapping)
data['protein'].num_nodes = len(cp_mapping_p)

data['chem', 'CPI', 'protein'].edge_index = edge_index
data['chem', 'CPI', 'protein'].edge_label = edge_label

data['chem'].x = torch.eye(data['chem'].num_nodes)
del data['chem'].num_nodes
data['protein'].x = torch.eye(data['protein'].num_nodes)
del data['protein'].num_nodes
`

Here is the result:
HeteroData( chem={ x=[1206, 1206] }, protein={ x=[152, 152] }, (chem, CPI, protein)={ edge_index=[2, 20000], edge_label=[20000] } )

Here is how I use RandomLinkSplit:
`

data = T.ToUndirected()(data)
del data['protein', 'rev_CPI', 'chem'].edge_label
train_data, val_data, test_data = T.RandomLinkSplit(
num_val=0.2,
num_test=0.2,
neg_sampling_ratio=1.0,
edge_types=[('chem', 'CPI', 'protein')],
rev_edge_types=[('protein', 'rev_CPI', 'chem')]
)(data)

`

Would you please take a look and correct me, or suggest me a better way?
@rusty1s

[rusty1s]

We can only add negative edges in case you want to perform classification, in which case edge_label needs to be of type torch.long. I assume it is a floating-point type for you.

[HamidHadipour]

Thanks for your reply.
Would you take a look again?
Actually, my problem is classification and the labels are just "1".
Here is what I have as edge_label and edge_index:

Thank you.

[rusty1s]

Ah, it looks like edge_label_index and edge_label don't match in size. You need a label for each edge, e.g., edge_label_index.size(1) == edge_label.size(0).