A question about metapath generation - AddMetaPaths package

[stupidoge]

Dear Author,

I meet a difficulty when running AddMetaPaths on my own datasets.

PS: my hetero_data orginal shape.

  company={
    x=[4264, 35],
    y=[4264]
  },
  audit_company={ x=[183, 1] },
  (company, to, audit_company)={ edge_index=[2, 10616] },
  (audit_company, to, company)={ edge_index=[2, 10616] }
)

Coding:

data_auditor_hetero = AddMetaPaths(metapaths=metapaths,drop_orig_edges=True,drop_unconnected_nodes=True)(data_auditor_hetero)

data_auditor_hetero = data_auditor_hetero.to(device)
print(data_auditor_hetero)

Here is the error feedback.

AssertionError                            Traceback (most recent call last)
~\AppData\Local\Temp\ipykernel_4872\3393051197.py in <module>
      2 # metapaths = [[('movie', 'actor'), ('actor', 'movie')],
      3 #              [('movie', 'director'), ('director', 'movie')]]
----> 4 data_auditor_hetero = AddMetaPaths(metapaths=metapaths,drop_orig_edges=True,drop_unconnected_nodes=True)(data_auditor_hetero)
      5 # data_auditor_hetero = AddMetaPaths(metapaths=metapaths)(data_auditor_hetero)
      6 

~\.conda\envs\gnn\lib\site-packages\torch_geometric\transforms\add_metapaths.py in __call__(self, data)
    130             adj1 = SparseTensor.from_edge_index(
    131                 edge_index=data[edge_type].edge_index,
--> 132                 sparse_sizes=data[edge_type].size(), edge_attr=edge_weight)
    133 
    134             if self.max_sample is not None:

~\.conda\envs\gnn\lib\site-packages\torch_sparse\tensor.py in from_edge_index(self, edge_index, edge_attr, sparse_sizes, is_sorted, trust_data)
     68         return SparseTensor(row=edge_index[0], rowptr=None, col=edge_index[1],
     69                             value=edge_attr, sparse_sizes=sparse_sizes,
---> 70                             is_sorted=is_sorted, trust_data=trust_data)
     71 
     72     @classmethod

~\.conda\envs\gnn\lib\site-packages\torch_sparse\tensor.py in __init__(self, row, rowptr, col, value, sparse_sizes, is_sorted, trust_data)
     36             csc2csr=None,
     37             is_sorted=is_sorted,
---> 38             trust_data=trust_data,
     39         )
     40 

~\.conda\envs\gnn\lib\site-packages\torch_sparse\storage.py in __init__(self, row, rowptr, col, value, sparse_sizes, rowcount, colptr, colcount, csr2csc, csc2csr, is_sorted, trust_data)
     64                 assert rowptr.numel() - 1 == M
     65             elif row is not None and row.numel() > 0:
---> 66                 assert trust_data or int(row.max()) < M
     67 
     68         N: int = 0

AssertionError: 

Could you help me out? I check tens of times, but still couldn't find where is the mistake.

[EdisonLeeeee]

Hi, what does your metapaths look like?

[stupidoge]

Should I just copy your code into console?

[EdisonLeeeee]

That's the problem. Your edge_index contains out-of-bound index. You need to carefully check your data again.

[stupidoge]

Actually, I re-labeled those index. The company has 4338 nodes, the audit company has 183 nodes. I give company from index 1 - 4338, and give audit company index from 4339 - 4520. Am i wrong?

Does the package could recognize same index from different rows?

data_auditor_hetero['company','audit_company'].edge_index

tensor([[   0,    0,    0,  ..., 4336, 4337, 4337],
        [4338, 4339, 4340,  ..., 4357, 4436, 4371]], device='cuda:0')

I mean, index from row 1 means company, and index from row 2 means audit company?

[EdisonLeeeee]

Your index in row 2 should be in the range of [0, 183]. You don't need to increment the edge index. It's treated as a bipartite graph in PyG.

[stupidoge]

Thanks for your kindness! Solved!