[Python] How to handel heterogenous gnn?

I have created this data:

HeteroData(
user={ x=[100, 16] },
keyword={ x=[321, 16] },
tweet={ x=[1000, 16] },
(user, follow, user)={ edge_index=[2, 291] },
(user, tweetedby, tweet)={ edge_index=[2, 1000] },
(keyword, haskeyword, tweet)={ edge_index=[2, 3752] }
)


And these two varibales based on that:

x_dict:
user: torch.Size([100, 16])
keyword: torch.Size([321, 16])
tweet: torch.Size([1000, 16])

edge_index_dict:
('user', 'follow', 'user'): torch.Size([2, 291])
('user', 'tweetedby', 'tweet'): torch.Size([2, 1000])
('keyword', 'haskeyword', 'tweet'): torch.Size([2, 3752])


My nodes were indexed from 0 to 99 for users, from 100 to 420 for keywords, and next for tweet nodes.

When I want to run this model:

class HeteroGATBinaryClassifier(torch.nn.Module):
def __init__(self, metadata, in_channels, hidden_channels, heads=1):
super().__init__()
self.metadata = metadata # Metadata about node and edge types

# Define GNN layers for each edge type
self.conv1 = HeteroConv({
edge_type: GATConv(in_channels, hidden_channels, heads=heads, add_self_loops = False)
for edge_type in metadata[1]
}, aggr='mean') # Aggregate using mean

self.conv2 = HeteroConv({
edge_type: GATConv(hidden_channels * heads, hidden_channels, heads=heads, add_self_loops = False)
for edge_type in metadata[1]
}, aggr='mean')

# Linear layer for classification (binary output)
self.classifier = Linear(hidden_channels * heads, 1) # Single output for binary classification

def forward(self, x_dict, edge_index_dict, target_node_type):
# First GAT layer
x_dict = self.conv1(x_dict, edge_index_dict)
x_dict = {key: F.elu(x) for key, x in x_dict.items()}

# Second GAT layer
x_dict = self.conv2(x_dict, edge_index_dict)
x_dict = {key: F.elu(x) for key, x in x_dict.items()}

# Apply the classifier only on the target node type
logits = self.classifier(x_dict[target_node_type]) # Logits for target node type
return torch.sigmoid(logits) # Output probabilities for binary classification


I see this error:

IndexError: Found indices in 'edge_index' that are larger than 999 (got 1420). Please ensure that all indices in 'edge_index' point to valid indices in the interval [0, 1000) in your node feature matrix and try again.
​

Do you have any idea how to handle it?

Continue reading...