python - How to handle heterogenous GNN? - Stack Overflow

admin管理员组

文章数量:1328950

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

But I see the following 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?

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

But I see the following 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?

• python
• pytorch
• pytorch-geometric
• gnn

Share Improve this question Follow edited Mar 10 at 10:16 daqh 1441212 bronze badges asked Dec 10, 2024 at 0:42 aliiiiiiiiiiiiiiiiiiiiialiiiiiiiiiiiiiiiiiiiii 31733 silver badges1010 bronze badges 2

• The edge_index tensors are supposed to contain tuples of node indices denoting what two nodes are connected by an edge. This means all the edge values should be less than the number of nodes. Your edge index has values outside the node count, ie pointing to nodes that don't exist. Requires more information to say why that is happening – Karl Commented Dec 10, 2024 at 6:40
• Dear @Karl, thank you for your comment. I think the issue lies in the fact that, for example, the indexes of the keyword nodes start from 321, but when referring to them during access time, the indexing begins from zero. Would it be correct to define an offset and update the edge indexes accordingly? – aliiiiiiiiiiiiiiiiiiiii Commented Dec 10, 2024 at 9:46

Add a comment  | 

1 Answer 1

Sorted by: Reset to default 0

Of course you are having troubles with indices mismatch between node feature matrix and the edge_index.

The edge index must be a tensor with shape (2, number_of_edges) and with values < num_nodes.
Each column of the edge index represent and edge and it is used to access the matrix of node features through the convolution process.

Probably, in the program you are running, you have 1000 nodes, and you didn't aligned edge indices correctly because you removed node features without updating the edge index or added nodes to the edge index without updating the node features.

It is very important that indices of edge index are aligned and consistent with node features, if not, you must add an offset to node features or normalize edge indices depending on what is your issue:

I usually do something like this on dim 0 or 1 to normalize src or dst of the edge index:

_, edge_index[0] = torch.unique(edge_index[0], return_inverse=True)

本文标签： pythonHow to handle heterogenous GNNStack Overflow