ValueError: Target size (torch.Size([1490])) must be the same as input size (torch.Size([100, 1490]))

[joshuajennysibbu]

I have a few questions:

• Is negative sampling is required for edge level prediction
• Why model return torch.Size([100, 1490], My batch size 32 why it return 100. How to fix this issue.

 class GNN(torch.nn.Module):
              def __init__(self, hidden_channels, out_channels):
                    super().__init__()
                    self.conv1 = SAGEConv((-1, -1), hidden_channels,normalize=True)
                    self.conv2 = SAGEConv((-1, -1), out_channels,normalize=True)

           def forward(self, source_,target_, edge_index_):
        
            source=source_.to(torch.float32)
        target=target_.to(torch.float32)
        edge_index_=edge_index_.to(torch.long)
        
        edge_index_=edge_index_.T
        
        print(self.conv1)
        source_conv = self.conv1((source,target),edge_index_)
        source_conv = self.conv1((source,target),edge_index_)
        print(source_conv.shape)
        source_conv = self.conv2(source_conv,edge_index_)
        
        return source_conv

loss = F.binary_cross_entropy_with_logits(output, link_labels)

[rusty1s]

• In general, negative sampling is required if you want to infer missing links, since you need to train against both existing links and non-existing links. For edge-level property prediction, you do not need to make use of negative sampling.
• This is hard to tell. I assume you are using the LinkNeighborLoader? In that case, the loader will sample nodes outside of your current mini-batch that are used to aggregate information inside your GNN. You need to slice your output embeddings afterwards to fit to the correct shape: x = x[:batch_size].

[rusty1s]

Can you clarify? For a bipartite graph, you could do:

data = HeteroData()
data['user'].x = ...
data['product'].x = ...
data['user, 'product'].edge_index = ...
data['product', 'user'].edge_index = ...

and then pass it to the LinkNeighborLoader:

loader = LingNeighborLoader(data, num_neighbors=[10, 10],
                            edge_label_index=('user', 'product'), ...)

[joshuajennysibbu]

Okay... and What will be the output label for the bipartite graph?

node=[-1,0,1] #INDEX : 0,1,2
edge_index=[[0,1],[1,2]]

Is this the right way for the edge index? I use the node index as an edge index

[rusty1s]

I am not sure I understand. What do you mean with output label? For the HeteroData, the edge_index attribute will point from the nodes in the source node type to indices in the destination node type, e.g.:

data = HeteroData()
data['user'].x = torch.randn(5, 16)  # 5 nodes
data['product'].x = torch.randn(2, 16)  # 2 nodes
data['user, 'product'].edge_index = torch.tensor([[0, 1, 2, 3, 4], [0, 1, 0, 1, 0]])

[joshuajennysibbu]

I am not sure I understand. What do you mean with output label? For the HeteroData, the edge_index attribute will point from the nodes in the source node type to indices in the destination node type, e.g.:

data = HeteroData()
data['user'].x = torch.randn(5, 16)  # 5 nodes
data['product'].x = torch.randn(2, 16)  # 2 nodes
data['user, 'product'].edge_index = torch.tensor([[0, 1, 2, 3, 4], [0, 1, 0, 1, 0]])

I'm new to Graph Networks. how to do link prediction. Can you give me a simple code.

[rusty1s]

Please take a look at https://github.com/pyg-team/pytorch_geometric/blob/master/examples/hetero/hetero_link_pred.py.