Will HeteroData be supported in Batch?

[YAndrewL]

Hi, I've been using torch_geometric.data.Batch to generate disconnected graph batches, but it seems to not work with HeteroData and raised error 'index out of range'. I'm wondering will the function will be supported with HeteroData in this version?

[rusty1s]

This should be supported, but it only got added in recently. Can you pull from master and see if that fixes your issues? Other-wise, please provide a minimal code snippet to reproduce so I can look into it.

[YAndrewL]

Got it! Very clear explanation, thanks a lot!

[YAndrewL]

Sorry 😭 Still have a question.
According to this example:

from torch_geometric.nn import SAGEConv, to_hetero

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

    def forward(self, x, edge_index):
        x = self.conv1(x, edge_index).relu()
        x = self.conv2(x, edge_index)
        return x


model = GNN(in_channels=-1, hidden_channels=64, out_channels=dataset.num_classes)
model = to_hetero(model, data.metadata(), aggr='sum')

I simple substitute SAGEConv to GATConv, but it raised an error: AttributeError: 'GATConv' object has no attribute 'lin_src'
Why did this happen?

[rusty1s]

Thanks for spotting and beta-testing our release :) This should be now fixed in master. I also tried to make it more clear in the documentation, see here. The issue is that we need to pass in a tuple as in_channels to each GNN operator to allow for message passing in bipartite graphs. You can create a heterogeneous graph attention network as follows:

from torch_geometric.nn import GATConv, Linear, to_hetero

class GAT(torch.nn.Module):
    def __init__(self, hidden_channels, out_channels):
        super().__init__()
        self.conv1 = GATConv((-1, -1), hidden_channels, add_self_loops=False)
        self.lin1 = Linear(-1, hidden_channels)
        self.conv2 = GATConv((-1, -1), out_channels, add_self_loops=False)
        self.lin2 = Linear(-1, out_channels)

    def forward(self, x, edge_index):
        x = self.conv1(x, edge_index) + self.lin1(x)
        x = x.relu()
        x = self.conv2(x, edge_index) + self.lin2(x)
        return x


model = GAT(hidden_channels=64, out_channels=dataset.num_classes)
model = to_hetero(model, data.metadata(), aggr='sum')

[YAndrewL]

Thanks, I'm unfamiliar with the paper, is this model consistent with this paper?

[rusty1s]

The above example describes a variant of R-GCN in which the GNN operator is described via graph attention. The HAN operator is quite different, as it computes attention across metapaths and uses a semantic-level aggregation to combine representations along the metapath. We have plans to add HAN to PyG as well in a future release.