GraphSAGE with edge features?

[j-adamczyk]

I am trying to implement GraphSAGE variant using edge features. I am using OGB implementation of GCN and GIN with edge features as a starting point: https://github.com/snap-stanford/ogb/blob/master/examples/graphproppred/mol/conv.py.

In PyG SAGEConv we have:

    def forward(self, x: Union[Tensor, OptPairTensor], edge_index: Adj,
                size: Size = None) -> Tensor:
        """"""
        if isinstance(x, Tensor):
            x: OptPairTensor = (x, x)

        if self.project and hasattr(self, 'lin'):
            x = (self.lin(x[0]).relu(), x[1])

        # propagate_type: (x: OptPairTensor)
        out = self.propagate(edge_index, x=x, size=size)
        out = self.lin_l(out)

        x_r = x[1]
        if self.root_weight and x_r is not None:
            out += self.lin_r(x_r)

        if self.normalize:
            out = F.normalize(out, p=2., dim=-1)

        return out


    def message(self, x_j: Tensor) -> Tensor:
        return x_j

    def message_and_aggregate(self, adj_t: SparseTensor,
                              x: OptPairTensor) -> Tensor:
        adj_t = adj_t.set_value(None, layout=None)
        return matmul(adj_t, x[0], reduce=self.aggr)

    def aggregate(self, x: Tensor, index: Tensor, ptr: Optional[Tensor] = None,
                  dim_size: Optional[int] = None) -> Tensor:
        if self.aggr is not None:
            return scatter(x, index, dim=self.node_dim, dim_size=dim_size,
                           reduce=self.aggr)

        # LSTM aggregation:
        if ptr is None and not torch.all(index[:-1] <= index[1:]):
            raise ValueError(f"Can not utilize LSTM-style aggregation inside "
                             f"'{self.__class__.__name__}' in case the "
                             f"'edge_index' tensor is not sorted by columns. "
                             f"Run 'sort_edge_index(..., sort_by_row=False)' "
                             f"in a pre-processing step.")

        x, mask = to_dense_batch(x, batch=index, batch_size=dim_size)
        out, _ = self.lstm(x)
        return out[:, -1]

So I guess I have to change 2 things:

1. Add BondEncoder from OGB to embed edge features to the same dimension as node features
2. Overwrite message and aggregate methods

I tried, basing on #3544 (reply in thread):

from ogb.graphproppred.mol_encoder import BondEncoder
from torch_geometric.nn import SAGEConv, MessagePassing
from torch_geometric.nn.models.basic_gnn import BasicGNN


class SAGEConvEdge(SAGEConv):
    def __init__(self, emb_dim: int, *args, **kwargs):
        super(SAGEConvE).__init__(*args, **kwargs)
        self.bond_encoder = BondEncoder(emb_dim)

    def forward(self, x, edge_index, edge_attr):
        out = self.propagate(edge_index, x=x, edge_attr=edge_attr)
        return self.lin2(x) + out

    def message(self, x_j, edge_attr):
        return self.lin1(x_j + edge_attr)


class GraphSageEdge(BasicGNN):
    def init_conv(
        self, in_channels: int, out_channels: int, **kwargs
    ) -> MessagePassing:
        return GraphSageE(in_channels, out_channels, **kwargs)

But this does not work, as there are no lin1 and lin2 in SAGEConv. How can I do this?

[rusty1s]

    def forward(self, x, edge_index, edge_attr):
        out = self.propagate(edge_index, x=x, edge_attr=edge_attr)
        return self.lin_r(x) + out

    def message(self, x_j, edge_attr):
        return self.lin_l(x_j) + self.bond_encoder(edge_attr)

should do the trick.

[j-adamczyk]

It works, thank you