Add option for relation to have different hidden channels

test/nn/kge/test_transd.py

@@ -4,8 +4,10 @@
 
 
 def test_transe():
-    model = TransD(num_nodes=10, num_relations=5, hidden_channels=32)
-    assert str(model) == 'TransD(10, num_relations=5, hidden_channels=32)'
+    model = TransD(num_nodes=10, num_relations=5, hidden_channels=32,
+                   rel_hidden_channels=64)
+    assert str(model) == ('TransD(10, num_relations=5, hidden_channels=32, '
+                          'rel_hidden_channels=64)')
 
     head_index = torch.tensor([0, 2, 4, 6, 8])
     rel_type = torch.tensor([0, 1, 2, 3, 4])

torch_geometric/nn/kge/transd.py

@@ -50,6 +50,9 @@ class TransD(KGEModel):
             (default: :obj:`1.0`)
         sparse (bool, optional): If set to :obj:`True`, gradients w.r.t. to the
             embedding matrices will be sparse. (default: :obj:`False`)
+        rel_hidden_channels (int, optional): The hidden embedding size of
+            relation if provided. Otherwise, relation embedding size has the
+            same size as :obj:`hidden_channels`. (default: :obj:`None`)
     """
     def __init__(
         self,
@@ -59,15 +62,22 @@ def __init__(
         margin: float = 1.0,
         p_norm: float = 1.0,
         sparse: bool = False,
+        rel_hidden_channels: int = None,
     ):
         super().__init__(num_nodes, num_relations, hidden_channels, sparse)
 
         self.p_norm = p_norm
         self.margin = margin
+        self.rel_hidden_channels = hidden_channels
+
+        if rel_hidden_channels is not None:
+            self.rel_hidden_channels = rel_hidden_channels
+            self.rel_emb = Embedding(num_relations, rel_hidden_channels,
+                                     sparse=sparse)
 
         self.node_proj_emb = Embedding(num_nodes, hidden_channels,
                                        sparse=sparse)
-        self.rel_proj_emb = Embedding(num_relations, hidden_channels,
+        self.rel_proj_emb = Embedding(num_relations, self.rel_hidden_channels,
                                       sparse=sparse)
 
         self.reset_parameters()
@@ -120,6 +130,12 @@ def loss(
             margin=self.margin,
         )
 
+    def __repr__(self) -> str:
+        return (f'{self.__class__.__name__}({self.num_nodes}, '
+                f'num_relations={self.num_relations}, '
+                f'hidden_channels={self.hidden_channels}, '
+                f'rel_hidden_channels={self.rel_hidden_channels})')
+
 
 def project_vector_ops(
     node_emb: Tensor,
@@ -141,5 +157,13 @@ def project_vector_ops(
         Tensor: The projected vector of node to relation embedding space.
     """
     out = (node_emb * node_proj_emb).sum(dim=-1).reshape(-1, 1)
-    out = out * rel_proj_emb + node_emb
+    out = out * rel_proj_emb
+
+    # Efficient identity matrix and node_emb multiplication
+    node_dim, rel_dim = node_emb.size(1), rel_proj_emb.size(1)
+    dim = min(node_dim, rel_dim)
+    node_identity = torch.zeros_like(rel_proj_emb)
+    node_identity[:, :dim] = node_emb[:, :dim]
+
+    out += node_identity
     return out