D2-equivariant GATv2.

PiperOrigin-RevId: 602424499

clrs/_src/processors.py

@@ -289,6 +289,64 @@ def __call__(  # pytype: disable=signature-mismatch  # numpy-scalars
     return ret, None  # pytype: disable=bad-return-type  # numpy-scalars
 
 
+class GATv2FullD2(GATv2):
+  """Graph Attention Network v2 with full adjacency matrix and D2 symmetry."""
+
+  def d2_forward(self,
+                 node_fts: List[_Array],
+                 edge_fts: List[_Array],
+                 graph_fts: List[_Array],
+                 adj_mat: _Array,
+                 hidden: _Array,
+                 **unused_kwargs) -> List[_Array]:
+    num_d2_actions = 4
+
+    d2_inverses = [
+        0, 1, 2, 3  # All members of D_2 are self-inverses!
+    ]
+
+    d2_multiply = [
+        [0, 1, 2, 3],
+        [1, 0, 3, 2],
+        [2, 3, 0, 1],
+        [3, 2, 1, 0],
+    ]
+
+    assert len(node_fts) == num_d2_actions
+    assert len(edge_fts) == num_d2_actions
+    assert len(graph_fts) == num_d2_actions
+
+    ret_nodes = []
+    adj_mat = jnp.ones_like(adj_mat)
+
+    for g in range(num_d2_actions):
+      emb_values = []
+      for h in range(num_d2_actions):
+        gh = d2_multiply[d2_inverses[g]][h]
+        node_features = jnp.concatenate(
+            (node_fts[g], node_fts[gh]),
+            axis=-1)
+        edge_features = jnp.concatenate(
+            (edge_fts[g], edge_fts[gh]),
+            axis=-1)
+        graph_features = jnp.concatenate(
+            (graph_fts[g], graph_fts[gh]),
+            axis=-1)
+        cell_embedding = super().__call__(
+            node_fts=node_features,
+            edge_fts=edge_features,
+            graph_fts=graph_features,
+            adj_mat=adj_mat,
+            hidden=hidden
+        )
+        emb_values.append(cell_embedding)
+      ret_nodes.append(
+          jnp.mean(jnp.stack(emb_values, axis=0), axis=0)
+      )
+
+    return ret_nodes
+
+
 class GATv2Full(GATv2):
   """Graph Attention Network v2 with full adjacency matrix."""