[API] remove the expand_as_pair call (#1929)

examples/pytorch/rgcn-hetero/model.py

@@ -101,7 +101,13 @@ class RelGraphConvLayer(nn.Module):
                      for i, w in enumerate(th.split(weight, 1, dim=0))}
         else:
             wdict = {}
-        inputs_src, inputs_dst = dglnn.expand_as_pair(inputs, g)
+
+        if g.is_block:
+            inputs_src = inputs
+            inputs_dst = {k: v[:g.number_of_dst_nodes(k)] for k, v in inputs.items()}
+        else:
+            inputs_src, inputs_dst = inputs
+
         hs = self.conv(g, inputs, mod_kwargs=wdict)
 
         def _apply(ntype, h):