[NN] Fix GATConv for Broadcasting with Residual Connections (#2867)

* Update

* update
Co-authored-by: Ubuntu <ubuntu@ip-172-31-59-108.us-west-2.compute.internal>

python/dgl/nn/pytorch/conv/gatconv.py

@@ -198,7 +198,8 @@ class GATConv(nn.Module):
             nn.init.xavier_normal_(self.fc_dst.weight, gain=gain)
         nn.init.xavier_normal_(self.attn_l, gain=gain)
         nn.init.xavier_normal_(self.attn_r, gain=gain)
-        nn.init.constant_(self.bias, 0)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0)
         if isinstance(self.res_fc, nn.Linear):
             nn.init.xavier_normal_(self.res_fc.weight, gain=gain)
 
@@ -304,7 +305,8 @@ class GATConv(nn.Module):
             rst = graph.dstdata['ft']
             # residual
             if self.res_fc is not None:
-                resval = self.res_fc(h_dst).view(h_dst.shape[0], self._num_heads, self._out_feats)
+                # Use -1 rather than self._num_heads to handle broadcasting
+                resval = self.res_fc(h_dst).view(h_dst.shape[0], -1, self._out_feats)
                 rst = rst + resval
             # bias
             if self.bias is not None:

tests/pytorch/test_nn.py

@@ -527,7 +527,7 @@ def test_rgcn_sorted(O):
 
 @parametrize_dtype
 @pytest.mark.parametrize('g', get_cases(['homo', 'block-bipartite'], exclude=['zero-degree']))
-@pytest.mark.parametrize('out_dim', [1, 2])
+@pytest.mark.parametrize('out_dim', [1, 5])
 @pytest.mark.parametrize('num_heads', [1, 4])
 def test_gat_conv(g, idtype, out_dim, num_heads):
     g = g.astype(idtype).to(F.ctx())
@@ -544,6 +544,11 @@ def test_gat_conv(g, idtype, out_dim, num_heads):
     _, a = gat(g, feat, get_attention=True)
     assert a.shape == (g.number_of_edges(), num_heads, 1)
 
+    # test residual connection
+    gat = nn.GATConv(5, out_dim, num_heads, residual=True)
+    gat = gat.to(ctx)
+    h = gat(g, feat)
+
 @parametrize_dtype
 @pytest.mark.parametrize('g', get_cases(['bipartite'], exclude=['zero-degree']))
 @pytest.mark.parametrize('out_dim', [1, 2])