fix: seed regional decoder observation nodes with encoder features

graph_weather/models/regional_forecast.py

@@ -274,14 +274,17 @@ def forward(
             nodes = torch.cat([features[i], regional_h3], dim=0)
             nodes = self.node_encoder(nodes)
             nodes, _ = self.encoder_gnn(nodes, enc_graph.edge_index, enc_edge_attr)
+            obs_features = nodes[:num_obs]
             h3_features = nodes[num_obs:]
 
             # Process: N rounds of H3 message passing
             h3_features = self.processor(h3_features, lat_graph.edge_index, latent_edge_attr)
 
-            # Decode: H3 -> obs through reversed bipartite GNN
-            obs_placeholders = torch.zeros(num_obs, self.config.node_dim, device=features.device)
-            dec_nodes = torch.cat([obs_placeholders, h3_features], dim=0)
+            # Decode: H3 -> obs through reversed bipartite GNN.
+            # Seed the decoder's observation nodes with their own encoded features rather
+            # than zeros, so the predicted delta can specialize per observation instead of
+            # only seeing its cell's pooled feature (mirrors the StretchedForecaster fix).
+            dec_nodes = torch.cat([obs_features, h3_features], dim=0)
             dec_nodes, _ = self.decoder_gnn(dec_nodes, dec_edge_index, dec_edge_attr)
             obs_out = self.node_decoder(dec_nodes[:num_obs])
             batch_outputs.append(obs_out)

tests/test_regional_forecast.py

@@ -125,6 +125,28 @@ def test_residual_connection():
     assert torch.allclose(out, expected_residual, atol=1e-5)
 
 
+def test_decoder_seeded_with_encoder_obs_features():
+    """The decoder's observation nodes carry the encoder's per-obs features, not zeros.
+
+    Seeding them with zeros collapses the model toward persistence, since the predicted delta
+    then only sees cell-level context. This guards that regression.
+    """
+    model = _small_config().build()
+    features = torch.randn(1, 5, 16)
+
+    captured = {}
+
+    def capture(_module, inputs, _output):
+        captured["dec_input"] = inputs[0].detach()
+
+    handle = model.decoder_gnn.register_forward_hook(capture)
+    model(features, _uk_latlons())
+    handle.remove()
+
+    obs_rows = captured["dec_input"][:5]
+    assert obs_rows.abs().sum() > 0
+
+
 def _nudging_config():
     """Config with nudging enabled."""
     return RegionalForecasterConfig(