fix: layers were not deserialized for Sequential and Residual

bayesflow/networks/residual/residual.py

@@ -12,7 +12,12 @@
 @serializable("bayesflow.networks")
 class Residual(Sequential):
     def __init__(self, *layers: keras.Layer, **kwargs):
-        if len(layers) == 1 and isinstance(layers[0], Sequence):
+        if len(layers) == 0 and "layers" in kwargs:
+            # extract layers from kwargs, in case they were passed as a keyword argument
+            layers = kwargs.pop("layers")
+        elif len(layers) > 0 and "layers" in kwargs:
+            raise ValueError("Layers passed both as positional argument and as keyword argument")
+        elif len(layers) == 1 and isinstance(layers[0], Sequence):
             layers = layers[0]
         super().__init__(list(layers), **sequential_kwargs(kwargs))
         self.projector = keras.layers.Dense(units=None, name="projector")
@@ -43,6 +48,8 @@ def build(self, input_shape=None):
         # this is a work-around for https://github.com/keras-team/keras/issues/21158
         output_shape = input_shape
         for layer in self._layers:
+            if layer.built:
+                continue
             layer.build(output_shape)
             output_shape = layer.compute_output_shape(output_shape)
 

bayesflow/networks/sequential/sequential.py

@@ -31,9 +31,14 @@ class Sequential(keras.Layer):
     """
 
     def __init__(self, *layers: keras.Layer | Sequence[keras.Layer], **kwargs):
-        super().__init__(**layer_kwargs(kwargs))
-        if len(layers) == 1 and isinstance(layers[0], Sequence):
+        if len(layers) == 0 and "layers" in kwargs:
+            # extract layers from kwargs, in case they were passed as a keyword argument
+            layers = kwargs.pop("layers")
+        elif len(layers) > 0 and "layers" in kwargs:
+            raise ValueError("Layers passed both as positional argument and as keyword argument")
+        elif len(layers) == 1 and isinstance(layers[0], Sequence):
             layers = layers[0]
+        super().__init__(**layer_kwargs(kwargs))
 
         self._layers = layers
 
@@ -44,6 +49,8 @@ def build(self, input_shape):
             return
 
         for layer in self._layers:
+            if layer.built:
+                continue
             layer.build(input_shape)
             input_shape = layer.compute_output_shape(input_shape)
 

tests/test_networks/test_residual/conftest.py

@@ -0,0 +1,15 @@
+import pytest
+
+from bayesflow.networks.residual import Residual
+
+
+@pytest.fixture()
+def residual():
+    import keras
+
+    return Residual(keras.layers.Flatten(), keras.layers.Dense(2))
+
+
+@pytest.fixture()
+def build_shapes():
+    return {"input_shape": (32, 2)}

tests/test_networks/test_residual/test_residual.py

@@ -0,0 +1,24 @@
+import keras
+
+from bayesflow.utils.serialization import deserialize, serialize
+
+from ...utils import assert_layers_equal
+
+
+def test_serialize_deserialize(residual, build_shapes):
+    residual.build(**build_shapes)
+
+    serialized = serialize(residual)
+    deserialized = deserialize(serialized)
+    reserialized = serialize(deserialized)
+
+    assert reserialized == serialized
+
+
+def test_save_and_load(tmp_path, residual, build_shapes):
+    residual.build(**build_shapes)
+
+    keras.saving.save_model(residual, tmp_path / "model.keras")
+    loaded = keras.saving.load_model(tmp_path / "model.keras")
+
+    assert_layers_equal(residual, loaded)

tests/test_networks/test_sequential/conftest.py

@@ -0,0 +1,15 @@
+import pytest
+
+from bayesflow.networks import Sequential
+
+
+@pytest.fixture()
+def sequential():
+    import keras
+
+    return Sequential(keras.layers.Flatten(), keras.layers.Dense(2))
+
+
+@pytest.fixture()
+def build_shapes():
+    return {"input_shape": (32, 2)}

tests/test_networks/test_sequential/test_sequential.py

@@ -0,0 +1,24 @@
+import keras
+
+from bayesflow.utils.serialization import deserialize, serialize
+
+from ...utils import assert_layers_equal
+
+
+def test_serialize_deserialize(sequential, build_shapes):
+    sequential.build(**build_shapes)
+
+    serialized = serialize(sequential)
+    deserialized = deserialize(serialized)
+    reserialized = serialize(deserialized)
+
+    assert reserialized == serialized
+
+
+def test_save_and_load(tmp_path, sequential, build_shapes):
+    sequential.build(**build_shapes)
+
+    keras.saving.save_model(sequential, tmp_path / "model.keras")
+    loaded = keras.saving.load_model(tmp_path / "model.keras")
+
+    assert_layers_equal(sequential, loaded)