Fix export for subclass models with multiple inputs. (#19720)

The export now supports subclasses of `Model` for which the `call` method takes more than one input argument. Note that it is required for the model class to implement a `build` method with a signature that matches the `call` method.

Author: hertschuh

keras/src/export/export_lib.py

@@ -621,18 +621,17 @@ def export_model(model, filepath):
             input_signature = [input_signature]
         export_archive.add_endpoint("serve", model.__call__, input_signature)
     else:
-        save_spec = _get_save_spec(model)
-        if not save_spec or not model._called:
+        input_signature = _get_input_signature(model)
+        if not input_signature or not model._called:
             raise ValueError(
                 "The model provided has never called. "
                 "It must be called at least once before export."
             )
-        input_signature = [save_spec]
         export_archive.add_endpoint("serve", model.__call__, input_signature)
     export_archive.write_out(filepath)
 
 
-def _get_save_spec(model):
+def _get_input_signature(model):
     shapes_dict = getattr(model, "_build_shapes_dict", None)
     if not shapes_dict:
         return None
@@ -654,16 +653,7 @@ def make_tensor_spec(structure):
                 f"Unsupported type {type(structure)} for {structure}"
             )
 
-    if len(shapes_dict) == 1:
-        value = list(shapes_dict.values())[0]
-        return make_tensor_spec(value)
-
-    specs = {}
-    for key, value in shapes_dict.items():
-        key = key.rstrip("_shape")
-        specs[key] = make_tensor_spec(value)
-
-    return specs
+    return [make_tensor_spec(value) for value in shapes_dict.values()]
 
 
 @keras_export("keras.layers.TFSMLayer")

keras/src/export/export_lib_test.py

@@ -145,6 +145,31 @@ def call(self, inputs):
         )
         revived_model.serve(bigger_input)
 
+    def test_model_with_multiple_inputs(self):
+
+        class TwoInputsModel(models.Model):
+            def call(self, x, y):
+                return x + y
+
+            def build(self, y_shape, x_shape):
+                self.built = True
+
+        temp_filepath = os.path.join(self.get_temp_dir(), "exported_model")
+        model = TwoInputsModel()
+        ref_input_x = tf.random.normal((3, 10))
+        ref_input_y = tf.random.normal((3, 10))
+        ref_output = model(ref_input_x, ref_input_y)
+
+        export_lib.export_model(model, temp_filepath)
+        revived_model = tf.saved_model.load(temp_filepath)
+        self.assertAllClose(
+            ref_output, revived_model.serve(ref_input_x, ref_input_y)
+        )
+        # Test with a different batch size
+        revived_model.serve(
+            tf.random.normal((6, 10)), tf.random.normal((6, 10))
+        )
+
     @parameterized.named_parameters(
         named_product(model_type=["sequential", "functional", "subclass"])
     )