Fix/20260 layers property setter

keras/src/layers/layer.py

@@ -1575,7 +1575,44 @@ def __repr__(self):
     def __str__(self):
         return self.__repr__()
 
+    # Attribute names reserved for internal tracking by the Layer class.
+    # Overriding these in subclasses will silently break weight
+    # saving/loading and other functionality.
+    _RESERVED_LAYER_ATTRIBUTES = frozenset(
+        {
+            "_layers",
+            "_metrics",
+            "_trainable_variables",
+            "_non_trainable_variables",
+            "_seed_generators",
+        }
+    )
+
+    # Internal modules that legitimately reassign reserved attributes.
+    _RESERVED_ATTR_EXEMPT_MODULES = frozenset(
+        {
+            "keras.src.models.sequential",
+            "keras.src.models.functional",
+        }
+    )
+
     def __setattr__(self, name, value):
+        # Warn if user code reassigns a reserved tracked attribute.
+        if (
+            name in self._RESERVED_LAYER_ATTRIBUTES
+            and hasattr(self, "_tracker")
+            and hasattr(self, name)
+            and tracking.is_tracking_enabled()
+            and type(self).__module__ not in self._RESERVED_ATTR_EXEMPT_MODULES
+        ):
+            warnings.warn(
+                f"`{name}` is a reserved attribute in Keras layers and "
+                "should not be used as a variable name in a Layer "
+                "subclass. Assigning to it can break weight saving, "
+                "metric tracking, and other functionality. "
+                f"Please use a different attribute name.",
+                stacklevel=2,
+            )
         # Track Variables, Layers, Metrics, SeedGenerators.
         name, value = self._setattr_hook(name, value)
         if name != "_tracker":

keras/src/layers/layer_test.py

@@ -1247,6 +1247,27 @@ def call(self, x):
         ):
             layer(np.random.random((3, 2)))
 
+    def test_reserved_attribute_warning(self):
+        """Warn when user code overrides reserved tracked attributes."""
+
+        class BadLayer(layers.Layer):
+            def __init__(self):
+                super().__init__()
+                self._layers = [layers.Dense(4)]
+
+        with self.assertWarnsRegex(
+            UserWarning,
+            "`_layers` is a reserved attribute",
+        ):
+            BadLayer()
+
+        # Internal Keras classes like Sequential should NOT warn.
+        import warnings
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("error", UserWarning)
+            models.Sequential([layers.Dense(4)])
+
     def test_init_after_state_tracking(self):
         class MyLayer(layers.Layer):
             def __init__(self):