Delete tracking of model's optimizer if include_optimizer=False when saving.

PiperOrigin-RevId: 359145092
Change-Id: Id337879de49ba9197876c487e4ea7c76318d8959

tensorflow/python/keras/engine/base_layer.py

@@ -2794,8 +2794,10 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
     # For any super.__delattr__() call, we will directly use the implementation
     # in Trackable and skip the behavior in AutoTrackable. The Layer was
     # originally use Trackable as base class, the change of using Module as base
-    # class forced us to have AutoTrackable in the class hierarchy. Skipping
+    # class forced us to have AutoTrackable in the class hierarchy.
-    # the __delattr__ and __setattr__ in AutoTrackable will keep the status quo.
+    #
+    # TODO(b/180760306) Keeping the status quo of skipping _delattr__ and
+    # __setattr__ in AutoTrackable may be unsustainable.
     existing_value = getattr(self, name, None)
 
     # If this value is replacing an existing object assigned to an attribute, we
@@ -2901,8 +2903,8 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
 
       backend.track_variable(val)
 
-    # Skip the auto trackable from tf.Module to keep status quo. See the comment
+    # TODO(b/180760306) Skip the auto trackable from tf.Module to keep status
-    # at __delattr__.
+    # quo. See the comment at __delattr__.
     super(tracking.AutoTrackable, self).__setattr__(name, value)
 
   def _gather_children_attribute(self, attribute):

tensorflow/python/keras/engine/base_layer_v1.py

@@ -2148,8 +2148,10 @@ class Layer(base_layer.Layer):
     # For any super.__delattr__() call, we will directly use the implementation
     # in Trackable and skip the behavior in AutoTrackable. The Layer was
     # originally use Trackable as base class, the change of using Module as base
-    # class forced us to have AutoTrackable in the class hierarchy. Skipping
+    # class forced us to have AutoTrackable in the class hierarchy.
-    # the __delattr__ and __setattr__ in AutoTrackable will keep the status quo.
+    #
+    # TODO(b/180760306) Keeping the status quo of skipping _delattr__ and
+    # __setattr__ in AutoTrackable may be unsustainable.
     existing_value = getattr(self, name, None)
 
     # If this value is replacing an existing object assigned to an attribute, we
@@ -2257,8 +2259,8 @@ class Layer(base_layer.Layer):
 
       backend.track_variable(val)
 
-    # Skip the auto trackable from tf.Module to keep status quo. See the comment
+    # TODO(b/180760306) Skip the auto trackable from tf.Module to keep status
-    # at __delattr__.
+    # quo. See the comment at __delattr__.
     super(tracking.AutoTrackable, self).__setattr__(name, value)
 
   # This is a hack so that the is_layer (within

tensorflow/python/keras/saving/save_test.py

@@ -303,6 +303,28 @@ class TestSaveModel(test.TestCase, parameterized.TestCase):
 
     self.assertAllClose(batch_loss, new_batch_loss)
 
+  @combinations.generate(combinations.combine(mode=['eager', 'graph']))
+  def test_save_include_optimizer_false(self):
+
+    def get_variables(file_name):
+      reader = training_module.load_checkpoint(
+          os.path.join(file_name, 'variables/variables'))
+      shape_from_key = reader.get_variable_to_shape_map()
+      return sorted(shape_from_key.keys())
+
+    model = keras.models.Sequential()
+    model.add(keras.layers.Dense(1))
+    model.compile('adam', loss='mse')
+    x, y = np.ones((10, 10)), np.ones((10, 1))
+    model.train_on_batch(x, y)
+
+    path = os.path.join(self.get_temp_dir(), 'no_optimizer')
+    model.save(path, save_format='tf', include_optimizer=False)
+    variables = get_variables(path)
+
+    for v in variables:
+      self.assertNotIn('optimizer', v)
+
   @combinations.generate(combinations.combine(mode=['graph', 'eager']))
   def test_saving_model_with_custom_object(self):
     with generic_utils.custom_object_scope(), self.cached_session():

tensorflow/python/keras/saving/saved_model/load.py

@@ -50,7 +50,6 @@ from tensorflow.python.saved_model import nested_structure_coder
 from tensorflow.python.saved_model import revived_types
 from tensorflow.python.training.tracking import base as trackable
 from tensorflow.python.training.tracking import data_structures
-from tensorflow.python.training.tracking.tracking import delete_tracking
 from tensorflow.python.util import compat
 from tensorflow.python.util import nest
 
@@ -284,7 +283,7 @@ class KerasObjectLoader(object):
         # loading layers from the config, such as variables.
         continue
       for name in PUBLIC_ATTRIBUTES:
-        delete_tracking(node, name)
+        node._delete_tracking(name)  # pylint: disable=protected-access
 
       if isinstance(node, functional_lib.Functional):
         # Delete the temporary layer dependencies, which were used to restore
@@ -294,7 +293,7 @@ class KerasObjectLoader(object):
         dependencies = list(node._self_unconditional_dependency_names)  # pylint: disable=protected-access
         for name in dependencies:
           if re.match(r'^layer(_with_weights)?-[\d+]', name) is not None:
-            delete_tracking(node, name)
+            node._delete_tracking(name)  # pylint: disable=protected-access
 
   def _add_children_recreated_from_config(self, obj, proto, node_id):
     """Recursively records objects recreated from config."""

tensorflow/python/keras/saving/saved_model/save.py

@@ -32,7 +32,6 @@ from tensorflow.python.keras.utils.io_utils import ask_to_proceed_with_overwrite
 from tensorflow.python.platform import gfile
 from tensorflow.python.saved_model import save as save_lib
 
-
 # To avoid circular dependencies between keras/engine and keras/saving,
 # code in keras/saving must delay imports.
 
@@ -81,6 +80,9 @@ def save(model, filepath, overwrite, include_optimizer, signatures=None,
   if not include_optimizer:
     orig_optimizer = model.optimizer
     model.optimizer = None
+    # TODO(b/180760306) Change to del model.optimizer if Layer's __delattr__
+    # calls AutoTrackable's __delattr__.
+    model._delete_tracking("optimizer")  # pylint: disable=protected-access
 
   # Trace all functions and signatures with `training=0` instead of using an
   # already-set learning phase placeholder.

tensorflow/python/training/tracking/tracking.py

@@ -92,8 +92,7 @@ class AutoTrackable(base.Trackable):
     super(AutoTrackable, self).__setattr__(name, value)
 
   def __delattr__(self, name):
-    self._maybe_initialize_trackable()
+    self._delete_tracking(name)
-    delete_tracking(self, name)
     super(AutoTrackable, self).__delattr__(name)
 
   def _no_dependency(self, value):
@@ -125,6 +124,17 @@ class AutoTrackable(base.Trackable):
         functions[attribute_name] = attribute_value
     return functions
 
+  def _delete_tracking(self, name):
+    """Removes the tracking of name."""
+    self._maybe_initialize_trackable()
+    if name in self._unconditional_dependency_names:
+      del self._unconditional_dependency_names[name]
+      for index, (dep_name, _) in enumerate(
+          self._unconditional_checkpoint_dependencies):
+        if dep_name == name:
+          del self._unconditional_checkpoint_dependencies[index]
+          break
+
 
 def delete_tracking(obj, name):
   """Removes the tracking of name from object."""