Checks if input tf.keras.Model is compiled

Instead of checking for attribute, check that attribute value is not None.

PiperOrigin-RevId: 264409455

tensorflow_federated/python/learning/keras_utils.py

@@ -194,15 +194,15 @@ def from_compiled_keras_model(keras_model, dummy_batch):
     ValueError: If `keras_model` was *not* compiled.
   """
   py_typecheck.check_type(keras_model, tf.keras.Model)
+  # Optimizer attribute is only set after calling tf.keras.Model.compile().
+  if not keras_model.optimizer:
+    raise ValueError('`keras_model` must be compiled. Use from_keras_model() '
+                     'instead.')
   dummy_tensors = _preprocess_dummy_batch(dummy_batch)
   # NOTE: A sub-classed tf.keras.Model does not produce the compiled metrics
   # until the model has been called on input. The work-around is to call
   # Model.test_on_batch() once before asking for metrics.
   keras_model.test_on_batch(**dummy_tensors)
-  # Optimizer attribute is only set after calling tf.keras.Model.compile().
-  if not hasattr(keras_model, 'optimizer'):
-    raise ValueError('`keras_model` must be compiled. Use from_keras_model() '
-                     'instead.')
   return model_utils.enhance(_TrainableKerasModel(keras_model, dummy_tensors))
 
 

tensorflow_federated/python/learning/keras_utils_test.py

@@ -117,6 +117,15 @@ class KerasUtilsTest(test.TestCase, parameterized.TestCase):
           dummy_batch=_create_dummy_batch(1),
           loss=tf.keras.losses.MeanSquaredError())
 
+  def test_from_compiled_keras_model_fails_on_uncompiled_model(self):
+    keras_model = model_examples.build_linear_regression_keras_functional_model(
+        feature_dims=1)
+
+    with self.assertRaisesRegex(ValueError, '`keras_model` must be compiled'):
+      keras_utils.from_compiled_keras_model(
+          keras_model=keras_model,
+          dummy_batch=_create_dummy_batch(feature_dims=1))
+
   # Test class for batches using namedtuple.
   _make_test_batch = collections.namedtuple('TestBatch', ['x', 'y'])