Enable name based definition of keras initializers in hyperparams.

PiperOrigin-RevId: 365913722

research/object_detection/builders/hyperparams_builder.py

@@ -359,7 +359,7 @@ def _build_initializer(initializer, build_for_keras=False):
       operators. If false builds for Slim.
 
   Returns:
-    tf initializer.
+    tf initializer or string corresponding to the tf keras initializer name.
 
   Raises:
     ValueError: On unknown initializer.
@@ -415,6 +415,13 @@ def _build_initializer(initializer, build_for_keras=False):
           factor=initializer.variance_scaling_initializer.factor,
           mode=mode,
           uniform=initializer.variance_scaling_initializer.uniform)
+  if initializer_oneof == 'keras_initializer_by_name':
+    if build_for_keras:
+      return initializer.keras_initializer_by_name
+    else:
+      raise ValueError(
+          'Unsupported non-Keras usage of keras_initializer_by_name: {}'.format(
+              initializer.keras_initializer_by_name))
   if initializer_oneof is None:
     return None
   raise ValueError('Unknown initializer function: {}'.format(

research/object_detection/builders/hyperparams_builder_test.py

@@ -1030,5 +1030,26 @@ def test_variance_in_range_with_random_normal_initializer_keras(self):
     self._assert_variance_in_range(initializer, shape=[100, 40],
                                    variance=0.64, tol=1e-1)
 
+  def test_keras_initializer_by_name(self):
+    conv_hyperparams_text_proto = """
+      regularizer {
+        l2_regularizer {
+        }
+      }
+      initializer {
+        keras_initializer_by_name: "glorot_uniform"
+      }
+    """
+    conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
+    text_format.Parse(conv_hyperparams_text_proto, conv_hyperparams_proto)
+    keras_config = hyperparams_builder.KerasLayerHyperparams(
+        conv_hyperparams_proto)
+    initializer_arg = keras_config.params()['kernel_initializer']
+    conv_layer = tf.keras.layers.Conv2D(
+        filters=16, kernel_size=3, **keras_config.params())
+    self.assertEqual(initializer_arg, 'glorot_uniform')
+    self.assertIsInstance(conv_layer.kernel_initializer,
+                          type(tf.keras.initializers.get('glorot_uniform')))
+
 if __name__ == '__main__':
   tf.test.main()

research/object_detection/protos/hyperparams.proto

@@ -88,6 +88,11 @@ message Initializer {
     TruncatedNormalInitializer truncated_normal_initializer = 1;
     VarianceScalingInitializer variance_scaling_initializer = 2;
     RandomNormalInitializer random_normal_initializer = 3;
+    // Allows specifying initializers by name, as a string, which will be passed
+    // directly as an argument during layer construction. Currently, this is
+    // only supported when using KerasLayerHyperparams, and for valid Keras
+    // initializers, e.g. `glorot_uniform`, `variance_scaling`, etc.
+    string keras_initializer_by_name = 4;
   }
 }