Provide option to perform in-place batch norm updates for ssd feature extractors.

PiperOrigin-RevId: 190688309

research/object_detection/builders/model_builder.py

@@ -95,13 +95,19 @@ def build(model_config, is_training, add_summaries=True):
 
 
 def _build_ssd_feature_extractor(feature_extractor_config, is_training,
-                                 reuse_weights=None):
+                                 reuse_weights=None,
+                                 inplace_batchnorm_update=False):
   """Builds a ssd_meta_arch.SSDFeatureExtractor based on config.
 
   Args:
     feature_extractor_config: A SSDFeatureExtractor proto config from ssd.proto.
     is_training: True if this feature extractor is being built for training.
     reuse_weights: if the feature extractor should reuse weights.
+    inplace_batchnorm_update: Whether to update batch_norm inplace during
+      training. This is required for batch norm to work correctly on TPUs. When
+      this is false, user must add a control dependency on
+      tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+      norm moving average parameters.
 
   Returns:
     ssd_meta_arch.SSDFeatureExtractor based on config.
@@ -126,7 +132,8 @@ def _build_ssd_feature_extractor(feature_extractor_config, is_training,
   return feature_extractor_class(is_training, depth_multiplier, min_depth,
                                  pad_to_multiple, conv_hyperparams,
                                  batch_norm_trainable, reuse_weights,
-                                 use_explicit_padding, use_depthwise)
+                                 use_explicit_padding, use_depthwise,
+                                 inplace_batchnorm_update)
 
 
 def _build_ssd_model(ssd_config, is_training, add_summaries):
@@ -140,6 +147,7 @@ def _build_ssd_model(ssd_config, is_training, add_summaries):
 
   Returns:
     SSDMetaArch based on the config.
+
   Raises:
     ValueError: If ssd_config.type is not recognized (i.e. not registered in
       model_class_map).
@@ -147,8 +155,9 @@ def _build_ssd_model(ssd_config, is_training, add_summaries):
   num_classes = ssd_config.num_classes
 
   # Feature extractor
-  feature_extractor = _build_ssd_feature_extractor(ssd_config.feature_extractor,
-                                                   is_training)
+  feature_extractor = _build_ssd_feature_extractor(
+      ssd_config.feature_extractor, is_training,
+      ssd_config.inplace_batchnorm_update)
 
   box_coder = box_coder_builder.build(ssd_config.box_coder)
   matcher = matcher_builder.build(ssd_config.matcher)
@@ -194,7 +203,8 @@ def _build_ssd_model(ssd_config, is_training, add_summaries):
 
 
 def _build_faster_rcnn_feature_extractor(
-    feature_extractor_config, is_training, reuse_weights=None):
+    feature_extractor_config, is_training, reuse_weights=None,
+    inplace_batchnorm_update=False):
   """Builds a faster_rcnn_meta_arch.FasterRCNNFeatureExtractor based on config.
 
   Args:
@@ -202,6 +212,11 @@ def _build_faster_rcnn_feature_extractor(
       faster_rcnn.proto.
     is_training: True if this feature extractor is being built for training.
     reuse_weights: if the feature extractor should reuse weights.
+    inplace_batchnorm_update: Whether to update batch_norm inplace during
+      training. This is required for batch norm to work correctly on TPUs. When
+      this is false, user must add a control dependency on
+      tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+      norm moving average parameters.
 
   Returns:
     faster_rcnn_meta_arch.FasterRCNNFeatureExtractor based on config.
@@ -209,6 +224,8 @@ def _build_faster_rcnn_feature_extractor(
   Raises:
     ValueError: On invalid feature extractor type.
   """
+  if inplace_batchnorm_update:
+    raise ValueError('inplace batchnorm updates not supported.')
   feature_type = feature_extractor_config.type
   first_stage_features_stride = (
       feature_extractor_config.first_stage_features_stride)
@@ -238,6 +255,7 @@ def _build_faster_rcnn_model(frcnn_config, is_training, add_summaries):
 
   Returns:
     FasterRCNNMetaArch based on the config.
+
   Raises:
     ValueError: If frcnn_config.type is not recognized (i.e. not registered in
       model_class_map).
@@ -246,7 +264,8 @@ def _build_faster_rcnn_model(frcnn_config, is_training, add_summaries):
   image_resizer_fn = image_resizer_builder.build(frcnn_config.image_resizer)
 
   feature_extractor = _build_faster_rcnn_feature_extractor(
-      frcnn_config.feature_extractor, is_training)
+      frcnn_config.feature_extractor, is_training,
+      frcnn_config.inplace_batchnorm_update)
 
   number_of_stages = frcnn_config.number_of_stages
   first_stage_anchor_generator = anchor_generator_builder.build(

research/object_detection/builders/model_builder_test.py

@@ -297,6 +297,7 @@ class ModelBuilderTest(tf.test.TestCase):
   def test_create_ssd_mobilenet_v1_model_from_config(self):
     model_text_proto = """
       ssd {
+        inplace_batchnorm_update: true
         feature_extractor {
           type: 'ssd_mobilenet_v1'
           conv_hyperparams {
@@ -519,6 +520,7 @@ class ModelBuilderTest(tf.test.TestCase):
   def test_create_faster_rcnn_resnet_v1_models_from_config(self):
     model_text_proto = """
       faster_rcnn {
+        inplace_batchnorm_update: true
         num_classes: 3
         image_resizer {
           keep_aspect_ratio_resizer {

research/object_detection/meta_architectures/ssd_meta_arch.py

@@ -46,7 +46,8 @@ class SSDFeatureExtractor(object):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """Constructor.
 
     Args:
@@ -64,6 +65,10 @@ class SSDFeatureExtractor(object):
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False.
       use_depthwise: Whether to use depthwise convolutions. Default is False.
+      inplace_batchnorm_update: Whether to update batch norm moving average
+        values inplace. When this is false train op must add a control
+        dependency on tf.graphkeys.UPDATE_OPS collection in order to update
+        batch norm statistics.
     """
     self._is_training = is_training
     self._depth_multiplier = depth_multiplier
@@ -71,6 +76,7 @@ class SSDFeatureExtractor(object):
     self._pad_to_multiple = pad_to_multiple
     self._conv_hyperparams = conv_hyperparams
     self._batch_norm_trainable = batch_norm_trainable
+    self._inplace_batchnorm_update = inplace_batchnorm_update
     self._reuse_weights = reuse_weights
     self._use_explicit_padding = use_explicit_padding
     self._use_depthwise = use_depthwise
@@ -108,7 +114,29 @@ class SSDFeatureExtractor(object):
       feature_maps: a list of tensors where the ith tensor has shape
         [batch, height_i, width_i, depth_i]
     """
-    pass
+    batchnorm_updates_collections = (None if self._inplace_batchnorm_update
+                                     else tf.GraphKeys.UPDATE_OPS)
+
+    with slim.arg_scope([slim.batch_norm],
+                        updates_collections=batchnorm_updates_collections):
+      return self._extract_features(preprocessed_inputs)
+
+  @abstractmethod
+  def _extract_features(self, preprocessed_inputs):
+    """Extracts features from preprocessed inputs.
+
+    This function is responsible for extracting feature maps from preprocessed
+    images.
+
+    Args:
+      preprocessed_inputs: a [batch, height, width, channels] float tensor
+        representing a batch of images.
+
+    Returns:
+      feature_maps: a list of tensors where the ith tensor has shape
+        [batch, height_i, width_i, depth_i]
+    """
+    raise NotImplementedError
 
 
 class SSDMetaArch(model.DetectionModel):

research/object_detection/models/embedded_ssd_mobilenet_v1_feature_extractor.py

@@ -53,7 +53,8 @@ class EmbeddedSSDMobileNetV1FeatureExtractor(
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """MobileNetV1 Feature Extractor for Embedded-friendly SSD Models.
 
     Args:
@@ -71,6 +72,11 @@ class EmbeddedSSDMobileNetV1FeatureExtractor(
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False.
       use_depthwise: Whether to use depthwise convolutions. Default is False.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
 
     Raises:
       ValueError: upon invalid `pad_to_multiple` values.
@@ -82,9 +88,9 @@ class EmbeddedSSDMobileNetV1FeatureExtractor(
     super(EmbeddedSSDMobileNetV1FeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, batch_norm_trainable, reuse_weights,
-        use_explicit_padding, use_depthwise)
+        use_explicit_padding, use_depthwise, inplace_batchnorm_update)
 
-  def extract_features(self, preprocessed_inputs):
+  def _extract_features(self, preprocessed_inputs):
     """Extract features from preprocessed inputs.
 
     Args:

research/object_detection/models/ssd_inception_v2_feature_extractor.py

@@ -37,7 +37,8 @@ class SSDInceptionV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """InceptionV2 Feature Extractor for SSD Models.
 
     Args:
@@ -55,11 +56,16 @@ class SSDInceptionV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False.
       use_depthwise: Whether to use depthwise convolutions. Default is False.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
     """
     super(SSDInceptionV2FeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, batch_norm_trainable, reuse_weights,
-        use_explicit_padding, use_depthwise)
+        use_explicit_padding, use_depthwise, inplace_batchnorm_update)
 
   def preprocess(self, resized_inputs):
     """SSD preprocessing.
@@ -76,7 +82,7 @@ class SSDInceptionV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
     """
     return (2.0 / 255.0) * resized_inputs - 1.0
 
-  def extract_features(self, preprocessed_inputs):
+  def _extract_features(self, preprocessed_inputs):
     """Extract features from preprocessed inputs.
 
     Args:

research/object_detection/models/ssd_inception_v3_feature_extractor.py

@@ -37,7 +37,8 @@ class SSDInceptionV3FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """InceptionV3 Feature Extractor for SSD Models.
 
     Args:
@@ -55,11 +56,16 @@ class SSDInceptionV3FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False.
       use_depthwise: Whether to use depthwise convolutions. Default is False.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
     """
     super(SSDInceptionV3FeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, batch_norm_trainable, reuse_weights,
-        use_explicit_padding, use_depthwise)
+        use_explicit_padding, use_depthwise, inplace_batchnorm_update)
 
   def preprocess(self, resized_inputs):
     """SSD preprocessing.
@@ -76,7 +82,7 @@ class SSDInceptionV3FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
     """
     return (2.0 / 255.0) * resized_inputs - 1.0
 
-  def extract_features(self, preprocessed_inputs):
+  def _extract_features(self, preprocessed_inputs):
     """Extract features from preprocessed inputs.
 
     Args:

research/object_detection/models/ssd_mobilenet_v1_feature_extractor.py

@@ -38,7 +38,8 @@ class SSDMobileNetV1FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """MobileNetV1 Feature Extractor for SSD Models.
 
     Args:
@@ -57,11 +58,16 @@ class SSDMobileNetV1FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
         inputs so that the output dimensions are the same as if 'SAME' padding
         were used.
       use_depthwise: Whether to use depthwise convolutions. Default is False.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
     """
     super(SSDMobileNetV1FeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, batch_norm_trainable, reuse_weights,
-        use_explicit_padding, use_depthwise)
+        use_explicit_padding, use_depthwise, inplace_batchnorm_update)
 
   def preprocess(self, resized_inputs):
     """SSD preprocessing.
@@ -78,7 +84,7 @@ class SSDMobileNetV1FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
     """
     return (2.0 / 255.0) * resized_inputs - 1.0
 
-  def extract_features(self, preprocessed_inputs):
+  def _extract_features(self, preprocessed_inputs):
     """Extract features from preprocessed inputs.
 
     Args:

research/object_detection/models/ssd_mobilenet_v2_feature_extractor.py

@@ -39,7 +39,8 @@ class SSDMobileNetV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """MobileNetV2 Feature Extractor for SSD Models.
 
     Mobilenet v2 (experimental), designed by sandler@. More details can be found
@@ -60,11 +61,16 @@ class SSDMobileNetV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False.
       use_depthwise: Whether to use depthwise convolutions. Default is False.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
     """
     super(SSDMobileNetV2FeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, batch_norm_trainable, reuse_weights,
-        use_explicit_padding, use_depthwise)
+        use_explicit_padding, use_depthwise, inplace_batchnorm_update)
 
   def preprocess(self, resized_inputs):
     """SSD preprocessing.
@@ -81,7 +87,7 @@ class SSDMobileNetV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
     """
     return (2.0 / 255.0) * resized_inputs - 1.0
 
-  def extract_features(self, preprocessed_inputs):
+  def _extract_features(self, preprocessed_inputs):
     """Extract features from preprocessed inputs.
 
     Args:

research/object_detection/models/ssd_resnet_v1_fpn_feature_extractor.py

@@ -43,7 +43,8 @@ class _SSDResnetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """SSD FPN feature extractor based on Resnet v1 architecture.
 
     Args:
@@ -66,6 +67,11 @@ class _SSDResnetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False. UNUSED currently.
       use_depthwise: Whether to use depthwise convolutions. UNUSED currently.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
 
     Raises:
       ValueError: On supplying invalid arguments for unused arguments.
@@ -73,7 +79,7 @@ class _SSDResnetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
     super(_SSDResnetV1FpnFeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, batch_norm_trainable, reuse_weights,
-        use_explicit_padding)
+        use_explicit_padding, inplace_batchnorm_update)
     if self._depth_multiplier != 1.0:
       raise ValueError('Only depth 1.0 is supported, found: {}'.
                        format(self._depth_multiplier))
@@ -110,7 +116,7 @@ class _SSDResnetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
         filtered_image_features[feature_name] = feature
     return filtered_image_features
 
-  def extract_features(self, preprocessed_inputs):
+  def _extract_features(self, preprocessed_inputs):
     """Extract features from preprocessed inputs.
 
     Args:
@@ -176,7 +182,8 @@ class SSDResnet50V1FpnFeatureExtractor(_SSDResnetV1FpnFeatureExtractor):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """Resnet50 v1 FPN Feature Extractor for SSD Models.
 
     Args:
@@ -194,11 +201,17 @@ class SSDResnet50V1FpnFeatureExtractor(_SSDResnetV1FpnFeatureExtractor):
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False. UNUSED currently.
       use_depthwise: Whether to use depthwise convolutions. UNUSED currently.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
     """
     super(SSDResnet50V1FpnFeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, resnet_v1.resnet_v1_50, 'resnet_v1_50', 'fpn',
-        batch_norm_trainable, reuse_weights, use_explicit_padding)
+        batch_norm_trainable, reuse_weights, use_explicit_padding,
+        inplace_batchnorm_update)
 
 
 class SSDResnet101V1FpnFeatureExtractor(_SSDResnetV1FpnFeatureExtractor):
@@ -212,7 +225,8 @@ class SSDResnet101V1FpnFeatureExtractor(_SSDResnetV1FpnFeatureExtractor):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """Resnet101 v1 FPN Feature Extractor for SSD Models.
 
     Args:
@@ -230,11 +244,17 @@ class SSDResnet101V1FpnFeatureExtractor(_SSDResnetV1FpnFeatureExtractor):
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False. UNUSED currently.
       use_depthwise: Whether to use depthwise convolutions. UNUSED currently.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
     """
     super(SSDResnet101V1FpnFeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, resnet_v1.resnet_v1_101, 'resnet_v1_101', 'fpn',
-        batch_norm_trainable, reuse_weights, use_explicit_padding)
+        batch_norm_trainable, reuse_weights, use_explicit_padding,
+        inplace_batchnorm_update)
 
 
 class SSDResnet152V1FpnFeatureExtractor(_SSDResnetV1FpnFeatureExtractor):
@@ -248,7 +268,8 @@ class SSDResnet152V1FpnFeatureExtractor(_SSDResnetV1FpnFeatureExtractor):
                batch_norm_trainable=True,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False):
+               use_depthwise=False,
+               inplace_batchnorm_update=False):
     """Resnet152 v1 FPN Feature Extractor for SSD Models.
 
     Args:
@@ -266,8 +287,14 @@ class SSDResnet152V1FpnFeatureExtractor(_SSDResnetV1FpnFeatureExtractor):
       use_explicit_padding: Whether to use explicit padding when extracting
         features. Default is False. UNUSED currently.
       use_depthwise: Whether to use depthwise convolutions. UNUSED currently.
+      inplace_batchnorm_update: Whether to update batch_norm inplace during
+        training. This is required for batch norm to work correctly on TPUs.
+        When this is false, user must add a control dependency on
+        tf.GraphKeys.UPDATE_OPS for train/loss op in order to update the batch
+        norm moving average parameters.
     """
     super(SSDResnet152V1FpnFeatureExtractor, self).__init__(
         is_training, depth_multiplier, min_depth, pad_to_multiple,
         conv_hyperparams, resnet_v1.resnet_v1_152, 'resnet_v1_152', 'fpn',
-        batch_norm_trainable, reuse_weights, use_explicit_padding)
+        batch_norm_trainable, reuse_weights, use_explicit_padding,
+        inplace_batchnorm_update)

research/object_detection/protos/faster_rcnn.proto

@@ -131,6 +131,12 @@ message FasterRcnn {
   // to use sigmoid loss and enable merge_multiple_label_boxes.
   // If not specified, Softmax loss is used as default.
   optional ClassificationLoss second_stage_classification_loss = 29;
+
+  // Whether to update batch_norm inplace during training. This is required
+  // for batch norm to work correctly on TPUs. When this is false, user must add
+  // a control dependency on tf.GraphKeys.UPDATE_OPS for train/loss op in order
+  // to update the batch norm moving average parameters.
+  optional bool inplace_batchnorm_update = 30 [default = false];
 }
 
 

research/object_detection/protos/ssd.proto

@@ -59,6 +59,12 @@ message Ssd {
 
   // Loss configuration for training.
   optional Loss loss = 11;
+
+  // Whether to update batch_norm inplace during training. This is required
+  // for batch norm to work correctly on TPUs. When this is false, user must add
+  // a control dependency on tf.GraphKeys.UPDATE_OPS for train/loss op in order
+  // to update the batch norm moving average parameters.
+  optional bool inplace_batchnorm_update = 15 [default = false];
 }