Merged commit includes the following changes:

191649512  by Zhichao Lu:

    Introduce two parameters in ssd.proto - freeze_batchnorm, inplace_batchnorm_update - and set up slim arg_scopes in ssd_meta_arch.py such that applies it to all batchnorm ops in the predict() method.

    This centralizes the control of freezing and doing inplace batchnorm updates.

--
191620303  by Zhichao Lu:

    Modifications to the preprocessor to support multiclass scores

--
191610773  by Zhichao Lu:

    Adding multiclass_scores to InputDataFields and adding padding for multiclass_scores.

--
191595011  by Zhichao Lu:

    Contains implementation of the detection metric for the Open Images Challenge.

--
191449408  by Zhichao Lu:

    Change hyperparams_builder to return a callable so the users can inherit values from outer arg_scopes. This allows us to easily set batch_norm parameters like "is_training" and "inplace_batchnorm_update" for all feature extractors from the base class and propagate it correctly to the nested scopes.

--
191437008  by Zhichao Lu:

    Contains implementation of the Recall@N and MedianRank@N metrics.

--
191385254  by Zhichao Lu:

    Add config rewrite flag to eval.py

--
191382500  by Zhichao Lu:

    Fix bug for config_util.

--

PiperOrigin-RevId: 191649512

research/object_detection/builders/box_predictor_builder.py

@@ -49,12 +49,12 @@ def build(argscope_fn, box_predictor_config, is_training, num_classes):
 
   if  box_predictor_oneof == 'convolutional_box_predictor':
     conv_box_predictor = box_predictor_config.convolutional_box_predictor
-    conv_hyperparams = argscope_fn(conv_box_predictor.conv_hyperparams,
-                                   is_training)
+    conv_hyperparams_fn = argscope_fn(conv_box_predictor.conv_hyperparams,
+                                      is_training)
     box_predictor_object = box_predictor.ConvolutionalBoxPredictor(
         is_training=is_training,
         num_classes=num_classes,
-        conv_hyperparams=conv_hyperparams,
+        conv_hyperparams_fn=conv_hyperparams_fn,
         min_depth=conv_box_predictor.min_depth,
         max_depth=conv_box_predictor.max_depth,
         num_layers_before_predictor=(conv_box_predictor.
@@ -73,12 +73,12 @@ def build(argscope_fn, box_predictor_config, is_training, num_classes):
   if  box_predictor_oneof == 'weight_shared_convolutional_box_predictor':
     conv_box_predictor = (box_predictor_config.
                           weight_shared_convolutional_box_predictor)
-    conv_hyperparams = argscope_fn(conv_box_predictor.conv_hyperparams,
-                                   is_training)
+    conv_hyperparams_fn = argscope_fn(conv_box_predictor.conv_hyperparams,
+                                      is_training)
     box_predictor_object = box_predictor.WeightSharedConvolutionalBoxPredictor(
         is_training=is_training,
         num_classes=num_classes,
-        conv_hyperparams=conv_hyperparams,
+        conv_hyperparams_fn=conv_hyperparams_fn,
         depth=conv_box_predictor.depth,
         num_layers_before_predictor=(conv_box_predictor.
                                      num_layers_before_predictor),
@@ -90,20 +90,20 @@ def build(argscope_fn, box_predictor_config, is_training, num_classes):
 
   if box_predictor_oneof == 'mask_rcnn_box_predictor':
     mask_rcnn_box_predictor = box_predictor_config.mask_rcnn_box_predictor
-    fc_hyperparams = argscope_fn(mask_rcnn_box_predictor.fc_hyperparams,
-                                 is_training)
-    conv_hyperparams = None
+    fc_hyperparams_fn = argscope_fn(mask_rcnn_box_predictor.fc_hyperparams,
+                                    is_training)
+    conv_hyperparams_fn = None
     if mask_rcnn_box_predictor.HasField('conv_hyperparams'):
-      conv_hyperparams = argscope_fn(mask_rcnn_box_predictor.conv_hyperparams,
-                                     is_training)
+      conv_hyperparams_fn = argscope_fn(
+          mask_rcnn_box_predictor.conv_hyperparams, is_training)
     box_predictor_object = box_predictor.MaskRCNNBoxPredictor(
         is_training=is_training,
         num_classes=num_classes,
-        fc_hyperparams=fc_hyperparams,
+        fc_hyperparams_fn=fc_hyperparams_fn,
         use_dropout=mask_rcnn_box_predictor.use_dropout,
         dropout_keep_prob=mask_rcnn_box_predictor.dropout_keep_probability,
         box_code_size=mask_rcnn_box_predictor.box_code_size,
-        conv_hyperparams=conv_hyperparams,
+        conv_hyperparams_fn=conv_hyperparams_fn,
         predict_instance_masks=mask_rcnn_box_predictor.predict_instance_masks,
         mask_height=mask_rcnn_box_predictor.mask_height,
         mask_width=mask_rcnn_box_predictor.mask_width,
@@ -116,12 +116,12 @@ def build(argscope_fn, box_predictor_config, is_training, num_classes):
 
   if box_predictor_oneof == 'rfcn_box_predictor':
     rfcn_box_predictor = box_predictor_config.rfcn_box_predictor
-    conv_hyperparams = argscope_fn(rfcn_box_predictor.conv_hyperparams,
-                                   is_training)
+    conv_hyperparams_fn = argscope_fn(rfcn_box_predictor.conv_hyperparams,
+                                      is_training)
     box_predictor_object = box_predictor.RfcnBoxPredictor(
         is_training=is_training,
         num_classes=num_classes,
-        conv_hyperparams=conv_hyperparams,
+        conv_hyperparams_fn=conv_hyperparams_fn,
         crop_size=[rfcn_box_predictor.crop_height,
                    rfcn_box_predictor.crop_width],
         num_spatial_bins=[rfcn_box_predictor.num_spatial_bins_height,

research/object_detection/builders/box_predictor_builder_test.py

@@ -54,7 +54,7 @@ class ConvolutionalBoxPredictorBuilderTest(tf.test.TestCase):
         box_predictor_config=box_predictor_proto,
         is_training=False,
         num_classes=10)
-    (conv_hyperparams_actual, is_training) = box_predictor._conv_hyperparams
+    (conv_hyperparams_actual, is_training) = box_predictor._conv_hyperparams_fn
     self.assertAlmostEqual((hyperparams_proto.regularizer.
                             l1_regularizer.weight),
                            (conv_hyperparams_actual.regularizer.l1_regularizer.
@@ -183,7 +183,7 @@ class WeightSharedConvolutionalBoxPredictorBuilderTest(tf.test.TestCase):
         box_predictor_config=box_predictor_proto,
         is_training=False,
         num_classes=10)
-    (conv_hyperparams_actual, is_training) = box_predictor._conv_hyperparams
+    (conv_hyperparams_actual, is_training) = box_predictor._conv_hyperparams_fn
     self.assertAlmostEqual((hyperparams_proto.regularizer.
                             l1_regularizer.weight),
                            (conv_hyperparams_actual.regularizer.l1_regularizer.
@@ -297,7 +297,7 @@ class MaskRCNNBoxPredictorBuilderTest(tf.test.TestCase):
         is_training=False,
         num_classes=10)
     mock_argscope_fn.assert_called_with(hyperparams_proto, False)
-    self.assertEqual(box_predictor._fc_hyperparams, 'arg_scope')
+    self.assertEqual(box_predictor._fc_hyperparams_fn, 'arg_scope')
 
   def test_non_default_mask_rcnn_box_predictor(self):
     fc_hyperparams_text_proto = """
@@ -417,7 +417,7 @@ class RfcnBoxPredictorBuilderTest(tf.test.TestCase):
         box_predictor_config=box_predictor_proto,
         is_training=False,
         num_classes=10)
-    (conv_hyperparams_actual, is_training) = box_predictor._conv_hyperparams
+    (conv_hyperparams_actual, is_training) = box_predictor._conv_hyperparams_fn
     self.assertAlmostEqual((hyperparams_proto.regularizer.
                             l1_regularizer.weight),
                            (conv_hyperparams_actual.regularizer.l1_regularizer.

research/object_detection/builders/dataset_builder.py

@@ -72,7 +72,9 @@ def _get_padding_shapes(dataset, max_num_boxes=None, num_classes=None,
       fields.InputDataFields.num_groundtruth_boxes: [],
       fields.InputDataFields.groundtruth_label_types: [max_num_boxes],
       fields.InputDataFields.groundtruth_label_scores: [max_num_boxes],
-      fields.InputDataFields.true_image_shape: [3]
+      fields.InputDataFields.true_image_shape: [3],
+      fields.InputDataFields.multiclass_scores: [
+          max_num_boxes, num_classes + 1 if num_classes is not None else None],
   }
   # Determine whether groundtruth_classes are integers or one-hot encodings, and
   # apply batching appropriately.

research/object_detection/builders/hyperparams_builder.py

@@ -43,7 +43,8 @@ def build(hyperparams_config, is_training):
     is_training: Whether the network is in training mode.
 
   Returns:
-    arg_scope: tf-slim arg_scope containing hyperparameters for ops.
+    arg_scope_fn: A function to construct tf-slim arg_scope containing
+      hyperparameters for ops.
 
   Raises:
     ValueError: if hyperparams_config is not of type hyperparams.Hyperparams.
@@ -64,16 +65,18 @@ def build(hyperparams_config, is_training):
   if hyperparams_config.HasField('op') and (
       hyperparams_config.op == hyperparams_pb2.Hyperparams.FC):
     affected_ops = [slim.fully_connected]
-  with slim.arg_scope(
-      affected_ops,
-      weights_regularizer=_build_regularizer(
-          hyperparams_config.regularizer),
-      weights_initializer=_build_initializer(
-          hyperparams_config.initializer),
-      activation_fn=_build_activation_fn(hyperparams_config.activation),
-      normalizer_fn=batch_norm,
-      normalizer_params=batch_norm_params) as sc:
-    return sc
+  def scope_fn():
+    with slim.arg_scope(
+        affected_ops,
+        weights_regularizer=_build_regularizer(
+            hyperparams_config.regularizer),
+        weights_initializer=_build_initializer(
+            hyperparams_config.initializer),
+        activation_fn=_build_activation_fn(hyperparams_config.activation),
+        normalizer_fn=batch_norm,
+        normalizer_params=batch_norm_params) as sc:
+      return sc
+  return scope_fn
 
 
 def _build_activation_fn(activation_fn):

research/object_detection/builders/hyperparams_builder_test.py

@@ -45,7 +45,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     self.assertTrue(self._get_scope_key(slim.conv2d) in scope)
 
   def test_default_arg_scope_has_separable_conv2d_op(self):
@@ -61,7 +63,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     self.assertTrue(self._get_scope_key(slim.separable_conv2d) in scope)
 
   def test_default_arg_scope_has_conv2d_transpose_op(self):
@@ -77,7 +81,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     self.assertTrue(self._get_scope_key(slim.conv2d_transpose) in scope)
 
   def test_explicit_fc_op_arg_scope_has_fully_connected_op(self):
@@ -94,7 +100,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     self.assertTrue(self._get_scope_key(slim.fully_connected) in scope)
 
   def test_separable_conv2d_and_conv2d_and_transpose_have_same_parameters(self):
@@ -110,7 +118,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     kwargs_1, kwargs_2, kwargs_3 = scope.values()
     self.assertDictEqual(kwargs_1, kwargs_2)
     self.assertDictEqual(kwargs_1, kwargs_3)
@@ -129,7 +139,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     regularizer = conv_scope_arguments['weights_regularizer']
     weights = np.array([1., -1, 4., 2.])
@@ -151,7 +163,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
 
     regularizer = conv_scope_arguments['weights_regularizer']
@@ -180,7 +194,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     self.assertEqual(conv_scope_arguments['normalizer_fn'], slim.batch_norm)
     batch_norm_params = conv_scope_arguments['normalizer_params']
@@ -210,7 +226,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=False)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=False)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     self.assertEqual(conv_scope_arguments['normalizer_fn'], slim.batch_norm)
     batch_norm_params = conv_scope_arguments['normalizer_params']
@@ -240,7 +258,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     self.assertEqual(conv_scope_arguments['normalizer_fn'], slim.batch_norm)
     batch_norm_params = conv_scope_arguments['normalizer_params']
@@ -263,7 +283,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     self.assertEqual(conv_scope_arguments['normalizer_fn'], None)
     self.assertEqual(conv_scope_arguments['normalizer_params'], None)
@@ -282,7 +304,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     self.assertEqual(conv_scope_arguments['activation_fn'], None)
 
@@ -300,7 +324,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     self.assertEqual(conv_scope_arguments['activation_fn'], tf.nn.relu)
 
@@ -318,7 +344,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     self.assertEqual(conv_scope_arguments['activation_fn'], tf.nn.relu6)
 
@@ -351,7 +379,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     initializer = conv_scope_arguments['weights_initializer']
     self._assert_variance_in_range(initializer, shape=[100, 40],
@@ -373,7 +403,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     initializer = conv_scope_arguments['weights_initializer']
     self._assert_variance_in_range(initializer, shape=[100, 40],
@@ -395,7 +427,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     initializer = conv_scope_arguments['weights_initializer']
     self._assert_variance_in_range(initializer, shape=[100, 40],
@@ -417,7 +451,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     initializer = conv_scope_arguments['weights_initializer']
     self._assert_variance_in_range(initializer, shape=[100, 40],
@@ -438,7 +474,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     initializer = conv_scope_arguments['weights_initializer']
     self._assert_variance_in_range(initializer, shape=[100, 40],
@@ -459,7 +497,9 @@ class HyperparamsBuilderTest(tf.test.TestCase):
     """
     conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
-    scope = hyperparams_builder.build(conv_hyperparams_proto, is_training=True)
+    scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
+                                         is_training=True)
+    scope = scope_fn()
     conv_scope_arguments = scope.values()[0]
     initializer = conv_scope_arguments['weights_initializer']
     self._assert_variance_in_range(initializer, shape=[100, 40],

research/object_detection/builders/model_builder.py

@@ -98,19 +98,13 @@ def build(model_config, is_training, add_summaries=True):
 
 
 def _build_ssd_feature_extractor(feature_extractor_config, is_training,
-                                 reuse_weights=None,
-                                 inplace_batchnorm_update=False):
+                                 reuse_weights=None):
   """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.
@@ -122,7 +116,6 @@ def _build_ssd_feature_extractor(feature_extractor_config, is_training,
   depth_multiplier = feature_extractor_config.depth_multiplier
   min_depth = feature_extractor_config.min_depth
   pad_to_multiple = feature_extractor_config.pad_to_multiple
-  batch_norm_trainable = feature_extractor_config.batch_norm_trainable
   use_explicit_padding = feature_extractor_config.use_explicit_padding
   use_depthwise = feature_extractor_config.use_depthwise
   conv_hyperparams = hyperparams_builder.build(
@@ -132,11 +125,9 @@ def _build_ssd_feature_extractor(feature_extractor_config, is_training,
     raise ValueError('Unknown ssd feature_extractor: {}'.format(feature_type))
 
   feature_extractor_class = SSD_FEATURE_EXTRACTOR_CLASS_MAP[feature_type]
-  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,
-                                 inplace_batchnorm_update)
+  return feature_extractor_class(
+      is_training, depth_multiplier, min_depth, pad_to_multiple,
+      conv_hyperparams, reuse_weights, use_explicit_padding, use_depthwise)
 
 
 def _build_ssd_model(ssd_config, is_training, add_summaries):
@@ -160,8 +151,7 @@ def _build_ssd_model(ssd_config, is_training, add_summaries):
   # Feature extractor
   feature_extractor = _build_ssd_feature_extractor(
       feature_extractor_config=ssd_config.feature_extractor,
-      is_training=is_training,
-      inplace_batchnorm_update=ssd_config.inplace_batchnorm_update)
+      is_training=is_training)
 
   box_coder = box_coder_builder.build(ssd_config.box_coder)
   matcher = matcher_builder.build(ssd_config.matcher)
@@ -203,7 +193,9 @@ def _build_ssd_model(ssd_config, is_training, add_summaries):
       normalize_loss_by_num_matches,
       hard_example_miner,
       add_summaries=add_summaries,
-      normalize_loc_loss_by_codesize=normalize_loc_loss_by_codesize)
+      normalize_loc_loss_by_codesize=normalize_loc_loss_by_codesize,
+      freeze_batchnorm=ssd_config.freeze_batchnorm,
+      inplace_batchnorm_update=ssd_config.inplace_batchnorm_update)
 
 
 def _build_faster_rcnn_feature_extractor(
@@ -276,7 +268,7 @@ def _build_faster_rcnn_model(frcnn_config, is_training, add_summaries):
       frcnn_config.first_stage_anchor_generator)
 
   first_stage_atrous_rate = frcnn_config.first_stage_atrous_rate
-  first_stage_box_predictor_arg_scope = hyperparams_builder.build(
+  first_stage_box_predictor_arg_scope_fn = hyperparams_builder.build(
       frcnn_config.first_stage_box_predictor_conv_hyperparams, is_training)
   first_stage_box_predictor_kernel_size = (
       frcnn_config.first_stage_box_predictor_kernel_size)
@@ -329,8 +321,8 @@ def _build_faster_rcnn_model(frcnn_config, is_training, add_summaries):
       'number_of_stages': number_of_stages,
       'first_stage_anchor_generator': first_stage_anchor_generator,
       'first_stage_atrous_rate': first_stage_atrous_rate,
-      'first_stage_box_predictor_arg_scope':
-      first_stage_box_predictor_arg_scope,
+      'first_stage_box_predictor_arg_scope_fn':
+      first_stage_box_predictor_arg_scope_fn,
       'first_stage_box_predictor_kernel_size':
       first_stage_box_predictor_kernel_size,
       'first_stage_box_predictor_depth': first_stage_box_predictor_depth,

research/object_detection/builders/model_builder_test.py

@@ -225,7 +225,6 @@ class ModelBuilderTest(tf.test.TestCase):
                 }
               }
           }
-          batch_norm_trainable: true
         }
         box_coder {
           faster_rcnn_box_coder {
@@ -298,6 +297,7 @@ class ModelBuilderTest(tf.test.TestCase):
   def test_create_ssd_mobilenet_v1_model_from_config(self):
     model_text_proto = """
       ssd {
+        freeze_batchnorm: true
         inplace_batchnorm_update: true
         feature_extractor {
           type: 'ssd_mobilenet_v1'
@@ -311,7 +311,6 @@ class ModelBuilderTest(tf.test.TestCase):
                 }
               }
           }
-          batch_norm_trainable: true
         }
         box_coder {
           faster_rcnn_box_coder {
@@ -368,8 +367,9 @@ class ModelBuilderTest(tf.test.TestCase):
     self.assertIsInstance(model, ssd_meta_arch.SSDMetaArch)
     self.assertIsInstance(model._feature_extractor,
                           SSDMobileNetV1FeatureExtractor)
-    self.assertTrue(model._feature_extractor._batch_norm_trainable)
     self.assertTrue(model._normalize_loc_loss_by_codesize)
+    self.assertTrue(model._freeze_batchnorm)
+    self.assertTrue(model._inplace_batchnorm_update)
 
   def test_create_ssd_mobilenet_v2_model_from_config(self):
     model_text_proto = """
@@ -386,7 +386,6 @@ class ModelBuilderTest(tf.test.TestCase):
                 }
               }
           }
-          batch_norm_trainable: true
         }
         box_coder {
           faster_rcnn_box_coder {
@@ -443,7 +442,6 @@ class ModelBuilderTest(tf.test.TestCase):
     self.assertIsInstance(model, ssd_meta_arch.SSDMetaArch)
     self.assertIsInstance(model._feature_extractor,
                           SSDMobileNetV2FeatureExtractor)
-    self.assertTrue(model._feature_extractor._batch_norm_trainable)
     self.assertTrue(model._normalize_loc_loss_by_codesize)
 
   def test_create_embedded_ssd_mobilenet_v1_model_from_config(self):
@@ -461,7 +459,6 @@ class ModelBuilderTest(tf.test.TestCase):
                 }
               }
           }
-          batch_norm_trainable: true
         }
         box_coder {
           faster_rcnn_box_coder {

research/object_detection/core/box_predictor.py

@@ -147,7 +147,7 @@ class RfcnBoxPredictor(BoxPredictor):
   def __init__(self,
                is_training,
                num_classes,
-               conv_hyperparams,
+               conv_hyperparams_fn,
                num_spatial_bins,
                depth,
                crop_size,
@@ -160,8 +160,8 @@ class RfcnBoxPredictor(BoxPredictor):
         include the background category, so if groundtruth labels take values
         in {0, 1, .., K-1}, num_classes=K (and not K+1, even though the
         assigned classification targets can range from {0,... K}).
-      conv_hyperparams: Slim arg_scope with hyperparameters for conolutional
-        layers.
+      conv_hyperparams_fn: A function to construct tf-slim arg_scope with
+        hyperparameters for convolutional layers.
       num_spatial_bins: A list of two integers `[spatial_bins_y,
         spatial_bins_x]`.
       depth: Target depth to reduce the input feature maps to.
@@ -169,7 +169,7 @@ class RfcnBoxPredictor(BoxPredictor):
       box_code_size: Size of encoding for each box.
     """
     super(RfcnBoxPredictor, self).__init__(is_training, num_classes)
-    self._conv_hyperparams = conv_hyperparams
+    self._conv_hyperparams_fn = conv_hyperparams_fn
     self._num_spatial_bins = num_spatial_bins
     self._depth = depth
     self._crop_size = crop_size
@@ -227,7 +227,7 @@ class RfcnBoxPredictor(BoxPredictor):
       return tf.reshape(ones_mat * multiplier, [-1])
 
     net = image_feature
-    with slim.arg_scope(self._conv_hyperparams):
+    with slim.arg_scope(self._conv_hyperparams_fn()):
       net = slim.conv2d(net, self._depth, [1, 1], scope='reduce_depth')
       # Location predictions.
       location_feature_map_depth = (self._num_spatial_bins[0] *
@@ -297,11 +297,11 @@ class MaskRCNNBoxPredictor(BoxPredictor):
   def __init__(self,
                is_training,
                num_classes,
-               fc_hyperparams,
+               fc_hyperparams_fn,
                use_dropout,
                dropout_keep_prob,
                box_code_size,
-               conv_hyperparams=None,
+               conv_hyperparams_fn=None,
                predict_instance_masks=False,
                mask_height=14,
                mask_width=14,
@@ -316,16 +316,16 @@ class MaskRCNNBoxPredictor(BoxPredictor):
         include the background category, so if groundtruth labels take values
         in {0, 1, .., K-1}, num_classes=K (and not K+1, even though the
         assigned classification targets can range from {0,... K}).
-      fc_hyperparams: Slim arg_scope with hyperparameters for fully
-        connected ops.
+      fc_hyperparams_fn: A function to generate tf-slim arg_scope with
+        hyperparameters for fully connected ops.
       use_dropout: Option to use dropout or not.  Note that a single dropout
         op is applied here prior to both box and class predictions, which stands
         in contrast to the ConvolutionalBoxPredictor below.
       dropout_keep_prob: Keep probability for dropout.
         This is only used if use_dropout is True.
       box_code_size: Size of encoding for each box.
-      conv_hyperparams: Slim arg_scope with hyperparameters for convolution
-        ops.
+      conv_hyperparams_fn: A function to generate tf-slim arg_scope with
+        hyperparameters for convolution ops.
       predict_instance_masks: Whether to predict object masks inside detection
         boxes.
       mask_height: Desired output mask height. The default value is 14.
@@ -347,11 +347,11 @@ class MaskRCNNBoxPredictor(BoxPredictor):
       ValueError: If mask_prediction_num_conv_layers is smaller than two.
     """
     super(MaskRCNNBoxPredictor, self).__init__(is_training, num_classes)
-    self._fc_hyperparams = fc_hyperparams
+    self._fc_hyperparams_fn = fc_hyperparams_fn
     self._use_dropout = use_dropout
     self._box_code_size = box_code_size
     self._dropout_keep_prob = dropout_keep_prob
-    self._conv_hyperparams = conv_hyperparams
+    self._conv_hyperparams_fn = conv_hyperparams_fn
     self._predict_instance_masks = predict_instance_masks
     self._mask_height = mask_height
     self._mask_width = mask_width
@@ -361,7 +361,7 @@ class MaskRCNNBoxPredictor(BoxPredictor):
     if self._predict_keypoints:
       raise ValueError('Keypoint prediction is unimplemented.')
     if ((self._predict_instance_masks or self._predict_keypoints) and
-        self._conv_hyperparams is None):
+        self._conv_hyperparams_fn is None):
       raise ValueError('`conv_hyperparams` must be provided when predicting '
                        'masks.')
     if self._mask_prediction_num_conv_layers < 2:
@@ -399,7 +399,7 @@ class MaskRCNNBoxPredictor(BoxPredictor):
       flattened_image_features = slim.dropout(flattened_image_features,
                                               keep_prob=self._dropout_keep_prob,
                                               is_training=self._is_training)
-    with slim.arg_scope(self._fc_hyperparams):
+    with slim.arg_scope(self._fc_hyperparams_fn()):
       box_encodings = slim.fully_connected(
           flattened_image_features,
           self._num_classes * self._box_code_size,
@@ -463,7 +463,7 @@ class MaskRCNNBoxPredictor(BoxPredictor):
       num_feature_channels = image_features.get_shape().as_list()[3]
       num_conv_channels = self._get_mask_predictor_conv_depth(
           num_feature_channels, self.num_classes)
-    with slim.arg_scope(self._conv_hyperparams):
+    with slim.arg_scope(self._conv_hyperparams_fn()):
       upsampled_features = tf.image.resize_bilinear(
           image_features,
           [self._mask_height, self._mask_width],
@@ -578,7 +578,7 @@ class ConvolutionalBoxPredictor(BoxPredictor):
   def __init__(self,
                is_training,
                num_classes,
-               conv_hyperparams,
+               conv_hyperparams_fn,
                min_depth,
                max_depth,
                num_layers_before_predictor,
@@ -597,8 +597,9 @@ class ConvolutionalBoxPredictor(BoxPredictor):
         include the background category, so if groundtruth labels take values
         in {0, 1, .., K-1}, num_classes=K (and not K+1, even though the
         assigned classification targets can range from {0,... K}).
-      conv_hyperparams: Slim arg_scope with hyperparameters for convolution ops.
-      min_depth: Minumum feature depth prior to predicting box encodings
+      conv_hyperparams_fn: A function to generate tf-slim arg_scope with
+        hyperparameters for convolution ops.
+      min_depth: Minimum feature depth prior to predicting box encodings
         and class predictions.
       max_depth: Maximum feature depth prior to predicting box encodings
         and class predictions. If max_depth is set to 0, no additional
@@ -626,7 +627,7 @@ class ConvolutionalBoxPredictor(BoxPredictor):
     super(ConvolutionalBoxPredictor, self).__init__(is_training, num_classes)
     if min_depth > max_depth:
       raise ValueError('min_depth should be less than or equal to max_depth')
-    self._conv_hyperparams = conv_hyperparams
+    self._conv_hyperparams_fn = conv_hyperparams_fn
     self._min_depth = min_depth
     self._max_depth = max_depth
     self._num_layers_before_predictor = num_layers_before_predictor
@@ -679,7 +680,7 @@ class ConvolutionalBoxPredictor(BoxPredictor):
         # Add a slot for the background class.
         num_class_slots = self.num_classes + 1
         net = image_feature
-        with slim.arg_scope(self._conv_hyperparams), \
+        with slim.arg_scope(self._conv_hyperparams_fn()), \
              slim.arg_scope([slim.dropout], is_training=self._is_training):
           # Add additional conv layers before the class predictor.
           features_depth = static_shape.get_depth(image_feature.get_shape())
@@ -767,7 +768,7 @@ class WeightSharedConvolutionalBoxPredictor(BoxPredictor):
   def __init__(self,
                is_training,
                num_classes,
-               conv_hyperparams,
+               conv_hyperparams_fn,
                depth,
                num_layers_before_predictor,
                box_code_size,
@@ -781,7 +782,8 @@ class WeightSharedConvolutionalBoxPredictor(BoxPredictor):
         include the background category, so if groundtruth labels take values
         in {0, 1, .., K-1}, num_classes=K (and not K+1, even though the
         assigned classification targets can range from {0,... K}).
-      conv_hyperparams: Slim arg_scope with hyperparameters for convolution ops.
+      conv_hyperparams_fn: A function to generate tf-slim arg_scope with
+        hyperparameters for convolution ops.
       depth: depth of conv layers.
       num_layers_before_predictor: Number of the additional conv layers before
         the predictor.
@@ -792,7 +794,7 @@ class WeightSharedConvolutionalBoxPredictor(BoxPredictor):
     """
     super(WeightSharedConvolutionalBoxPredictor, self).__init__(is_training,
                                                                 num_classes)
-    self._conv_hyperparams = conv_hyperparams
+    self._conv_hyperparams_fn = conv_hyperparams_fn
     self._depth = depth
     self._num_layers_before_predictor = num_layers_before_predictor
     self._box_code_size = box_code_size
@@ -846,7 +848,7 @@ class WeightSharedConvolutionalBoxPredictor(BoxPredictor):
         num_class_slots = self.num_classes + 1
         box_encodings_net = image_feature
         class_predictions_net = image_feature
-        with slim.arg_scope(self._conv_hyperparams):
+        with slim.arg_scope(self._conv_hyperparams_fn()):
           for i in range(self._num_layers_before_predictor):
             box_encodings_net = slim.conv2d(
                 box_encodings_net,

research/object_detection/core/box_predictor_test.py

@@ -49,7 +49,7 @@ class MaskRCNNBoxPredictorTest(tf.test.TestCase):
     mask_box_predictor = box_predictor.MaskRCNNBoxPredictor(
         is_training=False,
         num_classes=5,
-        fc_hyperparams=self._build_arg_scope_with_hyperparams(),
+        fc_hyperparams_fn=self._build_arg_scope_with_hyperparams(),
         use_dropout=False,
         dropout_keep_prob=0.5,
         box_code_size=4,
@@ -75,7 +75,7 @@ class MaskRCNNBoxPredictorTest(tf.test.TestCase):
       box_predictor.MaskRCNNBoxPredictor(
           is_training=False,
           num_classes=5,
-          fc_hyperparams=self._build_arg_scope_with_hyperparams(),
+          fc_hyperparams_fn=self._build_arg_scope_with_hyperparams(),
           use_dropout=False,
           dropout_keep_prob=0.5,
           box_code_size=4,
@@ -86,11 +86,11 @@ class MaskRCNNBoxPredictorTest(tf.test.TestCase):
     mask_box_predictor = box_predictor.MaskRCNNBoxPredictor(
         is_training=False,
         num_classes=5,
-        fc_hyperparams=self._build_arg_scope_with_hyperparams(),
+        fc_hyperparams_fn=self._build_arg_scope_with_hyperparams(),
         use_dropout=False,
         dropout_keep_prob=0.5,
         box_code_size=4,
-        conv_hyperparams=self._build_arg_scope_with_hyperparams(
+        conv_hyperparams_fn=self._build_arg_scope_with_hyperparams(
             op_type=hyperparams_pb2.Hyperparams.CONV),
         predict_instance_masks=True)
     box_predictions = mask_box_predictor.predict(
@@ -108,7 +108,7 @@ class MaskRCNNBoxPredictorTest(tf.test.TestCase):
     mask_box_predictor = box_predictor.MaskRCNNBoxPredictor(
         is_training=False,
         num_classes=5,
-        fc_hyperparams=self._build_arg_scope_with_hyperparams(),
+        fc_hyperparams_fn=self._build_arg_scope_with_hyperparams(),
         use_dropout=False,
         dropout_keep_prob=0.5,
         box_code_size=4)
@@ -125,7 +125,7 @@ class MaskRCNNBoxPredictorTest(tf.test.TestCase):
       box_predictor.MaskRCNNBoxPredictor(
           is_training=False,
           num_classes=5,
-          fc_hyperparams=self._build_arg_scope_with_hyperparams(),
+          fc_hyperparams_fn=self._build_arg_scope_with_hyperparams(),
           use_dropout=False,
           dropout_keep_prob=0.5,
           box_code_size=4,
@@ -155,7 +155,7 @@ class RfcnBoxPredictorTest(tf.test.TestCase):
     rfcn_box_predictor = box_predictor.RfcnBoxPredictor(
         is_training=False,
         num_classes=2,
-        conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+        conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
         num_spatial_bins=[3, 3],
         depth=4,
         crop_size=[12, 12],
@@ -205,7 +205,7 @@ class ConvolutionalBoxPredictorTest(test_case.TestCase):
       conv_box_predictor = box_predictor.ConvolutionalBoxPredictor(
           is_training=False,
           num_classes=0,
-          conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+          conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
           min_depth=0,
           max_depth=32,
           num_layers_before_predictor=1,
@@ -234,7 +234,7 @@ class ConvolutionalBoxPredictorTest(test_case.TestCase):
       conv_box_predictor = box_predictor.ConvolutionalBoxPredictor(
           is_training=False,
           num_classes=0,
-          conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+          conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
           min_depth=0,
           max_depth=32,
           num_layers_before_predictor=1,
@@ -265,7 +265,7 @@ class ConvolutionalBoxPredictorTest(test_case.TestCase):
       conv_box_predictor = box_predictor.ConvolutionalBoxPredictor(
           is_training=False,
           num_classes=num_classes_without_background,
-          conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+          conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
           min_depth=0,
           max_depth=32,
           num_layers_before_predictor=1,
@@ -297,7 +297,7 @@ class ConvolutionalBoxPredictorTest(test_case.TestCase):
     conv_box_predictor = box_predictor.ConvolutionalBoxPredictor(
         is_training=False,
         num_classes=0,
-        conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+        conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
         min_depth=0,
         max_depth=32,
         num_layers_before_predictor=1,
@@ -344,7 +344,7 @@ class ConvolutionalBoxPredictorTest(test_case.TestCase):
     conv_box_predictor = box_predictor.ConvolutionalBoxPredictor(
         is_training=False,
         num_classes=0,
-        conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+        conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
         min_depth=0,
         max_depth=32,
         num_layers_before_predictor=1,
@@ -416,7 +416,7 @@ class WeightSharedConvolutionalBoxPredictorTest(test_case.TestCase):
       conv_box_predictor = box_predictor.WeightSharedConvolutionalBoxPredictor(
           is_training=False,
           num_classes=0,
-          conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+          conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
           depth=32,
           num_layers_before_predictor=1,
           box_code_size=4)
@@ -442,7 +442,7 @@ class WeightSharedConvolutionalBoxPredictorTest(test_case.TestCase):
       conv_box_predictor = box_predictor.WeightSharedConvolutionalBoxPredictor(
           is_training=False,
           num_classes=num_classes_without_background,
-          conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+          conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
           depth=32,
           num_layers_before_predictor=1,
           box_code_size=4)
@@ -471,7 +471,7 @@ class WeightSharedConvolutionalBoxPredictorTest(test_case.TestCase):
       conv_box_predictor = box_predictor.WeightSharedConvolutionalBoxPredictor(
           is_training=False,
           num_classes=num_classes_without_background,
-          conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+          conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
           depth=32,
           num_layers_before_predictor=1,
           box_code_size=4)
@@ -500,7 +500,7 @@ class WeightSharedConvolutionalBoxPredictorTest(test_case.TestCase):
       conv_box_predictor = box_predictor.WeightSharedConvolutionalBoxPredictor(
           is_training=False,
           num_classes=num_classes_without_background,
-          conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+          conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
           depth=32,
           num_layers_before_predictor=2,
           box_code_size=4)
@@ -553,7 +553,7 @@ class WeightSharedConvolutionalBoxPredictorTest(test_case.TestCase):
     conv_box_predictor = box_predictor.WeightSharedConvolutionalBoxPredictor(
         is_training=False,
         num_classes=0,
-        conv_hyperparams=self._build_arg_scope_with_conv_hyperparams(),
+        conv_hyperparams_fn=self._build_arg_scope_with_conv_hyperparams(),
         depth=32,
         num_layers_before_predictor=1,
         box_code_size=4)

research/object_detection/core/preprocessor_test.py

@@ -119,6 +119,9 @@ class PreprocessorTest(tf.test.TestCase):
         [[-0.1, 0.25, 0.75, 1], [0.25, 0.5, 0.75, 1.1]], dtype=tf.float32)
     return boxes
 
+  def createTestMultiClassScores(self):
+    return tf.constant([[1.0, 0.0], [0.5, 0.5]], dtype=tf.float32)
+
   def expectedImagesAfterNormalization(self):
     images_r = tf.constant([[[0, 0, 0, 0], [-1, -1, 0, 0],
                              [-1, 0, 0, 0], [0.5, 0.5, 0, 0]]],
@@ -269,6 +272,9 @@ class PreprocessorTest(tf.test.TestCase):
   def expectedLabelsAfterThresholding(self):
     return tf.constant([1], dtype=tf.float32)
 
+  def expectedMultiClassScoresAfterThresholding(self):
+    return tf.constant([[1.0, 0.0]], dtype=tf.float32)
+
   def expectedMasksAfterThresholding(self):
     mask = np.array([
         [[255.0, 0.0, 0.0],
@@ -345,6 +351,28 @@ class PreprocessorTest(tf.test.TestCase):
       self.assertAllClose(
           retained_label_scores_, expected_retained_label_scores_)
 
+  def testRetainBoxesAboveThresholdWithMultiClassScores(self):
+    boxes = self.createTestBoxes()
+    labels = self.createTestLabels()
+    label_scores = self.createTestLabelScores()
+    multiclass_scores = self.createTestMultiClassScores()
+    (_, _, _,
+     retained_multiclass_scores) = preprocessor.retain_boxes_above_threshold(
+         boxes,
+         labels,
+         label_scores,
+         multiclass_scores=multiclass_scores,
+         threshold=0.6)
+    with self.test_session() as sess:
+      (retained_multiclass_scores_,
+       expected_retained_multiclass_scores_) = sess.run([
+           retained_multiclass_scores,
+           self.expectedMultiClassScoresAfterThresholding()
+       ])
+
+      self.assertAllClose(retained_multiclass_scores_,
+                          expected_retained_multiclass_scores_)
+
   def testRetainBoxesAboveThresholdWithMasks(self):
     boxes = self.createTestBoxes()
     labels = self.createTestLabels()
@@ -1264,6 +1292,48 @@ class PreprocessorTest(tf.test.TestCase):
         self.assertAllClose(distorted_boxes_, expected_boxes_)
         self.assertAllEqual(distorted_labels_, expected_labels_)
 
+  def testRandomCropImageWithMultiClassScores(self):
+    preprocessing_options = []
+    preprocessing_options.append((preprocessor.normalize_image, {
+        'original_minval': 0,
+        'original_maxval': 255,
+        'target_minval': 0,
+        'target_maxval': 1
+    }))
+    preprocessing_options.append((preprocessor.random_crop_image, {}))
+    images = self.createTestImages()
+    boxes = self.createTestBoxes()
+    labels = self.createTestLabels()
+    multiclass_scores = self.createTestMultiClassScores()
+
+    tensor_dict = {
+        fields.InputDataFields.image: images,
+        fields.InputDataFields.groundtruth_boxes: boxes,
+        fields.InputDataFields.groundtruth_classes: labels,
+        fields.InputDataFields.multiclass_scores: multiclass_scores
+    }
+    distorted_tensor_dict = preprocessor.preprocess(tensor_dict,
+                                                    preprocessing_options)
+    distorted_images = distorted_tensor_dict[fields.InputDataFields.image]
+    distorted_boxes = distorted_tensor_dict[
+        fields.InputDataFields.groundtruth_boxes]
+    distorted_multiclass_scores = distorted_tensor_dict[
+        fields.InputDataFields.multiclass_scores]
+    boxes_rank = tf.rank(boxes)
+    distorted_boxes_rank = tf.rank(distorted_boxes)
+    images_rank = tf.rank(images)
+    distorted_images_rank = tf.rank(distorted_images)
+
+    with self.test_session() as sess:
+      (boxes_rank_, distorted_boxes_rank_, images_rank_, distorted_images_rank_,
+       multiclass_scores_, distorted_multiclass_scores_) = sess.run([
+           boxes_rank, distorted_boxes_rank, images_rank, distorted_images_rank,
+           multiclass_scores, distorted_multiclass_scores
+       ])
+      self.assertAllEqual(boxes_rank_, distorted_boxes_rank_)
+      self.assertAllEqual(images_rank_, distorted_images_rank_)
+      self.assertAllEqual(multiclass_scores_, distorted_multiclass_scores_)
+
   def testStrictRandomCropImageWithLabelScores(self):
     image = self.createColorfulTestImage()[0]
     boxes = self.createTestBoxes()
@@ -2510,6 +2580,49 @@ class PreprocessorTest(tf.test.TestCase):
       self.assertAllEqual(boxes_rank_, distorted_boxes_rank_)
       self.assertAllEqual(images_rank_, distorted_images_rank_)
 
+  def testSSDRandomCropWithMultiClassScores(self):
+    preprocessing_options = [(preprocessor.normalize_image, {
+        'original_minval': 0,
+        'original_maxval': 255,
+        'target_minval': 0,
+        'target_maxval': 1
+    }), (preprocessor.ssd_random_crop, {})]
+    images = self.createTestImages()
+    boxes = self.createTestBoxes()
+    labels = self.createTestLabels()
+    multiclass_scores = self.createTestMultiClassScores()
+
+    tensor_dict = {
+        fields.InputDataFields.image: images,
+        fields.InputDataFields.groundtruth_boxes: boxes,
+        fields.InputDataFields.groundtruth_classes: labels,
+        fields.InputDataFields.multiclass_scores: multiclass_scores,
+    }
+    preprocessor_arg_map = preprocessor.get_default_func_arg_map(
+        include_multiclass_scores=True)
+    distorted_tensor_dict = preprocessor.preprocess(
+        tensor_dict, preprocessing_options, func_arg_map=preprocessor_arg_map)
+    distorted_images = distorted_tensor_dict[fields.InputDataFields.image]
+    distorted_boxes = distorted_tensor_dict[
+        fields.InputDataFields.groundtruth_boxes]
+    distorted_multiclass_scores = distorted_tensor_dict[
+        fields.InputDataFields.multiclass_scores]
+
+    images_rank = tf.rank(images)
+    distorted_images_rank = tf.rank(distorted_images)
+    boxes_rank = tf.rank(boxes)
+    distorted_boxes_rank = tf.rank(distorted_boxes)
+
+    with self.test_session() as sess:
+      (boxes_rank_, distorted_boxes_rank_, images_rank_, distorted_images_rank_,
+       multiclass_scores_, distorted_multiclass_scores_) = sess.run([
+           boxes_rank, distorted_boxes_rank, images_rank, distorted_images_rank,
+           multiclass_scores, distorted_multiclass_scores
+       ])
+      self.assertAllEqual(boxes_rank_, distorted_boxes_rank_)
+      self.assertAllEqual(images_rank_, distorted_images_rank_)
+      self.assertAllEqual(multiclass_scores_, distorted_multiclass_scores_)
+
   def testSSDRandomCropPad(self):
     images = self.createTestImages()
     boxes = self.createTestBoxes()
@@ -2562,28 +2675,31 @@ class PreprocessorTest(tf.test.TestCase):
 
   def _testSSDRandomCropFixedAspectRatio(self,
                                          include_label_scores,
+                                         include_multiclass_scores,
                                          include_instance_masks,
                                          include_keypoints):
     images = self.createTestImages()
     boxes = self.createTestBoxes()
     labels = self.createTestLabels()
-    preprocessing_options = [
-        (preprocessor.normalize_image, {
-            'original_minval': 0,
-            'original_maxval': 255,
-            'target_minval': 0,
-            'target_maxval': 1
-        }),
-        (preprocessor.ssd_random_crop_fixed_aspect_ratio, {})]
+    preprocessing_options = [(preprocessor.normalize_image, {
+        'original_minval': 0,
+        'original_maxval': 255,
+        'target_minval': 0,
+        'target_maxval': 1
+    }), (preprocessor.ssd_random_crop_fixed_aspect_ratio, {})]
     tensor_dict = {
         fields.InputDataFields.image: images,
         fields.InputDataFields.groundtruth_boxes: boxes,
-        fields.InputDataFields.groundtruth_classes: labels
+        fields.InputDataFields.groundtruth_classes: labels,
     }
     if include_label_scores:
       label_scores = self.createTestLabelScores()
       tensor_dict[fields.InputDataFields.groundtruth_label_scores] = (
           label_scores)
+    if include_multiclass_scores:
+      multiclass_scores = self.createTestMultiClassScores()
+      tensor_dict[fields.InputDataFields.multiclass_scores] = (
+          multiclass_scores)
     if include_instance_masks:
       masks = self.createTestMasks()
       tensor_dict[fields.InputDataFields.groundtruth_instance_masks] = masks
@@ -2593,6 +2709,7 @@ class PreprocessorTest(tf.test.TestCase):
 
     preprocessor_arg_map = preprocessor.get_default_func_arg_map(
         include_label_scores=include_label_scores,
+        include_multiclass_scores=include_multiclass_scores,
         include_instance_masks=include_instance_masks,
         include_keypoints=include_keypoints)
     distorted_tensor_dict = preprocessor.preprocess(
@@ -2615,16 +2732,25 @@ class PreprocessorTest(tf.test.TestCase):
 
   def testSSDRandomCropFixedAspectRatio(self):
     self._testSSDRandomCropFixedAspectRatio(include_label_scores=False,
+                                            include_multiclass_scores=False,
+                                            include_instance_masks=False,
+                                            include_keypoints=False)
+
+  def testSSDRandomCropFixedAspectRatioWithMultiClassScores(self):
+    self._testSSDRandomCropFixedAspectRatio(include_label_scores=False,
+                                            include_multiclass_scores=True,
                                             include_instance_masks=False,
                                             include_keypoints=False)
 
   def testSSDRandomCropFixedAspectRatioWithMasksAndKeypoints(self):
     self._testSSDRandomCropFixedAspectRatio(include_label_scores=False,
+                                            include_multiclass_scores=False,
                                             include_instance_masks=True,
                                             include_keypoints=True)
 
   def testSSDRandomCropFixedAspectRatioWithLabelScoresMasksAndKeypoints(self):
     self._testSSDRandomCropFixedAspectRatio(include_label_scores=True,
+                                            include_multiclass_scores=False,
                                             include_instance_masks=True,
                                             include_keypoints=True)
 

research/object_detection/core/standard_fields.py

@@ -61,6 +61,9 @@ class InputDataFields(object):
     num_groundtruth_boxes: number of groundtruth boxes.
     true_image_shapes: true shapes of images in the resized images, as resized
       images can be padded with zeros.
+    verified_labels: list of human-verified image-level labels (note, that a
+      label can be verified both as positive and negative).
+    multiclass_scores: the label score per class for each box.
   """
   image = 'image'
   original_image = 'original_image'
@@ -86,6 +89,8 @@ class InputDataFields(object):
   groundtruth_weights = 'groundtruth_weights'
   num_groundtruth_boxes = 'num_groundtruth_boxes'
   true_image_shape = 'true_image_shape'
+  verified_labels = 'verified_labels'
+  multiclass_scores = 'multiclass_scores'
 
 
 class DetectionResultFields(object):

research/object_detection/leakr.dic

+confidential;1;confidentialit,confidentiality
+dogfood;1;
+fishfood;1;
+catfood;1;
+teamfood;1;
+droidfood;1;
+//go/;1;
+//sites/;1;
+a/google.com;1;
+corp.google.com;1;
+.googleplex.com;1;
+sandbox.;1;wallet-web.sandbox.,sandbox.google.com/checkout, sandbox.,paymentssandbox
+stupid;1;astupidi
+caution:;2;
+fixme:;2;
+fixme(;2;
+internal only;2;
+internal_only;2;
+backdoor;2;
+STOPSHIP;2;
+ridiculous;1;
+notasecret;1;
+@google.com;1;noreply@google.com
+$RE:chmod [0-9]?777;3;chmod (0)777
+mactruck;2;
+seastar;2;

research/object_detection/meta_architectures/faster_rcnn_meta_arch.py

@@ -229,7 +229,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
                number_of_stages,
                first_stage_anchor_generator,
                first_stage_atrous_rate,
-               first_stage_box_predictor_arg_scope,
+               first_stage_box_predictor_arg_scope_fn,
                first_stage_box_predictor_kernel_size,
                first_stage_box_predictor_depth,
                first_stage_minibatch_size,
@@ -291,8 +291,9 @@ class FasterRCNNMetaArch(model.DetectionModel):
         denser resolutions.  The atrous rate is used to compensate for the
         denser feature maps by using an effectively larger receptive field.
         (This should typically be set to 1).
-      first_stage_box_predictor_arg_scope: Slim arg_scope for conv2d,
-        separable_conv2d and fully_connected ops for the RPN box predictor.
+      first_stage_box_predictor_arg_scope_fn: A function to construct tf-slim
+        arg_scope for conv2d, separable_conv2d and fully_connected ops for the
+        RPN box predictor.
       first_stage_box_predictor_kernel_size: Kernel size to use for the
         convolution op just prior to RPN box predictions.
       first_stage_box_predictor_depth: Output depth for the convolution op
@@ -396,8 +397,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
     # (First stage) Region proposal network parameters
     self._first_stage_anchor_generator = first_stage_anchor_generator
     self._first_stage_atrous_rate = first_stage_atrous_rate
-    self._first_stage_box_predictor_arg_scope = (
-        first_stage_box_predictor_arg_scope)
+    self._first_stage_box_predictor_arg_scope_fn = (
+        first_stage_box_predictor_arg_scope_fn)
     self._first_stage_box_predictor_kernel_size = (
         first_stage_box_predictor_kernel_size)
     self._first_stage_box_predictor_depth = first_stage_box_predictor_depth
@@ -406,7 +407,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
         positive_fraction=first_stage_positive_balance_fraction)
     self._first_stage_box_predictor = box_predictor.ConvolutionalBoxPredictor(
         self._is_training, num_classes=1,
-        conv_hyperparams=self._first_stage_box_predictor_arg_scope,
+        conv_hyperparams_fn=self._first_stage_box_predictor_arg_scope_fn,
         min_depth=0, max_depth=0, num_layers_before_predictor=0,
         use_dropout=False, dropout_keep_prob=1.0, kernel_size=1,
         box_code_size=self._box_coder.code_size)
@@ -914,7 +915,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
     anchors = box_list_ops.concatenate(
         self._first_stage_anchor_generator.generate([(feature_map_shape[1],
                                                       feature_map_shape[2])]))
-    with slim.arg_scope(self._first_stage_box_predictor_arg_scope):
+    with slim.arg_scope(self._first_stage_box_predictor_arg_scope_fn()):
       kernel_size = self._first_stage_box_predictor_kernel_size
       rpn_box_predictor_features = slim.conv2d(
           rpn_features_to_crop,

research/object_detection/meta_architectures/faster_rcnn_meta_arch_test_lib.py

@@ -196,7 +196,7 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
         }
       }
     """
-    first_stage_box_predictor_arg_scope = (
+    first_stage_box_predictor_arg_scope_fn = (
         self._build_arg_scope_with_hyperparams(
             first_stage_box_predictor_hyperparams_text_proto, is_training))
 
@@ -255,8 +255,8 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
         'number_of_stages': number_of_stages,
         'first_stage_anchor_generator': first_stage_anchor_generator,
         'first_stage_atrous_rate': first_stage_atrous_rate,
-        'first_stage_box_predictor_arg_scope':
-        first_stage_box_predictor_arg_scope,
+        'first_stage_box_predictor_arg_scope_fn':
+        first_stage_box_predictor_arg_scope_fn,
         'first_stage_box_predictor_kernel_size':
         first_stage_box_predictor_kernel_size,
         'first_stage_box_predictor_depth': first_stage_box_predictor_depth,

research/object_detection/meta_architectures/rfcn_meta_arch.py

@@ -56,7 +56,7 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
                number_of_stages,
                first_stage_anchor_generator,
                first_stage_atrous_rate,
-               first_stage_box_predictor_arg_scope,
+               first_stage_box_predictor_arg_scope_fn,
                first_stage_box_predictor_kernel_size,
                first_stage_box_predictor_depth,
                first_stage_minibatch_size,
@@ -103,8 +103,9 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
         denser resolutions.  The atrous rate is used to compensate for the
         denser feature maps by using an effectively larger receptive field.
         (This should typically be set to 1).
-      first_stage_box_predictor_arg_scope: Slim arg_scope for conv2d,
-        separable_conv2d and fully_connected ops for the RPN box predictor.
+      first_stage_box_predictor_arg_scope_fn: A function to generate tf-slim
+        arg_scope for conv2d, separable_conv2d and fully_connected ops for the
+        RPN box predictor.
       first_stage_box_predictor_kernel_size: Kernel size to use for the
         convolution op just prior to RPN box predictions.
       first_stage_box_predictor_depth: Output depth for the convolution op
@@ -174,7 +175,7 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
         number_of_stages,
         first_stage_anchor_generator,
         first_stage_atrous_rate,
-        first_stage_box_predictor_arg_scope,
+        first_stage_box_predictor_arg_scope_fn,
         first_stage_box_predictor_kernel_size,
         first_stage_box_predictor_depth,
         first_stage_minibatch_size,

research/object_detection/meta_architectures/ssd_meta_arch.py

@@ -42,12 +42,10 @@ class SSDFeatureExtractor(object):
                depth_multiplier,
                min_depth,
                pad_to_multiple,
-               conv_hyperparams,
-               batch_norm_trainable=True,
+               conv_hyperparams_fn,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False,
-               inplace_batchnorm_update=False):
+               use_depthwise=False):
     """Constructor.
 
     Args:
@@ -56,27 +54,19 @@ class SSDFeatureExtractor(object):
       min_depth: minimum feature extractor depth.
       pad_to_multiple: the nearest multiple to zero pad the input height and
         width dimensions to.
-      conv_hyperparams: tf slim arg_scope for conv2d and separable_conv2d ops.
-      batch_norm_trainable: Whether to update batch norm parameters during
-        training or not. When training with a small batch size
-        (e.g. 1), it is desirable to disable batch norm update and use
-        pretrained batch norm params.
+      conv_hyperparams_fn: A function to construct tf slim arg_scope for conv2d
+        and separable_conv2d ops.
       reuse_weights: whether to reuse variables. Default is None.
       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
     self._min_depth = min_depth
     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._conv_hyperparams_fn = conv_hyperparams_fn
     self._reuse_weights = reuse_weights
     self._use_explicit_padding = use_explicit_padding
     self._use_depthwise = use_depthwise
@@ -106,28 +96,6 @@ class SSDFeatureExtractor(object):
     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]
-    """
-    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.
@@ -162,7 +130,9 @@ class SSDMetaArch(model.DetectionModel):
                normalize_loss_by_num_matches,
                hard_example_miner,
                add_summaries=True,
-               normalize_loc_loss_by_codesize=False):
+               normalize_loc_loss_by_codesize=False,
+               freeze_batchnorm=False,
+               inplace_batchnorm_update=False):
     """SSDMetaArch Constructor.
 
     TODO(rathodv,jonathanhuang): group NMS parameters + score converter into
@@ -209,9 +179,19 @@ class SSDMetaArch(model.DetectionModel):
         should be added to tensorflow graph.
       normalize_loc_loss_by_codesize: whether to normalize localization loss
         by code size of the box encoder.
+      freeze_batchnorm: Whether to freeze batch norm parameters during
+        training or not. When training with a small batch size (e.g. 1), it is
+        desirable to freeze batch norm update and use pretrained batch norm
+        params.
+      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.
     """
     super(SSDMetaArch, self).__init__(num_classes=box_predictor.num_classes)
     self._is_training = is_training
+    self._freeze_batchnorm = freeze_batchnorm
+    self._inplace_batchnorm_update = inplace_batchnorm_update
 
     # Needed for fine-tuning from classification checkpoints whose
     # variables do not have the feature extractor scope.
@@ -372,32 +352,40 @@ class SSDMetaArch(model.DetectionModel):
         5) anchors: 2-D float tensor of shape [num_anchors, 4] containing
           the generated anchors in normalized coordinates.
     """
-    with tf.variable_scope(None, self._extract_features_scope,
-                           [preprocessed_inputs]):
-      feature_maps = self._feature_extractor.extract_features(
+    batchnorm_updates_collections = (None if self._inplace_batchnorm_update
+                                     else tf.GraphKeys.UPDATE_OPS)
+    with slim.arg_scope([slim.batch_norm],
+                        is_training=(self._is_training and
+                                     not self._freeze_batchnorm),
+                        updates_collections=batchnorm_updates_collections):
+      with tf.variable_scope(None, self._extract_features_scope,
+                             [preprocessed_inputs]):
+        feature_maps = self._feature_extractor.extract_features(
+            preprocessed_inputs)
+      feature_map_spatial_dims = self._get_feature_map_spatial_dims(
+          feature_maps)
+      image_shape = shape_utils.combined_static_and_dynamic_shape(
           preprocessed_inputs)
-    feature_map_spatial_dims = self._get_feature_map_spatial_dims(feature_maps)
-    image_shape = shape_utils.combined_static_and_dynamic_shape(
-        preprocessed_inputs)
-    self._anchors = box_list_ops.concatenate(
-        self._anchor_generator.generate(
-            feature_map_spatial_dims,
-            im_height=image_shape[1],
-            im_width=image_shape[2]))
-    prediction_dict = self._box_predictor.predict(
-        feature_maps, self._anchor_generator.num_anchors_per_location())
-    box_encodings = tf.squeeze(
-        tf.concat(prediction_dict['box_encodings'], axis=1), axis=2)
-    class_predictions_with_background = tf.concat(
-        prediction_dict['class_predictions_with_background'], axis=1)
-    predictions_dict = {
-        'preprocessed_inputs': preprocessed_inputs,
-        'box_encodings': box_encodings,
-        'class_predictions_with_background': class_predictions_with_background,
-        'feature_maps': feature_maps,
-        'anchors': self._anchors.get()
-    }
-    return predictions_dict
+      self._anchors = box_list_ops.concatenate(
+          self._anchor_generator.generate(
+              feature_map_spatial_dims,
+              im_height=image_shape[1],
+              im_width=image_shape[2]))
+      prediction_dict = self._box_predictor.predict(
+          feature_maps, self._anchor_generator.num_anchors_per_location())
+      box_encodings = tf.squeeze(
+          tf.concat(prediction_dict['box_encodings'], axis=1), axis=2)
+      class_predictions_with_background = tf.concat(
+          prediction_dict['class_predictions_with_background'], axis=1)
+      predictions_dict = {
+          'preprocessed_inputs': preprocessed_inputs,
+          'box_encodings': box_encodings,
+          'class_predictions_with_background':
+          class_predictions_with_background,
+          'feature_maps': feature_maps,
+          'anchors': self._anchors.get()
+      }
+      return predictions_dict
 
   def _get_feature_map_spatial_dims(self, feature_maps):
     """Return list of spatial dimensions for each feature map in a list.

research/object_detection/meta_architectures/ssd_meta_arch_test.py

@@ -38,8 +38,7 @@ class FakeSSDFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
         depth_multiplier=0,
         min_depth=0,
         pad_to_multiple=1,
-        batch_norm_trainable=True,
-        conv_hyperparams=None)
+        conv_hyperparams_fn=None)
 
   def preprocess(self, resized_inputs):
     return tf.identity(resized_inputs)
@@ -124,7 +123,8 @@ class SsdMetaArchTest(test_case.TestCase):
         non_max_suppression_fn, tf.identity, classification_loss,
         localization_loss, classification_loss_weight, localization_loss_weight,
         normalize_loss_by_num_matches, hard_example_miner, add_summaries=False,
-        normalize_loc_loss_by_codesize=normalize_loc_loss_by_codesize)
+        normalize_loc_loss_by_codesize=normalize_loc_loss_by_codesize,
+        freeze_batchnorm=False, inplace_batchnorm_update=False)
     return model, num_classes, mock_anchor_generator.num_anchors(), code_size
 
   def test_preprocess_preserves_shapes_with_dynamic_input_image(self):

research/object_detection/models/embedded_ssd_mobilenet_v1_feature_extractor.py

@@ -49,12 +49,10 @@ class EmbeddedSSDMobileNetV1FeatureExtractor(
                depth_multiplier,
                min_depth,
                pad_to_multiple,
-               conv_hyperparams,
-               batch_norm_trainable=True,
+               conv_hyperparams_fn,
                reuse_weights=None,
                use_explicit_padding=False,
-               use_depthwise=False,
-               inplace_batchnorm_update=False):
+               use_depthwise=False):
     """MobileNetV1 Feature Extractor for Embedded-friendly SSD Models.
 
     Args:
@@ -63,20 +61,12 @@ class EmbeddedSSDMobileNetV1FeatureExtractor(
       min_depth: minimum feature extractor depth.
       pad_to_multiple: the nearest multiple to zero pad the input height and
         width dimensions to. For EmbeddedSSD it must be set to 1.
-      conv_hyperparams: tf slim arg_scope for conv2d and separable_conv2d ops.
-      batch_norm_trainable:  Whether to update batch norm parameters during
-        training or not. When training with a small batch size
-        (e.g. 1), it is desirable to disable batch norm update and use
-        pretrained batch norm params.
+      conv_hyperparams_fn: A function to construct tf slim arg_scope for conv2d
+        and separable_conv2d ops.
       reuse_weights: Whether to reuse variables. Default is None.
       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.
@@ -87,10 +77,9 @@ 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, inplace_batchnorm_update)
+        conv_hyperparams_fn, reuse_weights, use_explicit_padding, use_depthwise)
 
-  def _extract_features(self, preprocessed_inputs):
+  def extract_features(self, preprocessed_inputs):
     """Extract features from preprocessed inputs.
 
     Args:
@@ -130,7 +119,7 @@ class EmbeddedSSDMobileNetV1FeatureExtractor(
         'use_depthwise': self._use_depthwise,
     }
 
-    with slim.arg_scope(self._conv_hyperparams):
+    with slim.arg_scope(self._conv_hyperparams_fn()):
       with slim.arg_scope([slim.batch_norm], fused=False):
         with tf.variable_scope('MobilenetV1',
                                reuse=self._reuse_weights) as scope:

research/object_detection/models/embedded_ssd_mobilenet_v1_feature_extractor_test.py

@@ -25,7 +25,7 @@ class EmbeddedSSDMobileNetV1FeatureExtractorTest(
     ssd_feature_extractor_test.SsdFeatureExtractorTestBase):
 
   def _create_feature_extractor(self, depth_multiplier, pad_to_multiple,
-                                is_training=True, batch_norm_trainable=True):
+                                is_training=True):
     """Constructs a new feature extractor.
 
     Args:
@@ -33,18 +33,15 @@ class EmbeddedSSDMobileNetV1FeatureExtractorTest(
       pad_to_multiple: the nearest multiple to zero pad the input height and
         width dimensions to.
       is_training: whether the network is in training mode.
-      batch_norm_trainable: whether to update batch norm parameters during
-        training.
 
     Returns:
       an ssd_meta_arch.SSDFeatureExtractor object.
     """
     min_depth = 32
-    conv_hyperparams = {}
     return (embedded_ssd_mobilenet_v1_feature_extractor.
             EmbeddedSSDMobileNetV1FeatureExtractor(
                 is_training, depth_multiplier, min_depth, pad_to_multiple,
-                conv_hyperparams, batch_norm_trainable))
+                self.conv_hyperparams_fn))
 
   def test_extract_features_returns_correct_shapes_256(self):
     image_height = 256