Object detection changes: (#7208)

257914648  by lzc:

    Internal changes

--
257525973  by Zhichao Lu:

    Fixes bug that silently prevents checkpoints from loading when training w/ eager + functions. Also sets up scripts to run training.

--
257296614  by Zhichao Lu:

    Adding detection_features to model outputs

--
257234565  by Zhichao Lu:

    Fix wrong order of `classes_with_max_scores` in class-agnostic NMS caused by
    sorting in partitioned-NMS.

--
257232002  by ronnyvotel:

    Supporting `filter_nonoverlapping` option in np_box_list_ops.clip_to_window().

--
257198282  by Zhichao Lu:

    Adding the focal loss and l1 loss from the Objects as Points paper.

--
257089535  by Zhichao Lu:

    Create Keras based ssd + resnetv1 + fpn.

--
257087407  by Zhichao Lu:

    Make object_detection/data_decoders Python3-compatible.

--
257004582  by Zhichao Lu:

    Updates _decode_raw_data_into_masks_and_boxes to the latest binary masks-to-string encoding fo...

research/object_detection/models/faster_rcnn_pnas_feature_extractor.py

@@ -21,6 +21,7 @@ Based on PNASNet model: https://arxiv.org/abs/1712.00559
 import tensorflow as tf
 
 from object_detection.meta_architectures import faster_rcnn_meta_arch
+from object_detection.utils import variables_helper
 from nets.nasnet import nasnet_utils
 from nets.nasnet import pnasnet
 
@@ -302,7 +303,7 @@ class FasterRCNNPNASFeatureExtractor(
       the model graph.
     """
     variables_to_restore = {}
-    for variable in tf.global_variables():
+    for variable in variables_helper.get_global_variables_safely():
       if variable.op.name.startswith(
           first_stage_feature_extractor_scope):
         var_name = variable.op.name.replace(

research/object_detection/models/faster_rcnn_resnet_v1_feature_extractor_test.py

@@ -25,6 +25,7 @@ class FasterRcnnResnetV1FeatureExtractorTest(tf.test.TestCase):
 
   def _build_feature_extractor(self,
                                first_stage_features_stride,
+                               activation_fn=tf.nn.relu,
                                architecture='resnet_v1_101'):
     feature_extractor_map = {
         'resnet_v1_50':
@@ -37,6 +38,7 @@ class FasterRcnnResnetV1FeatureExtractorTest(tf.test.TestCase):
     return feature_extractor_map[architecture](
         is_training=False,
         first_stage_features_stride=first_stage_features_stride,
+        activation_fn=activation_fn,
         batch_norm_trainable=False,
         reuse_weights=None,
         weight_decay=0.0)
@@ -132,6 +134,32 @@ class FasterRcnnResnetV1FeatureExtractorTest(tf.test.TestCase):
       features_shape_out = sess.run(features_shape)
       self.assertAllEqual(features_shape_out, [3, 7, 7, 2048])
 
+  def test_overwriting_activation_fn(self):
+    for architecture in ['resnet_v1_50', 'resnet_v1_101', 'resnet_v1_152']:
+      feature_extractor = self._build_feature_extractor(
+          first_stage_features_stride=16,
+          architecture=architecture,
+          activation_fn=tf.nn.relu6)
+      preprocessed_inputs = tf.random_uniform([4, 224, 224, 3],
+                                              maxval=255,
+                                              dtype=tf.float32)
+      rpn_feature_map, _ = feature_extractor.extract_proposal_features(
+          preprocessed_inputs, scope='TestStage1Scope')
+      _ = feature_extractor.extract_box_classifier_features(
+          rpn_feature_map, scope='TestStaget2Scope')
+      conv_ops = [
+          op for op in tf.get_default_graph().get_operations()
+          if op.type == 'Relu6'
+      ]
+      op_names = [op.name for op in conv_ops]
+
+      self.assertIsNotNone(conv_ops)
+      self.assertIn('TestStage1Scope/resnet_v1_50/resnet_v1_50/conv1/Relu6',
+                    op_names)
+      self.assertIn(
+          'TestStaget2Scope/resnet_v1_50/block4/unit_1/bottleneck_v1/conv1/Relu6',
+          op_names)
+
 
 if __name__ == '__main__':
   tf.test.main()

research/object_detection/models/feature_map_generators.py

@@ -79,14 +79,19 @@ def create_conv_block(
   """
   layers = []
   if use_depthwise:
-    layers.append(tf.keras.layers.SeparableConv2D(
-        depth,
-        [kernel_size, kernel_size],
-        depth_multiplier=1,
-        padding=padding,
-        strides=stride,
-        name=layer_name + '_depthwise_conv',
-        **conv_hyperparams.params()))
+    kwargs = conv_hyperparams.params()
+    # Both the regularizer and initializer apply to the depthwise layer,
+    # so we remap the kernel_* to depthwise_* here.
+    kwargs['depthwise_regularizer'] = kwargs['kernel_regularizer']
+    kwargs['depthwise_initializer'] = kwargs['kernel_initializer']
+    layers.append(
+        tf.keras.layers.SeparableConv2D(
+            depth, [kernel_size, kernel_size],
+            depth_multiplier=1,
+            padding=padding,
+            strides=stride,
+            name=layer_name + '_depthwise_conv',
+            **kwargs))
   else:
     layers.append(tf.keras.layers.Conv2D(
         depth,

research/object_detection/models/keras_models/mobilenet_v2.py

@@ -160,7 +160,12 @@ class _LayersOverride(object):
     """
     if self._conv_hyperparams:
       kwargs = self._conv_hyperparams.params(**kwargs)
+      # Both the regularizer and initializer apply to the depthwise layer in
+      # MobilenetV1, so we remap the kernel_* to depthwise_* here.
+      kwargs['depthwise_regularizer'] = kwargs['kernel_regularizer']
+      kwargs['depthwise_initializer'] = kwargs['kernel_initializer']
     else:
+      kwargs['depthwise_regularizer'] = self.regularizer
       kwargs['depthwise_initializer'] = self.initializer
 
     kwargs['padding'] = 'same'

research/object_detection/models/ssd_mobilenet_v1_fpn_feature_extractor_test.py

@@ -220,7 +220,7 @@ class SsdMobilenetV1FpnFeatureExtractorTest(
       _ = feature_extractor.extract_features(preprocessed_image)
 
     self.assertTrue(
-        any(op.type == 'FusedBatchNorm'
+        any('FusedBatchNorm' in op.type
             for op in tf.get_default_graph().get_operations()))