Release MobileDet code and model, and require tf_slim installation for OD API. (#8562)

* Merged commit includes the following changes:
311933687  by Sergio Guadarrama:

    Removes spurios use of tf.compat.v2, which results in spurious tf.compat.v1.compat.v2. Adds basic test to nasnet_utils.
    Replaces all remaining import tensorflow as tf with import tensorflow.compat.v1 as tf

--
311766063  by Sergio Guadarrama:

    Removes explicit tf.compat.v1 in all call sites (we already import tf.compat.v1, so this code was  doing tf.compat.v1.compat.v1). The existing code worked in latest version of tensorflow, 2.2, (and 1.15) but not in 1.14 or in 2.0.0a, this CL fixes it.

--
311624958  by Sergio Guadarrama:

    Updates README that doesn't render properly in github documentation

--
310980959  by Sergio Guadarrama:

    Moves research_models/slim off tf.contrib.slim/layers/framework to tf_slim

--
310263156  by Sergio Guadarrama:

    Adds model breakdown for MobilenetV3

--
308640...

research/object_detection/builders/graph_rewriter_builder_test.py

@@ -14,13 +14,14 @@
 # ==============================================================================
 """Tests for graph_rewriter_builder."""
 import mock
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
+import tf_slim as slim
+
 from object_detection.builders import graph_rewriter_builder
 from object_detection.protos import graph_rewriter_pb2
 
 # pylint: disable=g-import-not-at-top
 try:
-  from tensorflow.contrib import layers as contrib_layers
   from tensorflow.contrib import quantize as contrib_quantize
 except ImportError:
   # TF 2.0 doesn't ship with contrib.
@@ -34,7 +35,7 @@ class QuantizationBuilderTest(tf.test.TestCase):
     with mock.patch.object(
         contrib_quantize,
         'experimental_create_training_graph') as mock_quant_fn:
-      with mock.patch.object(contrib_layers,
+      with mock.patch.object(slim,
                              'summarize_collection') as mock_summarize_col:
         graph_rewriter_proto = graph_rewriter_pb2.GraphRewriter()
         graph_rewriter_proto.quantization.delay = 10
@@ -51,7 +52,7 @@ class QuantizationBuilderTest(tf.test.TestCase):
   def testQuantizationBuilderSetsUpCorrectEvalArguments(self):
     with mock.patch.object(contrib_quantize,
                            'experimental_create_eval_graph') as mock_quant_fn:
-      with mock.patch.object(contrib_layers,
+      with mock.patch.object(slim,
                              'summarize_collection') as mock_summarize_col:
         graph_rewriter_proto = graph_rewriter_pb2.GraphRewriter()
         graph_rewriter_proto.quantization.delay = 10

research/object_detection/builders/hyperparams_builder.py

@@ -14,19 +14,13 @@
 # ==============================================================================
 
 """Builder function to construct tf-slim arg_scope for convolution, fc ops."""
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
+import tf_slim as slim
 
 from object_detection.core import freezable_batch_norm
 from object_detection.protos import hyperparams_pb2
 from object_detection.utils import context_manager
 
-# pylint: disable=g-import-not-at-top
-try:
-  from tensorflow.contrib import slim
-  from tensorflow.contrib import layers as contrib_layers
-except ImportError:
-  # TF 2.0 doesn't ship with contrib.
-  pass
 # pylint: enable=g-import-not-at-top
 
 
@@ -223,7 +217,7 @@ def build(hyperparams_config, is_training):
     batch_norm_params = _build_batch_norm_params(
         hyperparams_config.batch_norm, is_training)
   if hyperparams_config.HasField('group_norm'):
-    normalizer_fn = contrib_layers.group_norm
+    normalizer_fn = slim.group_norm
   affected_ops = [slim.conv2d, slim.separable_conv2d, slim.conv2d_transpose]
   if hyperparams_config.HasField('op') and (
       hyperparams_config.op == hyperparams_pb2.Hyperparams.FC):

research/object_detection/builders/hyperparams_builder_test.py

@@ -17,22 +17,14 @@
 """Tests object_detection.core.hyperparams_builder."""
 
 import numpy as np
-import tensorflow as tf
-
+import tensorflow.compat.v1 as tf
+import tf_slim as slim
 from google.protobuf import text_format
 
 from object_detection.builders import hyperparams_builder
 from object_detection.core import freezable_batch_norm
 from object_detection.protos import hyperparams_pb2
 
-# pylint: disable=g-import-not-at-top
-try:
-  from tensorflow.contrib import slim
-except ImportError:
-  # TF 2.0 doesn't ship with contrib.
-  pass
-# pylint: enable=g-import-not-at-top
-
 
 def _get_scope_key(op):
   return getattr(op, '_key_op', str(op))

research/object_detection/builders/image_resizer_builder.py

@@ -14,7 +14,7 @@
 # ==============================================================================
 """Builder function for image resizing operations."""
 import functools
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from object_detection.core import preprocessor
 from object_detection.protos import image_resizer_pb2

research/object_detection/builders/image_resizer_builder_test.py

@@ -14,7 +14,7 @@
 # ==============================================================================
 """Tests for object_detection.builders.image_resizer_builder."""
 import numpy as np
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 from google.protobuf import text_format
 from object_detection.builders import image_resizer_builder
 from object_detection.protos import image_resizer_pb2

research/object_detection/builders/input_reader_builder.py

@@ -28,20 +28,21 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from object_detection.data_decoders import tf_example_decoder
+from object_detection.data_decoders import tf_sequence_example_decoder
 from object_detection.protos import input_reader_pb2
 
 # pylint: disable=g-import-not-at-top
 try:
-  from tensorflow.contrib import slim as contrib_slim
+  import tf_slim as slim
 except ImportError:
   # TF 2.0 doesn't ship with contrib.
   pass
 # pylint: enable=g-import-not-at-top
 
-parallel_reader = contrib_slim.parallel_reader
+parallel_reader = slim.parallel_reader
 
 
 def build(input_reader_config):
@@ -80,11 +81,18 @@ def build(input_reader_config):
     label_map_proto_file = None
     if input_reader_config.HasField('label_map_path'):
       label_map_proto_file = input_reader_config.label_map_path
-    decoder = tf_example_decoder.TfExampleDecoder(
-        load_instance_masks=input_reader_config.load_instance_masks,
-        instance_mask_type=input_reader_config.mask_type,
-        label_map_proto_file=label_map_proto_file,
-        load_context_features=input_reader_config.load_context_features)
-    return decoder.decode(string_tensor)
-
+    input_type = input_reader_config.input_type
+    if input_type == input_reader_pb2.InputType.TF_EXAMPLE:
+      decoder = tf_example_decoder.TfExampleDecoder(
+          load_instance_masks=input_reader_config.load_instance_masks,
+          instance_mask_type=input_reader_config.mask_type,
+          label_map_proto_file=label_map_proto_file,
+          load_context_features=input_reader_config.load_context_features)
+      return decoder.decode(string_tensor)
+    elif input_type == input_reader_pb2.InputType.TF_SEQUENCE_EXAMPLE:
+      decoder = tf_sequence_example_decoder.TfSequenceExampleDecoder(
+          label_map_proto_file=label_map_proto_file,
+          load_context_features=input_reader_config.load_context_features)
+      return decoder.decode(string_tensor)
+    raise ValueError('Unsupported input_type.')
   raise ValueError('Unsupported input_reader_config.')

research/object_detection/builders/input_reader_builder_test.py

@@ -17,16 +17,24 @@
 
 import os
 import numpy as np
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from google.protobuf import text_format
 
 from object_detection.builders import input_reader_builder
 from object_detection.core import standard_fields as fields
+from object_detection.dataset_tools import seq_example_util
 from object_detection.protos import input_reader_pb2
 from object_detection.utils import dataset_util
 
 
+def _get_labelmap_path():
+  """Returns an absolute path to label map file."""
+  parent_path = os.path.dirname(tf.resource_loader.get_data_files_path())
+  return os.path.join(parent_path, 'data',
+                      'pet_label_map.pbtxt')
+
+
 class InputReaderBuilderTest(tf.test.TestCase):
 
   def create_tf_record(self):
@@ -54,6 +62,56 @@ class InputReaderBuilderTest(tf.test.TestCase):
 
     return path
 
+  def _make_random_serialized_jpeg_images(self, num_frames, image_height,
+                                          image_width):
+    images = tf.cast(tf.random.uniform(
+        [num_frames, image_height, image_width, 3],
+        maxval=256,
+        dtype=tf.int32), dtype=tf.uint8)
+    images_list = tf.unstack(images, axis=0)
+    encoded_images_list = [tf.io.encode_jpeg(image) for image in images_list]
+    with tf.Session() as sess:
+      encoded_images = sess.run(encoded_images_list)
+    return encoded_images
+
+  def create_tf_record_sequence_example(self):
+    path = os.path.join(self.get_temp_dir(), 'tfrecord')
+    writer = tf.python_io.TFRecordWriter(path)
+    num_frames = 4
+    image_height = 20
+    image_width = 30
+    image_source_ids = [str(i) for i in range(num_frames)]
+    with self.test_session():
+      encoded_images = self._make_random_serialized_jpeg_images(
+          num_frames, image_height, image_width)
+      sequence_example_serialized = seq_example_util.make_sequence_example(
+          dataset_name='video_dataset',
+          video_id='video',
+          encoded_images=encoded_images,
+          image_height=image_height,
+          image_width=image_width,
+          image_source_ids=image_source_ids,
+          image_format='JPEG',
+          is_annotated=[[1], [1], [1], [1]],
+          bboxes=[
+              [[]],  # Frame 0.
+              [[0., 0., 1., 1.]],  # Frame 1.
+              [[0., 0., 1., 1.],
+               [0.1, 0.1, 0.2, 0.2]],  # Frame 2.
+              [[]],  # Frame 3.
+          ],
+          label_strings=[
+              [],  # Frame 0.
+              ['Abyssinian'],  # Frame 1.
+              ['Abyssinian', 'american_bulldog'],  # Frame 2.
+              [],  # Frame 3
+          ]).SerializeToString()
+
+    writer.write(sequence_example_serialized)
+    writer.close()
+
+    return path
+
   def create_tf_record_with_context(self):
     path = os.path.join(self.get_temp_dir(), 'tfrecord')
     writer = tf.python_io.TFRecordWriter(path)
@@ -124,6 +182,46 @@ class InputReaderBuilderTest(tf.test.TestCase):
         [0.0, 0.0, 1.0, 1.0],
         output_dict[fields.InputDataFields.groundtruth_boxes][0])
 
+  def test_build_tf_record_input_reader_sequence_example(self):
+    tf_record_path = self.create_tf_record_sequence_example()
+
+    input_reader_text_proto = """
+      shuffle: false
+      num_readers: 1
+      input_type: TF_SEQUENCE_EXAMPLE
+      tf_record_input_reader {{
+        input_path: '{0}'
+      }}
+    """.format(tf_record_path)
+    input_reader_proto = input_reader_pb2.InputReader()
+    input_reader_proto.label_map_path = _get_labelmap_path()
+    text_format.Merge(input_reader_text_proto, input_reader_proto)
+    tensor_dict = input_reader_builder.build(input_reader_proto)
+
+    with tf.train.MonitoredSession() as sess:
+      output_dict = sess.run(tensor_dict)
+
+    expected_groundtruth_classes = [[-1, -1], [1, -1], [1, 2], [-1, -1]]
+    expected_groundtruth_boxes = [[[0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0]],
+                                  [[0.0, 0.0, 1.0, 1.0], [0.0, 0.0, 0.0, 0.0]],
+                                  [[0.0, 0.0, 1.0, 1.0], [0.1, 0.1, 0.2, 0.2]],
+                                  [[0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0]]]
+    expected_num_groundtruth_boxes = [0, 1, 2, 0]
+
+    self.assertNotIn(
+        fields.InputDataFields.groundtruth_instance_masks, output_dict)
+    # sequence example images are encoded
+    self.assertEqual((4,), output_dict[fields.InputDataFields.image].shape)
+    self.assertAllEqual(expected_groundtruth_classes,
+                        output_dict[fields.InputDataFields.groundtruth_classes])
+    self.assertEqual(
+        (4, 2, 4), output_dict[fields.InputDataFields.groundtruth_boxes].shape)
+    self.assertAllClose(expected_groundtruth_boxes,
+                        output_dict[fields.InputDataFields.groundtruth_boxes])
+    self.assertAllClose(
+        expected_num_groundtruth_boxes,
+        output_dict[fields.InputDataFields.num_groundtruth_boxes])
+
   def test_build_tf_record_input_reader_with_context(self):
     tf_record_path = self.create_tf_record_with_context()
 

research/object_detection/builders/losses_builder_test.py

@@ -15,7 +15,7 @@
 
 """Tests for losses_builder."""
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from google.protobuf import text_format
 from object_detection.builders import losses_builder

research/object_detection/builders/matcher_builder_test.py

@@ -15,7 +15,7 @@
 
 """Tests for matcher_builder."""
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from google.protobuf import text_format
 from object_detection.builders import matcher_builder

research/object_detection/builders/model_builder.py

@@ -75,6 +75,9 @@ if tf_version.is_tf1():
   from object_detection.models.ssd_mobilenet_v2_feature_extractor import SSDMobileNetV2FeatureExtractor
   from object_detection.models.ssd_mobilenet_v3_feature_extractor import SSDMobileNetV3LargeFeatureExtractor
   from object_detection.models.ssd_mobilenet_v3_feature_extractor import SSDMobileNetV3SmallFeatureExtractor
+  from object_detection.models.ssd_mobiledet_feature_extractor import SSDMobileDetCPUFeatureExtractor
+  from object_detection.models.ssd_mobiledet_feature_extractor import SSDMobileDetDSPFeatureExtractor
+  from object_detection.models.ssd_mobiledet_feature_extractor import SSDMobileDetEdgeTPUFeatureExtractor
   from object_detection.models.ssd_pnasnet_feature_extractor import SSDPNASNetFeatureExtractor
   from object_detection.predictors import rfcn_box_predictor
 # pylint: enable=g-import-not-at-top
@@ -156,6 +159,9 @@ if tf_version.is_tf1():
           EmbeddedSSDMobileNetV1FeatureExtractor,
       'ssd_pnasnet':
           SSDPNASNetFeatureExtractor,
+      'ssd_mobiledet_cpu': SSDMobileDetCPUFeatureExtractor,
+      'ssd_mobiledet_dsp': SSDMobileDetDSPFeatureExtractor,
+      'ssd_mobiledet_edgetpu': SSDMobileDetEdgeTPUFeatureExtractor,
   }
 
   FASTER_RCNN_FEATURE_EXTRACTOR_CLASS_MAP = {
@@ -794,7 +800,25 @@ def object_center_proto_to_params(oc_config):
       object_center_loss_weight=oc_config.object_center_loss_weight,
       heatmap_bias_init=oc_config.heatmap_bias_init,
       min_box_overlap_iou=oc_config.min_box_overlap_iou,
-      max_box_predictions=oc_config.max_box_predictions)
+      max_box_predictions=oc_config.max_box_predictions,
+      use_labeled_classes=oc_config.use_labeled_classes)
+
+
+def mask_proto_to_params(mask_config):
+  """Converts CenterNet.MaskEstimation proto to parameter namedtuple."""
+  loss = losses_pb2.Loss()
+  # Add dummy localization loss to avoid the loss_builder throwing error.
+  loss.localization_loss.weighted_l2.CopyFrom(
+      losses_pb2.WeightedL2LocalizationLoss())
+  loss.classification_loss.CopyFrom(mask_config.classification_loss)
+  classification_loss, _, _, _, _, _, _ = (losses_builder.build(loss))
+  return center_net_meta_arch.MaskParams(
+      classification_loss=classification_loss,
+      task_loss_weight=mask_config.task_loss_weight,
+      mask_height=mask_config.mask_height,
+      mask_width=mask_config.mask_width,
+      score_threshold=mask_config.score_threshold,
+      heatmap_bias_init=mask_config.heatmap_bias_init)
 
 
 def _build_center_net_model(center_net_config, is_training, add_summaries):
@@ -844,6 +868,11 @@ def _build_center_net_model(center_net_config, is_training, add_summaries):
         keypoint_class_id_set.add(kp_params.class_id)
     if len(all_keypoint_indices) > len(set(all_keypoint_indices)):
       raise ValueError('Some keypoint indices are used more than once.')
+
+  mask_params = None
+  if center_net_config.HasField('mask_estimation_task'):
+    mask_params = mask_proto_to_params(center_net_config.mask_estimation_task)
+
   return center_net_meta_arch.CenterNetMetaArch(
       is_training=is_training,
       add_summaries=add_summaries,
@@ -852,7 +881,8 @@ def _build_center_net_model(center_net_config, is_training, add_summaries):
       image_resizer_fn=image_resizer_fn,
       object_center_params=object_center_params,
       object_detection_params=object_detection_params,
-      keypoint_params_dict=keypoint_params_dict)
+      keypoint_params_dict=keypoint_params_dict,
+      mask_params=mask_params)
 
 
 def _build_center_net_feature_extractor(

research/object_detection/builders/model_builder_tf1_test.py

@@ -16,7 +16,7 @@
 """Tests for model_builder under TensorFlow 1.X."""
 
 from absl.testing import parameterized
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from object_detection.builders import model_builder
 from object_detection.builders import model_builder_test

research/object_detection/builders/optimizer_builder.py

@@ -15,7 +15,7 @@
 
 """Functions to build DetectionModel training optimizers."""
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 
 from tensorflow.contrib import opt as tf_opt

research/object_detection/builders/optimizer_builder_tf1_test.py

@@ -21,7 +21,7 @@ from __future__ import division
 from __future__ import print_function
 
 import six
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from google.protobuf import text_format
 

research/object_detection/builders/post_processing_builder.py

@@ -16,7 +16,7 @@
 """Builder function for post processing operations."""
 import functools
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 from object_detection.builders import calibration_builder
 from object_detection.core import post_processing
 from object_detection.protos import post_processing_pb2
@@ -102,7 +102,8 @@ def _build_non_max_suppressor(nms_config):
       soft_nms_sigma=nms_config.soft_nms_sigma,
       use_partitioned_nms=nms_config.use_partitioned_nms,
       use_combined_nms=nms_config.use_combined_nms,
-      change_coordinate_frame=nms_config.change_coordinate_frame)
+      change_coordinate_frame=nms_config.change_coordinate_frame,
+      use_hard_nms=nms_config.use_hard_nms)
 
   return non_max_suppressor_fn
 

research/object_detection/builders/post_processing_builder_test.py

@@ -15,7 +15,7 @@
 
 """Tests for post_processing_builder."""
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 from google.protobuf import text_format
 from object_detection.builders import post_processing_builder
 from object_detection.protos import post_processing_pb2

research/object_detection/builders/preprocessor_builder.py

@@ -15,7 +15,7 @@
 
 """Builder for preprocessing steps."""
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from object_detection.core import preprocessor
 from object_detection.protos import preprocessor_pb2

research/object_detection/builders/preprocessor_builder_test.py

@@ -15,7 +15,7 @@
 
 """Tests for preprocessor_builder."""
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from google.protobuf import text_format
 

research/object_detection/builders/region_similarity_calculator_builder_test.py

@@ -15,7 +15,7 @@
 
 """Tests for region_similarity_calculator_builder."""
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from google.protobuf import text_format
 from object_detection.builders import region_similarity_calculator_builder

research/object_detection/builders/target_assigner_builder_test.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 # ==============================================================================
 
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 from google.protobuf import text_format
 

research/object_detection/core/anchor_generator.py

@@ -38,7 +38,7 @@ from abc import abstractmethod
 
 import six
 from six.moves import zip
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
 
 
 class AnchorGenerator(six.with_metaclass(ABCMeta, object)):