Merged commit includes the following changes: (#7800)

280142968  by Zhichao Lu:

    Opensource MobilenetEdgeTPU + ssdlite into third-party object detection APIs on EdgeTPU.

--
280134001  by Zhichao Lu:

    Adds MobilenetEdgeTpu + ssdlite into internal object detection APIs on EdgeTPU.

--
278941778  by Zhichao Lu:

    Add support for fixed input shapes for 'encoded_image_string_tensor' and 'tf_example' inputs.

--
278933274  by Zhichao Lu:

      Adding fool proof check to avoid using 1x1 depthwise conv op.

--
278762192  by Zhichao Lu:

    Ensure correct number of iterations after training resumes.

--
278746440  by Zhichao Lu:

    Internal change.

--
278006953  by Zhichao Lu:

    Internal changes to tf.contrib symbols

--
278006330  by Zhichao Lu:

    Internal changes to tf.contrib symbols

--
277593959  by Zhichao Lu:

      Make the ssd_feature_extractor_test.py PY3 compatible. The "six.zip" will use "itertools.izip" in Python 2 and "zip" in Python 3....

research/object_detection/models/ssd_mobilenet_v1_ppn_feature_extractor.py

@@ -16,6 +16,7 @@
 """SSDFeatureExtractor for MobilenetV1 PPN features."""
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.meta_architectures import ssd_meta_arch
 from object_detection.models import feature_map_generators
@@ -24,7 +25,7 @@ from object_detection.utils import ops
 from object_detection.utils import shape_utils
 from nets import mobilenet_v1
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class SSDMobileNetV1PpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):

research/object_detection/models/ssd_mobilenet_v1_ppn_feature_extractor_test.py

@@ -16,11 +16,12 @@
 """Tests for ssd_mobilenet_v1_ppn_feature_extractor."""
 import numpy as np
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.models import ssd_feature_extractor_test
 from object_detection.models import ssd_mobilenet_v1_ppn_feature_extractor
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class SsdMobilenetV1PpnFeatureExtractorTest(

research/object_detection/models/ssd_mobilenet_v2_feature_extractor.py

@@ -16,6 +16,7 @@
 """SSDFeatureExtractor for MobilenetV2 features."""
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.meta_architectures import ssd_meta_arch
 from object_detection.models import feature_map_generators
@@ -25,7 +26,7 @@ from object_detection.utils import shape_utils
 from nets.mobilenet import mobilenet
 from nets.mobilenet import mobilenet_v2
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class SSDMobileNetV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):

research/object_detection/models/ssd_mobilenet_v2_feature_extractor_test.py

@@ -18,12 +18,13 @@ from absl.testing import parameterized
 
 import numpy as np
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.models import ssd_feature_extractor_test
 from object_detection.models import ssd_mobilenet_v2_feature_extractor
 from object_detection.models import ssd_mobilenet_v2_keras_feature_extractor
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 @parameterized.parameters(

research/object_detection/models/ssd_mobilenet_v2_fpn_feature_extractor.py

@@ -18,6 +18,7 @@
 import copy
 import functools
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.meta_architectures import ssd_meta_arch
 from object_detection.models import feature_map_generators
@@ -27,7 +28,7 @@ from object_detection.utils import shape_utils
 from nets.mobilenet import mobilenet
 from nets.mobilenet import mobilenet_v2
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 # A modified config of mobilenet v2 that makes it more detection friendly.

research/object_detection/models/ssd_mobilenet_v2_fpn_feature_extractor_test.py

@@ -21,12 +21,13 @@ Keras-based Mobilenet V2 FPN feature extractors in SSD.
 from absl.testing import parameterized
 import numpy as np
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.models import ssd_feature_extractor_test
 from object_detection.models import ssd_mobilenet_v2_fpn_feature_extractor
 from object_detection.models import ssd_mobilenet_v2_fpn_keras_feature_extractor
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 @parameterized.parameters(

research/object_detection/models/ssd_mobilenet_v3_feature_extractor.py

@@ -15,6 +15,7 @@
 """SSDFeatureExtractor for MobileNetV3 features."""
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.meta_architectures import ssd_meta_arch
 from object_detection.models import feature_map_generators
@@ -24,10 +25,10 @@ from object_detection.utils import shape_utils
 from nets.mobilenet import mobilenet
 from nets.mobilenet import mobilenet_v3
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
-class _SSDMobileNetV3FeatureExtractorBase(ssd_meta_arch.SSDFeatureExtractor):
+class SSDMobileNetV3FeatureExtractorBase(ssd_meta_arch.SSDFeatureExtractor):
   """Base class of SSD feature extractor using MobilenetV3 features."""
 
   def __init__(self,
@@ -41,7 +42,8 @@ class _SSDMobileNetV3FeatureExtractorBase(ssd_meta_arch.SSDFeatureExtractor):
                reuse_weights=None,
                use_explicit_padding=False,
                use_depthwise=False,
-               override_base_feature_extractor_hyperparams=False):
+               override_base_feature_extractor_hyperparams=False,
+               scope_name='MobilenetV3'):
     """MobileNetV3 Feature Extractor for SSD Models.
 
     MobileNet v3. Details found in:
@@ -66,8 +68,9 @@ class _SSDMobileNetV3FeatureExtractorBase(ssd_meta_arch.SSDFeatureExtractor):
       override_base_feature_extractor_hyperparams: Whether to override
         hyperparameters of the base feature extractor with the one from
         `conv_hyperparams_fn`.
+      scope_name: scope name (string) of network variables.
     """
-    super(_SSDMobileNetV3FeatureExtractorBase, self).__init__(
+    super(SSDMobileNetV3FeatureExtractorBase, self).__init__(
         is_training=is_training,
         depth_multiplier=depth_multiplier,
         min_depth=min_depth,
@@ -80,6 +83,7 @@ class _SSDMobileNetV3FeatureExtractorBase(ssd_meta_arch.SSDFeatureExtractor):
     )
     self._conv_defs = conv_defs
     self._from_layer = from_layer
+    self._scope_name = scope_name
 
   def preprocess(self, resized_inputs):
     """SSD preprocessing.
@@ -129,7 +133,8 @@ class _SSDMobileNetV3FeatureExtractorBase(ssd_meta_arch.SSDFeatureExtractor):
         'use_explicit_padding': self._use_explicit_padding,
     }
 
-    with tf.variable_scope('MobilenetV3', reuse=self._reuse_weights) as scope:
+    with tf.variable_scope(
+        self._scope_name, reuse=self._reuse_weights) as scope:
       with slim.arg_scope(
           mobilenet_v3.training_scope(is_training=None, bn_decay=0.9997)), \
           slim.arg_scope(
@@ -137,7 +142,6 @@ class _SSDMobileNetV3FeatureExtractorBase(ssd_meta_arch.SSDFeatureExtractor):
         with (slim.arg_scope(self._conv_hyperparams_fn())
               if self._override_base_feature_extractor_hyperparams else
               context_manager.IdentityContextManager()):
-          # TODO(bochen): switch to v3 modules once v3 is properly refactored.
           _, image_features = mobilenet_v3.mobilenet_base(
               ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
               conv_defs=self._conv_defs,
@@ -156,7 +160,7 @@ class _SSDMobileNetV3FeatureExtractorBase(ssd_meta_arch.SSDFeatureExtractor):
     return feature_maps.values()
 
 
-class SSDMobileNetV3LargeFeatureExtractor(_SSDMobileNetV3FeatureExtractorBase):
+class SSDMobileNetV3LargeFeatureExtractor(SSDMobileNetV3FeatureExtractorBase):
   """Mobilenet V3-Large feature extractor."""
 
   def __init__(self,
@@ -168,7 +172,8 @@ class SSDMobileNetV3LargeFeatureExtractor(_SSDMobileNetV3FeatureExtractorBase):
                reuse_weights=None,
                use_explicit_padding=False,
                use_depthwise=False,
-               override_base_feature_extractor_hyperparams=False):
+               override_base_feature_extractor_hyperparams=False,
+               scope_name='MobilenetV3'):
     super(SSDMobileNetV3LargeFeatureExtractor, self).__init__(
         conv_defs=mobilenet_v3.V3_LARGE_DETECTION,
         from_layer=['layer_14/expansion_output', 'layer_17'],
@@ -180,11 +185,12 @@ class SSDMobileNetV3LargeFeatureExtractor(_SSDMobileNetV3FeatureExtractorBase):
         reuse_weights=reuse_weights,
         use_explicit_padding=use_explicit_padding,
         use_depthwise=use_depthwise,
-        override_base_feature_extractor_hyperparams=override_base_feature_extractor_hyperparams
+        override_base_feature_extractor_hyperparams=override_base_feature_extractor_hyperparams,
+        scope_name=scope_name
     )
 
 
-class SSDMobileNetV3SmallFeatureExtractor(_SSDMobileNetV3FeatureExtractorBase):
+class SSDMobileNetV3SmallFeatureExtractor(SSDMobileNetV3FeatureExtractorBase):
   """Mobilenet V3-Small feature extractor."""
 
   def __init__(self,
@@ -196,7 +202,8 @@ class SSDMobileNetV3SmallFeatureExtractor(_SSDMobileNetV3FeatureExtractorBase):
                reuse_weights=None,
                use_explicit_padding=False,
                use_depthwise=False,
-               override_base_feature_extractor_hyperparams=False):
+               override_base_feature_extractor_hyperparams=False,
+               scope_name='MobilenetV3'):
     super(SSDMobileNetV3SmallFeatureExtractor, self).__init__(
         conv_defs=mobilenet_v3.V3_SMALL_DETECTION,
         from_layer=['layer_10/expansion_output', 'layer_13'],
@@ -208,5 +215,6 @@ class SSDMobileNetV3SmallFeatureExtractor(_SSDMobileNetV3FeatureExtractorBase):
         reuse_weights=reuse_weights,
         use_explicit_padding=use_explicit_padding,
         use_depthwise=use_depthwise,
-        override_base_feature_extractor_hyperparams=override_base_feature_extractor_hyperparams
+        override_base_feature_extractor_hyperparams=override_base_feature_extractor_hyperparams,
+        scope_name=scope_name
     )

research/object_detection/models/ssd_mobilenet_v3_feature_extractor_test.py

@@ -15,12 +15,13 @@
 """Tests for ssd_mobilenet_v3_feature_extractor."""
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.models import ssd_mobilenet_v3_feature_extractor
 from object_detection.models import ssd_mobilenet_v3_feature_extractor_testbase
 
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class SsdMobilenetV3LargeFeatureExtractorTest(

research/object_detection/models/ssd_mobilenet_v3_feature_extractor_testbase.py

@@ -18,11 +18,12 @@ import abc
 
 import numpy as np
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.models import ssd_feature_extractor_test
 
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class _SsdMobilenetV3FeatureExtractorTestBase(

research/object_detection/models/ssd_pnasnet_feature_extractor.py

@@ -19,6 +19,7 @@ Based on PNASNet ImageNet model: https://arxiv.org/abs/1712.00559
 """
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.meta_architectures import ssd_meta_arch
 from object_detection.models import feature_map_generators
@@ -27,7 +28,7 @@ from object_detection.utils import ops
 from object_detection.utils import variables_helper
 from nets.nasnet import pnasnet
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 def pnasnet_large_arg_scope_for_detection(is_batch_norm_training=False):

research/object_detection/models/ssd_pnasnet_feature_extractor_test.py

@@ -16,11 +16,12 @@
 """Tests for ssd_pnas_feature_extractor."""
 import numpy as np
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.models import ssd_feature_extractor_test
 from object_detection.models import ssd_pnasnet_feature_extractor
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class SsdPnasNetFeatureExtractorTest(

research/object_detection/models/ssd_resnet_v1_fpn_feature_extractor.py

@@ -18,6 +18,7 @@ See https://arxiv.org/abs/1708.02002 for details.
 """
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.meta_architectures import ssd_meta_arch
 from object_detection.models import feature_map_generators
@@ -26,7 +27,7 @@ from object_detection.utils import ops
 from object_detection.utils import shape_utils
 from nets import resnet_v1
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class SSDResnetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):

research/object_detection/models/ssd_resnet_v1_ppn_feature_extractor.py

@@ -15,6 +15,7 @@
 """SSD feature extractors based on Resnet v1 and PPN architectures."""
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.meta_architectures import ssd_meta_arch
 from object_detection.models import feature_map_generators
@@ -23,7 +24,7 @@ from object_detection.utils import ops
 from object_detection.utils import shape_utils
 from nets import resnet_v1
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class _SSDResnetPpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):

research/object_detection/predictors/convolutional_box_predictor_test.py

@@ -198,7 +198,7 @@ class ConvolutionalBoxPredictorTest(test_case.TestCase):
             max_depth=32,
             num_layers_before_predictor=1,
             dropout_keep_prob=0.8,
-            kernel_size=1,
+            kernel_size=3,
             box_code_size=4,
             use_dropout=True,
             use_depthwise=True))
@@ -250,7 +250,7 @@ class ConvolutionalBoxPredictorTest(test_case.TestCase):
             max_depth=32,
             num_layers_before_predictor=1,
             dropout_keep_prob=0.8,
-            kernel_size=1,
+            kernel_size=3,
             box_code_size=4,
             use_dropout=True,
             use_depthwise=True))

research/object_detection/predictors/convolutional_keras_box_predictor_test.py

@@ -209,7 +209,7 @@ class ConvolutionalKerasBoxPredictorTest(test_case.TestCase):
             num_layers_before_predictor=1,
             use_dropout=True,
             dropout_keep_prob=0.8,
-            kernel_size=1,
+            kernel_size=3,
             box_code_size=4,
             use_depthwise=True
         ))

research/object_detection/predictors/heads/box_head.py

@@ -21,10 +21,11 @@ All the box prediction heads have a predict function that receives the
 """
 import functools
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.predictors.heads import head
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class MaskRCNNBoxHead(head.Head):
@@ -137,7 +138,11 @@ class ConvolutionalBoxHead(head.Head):
 
     Raises:
       ValueError: if min_depth > max_depth.
+      ValueError: if use_depthwise is True and kernel_size is 1.
     """
+    if use_depthwise and (kernel_size == 1):
+      raise ValueError('Should not use 1x1 kernel when using depthwise conv')
+
     super(ConvolutionalBoxHead, self).__init__()
     self._is_training = is_training
     self._box_code_size = box_code_size
@@ -221,7 +226,13 @@ class WeightSharedConvolutionalBoxHead(head.Head):
         box_coder]. Otherwise returns the prediction tensor before reshaping,
         whose shape is [batch, height, width, num_predictions_per_location *
         num_class_slots].
+
+    Raises:
+      ValueError: if use_depthwise is True and kernel_size is 1.
     """
+    if use_depthwise and (kernel_size == 1):
+      raise ValueError('Should not use 1x1 kernel when using depthwise conv')
+
     super(WeightSharedConvolutionalBoxHead, self).__init__()
     self._box_code_size = box_code_size
     self._kernel_size = kernel_size

research/object_detection/predictors/heads/class_head.py

@@ -21,10 +21,11 @@ All the class prediction heads have a predict function that receives the
 """
 import functools
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.predictors.heads import head
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class MaskRCNNClassHead(head.Head):
@@ -143,7 +144,11 @@ class ConvolutionalClassHead(head.Head):
 
     Raises:
       ValueError: if min_depth > max_depth.
+      ValueError: if use_depthwise is True and kernel_size is 1.
     """
+    if use_depthwise and (kernel_size == 1):
+      raise ValueError('Should not use 1x1 kernel when using depthwise conv')
+
     super(ConvolutionalClassHead, self).__init__()
     self._is_training = is_training
     self._num_class_slots = num_class_slots
@@ -247,7 +252,13 @@ class WeightSharedConvolutionalClassHead(head.Head):
         whose shape is [batch, height, width, num_predictions_per_location *
         num_class_slots].
       scope: Scope name for the convolution operation.
+
+    Raises:
+      ValueError: if use_depthwise is True and kernel_size is 1.
     """
+    if use_depthwise and (kernel_size == 1):
+      raise ValueError('Should not use 1x1 kernel when using depthwise conv')
+
     super(WeightSharedConvolutionalClassHead, self).__init__()
     self._num_class_slots = num_class_slots
     self._kernel_size = kernel_size
@@ -303,4 +314,3 @@ class WeightSharedConvolutionalClassHead(head.Head):
           class_predictions_with_background,
           [batch_size, -1, self._num_class_slots])
     return class_predictions_with_background
-

research/object_detection/predictors/heads/keras_box_head.py

@@ -62,7 +62,11 @@ class ConvolutionalBoxHead(head.KerasHead):
 
     Raises:
       ValueError: if min_depth > max_depth.
+      ValueError: if use_depthwise is True and kernel_size is 1.
     """
+    if use_depthwise and (kernel_size == 1):
+      raise ValueError('Should not use 1x1 kernel when using depthwise conv')
+
     super(ConvolutionalBoxHead, self).__init__(name=name)
     self._is_training = is_training
     self._box_code_size = box_code_size
@@ -266,7 +270,13 @@ class WeightSharedConvolutionalBoxHead(head.KerasHead):
         num_class_slots].
       name: A string name scope to assign to the model. If `None`, Keras
         will auto-generate one from the class name.
+
+    Raises:
+      ValueError: if use_depthwise is True and kernel_size is 1.
     """
+    if use_depthwise and (kernel_size == 1):
+      raise ValueError('Should not use 1x1 kernel when using depthwise conv')
+
     super(WeightSharedConvolutionalBoxHead, self).__init__(name=name)
     self._box_code_size = box_code_size
     self._kernel_size = kernel_size

research/object_detection/predictors/heads/keras_class_head.py

@@ -71,7 +71,11 @@ class ConvolutionalClassHead(head.KerasHead):
 
     Raises:
       ValueError: if min_depth > max_depth.
+      ValueError: if use_depthwise is True and kernel_size is 1.
     """
+    if use_depthwise and (kernel_size == 1):
+      raise ValueError('Should not use 1x1 kernel when using depthwise conv')
+
     super(ConvolutionalClassHead, self).__init__(name=name)
     self._is_training = is_training
     self._use_dropout = use_dropout
@@ -274,7 +278,13 @@ class WeightSharedConvolutionalClassHead(head.KerasHead):
         num_class_slots].
       name: A string name scope to assign to the model. If `None`, Keras
         will auto-generate one from the class name.
+
+    Raises:
+      ValueError: if use_depthwise is True and kernel_size is 1.
     """
+    if use_depthwise and (kernel_size == 1):
+      raise ValueError('Should not use 1x1 kernel when using depthwise conv')
+
     super(WeightSharedConvolutionalClassHead, self).__init__(name=name)
     self._num_class_slots = num_class_slots
     self._kernel_size = kernel_size

research/object_detection/predictors/heads/keypoint_head.py

@@ -23,9 +23,10 @@ Mask RCNN paper. Or they could be used to represent different part locations of
 objects.
 """
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.predictors.heads import head
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class MaskRCNNKeypointHead(head.Head):