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

--
308640516  by Sergio Guadarrama:

    Internal change

308244396  by Sergio Guadarrama:

    GroupNormalization support for MobilenetV3.

--
307475800  by Sergio Guadarrama:

    Internal change

--
302077708  by Sergio Guadarrama:

    Remove `disable_tf2` behavior from slim py_library targets

--
301208453  by Sergio Guadarrama:

    Automated refactoring to make code Python 3 compatible.

--
300816672  by Sergio Guadarrama:

    Internal change

299433840  by Sergio Guadarrama:

    Internal change

299221609  by Sergio Guadarrama:

    Explicitly disable Tensorflow v2 behaviors for all TF1.x binaries and tests

--
299179617  by Sergio Guadarrama:

    Internal change

299040784  by Sergio Guadarrama:

    Internal change

299036699  by Sergio Guadarrama:

    Internal change

298736510  by Sergio Guadarrama:

    Internal change

298732599  by Sergio Guadarrama:

    Internal change

298729507  by Sergio Guadarrama:

    Internal change

298253328  by Sergio Guadarrama:

    Internal change

297788346  by Sergio Guadarrama:

    Internal change

297785278  by Sergio Guadarrama:

    Internal change

297783127  by Sergio Guadarrama:

    Internal change

297725870  by Sergio Guadarrama:

    Internal change

297721811  by Sergio Guadarrama:

    Internal change

297711347  by Sergio Guadarrama:

    Internal change

297708059  by Sergio Guadarrama:

    Internal change

297701831  by Sergio Guadarrama:

    Internal change

297700038  by Sergio Guadarrama:

    Internal change

297670468  by Sergio Guadarrama:

    Internal change.

--
297350326  by Sergio Guadarrama:

    Explicitly replace "import tensorflow" with "tensorflow.compat.v1" for TF2.x migration

--
297201668  by Sergio Guadarrama:

    Explicitly replace "import tensorflow" with "tensorflow.compat.v1" for TF2.x migration

--
294483372  by Sergio Guadarrama:

    Internal change

PiperOrigin-RevId: 311933687

* Merged commit includes the following changes:
312578615  by Menglong Zhu:

    Modify the LSTM feature extractors to be python 3 compatible.

--
311264357  by Menglong Zhu:

    Removes contrib.slim

--
308957207  by Menglong Zhu:

    Automated refactoring to make code Python 3 compatible.

--
306976470  by yongzhe:

    Internal change

306777559  by Menglong Zhu:

    Internal change

--
299232507  by lzyuan:

    Internal update.

--
299221735  by lzyuan:

    Add small epsilon on max_range for quantize_op to prevent range collapse.

--

PiperOrigin-RevId: 312578615

* Merged commit includes the following changes:
310447280  by lzc:

    Internal changes.

--

PiperOrigin-RevId: 310447280

Co-authored-by: Sergio Guadarrama <[email protected]>
Co-authored-by: Menglong Zhu <[email protected]>

Author: 3 people

research/lstm_object_detection/README.md

@@ -1,6 +1,3 @@
-![TensorFlow Requirement: 1.x](https://img.shields.io/badge/TensorFlow%20Requirement-1.x-brightgreen)
-![TensorFlow 2 Not Supported](https://img.shields.io/badge/TensorFlow%202%20Not%20Supported-%E2%9C%95-red.svg)
-
 # Tensorflow Mobile Video Object Detection
 
 Tensorflow mobile video object detection implementation proposed in the
@@ -35,6 +32,7 @@ https://scholar.googleusercontent.com/scholar.bib?q=info:rLqvkztmWYgJ:scholar.go
 * [email protected]
 * [email protected]
 * [email protected]
+* [email protected]
 
 
 ## Table of Contents

research/lstm_object_detection/inputs/seq_dataset_builder.py

@@ -23,7 +23,8 @@
 that wraps the build function.
 """
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim as contrib_slim
+import tf_slim as slim
+
 from tensorflow.contrib.training.python.training import sequence_queueing_state_saver as sqss
 from lstm_object_detection.inputs import tf_sequence_example_decoder
 from lstm_object_detection.protos import input_reader_google_pb2
@@ -33,7 +34,7 @@
 from object_detection.protos import input_reader_pb2
 from object_detection.utils import ops as util_ops
 
-parallel_reader = contrib_slim.parallel_reader
+parallel_reader = slim.parallel_reader
 # TODO(yinxiao): Make the following variable into configurable proto.
 # Padding size for the labeled objects in each frame. Here we assume each
 # frame has a total number of objects less than _PADDING_SIZE.

research/lstm_object_detection/inputs/tf_sequence_example_decoder.py

@@ -19,11 +19,11 @@
 tensorflow.SequenceExample protos.
 """
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim as contrib_slim
+import tf_slim as slim
 from object_detection.core import data_decoder
 from object_detection.core import standard_fields as fields
 
-tfexample_decoder = contrib_slim.tfexample_decoder
+tfexample_decoder = slim.tfexample_decoder
 
 
 class BoundingBoxSequence(tfexample_decoder.ItemHandler):

research/lstm_object_detection/lstm/lstm_cells.py

@@ -15,10 +15,8 @@
 """BottleneckConvLSTMCell implementation."""
 
 import tensorflow.compat.v1 as tf
-
-from tensorflow.contrib import layers as contrib_layers
+import tf_slim as slim
 from tensorflow.contrib import rnn as contrib_rnn
-from tensorflow.contrib import slim
 from tensorflow.contrib.framework.python.ops import variables as contrib_variables
 import lstm_object_detection.lstm.utils as lstm_utils
 
@@ -121,7 +119,7 @@ def __call__(self, inputs, state, scope=None):
       if self._pre_bottleneck:
         bottleneck = inputs
       else:
-        bottleneck = contrib_layers.separable_conv2d(
+        bottleneck = slim.separable_conv2d(
             tf.concat([inputs, h], 3),
             self._num_units,
             self._filter_size,
@@ -133,7 +131,7 @@ def __call__(self, inputs, state, scope=None):
         if self._viz_gates:
           slim.summaries.add_histogram_summary(bottleneck, 'bottleneck')
 
-      concat = contrib_layers.separable_conv2d(
+      concat = slim.separable_conv2d(
           bottleneck,
           4 * self._num_units,
           self._filter_size,
@@ -243,7 +241,7 @@ def pre_bottleneck(self, inputs, state, input_index):
       state = tf.reshape(state, [batch_size, height, width, -1])
     with tf.variable_scope('conv_lstm_cell', reuse=tf.AUTO_REUSE):
       scope_name = 'bottleneck_%d' % input_index
-      inputs = contrib_layers.separable_conv2d(
+      inputs = slim.separable_conv2d(
           tf.concat([inputs, state], 3),
           self.output_size[-1],
           self._filter_size,

research/lstm_object_detection/lstm/utils.py

@@ -217,7 +217,8 @@ def quantize_op(inputs,
     # While training, collect EMAs of ranges seen, store in min_var, max_var.
     # TFLite requires that 0.0 is always in the [min; max] range.
     range_min = tf.minimum(tf.reduce_min(inputs), 0.0, 'SafeQuantRangeMin')
-    range_max = tf.maximum(tf.reduce_max(inputs), 0.0, 'SafeQuantRangeMax')
+    # We set the lower_bound of max_range to prevent range collapse.
+    range_max = tf.maximum(tf.reduce_max(inputs), 1e-5, 'SafeQuantRangeMax')
     min_val = moving_averages.assign_moving_average(
         min_var, range_min, ema_decay, name='AssignMinEma')
     max_val = moving_averages.assign_moving_average(

research/lstm_object_detection/meta_architectures/lstm_ssd_meta_arch.py

@@ -25,7 +25,6 @@
 import abc
 import re
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim as contrib_slim
 
 from object_detection.core import box_list_ops
 from object_detection.core import matcher
@@ -34,8 +33,6 @@
 from object_detection.utils import ops
 from object_detection.utils import shape_utils
 
-slim = contrib_slim
-
 
 class LSTMSSDMetaArch(ssd_meta_arch.SSDMetaArch):
   """LSTM Meta-architecture definition."""

research/lstm_object_detection/meta_architectures/lstm_ssd_meta_arch_test.py

@@ -23,7 +23,7 @@
 
 import numpy as np
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim as contrib_slim
+import tf_slim as slim
 
 from lstm_object_detection.lstm import lstm_cells
 from lstm_object_detection.meta_architectures import lstm_ssd_meta_arch
@@ -39,8 +39,6 @@
 from object_detection.utils import test_utils
 
 
-slim = contrib_slim
-
 MAX_TOTAL_NUM_BOXES = 5
 NUM_CLASSES = 1
 

research/lstm_object_detection/models/lstm_ssd_interleaved_mobilenet_v2_feature_extractor.py

@@ -16,7 +16,7 @@
 """LSTDInterleavedFeatureExtractor which interleaves multiple MobileNet V2."""
 
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim
+import tf_slim as slim
 
 from tensorflow.python.framework import ops as tf_ops
 from lstm_object_detection.lstm import lstm_cells
@@ -134,9 +134,10 @@ def extract_base_features_small(self, preprocessed_inputs):
     scope_name = self._base_network_scope + '_2'
     with tf.variable_scope(scope_name, reuse=self._reuse_weights) as base_scope:
       if self._low_res:
-        size_small = preprocessed_inputs.get_shape().as_list()[1] / 2
+        height_small = preprocessed_inputs.get_shape().as_list()[1] // 2
+        width_small = preprocessed_inputs.get_shape().as_list()[2] // 2
         inputs_small = tf.image.resize_images(preprocessed_inputs,
-                                              [size_small, size_small])
+                                              [height_small, width_small])
         # Create end point handle for tflite deployment.
         with tf.name_scope(None):
           inputs_small = tf.identity(
@@ -152,14 +153,17 @@ def extract_base_features_small(self, preprocessed_inputs):
           scope=base_scope)
       return net, end_points
 
-  def create_lstm_cell(self, batch_size, output_size, state_saver, state_name):
+  def create_lstm_cell(self, batch_size, output_size, state_saver, state_name,
+                       dtype=tf.float32):
     """Create the LSTM cell, and initialize state if necessary.
 
     Args:
       batch_size: input batch size.
       output_size: output size of the lstm cell, [width, height].
       state_saver: a state saver object with methods `state` and `save_state`.
       state_name: string, the name to use with the state_saver.
+      dtype: dtype to initialize lstm state.
+
     Returns:
       lstm_cell: the lstm cell unit.
       init_state: initial state representations.
@@ -180,7 +184,7 @@ def create_lstm_cell(self, batch_size, output_size, state_saver, state_name):
         visualize_gates=False)
 
     if state_saver is None:
-      init_state = lstm_cell.init_state('lstm_state', batch_size, tf.float32)
+      init_state = lstm_cell.init_state('lstm_state', batch_size, dtype)
       step = None
     else:
       step = state_saver.state(state_name + '_step')
@@ -222,7 +226,7 @@ def extract_features(self, preprocessed_inputs, state_saver=None,
         33, preprocessed_inputs)
     preprocessed_inputs = ops.pad_to_multiple(
         preprocessed_inputs, self._pad_to_multiple)
-    batch_size = preprocessed_inputs.shape[0].value / unroll_length
+    batch_size = preprocessed_inputs.shape[0].value // unroll_length
     batch_axis = 0
     nets = []
 
@@ -250,7 +254,8 @@ def extract_features(self, preprocessed_inputs, state_saver=None,
       with tf.variable_scope('LSTM', reuse=self._reuse_weights):
         output_size = (large_base_feature_shape[1], large_base_feature_shape[2])
         lstm_cell, init_state, step = self.create_lstm_cell(
-            batch_size, output_size, state_saver, state_name)
+            batch_size, output_size, state_saver, state_name,
+            dtype=preprocessed_inputs.dtype)
 
         nets_seq = [
             tf.split(net, unroll_length, axis=batch_axis) for net in nets
@@ -269,15 +274,16 @@ def extract_features(self, preprocessed_inputs, state_saver=None,
             scope=None)
         self._states_out = states_out
 
-      batcher_ops = None
+      image_features = {}
       if state_saver is not None:
         self._step = state_saver.state(state_name + '_step')
         batcher_ops = [
             state_saver.save_state(state_name + '_c', states_out[-1][0]),
             state_saver.save_state(state_name + '_h', states_out[-1][1]),
             state_saver.save_state(state_name + '_step', self._step + 1)]
-      image_features = {}
-      with tf_ops.control_dependencies(batcher_ops):
+        with tf_ops.control_dependencies(batcher_ops):
+          image_features['layer_19'] = tf.concat(net_seq, 0)
+      else:
         image_features['layer_19'] = tf.concat(net_seq, 0)
 
       # SSD layers.
@@ -289,4 +295,4 @@ def extract_features(self, preprocessed_inputs, state_saver=None,
             insert_1x1_conv=True,
             image_features=image_features,
             pool_residual=True)
-    return feature_maps.values()
+    return list(feature_maps.values())

research/lstm_object_detection/models/lstm_ssd_interleaved_mobilenet_v2_feature_extractor_test.py

@@ -15,10 +15,9 @@
 
 """Tests for lstm_ssd_interleaved_mobilenet_v2_feature_extractor."""
 
-import itertools
 import numpy as np
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim
+import tf_slim as slim
 from tensorflow.contrib import training as contrib_training
 
 from lstm_object_detection.models import lstm_ssd_interleaved_mobilenet_v2_feature_extractor
@@ -261,16 +260,16 @@ def test_lstm_states(self):
     state_channel = 320
     init_state1 = {
         'lstm_state_c': tf.zeros(
-            [image_height/32, image_width/32, state_channel]),
+            [image_height // 32, image_width // 32, state_channel]),
         'lstm_state_h': tf.zeros(
-            [image_height/32, image_width/32, state_channel]),
+            [image_height // 32, image_width // 32, state_channel]),
         'lstm_state_step': tf.zeros([1])
     }
     init_state2 = {
         'lstm_state_c': tf.random_uniform(
-            [image_height/32, image_width/32, state_channel]),
+            [image_height // 32, image_width // 32, state_channel]),
         'lstm_state_h': tf.random_uniform(
-            [image_height/32, image_width/32, state_channel]),
+            [image_height // 32, image_width // 32, state_channel]),
         'lstm_state_step': tf.zeros([1])
     }
     seq = {'dummy': tf.random_uniform([2, 1, 1, 1])}
@@ -326,7 +325,7 @@ def graph_fn(image_tensor):
     image_tensor = np.random.rand(batch_size, image_height, image_width,
                                   3).astype(np.float32)
     feature_maps = self.execute(graph_fn, [image_tensor])
-    for feature_map, expected_shape in itertools.izip(
+    for feature_map, expected_shape in zip(
         feature_maps, expected_feature_map_shapes):
       self.assertAllEqual(feature_map.shape, expected_shape)
 

research/lstm_object_detection/models/lstm_ssd_mobilenet_v1_feature_extractor.py

@@ -16,7 +16,7 @@
 """LSTMSSDFeatureExtractor for MobilenetV1 features."""
 
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim as contrib_slim
+import tf_slim as slim
 from tensorflow.python.framework import ops as tf_ops
 from lstm_object_detection.lstm import lstm_cells
 from lstm_object_detection.lstm import rnn_decoder
@@ -27,8 +27,6 @@
 from object_detection.utils import shape_utils
 from nets import mobilenet_v1
 
-slim = contrib_slim
-
 
 class LSTMSSDMobileNetV1FeatureExtractor(
     lstm_ssd_meta_arch.LSTMSSDFeatureExtractor):
@@ -85,14 +83,16 @@ def __init__(self,
     self._base_network_scope = 'MobilenetV1'
     self._lstm_state_depth = lstm_state_depth
 
-  def create_lstm_cell(self, batch_size, output_size, state_saver, state_name):
+  def create_lstm_cell(self, batch_size, output_size, state_saver, state_name,
+                       dtype=tf.float32):
     """Create the LSTM cell, and initialize state if necessary.
 
     Args:
       batch_size: input batch size.
       output_size: output size of the lstm cell, [width, height].
       state_saver: a state saver object with methods `state` and `save_state`.
       state_name: string, the name to use with the state_saver.
+      dtype: dtype to initialize lstm state.
 
     Returns:
       lstm_cell: the lstm cell unit.
@@ -107,7 +107,7 @@ def create_lstm_cell(self, batch_size, output_size, state_saver, state_name):
         visualize_gates=False)
 
     if state_saver is None:
-      init_state = lstm_cell.init_state(state_name, batch_size, tf.float32)
+      init_state = lstm_cell.init_state(state_name, batch_size, dtype)
       step = None
     else:
       step = state_saver.state(state_name + '_step')
@@ -166,11 +166,14 @@ def extract_features(self,
       with slim.arg_scope(
           [slim.batch_norm], fused=False, is_training=self._is_training):
         # ConvLSTM layers.
-        batch_size = net.shape[0].value / unroll_length
+        batch_size = net.shape[0].value // unroll_length
         with tf.variable_scope('LSTM', reuse=self._reuse_weights) as lstm_scope:
           lstm_cell, init_state, _ = self.create_lstm_cell(
-              batch_size, (net.shape[1].value, net.shape[2].value), state_saver,
-              state_name)
+              batch_size,
+              (net.shape[1].value, net.shape[2].value),
+              state_saver,
+              state_name,
+              dtype=preprocessed_inputs.dtype)
           net_seq = list(tf.split(net, unroll_length))
 
           # Identities added for inputing state tensors externally.
@@ -205,4 +208,4 @@ def extract_features(self,
               insert_1x1_conv=True,
               image_features=image_features)
 
-    return feature_maps.values()
+    return list(feature_maps.values())

research/lstm_object_detection/models/lstm_ssd_mobilenet_v1_feature_extractor_test.py

@@ -17,14 +17,12 @@
 
 import numpy as np
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim as contrib_slim
+import tf_slim as slim
 from tensorflow.contrib import training as contrib_training
 
 from lstm_object_detection.models import lstm_ssd_mobilenet_v1_feature_extractor as feature_extractor
 from object_detection.models import ssd_feature_extractor_test
 
-slim = contrib_slim
-
 
 class LstmSsdMobilenetV1FeatureExtractorTest(
     ssd_feature_extractor_test.SsdFeatureExtractorTestBase):

research/lstm_object_detection/models/mobilenet_defs.py

@@ -15,14 +15,11 @@
 """Definitions for modified MobileNet models used in LSTD."""
 
 import tensorflow.compat.v1 as tf
-from tensorflow.contrib import slim as contrib_slim
-
+import tf_slim as slim
 from nets import mobilenet_v1
 from nets.mobilenet import conv_blocks as mobilenet_convs
 from nets.mobilenet import mobilenet
 
-slim = contrib_slim
-
 
 def mobilenet_v1_lite_def(depth_multiplier, low_res=False):
   """Conv definitions for a lite MobileNet v1 model.