Merge remote-tracking branch 'upstream/master' into context_tf2

research/attention_ocr/python/model_export_lib.py

@@ -36,7 +36,7 @@ def normalize_image(image, original_minval, original_maxval, target_minval,
   Returns:
     image: image which is the same shape as input image.
   """
-  with tf.name_scope('NormalizeImage', values=[image]):
+  with tf.compat.v1.name_scope('NormalizeImage', values=[image]):
     original_minval = float(original_minval)
     original_maxval = float(original_maxval)
     target_minval = float(target_minval)
@@ -68,16 +68,17 @@ def generate_tfexample_image(input_example_strings,
     A tensor with shape [batch_size, height, width, channels] of type float32
     with values in the range [0..1]
   """
-  batch_size = tf.shape(input_example_strings)[0]
+  batch_size = tf.shape(input=input_example_strings)[0]
   images_shape = tf.stack(
       [batch_size, image_height, image_width, image_channels])
   tf_example_image_key = 'image/encoded'
   feature_configs = {
       tf_example_image_key:
-          tf.FixedLenFeature(
+          tf.io.FixedLenFeature(
               image_height * image_width * image_channels, dtype=tf.float32)
   }
-  feature_tensors = tf.parse_example(input_example_strings, feature_configs)
+  feature_tensors = tf.io.parse_example(
+      serialized=input_example_strings, features=feature_configs)
   float_images = tf.reshape(
       normalize_image(
           feature_tensors[tf_example_image_key],
@@ -97,11 +98,11 @@ def attention_ocr_attention_masks(num_characters):
   names = ['%s/Softmax:0' % (prefix)]
   for i in range(1, num_characters):
     names += ['%s_%d/Softmax:0' % (prefix, i)]
-  return [tf.get_default_graph().get_tensor_by_name(n) for n in names]
+  return [tf.compat.v1.get_default_graph().get_tensor_by_name(n) for n in names]
 
 
 def build_tensor_info(tensor_dict):
   return {
-      k: tf.saved_model.utils.build_tensor_info(t)
+      k: tf.compat.v1.saved_model.utils.build_tensor_info(t)
       for k, t in tensor_dict.items()
   }

research/attention_ocr/python/model_export_test.py

@@ -29,7 +29,7 @@ _CHECKPOINT_URL = (
 
 
 def _clean_up():
-  tf.gfile.DeleteRecursively(tf.test.get_temp_dir())
+  tf.io.gfile.rmtree(tf.compat.v1.test.get_temp_dir())
 
 
 def _create_tf_example_string(image):
@@ -47,7 +47,7 @@ class AttentionOcrExportTest(tf.test.TestCase):
     for suffix in ['.meta', '.index', '.data-00000-of-00001']:
       filename = _CHECKPOINT + suffix
       self.assertTrue(
-          tf.gfile.Exists(filename),
+          tf.io.gfile.exists(filename),
           msg='Missing checkpoint file %s. '
           'Please download and extract it from %s' %
           (filename, _CHECKPOINT_URL))
@@ -57,7 +57,8 @@ class AttentionOcrExportTest(tf.test.TestCase):
         os.path.dirname(__file__), 'datasets/testdata/fsns')
     tf.test.TestCase.setUp(self)
     _clean_up()
-    self.export_dir = os.path.join(tf.test.get_temp_dir(), 'exported_model')
+    self.export_dir = os.path.join(
+        tf.compat.v1.test.get_temp_dir(), 'exported_model')
     self.minimal_output_signature = {
         'predictions': 'AttentionOcr_v1/predicted_chars:0',
         'scores': 'AttentionOcr_v1/predicted_scores:0',
@@ -93,10 +94,10 @@ class AttentionOcrExportTest(tf.test.TestCase):
                               size=self.dataset.image_shape).astype('uint8'),
     }
     signature_def = graph_def.signature_def[
-        tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY]
+        tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY]
     if serving:
       input_name = signature_def.inputs[
-          tf.saved_model.signature_constants.CLASSIFY_INPUTS].name
+          tf.saved_model.CLASSIFY_INPUTS].name
       # Model for serving takes input: inputs['inputs'] = 'tf_example:0'
       feed_dict = {
           input_name: [
@@ -126,11 +127,11 @@ class AttentionOcrExportTest(tf.test.TestCase):
       export_for_serving: True if the model was exported for Serving. This
         affects how input is fed into the model.
     """
-    tf.reset_default_graph()
-    sess = tf.Session()
-    graph_def = tf.saved_model.loader.load(
+    tf.compat.v1.reset_default_graph()
+    sess = tf.compat.v1.Session()
+    graph_def = tf.compat.v1.saved_model.loader.load(
         sess=sess,
-        tags=[tf.saved_model.tag_constants.SERVING],
+        tags=[tf.saved_model.SERVING],
         export_dir=self.export_dir)
     feed_dict = self.create_input_feed(graph_def, export_for_serving)
     results = sess.run(self.minimal_output_signature, feed_dict=feed_dict)

research/attention_ocr/python/model_test.py

@@ -52,7 +52,7 @@ class ModelTest(tf.test.TestCase):
                               self.num_char_classes)
     self.length_logit_shape = (self.batch_size, self.seq_length + 1)
     # Placeholder knows image dimensions, but not batch size.
-    self.input_images = tf.placeholder(
+    self.input_images = tf.compat.v1.placeholder(
         tf.float32,
         shape=(None, self.image_height, self.image_width, 3),
         name='input_node')
@@ -89,8 +89,8 @@ class ModelTest(tf.test.TestCase):
     with self.test_session() as sess:
       endpoints_tf = ocr_model.create_base(
           images=self.input_images, labels_one_hot=None)
-      sess.run(tf.global_variables_initializer())
-      tf.tables_initializer().run()
+      sess.run(tf.compat.v1.global_variables_initializer())
+      tf.compat.v1.tables_initializer().run()
       endpoints = sess.run(
           endpoints_tf, feed_dict={self.input_images: self.fake_images})
 
@@ -127,7 +127,7 @@ class ModelTest(tf.test.TestCase):
       ocr_model = self.create_model()
       conv_tower = ocr_model.conv_tower_fn(self.input_images)
 
-      sess.run(tf.global_variables_initializer())
+      sess.run(tf.compat.v1.global_variables_initializer())
       conv_tower_np = sess.run(
           conv_tower, feed_dict={self.input_images: self.fake_images})
 
@@ -141,9 +141,9 @@ class ModelTest(tf.test.TestCase):
     ocr_model = self.create_model()
     ocr_model.create_base(images=self.input_images, labels_one_hot=None)
     with self.test_session() as sess:
-      tfprof_root = tf.profiler.profile(
+      tfprof_root = tf.compat.v1.profiler.profile(
           sess.graph,
-          options=tf.profiler.ProfileOptionBuilder
+          options=tf.compat.v1.profiler.ProfileOptionBuilder
           .trainable_variables_parameter())
 
       model_size_bytes = 4 * tfprof_root.total_parameters
@@ -163,9 +163,9 @@ class ModelTest(tf.test.TestCase):
     summaries = ocr_model.create_summaries(
         data, endpoints, charset, is_training=False)
     with self.test_session() as sess:
-      sess.run(tf.global_variables_initializer())
-      sess.run(tf.local_variables_initializer())
-      tf.tables_initializer().run()
+      sess.run(tf.compat.v1.global_variables_initializer())
+      sess.run(tf.compat.v1.local_variables_initializer())
+      tf.compat.v1.tables_initializer().run()
       sess.run(summaries)  # just check it is runnable
 
   def test_sequence_loss_function_without_label_smoothing(self):
@@ -188,7 +188,7 @@ class ModelTest(tf.test.TestCase):
     Returns:
       a list of tensors with encoded image coordinates in them.
     """
-    batch_size = tf.shape(net)[0]
+    batch_size = tf.shape(input=net)[0]
     _, h, w, _ = net.shape.as_list()
     h_loc = [
         tf.tile(
@@ -200,7 +200,8 @@ class ModelTest(tf.test.TestCase):
     h_loc = tf.concat([tf.expand_dims(t, 2) for t in h_loc], 2)
     w_loc = [
         tf.tile(
-            tf.contrib.layers.one_hot_encoding(tf.constant([i]), num_classes=w),
+            tf.contrib.layers.one_hot_encoding(
+                tf.constant([i]), num_classes=w),
             [h, 1]) for i in range(w)
     ]
     w_loc = tf.concat([tf.expand_dims(t, 2) for t in w_loc], 2)
@@ -272,8 +273,8 @@ class ModelTest(tf.test.TestCase):
       endpoints_tf = ocr_model.create_base(
           images=self.fake_images, labels_one_hot=None)
 
-      sess.run(tf.global_variables_initializer())
-      tf.tables_initializer().run()
+      sess.run(tf.compat.v1.global_variables_initializer())
+      tf.compat.v1.tables_initializer().run()
       endpoints = sess.run(endpoints_tf)
 
       self.assertEqual(endpoints.predicted_text.shape, (self.batch_size,))
@@ -289,7 +290,7 @@ class CharsetMapperTest(tf.test.TestCase):
     charset_mapper = model.CharsetMapper(charset)
 
     with self.test_session() as sess:
-      tf.tables_initializer().run()
+      tf.compat.v1.tables_initializer().run()
       text = sess.run(charset_mapper.get_text(ids))
 
     self.assertAllEqual(text, [b'hello', b'world'])

research/attention_ocr/python/sequence_layers_test.py

@@ -29,13 +29,13 @@ import sequence_layers
 
 def fake_net(batch_size, num_features, feature_size):
   return tf.convert_to_tensor(
-      np.random.uniform(size=(batch_size, num_features, feature_size)),
+      value=np.random.uniform(size=(batch_size, num_features, feature_size)),
       dtype=tf.float32)
 
 
 def fake_labels(batch_size, seq_length, num_char_classes):
   labels_np = tf.convert_to_tensor(
-      np.random.randint(
+      value=np.random.randint(
           low=0, high=num_char_classes, size=(batch_size, seq_length)))
   return slim.one_hot_encoding(labels_np, num_classes=num_char_classes)
 

research/object_detection/README.md

@@ -109,6 +109,18 @@ Sergi Caelles Prat, Shan Yang, Sudheendra Vijayanarasimhan, Tina Tian, Tomer
 Kaftan, Vighnesh Birodkar, Vishnu Banna, Vivek Rathod, Yanhui Liang, Yiming Shi,
 Yixin Shi, Yu-hui Chen, Zhichao Lu.
 
+### MobileDet GPU
+
+We have released SSDLite with MobileDet GPU backbone, which achieves 17% mAP
+higher than the MobileNetV2 SSDLite (27.5 mAP vs 23.5 mAP) on a NVIDIA Jetson
+Xavier at comparable latency (3.2ms vs 3.3ms).
+
+Along with the model definition, we are also releasing model checkpoints trained
+on the COCO dataset.
+
+<b>Thanks to contributors</b>: Yongzhe Wang, Bo Chen, Hanxiao Liu, Le An
+(NVIDIA), Yu-Te Cheng (NVIDIA), Oliver Knieps (NVIDIA), and Josh Park (NVIDIA).
+
 ### Context R-CNN
 
 We have released [Context R-CNN](https://arxiv.org/abs/1912.03538), a model that

research/object_detection/meta_architectures/context_rcnn_meta_arch.py

@@ -332,7 +332,7 @@ class ContextRCNNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
     """
 
     box_features = self._crop_and_resize_fn(
-        features_to_crop, proposal_boxes_normalized,
+        [features_to_crop], proposal_boxes_normalized, None,
         [self._initial_crop_size, self._initial_crop_size])
 
     attention_features = self._context_feature_extract_fn(