Merged commit includes the following changes: (#6932)

250447559  by Zhichao Lu:

    Update expected files format for Instance Segmentation challenge:
    - add fields ImageWidth, ImageHeight and store the values per prediction
    - as mask, store only encoded image and assume its size is ImageWidth x ImageHeight

--
250402780  by rathodv:

    Fix failing Mask R-CNN TPU convergence test.

    Cast second stage prediction tensors from bfloat16 to float32 to prevent errors in third target assignment (Mask Prediction) - Concat with different types bfloat16 and bfloat32 isn't allowed.

--
250300240  by Zhichao Lu:

    Addion Open Images Challenge 2019 object detection and instance segmentation
    support into Estimator framework.

--
249944839  by rathodv:

    Modify exporter.py to add multiclass score nodes in exported inference graphs.

--
249935201  by rathodv:

    Modify postprocess methods to preserve multiclass scores after non max suppression.

--
249878079  by Zhich...

research/object_detection/protos/post_processing.proto

@@ -22,6 +22,20 @@ message BatchNonMaxSuppression {
 
   // Whether to use the implementation of NMS that guarantees static shapes.
   optional bool use_static_shapes = 6 [default = false];
+
+  // Whether to use class agnostic NMS.
+  // Class-agnostic NMS function implements a class-agnostic version
+  // of Non Maximal Suppression where if max_classes_per_detection=k,
+  // 1) we keep the top-k scores for each detection and
+  // 2) during NMS, each detection only uses the highest class score for sorting.
+  // 3) Compared to regular NMS, the worst runtime of this version is O(N^2)
+  // instead of O(KN^2) where N is the number of detections and K the number of
+  // classes.
+  optional bool use_class_agnostic_nms = 7 [default = false];
+
+  // Number of classes retained per detection in class agnostic NMS.
+
+  optional int32 max_classes_per_detection = 8 [default = 1];
 }
 
 // Configuration proto for post-processing predicted boxes and

research/object_detection/tpu_exporters/faster_rcnn.py

@@ -87,7 +87,7 @@ def get_prediction_tensor_shapes(pipeline_config):
 
   _, input_tensors = exporter.input_placeholder_fn_map['image_tensor']()
 
-  inputs = tf.to_float(input_tensors)
+  inputs = tf.cast(input_tensors, dtype=tf.float32)
   preprocessed_inputs, true_image_shapes = detection_model.preprocess(inputs)
 
   prediction_dict = detection_model.predict(preprocessed_inputs,
@@ -125,7 +125,7 @@ def build_graph(pipeline_config,
       exporter.input_placeholder_fn_map[input_type]()
 
   # CPU pre-processing
-  inputs = tf.to_float(input_tensors)
+  inputs = tf.cast(input_tensors, dtype=tf.float32)
   preprocessed_inputs, true_image_shapes = detection_model.preprocess(inputs)
 
   # Dimshuffle: [b, h, w, c] -> [b, c, h, w]

research/object_detection/tpu_exporters/ssd.py

@@ -57,7 +57,7 @@ def get_prediction_tensor_shapes(pipeline_config):
   detection_model = model_builder.build(
       pipeline_config.model, is_training=False)
   _, input_tensors = exporter.input_placeholder_fn_map['image_tensor']()
-  inputs = tf.to_float(input_tensors)
+  inputs = tf.cast(input_tensors, dtype=tf.float32)
   preprocessed_inputs, true_image_shapes = detection_model.preprocess(inputs)
   prediction_dict = detection_model.predict(preprocessed_inputs,
                                             true_image_shapes)
@@ -138,7 +138,7 @@ def build_graph(pipeline_config,
   placeholder_tensor, input_tensors = \
       exporter.input_placeholder_fn_map[input_type]()
 
-  inputs = tf.to_float(input_tensors)
+  inputs = tf.cast(input_tensors, dtype=tf.float32)
   preprocessed_inputs, true_image_shapes = detection_model.preprocess(inputs)
 
   # Dimshuffle: (b, h, w, c) -> (b, c, h, w)

research/object_detection/utils/shape_utils_test.py

@@ -333,5 +333,79 @@ class AssertShapeEqualTest(tf.test.TestCase):
                               tensor_b: np.zeros([5])})
 
 
+class FlattenExpandDimensionTest(tf.test.TestCase):
+
+  def test_flatten_given_dims(self):
+    inputs = tf.random_uniform([5, 2, 10, 10, 3])
+    actual_flattened = shape_utils.flatten_dimensions(inputs, first=1, last=3)
+    expected_flattened = tf.reshape(inputs, [5, 20, 10, 3])
+    with self.test_session() as sess:
+      (actual_flattened_np,
+       expected_flattened_np) = sess.run([actual_flattened, expected_flattened])
+    self.assertAllClose(expected_flattened_np, actual_flattened_np)
+
+  def test_raises_value_error_incorrect_dimensions(self):
+    inputs = tf.random_uniform([5, 2, 10, 10, 3])
+    with self.assertRaises(ValueError):
+      shape_utils.flatten_dimensions(inputs, first=0, last=6)
+
+  def test_flatten_first_two_dimensions(self):
+    inputs = tf.constant(
+        [
+            [[1, 2], [3, 4]],
+            [[5, 6], [7, 8]],
+            [[9, 10], [11, 12]]
+        ], dtype=tf.int32)
+    flattened_tensor = shape_utils.flatten_first_n_dimensions(
+        inputs, 2)
+    with self.test_session() as sess:
+      flattened_tensor_out = sess.run(flattened_tensor)
+
+    expected_output = [[1, 2],
+                       [3, 4],
+                       [5, 6],
+                       [7, 8],
+                       [9, 10],
+                       [11, 12]]
+    self.assertAllEqual(expected_output, flattened_tensor_out)
+
+  def test_expand_first_dimension(self):
+    inputs = tf.constant(
+        [
+            [1, 2],
+            [3, 4],
+            [5, 6],
+            [7, 8],
+            [9, 10],
+            [11, 12]
+        ], dtype=tf.int32)
+    dims = [3, 2]
+    expanded_tensor = shape_utils.expand_first_dimension(
+        inputs, dims)
+    with self.test_session() as sess:
+      expanded_tensor_out = sess.run(expanded_tensor)
+
+    expected_output = [
+        [[1, 2], [3, 4]],
+        [[5, 6], [7, 8]],
+        [[9, 10], [11, 12]]]
+    self.assertAllEqual(expected_output, expanded_tensor_out)
+
+  def test_expand_first_dimension_with_incompatible_dims(self):
+    inputs_default = tf.constant(
+        [
+            [[1, 2]],
+            [[3, 4]],
+            [[5, 6]],
+        ], dtype=tf.int32)
+    inputs = tf.placeholder_with_default(inputs_default, [None, 1, 2])
+    dims = [3, 2]
+    expanded_tensor = shape_utils.expand_first_dimension(
+        inputs, dims)
+    with self.test_session() as sess:
+      with self.assertRaises(tf.errors.InvalidArgumentError):
+        sess.run(expanded_tensor)
+
+
 if __name__ == '__main__':
   tf.test.main()

research/object_detection/utils/visualization_utils.py

@@ -1005,6 +1005,10 @@ class EvalMetricOpsVisualization(object):
           lambda: tf.summary.image(summary_name, image),
           lambda: tf.constant(''))
 
+    if tf.executing_eagerly():
+      update_op = self.add_images([[images[0]]])
+      image_tensors = get_images()
+    else:
       update_op = tf.py_func(self.add_images, [[images[0]]], [])
       image_tensors = tf.py_func(
           get_images, [], [tf.uint8] * self._max_examples_to_draw)