Internal change

PiperOrigin-RevId: 397809846

Author: xianzhidu

official/vision/beta/configs/maskrcnn.py

@@ -131,7 +131,7 @@ class DetectionGenerator(hyperparams.Config):
   pre_nms_score_threshold: float = 0.05
   nms_iou_threshold: float = 0.5
   max_num_detections: int = 100
-  use_batched_nms: bool = False
+  nms_version: str = 'v2'  # `v2`, `v1`, `batched`
   use_cpu_nms: bool = False
 
 

official/vision/beta/configs/retinanet.py

@@ -112,7 +112,7 @@ class DetectionGenerator(hyperparams.Config):
   pre_nms_score_threshold: float = 0.05
   nms_iou_threshold: float = 0.5
   max_num_detections: int = 100
-  use_batched_nms: bool = False
+  nms_version: str = 'v2'  # `v2`, `v1`, `batched`.
   use_cpu_nms: bool = False
 
 

official/vision/beta/modeling/factory.py

@@ -197,7 +197,7 @@ def build_maskrcnn(
       pre_nms_score_threshold=generator_config.pre_nms_score_threshold,
       nms_iou_threshold=generator_config.nms_iou_threshold,
       max_num_detections=generator_config.max_num_detections,
-      use_batched_nms=generator_config.use_batched_nms,
+      nms_version=generator_config.nms_version,
       use_cpu_nms=generator_config.use_cpu_nms)
 
   if model_config.include_mask:
@@ -300,7 +300,7 @@ def build_retinanet(
       pre_nms_score_threshold=generator_config.pre_nms_score_threshold,
       nms_iou_threshold=generator_config.nms_iou_threshold,
       max_num_detections=generator_config.max_num_detections,
-      use_batched_nms=generator_config.use_batched_nms,
+      nms_version=generator_config.nms_version,
       use_cpu_nms=generator_config.use_cpu_nms)
 
   model = retinanet_model.RetinaNetModel(

official/vision/beta/modeling/layers/detection_generator.py

@@ -404,7 +404,7 @@ def __init__(self,
                pre_nms_score_threshold: float = 0.05,
                nms_iou_threshold: float = 0.5,
                max_num_detections: int = 100,
-               use_batched_nms: bool = False,
+               nms_version: str = 'v2',
                use_cpu_nms: bool = False,
                **kwargs):
     """Initializes a detection generator.
@@ -420,8 +420,7 @@ def __init__(self,
       nms_iou_threshold: A `float` in [0, 1], the NMS IoU threshold.
       max_num_detections: An `int` of the final number of total detections to
         generate.
-      use_batched_nms: A `bool` of whether or not use
-        `tf.image.combined_non_max_suppression`.
+      nms_version: A string of `batched`, `v1` or `v2` specifies NMS version.
       use_cpu_nms: A `bool` of whether or not enforce NMS to run on CPU.
       **kwargs: Additional keyword arguments passed to Layer.
     """
@@ -431,7 +430,7 @@ def __init__(self,
         'pre_nms_score_threshold': pre_nms_score_threshold,
         'nms_iou_threshold': nms_iou_threshold,
         'max_num_detections': max_num_detections,
-        'use_batched_nms': use_batched_nms,
+        'nms_version': nms_version,
         'use_cpu_nms': use_cpu_nms,
     }
     super(DetectionGenerator, self).__init__(**kwargs)
@@ -524,14 +523,14 @@ def __call__(self,
       nms_context = contextlib.nullcontext()
 
     with nms_context:
-      if self._config_dict['use_batched_nms']:
+      if self._config_dict['nms_version'] == 'batched':
         (nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections) = (
             _generate_detections_batched(
                 decoded_boxes, box_scores,
                 self._config_dict['pre_nms_score_threshold'],
                 self._config_dict['nms_iou_threshold'],
                 self._config_dict['max_num_detections']))
-      else:
+      elif self._config_dict['nms_version'] == 'v1':
         (nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections, _) = (
             _generate_detections_v1(
                 decoded_boxes,
@@ -541,6 +540,19 @@ def __call__(self,
                 ._config_dict['pre_nms_score_threshold'],
                 nms_iou_threshold=self._config_dict['nms_iou_threshold'],
                 max_num_detections=self._config_dict['max_num_detections']))
+      elif self._config_dict['nms_version'] == 'v2':
+        (nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections) = (
+            _generate_detections_v2(
+                decoded_boxes,
+                box_scores,
+                pre_nms_top_k=self._config_dict['pre_nms_top_k'],
+                pre_nms_score_threshold=self
+                ._config_dict['pre_nms_score_threshold'],
+                nms_iou_threshold=self._config_dict['nms_iou_threshold'],
+                max_num_detections=self._config_dict['max_num_detections']))
+      else:
+        raise ValueError('NMS version {} not supported.'.format(
+            self._config_dict['nms_version']))
 
     # Adds 1 to offset the background class which has index 0.
     nmsed_classes += 1
@@ -570,7 +582,7 @@ def __init__(self,
                pre_nms_score_threshold: float = 0.05,
                nms_iou_threshold: float = 0.5,
                max_num_detections: int = 100,
-               use_batched_nms: bool = False,
+               nms_version: str = 'v1',
                use_cpu_nms: bool = False,
                **kwargs):
     """Initializes a multi-level detection generator.
@@ -586,8 +598,7 @@ def __init__(self,
       nms_iou_threshold: A `float` in [0, 1], the NMS IoU threshold.
       max_num_detections: An `int` of the final number of total detections to
         generate.
-      use_batched_nms: A `bool` of whether or not use
-        `tf.image.combined_non_max_suppression`.
+      nms_version: A string of `batched`, `v1` or `v2` specifies NMS version
       use_cpu_nms: A `bool` of whether or not enforce NMS to run on CPU.
       **kwargs: Additional keyword arguments passed to Layer.
     """
@@ -597,7 +608,7 @@ def __init__(self,
         'pre_nms_score_threshold': pre_nms_score_threshold,
         'nms_iou_threshold': nms_iou_threshold,
         'max_num_detections': max_num_detections,
-        'use_batched_nms': use_batched_nms,
+        'nms_version': nms_version,
         'use_cpu_nms': use_cpu_nms,
     }
     super(MultilevelDetectionGenerator, self).__init__(**kwargs)
@@ -731,19 +742,19 @@ def __call__(self,
       nms_context = contextlib.nullcontext()
 
     with nms_context:
-      if self._config_dict['use_batched_nms']:
-        if raw_attributes:
-          raise ValueError(
-              'Attribute learning is not supported for batched NMS.')
-
+      if raw_attributes and (self._config_dict['nms_version'] != 'v1'):
+        raise ValueError(
+            'Attribute learning is only supported for NMSv1 but NMS {} is used.'
+            .format(self._config_dict['nms_version']))
+      if self._config_dict['nms_version'] == 'batched':
         (nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections) = (
             _generate_detections_batched(
                 boxes, scores, self._config_dict['pre_nms_score_threshold'],
                 self._config_dict['nms_iou_threshold'],
                 self._config_dict['max_num_detections']))
         # Set `nmsed_attributes` to None for batched NMS.
         nmsed_attributes = {}
-      else:
+      elif self._config_dict['nms_version'] == 'v1':
         (nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections,
          nmsed_attributes) = (
              _generate_detections_v1(
@@ -755,6 +766,21 @@ def __call__(self,
                  ._config_dict['pre_nms_score_threshold'],
                  nms_iou_threshold=self._config_dict['nms_iou_threshold'],
                  max_num_detections=self._config_dict['max_num_detections']))
+      elif self._config_dict['nms_version'] == 'v2':
+        (nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections) = (
+            _generate_detections_v2(
+                boxes,
+                scores,
+                pre_nms_top_k=self._config_dict['pre_nms_top_k'],
+                pre_nms_score_threshold=self
+                ._config_dict['pre_nms_score_threshold'],
+                nms_iou_threshold=self._config_dict['nms_iou_threshold'],
+                max_num_detections=self._config_dict['max_num_detections']))
+        # Set `nmsed_attributes` to None for v2.
+        nmsed_attributes = {}
+      else:
+        raise ValueError('NMS version {} not supported.'.format(
+            self._config_dict['nms_version']))
 
     # Adds 1 to offset the background class which has index 0.
     nmsed_classes += 1

official/vision/beta/modeling/layers/detection_generator_test.py

@@ -44,8 +44,8 @@ class DetectionGeneratorTest(
     parameterized.TestCase, tf.test.TestCase):
 
   @parameterized.product(
-      use_batched_nms=[True, False], use_cpu_nms=[True, False])
-  def testDetectionsOutputShape(self, use_batched_nms, use_cpu_nms):
+      nms_version=['batched', 'v1', 'v2'], use_cpu_nms=[True, False])
+  def testDetectionsOutputShape(self, nms_version, use_cpu_nms):
     max_num_detections = 100
     num_classes = 4
     pre_nms_top_k = 5000
@@ -57,7 +57,7 @@ def testDetectionsOutputShape(self, use_batched_nms, use_cpu_nms):
         'pre_nms_score_threshold': pre_nms_score_threshold,
         'nms_iou_threshold': 0.5,
         'max_num_detections': max_num_detections,
-        'use_batched_nms': use_batched_nms,
+        'nms_version': nms_version,
         'use_cpu_nms': use_cpu_nms,
     }
     generator = detection_generator.DetectionGenerator(**kwargs)
@@ -97,7 +97,7 @@ def test_serialize_deserialize(self):
         'pre_nms_score_threshold': 0.1,
         'nms_iou_threshold': 0.5,
         'max_num_detections': 10,
-        'use_batched_nms': False,
+        'nms_version': 'v2',
         'use_cpu_nms': False,
     }
     generator = detection_generator.DetectionGenerator(**kwargs)
@@ -116,15 +116,14 @@ class MultilevelDetectionGeneratorTest(
     parameterized.TestCase, tf.test.TestCase):
 
   @parameterized.parameters(
-      (True, False, True),
-      (True, False, False),
-      (False, False, True),
-      (False, False, False),
-      (False, True, True),
-      (False, True, False),
+      ('batched', False, True),
+      ('batched', False, False),
+      ('v2', False, True),
+      ('v2', False, False),
+      ('v1', True, True),
+      ('v1', True, False),
   )
-  def testDetectionsOutputShape(self, use_batched_nms, has_att_heads,
-                                use_cpu_nms):
+  def testDetectionsOutputShape(self, nms_version, has_att_heads, use_cpu_nms):
     min_level = 4
     max_level = 6
     num_scales = 2
@@ -142,7 +141,7 @@ def testDetectionsOutputShape(self, use_batched_nms, has_att_heads,
         'pre_nms_score_threshold': pre_nms_score_threshold,
         'nms_iou_threshold': 0.5,
         'max_num_detections': max_num_detections,
-        'use_batched_nms': use_batched_nms,
+        'nms_version': nms_version,
         'use_cpu_nms': use_cpu_nms,
     }
 
@@ -223,7 +222,7 @@ def test_serialize_deserialize(self):
         'pre_nms_score_threshold': 0.1,
         'nms_iou_threshold': 0.5,
         'max_num_detections': 10,
-        'use_batched_nms': False,
+        'nms_version': 'v2',
         'use_cpu_nms': False,
     }
     generator = detection_generator.MultilevelDetectionGenerator(**kwargs)

official/vision/beta/modeling/retinanet_model_test.py

@@ -193,7 +193,7 @@ def test_forward(self, strategy, image_size, training, has_att_heads,
           attribute_heads=attribute_heads,
           num_anchors_per_location=num_anchors_per_location)
       generator = detection_generator.MultilevelDetectionGenerator(
-          max_num_detections=10)
+          max_num_detections=10, nms_version='v1')
       model = retinanet_model.RetinaNetModel(
           backbone=backbone,
           decoder=decoder,

official/vision/beta/projects/deepmac_maskrcnn/serving/detection.py

@@ -28,7 +28,7 @@ def _build_model(self):
 
     if self._batch_size is None:
       ValueError("batch_size can't be None for detection models")
-    if not self.params.task.model.detection_generator.use_batched_nms:
+    if self.params.task.model.detection_generator.nms_version != 'batched':
       ValueError('Only batched_nms is supported.')
     input_specs = tf.keras.layers.InputSpec(shape=[self._batch_size] +
                                             self._input_image_size + [3])

official/vision/beta/projects/deepmac_maskrcnn/tasks/deep_mask_head_rcnn.py

@@ -120,7 +120,7 @@ def build_maskrcnn(input_specs: tf.keras.layers.InputSpec,
       pre_nms_score_threshold=generator_config.pre_nms_score_threshold,
       nms_iou_threshold=generator_config.nms_iou_threshold,
       max_num_detections=generator_config.max_num_detections,
-      use_batched_nms=generator_config.use_batched_nms)
+      nms_version=generator_config.nms_version)
 
   if model_config.include_mask:
     mask_head = deep_instance_heads.DeepMaskHead(

official/vision/beta/serving/detection_test.py

@@ -33,7 +33,7 @@ class DetectionExportTest(tf.test.TestCase, parameterized.TestCase):
   def _get_detection_module(self, experiment_name):
     params = exp_factory.get_exp_config(experiment_name)
     params.task.model.backbone.resnet.model_id = 18
-    params.task.model.detection_generator.use_batched_nms = True
+    params.task.model.detection_generator.nms_version = 'batched'
     detection_module = detection.DetectionModule(
         params, batch_size=1, input_image_size=[640, 640])
     return detection_module