context rcnn meta arch

research/object_detection/meta_architectures/context_rcnn_lib_v2.py

@@ -52,7 +52,8 @@ class ContextProjection(tf.keras.layers.Layer):
 class AttentionBlock(tf.keras.layers.Layer):
   """Custom layer to perform all attention."""
   def __init__(self, bottleneck_dimension, attention_temperature,
-               freeze_batchnorm, output_dimension=None, is_training=False, **kwargs):
+               freeze_batchnorm, output_dimension=None,
+               is_training=False, **kwargs):
     self._key_proj = ContextProjection(bottleneck_dimension, freeze_batchnorm)
     self._val_proj = ContextProjection(bottleneck_dimension, freeze_batchnorm)
     self._query_proj = ContextProjection(bottleneck_dimension, freeze_batchnorm)
@@ -77,16 +78,16 @@ class AttentionBlock(tf.keras.layers.Layer):
     channels = input_features.shape[-1]
 
     #Build the feature projection layer
-    if (not self._output_dimension):
+    if not self._output_dimension:
       self._output_dimension = channels
-    if (not self._feature_proj):
+    if not self._feature_proj:
       self._feature_proj = ContextProjection(self._output_dimension,
-                                            self._freeze_batchnorm)
+                                             self._freeze_batchnorm)
 
     # Average pools over height and width dimension so that the shape of
     # box_features becomes [batch_size, max_num_proposals, channels].
     input_features = tf.reduce_mean(input_features, [2, 3])
-    
+
     with tf.variable_scope("AttentionBlock"):
       queries = project_features(
           input_features, self._bottleneck_dimension, self._is_training,

research/object_detection/meta_architectures/context_rcnn_meta_arch.py

@@ -337,7 +337,8 @@ class ContextRCNNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
         [self._initial_crop_size, self._initial_crop_size])
 
     attention_features = self._context_feature_extract_fn(
-        box_features, context_features,
+        box_features,
+        context_features,
         valid_context_size=valid_context_size)
 
     # Adds box features with attention features.