make minor changes

research/object_detection/meta_architectures/context_rcnn_lib_tf2.py

@@ -71,11 +71,11 @@ class AttentionBlock(tf.keras.layers.Layer):
   def build(self, input_shapes):
     pass
 
-  def call(self, input_features, context_features, valid_context_size):
+  def call(self, box_features, context_features, valid_context_size):
     """Handles a call by performing attention."""
     _, context_size, _ = context_features.shape
     valid_mask = compute_valid_mask(valid_context_size, context_size)
-    channels = input_features.shape[-1]
+    channels = box_features.shape[-1]
 
     #Build the feature projection layer
     if not self._output_dimension:
@@ -86,11 +86,11 @@ class AttentionBlock(tf.keras.layers.Layer):
 
     # 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])
+    box_features = tf.reduce_mean(box_features, [2, 3])
 
     with tf.name_scope("AttentionBlock"):
       queries = project_features(
-          input_features, self._bottleneck_dimension, self._is_training,
+          box_features, self._bottleneck_dimension, self._is_training,
           self._query_proj, normalize=True)
       keys = project_features(
           context_features, self._bottleneck_dimension, self._is_training,