Rename AnchorTarget to RPNAnchorTarget

frcnn/rcnn_target.py

@@ -9,8 +9,6 @@
 class RCNNTarget(snt.AbstractModule):
     """
     Generate RCNN target tensors for both probabilities and bounding boxes.
-
-    TODO: We should unify this module with AnchorTarget.
     """
     def __init__(self, num_classes, name='rcnn_proposal'):
         super(RCNNTarget, self).__init__(name=name)

frcnn/rpn.py

@@ -7,7 +7,7 @@
 
 from sonnet.python.modules.conv import Conv2D
 
-from .anchor_target import AnchorTarget
+from .rpn_anchor_target import RPNAnchorTarget
 from .rpn_proposal import RPNProposal
 from .utils.generate_anchors import generate_anchors
 from .utils.losses import smooth_l1_loss
@@ -63,7 +63,7 @@ def _instantiate_layers(self):
 
 
             self._proposal = RPNProposal(self._num_anchors)
-            self._anchor_target = AnchorTarget(self._num_anchors)
+            self._anchor_target = RPNAnchorTarget(self._num_anchors)
 
     def _build(self, pretrained_feature_map, gt_boxes, image_shape, all_anchors,
                is_training=True):
@@ -188,7 +188,7 @@ def loss(self, prediction_dict):
 
                 # Finally, we need to calculate the regression loss over
                 # `rpn_bbox_target` and `rpn_bbox_pred`.
-                # Since `rpn_bbox_target` is obtained from AnchorTargetLayer then we
+                # Since `rpn_bbox_target` is obtained from RPNAnchorTarget then we
                 # just need to apply SmoothL1Loss.
                 rpn_bbox_target = tf.reshape(rpn_bbox_target, [-1, 4])
                 rpn_bbox_pred = tf.reshape(rpn_bbox_pred, [-1, 4])

frcnn/anchor_target.py → frcnn/rpn_anchor_target.py

@@ -6,9 +6,9 @@
 from .utils.bbox_transform import bbox_transform, unmap
 
 
-class AnchorTarget(snt.AbstractModule):
+class RPNAnchorTarget(snt.AbstractModule):
     """
-    AnchorTarget
+    RPNAnchorTarget
 
     TODO: (copied) Assign anchors to ground-truth targets. Produces anchor
     classification labels and bounding-box regression targets.
@@ -40,7 +40,7 @@ class AnchorTarget(snt.AbstractModule):
 
     """
     def __init__(self, num_anchors, feat_stride=[16], name='anchor_target'):
-        super(AnchorTarget, self).__init__(name=name)
+        super(RPNAnchorTarget, self).__init__(name=name)
         self._num_anchors = num_anchors
         self._feat_stride = feat_stride
 

frcnn/anchor_target_test.py → frcnn/rpn_anchor_target_test.py

@@ -4,22 +4,22 @@
 
 from sonnet.testing.parameterized import parameterized
 
-from .anchor_target import AnchorTarget
+from .rpn_anchor_target import RPNAnchorTarget
 from .utils.generate_anchors import generate_anchors
 
 
-class AnchorTargetTest(parameterized.ParameterizedTestCase, tf.test.TestCase):
+class RPNnchorTargetTest(parameterized.ParameterizedTestCase, tf.test.TestCase):
 
     def setUp(self):
-        super(AnchorTargetTest, self).setUp()
+        super(RPNAnchorTargetTest, self).setUp()
         # Setup anchors
         self.anchor_scales = np.array([8, 16, 32])
         self.anchor_ratios = np.array([0.5, 1, 2])
         self.anchors = generate_anchors(
             ratios=self.anchor_ratios, scales=self.anchor_scales)
 
     def testBasic(self):
-        model = AnchorTarget(self.anchors)
+        model = RPNAnchorTarget(self.anchors)
         rpn_cls_score_shape = (1, 32, 32, model._num_anchors * 2)
         gt_boxes_shape = (1, 4)  # 1 ground truth boxes.
         im_info_shape = (2,)