[FasterRCNN] rename a config; prefer arguments over constants

examples/FasterRCNN/config.py

@@ -22,7 +22,7 @@ ANCHOR_STRIDE = 16
 # sqrtarea of the anchor box
 ANCHOR_SIZES = (32, 64, 128, 256, 512)
 ANCHOR_RATIOS = (0.5, 1., 2.)
-NR_ANCHOR = len(ANCHOR_SIZES) * len(ANCHOR_RATIOS)
+NUM_ANCHOR = len(ANCHOR_SIZES) * len(ANCHOR_RATIOS)
 POSITIVE_ANCHOR_THRES = 0.7
 NEGATIVE_ANCHOR_THRES = 0.3
 

examples/FasterRCNN/data.py

@@ -29,25 +29,29 @@ class MalformedData(BaseException):
 
 
 @memoized
-def get_all_anchors():
+def get_all_anchors(
+        stride=config.ANCHOR_STRIDE,
+        sizes=config.ANCHOR_SIZES):
     """
     Get all anchors in the largest possible image, shifted, floatbox
 
     Returns:
-        anchors: SxSxNR_ANCHORx4, where S == MAX_SIZE//STRIDE, floatbox
+        anchors: SxSxNUM_ANCHORx4, where S == MAX_SIZE//STRIDE, floatbox
+        The layout in the NUM_ANCHOR dim is NUM_RATIO x NUM_SCALE.
+
     """
     # Generates a NAx4 matrix of anchor boxes in (x1, y1, x2, y2) format. Anchors
     # are centered on stride / 2, have (approximate) sqrt areas of the specified
     # sizes, and aspect ratios as given.
     cell_anchors = generate_anchors(
-        config.ANCHOR_STRIDE,
-        scales=np.array(config.ANCHOR_SIZES, dtype=np.float) / config.ANCHOR_STRIDE,
+        stride,
+        scales=np.array(sizes, dtype=np.float) / stride,
         ratios=np.array(config.ANCHOR_RATIOS, dtype=np.float))
     # anchors are intbox here.
     # anchors at featuremap [0,0] are centered at fpcoor (8,8) (half of stride)
 
-    field_size = config.MAX_SIZE // config.ANCHOR_STRIDE
-    shifts = np.arange(0, field_size) * config.ANCHOR_STRIDE
+    field_size = config.MAX_SIZE // stride
+    shifts = np.arange(0, field_size) * stride
     shift_x, shift_y = np.meshgrid(shifts, shifts)
     shift_x = shift_x.flatten()
     shift_y = shift_y.flatten()
@@ -179,12 +183,12 @@ def get_rpn_anchor_input(im, boxes, klass, is_crowd):
     anchor_labels, anchor_boxes = get_anchor_labels(inside_anchors, non_crowd_boxes, crowd_boxes)
 
     # Fill them back to original size: fHxfWx1, fHxfWx4
-    featuremap_labels = -np.ones((featureH * featureW * config.NR_ANCHOR, ), dtype='int32')
+    featuremap_labels = -np.ones((featureH * featureW * config.NUM_ANCHOR, ), dtype='int32')
     featuremap_labels[inside_ind] = anchor_labels
-    featuremap_labels = featuremap_labels.reshape((featureH, featureW, config.NR_ANCHOR))
-    featuremap_boxes = np.zeros((featureH * featureW * config.NR_ANCHOR, 4), dtype='float32')
+    featuremap_labels = featuremap_labels.reshape((featureH, featureW, config.NUM_ANCHOR))
+    featuremap_boxes = np.zeros((featureH * featureW * config.NUM_ANCHOR, 4), dtype='float32')
     featuremap_boxes[inside_ind, :] = anchor_boxes
-    featuremap_boxes = featuremap_boxes.reshape((featureH, featureW, config.NR_ANCHOR, 4))
+    featuremap_boxes = featuremap_boxes.reshape((featureH, featureW, config.NUM_ANCHOR, 4))
     return featuremap_labels, featuremap_boxes
 
 

examples/FasterRCNN/model.py

@@ -96,11 +96,12 @@ def rpn_losses(anchor_labels, anchor_boxes, label_logits, box_logits):
 
 
 @under_name_scope()
-def decode_bbox_target(box_predictions, anchors):
+def decode_bbox_target(box_predictions, anchors, stride):
     """
     Args:
         box_predictions: fHxfWxNAx4, logits
         anchors: fHxfWxNAx4, floatbox
+        stride (int): the stride of the anchors
 
     Returns:
         box_decoded: (fHxfWxNA)x4, float32
@@ -115,7 +116,7 @@ def decode_bbox_target(box_predictions, anchors):
     xaya = tf.to_float(anchors_x2y2 + anchors_x1y1) * 0.5
 
     wbhb = tf.exp(tf.minimum(
-        box_pred_twth, np.log(config.MAX_SIZE * 1.0 / config.ANCHOR_STRIDE))) * waha
+        box_pred_twth, np.log(config.MAX_SIZE * 1.0 / stride))) * waha
     xbyb = box_pred_txty * waha + xaya
     x1y1 = xbyb - wbhb * 0.5
     x2y2 = xbyb + wbhb * 0.5

examples/FasterRCNN/train.py

@@ -54,8 +54,8 @@ class Model(ModelDesc):
     def _get_inputs(self):
         return [
             InputDesc(tf.float32, (None, None, 3), 'image'),
-            InputDesc(tf.int32, (None, None, config.NR_ANCHOR), 'anchor_labels'),
-            InputDesc(tf.float32, (None, None, config.NR_ANCHOR, 4), 'anchor_boxes'),
+            InputDesc(tf.int32, (None, None, config.NUM_ANCHOR), 'anchor_labels'),
+            InputDesc(tf.float32, (None, None, config.NUM_ANCHOR, 4), 'anchor_boxes'),
             InputDesc(tf.float32, (None, 4), 'gt_boxes'),
             InputDesc(tf.int64, (None,), 'gt_labels'),
         ]
@@ -88,11 +88,11 @@ class Model(ModelDesc):
 
         anchor_boxes_encoded = encode_bbox_target(anchor_boxes, fm_anchors)
         featuremap = pretrained_resnet_conv4(image, config.RESNET_NUM_BLOCK[:3])
-        rpn_label_logits, rpn_box_logits = rpn_head(featuremap, 1024, config.NR_ANCHOR)
+        rpn_label_logits, rpn_box_logits = rpn_head(featuremap, 1024, config.NUM_ANCHOR)
         rpn_label_loss, rpn_box_loss = rpn_losses(
             anchor_labels, anchor_boxes_encoded, rpn_label_logits, rpn_box_logits)
 
-        decoded_boxes = decode_bbox_target(rpn_box_logits, fm_anchors)  # (fHxfWxNA)x4, floatbox
+        decoded_boxes = decode_bbox_target(rpn_box_logits, fm_anchors, config.ANCHOR_STRIDE)  # (fHxfWxNA)x4, floatbox
         proposal_boxes, proposal_scores = generate_rpn_proposals(
             decoded_boxes,
             tf.reshape(rpn_label_logits, [-1]),
@@ -131,7 +131,7 @@ class Model(ModelDesc):
             fg_boxes = tf.gather(proposal_boxes, fg_ind)
 
             fg_box_logits = fg_box_logits / tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS)
-            decoded_boxes = decode_bbox_target(fg_box_logits, fg_boxes)  # Nfx4, floatbox
+            decoded_boxes = decode_bbox_target(fg_box_logits, fg_boxes, config.ANCHOR_STRIDE)  # Nfx4, floatbox
             decoded_boxes = tf.identity(decoded_boxes, name='fastrcnn_fg_boxes')
 
     def _get_optimizer(self):