[MaskRCNN] improve name scope

examples/FasterRCNN/README.md

@@ -62,15 +62,15 @@ Evaluation or prediction will need the same config used during training.
 These models are trained with different configurations on trainval35k and evaluated on minival using mAP@IoU=0.50:0.95.
 MaskRCNN results contain both bbox and segm mAP.
 
- | Backbone | mAP<br/>(box/mask) | Detectron mAP <br/> (box/mask) | Time           | Configurations <br/> (click to expand)                                                                                                                                              |
- | -        | -                  | -                              | -              | -                                                                                                                                                                                   |
- | R50-C4   | 33.1               |                                | 18h on 8 V100s | <details><summary>super quick</summary>`MODE_MASK=False FRCNN.BATCH_PER_IM=64 PREPROC.SHORT_EDGE_SIZE=600 PREPROC.MAX_SIZE=1024 TRAIN.LR_SCHEDULE=[150000,230000,280000]` </details> |
- | R50-C4   | 36.6               | 36.5                           | 49h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=False` </details>                                                                                                                    |
- | R50-FPN  | 37.5               | 37.9<sup>[1](#ft1)</sup>       | 28h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=False MODE_FPN=True` </details>                                                                                                      |
- | R50-C4   | 36.8/32.1          |                                | 39h on 8 P100s | <details><summary>quick</summary>`MODE_MASK=True FRCNN.BATCH_PER_IM=256 TRAIN.LR_SCHEDULE=[150000,230000,280000]` </details>                                                         |
- | R50-C4   | 37.8/33.1          | 37.8/32.8                      | 51h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True` </details>                                                                                                                     |
- | R50-FPN  | 38.1/34.9          | 38.6/34.5<sup>[1](#ft1)</sup>  | 38h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True MODE_FPN=True` </details>                                                                                                       |
- | R101-C4  | 40.8/35.1          |                                | 63h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True BACKBONE.RESNET_NUM_BLOCK=[3,4,23,3]` </details>                                                                                |
+ | Backbone | mAP<br/>(box/mask) | Detectron mAP <br/> (box/mask) | Time           | Configurations <br/> (click to expand)                                                                                                                                                           |
+ | -        | -                  | -                              | -              | -                                                                                                                                                                                                |
+ | R50-C4   | 33.1               |                                | 18h on 8 V100s | <details><summary>super quick</summary>`MODE_MASK=False FRCNN.BATCH_PER_IM=64`<br/>`PREPROC.SHORT_EDGE_SIZE=600 PREPROC.MAX_SIZE=1024`<br/>`TRAIN.LR_SCHEDULE=[150000,230000,280000]` </details> |
+ | R50-C4   | 36.6               | 36.5                           | 49h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=False` </details>                                                                                                                                 |
+ | R50-FPN  | 37.5               | 37.9<sup>[1](#ft1)</sup>       | 28h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=False MODE_FPN=True` </details>                                                                                                                   |
+ | R50-C4   | 36.8/32.1          |                                | 39h on 8 P100s | <details><summary>quick</summary>`MODE_MASK=True FRCNN.BATCH_PER_IM=256`<br/>`TRAIN.LR_SCHEDULE=[150000,230000,280000]` </details>                                                               |
+ | R50-C4   | 37.8/33.1          | 37.8/32.8                      | 51h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True` </details>                                                                                                                                  |
+ | R50-FPN  | 38.1/34.9          | 38.6/34.5<sup>[1](#ft1)</sup>  | 38h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True MODE_FPN=True` </details>                                                                                                                    |
+ | R101-C4  | 40.8/35.1          |                                | 63h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True `<br/>`BACKBONE.RESNET_NUM_BLOCK=[3,4,23,3]` </details>                                                                                      |
  
  <a id="ft1">1</a>: Slightly different configurations.
 

examples/FasterRCNN/config.py

@@ -60,7 +60,7 @@ _C.DATA.NUM_CATEGORY = 80    # 80 categories
 _C.DATA.CLASS_NAMES = []  # NUM_CLASS strings. Needs to be populated later by data loader
 
 # basemodel ----------------------
-_C.BACKBONE.WEIGHTS = '/path/to/ImageNet-ResNet50.npz'
+_C.BACKBONE.WEIGHTS = ''   # /path/to/weights.npz
 _C.BACKBONE.RESNET_NUM_BLOCK = [3, 4, 6, 3]     # for resnet50
 # RESNET_NUM_BLOCK = [3, 4, 23, 3]    # for resnet101
 _C.BACKBONE.FREEZE_AFFINE = False   # do not train affine parameters inside BN

examples/FasterRCNN/model_fpn.py

@@ -141,12 +141,12 @@ def multilevel_rpn_losses(
     losses = []
     with tf.name_scope('rpn_losses'):
         for lvl in range(num_lvl):
-            with tf.name_scope('Level{}'.format(lvl + 2)):
-                anchors = multilevel_anchors[lvl]
-                label_loss, box_loss = rpn_losses(
-                    anchors.gt_labels, anchors.encoded_gt_boxes(),
-                    multilevel_label_logits[lvl], multilevel_box_logits[lvl])
-                losses.extend([label_loss, box_loss])
+            anchors = multilevel_anchors[lvl]
+            label_loss, box_loss = rpn_losses(
+                anchors.gt_labels, anchors.encoded_gt_boxes(),
+                multilevel_label_logits[lvl], multilevel_box_logits[lvl],
+                name_scope='Level{}'.format(lvl + 2))
+            losses.extend([label_loss, box_loss])
 
         total_label_loss = tf.add_n(losses[::2], name='label_loss')
         total_box_loss = tf.add_n(losses[1::2], name='box_loss')

examples/FasterRCNN/model_rpn.py

@@ -37,6 +37,7 @@ def rpn_head(featuremap, channel, num_anchors):
     return label_logits, box_logits
 
 
+@under_name_scope()
 def rpn_losses(anchor_labels, anchor_boxes, label_logits, box_logits):
     """
     Args:

examples/FasterRCNN/train.py

@@ -200,9 +200,8 @@ class ResNetC4Model(DetectionModel):
 
         if is_training:
             # rpn loss
-            with tf.name_scope('rpn_losses'):
-                rpn_label_loss, rpn_box_loss = rpn_losses(
-                    anchors.gt_labels, anchors.encoded_gt_boxes(), rpn_label_logits, rpn_box_logits)
+            rpn_label_loss, rpn_box_loss = rpn_losses(
+                anchors.gt_labels, anchors.encoded_gt_boxes(), rpn_label_logits, rpn_box_logits)
 
             # fastrcnn loss
             matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)

tensorpack/tfutils/scope_utils.py

@@ -54,11 +54,17 @@ def auto_reuse_variable_scope(func):
     return wrapper
 
 
-def under_name_scope(name=None):
+def under_name_scope(name_scope=None):
     """
+    Args:
+        name_scope(str): the default scope to use. If None, will use the name of the function.
+
     Returns:
-        A decorator which makes the function happen under a name scope.
-        The default name is the function itself.
+        A decorator which makes the function run under a name scope.
+        The name scope is obtained by the following:
+        1. The 'name_scope' keyword argument when the decorated function is called.
+        2. The 'name_scope' argument of the decorator.
+        3. (default) The name of the decorated function itself.
 
     Example:
 
@@ -66,9 +72,13 @@ def under_name_scope(name=None):
 
         @under_name_scope()
         def rms(x):
-            return tf.sqrt(  # will be under name scope 'rms'
+            return tf.sqrt(
                 tf.reduce_mean(tf.square(x)))
 
+        rms(tensor)  # will be called under name scope 'rms'
+        rms(tensor, name_scope='scope')  # will be called under name scope 'scope'
+
+
     Todo:
         Add a reuse option.
     """
@@ -76,10 +86,10 @@ def under_name_scope(name=None):
     def _impl(func):
         @functools.wraps(func)
         def wrapper(*args, **kwargs):
-            if name is None:
+            scopename = kwargs.pop('name_scope', name_scope)
+            if scopename is None:
                 scopename = func.__name__
-            else:
-                scopename = name
+
             with tf.name_scope(scopename):
                 return func(*args, **kwargs)
         return wrapper