update docs and sotabench

examples/FasterRCNN/config.py

@@ -43,6 +43,15 @@ class AttrDict():
         return {k: v.to_dict() if isinstance(v, AttrDict) else v
                 for k, v in self.__dict__.items() if not k.startswith('_')}
 
+    def from_dict(self, d):
+        self.freeze(False)
+        for k, v in d.items():
+            self_v = getattr(self, k)
+            if isinstance(self_v, AttrDict):
+                self_v.from_dict(v)
+            else:
+                setattr(self, k, v)
+
     def update_args(self, args):
         """Update from command line args. """
         for cfg in args:

sotabench/sotabench.py

@@ -3,6 +3,7 @@
 import os
 import sys
 import tqdm
+from contextlib import contextmanager
 
 from tensorpack.predict import OfflinePredictor, PredictConfig
 from tensorpack.tfutils import SmartInit
@@ -31,6 +32,13 @@ COCO_ROOT = os.path.join(DATA_ROOT, "coco")
 register_coco(COCO_ROOT)
 
 
+@contextmanager
+def backup_cfg():
+    orig_config = cfg.to_dict()
+    yield
+    cfg.from_dict(orig_config)
+
+
 def evaluate_rcnn(model_name, paper_arxiv_id, cfg_list, model_file):
     evaluator = COCOEvaluator(
         root=COCO_ROOT, model_name=model_name, paper_arxiv_id=paper_arxiv_id
@@ -77,16 +85,41 @@ def evaluate_rcnn(model_name, paper_arxiv_id, cfg_list, model_file):
     evaluator.save()
 
 
+download(
+    "http://models.tensorpack.com/FasterRCNN/COCO-MaskRCNN-R50FPN2x.npz",
+    "./",
+    expect_size=165362754)
+with backup_cfg():
+    evaluate_rcnn(
+        "Mask R-CNN (ResNet-50-FPN, 2x)", "1703.06870", [],
+        "COCO-MaskRCNN-R50FPN2x.npz",
+    )
+
+
+download(
+    "http://models.tensorpack.com/FasterRCNN/COCO-MaskRCNN-R50FPN2xGN.npz",
+    "./",
+    expect_size=167363872)
+with backup_cfg():
+    evaluate_rcnn(
+        "Mask R-CNN (ResNet-50-FPN, GroupNorm)", "1803.08494",
+        """FPN.NORM=GN BACKBONE.NORM=GN
+FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head
+FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head""".split(),
+        "COCO-MaskRCNN-R50FPN2xGN.npz",
+    )
+
+
 download(
     "http://models.tensorpack.com/FasterRCNN/COCO-MaskRCNN-R101FPN9xGNCasAugScratch.npz",
     "./",
     expect_size=355680386)
-evaluate_rcnn(
-    "Mask R-CNN (ResNet-101-FPN, GN, Cascade)",
-    "1811.08883",
+with backup_cfg():
+    evaluate_rcnn(
+        "Mask R-CNN (ResNet-101-FPN, GN, Cascade)", "1811.08883",
         """
-FPN.CASCADE=True BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3] FPN.NORM=GN
-BACKBONE.NORM=GN FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head
-FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head""".split(),
+    FPN.CASCADE=True BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3] FPN.NORM=GN
+    BACKBONE.NORM=GN FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head
+    FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head""".split(),
         "COCO-MaskRCNN-R101FPN9xGNCasAugScratch.npz",
-)
+    )

tensorpack/models/batch_norm.py

@@ -98,7 +98,7 @@ def BatchNorm(inputs, axis=None, training=None, momentum=0.9, epsilon=1e-5,
             This is not a good argument name, but it is what the Tensorflow layer uses.
         ema_update (str): Only effective when ``training=True``. It has the following options:
 
-          * "default": same as "collection". Because this is the default behavior in tensorflow.
+          * "default": same as "collection". Because this is the default behavior in TensorFlow.
           * "skip": do not update EMA. This can be useful when you reuse a batch norm layer in several places
             but do not want them to all update your EMA.
           * "collection": Add EMA update ops to collection `tf.GraphKeys.UPDATE_OPS`.
@@ -106,7 +106,7 @@ def BatchNorm(inputs, axis=None, training=None, momentum=0.9, epsilon=1e-5,
             your training iterations. This can waste compute if your training iterations do not always depend
             on the BatchNorm layer.
           * "internal": EMA is updated inside this layer itself by control dependencies.
-            In common cases, it has similar speed to "collection". But it covers more cases, e.g.:
+            In standard scenarios, it has similar speed to "collection". But it has some more benefits:
 
             1. BatchNorm is used inside dynamic control flow.
                The collection-based update does not support dynamic control flows.
@@ -114,7 +114,9 @@ def BatchNorm(inputs, axis=None, training=None, momentum=0.9, epsilon=1e-5,
                Putting all update ops into a single collection will waste a lot of compute.
             3. Other part of the model relies on the "updated" EMA. The collection-based method does not update
                EMA immediately.
+            4. It has less chance to cause TensorFlow bugs in a graph with complicated control flow.
 
+            Therefore this option is preferred over TensorFlow default.
             Corresponding TF issue: https://github.com/tensorflow/tensorflow/issues/14699
         sync_statistics (str or None): one of None, "nccl", or "horovod". It determines how to compute the
           "per-batch statistics" when ``training==True``.