Reenter the same name_scope when reusing a variable_scope

tensorpack/callbacks/summary.py

@@ -27,7 +27,8 @@ class MovingAverageSummary(Callback):
 
     def _setup_graph(self):
         ops = tf.get_collection(self._collection)
-        logger.info("Maintain moving averages of {} tensors.".format(len(ops)))
+        logger.info("Maintain moving average summary of {} tensors.".format(len(ops)))
+
         self.ema_op = tf.group(*ops, name='summary_moving_averages')
         self._fetch = tf.train.SessionRunArgs(fetches=self.ema_op)
 

tensorpack/tfutils/scope_utils.py

@@ -44,8 +44,10 @@ def auto_reuse_variable_scope(func):
         h = hash((tf.get_default_graph(), scope.name))
         # print("Entering " + scope.name + " reuse: " + str(h in used_scope))
         if h in used_scope:
+            ns = scope.original_name_scope
             with tf.variable_scope(scope, reuse=True):
-                return func(*args, **kwargs)
+                with tf.name_scope(ns):
+                    return func(*args, **kwargs)
         else:
             used_scope.add(h)
             return func(*args, **kwargs)

tensorpack/tfutils/summary.py

@@ -166,8 +166,10 @@ def add_moving_summary(v, *args, **kwargs):
     # TODO will produce variable tower0/xxx?
     # TODO not saved under distributed
     # TODO use zero_debias
+    # TODO create EMA for each variable separately, so that the maintain ops
+    # have a decent name (rather than EMA)
     gs = get_global_step_var()
-    with tf.name_scope(None), tf.device(gs.device):
+    with tf.device(gs.device):
         averager = tf.train.ExponentialMovingAverage(
             decay, num_updates=gs, name='EMA')
         avg_maintain_op = averager.apply(v)