Let MovingAverageSummary bind the train_op instead of the session.

.travis.yml

@@ -55,6 +55,7 @@ before_script:
 # Check that these private names can be imported because tensorpack is using them
 - python -c "from tensorflow.python.client.session import _FetchHandler"
 - python -c "from tensorflow.python.training.monitored_session import _HookedSession"
+- python -c "import tensorflow as tf; tf.Operation._add_control_input"
 
 script:
 - flake8 .

tensorpack/callbacks/summary.py

@@ -19,28 +19,46 @@ class MovingAverageSummary(Callback):
     This callback is enabled by default.
     Maintain the moving average of summarized tensors in every step,
     by ops added to the collection.
-    Note that it only __maintains__ the moving averages in the graph,
+    Note that it only __maintains__ the moving averages by updating
+    the relevant variables in the graph,
     the actual summary should be done in other callbacks.
     """
-    def __init__(self, collection=MOVING_SUMMARY_OPS_KEY):
+    def __init__(self, collection=MOVING_SUMMARY_OPS_KEY, train_op=None):
         """
         Args:
             collection(str): the collection of EMA-maintaining ops.
                 The default value would work with
                 the tensors you added by :func:`tfutils.summary.add_moving_summary()`,
                 but you can use other collections as well.
+            train_op (tf.Operation or str): the (name of) training op to associate the maintaing ops with.
+                If not provided, the EMA-maintaining ops will be hooked to
+                `trainer.hooked_session` and be executed in every iteration.
+                Otherwise, the EMA-maintaining ops will be executed whenever
+                the training op is executed.
         """
         self._collection = collection
+        self._train_op = train_op
 
     def _setup_graph(self):
-        ops = tf.get_collection(self._collection)
-        logger.info("Maintain moving average summary of {} tensors in collection {}.".format(
-            len(ops), self._collection))
+        ops = [k.op for k in tf.get_collection(self._collection)]
+        if self._train_op is None:
+            logger.info("[MovingAverageSummary] {} operations in collection '{}' "
+                        "will be run with session hooks.".format(len(ops), self._collection))
 
             self.ema_op = tf.group(*ops, name='maintain_moving_average_summary')
             self._fetch = tf.train.SessionRunArgs(fetches=self.ema_op)
+        else:
+            if isinstance(self._train_op, tf.Tensor):
+                self._train_op = self._train_op.op
+            if not isinstance(self._train_op, tf.Operation):
+                self._train_op = self.graph.get_operation_by_name(self._train_op)
+            self._train_op._add_control_inputs(ops)
+            logger.info("[MovingAverageSummary] {} operations in collection '{}'"
+                        " will be run together with operation '{}'.".format(
+                            len(ops), self._collection, self._train_op.name))
 
     def _before_run(self, _):
+        if self._train_op is None:
             return self._fetch
 
 

tensorpack/graph_builder/training.py

@@ -77,7 +77,7 @@ class DataParallelBuilder(GraphBuilder):
             raise ValueError("Number of gradients from each tower is different! " + str(nvars))
 
     @staticmethod
-    def build_on_towers(
+    def call_for_each_tower(
             towers, func, devices=None, use_vs=None):
         """
         Run `func` on all GPUs (towers) and return the results.
@@ -119,6 +119,10 @@ class DataParallelBuilder(GraphBuilder):
                     ret.append(func())
         return ret
 
+    @staticmethod
+    def build_on_towers(*args, **kwargs):
+        return DataParallelBuilder.call_for_each_tower(*args, **kwargs)
+
 
 class SyncMultiGPUParameterServerBuilder(DataParallelBuilder):
     """

tensorpack/train/trainers.py

@@ -53,7 +53,7 @@ class SimpleTrainer(SingleCostTrainer):
         with TrainTowerContext(''):
             grads = self._make_get_grad_fn(input, get_cost_fn, get_opt_fn)()
             opt = get_opt_fn()
-            self.train_op = opt.apply_gradients(grads, name='min_op')
+            self.train_op = opt.apply_gradients(grads, name='train_op')
         return []
 
 
@@ -404,7 +404,7 @@ class HorovodTrainer(SingleCostTrainer):
             grads = self.allreduce(grads)
 
             opt = get_opt_fn()
-            self.train_op = opt.apply_gradients(grads, name='min_op')
+            self.train_op = opt.apply_gradients(grads, name='train_op')
 
         def broadcast(self):
             logger.info("Running horovod broadcast ...")