bug fix

examples/DeepQNetwork/README.md

@@ -24,8 +24,7 @@ My Batch-A3C implementation only took <2 hours.
 Both were trained on one GPU with an extra GPU for simulation.
 
 The x-axis is the number of iterations, not wall time.
-Iteration speed on Tesla M40 is about 9.7it/s for B-A3C.
-D-DQN is faster at the beginning but will converge to 12it/s due of exploration annealing.
+D-DQN is faster at the beginning but will converge to 12 batches/s (768 frames/s) due of exploration annealing.
 
 ## How to use
 

tensorpack/tfutils/optimizer.py

@@ -4,6 +4,7 @@
 # Author: Yuxin Wu <ppwwyyxxc@gmail.com>
 
 import tensorflow as tf
+from contextlib import contextmanager
 from .gradproc import apply_grad_processors as apply_gradproc
 
 __all__ = ['apply_grad_processors', 'ProxyOptimizer',
@@ -86,6 +87,7 @@ class PostProcessVariablesOptimizer(ProxyOptimizer):
         update_op = tf.group(update_op, *ops, name=name)
         return update_op
 
+    @contextmanager
     def _maybe_colocate(self, var):
         G = tf.get_default_graph()
         if self._colocate: