check tf version in multigpu.

tensorpack/libinfo.py

@@ -10,4 +10,4 @@ os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'  # issue#9339
 os.environ['TF_AUTOTUNE_THRESHOLD'] = '3'   # use more warm-up
 os.environ['TF_AVGPOOL_USE_CUDNN'] = '1'   # issue#8566
 
-__version__ = '0.1.9'
+__version__ = '0.2.0'

tensorpack/tfutils/common.py

@@ -10,7 +10,6 @@ from ..utils.argtools import graph_memoized
 from ..utils.naming import GLOBAL_STEP_OP_NAME
 
 __all__ = ['get_default_sess_config',
-
            'get_global_step_value',
            'get_global_step_var',
            'get_op_tensor_name',

tensorpack/train/multigpu.py

@@ -27,6 +27,11 @@ __all__ = ['MultiGPUTrainerBase', 'SyncMultiGPUTrainer',
            'SyncMultiGPUTrainerParameterServer']
 
 
+def _check_tf_version():
+    ver = float('.'.join(tf.VERSION.split('.')[:2]))
+    assert ver >= 1.1, "TF version {} is too old to run multi GPU training!".format(ver)
+
+
 def apply_prefetch_policy(config, use_stage=True):
     if config.data is None and config.dataflow is not None:
         config.data = QueueInput(config.dataflow)
@@ -55,6 +60,8 @@ class MultiGPUTrainerBase(Trainer):
             List of outputs of ``func``, evaluated on each tower.
         """
         logger.info("Training a model of {} tower".format(len(towers)))
+        if len(towers) > 1:
+            _check_tf_version()
 
         ret = []
         if devices is not None:

tensorpack/train/predict.py

@@ -3,7 +3,6 @@
 # File: predict.py
 # Author: Yuxin Wu <ppwwyyxxc@gmail.com>
 
-import tensorflow as tf
 from ..predict import (OnlinePredictor,
                        PredictorTowerBuilder)
 
@@ -34,9 +33,8 @@ class PredictorFactory(object):
             an online predictor (which has to be used under a default session)
         """
         tower = self.towers[tower]
-        with tf.variable_scope(tf.get_variable_scope(), reuse=True):
-            # just ensure the tower exists. won't rebuild
-            self._tower_builder.build(tower)
+        # just ensure the tower exists. won't rebuild (memoized)
+        self._tower_builder.build(tower)
 
         placeholder_names = set([k.name for k in self.model.get_inputs_desc()])
         get_tensor_fn = PredictorTowerBuilder.get_tensors_maybe_in_tower