use reuse context instead of reuse_variables()

tensorpack/callbacks/inference_runner.py

@@ -182,10 +182,9 @@ class FeedfreeInferenceRunner(Callback):
     def _setup_graph(self):
         self._find_input_tensors()  # tensors
 
-        tf.get_variable_scope().reuse_variables()
-
-        # overwrite the FeedfreeInferenceRunner scope
-        with tf.name_scope(None), \
+        # overwrite the FeedfreeInferenceRunner name scope
+        with tf.variable_scope(tf.get_variable_scope(), reuse=True), \
+                tf.name_scope(None), \
                 freeze_collection(SUMMARY_BACKUP_KEYS):
             def fn(_):
                 self.trainer.model.build_graph(self._input_tensors)

tensorpack/predict/base.py

@@ -152,11 +152,12 @@ def build_prediction_graph(build_tower_fn, towers=[0], prefix=''):
         prefix: an extra prefix in tower name. The final tower prefix will be
             determined by :meth:`TowerContext.get_predict_tower_name`.
     """
-    for k in towers:
+    for idx, k in enumerate(towers):
         logger.info(
             "Building prediction graph for towerid={} with prefix='{}' ...".format(k, prefix))
         towername = TowerContext.get_predict_tower_name(prefix, k)
         with tf.device('/gpu:{}'.format(k) if k >= 0 else '/cpu:0'), \
-                TowerContext(towername, is_training=False):
+                TowerContext(towername, is_training=False), \
+                tf.variable_scope(tf.get_variable_scope(),
+                                  reuse=True if idx > 0 else None):
             build_tower_fn(k)
-            tf.get_variable_scope().reuse_variables()

tensorpack/predict/multigpu.py

@@ -71,16 +71,17 @@ class DataParallelOfflinePredictor(OnlinePredictor):
             sess = tf.Session(config=config.session_config)
             input_var_names = []
             output_vars = []
-            for k in towers:
+            for idx, k in enumerate(towers):
                 towername = PREDICT_TOWER + str(k)
                 input_vars = config.model.build_placeholders(
                     prefix=towername + '-')
                 logger.info(
                     "Building graph for predictor tower {}...".format(k))
                 with tf.device('/gpu:{}'.format(k) if k >= 0 else '/cpu:0'), \
-                        TowerContext(towername, is_training=False):
+                        TowerContext(towername, is_training=False), \
+                        tf.variable_scope(tf.get_variable_scope(),
+                                          reuse=True if idx > 0 else None):
                     config.model.build_graph(input_vars)
-                    tf.get_variable_scope().reuse_variables()
                 input_var_names.extend([k.name for k in input_vars])
                 output_vars.extend(get_tensors_by_names(
                     [towername + '/' + n