bugfix in replicated training varscope

tensorpack/callbacks/saver.py

@@ -43,6 +43,7 @@ class ModelSaver(Callback):
         vars = []
         for key in self.var_collections:
             vars.extend(tf.get_collection(key))
+        vars = list(set(vars))
         self.path = os.path.join(self.checkpoint_dir, 'model')
         if get_tf_version_number() <= 1.1:
             self.saver = tf.train.Saver(

tensorpack/tfutils/tower.py

@@ -43,6 +43,8 @@ class TowerContext(object):
             assert self._name
             if vs_name is None:
                 self._vs_name = self._name
+            else:
+                self._vs_name = vs_name
         else:
             assert vs_name is None, "vs_name is only valid in 'replicated' mode!"
             self._vs_name = ''

tensorpack/train/multigpu.py

@@ -259,7 +259,7 @@ class SyncMultiGPUTrainerReplicated(MultiGPUTrainerBase, SingleCostFeedfreeTrain
             lambda: self._get_cost_and_grad()[1],
             var_strategy='replicated',
             # use no variable scope for the first tower
-            vs_names=[''] + [None] * self.config.nr_tower - 1)
+            vs_names=[''] + [None] * (self.config.nr_tower - 1))
         grads = self._allreduce_grads(grad_list)
 
         train_ops = []