fix #1002

tensorpack/dataflow/imgaug/base.py

@@ -49,8 +49,7 @@ class Augmentor(object):
         This can be used in, e.g. augmentation for image, masks, keypoints altogether.
 
         Returns:
-            augmented data
-            augmentation params: can be any type
+            (augmented data, augmentation params)
         """
         return self._augment_return_params(d)
 

tensorpack/train/trainers.py

@@ -165,6 +165,7 @@ class SyncMultiGPUTrainerReplicated(SingleCostTrainer):
                 Supported values: ['nccl', 'hierarchical', 'cpu'].
                 Default to pick automatically by heuristics.
                 These modes may have slight (within 5%) differences in speed.
+                "hierarchical" mode was designed for DGX-like 8GPU machines.
             use_nccl: deprecated option
         """
         self.devices = gpus
@@ -173,7 +174,7 @@ class SyncMultiGPUTrainerReplicated(SingleCostTrainer):
             mode = 'nccl' if use_nccl else None
             log_deprecated("use_nccl option", "Use the `mode` option instead!", "2019-01-31")
         if mode is None:
-            mode = 'hierarchical' if len(gpus) >= 8 else 'nccl'
+            mode = 'hierarchical' if len(gpus) == 8 else 'nccl'
         mode = mode.lower()
 
         self._builder = SyncMultiGPUReplicatedBuilder(gpus, average, mode)