diff --git a/dygraph/bmn/README.md b/dygraph/bmn/README.md
index c2de8e8c9b0bab0c6c0e024faae472dc4e655b46..363287fcaf3b916708f0bf2bdf2e8de2128d6a8d 100644
--- a/dygraph/bmn/README.md
+++ b/dygraph/bmn/README.md
@@ -55,7 +55,7 @@ BMN的训练数据采用ActivityNet1.3提供的数据集，我们提供了处理
 
     bash run.sh
 
-若使用单卡训练，请将配置文件bmn.yaml中`TRAIN`和`VALID`对应的num\_gpus调整为1，启动方式如下:
+若使用单卡训练，请将配置文件bmn.yaml中`TRAIN`和`VALID`对应的num\_gpus调整为1，并将train.py文件最后一行的`nprocs`参数值设为1，启动方式如下:
 
     export CUDA_VISIBLE_DEVICES=0
     python train.py
diff --git a/dygraph/bmn/train.py b/dygraph/bmn/train.py
index 21591bb962d80b6224fd449d3b21dcf18aacb2df..bbe5783b15e38d9059544477b5a09eb0140ccc4e 100644
--- a/dygraph/bmn/train.py
+++ b/dygraph/bmn/train.py
@@ -262,4 +262,4 @@ def train_bmn(args):
 
 if __name__ == "__main__":
     args = parse_args()
-    dist.spawn(train_bmn, args=(args, ), nprocs=4)
+    dist.spawn(train_bmn, args=(args, ), nprocs=4)  #nprocs=1 when single card