diff --git a/PaddleCV/image_classification/README_ngraph.md b/PaddleCV/image_classification/README_ngraph.md
index bb8190758d876244df931a090134f8410b6d38b3..09abb8c5c44c6340041b4339d4653529fdac596b 100644
--- a/PaddleCV/image_classification/README_ngraph.md
+++ b/PaddleCV/image_classification/README_ngraph.md
@@ -16,7 +16,7 @@ Only support Adam optimizer yet.
 Short description of aforementioned steps:
 
 ## 1. Install PaddlePaddle
-Follow PaddlePaddle [installation instruction](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification#installation) to install PaddlePaddle. If you [build from source](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/beginners_guide/install/compile/compile_Ubuntu_en.md), please use the following cmake arguments and ensure to set `-DWITH_NGRAPH=ON`. 
+Follow PaddlePaddle [installation instruction](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification#installation) to install PaddlePaddle. If you [build from source](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/beginners_guide/install/compile/compile_Ubuntu_en.md), please use the following cmake arguments and ensure to set `-DWITH_NGRAPH=ON`.
 ```
 cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_GPU=OFF -DWITH_MKL=ON -DWITH_MKLDNN=ON  -DWITH_NGRAPH=ON
 ```
@@ -35,9 +35,8 @@ export KMP_AFFINITY=granularity=fine,compact,1,0
 ```
 
 ## 3. How the benchmark script might be run.
-If everything built successfully, you can run command in ResNet50 nGraph session in script [run.sh](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleCV/image_classification/run.sh) to start the benchmark job locally. You will need to uncomment the `#ResNet50 nGraph` part of script.
+If everything built successfully, you can run command in ResNet50 nGraph session in script [run.sh](https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/image_classification/run.sh) to start the benchmark job locally. You will need to uncomment the `#ResNet50 nGraph` part of script.
 
 Above is training job using the nGraph, to run the inference job using the nGraph:
 
 Please download the pre-trained resnet50 model from [supported models](https://github.com/PaddlePaddle/models/tree/72dcc7c1a8d5de9d19fbd65b4143bd0d661eee2c/fluid/PaddleCV/image_classification#supported-models-and-performances) for inference script.
-
diff --git a/PaddleCV/image_classification/train.py b/PaddleCV/image_classification/train.py
index 7707bf4a723f8cba0f406acb8c4b280643c0ddb7..d53eefb2c93565c14e3511bdea023e34335eab0b 100644
--- a/PaddleCV/image_classification/train.py
+++ b/PaddleCV/image_classification/train.py
@@ -323,6 +323,7 @@ def train(args):
     train_py_reader.decorate_paddle_reader(train_reader)
     test_py_reader.decorate_paddle_reader(test_reader)
 
+    # use_ngraph is for CPU only, please refer to README_ngraph.md for details
     use_ngraph = os.getenv('FLAGS_use_ngraph')
     if not use_ngraph:
         train_exe = fluid.ParallelExecutor(