diff --git a/PaddleCV/face_detection/README_cn.md b/PaddleCV/face_detection/README_cn.md
index 75df49f1582559fc2d3cfdd1235b1d4c49d46cd0..3329a914af4fd12ae943fdc09ba7a6a15995db77 100644
--- a/PaddleCV/face_detection/README_cn.md
+++ b/PaddleCV/face_detection/README_cn.md
@@ -111,6 +111,7 @@ python -u train.py --batch_size=16 --pretrained_model=vgg_ilsvrc_16_fc_reduced
 **注意**：
   - 本次开源模型中CPM模块与论文中有些许不同，相比论文中CPM模块训练和测试速度更快。
   - Pyramid Anchors模块的body部分可以针对不同情况，进行相应的长宽设置来调参。同时face、head、body部分的loss对应的系数也可以通过调参优化。
+  - 针对GPU资源有限时，可采取显存优化策略，`train.py` 与 `derface_eval.py` 已经加入部分显存优化FLAGS，详情更多内容请参考[API文档](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/flags/memory_cn.html#flags-fraction-of-gpu-memory-to-use)
 
 
 ### 模型评估
diff --git a/PaddleCV/face_detection/widerface_eval.py b/PaddleCV/face_detection/widerface_eval.py
index 8ce7c3f37dca5399087b2311a02cb455b99a0bdf..27a0d28562ff3172517d2ce5e679f1ad18d72927 100644
--- a/PaddleCV/face_detection/widerface_eval.py
+++ b/PaddleCV/face_detection/widerface_eval.py
@@ -22,6 +22,19 @@ import argparse
 import functools
 from PIL import Image
 
+def set_paddle_flags(**kwargs):
+    for key, value in kwargs.items():
+        if os.environ.get(key, None) is None:
+            os.environ[key] = str(value)
+
+# NOTE(paddle-dev): All of these flags should be
+# set before `import paddle`. Otherwise, it would
+# not take any effect.
+set_paddle_flags(
+    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
+)
+
+
 import paddle.fluid as fluid
 import reader
 from pyramidbox import PyramidBox
@@ -324,6 +337,7 @@ if __name__ == '__main__':
             is_infer=True)
         infer_program, nmsed_out = network.infer(main_program)
         fetches = [nmsed_out]
+        exe.run(startup_program)
         fluid.io.load_persistables(
             exe, args.model_dir, main_program=infer_program)
         # save model and program