diff --git a/slim/quantization/README.md b/slim/quantization/README.md
index 48644c1de708e5a84a5ed040abe5144eb7940e77..0db1e37b048e168b59e8edd7c4e025fbd9cf2ea0 100644
--- a/slim/quantization/README.md
+++ b/slim/quantization/README.md
@@ -67,6 +67,8 @@ python slim/quantization/train.py --not_quant_pattern yolo_output \
     pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar
 ```
 
+**注意：** 目前量化训练中只支持边训练边评估，所以训练时必需携带`--eval`参数，否则无法训练与保存模型。这个量化中的问题未来会持续改进。
+
 >通过命令行覆设置max_iters选项，因为量化的训练轮次比正常训练小很多，所以需要修改此选项。
 如果要调整训练卡数，可根据需要调整配置文件`yolov3_mobilenet_v1_voc.yml`中的以下参数：
 
diff --git a/slim/quantization/train.py b/slim/quantization/train.py
index f8a7d6faefc1c01ae1b7538facaee4cb7b2fad2b..429a92b083a1e5227481602c9bb0759bf71225c4 100644
--- a/slim/quantization/train.py
+++ b/slim/quantization/train.py
@@ -59,6 +59,10 @@ def load_global_step(exe, prog, path):
 
 
 def main():
+    if FLAGS.eval is False:
+        raise ValueError(
+            "Currently only supports `--eval==True` while training in `quantization`."
+        )
     env = os.environ
     FLAGS.dist = 'PADDLE_TRAINER_ID' in env and 'PADDLE_TRAINERS_NUM' in env
     if FLAGS.dist:
@@ -202,7 +206,6 @@ def main():
     if FLAGS.eval:
         # insert quantize op in eval_prog
         eval_prog = quant_aware(eval_prog, place, config, for_test=True)
-
         compiled_eval_prog = fluid.compiler.CompiledProgram(eval_prog)
 
     start_iter = 0