diff --git a/ppdet/optimizer.py b/ppdet/optimizer.py
index dfb49e9cc447ed6cfdc20b7d50a85ded1d84cac8..1fcfbf856a9f0bb110607e2dfe0fd768d511a966 100644
--- a/ppdet/optimizer.py
+++ b/ppdet/optimizer.py
@@ -248,6 +248,8 @@ class ModelEMA(object):
         self.step += 1
 
     def apply(self):
+        if self.step == 0:
+            return self.state_dict
         state_dict = dict()
         for k, v in self.state_dict.items():
             v = v / (1 - self._decay**self.step)