diff --git a/ppdet/optimizer.py b/ppdet/optimizer.py
index c476e2edb7b01a64795e099b4e3de1dad6141841..5334eba724d9b7ecfd5ec080ba58a02c20bb6434 100644
--- a/ppdet/optimizer.py
+++ b/ppdet/optimizer.py
@@ -249,6 +249,8 @@ 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)