diff --git a/configs/anchor_free/fcos_dcn_r50_fpn_1x.yml b/configs/anchor_free/fcos_dcn_r50_fpn_1x.yml
index ff46e744e48aa52a0b5cd58a85eb829b0574c70d..7a4d770cd24188a57ba39430026545f0cd9a0d40 100644
--- a/configs/anchor_free/fcos_dcn_r50_fpn_1x.yml
+++ b/configs/anchor_free/fcos_dcn_r50_fpn_1x.yml
@@ -180,4 +180,3 @@ TestReader:
     use_padded_im_info: true
   batch_size: 1
   shuffle: false
-  
diff --git a/ppdet/modeling/ops.py b/ppdet/modeling/ops.py
index 1f758ec0d5ea7f4a1515e459ac70f4ff6bd24ef8..f4787cf7007665ec0291dd728e56ea73f678b290 100644
--- a/ppdet/modeling/ops.py
+++ b/ppdet/modeling/ops.py
@@ -29,8 +29,8 @@ __all__ = [
     'AnchorGenerator', 'AnchorGrid', 'DropBlock', 'RPNTargetAssign',
     'GenerateProposals', 'MultiClassNMS', 'BBoxAssigner', 'MaskAssigner',
     'RoIAlign', 'RoIPool', 'MultiBoxHead', 'SSDLiteMultiBoxHead',
-    'SSDOutputDecoder', 'RetinaTargetAssign', 'RetinaOutputDecoder',
-    'ConvNorm', 'DeformConvNorm', 'MultiClassSoftNMS', 'LibraBBoxAssigner'
+    'SSDOutputDecoder', 'RetinaTargetAssign', 'RetinaOutputDecoder', 'ConvNorm',
+    'DeformConvNorm', 'MultiClassSoftNMS', 'LibraBBoxAssigner'
 ]
 
 
diff --git a/ppdet/optimizer.py b/ppdet/optimizer.py
index 21920bb66e168ed00518a6bb799f12670b8fdb89..5f7cfefabcacd0a5412d49d65d177458d1e2b713 100644
--- a/ppdet/optimizer.py
+++ b/ppdet/optimizer.py
@@ -197,6 +197,10 @@ class OptimizerBuilder():
         self.optimizer = optimizer
 
     def __call__(self, learning_rate):
+        if self.clip_grad_by_norm is not None:
+            fluid.clip.set_gradient_clip(
+                clip=fluid.clip.GradientClipByGlobalNorm(
+                    clip_norm=self.clip_grad_by_norm))
         if self.regularizer:
             reg_type = self.regularizer['type'] + 'Decay'
             reg_factor = self.regularizer['factor']
diff --git a/tools/train.py b/tools/train.py
index 477fe669e1301c2c5faa914d3f1ff4d8e3f84c67..e2d21cf80b11ed0fb19e29d9709f87e68607f641 100644
--- a/tools/train.py
+++ b/tools/train.py
@@ -126,11 +126,7 @@ def main():
                     loss *= ctx.get_loss_scale_var()
                 lr = lr_builder()
                 optimizer = optim_builder(lr)
-                clip = None
-                if optim_builder.clip_grad_by_norm is not None:
-                    clip = fluid.clip.GradientClipByGlobalNorm(
-                        clip_norm=optim_builder.clip_grad_by_norm)
-                optimizer.minimize(loss, grad_clip=clip)
+                optimizer.minimize(loss)
 
                 if FLAGS.fp16:
                     loss /= ctx.get_loss_scale_var()