diff --git a/ppdet/modeling/heads/detr_head.py b/ppdet/modeling/heads/detr_head.py
index ec511654bd32939bb7e855fae396cefcfe9218e0..5b55642e4d8297b286e8fb414162ddc0485d8dfa 100644
--- a/ppdet/modeling/heads/detr_head.py
+++ b/ppdet/modeling/heads/detr_head.py
@@ -16,13 +16,12 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import math
 import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
 from ppdet.core.workspace import register
 import pycocotools.mask as mask_util
-from ..initializer import *
+from ..initializer import linear_init_
 
 __all__ = ['DETRHead']
 
diff --git a/ppdet/modeling/initializer.py b/ppdet/modeling/initializer.py
index 0e9aef403d9475e109dfa65373b2a55fd24b720c..f7e83098b90fa930fcb1dfcce3961f7e0bfcd2ae 100644
--- a/ppdet/modeling/initializer.py
+++ b/ppdet/modeling/initializer.py
@@ -50,9 +50,7 @@ def _no_grad_normal_(tensor, mean=0., std=1.):
 
 def _no_grad_fill_(tensor, value=0.):
     with paddle.no_grad():
-        v = paddle.rand(shape=tensor.shape, dtype=tensor.dtype)
-        v[...] = value
-        tensor.set_value(v)
+        tensor.set_value(paddle.full_like(tensor, value, dtype=tensor.dtype))
     return tensor
 
 
diff --git a/ppdet/modeling/losses/detr_loss.py b/ppdet/modeling/losses/detr_loss.py
index 24eed2e78f8d5073384cc38672d64df67316c9ce..5a589d4a2b4dae5644dc8b8ecf6f839c68559bdb 100644
--- a/ppdet/modeling/losses/detr_loss.py
+++ b/ppdet/modeling/losses/detr_loss.py
@@ -21,7 +21,7 @@ import paddle.nn as nn
 import paddle.nn.functional as F
 from ppdet.core.workspace import register
 from .iou_loss import GIoULoss
-from ..transformers import bbox_cxcywh_to_xyxy, bbox_overlaps, sigmoid_focal_loss
+from ..transformers import bbox_cxcywh_to_xyxy, sigmoid_focal_loss
 
 __all__ = ['DETRLoss']
 
@@ -211,7 +211,7 @@ class DETRLoss(nn.Layer):
             num_gts = paddle.clip(
                 num_gts / paddle.distributed.get_world_size(), min=1).item()
         except:
-            num_gts = max(num_gts, 1)
+            num_gts = max(num_gts.item(), 1)
         total_loss = dict()
         total_loss.update(
             self._get_loss_class(logits[-1], gt_class, match_indices,
diff --git a/ppdet/modeling/transformers/detr_transformer.py b/ppdet/modeling/transformers/detr_transformer.py
index 9069ee8c4edb401be3b03c53a5e0369f7bcc4489..65eb1285561bdbdcf9ccbac824498548d5e63a79 100644
--- a/ppdet/modeling/transformers/detr_transformer.py
+++ b/ppdet/modeling/transformers/detr_transformer.py
@@ -23,8 +23,8 @@ import paddle.nn.functional as F
 from ppdet.core.workspace import register
 from ..layers import MultiHeadAttention, _convert_attention_mask
 from .position_encoding import PositionEmbedding
-from .utils import *
-from ..initializer import *
+from .utils import _get_clones
+from ..initializer import linear_init_, conv_init_, xavier_uniform_, normal_
 
 __all__ = ['DETRTransformer']
 
diff --git a/ppdet/modeling/transformers/utils.py b/ppdet/modeling/transformers/utils.py
index d8abad9fbd71166b5324e368f5c7d8f8a2fd3ecd..5756cfe859e2dfbf22e73abf9616b857a54919cf 100644
--- a/ppdet/modeling/transformers/utils.py
+++ b/ppdet/modeling/transformers/utils.py
@@ -54,5 +54,4 @@ def sigmoid_focal_loss(logit, label, normalizer=1.0, alpha=0.25, gamma=2.0):
     if alpha >= 0:
         alpha_t = alpha * label + (1 - alpha) * (1 - label)
         loss = alpha_t * loss
-    return loss.mean(1).sum() / normalizer if normalizer > 1. else loss.mean(
-        1).sum()
+    return loss.mean(1).sum() / normalizer