diff --git a/configs/dota/_base_/s2anet_optimizer_1x.yml b/configs/dota/_base_/s2anet_optimizer_1x.yml
index 65f794dc34c55..91069a3bbace2 100644
--- a/configs/dota/_base_/s2anet_optimizer_1x.yml
+++ b/configs/dota/_base_/s2anet_optimizer_1x.yml
@@ -1,7 +1,7 @@
 epoch: 12
 
 LearningRate:
-  base_lr: 0.01
+  base_lr: 0.005
   schedulers:
   - !PiecewiseDecay
     gamma: 0.1
diff --git a/configs/dota/s2anet_1x_spine.yml b/configs/dota/s2anet_1x_spine.yml
index 017ee28e51f12..440ac143afbec 100644
--- a/configs/dota/s2anet_1x_spine.yml
+++ b/configs/dota/s2anet_1x_spine.yml
@@ -8,9 +8,9 @@ _BASE_: [
 
 weights: output/s2anet_1x_spine/model_final
 
-# for 8 card
+# for 4 card
 LearningRate:
-  base_lr: 0.01
+  base_lr: 0.005
 
 S2ANetHead:
   anchor_strides: [8, 16, 32, 64, 128]
@@ -26,5 +26,4 @@ S2ANetHead:
   use_sigmoid_cls: True
   reg_loss_weight: [1.0, 1.0, 1.0, 1.0, 1.05]
   cls_loss_weight: [1.05, 1.0]
-  reg_loss_type: gwd
-  use_paddle_anchor: False
+  reg_loss_type: 'l1'
diff --git a/configs/dota/s2anet_alignconv_2x_dota.yml b/configs/dota/s2anet_alignconv_2x_dota.yml
index 7355a0122aa5a..48e2d2b38a365 100644
--- a/configs/dota/s2anet_alignconv_2x_dota.yml
+++ b/configs/dota/s2anet_alignconv_2x_dota.yml
@@ -23,4 +23,4 @@ S2ANetHead:
   use_sigmoid_cls: True
   reg_loss_weight: [1.0, 1.0, 1.0, 1.0, 1.05]
   cls_loss_weight: [1.05, 1.0]
-  #reg_loss_type: 'l1' # 'l1' 'gwd'
+  reg_loss_type: 'l1'
diff --git a/ppdet/engine/export_utils.py b/ppdet/engine/export_utils.py
index 0fe932af447f8..a2c19c8c5fa7e 100644
--- a/ppdet/engine/export_utils.py
+++ b/ppdet/engine/export_utils.py
@@ -136,10 +136,5 @@ def _dump_infer_config(config, path, image_shape, model):
     infer_cfg['Preprocess'], infer_cfg['label_list'] = _parse_reader(
         reader_cfg, dataset_cfg, config['metric'], label_arch, image_shape)
 
-    if infer_arch == 'S2ANet':
-        # TODO: move background to num_classes
-        if infer_cfg['label_list'][0] != 'background':
-            infer_cfg['label_list'].insert(0, 'background')
-
     yaml.dump(infer_cfg, open(path, 'w'))
     logger.info("Export inference config file to {}".format(os.path.join(path)))
diff --git a/ppdet/modeling/heads/s2anet_head.py b/ppdet/modeling/heads/s2anet_head.py
index d578bbef4c02c..0245923727284 100644
--- a/ppdet/modeling/heads/s2anet_head.py
+++ b/ppdet/modeling/heads/s2anet_head.py
@@ -69,7 +69,7 @@ def gen_base_anchors(self):
         return base_anchors
 
     def _meshgrid(self, x, y, row_major=True):
-        yy, xx = paddle.meshgrid(x, y)
+        yy, xx = paddle.meshgrid(y, x)
         yy = yy.reshape([-1])
         xx = xx.reshape([-1])
         if row_major:
@@ -264,7 +264,7 @@ def __init__(self,
         for anchor_base in self.anchor_base_sizes:
             self.anchor_generators.append(
                 S2ANetAnchorGenerator(anchor_base, anchor_scales,
-                                          anchor_ratios))
+                                      anchor_ratios))
 
         self.anchor_generators = nn.LayerList(self.anchor_generators)
         self.fam_cls_convs = nn.Sequential()
@@ -577,7 +577,6 @@ def get_fam_loss(self, fam_target, s2anet_head_out, reg_loss_type='gwd'):
             fam_cls_losses.append(fam_cls_total)
 
             # step3: regression loss
-            fam_bbox_pred = fam_reg_branch_list[idx]
             feat_bbox_targets = paddle.to_tensor(
                 feat_bbox_targets, dtype='float32', stop_gradient=True)
             feat_bbox_targets = paddle.reshape(feat_bbox_targets, [-1, 5])
@@ -585,8 +584,6 @@ def get_fam_loss(self, fam_target, s2anet_head_out, reg_loss_type='gwd'):
             fam_bbox_pred = fam_reg_branch_list[idx]
             fam_bbox_pred = paddle.squeeze(fam_bbox_pred, axis=0)
             fam_bbox_pred = paddle.reshape(fam_bbox_pred, [-1, 5])
-            fam_bbox = self.smooth_l1_loss(fam_bbox_pred, feat_bbox_targets)
-
             fam_bbox = self.smooth_l1_loss(fam_bbox_pred, feat_bbox_targets)
             loss_weight = paddle.to_tensor(
                 self.reg_loss_weight, dtype='float32', stop_gradient=True)
@@ -846,7 +843,7 @@ def get_bboxes(self, cls_score_list, bbox_pred_list, mlvl_anchors, nms_pre,
             bbox_pred = paddle.reshape(bbox_pred, [-1, 5])
             anchors = paddle.reshape(anchors, [-1, 5])
 
-            if nms_pre > 0 and scores.shape[0] > nms_pre:
+            if scores.shape[0] > nms_pre:
                 # Get maximum scores for foreground classes.
                 if use_sigmoid_cls:
                     max_scores = paddle.max(scores, axis=1)
diff --git a/ppdet/modeling/post_process.py b/ppdet/modeling/post_process.py
index 95f51a9a80e37..d9bda6bd727d7 100644
--- a/ppdet/modeling/post_process.py
+++ b/ppdet/modeling/post_process.py
@@ -230,7 +230,7 @@ class S2ANetBBoxPostProcess(nn.Layer):
     def __init__(self, num_classes=15, nms_pre=2000, min_bbox_size=0, nms=None):
         super(S2ANetBBoxPostProcess, self).__init__()
         self.num_classes = num_classes
-        self.nms_pre = nms_pre
+        self.nms_pre = paddle.to_tensor(nms_pre)
         self.min_bbox_size = min_bbox_size
         self.nms = nms
         self.origin_shape_list = []