fix detach bug (#3912)

configs/dota/README.md

@@ -121,7 +121,7 @@ python3.7 tools/infer.py -c configs/dota/s2anet_1x_spine.yml -o weights=output/s
 ### 5. DOTA数据评估
 执行如下命令，会在`output`文件夹下将每个图像预测结果保存到同文件夹名的txt文本中。
 ```
-python3.7 tools/infer.py -c configs/dota/s2anet_1x_dota.yml -o weights=./weights/s2anet_1x_dota.pdparams  --infer_dir=dota_test_images --draw_threshold=0.05 --save_txt=True --output_dir=output
+python3.7 tools/infer.py -c configs/dota/s2anet_alignconv_2x_dota.yml -o weights=./weights/s2anet_alignconv_2x_dota.pdparams  --infer_dir=dota_test_images --draw_threshold=0.05 --save_txt=True --output_dir=output
 ```
 
 请参考[DOTA_devkit](https://github.com/CAPTAIN-WHU/DOTA_devkit) 生成评估文件，评估文件格式请参考[DOTA Test](http://captain.whu.edu.cn/DOTAweb/tasks.html) ，生成zip文件，每个类一个txt文件，txt文件中每行格式为：`image_id score x1 y1 x2 y2 x3 y3 x4 y4`，提交服务器进行评估。
@@ -144,6 +144,7 @@ Paddle中`multiclass_nms`算子的输入支持四边形输入，因此部署时
 
 部署教程请参考[预测部署](../../deploy/README.md)
 
+**注意：** 由于paddle.detach函数动转静时会导致导出模型尺寸错误，因此在配置文件中增加了`is_training`参数，导出模型预测部署时需要将改参数设置为`False`
 
 ## Citations
 ```

configs/dota/s2anet_1x_spine.yml

@@ -27,3 +27,4 @@ S2ANetHead:
   reg_loss_weight: [1.0, 1.0, 1.0, 1.0, 1.05]
   cls_loss_weight: [1.05, 1.0]
   reg_loss_type: 'l1'
+  is_training: True

configs/dota/s2anet_alignconv_2x_dota.yml

@@ -24,3 +24,4 @@ S2ANetHead:
   reg_loss_weight: [1.0, 1.0, 1.0, 1.0, 1.05]
   cls_loss_weight: [1.05, 1.0]
   reg_loss_type: 'l1'
+  is_training: True

configs/dota/s2anet_conv_2x_dota.yml

@@ -21,3 +21,4 @@ S2ANetHead:
   use_sigmoid_cls: True
   reg_loss_weight: [1.0, 1.0, 1.0, 1.0, 1.1]
   cls_loss_weight: [1.1, 1.05]
+  is_training: True

ppdet/modeling/heads/s2anet_head.py

@@ -228,9 +228,10 @@ class S2ANetHead(nn.Layer):
                  align_conv_size=3,
                  use_sigmoid_cls=True,
                  anchor_assign=RBoxAssigner().__dict__,
-                 reg_loss_weight=[1.0, 1.0, 1.0, 1.0, 1.0],
-                 cls_loss_weight=[1.0, 1.0],
-                 reg_loss_type='l1'):
+                 reg_loss_weight=[1.0, 1.0, 1.0, 1.0, 1.1],
+                 cls_loss_weight=[1.1, 1.05],
+                 reg_loss_type='l1',
+                 is_training=True):
         super(S2ANetHead, self).__init__()
         self.stacked_convs = stacked_convs
         self.feat_in = feat_in
@@ -256,6 +257,7 @@ class S2ANetHead(nn.Layer):
         self.alpha = 1.0
         self.beta = 1.0
         self.reg_loss_type = reg_loss_type
+        self.is_training = is_training
 
         self.s2anet_head_out = None
 
@@ -446,10 +448,12 @@ class S2ANetHead(nn.Layer):
             init_anchors = self.rect2rbox(init_anchors)
             self.base_anchors_list.append(init_anchors)
 
-            fam_reg1 = fam_reg
+            if self.is_training:
+                refine_anchor = self.bbox_decode(fam_reg.detach(), init_anchors)
+            else:
+                fam_reg1 = fam_reg.clone()
                 fam_reg1.stop_gradient = True
                 refine_anchor = self.bbox_decode(fam_reg1, init_anchors)
-            #refine_anchor = self.bbox_decode(fam_reg.detach(), init_anchors)
 
             self.refine_anchor_list.append(refine_anchor)
 
@@ -615,19 +619,13 @@ class S2ANetHead(nn.Layer):
                 iou = rbox_iou(fam_bbox_decode, bbox_gt_bboxes)
                 iou = paddle.diag(iou)
 
-                if reg_loss_type == 'iou':
-                    EPS = paddle.to_tensor(
-                        1e-8, dtype='float32', stop_gradient=True)
-                    iou_factor = -1.0 * paddle.log(iou + EPS) / (fam_bbox + EPS)
-                    iou_factor.stop_gradient = True
-                    #fam_bbox = fam_bbox * iou_factor
-                elif reg_loss_type == 'gwd':
+                if reg_loss_type == 'gwd':
                     bbox_gt_bboxes_level = bbox_gt_bboxes[st_idx:st_idx +
                                                           feat_anchor_num, :]
                     fam_bbox_total = self.gwd_loss(fam_bbox_decode,
                                                    bbox_gt_bboxes_level)
                     fam_bbox_total = fam_bbox_total * feat_bbox_weights
-                    fam_bbox_total = paddle.sum(fam_bbox_total)
+                    fam_bbox_total = paddle.sum(fam_bbox_total) / num_total_samples
 
             fam_bbox_losses.append(fam_bbox_total)
             st_idx += feat_anchor_num
@@ -735,19 +733,13 @@ class S2ANetHead(nn.Layer):
                 iou = rbox_iou(odm_bbox_decode, bbox_gt_bboxes)
                 iou = paddle.diag(iou)
 
-                if reg_loss_type == 'iou':
-                    EPS = paddle.to_tensor(
-                        1e-8, dtype='float32', stop_gradient=True)
-                    iou_factor = -1.0 * paddle.log(iou + EPS) / (odm_bbox + EPS)
-                    iou_factor.stop_gradient = True
-                    # odm_bbox = odm_bbox * iou_factor
-                elif reg_loss_type == 'gwd':
+                if reg_loss_type == 'gwd':
                     bbox_gt_bboxes_level = bbox_gt_bboxes[st_idx:st_idx +
                                                           feat_anchor_num, :]
                     odm_bbox_total = self.gwd_loss(odm_bbox_decode,
                                                    bbox_gt_bboxes_level)
                     odm_bbox_total = odm_bbox_total * feat_bbox_weights
-                    odm_bbox_total = paddle.sum(odm_bbox_total)
+                    odm_bbox_total = paddle.sum(odm_bbox_total) / num_total_samples
 
             odm_bbox_losses.append(odm_bbox_total)
             st_idx += feat_anchor_num