add mainKD for ppyoloe distill (#7708)

* add mainkd for ppyoloe distill

* refine code style

configs/slim/distill/ppyoloe_plus_distill_m_distill_s.yml

@@ -43,6 +43,8 @@ DistillPPYOLOELoss: # M -> S
   loss_weight: {'logits': 4.0, 'feat': 1.0}
   logits_distill: True
   logits_loss_weight: {'class': 1.0, 'iou': 2.5, 'dfl': 0.5}
+  logits_ld_distill: True
+  logits_ld_params: {'weight': 20000, 'T': 10}
   feat_distill: True
   feat_distiller: 'fgd' # ['cwd', 'fgd', 'pkd', 'mgd', 'mimic']
   feat_distill_place: 'neck_feats'

ppdet/modeling/heads/ppyoloe_head.py

@@ -323,6 +323,7 @@ class PPYOLOEHead(nn.Layer):
                    assigned_bboxes, assigned_scores, assigned_scores_sum):
         # select positive samples mask
         mask_positive = (assigned_labels != self.num_classes)
+        self.distill_pairs['mask_positive_select'] = mask_positive
         num_pos = mask_positive.sum()
         # pos/neg loss
         if num_pos > 0:

ppdet/slim/distill_loss.py

@@ -212,6 +212,9 @@ class DistillPPYOLOELoss(nn.Layer):
             logits_loss_weight={'class': 1.0,
                                 'iou': 2.5,
                                 'dfl': 0.5},
+            logits_ld_distill=False,
+            logits_ld_params={'weight': 20000,
+                              'T': 10},
             feat_distill=True,
             feat_distiller='fgd',
             feat_distill_place='neck_feats',
@@ -222,6 +225,7 @@ class DistillPPYOLOELoss(nn.Layer):
         self.loss_weight_logits = loss_weight['logits']
         self.loss_weight_feat = loss_weight['feat']
         self.logits_distill = logits_distill
+        self.logits_ld_distill = logits_ld_distill
         self.feat_distill = feat_distill
 
         if logits_distill and self.loss_weight_logits > 0:
@@ -230,6 +234,10 @@ class DistillPPYOLOELoss(nn.Layer):
             self.qfl_loss_weight = logits_loss_weight['class']
             self.loss_bbox = GIoULoss()
 
+        if logits_ld_distill:
+            self.loss_kd = KnowledgeDistillationKLDivLoss(
+                loss_weight=logits_ld_params['weight'], T=logits_ld_params['T'])
+
         if feat_distill and self.loss_weight_feat > 0:
             assert feat_distiller in ['cwd', 'fgd', 'pkd', 'mgd', 'mimic']
             assert feat_distill_place in ['backbone_feats', 'neck_feats']
@@ -334,6 +342,20 @@ class DistillPPYOLOELoss(nn.Layer):
             loss_dfl = loss_dfl.mean(-1)
         return loss_dfl / 4.0  # 4 direction
 
+    def main_kd(self, mask_positive, pred_scores, soft_cls, num_classes):
+        num_pos = mask_positive.sum()
+        if num_pos > 0:
+            cls_mask = mask_positive.unsqueeze(-1).tile([1, 1, num_classes])
+            pred_scores_pos = paddle.masked_select(
+                pred_scores, cls_mask).reshape([-1, num_classes])
+            soft_cls_pos = paddle.masked_select(
+                soft_cls, cls_mask).reshape([-1, num_classes])
+            loss_kd = self.loss_kd(
+                pred_scores_pos, soft_cls_pos, avg_factor=num_pos)
+        else:
+            loss_kd = paddle.zeros([1])
+        return loss_kd
+
     def forward(self, teacher_model, student_model):
         teacher_distill_pairs = teacher_model.yolo_head.distill_pairs
         student_distill_pairs = student_model.yolo_head.distill_pairs
@@ -373,8 +395,15 @@ class DistillPPYOLOELoss(nn.Layer):
             distill_cls_loss = paddle.add_n(distill_cls_loss)
             distill_bbox_loss = paddle.add_n(distill_bbox_loss)
             distill_dfl_loss = paddle.add_n(distill_dfl_loss)
-
             logits_loss = distill_bbox_loss * self.bbox_loss_weight + distill_cls_loss * self.qfl_loss_weight + distill_dfl_loss * self.dfl_loss_weight
+
+            if self.logits_ld_distill:
+                loss_kd = self.main_kd(
+                    student_distill_pairs['mask_positive_select'],
+                    student_distill_pairs['pred_cls_scores'],
+                    teacher_distill_pairs['pred_cls_scores'],
+                    student_model.yolo_head.num_classes, )
+                logits_loss += loss_kd
         else:
             logits_loss = paddle.zeros([1])