关于Variable-speed inference的疑问

[Yang Yahu]

作者您好：

通过(14)求出来的是分类为“hard”图像的比例，假如0.75，也就是75%的图像被分类为“hard”。那

计算的不应该是25%的分位数嘛，这样计算出来的阈值才有75%的图像的难度分数大于它呀。如果直接将0.75放进percentile函数，计算的好像不对啊

[Yang Yahu]

这里First和Second是不是弄反了呀，阈值越大，通过的“hard”图像越少的应该是Second啊

[Zhihao Lin]

DynamicDet/models/yolo.py

Lines 236 to 245 in 4dbb9b2

 if self.dynamic: 
     score = self.router([y[i] for i in self.router_ins]) # 'score' denotes the (1 - difficulty score) 
  
     if not hasattr(self, 'get_score'): 
         self.get_score = False 
     if self.get_score: 
         return score 
  
 need_second = self.training or (not self.dynamic) or score[:, 0] < self.dy_thres 
 need_first_head = self.training or (self.dynamic and score[:, 0] >= self.dy_thres) 

That is, 'score' in our code denotes the (1 - difficulty score). So, it is ok~

[Yang Yahu]

DynamicDet/models/yolo.py

Lines 236 to 245 in 4dbb9b2

 if self.dynamic: 
     score = self.router([y[i] for i in self.router_ins]) # 'score' denotes the (1 - difficulty score) 
  
     if not hasattr(self, 'get_score'): 
         self.get_score = False 
     if self.get_score: 
         return score 
  
 need_second = self.training or (not self.dynamic) or score[:, 0] < self.dy_thres 
 need_first_head = self.training or (self.dynamic and score[:, 0] >= self.dy_thres) 

That is, 'score' in our code denotes the (1 - difficulty score). So, it is ok~

作者您好：
score = self.router([y[i] for i in self.router_ins]) 计算出来的就是difficulty score啊，怎么会是1 - difficulty score呢？ @LZHgrla

[Zhihao Lin]

DynamicDet/models/yolo.py

Lines 244 to 245 in 4dbb9b2

 need_second = self.training or (not self.dynamic) or score[:, 0] < self.dy_thres 
 need_first_head = self.training or (self.dynamic and score[:, 0] >= self.dy_thres) 

分数足够低会选择second，分数足够高则会选择first。因此，这是由训练决定的，只要训练与推理统一即可。

[Yang Yahu]

DynamicDet/models/yolo.py

Lines 244 to 245 in 4dbb9b2

 need_second = self.training or (not self.dynamic) or score[:, 0] < self.dy_thres 
 need_first_head = self.training or (self.dynamic and score[:, 0] >= self.dy_thres) 

分数足够低会选择second，分数足够高则会选择first。因此，这是由训练决定的，只要训练与推理统一即可。

由这里决定？奖励的检测器乘的是1- difficulty score，所以拟合出来的score实际上是1- difficulty score？
那我把这里的score和1-score调换后，计算出来的score就表示difficulty score了吧？

[Zhihao Lin]

DynamicDet/models/yolo.py

Lines 244 to 245 in 4dbb9b2

 need_second = self.training or (not self.dynamic) or score[:, 0] < self.dy_thres 
 need_first_head = self.training or (self.dynamic and score[:, 0] >= self.dy_thres) 

除了loss要修改，模型的forward也需要修改（分数足够大则走second，足够小则走first）。修改后即可认为score是difficulty score

[Yang Yahu]

DynamicDet/models/yolo.py

Lines 244 to 245 in 4dbb9b2

 need_second = self.training or (not self.dynamic) or score[:, 0] < self.dy_thres 
 need_first_head = self.training or (self.dynamic and score[:, 0] >= self.dy_thres) 

除了loss要修改，模型的forward也需要修改（分数足够大则走second，足够小则走first）。修改后即可认为score是difficulty score

原来如此，感谢您的耐心解答。