some question about focal loss in code

[liushuai839]

Hi,Guys! When I read the code, I found a problem that the two inputs of focal loss ,which are cls_pred an label , the dimension of them is not the same. e.g. the cls_pred is [nums_points, nums_class](nums_class is 80 in coco) but the label is [nums_points]. So I wonder that for the every point, the label that has the largest scores is the result???

[tianzhi0549]

@liushuai839 label contains the label num and will be converted to one-hot vectors.

[liushuai839]

Thank yiou for your reply. And I also want to know that If I can determine that certain sample is positive or nagative in training through the output of class_score subnet.Because I see that the output of class_score subnet is [num_points,num_classes]. It means that the output is the score of every sample coresponding to the every class. But I cannot make sure that it is positive or negative.