Any improvement?

[John1231983]

Thanks for sharing the repo. Do you achieve any improvement in comparison with training student only?

[John1231983]

If I have detect hand, so the number of class will be 1 or 2? Am I right to change it to support hand detection?

            c_obj_scale = t_obj_scale.unsqueeze(-1).repeat(1,1,1,1,2)
            # if dist_loss == "kl":
            kl_loss = DclsLoss(F.log_softmax(pi[..., 5:]/T,dim=-1), F.softmax(t_pi[..., 5:]/T, dim=-1))*(T*T)

[xiao-hua-sheng]

Thanks for sharing the repo. Do you achieve any improvement in comparison with training student only?

mAP will increase by 1-2%

If I have detect hand, so the number of class will be 1 or 2? Am I right to change it to support hand detection?

            c_obj_scale = t_obj_scale.unsqueeze(-1).repeat(1,1,1,1,2)
            # if dist_loss == "kl":
            kl_loss = DclsLoss(F.log_softmax(pi[..., 5:]/T,dim=-1), F.softmax(t_pi[..., 5:]/T, dim=-1))*(T*T)

right

[John1231983]

In your code, you setup it as 5 as number of classes. Could you list 5 classes? Do you count background?

[972821054]

@John1231983 @xiao-hua-sheng Did you solve this problem? I have the same problem, do I need to count the background class?