CIOU loss gives NaN values

[notabee]

I am using CIOU for training where there can be multiple boxes at once, and I am getting NaN by your compute_ciou() loss function.

[Zzh-tju]

Do u use this repo？or the others？what is the learning rate do u use？

[WWW2323]

hello,i am also getting in trouble with the NaN loss ，How can i replace step size 1/(w^2+h^2) by 1 in my pytorch code as the paper said， “And thus in our implementation, the dominator w^ 2 +h^ 2 is simply removed for stable convergence, by which the step size 1/(w^2+h^2) is replaced by 1 and the gradient direc-tion is still consistent with Eqn. (12)”

[Zzh-tju]

utilize with torch no grad

[WWW2323]

wrap what with torch no grad? forgive me for being a rookie , orz is there any code about this problem in your project? thanks very much

[Zzh-tju]

@WXP2314

......
u = d / c
v = torch.atan(w_gt / h_gt) - torch.atan(w_pred / h_pred)
with torch.no_grad():
    w_pred1 = w_pred * v
    h_pred1 = h_pred * v
ciou_loss =(1-iou+u+8/(math.pi**2) * (w_pred1*h_pred- h_pred1*w_pred)).sum()

[jiangnanLZP]

Hello, which file the ciou_loss code in?

[Zzh-tju]

@jiangnanLZP compute_iou function in lib/utils/net.py

[jiangnanLZP]

@Zzh-tju thank you

[henbucuoshanghai]

w_pred1 = w_pred * v h_pred1 = h_pred * v ciou_loss =(1-iou+u+8/(math.pi**2) * (w_pred1h_pred- h_pred1w_pred)).sum()

where is this code???