CIOULoss is Nan

After using ciouloss in yolox, ciouloss is nan，ciouloss code is as follows

elif self.loss_type == 'ciou':
c_tl = torch.min((pred[:, :2] - pred[:, 2:] / 2), (target[:, :2] - target[:, 2:] / 2))
c_br = torch.max((pred[:, :2] + pred[:, 2:] / 2), (target[:, :2] + target[:, 2:] / 2))
w_c = (c_br - c_tl)[:, 0]
h_c = (c_br - c_tl)[:, 1]
c = w_c ** 2 + h_c ** 2
w_d = (pred[:, :2] - target[:, :2])[:, 0]
h_d = (pred[:, :2] - target[:, :2])[:, 1]
d = w_d ** 2 + h_d ** 2
diou = iou - d / c.clamp(1e-16)
w_gt = target[:, 2]
h_gt = target[:, 3]
w = pred[:, 2]
h = pred[:, 3]
with torch.no_grad():
arctan = torch.atan(w_gt / h_gt.clamp(1e-16)) - torch.atan(w / h.clamp(1e-16))
v = (4 / (math.pi ** 2)) * torch.pow(arctan, 2)
s = 1 - iou
alpha = v / (s + v).clamp(1e-16)
ciou = diou - alpha * v
loss = 1 - ciou.clamp(min=-1.0, max=1.0)

Start by reproducing the NaN with the CIoU code shown in the issue and inspect the prediction and target values entering the calculation. Trace the first non-finite intermediate through the width, height, IoU, and distance terms. Done means the cause is identified and CIoU produces finite results for the reported YOLOX use case.