how to resolve loss=-inf

[YHaooo-4508]

I tried to add variance prediction in my own loss, so I tried to reproduce step by step.
so, at the first step, i just added the var prediction in my loss function calculation(without nf_loss).

`loss_body_pose =((torch.log(body_pose_sigma/amp) + torch.abs(body_pose_gt - body_pose_pred)) / (math.sqrt(2) * body_pose_sigma + 1e-9))

loss_body_pose = loss_body_pose[has_gt==1].sum()

loss_all = loss_keypoints_3d + loss_keypoints_2d + loss_body_pose
`
In the first 200+ iterations，the loss_all>0 and it's was normally decreasing. Then suddenly it changed from 50 to a negative number, and became smaller and smaller until it became -inf.

I would like to know if this issue has also occurred in your experiment and how to solve it.