The problem of gradient overflow during ImageNet training

[erhul]

Why does gradient overflow occur? It seems that it doesn't affect the experimental results. Have you encountered this situation during training? How did you solve it?