when training, the loss will be nan in a while

[fnzhan]

I continue to train in another dataset using your pretrained model.
I set a very small learning rate and batch size, but the loss often became inf in trainin, and later the loss will be nan.
Can you provide your training strategy? @bear63 thank you.