FloatingPointError: Loss became infinite or NaN at iteration=5593

[MishimaR]

It seems that the loss is too large. I used only 1 gpu to train and set batchsize to 1, and I did not change any other configs. I have tried to set learning rate to 1e-3 and 1e-4, but this error still happen.