When the loss is NaN during training

[Sslishy]

I use batch size = 32 ir = 1e-5
who can help me solve this question?thx