training loss increase periodically

Hi Thanks for your work.
When I'm training the model I found that the train loss of some steps after _**eval_batch**_ would increase slightly and then return to normal.
Does anyone have similar problem?