in my training all loss is nan , I made data using pakage.py 
