rewrite the model function to include device information

[zhangyu2ustc]

add averge_grad to class base_model

[zhangyu2ustc]

rewrite training function to include average_grad

[zhangyu2ustc]

to select the least loaded GPU for the script:
check this repo:
https://github.com/bamos/setGPU

add this line to your script:
'''
import setGPU
'''