Useful trainning tips: small bug for Python >= 3.5

[reverseuniverse]

For Python version >= 3.5, async becomes a keyword in Python. So the target.cuda(async=True) will cause error while running the scripts.

In order to fix this,
change line 276 and 325 in train.py
from

target = target.cuda(async=True)

to

target = target.cuda(non_blocking=True)

Also, remind to suitable code for different Pytorch version in train.py from line 327 to line 338