gaodalii

I am using a single A800(80G), but I can only train it with batch_size=1, if I set batch_size=2, there also would be a cuda out of memory error.