About train time

Open xLuge opened this issue 4 years ago • 1 comments

I train deeplab on 4gpus，each gpu 16g, but Each EPOCH training time takes more than an hour.Is this normal?

Mar 25 '21 07:03 xLuge

The training using DataParallel is really slow (due to syncbn). You may try using train_dist.py, which is much faster.

Mar 25 '21 20:03 zhanghang1989