why is loss reduced when distributed training

[LukeCho-8810]

Hi, thanks for your great work.

I dont know why loss is reduced explicitly when distributed training. Is this not be done in pytorch ddp

mmdetection/mmdet/models/detectors/base.py

Line 214 in ca11860

if dist.is_available() and dist.is_initialized():

[jbwang1997]

Sorry for not getting your question. mmdetection is built on pytorch. We directly use pytorch DDP in distributed training.