Default process group has not been initialized, please make sure to call init_process_group.

[sihan-shao]

When I run train.py, I get this problem. How can I slove it?

[chumingqian]

comment the line 28 in the train.py

following prepare for the distributed training

# if 'coco' in cfg.DATASETS.TRAIN[0]:
#    model = nn.SyncBatchNorm.convert_sync_batchnorm(model)

and modify this line: model = torch.nn.parallel.DataParallel(model.cuda(), device_ids=[args.local_rank], output_device=args.local_rank)