GPU utilization rate and distribution training

[byrsongyuxin]

when I use distribution training on GPUs, the process ids are the same and the GPU utilization rate is zero.

[byrsongyuxin]

CUDA_VISIBLE_DEVICES=0,1 python main.py
--root_path /dssg/syx/action-recognition/data
--video_path /dssg/syx/action-recognition/data/UCF101/UCF-101-jpg
--annotation_path /dssg/syx/action-recognition/data/UCF101/ucf101_01.json
--result_path /home/wuming/syx/action-recognition/3D-ResNets-PyTorch-master/result
--dataset ucf101 --n_classes 101 --n_pretrain_classes 700
--pretrain_path /home/wuming/syx/action-recognition/3D-ResNets-PyTorch-master/pretrain/r3d50_K_200ep.pth
--ft_begin_module fc
--model resnet --model_depth 50
--batch_size 128 --n_threads 4 --checkpoint 5
--batchnorm_sync --distributed