分布式训练的项目出现问题：

nprocs=torch.cuda.device_count(),

/python3.7/site-packages/torch/multiprocessing/spawn.py", line 199, in spawn
    return start_processes(fn, args, nprocs, join, daemon, start_method='spawn')

torch/multiprocessing/spawn.py", line 157, in start_processes
    while not context.join():

/torch/multiprocessing/spawn.py", line 118, in join
    raise Exception(msg)

出现上述问题，暂时无法从根源上解决，所以采用仅使用一个gpu的策略：

export CUDA_VISIBLE_DEVICES=0

解除上述限制：

export CUDA_VISIBLE_DEVICES=0,1,2,3