def main():
args = Args.from_args()
if args.debug:
pass
elif args.world_size < 2:
warnings.warn('World size must be larger than 1')
exit()
if args.seed is not None:
utils.reproduction.initialize_seed(args.seed)
utils.environment.ulimit_n_max()
# Run on main process to avoid conflict
args.resolve_continue()
args.make_run_dir()
args.save()
utils.pack_code(args.run_dir)
free_port = utils.distributed.find_free_port()
dist_url = f'tcp://127.0.0.1:{free_port}'
print(f'world_size={args.world_size} Using dist_url={dist_url}')
args.parser = None
# Only single node distributed training is supported
mp.spawn(main_worker, args=(
args,
dist_url,
), nprocs=args.world_size)
def main():
args = Args.from_args()
if args.seed is not None:
utils.reproduction.initialize_seed(args.seed)
# run in main process for preventing concurrency conflict
args.resolve_continue()
args.make_run_dir()
args.save()
pack_code(args.run_dir)
utils.environment.ulimit_n_max()
free_port = utils.distributed.find_free_port()
dist_url = f'tcp://127.0.0.1:{free_port}'
print(f'world_size={args.world_size} Using dist_url={dist_url}')
"""
We only consider single node here. 'world_size' is the number of processes.
"""
args.parser = None
mp.spawn(main_worker, args=(
args,
dist_url,
), nprocs=args.world_size)