bogcn_distributed_singlecard.py

'''
To run search algorithm on 8 gpus separately.
This reduces communication time between gpus comparing to DataParallel and Distributed Dataparallel
To search on a dataset such as imagenet, which requires more memory, DDP is expected to be implemented
'''
import subprocess
import pickle
import os
from settings import local_root_dir
from opendomain_utils.ioutils import create_dirs
create_dirs()
python_path = 'python' #path of python command
gpu_num = 8
tasks = []
for gpu in range(gpu_num):
    cmd = [python_path, os.path.join(local_root_dir, 'BOGCN_opendomain.py'), f"--gpu={gpu}"]
    train = subprocess.Popen(cmd)
    tasks.append(train)
for task in tasks:
    task.wait()
print(f"Num of Tasks finished:{len(tasks)}")