def main():
mlperf_log.ROOT_DIR_MASKRCNN = os.path.dirname(os.path.abspath(__file__))
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Training")
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
if is_main_process:
# Setting logging file parameters for compliance logging
os.environ["COMPLIANCE_FILE"] = './MASKRCNN_complVv0.5.0_' + str(
datetime.datetime.now())
mlperf_log.LOG_FILE = os.getenv("COMPLIANCE_FILE")
mlperf_log._FILE_HANDLER = logging.FileHandler(mlperf_log.LOG_FILE)
mlperf_log._FILE_HANDLER.setLevel(logging.DEBUG)
mlperf_log.LOGGER.addHandler(mlperf_log._FILE_HANDLER)
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
print_mlperf(key=mlperf_log.RUN_START)
# setting seeds - needs to be timed, so after RUN_START
if is_main_process():
master_seed = random.SystemRandom().randint(0, 2**32 - 1)
seed_tensor = torch.tensor(master_seed,
dtype=torch.float32,
device=torch.device("cuda"))
else:
seed_tensor = torch.tensor(0,
dtype=torch.float32,
device=torch.device("cuda"))
torch.distributed.broadcast(seed_tensor, 0)
master_seed = int(seed_tensor.item())
else:
print_mlperf(key=mlperf_log.RUN_START)
# random master seed, random.SystemRandom() uses /dev/urandom on Unix
master_seed = random.SystemRandom().randint(0, 2**32 - 1)
# actually use the random seed
args.seed = master_seed
# random number generator with seed set to master_seed
random_number_generator = random.Random(master_seed)
print_mlperf(key=mlperf_log.RUN_SET_RANDOM_SEED, value=master_seed)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
output_dir = cfg.OUTPUT_DIR
if output_dir:
mkdir(output_dir)
logger = setup_logger("maskrcnn_benchmark", output_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
# generate worker seeds, one seed for every distributed worker
worker_seeds = generate_seeds(
random_number_generator,
torch.distributed.get_world_size()
if torch.distributed.is_initialized() else 1)
# todo sharath what if CPU
# broadcast seeds from rank=0 to other workers
worker_seeds = broadcast_seeds(worker_seeds, device='cuda')
# Setting worker seeds
logger.info("Worker {}: Setting seed {}".format(
args.local_rank, worker_seeds[args.local_rank]))
torch.manual_seed(worker_seeds[args.local_rank])
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
model = train(cfg, args.local_rank, args.distributed)
print_mlperf(key=mlperf_log.RUN_FINAL)
def main():
configure_logger(constants.MASKRCNN)
log_start(key=constants.INIT_START)
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Training")
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument("--local_rank",
type=int,
default=herring.get_local_rank())
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = herring.get_world_size()
args.distributed = num_gpus > 1
args.local_rank = herring.get_local_rank()
# if is_main_process:
# # Setting logging file parameters for compliance logging
# os.environ["COMPLIANCE_FILE"] = '/MASKRCNN_complVv0.5.0_' + str(datetime.datetime.now())
# constants.LOG_FILE = os.getenv("COMPLIANCE_FILE")
# constants._FILE_HANDLER = logging.FileHandler(constants.LOG_FILE)
# constants._FILE_HANDLER.setLevel(logging.DEBUG)
# constants.LOGGER.addHandler(constants._FILE_HANDLER)
if args.distributed:
torch.cuda.set_device(args.local_rank)
# setting seeds - needs to be timed, so after RUN_START
if is_main_process():
master_seed = random.SystemRandom().randint(0, 2**32 - 1)
seed_tensor = torch.tensor(master_seed,
dtype=torch.float32,
device=torch.device("cuda"))
else:
seed_tensor = torch.tensor(0,
dtype=torch.float32,
device=torch.device("cuda"))
herring.broadcast(seed_tensor, 0)
master_seed = int(seed_tensor.item())
else:
# random master seed, random.SystemRandom() uses /dev/urandom on Unix
master_seed = random.SystemRandom().randint(0, 2**32 - 1)
# actually use the random seed
args.seed = master_seed
# random number generator with seed set to master_seed
random_number_generator = random.Random(master_seed)
log_event(key=constants.SEED, value=master_seed)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
output_dir = cfg.OUTPUT_DIR
if output_dir:
mkdir(output_dir)
logger = setup_logger("maskrcnn_benchmark", output_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
# generate worker seeds, one seed for every distributed worker
worker_seeds = generate_seeds(random_number_generator,
herring.get_world_size())
# todo sharath what if CPU
# broadcast seeds from rank=0 to other workers
worker_seeds = broadcast_seeds(worker_seeds, device='cuda')
# Setting worker seeds
logger.info("Worker {}: Setting seed {}".format(
args.local_rank, worker_seeds[args.local_rank]))
torch.manual_seed(worker_seeds[args.local_rank])
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
# Initialise async eval
init()
model, success = train(cfg, args.local_rank, args.distributed,
random_number_generator)
if success is not None:
if success:
log_end(key=constants.RUN_STOP, metadata={"status": "success"})
else:
log_end(key=constants.RUN_STOP, metadata={"status": "aborted"})
def main():
mlperf_log.ROOT_DIR_MASKRCNN = os.path.dirname(os.path.abspath(__file__))
# mlperf_log.LOGGER.propagate = False
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Training")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch/configs/rpn_r50.py",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument('--skip-test',
dest='skip_test',
help='Do not test the model',
action='store_true')
parser.add_argument("--fp16",
action="store_true",
help="Enable multi-precision training")
parser.add_argument("--min_bbox_map",
type=float,
default=0.377,
help="Target BBOX MAP")
parser.add_argument("--min_mask_map",
type=float,
default=0.339,
help="Target SEGM/MASK MAP")
parser.add_argument("--seed", type=int, default=1, help="Seed")
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
args.distributed = (int(os.environ["WORLD_SIZE"]) > 1
if "WORLD_SIZE" in os.environ else False)
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
# to synchronize start of time
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
if torch.distributed.get_rank() == 0:
# Setting logging file parameters for compliance logging
os.environ["COMPLIANCE_FILE"] = '/MASKRCNN_complVv0.5.0_' + str(
datetime.datetime.now())
mlperf_log.LOG_FILE = os.getenv("COMPLIANCE_FILE")
mlperf_log._FILE_HANDLER = logging.FileHandler(mlperf_log.LOG_FILE)
mlperf_log._FILE_HANDLER.setLevel(logging.DEBUG)
mlperf_log.LOGGER.addHandler(mlperf_log._FILE_HANDLER)
print_mlperf(key=mlperf_log.RUN_START)
# Setting seed
seed_tensor = torch.tensor(0,
dtype=torch.float32,
device=torch.device("cuda"))
if torch.distributed.get_rank() == 0:
# seed = int(time.time())
# random master seed, random.SystemRandom() uses /dev/urandom on Unix
master_seed = random.SystemRandom().randint(0, 2**32 - 1)
seed_tensor = torch.tensor(master_seed,
dtype=torch.float32,
device=torch.device("cuda"))
torch.distributed.broadcast(seed_tensor, 0)
master_seed = int(seed_tensor.item())
else:
# Setting logging file parameters for compliance logging
os.environ["COMPLIANCE_FILE"] = '/MASKRCNN_complVv0.5.0_' + str(
datetime.datetime.now())
mlperf_log.LOG_FILE = os.getenv("COMPLIANCE_FILE")
mlperf_log._FILE_HANDLER = logging.FileHandler(mlperf_log.LOG_FILE)
mlperf_log._FILE_HANDLER.setLevel(logging.DEBUG)
mlperf_log.LOGGER.addHandler(mlperf_log._FILE_HANDLER)
print_mlperf(key=mlperf_log.RUN_START)
# random master seed, random.SystemRandom() uses /dev/urandom on Unix
master_seed = random.SystemRandom().randint(0, 2**32 - 1)
args.seed = master_seed
# random number generator with seed set to master_seed
random_number_generator = random.Random(master_seed)
print_mlperf(key=mlperf_log.RUN_SET_RANDOM_SEED, value=master_seed)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
if args.skip_test:
cfg.DO_ONLINE_MAP_EVAL = False
cfg.freeze()
output_dir = cfg.OUTPUT_DIR
if output_dir:
mkdir(output_dir)
#logger = setup_logger("maskrcnn_benchmark", output_dir, args.local_rank)
logger = setup_logger("maskrcnn_benchmark", None, args.local_rank)
logger.info(args)
# generate worker seeds, one seed for every distributed worker
worker_seeds = generate_seeds(
random_number_generator,
torch.distributed.get_world_size()
if torch.distributed.is_initialized() else 1)
# todo sharath what if CPU
# broadcast seeds from rank=0 to other workers
worker_seeds = broadcast_seeds(worker_seeds, device='cuda')
# Setting worker seeds
logger.info("Worker {}: Setting seed {}".format(
args.local_rank, worker_seeds[args.local_rank]))
torch.manual_seed(worker_seeds[args.local_rank])
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
model = train(cfg, random_number_generator, args.local_rank,
args.distributed, args, args.fp16)
print_mlperf(key=mlperf_log.RUN_FINAL)
def main():
configure_logger(constants.MASKRCNN)
log_start(key=constants.INIT_START)
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Training")
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
# setting seeds - needs to be timed, so after RUN_START
if is_main_process():
master_seed = random.SystemRandom().randint(0, 2**32 - 1)
seed_tensor = torch.tensor(master_seed,
dtype=torch.float32,
device=torch.device("cuda"))
else:
seed_tensor = torch.tensor(0,
dtype=torch.float32,
device=torch.device("cuda"))
torch.distributed.broadcast(seed_tensor, 0)
master_seed = int(seed_tensor.item())
else:
# random master seed, random.SystemRandom() uses /dev/urandom on Unix
master_seed = random.SystemRandom().randint(0, 2**32 - 1)
# actually use the random seed
args.seed = master_seed
# random number generator with seed set to master_seed
random_number_generator = random.Random(master_seed)
log_event(key=constants.SEED, value=master_seed)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
output_dir = cfg.OUTPUT_DIR
if output_dir:
mkdir(output_dir)
logger = setup_logger("maskrcnn_benchmark", output_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
# generate worker seeds, one seed for every distributed worker
worker_seeds = generate_seeds(
random_number_generator,
torch.distributed.get_world_size()
if torch.distributed.is_initialized() else 1)
# todo sharath what if CPU
# broadcast seeds from rank=0 to other workers
worker_seeds = broadcast_seeds(worker_seeds, device='cuda')
# Setting worker seeds
logger.info("Worker {}: Setting seed {}".format(
args.local_rank, worker_seeds[args.local_rank]))
torch.manual_seed(worker_seeds[args.local_rank])
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
model, success = train(cfg, args.local_rank, args.distributed,
args.disable_allreduce_for_logging,
random_number_generator)
if success is not None:
if success:
log_end(key=constants.RUN_STOP, metadata={"status": "success"})
else:
log_end(key=constants.RUN_STOP, metadata={"status": "aborted"})