def mlperf_test_early_exit(iteration, iters_per_epoch, tester, model,
distributed, min_bbox_map, min_segm_map):
# Note: let iters / epoch == 10k, at iter 9999 we've finished epoch 0 and need to test
if iteration > 0 and (iteration + 1) % iters_per_epoch == 0:
epoch = iteration // iters_per_epoch
log_end(key=constants.EPOCH_STOP, metadata={"epoch_num": epoch})
log_end(key=constants.BLOCK_STOP, metadata={"first_epoch_num": epoch})
log_start(key=constants.EVAL_START, metadata={"epoch_num": epoch})
bbox_map, segm_map = test_and_exchange_map(tester, model, distributed)
# necessary for correctness
model.train()
logger = logging.getLogger('maskrcnn_benchmark.trainer')
logger.info('bbox mAP: {}, segm mAP: {}'.format(bbox_map, segm_map))
log_event(key=constants.EVAL_ACCURACY,
value={
"BBOX": bbox_map,
"SEGM": segm_map
},
metadata={"epoch_num": epoch})
log_end(key=constants.EVAL_STOP, metadata={"epoch_num": epoch})
# terminating condition
if bbox_map >= min_bbox_map and segm_map >= min_segm_map:
logger.info("Target mAP reached, exiting...")
return True
return False
def mlperf_test_early_exit(iteration, iters_per_epoch, tester, model, distributed, min_bbox_map, min_segm_map):
# Note: let iters / epoch == 10k, at iter 9999 we've finished epoch 0 and need to test
if iteration > 0 and (iteration + 1)% iters_per_epoch == 0:
synchronize()
epoch = iteration // iters_per_epoch + 1
log_end(key=constants.EPOCH_STOP, metadata={"epoch_num": epoch})
log_end(key=constants.BLOCK_STOP, metadata={"first_epoch_num": epoch})
log_start(key=constants.EVAL_START, metadata={"epoch_num":epoch})
# set the async evaluator's tag correctly
set_epoch_tag(epoch)
# Note: No longer returns anything, underlying future is in another castle
tester(model=model, distributed=distributed)
# necessary for correctness
model.train()
else:
# Otherwise, check for finished async results
results = check_completed_tags()
# on master process, check each result for terminating condition
# sentinel for run finishing
finished = 0
if is_main_process():
for result_epoch, (bbox_map, segm_map) in results.items():
logger = logging.getLogger('maskrcnn_benchmark.trainer')
logger.info('bbox mAP: {}, segm mAP: {}'.format(bbox_map, segm_map))
log_event(key=constants.EVAL_ACCURACY, value={"BBOX" : bbox_map, "SEGM" : segm_map}, metadata={"epoch_num" : result_epoch} )
log_end(key=constants.EVAL_STOP, metadata={"epoch_num": result_epoch})
# terminating condition
if bbox_map >= min_bbox_map and segm_map >= min_segm_map:
logger.info("Target mAP reached, exiting...")
finished = 1
#return True
# We now know on rank 0 whether or not we should terminate
# Bcast this flag on multi-GPU
if get_world_size() > 1:
with torch.no_grad():
finish_tensor = torch.tensor([finished], dtype=torch.int32, device = torch.device('cuda'))
torch.distributed.broadcast(finish_tensor, 0)
# If notified, end.
if finish_tensor.item() == 1:
return True
else:
# Single GPU, don't need to create tensor to bcast, just use value directly
if finished == 1:
return True
# Otherwise, default case, continue
return False
def mlperf_log_epoch_start(iteration, iters_per_epoch):
# First iteration:
# Note we've started training & tag first epoch start
if iteration == 0:
log_start(key=constants.BLOCK_START, metadata={"first_epoch_num":1, "epoch_count":1})
log_start(key=constants.EPOCH_START, metadata={"epoch_num":1})
return
if iteration % iters_per_epoch == 0:
epoch = iteration // iters_per_epoch + 1
log_start(key=constants.BLOCK_START, metadata={"first_epoch_num": epoch, "epoch_count": 1})
log_start(key=constants.EPOCH_START, metadata={"epoch_num": epoch})
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=os.getenv('LOCAL_RANK', 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())
# 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)
torch.distributed.init_process_group(
backend="nccl", init_method="env://"
)
# 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))
# 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 train(cfg, local_rank, distributed, random_number_generator=None):
if (torch._C, '_jit_set_profiling_executor') :
torch._C._jit_set_profiling_executor(False)
if (torch._C, '_jit_set_profiling_mode') :
torch._C._jit_set_profiling_mode(False)
# Model logging
log_event(key=constants.GLOBAL_BATCH_SIZE, value=cfg.SOLVER.IMS_PER_BATCH)
log_event(key=constants.NUM_IMAGE_CANDIDATES, value=cfg.MODEL.RPN.FPN_POST_NMS_TOP_N_TRAIN)
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
# Initialize mixed-precision training
is_fp16 = (cfg.DTYPE == "float16")
if is_fp16:
# convert model to FP16
model.half()
optimizer = make_optimizer(cfg, model)
# Optimizer logging
log_event(key=constants.OPT_NAME, value="sgd_with_momentum")
log_event(key=constants.OPT_BASE_LR, value=cfg.SOLVER.BASE_LR)
log_event(key=constants.OPT_LR_WARMUP_STEPS, value=cfg.SOLVER.WARMUP_ITERS)
log_event(key=constants.OPT_LR_WARMUP_FACTOR, value=cfg.SOLVER.WARMUP_FACTOR)
log_event(key=constants.OPT_LR_DECAY_FACTOR, value=cfg.SOLVER.GAMMA)
log_event(key=constants.OPT_LR_DECAY_STEPS, value=cfg.SOLVER.STEPS)
log_event(key=constants.MIN_IMAGE_SIZE, value=cfg.INPUT.MIN_SIZE_TRAIN[0])
log_event(key=constants.MAX_IMAGE_SIZE, value=cfg.INPUT.MAX_SIZE_TRAIN)
scheduler = make_lr_scheduler(cfg, optimizer)
# disable the garbage collection
gc.disable()
if distributed:
model = DDP(model, delay_allreduce=True)
arguments = {}
arguments["iteration"] = 0
arguments["nhwc"] = cfg.NHWC
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
arguments["save_checkpoints"] = cfg.SAVE_CHECKPOINTS
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT, cfg.NHWC)
arguments.update(extra_checkpoint_data)
if is_fp16:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
log_end(key=constants.INIT_STOP)
barrier()
log_start(key=constants.RUN_START)
barrier()
data_loader, iters_per_epoch = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
random_number_generator=random_number_generator,
)
log_event(key=constants.TRAIN_SAMPLES, value=len(data_loader))
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
# set the callback function to evaluate and potentially
# early exit each epoch
if cfg.PER_EPOCH_EVAL:
per_iter_callback_fn = functools.partial(
mlperf_test_early_exit,
iters_per_epoch=iters_per_epoch,
tester=functools.partial(test, cfg=cfg),
model=model,
distributed=distributed,
min_bbox_map=cfg.MLPERF.MIN_BBOX_MAP,
min_segm_map=cfg.MLPERF.MIN_SEGM_MAP)
else:
per_iter_callback_fn = None
start_train_time = time.time()
success = do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
cfg.DISABLE_REDUCED_LOGGING,
per_iter_start_callback_fn=functools.partial(mlperf_log_epoch_start, iters_per_epoch=iters_per_epoch),
per_iter_end_callback_fn=per_iter_callback_fn,
)
end_train_time = time.time()
total_training_time = end_train_time - start_train_time
print(
"&&&& MLPERF METRIC THROUGHPUT={:.4f} iterations / s".format((arguments["iteration"] * cfg.SOLVER.IMS_PER_BATCH) / total_training_time)
)
return model, success
def train(cfg, local_rank, distributed, disable_allreduce_for_logging,
random_number_generator):
# Model logging
log_event(key=constants.GLOBAL_BATCH_SIZE, value=cfg.SOLVER.IMS_PER_BATCH)
log_event(key=constants.NUM_IMAGE_CANDIDATES,
value=cfg.MODEL.RPN.FPN_POST_NMS_TOP_N_TRAIN)
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
# Optimizer logging
log_event(key=constants.OPT_NAME, value="sgd_with_momentum")
log_event(key=constants.OPT_BASE_LR, value=cfg.SOLVER.BASE_LR)
log_event(key=constants.OPT_LR_WARMUP_STEPS, value=cfg.SOLVER.WARMUP_ITERS)
log_event(key=constants.OPT_LR_WARMUP_FACTOR,
value=cfg.SOLVER.WARMUP_FACTOR)
log_event(key=constants.OPT_LR_DECAY_FACTOR, value=cfg.SOLVER.GAMMA)
log_event(key=constants.OPT_LR_DECAY_STEPS, value=cfg.SOLVER.STEPS)
log_event(key=constants.MIN_IMAGE_SIZE, value=cfg.INPUT.MIN_SIZE_TRAIN[0])
log_event(key=constants.MAX_IMAGE_SIZE, value=cfg.INPUT.MAX_SIZE_TRAIN)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(cfg, model, optimizer, scheduler,
output_dir, save_to_disk)
arguments["save_checkpoints"] = cfg.SAVE_CHECKPOINTS
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
log_end(key=constants.INIT_STOP)
log_start(key=constants.RUN_START)
data_loader, iters_per_epoch = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
random_number_generator=random_number_generator)
log_event(key=constants.TRAIN_SAMPLES, value=len(data_loader))
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
# set the callback function to evaluate and potentially
# early exit each epoch
if cfg.PER_EPOCH_EVAL:
per_iter_callback_fn = functools.partial(
mlperf_test_early_exit,
iters_per_epoch=iters_per_epoch,
tester=functools.partial(test, cfg=cfg),
model=model,
distributed=distributed,
min_bbox_map=cfg.MLPERF.MIN_BBOX_MAP,
min_segm_map=cfg.MLPERF.MIN_SEGM_MAP)
else:
per_iter_callback_fn = None
start_train_time = time.time()
success = do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
per_iter_start_callback_fn=functools.partial(
mlperf_log_epoch_start, iters_per_epoch=iters_per_epoch),
per_iter_end_callback_fn=per_iter_callback_fn,
)
end_train_time = time.time()
total_training_time = end_train_time - start_train_time
print("&&&& MLPERF METRIC THROUGHPUT per GPU={:.4f} iterations / s".format(
(arguments["iteration"] * 1.0) / total_training_time))
return model, success
