def main():
args = get_args()
args.local_rank = int(os.environ["LOCAL_RANK"]) if args.distributed else 0
if args.local_rank == 0 or args.local_rank is None:
print("Command line arguments: ", args)
manual_seed = 0 if args.local_rank is None else args.local_rank
init_seed(manual_seed)
cfg = load_cfg(args)
if args.local_rank == 0 or args.local_rank is None:
# In distributed training, only print and save the
# configurations using the node with local_rank=0.
print("PyTorch: ", torch.__version__)
print(cfg)
if not os.path.exists(cfg.DATASET.OUTPUT_PATH):
print('Output directory: ', cfg.DATASET.OUTPUT_PATH)
os.makedirs(cfg.DATASET.OUTPUT_PATH)
save_all_cfg(cfg, cfg.DATASET.OUTPUT_PATH)
if args.distributed:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
assert torch.cuda.is_available(), \
"Distributed training without GPUs is not supported!"
dist.init_process_group("nccl", init_method='env://')
else:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Rank: {}. Device: {}".format(args.local_rank, device))
cudnn.enabled = True
cudnn.benchmark = True
mode = 'test' if args.inference else 'train'
trainer = Trainer(cfg, device, mode,
rank=args.local_rank,
checkpoint=args.checkpoint)
# Start training or inference:
if cfg.DATASET.DO_CHUNK_TITLE == 0:
test_func = trainer.test_singly if cfg.INFERENCE.DO_SINGLY else trainer.test
test_func() if args.inference else trainer.train()
else:
trainer.run_chunk(mode)
print("Rank: {}. Device: {}. Process is finished!".format(
args.local_rank, device))
def main():
r"""Main function.
"""
# arguments
args = get_args()
print("Command line arguments:")
print(args)
# configurations
cfg = get_cfg_defaults()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
if args.inference:
update_inference_config(cfg)
cfg.freeze()
print("Configuration details:")
print(cfg)
if not os.path.exists(cfg.DATASET.OUTPUT_PATH):
print('Output directory: ', cfg.DATASET.OUTPUT_PATH)
os.makedirs(cfg.DATASET.OUTPUT_PATH)
save_all_cfg(cfg, cfg.DATASET.OUTPUT_PATH)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Device: ", device)
mode = 'test' if args.inference else 'train'
trainer = Trainer(cfg, device, mode, args.checkpoint)
if args.inference:
trainer.test()
else:
trainer.train()
def main():
r"""Main function.
"""
# arguments
args = get_args()
print("Command line arguments:")
print(args)
# configurations
cfg = get_cfg_defaults()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
if args.inference:
update_inference_cfg(cfg)
# Overwrite options given configs with higher priority.
overwrite_cfg(cfg, args)
cfg.freeze()
print("Configuration details:")
print(cfg)
if not os.path.exists(cfg.DATASET.OUTPUT_PATH):
print('Output directory: ', cfg.DATASET.OUTPUT_PATH)
os.makedirs(cfg.DATASET.OUTPUT_PATH)
save_all_cfg(cfg, cfg.DATASET.OUTPUT_PATH)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Device: ", device)
cudnn.enabled = True
cudnn.benchmark = True
mode = 'test' if args.inference else 'train'
trainer = Trainer(cfg, device, mode, args.checkpoint)
if cfg.DATASET.DO_CHUNK_TITLE == 0:
if args.inference:
trainer.test()
else:
trainer.train()
else:
trainer.run_chunk(mode)