| from src.models.build_sam3D import sam_model_registry3D | |
| from src.dataset.dataloader import Dataset_promise, Dataloader_promise | |
| import torchio as tio | |
| from torch.nn.parallel import DistributedDataParallel as DDP | |
| from torch.utils.data.distributed import DistributedSampler | |
| import torch | |
| def get_dataloader(args, split='', use_small=False): | |
| transforms_list = [tio.ToCanonical(), tio.Resample(1), ] | |
| if split == 'train': | |
| transforms_list.append(tio.RandomFlip(axes=(0, 1, 2))) | |
| transforms = tio.Compose(transforms_list) | |
| dataset = Dataset_promise( | |
| data=args.data, | |
| data_dir=args.data_dir, | |
| split=split, | |
| transform=transforms, | |
| image_size=args.image_size, | |
| args=args, | |
| ) | |
| batch_size = args.batch_size if split == 'train' else 1 | |
| if split == 'train': | |
| train_sampler = None | |
| shuffle = True | |
| if args.ddp: | |
| train_sampler = DistributedSampler(dataset) | |
| shuffle = False | |
| else: | |
| train_sampler = None | |
| shuffle = False | |
| pin_memory = True | |
| if split != 'train' and args.data == 'lits': | |
| pin_memory = False | |
| dataloader = Dataloader_promise( | |
| dataset=dataset, | |
| sampler=train_sampler, | |
| batch_size=batch_size, | |
| shuffle=shuffle, | |
| num_workers=args.num_workers, | |
| pin_memory=pin_memory, | |
| ) | |
| return dataloader | |
| def build_model(args, checkpoint=None): | |
| sam_model = sam_model_registry3D[args.model_type](checkpoint=checkpoint, args=args).to(args.device) | |
| if args.ddp: | |
| sam_model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(sam_model) | |
| sam_model = DDP(sam_model, device_ids=[args.rank], output_device=args.rank) | |
| return sam_model | |