File size: 6,225 Bytes
e5ccff5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 | import torch.utils.data as data
import torch
import logging
from timm.data.distributed_sampler import RepeatAugSampler
from .xray_transform import build_transform_xray, build_seg_transform
from .chest_dset import MedFMC_Chest, ShenZhenCXR
from .vindr import VinDrNew
from .rsna import RSNA, RSNA_3Class, RSNASegmentDataset
from .mimic import MIMIC_Pretrain
from .world_model_transform import build_world_model_transform
from .xray_transform_add import build_world_model_dual_transform
from .nih import NIH
from .chexpert import CheXpert
from .siim import SIIM
def build_dataset(args, split='train', ten_crop=False):
transform = build_transform_xray(args, is_train=(split == 'train'), ten_crop=ten_crop)
if args.dataset == 'vindr_new':
logging.info(f'Dataset VinDr New: Pct={args.data_pct}')
dataset = VinDrNew(transform=transform, data_pct=args.data_pct, split=split, dataset_seed=args.dataset_seed)
elif args.dataset == 'rsna':
logging.info(f'Dataset RSNA: Pct={args.data_pct}')
dataset = RSNA(transform=transform, data_pct=args.data_pct, split=split)
elif args.dataset == 'rsna_3cls':
dataset = RSNA_3Class(transform=transform, data_pct=args.data_pct, split=split)
elif args.dataset == 'medfmc':
logging.info(f'Dataset MedFMC: {args.fmc_shots} Shots, No.{args.fmc_exp_num}')
dataset = MedFMC_Chest(exp_num=args.fmc_exp_num, shots=args.fmc_shots, split=split, transform=transform)
elif args.dataset == 'nih':
logging.info(f'Dataset NIH: Pct={args.data_pct}')
dataset = NIH(split=split, data_pct=args.data_pct, transform=transform, return_label=True)
elif args.dataset == 'shenzhen':
logging.info(f'Dataset ShenZhen')
dataset = ShenZhenCXR(split=split, transform=transform)
elif args.dataset == 'chex':
dataset = CheXpert(transform=transform, return_label=True, data_pct=args.data_pct, split=split)
else:
raise NotImplementedError
if args.dataset_cat > 1 and split == 'train':
logging.info(f'Concat Dataset x{args.dataset_cat}')
dataset = data.ConcatDataset([dataset for _ in range(args.dataset_cat)])
dataset.num_classes = dataset.datasets[0].num_classes
return dataset
def build_seg_dataset(args, split='train'):
transform = build_seg_transform(args, is_train=(split=='train'))
if args.dataset == 'siim':
dataset = SIIM(transform=transform, split=split, data_pct=args.data_pct, balance=args.dataset_balance)
elif args.dataset == 'rsna':
dataset = RSNASegmentDataset(split=split, transform=transform, data_pct=args.data_pct)
else:
raise NotImplementedError
if args.dataset_cat > 1 and split == 'train':
logging.info(f'Concat Dataset x{args.dataset_cat}')
dataset = data.ConcatDataset([dataset for _ in range(args.dataset_cat)])
dataset.num_classes = dataset.datasets[0].num_classes
return dataset
def build_pretrain_dataset(args, ssl_type):
if ssl_type == 'iwm':
logging.info('Build World Model Transform')
transform = build_world_model_transform(args)
elif 'iwm_dual' in ssl_type:
transform = build_world_model_dual_transform(args)
else:
transform = build_transform_xray(args, is_train=True)
if args.dataset == 'mimic':
logging.info('Use MIMIC dataset')
dataset = MIMIC_Pretrain(transform=transform, split='all', data_pct=args.data_pct, preload=args.preload_dataset, data_postfix=args.data_postfix, include_lateral=args.include_lateral)
else:
logging.info(f'Use Hybrid dataset {args.dataset}')
d_list = []
if 'nih' in args.dataset:
d_list.append(NIH(transform=transform, data_pct=args.data_pct, return_label=False))
if 'chex' in args.dataset:
d_list.append(CheXpert(transform=transform, data_pct=args.data_pct, return_label=False, include_lateral=args.include_lateral))
if 'mimic' in args.dataset:
d_list.append(MIMIC_Pretrain(transform=transform, split='all', data_pct=args.data_pct, data_postfix=args.data_postfix, include_lateral=args.include_lateral))
dataset = data.ConcatDataset(d_list)
logging.info(f'Total Images: {len(dataset)}')
return dataset
def get_loader(args, dataset, is_train=True, persistent_workers=True, custom_bs=None):
if is_train:
if args.num_aug_repeats:
sampler = RepeatAugSampler(dataset, num_repeats=args.num_aug_repeats)
logging.info(f'Use Repeated Aug Sampler: {args.num_aug_repeats}')
else:
num_tasks = args.world_size
global_rank = args.rank
sampler = data.DistributedSampler(
dataset, num_replicas=num_tasks, rank=global_rank, shuffle=True
)
else:
sampler = data.SequentialSampler(dataset)
this_bs = args.batch_size if is_train else args.batch_size * 2
this_bs = custom_bs if custom_bs is not None else this_bs
return data.DataLoader(
dataset, sampler=sampler,
batch_size=this_bs,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=is_train,
persistent_workers=persistent_workers
)
def get_infinite_loader(args, dataset, batch_size, custom_iter=None):
# num_tasks = args.world_size
# global_rank = args.rank
num_iters = custom_iter if custom_iter is not None else args.num_iters
num_samples = batch_size * num_iters
if args.data_seed is not None:
generator = torch.Generator()
generator.manual_seed(args.data_seed)
logging.info(f'Use Data Seed {args.data_seed}')
else:
generator = None
logging.info(f'No Data Seed')
sampler = data.RandomSampler(dataset, replacement=(not args.data_no_replace), num_samples=num_samples, generator=generator)
# sampler = DistributedProxySampler(sampler, num_replicas=num_tasks, rank=global_rank)
sampler = data.BatchSampler(sampler, batch_size, drop_last=True)
return data.DataLoader(
dataset,
batch_sampler=sampler,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
persistent_workers=True
) |