def __init__(self, cfg: Configurer):
super().__init__(cfg)
cfg_dataset = cfg.DATASET
self.normalization = cfg_dataset.normalization
self.color_mode = cfg_dataset.color_mode
self.img_size = cfg_dataset.image_size
assert 'setting' not in cfg_dataset
self.zl_fabric_src_domain = ZLFabric(dir=cfg_dataset.dataset_dir,
fabric=cfg_dataset.fabric,
setting='setting1')
self.zl_fabric_tgt_domain = ZLFabric(dir=cfg_dataset.dataset_dir,
fabric=cfg_dataset.fabric,
setting='setting2')
assert self.normalization in [
'[0,1]',
'[-1,1]',
'zscore',
'histeq',
]
if self.normalization == '[-1,1]':
warnings.warn(
'Network using Sigmoid as output activation, you should use `[0,1]` normalization!'
)
elif self.normalization == 'histeq':
warnings.warn(
'histeq have problem to keep the background source and target image as same'
)
def __init__(self, cfg: Configurer):
super().__init__(cfg)
cfg_dataset = cfg.DATASET
self.normalization = cfg_dataset.normalization
self.color_mode = cfg_dataset.color_mode
self.img_size = cfg_dataset.image_size
self.setting = cfg_dataset.setting
self.patch_size = cfg_dataset.patch_size
self.patch_stride = cfg_dataset.patch_stride
self.zl_fabric = ZLFabric(dir=cfg_dataset.dataset_dir, fabric=cfg_dataset.fabric, setting=self.setting)
def __init__(self,
cfg: Configurer,
TrainDataset=_MaskDataset,
EvalDataset=_MaskDataset,
rnd_seed=None):
super().__init__(cfg)
cfg_dataset = cfg.DATASET
self.normalization = cfg_dataset.normalization
self.color_mode = cfg_dataset.color_mode
self.img_size = cfg_dataset.image_size
self.setting = cfg_dataset.setting
self.zl_fabric = ZLFabric(dir=cfg_dataset.dataset_dir,
fabric=cfg_dataset.fabric,
setting=self.setting,
seed=rnd_seed)
self.TrainDS = TrainDataset
self.EvalDS = EvalDataset
class PatchDSB(_DSBuilder):
def __init__(self, cfg: Configurer):
super().__init__(cfg)
cfg_dataset = cfg.DATASET
self.normalization = cfg_dataset.normalization
self.color_mode = cfg_dataset.color_mode
self.img_size = cfg_dataset.image_size
self.setting = cfg_dataset.setting
self.patch_size = cfg_dataset.patch_size
self.patch_stride = cfg_dataset.patch_stride
self.zl_fabric = ZLFabric(dir=cfg_dataset.dataset_dir, fabric=cfg_dataset.fabric, setting=self.setting)
def build_train(self) -> Tuple[Dataset, Dataset]:
zlimgs_train_normal, zlimgs_train_defect, zlimgs_train_eval_normal, zlimgs_train_eval_defect \
= self.zl_fabric.prepare_train()
dt_train = PatchDatasetTrain(zl_imgs=zlimgs_train_normal, normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
dt_train_eval = PatchDatasetEval(zl_imgs=zlimgs_train_eval_normal + zlimgs_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
return dt_train, dt_train_eval
def build_train_dev(self) -> Tuple[Dataset, Dataset, Dataset]:
zlimgs_folds = self.zl_fabric.prepare_k_fold(k_fold=5, shuffle=True)
zlimgs_k_train_normal, zlimgs_k_train_defect, \
zlimgs_k_train_eval_normal, zlimgs_k_train_eval_defect, \
zlimgs_k_dev_normal, zlimgs_k_dev_defect = zlimgs_folds[0]
dt_train = PatchDatasetTrain(zl_imgs=zlimgs_k_train_normal, normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
dt_train_eval = PatchDatasetEval(zl_imgs=zlimgs_k_train_eval_normal + zlimgs_k_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
dt_dev_eval = PatchDatasetEval(zl_imgs=zlimgs_k_dev_normal + zlimgs_k_dev_defect,
normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
return dt_train, dt_train_eval, dt_dev_eval
def build_k_fold(self) -> Tuple[List[Dataset], List[Dataset], List[Dataset]]:
k_fold = self.cfg.DATASET.k_fold
zlimgs_folds = self.zl_fabric.prepare_k_fold(k_fold=k_fold, shuffle=True)
dt_train_list = []
dt_train_eval_list = []
dt_dev_list = []
for zlimgs_k_train_normal, zlimgs_k_train_defect, \
zlimgs_k_train_eval_normal, zlimgs_k_train_eval_defect, \
zlimgs_k_dev_normal, zlimgs_k_dev_defect in zlimgs_folds:
dt_train = PatchDatasetTrain(zl_imgs=zlimgs_k_train_normal, normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
dt_train_eval = PatchDatasetEval(zl_imgs=zlimgs_k_train_eval_normal + zlimgs_k_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
dt_dev_eval = PatchDatasetEval(zl_imgs=zlimgs_k_dev_normal + zlimgs_k_dev_defect,
normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
dt_train_list.append(dt_train)
dt_train_eval_list.append(dt_train_eval)
dt_dev_list.append(dt_dev_eval)
return dt_train_list, dt_train_eval_list, dt_dev_list
def build_test(self) -> Dataset:
zlimgs_test_normal, zlimgs_test_defect = self.zl_fabric.prepare_test()
dt_test = PatchDatasetEval(zl_imgs=zlimgs_test_normal + zlimgs_test_defect,
normalization=self.normalization,
color_mode=self.color_mode, img_size=self.img_size,
patch_size=self.patch_size, patch_stride=self.patch_stride)
return dt_test
class PairAndMaskDSB(_DSBuilder):
def __init__(self, cfg: Configurer):
super().__init__(cfg)
cfg_dataset = cfg.DATASET
self.normalization = cfg_dataset.normalization
self.color_mode = cfg_dataset.color_mode
self.img_size = cfg_dataset.image_size
assert 'setting' not in cfg_dataset
self.zl_fabric_src_domain = ZLFabric(dir=cfg_dataset.dataset_dir,
fabric=cfg_dataset.fabric,
setting='setting1')
self.zl_fabric_tgt_domain = ZLFabric(dir=cfg_dataset.dataset_dir,
fabric=cfg_dataset.fabric,
setting='setting2')
assert self.normalization in [
'[0,1]',
'[-1,1]',
'zscore',
'histeq',
]
if self.normalization == '[-1,1]':
warnings.warn(
'Network using Sigmoid as output activation, you should use `[0,1]` normalization!'
)
elif self.normalization == 'histeq':
warnings.warn(
'histeq have problem to keep the background source and target image as same'
)
def build_train(self) -> Tuple[List[Dataset], Dataset]:
zlimgs_train_normal, _, _, _ = self.zl_fabric_src_domain.prepare_train(
)
dt_train_src = _PairDataset(zl_imgs=zlimgs_train_normal,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
_, zlimgs_train_defect, zlimgs_train_eval_normal, zlimgs_train_eval_defect \
= self.zl_fabric_tgt_domain.prepare_train()
dt_train_tgt = _MaskDataset(zl_imgs=zlimgs_train_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_train_eval = _MaskDataset(zl_imgs=zlimgs_train_eval_normal +
zlimgs_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
return [dt_train_src, dt_train_tgt], dt_train_eval
def build_train_dev(self) -> Tuple[List[Dataset], Dataset, Dataset]:
zlimgs_folds_src = self.zl_fabric_src_domain.prepare_k_fold(
k_fold=5, shuffle=True)
zlimgs_k_train_normal, _, _, _, _, _ = zlimgs_folds_src[0]
dt_train_src = _PairDataset(zl_imgs=zlimgs_k_train_normal,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
zlimgs_folds_tgt = self.zl_fabric_tgt_domain.prepare_k_fold(
k_fold=5, shuffle=True)
_, zlimgs_k_train_defect, \
zlimgs_k_train_eval_normal, zlimgs_k_train_eval_defect, \
zlimgs_k_dev_normal, zlimgs_k_dev_defect = zlimgs_folds_tgt[0]
dt_train_tgt = _MaskDataset(zl_imgs=zlimgs_k_train_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_train_eval = _MaskDataset(zl_imgs=zlimgs_k_train_eval_normal +
zlimgs_k_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_dev_eval = _MaskDataset(zl_imgs=zlimgs_k_dev_normal +
zlimgs_k_dev_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
return [dt_train_src, dt_train_tgt], dt_train_eval, dt_dev_eval
def build_k_fold(self):
k_fold = self.cfg.DATASET.k_fold
assert k_fold > 1
zlimgs_folds_src = self.zl_fabric_src_domain.prepare_k_fold(
k_fold=k_fold, shuffle=True)
zlimgs_folds_tgt = self.zl_fabric_tgt_domain.prepare_k_fold(
k_fold=k_fold, shuffle=True)
dt_train_list = []
dt_train_eval_list = []
dt_dev_list = []
for zlimgs_k_fold_src, zlimgs_k_fold_tgt in zip(
zlimgs_folds_src, zlimgs_folds_tgt):
zlimgs_k_train_normal, _, _, _, _, _ = zlimgs_k_fold_src
dt_train_src = _PairDataset(zl_imgs=zlimgs_k_train_normal,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
_, zlimgs_k_train_defect, \
zlimgs_k_train_eval_normal, zlimgs_k_train_eval_defect, \
zlimgs_k_dev_normal, zlimgs_k_dev_defect = zlimgs_k_fold_tgt
dt_train_tgt = _MaskDataset(zl_imgs=zlimgs_k_train_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_train_eval = _MaskDataset(zl_imgs=zlimgs_k_train_eval_normal +
zlimgs_k_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_dev_eval = _MaskDataset(zl_imgs=zlimgs_k_dev_normal +
zlimgs_k_dev_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_train_list.append([dt_train_src, dt_train_tgt])
dt_train_eval_list.append(dt_train_eval)
dt_dev_list.append(dt_dev_eval)
return dt_train_list, dt_train_eval_list, dt_dev_list
def build_test(self) -> Dataset:
zlimgs_test_normal, zlimgs_test_defect = self.zl_fabric_tgt_domain.prepare_test(
)
dt_test = _MaskDataset(zl_imgs=zlimgs_test_normal + zlimgs_test_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
return dt_test
class AugDSB(_DSBuilder):
def __init__(self, cfg) -> None:
super().__init__(cfg)
cfg_dataset = cfg.DATASET
self.normalization = cfg_dataset.normalization
self.color_mode = cfg_dataset.color_mode
self.img_size = cfg_dataset.image_size
self.setting = cfg_dataset.setting
self.zl_fabric = ZLFabric(dir=cfg_dataset.dataset_dir,
fabric=cfg_dataset.fabric,
setting=self.setting)
assert self.normalization in [
'[0,1]',
'[-1,1]',
'zscore',
'histeq',
]
if self.normalization == '[-1,1]':
warnings.warn(
'Network using Sigmoid as output activation, you should use `[0,1]` normalization!'
)
elif self.normalization == 'histeq':
warnings.warn(
'histeq have problem to keep the background source and target image as same'
)
def build_train(self) -> Tuple[Dataset, Dataset]:
zlimgs_train_normal, zlimgs_train_defect, zlimgs_train_eval_normal, zlimgs_train_eval_defect \
= self.zl_fabric.prepare_train()
dt_train = AugDataset(
good_zl_imgs=zlimgs_train_normal,
bad_zl_imgs=zlimgs_train_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size,
)
dt_train_eval = _MaskDataset(zl_imgs=zlimgs_train_eval_normal +
zlimgs_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
return dt_train, dt_train_eval
def build_train_dev(self) -> Tuple[Dataset, Dataset, Dataset]:
zlimgs_folds = self.zl_fabric.prepare_k_fold(k_fold=5, shuffle=True)
zlimgs_k_train_normal, zlimgs_k_train_defect, \
zlimgs_k_train_eval_normal, zlimgs_k_train_eval_defect, \
zlimgs_k_dev_normal, zlimgs_k_dev_defect = zlimgs_folds[0]
dt_train = AugDataset(good_zl_imgs=zlimgs_k_train_normal,
bad_zl_imgs=zlimgs_k_train_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_train_eval = _MaskDataset(zl_imgs=zlimgs_k_train_eval_normal +
zlimgs_k_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_dev_eval = _MaskDataset(zl_imgs=zlimgs_k_dev_normal +
zlimgs_k_dev_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
return dt_train, dt_train_eval, dt_dev_eval
def build_k_fold(
self) -> Tuple[List[Dataset], List[Dataset], List[Dataset]]:
k_fold = self.cfg.DATASET.k_fold
zlimgs_folds = self.zl_fabric.prepare_k_fold(k_fold=k_fold,
shuffle=True)
dt_train_list = []
dt_train_eval_list = []
dt_dev_list = []
for zlimgs_k_train_normal, zlimgs_k_train_defect, \
zlimgs_k_train_eval_normal, zlimgs_k_train_eval_defect, \
zlimgs_k_dev_normal, zlimgs_k_dev_defect in zlimgs_folds:
dt_train = AugDataset(good_zl_imgs=zlimgs_k_train_normal,
bad_zl_imgs=zlimgs_k_train_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_train_eval = _MaskDataset(zl_imgs=zlimgs_k_train_eval_normal +
zlimgs_k_train_eval_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_dev_eval = _MaskDataset(zl_imgs=zlimgs_k_dev_normal +
zlimgs_k_dev_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
dt_train_list.append(dt_train)
dt_train_eval_list.append(dt_train_eval)
dt_dev_list.append(dt_dev_eval)
return dt_train_list, dt_train_eval_list, dt_dev_list
def build_test(self) -> Dataset:
zlimgs_test_normal, zlimgs_test_defect = self.zl_fabric.prepare_test()
dt_test = _MaskDataset(zl_imgs=zlimgs_test_normal + zlimgs_test_defect,
normalization=self.normalization,
color_mode=self.color_mode,
img_size=self.img_size)
return dt_test