def get_training_augmentation():
"""Builds random transformations we want to apply to our dataset.
Arguments:
Returns:
A albumentation functions to pass our images to.
Raises:
"""
train_transform = [
A.IAAAdditiveGaussianNoise(p=0.2),
A.IAAPerspective(p=0.5),
A.OneOf([A.CLAHE(p=1), A.RandomBrightness(p=1), A.RandomGamma(p=1),], p=0.9,),
A.OneOf(
[
A.IAASharpen(p=1),
A.Blur(blur_limit=3, p=1),
A.MotionBlur(blur_limit=3, p=1),
],
p=0.9,
),
A.OneOf([A.RandomContrast(p=1), A.HueSaturationValue(p=1),], p=0.9,),
A.Lambda(mask=round_clip_0_1),
]
return A.Compose(train_transform)
def _get_transforms(self):
resize_transform: list = [
A.PadIfNeeded(
min_height=self.h, min_width=self.w, border_mode=cv2.BORDER_CONSTANT
),
A.RandomCrop(self.h, self.w),
]
augmentations: list = [
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(p=0.2),
]
def _normalizer(image, **_):
return image / 255.0
compatible: list = [
ToTensorV2(always_apply=True),
A.Lambda(image=_normalizer),
]
transforms_train_val_ = A.Compose(resize_transform + compatible)
transforms_test_ = A.Compose(resize_transform + compatible)
def transforms_train_val(image):
return transforms_train_val_(image=np.array(image))["image"]
def transforms_test(image):
return transforms_test_(image=np.array(image))["image"]
return transforms_train_val, transforms_test
def get_train_transforms_simple_bright_pad(input_size,
use_crop=False,
use_no_color_aug=False,
use_center_crop=False,
center_crop_ratio=0.9,
use_gray=False):
def longest_max_size(img, interpolation=cv2.INTER_LINEAR, **params):
img = F.longest_max_size(img,
max_size=input_size,
interpolation=interpolation)
return img
return al.Compose([
al.Lambda(longest_max_size),
al.PadIfNeeded(min_height=input_size,
min_width=input_size,
always_apply=True,
border_mode=0),
al.HorizontalFlip(p=0.5),
al.RandomBrightness(p=0.2, limit=0.2),
al.RandomContrast(p=0.1, limit=0.2),
al.Normalize(),
ToTensorV2()
])
def get_preprocessing(preprocessing_fn):
"""构造预处理转换
Args:
preprocessing_fn (callbale): 标准化函数
(可以针对每个预先训练的神经网络)
Return:
transform: albumentations.Compose
"""
_transform = [
albu.Lambda(image=preprocessing_fn),
albu.Lambda(image=to_tensor, mask=to_tensor),
]
return albu.Compose(_transform)
def test_lambda_serialization(image, mask, albumentations_bboxes, keypoints, seed, p):
def vflip_image(image, **kwargs):
return F.vflip(image)
def vflip_mask(mask, **kwargs):
return F.vflip(mask)
def vflip_bbox(bbox, **kwargs):
return F.bbox_vflip(bbox, **kwargs)
def vflip_keypoint(keypoint, **kwargs):
return F.keypoint_vflip(keypoint, **kwargs)
aug = A.Lambda(name="vflip", image=vflip_image, mask=vflip_mask, bbox=vflip_bbox, keypoint=vflip_keypoint, p=p)
serialized_aug = A.to_dict(aug)
deserialized_aug = A.from_dict(serialized_aug, lambda_transforms={"vflip": aug})
set_seed(seed)
aug_data = aug(image=image, mask=mask, bboxes=albumentations_bboxes, keypoints=keypoints)
set_seed(seed)
deserialized_aug_data = deserialized_aug(image=image, mask=mask, bboxes=albumentations_bboxes, keypoints=keypoints)
assert np.array_equal(aug_data["image"], deserialized_aug_data["image"])
assert np.array_equal(aug_data["mask"], deserialized_aug_data["mask"])
assert np.array_equal(aug_data["bboxes"], deserialized_aug_data["bboxes"])
assert np.array_equal(aug_data["keypoints"], deserialized_aug_data["keypoints"])
def __init__(self,
images,
labels,
batch_size=16,
image_shape=(256, 512, 1),
do_shuffle_at_epoch_end=True,
length=None,
do_augment=True):
super().__init__(images, labels, batch_size, image_shape,
do_shuffle_at_epoch_end, length, do_augment)
self.augmenting_pipeline = A.Compose([
A.HorizontalFlip(),
A.IAAAffine(translate_percent={"x": (-1, 1)}, mode="reflect", p=1),
A.PadIfNeeded(min_width=int(self.input_shape[1] * 2),
min_height=self.input_shape[0]),
A.GridDistortion(p=0.8, distort_limit=0.5),
A.ElasticTransform(p=0.5,
alpha=10,
sigma=100 * 0.03,
alpha_affine=0),
A.CenterCrop(width=self.input_shape[1],
height=self.input_shape[0]),
A.IAAPerspective(scale=(0, 0.10), p=1),
A.ShiftScaleRotate(shift_limit=0,
scale_limit=(.0, 0.4),
rotate_limit=0,
p=0.5),
A.CLAHE(clip_limit=2.0, p=0.5),
A.Lambda(
image=self.convert_image,
mask=self.convert_segmentations,
),
])
def _get_augmentations(h, w, augment):
def normalize(x, **_):
return x / 255.0
resizing: list = [
# A.LongestMaxSize(max_size=WIDTH, always_apply=True),
A.PadIfNeeded(min_height=h,
min_width=w,
border_mode=cv2.BORDER_CONSTANT),
A.RandomCrop(h, w),
# A.Resize(height=HEIGHT, width=WIDTH, always_apply=True),
]
compatibility: list = [
ToTensorV2(always_apply=True),
A.Lambda(image=normalize),
]
augmentations: list = []
if augment:
augmentations = [
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(p=0.2),
]
return A.Compose(
resizing + augmentations + compatibility,
bbox_params=A.BboxParams(format="pascal_voc",
min_visibility=0.05,
label_fields=["class_labels"]),
)
def get_inference_augmentation():
"""Add paddings to make image shape divisible by 32"""
test_transform = [
albu.Normalize(),
albu.Lambda(image=to_tensor, mask=to_tensor),
]
return albu.Compose(test_transform)
def test_lambda_serialization(image, mask, bboxes, keypoints, seed, p):
def vflip_image(image, **kwargs):
return F.vflip(image)
def vflip_mask(mask, **kwargs):
return F.vflip(mask)
def vflip_bbox(bbox, **kwargs):
return F.bbox_vflip(bbox, **kwargs)
def vflip_keypoint(keypoint, **kwargs):
return F.keypoint_vflip(keypoint, **kwargs)
aug = A.Lambda(name='vflip',
image=vflip_image,
mask=vflip_mask,
bbox=vflip_bbox,
keypoint=vflip_keypoint,
p=p)
serialized_aug = A.to_dict(aug)
deserialized_aug = A.from_dict(serialized_aug,
lambda_transforms={'vflip': aug})
random.seed(seed)
aug_data = aug(image=image, mask=mask, bboxes=bboxes, keypoints=keypoints)
random.seed(seed)
deserialized_aug_data = deserialized_aug(image=image,
mask=mask,
bboxes=bboxes,
keypoints=keypoints)
assert np.array_equal(aug_data['image'], deserialized_aug_data['image'])
assert np.array_equal(aug_data['mask'], deserialized_aug_data['mask'])
assert np.array_equal(aug_data['bboxes'], deserialized_aug_data['bboxes'])
assert np.array_equal(aug_data['keypoints'],
deserialized_aug_data['keypoints'])
def get_preprocessing(preprocessing_fn):
"""Construct preprocessing transform
Args:
preprocessing_fn (callbale): data normalization function
(can be specific for each pretrained neural network)
Return:
transform: albumentations.Compose
"""
_transform = [
albu.Lambda(image=preprocessing_fn),
albu.Lambda(image=Transforms.to_tensor, mask=Transforms.to_tensor),
]
return albu.Compose(_transform)
def test_lambda_transform():
def negate_image(image, **kwargs):
return -image
def one_hot_mask(mask, num_channels, **kwargs):
new_mask = np.eye(num_channels, dtype=np.uint8)[mask]
return new_mask
def vflip_bbox(bbox, **kwargs):
return F.bbox_vflip(bbox, **kwargs)
def vflip_keypoint(keypoint, **kwargs):
return F.keypoint_vflip(keypoint, **kwargs)
aug = A.Lambda(
image=negate_image, mask=partial(one_hot_mask, num_channels=16), bbox=vflip_bbox, keypoint=vflip_keypoint, p=1
)
output = aug(
image=np.ones((10, 10, 3), dtype=np.float32),
mask=np.tile(np.arange(0, 10), (10, 1)),
bboxes=[(10, 15, 25, 35)],
keypoints=[(20, 30, 40, 50)],
)
assert (output["image"] < 0).all()
assert output["mask"].shape[2] == 16 # num_channels
assert output["bboxes"] == [F.bbox_vflip((10, 15, 25, 35), 10, 10)]
assert output["keypoints"] == [F.keypoint_vflip((20, 30, 40, 50), 10, 10)]
def get_preprocessing(self, pre_processing_fn):
"""
Construct preprocessing transform
from https://github.com/qubvel/segmentation_models.pytorch/blob/master/examples/cars%20segmentation%20(camvid).ipynb
[Cell 11]
Args:
preprocessing_fn (callbale): data normalization function
(can be specific for each pretrained neural network)
Return:
transform: albumentations.Compose
"""
_transform = [
A.Lambda(name='pre_process', image=pre_processing_fn),
A.Lambda(name='to_tensor', image=self.to_tensor, mask=self.to_tensor),
]
return A.Compose(_transform)
def get_preprocessing(preprocessing_fn):
"""Construct preprocessing transform
Args:
preprocessing_fn (callbale): data normalization function
(can be specific for each pretrained neural network)
Return:
transform: albumentations.Compose
"""
# TODO: UserWarning: Using lambda is incompatible with multiprocessing. Consider using regular functions or partial().
# warnings.warn('Using lambda is incompatible with multiprocessing.
_transform = [
albu.Lambda(image=preprocessing_fn),
albu.Lambda(image=to_tensor, mask=to_tensor),
]
return albu.Compose(_transform)
def get_preprocessing(preprocessing_fn, downsample_factor=1):
"""Construct preprocessing transform
Args:
preprocessing_fn (callbale): data normalization function
(can be specific for each pretrained neural network)
Return:
transform: albumentations.Compose
"""
downsample_fn = partial(downsample, factor=downsample_factor)
_transform = [
albu.Lambda(image=downsample_fn, mask=downsample_fn),
albu.Lambda(image=preprocessing_fn),
albu.Lambda(image=to_tensor, mask=to_tensor),
]
return albu.Compose(_transform)
def get_preprocessing(preprocessing_fn):
def to_tensor(x, **kwargs):
return x.transpose(2, 0, 1).astype('float32')
"""Construct preprocessing transform
Args:
preprocessing_fn (callbale): data normalization functio
(can be specific for each pretrained neural network)
Return:
transform: albumentations.Compose
"""
_transform = [
albu.Lambda(image=preprocessing_fn),
albu.Lambda(image=to_tensor, mask=to_tensor),
]
return albu.Compose(_transform)
def build_test_transform(size=244):
"""Build argmentation process for testing data
Arg:
size (int): the output size for images (3, size, size).
"""
_transform = [albu.Resize(size, size), albu.Lambda(image=to_tensor)]
return albu.Compose(_transform)
def get_preprocessing():
'''
Preprocess the image and the mask
'''
_transform = [
A.Lambda(image=to_tensor, mask=to_tensor_mask),
]
return A.Compose(_transform)
def get_preprocessing():
import torch
def norm(img, **params):
return F.normalize(img,
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
max_pixel_value=255.0)
def to_tensor(x, **kwargs):
return torch.from_numpy(x.transpose(2, 0, 1).astype('float32'))
_transform = [
al.Lambda(image=norm),
al.Lambda(image=to_tensor, mask=to_tensor)
]
return al.Compose(_transform)
def build_test(self):
test_transforms = A.Compose([
A.Normalize(mean=self.mean, std=self.std),
AT.ToTensor(),
A.Lambda(lambda x: torch.cat([x, x, x], 0), always_apply=True)
])
return test_transforms
def get_validation_augmentation():
test_transform = [
albu.Lambda(image=pad_image_to_multiplys_of(32), mask=pad_image_to_multiplys_of(32))
# albu.LongestMaxSize(224),
# albu.Lambda(image=pad_image_to_multiplys_of(32), mask=pad_image_to_multiplys_of(32))
# albu.RandomCrop(height=320, width=320, always_apply=True)
]
return albu.Compose(test_transform)
def get_validation_augmentation(height, width, downsample_factor=1):
"""Add paddings to make image shape divisible by 32"""
downsample_fn = partial(downsample, factor=downsample_factor)
test_transform = [
albu.Lambda(image=downsample_fn, mask=downsample_fn),
albu.PadIfNeeded(min_height=height, min_width=width),
albu.CenterCrop(height=height, width=width)
]
return albu.Compose(test_transform)
def __init__(self, config, **kwargs):
super(DataManger_Epoch, self).__init__(config)
transform = dict()
transform['train'] = A.Compose([
A.Resize(config['image_size'][0], config['image_size'][1]),
A.PadIfNeeded(min_height=config['image_size'][0] + 10,
min_width=config['image_size'][1] + 10),
A.RandomCrop(config['image_size'][0], config['image_size'][1]),
A.HorizontalFlip(p=0.5),
A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
ToTensorV2(),
A.Lambda(image=RandomErasing(probability=0.5,
mean=[0.485, 0.456, 0.406])),
])
transform['val'] = A.Compose([
A.Resize(config['image_size'][0], config['image_size'][1]),
A.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
ToTensorV2()
])
transform['test'] = A.Compose([
A.Resize(config['image_size'][0], config['image_size'][1]),
A.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
ToTensorV2()
])
dataset = dict()
for _phase in self.datasource.get_phase():
dataset[_phase] = ImageDataset(self.datasource.get_data(_phase),
transform=transform[_phase])
self.dataloader['train'] = DataLoader(
dataset=dataset['train'],
batch_size=config['batch_size'],
shuffle=config['shuffle'],
num_workers=config['num_workers'],
pin_memory=config['pin_memory'],
drop_last=config['drop_last'])
self.dataloader['val'] = DataLoader(dataset=dataset['val'],
batch_size=config['batch_size'],
shuffle=False,
num_workers=config['num_workers'],
pin_memory=config['pin_memory'],
drop_last=config['drop_last'])
self.dataloader['test'] = DataLoader(dataset['test'],
batch_size=config['batch_size'],
shuffle=False,
drop_last=False)
def get_preprocessing(preprocessing_fn, apply_augmentation=False):
"""Construct preprocessing transform
Args:
preprocessing_fn (callable): data normalization function (can be specific for each pretrained neural network)
apply_augmentation (boolean): apply data augmentation or not
Return:
transform: albumentations.Compose
"""
_transform = [A.Resize(384, 480)]
if apply_augmentation:
_transform += [
A.HorizontalFlip(p=0.5),
A.ShiftScaleRotate(scale_limit=0.5,
rotate_limit=0,
shift_limit=0.1,
p=1,
border_mode=0),
A.IAAAdditiveGaussianNoise(p=0.2),
A.IAAPerspective(p=0.5),
A.OneOf(
[A.CLAHE(p=1),
A.RandomBrightness(p=1),
A.RandomGamma(p=1)],
p=0.9),
A.OneOf([
A.IAASharpen(p=1),
A.Blur(blur_limit=3, p=1),
A.MotionBlur(blur_limit=3, p=1)
],
p=0.9),
A.OneOf([A.RandomContrast(p=1),
A.HueSaturationValue(p=1)], p=0.9)
]
_transform += [
A.Lambda(image=preprocessing_fn),
A.Lambda(image=to_CHW, mask=to_CHW)
]
return A.Compose(_transform)
def numpy_to_torch():
conversion_transformation = albu.Compose([albu.Lambda(image=g.to_tensor)],
additional_targets={
'quaternion': 'image',
'scales': 'image',
'xy': 'image'
})
return conversion_transformation
def get_preprocessing(size, mean=(0., 0., 0.), std=(1., 1., 1.)):
preprocessing_fn = partial(
normalize_img, mean=mean, std=std,
)
transform = [
albu.Lambda(image=preprocessing_fn),
albu.Resize(size, size),
ToTensorV2(),
]
return albu.Compose(transform)
def sequence_preprocessing(self):
transform = A.Compose([
A.Lambda(image=self.aug_morph_close,
keypoint=self.aug_keypoints,
p=1.0),
],
additional_targets={
'image1': 'image',
'image2': 'image'
})
return transform
def default_train_transform(self) -> Callable:
"""
Default mnist training transforms. Performs
* Random Rotation between -5 and 5 degrees
* Normalization using mean: (0.1307,) and std: (0.3081,)
Returns:
Callable: The transform, an ``albumentations.Compose`` object.
"""
return A.Compose([
A.Lambda(ToNumpy),
A.Rotate(
limit=5
), # Randomly rotating the image in the range -5,5 degrees
A.Normalize(mean=mnist.mean, std=mnist.std,
max_pixel_value=1.0), # Normalizing
A.Lambda(ToTensor),
])
def get_training_augmentation(augment):
if augment > 0:
train_transform = [
A.GridDistortion(distort_limit=args.augment,
interpolation=cv2.INTER_AREA,
border_mode=cv2.BORDER_CONSTANT,
value=(240, 240, 240),
mask_value=0)
]
else:
train_transform = []
train_transform += [
#A.ShiftScaleRotate(rotate_limit=10, scale_limit=0, interpolation=cv2.INTER_AREA, border_mode=cv2.BORDER_CONSTANT, value=1),
#A.ElasticTransform(
# alpha=0.4, alpha_affine=0.3, interpolation=cv2.INTER_AREA, border_mode=cv2.BORDER_CONSTANT, value=(240, 240, 240), mask_value=0),
A.Rotate(limit=10,
interpolation=cv2.INTER_AREA,
border_mode=cv2.BORDER_CONSTANT,
value=(240, 240, 240),
mask_value=0),
#A.IAAAdditiveGaussianNoise(p=0.1),
#A.OneOf(
# [
# A.CLAHE(p=1),
# A.RandomBrightness(p=1),
# A.RandomGamma(p=1),
# ],
# p=0.9,
#),
#A.OneOf(
# [
# A.IAASharpen(p=1),
# A.Blur(blur_limit=3, p=1),
# ],
# p=0.9,
#),
A.OneOf(
[
A.RandomContrast(p=1),
A.HueSaturationValue(p=1),
],
p=0.9,
),
A.Lambda(mask=round_clip_0_1)
]
return A.Compose(train_transform)
def predict(model: nn.Module,
image: np.ndarray,
image_size,
tta=None,
normalize=A.Normalize(),
batch_size=1,
activation='sigmoid') -> np.ndarray:
model.eval()
tile_step = (image_size[0] // 2, image_size[1] // 2)
tile_slicer = ImageSlicer(image.shape,
image_size,
tile_step,
weight='pyramid')
tile_merger = CudaTileMerger(tile_slicer.target_shape, 1,
tile_slicer.weight)
patches = tile_slicer.split(image)
transform = A.Compose([normalize, A.Lambda(image=_tensor_from_rgb_image)])
if tta == 'fliplr':
model = TTAWrapperFlipLR(model)
print('Using FlipLR TTA')
if tta == 'd4':
model = TTAWrapperD4(model)
print('Using D4 TTA')
with torch.no_grad():
data = list({
'image': patch,
'coords': np.array(coords, dtype=np.int)
} for (patch, coords) in zip(patches, tile_slicer.crops))
for batch in DataLoader(InMemoryDataset(data, transform),
pin_memory=True,
batch_size=batch_size):
image = batch['image'].cuda(non_blocking=True)
coords = batch['coords']
mask_batch = model(image)
tile_merger.integrate_batch(mask_batch, coords)
mask = tile_merger.merge()
if activation == 'sigmoid':
mask = mask.sigmoid()
if isinstance(activation, float):
mask = F.relu(mask_batch - activation, inplace=True)
mask = np.moveaxis(to_numpy(mask), 0, -1)
mask = tile_slicer.crop_to_orignal_size(mask)
return mask
def get_training_augmentation(dim = 512, rot_limit = 45):
train_transform = [
A.HorizontalFlip(p=0.5),
A.ShiftScaleRotate(scale_limit=0.5, rotate_limit=rot_limit, shift_limit=0.1, p=1, border_mode=0),
A.PadIfNeeded(min_height=dim, min_width=dim, always_apply=True, border_mode=0),
A.RandomCrop(height=dim, width=dim, always_apply=True),
A.Lambda(mask=round_clip_0_1)
]
return A.Compose(train_transform)