def predefined_transform() -> None:
"""
Example from docs
https://github.com/albumentations-team/albumentations_examples/blob/master/notebooks/example.ipynb
:return:
"""
return A.Compose([
A.RandomRotate90(),
A.Flip(),
A.Transpose(),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.MotionBlur(p=.2),
A.MedianBlur(blur_limit=3, p=0.1),
A.Blur(blur_limit=3, p=0.1),
], p=0.2),
A.ShiftScaleRotate(shift_limit=0.0625, scale_limit=0.2, rotate_limit=45, p=0.2),
A.OneOf([
A.OpticalDistortion(p=0.3),
A.GridDistortion(p=.1),
A.IAAPiecewiseAffine(p=0.3),
], p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
A.RandomBrightnessContrast(),
], p=0.3),
A.HueSaturationValue(p=0.3),
])
def album(self): #이미지 변환
transform = A.Compose([
#A.RandomRotate90(),
A.Flip(p=0.2),
#A.Transpose(),
A.ChannelShuffle(p=0.3),
A.ElasticTransform(p=0.3,border_mode=cv2.BORDER_REFLECT_101,alpha_affine=40),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.MotionBlur(p=.2),
A.MedianBlur(blur_limit=3, p=0.1),
A.Blur(blur_limit=3, p=0.1),
], p=0.2),
A.ShiftScaleRotate(shift_limit=0.0625, scale_limit=0.2, rotate_limit=45, p=0.2),
A.OneOf([
A.OpticalDistortion(p=0.3),
A.GridDistortion(p=.1),
A.IAAPiecewiseAffine(p=0.3),
], p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
A.RandomBrightnessContrast(),
], p=0.3),
A.HueSaturationValue(p=0.3),
])
image = cv2.cvtColor(self.srcResize, cv2.COLOR_BGR2RGB)
transformed = transform(image=image)['image']
self.update(transformed)
def augmentation_hard(image, p=1., sub_p=0.3):
augmentation_fun = al.Compose([
al.ShiftScaleRotate(shift_limit=0.1,
scale_limit=0.1,
rotate_limit=8,
p=sub_p,
border_mode=cv2.BORDER_CONSTANT),
al.ElasticTransform(sub_p),
al.OneOf([
al.IAAAdditiveGaussianNoise(),
al.GaussNoise(),
], p=sub_p),
al.OneOf([
al.MotionBlur(p=sub_p),
al.MedianBlur(blur_limit=3, p=sub_p),
al.Blur(blur_limit=3, p=sub_p),
],
p=sub_p),
al.OneOf([
al.OpticalDistortion(p=sub_p),
al.GridDistortion(p=sub_p),
al.IAAPiecewiseAffine(p=sub_p),
],
p=sub_p),
al.OneOf([
al.CLAHE(clip_limit=3),
al.IAASharpen(),
al.IAAEmboss(),
al.RandomBrightnessContrast()
],
p=sub_p)
],
p=p)
return augmentation_fun(image)
def get_aug(mode="train"):
if mode=="Nor":
aug=A.Compose([
ToTensor(),
])
elif mode =="train":
print("train aug")
mean = (0.485,0.456,0.406)
std = (0.229,0.224,0.225)
aug=A.Compose([
A.Flip(),
A.ShiftScaleRotate(rotate_limit=1.0, p=0.8),
# Pixels
A.OneOf([
A.IAAEmboss(p=1.0),
A.IAASharpen(p=1.0),
A.Blur(p=1.0),
], p=0.5),
# Affine
A.OneOf([
A.ElasticTransform(p=1.0),
A.IAAPiecewiseAffine(p=1.0)
], p=0.5),
A.Normalize(mean=mean,std=std,max_pixel_value=255.0,always_apply=True),
])
else:
print("valid/test aug")
mean = (0.485,0.456,0.406)
std = (0.229,0.224,0.225)
aug=A.Compose([
A.Normalize(mean=mean,std=std,max_pixel_value=255.0,always_apply=True),
])
return aug
def augment_image(self, image):
transform = A.Compose([
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.3),
A.OneOf([
A.MotionBlur(p=.4),
A.MedianBlur(blur_limit=3, p=0.3),
A.Blur(blur_limit=3, p=0.3),
],
p=0.4),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
A.RandomBrightnessContrast(),
],
p=0.3),
A.HueSaturationValue(p=0.3),
])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
augmented_image = transform(image=image)['image']
augmented_image = cv2.cvtColor(augmented_image, cv2.COLOR_RGB2BGR)
return augmented_image
def augmentation_simple(image, p=1., sub_p=0.3):
augmentation_fun = al.Compose([
al.OneOf([
al.IAAAdditiveGaussianNoise(),
al.GaussNoise(),
], p=sub_p),
al.OneOf([
al.MotionBlur(p=sub_p),
al.MedianBlur(blur_limit=3, p=sub_p),
al.Blur(blur_limit=3, p=sub_p),
],
p=sub_p),
al.OneOf([
al.OpticalDistortion(p=sub_p),
al.GridDistortion(p=sub_p),
al.IAAPiecewiseAffine(p=sub_p),
],
p=sub_p),
al.OneOf([
al.CLAHE(clip_limit=3),
al.IAASharpen(),
al.IAAEmboss(),
al.RandomBrightnessContrast()
],
p=sub_p)
],
p=p)
return augmentation_fun(image)
def _strong_aug(p=0.5):
return albumentations.Compose([
albumentations.HorizontalFlip(),
albumentations.VerticalFlip(),
albumentations.ShiftScaleRotate(
shift_limit=0, scale_limit=0, rotate_limit=15, p=0.3),
albumentations.OneOf([
albumentations.OpticalDistortion(p=0.3),
albumentations.GridDistortion(p=0.1),
albumentations.IAAPiecewiseAffine(p=0.3),
],
p=0.2),
albumentations.OneOf([
albumentations.CLAHE(clip_limit=2),
albumentations.IAASharpen(),
albumentations.IAAEmboss(),
],
p=0.3),
albumentations.OneOf([
albumentations.RandomBrightnessContrast(p=0.3),
],
p=0.4),
albumentations.HueSaturationValue(p=0.3),
],
p=p)
def _strong_aug(p=0.5):
import albumentations
return albumentations.Compose([
albumentations.HorizontalFlip(p=0.5),
albumentations.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=0,
p=0.5,
border_mode=cv2.BORDER_CONSTANT),
albumentations.OneOf([
albumentations.OpticalDistortion(p=0.5,
border_mode=cv2.BORDER_CONSTANT),
albumentations.GridDistortion(p=0.5,
border_mode=cv2.BORDER_CONSTANT),
albumentations.IAAPiecewiseAffine(p=0.5),
albumentations.ElasticTransform(p=0.5,
border_mode=cv2.BORDER_CONSTANT),
],
p=0.5),
albumentations.OneOf([
albumentations.CLAHE(clip_limit=2),
albumentations.IAASharpen(),
albumentations.IAAEmboss(),
],
p=0.5),
albumentations.OneOf([
albumentations.RandomBrightnessContrast(p=0.5),
],
p=0.4),
albumentations.HueSaturationValue(p=0.5),
],
p=p)
def get_train_transform():
crop_height = 256
crop_width = 256
return albu.Compose([
albu.PadIfNeeded(min_height=crop_height, min_width=crop_width, p=1),
albu.RandomSizedCrop((int(0.3 * crop_height), 288), crop_height, crop_width, p=1),
albu.HorizontalFlip(p=0.5),
albu.OneOf([
albu.IAAAdditiveGaussianNoise(p=0.5),
albu.GaussNoise(p=0.5),
], p=0.2),
albu.OneOf([
albu.MotionBlur(p=0.2),
albu.MedianBlur(blur_limit=3, p=0.1),
albu.Blur(blur_limit=3, p=0.1),
], p=0.2),
albu.ShiftScaleRotate(shift_limit=0.0625, scale_limit=0, rotate_limit=20, p=0.1),
albu.OneOf([
albu.OpticalDistortion(p=0.3),
albu.GridDistortion(p=0.1),
albu.IAAPiecewiseAffine(p=0.3),
], p=0.2),
albu.OneOf([
albu.CLAHE(clip_limit=2, p=0.5),
albu.IAASharpen(p=0.5),
albu.IAAEmboss(p=0.5),
albu.RandomBrightnessContrast(p=0.5),
], p=0.3),
albu.HueSaturationValue(p=0.3),
albu.JpegCompression(p=0.2, quality_lower=20, quality_upper=99),
albu.ElasticTransform(p=0.1),
albu.Normalize(p=1)
], p=1)
def add_transforms(self):
if self.train:
self.transforms += [
A.Resize(int(self.img_size[0] * 1.1), int(self.img_size[1] * 1.1)),
A.RandomCrop(self.img_size[0], self.img_size[1]),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.5),
A.Rotate(p=0.5, border_mode=BORDER_REFLECT, value=0),
# Pixels
A.OneOf([
A.IAAEmboss(p=1.0),
A.IAASharpen(p=1.0),
A.Blur(p=1.0),
], p=0.5),
# Affine
A.OneOf([
A.ElasticTransform(p=1.0),
A.IAAPiecewiseAffine(p=1.0)
], p=0.5),
]
else:
self.transforms += [
A.Resize(self.img_size[0], self.img_size[1]),
]
def augment(image):
transform = A.Compose([
A.HorizontalFlip(p=0.5),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.MotionBlur(p=.2),
A.MedianBlur(blur_limit=3, p=0.1),
A.Blur(blur_limit=3, p=0.1),
],
p=0.2),
A.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=15,
p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
A.RandomBrightnessContrast(),
A.RandomGamma(),
],
p=0.5),
A.HueSaturationValue(p=0.3),
])
augmented_image = transform(image=image)['image']
return augmented_image
def albumentation():
transform = albumentations.Compose([
albumentations.OneOf([
albumentations.GaussNoise(),
albumentations.IAAAdditiveGaussianNoise()
]),
albumentations.OneOf([
albumentations.MotionBlur(blur_limit=3, p=0.2),
albumentations.MedianBlur(blur_limit=3, p=0.1),
albumentations.Blur(blur_limit=2, p=0.1)
]),
albumentations.OneOf([
albumentations.RandomBrightness(limit=(0.1, 0.4)),
albumentations.HueSaturationValue(hue_shift_limit=(0, 128), sat_shift_limit=(0, 60), val_shift_limit=(0, 20)),
albumentations.RGBShift(r_shift_limit=30, g_shift_limit=30, b_shift_limit=30)
]),
albumentations.OneOf([
albumentations.CLAHE(),
albumentations.ChannelShuffle(),
albumentations.IAASharpen(),
albumentations.IAAEmboss(),
albumentations.RandomBrightnessContrast(),
]),
albumentations.OneOf([
albumentations.RandomGamma(gamma_limit=(35,255)),
albumentations.OpticalDistortion(),
albumentations.GridDistortion(),
albumentations.IAAPiecewiseAffine()
]),
A_torch.ToTensor(normalize={
"mean": [0.485, 0.456, 0.406],
"std" : [0.229, 0.224, 0.225]})
])
return transform
def augmentations(image_size: int):
channel_augs = [
A.HueSaturationValue(p=0.5),
A.ChannelShuffle(p=0.5),
]
result = [
# *pre_transform(image_size),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.5),
A.OneOf([
A.MotionBlur(blur_limit=3, p=0.7),
A.MedianBlur(blur_limit=3, p=1.0),
A.Blur(blur_limit=3, p=0.7),
],
p=0.5),
A.OneOf(channel_augs),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
],
p=0.5),
A.RandomBrightnessContrast(brightness_limit=0.5,
contrast_limit=0.5,
p=0.5),
A.RandomGamma(p=0.5),
A.OneOf([A.MedianBlur(p=0.5), A.MotionBlur(p=0.5)]),
A.RandomGamma(gamma_limit=(85, 115), p=0.5),
]
return A.Compose(result, bbox_params=BBOX_PARAMS)
def weak_aug(self, p=0.5):
'''Create a weakly augmented image framework'''
return A.Compose([
A.HorizontalFlip(),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.MotionBlur(p=0.2),
A.MedianBlur(blur_limit=3, p=0.1),
A.Blur(blur_limit=3, p=0.1),
],
p=0.2),
A.ShiftScaleRotate(
shift_limit=0.0625, scale_limit=0.2, rotate_limit=10, p=0.2),
A.OpticalDistortion(p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
],
p=0.3),
],
p=p)
def augment_image(image):
# Works with single image
transform = A.Compose([
A.HorizontalFlip(),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.MotionBlur(),
A.MedianBlur(blur_limit=3),
A.Blur(blur_limit=3),
],
p=0.2),
A.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=45,
p=0.2),
A.OneOf([
A.OpticalDistortion(p=0.3),
], p=0.2),
A.OneOf([
A.IAASharpen(p=1.),
A.IAAEmboss(p=1.),
A.RandomBrightnessContrast(p=1.),
],
p=0.3),
A.HueSaturationValue(hue_shift_limit=5,
sat_shift_limit=5,
val_shift_limit=5,
p=0.3),
])
return transform(image=image)['image']
def strong_aug(p=.5):
return A.Compose([
A.RandomRotate90(),
A.Flip(),
A.Transpose(),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.MotionBlur(p=.2),
A.MedianBlur(blur_limit=3, p=.1),
A.Blur(blur_limit=3, p=.1),
],
p=0.2),
A.ShiftScaleRotate(
shift_limit=0.0625, scale_limit=0.2, rotate_limit=45, p=.2),
A.OneOf([
A.OpticalDistortion(p=0.3),
A.GridDistortion(p=.1),
A.IAAPiecewiseAffine(p=0.3),
],
p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
A.RandomContrast(),
A.RandomBrightness(),
],
p=0.3),
A.HueSaturationValue(p=0.3),
],
p=p)
def get_aug(mode="train"):
if mode == "Nor":
aug = A.Compose([
ToTensor(),
])
elif mode == "train":
aug = A.Compose([
A.Flip(),
A.ShiftScaleRotate(rotate_limit=1.0, p=0.8),
# Pixels
A.OneOf([
A.IAAEmboss(p=1.0),
A.IAASharpen(p=1.0),
A.Blur(p=1.0),
],
p=0.5),
# Affine
A.OneOf(
[A.ElasticTransform(p=1.0),
A.IAAPiecewiseAffine(p=1.0)],
p=0.5),
A.Normalize(p=1.0),
ToTensor(),
])
else:
aug = A.Compose([
A.Normalize(p=1.0),
ToTensor(),
])
return aug
def augmentation(image_size, train=True):
max_crop = image_size // 5
if train:
data_transform = A.Compose([
A.Resize(image_size, image_size),
A.Compose([
A.OneOf([
A.RandomRain(p=0.1),
A.GaussNoise(mean=15),
A.GaussianBlur(blur_limit=10, p=0.4),
A.MotionBlur(p=0.2)
]),
A.OneOf([
A.RGBShift(p=1.0,
r_shift_limit=(-10, 10),
g_shift_limit=(-10, 10),
b_shift_limit=(-10, 10)),
A.RandomBrightnessContrast(
brightness_limit=0.3, contrast_limit=0.1, p=1),
A.HueSaturationValue(hue_shift_limit=20, p=1),
],
p=0.6),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
]),
A.OneOf([A.IAAPerspective(p=0.3),
A.ElasticTransform(p=0.1)]),
A.OneOf([
A.Rotate(limit=25, p=0.6),
A.IAAAffine(
scale=0.9,
translate_px=15,
rotate=25,
shear=0.2,
)
],
p=1),
A.Cutout(num_holes=1,
max_h_size=max_crop,
max_w_size=max_crop,
p=0.2)
],
p=1),
A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
AT.ToTensor()
])
else:
data_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((image_size, image_size)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
return data_transform
def get_transforms(size: int, scope: str = 'geometric', crop='random'):
augs = {
'strong':
albu.Compose([
albu.HorizontalFlip(),
albu.ShiftScaleRotate(shift_limit=0.0,
scale_limit=0.2,
rotate_limit=20,
p=.4),
albu.ElasticTransform(),
albu.OpticalDistortion(),
albu.OneOf([
albu.CLAHE(clip_limit=2),
albu.IAASharpen(),
albu.IAAEmboss(),
albu.RandomBrightnessContrast(),
albu.RandomGamma()
],
p=0.5),
albu.OneOf([
albu.RGBShift(),
albu.HueSaturationValue(),
], p=0.5),
]),
'weak':
albu.Compose([
albu.HorizontalFlip(),
]),
'geometric':
albu.OneOf([
albu.HorizontalFlip(always_apply=True),
albu.ShiftScaleRotate(always_apply=True),
albu.Transpose(always_apply=True),
albu.OpticalDistortion(always_apply=True),
albu.ElasticTransform(always_apply=True),
])
}
crop_fn = {
'random': albu.RandomCrop(size, size, always_apply=True),
'center': albu.CenterCrop(size, size, always_apply=True)
}[crop]
pad = albu.PadIfNeeded(size, size)
if scope == "nothing":
pipeline = albu.Compose([crop_fn, pad],
additional_targets={'target': 'image'})
else:
aug_fn = augs[scope]
pipeline = albu.Compose([aug_fn, crop_fn, pad],
additional_targets={'target': 'image'})
def process(a, b):
r = pipeline(image=a, target=b)
return r['image'], r['target']
return process
def get_training_augmentation(min_area=0., min_visibility=0.):
train_transform = [
albu.OneOf([
albu.MotionBlur(p=.2),
albu.MedianBlur(blur_limit=3, p=0.1),
albu.Blur(blur_limit=3, p=0.1),
],
p=0.2),
albu.ShiftScaleRotate(shift_limit=0,
scale_limit=0,
rotate_limit=15,
p=0.5),
albu.OneOf([
albu.CLAHE(clip_limit=2),
albu.IAASharpen(),
albu.IAAEmboss(),
],
p=0.3),
albu.OneOf([
albu.RandomFog(fog_coef_lower=0.1, fog_coef_upper=0.15, p=0.1),
albu.RandomShadow(p=0.1),
albu.RandomBrightness(limit=0.3, p=0.2),
albu.RandomRain(slant_lower=0,
slant_upper=8,
drop_length=0,
blur_value=4,
brightness_coefficient=0.8,
rain_type='heavy',
p=0.1),
albu.RandomSunFlare(p=0.2),
]),
albu.OneOf([
albu.RGBShift(p=0.1),
albu.HueSaturationValue(p=0.3),
]),
albu.OneOf([
albu.HorizontalFlip(p=0.5),
albu.RandomSizedCrop(min_max_height=(720, 1380),
height=1380,
width=720,
interpolation=cv2.INTER_AREA)
],
p=0.2)
]
return albu.Compose(train_transform,
bbox_params={
'format': 'coco',
'min_area': min_area,
'min_visibility': min_visibility,
'label_fields': ['category_id']
})
def run_train(data_path: Path):
cfg = BasicConfig(
seed=444,
name='test_center_vgg_retrain2',
num_workers=4,
gpus=(0,),
batch_size=24,
num_epochs=15,
steps=(4, 9, np.inf),
warmup_epoch=1,
cooldown_epoch=2,
train_augmentations=alb.Compose([
alb.Resize(128, 128),
alb.OneOf([
alb.MotionBlur(blur_limit=5, p=0.2),
alb.MedianBlur(blur_limit=3, p=0.1),
alb.Blur(blur_limit=5, p=0.1)
], p=0.2),
alb.OneOf([
alb.ImageCompression(70, compression_type=alb.ImageCompression.ImageCompressionType.JPEG),
alb.ImageCompression(70, compression_type=alb.ImageCompression.ImageCompressionType.WEBP)
], p=0.2),
alb.OneOf([
alb.CLAHE(clip_limit=2),
alb.IAASharpen(),
alb.IAAEmboss(),
alb.RandomBrightnessContrast(),
], p=0.1),
alb.Rotate(5, border_mode=cv2.BORDER_REFLECT, p=0.2),
alb.OneOf([
alb.RandomResizedCrop(112, 112, scale=(0.9, 1.0), ratio=(0.8, 1.1), p=0.5),
alb.Resize(112, 112, p=0.5),
], p=1.0),
alb.HorizontalFlip(p=0.5),
alb.HueSaturationValue(p=0.7),
alb.ChannelShuffle(p=0.5)
]),
normalize=True,
weight_normalize=True,
uniform_subjects=True,
classifier_mult=3,
lr_factor=0.1,
initial_lr=1e-2,
# extra_rec=(Path('/run/media/andrey/Fast/FairFace/faces_emore/train.rec'),)
)
np.random.seed(cfg.seed)
mx.random.seed(cfg.seed)
train_df = load_info(data_path, Path('data/train_df.csv'))
lib.train.train(cfg, train_df)
def get_transforms(config):
size = config['size']
scope = config['scope']
crop = config['crop']
p = config['p']
augs = {'strong': albu.Compose([albu.HorizontalFlip(),
albu.VerticalFlip(),
albu.ShiftScaleRotate(shift_limit=0.0, scale_limit=0.2, rotate_limit=20, p=.4),
albu.ElasticTransform(),
albu.OpticalDistortion(),
albu.OneOf([
albu.CLAHE(clip_limit=2),
albu.IAASharpen(),
albu.IAAEmboss(),
albu.RandomBrightnessContrast(),
albu.RandomGamma()
], p=0.5),
albu.OneOf([
albu.RGBShift(),
albu.HueSaturationValue(),
], p=0.5),
]),
'weak': albu.Compose([albu.HorizontalFlip(),
]),
'geometric': albu.OneOf([albu.HorizontalFlip(p=p),
albu.VerticalFlip(p=p),
albu.ShiftScaleRotate(p=p),
albu.Transpose(p=p),
# albu.OpticalDistortion(p=p),
# albu.ElasticTransform(p=p),
])
}
aug_fn = augs[scope]
crop_fn = {'random': albu.RandomCrop(size[0], size[1], always_apply=True),
'center': albu.CenterCrop(size[0], size[1], always_apply=True), 'None': None}[crop]
pad = albu.PadIfNeeded(size[0], size[1])
pipeline = albu.Compose([aug_fn, crop_fn, pad])
def process(a, b):
r = pipeline(image=a, mask=b)
return r['image'], r['mask']
return process
def __init__(self):
self.policy = A.Compose([
A.OneOf([
A.Rotate(180),
A.Flip(),
], p=0.3),
A.ShiftScaleRotate(shift_limit=0.2, scale_limit=0.5, rotate_limit=0, p=0.2),
A.OneOf([
A.CoarseDropout(max_holes=16, max_height=16, max_width=16, p=0.3),
A.GridDropout(ratio=0.3, p=0.3),
]),
A.OneOf([
A.ElasticTransform(sigma=10, alpha_affine=25, p=0.3),
A.RandomFog(fog_coef_lower=0.2, fog_coef_upper=0.7, p=0.2),
], p=0.2),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
A.ISONoise()
], p=0.2),
A.OneOf([
A.MotionBlur(p=.3),
A.MedianBlur(blur_limit=5, p=0.3),
A.Blur(blur_limit=5, p=0.3),
A.GaussianBlur(p=0.3)
], p=0.2),
A.OneOf([
A.ChannelShuffle(p=.3),
A.HueSaturationValue(p=0.3),
A.ToGray(p=0.3),
A.ChannelDropout(p=0.3),
A.InvertImg(p=0.1)
], p=0.2),
A.OneOf([
A.OpticalDistortion(p=0.3),
A.GridDistortion(p=.2),
A.IAAPiecewiseAffine(p=0.3),
], p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
], p=0.2),
A.RandomBrightnessContrast(brightness_limit=0.3, contrast_limit=0.3, p=0.3),
A.Solarize(p=0.2),
])
def hard_transforms(crop_size=512):
return albu.Compose([
albu.ShiftScaleRotate(shift_limit=0,
scale_limit=0.1,
rotate_limit=180,
border_mode=cv2.BORDER_CONSTANT,
value=0,
mask_value=0),
albu.CropNonEmptyMaskIfExists(crop_size, crop_size, p=1),
albu.RandomShadow(shadow_roi=(0, 0, 1, 1),
num_shadows_lower=1,
num_shadows_upper=4,
shadow_dimension=7,
always_apply=False,
p=0.5),
albu.HueSaturationValue(p=0.3),
albu.OneOf([
albu.IAAAdditiveGaussianNoise(),
albu.GaussNoise(),
albu.MultiplicativeNoise(
multiplier=[0.5, 1.5], per_channel=True, p=1)
],
p=0.3),
albu.OneOf([
albu.MotionBlur(p=0.2),
albu.MedianBlur(blur_limit=3, p=0.1),
albu.Blur(blur_limit=3, p=0.1),
],
p=0.2),
albu.OneOf([
albu.CLAHE(clip_limit=2),
albu.IAASharpen(),
albu.IAAEmboss(),
albu.RandomBrightnessContrast(
brightness_limit=0.2, contrast_limit=0.2, p=0.3),
albu.RandomGamma(gamma_limit=(85, 115), p=0.3),
],
p=0.3),
albu.JpegCompression(quality_lower=40, quality_upper=100, p=0.5),
albu.Cutout(
num_holes=25, max_h_size=5, max_w_size=5, fill_value=0, p=0.3),
],
p=1)
def _pre_aug(p=0.5):
return albumentations.Compose([
albumentations.HorizontalFlip(),
albumentations.ShiftScaleRotate(shift_limit=0, scale_limit=0, rotate_limit=15, p=0.8),
albumentations.GridDistortion(distort_limit=0.3, p=0.3),
albumentations.OneOf([
albumentations.MotionBlur(p=0.5),
albumentations.Blur(blur_limit=3, p=0.2),
], p=0.2),
albumentations.OneOf([
albumentations.CLAHE(clip_limit=2),
albumentations.IAASharpen(),
albumentations.IAAEmboss(),
], p=0.3),
albumentations.OneOf([
albumentations.RandomBrightnessContrast(p=0.3),
], p=0.4),
albumentations.HueSaturationValue(p=0.3),
], p=p)
def __init__(self, phase, train_file, image_file_path, image_height, image_width, mean, std, binclass):
self.image_file_path = image_file_path
df = pd.read_csv(train_file)
if binclass == 'A':
class_map = {'A':1,'B':0,'C':0}
elif binclass == 'B':
class_map = {'A':0,'B':1,'C':0}
elif binclass == 'C':
class_map = {'A':0,'B':0,'C':1}
else:
class_map = {'A':0,'B':1,'C':2}
self.img_id = df['image_id'].apply(lambda x: x.split('.')[0]).values # just take id of image_id
self.labels = df['label'].apply(lambda x: x[-1]).map(class_map).values # encoding labels
if phase == 'valid':
# validation set
self.aug = albumentations.Compose([
albumentations.Resize(image_height, image_width),
albumentations.Normalize(mean, std),
# albumentations.ToFloat()
])
elif phase == 'train':
# training set
self.aug = albumentations.Compose([
albumentations.Resize(image_height, image_width),
albumentations.RandomRotate90(p=0.5),
albumentations.Transpose(p=0.5),
albumentations.Flip(p=0.5),
albumentations.OneOf([
albumentations.CLAHE(clip_limit=2), albumentations.IAASharpen(), albumentations.IAAEmboss(),
albumentations.RandomBrightness(), albumentations.RandomContrast(),
albumentations.JpegCompression(), albumentations.Blur(), albumentations.GaussNoise()], p=0.5),
# albumentations.HueSaturationValue(p=0.5),
albumentations.ShiftScaleRotate(shift_limit=0.15, scale_limit=0.15, rotate_limit=45, p=0.5),
albumentations.Normalize(mean, std),
# albumentations.ToFloat()
])
def classification_augmentation():
"""Rationalized version of the one above.
"""
return alb.Compose(
[
alb.HorizontalFlip(p=.5), # 128 mus
alb.OneOf(
[
# These two do the same thing I think. Keeping only
# the faster one.
alb.IAAAdditiveGaussianNoise(p=1.), # 484 mus
# alb.GaussNoise(p=1.), # 1.11 ms
],
p=0.2), # 1.03 ms with both
alb.OneOf([
alb.MotionBlur(p=1.),
alb.MedianBlur(blur_limit=3, p=1.),
alb.Blur(blur_limit=3, p=1.),
],
p=0.2), # 40 mus
alb.ShiftScaleRotate(
shift_limit=0.1, scale_limit=0.1, rotate_limit=30, p=0.8),
# (above) 348 mus
alb.OneOf(
[
alb.OpticalDistortion(p=1.), # 95 mus
alb.GridDistortion(p=1.), # 101 mus
# alb.IAAPiecewiseAffine(p=1.), # 5.61 ms
],
p=0.2), # 2.48 ms -> 113 mus with the 2 first ones
alb.OneOf([
alb.CLAHE(clip_limit=2, p=1.),
alb.IAASharpen(p=1.),
alb.IAAEmboss(p=1.),
alb.RandomContrast(p=1.),
alb.RandomBrightness(p=1.),
],
p=0.3), # 257 mus
alb.HueSaturationValue(p=0.3), # 395 mus
],
p=0.9) # 3.84 ms -> 1.52 ms
def hard_transforms():
result = [
# Random shifts, stretches and turns with a 50% probability
A.Flip(),
A.ShiftScaleRotate(rotate_limit=1.0, p=0.8),
# Pixels
A.OneOf([
A.IAAEmboss(p=1.0),
A.IAASharpen(p=1.0),
A.Blur(p=1.0),
], p=0.5),
# Affine
A.OneOf([
A.ElasticTransform(p=1.0),
A.IAAPiecewiseAffine(p=1.0)
], p=0.5),
]
return result
def data_augmentation(image_path, aug_num):
image = cv2.imread(image_path, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
augmentation = A.Compose(
[
A.RandomRotate90(),
A.Flip(),
A.Transpose(),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.MotionBlur(p=0.2),
A.MedianBlur(blur_limit=3, p=0.1),
A.Blur(blur_limit=3, p=0.1),
],
p=0.2),
A.ShiftScaleRotate(
shift_limit=0.0625, scale_limit=0.2, rotate_limit=45, p=0.2),
A.OneOf([
A.OpticalDistortion(p=0.3),
A.GridDistortion(p=0.1),
A.IAAPiecewiseAffine(p=0.3),
],
p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
A.RandomBrightnessContrast(),
],
p=0.3),
# A.HueSaturationValue(p=0.3),
],
p=0.5)
patches = []
for _ in range(aug_num):
patches.append(augmentation(image=image)['image'])
return patches
def _setup_transform(self, cfg):
# Albumentation example: https://albumentations.readthedocs.io/en/latest/examples.html
self.img_mask_transform = A.Compose([
A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.2, rotate_limit=175, p=0.8, border_mode=cv2.BORDER_CONSTANT),
A.Flip(),
A.Transpose(),
A.OneOf([
A.ElasticTransform(),
A.OpticalDistortion(),
A.GridDistortion(),
A.IAAPiecewiseAffine(),
]),
A.OneOf([
A.RandomCrop(height=self.size_crop,width=self.size_crop,p=0.5),
A.CenterCrop(height=self.size_crop,width=self.size_crop,p=0.5)
]),
A.Cutout(num_holes=8, max_h_size=8, max_w_size=8, fill_value=0,p=0.5),
],p=0.8)
self.img_pixel_transform = A.Compose([
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.MotionBlur(p=0.2),
A.MedianBlur(blur_limit=3, p=0.1),
A.Blur(blur_limit=3, p=0.1),
], p=0.2),
A.OneOf([
A.IAASharpen(),
A.IAAEmboss(),
# A.RandomBrightnessContrast(),
], p=0.3),
A.HueSaturationValue(hue_shift_limit=3,sat_shift_limit=20,val_shift_limit=3 ,p=0.2),
],p=0.5)
# Torch transform
self.resize_transform = transforms.Resize(cfg.MODEL.IMAGE_SIZE, Image.NEAREST)
self.to_tensor_transform = transforms.ToTensor()
self.normalize_transform = transforms.Normalize(mean=cfg.TRAIN.NORMALIZE_MEAN, std=cfg.TRAIN.NORMALIZE_STD)