ds_train = Vinmec(folder=args.data,
is_train='train',
fname='train.csv',
types=args.types,
pathology=args.pathology,
resize=int(args.shape))
# ds_train = ConcatData([ds_chexpert, ds_vinmec])
ag_train = [
imgaug.ColorSpace(mode=cv2.COLOR_GRAY2RGB),
imgaug.Affine(shear=10),
imgaug.RandomChooseAug([
imgaug.Albumentations(AB.Blur(blur_limit=4, p=0.25)),
imgaug.Albumentations(AB.MotionBlur(blur_limit=4, p=0.25)),
imgaug.Albumentations(AB.MedianBlur(blur_limit=4, p=0.25)),
]),
imgaug.RandomOrderAug(
[imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from
# fb.resnet.torch
imgaug.Lighting(0.1,
eigval=np.asarray(
[0.2175, 0.0188, 0.0045][::-1]) * 255.0,
eigvec=np.array(
[[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1]
def load_data(fold: int, params: Dict[str, Any]) -> Any:
torch.multiprocessing.set_sharing_strategy('file_system')
cudnn.benchmark = True
logger.info('Options:')
logger.info(pprint.pformat(opt))
full_df = pd.read_csv(opt.TRAIN.CSV)
print('full_df', full_df.shape)
train_df, val_df = train_val_split(full_df, fold)
print('train_df', train_df.shape, 'val_df', val_df.shape)
test_df = pd.read_csv(opt.TEST.CSV)
# transform_train = transforms.Compose([
# # transforms.Resize((opt.MODEL.IMAGE_SIZE)), # smaller edge
# transforms.RandomCrop(opt.MODEL.INPUT_SIZE),
# transforms.RandomHorizontalFlip(),
# # transforms.ColorJitter(brightness=0.2, contrast=0.2),
# # transforms.RandomAffine(degrees=20, scale=(0.8, 1.2), shear=10, resample=PIL.Image.BILINEAR),
# # transforms.RandomCrop(opt.MODEL.INPUT_SIZE),
# ])
augs = []
augs.append(albu.HorizontalFlip(.5))
if int(params['vflip']):
augs.append(albu.VerticalFlip(.5))
if int(params['rotate90']):
augs.append(albu.RandomRotate90())
if params['affine'] == 'soft':
augs.append(
albu.ShiftScaleRotate(shift_limit=0.075,
scale_limit=0.15,
rotate_limit=10,
p=.75))
elif params['affine'] == 'medium':
augs.append(
albu.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=45,
p=0.2))
elif params['affine'] == 'hard':
augs.append(
albu.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.50,
rotate_limit=45,
p=.75))
if float(params['noise']) > 0.1:
augs.append(
albu.OneOf([
albu.IAAAdditiveGaussianNoise(),
albu.GaussNoise(),
],
p=float(params['noise'])))
if float(params['blur']) > 0.1:
augs.append(
albu.OneOf([
albu.MotionBlur(p=.2),
albu.MedianBlur(blur_limit=3, p=0.1),
albu.Blur(blur_limit=3, p=0.1),
],
p=float(params['blur'])))
if float(params['distortion']) > 0.1:
augs.append(
albu.OneOf([
albu.OpticalDistortion(p=0.3),
albu.GridDistortion(p=.1),
albu.IAAPiecewiseAffine(p=0.3),
],
p=float(params['distortion'])))
if float(params['color']) > 0.1:
augs.append(
albu.OneOf([
albu.CLAHE(clip_limit=2),
albu.IAASharpen(),
albu.IAAEmboss(),
albu.RandomBrightnessContrast(),
],
p=float(params['color'])))
transform_train = albu.Compose([
albu.PadIfNeeded(opt.MODEL.INPUT_SIZE, opt.MODEL.INPUT_SIZE),
albu.RandomCrop(height=opt.MODEL.INPUT_SIZE,
width=opt.MODEL.INPUT_SIZE),
albu.Compose(augs, p=float(params['aug_global_prob'])),
])
if opt.TEST.NUM_TTAS > 1:
transform_test = albu.Compose([
albu.PadIfNeeded(opt.MODEL.INPUT_SIZE, opt.MODEL.INPUT_SIZE),
albu.RandomCrop(height=opt.MODEL.INPUT_SIZE,
width=opt.MODEL.INPUT_SIZE),
albu.HorizontalFlip(),
])
else:
transform_test = albu.Compose([
albu.PadIfNeeded(opt.MODEL.INPUT_SIZE, opt.MODEL.INPUT_SIZE),
albu.CenterCrop(height=opt.MODEL.INPUT_SIZE,
width=opt.MODEL.INPUT_SIZE),
])
train_dataset = Dataset(train_df,
path=opt.TRAIN.PATH,
mode='train',
num_classes=opt.MODEL.NUM_CLASSES,
resize=False,
augmentor=transform_train)
val_dataset = Dataset(
val_df,
path=opt.TRAIN.PATH,
mode='val',
# image_size=opt.MODEL.INPUT_SIZE,
num_classes=opt.MODEL.NUM_CLASSES,
resize=False,
num_tta=1, # opt.TEST.NUM_TTAS,
augmentor=transform_test)
test_dataset = Dataset(
test_df,
path=opt.TEST.PATH,
mode='test',
# image_size=opt.MODEL.INPUT_SIZE,
num_classes=opt.MODEL.NUM_CLASSES,
resize=False,
num_tta=opt.TEST.NUM_TTAS,
augmentor=transform_test)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=opt.TRAIN.BATCH_SIZE,
shuffle=True,
num_workers=opt.TRAIN.WORKERS)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=opt.TRAIN.BATCH_SIZE,
shuffle=False,
num_workers=opt.TRAIN.WORKERS)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=opt.TRAIN.BATCH_SIZE,
shuffle=False,
num_workers=opt.TRAIN.WORKERS)
return train_loader, val_loader, test_loader
def get_transform(size, transform_type="weak", min_visibility=0):
"""Creates transformation for COCO dataset
Args:
size (int): image size to return
transform_type (str):
'weak': resizes and normalizes image and bbox
'strong': performs different image effects, resizes and normalizes image and bbox
min_visibility (int): minimum fraction of area for a bounding box
Returns:
albu.core.transforms_interface.BasicTransform: image and bbox transformation
"""
bbox_params = {
'format': 'coco',
'min_visibility': min_visibility,
'label_fields': ['category_id']
}
augs = {
'strong':
albu.Compose(
[
albu.Resize(size, size),
albu.HorizontalFlip(),
# albu.VerticalFlip(p=0.1),
albu.ShiftScaleRotate(shift_limit=0.05,
scale_limit=0.15,
rotate_limit=20,
p=.4,
border_mode=cv2.BORDER_CONSTANT),
albu.OneOf([
albu.Blur(),
albu.MotionBlur(),
albu.MedianBlur(),
albu.GaussianBlur(),
],
p=0.2),
albu.OneOf([
albu.GaussNoise(var_limit=(10, 35)),
albu.IAAAdditiveGaussianNoise(),
albu.JpegCompression(quality_lower=50),
],
p=0.2),
albu.OneOf(
[
albu.RandomRain(),
# albu.RandomSunFlare(),
albu.RandomShadow()
],
p=0.15),
albu.OneOf([
albu.CLAHE(clip_limit=2),
albu.IAASharpen(alpha=(0.1, 0.3)),
albu.IAAEmboss(alpha=(0.1, 0.4)),
albu.RandomGamma(),
albu.RandomBrightnessContrast(),
],
p=0.2),
# albu.HueSaturationValue(p=0.25),
],
bbox_params=bbox_params),
'weak':
albu.Compose(
[
albu.Resize(size, size),
# albu.HorizontalFlip(),
],
bbox_params=bbox_params),
}
aug_fn = augs[transform_type]
normalize = albu.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
pipeline = albu.Compose([
aug_fn,
# ])
normalize
])
return pipeline
import os
import json
import glob
import SimpleITK as sitk
import pandas as pd
import cv2
import albumentations as A
from aug_transforms import aug_transform
a_transform = A.Compose([
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.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
A.RandomBrightnessContrast(),
],
p=0.3)
])
"lookahead":False,
"k_param":5,
"alpha_param":0.5,
"patience":3,
"DEBUG":False,
}
args['trn_transforms'] = A.Compose([
# A.Transpose(p=0.5),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.5),
# A.RandomBrightness(limit=0.2, p=0.75),
# A.RandomContrast(limit=0.2, p=0.75),
A.OneOf([
A.MotionBlur(blur_limit=5),
A.MedianBlur(blur_limit=5),
A.GaussianBlur(blur_limit=5),
A.GaussNoise(var_limit=(5.0, 30.0)),
], p=0.7),
# A.OneOf([
# A.OpticalDistortion(distort_limit=1.0),
# A.GridDistortion(num_steps=5, distort_limit=1.),
# A.ElasticTransform(alpha=3),
# ], p=0.7),
A.CLAHE(clip_limit=4.0, p=0.7),
# A.HueSaturationValue(hue_shift_limit=10, sat_shift_limit=20, val_shift_limit=10, p=0.5),
# A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.1, rotate_limit=15, border_mode=0, p=0.85),
# A.Cutout(max_h_size=int(256 * 0.375), max_w_size=int(256 * 0.375), num_holes=1, p=0.7),
])
args['val_transforms'] = A.Compose([
N = 2
M = 3
train_transforms = A.Compose([
A.CenterCrop(img_size[0], img_size[1], p=1),
# A.RandomResizedCrop(img_size[0], img_size[1], p=1.0),
# A.Resize(img_size[0], img_size[1], p=1.0),
RandAugment(N, M),
A.OneOf(
[A.RandomBrightness(limit=0.1, p=1),
A.RandomContrast(limit=0.1, p=1)],
p=0.5),
A.OneOf(
[
A.MotionBlur(blur_limit=3),
A.MedianBlur(blur_limit=3),
A.GaussianBlur(blur_limit=3),
],
p=0.5,
),
A.VerticalFlip(p=0.5),
A.HorizontalFlip(p=0.5),
A.ShiftScaleRotate(
shift_limit=0.2,
scale_limit=0.2,
rotate_limit=20,
interpolation=cv2.INTER_LINEAR,
border_mode=cv2.BORDER_REFLECT_101,
p=.75,
),
A.Cutout(num_holes=8, max_h_size=64, max_w_size=64, fill_value=0, p=0.5),
def get_transforms(image_size):
'''
albumentations 라이브러리 사용함
https://github.com/albumentations-team/albumentations
TODO: FAST AUTO AUGMENT
https://github.com/kakaobrain/fast-autoaugment
DATASET의 AUGMENT POLICY를 탐색해주는 알고리즘
TODO: Unsupervised Data Augmentation for Consistency Training
https://github.com/google-research/uda
TODO: Cutmix vs Mixup vs Gridmask vs Cutout
https://www.kaggle.com/saife245/cutmix-vs-mixup-vs-gridmask-vs-cutout
'''
transforms_train = albumentations.Compose([
# albumentations.Transpose(p=0.5),
# albumentations.VerticalFlip(p=0.5),
albumentations.HorizontalFlip(p=0.5),
albumentations.RandomBrightness(limit=0.2, p=0.75),
albumentations.RandomContrast(limit=0.2, p=0.75),
# one of 의 경우 하나를 랜덤하게 뽑아서 쓰게 해줌
albumentations.OneOf([
albumentations.MotionBlur(blur_limit=5),
albumentations.MedianBlur(blur_limit=5),
albumentations.GaussianBlur(blur_limit=5),
albumentations.GaussNoise(var_limit=(5.0, 30.0)),
],
p=0.7),
albumentations.OneOf([
albumentations.OpticalDistortion(distort_limit=1.0),
albumentations.GridDistortion(num_steps=5, distort_limit=1.),
albumentations.ElasticTransform(alpha=3),
],
p=0.7),
albumentations.CLAHE(clip_limit=4.0, p=0.7),
albumentations.HueSaturationValue(hue_shift_limit=10,
sat_shift_limit=20,
val_shift_limit=10,
p=0.5),
albumentations.ShiftScaleRotate(shift_limit=0.1,
scale_limit=0.1,
rotate_limit=15,
border_mode=0,
p=0.85),
albumentations.Resize(image_size, image_size),
albumentations.Cutout(max_h_size=int(image_size * 0.375),
max_w_size=int(image_size * 0.375),
num_holes=1,
p=0.7),
albumentations.Normalize()
])
transforms_val = albumentations.Compose([
albumentations.Resize(image_size, image_size),
# albumentations.GaussianBlur(blur_limit=3, p=1.0),
albumentations.Normalize()
])
return transforms_val, transforms_val
alb.RandomRotate90(),
alb.ShiftScaleRotate(),
alb.RandomGamma(),
alb.RandomBrightnessContrast(),
alb.HueSaturationValue(),
alb.RGBShift(),
alb.OneOf([
alb.ElasticTransform(),
alb.GridDistortion(),
alb.OpticalDistortion(distort_limit=0.1)
],
p=0.8),
alb.OneOf([
alb.Blur(),
alb.MotionBlur(),
alb.MedianBlur(),
alb.GaussianBlur()
]),
alb.GaussNoise(p=0.25)
])
for batch_idx, (augm_img_batch, augm_mask_batch) in \
enumerate(AugmentedSequence(root_path=TEST_PATH, img_size=IMG_SIZE, augm_compose=augm, augm_count=3, batch_size=4)):
print(f"{batch_idx}: {len(augm_img_batch)}")
print("---")
for augm_img_idx, augm_img in enumerate(augm_img_batch):
cv2.imwrite(
os.path.join(test_augm_img_path,
f"batch_{batch_idx}-{augm_img_idx}.png"),
import albumentations as A
train_tr = A.Compose([
A.CLAHE(p=0.2),
A.Flip(p=0.3),
A.OneOf([A.GaussianBlur(),
A.MedianBlur(),
A.GaussNoise()], p=0.2),
A.OpticalDistortion(p=0.2),
A.OneOf([
A.IAASharpen(),
A.RandomBrightnessContrast()], p=0.2)
])
val_tr = A.Compose([
])
def get_images(self, imgpath):
# Pick random clone, crypt or fufi
u01 = np.random.uniform()
if u01 < self.cpfr_frac[0]:
img, mask = self.all_svs_opened[imgpath].fetch_clone(
prop_displ=0.45)
elif u01 < np.sum(self.cpfr_frac[0:2]):
img, mask = self.all_svs_opened[imgpath].fetch_partial(
prop_displ=0.45)
elif u01 < np.sum(self.cpfr_frac[0:3]):
img, mask = self.all_svs_opened[imgpath].fetch_fufi(
prop_displ=0.45)
else:
img, mask = self.all_svs_opened[imgpath].fetch_rndmtile()
if self.dilate_masks == True:
n_dil = int(5 / self.um_per_pixel) # if mpp is one or less than 1
# dilate if desired
for i in range(mask.shape[2]):
mask[:, :, i] = cv2.morphologyEx(mask[:, :, i].copy(),
cv2.MORPH_DILATE,
st_3,
iterations=n_dil)
if self.aug == True:
composition = A.Compose([
A.HorizontalFlip(),
A.VerticalFlip(),
A.Rotate(border_mode=cv2.BORDER_CONSTANT),
A.OneOf([
A.ElasticTransform(alpha=1000,
sigma=30,
alpha_affine=30,
border_mode=cv2.BORDER_CONSTANT,
p=1),
A.GridDistortion(border_mode=cv2.BORDER_CONSTANT, p=1),
],
p=0.5),
A.CLAHE(p=0.2),
A.HueSaturationValue(hue_shift_limit=12,
sat_shift_limit=12,
val_shift_limit=12,
p=0.3),
A.RandomBrightnessContrast(p=0.3),
A.Posterize(p=0.1, num_bits=4),
A.OneOf([
A.JpegCompression(p=1),
A.MedianBlur(p=1),
A.Blur(p=1),
A.GlassBlur(p=1, max_delta=2, sigma=0.4),
A.IAASharpen(p=1)
],
p=0.3)
],
p=1)
transformed = composition(image=img, mask=mask)
img, mask = transformed['image'], transformed['mask']
mask_list = [mask[:, :, ii] for ii in range(mask.shape[2])]
if self.stride_bool:
mask_list = [cv2.pyrDown(mask_ii.copy()) for mask_ii in mask_list]
mask_list = [
cv2.threshold(mask_ii, 120, 255, cv2.THRESH_BINARY)[1]
for mask_ii in mask_list
]
## convert to floating point space, normalize and mask non-used clones
img = img.astype(np.float32) / 255
mask_list = [mask_ii.astype(np.float32) / 255 for mask_ii in mask_list]
if self.normalize:
img = (img - self.norm_mean) / self.norm_std
return img, np.stack(mask_list, axis=2)
from dataset import MyDataset, AlbuWrapper, tensor2img
import albumentations as alb
import albumentations.augmentations.transforms as aat
import albumentations
import pandas as pd
import matplotlib.pyplot as plt
from tqdm.auto import tqdm
import pathlib
from dataset import TRAFFIC_LABELS_TO_NUM
alb_transforms = [
alb.IAAAdditiveGaussianNoise(p=1),
alb.GaussNoise(p=1),
alb.MotionBlur(p=1),
alb.MedianBlur(blur_limit=3, p=1),
alb.Blur(blur_limit=3, p=1),
alb.OpticalDistortion(p=1),
alb.GridDistortion(p=1),
alb.IAAPiecewiseAffine(p=1),
aat.CLAHE(clip_limit=2, p=1),
alb.IAASharpen(p=1),
alb.IAAEmboss(p=1),
aat.HueSaturationValue(p=0.3),
aat.HorizontalFlip(p=1),
aat.RGBShift(),
aat.RandomBrightnessContrast(),
aat.RandomGamma(p=1),
aat.Cutout(2, 10, 10, p=1),
aat.Equalize(mode='cv', p=1),
aat.FancyPCA(p=1),
aat.RandomFog(p=1),
def get_augmentation_fcn2(mode, p=0.75):
if not mode or mode.lower() == 'none':
return None
augmentation_dict = {
'no_interpolation_necessary':
al.OneOf([al.RandomRotate90(p=1.),
al.Flip(p=1.)]),
'interpolation_necessary':
al.OneOf(
[al.Rotate(p=1.),
al.RandomScale(p=1.),
al.ShiftScaleRotate(p=1.)]),
'affine':
al.Compose([al.ShiftScaleRotate(p=1.),
al.HorizontalFlip(p=0.5)]),
'rot':
al.Rotate(p=1.),
'rot90':
al.RandomRotate90(p=1.),
'flip':
al.Flip(p=1.),
'hflip':
al.HorizontalFlip(p=1.),
'vflip':
al.VerticalFlip(p=1.),
'scale':
al.RandomScale(p=1.),
'ssr':
al.ShiftScaleRotate(p=1.),
'strong':
al.Compose(
[
# al.RandomRotate90(),
# al.Flip(),
# al.Transpose(),
al.OneOf([
al.IAAAdditiveGaussianNoise(),
al.GaussNoise(),
],
p=0.2),
al.OneOf([
al.MotionBlur(p=0.2),
al.MedianBlur(blur_limit=3, p=0.1),
al.Blur(blur_limit=3, p=0.1),
],
p=0.2),
al.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=10,
p=0.2),
# al.OneOf([
# al.OpticalDistortion(p=0.3),
# al.GridDistortion(p=0.1),
# al.IAAPiecewiseAffine(p=0.3),
# ], p=0.2),
al.OneOf(
[
# al.CLAHE(clip_limit=2),
al.IAASharpen(),
al.IAAEmboss(),
al.RandomBrightnessContrast(),
],
p=0.3),
al.HueSaturationValue(p=0.3),
],
p=p),
'truss_points':
al.Compose([
al.OneOf([
al.IAAAdditiveGaussianNoise(),
al.GaussNoise(),
], p=0.2),
al.OneOf([
al.MotionBlur(p=0.2),
al.MedianBlur(blur_limit=3, p=0.1),
al.Blur(blur_limit=3, p=0.1),
],
p=0.2),
al.ShiftScaleRotate(
shift_limit=0.0625, scale_limit=0.2, rotate_limit=10, p=0.2),
al.OneOf([
al.IAASharpen(),
al.IAAEmboss(),
al.RandomBrightnessContrast(),
],
p=0.3),
al.HueSaturationValue(p=0.3),
],
p=p)
}
def aug_fcn(**kwargs):
# params: image, mask, masks, bboxes, keypoints
# note keypoints of form (x,y,a,s) ... I think really tangent vectors
return augmentation_dict[mode](**kwargs)
return aug_fcn
def get_augmentation_fcn(mode, p=0.75):
if not mode or mode.lower() == 'none':
return None
augmentation_dict = {
'no_interpolation_necessary':
al.OneOf([al.RandomRotate90(p=1.),
al.Flip(p=1.)]),
'interpolation_necessary':
al.OneOf(
[al.Rotate(p=1.),
al.RandomScale(p=1.),
al.ShiftScaleRotate(p=1.)]),
'affine':
al.Compose([al.ShiftScaleRotate(p=1.),
al.HorizontalFlip(p=0.5)]),
'rot':
al.Rotate(p=1.),
'rot90':
al.RandomRotate90(p=1.),
'flip':
al.Flip(p=1.),
'hflip':
al.HorizontalFlip(p=1.),
'vflip':
al.VerticalFlip(p=1.),
'scale':
al.RandomScale(p=1.),
'ssr':
al.ShiftScaleRotate(p=1.),
'strong':
al.Compose(
[
# al.RandomRotate90(),
# al.Flip(),
# al.Transpose(),
al.OneOf([
al.IAAAdditiveGaussianNoise(),
al.GaussNoise(),
],
p=0.2),
al.OneOf([
al.MotionBlur(p=0.2),
al.MedianBlur(blur_limit=3, p=0.1),
al.Blur(blur_limit=3, p=0.1),
],
p=0.2),
al.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=10,
p=0.2),
# al.OneOf([
# al.OpticalDistortion(p=0.3),
# al.GridDistortion(p=0.1),
# al.IAAPiecewiseAffine(p=0.3),
# ], p=0.2),
al.OneOf(
[
# al.CLAHE(clip_limit=2),
al.IAASharpen(),
al.IAAEmboss(),
al.RandomBrightnessContrast(),
],
p=0.3),
al.HueSaturationValue(p=0.3),
],
p=p)
}
def aug_fcn(x):
try:
return augmentation_dict[mode](image=x)['image']
except Exception as e:
print("Exception caught in augmentation stage:", e)
return x
return aug_fcn
import json
RESIZE_SIZE = 1024
train_transform = albumentations.Compose([
albumentations.Resize(RESIZE_SIZE, RESIZE_SIZE),
albumentations.OneOf([
albumentations.RandomGamma(gamma_limit=(60, 120), p=0.9),
albumentations.RandomBrightnessContrast(brightness_limit=0.2,
contrast_limit=0.2,
p=0.9),
albumentations.CLAHE(clip_limit=4.0, tile_grid_size=(4, 4), p=0.9),
]),
albumentations.OneOf([
albumentations.Blur(blur_limit=4, p=1),
albumentations.MotionBlur(blur_limit=4, p=1),
albumentations.MedianBlur(blur_limit=4, p=1)
],
p=0.5),
albumentations.HorizontalFlip(p=0.5),
albumentations.ShiftScaleRotate(shift_limit=0.2,
scale_limit=0.2,
rotate_limit=20,
interpolation=cv2.INTER_LINEAR,
border_mode=cv2.BORDER_CONSTANT,
p=1),
albumentations.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
max_pixel_value=255.0,
p=1.0)
])
def get_augmentation(_C, is_train):
""""""
if is_train:
augmentation = [
# random flip
A.HorizontalFlip(p=_C.TRANSFORM.TRAIN_HORIZONTAL_FLIP_PROB),
A.VerticalFlip(p=_C.TRANSFORM.TRAIN_VERTICAL_FLIP_PROB),
# random rotate
A.ShiftScaleRotate(
scale_limit=0.0,
rotate_limit=_C.TRANSFORM.TRAIN_RANDOM_ROTATE_DEG,
shift_limit=0.0,
p=_C.TRANSFORM.TRAIN_RANDOM_ROTATE_PROB,
border_mode=0,
),
# random crop
A.RandomCrop(
width=_C.TRANSFORM.TRAIN_RANDOM_CROP_SIZE[0],
height=_C.TRANSFORM.TRAIN_RANDOM_CROP_SIZE[1],
p=_C.TRANSFORM.TRAIN_RANDOM_CROP_PROB,
),
# speckle noise
A.Lambda(image=functools.partial(
_random_speckle_noise,
speckle_std=_C.TRANSFORM.TRAIN_SPECKLE_NOISE_STD,
p=_C.TRANSFORM.TRAIN_SPECKLE_NOISE_PROB,
)),
# blur
A.OneOf(
[
A.MotionBlur(p=_C.TRANSFORM.TRAIN_BLUR_MOTION_PROB),
A.MedianBlur(
blur_limit=_C.TRANSFORM.TRAIN_BLUR_MEDIAN_LIMIT,
p=_C.TRANSFORM.TRAIN_BLUR_MEDIAN_PROB,
),
A.Blur(
blur_limit=_C.TRANSFORM.TRAIN_BLUR_LIMIT,
p=_C.TRANSFORM.TRAIN_BLUR_PROB,
),
],
p=_C.TRANSFORM.TRAIN_BLUR_ONEOF,
),
# random brightness
A.Lambda(
image=functools.partial(
_random_brightness,
brightness_std=_C.TRANSFORM.TRAIN_RANDOM_BRIGHT_STD_STD,
p=_C.TRANSFORM.TRAIN_RANDOM_BRIGHT_STD_PROB,
),
# transforms.ToTensor(),
# transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
# ])
),
]
else:
augmentation = [
A.PadIfNeeded(
min_width=_C.TRANSFORM.TEST_SIZE[0],
min_height=_C.TRANSFORM.TEST_SIZE[1],
always_apply=True,
border_mode=0,
)
]
return A.Compose(augmentation)
def train(root_path: str = "./data/",
train_dir: str = "training",
val_dir: str = "validation",
img_dir: str = "image",
mask_dir: str = "mask",
backbone: str = "efficientnetb4",
img_size: int = 192,
augm_count: int = 19,
batch_size: int = 5,
LR: float = 0.0001,
epochs: int = 10,
model_file: str = "./last_best_model.h5",
plot_history: bool = False):
"""
Train unet on provided dataset, saving the best weights.
Parameters
----------
root_path : string
The path of the directory which contains the training directory and the validation directory.
train_dir : string
The path of the training directory relative to 'root_path'.
val_dir : string
The path of the validation directory relative to 'root_path'.
img_dir : string
The path of the image directory relative to both 'img_dir' and 'mask_dir'.
mask_dir : string
The path of the mask directory relative to both 'img_dir' and 'mask_dir'.
backbone : string
The backbone used for unet.
img_size : int
Size of the crop for each image fed to the model.
augm_count : int
Number of augmentations for each image in the training dataset.
batch_size : int
Size of the batch of images fed to the model.
LR : float
Learning rate of the model.
epochs : int
Number of training epochs.
model_file : string
Path of the file in which to save the best weights of the model.
plot_history : bool
If True, save the training history to file.
The file name has the pattern '{number_of_images}i_{img_size}_{augm_count}a_{epochs}e.png'.
Returns
-------
-
"""
# define augmentation
augm = alb.Compose([
alb.Flip(),
alb.Transpose(),
alb.RandomRotate90(),
alb.ShiftScaleRotate(),
alb.RandomGamma(),
alb.RandomBrightnessContrast(),
alb.HueSaturationValue(),
alb.RGBShift(),
alb.OneOf([
alb.ElasticTransform(),
alb.GridDistortion(),
alb.OpticalDistortion(distort_limit=0.1)
],
p=0.8),
alb.OneOf([
alb.Blur(),
alb.MotionBlur(),
alb.MedianBlur(),
alb.GaussianBlur()
]),
alb.GaussNoise(p=0.25)
])
# create model
model = sm.Unet(backbone, encoder_weights="imagenet"
) # activation sigmoid by default, ok for 1 class
# define optimizer
optim = keras.optimizers.Adam(learning_rate=LR)
# define metrics
metrics = [
sm.metrics.IOUScore(threshold=0.5),
sm.metrics.FScore(threshold=0.5)
]
# compile keras model with defined optimizer, loss and metrics
model.compile(optim, sm.losses.binary_focal_dice_loss, metrics)
# define train and validation generators
train_generator = AugmentedSequence(root_path=os.path.join(
root_path, train_dir),
img_dir=img_dir,
mask_dir=mask_dir,
img_size=img_size,
batch_size=batch_size,
augm_count=augm_count,
augm_compose=augm)
valid_generator = AugmentedSequence(root_path=os.path.join(
root_path, val_dir),
img_dir=img_dir,
mask_dir=mask_dir,
img_size=img_size,
batch_size=batch_size,
augm_count=0)
# check shapes for errors
assert train_generator[0][0].shape == (batch_size, img_size, img_size, 3)
assert train_generator[0][1].shape == (batch_size, img_size, img_size,
1), train_generator[0][1].shape
# define callbacks for learning rate reduction and best checkpoint saving
callbacks = [
keras.callbacks.ModelCheckpoint(model_file,
save_best_only=True,
save_weights_only=True,
mode="min"),
keras.callbacks.ReduceLROnPlateau()
]
# train model
history = model.fit_generator(
train_generator,
epochs=epochs,
callbacks=callbacks,
validation_data=valid_generator,
#use_multiprocessing=True # crashes, BSODs, the apocalypse
)
if plot_history:
plot = plot_training_history(history)
plot.savefig(
f"{len(train_generator.img_files)}i_{img_size}_{augm_count}a_{epochs}e.png"
)
from torch.utils.data import Dataset, DataLoader
from configure import SPLIT_FOLDER, TRAIN_DATA_FOLDER
import albumentations as albu
train_aug = albu.Compose([
albu.OneOf([
albu.RandomBrightnessContrast(
brightness_limit=0.2, contrast_limit=0.2, p=1),
albu.CLAHE(clip_limit=4.0, tile_grid_size=(4, 4), p=1),
],
p=0.5),
albu.OneOf([
albu.Blur(blur_limit=4, p=1),
albu.MotionBlur(blur_limit=4, p=1),
albu.MedianBlur(blur_limit=4, p=1)
],
p=0.5),
albu.OneOf([albu.GridDistortion(p=1),
albu.OpticalDistortion(p=1)], p=0.5),
albu.HorizontalFlip(p=0.5),
albu.VerticalFlip(p=0.5),
albu.RandomSizedCrop(min_max_height=(160, 320),
height=320,
width=640,
w2h_ratio=2,
p=0.5),
albu.ShiftScaleRotate(shift_limit=0.2,
scale_limit=0.2,
rotate_limit=45,
interpolation=cv2.INTER_LINEAR,
use_validation_loss = False
if student_learn_from_sfm:
training_transforms = albu.Compose([
# Color augmentation
albu.OneOf([
albu.Compose([
albu.RandomBrightnessContrast(brightness_limit=0.3, contrast_limit=0.3, p=0.5),
albu.RandomGamma(gamma_limit=(50, 180), p=0.5),
albu.HueSaturationValue(hue_shift_limit=30, sat_shift_limit=0, val_shift_limit=0, p=0.5)]),
albu.HueSaturationValue(hue_shift_limit=30, sat_shift_limit=30, val_shift_limit=30, p=0.5)
]),
# Image quality augmentation
albu.OneOf([
albu.Blur(p=0.5),
albu.MedianBlur(p=0.5),
albu.MotionBlur(p=0.5),
albu.JpegCompression(quality_lower=20, quality_upper=100, p=0.5)
]),
# Noise augmentation
albu.OneOf([
albu.GaussNoise(var_limit=(10, 50), p=0.5),
albu.IAAAdditiveGaussianNoise(loc=0, scale=(0.01 * 255, 0.05 * 255))
]),
], p=1.)
else:
training_transforms = albu.Compose(
[albu.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.5)], p=1.)
validation_transforms = albu.Compose([
albu.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), max_pixel_value=255.0, p=1.)], p=1.)
def transform(image, mask=None, opt=None, paired=False):
assert opt is not None, 'opt is not optional (it cannot be None)'
#[ ab.Resize(height=opt.load_height, width=opt.load_width)] if opt.load_width!=-1 else [])
resize = ab.Resize(height=opt.load_height, width=opt.load_width)
T_stack_pre = ab.Compose([
#ab.ChannelDropout((1, 4 if paired else 2 )),
#ab.ChannelDropout(),
ab.OpticalDistortion(),
ab.RandomCrop(height=opt.crop_size, width=opt.crop_size),
# ab.CoarseDropout()
])
T_pair = ab.Compose([
ab.ShiftScaleRotate(shift_limit=0.005,
scale_limit=0.01,
rotate_limit=5,
interpolation=1),
ab.OneOf([
ab.MotionBlur(),
ab.MedianBlur(5),
ab.GaussianBlur(),
]),
#ab.ToFloat(),
#ab.HueSaturationValue(),
#ab.RandomBrightness(),
#ab.RandomContrast(),
ab.Normalize((.5, ) * 3, (.5, ) * 3)
])
#nc = 6 if paired else 3
T_stack_post = ab.Compose([
#ab.Normalize((0.5,)*nc, (0.5,)*nc),
ab.pytorch.ToTensor(),
])
if paired:
#i = 0
#print('\n\n')
#print(str(i), image.mean(), image.std()); i += 1
aug = resize(image=image, mask=mask)
image, mask = aug['image'], aug['mask']
#print(str(i), image.mean(), image.std()); i += 1
cat = T_stack_pre(
image=np.concatenate((image, mask), axis=-1))['image']
image, mask = cat[..., :3], cat[..., 3:]
#print(str(i), image.mean(), image.std()); i += 1
aug = T_pair(image=image, mask=mask)
image, mask = aug['image'], aug['mask']
#print(str(i), image.mean(), image.std()); i += 1
cat = T_stack_post(
image=np.concatenate((image, mask), axis=-1))['image']
image, mask = cat[:3], cat[3:]
#print(str(i), image.mean(), image.std()); i += 1
#print('\n\n')
return image, mask
else:
#print('IMG', type(image))
#print('1',image.shape, end=' -> ',flush=True)
image = resize(image=image)['image'],
#print('\n\n\n\n\n\n\n')
#print(image)
#print('\n\n\n\n\n\n\n')
#print('2',image.shape, end=' -> ', flush=True)
#print('IMG TYPE',type(image), len(image))
if type(image) is tuple:
image = image[0]
image = T_stack_pre(image=image)['image']
#print(image.shape, end=' -> ',flush=True)
image = T_pair(image=image)['image']
#print(image.shape, end=' -> ')
image = T_stack_post(image=image)['image']
#print(image.shape)
return image
def _add_augmentation(self, aug_name, params):
if aug_name == 'hflip':
return album.HorizontalFlip(p=params['p'])
elif aug_name == 'vflip':
return album.VerticalFlip(p=params['p'])
elif aug_name == 'blur':
return album.Blur(p=params['p'])
elif aug_name == 'mblur':
return album.MedianBlur(p=params['p'])
elif aug_name == 'rbright':
return album.RandomBrightness(p=params['p'])
elif aug_name == 'rotate':
return album.Rotate(p=params['p'], limit=params['limit'])
elif aug_name == 'rotate90':
return album.RandomRotate90(p=params['p'])
elif aug_name == 'ssrotate':
return album.ShiftScaleRotate(p=params['p'],
shift_limit=params['shift_limit'],
scale_limit=params['scale_limit'],
rotate_limit=params['rotate_limit'])
elif aug_name == 'cutout':
return album.Cutout(p=params['p'],
num_holes=params['num_holes'],
max_h_size=params['max_h_size'],
max_w_size=params['max_w_size'])
elif aug_name == 'cdropout':
return album.CoarseDropout(p=params['p'],
max_holes=params['max_holes'],
max_height=params['max_height'],
max_width=params['max_width'],
min_holes=params['min_holes'],
min_height=params['min_height'],
min_width=params['min_width'])
elif aug_name == 'rrcrop':
return album.RandomResizedCrop(p=params['p'],
height=params['height'],
width=params['width'],
scale=tuple(params['scale']),
ratio=tuple(params['ratio']))
elif aug_name == 'elastic':
return album.ElasticTransform(
p=params['p'],
alpha=params['alpha'],
sigma=params['sigma'],
alpha_affine=params['alpha_affine'],
interpolation=params['interpolation'],
border_mode=params['border_mode'])
elif aug_name == 'rmorph':
return RandomMorph(p=params['p'])
elif aug_name == 'augmix':
return RandomAugMix(p=params['p'])
elif aug_name == 'gridmask':
return GridMask(p=params['p'],
num_grid=tuple(params['num_grid']),
rotate=params['rotate_limit'],
mode=params['mode'])
else:
raise (NotImplementedError)
A.Flip(p=1),
A.RandomRotate90(p=1),
A.Rotate(p=1, limit=45, interpolation=3),
A.RandomResizedCrop(input_size[0],
input_size[1],
scale=(0.7, 1.0),
ratio=(1.0, 1.0),
interpolation=3,
p=1),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
], p=0.3),
A.OneOf([
A.MotionBlur(p=0.25),
A.MedianBlur(blur_limit=3, p=0.25),
A.Blur(blur_limit=3, p=0.25),
A.GaussianBlur(p=0.25)
],
p=0.1),
A.OneOf([
A.OpticalDistortion(interpolation=3, p=0.1),
A.GridDistortion(interpolation=3, p=0.1),
A.IAAPiecewiseAffine(p=0.5),
],
p=0.05),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
A.RandomBrightnessContrast(),
def get_augs(new_augs, size, size_after_reshape):
ALBUMENTATIONS_VAL = albu.Compose([
albu.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
max_pixel_value=255.0,
p=1.0),
albu.Resize(size_after_reshape, size_after_reshape)
])
if not new_augs:
ALBUMENTATIONS_TRAIN = albu.Compose([
albu.HorizontalFlip(),
albu.VerticalFlip(),
albu.RandomRotate90(),
albu.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=15,
p=0.9,
border_mode=cv2.BORDER_REFLECT),
albu.OneOf(
[
albu.OpticalDistortion(p=0.3),
albu.GridDistortion(p=.1),
#albu.IAAPiecewiseAffine(p=0.3),
],
p=0.3),
albu.OneOf([
albu.HueSaturationValue(10, 15, 10),
albu.CLAHE(clip_limit=2),
albu.RandomBrightnessContrast(),
],
p=0.3),
albu.Resize(size_after_reshape, size_after_reshape),
])
else:
ALBUMENTATIONS_TRAIN = albu.Compose([
albu.HorizontalFlip(),
albu.VerticalFlip(),
albu.RandomRotate90(),
albu.Cutout(num_holes=10,
max_h_size=int(.1 * size),
max_w_size=int(.1 * size),
p=.25),
albu.OneOf([
albu.GaussNoise(0.002, p=.5),
albu.IAAAffine(p=.5),
],
p=.25),
albu.OneOf([
albu.Blur(blur_limit=3, p=1),
albu.MedianBlur(blur_limit=3, p=1)
],
p=.25),
albu.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=15,
p=0.9,
border_mode=cv2.BORDER_REFLECT),
albu.HueSaturationValue(20, 30, 20),
albu.RandomBrightnessContrast(),
albu.OneOf(
[
albu.OpticalDistortion(1.2, 0.5, p=.3),
albu.GridDistortion(5, 0.2, p=.3),
# albu.IAAPiecewiseAffine(p=0.3),
],
p=0.3),
albu.OneOf([
albu.CLAHE(clip_limit=2),
albu.RandomGamma(),
], p=0.5),
albu.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
max_pixel_value=255.0,
p=1.0),
albu.Resize(size_after_reshape, size_after_reshape),
])
return ALBUMENTATIONS_TRAIN, ALBUMENTATIONS_VAL
def get_transforms(stage: str = None, mode: str = None):
if mode == 'train':
return albumentations.Compose([
# blur
albumentations.OneOf([
albumentations.Blur((1, 4), p=1.0),
albumentations.GaussianBlur(3, p=1.0),
albumentations.MedianBlur(blur_limit=5, p=1.0),
],
p=3 / 4),
# transformations
albumentations.ShiftScaleRotate(
scale_limit=0.2,
rotate_limit=25,
border_mode=cv2.BORDER_CONSTANT,
value=0,
p=1.0),
# cut and drop
albumentations.OneOf([
albumentations.Cutout(num_holes=10,
max_h_size=SIZE // 6,
max_w_size=SIZE // 6,
p=1.0),
albumentations.CoarseDropout(
max_holes=8, max_height=10, max_width=10, p=1.0),
],
p=2 / 3),
# distortion
albumentations.OneOf([
albumentations.OpticalDistortion(0.6, p=1.0),
albumentations.GridDistortion(
8,
0.06,
border_mode=cv2.BORDER_CONSTANT,
value=0,
p=1.0),
albumentations.ElasticTransform(
sigma=10,
alpha=1,
alpha_affine=10,
border_mode=cv2.BORDER_CONSTANT,
value=0,
p=1.0),
],
p=3 / 4),
# add noise
albumentations.OneOf([
albumentations.GaussNoise((0, 250), p=1.0),
albumentations.MultiplicativeNoise(p=1.0),
],
p=2 / 3),
# common
albumentations.Normalize(TRAIN_MEAN, TRAIN_STD),
ToTensorV2(),
])
elif mode == 'valid':
return albumentations.Compose([
albumentations.Normalize(TRAIN_MEAN, TRAIN_STD),
ToTensorV2(),
])
else:
raise ValueError('mode is %s' % mode)
"strong":
albu.Compose([
albu.HorizontalFlip(),
albu.ShiftScaleRotate(shift_limit=0.0,
scale_limit=0.2,
rotate_limit=30,
p=0.4),
albu.ElasticTransform(),
albu.GaussNoise(),
albu.OneOf(
[
albu.CLAHE(clip_limit=2),
albu.IAASharpen(),
albu.RandomBrightnessContrast(),
albu.RandomGamma(),
albu.MedianBlur(),
],
p=0.5,
),
albu.OneOf(
[albu.RGBShift(), albu.HueSaturationValue()], p=0.5),
]),
"weak":
albu.Compose([albu.HorizontalFlip()]),
"horizontal_flip":
albu.HorizontalFlip(p=0.5),
"none":
albu.Compose([]),
}
A.OneOf([
A.RandomBrightness(limit=(0.8, 1.2), always_apply=True),
A.RGBShift(r_shift_limit=(-10, 10),
g_shift_limit=(-10, 10),
b_shift_limit=(-10, 10),
always_apply=True),
A.HueSaturationValue(hue_shift_limit=(-10, 10),
sat_shift_limit=(-10, 10),
val_shift_limit=(-10, 10),
always_apply=True),
A.RandomContrast(limit=(0.9, 1.1), always_apply=True)
],
p=0.3),
A.OneOf([
A.Blur(blur_limit=3, always_apply=True),
A.MedianBlur(blur_limit=3, always_apply=True),
A.GaussNoise(var_limit=(10, 50), always_apply=True)
])
])
train_ds = D.PKUSingleObjectTrainDataset(
json_annotations=os.path.join(cfg.CV_DIR, 'fold-1',
'train_objects_both_train.json'),
images_dir=cfg.TRAIN_IMAGES,
color_augment_fn=lambda dct: D.augment_fn_albu_color(
dct, albu_train_transforms),
geom_augment_fn=lambda dct: D.augment_fn_flip_rotate(D.augment_fn_bbox(
dct, s=0.02),
p_flip=0.5,
p_rotate=0.5,
rotatelim=(-5, 5)),
FINALIZE_T = albumentations.Compose([
albumentations.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
ToTensor(),
])
BASIC_IMAGE_T = albumentations.Compose([
albumentations.Resize(128, 128),
FINALIZE_T
])
AUG_IMAGE_T = albumentations.Compose([
albumentations.GaussNoise(),
albumentations.OneOf([
albumentations.MotionBlur(),
albumentations.MedianBlur(blur_limit=3),
albumentations.Blur(blur_limit=3),
albumentations.JpegCompression(quality_lower=90),
]),
albumentations.OneOf([
albumentations.CLAHE(clip_limit=2),
albumentations.IAASharpen(),
albumentations.RandomBrightnessContrast(),
]),
albumentations.HueSaturationValue(p=0.3),
])
TRAIN_IMAGE_T = albumentations.Compose([
albumentations.Resize(128, 128),
AUG_IMAGE_T,
seed_torch(seed=CFG.seed)
base_transform = A.Compose([
A.OneOf(
[
# 加了hsv
A.HueSaturationValue(10, 15, 10),
A.RandomBrightness(limit=.2, p=1),
A.RandomContrast(limit=.2, p=1),
A.RandomGamma(p=1)
],
p=.5),
A.OneOf([A.Blur(blur_limit=3, p=1),
A.MedianBlur(blur_limit=3, p=1)],
p=.25),
A.OneOf([
A.GaussNoise(0.002, p=.5),
A.IAAAffine(p=.5),
], p=.25),
# A.OneOf([
# A.ElasticTransform(alpha=120, sigma=120 * .05, alpha_affine=120 * .03, p=.5),
# A.GridDistortion(p=.5),
# A.OpticalDistortion(distort_limit=2, shift_limit=.5, p=1)
# ], p=.25),
A.RandomRotate90(p=.5),
A.HorizontalFlip(p=.5),
A.VerticalFlip(p=.5),
A.Cutout(num_holes=10,
max_h_size=int(.1 * CFG.data),
p=0.7),
albumentations.RandomBrightnessContrast(brightness_limit=(-0.2, 0.2),
contrast_limit=(-0.2, 0.2),
p=0.7),
albumentations.CLAHE(clip_limit=(1, 4), p=0.5),
albumentations.OneOf([
albumentations.OpticalDistortion(distort_limit=1.0),
albumentations.GridDistortion(num_steps=5, distort_limit=1.),
albumentations.ElasticTransform(alpha=3),
],
p=0.2),
albumentations.OneOf([
albumentations.GaussNoise(var_limit=[10, 50]),
albumentations.GaussianBlur(),
albumentations.MotionBlur(),
albumentations.MedianBlur(),
],
p=0.2),
albumentations.Resize(CFG.img_size, CFG.img_size),
albumentations.OneOf([
albumentations.ImageCompression(),
albumentations.Downscale(scale_min=0.1, scale_max=0.15),
],
p=0.2),
albumentations.IAAPiecewiseAffine(p=0.2),
albumentations.IAASharpen(p=0.2),
albumentations.CoarseDropout(max_height=int(CFG.img_size * 0.1),
max_width=int(CFG.img_size * 0.1),
max_holes=5,
p=0.5),
])
def __call__(self, example):
if self.train:
x, y = example
else:
x = example
# --- Augmentation ---
if self.affine:
x = affine_image(x)
# --- Train/Test common preprocessing ---
if self.crop:
x = crop_char_image(x, threshold=self.threshold)
if self.size is not None:
x = resize(x, size=self.size)
if self.sigma > 0.:
x = add_gaussian_noise(x, sigma=self.sigma)
# albumentations...
x = x.astype(np.float32)
assert x.ndim == 2
# 1. blur
if _evaluate_ratio(self.blur_ratio):
r = np.random.uniform()
if r < 0.25:
x = apply_aug(A.Blur(p=1.0), x)
elif r < 0.5:
x = apply_aug(A.MedianBlur(blur_limit=5, p=1.0), x)
elif r < 0.75:
x = apply_aug(A.GaussianBlur(p=1.0), x)
else:
x = apply_aug(A.MotionBlur(p=1.0), x)
if _evaluate_ratio(self.noise_ratio):
r = np.random.uniform()
if r < 0.50:
x = apply_aug(A.GaussNoise(var_limit=5. / 255., p=1.0), x)
else:
x = apply_aug(A.MultiplicativeNoise(p=1.0), x)
if _evaluate_ratio(self.cutout_ratio):
# A.Cutout(num_holes=2, max_h_size=2, max_w_size=2, p=1.0) # Deprecated...
x = apply_aug(
A.CoarseDropout(max_holes=8, max_height=8, max_width=8, p=1.0),
x)
if _evaluate_ratio(self.grid_distortion_ratio):
x = apply_aug(A.GridDistortion(p=1.0), x)
if _evaluate_ratio(self.elastic_distortion_ratio):
x = apply_aug(
A.ElasticTransform(sigma=50, alpha=1, alpha_affine=10, p=1.0),
x)
if _evaluate_ratio(self.random_brightness_ratio):
# A.RandomBrightness(p=1.0) # Deprecated...
# A.RandomContrast(p=1.0) # Deprecated...
x = apply_aug(A.RandomBrightnessContrast(p=1.0), x)
if _evaluate_ratio(self.piece_affine_ratio):
x = apply_aug(A.IAAPiecewiseAffine(p=1.0), x)
if _evaluate_ratio(self.ssr_ratio):
x = apply_aug(
A.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.1,
rotate_limit=30,
p=1.0), x)
if self.normalize:
x = (x.astype(np.float32) - 0.0692) / 0.2051
if x.ndim == 2:
x = x[None, :, :]
x = x.astype(np.float32)
if self.train:
y = y.astype(np.int64)
return x, y
else:
return x
def __init__(self, folds, img_height, img_width, mean, std):
df = pd.read_csv('../input/dataset/train_folds.csv')
df = df[['Image', 'label', 'kfold']]
df = df[df.kfold.isin(folds)].reset_index(drop=True)
self.image_ids = df.Image.values
self.labels = df.label.values
if len(folds) == 1:
self.aug = albumentations.Compose([
albumentations.Resize(img_height, img_width),
albumentations.Normalize(mean, std, always_apply=True)
])
else:
self.aug = albumentations.Compose([
albumentations.Resize(img_height, img_width),
albumentations.OneOf([
albumentations.ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.1,
rotate_limit=45),
albumentations.Rotate(limit=5),
albumentations.RandomGamma(),
albumentations.RandomShadow(),
albumentations.RandomGridShuffle(),
albumentations.ElasticTransform(),
albumentations.RGBShift(),
]),
albumentations.OneOf([
albumentations.OneOf([
albumentations.Blur(),
albumentations.MedianBlur(),
albumentations.MotionBlur(),
albumentations.GaussianBlur(),
]),
albumentations.OneOf([
albumentations.GaussNoise(),
albumentations.IAAAdditiveGaussianNoise(),
albumentations.ISONoise()
]),
]),
albumentations.OneOf([
albumentations.RandomBrightness(),
albumentations.RandomContrast(),
albumentations.RandomBrightnessContrast(),
]),
albumentations.OneOf([
albumentations.OneOf([
albumentations.Cutout(),
albumentations.CoarseDropout(),
albumentations.GridDistortion(),
albumentations.GridDropout(),
albumentations.OpticalDistortion()
]),
albumentations.OneOf([
albumentations.HorizontalFlip(),
albumentations.VerticalFlip(),
albumentations.RandomRotate90(),
albumentations.Transpose()
]),
]),
# albumentations.OneOf([
# albumentations.RandomSnow(),
# albumentations.RandomRain(),
# albumentations.RandomFog(),
# ]),
albumentations.Normalize(mean, std, always_apply=True)
])