def __init__(self,
root,
input_transform=None,
target_transform=None,
_train=False):
self.images_root = os.path.join(root, 'Images')
self.labels_root = os.path.join(root, 'Labels')
self._train = _train
self.filenames = [
image_basename(f) for f in os.listdir(self.labels_root)
if is_image(f)
]
random.shuffle(self.filenames)
self.input_transform = input_transform
self.target_transform = target_transform
self.A_transform = A.Compose([
A.Rotate(limit=40, p=0.9),
A.HorizontalFlip(p=0.5),
A.RGBShift(r_shift_limit=25,
g_shift_limit=25,
b_shift_limit=25,
p=0.9),
A.OneOf([A.Blur(blur_limit=2, p=0.5),
A.ColorJitter(p=0.5)], p=1.0)
])
def common_aug(mode, params):
'''
:param mode: 'train', 'test', 'inference'
:param params:
'''
#aug_params = params.get('augm_params', dict())
augs_list = []
assert mode in {'train', 'debug', 'inference'}
if mode == 'train':
augs_list.append(albumentations.PadIfNeeded(min_height=params.data.net_hw[0], min_width=params.data.net_hw[1],
border_mode=cv2.BORDER_REPLICATE,
always_apply=True))
augs_list.append(albumentations.RandomCrop(height=params.data.net_hw[0], width=params.data.net_hw[1], always_apply=True))
if params.augmentation.rotate_limit:
augs_list.append(T.Rotate(limit=params.augmentation.rotate_limit, border_mode=cv2.BORDER_CONSTANT, always_apply=True))
# augs_list.append(T.OpticalDistortion(border_mode=cv2.BORDER_CONSTANT)) - can't handle boundboxes
elif mode == 'debug':
augs_list.append(albumentations.CenterCrop(height=params.data.net_hw[0], width=params.data.net_hw[1], always_apply=True))
if mode != 'inference':
if params.augmentation.get('blur_limit', 4):
augs_list.append(T.Blur(blur_limit=params.augmentation.get('blur_limit', 4)))
if params.augmentation.get('RandomBrightnessContrast', True):
augs_list.append(T.RandomBrightnessContrast())
#augs_list.append(T.MotionBlur())
if params.augmentation.get('JpegCompression', True):
augs_list.append(T.JpegCompression(quality_lower=30, quality_upper=100))
#augs_list.append(T.VerticalFlip())
if params.augmentation.get('HorizontalFlip', True):
augs_list.append(T.HorizontalFlip())
return albumentations.ReplayCompose(augs_list, p=1., bbox_params = {'format':'albumentations', 'min_visibility':0.5})
def augmentation(mode, target_size, prob=0.5, aug_m=2):
'''
description: augmentation
mode: 'train' 'test'
target_size: int or list, the shape of image ,
aug_m: Strength of transform
'''
high_p = prob
low_p = high_p / 2.0
M = aug_m
first_size = [int(x / 0.7) for x in target_size]
if mode == 'train':
return composition.Compose([
transforms.Resize(first_size[0], first_size[1], interpolation=3),
transforms.Flip(p=0.5),
composition.OneOf([
RandomCenterCut(scale=0.1 * M),
transforms.ShiftScaleRotate(shift_limit=0.05 * M,
scale_limit=0.1 * M,
rotate_limit=180,
border_mode=cv2.BORDER_CONSTANT,
value=0),
albumentations.imgaug.transforms.IAAAffine(
shear=(-10 * M, 10 * M), mode='constant')
],
p=high_p),
transforms.RandomBrightnessContrast(
brightness_limit=0.1 * M, contrast_limit=0.03 * M, p=high_p),
transforms.HueSaturationValue(hue_shift_limit=5 * M,
sat_shift_limit=15 * M,
val_shift_limit=10 * M,
p=high_p),
transforms.OpticalDistortion(distort_limit=0.03 * M,
shift_limit=0,
border_mode=cv2.BORDER_CONSTANT,
value=0,
p=low_p),
composition.OneOf([
transforms.Blur(blur_limit=7),
albumentations.imgaug.transforms.IAASharpen(),
transforms.GaussNoise(var_limit=(2.0, 10.0), mean=0),
transforms.ISONoise()
],
p=low_p),
transforms.Resize(target_size[0], target_size[1], interpolation=3),
transforms.Normalize(mean=(0.5, 0.5, 0.5),
std=(0.5, 0.5, 0.5),
max_pixel_value=255.0)
],
p=1)
else:
return composition.Compose([
transforms.Resize(target_size[0], target_size[1], interpolation=3),
transforms.Normalize(mean=(0.5, 0.5, 0.5),
std=(0.5, 0.5, 0.5),
max_pixel_value=255.0)
],
p=1)
def augment(im, params=None):
"""
Perform data augmentation on some image using the albumentations package.
Parameters
----------
im : Numpy array
params : dict or None
Contains the data augmentation parameters
Mandatory keys:
- h_flip ([0,1] float): probability of performing an horizontal left-right mirroring.
- v_flip ([0,1] float): probability of performing an vertical up-down mirroring.
- rot ([0,1] float): probability of performing a rotation to the image.
- rot_lim (int): max degrees of rotation.
- stretch ([0,1] float): probability of randomly stretching an image.
- crop ([0,1] float): randomly take an image crop.
- zoom ([0,1] float): random zoom applied to crop_size.
--> Therefore the effective crop size at each iteration will be a
random number between 1 and crop*(1-zoom). For example:
* crop=1, zoom=0: no crop of the image
* crop=1, zoom=0.1: random crop of random size between 100% image and 90% of the image
* crop=0.9, zoom=0.1: random crop of random size between 90% image and 80% of the image
* crop=0.9, zoom=0: random crop of always 90% of the image
Image size refers to the size of the shortest side.
- blur ([0,1] float): probability of randomly blurring an image.
- pixel_noise ([0,1] float): probability of randomly adding pixel noise to an image.
- pixel_sat ([0,1] float): probability of randomly using HueSaturationValue in the image.
- cutout ([0,1] float): probability of using cutout in the image.
Returns
-------
Numpy array
"""
## 1) Crop the image
effective_zoom = np.random.rand() * params['zoom']
crop = params['crop'] - effective_zoom
ly, lx, channels = im.shape
crop_size = int(crop * min([ly, lx]))
rand_x = np.random.randint(low=0, high=lx - crop_size + 1)
rand_y = np.random.randint(low=0, high=ly - crop_size + 1)
crop = transforms.Crop(x_min=rand_x,
y_min=rand_y,
x_max=rand_x + crop_size,
y_max=rand_y + crop_size)
im = crop(image=im)['image']
## 2) Now add the transformations for augmenting the image pixels
transform_list = []
# Add random stretching
if params['stretch']:
transform_list.append(
imgaug_transforms.IAAPerspective(scale=0.1, p=params['stretch']))
# Add random rotation
if params['rot']:
transform_list.append(
transforms.Rotate(limit=params['rot_lim'], p=params['rot']))
# Add horizontal flip
if params['h_flip']:
transform_list.append(transforms.HorizontalFlip(p=params['h_flip']))
# Add vertical flip
if params['v_flip']:
transform_list.append(transforms.VerticalFlip(p=params['v_flip']))
# Add some blur to the image
if params['blur']:
transform_list.append(
albumentations.OneOf([
transforms.MotionBlur(blur_limit=7, p=1.),
transforms.MedianBlur(blur_limit=7, p=1.),
transforms.Blur(blur_limit=7, p=1.),
],
p=params['blur']))
# Add pixel noise
if params['pixel_noise']:
transform_list.append(
albumentations.OneOf(
[
transforms.CLAHE(clip_limit=2, p=1.),
imgaug_transforms.IAASharpen(p=1.),
imgaug_transforms.IAAEmboss(p=1.),
transforms.RandomBrightnessContrast(contrast_limit=0,
p=1.),
transforms.RandomBrightnessContrast(brightness_limit=0,
p=1.),
transforms.RGBShift(p=1.),
transforms.RandomGamma(p=1.) #,
# transforms.JpegCompression(),
# transforms.ChannelShuffle(),
# transforms.ToGray()
],
p=params['pixel_noise']))
# Add pixel saturation
if params['pixel_sat']:
transform_list.append(
transforms.HueSaturationValue(p=params['pixel_sat']))
# Remove randomly remove some regions from the image
if params['cutout']:
ly, lx, channels = im.shape
scale_low, scale_high = 0.05, 0.25 # min and max size of the squares wrt the full image
scale = np.random.uniform(scale_low, scale_high)
transform_list.append(
transforms.Cutout(num_holes=8,
max_h_size=int(scale * ly),
max_w_size=int(scale * lx),
p=params['cutout']))
# Compose all image transformations and augment the image
augmentation_fn = albumentations.Compose(transform_list)
im = augmentation_fn(image=im)['image']
return im
def augment(im, params=None):
"""
Perform data augmentation on some image using the albumentations package.
Parameters
----------
im : Numpy array
params : dict or None
Contains the data augmentation parameters
Mandatory keys:
- h_flip ([0,1] float): probability of performing an horizontal left-right mirroring.
- v_flip ([0,1] float): probability of performing an vertical up-down mirroring.
- rot ([0,1] float): probability of performing a rotation to the image.
- rot_lim (int): max degrees of rotation.
- stretch ([0,1] float): probability of randomly stretching an image.
- expand ([True, False] bool): whether to pad the image to a square shape with background color canvas.
- crop ([0,1] float): randomly take an image crop.
- invert_col ([0, 1] float): randomly invert the colors of the image. p=1 -> invert colors (VPR)
- zoom ([0,1] float): random zoom applied to crop_size.
--> Therefore the effective crop size at each iteration will be a
random number between 1 and crop*(1-zoom). For example:
* crop=1, zoom=0: no crop of the image
* crop=1, zoom=0.1: random crop of random size between 100% image and 90% of the image
* crop=0.9, zoom=0.1: random crop of random size between 90% image and 80% of the image
* crop=0.9, zoom=0: random crop of always 90% of the image
Image size refers to the size of the shortest side.
- blur ([0,1] float): probability of randomly blurring an image.
- pixel_noise ([0,1] float): probability of randomly adding pixel noise to an image.
- pixel_sat ([0,1] float): probability of randomly using HueSaturationValue in the image.
- cutout ([0,1] float): probability of using cutout in the image.
Returns
-------
Numpy array
"""
## 1) Expand the image by padding it with bg-color canvas
if params["expand"]:
desired_size = max(im.shape)
# check bg
if np.argmax(im.shape) > 0:
bgcol = tuple(np.repeat(int(np.mean(im[[0, -1], :, :])), 3))
else:
bgcol = tuple(np.repeat(int(np.mean(im[:, [0, -1], :])), 3))
im = Image.fromarray(im)
old_size = im.size # old_size[0] is in (width, height) format
ratio = float(desired_size) / max(old_size)
new_size = tuple([int(x * ratio) for x in old_size])
im = im.resize(new_size, Image.ANTIALIAS)
# create a new image and paste the resized on it
new_im = Image.new("RGB", (desired_size, desired_size), color=bgcol)
new_im.paste(im, ((desired_size - new_size[0]) // 2,
(desired_size - new_size[1]) // 2))
im = np.array(new_im)
## 2) Crop the image
if params["crop"] and params["crop"] != 1:
effective_zoom = np.random.rand() * params['zoom']
crop = params['crop'] - effective_zoom
ly, lx, channels = im.shape
crop_size = int(crop * min([ly, lx]))
rand_x = np.random.randint(low=0, high=lx - crop_size + 1)
rand_y = np.random.randint(low=0, high=ly - crop_size + 1)
crop = transforms.Crop(x_min=rand_x,
y_min=rand_y,
x_max=rand_x + crop_size,
y_max=rand_y + crop_size)
im = crop(image=im)['image']
if params["enhance"]:
im = Image.fromarray(im)
enhancer = ImageEnhance.Contrast(im)
im = np.array(enhancer.enhance(params["enhance"]))
## 3) Now add the transformations for augmenting the image pixels
transform_list = []
if params['invert_col']:
transform_list.append(transforms.InvertImg(p=params['invert_col']))
# Add random stretching
if params['stretch']:
transform_list.append(
imgaug_transforms.IAAPerspective(scale=0.1, p=params['stretch']))
# Add random rotation
if params['rot']:
transform_list.append(
transforms.Rotate(limit=params['rot_lim'], p=params['rot']))
# Add horizontal flip
if params['h_flip']:
transform_list.append(transforms.HorizontalFlip(p=params['h_flip']))
# Add vertical flip
if params['v_flip']:
transform_list.append(transforms.VerticalFlip(p=params['v_flip']))
# Add some blur to the image
if params['blur']:
transform_list.append(
albumentations.OneOf([
transforms.MotionBlur(blur_limit=7, p=1.),
transforms.MedianBlur(blur_limit=7, p=1.),
transforms.Blur(blur_limit=7, p=1.),
],
p=params['blur']))
# Add pixel noise
if params['pixel_noise']:
transform_list.append(
albumentations.OneOf(
[
transforms.CLAHE(clip_limit=2, p=1.),
imgaug_transforms.IAASharpen(p=1.),
imgaug_transforms.IAAEmboss(p=1.),
transforms.RandomBrightnessContrast(contrast_limit=0,
p=1.),
transforms.RandomBrightnessContrast(brightness_limit=0,
p=1.),
transforms.RGBShift(p=1.),
transforms.RandomGamma(p=1.) #,
# transforms.JpegCompression(),
# transforms.ChannelShuffle(),
# transforms.ToGray()
],
p=params['pixel_noise']))
# Add pixel saturation
if params['pixel_sat']:
transform_list.append(
transforms.HueSaturationValue(p=params['pixel_sat']))
# Remove randomly remove some regions from the image
if params['cutout']:
ly, lx, channels = im.shape
scale_low, scale_high = 0.05, 0.25 # min and max size of the squares wrt the full image
scale = np.random.uniform(scale_low, scale_high)
transform_list.append(
transforms.Cutout(num_holes=8,
max_h_size=int(scale * ly),
max_w_size=int(scale * lx),
p=params['cutout']))
# Compose all image transformations and augment the image
augmentation_fn = albumentations.Compose(transform_list)
im = augmentation_fn(image=im)['image']
return im