def get_transformer(face_policy: str, patch_size: int, net_normalizer: transforms.Normalize, train: bool):
# Transformers and traindb
if face_policy == 'scale':
# The loader crops the face isotropically then scales to a square of size patch_size_load
loading_transformations = [
A.PadIfNeeded(min_height=patch_size, min_width=patch_size,
border_mode=cv2.BORDER_CONSTANT, value=0,always_apply=True),
A.Resize(height=patch_size,width=patch_size,always_apply=True),
]
if train:
downsample_train_transformations = [
A.Downscale(scale_max=0.5, scale_min=0.5, p=0.5), # replaces scaled dataset
]
else:
downsample_train_transformations = []
elif face_policy == 'tight':
# The loader crops the face tightly without any scaling
loading_transformations = [
A.LongestMaxSize(max_size=patch_size, always_apply=True),
A.PadIfNeeded(min_height=patch_size, min_width=patch_size,
border_mode=cv2.BORDER_CONSTANT, value=0,always_apply=True),
]
if train:
downsample_train_transformations = [
A.Downscale(scale_max=0.5, scale_min=0.5, p=0.5), # replaces scaled dataset
]
else:
downsample_train_transformations = []
else:
raise ValueError('Unknown value for face_policy: {}'.format(face_policy))
if train:
aug_transformations = [
A.Compose([
A.HorizontalFlip(),
A.OneOf([
A.RandomBrightnessContrast(),
A.HueSaturationValue(hue_shift_limit=10, sat_shift_limit=30, val_shift_limit=20),
]),
A.OneOf([
A.ISONoise(),
A.IAAAdditiveGaussianNoise(scale=(0.01 * 255, 0.03 * 255)),
]),
A.Downscale(scale_min=0.7, scale_max=0.9, interpolation=cv2.INTER_LINEAR),
A.ImageCompression(quality_lower=50, quality_upper=99),
], )
]
else:
aug_transformations = []
# Common final transformations
final_transformations = [
A.Normalize(mean=net_normalizer.mean, std=net_normalizer.std, ),
ToTensorV2(),
]
transf = A.Compose(
loading_transformations + downsample_train_transformations + aug_transformations + final_transformations)
return transf
def sequence_augmentation(self,
p_apply=0.5,
limit_rotation=40,
limit_translation=0.1,
limit_scale=(-0.2, 0.2)):
if self.rand_choice == 1:
augm = A.Lambda(image=self.aug_dilate, keypoint=self.aug_keypoints)
elif self.rand_choice == 2:
augm = A.Lambda(image=self.aug_erode, keypoint=self.aug_keypoints)
else:
augm = A.NoOp()
transform = A.Compose([
A.Lambda(image=self.aug_morph_close,
keypoint=self.aug_keypoints,
p=1.0), augm,
A.Downscale(scale_min=0.5,
scale_max=0.9,
p=p_apply,
interpolation=cv2.INTER_NEAREST_EXACT),
A.ShiftScaleRotate(limit_translation,
limit_scale,
limit_rotation,
p=p_apply,
border_mode=cv2.BORDER_REFLECT101,
value=-1.0)
],
additional_targets={
'image1': 'image',
'image2': 'image'
},
keypoint_params=A.KeypointParams(
"xy", remove_invisible=False))
return transform
def __init__(self,
metadata: DataFrame,
params: dict,
transform,
data_filter,
frame_num=8):
self.metadata_df = metadata
self.real_filename = list(data_filter(metadata).index)
self.transform = transform
self.frame_num = frame_num
self.same_transform = params['same_transform']
self.smooth = params['label_smoothing']
self.trans = aug.OneOf([
aug.Downscale(0.5, 0.5, p=0.666),
aug.JpegCompression(quality_lower=20, quality_upper=20, p=0.666),
aug.Flip(p=0)
])
self.video_path = pathlib.Path(params['data_path'])
self.cached_path = pathlib.Path(params['cache_path'])
self.cached_path.mkdir(exist_ok=True)
self.data_dropout = params['data_dropout']
self.input_mix = params['input_mix']
np.random.shuffle(self.real_filename)
self.real_filename = self.real_filename[:int(
len(self.real_filename) * (1 - self.data_dropout))]
self.real2fakes = {fn: [] for fn in self.real_filename}
filename_set = set(self.real_filename)
for fn, row in metadata.iterrows():
if row['label'] == 'FAKE' and row['original'] in filename_set:
self.real2fakes[row['original']].append(fn)
def __init__(self, metadata: DataFrame, bbox: DataFrame, params: dict,
transform, data_filter, diff):
self.metadata_df = metadata
self.real_filename = list(data_filter(metadata).index)
self.bbox_df = bbox
self.bbox_index_fn = set(bbox.index.get_level_values(0))
self.transform = transform
self.same_transform = params['same_transform']
self.diff = diff
self.use_diff = params["img_diff"]
self.smooth = params["smooth"]
self.fix_fake = params["fix_fake"]
self.video_path = pathlib.Path(params['data_path'])
self.cached_path = pathlib.Path(params['cache_path'])
self.cached_path.mkdir(exist_ok=True)
self.real2fakes = {fn: [] for fn in self.real_filename}
filename_set = set(self.real_filename)
for fn, row in metadata.iterrows():
if row['label'] == 'FAKE' and row['original'] in filename_set:
self.real2fakes[row['original']].append(fn)
if self.fix_fake == 1:
for key in self.real2fakes.keys():
if len(self.real2fakes[key]) > 0:
self.real2fakes[key] = np.random.choice(
self.real2fakes[key])
import albumentations as aug
self.trans1 = aug.Downscale(0.5, 0.5, p=1)
self.trans2 = aug.JpegCompression(quality_lower=20,
quality_upper=20,
p=1)
def oversampling(self):
self.transform = A.Compose([
A.RandomBrightnessContrast(
always_apply=False,
p=0.4,
brightness_limit=(-0.2, 0.2),
contrast_limit=(-0.2, 0.2),
),
# A.CLAHE(always_apply=False,
# p=0.5,
# clip_limit=(1, 4),
# tile_grid_size=(8, 8)),
A.GaussNoise(always_apply=False, p=0.4, var_limit=(20, 40)),
A.Downscale(
always_apply=False,
p=0.4,
scale_min=0.5,
scale_max=0.8,
interpolation=0,
),
A.HueSaturationValue(
always_apply=False,
p=0.4,
hue_shift_limit=(-20, 20),
sat_shift_limit=(-30, 30),
val_shift_limit=(-20, 20),
),
])
def get_transform(is_train):
if is_train:
return albumentations.Compose(
[
albumentations.Resize(224,224),
albumentations.OneOf([
albumentations.JpegCompression(quality_lower=20, quality_upper=70, p=0.5),
albumentations.Downscale(scale_min=0.25, scale_max=0.50, interpolation=1, p=0.5),
], p=0.6),
albumentations.HorizontalFlip(p=0.5),
albumentations.VerticalFlip(p=0.5),
# albumentations.ShiftScaleRotate(shift_limit=0.0625, scale_limit=0.1, rotate_limit=45),
albumentations.GaussNoise(p=0.2),
albumentations.RandomBrightnessContrast(0.3,0.3, p=0.7),
albumentations.RandomGamma(p=0.2),
albumentations.CLAHE(p=0.2),
albumentations.ChannelShuffle(p=0.2),
albumentations.MultiplicativeNoise(multiplier=[0.5, 1.5], elementwise=True, p=0.3),
albumentations.HueSaturationValue(hue_shift_limit=10, sat_shift_limit=10, val_shift_limit=10, p=0.7),
albumentations.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), max_pixel_value=255.0)
])
else:
return albumentations.Compose(
[
albumentations.Resize(224,224),
albumentations.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), max_pixel_value=255.0)
])
def __init__(self, metadata: DataFrame, bbox: DataFrame, params: dict,
transform, data_filter, diff):
self.metadata_df = metadata
self.real_filename = list(data_filter(metadata).index)
self.bbox_df = bbox
self.bbox_index_fn = set(bbox.index.get_level_values(0))
self.transform = transform
self.same_transform = params['same_transform']
self.diff = diff
self.video_path = pathlib.Path(params['data_path'])
self.cached_path = pathlib.Path(params['cache_path'])
self.cached_path.mkdir(exist_ok=True)
self.real2fakes = {fn: [] for fn in self.real_filename}
filename_set = set(self.real_filename)
for fn, row in metadata.iterrows():
if row['label'] == 'FAKE' and row['original'] in filename_set:
self.real2fakes[row['original']].append(fn)
import albumentations as aug
self.trans = aug.OneOf([
aug.Downscale(0.5, 0.5, p=0.66),
aug.JpegCompression(quality_lower=20, quality_upper=20, p=0.66),
aug.Flip(p=0)
])
def augmentation(p_apply=0.5,
limit_rotation=40,
limit_translation=0.1,
limit_scale=(-0.2, 0.2)):
transform = A.Compose(
[
A.Lambda(image=aug_morph_close, keypoint=aug_keypoints, p=1.0),
A.OneOf([
A.Lambda(image=aug_dilate, keypoint=aug_keypoints),
A.Lambda(image=aug_erode, keypoint=aug_keypoints),
A.NoOp()
],
p=p_apply),
# A.Lambda(image=aug_erode_or_dilate, keypoint=aug_keypoints, p=p_apply),
A.Downscale(scale_min=0.5,
scale_max=0.9,
p=p_apply,
interpolation=cv2.INTER_NEAREST_EXACT),
A.ShiftScaleRotate(limit_translation,
limit_scale,
limit_rotation,
p=p_apply,
border_mode=cv2.BORDER_REFLECT101,
value=-1.0),
A.Lambda(image=cropout, keypoint=aug_keypoints, p=p_apply),
],
keypoint_params=A.KeypointParams("xy", remove_invisible=False))
return transform
def generate_train_transformes(target_size=(50, 50)):
transforms = [
A.Resize(height=target_size[0], width=target_size[1]),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.25),
A.Rotate(limit=(-90, 90), p=1.),
A.RandomResizedCrop(height=target_size[0],
width=target_size[0],
scale=(0.5, 1.),
p=0.5),
A.OneOf([
A.Blur(p=ColorEffectP),
A.GaussNoise(p=ColorEffectP),
A.Downscale(p=ColorEffectP),
A.RGBShift(p=ColorEffectP,
r_shift_limit=20,
g_shift_limit=20,
b_shift_limit=20),
A.RandomGamma(p=ColorEffectP),
A.RandomBrightnessContrast(p=ColorEffectP)
],
p=0.5),
ToTensor()
]
return A.Compose(transforms)
def __down_scale(self,img):
transform = A.Compose([
A.Downscale(scale_min=0.3,scale_max=0.8)
])
transformed = transform(image=img)
transformed_image = transformed["image"]
return transformed_image
def test_downscale(interpolation):
img_float = np.random.rand(100, 100, 3)
img_uint = (img_float * 255).astype("uint8")
aug = A.Downscale(scale_min=0.5, scale_max=0.5, interpolation=interpolation, always_apply=True)
for img in (img_float, img_uint):
transformed = aug(image=img)["image"]
func_applied = F.downscale(img, scale=0.5, interpolation=interpolation)
np.testing.assert_almost_equal(transformed, func_applied)
def set_augmentation(self, flags):
aug_list = []
if flags.blur["p"] > 0.0:
aug_list.append(A.Blur(**flags.blur))
if flags.cutout["p"] > 0.0:
aug_list.append(A.Cutout(**flags.cutout))
if flags.downscale["p"] > 0.0:
aug_list.append(A.Downscale(**flags.downscale))
if flags.crossdrop["p"] > 0.0:
aug_list.append(CrossDrop(**flags.crossdrop))
self.aug = Compose(aug_list) if len(aug_list) != 0 else None
def get_training_augmentation(y=256, x=256):
train_transform = [
albu.RandomBrightnessContrast(p=0.3),
albu.VerticalFlip(p=0.5),
albu.HorizontalFlip(p=0.5),
albu.Downscale(
p=1.0,
scale_min=0.35,
scale_max=0.75,
),
albu.Resize(y, x)
]
return albu.Compose(train_transform)
def acase2_augs(name, **kwargs):
return [
A.Compose(
[A.Posterize(),
A.GridDistortion(num_steps=4),
A.Normalize()],
p=1.0),
A.Compose(
[A.Downscale(),
A.GridDistortion(num_steps=4),
A.Normalize()],
p=1.0),
]
def get_obliterate_augs():
"""
Get the augmentation that can obliterate the hidden signal.
This is used as augmentation to create negative sample from positive one.
:return:
"""
return A.OneOf(
[
A.ImageCompression(quality_lower=70, quality_upper=95, p=1),
A.Downscale(p=1),
A.GaussianBlur(blur_limit=(5, 9), p=1),
],
p=1,
)
def generate_transforms3(img_size):
train_transform = Compose([
A.RandomResizedCrop(img_size, img_size, scale=(0.9, 1), p=1),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.5),
A.ShiftScaleRotate(p=0.5),
A.HueSaturationValue(hue_shift_limit=10,
sat_shift_limit=10,
val_shift_limit=10,
p=0.7),
A.RandomBrightnessContrast(brightness_limit=(-0.2, 0.2),
contrast_limit=(-0.2, 0.2),
p=0.7),
A.CLAHE(clip_limit=(1, 4), p=0.5),
A.OneOf([
A.OpticalDistortion(distort_limit=1.0),
A.GridDistortion(num_steps=5, distort_limit=1.0),
A.ElasticTransform(alpha=3),
],
p=0.2),
A.Resize(img_size, img_size),
A.OneOf([
A.JpegCompression(),
A.Downscale(scale_min=0.1, scale_max=0.15),
],
p=0.2),
A.IAAPerspective(p=0.2),
A.IAASharpen(p=0.2),
A.Cutout(max_h_size=int(img_size * 0.1),
max_w_size=int(img_size * 0.1),
num_holes=5,
p=0.5),
Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
max_pixel_value=255.0,
p=1.0),
ToTensorV2(),
])
val_transform = Compose([
Resize(height=img_size, width=img_size),
Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
max_pixel_value=255.0,
p=1.0),
ToTensorV2(),
])
return {"train": train_transform, "val": val_transform}
def da_policy_downscale(img_size):
print("Using Data Augmentation Downscale")
train_aug = []
train_aug = common_test_augmentation(img_size) + train_aug
train_aug_img = [
albumentations.Downscale(p=0.7,
scale_min=0.4,
scale_max=0.8,
interpolation=0)
]
val_aug = common_test_augmentation(img_size)
return train_aug, train_aug_img, val_aug
def main():
size = (432,432)
color_dic = {1:[255,255,255]}
img_paths = [p.replace('\\', '/') for p in glob('dataset/train/img_aug/**', recursive=True) if os.path.isfile(p)]
mask_paths = list(map(lambda x: x.replace('/img_aug/', '/mask_aug/'), img_paths))
batch_size = 16
splits = math.ceil(len(img_paths)/batch_size)
empty = []
# albumentation
# https://qiita.com/kurilab/items/b69e1be8d0224ae139ad
transforms = albu.OneOf([
albu.ShiftScaleRotate(shift_limit=0.2, scale_limit=0.2, rotate_limit=90),
albu.GaussNoise(),
albu.ISONoise(intensity=(0.7,0.9)),
albu.Downscale(),
albu.ElasticTransform(),
albu.GaussianBlur(),
albu.MultiplicativeNoise(multiplier=(2.0,3.0)),
])
for i in range(splits):
tfrecord_fname = '_record_' + str(i) + '.tfrecord'
save_path = os.path.join('dataset', tfrecord_fname)
# tfrecordのファイルは(画像データ数 / バッチサイズ)分作成する
with tf.io.TFRecordWriter(tfrecord_fname) as writer:
for img_d, mask_d in zip(img_paths[i::splits], mask_paths[i::splits]):
# 画像変形
img = cv2.imread(img_d)
mask = cv2.imread(mask_d)
#augmented = transforms(image=img, mask=mask)
#img, mask = augmented['image'], augmented['mask']
img = cv2.resize(img, (size[0], size[1]), cv2.INTER_NEAREST)
mask = cv2.resize(mask, (size[0], size[1]), cv2.INTER_NEAREST)
# byte列に変換
img = np2byte(img)
mask = np2byte(mask)
#img = np2byte(np.float32(img/127.5 - 1))
#mask = np2byte(convert_mask(mask, color_dic))
# シリアライズして書き出し
proto = serialize_example(img, mask)
writer.write(proto.SerializeToString())
if i>2 : break
def get_augmentation(input_size, train_flag=True, normalize_flag=True):
aug_list = []
###基本的なサイズ変更関数
aug_list.append(A.Resize(height=input_size[0], width=input_size[1], p=1))
##学習用の水増し方法
if train_flag:
aug_list.extend([
A.Flip(),
A.ShiftScaleRotate(shift_limit=(-0.02, 0.02), scale_limit=(-0.05,0.05),
rotate_limit=30, border_mode=0 ,value=[0,0,0],p=0.5),
#色変更
A.RandomBrightnessContrast(brightness_limit=0.5, contrast_limit=0.5,
brightness_by_max=False, p=0.5),
A.HueSaturationValue(hue_shift_limit=10,sat_shift_limit=10, val_shift_limit=10, p=0.5),
#画像品質変換
A.OneOf([
A.OneOf([
A.Blur(blur_limit=5, p=1),
A.GaussianBlur(blur_limit=5, p=1),
], p=1),
A.GaussNoise(var_limit=(10, 80), p=1),
A.Downscale(scale_min=0.5, scale_max=0.5, p=1),
], p=0.4),
A.CoarseDropout(max_holes=4, max_height=int(input_size[0]/8), max_width=int(input_size[0]/8),
min_holes=1, min_height=int(input_size[0]/10), min_width=int(input_size[0]/10),
fill_value=(255,255,255), p=0.3),
])
if normalize_flag:
aug_list.extend([
A.Normalize(
p=1.0),
ToTensorV2(p=1.0)
])
else:
aug_list.extend([
ToTensor(),
])
return A.Compose(aug_list)
def get_augmentation(prob, aug_level, size):
if isinstance(size, list):
size = tuple(size)
augs = []
augs.append(albumentations.HorizontalFlip(prob))
if aug_level > 0:
shift_limit = [None, 0.0625, 0.125, 0.1875][aug_level]
rotate_limit = [None, 20., 40., 60.][aug_level]
scale_limit = [None, 0.1, 0.2, 0.3][aug_level]
augs.append(
albumentations.ShiftScaleRotate(
shift_limit,
scale_limit,
rotate_limit,
p=prob,
border_mode=cv2.BORDER_CONSTANT))
augs.append(albumentations.Downscale(p=prob / 4))
return albumentations.Compose(augs)
def __init__(self, img, data, img_size):
"""
arguments
---------
img : list
list of images, in the original size (height, width, 3)
data : list of dict
Each dict has :
'image' : index of the image. The index should match with img
'mask' : [xx, yy]
IMPORTANT : (WIDTH, HEIGHT)
'box' : [[xmin, ymin], [xmax,ymax]]
'size' : the size of the image that the data was created with
IMPORTANT : (WIDTH, HEIGHT)
img_size : tuple
Desired output image size
The axes will be swapped to match pygame.
IMPORTANT : (WIDTH, HEIGHT)
"""
self.image = img
self.data = data
self.n = len(data)
self.output_size = img_size
self.aug = A.Compose([
A.OneOf([
A.RandomGamma((40, 200), p=1),
A.RandomBrightness(limit=0.5, p=1),
A.RandomContrast(limit=0.5, p=1),
A.RGBShift(40, 40, 40, p=1),
A.Downscale(scale_min=0.25, scale_max=0.5, p=1),
A.ChannelShuffle(p=1),
],
p=0.8),
A.InvertImg(p=0.5),
A.VerticalFlip(p=0.5),
A.RandomRotate90(p=1),
A.Resize(img_size[0], img_size[1]),
], )
for datum in data:
datum['mask_min'] = np.min(datum['mask'], axis=1)
datum['mask_max'] = np.max(datum['mask'], axis=1) + 1
def init_augmentation(self, CONFIGURATION):
"""
Initialization of augmentation function
Parameters stored in CONFIGURATION.
Int: CROP_SIZE_HEIGHT -> Height of Image crop
Int: CROP_SIZE_WIDTH -> Width of Image crop
Float: VERTICAL_FLIP_PROBA -> Probability of vertical flipping the image
Float: HORIZONTAL_FLIP_PROBA -> Probability of horizontal flipping the image
Returns: Augmentation function (albumentations.Compose)
"""
augmentation = albumentations.Compose([
InvertImg(p=CONFIGURATION.INVERT_IMG_PROBA),
albumentations.ShiftScaleRotate(
p=CONFIGURATION.SCR_PROBA,
shift_limit=CONFIGURATION.SCR_SHIFT_LIMIT,
scale_limit=CONFIGURATION.SCR_SCALE_LIMIT,
rotate_limit=CONFIGURATION.SCR_ROTATE_LIMIT),
albumentations.GaussianBlur(blur_limit=CONFIGURATION.BLUR_LIMIT,
p=CONFIGURATION.BLUR_PROBA),
albumentations.Cutout(num_holes=CONFIGURATION.NUM_HOLES,
max_h_size=CONFIGURATION.HOLE_SIZE,
max_w_size=CONFIGURATION.HOLE_SIZE,
p=CONFIGURATION.CUTOUT_PROBA),
albumentations.Downscale(scale_min=CONFIGURATION.SCALE_MIN,
scale_max=CONFIGURATION.SCALE_MAX,
p=CONFIGURATION.DOWNSCALE_PROBA),
albumentations.RandomCrop(CONFIGURATION.CROP_SIZE_HEIGHT,
CONFIGURATION.CROP_SIZE_WIDTH,
p=1.0),
albumentations.HorizontalFlip(
p=CONFIGURATION.HORIZONTAL_FLIP_PROBA),
])
return augmentation
def __init__(self, config, split):
"""
:param opt:
:param split: train/val
"""
super(DATASET_CUSTOM, self).__init__()
self.data_dir = config['dataset']['data_dir']
self.img_dir = os.path.join(self.data_dir, 'images')
self.input_h = config['model']['input_h']
self.input_w = config['model']['input_h']
self.pad = config['model']['pad']
self.down_ratio = config['model']['down_ratio']
self.mean = config['dataset']['mean']
self.std = config['dataset']['std']
self.max_objs = config['dataset']['max_object']
self.num_classes = config['dataset']['num_classes']
self.radius = config['dataset']['radius']
self.annot_path = os.path.join(
self.data_dir, 'annotations',
'{}.json').format(split)
# print(self.data_dir)
self.class_name = ['__background__'] + config['dataset']['label_name']
self._valid_ids = [_id for _id in range(1, self.num_classes+1)] # [1,2,..self.num_classes]
self.cat_ids = {v: i for i, v in enumerate(self._valid_ids)} # {1:0,
self.split = split
print('==> initializing {} data.'.format(split))
self.coco = coco.COCO(self.annot_path)
self.images = self.coco.getImgIds()
self.num_samples = len(self.images)
print('Loaded {} {} samples'.format(split, self.num_samples))
self.output_h = self.input_h // self.down_ratio # 512/4 = 128
self.output_w = self.input_w // self.down_ratio
self.transform_train = A.Compose(
[
A.OneOf([
A.RandomBrightnessContrast(brightness_limit=0.5,
contrast_limit=0.4),
A.RandomGamma(gamma_limit=(50, 150)),
A.NoOp()
]),
A.OneOf([
A.RGBShift(r_shift_limit=20, b_shift_limit=15,
g_shift_limit=15),
A.HueSaturationValue(hue_shift_limit=5,
sat_shift_limit=5),
A.NoOp()
]),
A.HorizontalFlip(p=0.5), #OK
A.ShiftScaleRotate(shift_limit=[0.1, 0.1], scale_limit=[0,0], rotate_limit=[-45, 45], p=0.5, border_mode=cv2.BORDER_CONSTANT, value=(255,255,255)), #OK
A.Downscale(scale_min=0.1, scale_max=0.2, p=0.3), # OK
# A.CoarseDropout(max_holes=5, max_height=100, max_width=100, min_holes=3, min_height=64, min_width=64, p=0.5), # error
A.CLAHE(p=0.5),
A.Resize(height=self.input_h, width=self.input_w, interpolation=cv2.INTER_LINEAR, always_apply=True),
A.Normalize(mean=self.mean, std=self.std, always_apply=True)
],
keypoint_params=A.KeypointParams(format='xy', label_fields=['class_labels'])
)
self.transform_heatmap = A.Compose(
[
A.Resize(height=self.output_h, width=self.output_w, interpolation=cv2.INTER_LINEAR, always_apply=True)
]
, keypoint_params=A.KeypointParams(format='xy')
)
self.transform_test = A.Compose(
[
A.Resize(height=self.input_h, width=self.input_w, interpolation=cv2.INTER_LINEAR, always_apply=True),
A.Normalize(mean=self.mean, std=self.std, always_apply=True)
],
keypoint_params=A.KeypointParams(format='xy')
)
def get_config(runner,
raw_uri,
processed_uri,
root_uri,
test=False,
external_model=False,
external_loss=False,
augment=False):
debug = False
train_scene_info = get_scene_info(join(processed_uri, 'train-scenes.csv'))
val_scene_info = get_scene_info(join(processed_uri, 'val-scenes.csv'))
log_tensorboard = True
run_tensorboard = True
class_config = ClassConfig(names=['no_building', 'building'])
if test:
debug = True
train_scene_info = train_scene_info[0:1]
val_scene_info = val_scene_info[0:1]
def make_scene(scene_info):
(raster_uri, label_uri) = scene_info
raster_uri = join(raw_uri, raster_uri)
label_uri = join(processed_uri, label_uri)
aoi_uri = join(raw_uri, aoi_path)
if test:
crop_uri = join(processed_uri, 'crops',
os.path.basename(raster_uri))
label_crop_uri = join(processed_uri, 'crops',
os.path.basename(label_uri))
save_image_crop(raster_uri,
crop_uri,
label_uri=label_uri,
label_crop_uri=label_crop_uri,
size=600,
min_features=20,
class_config=class_config)
raster_uri = crop_uri
label_uri = label_crop_uri
id = os.path.splitext(os.path.basename(raster_uri))[0]
raster_source = RasterioSourceConfig(channel_order=[0, 1, 2],
uris=[raster_uri])
label_source = ChipClassificationLabelSourceConfig(
vector_source=GeoJSONVectorSourceConfig(uri=label_uri,
default_class_id=1,
ignore_crs_field=True),
ioa_thresh=0.5,
use_intersection_over_cell=False,
pick_min_class_id=False,
background_class_id=0,
infer_cells=True)
return SceneConfig(id=id,
raster_source=raster_source,
label_source=label_source,
aoi_uris=[aoi_uri])
chip_sz = 200
train_scenes = [make_scene(info) for info in train_scene_info]
val_scenes = [make_scene(info) for info in val_scene_info]
dataset = DatasetConfig(class_config=class_config,
train_scenes=train_scenes,
validation_scenes=val_scenes)
if external_model:
model = ClassificationModelConfig(external_def=ExternalModuleConfig(
github_repo='lukemelas/EfficientNet-PyTorch',
# uri='s3://raster-vision-ahassan/models/EfficientNet-PyTorch.zip',
name='efficient_net',
entrypoint='efficientnet_b0',
force_reload=False,
entrypoint_kwargs={
'num_classes': len(class_config.names),
'pretrained': 'imagenet'
}))
else:
model = ClassificationModelConfig(backbone=Backbone.resnet50)
if external_loss:
external_loss_def = ExternalModuleConfig(
github_repo='AdeelH/pytorch-multi-class-focal-loss',
name='focal_loss',
entrypoint='focal_loss',
force_reload=False,
entrypoint_kwargs={
'alpha': [.75, .25],
'gamma': 2
})
else:
external_loss_def = None
solver = SolverConfig(lr=1e-4,
num_epochs=20,
test_num_epochs=4,
batch_sz=32,
one_cycle=True,
external_loss_def=external_loss_def)
if augment:
mu = np.array((0.485, 0.456, 0.406))
std = np.array((0.229, 0.224, 0.225))
aug_transform = A.Compose([
A.Flip(),
A.Transpose(),
A.RandomRotate90(),
A.ShiftScaleRotate(),
A.OneOf([
A.ChannelShuffle(),
A.CLAHE(),
A.FancyPCA(),
A.HueSaturationValue(),
A.RGBShift(),
A.ToGray(),
A.ToSepia(),
]),
A.OneOf([
A.RandomBrightness(),
A.RandomGamma(),
]),
A.OneOf([
A.GaussNoise(),
A.ISONoise(),
A.RandomFog(),
]),
A.OneOf([
A.Blur(),
A.MotionBlur(),
A.ImageCompression(),
A.Downscale(),
]),
A.CoarseDropout(max_height=32, max_width=32, max_holes=5)
])
base_transform = A.Normalize(mean=mu.tolist(), std=std.tolist())
plot_transform = A.Normalize(mean=(-mu / std).tolist(),
std=(1 / std).tolist(),
max_pixel_value=1.)
else:
aug_transform = None
base_transform = None
plot_transform = None
backend = PyTorchChipClassificationConfig(
model=model,
solver=solver,
log_tensorboard=log_tensorboard,
run_tensorboard=run_tensorboard,
test_mode=test,
base_transform=A.to_dict(base_transform),
aug_transform=A.to_dict(aug_transform),
plot_options=PlotOptions(transform=A.to_dict(plot_transform)))
config = ChipClassificationConfig(root_uri=root_uri,
dataset=dataset,
backend=backend,
train_chip_sz=chip_sz,
predict_chip_sz=chip_sz)
return config
def get_augmentations(name, img_size):
if name == 'training_none':
aug = A.Compose([
A.Resize(img_size, img_size),
A.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
ToTensorV2()
])
elif name == 'training_dropout':
aug = A.Compose([
A.Resize(img_size, img_size),
A.CoarseDropout(min_height=int(img_size * 0.05),
min_width=int(img_size * 0.05),
max_height=int(img_size * 0.1),
max_width=int(img_size * 0.1),
min_holes=1,
max_holes=20,
p=0),
A.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
ToTensorV2()
])
elif name == 'training_1':
aug = A.Compose([
A.RandomResizedCrop(img_size, img_size, scale=(0.9, 1), p=1),
A.ShiftScaleRotate(p=0.5),
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(brightness_limit=0.2,
contrast_limit=0.2,
p=0.7),
A.HueSaturationValue(hue_shift_limit=10,
val_shift_limit=10,
sat_shift_limit=10,
p=0.7),
A.CLAHE(clip_limit=(1, 4), p=0.5),
A.OneOf([
A.GaussNoise(var_limit=[10, 50]),
A.GaussianBlur(),
A.MotionBlur(),
A.MedianBlur(),
],
p=0.3),
A.OneOf([
A.OpticalDistortion(distort_limit=1.0),
A.GridDistortion(num_steps=5, distort_limit=1.),
A.ElasticTransform(alpha=3),
],
p=0.3),
A.OneOf([
A.ImageCompression(),
A.Downscale(scale_min=0.1, scale_max=0.15),
],
p=0.2),
A.IAAPiecewiseAffine(p=0.2),
A.IAASharpen(p=0.2),
A.CoarseDropout(max_height=int(img_size * 0.1),
max_width=int(img_size * 0.1),
min_holes=5,
max_holes=10,
p=0.5),
A.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
ToTensorV2()
])
elif name == 'training_2':
aug = A.Compose([
A.RandomResizedCrop(img_size, img_size, scale=(0.9, 1), p=1),
A.ShiftScaleRotate(p=0.5),
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(brightness_limit=0.2,
contrast_limit=0.2,
p=0.7),
A.HueSaturationValue(hue_shift_limit=10,
val_shift_limit=10,
sat_shift_limit=10,
p=0.7),
A.CLAHE(clip_limit=(1, 4), p=0.5),
A.OneOf([
A.GaussNoise(var_limit=[10, 50]),
A.GaussianBlur(),
A.MotionBlur(),
A.MedianBlur(),
],
p=0.3),
A.OneOf([
A.OpticalDistortion(distort_limit=1.0),
A.GridDistortion(num_steps=5, distort_limit=1.),
A.ElasticTransform(alpha=3),
],
p=0.3),
A.OneOf([
A.ImageCompression(),
A.Downscale(scale_min=0.1, scale_max=0.15),
],
p=0.2),
A.IAAPiecewiseAffine(p=0.2),
A.IAASharpen(p=0.2),
A.CoarseDropout(max_height=int(img_size * 0.1),
max_width=int(img_size * 0.1),
min_holes=5,
max_holes=10,
p=0.5),
A.Normalize(),
ToTensorV2()
])
elif name == 'training_2_bis':
aug = A.Compose([
A.RandomResizedCrop(img_size, img_size, scale=(0.9, 1), p=1),
A.ShiftScaleRotate(rotate_limit=30, p=0.5),
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(brightness_limit=0.2,
contrast_limit=0.2,
p=0.7),
A.HueSaturationValue(hue_shift_limit=10,
val_shift_limit=10,
sat_shift_limit=10,
p=0.7),
A.CLAHE(clip_limit=(1, 4), p=0.5),
A.OneOf([
A.GaussNoise(var_limit=[10, 50]),
A.GaussianBlur(),
A.MotionBlur(),
A.MedianBlur()
],
p=0.3),
#A.OneOf([A.OpticalDistortion(distort_limit=1.0), A.GridDistortion(num_steps=5, distort_limit=1.),
# A.ElasticTransform(alpha=3)], p=0.3),
A.OneOf([
A.ImageCompression(),
A.Downscale(scale_min=0.1, scale_max=0.15)
],
p=0.2),
#A.IAAPiecewiseAffine(p=0.2),
A.IAASharpen(p=0.2),
A.CoarseDropout(max_height=int(img_size * 0.1),
max_width=int(img_size * 0.1),
min_holes=5,
max_holes=10,
p=0.5),
A.Normalize(),
ToTensorV2()
])
elif name == 'training_3':
aug = A.Compose([
A.Rotate(limit=5),
A.RandomResizedCrop(img_size, img_size, scale=(0.9, 1), p=1),
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(brightness_limit=0.15,
contrast_limit=0.15,
p=0.5),
A.CoarseDropout(min_height=int(img_size * 0.05),
min_width=int(img_size * 0.05),
max_height=int(img_size * 0.1),
max_width=int(img_size * 0.1),
min_holes=1,
max_holes=10,
p=0.5),
A.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
ToTensorV2()
])
elif name == 'training_4':
aug = A.Compose([
A.Rotate(limit=5, p=1),
A.RandomResizedCrop(img_size, img_size, scale=(0.9, 1), p=1),
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(brightness_limit=(-0.15, +0.25),
contrast_limit=(-0.15, +0.25),
p=1),
A.CLAHE(clip_limit=(1, 4), p=0.5),
A.OneOf([
A.GaussNoise(var_limit=(10, 50)),
A.GaussianBlur(),
A.MotionBlur(),
A.MedianBlur(),
],
p=1),
A.IAASharpen(p=0.3),
A.CoarseDropout(min_height=int(img_size * 0.05),
min_width=int(img_size * 0.05),
max_height=int(img_size * 0.1),
max_width=int(img_size * 0.1),
min_holes=1,
max_holes=20,
p=0),
A.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
ToTensorV2()
])
elif name == 'validation':
aug = A.Compose(
[A.Resize(img_size, img_size),
A.Normalize(),
ToTensorV2()])
elif name == 'none':
aug = A.Compose([A.Resize(img_size, img_size)])
else:
raise ValueError(f"{name} is not a valid augmentations name")
return aug
def get_transform_imagenet(use_albu_aug):
if use_albu_aug:
train_transform = al.Compose([
# al.Flip(p=0.5),
al.Resize(256, 256, interpolation=2),
al.RandomResizedCrop(224,
224,
scale=(0.08, 1.0),
ratio=(3. / 4., 4. / 3.),
interpolation=2),
al.HorizontalFlip(),
al.OneOf(
[
al.OneOf(
[
al.ShiftScaleRotate(
border_mode=cv2.BORDER_CONSTANT,
rotate_limit=30), # , p=0.05),
al.OpticalDistortion(
border_mode=cv2.BORDER_CONSTANT,
distort_limit=5.0,
shift_limit=0.1),
# , p=0.05),
al.GridDistortion(border_mode=cv2.BORDER_CONSTANT
), # , p=0.05),
al.ElasticTransform(
border_mode=cv2.BORDER_CONSTANT,
alpha_affine=15), # , p=0.05),
],
p=0.1),
al.OneOf(
[
al.RandomGamma(), # p=0.05),
al.HueSaturationValue(), # p=0.05),
al.RGBShift(), # p=0.05),
al.CLAHE(), # p=0.05),
al.ChannelShuffle(), # p=0.05),
al.InvertImg(), # p=0.05),
],
p=0.1),
al.OneOf(
[
al.RandomSnow(), # p=0.05),
al.RandomRain(), # p=0.05),
al.RandomFog(), # p=0.05),
al.RandomSunFlare(num_flare_circles_lower=1,
num_flare_circles_upper=2,
src_radius=110),
# p=0.05, ),
al.RandomShadow(), # p=0.05),
],
p=0.1),
al.RandomBrightnessContrast(p=0.1),
al.OneOf(
[
al.GaussNoise(), # p=0.05),
al.ISONoise(), # p=0.05),
al.MultiplicativeNoise(), # p=0.05),
],
p=0.1),
al.OneOf(
[
al.ToGray(), # p=0.05),
al.ToSepia(), # p=0.05),
al.Solarize(), # p=0.05),
al.Equalize(), # p=0.05),
al.Posterize(), # p=0.05),
al.FancyPCA(), # p=0.05),
],
p=0.1),
al.OneOf(
[
# al.MotionBlur(blur_limit=1),
al.Blur(blur_limit=[3, 5]),
al.MedianBlur(blur_limit=[3, 5]),
al.GaussianBlur(blur_limit=[3, 5]),
],
p=0.1),
al.OneOf(
[
al.CoarseDropout(), # p=0.05),
al.Cutout(), # p=0.05),
al.GridDropout(), # p=0.05),
al.ChannelDropout(), # p=0.05),
al.RandomGridShuffle(), # p=0.05),
],
p=0.1),
al.OneOf(
[
al.Downscale(), # p=0.1),
al.ImageCompression(quality_lower=60), # , p=0.1),
],
p=0.1),
],
p=0.5),
al.Normalize(),
ToTensorV2()
])
else:
train_transform = transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
test_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
])
if use_albu_aug:
train_transform = MultiDataTransformAlbu(train_transform)
else:
train_transform = MultiDataTransform(train_transform)
return train_transform, test_transform
'train_transform':
A.Compose([
A.Rotate(border_mode=cv2.BORDER_CONSTANT,
interpolation=cv2.INTER_AREA,
always_apply=True),
A.Flip(),
A.OneOf([
A.CLAHE(tile_grid_size=(5, 5)),
A.RandomBrightnessContrast(),
A.RandomGamma()
]),
A.OneOf(
[A.RGBShift(),
A.HueSaturationValue(),
A.ChannelShuffle(p=0.25)]),
A.Downscale(scale_min=0.3, scale_max=0.5, always_apply=True),
A.GaussNoise(var_limit=(20., 90.), always_apply=True),
A.Blur(blur_limit=12, always_apply=True),
A.Resize(224, 224, always_apply=True),
A.ToFloat()
]),
'val_transform':
A.Compose([A.Resize(224, 224, always_apply=True),
A.ToFloat()])
}
def apply_augmentation(image, is_training):
if is_training:
data = transformations['train_transform'](image=image)
else:
def get_aug(aug_type: str = "val",
task: str = "denoise",
dataset: str = "cifar100",
size: int = 64):
"""
Args:
aug_type: {`val`, `test`, `light`, `medium`}
task: {"denoise", "deblur", "sr"}
dataset: Name of dataset to get MEAN and STD
size: final size of the crop
"""
assert aug_type in ["val", "test", "light", "medium"]
# Add the same noise for all channels for single-channel images
mean, std, max_value = MEAN_STD_BY_NAME[dataset]
if dataset == "medicaldecathlon":
singlechannel = True
normalization = albu.NoOp()
noise = GaussNoiseNoClipping(
singlechannel,
var_limit=0.1,
)
else:
singlechannel = False
normalization = albu.Normalize(mean=mean,
std=std,
max_pixel_value=max_value)
noise = albu.MultiplicativeNoise(multiplier=(0.75, 1.25),
per_channel=True,
elementwise=True)
NORM_TO_TENSOR = albu.Compose(
[normalization, albu_pt.ToTensorV2()],
additional_targets={"mask": "image"})
CROP_AUG = albu.Compose([
albu.PadIfNeeded(size, size),
albu.RandomResizedCrop(size, size, scale=(0.5, 1.)),
])
if task == "deblur":
TASK_AUG = albu.OneOf(
[
albu.Blur(blur_limit=(3, 5)),
albu.GaussianBlur(blur_limit=(3, 5)),
# albu.MotionBlur(),
# albu.MedianBlur(),
# albu.GlassBlur(),
],
p=1.0)
elif task == "denoise":
# TASK_AUG = noise
TASK_AUG = albu.OneOf(
[
noise,
# albu.GaussNoise(),
# GaussNoiseNoClipping(singlechannel, var_limit=0.1 if singlechannel else (20., 50.)),
# albu.GlassBlur(),
# albu.ISONoise(),
# albu.MultiplicativeNoise(),
],
p=1.0)
elif task == "sr":
TASK_AUG = albu.Downscale(scale_min=0.5,
scale_max=0.5,
interpolation=cv2.INTER_CUBIC,
always_apply=True)
else:
raise ValueError("Name of task must be in {'deblur', 'denosie', 'sr'}")
VAL_AUG = albu.Compose([
albu.PadIfNeeded(size, size),
albu.CenterCrop(size, size),
TASK_AUG,
NORM_TO_TENSOR,
])
LIGHT_AUG = albu.Compose([
CROP_AUG,
TASK_AUG,
NORM_TO_TENSOR,
])
MEDIUM_AUG = albu.Compose([
albu.Flip(),
albu.RandomRotate90(), CROP_AUG, TASK_AUG, NORM_TO_TENSOR
])
types = {
"val": VAL_AUG,
"light": LIGHT_AUG,
"medium": MEDIUM_AUG,
}
return types[aug_type]
A.RandomBrightnessContrast(brightness_limit=[0.3,0.3], contrast_limit=[0.3,0.3]), #A.CLAHE(), ]), ], p=1) medium_aug = A.Compose([ A.OneOf([ A.Flip(), A.Rotate(limit=180, border_mode=0, value=0, mask_value=0), A.ElasticTransform(border_mode=0, value=0, p=1), ]), A.OneOf([ A.GaussNoise(var_limit=(50,100), mean=0), A.Downscale(scale_min=0.8, scale_max=0.99), ]), A.OneOf([ A.GaussianBlur(), A.MotionBlur(), A.MedianBlur(blur_limit=(3,5)), ]), A.OneOf([ A.RandomBrightnessContrast(brightness_limit=0.3, contrast_limit=0.3), A.CLAHE(), ]), ], p=0.5) heavy_aug = A.Compose([ A.OneOf([ A.Flip(),
def get_transforms(*, data):
if data == 'train':
import albumentations
return Compose([
albumentations.RandomResizedCrop(CFG.size,
CFG.size,
scale=(0.9, 1),
p=1),
albumentations.HorizontalFlip(p=0.5),
albumentations.ShiftScaleRotate(p=0.5),
albumentations.HueSaturationValue(hue_shift_limit=10,
sat_shift_limit=10,
val_shift_limit=10,
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.size, CFG.size),
albumentations.OneOf([
albumentations.JpegCompression(),
albumentations.Downscale(scale_min=0.1, scale_max=0.15),
],
p=0.2),
albumentations.IAAPiecewiseAffine(p=0.2),
albumentations.IAASharpen(p=0.2),
albumentations.Cutout(max_h_size=int(CFG.size * 0.1),
max_w_size=int(CFG.size * 0.1),
num_holes=5,
p=0.5),
albumentations.Normalize(),
ToTensorV2()
])
# if data == 'train':
# return Compose([
# # Resize(int(CFG.size * 1.25), int(CFG.size * 1.25)),
# #Resize(CFG.size, CFG.size),
# RandomResizedCrop(CFG.size, CFG.size, scale=(0.85, 1.0)),
# # HorizontalFlip(p=0.5),
# # RandomBrightnessContrast(p=0.2, brightness_limit=(-0.2, 0.2), contrast_limit=(-0.2, 0.2)),
# # HueSaturationValue(p=0.2),
# ShiftScaleRotate(p=0.2, shift_limit=0.0625, scale_limit=0.2, rotate_limit=20),
# # # CoarseDropout(p=0.2), # max_holes=8, max_height=8, max_width=8
# # Cutout(p=0.2, max_h_size=16, max_w_size=16, fill_value=(0., 0., 0.), num_holes=16),
# Normalize(
# mean=[0.485, 0.456, 0.406],
# std=[0.229, 0.224, 0.225],
# ),
# ToTensorV2(),
# ])
elif data == 'valid':
return Compose([
Resize(CFG.size, CFG.size),
Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
),
ToTensorV2(),
])
