def aug_mega_hardcore(p=.95):
return Compose([
OneOf([CLAHE(clip_limit=2),
IAASharpen(p=.25),
IAAEmboss(p=.25)],
p=.35),
OneOf([
IAAAdditiveGaussianNoise(p=.3),
GaussNoise(p=.7),
], p=.5),
RandomRotate90(),
Flip(),
Transpose(),
OneOf([
MotionBlur(p=.2),
MedianBlur(blur_limit=3, p=.3),
Blur(blur_limit=3, p=.5),
],
p=.4),
OneOf([
RandomContrast(p=.5),
RandomBrightness(p=.5),
], p=.4),
ShiftScaleRotate(
shift_limit=.0, scale_limit=.45, rotate_limit=45, p=.7),
OneOf([
OpticalDistortion(p=0.3),
GridDistortion(p=0.2),
ElasticTransform(p=.2),
IAAPerspective(p=.2),
IAAPiecewiseAffine(p=.3),
],
p=.6),
HueSaturationValue(p=.5)
],
p=p)
def read_train_img(images_paths):
images = []
gts = []
for image_path in images_paths:
gt_path = image_path.replace('images', 'gt')
image = tifffile.imread(image_path)
gt = tifffile.imread(gt_path)
# 数据扩充
h, w = image.shape[0], image.shape[1]
aug = Compose([
VerticalFlip(p=0.5),
RandomRotate90(p=0.5),
HorizontalFlip(p=0.5),
RandomSizedCrop(min_max_height=(128, 512),
height=h,
width=w,
p=0.5)
])
augmented = aug(image=image, mask=gt)
image = augmented['image']
gt = augmented['mask']
# 数据预处理
image = image / 255.0
gt_temp = gt.copy()
gt[gt_temp == 255] = 1
gt = np.expand_dims(gt, axis=2)
# gt = np_utils.to_categorical(gt, num_classes=1)
images.append(image)
gts.append(gt)
return np.array(images), np.array(gts)
def __init__(self, data, img_size=384, aug=True, mode='train'):
self.data = data
self.mode = mode
if mode is 'train':
self.images = data.ImageId.unique()
self._aug = Compose([
Flip(),
RandomRotate90(),
ShiftScaleRotate(),
Normalize(),
# Resize(256, 256),
Resize(img_size, img_size),
])
elif mode is 'test' or mode is 'val':
self.images = data.ImageId.unique()
self._aug = Compose([
Normalize(),
# Resize(256, 256),
Resize(img_size, img_size),
# PadIfNeeded(768, 768)
])
else:
raise RuntimeError()
def get_aug(p=1.0):
return Compose([
HorizontalFlip(),
VerticalFlip(),
RandomRotate90(),
ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=15,
p=0.9,
border_mode=cv2.BORDER_REFLECT),
OneOf([
OpticalDistortion(p=0.3),
GridDistortion(p=.1),
IAAPiecewiseAffine(p=0.3),
],
p=0.3),
OneOf([
HueSaturationValue(10, 15, 10),
CLAHE(clip_limit=2),
RandomBrightnessContrast(),
],
p=0.3),
],
p=p)
def augmentation_hardcore(size_image, p=0.8):
'''
Only use for second model
About albumentation, p in compose mean the prob that all transform in Compose work
'''
return Compose([
Resize(size_image, size_image),
CenterCrop(height=200, width=200, p=0.5),
Cutout(),
RandomShadow(shadow_dimension=3),
OneOf([
Flip(),
VerticalFlip(),
HorizontalFlip(),
], p=0.5),
OneOf([
RandomRotate90(),
Transpose(),
], p=0.5),
OneOf([GaussNoise(), GaussianBlur(blur_limit=9),
Blur()], p=0.5),
OneOf([
HueSaturationValue(
hue_shift_limit=10, sat_shift_limit=25, val_shift_limit=20),
RGBShift(),
RandomBrightness(brightness_limit=0.4),
RandomContrast(),
RandomBrightnessContrast(),
],
p=0.5),
OneOf([ShiftScaleRotate(),
ElasticTransform(),
RandomGridShuffle()],
p=0.5)
],
p=p)
def __getitem__(self, idx):
label1 = int(self.df[idx][1])
c = str(self.df[idx][0])
image = cv2.imread(c)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = RandomRotate90()(image=image)['image']
image = Flip()(image=image)['image']
image = JpegCompression(quality_lower=9,
quality_upper=10)(image=image)['image']
image = Transpose()(image=image)['image']
image = Downscale()(image=image)['image']
image = IAAAdditiveGaussianNoise()(image=image)['image']
image = Blur(blur_limit=7)(image=image)['image']
image = ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=45)(image=image)['image']
image = IAAPiecewiseAffine()(image=image)['image']
image = RGBShift()(image=image)['image']
image = RandomBrightnessContrast()(image=image)['image']
image = HueSaturationValue()(image=image)['image']
image = transforms.ToPILImage()(image)
if self.transform:
image = self.transform(image)
return image, label1
def create_train_transforms(conf):
height = conf['crop_height']
width = conf['crop_width']
return Compose([
ShiftScaleRotate(shift_limit=0.2, scale_limit=0, rotate_limit=0),
OneOf([
RandomSizedCrop(min_max_height=(int(height * 0.8), int(
height * 1.2)),
w2h_ratio=1.,
height=height,
width=width,
p=0.9),
RandomCrop(height=height, width=width, p=0.1)
],
p=1),
Rotate(limit=10, p=0.2, border_mode=cv2.BORDER_CONSTANT, value=0),
HorizontalFlip(),
VerticalFlip(),
RandomRotate90(),
Transpose(),
OneOf(
[RGBShift(), RandomBrightnessContrast(),
RandomGamma()], p=0.5),
])
def create_train_transforms(size):
return Compose([
# ImageCompression(quality_lower=60, quality_upper=100, p=0.5),
GaussNoise(p=0.1),
GaussianBlur(blur_limit=3, p=0.05),
HorizontalFlip(),
RandomRotate90(),
Resize(height=size[0], width=size[1]),
PadIfNeeded(min_height=size[0],
min_width=size[1],
border_mode=cv2.BORDER_CONSTANT),
OneOf([RandomBrightnessContrast(),
FancyPCA(),
HueSaturationValue()],
p=0.7),
ToGray(p=0.1),
ShiftScaleRotate(shift_limit=0.1,
scale_limit=0.2,
rotate_limit=10,
border_mode=cv2.BORDER_CONSTANT,
p=0.5),
Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
ToTensorV2()
])
def __init__(
self,
image_path='../../data/training/images/',
groundtruth_path='../../data/training/groundtruth/',
val_split=0.1,
additional_images_path='../../additional_data_generation/additional_data/images/',
additional_masks_path='../../additional_data_generation/additional_data/masks/'
):
training_images = os.listdir(image_path)
training_truths = os.listdir(groundtruth_path)
# shuffling the images so to obtain a random train-test split
zipped = list(zip(training_images, training_truths))
random.shuffle(zipped)
training_images, training_truths = zip(*zipped)
self.images = []
self.truths = []
self.validation_images = []
self.validation_truths = []
self.additional_images = []
self.additional_masks = []
self.treshold = 0.25
counter = int((val_split) * len(training_images))
print('Reading images...', flush=True)
for i, t in tqdm(list(zip(training_images, training_truths))):
if counter > 0 and val_split != 0.0:
self.validation_images.append(imread(image_path + i))
self.validation_truths.append(
rgb2gray(imread(groundtruth_path + t)))
counter -= 1
self.images.append(imread(image_path + i))
self.truths.append(rgb2gray(imread(groundtruth_path + t)))
print('Done!')
additional_paths = [
p for p in list(os.listdir(additional_images_path))
if p.endswith('.png')
]
random.shuffle(additional_paths)
print('Reading additional data...', flush=True)
for p in tqdm(additional_paths):
self.additional_images.append(imread(additional_images_path + p))
self.additional_masks.append(
rgb2gray(imread(additional_masks_path + p)))
print('Done!')
self.albument_p = .5
self.albumenters_1 = [
VerticalFlip(p=self.albument_p),
HorizontalFlip(p=self.albument_p),
RandomRotate90(p=self.albument_p),
ElasticTransform(alpha=1,
sigma=50,
alpha_affine=50,
p=self.albument_p)
]
self.albumenters_2 = [
RandomContrast(limit=.6, p=self.albument_p),
HueSaturationValue(hue_shift_limit=20,
sat_shift_limit=30,
val_shift_limit=20,
p=self.albument_p),
RandomBrightness(limit=0.2, p=self.albument_p)
]
def policy_transform(split,
policies=None,
size=224,
per_image_norm=False,
mean_std=None,
**kwargs):
means = np.array([127.5, 127.5, 127.5, 127.5])
stds = np.array([255.0, 255.0, 255.0, 255.0])
base_aug = Compose([
RandomRotate90(),
Flip(),
Transpose(),
])
if policies is None:
policies = []
if isinstance(policies, str):
with open(policies, 'r') as fid:
policies = eval(fid.read())
policies = itertools.chain.from_iterable(policies)
aug_list = []
for policy in policies:
op_1, params_1 = policy[0]
op_2, params_2 = policy[1]
print('op_1 ', op_1, ' pa_1 ', params_1)
print('op_2 ', op_2, ' pa_2 ', params_2)
aug = Compose([
globals().get(op_1)(**params_1),
globals().get(op_2)(**params_2),
])
aug_list.append(aug)
print('len(aug_list):', len(aug_list))
resize = Resize(height=size, width=size, always_apply=True)
def transform(image):
image = np.array(image)
if split == 'train':
# image = base_aug(image=image)['image']
# if len(aug_list) > 0:
# aug = random.choice(aug_list)
# image = aug(image=image)['image']
# print('image shape ', image.shape)
# image = resize(image=image)['image']
# image = misc.imresize(image, (size, size)).astype('float32')
image = cv2.resize(image, (size, size))
transform = transforms.Compose([
transforms.ToPILImage(),
transforms.ToTensor(), # range [0, 255] -> [0.0,1.0]
transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5]),
transforms.Normalize(mean = (0.5, 0.5, 0.5), std = (0.5, 0.5, 0.5))
])
image = transform(image)
else:
# if size != image.shape[0]:
# image = resize(image=image)['image']
#image = misc.imresize(image, (size, size)).astype('float32')
image = cv2.resize(image, (size, size))
transform = transforms.Compose([
transforms.ToPILImage(),
transforms.ToTensor(), # range [0, 255] -> [0.0,1.0]
transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5]),
transforms.Normalize(mean = (0.5, 0.5, 0.5), std = (0.5, 0.5, 0.5))
])
image = transform(image)
# image = image.astype(np.float32)
# if per_image_norm:
# mean = np.mean(image.reshape(-1, 3), axis=0)
# std = np.std(image.reshape(-1, 3), axis=0)
# image -= mean
# image /= (std + 0.0000001)
# else:
# image -= means
# image /= stds
# image = np.transpose(image, (2, 0, 1))
return image
return transform
if __name__ == '__main__':
import matplotlib.pyplot as plt
from tqdm import tqdm
from albumentations import Compose, OneOf, Normalize
from albumentations import HorizontalFlip, VerticalFlip, RandomRotate90, RandomScale, RandomCrop
from albumentations.pytorch import ToTensorV2
dataset = Xview2(
r'D:\DATA\xView2\train\images',
r'D:\DATA\xView2\train\labels',
transforms=Compose([
OneOf([
HorizontalFlip(True),
VerticalFlip(True),
RandomRotate90(True)
],
p=0.5),
# RandomDiscreteScale([0.75, 1.25, 1.5], p=0.5),
RandomCrop(640, 640, True),
Normalize(mean=(0.485, 0.456, 0.406, 0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225, 0.229, 0.224, 0.225),
max_pixel_value=255),
ToTensorV2(True),
]),
include=('pre', 'post')).pairwise_mode()
print(len(dataset))
a = dataset[1]
print()
# img, mask = dataset[4]
def train(model, cfg, model_cfg, start_epoch=0):
cfg.batch_size = 28 if cfg.batch_size < 1 else cfg.batch_size
cfg.val_batch_size = cfg.batch_size
cfg.input_normalization = model_cfg.input_normalization
crop_size = model_cfg.crop_size
loss_cfg = edict()
loss_cfg.instance_loss = SigmoidBinaryCrossEntropyLoss()
loss_cfg.instance_loss_weight = 1.0
num_epochs = 120
num_masks = 1
train_augmentator = Compose([
Flip(),
RandomRotate90(),
ShiftScaleRotate(shift_limit=0.03,
scale_limit=0,
rotate_limit=(-3, 3),
border_mode=0,
p=0.75),
PadIfNeeded(
min_height=crop_size[0], min_width=crop_size[1], border_mode=0),
RandomCrop(*crop_size),
RandomBrightnessContrast(brightness_limit=(-0.25, 0.25),
contrast_limit=(-0.15, 0.4),
p=0.75),
RGBShift(r_shift_limit=10, g_shift_limit=10, b_shift_limit=10, p=0.75)
],
p=1.0)
val_augmentator = Compose([
PadIfNeeded(
min_height=crop_size[0], min_width=crop_size[1], border_mode=0),
RandomCrop(*crop_size)
],
p=1.0)
def scale_func(image_shape):
return random.uniform(0.75, 1.25)
points_sampler = MultiPointSampler(model_cfg.num_max_points,
prob_gamma=0.7,
merge_objects_prob=0.15,
max_num_merged_objects=2)
trainset = SBDDataset(
cfg.SBD_PATH,
split='train',
num_masks=num_masks,
augmentator=train_augmentator,
points_from_one_object=False,
input_transform=model_cfg.input_transform,
min_object_area=80,
keep_background_prob=0.0,
image_rescale=scale_func,
points_sampler=points_sampler,
samples_scores_path='./models/sbd/sbd_samples_weights.pkl',
samples_scores_gamma=1.25)
valset = SBDDataset(cfg.SBD_PATH,
split='val',
augmentator=val_augmentator,
num_masks=num_masks,
points_from_one_object=False,
input_transform=model_cfg.input_transform,
min_object_area=80,
image_rescale=scale_func,
points_sampler=points_sampler)
optimizer_params = {'lr': 5e-4, 'betas': (0.9, 0.999), 'eps': 1e-8}
lr_scheduler = partial(torch.optim.lr_scheduler.MultiStepLR,
milestones=[100],
gamma=0.1)
trainer = ISTrainer(model,
cfg,
model_cfg,
loss_cfg,
trainset,
valset,
optimizer_params=optimizer_params,
lr_scheduler=lr_scheduler,
checkpoint_interval=5,
image_dump_interval=200,
metrics=[AdaptiveIoU()],
max_interactive_points=model_cfg.num_max_points)
logger.info(f'Starting Epoch: {start_epoch}')
logger.info(f'Total Epochs: {num_epochs}')
for epoch in range(start_epoch, num_epochs):
trainer.training(epoch)
trainer.validation(epoch)
def policy_transform(split,
policies=None,
size=512,
per_image_norm=False,
mean_std=None,
**kwargs):
means = np.array([127.5, 127.5, 127.5, 127.5])
stds = np.array([255.0, 255.0, 255.0, 255.0])
base_aug = Compose([
RandomRotate90(),
Flip(),
Transpose(),
])
if policies is None:
policies = []
if isinstance(policies, str):
with open(policies, 'r') as fid:
policies = eval(fid.read())
policies = itertools.chain.from_iterable(policies)
aug_list = []
for policy in policies:
op_1, params_1 = policy[0]
op_2, params_2 = policy[1]
print('op_1 ', op_1, ' pa_1 ', params_1)
print('op_2 ', op_2, ' pa_2 ', params_2)
aug = Compose([
globals().get(op_1)(**params_1),
globals().get(op_2)(**params_2),
])
aug_list.append(aug)
print('len(aug_list):', len(aug_list))
resize = Resize(height=size, width=size, always_apply=True)
def transform(image):
if split == 'train':
image = base_aug(image=image)['image']
if len(aug_list) > 0:
aug = random.choice(aug_list)
image = aug(image=image)['image']
image = resize(image=image)['image']
else:
if size != image.shape[0]:
image = resize(image=image)['image']
image = image.astype(np.float32)
if per_image_norm:
mean = np.mean(image.reshape(-1, 4), axis=0)
std = np.std(image.reshape(-1, 4), axis=0)
image -= mean
image /= (std + 0.0000001)
else:
image -= means
image /= stds
image = np.transpose(image, (2, 0, 1))
return image
return transform
from albumentations import Compose, Resize, RandomCrop, Flip, HorizontalFlip, VerticalFlip, Transpose, RandomRotate90, \
ShiftScaleRotate, OneOf, OpticalDistortion
from albumentations.pytorch import ToTensor
train_aug = Compose([
RandomCrop(height=96, width=96, p=0.2),
OneOf([
VerticalFlip(p=0.2),
HorizontalFlip(p=0.3),
Transpose(p=0.2),
RandomRotate90(p=0.2),
],
p=0.3),
ShiftScaleRotate(p=0.2),
OpticalDistortion(p=0.2),
Resize(128, 128, always_apply=True),
ToTensor()
])
valid_aug = Compose([Resize(128, 128, always_apply=True), ToTensor()])
def __getitem__(self, idx):
if len(self.channels) < 2:
raise Exception('You have to specify at least two channels.')
data_info_row = self.df.iloc[idx]
instance_name = '_'.join(
[data_info_row['name'], data_info_row['position']])
images_array, masks_array = [], []
for k in range(1, self.num_images + 1):
image_path = get_filepath(self.dataset_path,
data_info_row['dataset_folder'],
self.images_folder,
instance_name + f'_{k}',
file_type=self.image_type)
img = filter_by_channels(read_tensor(image_path), self.channels, 1)
images_array.append(img)
mask_path = get_filepath(self.dataset_path,
data_info_row['dataset_folder'],
self.masks_folder,
instance_name,
file_type=self.mask_type)
masks_array = read_tensor(mask_path)
if self.phase == 'train':
aug = Compose([
RandomRotate90(),
Flip(),
OneOf(
[
RandomSizedCrop(min_max_height=(int(
self.image_size * 0.7), self.image_size),
height=self.image_size,
width=self.image_size),
RandomBrightnessContrast(brightness_limit=0.15,
contrast_limit=0.15),
#MedianBlur(blur_limit=3, p=0.2),
MaskDropout(p=0.6),
ElasticTransform(alpha=15, sigma=5, alpha_affine=5),
GridDistortion(p=0.6)
],
p=0.8),
ToTensor()
])
else:
aug = ToTensor()
'''
keys = ['image']
values = [images_array[0]]
for k in range(self.num_images-1):
keys.append(f'image{k}')
values.append(images_array[k+1])
keys.append('mask')
values.append(masks_array)
#{"image" : images_array[0], "image2" : images_array[1], ..., "mask": masks_array, ...}
aug_input = { keys[i] : values[i] for i in range(len(keys)) }
augmented = aug(**aug_input)
augmented_images = [augmented['image']]
for k in range(self.num_images-1):
augmented_images.append(np.transpose(augmented[f'image{k}'], ( 2, 0, 1))/255)
augmented_masks = [augmented['mask']]
return {'features': augmented_images, 'targets': augmented_masks, 'name': data_info_row['name'], 'position': data_info_row['position']}
'''
augmented = aug(image=np.concatenate(
(images_array[0], images_array[1]), axis=-1),
mask=masks_array)
augmented_images = [
augmented['image'][:count_channels(self.channels), :, :],
augmented['image'][count_channels(self.channels):, :, :]
]
augmented_masks = [augmented['mask']]
return {
'features': augmented_images,
'targets': augmented_masks,
'name': data_info_row['name'],
'position': data_info_row['position']
}
import re
import os
import glob
import numpy as np
import argparse
import matplotlib.pyplot as plt
from albumentations import (HorizontalFlip, VerticalFlip, Transpose,
RandomRotate90)
from PIL import Image
NUM_LABELS = 12
AUGMENTS = [
RandomRotate90(p=1),
HorizontalFlip(p=1),
VerticalFlip(p=1),
Transpose(p=1)
]
parser = argparse.ArgumentParser()
parser.add_argument('--dataroot',
required=True,
help='expriments results path')
opt = parser.parse_args()
def read(index, ext):
file_name = f'{index}.{ext}'
img_path = os.path.join(opt.dataroot, 'train', file_name)
from albumentations import Compose, ShiftScaleRotate, PadIfNeeded, RandomCrop, Resize, RandomSizedCrop, CLAHE, \
RandomRotate90, Flip, OneOf, MotionBlur, MedianBlur, Blur, CenterCrop, LongestMaxSize, HorizontalFlip, VerticalFlip, \
Transpose
from albumentations.pytorch import ToTensor
transform_train = Compose([
RandomRotate90(0.2),
Flip(p=0.2),
ShiftScaleRotate(),
OneOf([
MotionBlur(p=.2),
MedianBlur(blur_limit=3, p=0.1),
Blur(blur_limit=3, p=0.1),
],
p=0.2),
ToTensor()
])
transform_test = Compose([ToTensor()])
IMG_SIZE_RAW = 224
RAW_CROP_SIZE = 448
transform_train_raw = Compose([
Resize(RAW_CROP_SIZE, RAW_CROP_SIZE),
# CenterCrop(width=IMG_SIZE_RAW, height=IMG_SIZE_RAW),
RandomRotate90(0.2),
Flip(p=0.2),
ShiftScaleRotate(),
ToTensor()
])
def __init__(self, config):
super(AugmentedPair3, self).__init__(config)
additional_targets = {
"image{}".format(i): "image" for i in range(1, self.n_images)
}
p = 0.9
appearance_augmentation = Compose(
[
OneOf(
[
MedianBlur(blur_limit=3, p=0.1),
Blur(blur_limit=3, p=0.1),
],
p=0.5,
),
OneOf(
[
RandomBrightnessContrast(p=0.3),
RGBShift(p=0.3),
HueSaturationValue(p=0.3),
],
p=0.8,
),
OneOf(
[
RandomBrightnessContrast(p=0.3),
RGBShift(p=0.3),
HueSaturationValue(p=0.3),
],
p=0.8,
),
OneOf(
[
RandomBrightnessContrast(p=0.3),
RGBShift(p=0.3),
HueSaturationValue(p=0.3),
],
p=0.8,
),
ToGray(p=0.1),
ChannelShuffle(p=0.3),
],
p=p,
additional_targets=additional_targets,
)
self.appearance_augmentation = appearance_augmentation
p = 0.9
shape_augmentation = Compose(
[
OneOf([
Transpose(p=0.5),
HorizontalFlip(p=0.5),
], p=0.9),
OneOf([
RandomRotate90(p=1.0),
], p=0.9),
ShiftScaleRotate(shift_limit=0.0625, scale_limit=0.25, rotate_limit=25, p=0.3, border_mode=cv2.BORDER_REPLICATE),
OneOf([
# OpticalDistortion(p=0.3),
# GridDistortion(p=0.1),
IAAPiecewiseAffine(p=0.5),
ElasticTransform(p=0.5, border_mode=cv2.BORDER_REPLICATE)
], p=0.3),
],
p=p,
additional_targets=additional_targets,
)
self.shape_augmentation = shape_augmentation
def __init__(self, prob):
self.horizontal = HorizontalFlip(p=prob)
self.vertical = VerticalFlip(p=prob)
self.rotate = RandomRotate90(p=prob)
self.transpose = Transpose(p=prob)
def __getitem__(self, idx):
if len(self.channels) < 2:
raise Exception('You have to specify at least two channels.')
data_info_row = self.df.iloc[idx]
instance_name = '_'.join(
[data_info_row['name'], data_info_row['position']])
images_array, masks_array = [], []
#for k in range(1,self.num_images+1):
for k in range(self.num_images, 0, -1):
image_path = get_filepath(self.dataset_path,
data_info_row['dataset_folder'],
self.images_folder,
instance_name + f'_{k}',
file_type=self.image_type)
img = filter_by_channels(read_tensor(image_path), self.channels, 1)
images_array.append(img)
mask_path = get_filepath(self.dataset_path,
data_info_row['dataset_folder'],
self.masks_folder,
instance_name + f'_{k}',
file_type=self.mask_type)
msk = read_tensor(mask_path)
masks_array.append(np.expand_dims(msk, axis=-1))
aug = Compose([
RandomRotate90(),
Flip(),
OneOf([
RandomSizedCrop(min_max_height=(int(
self.image_size * 0.7), self.image_size),
height=self.image_size,
width=self.image_size),
RandomBrightnessContrast(brightness_limit=0.15,
contrast_limit=0.15),
ElasticTransform(alpha=15, sigma=5, alpha_affine=5),
GridDistortion(p=0.6)
],
p=0.8),
ToTensor()
])
augmented = aug(image=np.concatenate(images_array, axis=-1),
mask=np.concatenate(masks_array, axis=-1))
augmented_images = torch.stack([
augmented['image'][num_img *
count_channels(self.channels):(num_img + 1) *
count_channels(self.channels), :, :]
for num_img in range(self.num_images)
])
if self.all_masks:
augmented_masks = torch.stack([
augmented['mask'][:, :, :, i]
for i in range(augmented['mask'].shape[-1])
]).squeeze()
else:
augmented_masks = torch.stack([augmented['mask'][:, :, :, -1]])
return {
'features': augmented_images,
'targets': augmented_masks,
'name': data_info_row['name'],
'position': data_info_row['position']
}
HorizontalFlip,
VerticalFlip,
CenterCrop,
Compose,
RandomRotate90,
RandomBrightnessContrast,
Normalize, RandomCrop, Blur
)
channel_max = [92.87763, 91.97153, 91.65466, 91.9873]
transform_rgb = {
'train': Compose(
[RandomCrop(512, 512, p=1.0),
HorizontalFlip(p=.5),
VerticalFlip(p=.5), RandomRotate90(p=0.5),
RandomBrightnessContrast(p=0.3),
Blur(p=0.3),
Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
ToTensor(),
]),
'valid': Compose(
[CenterCrop(512, 512),
Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
ToTensor()]),
'inference': Compose(
[
Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
def __init__(self,
imgs: Sequence[str] = None,
suffix: str = '.path',
line_width: int = 4,
im_transforms: Callable[[Any],
torch.Tensor] = transforms.Compose(
[]),
mode: str = 'path',
augmentation: bool = False,
valid_baselines: Sequence[str] = None,
merge_baselines: Dict[str, Sequence[str]] = None,
valid_regions: Sequence[str] = None,
merge_regions: Dict[str, Sequence[str]] = None):
"""
Reads a list of image-json pairs and creates a data set.
Args:
imgs (list):
suffix (int): Suffix to attach to image base name to load JSON
files from.
line_width (int): Height of the baseline in the scaled input.
target_size (tuple): Target size of the image as a (height, width) tuple.
mode (str): Either path, alto, page, xml, or None. In alto, page,
and xml mode the baseline paths and image data is
retrieved from an ALTO/PageXML file. In `None` mode
data is iteratively added through the `add` method.
augmentation (bool): Enable/disable augmentation.
valid_baselines (list): Sequence of valid baseline identifiers. If
`None` all are valid.
merge_baselines (dict): Sequence of baseline identifiers to merge.
Note that merging occurs after entities not
in valid_* have been discarded.
valid_regions (list): Sequence of valid region identifiers. If
`None` all are valid.
merge_regions (dict): Sequence of region identifiers to merge.
Note that merging occurs after entities not
in valid_* have been discarded.
"""
super().__init__()
self.mode = mode
self.im_mode = '1'
self.aug = None
self.targets = []
# n-th entry contains semantic of n-th class
self.class_mapping = {
'aux': {
'_start_separator': 0,
'_end_separator': 1
},
'baselines': {},
'regions': {}
}
self.class_stats = {
'baselines': defaultdict(int),
'regions': defaultdict(int)
}
self.num_classes = 2
self.mbl_dict = merge_baselines if merge_baselines is not None else {}
self.mreg_dict = merge_regions if merge_regions is not None else {}
self.valid_baselines = valid_baselines
self.valid_regions = valid_regions
if mode in ['alto', 'page', 'xml']:
if mode == 'alto':
fn = parse_alto
elif mode == 'page':
fn = parse_page
elif mode == 'xml':
fn = parse_xml
im_paths = []
self.targets = []
for img in imgs:
try:
data = fn(img)
im_paths.append(data['image'])
lines = defaultdict(list)
for line in data['lines']:
if valid_baselines is None or line[
'script'] in valid_baselines:
lines[self.mbl_dict.get(line['script'],
line['script'])].append(
line['baseline'])
self.class_stats['baselines'][self.mbl_dict.get(
line['script'], line['script'])] += 1
regions = defaultdict(list)
for k, v in data['regions'].items():
if valid_regions is None or k in valid_regions:
regions[self.mreg_dict.get(k, k)].extend(v)
self.class_stats['regions'][self.mreg_dict.get(
k, k)] += len(v)
data['regions'] = regions
self.targets.append({
'baselines': lines,
'regions': data['regions']
})
except KrakenInputException as e:
logger.warning(e)
continue
# get line types
imgs = im_paths
# calculate class mapping
line_types = set()
region_types = set()
for page in self.targets:
for line_type in page['baselines'].keys():
line_types.add(line_type)
for reg_type in page['regions'].keys():
region_types.add(reg_type)
idx = -1
for idx, line_type in enumerate(line_types):
self.class_mapping['baselines'][
line_type] = idx + self.num_classes
self.num_classes += idx + 1
idx = -1
for idx, reg_type in enumerate(region_types):
self.class_mapping['regions'][
reg_type] = idx + self.num_classes
self.num_classes += idx + 1
elif mode == 'path':
pass
elif mode is None:
imgs = []
else:
raise Exception('invalid dataset mode')
if augmentation:
from albumentations import (
Compose,
ToFloat,
FromFloat,
RandomRotate90,
Flip,
OneOf,
MotionBlur,
MedianBlur,
Blur,
ShiftScaleRotate,
OpticalDistortion,
ElasticTransform,
RandomBrightnessContrast,
HueSaturationValue,
)
self.aug = Compose([
ToFloat(),
RandomRotate90(),
Flip(),
OneOf([
MotionBlur(p=0.2),
MedianBlur(blur_limit=3, p=0.1),
Blur(blur_limit=3, p=0.1),
],
p=0.2),
ShiftScaleRotate(shift_limit=0.0625,
scale_limit=0.2,
rotate_limit=45,
p=0.2),
OneOf([
OpticalDistortion(p=0.3),
ElasticTransform(p=0.1),
],
p=0.2),
HueSaturationValue(hue_shift_limit=20,
sat_shift_limit=0.1,
val_shift_limit=0.1,
p=0.3),
],
p=0.5)
self.imgs = imgs
self.line_width = line_width
# split image transforms into two. one part giving the final PIL image
# before conversion to a tensor and the actual tensor conversion part.
self.head_transforms = transforms.Compose(im_transforms.transforms[:2])
self.tail_transforms = transforms.Compose(im_transforms.transforms[2:])
self.seg_type = None
# gray
norm_mean = [0.46152964]
norm_std = [0.10963361]
# RGB
# norm_mean = [0.4976264, 0.45133978, 0.3993562]
# norm_std = [0.11552592, 0.10886826, 0.10727626]
# Albumentations
train_Transform = Compose(
[
Resize(height=args.inputsize,
width=args.inputsize,
interpolation=Image.NEAREST,
p=1),
RandomRotate90(0.5),
Flip(p=0.5),
ShiftScaleRotate(p=0.2, interpolation=Image.NEAREST
), # , border_mode=cv2.BORDER_CONSTANT, value=0
# Normalize(mean=norm_mean, std=norm_std),
],
p=1.0)
valid_Transform = Compose(
[
Resize(height=args.inputsize,
width=args.inputsize,
interpolation=Image.NEAREST,
p=1),
# Normalize(mean=norm_mean, std=norm_std),
],
def train(inputs, working_dir, fold_id):
start_epoch, step = 0, 0
# TopCoder
num_workers, batch_size = 8, 4 * 8
gpus = [0, 1, 2, 3]
# My machine
# num_workers, batch_size = 8, 2 * 3
# gpus = [0, 1]
patience, n_epochs = 8, 150
lr, min_lr, lr_update_rate = 1e-4, 5e-5, 0.5
training_timelimit = 60 * 60 * 24 * 2 # 2 days
st_time = time.time()
model = unet_vgg16(pretrained=True)
model = nn.DataParallel(model, device_ids=gpus).cuda()
train_transformer = Compose([
HorizontalFlip(p=0.5),
RandomRotate90(p=0.5),
RandomCrop(512, 512, p=1.0),
Normalize(),
],
p=1.0)
val_transformer = Compose([
CenterCrop(512, 512, p=1.0),
Normalize(),
],
p=1.0)
# train/val loadrs
df_cvfolds = read_cv_splits(inputs)
trn_loader, val_loader = make_train_val_loader(train_transformer,
val_transformer, df_cvfolds,
fold_id, batch_size,
num_workers)
# train
criterion = binary_loss(jaccard_weight=0.25)
optimizer = Adam(model.parameters(), lr=lr)
report_epoch = 10
model_name = f'v12_f{fold_id}'
fh = open_log(model_name)
# vers for early stopping
best_score = 0
not_improved_count = 0
for epoch in range(start_epoch, n_epochs):
model.train()
tl = trn_loader # alias
trn_metrics = Metrics()
try:
tq = tqdm.tqdm(total=(len(tl) * trn_loader.batch_size))
tq.set_description(f'Ep{epoch:>3d}')
for i, (inputs, targets, labels, names) in enumerate(trn_loader):
inputs = inputs.cuda()
targets = targets.cuda()
outputs = model(inputs)
loss = criterion(outputs, targets)
optimizer.zero_grad()
# Increment step counter
batch_size = inputs.size(0)
loss.backward()
optimizer.step()
step += 1
tq.update(batch_size)
# Update eval metrics
trn_metrics.loss.append(loss.item())
trn_metrics.bce.append(criterion._stash_bce_loss.item())
trn_metrics.jaccard.append(criterion._stash_jaccard.item())
if i > 0 and i % report_epoch == 0:
report_metrics = Bunch(
epoch=epoch,
step=step,
trn_loss=np.mean(trn_metrics.loss[-report_epoch:]),
trn_bce=np.mean(trn_metrics.bce[-report_epoch:]),
trn_jaccard=np.mean(
trn_metrics.jaccard[-report_epoch:]),
)
write_event(fh, **report_metrics)
tq.set_postfix(loss=f'{report_metrics.trn_loss:.5f}',
bce=f'{report_metrics.trn_bce:.5f}',
jaccard=f'{report_metrics.trn_jaccard:.5f}')
# End of epoch
report_metrics = Bunch(
epoch=epoch,
step=step,
trn_loss=np.mean(trn_metrics.loss[-report_epoch:]),
trn_bce=np.mean(trn_metrics.bce[-report_epoch:]),
trn_jaccard=np.mean(trn_metrics.jaccard[-report_epoch:]),
)
write_event(fh, **report_metrics)
tq.set_postfix(loss=f'{report_metrics.trn_loss:.5f}',
bce=f'{report_metrics.trn_bce:.5f}',
jaccard=f'{report_metrics.trn_jaccard:.5f}')
tq.close()
save(model, epoch, step, model_name)
# Run validation
val_metrics = validation(model, criterion, val_loader, epoch, step,
fh)
report_val_metrics = Bunch(
epoch=epoch,
step=step,
val_loss=np.mean(val_metrics.loss[-report_epoch:]),
val_bce=np.mean(val_metrics.bce[-report_epoch:]),
val_jaccard=np.mean(val_metrics.jaccard[-report_epoch:]),
)
write_event(fh, **report_val_metrics)
if time.time() - st_time > training_timelimit:
tq.close()
break
if best_score < report_val_metrics.val_jaccard:
best_score = report_val_metrics.val_jaccard
not_improved_count = 0
copy_best(model, epoch, model_name, step)
else:
not_improved_count += 1
if not_improved_count >= patience:
# Update learning rate and optimizer
lr *= lr_update_rate
# Stop criterion
if lr < min_lr:
tq.close()
break
not_improved_count = 0
# Load best weight
del model
model = unet_vgg16(pretrained=False)
path = f'/root/working/models/{model_name}/{model_name}_best'
cp = torch.load(path)
model = nn.DataParallel(model).cuda()
epoch = cp['epoch']
model.load_state_dict(cp['model'])
model = model.module
model = nn.DataParallel(model, device_ids=gpus).cuda()
# Init optimizer
optimizer = Adam(model.parameters(), lr=lr)
except KeyboardInterrupt:
save(model, epoch, step, model_name)
tq.close()
fh.close()
sys.exit(1)
except Exception as e:
raise e
break
fh.close()
RandomGamma,
)
transformimg = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406][::-1], std=[0.225, 0.224, 0.225][::-1]
),
]
)
transformaug = Compose(
[
VerticalFlip(p=0.5),
RandomRotate90(p=0.5),
ISONoise(p=0.5),
RandomBrightnessContrast(p=0.5),
RandomGamma(p=0.5),
RandomFog(fog_coef_lower=0.025, fog_coef_upper=0.1, p=0.5),
]
)
class XViewDataset(Dataset):
def __init__(
self, size=None, aug=True, pattern="data/train/images1024/*pre_disaster*.png"
):
self.name = "train"
self.aug = aug
self.pre = glob(pattern)
def augment_data(images, masks, save_path, augment=True):
"""
Performing data augmentation.
"""
crop_size = (256, 256)
size = (2018, 2006)
# 将数据与标签组合
for image, mask in tqdm(zip(images, masks), total=len(images)):
image_name = image.split("/")[-1].split(".")[0]
mask_name = mask.split("/")[-1].split(".")[0]
x, y = read_data(image, mask)
# try except 使用
try:
h, w, c = x.shape # 获取图像的 w h z
except Exception as e:
image = image[:-1]
x, y = read_data(image, mask)
h, w, c = x.shape
# 进行数据增强
if augment == True:
# Center Crop
aug = CenterCrop(p=1, height=crop_size[1], width=crop_size[0])
augmented = aug(image=x, mask=y)
x1 = augmented['image']
y1 = augmented['mask']
# Crop
x_min = 0
y_min = 0
x_max = x_min + crop_size[0]
y_max = y_min + crop_size[1]
aug = Crop(p=1, x_min=x_min, x_max=x_max, y_min=y_min, y_max=y_max)
augmented = aug(image=x, mask=y)
x2 = augmented['image']
y2 = augmented['mask']
# Random Rotate 90 degree
aug = RandomRotate90(p=1)
augmented = aug(image=x, mask=y)
x3 = augmented['image']
y3 = augmented['mask']
# Transpose
aug = Transpose(p=1)
augmented = aug(image=x, mask=y)
x4 = augmented['image']
y4 = augmented['mask']
# ElasticTransform
aug = ElasticTransform(p=1, alpha=120, sigma=120 * 0.05, alpha_affine=120 * 0.03)
augmented = aug(image=x, mask=y)
x5 = augmented['image']
y5 = augmented['mask']
# Grid Distortion
aug = GridDistortion(p=1)
augmented = aug(image=x, mask=y)
x6 = augmented['image']
y6 = augmented['mask']
# Optical Distortion
aug = OpticalDistortion(p=1, distort_limit=2, shift_limit=0.5)
augmented = aug(image=x, mask=y)
x7 = augmented['image']
y7 = augmented['mask']
# Vertical Flip
aug = VerticalFlip(p=1)
augmented = aug(image=x, mask=y)
x8 = augmented['image']
y8 = augmented['mask']
# Horizontal Flip
aug = HorizontalFlip(p=1)
augmented = aug(image=x, mask=y)
x9 = augmented['image']
y9 = augmented['mask']
# Grayscale
x10 = cv2.cvtColor(x, cv2.COLOR_RGB2GRAY)
y10 = y
# Grayscale Vertical Flip
aug = VerticalFlip(p=1)
augmented = aug(image=x10, mask=y10)
x11 = augmented['image']
y11 = augmented['mask']
# Grayscale Horizontal Flip
aug = HorizontalFlip(p=1)
augmented = aug(image=x10, mask=y10)
x12 = augmented['image']
y12 = augmented['mask']
# Grayscale Center Crop
aug = CenterCrop(p=1, height=crop_size[1], width=crop_size[0])
augmented = aug(image=x10, mask=y10)
x13 = augmented['image']
y13 = augmented['mask']
# Random Brightness Contrast
aug = RandomBrightnessContrast(p=1)
augmented = aug(image=x, mask=y)
x14 = augmented['image']
y14 = augmented['mask']
# Random Gamma
aug = RandomGamma(p=1)
augmented = aug(image=x, mask=y)
x15 = augmented['image']
y15 = augmented['mask']
aug = HueSaturationValue(p=1)
augmented = aug(image=x, mask=y)
x16 = augmented['image']
y16 = augmented['mask']
aug = RGBShift(p=1)
augmented = aug(image=x, mask=y)
x17 = augmented['image']
y17 = augmented['mask']
aug = RandomBrightness(p=1)
augmented = aug(image=x, mask=y)
x18 = augmented['image']
y18 = augmented['mask']
aug = RandomContrast(p=1)
augmented = aug(image=x, mask=y)
x19 = augmented['image']
y19 = augmented['mask']
aug = MotionBlur(p=1, blur_limit=7)
augmented = aug(image=x, mask=y)
x20 = augmented['image']
y20 = augmented['mask']
aug = MedianBlur(p=1, blur_limit=10)
augmented = aug(image=x, mask=y)
x21 = augmented['image']
y21 = augmented['mask']
aug = GaussianBlur(p=1, blur_limit=10)
augmented = aug(image=x, mask=y)
x22 = augmented['image']
y22 = augmented['mask']
aug = GaussNoise(p=1)
augmented = aug(image=x, mask=y)
x23 = augmented['image']
y23 = augmented['mask']
aug = ChannelShuffle(p=1)
augmented = aug(image=x, mask=y)
x24 = augmented['image']
y24 = augmented['mask']
aug = CoarseDropout(p=1, max_holes=8, max_height=32, max_width=32)
augmented = aug(image=x, mask=y)
x25 = augmented['image']
y25 = augmented['mask']
images = [
x, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10,
x11, x12, x13, x14, x15, x16, x17, x18, x19, x20,
x21, x22, x23, x24, x25
]
masks = [
y, y1, y2, y3, y4, y5, y6, y7, y8, y9, y10,
y11, y12, y13, y14, y15, y16, y17, y18, y19, y20,
y21, y22, y23, y24, y25
]
else:
images = [x]
masks = [y]
idx = 0
# 数据增强之后数据
for i, m in zip(images, masks):
i = cv2.resize(i, size)
m = cv2.resize(m, size)
tmp_image_name = f"{image_name}_{idx}.jpg"
tmp_mask_name = f"{mask_name}_{idx}.jpg"
image_path = os.path.join(save_path, "image/", tmp_image_name)
mask_path = os.path.join(save_path, "mask/", tmp_mask_name)
# 保存数据
cv2.imwrite(image_path, i)
cv2.imwrite(mask_path, m)
idx += 1
from albumentations import (RandomRotate90, Transpose, Flip, Compose, Resize,
Normalize)
from albumentations.pytorch import ToTensor
SIZE = 320
p = 0.4
train_aug = Compose([
RandomRotate90(),
Flip(),
Transpose(),
Resize(width=SIZE, height=SIZE, always_apply=True),
ToTensor(normalize={
'mean': [0.485, 0.456, 0.406],
'std': [0.229, 0.224, 0.225]
})
],
p=p)
valid_aug = Compose([
Resize(width=SIZE, height=SIZE, always_apply=True),
ToTensor(normalize={
'mean': [0.485, 0.456, 0.406],
'std': [0.229, 0.224, 0.225]
}),
],
p=1.0)
def get_data(img_size, batch_size):
CSV_PATH = DATA_PATH / 'train_v2.csv'
IMG_FOLDER = DATA_PATH / 'train-jpg'
EXT = 'jpg'
SZ = img_size
BS = batch_size
MEAN, STD = np.array([0.485, 0.456,
0.406]), np.array([0.229, 0.224, 0.225])
# torch transforms
# transform = transforms.Compose([
# transforms.Resize(SZ),
# transforms.ToTensor(),
# transforms.Normalize(MEAN, STD)
# ])
transform = {
'train':
Compose([
Resize(height=SZ, width=SZ),
CLAHE(clip_limit=1.0, p=0.25),
Flip(p=0.5),
RandomRotate90(p=0.5),
RandomBrightnessContrast(brightness_limit=0.3,
contrast_limit=0.3,
p=0.5),
ShiftScaleRotate(shift_limit=0.1, scale_limit=0.1, rotate_limit=0),
RGBShift(p=0.25),
Normalize(mean=MEAN, std=STD),
ToTensor()
]),
'val':
Compose([
Resize(height=SZ, width=SZ),
Flip(p=0.5),
RandomRotate90(p=0.5),
Normalize(mean=MEAN, std=STD),
ToTensor()
])
}
train_ds = PlanetDataset(CSV_PATH / 'train.csv', IMG_FOLDER, EXT,
transform['train'])
val_ds = PlanetDataset(CSV_PATH / 'val.csv',
IMG_FOLDER,
EXT,
transform['val'],
val=True)
train_dl = DataLoader(train_ds,
batch_size=BS,
shuffle=True,
num_workers=4,
pin_memory=True,
drop_last=True)
val_dl = DataLoader(val_ds,
batch_size=BS * 2,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=True)
# Show the details in the console
print(f'''Train DS: {train_ds.img_folder} \t \
Ext: {train_ds.ext} \t \
x_train: {train_ds.x_train.shape} \t \
y_train: {train_ds.y_train.shape} \t''')
print(f'''Validation DS: {val_ds.img_folder} \t \
Ext: {val_ds.ext} \t \
x_train: {val_ds.x_train.shape} \t \
y_train: {val_ds.y_train.shape} \t''')
return (train_dl, val_dl)
def augment_data(images, masks, save_path, augment=True):
""" Performing data augmentation. """
size = (512, 512)
crop_size = (448, 448)
for idx, (x, y) in tqdm(enumerate(zip(images, masks)), total=len(images)):
image_name = x.split("/")[-1].split(".")[0]
mask_name = y.split("/")[-1].split(".")[0]
x = cv2.imread(x, cv2.IMREAD_COLOR)
y = cv2.imread(y, cv2.IMREAD_COLOR)
if x.shape[0] >= size[0] and x.shape[1] >= size[1]:
if augment == True:
## Crop
x_min = 0
y_min = 0
x_max = x_min + size[0]
y_max = y_min + size[1]
aug = Crop(p=1, x_min=x_min, x_max=x_max, y_min=y_min, y_max=y_max)
augmented = aug(image=x, mask=y)
x1 = augmented['image']
y1 = augmented['mask']
# Random Rotate 90 degree
aug = RandomRotate90(p=1)
augmented = aug(image=x, mask=y)
x2 = augmented['image']
y2 = augmented['mask']
## ElasticTransform
aug = ElasticTransform(p=1, alpha=120, sigma=120 * 0.05, alpha_affine=120 * 0.03)
augmented = aug(image=x, mask=y)
x3 = augmented['image']
y3 = augmented['mask']
## Grid Distortion
aug = GridDistortion(p=1)
augmented = aug(image=x, mask=y)
x4 = augmented['image']
y4 = augmented['mask']
## Optical Distortion
aug = OpticalDistortion(p=1, distort_limit=2, shift_limit=0.5)
augmented = aug(image=x, mask=y)
x5 = augmented['image']
y5 = augmented['mask']
## Vertical Flip
aug = VerticalFlip(p=1)
augmented = aug(image=x, mask=y)
x6 = augmented['image']
y6 = augmented['mask']
## Horizontal Flip
aug = HorizontalFlip(p=1)
augmented = aug(image=x, mask=y)
x7 = augmented['image']
y7 = augmented['mask']
## Grayscale
x8 = cv2.cvtColor(x, cv2.COLOR_RGB2GRAY)
y8 = y
## Grayscale Vertical Flip
aug = VerticalFlip(p=1)
augmented = aug(image=x8, mask=y8)
x9 = augmented['image']
y9 = augmented['mask']
## Grayscale Horizontal Flip
aug = HorizontalFlip(p=1)
augmented = aug(image=x8, mask=y8)
x10 = augmented['image']
y10 = augmented['mask']
# aug = RandomBrightnessContrast(p=1)
# augmented = aug(image=x, mask=y)
# x11 = augmented['image']
# y11 = augmented['mask']
#
# aug = RandomGamma(p=1)
# augmented = aug(image=x, mask=y)
# x12 = augmented['image']
# y12 = augmented['mask']
#
# aug = HueSaturationValue(p=1)
# augmented = aug(image=x, mask=y)
# x13 = augmented['image']
# y13 = augmented['mask']
aug = RGBShift(p=1)
augmented = aug(image=x, mask=y)
x14 = augmented['image']
y14 = augmented['mask']
# aug = RandomBrightness(p=1)
# augmented = aug(image=x, mask=y)
# x15 = augmented['image']
# y15 = augmented['mask']
#
# aug = RandomContrast(p=1)
# augmented = aug(image=x, mask=y)
# x16 = augmented['image']
# y16 = augmented['mask']
aug = ChannelShuffle(p=1)
augmented = aug(image=x, mask=y)
x17 = augmented['image']
y17 = augmented['mask']
aug = CoarseDropout(p=1, max_holes=10, max_height=32, max_width=32)
augmented = aug(image=x, mask=y)
x18 = augmented['image']
y18 = augmented['mask']
aug = GaussNoise(p=1)
augmented = aug(image=x, mask=y)
x19 = augmented['image']
y19 = augmented['mask']
# aug = MotionBlur(p=1, blur_limit=7)
# augmented = aug(image=x, mask=y)
# x20 = augmented['image']
# y20 = augmented['mask']
#
# aug = MedianBlur(p=1, blur_limit=11)
# augmented = aug(image=x, mask=y)
# x21 = augmented['image']
# y21 = augmented['mask']
#
# aug = GaussianBlur(p=1, blur_limit=11)
# augmented = aug(image=x, mask=y)
# x22 = augmented['image']
# y22 = augmented['mask']
##
aug = CenterCrop(256, 256, p=1)
augmented = aug(image=x, mask=y)
x23 = augmented['image']
y23 = augmented['mask']
aug = CenterCrop(384, 384, p=1)
augmented = aug(image=x, mask=y)
x24 = augmented['image']
y24 = augmented['mask']
aug = CenterCrop(448, 448, p=1)
augmented = aug(image=x, mask=y)
x25 = augmented['image']
y25 = augmented['mask']
## x23 Vertical Flip
aug = VerticalFlip(p=1)
augmented = aug(image=x23, mask=y23)
x26 = augmented['image']
y26 = augmented['mask']
## x23 Horizontal Flip
aug = HorizontalFlip(p=1)
augmented = aug(image=x23, mask=y23)
x27 = augmented['image']
y27 = augmented['mask']
## x24 Vertical Flip
aug = VerticalFlip(p=1)
augmented = aug(image=x24, mask=y24)
x28 = augmented['image']
y28 = augmented['mask']
## x24 Horizontal Flip
aug = HorizontalFlip(p=1)
augmented = aug(image=x24, mask=y24)
x29 = augmented['image']
y29 = augmented['mask']
## x25 Vertical Flip
aug = VerticalFlip(p=1)
augmented = aug(image=x25, mask=y25)
x30 = augmented['image']
y30 = augmented['mask']
## x25 Horizontal Flip
aug = HorizontalFlip(p=1)
augmented = aug(image=x25, mask=y25)
x31 = augmented['image']
y31 = augmented['mask']
images = [
x, x1, x2, x3, x4, x5, x6, x7, x8, x9,
x10,
# x11, x12, x13,
x14,
# x15, x16,
x17, x18, x19,
# x20, x21, x22,
x23, x24, x25, x26, x27, x28, x29,
x30, x31
]
masks = [
y, y1, y2, y3, y4, y5, y6, y7, y8, y9,
y10,
# y11, y12, y13,
y14,
# y15, y16,
y17, y18, y19,
# y20, y21, y22,
y23, y24, y25, y26, y27, y28, y29,
y30, y31
]
else:
images = [x]
masks = [y]
idx = 0
for i, m in zip(images, masks):
i = cv2.resize(i, size)
m = cv2.resize(m, size)
if len(images) == 1:
tmp_image_name = f"{image_name}.jpg"
tmp_mask_name = f"{mask_name}.jpg"
else:
tmp_image_name = f"{image_name}_{idx}.jpg"
tmp_mask_name = f"{mask_name}_{idx}.jpg"
image_path = os.path.join(save_path, "image/", tmp_image_name)
mask_path = os.path.join(save_path, "mask/", tmp_mask_name)
cv2.imwrite(image_path, i)
cv2.imwrite(mask_path, m)
idx += 1
def train(file_pattern,
train_num_batches=None,
train_aug=False,
train_batch_size=1,
val_batch_size=1,
learning_rate=1e-3,
epochs=1,
verbosity=2,
file_directory=None,
resume=None,
train_shuffle=True,
pre_image_mean=None,
post_image_mean=None):
"""
Function to train the UNet model
Parameters
----------
file_pattern : string
Location where the image folder is for the data. Example format:
"images/*pre_disaster*.png"
train_num_batches : int
Number of batches for the training set, if none, the full dataset will
be used.
train_aug : bool
If true, augmentations are performed.
train_batch_size : int, default 5
Batch size for the training set.
val_batch_size : int, default 5
Batch size for the validation set.
learning_rate : float, default 0.00001
Learning rate for the UNet.
epochs : int, default 1
How many epochs for the training to run.
verbosity : int, default 2
How verbose you'd like the output to be.
file_directory : string, default None:
Directory where you'd like the output files saved.
resume : string, default None
Enter in a string for the saved model file and training will resume
from this instance.
train_shuffle : bool
If True, the training data is shuffled for each epoch.
pre_image_mean : str
The filepath for the pre image mean numpy array file.
post_image_mean : str
The filepath for the post image mean numpy array file.
Returns
-------
Saves the model weights, csv logs, and tensorboard files in the original
directories specified.
"""
if file_directory is None:
file_directory = os.path.abspath(
os.path.join(os.getcwd(), "saved_models"))
tensorboard_path = os.path.join(
file_directory, "logs",
"tboard_{}".format(datetime.datetime.now().strftime("%Y%m%d-%H%M")))
weights_path = os.path.join(
file_directory, "unet_weights_{}".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M")))
csv_logger_path = os.path.join(
file_directory,
"log_unet_{}{}".format(datetime.datetime.now().strftime("%Y%m%d-%H%M"),
".csv"))
if train_aug:
train_augs = Compose([
VerticalFlip(p=0.5),
RandomRotate90(p=0.5),
ISONoise(p=0.5),
RandomBrightnessContrast(p=0.5),
RandomGamma(p=0.5),
RandomFog(fog_coef_lower=0.025, fog_coef_upper=0.1, p=0.5),
])
else:
train_augs = None
# Weighted categorical cross entropy weights
# class_weights = tf.constant([0.1, 1.0, 2.0, 2.0, 2.0])
# class_weights = tf.constant([1.0, 1.0, 0.5, 0.5, 0.5])
class_weights = tf.constant([1.0, 1.0, 3.0, 3.0, 3.0])
train_data = LabeledImageDataset(num_batches=train_num_batches,
augmentations=train_augs,
pattern=file_pattern,
shuffle=train_shuffle,
n_classes=5,
batch_size=train_batch_size,
normalize=True)
# Using random samples from train for validation
val_data = LabeledImageDataset(num_batches=100,
augmentations=train_augs,
pattern=file_pattern,
shuffle=train_shuffle,
n_classes=5,
batch_size=val_batch_size,
normalize=True)
if resume:
try:
print("the pretrained model was loaded")
model = UNet(num_classes=5).model((None, None, 3))
model.load_weights(resume)
except OSError:
print("The model file could not be found. "
"Starting from a new model instance")
model = UNet(num_classes=5).model((None, None, 3))
else:
model = UNet(num_classes=5).model((None, None, 3))
metrics = [tf.keras.metrics.CategoricalAccuracy()]
for i in range(5):
metrics.append(Precision(class_id=i, name=f"prec_class_{i}"))
metrics.append(Recall(class_id=i, name=f"rec_class_{i}"))
model.compile(optimizer=keras.optimizers.RMSprop(lr=learning_rate),
loss=CombinedLoss(class_weights),
metrics=metrics)
# Creating a checkpoint to save the model after every epoch if the
# validation loss has decreased
model_checkpoint = ModelCheckpoint("dual_unet_{epoch:02d}-{loss:.2f}.hdf5",
monitor='loss',
save_best_only=False,
mode='min',
save_weights_only=True,
verbose=verbosity)
csv_logger = CSVLogger(csv_logger_path, append=True, separator=',')
lr_logger = ReduceLROnPlateau(monitor='loss',
factor=0.2,
patience=1,
verbose=verbosity,
mode='min',
min_lr=1e-6)
tensorboard_cb = TensorBoard(log_dir=tensorboard_path, write_images=True)
try:
model.fit(train_data,
epochs=epochs,
verbose=verbosity,
callbacks=[
LossAndErrorPrintingCallback(), model_checkpoint,
csv_logger, lr_logger, tensorboard_cb
],
validation_data=val_data,
workers=6)
except KeyboardInterrupt:
save_model(model, pause=1)
sys.exit()
except Exception as exc:
save_model(model, pause=0)
raise exc