File size: 9,986 Bytes
1dc26c7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 |
import torch, random
import torchvision.transforms.functional as F
import torchvision.transforms as tf
import numpy as np
from PIL import Image
from typing import Tuple, List, Callable
class Compose:
def __init__(self,
transforms: List[Callable],
student_augs: bool = False):
self.transforms = transforms
self.student_augs = student_augs
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[torch.Tensor, torch.Tensor]:
for transform in self.transforms:
if pseudo is None:
img, gt = transform(img, gt)
else:
img, gt, pseudo = transform(img, gt, pseudo)
if self.student_augs:
aimg = img.clone()
aimg, _ = RandGaussianBlur()(aimg, gt)
if 0.5 > random.random():
aimg, _ = ColorJitter()(aimg, gt)
else:
aimg, _ = MaskGrayscale()(aimg, gt)
if pseudo is None and not self.student_augs:
return img, gt
elif pseudo is None and self.student_augs:
return img, gt, aimg
elif pseudo is not None and not self.student_augs:
return img, gt, pseudo
else:
return img, gt, aimg, pseudo
class ToTensor:
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[torch.Tensor, torch.Tensor]:
img = F.to_tensor(np.array(img))
gt = torch.from_numpy(np.array(gt)).unsqueeze(0)
if pseudo is not None:
pseudo = torch.from_numpy(np.array(pseudo)).unsqueeze(0)
if pseudo is None:
return img, gt
else:
return img, gt, pseudo
class Resize:
def __init__(self,
resize: Tuple[int]):
self.img_resize = tf.Resize(size=resize,
interpolation=tf.InterpolationMode.BILINEAR)
self.gt_resize = tf.Resize(size=resize,
interpolation=tf.InterpolationMode.NEAREST)
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[Image.Image, Image.Image]:
img = self.img_resize(img)
gt = self.gt_resize(gt)
if pseudo is None:
return img, gt
else:
return img, gt, self.gt_resize(pseudo)
class ImgResize:
def __init__(self,
resize: Tuple[int, int]):
self.resize = resize
self.num_pixels = self.resize[0]*self.resize[1]
def __call__(self,
img: torch.Tensor,
gt: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
if torch.prod(torch.tensor(img.shape[-2:])) > self.num_pixels:
img = torch.nn.functional.interpolate(img.unsqueeze(0), size=self.resize, mode='bilinear').squeeze(0)
return img, gt
class ImgResizePIL:
def __init__(self,
resize: Tuple[int]):
self.resize = resize
self.num_pixels = self.resize[0]*self.resize[1]
def __call__(self,
img: Image) -> Image:
if img.height*img.width > self.num_pixels:
img = img.resize((self.resize[1], self.resize[0]), tf.InterpolationMode.BILINEAR)
return img
class Normalize:
def __init__(self,
mean: List[float] = [0.485, 0.456, 0.406],
std: List[float] = [0.229, 0.224, 0.225]):
self.norm = tf.Normalize(mean=mean,
std=std)
def __call__(self,
img: torch.Tensor,
gt: torch.Tensor,
pseudo = None) -> Tuple[torch.Tensor, torch.Tensor]:
img = self.norm(img)
if pseudo is None:
return img, gt
else:
return img, gt, pseudo
class UnNormalize(object):
def __init__(self,
mean: List[float] = [0.485, 0.456, 0.406],
std: List[float] = [0.229, 0.224, 0.225]):
self.mean = mean
self.std = std
def __call__(self, image):
image2 = torch.clone(image)
for t, m, s in zip(image2, self.mean, self.std):
t.mul_(s).add_(m)
return image2
class RandomHFlip:
def __init__(self,
percentage: float = 0.5):
self.percentage = percentage
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[Image.Image, Image.Image]:
if random.random() < self.percentage:
img = F.hflip(img)
gt = F.hflip(gt)
if pseudo is not None:
pseudo = F.hflip(pseudo)
if pseudo is None:
return img, gt
else:
return img, gt, pseudo
class RandomResizedCrop:
def __init__(self,
crop_size: List[int],
crop_scale: List[float],
crop_ratio: List[float]):
print('RandomResizedCrop ratio modified!!!')
self.crop_scale = tuple(crop_scale)
self.crop_ratio = tuple(crop_ratio)
self.crop = tf.RandomResizedCrop(size=tuple(crop_size),
scale=self.crop_scale,
ratio=self.crop_ratio,)
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[Image.Image, Image.Image]:
i, j, h, w = self.crop.get_params(img=img,
scale=self.crop.scale,
ratio=self.crop.ratio)
img = F.resized_crop(img, i, j, h, w, self.crop.size, tf.InterpolationMode.BILINEAR)
gt = F.resized_crop(gt, i, j, h, w, self.crop.size, tf.InterpolationMode.NEAREST)
if pseudo is not None:
pseudo = F.resized_crop(pseudo, i, j, h, w, self.crop.size, tf.InterpolationMode.NEAREST)
if pseudo is None:
return img, gt
else:
return img, gt, pseudo
class CenterCrop:
def __init__(self,
crop_size: int):
self.crop = tf.CenterCrop(size=crop_size)
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[Image.Image, Image.Image]:
img = self.crop(img)
gt = self.crop(gt)
if pseudo is None:
return img, gt
else:
return img, gt, self.crop(pseudo)
class PyramidCenterCrop:
def __init__(self,
crop_size: List[int],
scales: List[float]):
self.crop_size = crop_size
self.scales = scales
self.crop = tf.CenterCrop(size=crop_size)
def __call__(self,
img: Image.Image,
gt: Image.Image) -> Tuple[Image.Image, Image.Image]:
imgs = []
gts = []
for s in self.scales:
new_size = (int(self.crop_size*1/s), int(self.crop_size*1/s*(img.shape[2]/img.shape[1])))
img = tf.Resize(size=new_size, interpolation=tf.InterpolationMode.BILINEAR)(img)
gt = tf.Resize(size=new_size, interpolation=tf.InterpolationMode.NEAREST)(gt)
imgs.append(self.crop(img))
gts.append(self.crop(gt))
return torch.stack(imgs), torch.stack(gts)
class IdsToTrainIds:
def __init__(self,
source: str):
self.source = source
self.first_nonvoid = 7
def __call__(self,
img: torch.Tensor,
gt: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
if self.source == 'cityscapes':
gt = gt.to(dtype=torch.int64) - self.first_nonvoid
gt[gt>26] = 255
gt[gt<0] = 255
elif self.source == 'cocostuff':
gt = gt.to(dtype=torch.int64)
elif self.source == 'potsdam':
gt = gt.to(dtype=torch.int64)
return img, gt
class ColorJitter:
def __init__(self, percentage: float = 0.3, brightness: float = 0.1,
contrast: float = 0.1, saturation: float = 0.1, hue: float = 0.1):
self.percentage = percentage
self.jitter = tf.ColorJitter(brightness=brightness,
contrast=contrast,
saturation=saturation,
hue=hue)
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[Image.Image, Image.Image]:
if random.random() < self.percentage:
img = self.jitter(img)
if pseudo is None:
return img, gt
else:
return img, gt, pseudo
class MaskGrayscale:
def __init__(self, percentage: float = 0.1):
self.percentage = percentage
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[Image.Image, Image.Image]:
if self.percentage > random.random():
img = tf.Grayscale(num_output_channels=3)(img)
if pseudo is None:
return img, gt
else:
return img, gt, pseudo
class RandGaussianBlur:
def __init__(self, radius: List[float] = [.1, 2.]):
self.radius = radius
def __call__(self,
img: Image.Image,
gt: Image.Image,
pseudo = None) -> Tuple[Image.Image, Image.Image]:
radius = random.uniform(self.radius[0], self.radius[1])
img = tf.GaussianBlur(kernel_size=21, sigma=radius)(img)
if pseudo is None:
return img, gt
else:
return img, gt, pseudo
|