---恢复内容开始---
1.首先对给的数据进行划分，类型为每个类单独放在一个文件夹中
1 import json 2 import shutil 3 import os 4 from glob import glob 5 from tqdm import tqdm 6 # 此文件的作用是创建每个类的文件夹，以及根据给出来的Json中已经做好的分类，对数据进行对号入座划分。 7 # 加载json文件得出一个字典，然后根据Key值来提取每个文件到相应的文件夹中，（注意去除了不合理数据） 8 9 try: 10 for i in range(0,59): 11 os.mkdir("./data/train/" + str(i)) 12 except: 13 pass 14 15 file_train = json.load(open("./data/temp/labels/AgriculturalDisease_train_annotations.json","r",encoding="utf-8")) 16 file_val = json.load(open("./data/temp/labels/AgriculturalDisease_validation_annotations.json","r",encoding="utf-8")) 17 18 file_list = file_train + file_val 19 20 for file in tqdm(file_list): 21 filename = file["image_id"] 22 origin_path = "./data/temp/images/" + filename 23 ids = file["disease_class"] 24 if ids == 44: 25 continue 26 if ids == 45: 27 continue 28 if ids > 45: 29 ids = ids -2 30 save_path = "./data/train/" + str(ids) + "/" 31 shutil.copy(origin_path,save_path)
2.加载数据类的定义
1.采用自定义数据加载类，此Dataset类中重新定义了对数据进行数据增强的多种方式，不仅限于pytorch中自带的增强方式。
首先附上自定义的数据增强的函数代码：
方式一，以重新定义重载方法类的方式定义多种增强方式，在dataset类中的get_item方法中的compose中加入自定义的方法，即可调用。
1 # 数据增强的多种方式，使用自定义的方法。调用只需在dataloader.py文件中的get_item函数中调用类自身参数 2 # transforms，transforms中集合了compose，compose中列出详细所使用的增强方式。 3 from __future__ import division 4 import cv2 5 import numpy as np 6 from numpy import random 7 import math 8 from sklearn.utils import shuffle 9 # 常用的增强方式几乎都在这里，只需在compose中列出类名即可 10 __all__ = ['Compose','RandomHflip', 'RandomUpperCrop', 'Resize', 'UpperCrop', 'RandomBottomCrop', 11 "RandomErasing",'BottomCrop', 'Normalize', 'RandomSwapChannels', 'RandomRotate', 12 'RandomHShift',"CenterCrop","RandomVflip",'ExpandBorder', 'RandomResizedCrop', 13 'RandomDownCrop', 'DownCrop', 'ResizedCrop',"FixRandomRotate"] 14 # 组合 15 # “随机翻转”，“随机顶部切割”，“调整大小”，“上切割”，“随机底部切割”、 16 # “随机擦除”，“底部切割”，“正则化”，“随机交换频道”，“随机旋转”， 17 # “随机HShift”，“中央切割”，“随机Vflip”，“扩展边界”，“随机调整切割”， 18 # “随机下降”，“下降切割”， “调整切割”，“固定随机化”。 19 20 21 # 每个增强方式类需要调用普通方法描述如下： 22 def rotate_nobound(image, angle, center=None, scale=1.): 23 (h, w) = image.shape[:2] 24 25 26 # if the center is None, initialize it as the center of 27 # the image 28 if center is None: 29 center = (w // 2, h // 2) 30 31 # perform the rotation 32 M = cv2.getRotationMatrix2D(center, angle, scale) 33 rotated = cv2.warpAffine(image, M, (w, h)) 34 35 return rotated 36 37 def scale_down(src_size, size): 38 w, h = size 39 sw, sh = src_size 40 if sh < h: 41 w, h = float(w * sh) / h, sh 42 if sw < w: 43 w, h = sw, float(h * sw) / w 44 return int(w), int(h) 45 46 47 def fixed_crop(src, x0, y0, w, h, size=None): 48 out = src[y0:y0 + h, x0:x0 + w] 49 if size is not None and (w, h) != size: 50 out = cv2.resize(out, (size[0], size[1]), interpolation=cv2.INTER_CUBIC) 51 return out 52 53 # 固定随机旋转 54 class FixRandomRotate(object): 55 def __init__(self, angles=[0,90,180,270], bound=False): 56 self.angles = angles 57 self.bound = bound 58 59 def __call__(self,img): 60 do_rotate = random.randint(0, 4) 61 angle=self.angles[do_rotate] 62 if self.bound: 63 img = rotate_bound(img, angle) 64 else: 65 img = rotate_nobound(img, angle) 66 return img 67 68 def center_crop(src, size): 69 h, w = src.shape[0:2] 70 new_w, new_h = scale_down((w, h), size) 71 72 x0 = int((w - new_w) / 2) 73 y0 = int((h - new_h) / 2) 74 75 out = fixed_crop(src, x0, y0, new_w, new_h, size) 76 return out 77 78 79 def bottom_crop(src, size): 80 h, w = src.shape[0:2] 81 new_w, new_h = scale_down((w, h), size) 82 83 x0 = int((w - new_w) / 2) 84 y0 = int((h - new_h) * 0.75) 85 86 out = fixed_crop(src, x0, y0, new_w, new_h, size) 87 return out 88 89 def rotate_bound(image, angle): 90 # grab the dimensions of the image and then determine the 91 # center 92 h, w = image.shape[:2] 93 94 (cX, cY) = (w // 2, h // 2) 95 96 M = cv2.getRotationMatrix2D((cX, cY), angle, 1.0) 97 cos = np.abs(M[0, 0]) 98 sin = np.abs(M[0, 1]) 99 100 # compute the new bounding dimensions of the image 101 nW = int((h * sin) + (w * cos)) 102 nH = int((h * cos) + (w * sin)) 103 104 # adjust the rotation matrix to take into account translation 105 M[0, 2] += (nW / 2) - cX 106 M[1, 2] += (nH / 2) - cY 107 108 rotated = cv2.warpAffine(image, M, (nW, nH)) 109 110 return rotated 111 112 113 114 # 常用增强方式，以类的方式体现： 115 # 将多个transform组合起来使用 116 crop切割 filp旋转 117 class Compose(object): 118 def __init__(self, transforms): 119 self.transforms = transforms 120 def __call__(self, img): 121 for t in self.transforms: 122 img = t(img) 123 return img 124 class RandomRotate(object): 125 def __init__(self, angles, bound=False): 126 self.angles = angles 127 self.bound = bound 128 129 def __call__(self,img): 130 do_rotate = random.randint(0, 2) 131 if do_rotate: 132 angle = np.random.uniform(self.angles[0], self.angles[1]) 133 if self.bound: 134 img = rotate_bound(img, angle) 135 else: 136 img = rotate_nobound(img, angle) 137 return img 138 class RandomBrightness(object): 139 def __init__(self, delta=10): 140 assert delta >= 0 141 assert delta <= 255 142 self.delta = delta 143 144 def __call__(self, image): 145 if random.randint(2): 146 delta = random.uniform(-self.delta, self.delta) 147 image = (image + delta).clip(0.0, 255.0) 148 # print('RandomBrightness,delta ',delta) 149 return image 150 151 152 class RandomContrast(object): 153 def __init__(self, lower=0.9, upper=1.05): 154 self.lower = lower 155 self.upper = upper 156 assert self.upper >= self.lower, "contrast upper must be >= lower." 157 assert self.lower >= 0, "contrast lower must be non-negative." 158 159 # expects float image 160 def __call__(self, image): 161 if random.randint(2): 162 alpha = random.uniform(self.lower, self.upper) 163 # print('contrast:', alpha) 164 image = (image * alpha).clip(0.0,255.0) 165 return image 166 167 168 class RandomSaturation(object): 169 def __init__(self, lower=0.8, upper=1.2): 170 self.lower = lower 171 self.upper = upper 172 assert self.upper >= self.lower, "contrast upper must be >= lower." 173 assert self.lower >= 0, "contrast lower must be non-negative." 174 175 def __call__(self, image): 176 if random.randint(2): 177 alpha = random.uniform(self.lower, self.upper) 178 image[:, :, 1] *= alpha 179 # print('RandomSaturation,alpha',alpha) 180 return image 181 182 183 class RandomHue(object): 184 def __init__(self, delta=18.0): 185 assert delta >= 0.0 and delta <= 360.0 186 self.delta = delta 187 188 def __call__(self, image): 189 if random.randint(2): 190 alpha = random.uniform(-self.delta, self.delta) 191 image[:, :, 0] += alpha 192 image[:, :, 0][image[:, :, 0] > 360.0] -= 360.0 193 image[:, :, 0][image[:, :, 0] < 0.0] += 360.0 194 # print('RandomHue,alpha:', alpha) 195 return image 196 197 198 class ConvertColor(object): 199 def __init__(self, current='BGR', transform='HSV'): 200 self.transform = transform 201 self.current = current 202 203 def __call__(self, image): 204 if self.current == 'BGR' and self.transform == 'HSV': 205 image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV) 206 elif self.current == 'HSV' and self.transform == 'BGR': 207 image = cv2.cvtColor(image, cv2.COLOR_HSV2BGR) 208 else: 209 raise NotImplementedError 210 return image 211 212 class RandomSwapChannels(object): 213 def __call__(self, img): 214 if np.random.randint(2): 215 order = np.random.permutation(3) 216 return img[:,:,order] 217 return img 218 219 class RandomCrop(object): 220 def __init__(self, size): 221 self.size = size 222 def __call__(self, image): 223 h, w, _ = image.shape 224 new_w, new_h = scale_down((w, h), self.size) 225 226 if w == new_w: 227 x0 = 0 228 else: 229 x0 = random.randint(0, w - new_w) 230 231 if h == new_h: 232 y0 = 0 233 else: 234 y0 = random.randint(0, h - new_h) 235 236 out = fixed_crop(image, x0, y0, new_w, new_h, self.size) 237 return out 238 239 240 241 class RandomResizedCrop(object): 242 def __init__(self, size,scale=(0.49, 1.0), ratio=(1., 1.)): 243 self.size = size 244 self.scale = scale 245 self.ratio = ratio 246 247 def __call__(self,img): 248 if random.random() < 0.2: 249 return cv2.resize(img,self.size) 250 h, w, _ = img.shape 251 area = h * w 252 d=1 253 for attempt in range(10): 254 target_area = random.uniform(self.scale[0], self.scale[1]) * area 255 aspect_ratio = random.uniform(self.ratio[0], self.ratio[1]) 256 257 258 new_w = int(round(math.sqrt(target_area * aspect_ratio))) 259 new_h = int(round(math.sqrt(target_area / aspect_ratio))) 260 261 if random.random() < 0.5: 262 new_h, new_w = new_w, new_h 263 264 if new_w < w and new_h < h: 265 x0 = random.randint(0, w - new_w) 266 y0 = (random.randint(0, h - new_h))//d 267 out = fixed_crop(img, x0, y0, new_w, new_h, self.size) 268 269 return out 270 271 # Fallback 272 return center_crop(img, self.size) 273 274 275 class DownCrop(): 276 def __init__(self, size, select, scale=(0.36,0.81)): 277 self.size = size 278 self.scale = scale 279 self.select = select 280 281 def __call__(self,img, attr_idx): 282 if attr_idx not in self.select: 283 return img, attr_idx 284 if attr_idx == 0: 285 self.scale=(0.64,1.0) 286 h, w, _ = img.shape 287 area = h * w 288 289 s = (self.scale[0]+self.scale[1])/2.0 290 291 target_area = s * area 292 293 new_w = int(round(math.sqrt(target_area))) 294 new_h = int(round(math.sqrt(target_area))) 295 296 if new_w < w and new_h < h: 297 dw = w-new_w 298 x0 = int(0.5*dw) 299 y0 = h-new_h 300 out = fixed_crop(img, x0, y0, new_w, new_h, self.size) 301 return out, attr_idx 302 303 # Fallback 304 return center_crop(img, self.size), attr_idx 305 306 307 class ResizedCrop(object): 308 def __init__(self, size, select,scale=(0.64, 1.0), ratio=(3. / 4., 4. / 3.)): 309 self.size = size 310 self.scale = scale 311 self.ratio = ratio 312 self.select = select 313 314 def __call__(self,img, attr_idx): 315 if attr_idx not in self.select: 316 return img, attr_idx 317 h, w, _ = img.shape 318 area = h * w 319 d=1 320 if attr_idx == 2: 321 self.scale=(0.36,0.81) 322 d=2 323 if attr_idx == 0: 324 self.scale=(0.81,1.0) 325 326 target_area = (self.scale[0]+self.scale[1])/2.0 * area 327 # aspect_ratio = random.uniform(self.ratio[0], self.ratio[1]) 328 329 330 new_w = int(round(math.sqrt(target_area))) 331 new_h = int(round(math.sqrt(target_area))) 332 333 # if random.random() < 0.5: 334 # new_h, new_w = new_w, new_h 335 336 if new_w < w and new_h < h: 337 x0 = (w - new_w)//2 338 y0 = (h - new_h)//d//2 339 out = fixed_crop(img, x0, y0, new_w, new_h, self.size) 340 # cv2.imshow('{}_img'.format(idx2attr_map[attr_idx]), img) 341 # cv2.imshow('{}_crop'.format(idx2attr_map[attr_idx]), out) 342 # 343 # cv2.waitKey(0) 344 return out, attr_idx 345 346 # Fallback 347 return center_crop(img, self.size), attr_idx 348 349 class RandomHflip(object): 350 def __call__(self, image): 351 if random.randint(2): 352 return cv2.flip(image, 1) 353 else: 354 return image 355 class RandomVflip(object): 356 def __call__(self, image): 357 if random.randint(2): 358 return cv2.flip(image, 0) 359 else: 360 return image 361 362 363 class Hflip(object): 364 def __init__(self,doHflip): 365 self.doHflip = doHflip 366 367 def __call__(self, image): 368 if self.doHflip: 369 return cv2.flip(image, 1) 370 else: 371 return image 372 373 374 class CenterCrop(object): 375 def __init__(self, size): 376 self.size = size 377 378 def __call__(self, image): 379 return center_crop(image, self.size) 380 381 class UpperCrop(): 382 def __init__(self, size, scale=(0.09, 0.64)): 383 self.size = size 384 self.scale = scale 385 386 def __call__(self,img): 387 h, w, _ = img.shape 388 area = h * w 389 390 s = (self.scale[0]+self.scale[1])/2.0 391 392 target_area = s * area 393 394 new_w = int(round(math.sqrt(target_area))) 395 new_h = int(round(math.sqrt(target_area))) 396 397 if new_w < w and new_h < h: 398 dw = w-new_w 399 x0 = int(0.5*dw) 400 y0 = 0 401 out = fixed_crop(img, x0, y0, new_w, new_h, self.size) 402 return out 403 404 # Fallback 405 return center_crop(img, self.size) 406 407 408 409 class RandomUpperCrop(object): 410 def __init__(self, size, select, scale=(0.09, 0.64), ratio=(3. / 4., 4. / 3.)): 411 self.size = size 412 self.scale = scale 413 self.ratio = ratio 414 self.select = select 415 416 def __call__(self,img, attr_idx): 417 if random.random() < 0.2: 418 return img, attr_idx 419 if attr_idx not in self.select: 420 return img, attr_idx 421 422 h, w, _ = img.shape 423 area = h * w 424 for attempt in range(10): 425 s = random.uniform(self.scale[0], self.scale[1]) 426 d = 0.1 + (0.3 - 0.1) / (self.scale[1] - self.scale[0]) * (s - self.scale[0]) 427 target_area = s * area 428 aspect_ratio = random.uniform(self.ratio[0], self.ratio[1]) 429 new_w = int(round(math.sqrt(target_area * aspect_ratio))) 430 new_h = int(round(math.sqrt(target_area / aspect_ratio))) 431 432 433 # new_w = int(round(math.sqrt(target_area))) 434 # new_h = int(round(math.sqrt(target_area))) 435 436 if new_w < w and new_h < h: 437 dw = w-new_w 438 x0 = random.randint(int((0.5-d)*dw), int((0.5+d)*dw)+1) 439 y0 = (random.randint(0, h - new_h))//10 440 out = fixed_crop(img, x0, y0, new_w, new_h, self.size) 441 return out, attr_idx 442 443 # Fallback 444 return center_crop(img, self.size), attr_idx 445 class RandomDownCrop(object): 446 def __init__(self, size, select, scale=(0.36, 0.81), ratio=(3. / 4., 4. / 3.)): 447 self.size = size 448 self.scale = scale 449 self.ratio = ratio 450 self.select = select 451 452 def __call__(self,img, attr_idx): 453 if random.random() < 0.2: 454 return img, attr_idx 455 if attr_idx not in self.select: 456 return img, attr_idx 457 if attr_idx == 0: 458 self.scale=(0.64,1.0) 459 460 h, w, _ = img.shape 461 area = h * w 462 for attempt in range(10): 463 s = random.uniform(self.scale[0], self.scale[1]) 464 d = 0.1 + (0.3 - 0.1) / (self.scale[1] - self.scale[0]) * (s - self.scale[0]) 465 target_area = s * area 466 aspect_ratio = random.uniform(self.ratio[0], self.ratio[1]) 467 new_w = int(round(math.sqrt(target_area * aspect_ratio))) 468 new_h = int(round(math.sqrt(target_area / aspect_ratio))) 469 # 470 # new_w = int(round(math.sqrt(target_area))) 471 # new_h = int(round(math.sqrt(target_area))) 472 473 if new_w < w and new_h < h: 474 dw = w-new_w 475 x0 = random.randint(int((0.5-d)*dw), int((0.5+d)*dw)+1) 476 y0 = (random.randint((h - new_h)*9//10, h - new_h)) 477 out = fixed_crop(img, x0, y0, new_w, new_h, self.size) 478 479 # cv2.imshow('{}_img'.format(idx2attr_map[attr_idx]), img) 480 # cv2.imshow('{}_crop'.format(idx2attr_map[attr_idx]), out) 481 # 482 # cv2.waitKey(0) 483 484 return out, attr_idx 485 486 # Fallback 487 return center_crop(img, self.size), attr_idx 488 489 class RandomHShift(object): 490 def __init__(self, select, scale=(0.0, 0.2)): 491 self.scale = scale 492 self.select = select 493 494 def __call__(self,img, attr_idx): 495 if attr_idx not in self.select: 496 return img, attr_idx 497 do_shift_crop = random.randint(0, 2) 498 if do_shift_crop: 499 h, w, _ = img.shape 500 min_shift = int(w*self.scale[0]) 501 max_shift = int(w*self.scale[1]) 502 shift_idx = random.randint(min_shift, max_shift) 503 direction = random.randint(0,2) 504 if direction: 505 right_part = img[:, -shift_idx:, :] 506 left_part = img[:, :-shift_idx, :] 507 else: 508 left_part = img[:, :shift_idx, :] 509 right_part = img[:, shift_idx:, :] 510 img = np.concatenate((right_part, left_part), axis=1) 511 512 # Fallback 513 return img, attr_idx 514 515 516 class RandomBottomCrop(object): 517 def __init__(self, size, select, scale=(0.4, 0.8)): 518 self.size = size 519 self.scale = scale 520 self.select = select 521 522 def __call__(self,img, attr_idx): 523 if attr_idx not in self.select: 524 return img, attr_idx 525 526 h, w, _ = img.shape 527 area = h * w 528 for attempt in range(10): 529 s = random.uniform(self.scale[0], self.scale[1]) 530 d = 0.25 + (0.45 - 0.25) / (self.scale[1] - self.scale[0]) * (s - self.scale[0]) 531 target_area = s * area 532 533 new_w = int(round(math.sqrt(target_area))) 534 new_h = int(round(math.sqrt(target_area))) 535 536 if new_w < w and new_h < h: 537 dw = w-new_w 538 dh = h - new_h 539 x0 = random.randint(int((0.5-d)*dw), min(int((0.5+d)*dw)+1,dw)) 540 y0 = (random.randint(max(0,int(0.8*dh)-1), dh)) 541 out = fixed_crop(img, x0, y0, new_w, new_h, self.size) 542 return out, attr_idx 543 544 # Fallback 545 return bottom_crop(img, self.size), attr_idx 546 547 548 class BottomCrop(): 549 def __init__(self, size, select, scale=(0.4, 0.8)): 550 self.size = size 551 self.scale = scale 552 self.select = select 553 554 def __call__(self,img, attr_idx): 555 if attr_idx not in self.select: 556 return img, attr_idx 557 558 h, w, _ = img.shape 559 area = h * w 560 561 s = (self.scale[0]+self.scale[1])/3.*2. 562 563 target_area = s * area 564 565 new_w = int(round(math.sqrt(target_area))) 566 new_h = int(round(math.sqrt(target_area))) 567 568 if new_w < w and new_h < h: 569 dw = w-new_w 570 dh = h-new_h 571 x0 = int(0.5*dw) 572 y0 = int(0.9*dh) 573 out = fixed_crop(img, x0, y0, new_w, new_h, self.size) 574 return out, attr_idx 575 576 # Fallback 577 return bottom_crop(img, self.size), attr_idx 578 579 580 581 class Resize(object): 582 def __init__(self, size, inter=cv2.INTER_CUBIC): 583 self.size = size 584 self.inter = inter 585 586 def __call__(self, image): 587 return cv2.resize(image, (self.size[0], self.size[0]), interpolation=self.inter) 588 589 class ExpandBorder(object): 590 def __init__(self, mode='constant', value=255, size=(336,336), resize=False): 591 self.mode = mode 592 self.value = value 593 self.resize = resize 594 self.size = size 595 596 def __call__(self, image): 597 h, w, _ = image.shape 598 if h > w: 599 pad1 = (h-w)//2 600 pad2 = h - w - pad1 601 if self.mode == 'constant': 602 image = np.pad(image, ((0, 0), (pad1, pad2), (0, 0)), 603 self.mode, constant_values=self.value) 604 else: 605 image = np.pad(image,((0,0), (pad1, pad2),(0,0)), self.mode) 606 elif h < w: 607 pad1 = (w-h)//2 608 pad2 = w-h - pad1 609 if self.mode == 'constant': 610 image = np.pad(image, ((pad1, pad2),(0, 0), (0, 0)), 611 self.mode,constant_values=self.value) 612 else: 613 image = np.pad(image, ((pad1, pad2), (0, 0), (0, 0)),self.mode) 614 if self.resize: 615 image = cv2.resize(image, (self.size[0], self.size[0]),interpolation=cv2.INTER_LINEAR) 616 return image 617 class AstypeToInt(): 618 def __call__(self, image, attr_idx): 619 return image.clip(0,255.0).astype(np.uint8), attr_idx 620 621 class AstypeToFloat(): 622 def __call__(self, image, attr_idx): 623 return image.astype(np.float32), attr_idx 624 625 import matplotlib.pyplot as plt 626 class Normalize(object): 627 def __init__(self,mean, std): 628 ''' 629 :param mean: RGB order 630 :param std: RGB order 631 ''' 632 self.mean = np.array(mean).reshape(3,1,1) 633 self.std = np.array(std).reshape(3,1,1) 634 def __call__(self, image): 635 ''' 636 :param image: (H,W,3) RGB 637 :return: 638 ''' 639 # plt.figure(1) 640 # plt.imshow(image) 641 # plt.show() 642 return (image.transpose((2, 0, 1)) / 255. - self.mean) / self.std 643 644 class RandomErasing(object): 645 def __init__(self, select,EPSILON=0.5,sl=0.02, sh=0.09, r1=0.3, mean=[0.485, 0.456, 0.406]): 646 self.EPSILON = EPSILON 647 self.mean = mean 648 self.sl = sl 649 self.sh = sh 650 self.r1 = r1 651 self.select = select 652 653 def __call__(self, img,attr_idx): 654 if attr_idx not in self.select: 655 return img,attr_idx 656 657 if random.uniform(0, 1) > self.EPSILON: 658 return img,attr_idx 659 660 for attempt in range(100): 661 area = img.shape[1] * img.shape[2] 662 663 target_area = random.uniform(self.sl, self.sh) * area 664 aspect_ratio = random.uniform(self.r1, 1 / self.r1) 665 666 h = int(round(math.sqrt(target_area * aspect_ratio))) 667 w = int(round(math.sqrt(target_area / aspect_ratio))) 668 669 if w <= img.shape[2] and h <= img.shape[1]: 670 x1 = random.randint(0, img.shape[1] - h) 671 y1 = random.randint(0, img.shape[2] - w) 672 if img.shape[0] == 3: 673 # img[0, x1:x1+h, y1:y1+w] = random.uniform(0, 1) 674 # img[1, x1:x1+h, y1:y1+w] = random.uniform(0, 1) 675 # img[2, x1:x1+h, y1:y1+w] = random.uniform(0, 1) 676 img[0, x1:x1 + h, y1:y1 + w] = self.mean[0] 677 img[1, x1:x1 + h, y1:y1 + w] = self.mean[1] 678 img[2, x1:x1 + h, y1:y1 + w] = self.mean[2] 679 # img[:, x1:x1+h, y1:y1+w] = torch.from_numpy(np.random.rand(3, h, w)) 680 else: 681 img[0, x1:x1 + h, y1:y1 + w] = self.mean[1] 682 # img[0, x1:x1+h, y1:y1+w] = torch.from_numpy(np.random.rand(1, h, w)) 683 return img,attr_idx 684 685 return img,attr_idx 686 687 if __name__ == '__main__': 688 import matplotlib.pyplot as plt 689 690 691 class FSAug(object): 692 def __init__(self): 693 self.augment = Compose([ 694 AstypeToFloat(), 695 # RandomHShift(scale=(0.,0.2),select=range(8)), 696 # RandomRotate(angles=(-20., 20.), bound=True), 697 ExpandBorder(select=range(8), mode='symmetric'),# symmetric 698 # Resize(size=(336, 336), select=[ 2, 7]), 699 AstypeToInt() 700 ]) 701 702 def __call__(self, spct,attr_idx): 703 return self.augment(spct,attr_idx) 704 705 706 trans = FSAug() 707 708 img_path = '/media/gserver/data/FashionAI/round2/train/Images/coat_length_labels/0b6b4a2146fc8616a19fcf2026d61d50.jpg' 709 img = cv2.cvtColor(cv2.imread(img_path),cv2.COLOR_BGR2RGB) 710 img_trans,_ = trans(img,5) 711 # img_trans2,_ = trans(img,6) 712 713 plt.figure() 714 plt.subplot(221) 715 plt.imshow(img) 716 717 plt.subplot(222) 718 plt.imshow(img_trans) 719 720 # plt.subplot(223) 721 # plt.imshow(img_trans2) 722 # plt.imshow(img_trans2) 723 plt.show()
方式二： 用于线下增强数据，采用的方法是
- 高斯噪声
- 亮度变化
- 左右翻转
- 上下翻转
- 色彩抖动
- 对化
- 锐度变化
-
1 from PIL import Image,ImageEnhance,ImageFilter,ImageOps 2 import os 3 import shutil 4 import numpy as np 5 import cv2 6 import random 7 from skimage.util import random_noise 8 from skimage import exposure 9 10 11 image_number = 0 12 13 raw_path = "./data/train/" 14 15 new_path = "./aug/train/" 16 17 # 加高斯噪声 18 def addNoise(img): 19 ''' 20 注意：输出的像素是[0,1]之间,所以乘以5得到[0,255]之间 21 ''' 22 return random_noise(img, mode='gaussian', seed=13, clip=True)*255 23 24 def changeLight(img): 25 rate = random.uniform(0.5, 1.5) 26 # print(rate) 27 img = exposure.adjust_gamma(img, rate) #大于1为调暗，小于1为调亮;1.05 28 return img 29 30 try: 31 for i in range(59): 32 os.makedirs(new_path + os.sep + str(i)) 33 except: 34 pass 35 36 for raw_dir_name in range(59): 37 38 raw_dir_name = str(raw_dir_name) 39 40 saved_image_path = new_path + raw_dir_name+"/" 41 42 raw_image_path = raw_path + raw_dir_name+"/" 43 44 if not os.path.exists(saved_image_path): 45 46 os.mkdir(saved_image_path) 47 48 raw_image_file_name = os.listdir(raw_image_path) 49 50 raw_image_file_path = [] 51 52 for i in raw_image_file_name: 53 54 raw_image_file_path.append(raw_image_path+i) 55 56 for x in raw_image_file_path: 57 58 img = Image.open(x) 59 cv_image = cv2.imread(x) 60 61 # 高斯噪声 62 gau_image = addNoise(cv_image) 63 # 随机改变 64 light = changeLight(cv_image) 65 light_and_gau = addNoise(light) 66 67 cv2.imwrite(saved_image_path + "gau_" + os.path.basename(x),gau_image) 68 cv2.imwrite(saved_image_path + "light_" + os.path.basename(x),light) 69 cv2.imwrite(saved_image_path + "gau_light" + os.path.basename(x),light_and_gau) 70 #img = img.resize((800,600)) 71 72 #1.翻转 73 74 img_flip_left_right = img.transpose(Image.FLIP_LEFT_RIGHT) 75 76 img_flip_top_bottom = img.transpose(Image.FLIP_TOP_BOTTOM) 77 78 #2.旋转 79 80 #img_rotate_90 = img.transpose(Image.ROTATE_90) 81 82 #img_rotate_180 = img.transpose(Image.ROTATE_180) 83 84 #img_rotate_270 = img.transpose(Image.ROTATE_270) 85 86 #img_rotate_90_left = img_flip_left_right.transpose(Image.ROTATE_90) 87 88 #img_rotate_270_left = img_flip_left_right.transpose(Image.ROTATE_270) 89 90 #3.亮度 91 92 #enh_bri = ImageEnhance.Brightness(img) 93 #brightness = 1.5 94 #image_brightened = enh_bri.enhance(brightness) 95 96 #4.色彩 97 98 #enh_col = ImageEnhance.Color(img) 99 #color = 1.5 100 101 #image_colored = enh_col.enhance(color) 102 103 #5.对比度 104 105 enh_con = ImageEnhance.Contrast(img) 106 107 contrast = 1.5 108 109 image_contrasted = enh_con.enhance(contrast) 110 111 #6.锐度 112 113 #enh_sha = ImageEnhance.Sharpness(img) 114 #sharpness = 3.0 115 116 #image_sharped = enh_sha.enhance(sharpness) 117 118 #保存 119 120 img.save(saved_image_path + os.path.basename(x)) 121 122 img_flip_left_right.save(saved_image_path + "left_right_" + os.path.basename(x)) 123 124 img_flip_top_bottom.save(saved_image_path + "top_bottom_" + os.path.basename(x)) 125 126 #img_rotate_90.save(saved_image_path + "rotate_90_" + os.path.basename(x)) 127 128 #img_rotate_180.save(saved_image_path + "rotate_180_" + os.path.basename(x)) 129 130 #img_rotate_270.save(saved_image_path + "rotate_270_" + os.path.basename(x)) 131 132 #img_rotate_90_left.save(saved_image_path + "rotate_90_left_" + os.path.basename(x)) 133 134 #img_rotate_270_left.save(saved_image_path + "rotate_270_left_" + os.path.basename(x)) 135 136 #image_brightened.save(saved_image_path + "brighted_" + os.path.basename(x)) 137 138 #image_colored.save(saved_image_path + "colored_" + os.path.basename(x)) 139 140 image_contrasted.save(saved_image_path + "contrasted_" + os.path.basename(x)) 141 142 #image_sharped.save(saved_image_path + "sharped_" + os.path.basename(x)) 143 144 image_number += 1 145 146 print("convert pictur" "es :%s size:%s mode:%s" % (image_number, img.size, img.mode)) 147 148
加载数据的类（自定义继承）
- 与pytorch中的加载数据类差不多，只是多了自己的某些功能。
-
1 from torch.utils.data import Dataset 2 from torchvision import transforms as T 3 from config import config 4 from PIL import Image 5 from itertools import chain 6 from glob import glob 7 from tqdm import tqdm 8 import random 9 import numpy as np 10 import pandas as pd 11 import os 12 import cv2 13 import torch 14 15 #1.set random seed 16 random.seed(config.seed) 17 np.random.seed(config.seed) 18 torch.manual_seed(config.seed) 19 torch.cuda.manual_seed_all(config.seed) 20 21 #2.define dataset 22 class ZiyiDataset(Dataset): 23 def __init__(self,label_list,transforms=None,train=True,test=False): 24 self.test = test 25 self.train = train 26 imgs = [] 27 if self.test: 28 for index,row in label_list.iterrows(): 29 imgs.append((row["filename"])) 30 self.imgs = imgs 31 else: 32 for index,row in label_list.iterrows(): 33 imgs.append((row["filename"],row["label"])) 34 self.imgs = imgs 35 if transforms is None: 36 if self.test or not train: 37 self.transforms = T.Compose([ 38 T.Resize((config.img_weight,config.img_height)), 39 T.ToTensor(), 40 T.Normalize(mean = [0.485,0.456,0.406], 41 std = [0.229,0.224,0.225])]) 42 else: 43 self.transforms = T.Compose([ 44 T.Resize((config.img_weight,config.img_height)), 45 T.RandomRotation(30), 46 T.RandomHorizontalFlip(), 47 T.RandomVerticalFlip(), 48 T.RandomAffine(45), 49 T.ToTensor(), 50 T.Normalize(mean = [0.485,0.456,0.406], 51 std = [0.229,0.224,0.225])]) 52 else: 53 self.transforms = transforms 54 def __getitem__(self,index): 55 if self.test: 56 filename = self.imgs[index] 57 img = Image.open(filename) 58 img = self.transforms(img) 59 return img,filename 60 else: 61 filename,label = self.imgs[index] 62 img = Image.open(filename) 63 img = self.transforms(img) 64 return img,label 65 def __len__(self): 66 return len(self.imgs) 67 68 def collate_fn(batch): 69 imgs = [] 70 label = [] 71 for sample in batch: 72 imgs.append(sample[0]) 73 label.append(sample[1]) 74 75 return torch.stack(imgs, 0), \ 76 label 77 78 def get_files(root,mode): 79 #for test 80 if mode == "test": 81 files = [] 82 for img in os.listdir(root): 83 files.append(root + img) 84 files = pd.DataFrame({"filename":files}) 85 return files 86 elif mode != "test": 87 #for train and val 88 all_data_path,labels = [],[] 89 image_folders = list(map(lambda x:root+x,os.listdir(root))) 90 jpg_image_1 = list(map(lambda x:glob(x+"/*.jpg"),image_folders)) 91 jpg_image_2 = list(map(lambda x:glob(x+"/*.JPG"),image_folders)) 92 all_images = list(chain.from_iterable(jpg_image_1 + jpg_image_2)) 93 print("loading train dataset") 94 for file in tqdm(all_images): 95 all_data_path.append(file) 96 labels.append(int(file.split("/")[-2])) 97 all_files = pd.DataFrame({"filename":all_data_path,"label":labels}) 98 return all_files 99 else: 100 print("check the mode please!") 101
3.获取模型
获取模型较为简单，单一模型采取pytorch中的预训练模型，添加所需要的层，进行微调然后迁移学习新数据。
1 import torchvision 2 import torch.nn.functional as F 3 from torch import nn 4 from config import config 5 6 def generate_model(): 7 class DenseModel(nn.Module): 8 def __init__(self, pretrained_model): 9 super(DenseModel, self).__init__() 10 self.classifier = nn.Linear(pretrained_model.classifier.in_features, config.num_classes) 11 12 for m in self.modules(): 13 if isinstance(m, nn.Conv2d): 14 nn.init.kaiming_normal(m.weight) 15 elif isinstance(m, nn.BatchNorm2d): 16 m.weight.data.fill_(1) 17 m.bias.data.zero_() 18 elif isinstance(m, nn.Linear): 19 m.bias.data.zero_() 20 21 self.features = pretrained_model.features 22 self.layer1 = pretrained_model.features._modules['denseblock1'] 23 self.layer2 = pretrained_model.features._modules['denseblock2'] 24 self.layer3 = pretrained_model.features._modules['denseblock3'] 25 self.layer4 = pretrained_model.features._modules['denseblock4'] 26 27 def forward(self, x): 28 features = self.features(x) 29 out = F.relu(features, inplace=True) 30 out = F.avg_pool2d(out, kernel_size=8).view(features.size(0), -1) 31 out = F.sigmoid(self.classifier(out)) 32 return out 33 34 return DenseModel(torchvision.models.densenet169(pretrained=True)) 35 36 def get_net(): 37 #return MyModel(torchvision.models.resnet101(pretrained = True)) 38 model = torchvision.models.resnet50(pretrained = True) 39 #for param in model.parameters(): 40 # param.requires_grad = False 41 # pytorch添加层的方式直接在Model.层名=层具体形式 42 model.avgpool = nn.AdaptiveAvgPool2d(1) 43 model.fc = nn.Linear(2048,config.num_classes) #添加全连接层以作分类任务，num_classes为分类个数 44 return model
---恢复内容结束---