You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1038 lines
45KB

  1. # YOLOv5 🚀 by Ultralytics, GPL-3.0 license
  2. """
  3. Dataloaders and dataset utils
  4. """
  5. import glob
  6. import hashlib
  7. import json
  8. import math
  9. import os
  10. import random
  11. import shutil
  12. import time
  13. from itertools import repeat
  14. from multiprocessing.pool import Pool, ThreadPool
  15. from pathlib import Path
  16. from threading import Thread
  17. from zipfile import ZipFile
  18. import cv2
  19. import numpy as np
  20. import torch
  21. import torch.nn.functional as F
  22. import yaml
  23. from PIL import ExifTags, Image, ImageOps
  24. from torch.utils.data import DataLoader, Dataset, dataloader, distributed
  25. from tqdm import tqdm
  26. from utils.augmentations import Albumentations, augment_hsv, copy_paste, letterbox, mixup, random_perspective
  27. from utils.general import (DATASETS_DIR, LOGGER, NUM_THREADS, check_dataset, check_requirements, check_yaml, clean_str,
  28. segments2boxes, xyn2xy, xywh2xyxy, xywhn2xyxy, xyxy2xywhn)
  29. from utils.torch_utils import torch_distributed_zero_first
  30. # Parameters
  31. HELP_URL = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data'
  32. IMG_FORMATS = ['bmp', 'dng', 'jpeg', 'jpg', 'mpo', 'png', 'tif', 'tiff', 'webp'] # include image suffixes
  33. VID_FORMATS = ['asf', 'avi', 'gif', 'm4v', 'mkv', 'mov', 'mp4', 'mpeg', 'mpg', 'wmv'] # include video suffixes
  34. # Get orientation exif tag
  35. for orientation in ExifTags.TAGS.keys():
  36. if ExifTags.TAGS[orientation] == 'Orientation':
  37. break
  38. def get_hash(paths):
  39. # Returns a single hash value of a list of paths (files or dirs)
  40. size = sum(os.path.getsize(p) for p in paths if os.path.exists(p)) # sizes
  41. h = hashlib.md5(str(size).encode()) # hash sizes
  42. h.update(''.join(paths).encode()) # hash paths
  43. return h.hexdigest() # return hash
  44. def exif_size(img):
  45. # Returns exif-corrected PIL size
  46. s = img.size # (width, height)
  47. try:
  48. rotation = dict(img._getexif().items())[orientation]
  49. if rotation == 6: # rotation 270
  50. s = (s[1], s[0])
  51. elif rotation == 8: # rotation 90
  52. s = (s[1], s[0])
  53. except Exception:
  54. pass
  55. return s
  56. def exif_transpose(image):
  57. """
  58. Transpose a PIL image accordingly if it has an EXIF Orientation tag.
  59. Inplace version of https://github.com/python-pillow/Pillow/blob/master/src/PIL/ImageOps.py exif_transpose()
  60. :param image: The image to transpose.
  61. :return: An image.
  62. """
  63. exif = image.getexif()
  64. orientation = exif.get(0x0112, 1) # default 1
  65. if orientation > 1:
  66. method = {2: Image.FLIP_LEFT_RIGHT,
  67. 3: Image.ROTATE_180,
  68. 4: Image.FLIP_TOP_BOTTOM,
  69. 5: Image.TRANSPOSE,
  70. 6: Image.ROTATE_270,
  71. 7: Image.TRANSVERSE,
  72. 8: Image.ROTATE_90,
  73. }.get(orientation)
  74. if method is not None:
  75. image = image.transpose(method)
  76. del exif[0x0112]
  77. image.info["exif"] = exif.tobytes()
  78. return image
  79. def create_dataloader(path, imgsz, batch_size, stride, single_cls=False, hyp=None, augment=False, cache=False, pad=0.0,
  80. rect=False, rank=-1, workers=8, image_weights=False, quad=False, prefix='', shuffle=False):
  81. if rect and shuffle:
  82. LOGGER.warning('WARNING: --rect is incompatible with DataLoader shuffle, setting shuffle=False')
  83. shuffle = False
  84. with torch_distributed_zero_first(rank): # init dataset *.cache only once if DDP
  85. dataset = LoadImagesAndLabels(path, imgsz, batch_size,
  86. augment=augment, # augmentation
  87. hyp=hyp, # hyperparameters
  88. rect=rect, # rectangular batches
  89. cache_images=cache,
  90. single_cls=single_cls,
  91. stride=int(stride),
  92. pad=pad,
  93. image_weights=image_weights,
  94. prefix=prefix)
  95. batch_size = min(batch_size, len(dataset))
  96. nd = torch.cuda.device_count() # number of CUDA devices
  97. nw = min([os.cpu_count() // max(nd, 1), batch_size if batch_size > 1 else 0, workers]) # number of workers
  98. sampler = None if rank == -1 else distributed.DistributedSampler(dataset, shuffle=shuffle)
  99. loader = DataLoader if image_weights else InfiniteDataLoader # only DataLoader allows for attribute updates
  100. return loader(dataset,
  101. batch_size=batch_size,
  102. shuffle=shuffle and sampler is None,
  103. num_workers=nw,
  104. sampler=sampler,
  105. pin_memory=True,
  106. collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn), dataset
  107. class InfiniteDataLoader(dataloader.DataLoader):
  108. """ Dataloader that reuses workers
  109. Uses same syntax as vanilla DataLoader
  110. """
  111. def __init__(self, *args, **kwargs):
  112. super().__init__(*args, **kwargs)
  113. object.__setattr__(self, 'batch_sampler', _RepeatSampler(self.batch_sampler))
  114. self.iterator = super().__iter__()
  115. def __len__(self):
  116. return len(self.batch_sampler.sampler)
  117. def __iter__(self):
  118. for i in range(len(self)):
  119. yield next(self.iterator)
  120. class _RepeatSampler:
  121. """ Sampler that repeats forever
  122. Args:
  123. sampler (Sampler)
  124. """
  125. def __init__(self, sampler):
  126. self.sampler = sampler
  127. def __iter__(self):
  128. while True:
  129. yield from iter(self.sampler)
  130. class LoadImages:
  131. # YOLOv5 image/video dataloader, i.e. `python detect.py --source image.jpg/vid.mp4`
  132. def __init__(self, path, img_size=640, stride=32, auto=True):
  133. p = str(Path(path).resolve()) # os-agnostic absolute path
  134. if '*' in p:
  135. files = sorted(glob.glob(p, recursive=True)) # glob
  136. elif os.path.isdir(p):
  137. files = sorted(glob.glob(os.path.join(p, '*.*'))) # dir
  138. elif os.path.isfile(p):
  139. files = [p] # files
  140. else:
  141. raise Exception(f'ERROR: {p} does not exist')
  142. images = [x for x in files if x.split('.')[-1].lower() in IMG_FORMATS]
  143. videos = [x for x in files if x.split('.')[-1].lower() in VID_FORMATS]
  144. ni, nv = len(images), len(videos)
  145. self.img_size = img_size
  146. self.stride = stride
  147. self.files = images + videos
  148. self.nf = ni + nv # number of files
  149. self.video_flag = [False] * ni + [True] * nv
  150. self.mode = 'image'
  151. self.auto = auto
  152. if any(videos):
  153. self.new_video(videos[0]) # new video
  154. else:
  155. self.cap = None
  156. assert self.nf > 0, f'No images or videos found in {p}. ' \
  157. f'Supported formats are:\nimages: {IMG_FORMATS}\nvideos: {VID_FORMATS}'
  158. def __iter__(self):
  159. self.count = 0
  160. return self
  161. def __next__(self):
  162. if self.count == self.nf:
  163. raise StopIteration
  164. path = self.files[self.count]
  165. if self.video_flag[self.count]:
  166. # Read video
  167. self.mode = 'video'
  168. ret_val, img0 = self.cap.read()
  169. while not ret_val:
  170. self.count += 1
  171. self.cap.release()
  172. if self.count == self.nf: # last video
  173. raise StopIteration
  174. else:
  175. path = self.files[self.count]
  176. self.new_video(path)
  177. ret_val, img0 = self.cap.read()
  178. self.frame += 1
  179. s = f'video {self.count + 1}/{self.nf} ({self.frame}/{self.frames}) {path}: '
  180. else:
  181. # Read image
  182. self.count += 1
  183. img0 = cv2.imread(path) # BGR
  184. assert img0 is not None, f'Image Not Found {path}'
  185. s = f'image {self.count}/{self.nf} {path}: '
  186. # Padded resize
  187. img = letterbox(img0, self.img_size, stride=self.stride, auto=self.auto)[0]
  188. # Convert
  189. img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB
  190. img = np.ascontiguousarray(img)
  191. return path, img, img0, self.cap, s
  192. def new_video(self, path):
  193. self.frame = 0
  194. self.cap = cv2.VideoCapture(path)
  195. self.frames = int(self.cap.get(cv2.CAP_PROP_FRAME_COUNT))
  196. def __len__(self):
  197. return self.nf # number of files
  198. class LoadWebcam: # for inference
  199. # YOLOv5 local webcam dataloader, i.e. `python detect.py --source 0`
  200. def __init__(self, pipe='0', img_size=640, stride=32):
  201. self.img_size = img_size
  202. self.stride = stride
  203. self.pipe = eval(pipe) if pipe.isnumeric() else pipe
  204. self.cap = cv2.VideoCapture(self.pipe) # video capture object
  205. self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 3) # set buffer size
  206. def __iter__(self):
  207. self.count = -1
  208. return self
  209. def __next__(self):
  210. self.count += 1
  211. if cv2.waitKey(1) == ord('q'): # q to quit
  212. self.cap.release()
  213. cv2.destroyAllWindows()
  214. raise StopIteration
  215. # Read frame
  216. ret_val, img0 = self.cap.read()
  217. img0 = cv2.flip(img0, 1) # flip left-right
  218. # Print
  219. assert ret_val, f'Camera Error {self.pipe}'
  220. img_path = 'webcam.jpg'
  221. s = f'webcam {self.count}: '
  222. # Padded resize
  223. img = letterbox(img0, self.img_size, stride=self.stride)[0]
  224. # Convert
  225. img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB
  226. img = np.ascontiguousarray(img)
  227. return img_path, img, img0, None, s
  228. def __len__(self):
  229. return 0
  230. class LoadStreams:
  231. # YOLOv5 streamloader, i.e. `python detect.py --source 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP streams`
  232. def __init__(self, sources='streams.txt', img_size=640, stride=32, auto=True):
  233. self.mode = 'stream'
  234. self.img_size = img_size
  235. self.stride = stride
  236. if os.path.isfile(sources):
  237. with open(sources) as f:
  238. sources = [x.strip() for x in f.read().strip().splitlines() if len(x.strip())]
  239. else:
  240. sources = [sources]
  241. n = len(sources)
  242. self.imgs, self.fps, self.frames, self.threads = [None] * n, [0] * n, [0] * n, [None] * n
  243. self.sources = [clean_str(x) for x in sources] # clean source names for later
  244. self.auto = auto
  245. for i, s in enumerate(sources): # index, source
  246. # Start thread to read frames from video stream
  247. st = f'{i + 1}/{n}: {s}... '
  248. if 'youtube.com/' in s or 'youtu.be/' in s: # if source is YouTube video
  249. check_requirements(('pafy', 'youtube_dl==2020.12.2'))
  250. import pafy
  251. s = pafy.new(s).getbest(preftype="mp4").url # YouTube URL
  252. s = eval(s) if s.isnumeric() else s # i.e. s = '0' local webcam
  253. cap = cv2.VideoCapture(s)
  254. assert cap.isOpened(), f'{st}Failed to open {s}'
  255. w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
  256. h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
  257. fps = cap.get(cv2.CAP_PROP_FPS) # warning: may return 0 or nan
  258. self.frames[i] = max(int(cap.get(cv2.CAP_PROP_FRAME_COUNT)), 0) or float('inf') # infinite stream fallback
  259. self.fps[i] = max((fps if math.isfinite(fps) else 0) % 100, 0) or 30 # 30 FPS fallback
  260. _, self.imgs[i] = cap.read() # guarantee first frame
  261. self.threads[i] = Thread(target=self.update, args=([i, cap, s]), daemon=True)
  262. LOGGER.info(f"{st} Success ({self.frames[i]} frames {w}x{h} at {self.fps[i]:.2f} FPS)")
  263. self.threads[i].start()
  264. LOGGER.info('') # newline
  265. # check for common shapes
  266. s = np.stack([letterbox(x, self.img_size, stride=self.stride, auto=self.auto)[0].shape for x in self.imgs])
  267. self.rect = np.unique(s, axis=0).shape[0] == 1 # rect inference if all shapes equal
  268. if not self.rect:
  269. LOGGER.warning('WARNING: Stream shapes differ. For optimal performance supply similarly-shaped streams.')
  270. def update(self, i, cap, stream):
  271. # Read stream `i` frames in daemon thread
  272. n, f, read = 0, self.frames[i], 1 # frame number, frame array, inference every 'read' frame
  273. while cap.isOpened() and n < f:
  274. n += 1
  275. # _, self.imgs[index] = cap.read()
  276. cap.grab()
  277. if n % read == 0:
  278. success, im = cap.retrieve()
  279. if success:
  280. self.imgs[i] = im
  281. else:
  282. LOGGER.warning('WARNING: Video stream unresponsive, please check your IP camera connection.')
  283. self.imgs[i] = np.zeros_like(self.imgs[i])
  284. cap.open(stream) # re-open stream if signal was lost
  285. time.sleep(1 / self.fps[i]) # wait time
  286. def __iter__(self):
  287. self.count = -1
  288. return self
  289. def __next__(self):
  290. self.count += 1
  291. if not all(x.is_alive() for x in self.threads) or cv2.waitKey(1) == ord('q'): # q to quit
  292. cv2.destroyAllWindows()
  293. raise StopIteration
  294. # Letterbox
  295. img0 = self.imgs.copy()
  296. img = [letterbox(x, self.img_size, stride=self.stride, auto=self.rect and self.auto)[0] for x in img0]
  297. # Stack
  298. img = np.stack(img, 0)
  299. # Convert
  300. img = img[..., ::-1].transpose((0, 3, 1, 2)) # BGR to RGB, BHWC to BCHW
  301. img = np.ascontiguousarray(img)
  302. return self.sources, img, img0, None, ''
  303. def __len__(self):
  304. return len(self.sources) # 1E12 frames = 32 streams at 30 FPS for 30 years
  305. def img2label_paths(img_paths):
  306. # Define label paths as a function of image paths
  307. sa, sb = os.sep + 'images' + os.sep, os.sep + 'labels' + os.sep # /images/, /labels/ substrings
  308. return [sb.join(x.rsplit(sa, 1)).rsplit('.', 1)[0] + '.txt' for x in img_paths]
  309. class LoadImagesAndLabels(Dataset):
  310. # YOLOv5 train_loader/val_loader, loads images and labels for training and validation
  311. cache_version = 0.6 # dataset labels *.cache version
  312. def __init__(self, path, img_size=640, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False,
  313. cache_images=False, single_cls=False, stride=32, pad=0.0, prefix=''):
  314. self.img_size = img_size
  315. self.augment = augment
  316. self.hyp = hyp
  317. self.image_weights = image_weights
  318. self.rect = False if image_weights else rect
  319. self.mosaic = self.augment and not self.rect # load 4 images at a time into a mosaic (only during training)
  320. self.mosaic_border = [-img_size // 2, -img_size // 2]
  321. self.stride = stride
  322. self.path = path
  323. self.albumentations = Albumentations() if augment else None
  324. try:
  325. f = [] # image files
  326. for p in path if isinstance(path, list) else [path]:
  327. p = Path(p) # os-agnostic
  328. if p.is_dir(): # dir
  329. f += glob.glob(str(p / '**' / '*.*'), recursive=True)
  330. # f = list(p.rglob('*.*')) # pathlib
  331. elif p.is_file(): # file
  332. with open(p) as t:
  333. t = t.read().strip().splitlines()
  334. parent = str(p.parent) + os.sep
  335. f += [x.replace('./', parent) if x.startswith('./') else x for x in t] # local to global path
  336. # f += [p.parent / x.lstrip(os.sep) for x in t] # local to global path (pathlib)
  337. else:
  338. raise Exception(f'{prefix}{p} does not exist')
  339. self.img_files = sorted(x.replace('/', os.sep) for x in f if x.split('.')[-1].lower() in IMG_FORMATS)
  340. # self.img_files = sorted([x for x in f if x.suffix[1:].lower() in IMG_FORMATS]) # pathlib
  341. assert self.img_files, f'{prefix}No images found'
  342. except Exception as e:
  343. raise Exception(f'{prefix}Error loading data from {path}: {e}\nSee {HELP_URL}')
  344. # Check cache
  345. self.label_files = img2label_paths(self.img_files) # labels
  346. cache_path = (p if p.is_file() else Path(self.label_files[0]).parent).with_suffix('.cache')
  347. try:
  348. cache, exists = np.load(cache_path, allow_pickle=True).item(), True # load dict
  349. assert cache['version'] == self.cache_version # same version
  350. assert cache['hash'] == get_hash(self.label_files + self.img_files) # same hash
  351. except Exception:
  352. cache, exists = self.cache_labels(cache_path, prefix), False # cache
  353. # Display cache
  354. nf, nm, ne, nc, n = cache.pop('results') # found, missing, empty, corrupt, total
  355. if exists:
  356. d = f"Scanning '{cache_path}' images and labels... {nf} found, {nm} missing, {ne} empty, {nc} corrupt"
  357. tqdm(None, desc=prefix + d, total=n, initial=n) # display cache results
  358. if cache['msgs']:
  359. LOGGER.info('\n'.join(cache['msgs'])) # display warnings
  360. assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {HELP_URL}'
  361. # Read cache
  362. [cache.pop(k) for k in ('hash', 'version', 'msgs')] # remove items
  363. labels, shapes, self.segments = zip(*cache.values())
  364. self.labels = list(labels)
  365. self.shapes = np.array(shapes, dtype=np.float64)
  366. self.img_files = list(cache.keys()) # update
  367. self.label_files = img2label_paths(cache.keys()) # update
  368. n = len(shapes) # number of images
  369. bi = np.floor(np.arange(n) / batch_size).astype(np.int) # batch index
  370. nb = bi[-1] + 1 # number of batches
  371. self.batch = bi # batch index of image
  372. self.n = n
  373. self.indices = range(n)
  374. # Update labels
  375. include_class = [] # filter labels to include only these classes (optional)
  376. include_class_array = np.array(include_class).reshape(1, -1)
  377. for i, (label, segment) in enumerate(zip(self.labels, self.segments)):
  378. if include_class:
  379. j = (label[:, 0:1] == include_class_array).any(1)
  380. self.labels[i] = label[j]
  381. if segment:
  382. self.segments[i] = segment[j]
  383. if single_cls: # single-class training, merge all classes into 0
  384. self.labels[i][:, 0] = 0
  385. if segment:
  386. self.segments[i][:, 0] = 0
  387. # Rectangular Training
  388. if self.rect:
  389. # Sort by aspect ratio
  390. s = self.shapes # wh
  391. ar = s[:, 1] / s[:, 0] # aspect ratio
  392. irect = ar.argsort()
  393. self.img_files = [self.img_files[i] for i in irect]
  394. self.label_files = [self.label_files[i] for i in irect]
  395. self.labels = [self.labels[i] for i in irect]
  396. self.shapes = s[irect] # wh
  397. ar = ar[irect]
  398. # Set training image shapes
  399. shapes = [[1, 1]] * nb
  400. for i in range(nb):
  401. ari = ar[bi == i]
  402. mini, maxi = ari.min(), ari.max()
  403. if maxi < 1:
  404. shapes[i] = [maxi, 1]
  405. elif mini > 1:
  406. shapes[i] = [1, 1 / mini]
  407. self.batch_shapes = np.ceil(np.array(shapes) * img_size / stride + pad).astype(np.int) * stride
  408. # Cache images into RAM/disk for faster training (WARNING: large datasets may exceed system resources)
  409. self.imgs, self.img_npy = [None] * n, [None] * n
  410. if cache_images:
  411. if cache_images == 'disk':
  412. self.im_cache_dir = Path(Path(self.img_files[0]).parent.as_posix() + '_npy')
  413. self.img_npy = [self.im_cache_dir / Path(f).with_suffix('.npy').name for f in self.img_files]
  414. self.im_cache_dir.mkdir(parents=True, exist_ok=True)
  415. gb = 0 # Gigabytes of cached images
  416. self.img_hw0, self.img_hw = [None] * n, [None] * n
  417. results = ThreadPool(NUM_THREADS).imap(self.load_image, range(n))
  418. pbar = tqdm(enumerate(results), total=n)
  419. for i, x in pbar:
  420. if cache_images == 'disk':
  421. if not self.img_npy[i].exists():
  422. np.save(self.img_npy[i].as_posix(), x[0])
  423. gb += self.img_npy[i].stat().st_size
  424. else: # 'ram'
  425. self.imgs[i], self.img_hw0[i], self.img_hw[i] = x # im, hw_orig, hw_resized = load_image(self, i)
  426. gb += self.imgs[i].nbytes
  427. pbar.desc = f'{prefix}Caching images ({gb / 1E9:.1f}GB {cache_images})'
  428. pbar.close()
  429. def cache_labels(self, path=Path('./labels.cache'), prefix=''):
  430. # Cache dataset labels, check images and read shapes
  431. x = {} # dict
  432. nm, nf, ne, nc, msgs = 0, 0, 0, 0, [] # number missing, found, empty, corrupt, messages
  433. desc = f"{prefix}Scanning '{path.parent / path.stem}' images and labels..."
  434. with Pool(NUM_THREADS) as pool:
  435. pbar = tqdm(pool.imap(verify_image_label, zip(self.img_files, self.label_files, repeat(prefix))),
  436. desc=desc, total=len(self.img_files))
  437. for im_file, lb, shape, segments, nm_f, nf_f, ne_f, nc_f, msg in pbar:
  438. nm += nm_f
  439. nf += nf_f
  440. ne += ne_f
  441. nc += nc_f
  442. if im_file:
  443. x[im_file] = [lb, shape, segments]
  444. if msg:
  445. msgs.append(msg)
  446. pbar.desc = f"{desc}{nf} found, {nm} missing, {ne} empty, {nc} corrupt"
  447. pbar.close()
  448. if msgs:
  449. LOGGER.info('\n'.join(msgs))
  450. if nf == 0:
  451. LOGGER.warning(f'{prefix}WARNING: No labels found in {path}. See {HELP_URL}')
  452. x['hash'] = get_hash(self.label_files + self.img_files)
  453. x['results'] = nf, nm, ne, nc, len(self.img_files)
  454. x['msgs'] = msgs # warnings
  455. x['version'] = self.cache_version # cache version
  456. try:
  457. np.save(path, x) # save cache for next time
  458. path.with_suffix('.cache.npy').rename(path) # remove .npy suffix
  459. LOGGER.info(f'{prefix}New cache created: {path}')
  460. except Exception as e:
  461. LOGGER.warning(f'{prefix}WARNING: Cache directory {path.parent} is not writeable: {e}') # not writeable
  462. return x
  463. def __len__(self):
  464. return len(self.img_files)
  465. # def __iter__(self):
  466. # self.count = -1
  467. # print('ran dataset iter')
  468. # #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF)
  469. # return self
  470. def __getitem__(self, index):
  471. index = self.indices[index] # linear, shuffled, or image_weights
  472. hyp = self.hyp
  473. mosaic = self.mosaic and random.random() < hyp['mosaic']
  474. if mosaic:
  475. # Load mosaic
  476. img, labels = self.load_mosaic(index)
  477. shapes = None
  478. # MixUp augmentation
  479. if random.random() < hyp['mixup']:
  480. img, labels = mixup(img, labels, *self.load_mosaic(random.randint(0, self.n - 1)))
  481. else:
  482. # Load image
  483. img, (h0, w0), (h, w) = self.load_image(index)
  484. # Letterbox
  485. shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size # final letterboxed shape
  486. img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment)
  487. shapes = (h0, w0), ((h / h0, w / w0), pad) # for COCO mAP rescaling
  488. labels = self.labels[index].copy()
  489. if labels.size: # normalized xywh to pixel xyxy format
  490. labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1])
  491. if self.augment:
  492. img, labels = random_perspective(img, labels,
  493. degrees=hyp['degrees'],
  494. translate=hyp['translate'],
  495. scale=hyp['scale'],
  496. shear=hyp['shear'],
  497. perspective=hyp['perspective'])
  498. nl = len(labels) # number of labels
  499. if nl:
  500. labels[:, 1:5] = xyxy2xywhn(labels[:, 1:5], w=img.shape[1], h=img.shape[0], clip=True, eps=1E-3)
  501. if self.augment:
  502. # Albumentations
  503. img, labels = self.albumentations(img, labels)
  504. nl = len(labels) # update after albumentations
  505. # HSV color-space
  506. augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v'])
  507. # Flip up-down
  508. if random.random() < hyp['flipud']:
  509. img = np.flipud(img)
  510. if nl:
  511. labels[:, 2] = 1 - labels[:, 2]
  512. # Flip left-right
  513. if random.random() < hyp['fliplr']:
  514. img = np.fliplr(img)
  515. if nl:
  516. labels[:, 1] = 1 - labels[:, 1]
  517. # Cutouts
  518. # labels = cutout(img, labels, p=0.5)
  519. # nl = len(labels) # update after cutout
  520. labels_out = torch.zeros((nl, 6))
  521. if nl:
  522. labels_out[:, 1:] = torch.from_numpy(labels)
  523. # Convert
  524. img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB
  525. img = np.ascontiguousarray(img)
  526. return torch.from_numpy(img), labels_out, self.img_files[index], shapes
  527. def load_image(self, i):
  528. # loads 1 image from dataset index 'i', returns (im, original hw, resized hw)
  529. im = self.imgs[i]
  530. if im is None: # not cached in RAM
  531. npy = self.img_npy[i]
  532. if npy and npy.exists(): # load npy
  533. im = np.load(npy)
  534. else: # read image
  535. f = self.img_files[i]
  536. im = cv2.imread(f) # BGR
  537. assert im is not None, f'Image Not Found {f}'
  538. h0, w0 = im.shape[:2] # orig hw
  539. r = self.img_size / max(h0, w0) # ratio
  540. if r != 1: # if sizes are not equal
  541. im = cv2.resize(im,
  542. (int(w0 * r), int(h0 * r)),
  543. interpolation=cv2.INTER_LINEAR if (self.augment or r > 1) else cv2.INTER_AREA)
  544. return im, (h0, w0), im.shape[:2] # im, hw_original, hw_resized
  545. else:
  546. return self.imgs[i], self.img_hw0[i], self.img_hw[i] # im, hw_original, hw_resized
  547. def load_mosaic(self, index):
  548. # YOLOv5 4-mosaic loader. Loads 1 image + 3 random images into a 4-image mosaic
  549. labels4, segments4 = [], []
  550. s = self.img_size
  551. yc, xc = (int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border) # mosaic center x, y
  552. indices = [index] + random.choices(self.indices, k=3) # 3 additional image indices
  553. random.shuffle(indices)
  554. for i, index in enumerate(indices):
  555. # Load image
  556. img, _, (h, w) = self.load_image(index)
  557. # place img in img4
  558. if i == 0: # top left
  559. img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles
  560. x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image)
  561. x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image)
  562. elif i == 1: # top right
  563. x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
  564. x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
  565. elif i == 2: # bottom left
  566. x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
  567. x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
  568. elif i == 3: # bottom right
  569. x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
  570. x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)
  571. img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax]
  572. padw = x1a - x1b
  573. padh = y1a - y1b
  574. # Labels
  575. labels, segments = self.labels[index].copy(), self.segments[index].copy()
  576. if labels.size:
  577. labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh) # normalized xywh to pixel xyxy format
  578. segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
  579. labels4.append(labels)
  580. segments4.extend(segments)
  581. # Concat/clip labels
  582. labels4 = np.concatenate(labels4, 0)
  583. for x in (labels4[:, 1:], *segments4):
  584. np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective()
  585. # img4, labels4 = replicate(img4, labels4) # replicate
  586. # Augment
  587. img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp['copy_paste'])
  588. img4, labels4 = random_perspective(img4, labels4, segments4,
  589. degrees=self.hyp['degrees'],
  590. translate=self.hyp['translate'],
  591. scale=self.hyp['scale'],
  592. shear=self.hyp['shear'],
  593. perspective=self.hyp['perspective'],
  594. border=self.mosaic_border) # border to remove
  595. return img4, labels4
  596. def load_mosaic9(self, index):
  597. # YOLOv5 9-mosaic loader. Loads 1 image + 8 random images into a 9-image mosaic
  598. labels9, segments9 = [], []
  599. s = self.img_size
  600. indices = [index] + random.choices(self.indices, k=8) # 8 additional image indices
  601. random.shuffle(indices)
  602. hp, wp = -1, -1 # height, width previous
  603. for i, index in enumerate(indices):
  604. # Load image
  605. img, _, (h, w) = self.load_image(index)
  606. # place img in img9
  607. if i == 0: # center
  608. img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles
  609. h0, w0 = h, w
  610. c = s, s, s + w, s + h # xmin, ymin, xmax, ymax (base) coordinates
  611. elif i == 1: # top
  612. c = s, s - h, s + w, s
  613. elif i == 2: # top right
  614. c = s + wp, s - h, s + wp + w, s
  615. elif i == 3: # right
  616. c = s + w0, s, s + w0 + w, s + h
  617. elif i == 4: # bottom right
  618. c = s + w0, s + hp, s + w0 + w, s + hp + h
  619. elif i == 5: # bottom
  620. c = s + w0 - w, s + h0, s + w0, s + h0 + h
  621. elif i == 6: # bottom left
  622. c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h
  623. elif i == 7: # left
  624. c = s - w, s + h0 - h, s, s + h0
  625. elif i == 8: # top left
  626. c = s - w, s + h0 - hp - h, s, s + h0 - hp
  627. padx, pady = c[:2]
  628. x1, y1, x2, y2 = (max(x, 0) for x in c) # allocate coords
  629. # Labels
  630. labels, segments = self.labels[index].copy(), self.segments[index].copy()
  631. if labels.size:
  632. labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padx, pady) # normalized xywh to pixel xyxy format
  633. segments = [xyn2xy(x, w, h, padx, pady) for x in segments]
  634. labels9.append(labels)
  635. segments9.extend(segments)
  636. # Image
  637. img9[y1:y2, x1:x2] = img[y1 - pady:, x1 - padx:] # img9[ymin:ymax, xmin:xmax]
  638. hp, wp = h, w # height, width previous
  639. # Offset
  640. yc, xc = (int(random.uniform(0, s)) for _ in self.mosaic_border) # mosaic center x, y
  641. img9 = img9[yc:yc + 2 * s, xc:xc + 2 * s]
  642. # Concat/clip labels
  643. labels9 = np.concatenate(labels9, 0)
  644. labels9[:, [1, 3]] -= xc
  645. labels9[:, [2, 4]] -= yc
  646. c = np.array([xc, yc]) # centers
  647. segments9 = [x - c for x in segments9]
  648. for x in (labels9[:, 1:], *segments9):
  649. np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective()
  650. # img9, labels9 = replicate(img9, labels9) # replicate
  651. # Augment
  652. img9, labels9 = random_perspective(img9, labels9, segments9,
  653. degrees=self.hyp['degrees'],
  654. translate=self.hyp['translate'],
  655. scale=self.hyp['scale'],
  656. shear=self.hyp['shear'],
  657. perspective=self.hyp['perspective'],
  658. border=self.mosaic_border) # border to remove
  659. return img9, labels9
  660. @staticmethod
  661. def collate_fn(batch):
  662. img, label, path, shapes = zip(*batch) # transposed
  663. for i, lb in enumerate(label):
  664. lb[:, 0] = i # add target image index for build_targets()
  665. return torch.stack(img, 0), torch.cat(label, 0), path, shapes
  666. @staticmethod
  667. def collate_fn4(batch):
  668. img, label, path, shapes = zip(*batch) # transposed
  669. n = len(shapes) // 4
  670. img4, label4, path4, shapes4 = [], [], path[:n], shapes[:n]
  671. ho = torch.tensor([[0.0, 0, 0, 1, 0, 0]])
  672. wo = torch.tensor([[0.0, 0, 1, 0, 0, 0]])
  673. s = torch.tensor([[1, 1, 0.5, 0.5, 0.5, 0.5]]) # scale
  674. for i in range(n): # zidane torch.zeros(16,3,720,1280) # BCHW
  675. i *= 4
  676. if random.random() < 0.5:
  677. im = F.interpolate(img[i].unsqueeze(0).float(), scale_factor=2.0, mode='bilinear', align_corners=False)[
  678. 0].type(img[i].type())
  679. lb = label[i]
  680. else:
  681. im = torch.cat((torch.cat((img[i], img[i + 1]), 1), torch.cat((img[i + 2], img[i + 3]), 1)), 2)
  682. lb = torch.cat((label[i], label[i + 1] + ho, label[i + 2] + wo, label[i + 3] + ho + wo), 0) * s
  683. img4.append(im)
  684. label4.append(lb)
  685. for i, lb in enumerate(label4):
  686. lb[:, 0] = i # add target image index for build_targets()
  687. return torch.stack(img4, 0), torch.cat(label4, 0), path4, shapes4
  688. # Ancillary functions --------------------------------------------------------------------------------------------------
  689. def create_folder(path='./new'):
  690. # Create folder
  691. if os.path.exists(path):
  692. shutil.rmtree(path) # delete output folder
  693. os.makedirs(path) # make new output folder
  694. def flatten_recursive(path=DATASETS_DIR / 'coco128'):
  695. # Flatten a recursive directory by bringing all files to top level
  696. new_path = Path(str(path) + '_flat')
  697. create_folder(new_path)
  698. for file in tqdm(glob.glob(str(Path(path)) + '/**/*.*', recursive=True)):
  699. shutil.copyfile(file, new_path / Path(file).name)
  700. def extract_boxes(path=DATASETS_DIR / 'coco128'): # from utils.datasets import *; extract_boxes()
  701. # Convert detection dataset into classification dataset, with one directory per class
  702. path = Path(path) # images dir
  703. shutil.rmtree(path / 'classifier') if (path / 'classifier').is_dir() else None # remove existing
  704. files = list(path.rglob('*.*'))
  705. n = len(files) # number of files
  706. for im_file in tqdm(files, total=n):
  707. if im_file.suffix[1:] in IMG_FORMATS:
  708. # image
  709. im = cv2.imread(str(im_file))[..., ::-1] # BGR to RGB
  710. h, w = im.shape[:2]
  711. # labels
  712. lb_file = Path(img2label_paths([str(im_file)])[0])
  713. if Path(lb_file).exists():
  714. with open(lb_file) as f:
  715. lb = np.array([x.split() for x in f.read().strip().splitlines()], dtype=np.float32) # labels
  716. for j, x in enumerate(lb):
  717. c = int(x[0]) # class
  718. f = (path / 'classifier') / f'{c}' / f'{path.stem}_{im_file.stem}_{j}.jpg' # new filename
  719. if not f.parent.is_dir():
  720. f.parent.mkdir(parents=True)
  721. b = x[1:] * [w, h, w, h] # box
  722. # b[2:] = b[2:].max() # rectangle to square
  723. b[2:] = b[2:] * 1.2 + 3 # pad
  724. b = xywh2xyxy(b.reshape(-1, 4)).ravel().astype(np.int)
  725. b[[0, 2]] = np.clip(b[[0, 2]], 0, w) # clip boxes outside of image
  726. b[[1, 3]] = np.clip(b[[1, 3]], 0, h)
  727. assert cv2.imwrite(str(f), im[b[1]:b[3], b[0]:b[2]]), f'box failure in {f}'
  728. def autosplit(path=DATASETS_DIR / 'coco128/images', weights=(0.9, 0.1, 0.0), annotated_only=False):
  729. """ Autosplit a dataset into train/val/test splits and save path/autosplit_*.txt files
  730. Usage: from utils.datasets import *; autosplit()
  731. Arguments
  732. path: Path to images directory
  733. weights: Train, val, test weights (list, tuple)
  734. annotated_only: Only use images with an annotated txt file
  735. """
  736. path = Path(path) # images dir
  737. files = sorted(x for x in path.rglob('*.*') if x.suffix[1:].lower() in IMG_FORMATS) # image files only
  738. n = len(files) # number of files
  739. random.seed(0) # for reproducibility
  740. indices = random.choices([0, 1, 2], weights=weights, k=n) # assign each image to a split
  741. txt = ['autosplit_train.txt', 'autosplit_val.txt', 'autosplit_test.txt'] # 3 txt files
  742. [(path.parent / x).unlink(missing_ok=True) for x in txt] # remove existing
  743. print(f'Autosplitting images from {path}' + ', using *.txt labeled images only' * annotated_only)
  744. for i, img in tqdm(zip(indices, files), total=n):
  745. if not annotated_only or Path(img2label_paths([str(img)])[0]).exists(): # check label
  746. with open(path.parent / txt[i], 'a') as f:
  747. f.write('./' + img.relative_to(path.parent).as_posix() + '\n') # add image to txt file
  748. def verify_image_label(args):
  749. # Verify one image-label pair
  750. im_file, lb_file, prefix = args
  751. nm, nf, ne, nc, msg, segments = 0, 0, 0, 0, '', [] # number (missing, found, empty, corrupt), message, segments
  752. try:
  753. # verify images
  754. im = Image.open(im_file)
  755. im.verify() # PIL verify
  756. shape = exif_size(im) # image size
  757. assert (shape[0] > 9) & (shape[1] > 9), f'image size {shape} <10 pixels'
  758. assert im.format.lower() in IMG_FORMATS, f'invalid image format {im.format}'
  759. if im.format.lower() in ('jpg', 'jpeg'):
  760. with open(im_file, 'rb') as f:
  761. f.seek(-2, 2)
  762. if f.read() != b'\xff\xd9': # corrupt JPEG
  763. ImageOps.exif_transpose(Image.open(im_file)).save(im_file, 'JPEG', subsampling=0, quality=100)
  764. msg = f'{prefix}WARNING: {im_file}: corrupt JPEG restored and saved'
  765. # verify labels
  766. if os.path.isfile(lb_file):
  767. nf = 1 # label found
  768. with open(lb_file) as f:
  769. lb = [x.split() for x in f.read().strip().splitlines() if len(x)]
  770. if any([len(x) > 8 for x in lb]): # is segment
  771. classes = np.array([x[0] for x in lb], dtype=np.float32)
  772. segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in lb] # (cls, xy1...)
  773. lb = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1) # (cls, xywh)
  774. lb = np.array(lb, dtype=np.float32)
  775. nl = len(lb)
  776. if nl:
  777. assert lb.shape[1] == 5, f'labels require 5 columns, {lb.shape[1]} columns detected'
  778. assert (lb >= 0).all(), f'negative label values {lb[lb < 0]}'
  779. assert (lb[:, 1:] <= 1).all(), f'non-normalized or out of bounds coordinates {lb[:, 1:][lb[:, 1:] > 1]}'
  780. _, i = np.unique(lb, axis=0, return_index=True)
  781. if len(i) < nl: # duplicate row check
  782. lb = lb[i] # remove duplicates
  783. if segments:
  784. segments = segments[i]
  785. msg = f'{prefix}WARNING: {im_file}: {nl - len(i)} duplicate labels removed'
  786. else:
  787. ne = 1 # label empty
  788. lb = np.zeros((0, 5), dtype=np.float32)
  789. else:
  790. nm = 1 # label missing
  791. lb = np.zeros((0, 5), dtype=np.float32)
  792. return im_file, lb, shape, segments, nm, nf, ne, nc, msg
  793. except Exception as e:
  794. nc = 1
  795. msg = f'{prefix}WARNING: {im_file}: ignoring corrupt image/label: {e}'
  796. return [None, None, None, None, nm, nf, ne, nc, msg]
  797. def dataset_stats(path='coco128.yaml', autodownload=False, verbose=False, profile=False, hub=False):
  798. """ Return dataset statistics dictionary with images and instances counts per split per class
  799. To run in parent directory: export PYTHONPATH="$PWD/yolov5"
  800. Usage1: from utils.datasets import *; dataset_stats('coco128.yaml', autodownload=True)
  801. Usage2: from utils.datasets import *; dataset_stats('path/to/coco128_with_yaml.zip')
  802. Arguments
  803. path: Path to data.yaml or data.zip (with data.yaml inside data.zip)
  804. autodownload: Attempt to download dataset if not found locally
  805. verbose: Print stats dictionary
  806. """
  807. def round_labels(labels):
  808. # Update labels to integer class and 6 decimal place floats
  809. return [[int(c), *(round(x, 4) for x in points)] for c, *points in labels]
  810. def unzip(path):
  811. # Unzip data.zip TODO: CONSTRAINT: path/to/abc.zip MUST unzip to 'path/to/abc/'
  812. if str(path).endswith('.zip'): # path is data.zip
  813. assert Path(path).is_file(), f'Error unzipping {path}, file not found'
  814. ZipFile(path).extractall(path=path.parent) # unzip
  815. dir = path.with_suffix('') # dataset directory == zip name
  816. return True, str(dir), next(dir.rglob('*.yaml')) # zipped, data_dir, yaml_path
  817. else: # path is data.yaml
  818. return False, None, path
  819. def hub_ops(f, max_dim=1920):
  820. # HUB ops for 1 image 'f': resize and save at reduced quality in /dataset-hub for web/app viewing
  821. f_new = im_dir / Path(f).name # dataset-hub image filename
  822. try: # use PIL
  823. im = Image.open(f)
  824. r = max_dim / max(im.height, im.width) # ratio
  825. if r < 1.0: # image too large
  826. im = im.resize((int(im.width * r), int(im.height * r)))
  827. im.save(f_new, 'JPEG', quality=75, optimize=True) # save
  828. except Exception as e: # use OpenCV
  829. print(f'WARNING: HUB ops PIL failure {f}: {e}')
  830. im = cv2.imread(f)
  831. im_height, im_width = im.shape[:2]
  832. r = max_dim / max(im_height, im_width) # ratio
  833. if r < 1.0: # image too large
  834. im = cv2.resize(im, (int(im_width * r), int(im_height * r)), interpolation=cv2.INTER_AREA)
  835. cv2.imwrite(str(f_new), im)
  836. zipped, data_dir, yaml_path = unzip(Path(path))
  837. with open(check_yaml(yaml_path), errors='ignore') as f:
  838. data = yaml.safe_load(f) # data dict
  839. if zipped:
  840. data['path'] = data_dir # TODO: should this be dir.resolve()?
  841. check_dataset(data, autodownload) # download dataset if missing
  842. hub_dir = Path(data['path'] + ('-hub' if hub else ''))
  843. stats = {'nc': data['nc'], 'names': data['names']} # statistics dictionary
  844. for split in 'train', 'val', 'test':
  845. if data.get(split) is None:
  846. stats[split] = None # i.e. no test set
  847. continue
  848. x = []
  849. dataset = LoadImagesAndLabels(data[split]) # load dataset
  850. for label in tqdm(dataset.labels, total=dataset.n, desc='Statistics'):
  851. x.append(np.bincount(label[:, 0].astype(int), minlength=data['nc']))
  852. x = np.array(x) # shape(128x80)
  853. stats[split] = {'instance_stats': {'total': int(x.sum()), 'per_class': x.sum(0).tolist()},
  854. 'image_stats': {'total': dataset.n, 'unlabelled': int(np.all(x == 0, 1).sum()),
  855. 'per_class': (x > 0).sum(0).tolist()},
  856. 'labels': [{str(Path(k).name): round_labels(v.tolist())} for k, v in
  857. zip(dataset.img_files, dataset.labels)]}
  858. if hub:
  859. im_dir = hub_dir / 'images'
  860. im_dir.mkdir(parents=True, exist_ok=True)
  861. for _ in tqdm(ThreadPool(NUM_THREADS).imap(hub_ops, dataset.img_files), total=dataset.n, desc='HUB Ops'):
  862. pass
  863. # Profile
  864. stats_path = hub_dir / 'stats.json'
  865. if profile:
  866. for _ in range(1):
  867. file = stats_path.with_suffix('.npy')
  868. t1 = time.time()
  869. np.save(file, stats)
  870. t2 = time.time()
  871. x = np.load(file, allow_pickle=True)
  872. print(f'stats.npy times: {time.time() - t2:.3f}s read, {t2 - t1:.3f}s write')
  873. file = stats_path.with_suffix('.json')
  874. t1 = time.time()
  875. with open(file, 'w') as f:
  876. json.dump(stats, f) # save stats *.json
  877. t2 = time.time()
  878. with open(file) as f:
  879. x = json.load(f) # load hyps dict
  880. print(f'stats.json times: {time.time() - t2:.3f}s read, {t2 - t1:.3f}s write')
  881. # Save, print and return
  882. if hub:
  883. print(f'Saving {stats_path.resolve()}...')
  884. with open(stats_path, 'w') as f:
  885. json.dump(stats, f) # save stats.json
  886. if verbose:
  887. print(json.dumps(stats, indent=2, sort_keys=False))
  888. return stats