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.

627 line
32KB

  1. # YOLOv5 🚀 by Ultralytics, GPL-3.0 license
  2. """
  3. Train a YOLOv5 model on a custom dataset
  4. Usage:
  5. $ python path/to/train.py --data coco128.yaml --weights yolov5s.pt --img 640
  6. """
  7. import argparse
  8. import math
  9. import os
  10. import random
  11. import sys
  12. import time
  13. from copy import deepcopy
  14. from datetime import datetime
  15. from pathlib import Path
  16. import numpy as np
  17. import torch
  18. import torch.distributed as dist
  19. import torch.nn as nn
  20. import yaml
  21. from torch.cuda import amp
  22. from torch.nn.parallel import DistributedDataParallel as DDP
  23. from torch.optim import SGD, Adam, lr_scheduler
  24. from tqdm import tqdm
  25. FILE = Path(__file__).resolve()
  26. ROOT = FILE.parents[0] # YOLOv5 root directory
  27. if str(ROOT) not in sys.path:
  28. sys.path.append(str(ROOT)) # add ROOT to PATH
  29. ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative
  30. import val # for end-of-epoch mAP
  31. from models.experimental import attempt_load
  32. from models.yolo import Model
  33. from utils.autoanchor import check_anchors
  34. from utils.autobatch import check_train_batch_size
  35. from utils.callbacks import Callbacks
  36. from utils.datasets import create_dataloader
  37. from utils.downloads import attempt_download
  38. from utils.general import (LOGGER, check_dataset, check_file, check_git_status, check_img_size, check_requirements,
  39. check_suffix, check_yaml, colorstr, get_latest_run, increment_path, init_seeds,
  40. intersect_dicts, labels_to_class_weights, labels_to_image_weights, methods, one_cycle,
  41. print_args, print_mutation, strip_optimizer)
  42. from utils.loggers import Loggers
  43. from utils.loggers.wandb.wandb_utils import check_wandb_resume
  44. from utils.loss import ComputeLoss
  45. from utils.metrics import fitness
  46. from utils.plots import plot_evolve, plot_labels
  47. from utils.torch_utils import EarlyStopping, ModelEMA, de_parallel, select_device, torch_distributed_zero_first
  48. LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1)) # https://pytorch.org/docs/stable/elastic/run.html
  49. RANK = int(os.getenv('RANK', -1))
  50. WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
  51. def train(hyp, # path/to/hyp.yaml or hyp dictionary
  52. opt,
  53. device,
  54. callbacks
  55. ):
  56. save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze, = \
  57. Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
  58. opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
  59. # Directories
  60. w = save_dir / 'weights' # weights dir
  61. (w.parent if evolve else w).mkdir(parents=True, exist_ok=True) # make dir
  62. last, best = w / 'last.pt', w / 'best.pt'
  63. # Hyperparameters
  64. if isinstance(hyp, str):
  65. with open(hyp, errors='ignore') as f:
  66. hyp = yaml.safe_load(f) # load hyps dict
  67. LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
  68. # Save run settings
  69. if not evolve:
  70. with open(save_dir / 'hyp.yaml', 'w') as f:
  71. yaml.safe_dump(hyp, f, sort_keys=False)
  72. with open(save_dir / 'opt.yaml', 'w') as f:
  73. yaml.safe_dump(vars(opt), f, sort_keys=False)
  74. # Loggers
  75. data_dict = None
  76. if RANK in [-1, 0]:
  77. loggers = Loggers(save_dir, weights, opt, hyp, LOGGER) # loggers instance
  78. if loggers.wandb:
  79. data_dict = loggers.wandb.data_dict
  80. if resume:
  81. weights, epochs, hyp = opt.weights, opt.epochs, opt.hyp
  82. # Register actions
  83. for k in methods(loggers):
  84. callbacks.register_action(k, callback=getattr(loggers, k))
  85. # Config
  86. plots = not evolve # create plots
  87. cuda = device.type != 'cpu'
  88. init_seeds(1 + RANK)
  89. with torch_distributed_zero_first(LOCAL_RANK):
  90. data_dict = data_dict or check_dataset(data) # check if None
  91. train_path, val_path = data_dict['train'], data_dict['val']
  92. nc = 1 if single_cls else int(data_dict['nc']) # number of classes
  93. names = ['item'] if single_cls and len(data_dict['names']) != 1 else data_dict['names'] # class names
  94. assert len(names) == nc, f'{len(names)} names found for nc={nc} dataset in {data}' # check
  95. is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt') # COCO dataset
  96. # Model
  97. check_suffix(weights, '.pt') # check weights
  98. pretrained = weights.endswith('.pt')
  99. if pretrained:
  100. with torch_distributed_zero_first(LOCAL_RANK):
  101. weights = attempt_download(weights) # download if not found locally
  102. ckpt = torch.load(weights, map_location=device) # load checkpoint
  103. model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create
  104. exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else [] # exclude keys
  105. csd = ckpt['model'].float().state_dict() # checkpoint state_dict as FP32
  106. csd = intersect_dicts(csd, model.state_dict(), exclude=exclude) # intersect
  107. model.load_state_dict(csd, strict=False) # load
  108. LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}') # report
  109. else:
  110. model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create
  111. # Freeze
  112. freeze = [f'model.{x}.' for x in range(freeze)] # layers to freeze
  113. for k, v in model.named_parameters():
  114. v.requires_grad = True # train all layers
  115. if any(x in k for x in freeze):
  116. LOGGER.info(f'freezing {k}')
  117. v.requires_grad = False
  118. # Image size
  119. gs = max(int(model.stride.max()), 32) # grid size (max stride)
  120. imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2) # verify imgsz is gs-multiple
  121. # Batch size
  122. if RANK == -1 and batch_size == -1: # single-GPU only, estimate best batch size
  123. batch_size = check_train_batch_size(model, imgsz)
  124. # Optimizer
  125. nbs = 64 # nominal batch size
  126. accumulate = max(round(nbs / batch_size), 1) # accumulate loss before optimizing
  127. hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay
  128. LOGGER.info(f"Scaled weight_decay = {hyp['weight_decay']}")
  129. g0, g1, g2 = [], [], [] # optimizer parameter groups
  130. for v in model.modules():
  131. if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): # bias
  132. g2.append(v.bias)
  133. if isinstance(v, nn.BatchNorm2d): # weight (no decay)
  134. g0.append(v.weight)
  135. elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): # weight (with decay)
  136. g1.append(v.weight)
  137. if opt.adam:
  138. optimizer = Adam(g0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum
  139. else:
  140. optimizer = SGD(g0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
  141. optimizer.add_param_group({'params': g1, 'weight_decay': hyp['weight_decay']}) # add g1 with weight_decay
  142. optimizer.add_param_group({'params': g2}) # add g2 (biases)
  143. LOGGER.info(f"{colorstr('optimizer:')} {type(optimizer).__name__} with parameter groups "
  144. f"{len(g0)} weight, {len(g1)} weight (no decay), {len(g2)} bias")
  145. del g0, g1, g2
  146. # Scheduler
  147. if opt.linear_lr:
  148. lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp['lrf'] # linear
  149. else:
  150. lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf']
  151. scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs)
  152. # EMA
  153. ema = ModelEMA(model) if RANK in [-1, 0] else None
  154. # Resume
  155. start_epoch, best_fitness = 0, 0.0
  156. if pretrained:
  157. # Optimizer
  158. if ckpt['optimizer'] is not None:
  159. optimizer.load_state_dict(ckpt['optimizer'])
  160. best_fitness = ckpt['best_fitness']
  161. # EMA
  162. if ema and ckpt.get('ema'):
  163. ema.ema.load_state_dict(ckpt['ema'].float().state_dict())
  164. ema.updates = ckpt['updates']
  165. # Epochs
  166. start_epoch = ckpt['epoch'] + 1
  167. if resume:
  168. assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.'
  169. if epochs < start_epoch:
  170. LOGGER.info(f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs.")
  171. epochs += ckpt['epoch'] # finetune additional epochs
  172. del ckpt, csd
  173. # DP mode
  174. if cuda and RANK == -1 and torch.cuda.device_count() > 1:
  175. LOGGER.warning('WARNING: DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n'
  176. 'See Multi-GPU Tutorial at https://github.com/ultralytics/yolov5/issues/475 to get started.')
  177. model = torch.nn.DataParallel(model)
  178. # SyncBatchNorm
  179. if opt.sync_bn and cuda and RANK != -1:
  180. model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
  181. LOGGER.info('Using SyncBatchNorm()')
  182. # Trainloader
  183. train_loader, dataset = create_dataloader(train_path, imgsz, batch_size // WORLD_SIZE, gs, single_cls,
  184. hyp=hyp, augment=True, cache=opt.cache, rect=opt.rect, rank=LOCAL_RANK,
  185. workers=workers, image_weights=opt.image_weights, quad=opt.quad,
  186. prefix=colorstr('train: '), shuffle=True)
  187. mlc = int(np.concatenate(dataset.labels, 0)[:, 0].max()) # max label class
  188. nb = len(train_loader) # number of batches
  189. assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
  190. # Process 0
  191. if RANK in [-1, 0]:
  192. val_loader = create_dataloader(val_path, imgsz, batch_size // WORLD_SIZE * 2, gs, single_cls,
  193. hyp=hyp, cache=None if noval else opt.cache, rect=True, rank=-1,
  194. workers=workers, pad=0.5,
  195. prefix=colorstr('val: '))[0]
  196. if not resume:
  197. labels = np.concatenate(dataset.labels, 0)
  198. # c = torch.tensor(labels[:, 0]) # classes
  199. # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency
  200. # model._initialize_biases(cf.to(device))
  201. if plots:
  202. plot_labels(labels, names, save_dir)
  203. # Anchors
  204. if not opt.noautoanchor:
  205. check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
  206. model.half().float() # pre-reduce anchor precision
  207. callbacks.run('on_pretrain_routine_end')
  208. # DDP mode
  209. if cuda and RANK != -1:
  210. model = DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK)
  211. # Model attributes
  212. nl = de_parallel(model).model[-1].nl # number of detection layers (to scale hyps)
  213. hyp['box'] *= 3 / nl # scale to layers
  214. hyp['cls'] *= nc / 80 * 3 / nl # scale to classes and layers
  215. hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl # scale to image size and layers
  216. hyp['label_smoothing'] = opt.label_smoothing
  217. model.nc = nc # attach number of classes to model
  218. model.hyp = hyp # attach hyperparameters to model
  219. model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc # attach class weights
  220. model.names = names
  221. # Start training
  222. t0 = time.time()
  223. nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations)
  224. # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
  225. last_opt_step = -1
  226. maps = np.zeros(nc) # mAP per class
  227. results = (0, 0, 0, 0, 0, 0, 0) # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)
  228. scheduler.last_epoch = start_epoch - 1 # do not move
  229. scaler = amp.GradScaler(enabled=cuda)
  230. stopper = EarlyStopping(patience=opt.patience)
  231. compute_loss = ComputeLoss(model) # init loss class
  232. LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
  233. f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
  234. f"Logging results to {colorstr('bold', save_dir)}\n"
  235. f'Starting training for {epochs} epochs...')
  236. for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
  237. model.train()
  238. # Update image weights (optional, single-GPU only)
  239. if opt.image_weights:
  240. cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc # class weights
  241. iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights
  242. dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx
  243. # Update mosaic border (optional)
  244. # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
  245. # dataset.mosaic_border = [b - imgsz, -b] # height, width borders
  246. mloss = torch.zeros(3, device=device) # mean losses
  247. if RANK != -1:
  248. train_loader.sampler.set_epoch(epoch)
  249. pbar = enumerate(train_loader)
  250. LOGGER.info(('\n' + '%10s' * 7) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'labels', 'img_size'))
  251. if RANK in [-1, 0]:
  252. pbar = tqdm(pbar, total=nb, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar
  253. optimizer.zero_grad()
  254. for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------
  255. ni = i + nb * epoch # number integrated batches (since train start)
  256. imgs = imgs.to(device, non_blocking=True).float() / 255 # uint8 to float32, 0-255 to 0.0-1.0
  257. # Warmup
  258. if ni <= nw:
  259. xi = [0, nw] # x interp
  260. # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
  261. accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
  262. for j, x in enumerate(optimizer.param_groups):
  263. # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
  264. x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
  265. if 'momentum' in x:
  266. x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
  267. # Multi-scale
  268. if opt.multi_scale:
  269. sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size
  270. sf = sz / max(imgs.shape[2:]) # scale factor
  271. if sf != 1:
  272. ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple)
  273. imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
  274. # Forward
  275. with amp.autocast(enabled=cuda):
  276. pred = model(imgs) # forward
  277. loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size
  278. if RANK != -1:
  279. loss *= WORLD_SIZE # gradient averaged between devices in DDP mode
  280. if opt.quad:
  281. loss *= 4.
  282. # Backward
  283. scaler.scale(loss).backward()
  284. # Optimize
  285. if ni - last_opt_step >= accumulate:
  286. scaler.step(optimizer) # optimizer.step
  287. scaler.update()
  288. optimizer.zero_grad()
  289. if ema:
  290. ema.update(model)
  291. last_opt_step = ni
  292. # Log
  293. if RANK in [-1, 0]:
  294. mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
  295. mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G' # (GB)
  296. pbar.set_description(('%10s' * 2 + '%10.4g' * 5) % (
  297. f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
  298. callbacks.run('on_train_batch_end', ni, model, imgs, targets, paths, plots, opt.sync_bn)
  299. # end batch ------------------------------------------------------------------------------------------------
  300. # Scheduler
  301. lr = [x['lr'] for x in optimizer.param_groups] # for loggers
  302. scheduler.step()
  303. if RANK in [-1, 0]:
  304. # mAP
  305. callbacks.run('on_train_epoch_end', epoch=epoch)
  306. ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
  307. final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
  308. if not noval or final_epoch: # Calculate mAP
  309. results, maps, _ = val.run(data_dict,
  310. batch_size=batch_size // WORLD_SIZE * 2,
  311. imgsz=imgsz,
  312. model=ema.ema,
  313. single_cls=single_cls,
  314. dataloader=val_loader,
  315. save_dir=save_dir,
  316. plots=False,
  317. callbacks=callbacks,
  318. compute_loss=compute_loss)
  319. # Update best mAP
  320. fi = fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
  321. if fi > best_fitness:
  322. best_fitness = fi
  323. log_vals = list(mloss) + list(results) + lr
  324. callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
  325. # Save model
  326. if (not nosave) or (final_epoch and not evolve): # if save
  327. ckpt = {'epoch': epoch,
  328. 'best_fitness': best_fitness,
  329. 'model': deepcopy(de_parallel(model)).half(),
  330. 'ema': deepcopy(ema.ema).half(),
  331. 'updates': ema.updates,
  332. 'optimizer': optimizer.state_dict(),
  333. 'wandb_id': loggers.wandb.wandb_run.id if loggers.wandb else None,
  334. 'date': datetime.now().isoformat()}
  335. # Save last, best and delete
  336. torch.save(ckpt, last)
  337. if best_fitness == fi:
  338. torch.save(ckpt, best)
  339. if (epoch > 0) and (opt.save_period > 0) and (epoch % opt.save_period == 0):
  340. torch.save(ckpt, w / f'epoch{epoch}.pt')
  341. del ckpt
  342. callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
  343. # Stop Single-GPU
  344. if RANK == -1 and stopper(epoch=epoch, fitness=fi):
  345. break
  346. # Stop DDP TODO: known issues shttps://github.com/ultralytics/yolov5/pull/4576
  347. # stop = stopper(epoch=epoch, fitness=fi)
  348. # if RANK == 0:
  349. # dist.broadcast_object_list([stop], 0) # broadcast 'stop' to all ranks
  350. # Stop DPP
  351. # with torch_distributed_zero_first(RANK):
  352. # if stop:
  353. # break # must break all DDP ranks
  354. # end epoch ----------------------------------------------------------------------------------------------------
  355. # end training -----------------------------------------------------------------------------------------------------
  356. if RANK in [-1, 0]:
  357. LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
  358. for f in last, best:
  359. if f.exists():
  360. strip_optimizer(f) # strip optimizers
  361. if f is best:
  362. LOGGER.info(f'\nValidating {f}...')
  363. results, _, _ = val.run(data_dict,
  364. batch_size=batch_size // WORLD_SIZE * 2,
  365. imgsz=imgsz,
  366. model=attempt_load(f, device).half(),
  367. iou_thres=0.65 if is_coco else 0.60, # best pycocotools results at 0.65
  368. single_cls=single_cls,
  369. dataloader=val_loader,
  370. save_dir=save_dir,
  371. save_json=is_coco,
  372. verbose=True,
  373. plots=True,
  374. callbacks=callbacks,
  375. compute_loss=compute_loss) # val best model with plots
  376. if is_coco:
  377. callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
  378. callbacks.run('on_train_end', last, best, plots, epoch, results)
  379. LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")
  380. torch.cuda.empty_cache()
  381. return results
  382. def parse_opt(known=False):
  383. parser = argparse.ArgumentParser()
  384. parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='initial weights path')
  385. parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
  386. parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
  387. parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch.yaml', help='hyperparameters path')
  388. parser.add_argument('--epochs', type=int, default=300)
  389. parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
  390. parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
  391. parser.add_argument('--rect', action='store_true', help='rectangular training')
  392. parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
  393. parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
  394. parser.add_argument('--noval', action='store_true', help='only validate final epoch')
  395. parser.add_argument('--noautoanchor', action='store_true', help='disable autoanchor check')
  396. parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
  397. parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
  398. parser.add_argument('--cache', type=str, nargs='?', const='ram', help='--cache images in "ram" (default) or "disk"')
  399. parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
  400. parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
  401. parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
  402. parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
  403. parser.add_argument('--adam', action='store_true', help='use torch.optim.Adam() optimizer')
  404. parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
  405. parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
  406. parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
  407. parser.add_argument('--name', default='exp', help='save to project/name')
  408. parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
  409. parser.add_argument('--quad', action='store_true', help='quad dataloader')
  410. parser.add_argument('--linear-lr', action='store_true', help='linear LR')
  411. parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
  412. parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
  413. parser.add_argument('--freeze', type=int, default=0, help='Number of layers to freeze. backbone=10, all=24')
  414. parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
  415. parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify')
  416. # Weights & Biases arguments
  417. parser.add_argument('--entity', default=None, help='W&B: Entity')
  418. parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='W&B: Upload data, "val" option')
  419. parser.add_argument('--bbox_interval', type=int, default=-1, help='W&B: Set bounding-box image logging interval')
  420. parser.add_argument('--artifact_alias', type=str, default='latest', help='W&B: Version of dataset artifact to use')
  421. opt = parser.parse_known_args()[0] if known else parser.parse_args()
  422. return opt
  423. def main(opt, callbacks=Callbacks()):
  424. # Checks
  425. if RANK in [-1, 0]:
  426. print_args(FILE.stem, opt)
  427. check_git_status()
  428. check_requirements(exclude=['thop'])
  429. # Resume
  430. if opt.resume and not check_wandb_resume(opt) and not opt.evolve: # resume an interrupted run
  431. ckpt = opt.resume if isinstance(opt.resume, str) else get_latest_run() # specified or most recent path
  432. assert os.path.isfile(ckpt), 'ERROR: --resume checkpoint does not exist'
  433. with open(Path(ckpt).parent.parent / 'opt.yaml', errors='ignore') as f:
  434. opt = argparse.Namespace(**yaml.safe_load(f)) # replace
  435. opt.cfg, opt.weights, opt.resume = '', ckpt, True # reinstate
  436. LOGGER.info(f'Resuming training from {ckpt}')
  437. else:
  438. opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
  439. check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project) # checks
  440. assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
  441. if opt.evolve:
  442. opt.project = str(ROOT / 'runs/evolve')
  443. opt.exist_ok, opt.resume = opt.resume, False # pass resume to exist_ok and disable resume
  444. opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
  445. # DDP mode
  446. device = select_device(opt.device, batch_size=opt.batch_size)
  447. if LOCAL_RANK != -1:
  448. assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
  449. assert opt.batch_size % WORLD_SIZE == 0, '--batch-size must be multiple of CUDA device count'
  450. assert not opt.image_weights, '--image-weights argument is not compatible with DDP training'
  451. assert not opt.evolve, '--evolve argument is not compatible with DDP training'
  452. torch.cuda.set_device(LOCAL_RANK)
  453. device = torch.device('cuda', LOCAL_RANK)
  454. dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
  455. # Train
  456. if not opt.evolve:
  457. train(opt.hyp, opt, device, callbacks)
  458. if WORLD_SIZE > 1 and RANK == 0:
  459. LOGGER.info('Destroying process group... ')
  460. dist.destroy_process_group()
  461. # Evolve hyperparameters (optional)
  462. else:
  463. # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
  464. meta = {'lr0': (1, 1e-5, 1e-1), # initial learning rate (SGD=1E-2, Adam=1E-3)
  465. 'lrf': (1, 0.01, 1.0), # final OneCycleLR learning rate (lr0 * lrf)
  466. 'momentum': (0.3, 0.6, 0.98), # SGD momentum/Adam beta1
  467. 'weight_decay': (1, 0.0, 0.001), # optimizer weight decay
  468. 'warmup_epochs': (1, 0.0, 5.0), # warmup epochs (fractions ok)
  469. 'warmup_momentum': (1, 0.0, 0.95), # warmup initial momentum
  470. 'warmup_bias_lr': (1, 0.0, 0.2), # warmup initial bias lr
  471. 'box': (1, 0.02, 0.2), # box loss gain
  472. 'cls': (1, 0.2, 4.0), # cls loss gain
  473. 'cls_pw': (1, 0.5, 2.0), # cls BCELoss positive_weight
  474. 'obj': (1, 0.2, 4.0), # obj loss gain (scale with pixels)
  475. 'obj_pw': (1, 0.5, 2.0), # obj BCELoss positive_weight
  476. 'iou_t': (0, 0.1, 0.7), # IoU training threshold
  477. 'anchor_t': (1, 2.0, 8.0), # anchor-multiple threshold
  478. 'anchors': (2, 2.0, 10.0), # anchors per output grid (0 to ignore)
  479. 'fl_gamma': (0, 0.0, 2.0), # focal loss gamma (efficientDet default gamma=1.5)
  480. 'hsv_h': (1, 0.0, 0.1), # image HSV-Hue augmentation (fraction)
  481. 'hsv_s': (1, 0.0, 0.9), # image HSV-Saturation augmentation (fraction)
  482. 'hsv_v': (1, 0.0, 0.9), # image HSV-Value augmentation (fraction)
  483. 'degrees': (1, 0.0, 45.0), # image rotation (+/- deg)
  484. 'translate': (1, 0.0, 0.9), # image translation (+/- fraction)
  485. 'scale': (1, 0.0, 0.9), # image scale (+/- gain)
  486. 'shear': (1, 0.0, 10.0), # image shear (+/- deg)
  487. 'perspective': (0, 0.0, 0.001), # image perspective (+/- fraction), range 0-0.001
  488. 'flipud': (1, 0.0, 1.0), # image flip up-down (probability)
  489. 'fliplr': (0, 0.0, 1.0), # image flip left-right (probability)
  490. 'mosaic': (1, 0.0, 1.0), # image mixup (probability)
  491. 'mixup': (1, 0.0, 1.0), # image mixup (probability)
  492. 'copy_paste': (1, 0.0, 1.0)} # segment copy-paste (probability)
  493. with open(opt.hyp, errors='ignore') as f:
  494. hyp = yaml.safe_load(f) # load hyps dict
  495. if 'anchors' not in hyp: # anchors commented in hyp.yaml
  496. hyp['anchors'] = 3
  497. opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir) # only val/save final epoch
  498. # ei = [isinstance(x, (int, float)) for x in hyp.values()] # evolvable indices
  499. evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
  500. if opt.bucket:
  501. os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {save_dir}') # download evolve.csv if exists
  502. for _ in range(opt.evolve): # generations to evolve
  503. if evolve_csv.exists(): # if evolve.csv exists: select best hyps and mutate
  504. # Select parent(s)
  505. parent = 'single' # parent selection method: 'single' or 'weighted'
  506. x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
  507. n = min(5, len(x)) # number of previous results to consider
  508. x = x[np.argsort(-fitness(x))][:n] # top n mutations
  509. w = fitness(x) - fitness(x).min() + 1E-6 # weights (sum > 0)
  510. if parent == 'single' or len(x) == 1:
  511. # x = x[random.randint(0, n - 1)] # random selection
  512. x = x[random.choices(range(n), weights=w)[0]] # weighted selection
  513. elif parent == 'weighted':
  514. x = (x * w.reshape(n, 1)).sum(0) / w.sum() # weighted combination
  515. # Mutate
  516. mp, s = 0.8, 0.2 # mutation probability, sigma
  517. npr = np.random
  518. npr.seed(int(time.time()))
  519. g = np.array([meta[k][0] for k in hyp.keys()]) # gains 0-1
  520. ng = len(meta)
  521. v = np.ones(ng)
  522. while all(v == 1): # mutate until a change occurs (prevent duplicates)
  523. v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
  524. for i, k in enumerate(hyp.keys()): # plt.hist(v.ravel(), 300)
  525. hyp[k] = float(x[i + 7] * v[i]) # mutate
  526. # Constrain to limits
  527. for k, v in meta.items():
  528. hyp[k] = max(hyp[k], v[1]) # lower limit
  529. hyp[k] = min(hyp[k], v[2]) # upper limit
  530. hyp[k] = round(hyp[k], 5) # significant digits
  531. # Train mutation
  532. results = train(hyp.copy(), opt, device, callbacks)
  533. # Write mutation results
  534. print_mutation(results, hyp.copy(), save_dir, opt.bucket)
  535. # Plot results
  536. plot_evolve(evolve_csv)
  537. LOGGER.info(f'Hyperparameter evolution finished\n'
  538. f"Results saved to {colorstr('bold', save_dir)}\n"
  539. f'Use best hyperparameters example: $ python train.py --hyp {evolve_yaml}')
  540. def run(**kwargs):
  541. # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolov5m.pt')
  542. opt = parse_opt(True)
  543. for k, v in kwargs.items():
  544. setattr(opt, k, v)
  545. main(opt)
  546. if __name__ == "__main__":
  547. opt = parse_opt()
  548. main(opt)