承接上一篇博客：https://blog.csdn.net/qq_38964360/article/details/128728145?spm=1001.2014.3001.5501
今天记录一下yolov8模型训练的调试过程。
在工程里创建训练脚本python_example.py
代码如下：
# filename: python_example.py
# dir: yolov8/python_example.py
from ultralytics import YOLO# build a new model from scratch
model = YOLO("yolov8/ultralytics/models/v8/yolov8m.yaml")
# train the model
results = model.train(data="yolov8/ultralytics/yolo/data/datasets/wider_face.yaml", epochs=100)以上示例首先是利用yolov8m.yaml文件初始化了YOLO类，
model = YOLO("yolov8/ultralytics/models/v8/yolov8m.yaml")接下来可以看看YOLO的__init__，
# dir: yolov8/ultralytics/yolo/engine/model.py
class YOLO:def __init__(self, model='yolov8n.yaml', type="v8") -> None:"""Initializes the YOLO object.Args:model (str, Path): model to load or createtype (str): Type/version of models to use. Defaults to "v8"."""self.type = typeself.ModelClass = None # model classself.TrainerClass = None # trainer classself.ValidatorClass = None # validator classself.PredictorClass = None # predictor classself.model = None # model objectself.trainer = None # trainer objectself.task = None # task typeself.ckpt = None # if loaded from *.ptself.cfg = None # if loaded from *.yamlself.ckpt_path = Noneself.overrides = {} # overrides for trainer object# Load or create new YOLO model{'.pt': self._load, '.yaml': self._new}[Path(model).suffix](model)这部分的重点是最后一句代码：
{'.pt': self._load, '.yaml': self._new}[Path(model).suffix](model)根据我们输入的model参数(yolov8m.yaml)的后缀，代码将跳到self._new中，根据yaml文件定义模型：
def _new(self, cfg: str, verbose=True):cfg = check_yaml(cfg) # cfg='yolov8/ultralytics/models/v8/yolov8m.yaml'cfg_dict = yaml_load(cfg, append_filename=True) # model dict'''cfg_dict={'nc': 80, 'depth_multiple': 0.33, 'width_multiple': 0.25, 'backbone': [[...], [...], [...], [...], [...], [...], [...], [...], [...], ...], 'head': [[...], [...], [...], [...], [...], [...], [...], [...], [...], ...], 'yaml_file': 'yolov8/ultralytics/models/v8/yolov8m.yaml','ch': 3}'''self.task = guess_task_from_head(cfg_dict["head"][-1][-2]) # self.task='detect'self.ModelClass, self.TrainerClass, self.ValidatorClass, self.PredictorClass = \self._guess_ops_from_task(self.task) # 根据task定义ModelClass、TrainerClass等, self.TrainerClass=DetectionTrainer()self.model = self.ModelClass(cfg_dict, verbose=verbose) # self.model=DetectionModel(), class DetectionModel的定义在yolov8/ultralytics/nn/tasks.pyself.cfg = cfg # self.cfg='yolov8/ultralytics/models/v8/yolov8m.yaml'def _guess_ops_from_task(self, task): # task='detect'model_class, train_lit, val_lit, pred_lit = MODEL_MAP[task] # model_class="<class 'ultralytics.nn.tasks.DetectionModel'>"# warning: eval is unsafe. Use with cautiontrainer_class = eval(train_lit.replace("TYPE", f"{self.type}")) # trainer_class="<class 'ultralytics.yolo.v8.detect.train.DetectionTrainer'>"validator_class = eval(val_lit.replace("TYPE", f"{self.type}")) # validator_class="<class 'ultralytics.yolo.v8.detect.val.DetectionValidator'>"predictor_class = eval(pred_lit.replace("TYPE", f"{self.type}")) # predictor_class="<class 'ultralytics.yolo.v8.detect.predict.DetectionPredictor'>"return model_class, trainer_class, validator_class, predictor_class实例化完模型(YOLO)后，就可以开始训练模型了，
results = model.train(data="yolov8/ultralytics/yolo/data/datasets/wider_face.yaml", epochs=100)跳转到class YOLO中的train，
def train(self, **kwargs):"""Trains the model on a given dataset.Args:**kwargs (Any): Any number of arguments representing the training configuration. List of all args can be found in 'config' section.You can pass all arguments as a yaml file in `cfg`. Other args are ignored if `cfg` file is passed"""overrides = self.overrides.copy() # overrides={}overrides.update(kwargs) # overrides={'data': 'wider_face.yaml', 'epochs': 100}if kwargs.get("cfg"):LOGGER.info(f"cfg file passed. Overriding default params with {kwargs['cfg']}.")overrides = yaml_load(check_yaml(kwargs["cfg"]), append_filename=True)overrides["task"] = self.taskoverrides["mode"] = "train" # overrides={'data': 'wider_face.yaml', 'epochs': 100, 'task': 'detect', 'mode': 'train'}if not overrides.get("data"):raise AttributeError("dataset not provided! Please define `data` in config.yaml or pass as an argument.")if overrides.get("resume"):overrides["resume"] = self.ckpt_pathself.trainer = self.TrainerClass(overrides=overrides)if not overrides.get("resume"): # manually set model only if not resumingself.trainer.model = self.trainer.get_model(weights=self.model if self.ckpt else None, cfg=self.model.yaml) # 如果有ckpt, 则直接加载; 没有则根据yolov8n.yaml新建一个模型self.model = self.trainer.modelself.trainer.train()# update model and configs after trainingself.model, _ = attempt_load_one_weight(str(self.trainer.best))self.overrides = self.model.args上述代码的重点是self.trainer.train()，self.trainer是<class 'ultralytics.yolo.v8.detect.train.DetectionTrainer'>，而DetectionTrainer继承了class BaseTrainer(dir: 'yolov8/ultralytics/yolo/engine/trainer.py')。利用overrides来初始化BaseTrainer，
self.trainer = self.TrainerClass(overrides=overrides)class BaseTrainer:"""BaseTrainerA base class for creating trainers.Attributes:args (OmegaConf): Configuration for the trainer.check_resume (method): Method to check if training should be resumed from a saved checkpoint.console (logging.Logger): Logger instance.validator (BaseValidator): Validator instance.model (nn.Module): Model instance.callbacks (defaultdict): Dictionary of callbacks.save_dir (Path): Directory to save results.wdir (Path): Directory to save weights.last (Path): Path to last checkpoint.best (Path): Path to best checkpoint.batch_size (int): Batch size for training.epochs (int): Number of epochs to train for.start_epoch (int): Starting epoch for training.device (torch.device): Device to use for training.amp (bool): Flag to enable AMP (Automatic Mixed Precision).scaler (amp.GradScaler): Gradient scaler for AMP.data (str): Path to data.trainset (torch.utils.data.Dataset): Training dataset.testset (torch.utils.data.Dataset): Testing dataset.ema (nn.Module): EMA (Exponential Moving Average) of the model.lf (nn.Module): Loss function.scheduler (torch.optim.lr_scheduler._LRScheduler): Learning rate scheduler.best_fitness (float): The best fitness value achieved.fitness (float): Current fitness value.loss (float): Current loss value.tloss (float): Total loss value.loss_names (list): List of loss names.csv (Path): Path to results CSV file."""def __init__(self, config=DEFAULT_CONFIG, overrides=None): # overrides={'data': 'wider_face.yaml', 'epochs': 100, 'task': 'detect', 'mode': 'train'}if overrides is None:overrides = {}self.args = get_config(config, overrides) # config='yolov8/ultralytics/yolo/configs/default.yaml''''self.args={'task': 'detect', 'mode': 'train', 'model': None, 'data': 'wider_face.yaml', 'epochs': 100, 'patience': 50, 'batch': 16, 'imgsz': 640, 'save': True, 'cache': False, 'device': None, 'workers': 8, 'project': None, 'name': None, 'exist_ok': False, 'pretrained': False, 'optimizer': 'SGD', 'verbose': False, 'seed': 0, 'deterministic': True, 'single_cls': False, 'image_weights': False, 'rect': False, 'cos_lr': False, 'close_mosaic': 10, 'resume': False, 'overlap_mask': True, 'mask_ratio': 4, 'dropout': 0.0, 'val': True, 'save_json': False, 'save_hybrid': False, 'conf': None, 'iou': 0.7, 'max_det': 300, 'half': False, 'dnn': False, 'plots': True, 'source': None, 'show': False, 'save_txt': False, 'save_conf': False, 'save_crop': False, 'hide_labels': False, 'hide_conf': False, 'vid_stride': 1, 'line_thickness': 3, 'visualize': False, 'augment': False, 'agnostic_nms': False, 'retina_masks': False, 'format': 'torchscript', 'keras': False, 'optimize': False, 'int8': False, 'dynamic': False, 'simplify': False, 'opset': 17, 'workspace': 4, 'nms': False, 'lr0': 0.01, 'lrf': 0.01, 'momentum': 0.937, 'weight_decay': 0.0005, 'warmup_epochs': 3.0, 'warmup_momentum': 0.8, 'warmup_bias_lr': 0.1, 'box': 7.5, 'cls': 0.5, 'dfl': 1.5, 'fl_gamma': 0.0, 'label_smoothing': 0.0, 'nbs': 64, 'hsv_h': 0.015, 'hsv_s': 0.7, 'hsv_v': 0.4, 'degrees': 0.0, 'translate': 0.1, 'scale': 0.5, 'shear': 0.0, 'perspective': 0.0, 'flipud': 0.0, 'fliplr': 0.5, 'mosaic': 1.0, 'mixup': 0.0, 'copy_paste': 0.0, 'cfg': None, 'hydra': {'output_subdir': None, 'run': {'dir': '.'}}, 'v5loader': False}'''self.device = utils.torch_utils.select_device(self.args.device, self.args.batch) # self.device=device(type='cuda', index=0)self.check_resume()self.console = LOGGERself.validator = Noneself.model = Noneself.callbacks = defaultdict(list)init_seeds(self.args.seed + 1 + RANK, deterministic=self.args.deterministic)# Dirsproject = self.args.project or Path(SETTINGS['runs_dir']) / self.args.task # project='yolov8/runs/detect'name = self.args.name or f"{self.args.mode}" # name='train'self.save_dir = Path(self.args.get("save_dir",increment_path(Path(project) / name, exist_ok=self.args.exist_ok if RANK in {-1, 0} else True))) # self.save_dir='yolov8/runs/detect/train'self.wdir = self.save_dir / 'weights' # self.wdir='yolov8/runs/detect/train/weights'if RANK in {-1, 0}:self.wdir.mkdir(parents=True, exist_ok=True) # make dirwith open_dict(self.args):self.args.save_dir = str(self.save_dir)yaml_save(self.save_dir / 'args.yaml', OmegaConf.to_container(self.args, resolve=True)) # save run argsself.last, self.best = self.wdir / 'last.pt', self.wdir / 'best.pt' # checkpoint pathsself.batch_size = self.args.batchself.epochs = self.args.epochs # self.epochs=100self.start_epoch = 0if RANK == -1:print_args(dict(self.args))# Deviceself.amp = self.device.type != 'cpu'self.scaler = amp.GradScaler(enabled=self.amp)if self.device.type == 'cpu':self.args.workers = 0 # faster CPU training as time dominated by inference, not dataloading# Model and Dataloaders.self.model = self.args.model # self.model=Noneself.data = self.args.data # self.data='wider_face.yaml'if self.data.endswith(".yaml"):self.data = check_dataset_yaml(self.data)'''self.data={'path': PosixPath('yolov8/datasets/wider_face'), 'train': 'yolov8/datasets/wider_face/images/train', 'val': 'yolov8/datasets/wider_face/images/val', 'test': None, 'names': {0: 'face'}, 'download': , 'yaml_file': 'yolov8/ultralytics/yolo/data/datasets/wider_face.yaml', 'nc': 1}'''else:self.data = check_dataset(self.data)self.trainset, self.testset = self.get_dataset(self.data) # self.trainset='yolov8/datasets/wider_face/images/train', self.testset='yolov8/datasets/wider_face/images/val'self.ema = None# Optimization utils initself.lf = Noneself.scheduler = None# Epoch level metricsself.best_fitness = Noneself.fitness = Noneself.loss = Noneself.tloss = Noneself.loss_names = ['Loss']self.csv = self.save_dir / 'results.csv' # self.csv='yolov8/runs/detect/train/results.csv'self.plot_idx = [0, 1, 2]# Callbacksself.callbacks = defaultdict(list, {k: [v] for k, v in callbacks.default_callbacks.items()}) # add callbacksif RANK in {0, -1}:callbacks.add_integration_callbacks(self)初始化完self.trainer后，便开始训练，
self.trainer.train()同样，跳转到BaseTrainer中的train()中，
def train(self):# Allow device='', device=None on Multi-GPU systems to default to device=0if isinstance(self.args.device, int) or self.args.device: # i.e. device=0 or device=[0,1,2,3]world_size = torch.cuda.device_count()elif torch.cuda.is_available(): # i.e. device=None or device=''world_size = 1 # default to device 0else: # i.e. device='cpu' or 'mps'world_size = 0# Run subprocess if DDP training, else train normallyif world_size > 1 and "LOCAL_RANK" not in os.environ:command = generate_ddp_command(world_size, self)try:subprocess.run(command)except Exception as e:self.console(e)finally:ddp_cleanup(command, self)else:self._do_train(int(os.getenv("RANK", -1)), world_size) # world_size=1因为world_size=1，所以直接进入到self._do_train中，
def _do_train(self, rank=-1, world_size=1): # rank=-1， world_size=1if world_size > 1:self._setup_ddp(rank, world_size)self._setup_train(rank, world_size) # 设置与训练相关的参数, 如: optimizer、scheduler、train_loader、test_loader、validator、metrics等self.epoch_time = Noneself.epoch_time_start = time.time()self.train_time_start = time.time()nb = len(self.train_loader) # number of batchesnw = max(round(self.args.warmup_epochs * nb), 100) # number of warmup iterationslast_opt_step = -1self.run_callbacks("on_train_start")self.log(f"Image sizes {self.args.imgsz} train, {self.args.imgsz} val\n"f'Using {self.train_loader.num_workers * (world_size or 1)} dataloader workers\n'f"Logging results to {colorstr('bold', self.save_dir)}\n"f"Starting training for {self.epochs} epochs...")if self.args.close_mosaic:base_idx = (self.epochs - self.args.close_mosaic) * nbself.plot_idx.extend([base_idx, base_idx + 1, base_idx + 2])for epoch in range(self.start_epoch, self.epochs):self.epoch = epochself.run_callbacks("on_train_epoch_start")self.model.train()if rank != -1:self.train_loader.sampler.set_epoch(epoch)pbar = enumerate(self.train_loader)# Update dataloader attributes (optional)if epoch == (self.epochs - self.args.close_mosaic):self.console.info("Closing dataloader mosaic")if hasattr(self.train_loader.dataset, 'mosaic'):self.train_loader.dataset.mosaic = Falseif hasattr(self.train_loader.dataset, 'close_mosaic'):self.train_loader.dataset.close_mosaic(hyp=self.args)if rank in {-1, 0}:self.console.info(self.progress_string())pbar = tqdm(enumerate(self.train_loader), total=nb, bar_format=TQDM_BAR_FORMAT)self.tloss = Noneself.optimizer.zero_grad()for i, batch in pbar:self.run_callbacks("on_train_batch_start")# Warmupni = i + nb * epochif ni <= nw:xi = [0, nw] # x interpself.accumulate = max(1, np.interp(ni, xi, [1, self.args.nbs / self.batch_size]).round())for j, x in enumerate(self.optimizer.param_groups):# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0x['lr'] = np.interp(ni, xi, [self.args.warmup_bias_lr if j == 0 else 0.0, x['initial_lr'] * self.lf(epoch)])if 'momentum' in x:x['momentum'] = np.interp(ni, xi, [self.args.warmup_momentum, self.args.momentum])# Forwardwith torch.cuda.amp.autocast(self.amp):batch = self.preprocess_batch(batch)preds = self.model(batch["img"])self.loss, self.loss_items = self.criterion(preds, batch)if rank != -1:self.loss *= world_sizeself.tloss = (self.tloss * i + self.loss_items) / (i + 1) if self.tloss is not None \else self.loss_items# Backwardself.scaler.scale(self.loss).backward()# Optimize - https://pytorch.org/docs/master/notes/amp_examples.htmlif ni - last_opt_step >= self.accumulate:self.optimizer_step()last_opt_step = ni# Logmem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G' # (GB)loss_len = self.tloss.shape[0] if len(self.tloss.size()) else 1losses = self.tloss if loss_len > 1 else torch.unsqueeze(self.tloss, 0)if rank in {-1, 0}:pbar.set_description(('%11s' * 2 + '%11.4g' * (2 + loss_len)) %(f'{epoch + 1}/{self.epochs}', mem, *losses, batch["cls"].shape[0], batch["img"].shape[-1]))self.run_callbacks('on_batch_end')if self.args.plots and ni in self.plot_idx:self.plot_training_samples(batch, ni)self.run_callbacks("on_train_batch_end")self.lr = {f"lr/pg{ir}": x['lr'] for ir, x in enumerate(self.optimizer.param_groups)} # for loggersself.scheduler.step()self.run_callbacks("on_train_epoch_end")if rank in {-1, 0}:# Validationself.ema.update_attr(self.model, include=['yaml', 'nc', 'args', 'names', 'stride', 'class_weights'])final_epoch = (epoch + 1 == self.epochs)if self.args.val or final_epoch:self.metrics, self.fitness = self.validate()self.save_metrics(metrics={**self.label_loss_items(self.tloss), **self.metrics, **self.lr})# Save modelif self.args.save or (epoch + 1 == self.epochs):self.save_model()self.run_callbacks('on_model_save')tnow = time.time()self.epoch_time = tnow - self.epoch_time_startself.epoch_time_start = tnowself.run_callbacks("on_fit_epoch_end")# TODO: termination conditionif rank in {-1, 0}:# Do final val with best.ptself.log(f'\n{epoch - self.start_epoch + 1} epochs completed in 'f'{(time.time() - self.train_time_start) / 3600:.3f} hours.')self.final_eval()if self.args.plots:self.plot_metrics()self.log(f"Results saved to {colorstr('bold', self.save_dir)}")self.run_callbacks('on_train_end')torch.cuda.empty_cache()self.run_callbacks('teardown')以上便是yolov8的训练过程。
