yolov8官方教程提供了2种训练方式，一种是通过命令行启动训练，一种是通过写代码启动。

命令行的方式启动方便，通过传入参数可以方便的调整训练参数，但这种方式不方便记录训练参数和调试训练代码。
自行写训练代码的方式更灵活，也比较方便调试，但官方的示例各种参数都是在代码中写死的方式，失去了灵活性。
其实我们可以结合这两种方法的优势，既能够通过命令行参数修改很容易变化的参数（如batch size, epoch, imgsz等），然后用配置文件保存很少需要变化的参数，或者这些变化需要保存下来方便对比（如各类增强比例）。

代码分析

首先我们需要知道我们能够设置哪些参数，尽管官方文档列出了命令行能够传入的参数列表，但每次设置大量参数还是不方便，而不设置的时候默认参数是多少我们也不知道，所以还是有必要分析一下代码。
通过模型的train接口我们会知道所有的Trainer均继承自BaseTrainer(yolo/engine/trainer.py)，该类的构造函数如下：

def __init__(self, cfg=DEFAULT_CFG, overrides=None, _callbacks=None):
        """
        Initializes the BaseTrainer class.

        Args:
            cfg (str, optional): Path to a configuration file. Defaults to DEFAULT_CFG.
            overrides (dict, optional): Configuration overrides. Defaults to None.
        """
        self.args = get_cfg(cfg, overrides)
        self.device = select_device(self.args.device, self.args.batch)
        self.check_resume()
        ...

其中overrides就是我们设置的参数，我们未设置的参数则来源于DEFAULT_CFG，继续跟踪我们会发现这个DEFAULT_CFG实际来源于yolo/cfg/default.yaml：

# Ultralytics YOLO 🚀, AGPL-3.0 license
# Default training settings and hyperparameters for medium-augmentation COCO training

task: detect  # YOLO task, i.e. detect, segment, classify, pose
mode: train  # YOLO mode, i.e. train, val, predict, export, track, benchmark

# Train settings -------------------------------------------------------------------------------------------------------
model:  # path to model file, i.e. yolov8n.pt, yolov8n.yaml
data:  # path to data file, i.e. coco128.yaml
epochs: 100  # number of epochs to train for
start_epoch: 0  # start epoch
patience: 50  # epochs to wait for no observable improvement for early stopping of training
batch: 16  # number of images per batch (-1 for AutoBatch)
imgsz: 640  # size of input images as integer or w,h
save: True  # save train checkpoints and predict results
save_period: -1 # Save checkpoint every x epochs (disabled if < 1)
cache: False  # True/ram, disk or False. Use cache for data loading
device:  # device to run on, i.e. cuda device=0 or device=0,1,2,3 or device=cpu
workers: 8  # number of worker threads for data loading (per RANK if DDP)
project:  # project name
name:  # experiment name, results saved to 'project/name' directory
exist_ok: False  # whether to overwrite existing experiment
pretrained: False  # whether to use a pretrained model
optimizer: SGD  # optimizer to use, choices=[SGD, Adam, Adamax, AdamW, NAdam, RAdam, RMSProp, auto]
verbose: True  # whether to print verbose output
seed: 0  # random seed for reproducibility
deterministic: True  # whether to enable deterministic mode
single_cls: False  # train multi-class data as single-class
rect: False  # rectangular training if mode='train' or rectangular validation if mode='val'
cos_lr: False  # use cosine learning rate scheduler
close_mosaic: 0  # (int) disable mosaic augmentation for final epochs
resume: False  # resume training from last checkpoint
amp: True  # Automatic Mixed Precision (AMP) training, choices=[True, False], True runs AMP check
fraction: 1.0  # dataset fraction to train on (default is 1.0, all images in train set)
profile: False  # profile ONNX and TensorRT speeds during training for loggers
# Segmentation
overlap_mask: True  # masks should overlap during training (segment train only)
mask_ratio: 4  # mask downsample ratio (segment train only)
# Classification
dropout: 0.0  # use dropout regularization (classify train only)
...

我们所有能设置的参数就在这个文件中，如果我们设置了不在其中的参数则会报错（下一篇介绍怎么增加参数）。

自定义参数配置文件

我们可以将训练会调整的参数单独保存到一个yaml文件,如hyp.scratch.yaml作为从头训练的配置，进行多次实验时，就可以建立不同的配置参数文件：

lr0: 0.01  # initial learning rate (i.e. SGD=1E-2, Adam=1E-3)
lrf: 0.001  # final learning rate (lr0 * lrf)
momentum: 0.937  # SGD momentum/Adam beta1
weight_decay: 0.0005  # optimizer weight decay 5e-4
warmup_epochs: 3.0  # warmup epochs (fractions ok)
warmup_momentum: 0.8  # warmup initial momentum
warmup_bias_lr: 0.1  # warmup initial bias lr
box: 7.5  # box loss gain
cls: 0.5  # cls loss gain (scale with pixels)
dfl: 1.5  # dfl loss gain
pose: 12.0  # pose loss gain
kobj: 1.0  # keypoint obj loss gain
label_smoothing: 0.0  # label smoothing (fraction)
nbs: 64  # nominal batch size
hsv_h: 0.015  # image HSV-Hue augmentation (fraction)
hsv_s: 0.7  # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4  # image HSV-Value augmentation (fraction)
degrees: 0.0  # image rotation (+/- deg)
translate: 0.1  # image translation (+/- fraction)
scale: 0.5  # image scale (+/- gain)
shear: 0.0  # image shear (+/- deg)
perspective: 0.0  # image perspective (+/- fraction), range 0-0.001
flipud: 0.0  # image flip up-down (probability)
fliplr: 0.5  # image flip left-right (probability)
mosaic: 0.1  # image mosaic (probability)
mixup: 0.05  # image mixup (probability)
copy_paste: 0.0  # segment copy-paste (probability)

workers: 12  # number of workers
# cache: disk

自定义训练脚本

建立了自定义参数文件，我们还要建立自己的训练脚本来载入配置文件，并且还有一些经常变化的参数需要通过命令行传入, 新建train.py：

from ultralytics import YOLO
import yaml
import argparse

parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, default='configs/data/phd.yaml', help='dataset.yaml path')
parser.add_argument('--epochs', type=int, default=300, help='number of epochs')
parser.add_argument('--hyp', type=str, default='configs/hyp.yaml', help='size of each image batch')
parser.add_argument('--model', type=str, default='weights/yolov8n.pt', help='pretrained weights or model.config path')
parser.add_argument('--batch-size', type=int, default=64, help='size of each image batch')
parser.add_argument('--img-size', type=int, default=320, help='size of each image dimension')
parser.add_argument('--device', type=str, default='0', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
parser.add_argument('--project', type=str, default='yolo', help='project name')
parser.add_argument('--name', type=str, default='pretrain', help='exp name')
parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')

args = parser.parse_args()

assert args.data, 'argument --data path is required'
assert args.model, 'argument --model path is required'

if __name__ == '__main__':
    # Initialize
    model = YOLO(args.model)
    hyperparams = yaml.safe_load(open(args.hyp))
    hyperparams['epochs'] = args.epochs
    hyperparams['batch'] = args.batch_size
    hyperparams['imgsz'] = args.img_size
    hyperparams['device'] = args.device
    hyperparams['project'] = args.project
    hyperparams['name'] = args.name
    hyperparams['resume'] = args.resume

    model.train(data= args.data, **hyperparams)

该脚本通过argparse来接受命令行参数，并设置到超参数字典，和yolov5的启动脚本类似。
主要有以下几个参数（可以根据个人需要增删）：

data: 数据集配置文件
hyp: 参数配置文件（上一节我们建立的）
model: 模型权重或者模型结构配置文件
其他参数根据名字就显而易见了。

模型训练（单卡）

python train.py --model weights/yolov8n.pt --data
configs/data/objects365.yaml --hyp configs/hyp.yaml --batch-size 512 --img-size 416 --device
0 --project object365 --name yolov8n

模型训练（多卡DDP）

理论上，我们只需要将device设置为多张卡就可以进行多卡并行了，但我们直接运行会发生一下错误：

assert args.model, 'argument --model path is required'

也就是我们设置的参数并没有接收到，进一步分析，DDP情况下，实际运行的命令是：

DDP command: ['/root/miniconda3/bin/python', '-m', 'torch.distributed.run', '--nproc_per_node', '4', '--master_port', '39083', 'xxx/code/yolov8/train.py']
WARNING:__main__:

也就是yolov8实际是用pytorch的ddp脚本启动了我们写得train.py脚本，但是却没有把我们设置的参数传过来（应该算是个bug吧···），这个过程发生在BaseTrainer的train接口中：

我们对generate_ddp_command进行修改，将命令行参数增加到train.py后（file后增加*sys.argv[1:]）：

cmd = [sys.executable, '-m', dist_cmd, '--nproc_per_node', f'{world_size}', '--master_port', f'{port}', file, *sys.argv[1:]]

完整的函数：

def generate_ddp_command(world_size, trainer):
    """Generates and returns command for distributed training."""
    import __main__  # noqa local import to avoid https://github.com/Lightning-AI/lightning/issues/15218
    if not trainer.resume:
        shutil.rmtree(trainer.save_dir)  # remove the save_dir
    file = str(Path(sys.argv[0]).resolve())

    safe_pattern = re.compile(r'^[a-zA-Z0-9_. /\\-]{1,128}$')  # allowed characters and maximum of 100 characters
    if not (safe_pattern.match(file) and Path(file).exists() and file.endswith('.py')):  # using CLI
        file = generate_ddp_file(trainer)
    dist_cmd = 'torch.distributed.run' if TORCH_1_9 else 'torch.distributed.launch'
    port = find_free_network_port()
    cmd = [sys.executable, '-m', dist_cmd, '--nproc_per_node', f'{world_size}', '--master_port', f'{port}', file, *sys.argv[1:]]
    return cmd, file

修改后，device设置多卡则能正常开启训练。