devbox
Cpp五条
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

YOLOv5代码解析——train.py

作者：Cpp五条 | 2024-02-29 19:05:03

赞

踩

YOLOv5代码解析——train.py

         train.py是训练YOLOV5使用的代码，后边根据函数调用展开。首先是主函数，显示有parse_opt()函数获得参数，然后把参数传给main函数。

  1.  
  2. if __name__ == '__main__':
  3.     opt = parse_opt() # 获得参数
  4.     main(opt) # 把参数传给main函数完成后续操作

1、parse_opt()函数主要用来加载参数，都有默认值，在使用的时候重新进行配置。

  1. def parse_opt(known=False):
  2.     parser = argparse.ArgumentParser()
  3.     # 加载预训练权重
  4.     parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='initial weights path')
  5.     # 加载cfg配置文件(网络结构)
  6.     parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
  7.     # 加载数据集
  8.     parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
  9.     # 配置超参数
  10.     parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
  11.     # epochs 训练轮次
  12.     parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
  13.     # batch-size 训练批次，默认16
  14.     parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
  15.     # 设置图片大小
  16.     parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
  17.     # 是否采用矩形训练，默认False
  18.     parser.add_argument('--rect', action='store_true', help='rectangular training')
  19.     # 是否接着上次的训练结构继续训练
  20.     parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
  21.     # 不保存训练结果
  22.     parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
  23.     # noval 最后进行测试, 设置了之后就是训练结束都测试一下， 不设置每轮都计算mAP, 建议不设置#
  24.     parser.add_argument('--noval', action='store_true', help='only validate final epoch')
  25.     # 不自动调整anchor，默认为false(yolov5会根据数据集自动计算anchor，这是特色之一)
  26.     parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
  27.     parser.add_argument('--noplots', action='store_true', help='save no plot files')
  28.     # 遗传算法调参
  29.     parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
  30.     # 谷歌优盘(一般用不到)
  31.     parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
  32.     # cache 是否提前缓存图片到内存，以加快训练速度，默认False
  33.     parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
  34.     # mage-weights 使用图片采样策略，默认不使用
  35.     parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
  36.     # gpu设备选择，
  37.     parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
  38.     # 多尺度训练
  39.     parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
  40.     # single-cls 数据集是否多类/默认True
  41.     parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
  42.     # # optimizer 优化器选择 / 提供了三种优化器
  43.     parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW'], default='SGD', help='optimizer')
  44.     # sync-bn:是否使用跨卡同步BN,在DDP模式使用
  45.     parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
  46.     # 使用的线程数量(谨慎选择，适度的workers有助于提升训练速度，workers过大反而会导致变慢)
  47.     parser.add_argument('--workers', type=int, default=12, help='max dataloader workers (per RANK in DDP mode)')
  48.     # 保存路径 / 默认保存路径 ./runs/ train
  49.     parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
  50.     # 实验名称
  51.     parser.add_argument('--name', default='exp', help='save to project/name')
  52.     # 项目位置是否存在 / 默认是都不存在
  53.     parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
  54.     parser.add_argument('--quad', action='store_true', help='quad dataloader')
  55.     # 余弦学习率
  56.     parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
  57.     # 标签平滑，默认不使用
  58.     parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
  59.     # 100次不更新就停止训练
  60.     parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
  61.     # --freeze冻结训练 可以设置 default = [0] 数据量大的情况下，建议不设置这个参数
  62.     parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
  63.     # 多少个epoch保存一下checkpoint
  64.     parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
  65.     # 随机数种子
  66.     parser.add_argument('--seed', type=int, default=0, help='Global training seed')
  67.     # --local_rank 进程编号 / 多卡使用
  68.     parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
  69.  
  70.     # Logger arguments
  71.     # # Weights & Biases arguments，类似于tensorboard的可视化工具
  72.     parser.add_argument('--entity', default=None, help='Entity')
  73.     parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
  74.     parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
  75.     parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')
  76.  
  77.     return parser.parse_known_args()[0] if known else parser.parse_args()

2、main函数这里没有介绍遗传净化算法，后边会更新博客单独介绍这一部分。

2.1 打印参数，检查环境
  1.  # Checks
  2.     if RANK in {-1, 0}:
  3.         # 输出所有训练参数
  4.         print_args(vars(opt))
  5.         # 检查代码版本是否更新
  6.         check_git_status()
  7.         # 检查所需要的包是否都安装了
  8.         check_requirements(ROOT / 'requirements.txt')
2.2 断点训练，判断是否要接着上一次的训练继续训练，如果是，就把参数替换为上一次的参数，如果不使用断点训练，直接加载参数并保存到一个文件中。
  1. # Resume (from specified or most recent last.pt) 断点训练
  2.     if opt.resume and not check_comet_resume(opt) and not opt.evolve:
  3.         # isinstance()是否是已经知道的类型
  4.         # 如果resume是True，则通过get_lastest_run()函数找到runs为文件夹中最近的权重文件last.pt
  5.         last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
  6.         opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
  7.         opt_data = opt.data  # original dataset
  8.         # 把opt的参数替换为last.pt中opt的参数
  9.         if opt_yaml.is_file():
  10.             with open(opt_yaml, errors='ignore') as f:
  11.                 d = yaml.safe_load(f)
  12.         else:
  13.             d = torch.load(last, map_location='cpu')['opt']
  14.         opt = argparse.Namespace(**d)  # replace
  15.         opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
  16.         if is_url(opt_data):
  17.             opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
  18.     # 不使用断点训练
  19.     else:
  20.         # 加载参数
  21.         opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
  22.             check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
  23.         assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
  24.         if opt.evolve:
  25.             if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
  26.                 opt.project = str(ROOT / 'runs/evolve')
  27.             opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
  28.         if opt.name == 'cfg':
  29.             opt.name = Path(opt.cfg).stem  # use model.yaml as name
  30.         # 保存相关信息到文件中
  31.         opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))

2.3 分布式训练

  1.  # DDP mode
  2.     # 选择device
  3.     device = select_device(opt.device, batch_size=opt.batch_size) 
  4.     if LOCAL_RANK != -1:
  5.         msg = 'is not compatible with YOLOv5 Multi-GPU DDP training'
  6.         assert not opt.image_weights, f'--image-weights {msg}'
  7.         assert not opt.evolve, f'--evolve {msg}'
  8.         assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
  9.         assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
  10.         assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
  11.         # 根据编号选择设备
  12.         #使用torch.cuda.set_device()可以更方便地将模型和数据加载到对应GPU上, 直接定义模型之前加入一行代码即可
  13.         # torch.cuda.set_device(gpu_id) #单卡
  14.         # torch.cuda.set_device('cuda:'+str(gpu_ids)) 
  15.         #可指定多卡
  16.         torch.cuda.set_device(LOCAL_RANK)
  17.         device = torch.device('cuda', LOCAL_RANK)
  18.         # 初始化多进程
  19.         dist.init_process_group(backend='nccl' if dist.is_nccl_available() else 'gloo',
  20.                                 timeout=timedelta(seconds=10800))

3、train函数

3.1 train函数——基本配置信息

解析了从opt传入的参数，创建了训练结果的保存路径，对绘图的参数进行了配置。

  1. ##################### 基本信息配置 ##################################
  2.     # 解析opt传入的参数
  3.     save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
  4.         Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
  5.         opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
  6.     callbacks.run('on_pretrain_routine_start')
  7.  
  8.     # Directories
  9.     w = save_dir / 'weights'  # weights dir
  10.     # 创建保存训练结果的文件夹
  11.     (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
  12.     # 保存训练结果的目录 ，如runs/train/exp1/weights/last.pt
  13.     last, best = w / 'last.pt', w / 'best.pt'
  14.  
  15.     # Hyperparameters
  16.     if isinstance(hyp, str):
  17.         with open(hyp, errors='ignore') as f:
  18.             hyp = yaml.safe_load(f)  # load hyps dict
  19.     # 打印超参数，彩色字体
  20.     LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
  21.     opt.hyp = hyp.copy()  # for saving hyps to checkpoints
  22.  
  23.     # Save run settings
  24.     if not evolve:
  25.         yaml_save(save_dir / 'hyp.yaml', hyp)
  26.         yaml_save(save_dir / 'opt.yaml', vars(opt))
  27.  
  28.     # Loggers
  29.     data_dict = None
  30.     if RANK in {-1, 0}:
  31.         loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance
  32.  
  33.         # Register actions
  34.         for k in methods(loggers):
  35.             callbacks.register_action(k, callback=getattr(loggers, k))
  36.  
  37.         # Process custom dataset artifact link
  38.         data_dict = loggers.remote_dataset
  39.         if resume:  # If resuming runs from remote artifact
  40.             weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
  41.  
  42.     # Config
  43.     plots = not evolve and not opt.noplots  # create plots
  44.     cuda = device.type != 'cpu'
  45.     # 随机种子
  46.     init_seeds(opt.seed + 1 + RANK, deterministic=True)
  47.     # 存在子进程-分布式训练
  48.     with torch_distributed_zero_first(LOCAL_RANK):
  49.         data_dict = data_dict or check_dataset(data)  # check if None
  50.     # 获取训练集和验证集的路径
  51.     train_path, val_path = data_dict['train'], data_dict['val']
  52.     # 设置类别，判断是否为蛋类
  53.     nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
  54.     # 类别对应的名称
  55.     names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
  56.     # 判断是否是coco数据集
  57.     is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
3.2 train函数——模型加载/断点训练
  1. # Model
  2.     # 检查文件后缀是否是.pt
  3.     check_suffix(weights, '.pt')  # check weights
  4.     pretrained = weights.endswith('.pt')
  5.     if pretrained:
  6.          #  torch_distributed_zero_first(RANK): 用于同步不同进程对数据读取的上下文管理器
  7.         with torch_distributed_zero_first(LOCAL_RANK):
  8.             # 如果不存在就从网站上下载
  9.             weights = attempt_download(weights)  # download if not found locally
  10.         # 加载预训练模型及参数
  11.         ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
  12.         """
  13.         两种加载模型的方式: opt.cfg / ckpt['model'].yaml
  14.         使用resume-断点训练: 选择ckpt['model']yaml创建模型, 且不加载anchor
  15.         使用断点训练时,保存的模型会保存anchor,所以不需要加载
  17.         """
  18.         model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
  19.         exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
  20.         # 筛选字典中断电键值对，把exclude删除
  21.         csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
  22.         csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
  23.         model.load_state_dict(csd, strict=False)  # load
  24.         LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
  25.     else:
  26.         # 不使用预训练权重
  27.         model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
  28.     amp = check_amp(model)  # check AMP
3.3 train函数——冻结训练层
  1.  # Freeze  冻结网络的训练层 
  2.     freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
  3.     for k, v in model.named_parameters():
  4.         v.requires_grad = True  # train all layers
  5.         # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
  6.         if any(x in k for x in freeze):
  7.             LOGGER.info(f'freezing {k}')
  8.             # 冻结的训练层梯度不更新
  9.             v.requires_grad = False
3.4 train函数——图片大小和batchsize设置
  1.  # Image size
  2.     gs = max(int(model.stride.max()), 32)  # grid size (max stride)
  3.     # 检查图片大小
  4.     imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
  5.  
  6.     # Batch size
  7.     if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
  8.         batch_size = check_train_batch_size(model, imgsz, amp)
  9.         loggers.on_params_update({'batch_size': batch_size})
3.5 train函数——优化器设置
  1. """
  2.     yolov5这里并不是根据batch size的大小去更新梯度,而是设置了一个固定的值
  3.     nbs = 64
  4.     batchsize = 16
  5.     accumulate = 64/16=4 
  6.     梯度累计accumlate次之后更新一次模型，相当于使用更大的batch_size
  7.     """
  8.     # Optimizer
  9.     nbs = 64  # nominal batch size
  10.     accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
  11.     # 权重衰减参数
  12.     hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
  13.     optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
3.6 train函数—— 学习率/EMA/显卡设置
  1. # Scheduler
  2.     # 是否使用余弦学习率调整方式
  3.     if opt.cos_lr:
  4.         lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
  5.     else:
  6.         lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
  7.     scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
  8.  
  9.     # EMA 对模型的参数做平均，给予近期数据更高权重的平均方法
  10.     ema = ModelEMA(model) if RANK in {-1, 0} else None
  11.  
  12.     # Resume
  13.     best_fitness, start_epoch = 0.0, 0
  14.     if pretrained:
  15.         if resume:
  16.             best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
  17.         del ckpt, csd
  18.  
  19.     # DP mode 单机多卡
  20.     if cuda and RANK == -1 and torch.cuda.device_count() > 1:
  21.         LOGGER.warning(
  22.             'WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n'
  23.             'See Multi-GPU Tutorial at https://docs.ultralytics.com/yolov5/tutorials/multi_gpu_training to get started.'
  24.         )
  25.         model = torch.nn.DataParallel(model)
  26.  
  27.     # SyncBatchNorm 多卡归一化
  28.     if opt.sync_bn and cuda and RANK != -1:
  29.         model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
  30.         LOGGER.info('Using SyncBatchNorm()')
3.7 train函数——数据加载/anchor 调整
  1. # Trainloader 训练集数据加载
  2.     train_loader, dataset = create_dataloader(train_path,
  3.                                               imgsz,
  4.                                               batch_size // WORLD_SIZE,
  5.                                               gs,
  6.                                               single_cls,
  7.                                               hyp=hyp,
  8.                                               augment=True,
  9.                                               cache=None if opt.cache == 'val' else opt.cache,
  10.                                               rect=opt.rect,
  11.                                               rank=LOCAL_RANK,
  12.                                               workers=workers,
  13.                                               image_weights=opt.image_weights,
  14.                                               quad=opt.quad,
  15.                                               prefix=colorstr('train: '),
  16.                                               shuffle=True,
  17.                                               seed=opt.seed)
  18.     # mlc 标签编号最大值
  19.     labels = np.concatenate(dataset.labels, 0)
  20.     mlc = int(labels[:, 0].max())  # max label class
  21.     # 判断编号是否正确
  22.     assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
  23.  
  24.     # Process 0
  25.     # 验证集数据加载
  26.     if RANK in {-1, 0}:
  27.         val_loader = create_dataloader(val_path,
  28.                                        imgsz,
  29.                                        batch_size // WORLD_SIZE * 2,
  30.                                        gs,
  31.                                        single_cls,
  32.                                        hyp=hyp,
  33.                                        cache=None if noval else opt.cache,
  34.                                        rect=True,
  35.                                        rank=-1,
  36.                                        workers=workers * 2,
  37.                                        pad=0.5,
  38.                                        prefix=colorstr('val: '))[0]
  39.  
  40.         if not resume: # 不使用断点训练
  41.             if not opt.noautoanchor:
  42.                 # dataset是在上边创建train_loader时生成的
  43.                 # hyp['anchor_t']是从配置文件hpy.scratch.yaml读取的超参数 anchor_t:4.0
  44.                 # 当配置文件中的anchor计算bpr（best possible recall）小于0.98时才会重新计算anchor
  45.                 # best possible recall最大值1，如果bpr小于0.98，程序会根据数据集的label自动学习anchor的尺寸
  46.                 check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
  47.             # 半精度
  48.             model.half().float()  # pre-reduce anchor precision
  49.  
  50.         callbacks.run('on_pretrain_routine_end', labels, names)
3.8 train函数——训练配置
  1. ############# 训练配置 ##############
  2.     # DDP mode 多机多卡
  3.     if cuda and RANK != -1:
  4.         model = smart_DDP(model)
  5.  
  6.     # Model attributes
  7.     # smart_DDP和de_parallel代码在utils.torch_utils中
  8.     # 对hpy字典中的一些值进行缩放和预设置，以适应不同的层级、类别、图像尺寸和标签平滑需求
  9.     # 默认 nl = 3
  10.     nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
  11.     # hyp-low中给出的 box=0.05; cls=0.5; obj=1.0 
  12.     # hyp['box'] = 0.05*3/3=0.05
  13.     hyp['box'] *= 3 / nl  # scale to layers
  14.     # hyp['cls'] = 0.5*20/80*3/3=0.125
  15.     hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
  16.     # hyp['obj']=1.0*(640/640)**2*3/nl=1
  17.     hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
  18.     hyp['label_smoothing'] = opt.label_smoothing
  19.     model.nc = nc  # attach number of classes to model
  20.     model.hyp = hyp  # attach hyperparameters to model
  21.     # 从训练样本标签得到类别权重（和类别中的目标数即类别频率成反比）
  22.     model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
  23.     model.names = names
3.9 train函数——训练

        开始训练的代码，使用先前生成的trian_loader读取图片，送入模型开始训练，并计算损失进行反向传播，以及每轮训练后进行验证计算P，R，mAP等。

  1.  ########################## Start training ##########################
  2.     t0 = time.time()
  3.     nb = len(train_loader)  # number of batches
  4.     nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
  5.     # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
  6.     last_opt_step = -1
  7.     # 初始化maps(每个类别的map)和results
  8.     maps = np.zeros(nc)  # mAP per class
  9.     results = (0, 0, 0, 0, 0, 0, 0)  # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)
  10.     # 设置学习率衰减所进行到的轮次，即使打断训练，使用resume接着训练也能正常衔接之前的训练进行学习率衰减
  11.     scheduler.last_epoch = start_epoch - 1  # do not move
  12.     # 设置amp混合精度训练
  13.     scaler = torch.cuda.amp.GradScaler(enabled=amp)
  14.     # 早停止
  15.     stopper, stop = EarlyStopping(patience=opt.patience), False
  16.     # 初始化损失
  17.     compute_loss = ComputeLoss(model)  # init loss class
  18.     callbacks.run('on_train_start')
  19.     LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
  20.                 f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
  21.                 f"Logging results to {colorstr('bold', save_dir)}\n"
  22.                 f'Starting training for {epochs} epochs...')
  23.     # 正式开始训练
  24.     for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
  25.         callbacks.run('on_train_epoch_start')
  26.         model.train()
  27.  
  28.         # Update image weights (optional, single-GPU only)
  29.         if opt.image_weights:
  30.             """
  31.             如果设置图片采样策略
  32.             则根据前面初始化的图片采样权重model.class_weights以及maps配合每张图片包含的类别数
  33.             通过random.choices生成图片所有indices从而进行采样
  34.             """
  35.             cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
  36.             iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
  37.             dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
  38.  
  39.         # Update mosaic border (optional)
  40.         # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
  41.         # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
  42.  
  43.         mloss = torch.zeros(3, device=device)  # mean losses
  44.         if RANK != -1:
  45.             train_loader.sampler.set_epoch(epoch)
  46.         pbar = enumerate(train_loader)
  47.         LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'obj_loss', 'cls_loss', 'Instances', 'Size'))
  48.         if RANK in {-1, 0}:
  49.             # 进度条显示
  50.             pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
  51.         optimizer.zero_grad() # 梯度清零
  52.         for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
  53.             callbacks.run('on_train_batch_start')
  54.             ni = i + nb * epoch  # number integrated batches (since train start)
  55.             imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
  56.             
  57.             """
  58.             热身训练(前nw次迭代，一般是3)
  59.             在前nw次迭代中，根据以下方式选取accumulate和学习率
  60.             """
  61.             # Warmup
  62.             if ni <= nw:
  63.                 xi = [0, nw]  # x interp
  64.                 # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
  65.                 accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
  66.                 for j, x in enumerate(optimizer.param_groups):
  67.                      """
  68.                         bias的学习率从0.1下降到基准学习率lr*lf(epoch)，
  69.                         其他的参数学习率从0增加到lr*lf(epoch).
  70.                         lf为上面设置的余弦退火的衰减函数
  71.                         动量momentum也从0.9慢慢变到hyp['momentum'](default=0.937)
  72.                     """
  73.                     # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
  74.                     x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
  75.                     if 'momentum' in x:
  76.                         x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
  77.  
  78.             # Multi-scale
  79.             if opt.multi_scale:
  80.                 """
  81.                     Multi-scale  设置多尺度训练，从imgsz * 0.5, imgsz * 1.5 + gs随机选取尺寸
  82.                 """
  83.                 sz = random.randrange(int(imgsz * 0.5), int(imgsz * 1.5) + gs) // gs * gs  # size
  84.                 sf = sz / max(imgs.shape[2:])  # scale factor
  85.                 if sf != 1:
  86.                     ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
  87.                     imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
  88.  
  89.             # Forward
  90.             with torch.cuda.amp.autocast(amp):
  91.                 pred = model(imgs)  # forward
  92.                 # loss是总损失 loss_items是一个元组，包含分类损失，obj损失，boundingbox的回归损失和总损失
  93.                 loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
  94.                 if RANK != -1:
  95.                     # 平均不同gpu之间的梯度
  96.                     loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
  97.                 if opt.quad:
  98.                     loss *= 4.
  99.  
  100.             # Backward
  101.             scaler.scale(loss).backward()
  102.  
  103.             # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
  104.             # 模型反向传播accumulate次之后再根据累计的梯度更新一次参数
  105.             if ni - last_opt_step >= accumulate:
  106.                 scaler.unscale_(optimizer)  # unscale gradients
  107.                 torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
  108.                 scaler.step(optimizer)  # optimizer.step
  109.                 scaler.update()
  110.                 optimizer.zero_grad()
  111.                 if ema:
  112.                     ema.update(model)
  113.                 last_opt_step = ni
  114.  
  115.             # Log
  116.             if RANK in {-1, 0}:
  117.                 mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
  118.                 mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
  119.                 pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
  120.                                      (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
  121.                 callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths, list(mloss))
  122.                 if callbacks.stop_training:
  123.                     return
  124.             # end batch ------------------------------------------------------------------------------------------------
  125.  
  126.         # Scheduler 学习率衰减
  127.         lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
  128.         scheduler.step()
  129.  
  130.         if RANK in {-1, 0}:
  131.             # mAP
  132.             callbacks.run('on_train_epoch_end', epoch=epoch)
  133.             # 把model中的属性赋值给ema
  134.             ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
  135.             # 判断是否是最后一轮
  136.             final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
  137.             # notest: 是否只测试最后一轮  True: 只测试最后一轮   False: 每轮训练完都测试mAP
  138.             if not noval or final_epoch:  # Calculate mAP
  139.                 # 测试使用的是ema(对模型的参数做平均)模型
  140.                 # verbose设置为true后，每轮的验证都输出每个类别的信息
  141.                 results, maps, _ = validate.run(data_dict,
  142.                                                 batch_size=batch_size // WORLD_SIZE * 2,
  143.                                                 imgsz=imgsz,
  144.                                                 half=amp,
  145.                                                 model=ema.ema,
  146.                                                 single_cls=single_cls,
  147.                                                 dataloader=val_loader,
  148.                                                 save_dir=save_dir,
  149.                                                 plots=False,
  150.                                                 callbacks=callbacks,
  151.                                                 compute_loss=compute_loss,
  152.                                                 verbose=True)
  153.  
  154.             # Update best mAP
  155.             fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
  156.             stop = stopper(epoch=epoch, fitness=fi)  # early stop check
  157.             if fi > best_fitness:
  158.                 best_fitness = fi
  159.             log_vals = list(mloss) + list(results) + lr
  160.             callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
  161.  
  162.             # Save model
  163.             """
  164.                 保存带checkpoint的模型用于inference或resuming training
  165.                 保存模型的同时还保存epoch，results，optimizer等信息
  166.                 optimizer在最后一轮不会报错
  167.                 model保存的是EMA后的模型
  168.             """
  169.             if (not nosave) or (final_epoch and not evolve):  # if save
  170.                 ckpt = {
  171.                     'epoch': epoch,
  172.                     'best_fitness': best_fitness,
  173.                     'model': deepcopy(de_parallel(model)).half(),
  174.                     'ema': deepcopy(ema.ema).half(),
  175.                     'updates': ema.updates,
  176.                     'optimizer': optimizer.state_dict(),
  177.                     'opt': vars(opt),
  178.                     'git': GIT_INFO,  # {remote, branch, commit} if a git repo
  179.                     'date': datetime.now().isoformat()}
  180.  
  181.                 # Save last, best and delete
  182.                 torch.save(ckpt, last)
  183.                 if best_fitness == fi:
  184.                     torch.save(ckpt, best)
  185.                 if opt.save_period > 0 and epoch % opt.save_period == 0:
  186.                     torch.save(ckpt, w / f'epoch{epoch}.pt')
  187.                 del ckpt
  188.                 callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
  189.  
  190.         # EarlyStopping
  191.         if RANK != -1:  # if DDP training
  192.             broadcast_list = [stop if RANK == 0 else None]
  193.             dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
  194.             if RANK != 0:
  195.                 stop = broadcast_list[0]
  196.         if stop:
  197.             break  # must break all DDP ranks
  198.  
  199.         # end epoch ----------------------------------------------------------------------------------------------------
  200.     # end training --------------------------

3.10 train函数——打印训练信息

  1.  ############################ 打印训练信息 #########################
  2.     if RANK in {-1, 0}:
  3.         LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
  4.         for f in last, best:
  5.             if f.exists():
  6.                 strip_optimizer(f)  # strip optimizers
  7.                 if f is best:
  8.                     LOGGER.info(f'\nValidating {f}...')
  9.                     results, _, _ = validate.run(
  10.                         data_dict,
  11.                         batch_size=batch_size // WORLD_SIZE * 2,
  12.                         imgsz=imgsz,
  13.                         model=attempt_load(f, device).half(),
  14.                         iou_thres=0.65 if is_coco else 0.60,  # best pycocotools at iou 0.65
  15.                         single_cls=single_cls,
  16.                         dataloader=val_loader,
  17.                         save_dir=save_dir,
  18.                         save_json=is_coco,
  19.                         verbose=True,
  20.                         plots=plots,
  21.                         callbacks=callbacks,
  22.                         compute_loss=compute_loss)  # val best model with plots
  23.                     if is_coco:
  24.                         callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
  25.  
  26.         callbacks.run('on_train_end', last, best, epoch, results)
  27.  
  28.     torch.cuda.empty_cache() # 释放显存
  29.     return results

4、代码使用

        使用的命令如下，主要指定数据 --data 、权重 --weights 和模型配置文件 --cfg，其他参数选择使用。为了方便建议写入到xxx.sh文件进行运行，如果权限不够无法使用./xxx.sh运行，使用chmod +777 xxx.sh修改权限，就可以运行了。

 python train.py --data ./data/mydata.yaml --weight ./weights/yolov5l.pt --cfg ./models/yolov5l.yaml
声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/Cpp五条/article/detail/167675
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号