?????????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。
if __name__ == '__main__':
opt = parse_opt() # 获得参数
main(opt) # 把参数传给main函数完成后续操作
def parse_opt(known=False):
parser = argparse.ArgumentParser()
# 加载预训练权重
parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='initial weights path')
# 加载cfg配置文件(网络结构)
parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
# 加载数据集
parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
# 配置超参数
parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
# epochs 训练轮次
parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
# batch-size 训练批次，默认16
parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
# 设置图片大小
parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
# 是否采用矩形训练，默认False
parser.add_argument('--rect', action='store_true', help='rectangular training')
# 是否接着上次的训练结构继续训练
parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
# 不保存训练结果
parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
# 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#
parser.add_argument('--noval', action='store_true', help='only validate final epoch')
# 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)
parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
parser.add_argument('--noplots', action='store_true', help='save no plot files')
# 遗传算法调参
parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
# 谷歌优盘(一般用不到)
parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
# cache 是否提前缓存图片到内存，以加快训练速度，默认False
parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
# mage-weights 使用图片采样策略，默认不使用
parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
# gpu设备选择，
parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
# 多尺度训练
parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
# single-cls 数据集是否多类/默认True
parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
# # optimizer 优化器选择 / 提供了三种优化器
parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW'], default='SGD', help='optimizer')
# sync-bn:是否使用跨卡同步BN,在DDP模式使用
parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
# 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)
parser.add_argument('--workers', type=int, default=12, help='max dataloader workers (per RANK in DDP mode)')
# 保存路径 / 默认保存路径 ./runs/ train
parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
# 实验名称
parser.add_argument('--name', default='exp', help='save to project/name')
# 项目位置是否存在 / 默认是都不存在
parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
parser.add_argument('--quad', action='store_true', help='quad dataloader')
# 余弦学习率
parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
# 标签平滑，默认不使用
parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
# 100次不更新就停止训练
parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
# --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
parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
# 多少个epoch保存一下checkpoint
parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
# 随机数种子
parser.add_argument('--seed', type=int, default=0, help='Global training seed')
# --local_rank 进程编号 / 多卡使用
parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
# Logger arguments
# # Weights & Biases arguments，类似于tensorboard的可视化工具
parser.add_argument('--entity', default=None, help='Entity')
parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')
return parser.parse_known_args()[0] if known else parser.parse_args() # Checks
if RANK in {-1, 0}:
# 输出所有训练参数
print_args(vars(opt))
# 检查代码版本是否更新
check_git_status()
# 检查所需要的包是否都安装了
check_requirements(ROOT / 'requirements.txt')# Resume (from specified or most recent last.pt) 断点训练
if opt.resume and not check_comet_resume(opt) and not opt.evolve:
# isinstance()是否是已经知道的类型
# 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
last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
opt_yaml = last.parent.parent / 'opt.yaml' # train options yaml
opt_data = opt.data # original dataset
# 把opt的参数替换为last.pt中opt的参数
if opt_yaml.is_file():
with open(opt_yaml, errors='ignore') as f:
d = yaml.safe_load(f)
else:
d = torch.load(last, map_location='cpu')['opt']
opt = argparse.Namespace(**d) # replace
opt.cfg, opt.weights, opt.resume = '', str(last), True # reinstate
if is_url(opt_data):
opt.data = check_file(opt_data) # avoid HUB resume auth timeout
# 不使用断点训练
else:
# 加载参数
opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project) # checks
assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
if opt.evolve:
if opt.project == str(ROOT / 'runs/train'): # if default project name, rename to runs/evolve
opt.project = str(ROOT / 'runs/evolve')
opt.exist_ok, opt.resume = opt.resume, False # pass resume to exist_ok and disable resume
if opt.name == 'cfg':
opt.name = Path(opt.cfg).stem # use model.yaml as name
# 保存相关信息到文件中
opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))2.3?分布式训练
# DDP mode
# 选择device
device = select_device(opt.device, batch_size=opt.batch_size)
if LOCAL_RANK != -1:
msg = 'is not compatible with YOLOv5 Multi-GPU DDP training'
assert not opt.image_weights, f'--image-weights {msg}'
assert not opt.evolve, f'--evolve {msg}'
assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
# 根据编号选择设备
#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
# torch.cuda.set_device(gpu_id) #单卡
# torch.cuda.set_device('cuda:'+str(gpu_ids))
#可指定多卡
torch.cuda.set_device(LOCAL_RANK)
device = torch.device('cuda', LOCAL_RANK)
# 初始化多进程
dist.init_process_group(backend='nccl' if dist.is_nccl_available() else 'gloo',
timeout=timedelta(seconds=10800))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。
##################### 基本信息配置 ##################################
# 解析opt传入的参数
save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
callbacks.run('on_pretrain_routine_start')
# Directories
w = save_dir / 'weights' # weights dir
# 创建保存训练结果的文件夹
(w.parent if evolve else w).mkdir(parents=True, exist_ok=True) # make dir
# 保存训练结果的目录 ，如runs/train/exp1/weights/last.pt
last, best = w / 'last.pt', w / 'best.pt'
# Hyperparameters
if isinstance(hyp, str):
with open(hyp, errors='ignore') as f:
hyp = yaml.safe_load(f) # load hyps dict
# 打印超参数，彩色字体
LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
opt.hyp = hyp.copy() # for saving hyps to checkpoints
# Save run settings
if not evolve:
yaml_save(save_dir / 'hyp.yaml', hyp)
yaml_save(save_dir / 'opt.yaml', vars(opt))
# Loggers
data_dict = None
if RANK in {-1, 0}:
loggers = Loggers(save_dir, weights, opt, hyp, LOGGER) # loggers instance
# Register actions
for k in methods(loggers):
callbacks.register_action(k, callback=getattr(loggers, k))
# Process custom dataset artifact link
data_dict = loggers.remote_dataset
if resume: # If resuming runs from remote artifact
weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
# Config
plots = not evolve and not opt.noplots # create plots
cuda = device.type != 'cpu'
# 随机种子
init_seeds(opt.seed + 1 + RANK, deterministic=True)
# 存在子进程-分布式训练
with torch_distributed_zero_first(LOCAL_RANK):
data_dict = data_dict or check_dataset(data) # check if None
# 获取训练集和验证集的路径
train_path, val_path = data_dict['train'], data_dict['val']
# 设置类别，判断是否为蛋类
nc = 1 if single_cls else int(data_dict['nc']) # number of classes
# 类别对应的名称
names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names'] # class names
# 判断是否是coco数据集
is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt') # COCO dataset# Model
# 检查文件后缀是否是.pt
check_suffix(weights, '.pt') # check weights
pretrained = weights.endswith('.pt')
if pretrained:
# 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
with torch_distributed_zero_first(LOCAL_RANK):
# 如果不存在就从网站上下载
weights = attempt_download(weights) # download if not found locally
# 加载预训练模型及参数
ckpt = torch.load(weights, map_location='cpu') # load checkpoint to CPU to avoid CUDA memory leak
"""
两种加载模型的方式: opt.cfg / ckpt['model'].yaml
使用resume-断点训练: 选择ckpt['model']yaml创建模型, 且不加载anchor
使用断点训练时,保存的模型会保存anchor,所以不需要加载
"""
model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create
exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else [] # exclude keys
# 筛选字典中断电键值对，把exclude删除
csd = ckpt['model'].float().state_dict() # checkpoint state_dict as FP32
csd = intersect_dicts(csd, model.state_dict(), exclude=exclude) # intersect
model.load_state_dict(csd, strict=False) # load
LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}') # report
else:
# 不使用预训练权重
model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create
amp = check_amp(model) # check AMP
# Freeze 冻结网络的训练层
freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))] # layers to freeze
for k, v in model.named_parameters():
v.requires_grad = True # train all layers
# v.register_hook(lambda x: torch.nan_to_num(x)) # NaN to 0 (commented for erratic training results)
if any(x in k for x in freeze):
LOGGER.info(f'freezing {k}')
# 冻结的训练层梯度不更新
v.requires_grad = False
# Image size
gs = max(int(model.stride.max()), 32) # grid size (max stride)
# 检查图片大小
imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2) # verify imgsz is gs-multiple
# Batch size
if RANK == -1 and batch_size == -1: # single-GPU only, estimate best batch size
batch_size = check_train_batch_size(model, imgsz, amp)
loggers.on_params_update({'batch_size': batch_size})"""
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
nbs = 64
batchsize = 16
accumulate = 64/16=4
梯度累计accumlate次之后更新一次模型，相当于使用更大的batch_size
"""
# Optimizer
nbs = 64 # nominal batch size
accumulate = max(round(nbs / batch_size), 1) # accumulate loss before optimizing
# 权重衰减参数
hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay
optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
# Scheduler
# 是否使用余弦学习率调整方式
if opt.cos_lr:
lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf']
else:
lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf'] # linear
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs)
# 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
ema = ModelEMA(model) if RANK in {-1, 0} else None
# Resume
best_fitness, start_epoch = 0.0, 0
if pretrained:
if resume:
best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
del ckpt, csd
# DP mode 单机多卡
if cuda and RANK == -1 and torch.cuda.device_count() > 1:
LOGGER.warning(
'WARNING ?? DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n'
'See Multi-GPU Tutorial at https://docs.ultralytics.com/yolov5/tutorials/multi_gpu_training to get started.'
)
model = torch.nn.DataParallel(model)
# SyncBatchNorm 多卡归一化
if opt.sync_bn and cuda and RANK != -1:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
LOGGER.info('Using SyncBatchNorm()')# Trainloader 训练集数据加载
train_loader, dataset = create_dataloader(train_path,
imgsz,
batch_size // WORLD_SIZE,
gs,
single_cls,
hyp=hyp,
augment=True,
cache=None if opt.cache == 'val' else opt.cache,
rect=opt.rect,
rank=LOCAL_RANK,
workers=workers,
image_weights=opt.image_weights,
quad=opt.quad,
prefix=colorstr('train: '),
shuffle=True,
seed=opt.seed)
# mlc 标签编号最大值
labels = np.concatenate(dataset.labels, 0)
mlc = int(labels[:, 0].max()) # max label class
# 判断编号是否正确
assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
# Process 0
# 验证集数据加载
if RANK in {-1, 0}:
val_loader = create_dataloader(val_path,
imgsz,
batch_size // WORLD_SIZE * 2,
gs,
single_cls,
hyp=hyp,
cache=None if noval else opt.cache,
rect=True,
rank=-1,
workers=workers * 2,
pad=0.5,
prefix=colorstr('val: '))[0]
if not resume: # 不使用断点训练
if not opt.noautoanchor:
# dataset是在上边创建train_loader时生成的
# hyp['anchor_t']是从配置文件hpy.scratch.yaml读取的超参数 anchor_t:4.0
# 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
# 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
check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # run AutoAnchor
# 半精度
model.half().float() # pre-reduce anchor precision
callbacks.run('on_pretrain_routine_end', labels, names)############# 训练配置 ##############
# DDP mode 多机多卡
if cuda and RANK != -1:
model = smart_DDP(model)
# Model attributes
# smart_DDP和de_parallel代码在utils.torch_utils中
# 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
# 默认 nl = 3
nl = de_parallel(model).model[-1].nl # number of detection layers (to scale hyps)
# hyp-low中给出的 box=0.05; cls=0.5; obj=1.0
# hyp['box'] = 0.05*3/3=0.05
hyp['box'] *= 3 / nl # scale to layers
# hyp['cls'] = 0.5*20/80*3/3=0.125
hyp['cls'] *= nc / 80 * 3 / nl # scale to classes and layers
# hyp['obj']=1.0*(640/640)**2*3/nl=1
hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl # scale to image size and layers
hyp['label_smoothing'] = opt.label_smoothing
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
# 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）
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc # attach class weights
model.names = names????????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。
########################## Start training ##########################
t0 = time.time()
nb = len(train_loader) # number of batches
nw = max(round(hyp['warmup_epochs'] * nb), 100) # number of warmup iterations, max(3 epochs, 100 iterations)
# nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training
last_opt_step = -1
# 初始化maps(每个类别的map)和results
maps = np.zeros(nc) # mAP per class
results = (0, 0, 0, 0, 0, 0, 0) # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)
# 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
scheduler.last_epoch = start_epoch - 1 # do not move
# 设置amp混合精度训练
scaler = torch.cuda.amp.GradScaler(enabled=amp)
# 早停止
stopper, stop = EarlyStopping(patience=opt.patience), False
# 初始化损失
compute_loss = ComputeLoss(model) # init loss class
callbacks.run('on_train_start')
LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
f"Logging results to {colorstr('bold', save_dir)}\n"
f'Starting training for {epochs} epochs...')
# 正式开始训练
for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
callbacks.run('on_train_epoch_start')
model.train()
# Update image weights (optional, single-GPU only)
if opt.image_weights:
"""
如果设置图片采样策略
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
通过random.choices生成图片所有indices从而进行采样
"""
cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc # class weights
iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights
dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx
# Update mosaic border (optional)
# b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
# dataset.mosaic_border = [b - imgsz, -b] # height, width borders
mloss = torch.zeros(3, device=device) # mean losses
if RANK != -1:
train_loader.sampler.set_epoch(epoch)
pbar = enumerate(train_loader)
LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'obj_loss', 'cls_loss', 'Instances', 'Size'))
if RANK in {-1, 0}:
# 进度条显示
pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT) # progress bar
optimizer.zero_grad() # 梯度清零
for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------
callbacks.run('on_train_batch_start')
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device, non_blocking=True).float() / 255 # uint8 to float32, 0-255 to 0.0-1.0
"""
热身训练(前nw次迭代，一般是3)
在前nw次迭代中，根据以下方式选取accumulate和学习率
"""
# Warmup
if ni <= nw:
xi = [0, nw] # x interp
# compute_loss.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou)
accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
"""
bias的学习率从0.1下降到基准学习率lr*lf(epoch)，
其他的参数学习率从0增加到lr*lf(epoch).
lf为上面设置的余弦退火的衰减函数
动量momentum也从0.9慢慢变到hyp['momentum'](default=0.937)
"""
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
# Multi-scale
if opt.multi_scale:
"""
Multi-scale 设置多尺度训练，从imgsz * 0.5, imgsz * 1.5 + gs随机选取尺寸
"""
sz = random.randrange(int(imgsz * 0.5), int(imgsz * 1.5) + gs) // gs * gs # size
sf = sz / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple)
imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
# Forward
with torch.cuda.amp.autocast(amp):
pred = model(imgs) # forward
# 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
loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size
if RANK != -1:
# 平均不同gpu之间的梯度
loss *= WORLD_SIZE # gradient averaged between devices in DDP mode
if opt.quad:
loss *= 4.
# Backward
scaler.scale(loss).backward()
# Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
# 模型反向传播accumulate次之后再根据累计的梯度更新一次参数
if ni - last_opt_step >= accumulate:
scaler.unscale_(optimizer) # unscale gradients
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0) # clip gradients
scaler.step(optimizer) # optimizer.step
scaler.update()
optimizer.zero_grad()
if ema:
ema.update(model)
last_opt_step = ni
# Log
if RANK in {-1, 0}:
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G' # (GB)
pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
(f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths, list(mloss))
if callbacks.stop_training:
return
# end batch ------------------------------------------------------------------------------------------------
# Scheduler 学习率衰减
lr = [x['lr'] for x in optimizer.param_groups] # for loggers
scheduler.step()
if RANK in {-1, 0}:
# mAP
callbacks.run('on_train_epoch_end', epoch=epoch)
# 把model中的属性赋值给ema
ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
# 判断是否是最后一轮
final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
# 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
if not noval or final_epoch: # Calculate mAP
# 测试使用的是ema(对模型的参数做平均)模型
# verbose设置为true后，每轮的验证都输出每个类别的信息
results, maps, _ = validate.run(data_dict,
batch_size=batch_size // WORLD_SIZE * 2,
imgsz=imgsz,
half=amp,
model=ema.ema,
single_cls=single_cls,
dataloader=val_loader,
save_dir=save_dir,
plots=False,
callbacks=callbacks,
compute_loss=compute_loss,
verbose=True)
# Update best mAP
fi = fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
stop = stopper(epoch=epoch, fitness=fi) # early stop check
if fi > best_fitness:
best_fitness = fi
log_vals = list(mloss) + list(results) + lr
callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
# Save model
"""
保存带checkpoint的模型用于inference或resuming training
保存模型的同时还保存epoch，results，optimizer等信息
optimizer在最后一轮不会报错
model保存的是EMA后的模型
"""
if (not nosave) or (final_epoch and not evolve): # if save
ckpt = {
'epoch': epoch,
'best_fitness': best_fitness,
'model': deepcopy(de_parallel(model)).half(),
'ema': deepcopy(ema.ema).half(),
'updates': ema.updates,
'optimizer': optimizer.state_dict(),
'opt': vars(opt),
'git': GIT_INFO, # {remote, branch, commit} if a git repo
'date': datetime.now().isoformat()}
# Save last, best and delete
torch.save(ckpt, last)
if best_fitness == fi:
torch.save(ckpt, best)
if opt.save_period > 0 and epoch % opt.save_period == 0:
torch.save(ckpt, w / f'epoch{epoch}.pt')
del ckpt
callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
# EarlyStopping
if RANK != -1: # if DDP training
broadcast_list = [stop if RANK == 0 else None]
dist.broadcast_object_list(broadcast_list, 0) # broadcast 'stop' to all ranks
if RANK != 0:
stop = broadcast_list[0]
if stop:
break # must break all DDP ranks
# end epoch ----------------------------------------------------------------------------------------------------
# end training --------------------------3.10 train函数——打印训练信息
############################ 打印训练信息 #########################
if RANK in {-1, 0}:
LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
for f in last, best:
if f.exists():
strip_optimizer(f) # strip optimizers
if f is best:
LOGGER.info(f'\nValidating {f}...')
results, _, _ = validate.run(
data_dict,
batch_size=batch_size // WORLD_SIZE * 2,
imgsz=imgsz,
model=attempt_load(f, device).half(),
iou_thres=0.65 if is_coco else 0.60, # best pycocotools at iou 0.65
single_cls=single_cls,
dataloader=val_loader,
save_dir=save_dir,
save_json=is_coco,
verbose=True,
plots=plots,
callbacks=callbacks,
compute_loss=compute_loss) # val best model with plots
if is_coco:
callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
callbacks.run('on_train_end', last, best, epoch, results)
torch.cuda.empty_cache() # 释放显存
return results
????????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