train_custom.py

import argparse
import logging
import math
import os
import random
import time
from copy import deepcopy
from pathlib import Path
from threading import Thread

import numpy as np
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
import torch.utils.data
import yaml
from torch.cuda import amp
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm

import test  # import test.py to get mAP after each epoch
from models.experimental import attempt_load
from models.yolo import Model
from utils.autoanchor import check_anchors
from utils.datasets import create_dataloader
from utils.general import labels_to_class_weights, increment_path, labels_to_image_weights, init_seeds, \
    fitness, fitness2, strip_optimizer, get_latest_run, check_dataset, check_file, check_git_status, check_img_size, \
    check_requirements, print_mutation, set_logging, one_cycle, colorstr
from utils.google_utils import attempt_download
from utils.loss import ComputeLoss, SegmentationLosses, SegFocalLoss, OhemCELoss, ProbOhemCrossEntropy2d
from utils.plots import plot_images, plot_labels, plot_results, plot_evolution
from utils.torch_utils import ModelEMA, select_device, intersect_dicts, torch_distributed_zero_first, is_parallel
from utils.wandb_logging.wandb_utils import WandbLogger, check_wandb_resume
import SegmentationDataset
import torch.backends.cudnn as cudnn


logger = logging.getLogger(__name__)


def train(hyp, opt, device, tb_writer=None):
    logger.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))  # 打印超参数
    save_dir, epochs, batch_size, total_batch_size, weights, rank = \
        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank

    # Directories
    wdir = save_dir / 'weights'
    wdir.mkdir(parents=True, exist_ok=True)  # make dir
    last = wdir / 'last.pt'
    best = wdir / 'best.pt'
    results_file = save_dir / 'results.txt'

    # Save run settings # 存超参数和优化器参数, 优化器参数,可用于resume
    with open(save_dir / 'hyp.yaml', 'w') as f:
        yaml.dump(hyp, f, sort_keys=False)
    with open(save_dir / 'opt.yaml', 'w') as f:
        yaml.dump(vars(opt), f, sort_keys=False)

    # Configure
    plots = not opt.evolve  # create plots  不进化就画图
    cuda = device.type != 'cpu'
    init_seeds(2 + rank)
    with open(opt.data) as f:
        data_dict = yaml.load(f, Loader=yaml.SafeLoader)  # data dict
    is_coco = opt.data.endswith('coco.yaml')

    # Logging- Doing this before checking the dataset. Might update data_dict
    loggers = {'wandb': None}  # loggers dict
    if rank in [-1, 0]:  # -1不开DDP, 0是DDP主进程
        opt.hyp = hyp  # add hyperparameters
        run_id = torch.load(weights).get('wandb_id') if weights.endswith('.pt') and os.path.isfile(weights) else None
        wandb_logger = WandbLogger(opt, Path(opt.save_dir).stem, run_id, data_dict)
        loggers['wandb'] = wandb_logger.wandb
        data_dict = wandb_logger.data_dict
        if wandb_logger.wandb:
            weights, epochs, hyp = opt.weights, opt.epochs, opt.hyp  # WandbLogger might update weights, epochs if resuming

    nc = 1 if opt.single_cls else int(data_dict['nc'])  # number of classes
    names = ['item'] if opt.single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
    assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (len(names), nc, opt.data)  # check

    # Model
    pretrained = weights.endswith('.pt')  # 有weights输入就用其初始化
    if pretrained:  # 有预训练参数
        with torch_distributed_zero_first(rank):
            attempt_download(weights)  # download if not found locally
        ckpt = torch.load(weights, map_location=device)  # load checkpoint
        model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
        exclude = ['anchor'] if (opt.cfg or hyp.get('anchors')) and not opt.resume else []  # exclude keys 初始化时候不使用的参数(非resume且有配置时按cfg和model初始化来指定anchor, 否则延用pretrained weights的anchor)
        state_dict = ckpt['model'].float().state_dict()  # to FP32 ckpt里model键对应的值才是模型,训练结束后保存的是float16(中间保存的float32),模型的state_dict是参数,包括可训练参数和register_buffer保存的buffer parameter(Detect的anchor和gird等)
        state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude)  # intersect 赋值参数,预训练的参数赋值到新建的模型,exclude除外(即anchor使用cfg的而不是预训练)
        model.load_state_dict(state_dict, strict=False)  # load 调整好的预训练参数加载到模型
        logger.info('Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights))  # report
    else:  # 无预训练参数,只建模型
        model = Model(opt.cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
    with torch_distributed_zero_first(rank):
        check_dataset(data_dict)  # check
    train_path = data_dict['train']
    test_path = data_dict['val']
    segtrain_path = data_dict['segtrain']
    segval_path = data_dict['segval']

    # Freeze 要冻结的参数 似乎只能在此代码处手动设置列表
    freeze = []  # parameter names to freeze (full or partial)
    for k, v in model.named_parameters():
        v.requires_grad = True  # train all layers 全部参数可导
        if any(x in k for x in freeze):  # 碰见freeze列表的层使其参数不可导
            print('freezing %s' % k)
            v.requires_grad = False

    # Optimizer
    nbs = 64  # nominal batch size
    accumulate = max(round(nbs / total_batch_size), 1)  # accumulate loss before optimizing
    hyp['weight_decay'] *= total_batch_size * accumulate / nbs  # scale weight_decay 权重衰减系数
    logger.info(f"Scaled weight_decay = {hyp['weight_decay']}")

    # 参数分组,pg0是BN,pg1是权重,pg2是偏置
    pg0, pg1, pg2 = [], [], []  # optimizer parameter groups
    for k, v in model.named_modules():
        if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter):
            pg2.append(v.bias)  # biases
        if isinstance(v, nn.BatchNorm2d):
            pg0.append(v.weight)  # no decay
        elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter):
            pg1.append(v.weight)  # apply decay
    # 优化器初始化, BN层参数不带权重衰减
    if opt.adam:
        optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999))  # adjust beta1 to momentum
    else:
        optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
    # 权重参数, 带权重衰减
    optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']})  # add pg1 with weight_decay
    # 偏置参数, 同样不带衰减
    optimizer.add_param_group({'params': pg2})  # add pg2 (biases)
    logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
    del pg0, pg1, pg2

    # Scheduler https://arxiv.org/pdf/1812.01187.pdf 设置好优化器后用其设置Learning Scheduler
    # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR
    if opt.linear_lr:
        lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
    else:
        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # 自定义lambda函数学习率衰减策略
    # plot_lr_scheduler(optimizer, scheduler, epochs)

    # EMA 指数滑动平均
    ema = ModelEMA(model) if rank in [-1, 0] else None

    # Resume
    start_epoch, best_fitness = 0, 0.0
    if pretrained:
        # Optimizer
        if ckpt['optimizer'] is not None:
            optimizer.load_state_dict(ckpt['optimizer'])
            best_fitness = ckpt['best_fitness']

        # EMA 指数平均
        if ema and ckpt.get('ema'):
            ema.ema.load_state_dict(ckpt['ema'].float().state_dict())
            ema.updates = ckpt['updates']

        # Results
        if ckpt.get('training_results') is not None:
            results_file.write_text(ckpt['training_results'])  # write results.txt

        # Epochs
        start_epoch = ckpt['epoch'] + 1  # 预训练模型的epoch是-1
        if opt.resume:  # resume参数epoch应该大于0
            assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (weights, epochs)
        if epochs < start_epoch:  # 总轮数比开始轮还小, 总轮数加上已训练轮(即再训练总轮数次而不是通常的 总轮数-开始轮 次)
            logger.info('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' %
                        (weights, ckpt['epoch'], epochs))
            epochs += ckpt['epoch']  # finetune additional epochs

        del ckpt, state_dict

    # Image sizes
    gs = max(int(model.stride.max()), 32)  # grid size (max stride)  至少32
    nl = model.model[-1].nl  # number of detection layers (used for scaling hyp['obj']) model最后一层是Detect, nl是其输出层数量
    imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size]  # verify imgsz are gs-multiples 检查图片尺寸是否合法,不合法就自动替换

    # DP mode DP多线程数据并行模式, 不使用, 并行推荐DDP多进程
    if cuda and rank == -1 and torch.cuda.device_count() > 1:
        model = torch.nn.DataParallel(model)

    # SyncBatchNorm 跨卡BN, 仅支持DDP
    if opt.sync_bn and cuda and rank != -1:
        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
        logger.info('Using SyncBatchNorm()')

    # 检测 Trainloader
    dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt,
                                            hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank,
                                            world_size=opt.world_size, workers=opt.workers,
                                            image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: '))
    mlc = np.concatenate(dataset.labels, 0)[:, 0].max()  # max label class 纵向连接了标签后找第一列最大值, mlc的值就是类别数-1
    nb = len(dataloader)  # number of batches
    # mlc=实际标签类别数-1 应该小于 nc模型结构支持的前景类别数 (不用等式关系, 因为结构类别多的模型可以支持训练标签类别少的数据, 反之不成立)
    assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (mlc, nc, opt.data, nc - 1)

    # Process 0  非DDP或DDP中的主进程
    if rank in [-1, 0]:
        testloader = create_dataloader(test_path, imgsz_test, batch_size * 2, gs, opt,  # testloader batch_size翻倍
                                       hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1,
                                       world_size=opt.world_size, workers=opt.workers,
                                       pad=0.5, prefix=colorstr('val: '))[0]  # [0]只要了loader没要dataset, 和train处理不一样

        if not opt.resume:  # 常规, 非resume
            labels = np.concatenate(dataset.labels, 0)
            c = torch.tensor(labels[:, 0])  # classes 所有对象类别(包括所有目标,不是图像)
            # cf = torch.bincount(c.long(), minlength=nc) + 1.  # frequency
            # model._initialize_biases(cf.to(device))
            if plots:
                plot_labels(labels, names, save_dir, loggers)
                if tb_writer:
                    tb_writer.add_histogram('classes', c, 0)  #

            # Anchors
            if not opt.noautoanchor:  # 用train dataset自动聚类选取最好anchor
                check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # anchor_t是最大放大倍数,yolov5公式不同于v3v4, 见核心Model推理时anchor偏移放缩公式和issue
            model.half().float()  # pre-reduce anchor precision 先转float16再转回32,虽然type是32,但此时参数的数值范围限到16了

    # 图尺寸相同时候用这个准确测指标（batch_size设为1可以支持图尺寸不同, 或者用下面的val mode）
        seg_valloader = SegmentationDataset.get_custom_loader(root=segval_path, batch_size=1,
                                                         split="val", mode="testval",  # 旧版为val新版训练中验证也用testval模式
                                                         base_size=imgsz,   # 原图按照长边resize到imgz输入后双线性插值到原图尺寸计算精度
                                                         # crop_size=640,  # testval 时候cropsize不起作用
                                                         workers=2, pin=True)  # 验证batch_size和workers得配合, 都太大会导致子进程死亡, 单进程龟速加载数据
                                                                     # 我电脑上(4,4)是最快的, 更大子进程会挂(现在图大了,怎么设都会挂, BUG)

    # # 图尺寸不同时候改用val模式,一般会比标准目标输入低一点
    #     seg_valloader = SegmentationDataset.get_citysbdd_loader(root=segval_path, batch_size=4,
    #                                                      split="val", mode="val",  # 和train.py不同，使用val模式
    #                                                      base_size=1024,   # val模式base_size无效
    #                                                      # crop_size手动取，建议目标输入的短边尺寸，如cityscapes取512
    #                                                      crop_size=512,  # 图尺寸不同，用val，按短边resize到cropsize再crop（cropsize，cropsize）
    #                                                      workers=4, pin=True)  # 验证batch_size和workers得配合, 都太大会导致子进程死亡, 单进程龟速加载数据
    #                                                                  # 我电脑上(4,4)是最快的, 更大子进程会挂(现在图大了,怎么设都会挂, BUG)
    
    # 分割 loader custom的base_size和crop_size的长边就是imgsz
    seg_trainloader = SegmentationDataset.get_custom_loader(root=segtrain_path,
                                                           split="train", mode="train",
                                                           base_size=imgsz,
                                                           # custom的crop_size就是(imgsz, imgsz)
                                                           batch_size=batch_size,
                                                           workers=opt.workers, pin=True)

    segnb = len(seg_trainloader)
    # DDP mode
    if cuda and rank != -1:  # 没禁用(-1)就开DDP模型
        model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank,
                    # nn.MultiheadAttention incompatibility with DDP https://github.com/pytorch/pytorch/issues/26698
                    find_unused_parameters=any(isinstance(layer, nn.MultiheadAttention) for layer in model.modules()))

    # Model parameters 根据输出层数,类别数等调整损失增益,模型超参数
    hyp['box'] *= 3. / nl  # scale to layers
    hyp['cls'] *= nc / 80. * 3. / nl  # scale to classes and layers
    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
    model.gr = 1.0  # iou loss ratio (obj_loss = 1.0 or iou)
    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
    model.names = names

    # Start training
    t0 = time.time()
    nw = max(round(hyp['warmup_epochs'] * nb), 500)  # number of warmup iterations, max(3 epochs, 1k iterations) 最少warmup三轮或500batch(原版1000,800就够了)
    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
    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)
    scheduler.last_epoch = start_epoch - 1  # do not move 配置lr_scheduler起始位置
    scaler = amp.GradScaler(enabled=cuda)  # 说明不是float16训练,而是16和32混合精度训练. 训练前初始化loss scaler 用于float16放大梯度后backward, optimizer.step之前自动转float32再缩回来
    compute_loss = ComputeLoss(model)  # init loss class 初始化检测criteria
  
# -----------------------------------------------------------------------------------------------------------
    # 无aux模型输出不用[],有aux几个结果输出用[]包装
    # Base，PSP和Lab用这个，无aux
    compute_seg_loss = SegmentationLosses(aux=False, ignore_index=-1, weight=None).cuda()
    # compute_seg_loss = SegFocalLoss(ignore_index=-1, gamma=2, reduction="mean").cuda()
    # BiSe用这个　两个aux
    # compute_seg_loss = SegmentationLosses(nclass=19, aux=True, aux_num=2, aux_weight=0.1, ignore_index=-1, weight=None).cuda()
    # 一个aux，没有用这个
    # compute_seg_loss = SegmentationLosses(nclass=19, aux=True, aux_num=1, aux_weight=0.1, ignore_index=-1, weight=None).cuda()
# -----------------------------------------------------------------------------------------------------------

    # focalloss别用，cityscapes效果不行
    # OHEM能用，理论上应该超过CE，但是目前实验效果不如CE(设成默认0.7收敛蛮快的但最终值不够好)，认为与计算像素个数和学习率有关，用的话循环损失计算的语句得改一下，接口和CE还没来得及保持一致
    # compute_seg_loss = OhemCELoss(thresh=0.7, ignore_index=-1, aux=False)
    # compute_seg_loss = OhemCELoss(thresh=0.7, ignore_index=-1, aux=True, aux_weight=[0.15, 0.1])

    detgain, seggain = 0.6, 0.35  # 检测, 分割比例  
    # CE、1/8单输入、batchsize13用0.65,0.35左右,注意64向下取整的梯度积累，比13*4=52大(12*5=64)通常应该降低分割损失比例或调小学习率



    logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n'
                f'Using {dataloader.num_workers} dataloader workers\n'
                f'Logging results to {save_dir}\n'
                f'Starting training for {epochs} epochs...')
    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
        mIoU = 0  # 每轮开始mIoU设置成０，因为选模型按ｍIoU选，为了加速训练可能ｎ轮才测一次mIoU，对没测mIoU的模型不会存为best.pt
        print(f'accumulate: {accumulate}')  # 显示epoch开始时梯度积累次数(第一个值忽略, 注意warmup期间按batch变化, 此处只是辅助观察防梯度爆炸)
        model.train()  # epoch开始, 确保train模式 注意validation时候可能会把模型.eval()因此开始的train()很有必要

        # Update image weights (optional) 更新image_weights权重, 默认不开image_weights忽略此块代码
        if opt.image_weights:
            # Generate indices
            if rank in [-1, 0]:
                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
            # Broadcast if DDP
            if rank != -1:
                indices = (torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n)).int()
                dist.broadcast(indices, 0)
                if rank != 0:
                    dataset.indices = indices.cpu().numpy()

        # Update mosaic border
        # 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(4, device=device)  # 检测 mean losses
        msegloss = torch.zeros(1, device=device)  # 混合的 mean losses, 两者计算也可知分割loss
        if rank != -1:
            dataloader.sampler.set_epoch(epoch)  # shuffle时, 保证每个epoch顺序不同
        pbar = enumerate(dataloader)
        segpbar = enumerate(seg_trainloader)
        logger.info(('\n' + '%10s' * 9) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'seg', 'labels', 'img_size'))
        if rank in [-1, 0]:
            pbar = tqdm(pbar, total=min(nb, segnb))  # progress bar # tqdm进度条迭代
            segpbar = tqdm(segpbar, total=min(nb, segnb))
        optimizer.zero_grad()  # 每轮前清空梯度

        # 暂时用zip, 每轮batch数以数量少的为准
        for det_batch, seg_batch in zip(pbar, segpbar):  # batch -------------------------------------------------------------
            i, (imgs, targets, paths, _) = det_batch  # 检测
            _, (segimgs, segtargets) = seg_batch   # 分割
            if len(imgs)==1 or len(segimgs)==1:  # 手动droplast,SE或者gloablpool后的bn不支持单个样本，检测loader调用地方太多不好droplast，这里手动
                continue
            # warmup等参数变化以检测为准
            ni = i + nb * epoch  # number integrated batches (since train start) 记录总iterations, 可以用于停止warmup
            imgs = imgs.to(device, non_blocking=True).float() / 255.0  # uint8 to float32, 0-255 to 0.0-1.0
            # Warmup
            if ni <= nw:
                xi = [0, nw]  # x interp
                # model.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)  # 修改了accumulate上限,使其不超过nbs(防止Nan)
                accumulate = max(1, np.interp(ni, xi, [1, math.floor(nbs / total_batch_size)]).round())  # 梯度积累 线性插值xi=[0, 1000], yi=[1, 64/batchsize], 插入点x=ni, 之后取整, 最小限1. warmup时accumulate会逐渐从1按整数增大到目标, warmup结束后稳定在目标值 round(nbs/accumulate), 例如batchsize32实际上两batch才更新一次,等效于64
                for j, x in enumerate(optimizer.param_groups):  # warmup过程中逐渐把三组参数的lr调到lr0
                    # 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 == 2 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 默认关multi scale
            if opt.multi_scale:
                sz = random.randrange(imgsz * 0.5, 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 = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)

            # Forward and Backward 对比原版yolov5此处修改, 否则batchsize只能取单检测时候的一半, 这种写法可以更大一点

            with amp.autocast(enabled=cuda):  # 混合精度训练中用来代替autograd
                pred = model(imgs)  # forward
                loss, loss_items = compute_loss(pred[0], targets.to(device))  # loss scaled by batch_size
                if rank != -1:  # DDP中loss * GPU数
                    loss *= opt.world_size  # gradient averaged between devices in DDP mode
                if opt.quad:
                    loss *= 4.
                loss *= detgain  # 检测loss比例
            scaler.scale(loss).backward()
            imgshape = imgs.shape[-1]
            if plots and ni >= 3:
                del imgs  # 前三个batch画图不能del
            else:
                imgs = imgs.to(torch.device('cpu'), non_blocking=True)  # 释放 segimgs输入后就没被调用会被pytorch自动回收不用手动释放(img后续有被调用要手动释放)
            
            segimgs = segimgs.to(device, non_blocking=True)  # 分割已经做过totensor了, 不用/255

            with amp.autocast(enabled=cuda):  # 混合精度训练中用来代替autograd
                pred = model(segimgs)
# -----------------------------------------------------------------------------------------------------------
                # 无aux模型输出不用[],有aux模型几个结果输出用[]包装
                # Base,PSP和Lab用这个,无aux
                segloss = compute_seg_loss(pred[1], segtargets.to(device)) * batch_size # 分割loss CE是平均loss, 配合检测做梯度积累, 因此乘以batchsize(注意有梯度积累其真实batchsize约是nbs=64)  
                # Bise用这个,两个aux   
                # segloss = compute_seg_loss(pred[1][0], pred[1][1], pred[1][2], segtargets.to(device)) * batch_size    
                # 一个aux，没有用这个
                # segloss = compute_seg_loss(pred[1][0], pred[1][1], segtargets.to(device)) * batch_size   
# -----------------------------------------------------------------------------------------------------------
                segloss *= seggain
            scaler.scale(segloss).backward()
            del segimgs

            # Optimize
            if ni % accumulate == 0:  # 梯度积累accumulate次后才优化,
                scaler.step(optimizer)  # optimizer.step  # 混合精度训练优化时用scaler
                scaler.update()
                optimizer.zero_grad()  # 每次更新完参数才清空梯度, 不更新时累计
                if ema:  # 不开DDP和DDP主进程中ema开启, 每次更新ema
                    ema.update(model)

            # Print
            if rank in [-1, 0]:
                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
                msegloss = (msegloss * i + segloss.detach()/total_batch_size) / (i + 1)
                mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0)  # (GB)
                s = ('%10s' * 2 + '%10.4g' * 7) % (
                    '%g/%g' % (epoch, epochs - 1), mem, *mloss, msegloss, targets.shape[0], imgshape)
                pbar.set_description(s)

                # Plot
                if plots and ni < 3:
                    f = save_dir / f'train_batch{ni}.jpg'  # filename
                    Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start()
                    # if tb_writer:
                    #     tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
                    #     tb_writer.add_graph(torch.jit.trace(model, imgs, strict=False), [])  # add model graph
                elif plots and ni == 10 and wandb_logger.wandb:
                    wandb_logger.log({"Mosaics": [wandb_logger.wandb.Image(str(x), caption=x.name) for x in
                                                  save_dir.glob('train*.jpg') if x.exists()]})

            # end batch ------------------------------------------------------------------------------------------------
        # end epoch ----------------------------------------------------------------------------------------------------

        # Scheduler
        lr = [x['lr'] for x in optimizer.param_groups]  # for tensorboard
        scheduler.step()  # 更新Scheduler


        # DDP process 0 or single-GPU
        if rank in [-1, 0]:
            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride', 'class_weights'])
            # pixACC, mIoU
            if epoch % 10 == 0 or (epochs - epoch) < 40:
                mIoU = test.seg_validation(model=ema.ema, valloader=seg_valloader, device=device, n_segcls=19,
                                half_precision=True)
            # mAP
            final_epoch = epoch + 1 == epochs  # 是否是最后一轮
            if not opt.notest or final_epoch:  # Calculate mAP
                wandb_logger.current_epoch = epoch + 1
                results, maps, times = test.test(data_dict,
                                                 batch_size=batch_size * 2,
                                                 imgsz=imgsz_test,
                                                 model=ema.ema,
                                                 single_cls=opt.single_cls,
                                                 dataloader=testloader,
                                                 save_dir=save_dir,
                                                 verbose=nc < 50 and final_epoch,
                                                 plots=plots and final_epoch,
                                                 wandb_logger=wandb_logger,
                                                 compute_loss=compute_loss,
                                                 is_coco=is_coco)

            # Write
            with open(results_file, 'a') as f:
                f.write(s + '%10.4g' * 7 % results + '\n')  # append metrics, val_loss
            if len(opt.name) and opt.bucket:
                os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name))

            # Log
            tags = ['train/box_loss', 'train/obj_loss', 'train/cls_loss',  # train loss
                    'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95',
                    'val/box_loss', 'val/obj_loss', 'val/cls_loss',  # val loss
                    'x/lr0', 'x/lr1', 'x/lr2']  # params
            for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):  # 写tensorboard
                if tb_writer:
                    tb_writer.add_scalar(tag, x, epoch)  # tensorboard
                if wandb_logger.wandb:
                    wandb_logger.log({tag: x})  # W&B

            # Update best mIoU  #mAP
            # fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95] 按0.1*AP.5+0.9*AP.5:.95指标衡量模型
            fi = fitness2(np.array(results).reshape(1, -1), mIoU)  # weighted combination of [P, R, mAP@.5, mAP@.5-.95] 按0.1*AP.5+0.9*AP.5:.95指标衡量模型

            if fi > best_fitness:
                best_fitness = fi
            wandb_logger.end_epoch(best_result=best_fitness == fi)

            # Save model
            if (not opt.nosave) or (final_epoch and not opt.evolve):  # if save
                ckpt = {'epoch': epoch,
                        'best_fitness': best_fitness,
                        'training_results': results_file.read_text(),
                        'model': deepcopy(model.module if is_parallel(model) else model).half(),
                        'ema': deepcopy(ema.ema).half(),
                        'updates': ema.updates,
                        'optimizer': optimizer.state_dict(),
                        'wandb_id': wandb_logger.wandb_run.id if wandb_logger.wandb else None}

                # Save last, best and delete
                torch.save(ckpt, last)
                if best_fitness == fi:
                    torch.save(ckpt, best)
                if wandb_logger.wandb:
                    if ((epoch + 1) % opt.save_period == 0 and not final_epoch) and opt.save_period != -1:
                        wandb_logger.log_model(
                            last.parent, opt, epoch, fi, best_model=best_fitness == fi)
                del ckpt

        # end epoch ----------------------------------------------------------------------------------------------------
    # end training
    if rank in [-1, 0]:
        # Plots
        if plots:  # 不进化就画图
            plot_results(save_dir=save_dir)  # save as results.png
            if wandb_logger.wandb:
                files = ['results.png', 'confusion_matrix.png', *[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')]]
                wandb_logger.log({"Results": [wandb_logger.wandb.Image(str(save_dir / f), caption=f) for f in files
                                              if (save_dir / f).exists()]})
        # Test best.pt
        logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
        if opt.data.endswith('coco.yaml') and nc == 80:  # if COCO
            for m in (last, best) if best.exists() else (last):  # speed, mAP tests
                results, _, _ = test.test(opt.data,
                                          batch_size=batch_size * 2,
                                          imgsz=imgsz_test,
                                          conf_thres=0.001,
                                          iou_thres=0.7,
                                          model=attempt_load(m, device).half(),
                                          single_cls=opt.single_cls,
                                          dataloader=testloader,
                                          save_dir=save_dir,
                                          save_json=True,
                                          plots=False,
                                          is_coco=is_coco)

        # Strip optimizers
        final = best if best.exists() else last  # final model
        for f in last, best:
            if f.exists():
                strip_optimizer(f)  # strip optimizers
        if opt.bucket:
            os.system(f'gsutil cp {final} gs://{opt.bucket}/weights')  # upload
        if wandb_logger.wandb and not opt.evolve:  # Log the stripped model
            wandb_logger.wandb.log_artifact(str(final), type='model',
                                            name='run_' + wandb_logger.wandb_run.id + '_model',
                                            aliases=['last', 'best', 'stripped'])
        wandb_logger.finish_run()
    else:
        dist.destroy_process_group()
    torch.cuda.empty_cache()
    return results


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='yolov5s.pt', help='initial weights path')
    parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
    parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path')
    parser.add_argument('--hyp', type=str, default='data/hyp.scratch.yaml', help='hyperparameters path')
    parser.add_argument('--epochs', type=int, default=300)
    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs')
    parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='[train, test] image sizes')
    parser.add_argument('--rect', action='store_true', help='rectangular training')
    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
    parser.add_argument('--notest', action='store_true', help='only test final epoch')
    parser.add_argument('--noautoanchor', action='store_true', help='disable autoanchor check')
    parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters')
    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
    parser.add_argument('--cache-images', action='store_true', help='cache images for faster training')
    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
    parser.add_argument('--adam', action='store_true', help='use torch.optim.Adam() optimizer')
    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
    parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify')
    parser.add_argument('--workers', type=int, default=8, help='maximum number of dataloader workers')
    parser.add_argument('--project', default='runs/train', help='save to project/name')
    parser.add_argument('--entity', default=None, help='W&B entity')
    parser.add_argument('--name', default='exp', help='save to project/name')
    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')
    parser.add_argument('--linear-lr', action='store_true', help='linear LR')
    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
    parser.add_argument('--upload_dataset', action='store_true', help='Upload dataset as W&B artifact table')
    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval for W&B')
    parser.add_argument('--save_period', type=int, default=-1, help='Log model after every "save_period" epoch')
    parser.add_argument('--artifact_alias', type=str, default="latest", help='version of dataset artifact to be used')

    opt = parser.parse_args()

    # Set DDP variables  DDP常规初始化
    opt.world_size = int(os.environ['WORLD_SIZE']) if 'WORLD_SIZE' in os.environ else 1  # 获取总进程数world_size
    opt.global_rank = int(os.environ['RANK']) if 'RANK' in os.environ else -1  # global_rank是所有进程可用的GPU号, local_rank是当前进程对应GPU号
    set_logging(opt.global_rank)
    if opt.global_rank in [-1, 0]:
        # check_git_status()  # 检测git版本,网络不好会卡住,手动关闭
        check_requirements()

    # Resume
    wandb_run = check_wandb_resume(opt)  # wandb有bug,没装
    # 断点重续且没有wandb库
    if opt.resume and not wandb_run:  # resume an interrupted run
        ckpt = opt.resume if isinstance(opt.resume, str) else get_latest_run()  # specified or most recent path 找要续的模型pt
        assert os.path.isfile(ckpt), 'ERROR: --resume checkpoint does not exist'
        apriori = opt.global_rank, opt.local_rank
        with open(Path(ckpt).parent.parent / 'opt.yaml') as f:  # 找 优化器 配置文件
            opt = argparse.Namespace(**yaml.load(f, Loader=yaml.SafeLoader))  # replace
        opt.cfg, opt.weights, opt.resume, opt.batch_size, opt.global_rank, opt.local_rank = '', ckpt, True, opt.total_batch_size, *apriori  # reinstate
        logger.info('Resuming training from %s' % ckpt)
    else:
        # opt.hyp = opt.hyp or ('hyp.finetune.yaml' if opt.weights else 'hyp.scratch.yaml')
        opt.data, opt.cfg, opt.hyp = check_file(opt.data), check_file(opt.cfg), check_file(opt.hyp)  # check files
        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'  # cfg和weights至少有一个
        opt.img_size.extend([opt.img_size[-1]] * (2 - len(opt.img_size)))  # extend to 2 sizes (train, test)
        opt.name = 'evolve' if opt.evolve else opt.name  # project名字,用于保存文件夹
        opt.save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok | opt.evolve)  # increment run

    # DDP mode 数据多进程并行
    opt.total_batch_size = opt.batch_size  # 总batchsize
    device = select_device(opt.device, batch_size=opt.batch_size)  # 设备数
    if opt.local_rank != -1:  # 默认是-1不开启DDP
        assert torch.cuda.device_count() > opt.local_rank
        torch.cuda.set_device(opt.local_rank)
        device = torch.device('cuda', opt.local_rank)
        dist.init_process_group(backend='nccl', init_method='env://')  # distributed backend DDP初始化进程组
        assert opt.batch_size % opt.world_size == 0, '--batch-size must be multiple of CUDA device count'  # 一般一卡开一进程, batchsize可被进程数整除
        opt.batch_size = opt.total_batch_size // opt.world_size  # 每个进程batchsize

    # Hyperparameters 配置超参数
    with open(opt.hyp) as f:
        hyp = yaml.load(f, Loader=yaml.SafeLoader)  # load hyps

    # Train
    logger.info(opt)
    if not opt.evolve:  # 没有用进化算法(默认)
        tb_writer = None  # init loggers
        if opt.global_rank in [-1, 0]:
            prefix = colorstr('tensorboard: ')
            logger.info(f"{prefix}Start with 'tensorboard --logdir {opt.project}', view at http://localhost:6006/")
            tb_writer = SummaryWriter(opt.save_dir)  # Tensorboard
        train(hyp, opt, device, tb_writer)

    # Evolve hyperparameters (optional)
    else:
        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
        meta = {'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
                'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
                'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
                'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
                'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
                'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
                'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
                'box': (1, 0.02, 0.2),  # box loss gain
                'cls': (1, 0.2, 4.0),  # cls loss gain
                'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
                'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
                'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
                'iou_t': (0, 0.1, 0.7),  # IoU training threshold
                'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
                'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
                'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
                'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
                'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
                'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
                'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
                'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
                'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
                'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
                'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
                'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
                'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
                'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
                'mixup': (1, 0.0, 1.0)}  # image mixup (probability)

        assert opt.local_rank == -1, 'DDP mode not implemented for --evolve'
        opt.notest, opt.nosave = True, True  # only test/save final epoch
        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
        yaml_file = Path(opt.save_dir) / 'hyp_evolved.yaml'  # save best result here
        if opt.bucket:
            os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket)  # download evolve.txt if exists

        for _ in range(300):  # generations to evolve
            if Path('evolve.txt').exists():  # if evolve.txt exists: select best hyps and mutate
                # Select parent(s)
                parent = 'single'  # parent selection method: 'single' or 'weighted'
                x = np.loadtxt('evolve.txt', ndmin=2)
                n = min(5, len(x))  # number of previous results to consider
                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
                w = fitness(x) - fitness(x).min()  # weights
                if parent == 'single' or len(x) == 1:
                    # x = x[random.randint(0, n - 1)]  # random selection
                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
                elif parent == 'weighted':
                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination

                # Mutate
                mp, s = 0.8, 0.2  # mutation probability, sigma
                npr = np.random
                npr.seed(int(time.time()))
                g = np.array([x[0] for x in meta.values()])  # gains 0-1
                ng = len(meta)
                v = np.ones(ng)
                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
                    hyp[k] = float(x[i + 7] * v[i])  # mutate

            # Constrain to limits
            for k, v in meta.items():
                hyp[k] = max(hyp[k], v[1])  # lower limit
                hyp[k] = min(hyp[k], v[2])  # upper limit
                hyp[k] = round(hyp[k], 5)  # significant digits

            # Train mutation
            results = train(hyp.copy(), opt, device)

            # Write mutation results
            print_mutation(hyp.copy(), results, yaml_file, opt.bucket)

        # Plot results
        plot_evolution(yaml_file)
        print(f'Hyperparameter evolution complete. Best results saved as: {yaml_file}\n'
              f'Command to train a new model with these hyperparameters: $ python train.py --hyp {yaml_file}')