YOLOV5训练代码train.py注释与解析

YOLOv5代码注释版更新啦，注释的是最近的2021.07.14的版本，且注释更全
github: https://github.com/Laughing-q/yolov5_annotations

YOLOV5训练代码train.py注释与解析

2020.8.31版本
- 超参数文件hyp解析
- 训练参数以及main函数解析
- train函数解析
2020.7.20版本
- 训练参数以及main函数解析
- train函数解析

本文主要对ultralytics\yolov5的训练代码train.py的解析，由于yolov5还在开发当中，平常多多少少都会修复一些bug或者有一些代码和功能的更新，但基本上不会有很大的改动，故以下注释与解析都是适用的；当然如果有大改动，笔者也会更新注释。

yolov5其他代码解析

2020.9.3
1.更新了最新的代码解析注释(其实也不算最最新的,是这周一(8.31)clone的代码, 最近比较忙,今天才把注释完成,主要在于添加了分布式计算的一些代码,以及更新了一些小细节的东西;
2.由于笔者目前还没试用过分布式训练的代码,可能对这方面代码理解不是很好,如有问题欢迎指正,谢谢;
3.以前版本(7.20)的注释我也会留着;

2020.8.31版本

超参数文件hyp解析

# Hyperparameters for VOC finetuning
# python train.py --batch 64 --weights yolov5m.pt --data voc.yaml --img 512 --epochs 50
# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials# Hyperparameter Evolution Results
# Generations: 51
#                   P         R     mAP.5 mAP.5:.95       box       obj       cls
# Metrics:      0.625     0.926      0.89     0.677    0.0111   0.00849   0.00124lr0: 0.00447  # 学习率
lrf: 0.114    # 余弦退火超参数
momentum: 0.873 # 学习率动量
weight_decay: 0.00047 # 权重衰减系数
giou: 0.0306 # giou损失的系数
cls: 0.211 # 分类损失的系数
cls_pw: 0.546 # 分类BCELoss中正样本的权重
obj: 0.421 # 有无物体损失的系数
obj_pw: 0.972 # 有无物体BCELoss中正样本的权重
iou_t: 0.2 # 标签与anchors的iou阈值iou training threshold
anchor_t: 2.26 # 标签的长h宽w/anchor的长h_a宽w_a阈值, 即h/h_a, w/w_a都要在(1/2.26, 2.26)之间anchor-multiple threshold
# anchors: 5.07
fl_gamma: 0.0  # 设为0则表示不使用focal loss(efficientDet default is gamma=1.5)
# 下面是一些数据增强的系数, 包括颜色空间和图片空间
hsv_h: 0.0154 # 色调
hsv_s: 0.9 # 饱和度
hsv_v: 0.619 # 明度
degrees: 0.404 #旋转角度
translate: 0.206　 # 水平和垂直平移
scale: 0.86　  # 缩放
shear: 0.795  # 剪切
perspective: 0.0  # 透视变换参数
flipud: 0.00756  # 上下翻转
fliplr: 0.5  # 左右翻转
mixup: 0.153  # mixup系数

训练参数以及main函数解析

if __name__ == '__main__':"""opt参数解析：cfg:模型配置文件，网络结构data:数据集配置文件，数据集路径，类名等hyp:超参数文件epochs:训练总轮次batch-size:批次大小img-size:输入图片分辨率大小rect:是否采用矩形训练，默认Falseresume:接着打断训练上次的结果接着训练nosave:不保存模型，默认Falsenotest:不进行test，默认Falsenoautoanchor:不自动调整anchor，默认Falseevolve:是否进行超参数进化，默认Falsebucket:谷歌云盘bucket，一般不会用到cache-images:是否提前缓存图片到内存，以加快训练速度，默认Falseweights:加载的权重文件name:数据集名字，如果设置：results.txt to results_name.txt，默认无device:训练的设备，cpu；0(表示一个gpu设备cuda:0)；0,1,2,3(多个gpu设备)multi-scale:是否进行多尺度训练，默认Falsesingle-cls:数据集是否只有一个类别，默认Falseadam:是否使用adam优化器sync-bn:是否使用跨卡同步BN,在DDP模式使用local_rank:gpu编号logdir:存放日志的目录workers:dataloader的最大worker数量"""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='', help='hyperparameters path, i.e. data/hyp.scratch.yaml')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 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('--name', default='', help='renames results.txt to results_name.txt if supplied')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 as single-class dataset')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('--logdir', type=str, default='runs/', help='logging directory')parser.add_argument('--workers', type=int, default=8, help='maximum number of dataloader workers')opt = parser.parse_args()# Set DDP variables"""设置DDP模式的参数world_size:表示全局进程个数global_rank:进程编号"""opt.total_batch_size = opt.batch_sizeopt.world_size = int(os.environ['WORLD_SIZE']) if 'WORLD_SIZE' in os.environ else 1opt.global_rank = int(os.environ['RANK']) if 'RANK' in os.environ else -1set_logging(opt.global_rank)if opt.global_rank in [-1, 0]:# 检查你的代码版本是否为最新的(不适用于windows系统)check_git_status()# Resume# 是否resumeif opt.resume:  # resume an interrupted run# 如果resume是str,则表示传入的是模型的路径地址# get_latest_run()函数获取runs文件夹中最近的last.ptckpt = opt.resume if isinstance(opt.resume, str) else get_latest_run()  # specified or most recent pathlog_dir = Path(ckpt).parent.parent  # runs/exp0assert os.path.isfile(ckpt), 'ERROR: --resume checkpoint does not exist'# opt参数也全部替换with open(log_dir / 'opt.yaml') as f:opt = argparse.Namespace(**yaml.load(f, Loader=yaml.FullLoader))  # replace# opt.cfg设置为'' 对应着train函数里面的操作(加载权重时是否加载权重里的anchor)opt.cfg, opt.weights, opt.resume = '', ckpt, Truelogger.info('Resuming training from %s' % ckpt)else:# 获取超参数列表opt.hyp = opt.hyp or ('data/hyp.finetune.yaml' if opt.weights else 'data/hyp.scratch.yaml')# 检查配置文件信息opt.data, opt.cfg, opt.hyp = check_file(opt.data), check_file(opt.cfg), check_file(opt.hyp)  # check filesassert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'# 扩展image_size为[image_size, image_size]一个是训练size，一个是测试sizeopt.img_size.extend([opt.img_size[-1]] * (2 - len(opt.img_size)))  # extend to 2 sizes (train, test)# 根据opt.logdir生成目录log_dir = increment_dir(Path(opt.logdir) / 'exp', opt.name)  # runs/exp1# 选择设备device = select_device(opt.device, batch_size=opt.batch_size)# DDP mode# DDP 模式if opt.local_rank != -1:assert torch.cuda.device_count() > opt.local_rank# 根据gpu编号选择设备torch.cuda.set_device(opt.local_rank)device = torch.device('cuda', opt.local_rank)# 初始化进程组dist.init_process_group(backend='nccl', init_method='env://')  # distributed backendassert opt.batch_size % opt.world_size == 0, '--batch-size must be multiple of CUDA device count'# 将总批次按照进程数分配给各个gpuopt.batch_size = opt.total_batch_size // opt.world_size# 打印opt参数信息logger.info(opt)# 加载超参数列表with open(opt.hyp) as f:hyp = yaml.load(f, Loader=yaml.FullLoader)  # load hyps# Train# 如果不进行超参数进化，则直接调用train()函数，开始训练if not opt.evolve:tb_writer = Noneif opt.global_rank in [-1, 0]:# 创建tensorboardlogger.info('Start Tensorboard with "tensorboard --logdir %s", view at http://localhost:6006/' % opt.logdir)tb_writer = SummaryWriter(log_dir=log_dir)  # runs/exp0train(hyp, opt, device, tb_writer)# Evolve hyperparameters (optional)else:# Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)# 超参数进化列表,括号里分别为(突变规模, 最小值,最大值)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': (0.1, 0.6, 0.98),  # SGD momentum/Adam beta1'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay'giou': (1, 0.02, 0.2),  # GIoU 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': (1, 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)'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 indicesyaml_file = Path('runs/evolve/hyp_evolved.yaml')  # save best result hereif opt.bucket:os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket)  # download evolve.txt if exists# 默认进化100次"""这里的进化算法是：根据之前训练时的hyp来确定一个base hyp再进行突变；如何根据？通过之前每次进化得到的results来确定之前每个hyp的权重有了每个hyp和每个hyp的权重之后有两种进化方式；1.根据每个hyp的权重随机选择一个之前的hyp作为base hyp，random.choices(range(n), weights=w)2.根据每个hyp的权重对之前所有的hyp进行融合获得一个base hyp，(x * w.reshape(n, 1)).sum(0) / w.sum()evolve.txt会记录每次进化之后的results+hyp每次进化时，hyp会根据之前的results进行从大到小的排序；再根据fitness函数计算之前每次进化得到的hyp的权重再确定哪一种进化方式，从而进行进化"""for _ in range(100):  # generations to evolveif os.path.exists('evolve.txt'):  # if evolve.txt exists: select best hyps and mutate# Select parent(s)# 选择进化方式parent = 'single'  # parent selection method: 'single' or 'weighted'# 加载evolve.txtx = np.loadtxt('evolve.txt', ndmin=2)# 选取至多前5次进化的结果n = min(5, len(x))  # number of previous results to considerx = x[np.argsort(-fitness(x))][:n]  # top n mutations# 根据results计算hyp的权重w = fitness(x) - fitness(x).min()  # weights# 根据不同进化方式获得base hypif parent == 'single' or len(x) == 1:# x = x[random.randint(0, n - 1)]  # random selectionx = x[random.choices(range(n), weights=w)[0]]  # weighted selectionelif parent == 'weighted':x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination# Mutate# 超参数进化mp, s = 0.9, 0.2  # mutation probability, sigmanpr = np.randomnpr.seed(int(time.time()))# 获取突变初始值g = np.array([x[0] for x in meta.values()])  # gains 0-1ng = len(meta)v = np.ones(ng)# 设置突变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)# 将突变添加到base hyp上# [i+7]是因为x中前七个数字为results的指标(P, R, mAP, F1, test_losses=(GIoU, obj, cls))，之后才是超参数hypfor i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)hyp[k] = float(x[i + 7] * v[i])  # mutate# Constrain to limits# 修剪hyp在规定范围里for k, v in meta.items():hyp[k] = max(hyp[k], v[1])  # lower limithyp[k] = min(hyp[k], v[2])  # upper limithyp[k] = round(hyp[k], 5)  # significant digits# Train mutation# 训练results = train(hyp.copy())# Write mutation results"""写入results和对应的hyp到evolve.txtevolve.txt文件每一行为一次进化的结果一行中前七个数字为(P, R, mAP, F1, test_losses=(GIoU, obj, cls))，之后为hyp保存hyp到yaml文件"""print_mutation(hyp.copy(), results, yaml_file, opt.bucket)# Plot resultsplot_evolution(yaml_file)print('Hyperparameter evolution complete. Best results saved as: %s\nCommand to train a new model with these ''hyperparameters: $ python train.py --hyp %s' % (yaml_file, yaml_file))

train函数解析

import argparse
import logging
import math
import os
import random
import shutil
import time
from pathlib import Pathimport numpy as np
import torch.distributed as dist
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 tqdmimport test  # import test.py to get mAP after each epoch
from models.yolo import Model
from utils.datasets import create_dataloader
from utils.general import (torch_distributed_zero_first, labels_to_class_weights, plot_labels, check_anchors, labels_to_image_weights,compute_loss, plot_images, fitness, strip_optimizer, plot_results, get_latest_run, check_dataset, check_file,check_git_status, check_img_size, increment_dir, print_mutation, plot_evolution, set_logging)
from utils.google_utils import attempt_download
from utils.torch_utils import init_seeds, ModelEMA, select_device, intersect_dictslogger = logging.getLogger(__name__)def train(hyp, opt, device, tb_writer=None):logger.info(f'Hyperparameters {hyp}')# 获取记录训练日志的路径"""训练日志包括：权重、tensorboard文件、超参数hyp、设置的训练参数opt(也就是epochs,batch_size等),result.txtresult.txt包括: 占GPU内存、训练集的GIOU loss, objectness loss, classification loss, 总loss, targets的数量, 输入图片分辨率, 准确率TP/(TP+FP),召回率TP/P ; 测试集的mAP50, mAP@0.5:0.95, GIOU loss, objectness loss, classification loss.还会保存batch<3的ground truth"""# 如果设置进化算法则不会传入tb_writer(则为None)，设置一个evolve文件夹作为日志目录log_dir = Path(tb_writer.log_dir) if tb_writer else Path(opt.logdir) / 'evolve'  # logging directory# 设置保存权重的路径wdir = log_dir / 'weights'  # weights directoryos.makedirs(wdir, exist_ok=True)last = wdir / 'last.pt'best = wdir / 'best.pt'# 设置保存results的路径results_file = str(log_dir / 'results.txt')# 获取轮次、批次、总批次(涉及到分布式训练)、权重、进程序号(主要用于分布式训练)epochs, batch_size, total_batch_size, weights, rank = \opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank# Save run settings# 保存hyp和optwith open(log_dir / 'hyp.yaml', 'w') as f:yaml.dump(hyp, f, sort_keys=False)with open(log_dir / 'opt.yaml', 'w') as f:yaml.dump(vars(opt), f, sort_keys=False)# Configurecuda = device.type != 'cpu'# 设置随机种子init_seeds(2 + rank)# 加载数据配置信息with open(opt.data) as f:data_dict = yaml.load(f, Loader=yaml.FullLoader)  # data dict# torch_distributed_zero_first同步所有进程# check_dataset检查数据集，如果没找到数据集则下载数据集(仅适用于项目中自带的yaml文件数据集)with torch_distributed_zero_first(rank):check_dataset(data_dict)  # check# 获取训练集、测试集图片路径train_path = data_dict['train']test_path = data_dict['val']# 获取类别数量和类别名字# 如果设置了opt.single_cls则为一类nc, names = (1, ['item']) if opt.single_cls else (int(data_dict['nc']), data_dict['names'])  # number classes, namesassert len(names) == nc, '%g names found for nc=%g dataset in %s' % (len(names), nc, opt.data)  # check# Modelpretrained = weights.endswith('.pt')# 如果采用预训练if pretrained:# 加载模型，从google云盘中自动下载模型# 但通常会下载失败，建议提前下载下来放进weights目录with torch_distributed_zero_first(rank):attempt_download(weights)  # download if not found locally# 加载检查点ckpt = torch.load(weights, map_location=device)  # load checkpoint# if hyp['anchors']:#     ckpt['model'].yaml['anchors'] = round(hyp['anchors'])  # force autoanchor"""这里模型创建，可通过opt.cfg，也可通过ckpt['model'].yaml这里的区别在于是否是resume，resume时会将opt.cfg设为空，则按照ckpt['model'].yaml创建模型；这也影响着下面是否除去anchor的key(也就是不加载anchor)，如果resume，则加载权重中保存的anchor来继续训练；主要是预训练权重里面保存了默认coco数据集对应的anchor，如果用户自定义了anchor，再加载预训练权重进行训练，会覆盖掉用户自定义的anchor；所以这里主要是设定一个，如果加载预训练权重进行训练的话，就去除掉权重中的anchor，采用用户自定义的；如果是resume的话，就是不去除anchor，就权重和anchor一起加载， 接着训练；参考https://github.com/ultralytics/yolov5/issues/459所以下面设置了intersect_dicts，该函数就是忽略掉exclude中的键对应的值"""model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device)  # create# 如果opt.cfg存在(表示采用预训练权重进行训练)就设置去除anchorexclude = ['anchor'] if opt.cfg else []  # exclude keysstate_dict = ckpt['model'].float().state_dict()  # to FP32state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude)  # intersectmodel.load_state_dict(state_dict, strict=False)  # load# 显示加载预训练权重的的键值对和创建模型的键值对# 如果设置了resume，则会少加载两个键值对(anchors,anchor_grid)logger.info('Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights))  # reportelse:# 创建模型，　ch为输入图片通道model = Model(opt.cfg, ch=3, nc=nc).to(device)  # create# Freeze"""冻结模型层,设置冻结层名字即可具体可以查看https://github.com/ultralytics/yolov5/issues/679但作者不鼓励冻结层,因为他的实验当中显示冻结层不能获得更好的性能,参照:https://github.com/ultralytics/yolov5/pull/707并且作者为了使得优化参数分组可以正常进行,在下面将所有参数的requires_grad设为了True其实这里只是给一个freeze的示例"""freeze = ['', ]  # parameter names to freeze (full or partial)if any(freeze):for k, v in model.named_parameters():if any(x in k for x in freeze):print('freezing %s' % k)v.requires_grad = False# Optimizer"""nbs为模拟的batch_size; 就比如默认的话上面设置的opt.batch_size为16,这个nbs就为64，也就是模型梯度累积了64/16=4(accumulate)次之后再更新一次模型，变相的扩大了batch_size"""nbs = 64  # nominal batch sizeaccumulate = max(round(nbs / total_batch_size), 1)  # accumulate loss before optimizing# 根据accumulate设置权重衰减系数hyp['weight_decay'] *= total_batch_size * accumulate / nbs  # scale weight_decaypg0, pg1, pg2 = [], [], []  # optimizer parameter groups# 将模型分成三组(weight、bn, bias, 其他所有参数)优化for k, v in model.named_parameters():v.requires_grad = Trueif '.bias' in k:pg2.append(v)  # biaseselif '.weight' in k and '.bn' not in k:pg1.append(v)  # apply weight decayelse:pg0.append(v)  # all else# 选用优化器，并设置pg0组的优化方式if opt.adam:optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999))  # adjust beta1 to momentumelse:optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)# 设置weight、bn的优化方式optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']})  # add pg1 with weight_decay# 设置biases的优化方式optimizer.add_param_group({'params': pg2})  # add pg2 (biases)# 打印优化信息logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0)))del pg0, pg1, pg2# 设置学习率衰减，这里为余弦退火方式进行衰减# 就是根据以下公式lf,epoch和超参数hyp['lrf']进行衰减# Scheduler https://arxiv.org/pdf/1812.01187.pdf# https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLRlf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf']  # cosinescheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)# plot_lr_scheduler(optimizer, scheduler, epochs)# Resume# 初始化开始训练的epoch和最好的结果# best_fitness是以[0.0, 0.0, 0.1, 0.9]为系数并乘以[精确度, 召回率, mAP@0.5, mAP@0.5:0.95]再求和所得# 根据best_fitness来保存best.ptstart_epoch, best_fitness = 0, 0.0if pretrained:# Optimizer# 加载优化器与best_fitnessif ckpt['optimizer'] is not None:optimizer.load_state_dict(ckpt['optimizer'])best_fitness = ckpt['best_fitness']# Results# 加载训练结果result.txtif ckpt.get('training_results') is not None:with open(results_file, 'w') as file:file.write(ckpt['training_results'])  # write results.txt# Epochs# 加载训练的轮次start_epoch = ckpt['epoch'] + 1"""如果resume,则备份权重尽管目前resume能够近似100%成功的起作用了,参照:https://github.com/ultralytics/yolov5/pull/756但为了防止resume时出现其他问题,把之前的权重覆盖了,所以这里进行备份,参照:https://github.com/ultralytics/yolov5/pull/765"""if opt.resume:assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (weights, epochs)shutil.copytree(wdir, wdir.parent / f'weights_backup_epoch{start_epoch - 1}')  # save previous weights"""如果新设置epochs小于加载的epoch，则视新设置的epochs为需要再训练的轮次数而不再是总的轮次数"""if epochs < start_epoch:logger.info('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' %(weights, ckpt['epoch'], epochs))epochs += ckpt['epoch']  # finetune additional epochsdel ckpt, state_dict# Image sizes# 获取模型总步长和模型输入图片分辨率gs = int(max(model.stride))  # grid size (max stride)# 检查输入图片分辨率确保能够整除总步长gsimgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size]  # verify imgsz are gs-multiples# DP mode# 分布式训练,参照:https://github.com/ultralytics/yolov5/issues/475# DataParallel模式,仅支持单机多卡# rank为进程编号, 这里应该设置为rank=-1则使用DataParallel模式# rank=-1且gpu数量=1时,不会进行分布式if cuda and rank == -1 and torch.cuda.device_count() > 1:model = torch.nn.DataParallel(model)# SyncBatchNorm# 使用跨卡同步BNif opt.sync_bn and cuda and rank != -1:model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)logger.info('Using SyncBatchNorm()')# Exponential moving average# 为模型创建EMA指数滑动平均,如果GPU进程数大于1,则不创建ema = ModelEMA(model) if rank in [-1, 0] else None# DDP mode# 如果rank不等于-1,则使用DistributedDataParallel模式# local_rank为gpu编号,rank为进程,例如rank=3，local_rank=0 表示第 3 个进程内的第 1 块 GPU。if cuda and rank != -1:model = DDP(model, device_ids=[opt.local_rank], output_device=(opt.local_rank))# Trainloader# 创建训练集dataloaderdataloader, 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)"""获取标签中最大的类别值，并于类别数作比较如果大于类别数则表示有问题"""mlc = np.concatenate(dataset.labels, 0)[:, 0].max()  # max label classnb = len(dataloader)  # number of batchesassert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (mlc, nc, opt.data, nc - 1)# Testloaderif rank in [-1, 0]:# 更新ema模型的updates参数,保持ema的平滑性ema.updates = start_epoch * nb // accumulate  # set EMA updates# 创建测试集dataloadertestloader = create_dataloader(test_path, imgsz_test, total_batch_size, gs, opt,hyp=hyp, augment=False, cache=opt.cache_images, rect=True, rank=-1,world_size=opt.world_size, workers=opt.workers)[0]  # only runs on process 0# Model parameters# 根据自己数据集的类别数设置分类损失的系数hyp['cls'] *= nc / 80.  # scale coco-tuned hyp['cls'] to current dataset# 设置类别数，超参数model.nc = nc  # attach number of classes to modelmodel.hyp = hyp  # attach hyperparameters to model"""设置giou的值在objectness loss中做标签的系数, 使用代码如下tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)这里model.gr=1，也就是说完全使用标签框与预测框的giou值来作为该预测框的objectness标签"""model.gr = 1.0  # giou loss ratio (obj_loss = 1.0 or giou)# 根据labels初始化图片采样权重model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device)  # attach class weights# 获取类别的名字model.names = names# Class frequencyif rank in [-1, 0]:# 将所有样本的标签拼接到一起shape为(total, 5)，统计后做可视化labels = np.concatenate(dataset.labels, 0)# 获得所有样本的类别c = torch.tensor(labels[:, 0])  # classes# cf = torch.bincount(c.long(), minlength=nc) + 1.# model._initialize_biases(cf.to(device))# 根据上面的统计对所有样本的类别，中心点xy位置，长宽wh做可视化plot_labels(labels, save_dir=log_dir)if tb_writer:# tb_writer.add_hparams(hyp, {})  # causes duplicate https://github.com/ultralytics/yolov5/pull/384tb_writer.add_histogram('classes', c, 0)# Check anchors"""计算默认锚点anchor与数据集标签框的长宽比值标签的长h宽w与anchor的长h_a宽w_a的比值, 即h/h_a, w/w_a都要在(1/hyp['anchor_t'], hyp['anchor_t'])是可以接受的如果标签框满足上面条件的数量小于总数的99%，则根据k-mean算法聚类新的锚点anchor"""if not opt.noautoanchor:check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)# Start trainingt0 = time.time()# 获取热身训练的迭代次数nw = max(3 * nb, 1e3)  # number of warmup iterations, max(3 epochs, 1k iterations)# nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training# 初始化mAP和resultsmaps = np.zeros(nc)  # mAP per classresults = (0, 0, 0, 0, 0, 0, 0)  # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'"""设置学习率衰减所进行到的轮次，目的是打断训练后，--resume接着训练也能正常的衔接之前的训练进行学习率衰减"""scheduler.last_epoch = start_epoch - 1  # do not move# 通过torch1.6自带的api设置混合精度训练scaler = amp.GradScaler(enabled=cuda)"""打印训练和测试输入图片分辨率加载图片时调用的cpu进程数从哪个epoch开始训练"""logger.info('Image sizes %g train, %g test' % (imgsz, imgsz_test))logger.info('Using %g dataloader workers' % dataloader.num_workers)logger.info('Starting training for %g epochs...' % epochs)# torch.autograd.set_detect_anomaly(True)# 训练for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------model.train()# Update image weights (optional)if dataset.image_weights:# Generate indices"""如果设置进行图片采样策略，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生成图片索引indices从而进行采样"""if rank in [-1, 0]:w = model.class_weights.cpu().numpy() * (1 - maps) ** 2  # class weightsimage_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)dataset.indices = random.choices(range(dataset.n), weights=image_weights,k=dataset.n)  # rand weighted idx# Broadcast if DDP# 如果是DDP模式,则广播采样策略if rank != -1:indices = torch.zeros([dataset.n], dtype=torch.int)if rank == 0:indices[:] = torch.tensor(dataset.indices, dtype=torch.int)# 广播索引到其他groupdist.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# 初始化训练时打印的平均损失信息mloss = torch.zeros(4, device=device)  # mean lossesif rank != -1:# DDP模式下打乱数据, ddp.sampler的随机采样数据是基于epoch+seed作为随机种子，# 每次epoch不同，随机种子就不同dataloader.sampler.set_epoch(epoch)pbar = enumerate(dataloader)logger.info(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size'))if rank in [-1, 0]:# tqdm 创建进度条，方便训练时 信息的展示pbar = tqdm(pbar, total=nb)  # progress baroptimizer.zero_grad()for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------# 计算迭代的次数iterationni = i + nb * epoch  # number integrated batches (since train start)imgs = imgs.to(device, non_blocking=True).float() / 255.0  # uint8 to float32, 0-255 to 0.0-1.0# Warmup"""热身训练(前nw次迭代)在前nw次迭代中，根据以下方式选取accumulate和学习率"""if ni <= nw:xi = [0, nw]  # x interp# model.gr = np.interp(ni, xi, [0.0, 1.0])  # giou loss ratio (obj_loss = 1.0 or giou)accumulate = max(1, np.interp(ni, xi, [1, nbs / total_batch_size]).round())for j, x in enumerate(optimizer.param_groups):# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0"""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"""x['lr'] = np.interp(ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])# 动量momentum也从0.9慢慢变到hyp['momentum'](default=0.937)if 'momentum' in x:x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']])# Multi-scale# 设置多尺度训练，从imgsz * 0.5, imgsz * 1.5 + gs随机选取尺寸if opt.multi_scale:sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # sizesf = sz / max(imgs.shape[2:])  # scale factorif 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# 混合精度with amp.autocast(enabled=cuda):# 前向传播pred = model(imgs)  # forward# Loss# 计算损失，包括分类损失，objectness损失，框的回归损失# loss为总损失值，loss_items为一个元组，包含分类损失，objectness损失，框的回归损失和总损失loss, loss_items = compute_loss(pred, targets.to(device), model)  # loss scaled by batch_sizeif rank != -1:# 平均不同gpu之间的梯度loss *= opt.world_size  # gradient averaged between devices in DDP mode# Backward# 反向传播scaler.scale(loss).backward()# Optimize# 模型反向传播accumulate次之后再根据累积的梯度更新一次参数if ni % accumulate == 0:scaler.step(optimizer)  # optimizer.stepscaler.update()optimizer.zero_grad()if ema:ema.update(model)# Printif rank in [-1, 0]:# 打印显存，进行的轮次，损失，target的数量和图片的size等信息mloss = (mloss * i + loss_items) / (i + 1)  # update mean lossesmem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0)  # (GB)s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1])# 进度条显示以上信息pbar.set_description(s)# Plot# 将前三次迭代batch的标签框在图片上画出来并保存if ni < 3:f = str(log_dir / ('train_batch%g.jpg' % ni))  # filenameresult = plot_images(images=imgs, targets=targets, paths=paths, fname=f)if tb_writer and result is not None:tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)# tb_writer.add_graph(model, imgs)  # add model to tensorboard# end batch ------------------------------------------------------------------------------------------------# Scheduler# 进行学习率衰减lr = [x['lr'] for x in optimizer.param_groups]  # for tensorboardscheduler.step()# DDP process 0 or single-GPUif rank in [-1, 0]:# mAPif ema:# 更新EMA的属性# 添加include的属性ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride'])# 判断该epoch是否为最后一轮final_epoch = epoch + 1 == epochs# 对测试集进行测试，计算mAP等指标# 测试时使用的是EMA模型if not opt.notest or final_epoch:  # Calculate mAPresults, maps, times = test.test(opt.data,batch_size=total_batch_size,imgsz=imgsz_test,model=ema.ema,single_cls=opt.single_cls,dataloader=testloader,save_dir=log_dir)# Write# 将指标写入result.txtwith open(results_file, 'a') as f:f.write(s + '%10.4g' * 7 % results + '\n')  # P, R, mAP, F1, test_losses=(GIoU, obj, cls)# 如果设置opt.bucket, 上传results.txt到谷歌云盘if len(opt.name) and opt.bucket:os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name))# Tensorboard# 添加指标，损失等信息到tensorboard显示if tb_writer:tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss',  # train loss'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95','val/giou_loss', 'val/obj_loss', 'val/cls_loss',  # val loss'x/lr0', 'x/lr1', 'x/lr2']  # paramsfor x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags):tb_writer.add_scalar(tag, x, epoch)# Update best mAP# 更新best_fitnessfi = fitness(np.array(results).reshape(1, -1))  # fitness_i = weighted combination of [P, R, mAP, F1]if fi > best_fitness:best_fitness = fi# Save model"""保存模型，还保存了epoch，results，optimizer等信息，optimizer将不会在最后一轮完成后保存model保存的是EMA的模型"""save = (not opt.nosave) or (final_epoch and not opt.evolve)if save:with open(results_file, 'r') as f:  # create checkpointckpt = {'epoch': epoch,'best_fitness': best_fitness,'training_results': f.read(),'model': ema.ema,'optimizer': None if final_epoch else optimizer.state_dict()}# Save last, best and deletetorch.save(ckpt, last)if best_fitness == fi:torch.save(ckpt, best)del ckpt# end epoch ----------------------------------------------------------------------------------------------------# end trainingif rank in [-1, 0]:# Strip optimizers"""模型训练完后，strip_optimizer函数将optimizer从ckpt中去除；并且对模型进行model.half(), 将Float32的模型->Float16，可以减少模型大小，提高inference速度"""n = ('_' if len(opt.name) and not opt.name.isnumeric() else '') + opt.namefresults, flast, fbest = 'results%s.txt' % n, wdir / f'last{n}.pt', wdir / f'best{n}.pt'for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', 'results.txt'], [flast, fbest, fresults]):if os.path.exists(f1):os.rename(f1, f2)  # renameif str(f2).endswith('.pt'):  # is *.ptstrip_optimizer(f2)  # strip optimizer# 上传结果到谷歌云盘os.system('gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket else None  # upload# Finish# 可视化results.txt文件if not opt.evolve:plot_results(save_dir=log_dir)  # save as results.pnglogger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))# 释放显存dist.destroy_process_group() if rank not in [-1, 0] else Nonetorch.cuda.empty_cache()return results

2020.7.20版本

训练参数以及main函数解析

训练的时候可以设置进行超参数进化算法(默认不使用)。
值得一提的是，由于现在yolov5还在开发当中，训练文件的–resume还不是100%的完善，不建议打断训练再resume。具体可以参照issue292。

if __name__ == '__main__':# 因为yolov5还在开发当中，check_git_status()检查你的代码版本是否为最新的(不适用于windows系统)check_git_status()"""opt参数解析：cfg:模型配置文件，网络结构data:数据集配置文件，数据集路径，类名等hyp:超参数文件epochs:训练总轮次batch-size:批次大小img-size:输入图片分辨率大小rect:是否采用矩形训练，默认Falseresume:接着打断训练上次的结果接着训练nosave:不保存模型，默认Falsenotest:不进行test，默认Falsenoautoanchor:不自动调整anchor，默认Falseevolve:是否进行超参数进化，默认Falsebucket:谷歌云盘bucket，一般不会用到cache-images:是否提前缓存图片到内存，以加快训练速度，默认Falseweights:加载的权重文件name:数据集名字，如果设置：results.txt to results_name.txt，默认无device:训练的设备，cpu；0(表示一个gpu设备cuda:0)；0,1,2,3(多个gpu设备)multi-scale:是否进行多尺度训练，默认Falsesingle-cls:数据集是否只有一个类别，默认False"""parser = argparse.ArgumentParser()parser.add_argument('--cfg', type=str, default='models/yolov5x_landslide.yaml', help='model.yaml path')parser.add_argument('--data', type=str, default='data/landslide.yaml', help='data.yaml path')parser.add_argument('--hyp', type=str, default='', help='hyp.yaml path (optional)')parser.add_argument('--epochs', type=int, default=300)parser.add_argument('--batch-size', type=int, default=8)parser.add_argument('--img-size', nargs='+', type=int, default=[416, 416], help='train,test sizes')parser.add_argument('--rect', action='store_true', help='rectangular training')parser.add_argument('--resume', nargs='?', const='get_last', default='runs/exp0/weights/last.pt',help='resume from given path/to/last.pt, or most recent run if blank.')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('--weights', type=str, default='', help='initial weights path')parser.add_argument('--name', default='', help='renames results.txt to results_name.txt if supplied')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 as single-class dataset')opt = parser.parse_args()"""resume时获取last.pt的路径get_latest_run()函数获取runs文件夹中最近的last.pt注意：进行resume时，不要设置opt.weights(除非设置opt.weights='last.pt')，否则会重新开始训练"""last = get_latest_run() if opt.resume == 'get_last' else opt.resume  # resume from most recent runif last and not opt.weights:print(f'Resuming training from {last}')opt.weights = last if opt.resume and not opt.weights else opt.weights# check_file检查文件是否存在opt.cfg = check_file(opt.cfg)  # check fileopt.data = check_file(opt.data)  # check fileopt.hyp = check_file(opt.hyp) if opt.hyp else ''  # check fileprint(opt)# 扩展image_size为[image_size, image_size]一个是训练size，一个是测试sizeopt.img_size.extend([opt.img_size[-1]] * (2 - len(opt.img_size)))  # extend to 2 sizes (train, test)# 选择设备device = torch_utils.select_device(opt.device, apex=mixed_precision, batch_size=opt.batch_size)if device.type == 'cpu':mixed_precision = False# Train# 如果不进行超参数进化，则直接调用train()函数，开始训练if not opt.evolve:print('Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/')# 创建tensorboardtb_writer = SummaryWriter(log_dir=increment_dir('runs' + os.sep + 'exp', opt.name))# 如果设置了超参数文件路径，则加载新的超参数文件if opt.hyp:  # update hypswith open(opt.hyp) as f:hyp.update(yaml.load(f, Loader=yaml.FullLoader))train(hyp)# Evolve hyperparameters (optional)# 根据训练结果进行超参数的进化else:tb_writer = None# 设置不测试不保存模型opt.notest, opt.nosave = True, True  # only test/save final epochif opt.bucket:os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket)  # download evolve.txt if exists# 默认进化十次"""这里的进化算法是：根据之前训练时的hyp来确定一个base hyp再进行突变；如何根据？通过之前每次进化得到的results来确定之前每个hyp的权重有了每个hyp和每个hyp的权重之后有两种进化方式；1.根据每个hyp的权重随机选择一个之前的hyp作为base hyp，random.choices(range(n), weights=w)2.根据每个hyp的权重对之前所有的hyp进行融合获得一个base hyp，(x * w.reshape(n, 1)).sum(0) / w.sum()evolve.txt会记录每次进化之后的results+hyp每次进化时，hyp会根据之前的results进行从大到小的排序；再根据fitness函数计算之前每次进化得到的hyp的权重再确定哪一种进化方式，从而进行进化"""for _ in range(10):  # generations to evolveif os.path.exists('evolve.txt'):  # if evolve.txt exists: select best hyps and mutate# Select parent(s)# 选择进化方式parent = 'single'  # parent selection method: 'single' or 'weighted'# 加载evolve.txtx = np.loadtxt('evolve.txt', ndmin=2)# 选取至多前5次进化的结果n = min(5, len(x))  # number of previous results to considerx = x[np.argsort(-fitness(x))][:n]  # top n mutations# 根据results计算hyp的权重w = fitness(x) - fitness(x).min()  # weights# 根据不同进化方式获得base hypif parent == 'single' or len(x) == 1:# x = x[random.randint(0, n - 1)]  # random selectionx = x[random.choices(range(n), weights=w)[0]]  # weighted selectionelif parent == 'weighted':x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination# Mutate# 超参数进化mp, s = 0.9, 0.2  # mutation probability, sigmanpr = np.randomnpr.seed(int(time.time()))g = np.array([1, 1, 1, 1, 1, 1, 1, 0, .1, 1, 0, 1, 1, 1, 1, 1, 1, 1])  # gainsng = len(g)v = np.ones(ng)# 设置突变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)# 将突变添加到base hyp上# [i+7]是因为x中前七个数字为results的指标(P, R, mAP, F1, test_losses=(GIoU, obj, cls))，之后才是超参数hypfor i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)hyp[k] = x[i + 7] * v[i]  # mutate# Clip to limits# 修剪hyp在规定范围里keys = ['lr0', 'iou_t', 'momentum', 'weight_decay', 'hsv_s', 'hsv_v', 'translate', 'scale', 'fl_gamma']limits = [(1e-5, 1e-2), (0.00, 0.70), (0.60, 0.98), (0, 0.001), (0, .9), (0, .9), (0, .9), (0, .9), (0, 3)]for k, v in zip(keys, limits):hyp[k] = np.clip(hyp[k], v[0], v[1])# Train mutation# 训练results = train(hyp.copy())# Write mutation results"""写入results和对应的hyp到evolve.txtevolve.txt文件每一行为一次进化的结果一行中前七个数字为(P, R, mAP, F1, test_losses=(GIoU, obj, cls))，之后为hyp"""print_mutation(hyp, results, opt.bucket)# Plot results# plot_evolution_results(hyp)

train函数解析

import argparseimport torch.distributed as dist
import torch.nn.functional as F
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
import torch.utils.data
from torch.utils.tensorboard import SummaryWriterimport test  # import test.py to get mAP after each epoch
from models.yolo import Model
from utils import google_utils
from utils.datasets import *
from utils.utils import *#  设置混精度训练，需要安装英伟达的apex，默认为True,笔者没用到就设置为False
mixed_precision = False
try:  # Mixed precision training https://github.com/NVIDIA/apexfrom apex import amp
except:print('Apex recommended for faster mixed precision training: https://github.com/NVIDIA/apex')mixed_precision = False  # not installed# 超参数
hyp = {'optimizer': 'SGD',  # 优化器['adam', 'SGD', None] if none, default is SGD'lr0': 0.01,  # 学习率initial learning rate (SGD=1E-2, Adam=1E-3)'momentum': 0.937,  # 学习率动量SGD momentum/Adam beta1'weight_decay': 5e-4,  # 权重衰减系数optimizer weight decay'giou': 0.05,  # giou损失的系数giou loss gain'cls': 0.58,  # 分类损失的系数cls loss gain'cls_pw': 1.0,  # 分类BCELoss中正样本的权重cls BCELoss positive_weight'obj': 1.0,  # 有无物体损失的系数obj loss gain (*=img_size/320 if img_size != 320)'obj_pw': 1.0,  # 有无物体BCELoss中正样本的权重obj BCELoss positive_weight'iou_t': 0.20,  # 标签与anchors的iou阈值iou training threshold'anchor_t': 4.0,  # 标签的长h宽w/anchor的长h_a宽w_a阈值, 即h/h_a, w/w_a都要在(1/4, 4)之间anchor-multiple threshold'fl_gamma': 0.0,  # focal loss gamma, 设为0则表示不使用focal loss(efficientDet default is gamma=1.5)# 下面是一些数据增强的系数, 包括颜色空间和图片空间'hsv_h': 0.014,  # image HSV-Hue augmentation (fraction)'hsv_s': 0.68,  # image HSV-Saturation augmentation (fraction)'hsv_v': 0.36,  # image HSV-Value augmentation (fraction)'degrees': 0.0,  # image rotation (+/- deg)'translate': 0.0,  # image translation (+/- fraction)'scale': 0.5,  # image scale (+/- gain)'shear': 0.0}  # image shear (+/- deg)def train(hyp):print(f'Hyperparameters {hyp}')# 获取记录训练日志的路径"""训练日志包括：权重、tensorboard文件、超参数hyp、设置的训练参数opt(也就是epochs,batch_size等),result.txtresult.txt包括: 占GPU内存、训练集的GIOU loss, objectness loss, classification loss, 总loss, targets的数量, 输入图片分辨率, 准确率TP/(TP+FP),召回率TP/P ; 测试集的mAP50, mAP@0.5:0.95, GIOU loss, objectness loss, classification loss.还会保存batch<3的ground truth"""log_dir = tb_writer.log_dir  # run directory# 设置保存权重的路径wdir = str(Path(log_dir) / 'weights') + os.sep  # weights directoryos.makedirs(wdir, exist_ok=True)last = wdir + 'last.pt'best = wdir + 'best.pt'# 设置保存results的路径results_file = log_dir + os.sep + 'results.txt'# Save run settings# 保存hyp和optwith open(Path(log_dir) / 'hyp.yaml', 'w') as f:yaml.dump(hyp, f, sort_keys=False)with open(Path(log_dir) / 'opt.yaml', 'w') as f:yaml.dump(vars(opt), f, sort_keys=False)# 设置轮次、批次、权重epochs = opt.epochs  # 300batch_size = opt.batch_size  # 64weights = opt.weights  # initial training weights# Configure# 设置随机种子init_seeds(1)# 加载数据配置信息with open(opt.data) as f:data_dict = yaml.load(f, Loader=yaml.FullLoader)  # model dict# 获取训练集、测试集图片路径train_path = data_dict['train']test_path = data_dict['val']# 获取类别数量和类别名字# 如果设置了opt.single_cls则为一类nc, names = (1, ['item']) if opt.single_cls else (int(data_dict['nc']), data_dict['names'])  # number classes, namesassert len(names) == nc, '%g names found for nc=%g dataset in %s' % (len(names), nc, opt.data)  # check# Remove previous results# 移除之前的图片结果for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):os.remove(f)# Create model# 创建模型model = Model(opt.cfg, nc=nc).to(device)# Image sizes# 获取模型总步长和模型输入图片分辨率gs = int(max(model.stride))  # grid size (max stride)# 检查输入图片分辨率确保能够整除总步长gsimgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size]  # verify imgsz are gs-multiples# Optimizer"""nbs为模拟的batch_size; 就比如默认的话上面设置的opt.batch_size为16,这个nbs就为64，也就是模型梯度累积了64/16=4(accumulate)次之后再更新一次模型，变相的扩大了batch_size"""nbs = 32  # nominal batch sizeaccumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing# 根据accumulate设置权重衰减系数hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decaypg0, pg1, pg2 = [], [], []  # optimizer parameter groups# 将模型分成三组(weight、bn, bias, 其他所有参数)优化for k, v in model.named_parameters():if v.requires_grad:if '.bias' in k:pg2.append(v)  # biaseselif '.weight' in k and '.bn' not in k:pg1.append(v)  # apply weight decayelse:pg0.append(v)  # all else# 选用优化器，并设置pg0组的优化方式if hyp['optimizer'] == 'adam':  # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLRoptimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999))  # adjust beta1 to momentumelse:optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)# 设置weight、bn的优化方式optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']})  # add pg1 with weight_decay# 设置biases的优化方式optimizer.add_param_group({'params': pg2})  # add pg2 (biases)print('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0)))del pg0, pg1, pg2# 设置学习率衰减，这里为余弦退火方式进行衰减# 就是根据以下公式lf与epoch进行衰减# Scheduler https://arxiv.org/pdf/1812.01187.pdflf = lambda x: (((1 + math.cos(x * math.pi / epochs)) / 2) ** 1.0) * 0.9 + 0.1  # cosinescheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)# plot_lr_scheduler(optimizer, scheduler, epochs, save_dir=log_dir)# Load Model# 加载模型，从google云盘中自动下载模型# 但通常会下载失败，建议提前下载下来放进weights目录google_utils.attempt_download(weights)# 初始化开始训练的epoch和最好的结果# best_fitness是以[0.0, 0.0, 0.1, 0.9]为系数并乘以[精确度, 召回率, mAP@0.5, mAP@0.5:0.95]再求和所得# 根据best_fitness来保存best.ptstart_epoch, best_fitness = 0, 0.0if weights.endswith('.pt'):  # pytorch format# 加载检查点ckpt = torch.load(weights, map_location=device)  # load checkpoint# load model# 加载模型try:ckpt['model'] = {k: v for k, v in ckpt['model'].float().state_dict().items()if model.state_dict()[k].shape == v.shape}  # to FP32, filtermodel.load_state_dict(ckpt['model'], strict=False)except KeyError as e:s = "%s is not compatible with %s. This may be due to model differences or %s may be out of date. " \"Please delete or update %s and try again, or use --weights '' to train from scratch." \% (opt.weights, opt.cfg, opt.weights, opt.weights)raise KeyError(s) from e# load optimizer# 加载优化器与best_fitnessif ckpt['optimizer'] is not None:optimizer.load_state_dict(ckpt['optimizer'])best_fitness = ckpt['best_fitness']# load results# 加载训练结果result.txtif ckpt.get('training_results') is not None:with open(results_file, 'w') as file:file.write(ckpt['training_results'])  # write results.txt# epochs# 加载训练的轮次start_epoch = ckpt['epoch'] + 1"""如果新设置epochs小于加载的epoch，则视新设置的epochs为需要再训练的轮次数而不再是总的轮次数"""if epochs < start_epoch:print('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' %(opt.weights, ckpt['epoch'], epochs))epochs += ckpt['epoch']  # finetune additional epochsdel ckpt# Mixed precision training https://github.com/NVIDIA/apex# 如果设置混精度训练，初始化混精度训练if mixed_precision:model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)# Distributed training# 如果不在cpu上计算且gpu数量大于1且pytorch允许分布式，则设置分布式训练if device.type != 'cpu' and torch.cuda.device_count() > 1 and torch.distributed.is_available():dist.init_process_group(backend='nccl',  # distributed backendinit_method='tcp://127.0.0.1:9999',  # init methodworld_size=1,  # number of nodesrank=0)  # node rankmodel = torch.nn.parallel.DistributedDataParallel(model)# pip install torch==1.4.0+cu100 torchvision==0.5.0+cu100 -f https://download.pytorch.org/whl/torch_stable.html# Trainloader# 创建训练集dataloaderdataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt,hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect)"""获取标签中最大的类别值，并于类别数作比较如果大于类别数则表示有问题"""mlc = np.concatenate(dataset.labels, 0)[:, 0].max()  # max label classnb = len(dataloader)  # number of batchesassert mlc < nc, 'Label class %g exceeds nc=%g in %s. Correct your labels or your model.' % (mlc, nc, opt.cfg)# Testloader# 创建测试集dataloadertestloader = create_dataloader(test_path, imgsz_test, batch_size, gs, opt,hyp=hyp, augment=False, cache=opt.cache_images, rect=True)[0]# Model parameters# 根据自己数据集的类别数设置分类损失的系数hyp['cls'] *= nc / 80.  # scale coco-tuned hyp['cls'] to current dataset# 设置类别数，超参数model.nc = nc  # attach number of classes to modelmodel.hyp = hyp  # attach hyperparameters to model"""设置giou的值在objectness loss中做标签的系数, 使用代码如下tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)这里model.gr=1，也就是说完全使用标签框与预测框的giou值来作为该预测框的objectness标签"""model.gr = 1.0  # giou loss ratio (obj_loss = 1.0 or giou)# 根据labels初始化图片采样权重model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device)  # attach class weights# 获取类别的名字model.names = names# Class frequency# 将所有样本的标签拼接到一起shape为(total, 5)，统计后做可视化labels = np.concatenate(dataset.labels, 0)# 获得所有样本的类别c = torch.tensor(labels[:, 0])  # classes# cf = torch.bincount(c.long(), minlength=nc) + 1.# model._initialize_biases(cf.to(device))# 根据上面的统计对所有样本的类别，中心点xy位置，长宽wh做可视化plot_labels(labels, save_dir=log_dir)# 添加类别的直方图到tensorboard中if tb_writer:tb_writer.add_histogram('classes', c, 0)# Check anchors"""计算默认锚点anchor与数据集标签框的长宽比值标签的长h宽w与anchor的长h_a宽w_a的比值, 即h/h_a, w/w_a都要在(1/hyp['anchor_t'], hyp['anchor_t'])是可以接受的如果标签框满足上面条件的数量小于总数的99%，则根据k-mean算法聚类新的锚点anchor"""if not opt.noautoanchor:check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)# Exponential moving average# 为模型创建EMA指数滑动平均ema = torch_utils.ModelEMA(model, updates=start_epoch * nb / accumulate)print(ema.updates)# Start trainingt0 = time.time()# 获取热身训练的迭代次数nw = max(3 * nb, 1e3)  # number of warmup iterations, max(3 epochs, 1k iterations)# 初始化mAP和resultsmaps = np.zeros(nc)  # mAP per classresults = (0, 0, 0, 0, 0, 0, 0)  # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'"""设置学习率衰减所进行到的轮次，目的是打断训练后，--resume接着训练也能正常的衔接之前的训练进行学习率衰减"""scheduler.last_epoch = start_epoch - 1  # do not move"""打印训练和测试输入图片分辨率加载图片时调用的cpu进程数从哪个epoch开始训练"""print('Image sizes %g train, %g test' % (imgsz, imgsz_test))print('Using %g dataloader workers' % dataloader.num_workers)print('Starting training for %g epochs...' % epochs)# torch.autograd.set_detect_anomaly(True)# 训练for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------# if epoch == 250:#     exit()model.train()# Update image weights (optional)"""如果设置进行图片采样策略，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生成图片索引indices从而进行采样"""if dataset.image_weights:w = model.class_weights.cpu().numpy() * (1 - maps) ** 2  # class weightsimage_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n)  # rand weighted idx# 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# 初始化训练时打印的平均损失信息mloss = torch.zeros(4, device=device)  # mean lossesprint(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size'))# tqdm 创建进度条，方便训练时 信息的展示pbar = tqdm(enumerate(dataloader), total=nb)  # progress barfor i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------# 计算迭代的次数iterationni = i + nb * epoch  # number integrated batches (since train start)imgs = imgs.to(device).float() / 255.0  # uint8 to float32, 0 - 255 to 0.0 - 1.0# Warmup"""热身训练(前nw次迭代)在前nw次迭代中，根据以下方式选取accumulate和学习率"""if ni <= nw:xi = [0, nw]  # x interp# model.gr = np.interp(ni, xi, [0.0, 1.0])  # giou loss ratio (obj_loss = 1.0 or giou)accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())for j, x in enumerate(optimizer.param_groups):# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0"""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"""x['lr'] = np.interp(ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])# 动量momentum也从0.9慢慢变到hyp['momentum'](default=0.937)if 'momentum' in x:x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']])# Multi-scale# 设置多尺度训练，从imgsz * 0.5, imgsz * 1.5 + gs随机选取尺寸if opt.multi_scale:sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # sizesf = sz / max(imgs.shape[2:])  # scale factorif 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)# Forwardpred = model(imgs)# Loss# 计算损失，包括分类损失，objectness损失，框的回归损失# loss为总损失值，loss_items为一个元组，包含分类损失，objectness损失，框的回归损失和总损失loss, loss_items = compute_loss(pred, targets.to(device), model)# 检查loss是否无穷大(可能时梯度爆炸，或者计算损失梯度时存在log(score)->log(0)->无穷大)if not torch.isfinite(loss):print('WARNING: non-finite loss, ending training ', loss_items)return results# Backward# 如果设置混精度训练，混合精度反向传播if mixed_precision:with amp.scale_loss(loss, optimizer) as scaled_loss:scaled_loss.backward()else:loss.backward()# Optimize# 模型反向传播accumulate次之后再根据累积的梯度更新一次参数if ni % accumulate == 0:optimizer.step()optimizer.zero_grad()ema.update(model)# Print# 打印显存，进行的轮次，损失，target的数量和图片的size等信息mloss = (mloss * i + loss_items) / (i + 1)  # update mean lossesmem = '%.3gG' % (torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available() else 0)  # (GB)s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1])# 进度条显示以上信息pbar.set_description(s)# Plot# 将前三次迭代batch的标签框在图片上画出来并保存if ni < 3:f = str(Path(log_dir) / ('train_batch%g.jpg' % ni))  # filenameresult = plot_images(images=imgs, targets=targets, paths=paths, fname=f)if tb_writer and result is not None:tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)# tb_writer.add_graph(model, imgs)  # add model to tensorboard# end batch ------------------------------------------------------------------------------------------------# Scheduler# 进行学习率衰减scheduler.step()# mAP# 更新EMA的属性ema.update_attr(model)# 判断该epoch是否为最后一轮final_epoch = epoch + 1 == epochs# 对测试集进行测试，计算mAP等指标# 测试时使用的是EMA模型if not opt.notest or final_epoch:  # Calculate mAPresults, maps, times = test.test(opt.data,batch_size=batch_size,imgsz=imgsz_test,save_json=final_epoch and opt.data.endswith(os.sep + 'coco.yaml'),model=ema.ema,single_cls=opt.single_cls,dataloader=testloader,save_dir=log_dir)# Write# 将指标写入result.txtwith open(results_file, 'a') as f:f.write(s + '%10.4g' * 7 % results + '\n')  # P, R, mAP, F1, test_losses=(GIoU, obj, cls)# 如果设置opt.bucket, 上传results.txt到谷歌云盘if len(opt.name) and opt.bucket:os.system('gsutil cp results.txt gs://%s/results/results%s.txt' % (opt.bucket, opt.name))# Tensorboard# 添加指标，损失等信息到tensorboard显示if tb_writer:tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss','metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/F1','val/giou_loss', 'val/obj_loss', 'val/cls_loss']for x, tag in zip(list(mloss[:-1]) + list(results), tags):tb_writer.add_scalar(tag, x, epoch)# Update best mAP# 更新best_fitnessfi = fitness(np.array(results).reshape(1, -1))  # fitness_i = weighted combination of [P, R, mAP, F1]if fi > best_fitness:best_fitness = fi# Save model"""保存模型，还保存了epoch，results，optimizer等信息，optimizer将不会在最后一轮完成后保存model保存的是EMA的模型"""save = (not opt.nosave) or (final_epoch and not opt.evolve)if save:with open(results_file, 'r') as f:  # create checkpointckpt = {'epoch': epoch,'best_fitness': best_fitness,'training_results': f.read(),'model': ema.ema,'optimizer': None if final_epoch else optimizer.state_dict()}# Save last, best and deletetorch.save(ckpt, last)if (best_fitness == fi) and not final_epoch:torch.save(ckpt, best)del ckpt# end epoch ----------------------------------------------------------------------------------------------------# end training# Strip optimizers"""模型训练完后，strip_optimizer函数将optimizer从ckpt中去除；并且对模型进行model.half(), 将Float32的模型->Float16，可以减少模型大小，提高inference速度"""n = ('_' if len(opt.name) and not opt.name.isnumeric() else '') + opt.namefresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % nfor f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'], [flast, fbest, fresults]):if os.path.exists(f1):os.rename(f1, f2)  # renameispt = f2.endswith('.pt')  # is *.ptstrip_optimizer(f2) if ispt else None  # strip optimizer# 上传结果到谷歌云盘os.system('gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket and ispt else None  # upload# Finish# 可视化results.txt文件if not opt.evolve:plot_results(save_dir=log_dir)  # save as results.pngprint('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))# 释放显存dist.destroy_process_group() if device.type != 'cpu' and torch.cuda.device_count() > 1 else Nonetorch.cuda.empty_cache()return results

以上我根据ultralytics\yolov5的train.py代码对其整体流程做一个梳理，讲解每个部分的代码的作用，但是对于一些细节函数还没做详细解析，就比如说计算损失的compute_loss()函数等，这些函数在utils.py文件里，之后更新解析utils.py。

YOLOV5训练代码train.py注释与解析相关推荐

YOLOV5测试代码test.py注释与解析
YOLOv5代码注释版更新啦,注释的是最近的2021.07.14的版本,且注释更全 github: https://github.com/Laughing-q/yolov5_annotations Y ...
YOLOV5检测代码detect.py注释与解析
YOLOv5代码注释版更新啦,注释的是最近的2021.07.14的版本,且注释更全 github: https://github.com/Laughing-q/yolov5_annotations Y ...
DeamNet||训练代码学习train.py注释与解析
目录 1. 导入各种库,设置运行环境 2. 训练设置,各种参数 3.测试 1. 导入各种库,设置运行环境 from __future__ import print_function import os ...
YOLOv5源码逐行超详细注释与解读（3）——训练部分train.py
前言本篇文章主要是对YOLOv5项目的训练部分train.py.通常这个文件主要是用来读取用户自己的数据集,加载模型并训练. 文章代码逐行手打注释,每个模块都有对应讲解,一文帮你梳理整个代码逻辑! ...
yolov5---detect.py注释解析
YOLOV5检测代码detect.py注释与解析检测参数以及main函数解析 if __name__ == '__main__':"""weights:训练的权重sou ...
图像去雨：超详细手把手写 pytorch 实现代码（带注释）
引导数据集准备训练数据集代码测试数据集代码网络模型代码训练代码测试代码参考文献其他数据集准备使用来自第一个参考文献的公开数据集Rain12600和Rain1400,下载链接.其中训 ...
1 PyTorch版YOLOv3 代码中文注释之训练 train.py test.py detect.py
文章目录 PyTorch版YOLOv3 代码中文注释 1. 相关链接: 2. 代码结构: 3. train.py 3.1. train.py 中包含的主要功能 4. test.py 4.1. test ...
图像分割套件PaddleSeg全面解析（一）train.py代码解读
首先祝贺百度团队百度斩获NeurIPS2020挑战赛冠军,https://www.jiqizhixin.com/articles/2020-12-09-2. 在此次比赛中使用的是基于飞桨深度学习框架开 ...
基于YOLOv5的手势识别系统(含手势识别数据集+训练代码)
基于YOLOv5的手势识别系统(含手势识别数据集+训练代码) 目录基于YOLOv5的手势识别系统(含手势识别数据集+训练代码) 1. 前言 2. 手势识别的方法 (1)基于多目标检测的手势识别方法 ...

YOLOV5训练代码train.py注释与解析

YOLOV5训练代码train.py注释与解析

2020.8.31版本

超参数文件hyp解析

训练参数以及main函数解析

train函数解析

2020.7.20版本

训练参数以及main函数解析

train函数解析

YOLOV5训练代码train.py注释与解析相关推荐

最新文章

热门文章