yolov3代码探究-train.py

github-yolov3
train.py

# https://github.com/ultralytics/yolov3/blob/master/train.py
import argparse
import time
import torch.distributed as dist
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
from torch.utils.data import DataLoader
import test  # import test.py to get mAP after each epoch
from models import *
from utils.datasets import *
from utils.utils import *
#      0.109      0.297       0.15      0.126       7.04      1.666      4.062     0.1845       42.6       3.34      12.61      8.338     0.2705      0.001         -4        0.9     0.0005   320 giou + best_anchor False
#      0.223      0.218      0.138      0.189       9.28      1.153      4.376    0.08263      24.28       3.05      20.93      2.842     0.2759   0.001357     -5.036     0.9158  0.0005722   mAP/F1 - 50/50 weighting
#      0.231      0.215      0.135      0.191       9.51      1.432      3.007    0.06082      24.87      3.477      24.13      2.802     0.3436   0.001127     -5.036     0.9232  0.0005874
#      0.246      0.194      0.128      0.192       8.12      1.101      3.954     0.0817      22.83      3.967      19.83      1.779     0.3352   0.000895     -5.036     0.9238  0.0007973
# 0.242    0.296    0.196    0.231    5.67    0.8541    4.286    0.1539    21.61    1.957    22.9    2.894    0.3689    0.001844    -4    0.913    0.000467
# 0.298    0.244    0.167    0.247    4.99    0.8896    4.067    0.1694    21.41    2.033    25.61    1.783    0.4115    0.00128        -4    0.950    0.000377
# 0.268    0.268    0.178    0.240    4.36    1.104    5.596    0.2087    14.47    2.599    16.27    2.406    0.4114    0.001585    -4    0.950    0.000524
# 0.161    0.327    0.190    0.193    7.82    1.153    4.062    0.1845    24.28    3.05    20.93    2.842    0.2759    0.001357    -4    0.916    0.000572  # 320 --epochs 2
# Training hyperparameters
hyp = {'giou': 0.8541,  # giou loss gain
       'xy': 4.062,  # xy loss gain
       'wh': 0.1845,  # wh loss gain
       'cls': 21.61,  # cls loss gain
       'cls_pw': 1.957,  # cls BCELoss positive_weight
       'obj': 22.9,  # obj loss gain
       'obj_pw': 2.894,  # obj BCELoss positive_weight
       'iou_t': 0.3689,  # iou target-anchor training threshold
       'lr0': 0.001844,  # initial learning rate
       'lrf': -4.,  # final learning rate = lr0 * (10 ** lrf)
       'momentum': 0.913,  # SGD momentum
       'weight_decay': 0.000467}  # optimizer weight decay

含金量超高的一个环节来了！

def train(cfg,
          data_cfg,
          img_size=416,
          epochs=100,  # 500200 batches at bs 16, 117263 images = 273 epochs
          batch_size=16,
          accumulate=4):  # effective bs = batch_size * accumulate = 8 * 8 = 64
    # Initialize
    init_seeds()
    weights = 'weights' + os.sep
    last = weights + 'last.pt'
    best = weights + 'best.pt'
    device = torch_utils.select_device()
    multi_scale = opt.multi_scale
    if multi_scale:
        img_size_min = round(img_size / 32 / 1.5)
        img_size_max = round(img_size / 32 * 1.5)
        img_size = img_size_max * 32  # initiate with maximum multi_scale size
    # Configure run
    data_dict = parse_data_cfg(data_cfg)
    train_path = data_dict['train']
    nc = int(data_dict['classes'])  # number of classes
    # Initialize model
    model = Darknet(cfg).to(device)
    # Optimizer
    optimizer = optim.SGD(model.parameters(), lr=hyp['lr0'], momentum=hyp['momentum'], weight_decay=hyp['weight_decay'])
    cutoff = -1  # backbone reaches to cutoff layer
    start_epoch = 0
    best_fitness = 0.0
    """
    这边的代码学习意义很大。具体包括：
    1. 如何迁移学习
    2. 如何在上一次训练的基础上继续训练
    3. 如何保存训练的中间结果
    """
    # 迁移学习或者加载之前训练成果的代码 
    if opt.resume or opt.transfer:  # Load previously saved model
        # 迁移学习
        if opt.transfer:  # Transfer learning
                        # 获得yolo_layer的size
            nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters'])  # yolo layer size (i.e. 255)
            # torch.load加载预训练好的模型，weights是文件夹路径, map_location可以根据情况选择加载到GPU还是CPU.  device = torch_utils.select_device()
            chkpt = torch.load(weights + 'yolov3-spp.pt', map_location=device)
            # k指层的名字name，v指权重/参数param。如果有权重且v不是yolo层的前一层，那就加载权重到model, strict，默认是True，表示预训练模型的层和自己定义的网络结构层严格对应相等（比如层名和维度）
            model.load_state_dict({k: v for k, v in chkpt['model'].items() if v.numel() > 1 and v.shape[0] != 255},
                                  strict=False)
            # 对model中的每层的参数，注意！不是每个参数
            for p in model.parameters():
                #requires_grad表示是否进行梯度更新。只有yolo层的前一层需要
                p.requires_grad = True if p.shape[0] == nf else False
        else:  # resume from last.pt
            if opt.bucket:
                os.system('gsutil cp gs://%s/last.pt %s' % (opt.bucket, last))  # download from bucket
            chkpt = torch.load(last, map_location=device)  # load checkpoint
            model.load_state_dict(chkpt['model'])
        # 除了要加载checkpoint中的模型，优化函数的参数可能也随着训练改变了，因此还需要加载优化函数的参数
        if chkpt['optimizer'] is not None:
            optimizer.load_state_dict(chkpt['optimizer'])
            best_fitness = chkpt['best_fitness']
        # 作者还设置了一个training_results用来保存之前的训练结果
        if chkpt['training_results'] is not None:
            with open('results.txt', 'w') as file:
                file.write(chkpt['training_results'])  # write results.txt
        start_epoch = chkpt['epoch'] + 1
        del chkpt
    #既不迁移学习也不加载过去的训练模型
    else:  # Initialize model with backbone (optional)
        #load_darknet_weights是作者写的函数，用来加载权重
        if '-tiny.cfg' in cfg:
            cutoff = load_darknet_weights(model, weights + 'yolov3-tiny.conv.15')
        else:
            cutoff = load_darknet_weights(model, weights + 'darknet53.conv.74')
        # Remove old results
        # glob.glob相加只是单纯的罗列所有文件，把所有过去的结果都删除
        for f in glob.glob('*_batch*.jpg') + glob.glob('results.txt'):
            os.remove(f)
    # Scheduler https://github.com/ultralytics/yolov3/issues/238
    # lf = lambda x: 1 - x / epochs  # linear ramp to zero
    # lf = lambda x: 10 ** (hyp['lrf'] * x / epochs)  # exp ramp
    # lf = lambda x: 1 - 10 ** (hyp['lrf'] * (1 - x / epochs))  # inverse exp ramp
    # scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
    #一个调整学习率的策略，milestones表示多少次变一次学习率
    scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[round(opt.epochs * x) for x in (0.8, 0.9)], gamma=0.1)
    #设置初始时的学习率
    scheduler.last_epoch = start_epoch - 1
    # # Plot lr schedule
    # y = []
    # for _ in range(epochs):
    #     scheduler.step()
    #     y.append(optimizer.param_groups[0]['lr'])
    # plt.plot(y, label='LambdaLR')
    # plt.xlabel('epoch')
    # plt.ylabel('LR')
    # plt.tight_layout()
    # plt.savefig('LR.png', dpi=300)
    # Dataset
    # 加载数据集 
    ''' 
    LoadImagesAndLabels是作者写的读取数据的类
    getitem返回的包括：
      img, labels, img_path, (h,w) 
      其中labels:image_id, class, x, y, w, h, 表示位置的大小都是0-1 
    '''
    dataset = LoadImagesAndLabels(train_path,
                                  img_size,
                                  batch_size,
                                  augment=True,
                                  rect=opt.rect)  # rectangular training
    # Initialize distributed training
    # 分布式训练, 我实际使用的是需要改代码的，有张显卡不能用所以要限制其使用的显卡
    if torch.cuda.device_count() > 1:
        dist.init_process_group(backend='nccl',  # 'distributed backend'
                                init_method='tcp://127.0.0.1:9999',  # distributed training init method
                                world_size=1,  # number of nodes for distributed training
                                rank=0)  # distributed training node rank
        model = torch.nn.parallel.DistributedDataParallel(model)
        # sampler = torch.utils.data.distributed.DistributedSampler(dataset)
    # Dataloader
    dataloader = DataLoader(dataset,
                            batch_size=batch_size,
                            num_workers=opt.num_workers,
                            shuffle=not opt.rect,  # Shuffle=True unless rectangular training is used
                            pin_memory=True,
                            collate_fn=dataset.collate_fn)
    # Mixed precision training https://github.com/NVIDIA/apex
    mixed_precision = True
    if mixed_precision:
        try:
            from apex import amp
            model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
        except:  # not installed: install help: https://github.com/NVIDIA/apex/issues/259
            mixed_precision = False
    # Start training
    model.hyp = hyp  # attach hyperparameters to model
    # model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device)  # attach class weights
    model_info(model, report='summary')  # 'full' or 'summary'
    nb = len(dataloader)
    maps = np.zeros(nc)  # mAP per class
    results = (0, 0, 0, 0, 0)  # P, R, mAP, F1, test_loss
    n_burnin = min(round(nb / 5 + 1), 1000)  # burn-in batches
    t, t0 = time.time(), time.time()
    torch.cuda.empty_cache()
    # 训练过程
    for epoch in range(start_epoch, epochs):
        # 将model.training设置成true
        model.train()
        print(('\n%8s%12s' + '%10s' * 7) %
              ('Epoch', 'Batch', 'GIoU/xy', 'wh', 'obj', 'cls', 'total', 'targets', 'img_size'))
        # Update scheduler
        scheduler.step()
        # Freeze backbone at epoch 0, unfreeze at epoch 1 (optional)
        freeze_backbone = False
        if freeze_backbone and epoch < 2:
            for name, p in model.named_parameters():
                if int(name.split('.')[1]) < cutoff:  # if layer < 75
                    p.requires_grad = False if epoch == 0 else True
        # # Update image weights (optional)
        # w = model.class_weights.cpu().numpy() * (1 - maps)  # class weights
        # image_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)  # random weighted index
        mloss = torch.zeros(5).to(device)  # mean losses
        pbar = tqdm(enumerate(dataloader), total=nb)  # progress bar
        for i, (imgs, targets, paths, _) in pbar:
            imgs = imgs.to(device)
            targets = targets.to(device)
            # Multi-Scale training TODO: short-side to 32-multiple https://github.com/ultralytics/yolov3/issues/358
            if multi_scale:
                if (i + nb * epoch) / accumulate % 10 == 0:  #  adjust (67% - 150%) every 10 batches
                    img_size = random.choice(range(img_size_min, img_size_max + 1)) * 32
                    # print('img_size = %g' % img_size)
                scale_factor = img_size / max(imgs.shape[-2:])
                imgs = F.interpolate(imgs, scale_factor=scale_factor, mode='bilinear', align_corners=False)
            # Plot images with bounding boxes
            if epoch == 0 and i == 0:
                plot_images(imgs=imgs, targets=targets, paths=paths, fname='train_batch%g.jpg' % i)
            # SGD burn-in
            if epoch == 0 and i <= n_burnin:
                lr = hyp['lr0'] * (i / n_burnin) ** 4
                for x in optimizer.param_groups:
                    x['lr'] = lr
            # Run model
            pred = model(imgs)
            # Compute loss
            # compute_loss是作者自己写的，是重点章节，哎哟真不容易，总算到重点章节了
            loss, loss_items = compute_loss(pred, targets, model, giou_loss=not opt.xywh)
            if torch.isnan(loss):
                print('WARNING: nan loss detected, ending training')
                return results
            # Compute gradient
            if mixed_precision:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()
            # Accumulate gradient for x batches before optimizing
            if (i + 1) % accumulate == 0 or (i + 1) == nb:
                optimizer.step()
                optimizer.zero_grad()
            # Print batch results
            # 这段蛮有教学意义的, tqdm怎么显示信息 
            mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
            # s = ('%8s%12s' + '%10.3g' * 7) % ('%g/%g' % (epoch, epochs - 1), '%g/%g' % (i, nb - 1), *mloss, len(targets), time.time() - t)
            s = ('%8s%12s' + '%10.3g' * 7) % (
                '%g/%g' % (epoch, epochs - 1), '%g/%g' % (i, nb - 1), *mloss, len(targets), img_size)
            t = time.time()
            pbar.set_description(s)  # print(s)
        # Report time
        # dt = (time.time() - t0) / 3600
        # print('%g epochs completed in %.3f hours.' % (epoch - start_epoch + 1, dt))
        # Calculate mAP (always test final epoch, skip first 5 if opt.nosave)
        if not (opt.notest or (opt.nosave and epoch < 10)) or epoch == epochs - 1:
            with torch.no_grad():
                # 这里用到了作者额外写的一个类 test.py 用于getmAP和loss after each epoch
                results, maps = test.test(cfg, data_cfg, batch_size=batch_size, img_size=opt.img_size, model=model,
                                          conf_thres=0.1)
        # Write epoch results
        with open('results.txt', 'a') as file:
            file.write(s + '%11.3g' * 5 % results + '\n')  # P, R, mAP, F1, test_loss
        # Update best map
        fitness = results[2]
        if fitness > best_fitness:
            best_fitness = fitness
        # Save training results
        # 保存训练结果和checkpoint
        save = (not opt.nosave) or ((not opt.evolve) and (epoch == epochs - 1))
        if save:
            with open('results.txt', 'r') as file:
                # Create checkpoint
                chkpt = {'epoch': epoch,
                         'best_fitness': best_fitness,
                         'training_results': file.read(),
                         'model': model.module.state_dict() if type(
                             model) is nn.parallel.DistributedDataParallel else model.state_dict(),
                         'optimizer': optimizer.state_dict()}
            # Save last checkpoint
            torch.save(chkpt, last)
            if opt.bucket:
                os.system('gsutil cp %s gs://%s' % (last, opt.bucket))  # upload to bucket
            # Save best checkpoint
            # mAP是评价检测性能的极佳指标，best_fitness用来记录当前最好的检测性能。best_fitness初始化为0.0
            if best_fitness == fitness:
                torch.save(chkpt, best)
            # Save backup every 10 epochs (optional)
            # 保存中间结果
            if epoch > 0 and epoch % 10 == 0:
                torch.save(chkpt, weights + 'backup%g.pt' % epoch)
            # Delete checkpoint
            del chkpt
    return results

上文的重点梳理。

1. torch.load(map_location=), cpu与gpu load时相互转化
2. model.load_state_dict(strict=), strict默认是True，表示预训练模型的层和自己定义的网络结构层严格对应相等（比如层名和维度）
3. model.parameters()，看源码可以发现本质上来源于named_parameters()，后者是(name, param)，前者只有param
4. requires_grad, 是否进行梯度更新

补充知识

1. 加载预训练模型的权重

resnet152 = models.resnet152(pretrained=True) 
pretrained_dict = resnet152.state_dict() 
"""
加载torchvision中的预训练模型和参数后通过state_dict()方法提取参数    
也可以直接从官方model_zoo下载：    
pretrained_dict = model_zoo.load_url(model_urls['resnet152'])""" 
model_dict = model.state_dict() 
# 将pretrained_dict里不属于model_dict的键剔除掉 
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict} 
# 更新现有的model_dict 
model_dict.update(pretrained_dict) 
# 加载我们真正需要的state_dict 
model.load_state_dict(model_dict)

下面的暂时没看，没我需要的重点

def print_mutation(hyp, results):
    # Write mutation results
    a = '%11s' * len(hyp) % tuple(hyp.keys())  # hyperparam keys
    b = '%11.4g' * len(hyp) % tuple(hyp.values())  # hyperparam values
    c = '%11.3g' * len(results) % results  # results (P, R, mAP, F1, test_loss)
    print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c))

    if opt.bucket:
        os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket)  # download evolve.txt
        with open('evolve.txt', 'a') as f:  # append result
            f.write(c + b + '\n')
        os.system('gsutil cp evolve.txt gs://%s' % opt.bucket)  # upload evolve.txt
    else:
        with open('evolve.txt', 'a') as f:
            f.write(c + b + '\n')


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--epochs', type=int, default=100, help='number of epochs')
    parser.add_argument('--batch-size', type=int, default=16, help='batch size')
    parser.add_argument('--accumulate', type=int, default=4, help='number of batches to accumulate before optimizing')
    parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='cfg file path')
    parser.add_argument('--data-cfg', type=str, default='data/coco_64img.data', help='coco.data file path')
    parser.add_argument('--multi-scale', action='store_true', help='train at (1/1.5)x - 1.5x sizes')
    parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)')
    parser.add_argument('--rect', action='store_true', help='rectangular training')
    parser.add_argument('--resume', action='store_true', help='resume training flag')
    parser.add_argument('--transfer', action='store_true', help='transfer learning flag')
    parser.add_argument('--num-workers', type=int, default=4, help='number of Pytorch DataLoader workers')
    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('--xywh', action='store_true', help='use xywh loss instead of GIoU loss')
    parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters')
    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
    parser.add_argument('--var', default=0, type=int, help='debug variable')
    opt = parser.parse_args()
    print(opt)

    if opt.evolve:
        opt.notest = True  # only test final epoch
        opt.nosave = True  # only save final checkpoint

    # Train
    results = train(opt.cfg,
                    opt.data_cfg,
                    img_size=opt.img_size,
                    epochs=opt.epochs,
                    batch_size=opt.batch_size,
                    accumulate=opt.accumulate)

    # Evolve hyperparameters (optional)
    if opt.evolve:
        gen = 1000  # generations to evolve
        print_mutation(hyp, results)  # Write mutation results

        for _ in range(gen):
            # Get best hyperparameters
            x = np.loadtxt('evolve.txt', ndmin=2)
            fitness = x[:, 2] * 0.5 + x[:, 3] * 0.5  # fitness as weighted combination of mAP and F1
            x = x[fitness.argmax()]  # select best fitness hyps
            for i, k in enumerate(hyp.keys()):
                hyp[k] = x[i + 5]

            # Mutate
            init_seeds(seed=int(time.time()))
            s = [.15, .15, .15, .15, .15, .15, .15, .15, .15, .00, .05, .10]  # fractional sigmas
            for i, k in enumerate(hyp.keys()):
                x = (np.random.randn(1) * s[i] + 1) ** 2.0  # plt.hist(x.ravel(), 300)
                hyp[k] *= float(x)  # vary by 20% 1sigma

            # Clip to limits
            keys = ['lr0', 'iou_t', 'momentum', 'weight_decay']
            limits = [(1e-4, 1e-2), (0.00, 0.70), (0.60, 0.95), (0, 0.01)]
            for k, v in zip(keys, limits):
                hyp[k] = np.clip(hyp[k], v[0], v[1])

            # Train mutation
            results = train(opt.cfg,
                            opt.data_cfg,
                            img_size=opt.img_size,
                            epochs=opt.epochs,
                            batch_size=opt.batch_size,
                            accumulate=opt.accumulate)

            # Write mutation results
            print_mutation(hyp, results)

            # # Plot results
            # import numpy as np
            # import matplotlib.pyplot as plt
            # a = np.loadtxt('evolve_1000val.txt')
            # x = a[:, 2] * a[:, 3]  # metric = mAP * F1
            # weights = (x - x.min()) ** 2
            # fig = plt.figure(figsize=(14, 7))
            # for i in range(len(hyp)):
            #     y = a[:, i + 5]
            #     mu = (y * weights).sum() / weights.sum()
            #     plt.subplot(2, 5, i+1)
            #     plt.plot(x.max(), mu, 'o')
            #     plt.plot(x, y, '.')
            #     print(list(hyp.keys())[i],'%.4g' % mu)