yolov5中增加了自适应锚定框(Auto Learning Bounding Box Anchors),而其他yolo系列是没有的。

一、默认锚定框

Yolov5 中默认保存了一些针对 coco数据集的预设锚定框,在 yolov5 的配置文件*.yaml 中已经预设了640×640图像大小下锚定框的尺寸(以 yolov5s.yaml 为例):

# anchors
anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 anchors参数共有三行,每行9个数值;且每一行代表应用不同的特征图;
1、第一行是在最大的特征图上的锚框

2、第二行是在中间的特征图上的锚框

3、第三行是在最小的特征图上的锚框;

在目标检测任务中,一般希望在大的特征图上去检测小目标,因为大特征图才含有更多小目标信息,因此大特征图上的anchor数值通常设置为小数值,而小特征图上数值设置为大数值检测大的目标。
 

二、自定义锚定框

1、训练时自动计算锚定框

yolov5 中不是只使用默认锚定框,在开始训练之前会对数据集中标注信息进行核查,计算此数据集标注信息针对默认锚定框的最佳召回率,当最佳召回率大于或等于0.98,则不需要更新锚定框;如果最佳召回率小于0.98,则需要重新计算符合此数据集的锚定框。

核查锚定框是否适合要求的函数在 /utils/autoanchor.py 文件中:

def check_anchors(dataset, model, thr=4.0, imgsz=640):

 其中 thr 是指 数据集中标注框宽高比最大阈值,默认是使用 超参文件 hyp.scratch.yaml 中的 “anchor_t” 参数值。

核查主要代码如下:

    def metric(k):  # compute metric
        r = wh[:, None] / k[None]
        x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
        best = x.max(1)[0]  # best_x
        aat = (x > 1. / thr).float().sum(1).mean()  # anchors above threshold
        bpr = (best > 1. / thr).float().mean()  # best possible recall
        return bpr, aat

    bpr, aat = metric(m.anchor_grid.clone().cpu().view(-1, 2))

其中两个指标需要解释一下(bpr 和 aat):

bpr(best possible recall) 

aat(anchors above threshold) 

 其中 bpr 参数就是判断是否需要重新计算锚定框的依据(是否小于 0.98)。

重新计算符合此数据集标注框的锚定框,是利用 kmean聚类方法实现的,代码在  /utils/autoanchor.py 文件中:

def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
    """ Creates kmeans-evolved anchors from training dataset

        Arguments:
            path: path to dataset *.yaml, or a loaded dataset
            n: number of anchors
            img_size: image size used for training
            thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0
            gen: generations to evolve anchors using genetic algorithm
            verbose: print all results

        Return:
            k: kmeans evolved anchors

        Usage:
            from utils.autoanchor import *; _ = kmean_anchors()
    """
    thr = 1. / thr
    prefix = colorstr('autoanchor: ')

    def metric(k, wh):  # compute metrics
        r = wh[:, None] / k[None]
        x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
        # x = wh_iou(wh, torch.tensor(k))  # iou metric
        return x, x.max(1)[0]  # x, best_x

    def anchor_fitness(k):  # mutation fitness
        _, best = metric(torch.tensor(k, dtype=torch.float32), wh)
        return (best * (best > thr).float()).mean()  # fitness

    def print_results(k):
        k = k[np.argsort(k.prod(1))]  # sort small to large
        x, best = metric(k, wh0)
        bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n  # best possible recall, anch > thr
        print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr')
        print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, '
              f'past_thr={x[x > thr].mean():.3f}-mean: ', end='')
        for i, x in enumerate(k):
            print('%i,%i' % (round(x[0]), round(x[1])), end=',  ' if i < len(k) - 1 else '\n')  # use in *.cfg
        return k

    if isinstance(path, str):  # *.yaml file
        with open(path) as f:
            data_dict = yaml.load(f, Loader=yaml.SafeLoader)  # model dict
        from utils.datasets import LoadImagesAndLabels
        dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
    else:
        dataset = path  # dataset

    # Get label wh
    shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
    wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])  # wh

    # Filter
    i = (wh0 < 3.0).any(1).sum()
    if i:
        print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.')
    wh = wh0[(wh0 >= 2.0).any(1)]  # filter > 2 pixels
    # wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1)  # multiply by random scale 0-1

    # Kmeans calculation
    print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...')
    s = wh.std(0)  # sigmas for whitening
    k, dist = kmeans(wh / s, n, iter=30)  # points, mean distance
    k *= s
    wh = torch.tensor(wh, dtype=torch.float32)  # filtered
    wh0 = torch.tensor(wh0, dtype=torch.float32)  # unfiltered
    k = print_results(k)

    # Plot
    # k, d = [None] * 20, [None] * 20
    # for i in tqdm(range(1, 21)):
    #     k[i-1], d[i-1] = kmeans(wh / s, i)  # points, mean distance
    # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True)
    # ax = ax.ravel()
    # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.')
    # fig, ax = plt.subplots(1, 2, figsize=(14, 7))  # plot wh
    # ax[0].hist(wh[wh[:, 0]<100, 0],400)
    # ax[1].hist(wh[wh[:, 1]<100, 1],400)
    # fig.savefig('wh.png', dpi=200)

    # Evolve
    npr = np.random
    f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
    pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:')  # progress bar
    for _ in pbar:
        v = np.ones(sh)
        while (v == 1).all():  # mutate until a change occurs (prevent duplicates)
            v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
        kg = (k.copy() * v).clip(min=2.0)
        fg = anchor_fitness(kg)
        if fg > f:
            f, k = fg, kg.copy()
            pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}'
            if verbose:
                print_results(k)

    return print_results(k)

对 kmean_anchors()函数中的参数做一下简单解释(代码中已经有了英文注释):

  • path:包含数据集文件路径等相关信息的 yaml 文件(比如 coco128.yaml), 或者 数据集张量(yolov5 自动计算锚定框时就是用的这种方式,先把数据集标签信息读取再处理)
  • n:锚定框的数量,即有几组;默认值是9
  • img_size:图像尺寸。计算数据集样本标签框的宽高比时,是需要缩放到 img_size 大小后再计算的;默认值是640
  • thr:数据集中标注框宽高比最大阈值,默认是使用 超参文件 hyp.scratch.yaml 中的 “anchor_t” 参数值;默认值是4.0;自动计算时,会自动根据你所使用的数据集,来计算合适的阈值。
  • gen:kmean聚类算法迭代次数,默认值是1000
  • verbose:是否打印输出所有计算结果,默认值是true
  • 如果你不想自动计算锚定框,可以在 train.py 中设置参数即可:

    parser.add_argument('--noautoanchor', action='store_true', help='disable autoanchor check')

     

    2、训练前手动计算锚定框

    如果使用 yolov5 训练效果并不好(排除其他原因,只考虑 “预设锚定框” 这个因素), yolov5在核查默认锚定框是否符合要求时,计算的最佳召回率大于0.98,没有自动计算锚定框;此时你可以自己手动计算锚定框。【即使自己的数据集中目标宽高比最大值小于4,默认锚定框也不一定是最合适的】

     首先可以自行编写一个程序,统计一下你所训练的数据集所有标签框宽高比,看下宽高比主要分布在哪个范围、最大宽高比是多少? 比如:你使用的数据集中目标宽高比最大达到了 5:1(甚至 10:1) ,那肯定需要重新计算锚定框了,针对coco数据集的最大宽高比是 4:1 。

    然后在 yolov5 程序中创建一个新的 python 文件 test.py,手动计算锚定框:

    import utils.autoanchor as autoAC
    
    # 对数据集重新计算 anchors
    new_anchors = autoAC.kmean_anchors('./data/mydata.yaml', 9, 640, 5.0, 1000, True)
    print(new_anchors)

    输入信息如下(只截取了部分):

    autoanchor: Evolving anchors with Genetic Algorithm: fitness = 0.6604:  87%|████████▋ | 866/1000 [00:00<00:00, 2124.00it/s]autoanchor: thr=0.25: 0.9839 best possible recall, 3.84 anchors past thr
    autoanchor: n=9, img_size=640, metric_all=0.267/0.662-mean/best, past_thr=0.476-mean: 15,20,  38,25,  55,65,  131,87,  97,174,  139,291,  256,242,  368,382,  565,422
    autoanchor: thr=0.25: 0.9849 best possible recall, 3.84 anchors past thr
    autoanchor: n=9, img_size=640, metric_all=0.267/0.663-mean/best, past_thr=0.476-mean: 15,20,  39,26,  54,64,  127,87,  97,176,  142,286,  257,245,  374,379,  582,424
    autoanchor: thr=0.25: 0.9849 best possible recall, 3.84 anchors past thr
    autoanchor: n=9, img_size=640, metric_all=0.267/0.663-mean/best, past_thr=0.476-mean: 15,20,  39,26,  54,63,  126,86,  97,176,  143,285,  258,241,  369,381,  583,424
    autoanchor: thr=0.25: 0.9849 best possible recall, 3.84 anchors past thr
    autoanchor: n=9, img_size=640, metric_all=0.267/0.663-mean/best, past_thr=0.476-mean: 15,20,  39,26,  54,63,  127,86,  97,176,  143,285,  258,241,  369,380,  583,424
    autoanchor: thr=0.25: 0.9849 best possible recall, 3.84 anchors past thr
    autoanchor: n=9, img_size=640, metric_all=0.267/0.663-mean/best, past_thr=0.476-mean: 15,20,  39,26,  53,63,  127,86,  97,175,  143,284,  257,243,  369,381,  582,422
    autoanchor: thr=0.25: 0.9849 best possible recall, 3.84 anchors past thr
    autoanchor: n=9, img_size=640, metric_all=0.267/0.663-mean/best, past_thr=0.476-mean: 15,20,  40,26,  53,62,  129,85,  96,175,  143,287,  256,240,  370,378,  582,419
    autoanchor: Evolving anchors with Genetic Algorithm: fitness = 0.6605: 100%|██████████| 1000/1000 [00:00<00:00, 2170.29it/s]
    Scanning '..\coco128\labels\train2017.cache' for images and labels... 128 found, 0 missing, 2 empty, 0 corrupted: 100%|██████████| 128/128 [00:00<?, ?it/s]
    autoanchor: thr=0.25: 0.9849 best possible recall, 3.84 anchors past thr
    autoanchor: n=9, img_size=640, metric_all=0.267/0.663-mean/best, past_thr=0.476-mean: 15,20,  40,26,  53,62,  129,85,  96,175,  143,287,  256,240,  370,378,  582,419
    [[     14.931      20.439]
     [     39.648       25.53]
     [     53.371       62.35]
     [     129.07      84.774]
     [     95.719      175.08]
     [     142.69      286.95]
     [     256.46      239.83]
     [      369.9       378.3]
     [     581.87      418.56]]
    
    Process finished with exit code 0

    输出的 9 组新的锚定框即是根据自己的数据集来计算的,可以按照顺序替换到你所使用的配置文件*.yaml中(比如 yolov5s.yaml)。就可以重新训练了。

    参考的博文(表示感谢!):

    https://github.com/ultralytics/yolov5

    https://blog.csdn.net/flyfish1986/article/details/117594265

    https://zhuanlan.zhihu.com/p/183838757

    https://blog.csdn.net/aabbcccddd01/article/details/109578614

    来源:高祥xiang

    物联沃分享整理
    物联沃-IOTWORD物联网 » yolov5 anchors设置详解

    发表评论