【Yolov系列】Yolov5学习（一）补充1.1：自适应锚框计算

已于 2024-04-19 17:23:50 修改
阅读量1.6k 收藏 35
点赞数 25
分类专栏： AI学习 Yolo系列 文章标签： YOLO 学习
于 2024-04-17 18:19:57 首次发布
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/qq_43737512/article/details/137867943
版权

1、Yolov5的网络结构

  • Yolov5中使用的Coco数据集输入图片的尺寸为640*640，但是训练过程的输入尺寸并不唯一，Yolov5可以采用Mosaic增强技术把4张图片的部分组成了一张尺寸一定的输入图片。如果需要使用预训练权重，最好将输入图片尺寸调整到与作者相同的尺寸，输入图片尺寸必须是32的倍数，这与anchor检测的阶段有关。

Yolov5s网络结构示意图：

  • 当输入尺寸为640*640时，会得到3个不同尺度的输出：80x80（640/8）、40x40（640/16）、20x20（640/32）。
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6个数值，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，80x80代表浅层的特征图（P3），包含较多的低层级信息，适合用于检测小目标，所以这一特征图所用的anchor尺度较小；
  2. 第二行是在中间的特征图上的锚框，40x40代表中间的特征图（P4），介于浅层和深层这两个尺度之间的anchor用来检测中等大小的目标；
  3. 第三行是在最小的特征图上的锚框，20x20代表深层的特征图（P5），包含更多高层级的信息，如轮廓、结构等信息，适合用于大目标的检测，所以这一特征图所用的anchor尺度较大。

待验证注释：

查阅其他博主博客发现，Yolov5也可以不预设anchor，直接写个3，此时yolov5就会自动按照训练集聚类anchor：

# Parameters
nc: 4 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.50 # layer channel multiple
anchors: 3

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，yolov5之所以能高效快速地检测跨尺度目标，这种对不同特征图使用不同尺度的anchor的思想功不可没。

2、自适应锚框计算

  • Yolov5 中并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：
# YOLOv5 🚀 by Ultralytics, AGPL-3.0 license
"""AutoAnchor utils."""

import random

import numpy as np
import torch
import yaml
from tqdm import tqdm

from utils import TryExcept
from utils.general import LOGGER, TQDM_BAR_FORMAT, colorstr

PREFIX = colorstr("AutoAnchor: ")


def check_anchor_order(m):
    """Checks and corrects anchor order against stride in YOLOv5 Detect() module if necessary."""
    a = m.anchors.prod(-1).mean(-1).view(-1)  # mean anchor area per output layer
    da = a[-1] - a[0]  # delta a
    ds = m.stride[-1] - m.stride[0]  # delta s
    if da and (da.sign() != ds.sign()):  # same order
        LOGGER.info(f"{PREFIX}Reversing anchor order")
        m.anchors[:] = m.anchors.flip(0)


@TryExcept(f"{PREFIX}ERROR")
def check_anchors(dataset, model, thr=4.0, imgsz=640):
    """Evaluates anchor fit to dataset and adjusts if necessary, supporting customizable threshold and image size."""
    m = model.module.model[-1] if hasattr(model, "module") else model.model[-1]  # Detect()
    shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
    scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1))  # augment scale
    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float()  # wh

    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

    stride = m.stride.to(m.anchors.device).view(-1, 1, 1)  # model strides
    anchors = m.anchors.clone() * stride  # current anchors
    bpr, aat = metric(anchors.cpu().view(-1, 2))
    s = f"\n{PREFIX}{aat:.2f} anchors/target, {bpr:.3f} Best Possible Recall (BPR). "
    if bpr > 0.98:  # threshold to recompute
        LOGGER.info(f"{s}Current anchors are a good fit to dataset ✅")
    else:
        LOGGER.info(f"{s}Anchors are a poor fit to dataset ⚠️, attempting to improve...")
        na = m.anchors.numel() // 2  # number of anchors
        anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
        new_bpr = metric(anchors)[0]
        if new_bpr > bpr:  # replace anchors
            anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
            m.anchors[:] = anchors.clone().view_as(m.anchors)
            check_anchor_order(m)  # must be in pixel-space (not grid-space)
            m.anchors /= stride
            s = f"{PREFIX}Done ✅ (optional: update model *.yaml to use these anchors in the future)"
        else:
            s = f"{PREFIX}Done ⚠️ (original anchors better than new anchors, proceeding with original anchors)"
        LOGGER.info(s)


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

    Arguments:
        dataset: path to data.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()
    """
    from scipy.cluster.vq import kmeans

    npr = np.random
    thr = 1 / thr

    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, verbose=True):
        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
        s = (
            f"{PREFIX}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr\n"
            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: "
        )
        for x in k:
            s += "%i,%i, " % (round(x[0]), round(x[1]))
        if verbose:
            LOGGER.info(s[:-2])
        return k

    if isinstance(dataset, str):  # *.yaml file
        with open(dataset, errors="ignore") as f:
            data_dict = yaml.safe_load(f)  # model dict
        from utils.dataloaders import LoadImagesAndLabels

        dataset = LoadImagesAndLabels(data_dict["train"], augment=True, rect=True)

    # 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:
        LOGGER.info(f"{PREFIX}WARNING ⚠️ Extremely small objects found: {i} of {len(wh0)} labels are <3 pixels in size")
    wh = wh0[(wh0 >= 2.0).any(1)].astype(np.float32)  # filter > 2 pixels
    # wh = wh * (npr.rand(wh.shape[0], 1) * 0.9 + 0.1)  # multiply by random scale 0-1

    # Kmeans init
    try:
        LOGGER.info(f"{PREFIX}Running kmeans for {n} anchors on {len(wh)} points...")
        assert n <= len(wh)  # apply overdetermined constraint
        s = wh.std(0)  # sigmas for whitening
        k = kmeans(wh / s, n, iter=30)[0] * s  # points
        assert n == len(k)  # kmeans may return fewer points than requested if wh is insufficient or too similar
    except Exception:
        LOGGER.warning(f"{PREFIX}WARNING ⚠️ switching strategies from kmeans to random init")
        k = np.sort(npr.rand(n * 2)).reshape(n, 2) * img_size  # random init
    wh, wh0 = (torch.tensor(x, dtype=torch.float32) for x in (wh, wh0))
    k = print_results(k, verbose=False)

    # 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
    f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
    pbar = tqdm(range(gen), bar_format=TQDM_BAR_FORMAT)  # 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) * random.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, verbose)

    return print_results(k).astype(np.float32)
  • 核查的主要代码：
@TryExcept(f"{PREFIX}ERROR")
def check_anchors(dataset, model, thr=4.0, imgsz=640):
    """Evaluates anchor fit to dataset and adjusts if necessary, supporting customizable threshold and image size."""
    m = model.module.model[-1] if hasattr(model, "module") else model.model[-1]  # Detect()
    shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
    scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1))  # augment scale
    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float()  # wh

    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

    stride = m.stride.to(m.anchors.device).view(-1, 1, 1)  # model strides
    anchors = m.anchors.clone() * stride  # current anchors
    bpr, aat = metric(anchors.cpu().view(-1, 2))
    s = f"\n{PREFIX}{aat:.2f} anchors/target, {bpr:.3f} Best Possible Recall (BPR). "
    if bpr > 0.98:  # threshold to recompute
        LOGGER.info(f"{s}Current anchors are a good fit to dataset ✅")
    else:
        LOGGER.info(f"{s}Anchors are a poor fit to dataset ⚠️, attempting to improve...")
        na = m.anchors.numel() // 2  # number of anchors
        anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
        new_bpr = metric(anchors)[0]
        if new_bpr > bpr:  # replace anchors
            anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
            m.anchors[:] = anchors.clone().view_as(m.anchors)
            check_anchor_order(m)  # must be in pixel-space (not grid-space)
            m.anchors /= stride
            s = f"{PREFIX}Done ✅ (optional: update model *.yaml to use these anchors in the future)"
        else:
            s = f"{PREFIX}Done ⚠️ (original anchors better than new anchors, proceeding with original anchors)"
        LOGGER.info(s)

ps：
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，主要代码如下：
def kmean_anchors(dataset="./data/coco128.yaml", n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
    """
    Creates kmeans-evolved anchors from training dataset.

    Arguments:
        dataset: path to data.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()
    """
    from scipy.cluster.vq import kmeans

    npr = np.random
    thr = 1 / thr

    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, verbose=True):
        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
        s = (
            f"{PREFIX}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr\n"
            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: "
        )
        for x in k:
            s += "%i,%i, " % (round(x[0]), round(x[1]))
        if verbose:
            LOGGER.info(s[:-2])
        return k

    if isinstance(dataset, str):  # *.yaml file
        with open(dataset, errors="ignore") as f:
            data_dict = yaml.safe_load(f)  # model dict
        from utils.dataloaders import LoadImagesAndLabels

        dataset = LoadImagesAndLabels(data_dict["train"], augment=True, rect=True)

    # 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:
        LOGGER.info(f"{PREFIX}WARNING ⚠️ Extremely small objects found: {i} of {len(wh0)} labels are <3 pixels in size")
    wh = wh0[(wh0 >= 2.0).any(1)].astype(np.float32)  # filter > 2 pixels
    # wh = wh * (npr.rand(wh.shape[0], 1) * 0.9 + 0.1)  # multiply by random scale 0-1

    # Kmeans init
    try:
        LOGGER.info(f"{PREFIX}Running kmeans for {n} anchors on {len(wh)} points...")
        assert n <= len(wh)  # apply overdetermined constraint
        s = wh.std(0)  # sigmas for whitening
        k = kmeans(wh / s, n, iter=30)[0] * s  # points
        assert n == len(k)  # kmeans may return fewer points than requested if wh is insufficient or too similar
    except Exception:
        LOGGER.warning(f"{PREFIX}WARNING ⚠️ switching strategies from kmeans to random init")
        k = np.sort(npr.rand(n * 2)).reshape(n, 2) * img_size  # random init
    wh, wh0 = (torch.tensor(x, dtype=torch.float32) for x in (wh, wh0))
    k = print_results(k, verbose=False)

    # 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
    f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
    pbar = tqdm(range(gen), bar_format=TQDM_BAR_FORMAT)  # 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) * random.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, verbose)

    return print_results(k).astype(np.float32)

参数释意：

  • dataset：包含数据集文件路径等相关信息的 yaml 文件，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）。默认 coco128.yaml
  • 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，默认使用超参文件./data/hyps/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
  • 如果不想自动计算锚定框，可以在train.py中设置参数：
parser.add_argument('--noautoanchor', action='store_true', help='disable autoanchor check')

3、手动锚框计算

  • 1. 在./data文件夹下复制VOC.yaml文件，自己命名，如train_data.yaml文件，修改文件路径为绝对路径
train: # train images (relative to 'path')  16551 images
  F:/dataset/yolo/yolov5_up_sum/yolov5-master/datasets/train_data/images/train

val: # val images (relative to 'path')  4952 images
  F:/dataset/yolo/yolov5_up_sum/yolov5-master/datasets/train_data/images/val
test: # test images (optional)


# Classes
names: ['Team1', 'Team2', 'Ball', 'Team3']
  • 数据集中需含有.cache文件

如果数据集中不存在.cache文件，查找Yolov5训练自己数据集的帖子，按照流程运行train.py文件，成功的话文件夹下会自动生成.cache文件

.cache文件：原始数据里没有该文件，yolov5自动生成的缓存文件，再下次读数据时，直接读取缓存文件，速度更快

  • 2. 在Yolov5目录下新建一个.py文件，调用kmeans算法计算anchor：
import utils.autoanchor as autoAC


if __name__ == '__main__':
    config = "./data/train_data.yaml"
    # 对数据集重新计算 anchors
    new_anchors = autoAC.kmean_anchors(config, 9, 640, 5.0, 1000, True)
    print(new_anchors)

运行结果展示：

albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8))
Scanning F:\dataset\yolo\yolov5_up_sum\yolov5-master\datasets\train_data\labels\train... 107 images, 0 backgrounds, 0 corrupt: 100%|██████████| 107/107 [00:15<00:00,  6.94it/s]
WARNING  Cache directory F:\dataset\yolo\yolov5_up_sum\yolov5-master\datasets\train_data\labels is not writeable: [WinError 183] : 'F:\\dataset\\yolo\\yolov5_up_sum\\yolov5-master\\datasets\\train_data\\labels\\train.cache.npy' -> 'F:\\dataset\\yolo\\yolov5_up_sum\\yolov5-master\\datasets\\train_data\\labels\\train.cache'
AutoAnchor: Running kmeans for 9 anchors on 855 points...
  0%|          | 0/1000 [00:00<?, ?it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.79 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.399/0.789-mean/best, past_thr=0.488-mean: 15,20, 25,23, 20,46, 35,39, 34,71, 62,60, 76,129, 100,213, 162,232
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.83 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.400/0.789-mean/best, past_thr=0.486-mean: 15,20, 25,23, 20,47, 36,40, 34,67, 63,62, 74,126, 102,216, 159,232
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.81 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.399/0.790-mean/best, past_thr=0.485-mean: 15,19, 25,23, 20,47, 36,39, 34,69, 63,62, 73,128, 97,212, 161,230
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.82 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.398/0.790-mean/best, past_thr=0.483-mean: 15,19, 25,23, 19,46, 36,39, 32,67, 64,63, 74,131, 101,205, 161,220
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.84 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.398/0.790-mean/best, past_thr=0.483-mean: 15,19, 25,23, 20,46, 35,39, 33,67, 64,63, 74,131, 101,205, 160,219
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.84 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.399/0.791-mean/best, past_thr=0.484-mean: 15,19, 25,23, 19,46, 35,39, 32,66, 64,65, 74,131, 100,205, 160,222
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.84 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.399/0.791-mean/best, past_thr=0.484-mean: 15,19, 25,23, 20,46, 35,39, 32,66, 64,64, 74,131, 100,205, 160,221
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.88 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.792-mean/best, past_thr=0.492-mean: 14,20, 25,24, 19,45, 34,38, 33,59, 60,64, 73,128, 97,209, 147,221
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.88 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.408/0.794-mean/best, past_thr=0.494-mean: 14,20, 25,22, 20,45, 30,36, 33,58, 60,59, 73,123, 90,203, 153,229
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.88 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.408/0.794-mean/best, past_thr=0.494-mean: 14,20, 25,22, 20,45, 30,36, 33,58, 60,59, 74,124, 89,204, 154,228
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.89 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.408/0.795-mean/best, past_thr=0.493-mean: 14,19, 25,22, 20,44, 31,36, 34,57, 60,59, 75,121, 89,211, 154,225
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.89 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.407/0.795-mean/best, past_thr=0.493-mean: 14,19, 25,22, 20,44, 31,36, 34,57, 59,59, 75,121, 89,212, 154,225
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.88 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.407/0.795-mean/best, past_thr=0.493-mean: 14,19, 25,22, 20,44, 31,36, 34,57, 59,59, 75,122, 89,213, 155,224
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.84 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.403/0.795-mean/best, past_thr=0.489-mean: 14,19, 24,21, 19,44, 31,37, 35,57, 60,49, 76,122, 93,226, 146,224
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.82 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.404/0.795-mean/best, past_thr=0.491-mean: 14,19, 24,21, 19,44, 31,37, 35,58, 60,50, 76,121, 94,227, 148,224
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.82 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.404/0.796-mean/best, past_thr=0.491-mean: 14,19, 24,21, 19,44, 31,37, 35,58, 59,50, 76,121, 94,228, 149,225
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.85 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.404/0.796-mean/best, past_thr=0.491-mean: 14,19, 25,21, 19,44, 30,36, 33,58, 60,52, 75,116, 92,228, 149,227
AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7959:  16%|█▌        | 156/1000 [00:00<00:00, 1559.97it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.84 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.404/0.796-mean/best, past_thr=0.492-mean: 14,19, 25,21, 19,44, 30,36, 33,58, 60,52, 75,117, 92,228, 148,227
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.85 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.404/0.796-mean/best, past_thr=0.490-mean: 14,18, 24,20, 18,44, 29,36, 33,58, 59,53, 73,116, 91,228, 149,225
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.88 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.490-mean: 14,18, 23,20, 19,44, 29,36, 33,58, 59,53, 74,118, 92,219, 150,226
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 23,20, 19,43, 30,36, 32,58, 59,52, 73,119, 92,220, 148,230
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 23,20, 19,43, 30,36, 33,58, 59,53, 73,122, 92,220, 148,228
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.87 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 23,20, 19,43, 30,37, 33,58, 59,52, 72,122, 93,220, 147,229
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.87 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 23,20, 19,43, 30,37, 33,58, 59,52, 72,122, 93,220, 147,229
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 23,20, 19,43, 30,36, 33,58, 59,52, 73,122, 93,220, 146,230
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 24,20, 19,43, 30,36, 33,58, 58,53, 73,122, 93,222, 147,227
AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7964:  32%|███▏      | 321/1000 [00:00<00:00, 1600.54it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 24,20, 19,43, 30,37, 33,58, 58,53, 73,122, 93,222, 147,227
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 24,20, 19,43, 30,36, 33,58, 58,53, 73,122, 93,221, 147,227
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 24,20, 19,43, 30,36, 33,58, 58,53, 73,122, 93,222, 147,227
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.796-mean/best, past_thr=0.491-mean: 14,18, 24,20, 19,43, 30,36, 33,58, 58,53, 73,123, 92,222, 147,227
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.87 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.492-mean: 14,18, 24,20, 19,42, 30,36, 33,57, 58,54, 72,122, 93,218, 147,230
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.87 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.493-mean: 14,18, 24,20, 19,42, 30,35, 33,57, 57,54, 72,122, 93,218, 146,236
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.407/0.797-mean/best, past_thr=0.493-mean: 14,18, 24,20, 19,43, 30,36, 32,57, 57,54, 72,122, 93,218, 143,234
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.85 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.493-mean: 14,18, 24,21, 19,42, 30,36, 32,57, 58,55, 72,123, 93,219, 143,234
AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7968:  48%|████▊     | 482/1000 [00:00<00:00, 1548.48it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.85 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.493-mean: 14,18, 24,21, 19,42, 30,36, 32,58, 58,55, 72,123, 92,220, 143,236
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.87 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.492-mean: 14,19, 24,21, 19,42, 31,36, 32,58, 59,55, 72,122, 93,218, 143,236
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.87 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.492-mean: 14,19, 24,21, 19,42, 31,36, 32,58, 59,55, 72,122, 93,218, 143,236
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.87 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.492-mean: 14,18, 24,21, 19,42, 31,37, 32,58, 59,56, 73,123, 93,218, 143,236
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.492-mean: 14,18, 24,21, 19,42, 31,37, 32,58, 59,56, 72,123, 92,218, 143,235
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.84 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.797-mean/best, past_thr=0.493-mean: 14,18, 24,21, 19,42, 31,36, 32,59, 57,57, 72,126, 92,217, 142,240
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.84 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.797-mean/best, past_thr=0.493-mean: 14,18, 24,21, 19,42, 31,36, 32,59, 58,57, 72,126, 91,217, 142,240
AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7971:  65%|██████▍   | 647/1000 [00:00<00:00, 1563.54it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.85 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.797-mean/best, past_thr=0.492-mean: 14,18, 24,21, 19,42, 31,36, 32,59, 57,57, 73,126, 92,216, 142,239
AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7971:  82%|████████▏ | 816/1000 [00:00<00:00, 1607.99it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.85 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.797-mean/best, past_thr=0.492-mean: 14,18, 24,20, 19,42, 31,36, 32,58, 57,57, 72,126, 92,215, 140,237
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.86 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.406/0.797-mean/best, past_thr=0.492-mean: 14,18, 24,20, 19,42, 31,36, 32,58, 57,57, 71,126, 92,214, 140,236
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.85 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.797-mean/best, past_thr=0.492-mean: 14,18, 24,20, 19,42, 31,36, 32,58, 57,57, 71,126, 92,214, 140,236
[[     14.077      18.166]
 [     23.833       20.45]
 [      18.59      42.161]
 [     30.656      36.243]
 [     32.122      58.356]
 [     57.303      56.757]
 [     71.018      126.44]
 [     91.503      214.03]
 [     140.23      235.73]]
AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7972: 100%|██████████| 1000/1000 [00:00<00:00, 1630.48it/s]
AutoAnchor: thr=0.20: 1.0000 best possible recall, 6.85 anchors past thr
AutoAnchor: n=9, img_size=640, metric_all=0.405/0.797-mean/best, past_thr=0.492-mean: 14,18, 24,20, 19,42, 31,36, 32,58, 57,57, 71,126, 92,214, 140,236

Process finished with exit code 0

输出的9个坐标即为锚框中心坐标，复制yolov5s.yaml文件，自己命名，如yolov5s_train.yaml，将计算所得值按顺序修改至模型配置文件./model/yolov5s_train.yaml中，重新训练即可：

# YOLOv5 🚀 by Ultralytics, AGPL-3.0 license

# Parameters
nc: 4 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.50 # layer channel multiple
anchors:
  - [14.077, 18.166, 23.833, 20.45, 18.59, 42.161] # P3/8
  - [30.656, 36.243, 32.122, 58.356, 57.303, 56.757] # P4/16
  - [71.018, 126.44, 91.503, 214.03, 140.23, 235.73] # P5/32

4. 检测模块

（没看太懂，后面再查些资料）

anchor在模型中的应用涉及到了yolo系列目标框回归的过程。yolov5中的detect模块沿用了v3检测方式。

  • 1. 检测到的不是框而是偏移量： tx，ty指的是针对所在grid的左上角坐标的偏移量， tw，th指的是相对于anchor的宽高的偏移量，通过如下图的计算方式，得到bx,by,bw,bh就是最终的检测结果。

  • 2. 前面经过backbone，neck，head是panet的三个分支，可见特征图size不同，每个特征图分了13个网格，同一尺度的特征图对应了3个anchor，检测了[c,x,y,w,h]和num_class个的one-hot类别标签。3个尺度的特征图，总共就有9个anchor。

参考：

Yolov5的anchors设置详解

Yolov5的anchor详解

YOLOv5的anchor设定

（20）目标检测算法之YOLOv5计算预选框、详解anchor计算

悸惓
关注
  • 25
    点赞
  • 踩
  • 35
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
专栏目录
anchor自适应锚框计算
ahelloyou的博客
12-19 1万+
自适应锚框计算 预定义边框就是一组预设的边框，在训练时，以真实的边框位置相对于预设边框的偏移来构建（也就是我们打下的标签） 训练样本。 这就相当于，预设边框先大致在可能的位置“框“出来目标，然后再在这些预设边框的基础上进行调整。 一个Anchor Box可以由:边框的纵横比和边框的面积（尺度)来定义，相当于一系列预设边框的生成规则，根据Anchor Box，可以在图像的任意位置，生成一系列的边框 由于Anchor box 通常是以CNN提取到的Feature Map 的点为中心位置，生成边框，所以一个An
【YOLO】yolov5的训练策略2 -- 自动锚框autoanchor
AI学长
07-13 2819
在yolov5 中训练开始前，计算数据集标注信息针对默认锚定框的最佳召回率，当最佳召回率大于等于0.98时，则不需要更新锚定框；如果最佳召回率小于0.98，则需要重新计算数据集的锚定框，如果计算处理更好则更新原理的anchors。
YOLO系列梳理（三）YOLOv5
CV技术指南（微信公众号）
04-24 1万+
前言YOLOv5 是在 YOLOv4 出来之后没多久就横空出世了。今天笔者介绍一下 YOLOv5 的相关知识。目前 YOLOv5 发布了新的版本，6.0版本。在这里，YOLOv5 也在5.0基础上集成了更多特性，同时也对模型做了微调，并且优化了模型大小，减少了模型的参数量。那么这样，就更加适合移动端了。 欢迎关注公众号CV技术指南，专注于计算机视觉的技术总结、最新技术跟踪、经典论文解读、CV招聘信息。 YOLOv5 网络模型结构 与之前的 YOLOv3、YOLOv4 不同，v3、v4 除..
yolov5s-5.0网络模型结构图
热门推荐
qianxun
01-04 4万+
看了很多yolov5方面的东西，最近需要yolov5得模型结构图，但是网上的最多的是大白老师的，但是大白老师的yolov5得模型结构图不知道是哪个版本得，肯定不是5.0和6.0版本得。参考了大白老师得模型结构图和其他大佬的模型结构图，以及参考了yolov5得onnx。画出了以下得结构图，初次画不知道有些地方是否对不对，如果有错误，请大家指出。。。 这个模型结构图是用思维导图画出来得。这里面有几个点可能与其他人画的不一样。 1、5.0采用的激活函数是SiLU(),不再是LeaKyReLU()，所以这里
YOLOv5算法原理与网络结构
babyai996的博客
02-07 3万+
以 YOLOv5s为例，第一个CSP1中，使用了1个残差组件，因此是CSP1_1，在YOLOv5m中，增加了网络的深度，在第一个CSP1中，使用了2个残差组件，因此是CSP1_2，YOLOv5l中，同样的位置，则使用了3个残差组件，YOLOv5x中，使用了4个，其余CSP也是同样的原理，使用残差组件的个数如图5所示。第一，提高了数据集的复杂度：对多张图片，进行随机的缩放以及剪裁，之后再随机分布，进行拼接，使数据集得到极大地丰富，特别是进行随机缩放操作，可以增加许多小目标，训练得到的模型，鲁棒性会更好；
AWC：自适应Web组件框架
02-06
自适应Web组件框架 基于以下规范的自适应用户界面组件的实现框架： 用法 使用以下必要的标志启动最新版本的： $ <Application> --enable-blink-features=CustomElementsV1,ShadowDOMV1 Mac OS X 手动或通过 安装 ...
基于yolov5的深度学习目标检测程序，配置完环境可直接训练模型
05-06
YOLOV5：You Only Look Once目标检测模型在pytorch当中的实现（edition v5.0 in Ultralytics）支持多GPU训练，新增各个种类目标数量计算，新增heatmap；将正样本匹配过程加入dataloader，加快了运算速度；加入EMA...
Yolov5结构图展示清晰易懂
08-03
Input部分完成数据增强、自适应图片缩放、锚框计算等基本处理任务；Backbone部分作为主干网络，主要使用CSP结构提取出输入样本中的主要信息，以供后续阶段使用；Neck部分使用FPN及PAN结构，利用Backbone部分提取到的...
HTML5+CSS3+JS小实例：自适应瀑布流布局
01-24
html + css + js 原生实现自适应瀑布流布局，可在多种需求下使用，具体样式也可以自行调整，内部代码有详细注释，更多好用的小示例可以关注专栏第一篇！！！
yolov5+改进策略+案例分析+资源合集
05-30
输入端：在模型训练阶段，提出了一些改进思路，主要包括Mosaic数据增强、自适应锚框计算、自适应图片缩放； 基准网络：融合其它检测算法中的一些新思路，主要包括：Focus结构与CSP结构； Neck网络：目标检测网络在...
黑科技：自适应循环航空发动机.mp4
04-29
黑科技：自适应循环航空发动机
自适应：自适应：并行主动学习数学函数
02-04
自适应：自适应：并行主动学习数学函数
RAdam:自适应学习率的方差及超越
04-06
在我们的研究中，我们提出一个根本原因是自适应学习率的巨大差异，并提供理论和经验支持证据。 除了解释为什么要使用预热之外，我们还提出RAdam ，这是Adam的理论上合理的变体。 动机 如图1所示，我们假定梯度遵循...
自适应k均值matlab代码-SMCL:自适应多原型竞争学习
06-07
种子点的数量逐渐增加，直到一个种子点被竞争学习赶走。 用法 以数据集名称作为参数直接运行 smcl.m。 例如： smcl gaussian 输出示例 这些代码输出四个数字和一组数字度量。 命令行输出： >> smcl gaussian SMCL ...
web前端工程师：5分钟学会Web自适应
02-25
特别说明：在开始这一切之前，请开发移动界面的工程师们在头部加上下面这条meta:所谓宽度自适应严格来说是一种PC端的自适应布局方式在移动端的延伸。在处理PC端的前端界面时候需要用到全屏布局时采用的就是此种布局...
【Yolov系列】Yolov5学习（一）补充1.2：自适应锚框计算详解+代码注释
最新发布
qq_43737512的博客
04-19 1395
①获取数据集中所有目标的宽和高。②将每张图片中按照等比例缩放的方式到 resize 指定大小，这里保证宽高中的最大值符合指定大小。③将 bboxes 从相对坐标改成绝对坐标，这里乘以的是缩放后的宽高。④筛选 bboxes，保留宽高都大于等于两个像素的 bboxes。⑤使用三方得到n个 anchors，与YOLOv3、YOLOv4 操作一样。⑥使用随机对 anchors 的宽高进行变异。倘若变异后的效果好，就将变异后的结果赋值给 anchors；
YOLOv5改进实战（导读篇） | YOLOv5 7.0 网络结构超详细解读
w94ghz的博客
09-21 9088
【快速入门】YOLOv5目标检测算法
不积跬步，无以至千里！
03-01 2万+
简单快速入门YOLOv5秘籍！
yolov5自适应锚框计算的优点
05-03
Yolov5的自适应锚框计算具有以下优点： 1. 减少手工调整锚框的工作量：传统的目标检测算法需要手动调整锚框的大小和位置，这个过程比较繁琐。而Yolov5的自适应锚框计算可以自动计算出最优的锚框大小和位置，减少了手动调整的工作量。 2. 提高模型的泛化能力：Yolov5的自适应锚框计算可以根据不同的目标自动调整锚框的大小和位置，使得模型更加适应不同尺寸和形状的目标，提高了模型的泛化能力。 3. 提高目标检测的准确率：Yolov5的自适应锚框计算可以根据目标的大小和形状自动调整锚框的大小和位置，从而提高了目标检测的准确率。 总之，Yolov5的自适应锚框计算可以自动计算出最优的锚框大小和位置，减少了手动调整的工作量，提高了模型的泛化能力和目标检测的准确率。

“相关推荐”对你有帮助么？

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值