devbox
你好赵伟
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

yolov5/7 anchors设置详解_yolov5的锚定框参数什么含义

作者：你好赵伟 | 2024-05-16 13:56:05

赞

踩

yolov5的锚定框参数什么含义

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）：

  1. # anchors
  2. anchors:
  3.   - [10,13, 16,30, 33,23]  # P3/8
  4.   - [30,61, 62,45, 59,119]  # P4/16
  5.   - [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：

  1.     def metric(k):  # compute metric
  2.         r = wh[:, None] / k[None]
  3.         x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
  4.         best = x.max(1)[0]  # best_x
  5.         aat = (x > 1. / thr).float().sum(1).mean()  # anchors above threshold
  6.         bpr = (best > 1. / thr).float().mean()  # best possible recall
  7.         return bpr, aat
  8.  
  9.     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）：

即bpr（best possible recall）和 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：自己找找。

其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）：

1.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）
2.n：锚定框的数量，即有几组；默认值是9
3.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
4.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。
5.gen：kmean聚类算法迭代次数，默认值是1000
6.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

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

2、训练前手动计算锚定框

重点在于如果使用 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】。在这可以自己在labels.jpg中查看数据集目标物的宽高大小，超过4：1时即可自定义anchors大小，因为默认coco数据集宽高不一定适合你的数据集标注的大小。

 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：

  1. import utils.autoanchor as autoAC
  2.  
  3. # 对数据集重新计算 anchors
  4. new_anchors = autoAC.kmean_anchors('./data/mydata.yaml', 9, 640, 5.0, 1000, True)
  5. print(new_anchors)

问题1：你可能会出现报错，例如下：

  1. RuntimeError:
  2. An attempt has been made to start a new process before the
  3. current process has finished its bootstrapping phase.
  4. This probably means that you are not using fork to start your
  5. child processes and you have forgotten to use the proper idiom
  6. in the main module:
  7. 　　if name == ‘main‘:
  8. 　　　　freeze_support()
  9. 　　　　…
  10. The “freeze_support()” line can be omitted if the program
  11. is not going to be frozen to produce an executable.

错误原因：

Python 解释器在 Windows平台 执行创建多进程的程序时，子进程会读取当前 Python 文件，用以创建进程。
在子进程读取当前文件时，读取到创建子进程的代码时又会创建新的子进程，这样程序就陷入递归创建进程的状态

解决办法：

按照提示将程序创建子进程的放进 if __name__ == '__main__': 语句内，该语句的作用是判断当前进程是否为主进程，是主进程才执行程序。

  1. import utils.autoanchor as autoAC
  2. if __name__ == '__main__':
  3. # 对数据集重新计算 anchors
  4.    new_anchors = autoAC.kmean_anchors('./data/pk.yaml', 9, 640, 5.0, 1000, True)
  5.    print(new_anchors)

运行脚本打印平均anchors。

  1. Scanning C:\Users\49626\Desktop\yolov5-master\pk\labels\train... 588 images, 8 backgrounds, 0 corrupt: 100%|██████████| 591/591 [00:08<00:00, 65.75it/s]
  2. WARNING  Cache directory C:\Users\49626\Desktop\yolov5-master\pk\labels is not writeable: [WinError 183] : 'C:\\Users\\49626\\Desktop\\yolov5-master\\pk\\labels\\train.cache.npy' -> 'C:\\Users\\49626\\Desktop\\yolov5-master\\pk\\labels\\train.cache'
  3. AutoAnchor: WARNING  Extremely small objects found: 5 of 1195 labels are <3 pixels in size
  4. AutoAnchor: Running kmeans for 9 anchors on 1195 points...
  5.   0%|          | 0/1000 [00:00<?, ?it/s]AutoAnchor: thr=0.20: 0.9808 best possible recall, 5.49 anchors past thr
  6. AutoAnchor: n=9, img_size=640, metric_all=0.317/0.708-mean/best, past_thr=0.455-mean: 19,13, 50,25, 87,41, 168,30, 152,74, 260,48, 269,106, 328,158, 510,220
  7. AutoAnchor: thr=0.20: 0.9808 best possible recall, 5.46 anchors past thr
  8. AutoAnchor: n=9, img_size=640, metric_all=0.316/0.709-mean/best, past_thr=0.455-mean: 19,13, 49,24, 86,41, 171,30, 155,74, 270,49, 269,104, 326,162, 512,215
  9. AutoAnchor: thr=0.20: 0.9841 best possible recall, 5.47 anchors past thr
  10. AutoAnchor: n=9, img_size=640, metric_all=0.316/0.710-mean/best, past_thr=0.453-mean: 18,12, 49,23, 86,43, 180,31, 256,44, 151,75, 267,98, 326,153, 505,215
  11. AutoAnchor: thr=0.20: 0.9841 best possible recall, 5.46 anchors past thr
  12. AutoAnchor: n=9, img_size=640, metric_all=0.315/0.710-mean/best, past_thr=0.453-mean: 18,12, 50,22, 87,43, 180,31, 151,75, 263,44, 262,95, 333,156, 535,207
  13. AutoAnchor: thr=0.20: 0.9841 best possible recall, 5.46 anchors past thr
  14. AutoAnchor: n=9, img_size=640, metric_all=0.315/0.710-mean/best, past_thr=0.454-mean: 18,11, 51,22, 92,42, 185,30, 151,74, 266,45, 260,93, 339,156, 539,207
  15. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.48 anchors past thr
  16. AutoAnchor: n=9, img_size=640, metric_all=0.316/0.711-mean/best, past_thr=0.454-mean: 17,11, 52,22, 91,41, 175,29, 151,75, 257,47, 265,92, 331,158, 539,207
  17. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.46 anchors past thr
  18. AutoAnchor: n=9, img_size=640, metric_all=0.316/0.712-mean/best, past_thr=0.455-mean: 17,11, 51,23, 91,41, 177,30, 153,74, 257,47, 267,92, 334,158, 547,206
  19. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.45 anchors past thr
  20. AutoAnchor: n=9, img_size=640, metric_all=0.315/0.712-mean/best, past_thr=0.454-mean: 17,11, 47,23, 86,40, 186,30, 145,74, 254,45, 260,88, 334,149, 517,223
  21. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.44 anchors past thr
  22. AutoAnchor: n=9, img_size=640, metric_all=0.315/0.713-mean/best, past_thr=0.453-mean: 17,11, 47,23, 85,40, 196,30, 246,45, 147,75, 263,87, 337,150, 512,222
  23. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.45 anchors past thr
  24. AutoAnchor: n=9, img_size=640, metric_all=0.314/0.713-mean/best, past_thr=0.453-mean: 17,11, 48,23, 86,40, 198,29, 246,45, 147,75, 263,88, 337,150, 510,222
  25. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.45 anchors past thr
  26. AutoAnchor: n=9, img_size=640, metric_all=0.314/0.713-mean/best, past_thr=0.453-mean: 17,11, 48,23, 85,41, 197,29, 149,72, 248,45, 269,88, 337,150, 495,227
  27. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.46 anchors past thr
  28. AutoAnchor: n=9, img_size=640, metric_all=0.315/0.714-mean/best, past_thr=0.453-mean: 17,12, 48,22, 87,39, 202,29, 149,72, 248,45, 260,88, 324,148, 489,224
  29. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.46 anchors past thr
  30. AutoAnchor: n=9, img_size=640, metric_all=0.315/0.714-mean/best, past_thr=0.453-mean: 17,12, 48,22, 87,39, 202,29, 149,72, 242,45, 264,89, 323,150, 489,226
  31. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.45 anchors past thr
  32. AutoAnchor: n=9, img_size=640, metric_all=0.315/0.714-mean/best, past_thr=0.453-mean: 17,12, 48,22, 86,40, 203,29, 149,71, 243,45, 264,90, 323,155, 484,223
  33. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.46 anchors past thr
  34. AutoAnchor: n=9, img_size=640, metric_all=0.316/0.714-mean/best, past_thr=0.454-mean: 17,12, 48,22, 86,39, 200,30, 151,72, 242,47, 260,90, 329,153, 468,230
  35. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.46 anchors past thr
  36. AutoAnchor: n=9, img_size=640, metric_all=0.316/0.715-mean/best, past_thr=0.455-mean: 17,12, 48,22, 89,38, 199,29, 233,47, 152,72, 263,91, 325,153, 491,225
  37. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.46 anchors past thr
  38. AutoAnchor: n=9, img_size=640, metric_all=0.316/0.715-mean/best, past_thr=0.455-mean: 16,12, 48,22, 89,38, 199,29, 232,47, 153,72, 262,91, 324,152, 493,225
  39. AutoAnchor: thr=0.20: 0.9900 best possible recall, 5.48 anchors past thr
  40. AutoAnchor: n=9, img_size=640, metric_all=0.317/0.715-mean/best, past_thr=0.455-mean: 17,12, 48,22, 87,39, 201,29, 153,71, 232,47, 261,92, 323,151, 495,223
  41. AutoAnchor: thr=0.20: 0.9958 best possible recall, 5.41 anchors past thr
  42. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.714-mean/best, past_thr=0.454-mean: 15,12, 46,21, 89,37, 213,29, 159,72, 238,51, 261,92, 323,151, 510,216
  43. AutoAnchor: thr=0.20: 0.9958 best possible recall, 5.42 anchors past thr
  44. AutoAnchor: n=9, img_size=640, metric_all=0.314/0.715-mean/best, past_thr=0.455-mean: 15,13, 45,21, 85,38, 210,29, 166,70, 234,51, 261,86, 321,151, 518,215
  45. AutoAnchor: thr=0.20: 0.9958 best possible recall, 5.42 anchors past thr
  46. AutoAnchor: n=9, img_size=640, metric_all=0.315/0.715-mean/best, past_thr=0.455-mean: 15,13, 45,21, 85,38, 207,30, 166,71, 235,50, 263,85, 315,150, 518,212
  47. AutoAnchor: thr=0.20: 0.9983 best possible recall, 5.35 anchors past thr
  48. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.714-mean/best, past_thr=0.455-mean: 13,13, 48,21, 84,38, 189,28, 261,50, 178,75, 257,75, 310,154, 518,211
  49. AutoAnchor: thr=0.20: 0.9992 best possible recall, 5.32 anchors past thr
  50. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.715-mean/best, past_thr=0.456-mean: 11,12, 47,21, 88,35, 184,26, 163,66, 261,48, 254,74, 313,153, 547,211
  51. AutoAnchor: thr=0.20: 0.9992 best possible recall, 5.32 anchors past thr
  52. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.715-mean/best, past_thr=0.457-mean: 11,12, 47,21, 88,35, 182,26, 163,66, 261,48, 254,74, 315,153, 551,210
  53. AutoAnchor: thr=0.20: 0.9992 best possible recall, 5.32 anchors past thr
  54. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.715-mean/best, past_thr=0.457-mean: 11,12, 47,21, 88,35, 182,26, 162,66, 262,48, 255,74, 316,153, 551,210
  55. AutoAnchor: thr=0.20: 0.9992 best possible recall, 5.33 anchors past thr
  56. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.716-mean/best, past_thr=0.457-mean: 11,12, 47,21, 88,35, 184,26, 162,66, 258,48, 254,75, 318,150, 552,208
  57. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.35 anchors past thr
  58. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.717-mean/best, past_thr=0.456-mean: 11,12, 48,20, 88,36, 184,26, 161,64, 258,47, 247,74, 317,150, 552,209
  59. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.34 anchors past thr
  60. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.717-mean/best, past_thr=0.457-mean: 11,12, 47,20, 88,36, 185,26, 161,64, 257,47, 247,74, 318,149, 552,209
  61. AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7172:  16%|█▋        | 165/1000 [00:00<00:00, 1640.33it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.31 anchors past thr
  62. AutoAnchor: n=9, img_size=640, metric_all=0.309/0.719-mean/best, past_thr=0.454-mean: 10,14, 46,19, 76,42, 177,26, 160,71, 256,47, 240,76, 313,153, 571,201
  63. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.32 anchors past thr
  64. AutoAnchor: n=9, img_size=640, metric_all=0.308/0.720-mean/best, past_thr=0.452-mean: 10,14, 50,20, 78,42, 183,26, 147,75, 275,44, 249,76, 325,145, 539,201
  65. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.31 anchors past thr
  66. AutoAnchor: n=9, img_size=640, metric_all=0.308/0.721-mean/best, past_thr=0.453-mean: 10,14, 50,20, 78,41, 184,26, 147,75, 278,44, 250,76, 327,144, 545,203
  67. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.28 anchors past thr
  68. AutoAnchor: n=9, img_size=640, metric_all=0.308/0.724-mean/best, past_thr=0.456-mean: 9,14, 44,20, 76,41, 174,25, 157,64, 282,47, 250,79, 298,144, 548,210
  69. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.28 anchors past thr
  70. AutoAnchor: n=9, img_size=640, metric_all=0.309/0.726-mean/best, past_thr=0.457-mean: 9,14, 44,19, 79,41, 171,24, 157,65, 280,48, 246,81, 302,140, 565,205
  71. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.29 anchors past thr
  72. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.731-mean/best, past_thr=0.459-mean: 9,15, 46,19, 79,39, 171,23, 157,64, 272,48, 252,82, 276,137, 525,209
  73. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.30 anchors past thr
  74. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.731-mean/best, past_thr=0.459-mean: 9,15, 47,19, 81,40, 171,24, 154,63, 269,48, 252,83, 282,135, 537,206
  75. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.27 anchors past thr
  76. AutoAnchor: n=9, img_size=640, metric_all=0.309/0.732-mean/best, past_thr=0.458-mean: 8,15, 45,18, 81,39, 179,24, 155,64, 269,50, 259,82, 276,140, 550,202
  77. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.30 anchors past thr
  78. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.732-mean/best, past_thr=0.460-mean: 8,16, 45,17, 84,36, 181,25, 143,65, 272,50, 258,76, 269,129, 551,208
  79. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.32 anchors past thr
  80. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.732-mean/best, past_thr=0.460-mean: 8,16, 46,18, 82,36, 179,25, 142,63, 262,51, 261,74, 270,129, 545,214
  81. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  82. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.733-mean/best, past_thr=0.460-mean: 9,17, 45,18, 81,36, 188,25, 143,63, 245,51, 262,74, 270,129, 516,215
  83. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  84. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.733-mean/best, past_thr=0.460-mean: 9,17, 45,18, 81,37, 189,25, 145,64, 244,50, 262,74, 267,130, 510,215
  85. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  86. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.733-mean/best, past_thr=0.460-mean: 9,17, 45,18, 81,37, 188,25, 144,64, 244,50, 262,74, 266,129, 511,215
  87. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  88. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.734-mean/best, past_thr=0.460-mean: 9,17, 45,18, 81,37, 187,25, 145,65, 245,50, 262,74, 266,129, 511,214
  89. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  90. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.734-mean/best, past_thr=0.460-mean: 9,17, 45,18, 81,37, 187,25, 145,65, 245,50, 262,74, 266,129, 511,214
  91. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  92. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.734-mean/best, past_thr=0.460-mean: 9,17, 45,18, 81,37, 188,25, 145,64, 245,50, 261,75, 265,128, 511,214
  93. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.32 anchors past thr
  94. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.734-mean/best, past_thr=0.459-mean: 8,16, 45,17, 80,36, 181,26, 142,68, 251,49, 260,76, 274,133, 526,204
  95. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  96. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.735-mean/best, past_thr=0.457-mean: 8,16, 43,17, 78,36, 188,27, 142,72, 241,43, 240,85, 293,138, 511,198
  97. AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7348:  38%|███▊      | 376/1000 [00:00<00:00, 1908.21it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.34 anchors past thr
  98. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.735-mean/best, past_thr=0.457-mean: 8,16, 43,17, 78,36, 188,27, 236,43, 142,72, 238,85, 293,138, 509,197
  99. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.34 anchors past thr
  100. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.735-mean/best, past_thr=0.458-mean: 8,16, 43,17, 78,36, 188,27, 142,72, 236,43, 237,85, 293,139, 508,196
  101. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  102. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.735-mean/best, past_thr=0.458-mean: 8,16, 43,17, 78,36, 188,28, 143,72, 237,44, 240,85, 295,137, 508,197
  103. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  104. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.735-mean/best, past_thr=0.457-mean: 8,16, 43,17, 78,36, 187,27, 236,43, 143,73, 242,85, 295,139, 505,197
  105. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  106. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.735-mean/best, past_thr=0.457-mean: 8,16, 43,17, 78,37, 187,27, 143,73, 237,44, 247,84, 295,139, 500,198
  107. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  108. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.735-mean/best, past_thr=0.457-mean: 8,16, 43,17, 78,36, 187,27, 142,73, 238,44, 248,84, 295,138, 499,198
  109. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.31 anchors past thr
  110. AutoAnchor: n=9, img_size=640, metric_all=0.311/0.736-mean/best, past_thr=0.459-mean: 8,15, 45,18, 81,37, 190,27, 232,43, 149,69, 267,86, 288,137, 471,203
  111. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.31 anchors past thr
  112. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.736-mean/best, past_thr=0.460-mean: 8,15, 45,18, 79,37, 190,28, 231,43, 152,69, 264,86, 289,138, 469,203
  113. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.32 anchors past thr
  114. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.736-mean/best, past_thr=0.459-mean: 8,15, 44,18, 79,37, 190,28, 231,43, 153,70, 262,86, 292,138, 466,203
  115. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.32 anchors past thr
  116. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.736-mean/best, past_thr=0.460-mean: 8,15, 45,18, 80,37, 189,28, 226,44, 153,70, 262,86, 291,139, 470,208
  117. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.32 anchors past thr
  118. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.736-mean/best, past_thr=0.460-mean: 8,15, 44,18, 79,36, 190,28, 224,43, 152,70, 264,86, 288,138, 467,208
  119. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  120. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.736-mean/best, past_thr=0.460-mean: 8,15, 44,18, 79,36, 192,27, 225,44, 152,69, 247,84, 289,138, 475,217
  121. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.32 anchors past thr
  122. AutoAnchor: n=9, img_size=640, metric_all=0.312/0.736-mean/best, past_thr=0.460-mean: 8,15, 44,18, 79,36, 192,27, 225,44, 152,69, 248,84, 289,138, 476,216
  123. AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7362:  59%|█████▉    | 594/1000 [00:00<00:00, 2025.15it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  124. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.736-mean/best, past_thr=0.460-mean: 8,15, 44,18, 79,36, 190,27, 224,44, 152,68, 247,84, 289,137, 473,215
  125. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  126. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.736-mean/best, past_thr=0.460-mean: 8,15, 44,18, 80,37, 189,27, 223,44, 152,68, 247,84, 288,137, 473,216
  127. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  128. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.736-mean/best, past_thr=0.461-mean: 8,15, 44,18, 80,37, 188,27, 224,43, 153,67, 247,83, 287,137, 475,216
  129. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  130. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.736-mean/best, past_thr=0.461-mean: 8,15, 44,18, 80,37, 188,27, 224,43, 153,67, 247,83, 287,137, 475,216
  131. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  132. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.736-mean/best, past_thr=0.461-mean: 8,15, 44,18, 80,37, 188,27, 224,43, 153,67, 247,83, 286,137, 475,216
  133. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  134. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.736-mean/best, past_thr=0.460-mean: 8,15, 44,18, 80,37, 188,27, 225,43, 153,67, 248,83, 286,137, 473,215
  135. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  136. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.736-mean/best, past_thr=0.460-mean: 8,15, 44,18, 80,37, 188,27, 224,43, 152,68, 248,83, 288,136, 476,215
  137. AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7364:  84%|████████▍ | 842/1000 [00:00<00:00, 2203.49it/s]AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.33 anchors past thr
  138. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.736-mean/best, past_thr=0.461-mean: 8,15, 44,18, 80,37, 188,27, 224,43, 153,68, 249,83, 288,136, 476,215
  139. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.34 anchors past thr
  140. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.737-mean/best, past_thr=0.461-mean: 8,15, 44,18, 80,37, 187,27, 221,43, 154,67, 249,83, 285,136, 480,213
  141. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.34 anchors past thr
  142. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.737-mean/best, past_thr=0.461-mean: 8,15, 44,18, 80,37, 186,27, 221,43, 154,67, 249,83, 285,136, 481,213
  143. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.34 anchors past thr
  144. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.737-mean/best, past_thr=0.460-mean: 8,15, 44,18, 81,38, 187,26, 219,43, 152,67, 247,84, 282,136, 498,212
  145. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.34 anchors past thr
  146. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.737-mean/best, past_thr=0.460-mean: 8,15, 43,18, 80,38, 190,26, 219,43, 152,68, 247,83, 283,132, 497,210
  147. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.34 anchors past thr
  148. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.737-mean/best, past_thr=0.460-mean: 8,15, 43,18, 81,38, 191,26, 221,43, 150,68, 246,82, 284,132, 492,212
  149. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.35 anchors past thr
  150. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.737-mean/best, past_thr=0.460-mean: 8,15, 43,18, 81,38, 191,26, 221,43, 149,67, 247,82, 284,131, 486,211
  151. AutoAnchor: Evolving anchors with Genetic Algorithm: fitness = 0.7370: 100%|██████████| 1000/1000 [00:00<00:00, 2132.71it/s]
  152. AutoAnchor: thr=0.20: 1.0000 best possible recall, 5.35 anchors past thr
  153. AutoAnchor: n=9, img_size=640, metric_all=0.313/0.737-mean/best, past_thr=0.460-mean: 8,15, 43,18, 81,38, 191,26, 221,43, 149,67, 247,82, 284,131, 486,211
  154. [[     8.0319      15.323]
  155.  [     42.802      17.647]
  156.  [     81.306       37.98]
  157.  [     191.47      26.115]
  158.  [     220.98      42.882]
  159.  [     149.14      66.916]
  160.  [     247.44      82.486]
  161.  [     283.74      131.36]
  162.  [     485.51      211.37]]

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。

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/你好赵伟/article/detail/579094
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号