参考：
    https://github.com/lars76/kmeans-anchor-boxes

    1. #coding=utf-8
    2. import xml.etree.ElementTree as ET
    3. import numpy as np
    4. import glob
    6. def iou(box, clusters):
    7.     """
    8.     计算一个 ground truth 边界盒和 k 个先验框(Anchor)的交并比(IOU)值。
    9.     参数box: 元组或者数据，代表 ground truth 的长宽。
    10.     参数clusters: 形如(k,2)的numpy数组，其中k是聚类Anchor框的个数
    11.     返回：ground truth和每个Anchor框的交并比。
    12.     """
    13.     x = np.minimum(clusters[:, 0], box[0])
    14.     y = np.minimum(clusters[:, 1], box[1])
    15.     if np.count_nonzero(x == 0) > 0 or np.count_nonzero(y == 0) > 0:
    16.         raise ValueError("Box has no area")
    17.     intersection = x * y
    18.     box_area = box[0] * box[1]
    19.     cluster_area = clusters[:, 0] * clusters[:, 1]
    20.     iou_ = intersection / (box_area + cluster_area - intersection)
    21.     return iou_
    24. def avg_iou(boxes, clusters):
    25.     """
    26.     计算一个ground truth和k个Anchor的交并比的均值。
    27.     """
    28.     return np.mean([np.max(iou(boxes[i], clusters)) for i in range(boxes.shape[0])])
    31. def kmeans(boxes, k, dist=np.median):
    32.     """
    33.     利用IOU值进行K-means聚类
    34.     参数boxes: 形状为(r, 2)的ground truth框，其中r是ground truth的个数
    35.     参数k: Anchor的个数
    36.     参数dist: 距离函数
    37.     返回值：形状为(k, 2)的k个Anchor框
    38.     """
    39.     # 即是上面提到的r
    40.     rows = boxes.shape[0]
    41.     # 距离数组，计算每个ground truth和k个Anchor的距离
    42.     distances = np.empty((rows, k))
    43.     # 上一次每个ground truth"距离"最近的Anchor索引
    44.     last_clusters = np.zeros((rows,))
    45.     # 设置随机数种子
    46.     np.random.seed()
    48.     # 初始化聚类中心，k个簇，从r个ground truth随机选k个
    49.     clusters = boxes[np.random.choice(rows, k, replace=False)]
    50.     # 开始聚类
    51.     while True:
    52.         # 计算每个ground truth和k个Anchor的距离，用1-IOU(box,anchor)来计算
    53.         for row in range(rows):
    54.             distances[row] = 1 - iou(boxes[row], clusters)
    55.         # 对每个ground truth，选取距离最小的那个Anchor，并存下索引
    56.         nearest_clusters = np.argmin(distances, axis=1)
    57.         # 如果当前每个ground truth"距离"最近的Anchor索引和上一次一样，聚类结束
    58.         if (last_clusters == nearest_clusters).all():
    59.             break
    60.         # 更新簇中心为簇里面所有的ground truth框的均值
    61.         for cluster in range(k):
    62.             clusters[cluster] = dist(boxes[nearest_clusters == cluster], axis=0)
    63.         # 更新每个ground truth"距离"最近的Anchor索引
    64.         last_clusters = nearest_clusters
    66.     return clusters
    70. # 加载自己的数据集，只需要所有 labelimg 标注出来的 xml 文件即可
    71. def load_dataset(path):
    72.     dataset = []
    73.     for xml_file in glob.glob("{}/*xml".format(path)):
    74.         tree = ET.parse(xml_file)
    75.         # 图片高度
    76.         height = int(tree.findtext("./size/height"))
    77.         # 图片宽度
    78.         width = int(tree.findtext("./size/width"))
    80.         for obj in tree.iter("object"):
    81.             # 偏移量
    82.             xmin = int(obj.findtext("bndbox/xmin")) / width
    83.             ymin = int(obj.findtext("bndbox/ymin")) / height
    84.             xmax = int(obj.findtext("bndbox/xmax")) / width
    85.             ymax = int(obj.findtext("bndbox/ymax")) / height
    86.             xmin = np.float64(xmin)
    87.             ymin = np.float64(ymin)
    88.             xmax = np.float64(xmax)
    89.             ymax = np.float64(ymax)
    90.             if xmax == xmin or ymax == ymin:
    91.                 print(xml_file)
    92.             # 将Anchor的宽和高放入dateset，运行kmeans获得Anchor
    93.             dataset.append([xmax - xmin, ymax - ymin])
    94.     return np.array(dataset)
    97. if __name__ == '__main__':
    99.     ANNOTATIONS_PATH = "./label_source" #xml文件所在文件夹
    100.     CLUSTERS = 9 #聚类数量，anchor数量
    101.     INPUTDIM = 416 #输入网络大小
    103.     data = load_dataset(ANNOTATIONS_PATH)
    104.     out = kmeans(data, k=CLUSTERS)
    105.     print('Boxes:')
    106.     print(np.array(out)*INPUTDIM)
    107.     print("Accuracy: {:.2f}%".format(avg_iou(data, out) * 100))
    108.     final_anchors = np.around(out[:, 0] / out[:, 1], decimals=2).tolist()
    109.     print("Before Sort Ratios:\n {}".format(final_anchors))
    110.     print("After Sort Ratios:\n {}".format(sorted(final_anchors)))

    得到的结果比如

    1. Boxes:
    2. [[131.456      126.464     ]
    3.  [159.744      293.57357357]
    4.  [ 16.64        28.84266667]
    5.  [ 26.624       77.65333333]
    6.  [ 48.256       44.37333333]
    7.  [ 59.072      106.496     ]
    8.  [266.98178313 187.43236036]
    9.  [ 82.368      212.16      ]
    10.  [346.112      348.33066667]]
    11. Accuracy: 67.29%
    12. Before Sort Ratios:
    13.  [1.04, 0.54, 0.58, 0.34, 1.09, 0.55, 1.42, 0.39, 0.99]
    14. After Sort Ratios:
    15.  [0.34, 0.39, 0.54, 0.55, 0.58, 0.99, 1.04, 1.09, 1.42]

    After Sort Ratios 从小到大分别是浅层，中层，高层特征图，然后对应上面原始的 Before Sort Ratios
    ，接着去找对应的 Boxes