[目标检测] 如何获取数据集对应的anchor size

背景：由于yolov5需要根据size生成候选anchor，因此用yolov5训练某个特定的数据集时，如果没有设置对应的anchor size。那么可能出现很多乱框的现象。

可以通过.pt文件就能读取到anchor size的参数。输入是.pt文件，输出是

const int anchor0[6] = {3, 5, 6, 5, 5, 11};
const int anchor1[6] = {10, 9, 10, 17, 21, 11};
const int anchor2[6] = {20, 25, 39, 21, 47, 47};

import torch
from models.experimental import attempt_load

model = attempt_load('/root/yolov5-master/runs/train/exp8/weights/best.pt')
m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1]

# Convert to CPU for easier manipulation
for i, grid in enumerate(m.anchor_grid):
    grid = grid.cpu()  # Move to CPU if on GPU
    anchor_sets = []

    # Traverse the second dimension (3 anchor sets)
    for j in range(3):  # Since the second dimension is of size 3
        # Extract the first two 'anchors' from each set (assuming we only want the first two for each set)
        anchor_pair = grid[0, j, 0, 0].tolist()  # [0, 0] for spatial dimensions, assuming we want the first anchor
        anchor_sets.append([round(x) for x in anchor_pair])

    # Flatten and format for C-style declaration
    flat_anchors = [item for pair in anchor_sets for item in pair]
    print(f"const int anchor{i}[6] = {{{', '.join(map(str, flat_anchors))}}};")