yolov5 自适应 anchor
yolov5 自适应 anchor
anchor（锚框）
在 yolos.yaml 文件中
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
之前在 yolov5s.yaml 增加一个检测头和 anchor , 对anchor的应用理解并不是很深刻
只知道，为了对小目标适应力，增加了一个小目标的 anchors [5,6,8,15,15,11]
anchors参数里面共9个数值，一共三行9列，每一行代表在不同的特征图上，
比如第一行是在最大的特征图上的锚框，
第二行是在中间的特征图上的锚框，
第三行是在最小的特征图上的锚框，
每一行后面的注释是 特征图的大小， 如P3/8:当前特征图的大小为P3/8
对于目标检测的任务来说，我们一般希望在更大的特征图上去检测小目标，这样的大特征图才含有更多的小目标信息。
事实上 不同的目标检测算法的锚框设置有很大区别, 不同的数据集标注所对的锚框也有区别
设置自动计算
设置位置：
trian.py
parser.add_argument('--noautoanchor',default='True', action='store_true', help='disable autoanchor check')
trian.py
if not opt.noautoanchor:
check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
源代码
检测 anchors参数 是否符合
yolov5-master\utils\autoanchor.py
def check_anchors(dataset, model, thr=4.0, imgsz=640):
"""
参数dataset代表的是训练集，
hyp['anchor_t']这个超参数是一个界定anchor与label匹配程度的阈值，
imgsz自然就是网络输入尺寸
"""
# Check anchor fit to data, recompute if necessary
prefix = colorstr('autoanchor: ')
print(f'\n{prefix}Analyzing anchors... ', end='')
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
anchors = m.anchor_grid.clone().cpu().view(-1, 2) # current anchors
bpr, aat = metric(anchors)
print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='')
if bpr < 0.98: # threshold to recompute
print('. Attempting to improve anchors, please wait...')
na = m.anchor_grid.numel() // 2 # number of anchors
try:
anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
except Exception as e:
print(f'{prefix}ERROR: {e}')
new_bpr = metric(anchors)[0]
if new_bpr > bpr: # replace anchors
anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid) # for inference
m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1) # loss
check_anchor_order(m)
print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.')
else:
print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.')
print('') # newline
yolov5模型中 只有当bpr小于0.98时才会重新计算anchor
具体计算bpr(best possible recall)的时候，会考虑这9类anchor的宽高和gt框的宽高之间的差距
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
变量wh用来存储训练数据中所有gt框的宽高，是一个shape为(N,2)的tensor，N为gt框的总的个数
输入k是一个shape为 (9,2) 表示 anchors的宽和高两个维度
通过anchor和wh来计算bpr,aat(anchors above threshold) 两个指标，来判断是否满足
- bpr(best possible recall): 最多能被召回的gt框数量 / 所有gt框数量 最大值为1 越大越好 小于0.98就需要使用k-means + 遗传进化算法选择出与数据集更匹配的anchors框。
wh[:, None]的shape是[N, 2]变[N, 1, 2]
k[None]的shape是[9,2]变[1, 9, 2]
x = torch.min(r, 1. / r).min(2)[0]
无论r大于1，还是小于等于1最后统一结果都要小于1
x的shape[N,9]
best的shape是[N]
重新计算
重新计算新的 anchors 框参数
使用 kmeans 对自定义数据进行分析，重新计算 anchors ，获得适合自定义数据集中对象边界框预测的预设锚定框。
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， 长宽比阈值
gen: generations to evolve anchors using genetic algorithm， 产生 anchors 迭代次数
verbose: print all results
Return:
k: kmeans evolved anchors
返回聚类后的数据结果
Usage:
from utils.autoanchor import *; _ = kmean_anchors()
"""
from scipy.cluster.vq import kmeans
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(dataset, str): # *.yaml file
with open(dataset, errors='ignore') as f:
data_dict = yaml.safe_load(f) # model dict
from utils.datasets 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:
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
assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}')
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) * 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)
return print_results(k)
主要思想是 将图像转化到resize640大小的长宽比下，再进行kmeans 聚类，得到9个聚类中心，每个聚类中的包含(x,y)坐标就是需要的 anchor
主要流程
- 载入数据集，得到数据集中所有数据的wh
- 将每张图片中wh的最大值等比例缩放到指定大小img_size，较小边也相应缩放
- 将bboxes从相对坐标改成绝对坐标（乘以缩放后的wh）
- 筛选bboxes，保留wh都大于等于两个像素的bboxes
- 使用k-means聚类得到n个anchors（掉k-means包 涉及一个白化操作）
- 使用遗传算法随机对anchors的wh进行变异，如果变异后效果变得更好（使用anchor_fitness方法计算得到的fitness（适应度）进行评估）就将变异后的结果赋值给anchors，如果变异后效果变差就跳过，默认变异1000次
参考博客：1、https://blog.csdn.net/aabbcccddd01/article/details/109578614
参考博客：2、(推荐) https://blog.csdn.net/qq_38253797/article/details/119713706?utm_medium=distribute.pc_relevant.none-task-blog-2~default~baidujs_title~default-1.no_search_link&spm=1001.2101.3001.4242
