本文主要讲述yolov5anchor匹配策略-跨网格预测以及损失函数计算的核心过程理解,网络部分相对容易这里不再赘述。
yolov5最重要的便是跨网格进行预测,从当前网格的上、下、左、右的四个网格中找到离目标中心点最近的两个网格,再加上当前网格共三个网格进行匹配。增大正样本的数量,加快模型收敛。
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
l, m = ((gxy % 1. > (1 - g)) & (gxy < (gain[[2, 3]] - 1.))).T
# a:anchor索引值,包含添加的正样本(让近于当前网格中心点的上下左右中的两个网格充当正样本)信息
# t: gt框信息, 包含添加的正样本
# offsets: 每个框中心点的偏移值
a, t = torch.cat((a, a[j], a[k], a[l], a[m]), 0), torch.cat((t, t[j], t[k], t[l], t[m]), 0)
offsets = torch.cat((z, z[j] + off[0], z[k] + off[1], z[l] + off[2], z[m] + off[3]), 0) * g
代码如下:
pxy = ps[:, :2].sigmoid() * 2. - 0.5
pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
def compute_loss(p, targets, model): # predictions, targets, model
# p 预测结果, list[torch.tensor * 3], p[i] = (b, 3, h, w, nc + 5[x,y,w,h,conf]) --- h,w (80, 80)、(40, 40)、 (20, 20)
ft = torch.cuda.FloatTensor if p[0].is_cuda else torch.Tensor
lcls, lbox, lobj = ft([0]), ft([0]), ft([0])
tcls, tbox, indices, anchors = build_targets(p, targets, model) # 类别,box, 索引, anchors
h = model.hyp # hyperparameters
red = 'mean' # Loss reduction (sum or mean)
# Define criteria, 类别和目标损失
BCEcls = nn.BCEWithLogitsLoss(pos_weight=ft([h['cls_pw']]), reduction=red)
BCEobj = nn.BCEWithLogitsLoss(pos_weight=ft([h['obj_pw']]), reduction=red)
# class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
cp, cn = smooth_BCE(eps=0.0)
# focal loss
g = h['fl_gamma'] # focal loss gamma
if g > 0:
BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
# per output
nt = 0 # targets
for i, pi in enumerate(p): # layer index, layer predictions
b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
tobj = torch.zeros_like(pi[..., 0]) # target obj
nb = b.shape[0] # number of targets
if nb:
nt += nb # cumulative targets
# 取出对应位置预测值
ps = pi[b, a, gj, gi] # prediction subset corresponding to targets
# GIoU
pxy = ps[:, :2].sigmoid() * 2. - 0.5
pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
pbox = torch.cat((pxy, pwh), 1) # predicted box
giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True) # giou(prediction, target)
lbox += (1.0 - giou).sum() if red == 'sum' else (1.0 - giou).mean() # giou loss
# Obj 有物体的conf分支权重
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype) # giou ratio
# Class
if model.nc > 1: # cls loss (only if multiple classes)
t = torch.full_like(ps[:, 5:], cn) # targets
t[range(nb), tcls[i]] = cp
lcls += BCEcls(ps[:, 5:], t) # BCE 每个类单独计算Loss
# Append targets to text file
# with open('targets.txt', 'a') as file:
# [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
lobj += BCEobj(pi[..., 4], tobj) # obj loss
lbox *= h['giou']
lobj *= h['obj']
lcls *= h['cls']
bs = tobj.shape[0] # batch size
if red == 'sum':
g = 3.0 # loss gain
lobj *= g / bs
if nt:
lcls *= g / nt / model.nc
lbox *= g / nt
loss = lbox + lobj + lcls
return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
def build_targets(p, targets, model):
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
# 获取检测层
det = model.module.model[-1] if type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) \
else model.model[-1] # Detect() module
na, nt = det.na, targets.shape[0] # number of anchors, targets
tcls, tbox, indices, anch = [], [], [], []
gain = torch.ones(6, device=targets.device) # normalized to gridspace gain
# 上下左右四个网格
off = torch.tensor([[1, 0], [0, 1], [-1, 0], [0, -1]], device=targets.device).float() # overlap offsets
# anchor索引,表示当前bbox和当前层哪个anchor匹配
at = torch.arange(na).view(na, 1).repeat(1, nt) # anchor tensor, same as .repeat_interleave(nt)
style = 'rect4'
# 处理每个检测层(三个)
for i in range(det.nl):
# 三个anchors,已经除以当前特征图对应的stride
anchors = det.anchors[i]
# 将target中归一化后的xywh映射到三个尺度(80,80, 40,40, 20,20)的输出需要的系数
gain[2:] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
# Match targets to anchors
# t 将标签框的xywh从基于0-1映射到基于特征图,target的xywh本身是归一化尺度,需要变成特征图尺寸
# t GT框 shape:(nt, 6) , 6 = icxywh, i 指第i+1张图片,c为类别
a, t, offsets = [], targets * gain, 0
# 判断是否存在目标
if nt:
# 计算当前tartget的wh和anchor的wh比值
# 如果最大比值大于预设值model.hyp['anchor_t']=4,则当前target和anchor匹配度不高,不强制回归,而把target丢弃
r = t[None, :, 4:6] / anchors[:, None] # wh ratio
# 筛选满足条件1/hyp['anchor_t] < target_wh / anchor_wh < hyp['anchor_t]的框
#.max(2)对第三维度的值进行max
j = torch.max(r, 1. / r).max(2)[0] < model.hyp['anchor_t'] # compare
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n) = wh_iou(anchors(3,2), gwh(n,2))
# t.repeat(na, 1, 1)是的t内的gtbox信息与at索引相对应, repeat(3,1,1)第一维度重复三次
# a 筛选后的gtbox的anchor索引, 筛选后的gtbox信息
a, t = at[j], t.repeat(na, 1, 1)[j] # filter
# overlaps
gxy = t[:, 2:4] # grid xy,label的中心点坐标
z = torch.zeros_like(gxy)
# 取得上下左右四个网格中近于当前网格中心点中的其中两个
"""把相对于各个网格左上角x<0.5,y<0.5和相对于右下角的x<0.5,y<0.5的框提取出来,就是j,k,l,m
在选取gij(标签分配的网格)的时候对这四个部分都做一个偏移(加去上面的off),
"""
if style == 'rect2':
g = 0.2 # offset
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
a, t = torch.cat((a, a[j], a[k]), 0), torch.cat((t, t[j], t[k]), 0)
offsets = torch.cat((z, z[j] + off[0], z[k] + off[1]), 0) * g
elif style == 'rect4':
g = 0.5 # offset
# 对于筛选后的bbox,计算其落在哪个网格内,同时找出邻近的网格,将这些网格都认为是负责预测该bbox的网格
# 浮点数取模的数学定义:对于两个浮点数a和b: a / b = a - n * b , 其中n为不超过 a/b的最大整数
# gxy > 1.和gxy < (gain[[2, 3]] - 1.)判断中心点是否落在四个角,如果落在四个角,就不取界外的框(因为超出了边界)
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
l, m = ((gxy % 1. > (1 - g)) & (gxy < (gain[[2, 3]] - 1.))).T
# a:anchor索引值,包含添加的正样本(让近于当前网格中心点的上下左右中的两个网格充当正样本)信息
# t: gt框信息, 包含添加的正样本
# offsets: 每个框中心点的偏移值
a, t = torch.cat((a, a[j], a[k], a[l], a[m]), 0), torch.cat((t, t[j], t[k], t[l], t[m]), 0)
offsets = torch.cat((z, z[j] + off[0], z[k] + off[1], z[l] + off[2], z[m] + off[3]), 0) * g
# Define
"""
对每个bbox找出对应的正样本anchor。
a 表示当前bbox和当前层的第几个anchor匹配
b 表示当前bbox属于batch内部的第几张图片,
c 是该bbox的类别
gi,gj 是对应的负责预测该bbox的网格坐标
gxy 负责预测网格中心点坐标xy
gwh 是对应的bbox的wh
"""
b, c = t[:, :2].long().T # image, class
gxy = t[:, 2:4] # grid xy
gwh = t[:, 4:6] # grid wh
# 当前label落在哪个网格上
gij = (gxy - offsets).long()
gi, gj = gij.T # grid xy indices
# Append 添加索引,方便计算损失的时候去除对应位置的输出
indices.append((b, a, gj, gi)) # image, anchor, grid indices
# gxy - gij: 预测中心点坐标相对于gi,gj的偏移值
tbox.append(torch.cat((gxy - gij, gwh), 1)) # box
anch.append(anchors[a]) # anchors,使用哪个anchor尺寸
tcls.append(c) # class
return tcls, tbox, indices, anch
yolov5增加正样本的方法,最多可增大到原来的三倍,大大增加了正样本的数量,加速了模型的收敛。
目标检测重中之重可以理解为anchor的匹配策略,当下流行的anchor-free不过换了一种匹配策略罢了。
我想当下真正可创新之处在于更优的匹配策略。
鄙人拙见,请不吝指教。