前一篇博客已经从源码yolo.py介绍了Yolov5的网络结构,这篇文章将从loss.py介绍yolov5的损失函数。
源码的版本是tagV5.0。
代码和注释如下:
def __init__(self, model, autobalance=False):
super(ComputeLoss, self).__init__()
device = next(model.parameters()).device # get model device
h = model.hyp # hyperparameters
'''
定义分类损失和置信度损失为带sigmoid的二值交叉熵损失,
即会先将输入进行sigmoid再计算BinaryCrossEntropyLoss(BCELoss)。
pos_weight参数是正样本损失的权重参数。
'''
# Define criteria
BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
'''
对标签做平滑,eps=0就代表不做标签平滑,那么默认cp=1,cn=0
后续对正类别赋值cp,负类别赋值cn
'''
# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
'''
超参设置g>0则计算FocalLoss
'''
# Focal loss
g = h['fl_gamma'] # focal loss gamma
if g > 0:
BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
'''
获取detect层
'''
det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module
'''
每一层预测值所占的权重比,分别代表浅层到深层,小特征到大特征,4.0对应着P3,1.0对应P4,0.4对应P5。
如果是自己设置的输出不是3层,则返回[4.0, 1.0, 0.25, 0.06, .02],可对应1-5个输出层P3-P7的情况。
'''
self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7
'''
autobalance 默认为 False,yolov5中目前也没有使用 ssi = 0即可
'''
self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index
'''
赋值各种参数,gr是用来设置IoU的值在objectness loss中做标签的系数,
使用代码如下:
tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio
train.py源码中model.gr=1,也就是说完全使用标签框与预测框的CIoU值来作为该预测框的objectness标签。
'''
self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
for k in 'na', 'nc', 'nl', 'anchors':
setattr(self, k, getattr(det, k))
代码和注释如下:
def build_targets(self, p, targets):
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
'''
na = 3,表示每个预测层anchors的个数
targets 为一个batch中所有的标签,包括标签所属的image,以及class,x,y,w,h
targets = [[image1,class1,x1,y1,w1,h1],
[image2,class2,x2,y2,w2,h2],
...
[imageN,classN,xN,yN,wN,hN]]
nt为一个batch中所有标签的数量
'''
na, nt = self.na, targets.shape[0] # number of anchors, targets
tcls, tbox, indices, anch = [], [], [], []
'''
gain是为了最终将坐标所属grid坐标限制在坐标系内,不要超出范围,
其中7是为了对应: image class x y w h ai,
但后续代码只对x y w h赋值,x,y,w,h = nx,ny,nx,ny,
nx和ny为当前输出层的grid大小。
'''
gain = torch.ones(7, device=targets.device) # normalized to gridspace gain
'''
ai.shape = [na,nt]
ai = [[0,0,0,.....],
[1,1,1,...],
[2,2,2,...]]
这么做的目的是为了给targets增加一个属性,即当前标签所属的anchor索引
'''
ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
'''
targets.repeat(na, 1, 1).shape = [na,nt,6]
ai[:, :, None].shape = [na,nt,1](None在list中的作用就是在插入维度1)
ai[:, :, None] = [[[0],[0],[0],.....],
[[1],[1],[1],...],
[[2],[2],[2],...]]
cat之后:
targets.shape = [na,nt,7]
targets = [[[image1,class1,x1,y1,w1,h1,0],
[image2,class2,x2,y2,w2,h2,0],
...
[imageN,classN,xN,yN,wN,hN]],
[[image1,class1,x1,y1,w1,h1,1],
[image2,class2,x2,y2,w2,h2,1],
...],
[[image1,class1,x1,y1,w1,h1,2],
[image2,class2,x2,y2,w2,h2,2],
...]]
这么做是为了纪录每个label对应的anchor。
'''
targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices
'''
定义每个grid偏移量,会根据标签在grid中的相对位置来进行偏移
'''
g = 0.5 # bias
'''
[0, 0]代表中间,
[1, 0] * g = [0.5, 0]代表往左偏移半个grid, [0, 1]*0.5 = [0, 0.5]代表往上偏移半个grid,与后面代码的j,k对应
[-1, 0] * g = [-0.5, 0]代代表往右偏移半个grid, [0, -1]*0.5 = [0, -0.5]代表往下偏移半个grid,与后面代码的l,m对应
具体原理在代码后讲述
'''
off = torch.tensor([[0, 0],
[1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
], device=targets.device).float() * g # offsets
for i in range(self.nl):
'''
原本yaml中加载的anchors.shape = [3,6],但在yolo.py的Detect中已经通过代码
a = torch.tensor(anchors).float().view(self.nl, -1, 2)
self.register_buffer('anchors', a)
将anchors进行了reshape。
self.anchors.shape = [3,3,2]
anchors.shape = [3,2]
'''
anchors = self.anchors[i]
'''
p.shape = [nl,bs,na,nx,ny,no]
p[i].shape = [bs,na,nx,ny,no]
gain = [1,1,nx,ny,nx,ny,1]
'''
gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
# Match targets to anchors
'''
因为targets进行了归一化,默认在w = 1, h =1 的坐标系中,
需要将其映射到当前输出层w = nx, h = ny的坐标系中。
'''
t = targets * gain
if nt:
# Matches
'''
t[:, :, 4:6].shape = [na,nt,2] = [3,nt,2],存放的是标签的w和h
anchor[:,None] = [3,1,2]
r.shape = [3,nt,2],存放的是标签和当前层anchor的长宽比
'''
r = t[:, :, 4:6] / anchors[:, None] # wh ratio
'''
torch.max(r, 1. / r)求出最大的宽比和最大的长比,shape = [3,nt,2]
再max(2)求出同一标签中宽比和长比较大的一个,shape = [2,3,nt],之所以第一个维度变成2,
因为torch.max如果不是比较两个tensor的大小,而是比较1个tensor某一维度的大小,则会返回values和indices:
torch.return_types.max(
values=tensor([...]),
indices=tensor([...]))
所以还需要加上索引0获取values,
torch.max(r, 1. / r).max(2)[0].shape = [3,nt],
将其和hyp.yaml中的anchor_t超参比较,小于该值则认为标签属于当前输出层的anchor
j = [[bool,bool,....],[bool,bool,...],[bool,bool,...]]
j.shape = [3,nt]
'''
j = torch.max(r, 1. / r).max(2)[0] < self.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.shape = [na,nt,7]
j.shape = [3,nt]
假设j中有NTrue个True值,则
t[j].shape = [NTrue,7]
返回的是na*nt的标签中,所有属于当前层anchor的标签。
'''
t = t[j] # filter
# Offsets
'''
下面这段代码和注释可以配合代码后的图片进行理解。
t.shape = [NTrue,7]
7:image,class,x,y,h,w,ai
gxy.shape = [NTrue,2] 存放的是x,y,相当于坐标到坐标系左边框和上边框的记录
gxi.shape = [NTrue,2] 存放的是w-x,h-y,相当于测量坐标到坐标系右边框和下边框的距离
'''
gxy = t[:, 2:4] # grid xy
gxi = gain[[2, 3]] - gxy # inverse
'''
因为grid单位为1,共nx*ny个gird
gxy % 1相当于求得标签在第gxy.long()个grid中以grid左上角为原点的相对坐标,
gxi % 1相当于求得标签在第gxy.long()个grid中以grid右下角为原点的相对坐标,
下面这两行代码作用在于
筛选中心坐标 左、上方偏移量小于0.5,并且中心点大于1的标签
筛选中心坐标 右、下方偏移量小于0.5,并且中心点大于1的标签
j.shape = [NTrue], j = [bool,bool,...]
k.shape = [NTrue], k = [bool,bool,...]
l.shape = [NTrue], l = [bool,bool,...]
m.shape = [NTrue], m = [bool,bool,...]
'''
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
l, m = ((gxi % 1. < g) & (gxi > 1.)).T
'''
j.shape = [5,NTrue]
t.repeat之后shape为[5,NTrue,7],
通过索引j后t.shape = [NOff,7],NOff表示NTrue + (j,k,l,m中True的总数量)
torch.zeros_like(gxy)[None].shape = [1,NTrue,2]
off[:, None].shape = [5,1,2]
相加之和shape = [5,NTrue,2]
通过索引j后offsets.shape = [NOff,2]
这段代码的表示当标签在grid左侧半部分时,会将标签往左偏移0.5个grid,上下右同理。
'''
j = torch.stack((torch.ones_like(j), j, k, l, m))
t = t.repeat((5, 1, 1))[j]
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
else:
t = targets[0]
offsets = 0
# Define
'''
t.shape = [NOff,7],(image,class,x,y,w,h,ai)
'''
b, c = t[:, :2].long().T # image, class
gxy = t[:, 2:4] # grid xy
gwh = t[:, 4:6] # grid wh
'''
offsets.shape = [NOff,2]
gxy - offsets为gxy偏移后的坐标,
gxi通过long()得到偏移后坐标所在的grid坐标
'''
gij = (gxy - offsets).long()
gi, gj = gij.T # grid xy indices
# Append
'''
a:所有anchor的索引 shape = [NOff]
b:标签所属image的索引 shape = [NOff]
gj.clamp_(0, gain[3] - 1)将标签所在grid的y限定在0到ny-1之间
gi.clamp_(0, gain[2] - 1)将标签所在grid的x限定在0到nx-1之间
indices = [image, anchor, gridy, gridx] 最终shape = [nl,4,NOff]
tbox存放的是标签在所在grid内的相对坐标,∈[0,1] 最终shape = [nl,NOff]
anch存放的是anchors 最终shape = [nl,NOff,2]
tcls存放的是标签的分类 最终shape = [nl,NOff]
'''
a = t[:, 6].long() # anchor indices
indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
tbox.append(torch.cat((gxy - gij, gwh), 1)) # box
anch.append(anchors[a]) # anchors
tcls.append(c) # class
return tcls, tbox, indices, anch
在上述论文中的代码中包含了标签偏移的代码部分:
g = 0.5 # bias
off = torch.tensor([[0, 0],
[1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
], device=targets.device).float() * g # offsets
# Offsets
gxy = t[:, 2:4] # grid xy
gxi = gain[[2, 3]] - gxy # inverse
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
l, m = ((gxi % 1. < g) & (gxi > 1.)).T
j = torch.stack((torch.ones_like(j), j, k, l, m))
t = t.repeat((5, 1, 1))[j]
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
gxy = t[:, 2:4] # grid xy
gij = (gxy - offsets).long()
在讲述yolo.py的时候也已经介绍过,这里再介绍一遍。
这段代码的大致意思是,当标签在grid左侧半部分时,会将标签往左偏移0.5个grid,在上、下、右侧同理。具体如图所示:
grid B中的标签在右上半部分,所以标签偏移0.5个gird到E中,A,B,C,D同理,即每个网格除了回归中心点在该网格的目标,还会回归中心点在该网格附近周围网格的目标。以E左上角为坐标(Cx,Cy),所以bx∈[Cx-0.5,Cx+1.5],by∈[Cy-0.5,Cy+1.5],而bw∈[0,4pw],bh∈[0,4ph]应该是为了限制anchor的大小。
代码和注释如下:
def __call__(self, p, targets): # predictions, targets, model
device = targets.device
lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
'''
从build_targets函数中构建目标标签,获取标签中的tcls, tbox, indices, anchors
tcls = [[cls1,cls2,...],[cls1,cls2,...],[cls1,cls2,...]]
tcls.shape = [nl,N]
tbox = [[[gx1,gy1,gw1,gh1],[gx2,gy2,gw2,gh2],...],
indices = [[image indices1,anchor indices1,gridj1,gridi1],
[image indices2,anchor indices2,gridj2,gridi2],
...]]
anchors = [[aw1,ah1],[aw2,ah2],...]
'''
tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets
# Losses
'''
p.shape = [nl,bs,na,nx,ny,no]
nl 为 预测层数,一般为3
na 为 每层预测层的anchor数,一般为3
nx,ny 为 grid的w和h
no 为 输出数,为5 + nc (5:x,y,w,h,obj,nc:分类数)
'''
for i, pi in enumerate(p): # layer index, layer predictions
'''
a:所有anchor的索引
b:标签所属image的索引
gridy:标签所在grid的y,在0到ny-1之间
gridy:标签所在grid的x,在0到nx-1之间
'''
b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
'''
pi.shape = [bs,na,nx,ny,no]
tobj.shape = [bs,na,nx,ny]
'''
tobj = torch.zeros_like(pi[..., 0], device=device) # target obj
n = b.shape[0] # number of targets
if n:
'''
ps为batch中第b个图像第a个anchor的第gj行第gi列的output
ps.shape = [N,5+nc],N = a[0].shape,即符合anchor大小的所有标签数
'''
ps = pi[b, a, gj, gi] # prediction subset corresponding to targets
'''
xy的预测范围为-0.5~1.5
wh的预测范围是0~4倍anchor的w和h,
原理在代码后讲述。
'''
# Regression
pxy = ps[:, :2].sigmoid() * 2. - 0.5
pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
pbox = torch.cat((pxy, pwh), 1) # predicted box
'''
只有当CIOU=True时,才计算CIOU,否则默认为GIOU
'''
iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # iou(prediction, target)
lbox += (1.0 - iou).mean() # iou loss
# Objectness
'''
通过gr用来设置IoU的值在objectness loss中做标签的比重,
'''
tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio
# Classification
if self.nc > 1: # cls loss (only if multiple classes)
'''
ps[:, 5:].shape = [N,nc],用 self.cn 来填充型为[N,nc]得Tensor。
self.cn通过smooth_BCE平滑标签得到的,使得负样本不再是0,而是0.5 * eps
'''
t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets
'''
self.cp 是通过smooth_BCE平滑标签得到的,使得正样本不再是1,而是1.0 - 0.5 * eps
'''
t[range(n), tcls[i]] = self.cp
'''
计算用sigmoid+BCE分类损失
'''
lcls += self.BCEcls(ps[:, 5:], t) # BCE
# 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)]
'''
pi[..., 4]所存储的是预测的obj
'''
obji = self.BCEobj(c, tobj)
'''
self.balance[i]为第i层输出层所占的权重,在init函数中已介绍
将每层的损失乘上权重计算得到obj损失
'''
lobj += obji * self.balance[i] # obj loss
if self.autobalance:
self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
if self.autobalance:
self.balance = [x / self.balance[self.ssi] for x in self.balance]
'''
hyp.yaml中设置了每种损失所占比重,分别对应相乘
'''
lbox *= self.hyp['box']
lobj *= self.hyp['obj']
lcls *= self.hyp['cls']
bs = tobj.shape[0] # batch size
loss = lbox + lobj + lcls
return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
在anchor回归时,对xywh进行了以下处理:
# Regression
pxy = ps[:, :2].sigmoid() * 2. - 0.5
pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
这和yolo.py Detect中的代码一致:
y = x[i].sigmoid()
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
可以先翻看yolov3论文中对于anchor box回归的介绍:
这里的bx∈[Cx,Cx+1],by∈[Cy,Cy+1],bw∈(0,+∞),bh∈(0,+∞)
而yolov5里这段公式变成了:
使得bx∈[Cx-0.5,Cx+1.5],by∈[Cy-0.5,Cy+1.5],bw∈[0,4pw],bh∈[0,4ph]。
原因在build_targets函数时已经进行了介绍。这一策略可以提高召回率(因为每个grid的预测范围变大了),但会略微降低精确度,总体提升mAP。