YOLO-V3-SPP 训练loss计算源码解析之compute_loss

前言

理论详解：YOLO-V3-SPP详细解析
该函数需要了解dataloader那边筛选出来的gt，即build_targets函数

compute_loss

主要讲解model的pred和筛选的gt进行loss计算过程，包括正负样本的区分，以及二值交叉熵loss和forcal loss的转换及使用代码，还有关于IOU的计算。
讲解形式结合图文信息，不会那么枯燥，尽量形象点。

源码

def compute_loss(p, targets, model):  # predictions, targets, modeldevice = p[0].devicelcls = torch.zeros(1, device=device)  # Tensor(0)lbox = torch.zeros(1, device=device)  # Tensor(0)lobj = torch.zeros(1, device=device)  # Tensor(0)tcls, tbox, indices, anchors = build_targets(p, targets, model)  # targetsh = model.hyp  # hyperparametersred = 'mean'  # Loss reduction (sum or mean)"""tcls：筛选出来的gt的类索引，shape(YoloLayer_num,targets_num)tbox：筛选出来的gt的box信息，tx,ty,w,h。其中tx,ty是偏移量；w,h是宽高,shape(YoloLayer_num,targets_num,txtywh)indices：(YoloLayer_num,img_index+anchor_index+grid_y+grid_x)anch：每个target对应使用的anchor尺度,shape(YoloLayer_num,targets_num,wh)"""# Define criteriaBCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device), reduction=red)BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device), reduction=red)# class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3cp, cn = smooth_BCE(eps=0.1)# focal lossg = h['fl_gamma']  # focal loss gammaif g > 0:BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)# per outputnt = 0  # targetsfor i, pi in enumerate(p):  # layer index, layer predictionsb, a, gj, gi = indices[i]  # image, anchor, gridy, gridxtobj = torch.zeros_like(pi[..., 0], device=device)  # target objnb = b.shape[0]  # number of targetsif nb:# 对应匹配到正样本的预测信息ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets# GIoUpxy = ps[:, :2].sigmoid()pwh = ps[:, 2:4].exp().clamp(max=1E3) * anchors[i]pbox = torch.cat((pxy, pwh), 1)  # predicted boxgiou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True)  # giou(prediction, target)lbox += (1.0 - giou).mean()  # giou loss# Objtobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)  # giou ratio# Classif model.nc > 1:  # cls loss (only if multiple classes)t = torch.full_like(ps[:, 5:], cn, device=device)  # targetst[range(nb), tcls[i]] = cplcls += 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)]lobj += BCEobj(pi[..., 4], tobj)  # obj loss# 乘上每种损失的对应权重lbox *= h['giou']lobj *= h['obj']lcls *= h['cls']# loss = lbox + lobj + lclsreturn {
    "box_loss": lbox,"obj_loss": lobj,"class_loss": lcls}

详解

def compute_loss(p, targets, model):  # predictions, targets, modeldevice = p[0].devicelcls = torch.zeros(1, device=device)  # Tensor(0)lbox = torch.zeros(1, device=device)  # Tensor(0)lobj = torch.zeros(1, device=device)  # Tensor(0)tcls, tbox, indices, anchors = build_targets(p, targets, model)  # targetsh = model.hyp  # hyperparameters

这里需要了解dataloader那边筛选出来的gt，即build_targets函数
tcls：筛选出来的gt的类索引，shape(YoloLayer_num,targets_num)
tbox：筛选出来的gt的box信息，tx,ty,w,h。其中tx,ty是偏移量；w,h是宽高,$shape(YoloLayer\_num,targets\_num,t_x{t}_{y}wh)$
indices：(YoloLayer_num,img_index+anchor_index+grid_y+grid_x)
anch：每个target对应使用的anchor尺度,shape(YoloLayer_num,targets_num,wh)

    red = 'mean'  # Loss reduction (sum or mean)# Define criteriaBCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device), reduction=red)BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device), reduction=red)

这里如果采用BCELOSS，即二值交叉熵损失函数，取的是平均二值交叉熵loss。
这里BCEWithLogitsLoss传入了一个参数reduction，关于reduction的说明如下：

这里采用的是平均二值交叉熵
关于YOLOV3应用BCELOSS的公式参考BCELOSS公式

	# class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3cp, cn = smooth_BCE(eps=0.1)# focal lossg = h['fl_gamma']  # focal loss gammaif g > 0:BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)

这里cp,cn是对样本标签进行平滑的参数，后续会用到再提到。这里贴出标签平滑的源码：

def smooth_BCE(eps=0.1):  # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441# return positive, negative label smoothing BCE targetsreturn 1.0 - 0.5 * eps, 0.5 * eps

这里如果超参数fl_gamma＞0，则将定义好的BCEloss module传入Foralloss module修改为Forcalloss。

这里贴出Forcalloss的源码:

class FocalLoss(nn.Module):# Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):super(FocalLoss, self).__init__()self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()self.gamma = gammaself.alpha = alphaself.reduction = loss_fcn.reductionself.loss_fcn.reduction = 'none'  # required to apply FL to each elementdef forward(self, pred, true):loss = self.loss_fcn(pred, true)# p_t = torch.exp(-loss)# loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability# TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.pypred_prob = torch.sigmoid(pred)  # prob from logitsp_t = true * pred_prob + (1 - true) * (1 - pred_prob)alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)modulating_factor = (1.0 - p_t) ** self.gammaloss *= alpha_factor * modulating_factorif self.reduction == 'mean':return loss.mean()elif self.reduction == 'sum':return loss.sum()else:  # 'none'return loss

output的处理

 	# per outputnt = 0  # targetsfor i, pi in enumerate(p):  # layer index, layer predictionsb, a, gj, gi = indices[i]  # image, anchor, gridy, gridxtobj = torch.zeros_like(pi[..., 0], device=device)  # target objnb = b.shape[0]  # number of targets

p的shape为(YoloLayer_num,batch_size,anchor_num,grid_x,grid_y,xywh+obj_confidence+classes_num)
pi的shape为(batch_size,anchor_num,grid_x,grid_y,xywh+obj_confidence+classes_num)
这里计算loss是训练阶段的输出，那么这里的输出的p是未归一化的信息。
$b,a,gj,gi$分别表示$image\_index,anchor\_index,grid\_y,gird\_x$
tobj生成一个和预测类别个数相同维度的，值全为0的tensor
nb获取这一批次筛选出来的gt数量，即targets数量

        if nb:# 对应匹配到正样本的预测信息ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets# GIoUpxy = ps[:, :2].sigmoid()pwh = ps[:, 2:4].exp().clamp(max=1E3) * anchors[i]pbox = torch.cat((pxy, pwh), 1)  # predicted boxgiou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True)  # giou(prediction, target)lbox += (1.0 - giou).mean()  # giou loss

对当前pi的shape为(batch_size,anchor_num,grid_x,grid_y,xywh+obj_confidence+classes_num)
b表示该batch的所有target的图片索引
a表示该batch的所有target的anchor索引
gj和gi表示所有target的grid_cell坐标
ps = pi[b, a, gj, gi]表示取该模型输出中前四个维度为[image_index,anchor_index,gj,gi]的[x,y,w,h,obj,cls]
ps的shape为(targets_num,xywh+obj+cls]

pxy对xy输出进行sigmoid处理，即归一化，shape为(targets_num,2)
pwh将预测的未处理的宽高维度的信息基于anchor映射到feature map尺度上的宽高信息，shape为(targets_num,2)
宽高映射在models.py的yololayer的前向传播中也定义了，两者没有什么太大的区别，感兴趣可以回去看看models.py的代码
注：pwh的映射调用一个clamp函数1E3表示$1\times 10^3$，表示将pwh的值限制在1000之内，不清楚有无必要，models的处理没有用到这个。

io[..., 2:4] = torch.exp(io[..., 2:4]) * self.anchor_wh

论文中，关于预测输出的映射关系如下：$\sigma$表示sigmoid处理

pbox将pxy和pwh在第二维度拼接，得到shape(targets_num,xywh)

关于iou计算中使用偏移量来计算的原因

pbox：预测输出中xy是对应grid_cell的中心偏移量，尺度均是feature map尺度
$shape(targets\_num,xywh)$
tbox：筛选出来的gt的box信息，tx,ty,w,h。其中tx,ty是偏移量；w,h是宽高，尺度均是feature map尺度
$shape(YoloLayer\_num,targets\_num,t_x{t}_{y}wh)$

pbox和tbox的尺度是feature map尺度
pbox为网络的输出，回顾之前代码：

    for i, pi in enumerate(p):  # layer index, layer predictionsb, a, gj, gi = indices[i]  # image, anchor, gridy, gridxtobj = torch.zeros_like(pi[..., 0], device=device)  # target objnb = b.shape[0]  # number of targetsif nb:# 对应匹配到正样本的预测信息ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets# GIoUpxy = ps[:, :2].sigmoid()pwh = ps[:, 2:4].exp().clamp(max=1E3) * anchors[i]pbox = torch.cat((pxy, pwh), 1)  # predicted box

pi的shape为(batch_size,anchor_num,grid_x,grid_y,xywh+obj_confidence+classes_num)，其中xy表示基于当前grid_x和grid_y的偏移量，pxy对该偏移量经过sigmoid处理后得到的xy就是feature map尺度上的对应grid_x和grid_y的偏移量（注：ps将gt对应图片，使用anchor，所在的gridcell的预测xywh筛选出来，和tbox一一对应）

build_targets函数中tbox是feature map尺度上的，详情见YOLO-V3-SPP utils.py build_targets函数-详细解读（ultralytic版本）

loss中giou的计算

giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True)
# giou(prediction, target)

这里pbox经过转置传入bbox_iou,方便bbox_iou计算iou

pbox：预测输出中恢复到feature map尺度的xywh，xy也是偏移量
$shape(targets\_num,xywh)$
tbox：筛选出来的gt的box信息，tx,ty,w,h。其中tx,ty是偏移量；w,h是宽高
$shape(YoloLayer\_num,targets\_num,t_x{t}_{y}wh)$

bbox_iou源码

def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False):# Returns the IoU of box1 to box2. box1 is 4, box2 is nx4box2 = box2.t()# Get the coordinates of bounding boxesif x1y1x2y2:  # x1, y1, x2, y2 = box1b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]else:  # transform from xywh to xyxyb1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2# Intersection areainter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \(torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)# Union Areaw1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1union = (w1 * h1 + 1e-16) + w2 * h2 - interiou = inter / union  # iouif GIoU or DIoU or CIoU:cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1)  # convex (smallest enclosing box) widthch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1)  # convex heightif GIoU:  # Generalized IoU https://arxiv.org/pdf/1902.09630.pdfc_area = cw * ch + 1e-16  # convex areareturn iou - (c_area - union) / c_area  # GIoUif DIoU or CIoU:  # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1# convex diagonal squaredc2 = cw ** 2 + ch ** 2 + 1e-16# centerpoint distance squaredrho2 = ((b2_x1 + b2_x2) - (b1_x1 + b1_x2)) ** 2 / 4 + ((b2_y1 + b2_y2) - (b1_y1 + b1_y2)) ** 2 / 4if DIoU:return iou - rho2 / c2  # DIoUelif CIoU:  # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)with torch.no_grad():alpha = v / (1 - iou + v)return iou - (rho2 / c2 + v * alpha)  # CIoUreturn iou

bbox_iou源码解析

# giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True)
def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False):# Returns the IoU of box1 to box2. box1 is 4, box2 is nx4box2 = box2.t()

第一个pbox传进来前进行了转置，第二个参数tbox传进来后也得进行转置方便计算iou，所以传参前后进行一次转置均可

函数的第三个参数x1y1x2y2表示计算iou的box坐标格式的布尔变量
其他参数表示所使用的iou属于哪种iou

	# Get the coordinates of bounding boxesif x1y1x2y2:  # x1, y1, x2, y2 = box1b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]else:  # transform from xywh to xyxyb1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2

这段代码处理box的格式

回顾下GIOU的算法公式：

    # Intersection area \为续行符inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \(torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)

计算两个框之间的交集区域面积$|A\cap B|$,其中clamp(0)约束当两框不相交时，IOU取0

    # Union Areaw1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1union = (w1 * h1 + 1e-16) + w2 * h2 - interiou = inter / union  # iou

计算两个框之间的并集区域面积$|A\cup B|$,其中1e-16是一个很小的正数，目的是防止iou的除法运算遇到分母为0导致计算逻辑错误的情况

    if GIoU or DIoU or CIoU:cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1)  # convex (smallest enclosing box) widthch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1)  # convex heightif GIoU:  # Generalized IoU https://arxiv.org/pdf/1902.09630.pdfc_area = cw * ch + 1e-16  # convex areareturn iou - (c_area - union) / c_area  # GIoUif DIoU or CIoU:  # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1# convex diagonal squaredc2 = cw ** 2 + ch ** 2 + 1e-16# centerpoint distance squaredrho2 = ((b2_x1 + b2_x2) - (b1_x1 + b1_x2)) ** 2 / 4 + ((b2_y1 + b2_y2) - (b1_y1 + b1_y2)) ** 2 / 4if DIoU:return iou - rho2 / c2  # DIoUelif CIoU:  # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)with torch.no_grad():alpha = v / (1 - iou + v)return iou - (rho2 / c2 + v * alpha)  # CIoU

这里只讲GIOU的实现部分，回顾下GIOU的公式：
$$GIOU=IOU?\frac{|C?(A\cup B)|}{|C|}$$
cw,ch分别表示A和B的最小闭包C的宽高
最终bbox_iou的返回值为tensor(targets_num,)表示每个预测输出和gt的giou值

loss中定位损失的计算

giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True)  # giou(prediction, target)
lbox += (1.0 - giou).mean()  # giou loss

这里giou状态为tensor(targets_num,)，表示每个预测输出和gt的giou值
回顾GIOU损失计算公式：$L_{GIOU}=1?GIOU$
lbox属于定位损失，这里做了mean()处理，即求giou的tensor里所有数值的和取平均：
$$lbox=\frac{{\sum }_{j=1}^{targets\_num}(1?gio{u}_j)}{targets\_num}$$

loss中置信度损失的计算

注：这里指的BCEloss均是BCEWithLogitsLoss，与BCEloss的区别就是BCEWithLogitsLoss在使用时内置了对input的sigmoid处理。
参考：Pytorch详解BCELoss和BCEWithLogitsLoss

# Obj
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)  # giou ratio
lobj += BCEobj(pi[..., 4], tobj)  # obj loss

detach的目的

使giou这个tensor的required_grad参数从true设为false
具体可以参考这个博客的解释：Pytorch之requires_grad

clamp（0）的目的

这里giou经过了clamp(0)，将giou的下限设为0，giou的理论范围是$[?1,1]$，在进行loss计算时，$L_{GIOU}=1?GIOU$使在pred和gt不重叠时得到的负值giou变为正值，从而能够训练。但在计算置信度损失时，对于不重叠的pred和gt，默认是将giou置0，表示置信度为0。

model.gr的作用

从代码就可以看出，对于求出来的giou乘上了model.gr参数，yolov3默认使用model.gr=1.0，这个参数在train.py文件有定义

model.gr = 1.0  # giou loss ratio (obj_loss = 1.0 or giou)

对于简单样本（pred和gt拟合效果好），giou会越接近1，置信度会越高。
但对于困难样本（pred和gt拟合效果差），giou经过clamp(0)会越接近0,置信度会越低。

这里gr是平衡简单样本的置信度和困难样本的置信度，yolov3-spp的gr=1，相当于没有平衡。对于困难样本较多的情况，我们可以适当设置这个gr来平衡。

置信度损失和分类损失计算

回顾下tobj的初始化

tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj

tobj[b, a, gj, gi]筛选了对应targets的tensor维度的数据进行置信度填充

tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)
lobj += BCEobj(pi[..., 4], tobj)  # obj loss

求出每个target的置信度之后，将置信度tobj（也是giou）和pred的置信度传入BCEobj
这里给出BCEloss的公式
$$L_{conf}(o,c)=?\frac{{\sum }_i(o_i\ln ({\hat{c}}_i)+(1?o_i)\ln (1?{\hat{c}}_i))}{N}$$
$${\hat{c}}_i=Sigmoid(c_i)$$
其中$o_i\in [0,1]$，表示预测目标边界框与真实目标边界框的IOU，
$c$为预测值，${\hat{c}}_i$为$c$通过$Sigmoid$函数得到的预测置信度（预测的IOU）
$N$为正负样本个数

注：$o_i$和${\hat{c}}_i$均指IOU，唯一区别是$o_i$是pred和gt的IOU（作为label指导训练），而${\hat{c}}_i$是网络pred的IOU。

注：对交叉熵不太了解的看看这篇博文交叉熵的理解

BCEloss能解决二分类问题，这里一类是pred的${\hat{c}}_i$，一类是pred和gt的$o_i$(作为label指导训练)。对于每个target，置信度的训练都是一个二分类问题，因此使用BCEloss作为loss计算，这里对每个targets求得的BCE进行求和取平均。
如果采用了Forcal loss，BCEloss将初始化为Forcal loss类
初始化代码:

    # focal lossg = h['fl_gamma']  # focal loss gammaif g > 0:BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)

Forcal loss类定义

注：BCE和forcal类均是torch计算图的module，属于网络传播的一部分。

class FocalLoss(nn.Module):# Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):super(FocalLoss, self).__init__()self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()self.gamma = gammaself.alpha = alphaself.reduction = loss_fcn.reductionself.loss_fcn.reduction = 'none'  # required to apply FL to each elementdef forward(self, pred, true):loss = self.loss_fcn(pred, true)# p_t = torch.exp(-loss)# loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability# TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.pypred_prob = torch.sigmoid(pred)  # prob from logitsp_t = true * pred_prob + (1 - true) * (1 - pred_prob)alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)modulating_factor = (1.0 - p_t) ** self.gammaloss *= alpha_factor * modulating_factorif self.reduction == 'mean':return loss.mean()elif self.reduction == 'sum':return loss.sum()else:  # 'none'return loss

Forcal loss源码详解

    # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):super(FocalLoss, self).__init__()self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()self.gamma = gammaself.alpha = alphaself.reduction = loss_fcn.reductionself.loss_fcn.reduction = 'none'  # required to apply FL to each element

init函数有一个loss_fcn参数，必须是nn.BCEWithLogitsLoss()对象，在调用Forcal loss的使用传入，gamma和alpha参数默认为1.5和0.2。

self.reduction = loss_fcn.reduction
self.loss_fcn.reduction = 'none'  # required to apply FL to each element

注：reduction是loss的结果处理参数，主要有以下三种状态：
reduction=mean：表示对loss的所有targets结果进行求均
reduction=sum：表示对loss的所有targets结果进行求和
reduction=none：表示不对loss的结果处理，输出一个包含所有targets结果的tensor

这里第一句代码的reduction是forcal loss的reduction，loss_fcn的reduciton是mean，表示forcal loss输出是对所有targets结果进行求均
第二句代码将BCE的reduction从mean修改为none，这样将传入进来的BCE的输出将是所有targets结果的tensor，将此结果去计算forcal loss

 def forward(self, pred, true):loss = self.loss_fcn(pred, true)

这里是前向传播的参数，pred是预测信息，true是“lable”（作label指导训练的）。

调用代码如下：

lobj += BCEobj(pi[..., 4], tobj)  # obj loss

前向传播细节

pred_prob = torch.sigmoid(pred)  # prob from logits
p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
modulating_factor = (1.0 - p_t) ** self.gamma
loss *= alpha_factor * modulating_factor

pred_prob将pi的置信度进行sigmoid处理，缩放到[0,1]区间，代表置信度概率。

这里给出forcal loss的公式

p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
p_t表示targets的预测正确或预测错误的概率，对应公式中的$p_t$

alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
alpha_factor表示平衡正负样本在loss的权重，对应公式中的${\alpha }_t$

modulating_factor = (1.0 - p_t) ** self.gamma
对应公式的$(1?p_t{)}^{\gamma }$

loss *= alpha_factor * modulating_factor
alpha_factor * modulating_factor对应公式中的${\alpha }_t(1?p_t{)}^{\gamma }$

loss = self.loss_fcn(pred, true)
这里返回得到的loss为shape(targets_num,pred_num）
pred_num指预测输出的数目，比如预测置信度是1个，那么pred_num=1,再比如预测类别，pred_num=类别数目。pred_num的每一个数值都表示当前预测的二值交叉熵。
loss得到的值对应公式中的$?log(p_t)$

loss *= alpha_factor * modulating_factor
这里将上述求得的${\alpha }_t(1?p_t{)}^{\gamma }$和$?log(p_t)$，相乘得$$FL(p_t)=?{\alpha }_t(1?p_t{)}^{\gamma }log(p_t)$$

        if self.reduction == 'mean':return loss.mean()elif self.reduction == 'sum':return loss.sum()else:  # 'none'return loss

这里对loss的所有维度取平均，得到的是一个数值

回到置信度和分类损失计算

# Class
if model.nc > 1:  # cls loss (only if multiple classes)t = torch.full_like(ps[:, 5:], cn, device=device)  # targetst[range(nb), tcls[i]] = cplcls += BCEcls(ps[:, 5:], t)  # BCE

分类损失和前面置信度的计算大同小异，这里不再赘述。

comupte_loss返回值

    # 乘上每种损失的对应权重lbox *= h['giou']lobj *= h['obj']lcls *= h['cls']# loss = lbox + lobj + lclsreturn {
    "box_loss": lbox,"obj_loss": lobj,"class_loss": lcls}

这里对lbox(定位损失)，lobj(置信度损失)，lcls(分类损失)乘以一个分配权重，作为超参，这个参数是作者优化得到的，不会轻易改动。

一些注意的地方

compute_loss，for循环三个预测器中，唯独置信度损失lobj是需要考虑正样本和负样本的，其他损失只需考虑正样本的计算。
置信度损失lbox在计算的时候需要计算正负样本的损失，所以在对lbox进行累加时，lbox的累加是放在if nb：之外的，即判断体之外，这样才能累加到正负样本。

还有一个值得一提的是，由于我们筛选的gt是经过whIOU得到的，导致得到的targets是不全的，在计算lbox时会将那些没有被筛选到的gt当成负样本来算，导致lbox可能会偏高一点。（实际上那些被whIOU筛掉的gt没有经过计算，lbox将这些被筛掉gt对应的tobj置0，会让lbox有点虚高）
当然这些是能够被一些trick平衡掉一些

下一步要研究的地方

对loss的计算以及正负样本的分配已经基本了解
后面会研究预测模块对数据的处理