2021-03-06 yolov5 compute_loss分析，我准备结合这张图分析一下，图是江小白的。

def compute_loss(p, targets, model):

    # predictions=[tensor[batch, 3(每个位置3个预测值), 80, 80, 类别+5],[...],[...]]

   # targets,传入的是1batch 图的所有框结构tensor([[图片索引,类别,中心x,中心y,宽,高])

#  model,

 

    #  model,

    device = targets.device   #targets.device “cpu” 或者“cuda”

 

    #print("预测结果：",p)

    lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)

    tcls, tbox, indices, anchors = build_targets(p, targets, model)  # targets

#build_targets收集了该批训练图片中的信息

tcls, tbox, indices, anchors结构都为[tensor1, tensor2, tensor3]

其中tensor1对应模型第一部分输出, tensor2对应第二部分……

 #tcls 真实目标类别0,1,2,3...., tbox  真实目标框(偏移x偏移x可能是每个格子大小作为1个单位,偏移y,比值宽,比值高), indices 图片索引, anchors每个真实框对应的锚框（每个预测结果都有3个锚框可供选择）

#indices =[ind，，]  ind=[[image索引], [anchor索引], [gridy 第几个格子第几行或列],[gridx 第几个格子第几列或行]] #indices完全对应预测模型每个数据位置

 

    #

 

    h = model.hyp  # hyperparameters 默认是data/hyp.scratch.yaml超参数设置

 

    # Define criteria h['cls_pw']类别权重 h['obj_pw'] 目标权重 可能是概率权重

    BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([h['cls_pw']])).to(device)

    BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([h['obj_pw']])).to(device)

 

    # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3

    cp, cn = smooth_BCE(eps=0.0)

 

    # Focal loss  #默认未启用

    g = h['fl_gamma']  # focal loss gamma

    if g > 0:

        BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)

 

    # Losses

    nt = 0  # number of targets

    np = len(p)  # number of outputs

    balance = [4.0, 1.0, 0.4] if np == 3 else [4.0, 1.0, 0.4, 0.1]  # P3-5 or P3-6

for i, pi in enumerate(p):  # layer index, layer predictions

 p是这3个预测矩阵

        b, a, gj, gi = indices[i]  # image, anchor, gridy, gridx

#indices[i]这里对应第i个输出方阵（总共3个）的所有训练数据

# indices的结构为[[该批图片索引，对应的anchor(0,1,2),对应的格子行，对应的格子列],[…],[…]]

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

 

        n = b.shape[0]  # number of targets

        if n:

            nt += n  # cumulative targets

            ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets

 

            # Regression

            pxy = ps[:, :2].sigmoid() * 2. - 0.5    #预测偏移 用于计算giou  ？？？？

pxy的范围改为图中蓝色的区域

            pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]  ##预测的宽高pwh先sigmodi把范围缩小到【0,1】*2然后平方，范围变为（0,4）意思是最大只能是anchor的4倍？可以无限缩小。没有彻底理解

            pbox = torch.cat((pxy, pwh), 1).to(device)  # predicted box

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

            #计算giou giou=1完全匹配，giou越大匹配度越高 (-1,+1]

 

            lbox += (1.0 - giou).mean()  # giou loss 框的损失 loss box

 

            # Objectness

            # Objectness

            # tobj.shape = (bs, 3, w, h), if model.gr=1, then tobj is the iou with shape(bs, 3, w, h)

            # tobj是shape=(bs, 3, w, h)的tensor,

            # 正样本处保存预测框和gt框的iou，负样本处仍然是0，用作obj损失的真值。

            tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype)  # giou ratio

 

            # Classification

            if model.nc > 1:  # cls loss (only if multiple classes)

                t = torch.full_like(ps[:, 5:], cn, device=device)  # targets

                t[range(n), tcls[i]] = cp

                lcls += 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) * balance[i]  # obj loss，balance[i]是3个输出权重

 

    s = 3 / np  # output count scaling，np就是输出数为3

    lbox *= h['giou'] * s

    lobj *= h['obj'] * s * (1.4 if np == 4 else 1.)

    lcls *= h['cls'] * s

    bs = tobj.shape[0]  # batch size

 

    loss = lbox + lobj + lcls

    return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()