YOLOV5源码解析-损失解释 compute_loss(), build_targtets()

  • Post author:
  • Post category:其他




1. 损失的特殊性

先说说YOLOV5的损失:一般检测的损失分为

分类损失



回归损失

  1. 一般的检测算法:

    • 回归损失只有正样本有
    • 分类损失的标签直接就非1即0,正样本的标签是1,负样本的标签是0(可以把背景作为一种类别一起算),

      这些标签根据anchor和GT框的比较就可以直接得到

      。比如标签是[0,1,0,0],代表当前样本属于背景或其他三种类别的标签。
  2. 而YOLOV5中:

    • 回归损失只有正样本有
    • 分类损失分为两种,

      • 一种是正样本的分类损失(不包括背景的onehot标签,比如[1,0,0])。
      • 另一种是

        正负

        样本的

        前景背景预测

        的分类损失,这里叫Objectness loss,比起以往算法前景背景损失的标签非1即0,这里采用预测框和GT框的IOU作为Objectness的标签,也就是说预测框和GT框重合度越高,当前样本属于前景的可能性就越大。一般算法对于不同的anchor接近GT框的程度不同,标签都是非1即0,统一对待。

        另外注意这里的Objectness loss的标签在得到预测值之前是不知道的,也就是说根据预测框的结果才实时的计算出来的,这种方法直观上来讲更合适些。

接下来细看代码



2. compute_loss


正负样本

指的是anchor!

中英文混合注释如下:

def compute_loss(p, targets, model):  # predictions, targets, model
    device = targets.device
    lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
    # build_targets主要为了拿到所有targets(扩充了周围grids)对应的类别,框,图片和anchor索引,以及具体的anchors。
    tcls, tbox, indices, anchors = build_targets(p, targets, model)  # targets
    h = model.hyp  # hyperparameters

    # Define criteria, 'cls_pw' = 1, 'obj_pw' = 1
    BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
    BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))

    # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
    # cp = 1, cn = 0, 正负样本的损失权重
    cp, cn = smooth_BCE(eps=0.0)

    # Focal loss, 没有采用focal loss, 所以g=0
    g = h['fl_gamma']  # focal loss gamma # TBD
    if g > 0:
        BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)

    # Losses
    nt = 0  # number of targets
    # P3-P5 feature map的特征点的数量比值为P3 : P4 : P5 = 4:1:0.25
    # 不过在求损失的时候用的是平均损失,所以似乎没必要让各自的损失权重再乘以对应层的特征点数量的比例了。 
    # So just the P3 is more significant than P4? TBD
    # balance = [4.0, 1.0, 0.3, 0.1, 0.03]  # P3-P7
    balance = [1.0, 1.0, 1.0, 1.0, 1.0]
    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], device=device)  # target obj

        n = b.shape[0]  # number of targets
        if n:
            nt += n  # cumulative targets
            # pi.shape = (bs, na, w, h, 4+1+class_num)
            # ps.shape = (220, 4+1+class_num), 某次运行中的??是220
            # 对应targets的预测值,pi是每一层的预测tensor
            # gj,gi是中心点所在feature map位置,是采用targets中心点所在位置的anchor来回归。
            ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets

            # Regression loss
            # 将预测值转换成box,并计算预测框和gt框的iou损失。
            pxy = ps[:, :2].sigmoid() * 2. - 0.5
            pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
            pbox = torch.cat((pxy, pwh), 1)  # predicted box
            iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True)  # iou(prediction, target)
            lbox += (1.0 - iou).mean()  # iou loss

            # 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(anchor)处保存预测框和gt框的iou,负样本(anchor)处仍然是0,用作obj损失的真值。
            tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * iou.detach().clamp(0).type(tobj.dtype)

            # Classification loss
            if model.nc > 1:  # cls loss (only if multiple classes)
            	# t.shape = (220, cls_num),t的值都是cn(0)
                t = torch.full_like(ps[:, 5:], cn, device=device)  # targets,
                # 为正样本分配标签,cp=1,变成onehot类型标枪。 t.shape = (220, cls_num)
                # tcls[i].shape = (220,), tcls[i]里面所有的值都是0~cls_num 
                t[range(n), tcls[i]] = cp   # n = 220
                # ps.shape = (220, 4+1+class_num), ??=220
                # 只有正样本的分类损失
                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)]
        # only balance P3-P5; iou of preds and targets is the target of BCE.
        # pi[..., 4] 是前景的预测值,‘1’代表肯定是前景。这里单独计算一个前景预测的损失,还要乘以一个系数。
        # lobj损失包括正负样本的损失,如此训练,使得负样本处的预测值接近0.
        lobj += BCEobj(pi[..., 4], tobj) * balance[i]  # obj loss

    # TBD: loss*bs, h para?, model.gr?
    lbox *= h['box']
    lobj *= h['obj']
    lcls *= h['cls']
    bs = tobj.shape[0]  # batch size

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



3. build_targets

中英文混合注释如下:

def build_targets(p, targets, model):
    '''Build targets for compute_loss(), input targets(image,class,x,y,w,h)
    Args: 
        p is output of model, list of 3 layers.
        targets is normalized label, targets.shape = (nt, 6).
        model is the model.
    '''
    det = model.module.model[-1] if is_parallel(model) else model.model[-1]  # Detect() module
    # na是每一层layer的每一个位置的anchor的数量,nt是当前batch图片中所有目标框的数量
    na, nt = det.na, targets.shape[0]  # number of anchors (3), targets (28)
    tcls, tbox, indices, anch = [], [], [], []
    gain = torch.ones(7, device=targets.device)  # normalized to gridspace gain
    # ai is matrix: [[0,0,...,0], [1,1,...,1], [2,2,...,2]], ai.shape = (na, nt)
    ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)
    # after repeat, targets.shape = (na, nt, 6), after cat, targets.shape = (na, nt, 7) 
    # 将targets扩充至(na, nt, 7),也就是每个anchor与每个targets都有对应,为了接下来计算损失用
    # targets的值[ [[image,class,x,y,w,h,0],
    #             [image,class,x,y,w,h,0],
    #               	...		共nt个   ]

	# 			  [[image,class,x,y,w,h,1],
	#              [image,class,x,y,w,h,1],
	#                   ...		共nt个    ]

	# 			  [[image,class,x,y,w,h,2],
	#              [image,class,x,y,w,h,2],
	#                   ...		共nt个    ]
	#          ]
    targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)  # append anchor indices

    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

	# 每一层layer(共三层)单独计算。
	# 先通过判定每个target和3个anchor的长宽比是否满足一定条件,来得到满足条件的anchor所对应的targets (t)。
	# 这时的anchor数量是3,并不是某个位置的anchor,而是当前层的anchor。
	# 这时的t是3个anchor对应的targets的值,也就是说如果一个target如果对应多个anchor,那么t就有重复的值。
	
	# 然后根据t的每个target的中心点的偏移情况,得到扩充3倍的t。
	# 这时的t就是3个anchor对应的targets的值的扩充。

	# 接下来indices保存每层targets对应的图片索引,对应的anchor索引(只有3个),以及中心点坐标。
	# 接下来计算损失的时候,要根据targets对应的anchor索引来选择在某个具体位置的anchors,用来回归。
    for i in range(det.nl):
    	# 每一层的anchor数量都是3个,三层共9个。
        anchors = det.anchors[i]  # det.anchors.shape = (3, 3, 2)
        # p[i].shape = (bs, na, h, w, 4+num_class+1)
        # gain[2:6]: [w, h, w, h], w & h is the width and height of the feature map
        gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]]  # xyxy gain

        # Match targets to anchors, before mul gain, targets is normalized. t.shape = (na, nt, 7)
        # 将提前归一化的targets转变成当前feature map下的绝对尺寸。
        t = targets * gain
        if nt:
            # Matches, t[:, :, 4:6].shape=(na, nt, 2), anchors[:, None].shape=(3, 1, 2)
            # 在当前feature map下,绝对尺寸的targets / 绝对尺寸的anchor,得到所有目标对于所有anchor的长宽的比例。
            r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
            # max of w & h ratio between targets and anchor that < 4.
            # it means that the targets match the anchors. j.shape = (na, nt)
            # 根据得到的比例来判定targets和当前layer的3个anchor是否匹配,最大边长的长度比小于4,则认为匹配上了。
            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))
            
            # before filter, t.shape = (3, nt, 7), j.shape = (3, nt)
            # after filter, t.shape = (?, 7) 
            # t经过[j]索引后,就是匹配上这3个anchor的targets,匹配的数量视情况而定,这里用?表示。
            # t经过[j]索引后的值有重复,因为每个target可能匹配到多个anchor
            # 到这一步其实指示根据anchor的长宽比来计算匹配的targets。
            t = t[j]  # filter

            # Offsets
            # gxy.shape = (?, 2), it means gt's x, y
            gxy = t[:, 2:4]  # grid xy
            gxi = gain[[2, 3]] - gxy  # inverse
            # j means if x lean in left, k means if y lean in top.
            # l means if x lean in right, m means if y lean in bottom.
            # 得到该点更向哪个方向偏移,j.shape = (?,),其他k,l,m的shape一致。
            j, k = ((gxy % 1. < g) & (gxy > 1.)).T
            l, m = ((gxi % 1. < g) & (gxi > 1.)).T
            # j.shape = (5, ?), ? means nrof filtered anchors
            j = torch.stack((torch.ones_like(j), j, k, l, m))
            # 在[j]索引之前,t.shape = (5, ?, 7), 
            # 5 means the 4+1 grids, which '1' is the center grid, '4' means the surounding grids.
            # 在[j]索引之后,t.shape = (??, 7), 得到了扩充了的偏移后的targets。
            # 除了targets的中心点落在在feature map上的grid位置外,还有'上下左右'四个方向中,两个更靠近的grid。
            t = t.repeat((5, 1, 1))[j]
            # shape(1, ?, 2) + shape(5, 1, 2) = shape(5, ?, 2)
            # after indexed by [j], offsets.shape = (??, 2), 
            # 举例来说,gxy.shape = (72, 2), offsets.shape = (220, 2)。
            # shape=(5, 72, 2)的tensor经过[j]的索引后,offsets.shape = (220, 2)
            offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
        else:
            t = targets[0]
            offsets = 0

        # Define
        b, c = t[:, :2].long().T  # image, class
        gxy = t[:, 2:4]  # grid xy
        gwh = t[:, 4:6]  # grid wh
        # which grid does the gt falls in. gij.shape = (??, 2), ?? ≈ ?*3
        # 对所有targets和match anchor的坐标集合做一个补充,不光是match的那个anchor的grid坐标,
        # 还包括所有targets在’上下左右‘四个方向中,更靠近的两个方向的位置的坐标。
        gij = (gxy - offsets).long()
        # 拆成横坐标和纵坐标,这些就代表正样本的坐标
        gi, gj = gij.T  # grid xy indices

        # Append
        a = t[:, 6].long()  # anchor indices (0或者1或者2), a.shape = (??,)
        # indices将每层的正样本的图片索引号(一个batch中的图片的索引号)
        # 对应的anchor索引号(只有3个),以及坐标保存下来。
        # image, anchor, grid indices
        indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))
        # tbox是[(??, 4), (??, 4), (??,4)]的list, 保存了每个正样本对应的gt box?
        tbox.append(torch.cat((gxy - gij, gwh), 1))  # box
        # anch是[(??, 2), (??, 2), (??, 2)]的list,保存了每个正样本anchor。
        anch.append(anchors[a])  # anchors
        # tcls是[(??,), (??,), (??,)]的list, 保存了每个正样本对应的类别。
        tcls.append(c)  # class

    return tcls, tbox, indices, anch

引用一张图片来解释扩充后的正样本对应targets的中心点的集合



版权声明:本文为lcb_coconut原创文章,遵循 CC 4.0 BY-SA 版权协议,转载请附上原文出处链接和本声明。