devbox
weixin_40725706
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

YOLOV7算法(三)损失函数ComputeLossOTA学习记录_yolov7损失函数

作者：weixin_40725706 | 2024-04-08 10:40:20

赞

踩

yolov7损失函数

YOLOV7正负样本策略及ComputeLossOTA学习笔记

  1. class ComputeLossOTA:
  2.     # Compute losses
  3.     def __init__(self, model, autobalance=False):
  4.         super(ComputeLossOTA, self).__init__()
  5.         device = next(model.parameters()).device  # get model device
  6.         h = model.hyp  # hyperparameters
  7.  
  8.         # Define criteria
  9.         BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
  10.         BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
  11.  
  12.         # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
  13.         self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0))  # positive, negative BCE targets, 1.0, 0.0
  14.         
  15.         # Focal loss
  16.         g = h['fl_gamma']  # focal loss gamma, 0.0
  17.         if g > 0:
  18.             BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
  19.  
  20.         det = model.module.model[-1] if is_parallel(model) else model.model[-1]  # Detect() module
  21.         self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02])  # P3-P7, [4.0, 1.0, 0.4]
  22.         self.ssi = list(det.stride).index(16) if autobalance else 0  # stride 16 index, 0
  23.         self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
  24.         for k in 'na', 'nc', 'nl', 'anchors', 'stride':
  25.             setattr(self, k, getattr(det, k))
  26.  
  27.     def __call__(self, p, targets, imgs):  # predictions, targets, model
  28.         import sys
  29.         device = targets.device
  30.         lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
  31.         bs, as_, gjs, gis, targets, anchors = self.build_targets(p, targets, imgs)  # 返回匹配到的image索引, anchor索引, gj, gi, GT, anchor
  32.         # pre_gen_gains=[tensor([80, 80, 80, 80], device='cuda:0'), tensor([40, 40, 40, 40], device='cuda:0'), tensor([20, 20, 20, 20], device='cuda:0')]
  33.         pre_gen_gains = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p]
  34.  
  35.         # Losses
  36.         for i, pi in enumerate(p):  # layer index, layer predictions
  37.             b, a, gj, gi = bs[i], as_[i], gjs[i], gis[i]  # image, anchor, gridy, gridx
  38.             tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj, tobj.shape=torch.Size([bs, 3, 80, 80]) or torch.Size([bs, 3, 40, 40]) or torch.Size([bs, 3, 20, 20])
  39.  
  40.             n = b.shape[0]  # number of targets
  41.             if n:
  42.                 ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets, 设匹配到GT的正样本数为p, ps.shape=torch.Size([p, 85])
  43.  
  44.                 # Regression
  45.                 grid = torch.stack([gi, gj], dim=1)  # grid.shape=torch.Size([p, 2])
  46.                 pxy = ps[:, :2].sigmoid() * 2. - 0.5  # pxy.shape=torch.Size([p, 2])
  47.                 pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]  # pwh.shape=torch.Size([p, 2])
  48.                 pbox = torch.cat((pxy, pwh), 1)  # predicted box, pbox.shape=torch.Size([p, 4])
  49.                 selected_tbox = targets[i][:, 2:6] * pre_gen_gains[i]  # selected_tbox.shape=torch.Size([p, 4])
  50.                 selected_tbox[:, :2] -= grid  # 将选中的tbox减去网格坐标，得到偏移量
  51.                 iou = bbox_iou(pbox.T, selected_tbox, x1y1x2y2=False, CIoU=True)  # iou(prediction, target)
  52.                 lbox += (1.0 - iou).mean()  # iou loss
  53.  
  54.                 # Objectness
  55.                 tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype)  # iou ratio, tobj[b, a, gj, gi]=iou.detach().clamp(0).type(tobj.dtype)
  56.  
  57.                 # Classification
  58.                 selected_tcls = targets[i][:, 1].long()
  59.                 if self.nc > 1:  # cls loss (only if multiple classes)
  60.                     t = torch.full_like(ps[:, 5:], self.cn, device=device)  # targets
  61.                     t[range(n), selected_tcls] = self.cp  # t相当于one-hot编码, 里面的元素仅在所属类别那一列为1, 其余为0 
  62.                     lcls += self.BCEcls(ps[:, 5:], t)  # BCE
  63.  
  64.             obji = self.BCEobj(pi[..., 4], tobj)
  65.             lobj += obji * self.balance[i]  # obj loss, self.balance=[4.0, 1.0, 0.4]
  66.             if self.autobalance:
  67.                 self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
  68.  
  69.         if self.autobalance:
  70.             self.balance = [x / self.balance[self.ssi] for x in self.balance]
  71.         lbox *= self.hyp['box']  # self.hyp['box']=0.05
  72.         lobj *= self.hyp['obj']  # self.hyp['obj']=0.7
  73.         lcls *= self.hyp['cls']  # self.hyp['cls']=0.3
  74.         bs = tobj.shape[0]  # batch size
  75.  
  76.         loss = lbox + lobj + lcls
  77.         return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
  78.  
  79.     def build_targets(self, p, targets, imgs):
  80.         #indices, anch = self.find_positive(p, targets)
  81.         indices, anch = self.find_3_positive(p, targets)  # 正样本
  82.         #indices, anch = self.find_4_positive(p, targets)
  83.         #indices, anch = self.find_5_positive(p, targets)
  84.         #indices, anch = self.find_9_positive(p, targets)
  85.         device = torch.device(targets.device)
  86.         matching_bs = [[] for pp in p]
  87.         matching_as = [[] for pp in p]
  88.         matching_gjs = [[] for pp in p]
  89.         matching_gis = [[] for pp in p]
  90.         matching_targets = [[] for pp in p]
  91.         matching_anchs = [[] for pp in p]
  92.         
  93.         nl = len(p)    
  94.     
  95.         for batch_idx in range(p[0].shape[0]):  # 遍历batch size中的每一张图片
  96.         
  97.             b_idx = targets[:, 0]==batch_idx  # 找出targets中与batch_idx相等的索引
  98.             this_target = targets[b_idx]  # 根据索引找出对应的GT
  99.             if this_target.shape[0] == 0:  # 如果GT数量为0，则处理下一张图片
  100.                 continue
  101.                 
  102.             txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]  # 将GT的坐标由0~1映射到与输入图片大小匹配的数值
  103.             txyxy = xywh2xyxy(txywh)  # 将坐标由[cx,cy,w,h]转换为[x1,y1,x2,y2](左上角及右下角坐标)
  104.  
  105.             pxyxys = []
  106.             p_cls = []
  107.             p_obj = []
  108.             from_which_layer = []
  109.             all_b = []
  110.             all_a = []
  111.             all_gj = []
  112.             all_gi = []
  113.             all_anch = []
  114.             
  115.             for i, pi in enumerate(p):
  116.                 
  117.                 b, a, gj, gi = indices[i]  # image, anchor, grid indices, 对gj、gi进行截断，不能超出特征图的范围
  118.                 idx = (b == batch_idx)  # 从b中找出与batch_idx相等的目标
  119.                 b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]
  120.                 all_b.append(b)
  121.                 all_a.append(a)
  122.                 all_gj.append(gj)
  123.                 all_gi.append(gi)
  124.                 all_anch.append(anch[i][idx])
  125.                 from_which_layer.append((torch.ones(size=(len(b),)) * i).to(device))
  126.                 
  127.                 fg_pred = pi[b, a, gj, gi]  # 选出相应的预测值，假设数量为n
  128.                 p_obj.append(fg_pred[:, 4:5])  # obj预测值
  129.                 p_cls.append(fg_pred[:, 5:])  # cls预测值
  130.                 
  131.                 grid = torch.stack([gi, gj], dim=1)
  132.                 pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i]  # 预测的cx、cy
  133.                 pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i]  # 预测的w、h
  134.                 pxywh = torch.cat([pxy, pwh], dim=-1)  # 预测的cx、cy、w、h
  135.                 pxyxy = xywh2xyxy(pxywh)  # 将cx、cy、w、h转换为x1、y1、x2、y2
  136.                 pxyxys.append(pxyxy)
  137.             
  138.             pxyxys = torch.cat(pxyxys, dim=0)  # pxyxys.shape=torch.Size([n, 4])
  139.             if pxyxys.shape[0] == 0:
  140.                 continue
  141.             p_obj = torch.cat(p_obj, dim=0)  # p_obj.shape=torch.Size([n, 1])
  142.             p_cls = torch.cat(p_cls, dim=0)  # p_cls.shape=torch.Size([n, 80])
  143.             from_which_layer = torch.cat(from_which_layer, dim=0)
  144.             all_b = torch.cat(all_b, dim=0)  # torch.Size([n])
  145.             all_a = torch.cat(all_a, dim=0)  # torch.Size([n])
  146.             all_gj = torch.cat(all_gj, dim=0)  # torch.Size([n])
  147.             all_gi = torch.cat(all_gi, dim=0)  # torch.Size([n])
  148.             all_anch = torch.cat(all_anch, dim=0)  # torch.Size([n, 2])
  149.         
  150.             pair_wise_iou = box_iou(txyxy, pxyxys)  # 计算GT与预测边界框之间的iou
  151.  
  152.             pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)  # iou loss
  153.  
  154.             top_k, _ = torch.topk(pair_wise_iou, min(10, pair_wise_iou.shape[1]), dim=1)  # 从大到小对iou进行排序，取前10个iou
  155.             dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)  # 对topk进行求和、取整，将该数值作为一个GT需要匹配到的正样本数，匹配的正样本数不能小于1
  156.  
  157.             gt_cls_per_image = (
  158.                 F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
  159.                 .float()
  160.                 .unsqueeze(1)
  161.                 .repeat(1, pxyxys.shape[0], 1)
  162.             )  # 将GT类别转换成one-hot编码
  163.  
  164.             num_gt = this_target.shape[0]  # GT数量, 假设为t
  165.             cls_preds_ = (
  166.                 p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
  167.                 * p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
  168.             )  # 各类别的预测分数
  169.  
  170.             y = cls_preds_.sqrt_()
  171.             pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
  172.                torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
  173.             ).sum(-1)  # cls loss
  174.             del cls_preds_
  175.         
  176.             cost = (
  177.                 pair_wise_cls_loss
  178.                 + 3.0 * pair_wise_iou_loss
  179.             )
  180.  
  181.             matching_matrix = torch.zeros_like(cost, device=device)  # torch.Size([t, n])
  182.  
  183.             for gt_idx in range(num_gt):
  184.                 _, pos_idx = torch.topk(
  185.                     cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
  186.                 )  # 对每个GT的loss由小到大排序，取排在前dynamic_ks个数值的索引
  187.                 matching_matrix[gt_idx][pos_idx] = 1.0  # 按照索引，给matching_matrix的相应元素置1
  188.  
  189.             del top_k, dynamic_ks
  190.             anchor_matching_gt = matching_matrix.sum(0)  # torch.Size([n]), 对所有正样本匹配的GT数进行求和
  191.             if (anchor_matching_gt > 1).sum() > 0:  # 如果大于0, 则认为一个正样本匹配多个GT
  192.                 _, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)  # 如果同一个正样本匹配到的GT数量大于1,则比较多个GT，取cost小作为正样本，其他的舍去
  193.                 matching_matrix[:, anchor_matching_gt > 1] *= 0.0  # 首先将大于1的列的元素置0
  194.                 matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0  # 再根据cost_argmin，将对应位置的元素置1
  195.             fg_mask_inboxes = (matching_matrix.sum(0) > 0.0).to(device)  # 找出与GT成功匹配的正样本
  196.             matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)  # 根据fg_mask_inboxes找出符合条件的列，再求每列的argmax。这步操作目的在于找出与GT匹配的正样本的索引值
  197.         
  198.             from_which_layer = from_which_layer[fg_mask_inboxes]
  199.             all_b = all_b[fg_mask_inboxes]  # 为匹配到正样本对应的图像索引
  200.             all_a = all_a[fg_mask_inboxes]  # 为匹配到正样本对应的anchor索引
  201.             all_gj = all_gj[fg_mask_inboxes]  # 为匹配到正样本对应的gj
  202.             all_gi = all_gi[fg_mask_inboxes]  # 为匹配到正样本对应的gi
  203.             all_anch = all_anch[fg_mask_inboxes]  # 为匹配到正样本对应的anchor
  204.  
  205.             this_target = this_target[matched_gt_inds]
  206.         
  207.             for i in range(nl):
  208.                 layer_idx = from_which_layer == i
  209.                 matching_bs[i].append(all_b[layer_idx])
  210.                 matching_as[i].append(all_a[layer_idx])
  211.                 matching_gjs[i].append(all_gj[layer_idx])
  212.                 matching_gis[i].append(all_gi[layer_idx])
  213.                 matching_targets[i].append(this_target[layer_idx])
  214.                 matching_anchs[i].append(all_anch[layer_idx])
  215.  
  216.         for i in range(nl):
  217.             if matching_targets[i] != []:
  218.                 matching_bs[i] = torch.cat(matching_bs[i], dim=0)
  219.                 matching_as[i] = torch.cat(matching_as[i], dim=0)
  220.                 matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
  221.                 matching_gis[i] = torch.cat(matching_gis[i], dim=0)
  222.                 matching_targets[i] = torch.cat(matching_targets[i], dim=0)
  223.                 matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
  224.             else:
  225.                 matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
  226.                 matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
  227.                 matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
  228.                 matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
  229.                 matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
  230.                 matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
  231.  
  232.         return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs           
  233.  
  234.     def find_3_positive(self, p, targets):
  235.         # p为预测值(p[0].shape=torch.Size([8, 3, 80, 80, 85]), p[1].shape=torch.Size([8, 3, 40, 40, 85]), p[2].shape=torch.Size([8, 3, 20, 20, 85])), targets=(image,class,x,y,w,h)
  236.         na, nt = self.na, targets.shape[0]  # anchors的数量, GT的数量; na=3, nt=n
  237.         indices, anch = [], []
  238.         gain = torch.ones(7, device=targets.device).long()  # normalized to gridspace gain, gain=tensor([1, 1, 1, 1, 1, 1, 1])
  239.         ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt), ai.shape=torch.Size([3, n])
  240.         targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)  # append anchor indices, targets.shape=torch.Size([3, n, 7])
  241.  
  242.         g = 0.5  # bias
  243.         off = torch.tensor([[0, 0],
  244.                             [1, 0], [0, 1], [-1, 0], [0, -1],  # j,k,l,m
  245.                             ], device=targets.device).float() * g  # offsets, [0,0]为中间网格，[1, 0]、[0, 1]、[-1, 0]、[0, -1]为相邻右、下、左、上的网格
  246.                 
  247.         for i in range(self.nl):  # self.nl=3
  248.             anchors = self.anchors[i]
  249.             '''
  250.             anchors=tensor([[1.50000, 2.00000],
  251.                             [2.37500, 4.50000],
  252.                             [5.00000, 3.50000]], device='cuda:0')
  253.             '''
  254.             gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]]  # xyxy gain, gain[2:6]=tensor([80, 80, 80, 80]) or tensor([40, 40, 40, 40]) or tensor([20, 20, 20, 20])
  255.             # gain = tensor([1, 1, 80, 80, 80, 80, 1]) or tensor([1, 1, 40, 40, 40, 40, 1]) or tensor([1, 1, 20, 20, 20, 20, 1])
  256.  
  257.             # Match targets to anchors
  258.             t = targets * gain  # 将targets里的坐标从0~1映射到与特征图大小匹配的坐标
  259.             if nt:
  260.                 # Matches
  261.                 r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
  262.                 j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t']  # compare
  263.                 t = t[j]  # filter, 去掉大于self.hyp['anchor_t']的GT, 设余下数量为m
  264.  
  265.                 # Offsets
  266.                 gxy = t[:, 2:4]  # grid xy, GT中心点坐标(以左上角为参考点)
  267.                 gxi = gain[[2, 3]] - gxy  # inverse, GT中心点坐标(以右下角为参考点)
  268.                 j, k = ((gxy % 1. < g) & (gxy > 1.)).T  # 对gxy取余，也就是把坐标的小数提出来与0.5做对比，对坐标做近似值处理
  269.                 l, m = ((gxi % 1. < g) & (gxi > 1.)).T  # 原理与上面的一致，l为横坐标，m为纵坐标
  270.                 j = torch.stack((torch.ones_like(j), j, k, l, m))  # j.shape=torch.Size([5, m])
  271.                 '''
  272.                 复制5个t, 选其中的3个, 第一个torch.ones_like(j)必选, 在剩下的4个相邻网格中选2个(j、l互斥, 点只能落在左右两边的其中一边; k、m互斥; 点只能落在上下两边的其中一边),
  273.                 因此总共选择了3个网格作为正样本的中心点。每个layer分配了3个anchor, 理论上一个GT最多可以匹配9个正样本。YOLOV7的输出有3个layer, 则一个GT最多可以匹配27个正样本。
  274.                 '''
  275.                 t = t.repeat((5, 1, 1))[j]  
  276.                 offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
  277.             else:
  278.                 t = targets[0]
  279.                 offsets = 0
  280.  
  281.             # Define
  282.             b, c = t[:, :2].long().T  # image, class
  283.             gxy = t[:, 2:4]  # grid xy
  284.             gwh = t[:, 4:6]  # grid wh
  285.             gij = (gxy - offsets).long()  # 坐标减去偏移量
  286.             gi, gj = gij.T  # grid xy indices
  287.  
  288.             # Append
  289.             a = t[:, 6].long()  # anchor indices
  290.             indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))  # image, anchor, grid indices, 对gj、gi进行截断，不能超出特征图的范围
  291.             anch.append(anchors[a])  # anchors
  292.  
  293.         return indices, anch

假设图中蓝色的点为GT的中心点，则YOLOV7中的ComputeLossOTA会把3个黄色的框视为正样本(对应着ComputeLossOTA类中的find_3_positive函数)，而ComputeLossAuxOTA会把黄色框以及橙色框，总共5个框视为正样本(对应着ComputeLossAuxOTA类中的find_5_positive函数)。

参考：

https://zhuanlan.zhihu.com/p/543160484

声明：本文内容由网友自发贡献，转载请注明出处：【wpsshop博客】
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号