开始讲解之前推荐一下我的专栏本专栏的内容支持(分类、检测、分割、追踪、关键点检测),专栏目前为限时折扣欢迎大家订阅本专栏本专栏每周更新5-7篇最新机制更有包含我所有改进的文件和交流群提供给大家本人定期在群内分享发表论文方法和经验。一、本文介绍本文给大家带来的改进机制是损失函数的改进机制标题虽然提到了ShapeIoU和InnnerShapeIoU但是本文的内容包括过去到现在的百分之九十以上的损失函数的实现同时使用方法非常简单在本文的末尾还会教大家在改进模型时何时添加损失函数才能达到最好的效果同时在开始讲解之前推荐一下我的专栏本专栏的内容支持(分类、检测、分割、追踪、关键点检测),专栏目前为限时折扣欢迎大家订阅本专栏本专栏每周更新3-5篇最新机制更有包含我所有改进的文件和交流群提供给大家本文支持的损失函数共有如下图片所示欢迎大家订阅我的专栏一起学习YOLO专栏链接YOLOv26有效涨点专栏包含Conv、注意力机制、主干/Backbone、损失函数、优化器、后处理等改进机制目录一、本文介绍二、ShapeIoU三、核心代码四、 损失函数使用方式4.1 修改一4.2 修改二4.3 步骤三未必要修改4.4 步骤四五、总结二、ShapeIoU官方论文地址官方论文地址官方代码地址官方代码地址这幅图展示了在目标检测任务中两种不同情况或方法下的边界框回归的对比。GT (Ground Truth): 用桃色框表示指的是图像中物体实际的位置和形状。在目标检测中算法试图尽可能准确地预测这个框。Anchor: 蓝色框代表一个预定义的框是算法预设的一系列框用于与GT框进行匹配寻找最佳的候选框。在图中我们看到四个不同的情况A、B、C、D每个都显示了一个anchor与GT的对比并给出了IoU交并比的数值。IoU是一个常用的度量用来评估预测边界框与真实边界框之间的重叠程度。论文中给了一堆公式大家有兴趣的可以看看。三、核心代码下面的代码的使用方式看章节四。def bbox_iou( box1: torch.Tensor, box2: torch.Tensor, xywh: bool True, GIoU: bool False, DIoU: bool False, CIoU: bool False, eps: float 1e-7, *, SIoU: bool False, EIoU: bool False, WIoU: bool False, ShapeIoU: bool False, Focal: bool False, Inner: bool False, ratio: float 0.7, alpha: float 1.0, gamma: float 0.5, scale: bool False, shape_scale: float 0.0, ) - Union[torch.Tensor, Tuple[torch.Tensor, ...]]: Calculate IoU / GIoU / DIoU / CIoU / EIoU / SIoU / WIoU / ShapeIoU between bounding boxes. Args: box1 (torch.Tensor): Boxes with last dimension 4. box2 (torch.Tensor): Boxes with last dimension 4. xywh (bool): If True, boxes are in (x, y, w, h) format. If False, boxes are in (x1, y1, x2, y2) format. GIoU (bool): Enable Generalized IoU. DIoU (bool): Enable Distance IoU. CIoU (bool): Enable Complete IoU. SIoU (bool): Enable SIoU. EIoU (bool): Enable EIoU. WIoU (bool): Enable WIoU. ShapeIoU (bool): Enable ShapeIoU. Focal (bool): Enable Focal-IoU style output. Inner (bool): Enable Inner-IoU. ratio (float): Inner box scale ratio. alpha (float): Power factor used by some extended losses. gamma (float): Focal exponent. scale (bool): Enable WIoU v2 / v3 dynamic scaling. shape_scale (float): ShapeIoU shape scale parameter. eps (float): Small value to avoid division by zero. Returns: torch.Tensor or Tuple[torch.Tensor, ...] # ------------------------------------------------------------ # 1. Get coordinates of bounding boxes # ------------------------------------------------------------ if xywh: # Input format: x, y, w, h (x1, y1, w1, h1), (x2, y2, w2, h2) box1.chunk(4, -1), box2.chunk(4, -1) w1 w1.clamp(mineps) h1 h1.clamp(mineps) w2 w2.clamp(mineps) h2 h2.clamp(mineps) else: # Input format: x1, y1, x2, y2 b1_x1_ori, b1_y1_ori, b1_x2_ori, b1_y2_ori box1.chunk(4, -1) b2_x1_ori, b2_y1_ori, b2_x2_ori, b2_y2_ori box2.chunk(4, -1) w1 (b1_x2_ori - b1_x1_ori).clamp(mineps) h1 (b1_y2_ori - b1_y1_ori).clamp(mineps) w2 (b2_x2_ori - b2_x1_ori).clamp(mineps) h2 (b2_y2_ori - b2_y1_ori).clamp(mineps) x1 (b1_x1_ori b1_x2_ori) / 2 y1 (b1_y1_ori b1_y2_ori) / 2 x2 (b2_x1_ori b2_x2_ori) / 2 y2 (b2_y1_ori b2_y2_ori) / 2 # ------------------------------------------------------------ # 2. Normal box or Inner box # ------------------------------------------------------------ if Inner: # Inner-IoU: # 用 ratio 缩小 box只在缩小后的 inner box 上计算 IoU。 # ratio0.7 表示使用原框中心区域的 70% 宽高。 r torch.as_tensor(ratio, dtypebox1.dtype, devicebox1.device).clamp(mineps) w1_half, h1_half (w1 * r) / 2, (h1 * r) / 2 w2_half, h2_half (w2 * r) / 2, (h2 * r) / 2 b1_x1, b1_x2 x1 - w1_half, x1 w1_half b1_y1, b1_y2 y1 - h1_half, y1 h1_half b2_x1, b2_x2 x2 - w2_half, x2 w2_half b2_y1, b2_y2 y2 - h2_half, y2 h2_half area1 w1 * h1 * r * r area2 w2 * h2 * r * r else: # Normal IoU box w1_half, h1_half w1 / 2, h1 / 2 w2_half, h2_half w2 / 2, h2 / 2 b1_x1, b1_x2 x1 - w1_half, x1 w1_half b1_y1, b1_y2 y1 - h1_half, y1 h1_half b2_x1, b2_x2 x2 - w2_half, x2 w2_half b2_y1, b2_y2 y2 - h2_half, y2 h2_half area1 w1 * h1 area2 w2 * h2 # ------------------------------------------------------------ # 3. Intersection and union # ------------------------------------------------------------ inter ( (b1_x2.minimum(b2_x2) - b1_x1.maximum(b2_x1)).clamp(min0) * (b1_y2.minimum(b2_y2) - b1_y1.maximum(b2_y1)).clamp(min0) ) union area1 area2 - inter eps iou inter / union iou torch.nan_to_num(iou, nan0.0, posinf1.0, neginf0.0) iou iou.clamp(min0.0, max1.0) # Focal weight focal_weight iou.clamp(min0, max1).pow(gamma) # ------------------------------------------------------------ # 4. Extended IoU losses # ------------------------------------------------------------ if CIoU or DIoU or GIoU or EIoU or SIoU or WIoU or ShapeIoU: cw b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1) ch b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1) if CIoU or DIoU or EIoU or SIoU or WIoU or ShapeIoU: c2 cw.pow(2) ch.pow(2) eps rho2 ( (b2_x1 b2_x2 - b1_x1 - b1_x2).pow(2) (b2_y1 b2_y2 - b1_y1 - b1_y2).pow(2) ) / 4 # CIoU if CIoU: v (4 / math.pi**2) * ( (w2 / (h2 eps)).atan() - (w1 / (h1 eps)).atan() ).pow(2) with torch.no_grad(): ciou_alpha v / (v - iou (1 eps)) out iou - (rho2 / c2 v * ciou_alpha) return (out, focal_weight) if Focal else out # EIoU if EIoU: rho_w2 ((b2_x2 - b2_x1) - (b1_x2 - b1_x1)).pow(2) rho_h2 ((b2_y2 - b2_y1) - (b1_y2 - b1_y1)).pow(2) cw2 (cw.pow(2) eps).pow(alpha) ch2 (ch.pow(2) eps).pow(alpha) out iou - (rho2 / c2 rho_w2 / cw2 rho_h2 / ch2) return (out, focal_weight) if Focal else out # SIoU if SIoU: s_cw (b2_x1 b2_x2 - b1_x1 - b1_x2) * 0.5 eps s_ch (b2_y1 b2_y2 - b1_y1 - b1_y2) * 0.5 eps sigma (s_cw.pow(2) s_ch.pow(2)).sqrt().clamp(mineps) sin_alpha_1 s_cw.abs() / sigma sin_alpha_2 s_ch.abs() / sigma threshold math.sqrt(2) / 2 sin_alpha torch.where( sin_alpha_1 threshold, sin_alpha_2, sin_alpha_1, ) sin_alpha sin_alpha.clamp(min-1 eps, max1 - eps) angle_cost torch.cos(torch.arcsin(sin_alpha) * 2 - math.pi / 2) rho_x (s_cw / (cw eps)).pow(2) rho_y (s_ch / (ch eps)).pow(2) distance_gamma angle_cost - 2 distance_cost ( 2 - torch.exp(distance_gamma * rho_x) - torch.exp(distance_gamma * rho_y) ) omega_w (w1 - w2).abs() / torch.max(w1, w2).clamp(mineps) omega_h (h1 - h2).abs() / torch.max(h1, h2).clamp(mineps) shape_cost ( (1 - torch.exp(-omega_w)).pow(4) (1 - torch.exp(-omega_h)).pow(4) ) out iou - (0.5 * (distance_cost shape_cost) eps).pow(alpha) return (out, focal_weight) if Focal else out # ShapeIoU if ShapeIoU: # Shape-Distance shape_scale_tensor torch.as_tensor( shape_scale, dtypebox1.dtype, devicebox1.device, ) w2_s w2.clamp(mineps).pow(shape_scale_tensor) h2_s h2.clamp(mineps).pow(shape_scale_tensor) shape_den (w2_s h2_s).clamp(mineps) ww 2 * w2_s / shape_den hh 2 * h2_s / shape_den cw_shape b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1) ch_shape b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1) c2_shape cw_shape.pow(2) ch_shape.pow(2) eps center_distance_x ( (b2_x1 b2_x2 - b1_x1 - b1_x2).pow(2) ) / 4 center_distance_y ( (b2_y1 b2_y2 - b1_y1 - b1_y2).pow(2) ) / 4 center_distance hh * center_distance_x ww * center_distance_y distance center_distance / c2_shape # Shape-Shape omega_w hh * (w1 - w2).abs() / torch.max(w1, w2).clamp(mineps) omega_h ww * (h1 - h2).abs() / torch.max(h1, h2).clamp(mineps) shape_cost ( (1 - torch.exp(-omega_w)).pow(4) (1 - torch.exp(-omega_h)).pow(4) ) out iou - distance - 0.5 * shape_cost return (out, focal_weight) if Focal else out # WIoU if WIoU: if Focal: raise RuntimeError(WIoU does not support FocalTrue.) distance_weight torch.exp(rho2 / c2) if scale: wiou_scale WIoU_Scale(1 - iou) return ( wiou_scale._scaled_loss(), (1 - iou) * distance_weight, iou, ) return iou, distance_weight # DIoU out iou - rho2 / c2 return (out, focal_weight) if Focal else out # GIoU c_area cw * ch eps out iou - (((c_area - union) / c_area) eps).pow(alpha) return (out, focal_weight) if Focal else out # Normal IoU return (iou, focal_weight) if Focal else iou四、 损失函数使用方式4.1 修改一第一步我们需要找到如下的文件ultralytics/utils/metrics.py,找到如下的代码下面的图片是原先的代码部分截图的正常样子然后我们将整个代码块一将下面的整个方法(这里这是部分截图)内容全部替换4.2 修改二第二步我们找到另一个文件如下-ultralytics/utils/loss.py我们按图进行修改.iou bbox_iou( pred_bboxes[fg_mask], target_bboxes[fg_mask], xywhFalse, GIoUFalse, DIoUFalse, CIoUTrue, EIoUFalse, SIoUFalse, WIoUFalse, ShapeIoUFalse, FocalFalse, InnerFalse, ratio0.7, alpha1.0, gamma0.5, scaleFalse, shape_scale0.0, eps1e-7) # 默认使用 CIoUGIoU / DIoU / CIoU / EIoU / SIoU / WIoU / ShapeIoU 建议一次只开一个。 # InnerFalse 表示不启用 Inner-IoU若使用 Inner-CIoU只需设置 InnerTrueratio 控制内框比例。 # ShapeIoUFalse 表示不启用 ShapeIoU若使用 ShapeIoU将 CIoUFalse, ShapeIoUTrue。 # Focal 和 WIoU 会改变返回值结构若开启需要同步修改后续 loss 计算。4.3 步骤三未必要修改我们找到另一个文件如下-ultralytics/utils/loss.py步骤二的下一行此处是使用Focus和WIoU时候需要修改的代码你不使用跳过此步骤直接进行步骤三即可if type(iou) is tuple: if len(iou) 2: # increased the weight of low/high IoU loss_iou ((1 - iou[1].detach().squeeze()) * (1 - iou[0].squeeze()) * weight).sum() / target_scores_sum # Focal # lbox (iou[1].detach().squeeze() * (1 - iou[0].squeeze())* weight).sum() / target_scores_sum # Focal-inv # 这里有两种方法大家可以自行尝试,这里的Focal-inv也是文章中提出的. else: loss_iou (iou[0] * iou[1] * weight).sum() / target_scores_sum else: loss_iou ((1.0 - iou) * weight).sum() / target_scores_sum # iou loss4.4 步骤四我们还需要修改一处找到如下的文件ultralytics/utils/tal.py然后找到其中下面图片的代码用我给的代码替换红框内的代码。return bbox_iou(gt_bboxes, pd_bboxes, xywhFalse, GIoUFalse, DIoUFalse, CIoUTrue, EIoUFalse, SIoUFalse, WIoUFalse, ShapeIoUFalse, FocalFalse, InnerTrue, ratio0.7, alpha1.0, gamma0.5, eps1e-7, scaleFalse, shape_scale0.0).squeeze(-1).clamp_(0)此处和loss.py里面的最好是使用同一个参数(但非必须)但Focus和WIoU此处不能使用需要注意.五、总结到此本文的正式分享内容就结束了在这里给大家推荐我的YOLOv26改进有效涨点专栏本专栏目前为新开的平均质量分98分后期我会根据各种最新的前沿顶会进行论文复现也会对一些老的改进机制进行补充目前本专栏免费阅读(暂时大家尽早关注不迷路~)如果大家觉得本文帮助到你了订阅本专栏关注后续更多的更新~专栏链接YOLOv26有效涨点专栏包含Conv、注意力机制、主干/Backbone、损失函数、优化器、后处理等改进机制
yolov26改进 | 损失函数改进篇 | 最新ShapeIoU、InnerShapeIoU损失助力细节涨点（含三十余种损失函数改进方法）
开始讲解之前推荐一下我的专栏本专栏的内容支持(分类、检测、分割、追踪、关键点检测),专栏目前为限时折扣欢迎大家订阅本专栏本专栏每周更新5-7篇最新机制更有包含我所有改进的文件和交流群提供给大家本人定期在群内分享发表论文方法和经验。一、本文介绍本文给大家带来的改进机制是损失函数的改进机制标题虽然提到了ShapeIoU和InnnerShapeIoU但是本文的内容包括过去到现在的百分之九十以上的损失函数的实现同时使用方法非常简单在本文的末尾还会教大家在改进模型时何时添加损失函数才能达到最好的效果同时在开始讲解之前推荐一下我的专栏本专栏的内容支持(分类、检测、分割、追踪、关键点检测),专栏目前为限时折扣欢迎大家订阅本专栏本专栏每周更新3-5篇最新机制更有包含我所有改进的文件和交流群提供给大家本文支持的损失函数共有如下图片所示欢迎大家订阅我的专栏一起学习YOLO专栏链接YOLOv26有效涨点专栏包含Conv、注意力机制、主干/Backbone、损失函数、优化器、后处理等改进机制目录一、本文介绍二、ShapeIoU三、核心代码四、 损失函数使用方式4.1 修改一4.2 修改二4.3 步骤三未必要修改4.4 步骤四五、总结二、ShapeIoU官方论文地址官方论文地址官方代码地址官方代码地址这幅图展示了在目标检测任务中两种不同情况或方法下的边界框回归的对比。GT (Ground Truth): 用桃色框表示指的是图像中物体实际的位置和形状。在目标检测中算法试图尽可能准确地预测这个框。Anchor: 蓝色框代表一个预定义的框是算法预设的一系列框用于与GT框进行匹配寻找最佳的候选框。在图中我们看到四个不同的情况A、B、C、D每个都显示了一个anchor与GT的对比并给出了IoU交并比的数值。IoU是一个常用的度量用来评估预测边界框与真实边界框之间的重叠程度。论文中给了一堆公式大家有兴趣的可以看看。三、核心代码下面的代码的使用方式看章节四。def bbox_iou( box1: torch.Tensor, box2: torch.Tensor, xywh: bool True, GIoU: bool False, DIoU: bool False, CIoU: bool False, eps: float 1e-7, *, SIoU: bool False, EIoU: bool False, WIoU: bool False, ShapeIoU: bool False, Focal: bool False, Inner: bool False, ratio: float 0.7, alpha: float 1.0, gamma: float 0.5, scale: bool False, shape_scale: float 0.0, ) - Union[torch.Tensor, Tuple[torch.Tensor, ...]]: Calculate IoU / GIoU / DIoU / CIoU / EIoU / SIoU / WIoU / ShapeIoU between bounding boxes. Args: box1 (torch.Tensor): Boxes with last dimension 4. box2 (torch.Tensor): Boxes with last dimension 4. xywh (bool): If True, boxes are in (x, y, w, h) format. If False, boxes are in (x1, y1, x2, y2) format. GIoU (bool): Enable Generalized IoU. DIoU (bool): Enable Distance IoU. CIoU (bool): Enable Complete IoU. SIoU (bool): Enable SIoU. EIoU (bool): Enable EIoU. WIoU (bool): Enable WIoU. ShapeIoU (bool): Enable ShapeIoU. Focal (bool): Enable Focal-IoU style output. Inner (bool): Enable Inner-IoU. ratio (float): Inner box scale ratio. alpha (float): Power factor used by some extended losses. gamma (float): Focal exponent. scale (bool): Enable WIoU v2 / v3 dynamic scaling. shape_scale (float): ShapeIoU shape scale parameter. eps (float): Small value to avoid division by zero. Returns: torch.Tensor or Tuple[torch.Tensor, ...] # ------------------------------------------------------------ # 1. Get coordinates of bounding boxes # ------------------------------------------------------------ if xywh: # Input format: x, y, w, h (x1, y1, w1, h1), (x2, y2, w2, h2) box1.chunk(4, -1), box2.chunk(4, -1) w1 w1.clamp(mineps) h1 h1.clamp(mineps) w2 w2.clamp(mineps) h2 h2.clamp(mineps) else: # Input format: x1, y1, x2, y2 b1_x1_ori, b1_y1_ori, b1_x2_ori, b1_y2_ori box1.chunk(4, -1) b2_x1_ori, b2_y1_ori, b2_x2_ori, b2_y2_ori box2.chunk(4, -1) w1 (b1_x2_ori - b1_x1_ori).clamp(mineps) h1 (b1_y2_ori - b1_y1_ori).clamp(mineps) w2 (b2_x2_ori - b2_x1_ori).clamp(mineps) h2 (b2_y2_ori - b2_y1_ori).clamp(mineps) x1 (b1_x1_ori b1_x2_ori) / 2 y1 (b1_y1_ori b1_y2_ori) / 2 x2 (b2_x1_ori b2_x2_ori) / 2 y2 (b2_y1_ori b2_y2_ori) / 2 # ------------------------------------------------------------ # 2. Normal box or Inner box # ------------------------------------------------------------ if Inner: # Inner-IoU: # 用 ratio 缩小 box只在缩小后的 inner box 上计算 IoU。 # ratio0.7 表示使用原框中心区域的 70% 宽高。 r torch.as_tensor(ratio, dtypebox1.dtype, devicebox1.device).clamp(mineps) w1_half, h1_half (w1 * r) / 2, (h1 * r) / 2 w2_half, h2_half (w2 * r) / 2, (h2 * r) / 2 b1_x1, b1_x2 x1 - w1_half, x1 w1_half b1_y1, b1_y2 y1 - h1_half, y1 h1_half b2_x1, b2_x2 x2 - w2_half, x2 w2_half b2_y1, b2_y2 y2 - h2_half, y2 h2_half area1 w1 * h1 * r * r area2 w2 * h2 * r * r else: # Normal IoU box w1_half, h1_half w1 / 2, h1 / 2 w2_half, h2_half w2 / 2, h2 / 2 b1_x1, b1_x2 x1 - w1_half, x1 w1_half b1_y1, b1_y2 y1 - h1_half, y1 h1_half b2_x1, b2_x2 x2 - w2_half, x2 w2_half b2_y1, b2_y2 y2 - h2_half, y2 h2_half area1 w1 * h1 area2 w2 * h2 # ------------------------------------------------------------ # 3. Intersection and union # ------------------------------------------------------------ inter ( (b1_x2.minimum(b2_x2) - b1_x1.maximum(b2_x1)).clamp(min0) * (b1_y2.minimum(b2_y2) - b1_y1.maximum(b2_y1)).clamp(min0) ) union area1 area2 - inter eps iou inter / union iou torch.nan_to_num(iou, nan0.0, posinf1.0, neginf0.0) iou iou.clamp(min0.0, max1.0) # Focal weight focal_weight iou.clamp(min0, max1).pow(gamma) # ------------------------------------------------------------ # 4. Extended IoU losses # ------------------------------------------------------------ if CIoU or DIoU or GIoU or EIoU or SIoU or WIoU or ShapeIoU: cw b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1) ch b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1) if CIoU or DIoU or EIoU or SIoU or WIoU or ShapeIoU: c2 cw.pow(2) ch.pow(2) eps rho2 ( (b2_x1 b2_x2 - b1_x1 - b1_x2).pow(2) (b2_y1 b2_y2 - b1_y1 - b1_y2).pow(2) ) / 4 # CIoU if CIoU: v (4 / math.pi**2) * ( (w2 / (h2 eps)).atan() - (w1 / (h1 eps)).atan() ).pow(2) with torch.no_grad(): ciou_alpha v / (v - iou (1 eps)) out iou - (rho2 / c2 v * ciou_alpha) return (out, focal_weight) if Focal else out # EIoU if EIoU: rho_w2 ((b2_x2 - b2_x1) - (b1_x2 - b1_x1)).pow(2) rho_h2 ((b2_y2 - b2_y1) - (b1_y2 - b1_y1)).pow(2) cw2 (cw.pow(2) eps).pow(alpha) ch2 (ch.pow(2) eps).pow(alpha) out iou - (rho2 / c2 rho_w2 / cw2 rho_h2 / ch2) return (out, focal_weight) if Focal else out # SIoU if SIoU: s_cw (b2_x1 b2_x2 - b1_x1 - b1_x2) * 0.5 eps s_ch (b2_y1 b2_y2 - b1_y1 - b1_y2) * 0.5 eps sigma (s_cw.pow(2) s_ch.pow(2)).sqrt().clamp(mineps) sin_alpha_1 s_cw.abs() / sigma sin_alpha_2 s_ch.abs() / sigma threshold math.sqrt(2) / 2 sin_alpha torch.where( sin_alpha_1 threshold, sin_alpha_2, sin_alpha_1, ) sin_alpha sin_alpha.clamp(min-1 eps, max1 - eps) angle_cost torch.cos(torch.arcsin(sin_alpha) * 2 - math.pi / 2) rho_x (s_cw / (cw eps)).pow(2) rho_y (s_ch / (ch eps)).pow(2) distance_gamma angle_cost - 2 distance_cost ( 2 - torch.exp(distance_gamma * rho_x) - torch.exp(distance_gamma * rho_y) ) omega_w (w1 - w2).abs() / torch.max(w1, w2).clamp(mineps) omega_h (h1 - h2).abs() / torch.max(h1, h2).clamp(mineps) shape_cost ( (1 - torch.exp(-omega_w)).pow(4) (1 - torch.exp(-omega_h)).pow(4) ) out iou - (0.5 * (distance_cost shape_cost) eps).pow(alpha) return (out, focal_weight) if Focal else out # ShapeIoU if ShapeIoU: # Shape-Distance shape_scale_tensor torch.as_tensor( shape_scale, dtypebox1.dtype, devicebox1.device, ) w2_s w2.clamp(mineps).pow(shape_scale_tensor) h2_s h2.clamp(mineps).pow(shape_scale_tensor) shape_den (w2_s h2_s).clamp(mineps) ww 2 * w2_s / shape_den hh 2 * h2_s / shape_den cw_shape b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1) ch_shape b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1) c2_shape cw_shape.pow(2) ch_shape.pow(2) eps center_distance_x ( (b2_x1 b2_x2 - b1_x1 - b1_x2).pow(2) ) / 4 center_distance_y ( (b2_y1 b2_y2 - b1_y1 - b1_y2).pow(2) ) / 4 center_distance hh * center_distance_x ww * center_distance_y distance center_distance / c2_shape # Shape-Shape omega_w hh * (w1 - w2).abs() / torch.max(w1, w2).clamp(mineps) omega_h ww * (h1 - h2).abs() / torch.max(h1, h2).clamp(mineps) shape_cost ( (1 - torch.exp(-omega_w)).pow(4) (1 - torch.exp(-omega_h)).pow(4) ) out iou - distance - 0.5 * shape_cost return (out, focal_weight) if Focal else out # WIoU if WIoU: if Focal: raise RuntimeError(WIoU does not support FocalTrue.) distance_weight torch.exp(rho2 / c2) if scale: wiou_scale WIoU_Scale(1 - iou) return ( wiou_scale._scaled_loss(), (1 - iou) * distance_weight, iou, ) return iou, distance_weight # DIoU out iou - rho2 / c2 return (out, focal_weight) if Focal else out # GIoU c_area cw * ch eps out iou - (((c_area - union) / c_area) eps).pow(alpha) return (out, focal_weight) if Focal else out # Normal IoU return (iou, focal_weight) if Focal else iou四、 损失函数使用方式4.1 修改一第一步我们需要找到如下的文件ultralytics/utils/metrics.py,找到如下的代码下面的图片是原先的代码部分截图的正常样子然后我们将整个代码块一将下面的整个方法(这里这是部分截图)内容全部替换4.2 修改二第二步我们找到另一个文件如下-ultralytics/utils/loss.py我们按图进行修改.iou bbox_iou( pred_bboxes[fg_mask], target_bboxes[fg_mask], xywhFalse, GIoUFalse, DIoUFalse, CIoUTrue, EIoUFalse, SIoUFalse, WIoUFalse, ShapeIoUFalse, FocalFalse, InnerFalse, ratio0.7, alpha1.0, gamma0.5, scaleFalse, shape_scale0.0, eps1e-7) # 默认使用 CIoUGIoU / DIoU / CIoU / EIoU / SIoU / WIoU / ShapeIoU 建议一次只开一个。 # InnerFalse 表示不启用 Inner-IoU若使用 Inner-CIoU只需设置 InnerTrueratio 控制内框比例。 # ShapeIoUFalse 表示不启用 ShapeIoU若使用 ShapeIoU将 CIoUFalse, ShapeIoUTrue。 # Focal 和 WIoU 会改变返回值结构若开启需要同步修改后续 loss 计算。4.3 步骤三未必要修改我们找到另一个文件如下-ultralytics/utils/loss.py步骤二的下一行此处是使用Focus和WIoU时候需要修改的代码你不使用跳过此步骤直接进行步骤三即可if type(iou) is tuple: if len(iou) 2: # increased the weight of low/high IoU loss_iou ((1 - iou[1].detach().squeeze()) * (1 - iou[0].squeeze()) * weight).sum() / target_scores_sum # Focal # lbox (iou[1].detach().squeeze() * (1 - iou[0].squeeze())* weight).sum() / target_scores_sum # Focal-inv # 这里有两种方法大家可以自行尝试,这里的Focal-inv也是文章中提出的. else: loss_iou (iou[0] * iou[1] * weight).sum() / target_scores_sum else: loss_iou ((1.0 - iou) * weight).sum() / target_scores_sum # iou loss4.4 步骤四我们还需要修改一处找到如下的文件ultralytics/utils/tal.py然后找到其中下面图片的代码用我给的代码替换红框内的代码。return bbox_iou(gt_bboxes, pd_bboxes, xywhFalse, GIoUFalse, DIoUFalse, CIoUTrue, EIoUFalse, SIoUFalse, WIoUFalse, ShapeIoUFalse, FocalFalse, InnerTrue, ratio0.7, alpha1.0, gamma0.5, eps1e-7, scaleFalse, shape_scale0.0).squeeze(-1).clamp_(0)此处和loss.py里面的最好是使用同一个参数(但非必须)但Focus和WIoU此处不能使用需要注意.五、总结到此本文的正式分享内容就结束了在这里给大家推荐我的YOLOv26改进有效涨点专栏本专栏目前为新开的平均质量分98分后期我会根据各种最新的前沿顶会进行论文复现也会对一些老的改进机制进行补充目前本专栏免费阅读(暂时大家尽早关注不迷路~)如果大家觉得本文帮助到你了订阅本专栏关注后续更多的更新~专栏链接YOLOv26有效涨点专栏包含Conv、注意力机制、主干/Backbone、损失函数、优化器、后处理等改进机制
