YOLO26 解耦检测头：分类回归分开做

class DecoupledHead(nn.Module):
    def __init__(self, nc=2, ch=[256, 512, 512], reg_max=16):
        super().__init__()
        self.nl = 3

        c2 = max((16, ch[0] // 4, reg_max * 4))
        c3 = max(ch[0], min(nc, 100))

        # Box 分支：回归
        self.cv2 = nn.ModuleList([
            nn.Sequential(Conv(ch[i], c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * reg_max, 1))
            for i in range(3)
        ])

        # Cls 分支：分类
        self.cv3 = nn.ModuleList([
            nn.Sequential(Conv(ch[i], c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, nc, 1))
            for i in range(3)
        ])

        self.dfl = DFL(reg_max)

    def forward(self, x):
        for i in range(self.nl):
            box = self.cv2[i](x[i])  # 各自独立计算
            cls = self.cv3[i](x[i])

            bs, _, h, w = box.shape
            # DFL 解码
            box = box.view(bs, 4, self.reg_max, h * w)
            box_decoded = self.dfl(box)

            decoded_boxes.append(box_decoded.transpose(1, 2))
            all_scores.append(cls.view(bs, self.nc, -1).permute(0, 2, 1))

        return cat([all_boxes, all_scores], dim=-1)

# backbone 配置
backbone_type: "yolo26m"

# 冻结策略
freeze_backbone: 0  # 或 2-4 层

# 学习率
lr0: 0.0005
weight_decay: 0.0005

# 数据增强
mosaic: 0.8
mixup: 0.1