本文介绍了Token Statistics Self-Attention（TSSA）机制，并将其集成到YOLO26中。传统自注意力计算复杂度高，TSSA进行了范式转变，基于token统计特征实现高效注意力交互。它通过“算法展开”推导得出，以“最大编码率降低”为目标，实现特征学习。TSSA包含动态分组和低秩投影优化两步创新，具备线性复杂度。我们将TSSA代码集成到YOLO26的C2PSA模块中。实验表明，改进后的YOLO26在目标检测任务中表现良好，验证了TSSA机制的有效性。

文章目录： YOLO26改进大全：卷积层、轻量化、注意力机制、损失函数、Backbone、SPPF、Neck、检测头全方位优化汇总

专栏链接: YOLO26改进专栏

介绍

摘要

注意力算子可以说是 Transformer 架构的关键特征，该架构在多种任务中都表现出了最先进的性能。然而，Transformer 的注意力算子通常会带来巨大的计算负担，其计算复杂度随 Token 数量呈二次方增长。在这项工作中，我们提出了一种新型的 Transformer 注意力算子，其计算复杂度随 Token 数量呈线性增长。我们将之前的研究成果进行了扩展，之前的研究表明，通过“白盒”架构设计可以自然地构建出 Transformer 风格的架构，即网络的每一层都被设计为实现最大编码率降低目标（$MCR^{2}$）的一个增量优化步骤。具体来说，我们推导了 $MCR^{2}$ 目标的一种新颖变分形式，并展示了基于该变分目标进行展开梯度下降所得到的架构，导出了一种新的注意力模块，称为 Token 统计自注意力（Token Statistics Self-Attention，TSSA）。TSSA 具有线性的计算和内存复杂度，并且与计算 Token 之间成对相似度的典型注意力架构截然不同。在视觉、语言和长序列任务上的实验表明，只需简单地用 TSSA 替换标准自注意力（我们将这种架构称为 Token 统计 Transformer，即 TOST），就能获得与传统 Transformer 相当的性能，同时计算效率更高且更具可解释性。我们的结果还在一定程度上质疑了“成对相似度风格的注意力机制是 Transformer 架构成功的关键”这一传统观念。代码将在 https://github.com/RobinWu218/ToST 开源。

文章链接

论文地址：论文地址

代码地址：代码地址

基本原理

TSSA（Token Statistics Self-Attention）的核心创新是彻底抛弃传统自注意力的“成对相似度计算”，转而基于token的统计特征实现高效注意力交互 ：

1. 从“逐对对比”到“统计聚合”的范式转变

传统自注意力需要计算所有token两两之间的相似度（如缩放点积），导致复杂度随token数量呈平方增长。TSSA跳出这一框架，认为注意力的本质是“基于数据关联的特征优化”，而这种关联无需逐对计算——只需捕捉token群体的统计规律（即“二阶矩”，可理解为token特征的分布集中程度），就能实现类似的特征聚合效果。

2. 基于“白盒设计”的目标导向优化

TSSA并非经验性设计，而是通过“算法展开”的白盒思路推导得出：以“最大编码率降低（MCR²）”为核心目标，先将该目标转化为更易计算的变分形式，再把优化过程拆分成网络的逐层操作。每一层的作用都是增量优化这个目标——让同一组内的token特征更集中（压缩），同时让所有token的整体特征更分散（扩展），最终实现 discriminative 特征学习。

3. 数据驱动的低秩投影与动态分组

TSSA的核心操作包含两步关键创新：

• 动态分组：通过计算token与不同子空间的匹配度，用软聚类（类似概率分配）将token分到K个组，无需人工定义分组规则，完全由数据自动决定。
• 低秩投影优化：对每个组，基于token特征的统计信息构建“重要性权重”，保留组内特征中“能量集中”（即多数token共同拥有）的方向，抑制冗余或噪声方向。这一过程不依赖任何成对相似度，仅通过矩阵投影和统计计算完成，天然具备线性复杂度。

YOLO26引入代码

在根目录下的ultralytics/nn/目录，新建一个C2PSA目录，然后新建一个以 C2PSA_TSSA为文件名的py文件， 把代码拷贝进去。

import torch
import torch.nn as nn
from einops import rearrange

class AttentionTSSA(nn.Module):
    def __init__(self, dim, num_heads = 8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()

        self.heads = num_heads
        self.dim = dim
        head_dim = dim // num_heads

        self.attend = nn.Softmax(dim = 1)
        self.attn_drop = nn.Dropout(attn_drop)

        self.qkv = nn.Linear(dim, dim, bias=qkv_bias)

        self.temp = nn.Parameter(torch.ones(num_heads, 1))

        self.to_out = nn.Sequential(
            nn.Linear(dim, dim),
            nn.Dropout(proj_drop)
        )

    def forward(self, x):
        # x: (B, C, H, W) - standard attention interface
        B, C, H, W = x.shape
        N = H * W
        x_flat = x.view(B, C, N).permute(0, 2, 1)  # (B, N, C)

        # Apply linear projection and reshape for multi-head
        w = self.qkv(x_flat)  # (B, N, C)
        w = w.view(B, N, self.heads, C // self.heads).permute(0, 2, 1, 3)  # (B, h, N, d)

        w_normed = torch.nn.functional.normalize(w, dim=-2)
        w_sq = w_normed ** 2

        # Pi from Eq. 10 in the paper
        Pi = self.attend(torch.sum(w_sq, dim=-1) * self.temp) # b * h * n

        dots = torch.matmul((Pi / (Pi.sum(dim=-1, keepdim=True) + 1e-8)).unsqueeze(-2), w ** 2)
        attn = 1. / (1 + dots)
        attn = self.attn_drop(attn)

        out = - torch.mul(w.mul(Pi.unsqueeze(-1)), attn)

        out = rearrange(out, 'b h n d -> b n (h d)')
        out = self.to_out(out)

        # Reshape back to (B, C, H, W)
        out = out.permute(0, 2, 1).view(B, C, H, W)

        return out

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'temp'}

class Attention(nn.Module):

    def __init__(self, dim: int, num_heads: int = 8, attn_ratio: float = 0.5):

        super().__init__()
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.key_dim = int(self.head_dim * attn_ratio)
        self.scale = self.key_dim**-0.5
        nh_kd = self.key_dim * num_heads
        h = dim + nh_kd * 2
        self.qkv = Conv(dim, h, 1, act=False)
        self.proj = Conv(dim, dim, 1, act=False)
        self.pe = Conv(dim, dim, 3, 1, g=dim, act=False)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        B, C, H, W = x.shape
        N = H * W
        qkv = self.qkv(x)
        q, k, v = qkv.view(B, self.num_heads, self.key_dim * 2 + self.head_dim, N).split(
            [self.key_dim, self.key_dim, self.head_dim], dim=2
        )

        attn = (q.transpose(-2, -1) @ k) * self.scale
        attn = attn.softmax(dim=-1)
        x = (v @ attn.transpose(-2, -1)).view(B, C, H, W) + self.pe(v.reshape(B, C, H, W))
        x = self.proj(x)
        return x

def autopad(k, p=None, d=1):
    """Pad to 'same' shape outputs."""
    if d > 1:
        k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k]
    if p is None:
        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
    return p

class Conv(nn.Module):
    default_act = nn.SiLU()

    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
        super().__init__()
        self.conv = nn.Conv2d(
            c1, c2, k, s, autopad(k, p, d), groups=g, dilation=d, bias=False
        )
        self.bn = nn.BatchNorm2d(c2)
        self.act = (
            self.default_act
            if act is True
            else act
            if isinstance(act, nn.Module)
            else nn.Identity()
        )

    def forward(self, x):
        c = self.conv(x)
        c = self.bn(c)
        c = self.act(c)
        return c

class PSABlock(nn.Module):
    def __init__(self, c: int, attn_ratio: float = 0.5, num_heads: int = 4, shortcut: bool = True) -> None:

        super().__init__()

        self.attn = Attention(c, attn_ratio=attn_ratio, num_heads=num_heads)
        self.ffn = nn.Sequential(Conv(c, c * 2, 1), Conv(c * 2, c, 1, act=False))
        self.add = shortcut

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = x + self.attn(x) if self.add else self.attn(x)
        x = x + self.ffn(x) if self.add else self.ffn(x)
        return x

class C2PSA(nn.Module):

    def __init__(self, c1, c2, n=1, e=0.5):

        super().__init__()
        assert c1 == c2
        self.c = int(c1 * e)
        self.cv1 = Conv(c1, 2 * self.c, 1, 1)
        self.cv2 = Conv(2 * self.c, c1, 1)

        self.m = nn.Sequential(*(PSABlock(self.c, attn_ratio=0.5, num_heads=self.c // 64) for _ in range(n)))

    def forward(self, x):

        a, b = self.cv1(x).split((self.c, self.c), dim=1)
        b = self.m(b)
        return self.cv2(torch.cat((a, b), 1))

class PSABlock_AttentionTSSA(PSABlock):
    def __init__(self, c: int, attn_ratio: float = 0.5, num_heads: int = 4, shortcut: bool = True) -> None:
        super().__init__(c, attn_ratio, num_heads, shortcut)

        self.attn = AttentionTSSA(c)

class C2PSA_TSSA(C2PSA):
    def __init__(self, c1: int, c2: int, n: int = 1, e: float = 0.5):
        super().__init__(c1, c2, n, e)

        self.m = nn.Sequential(*(PSABlock_AttentionTSSA(self.c, attn_ratio=0.5, num_heads=self.c // 64) for _ in range(n)))

注册

在ultralytics/nn/tasks.py中进行如下操作：

步骤1:

from ultralytics.nn.C2PSA.C2PSA_TSSA import C2PSA_TSSA

步骤2

修改def parse_model(d, ch, verbose=True):

C2PSA_TSSA

配置yolo26-C2PSA_TSSA.yaml

ultralytics/cfg/models/26/yolo26-C2PSA_TSSA.yaml

# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license

# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
# Model docs: https://docs.ultralytics.com/models/yolo26
# Task docs: https://docs.ultralytics.com/tasks/detect

# Parameters
nc: 80 # number of classes
end2end: True # whether to use end-to-end mode
reg_max: 1 # DFL bins
scales: # model compound scaling constants, i.e. 'model=yolo26n.yaml' will call yolo26.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.50, 0.25, 1024] # summary: 260 layers, 2,572,280 parameters, 2,572,280 gradients, 6.1 GFLOPs
  s: [0.50, 0.50, 1024] # summary: 260 layers, 10,009,784 parameters, 10,009,784 gradients, 22.8 GFLOPs
  m: [0.50, 1.00, 512] # summary: 280 layers, 21,896,248 parameters, 21,896,248 gradients, 75.4 GFLOPs
  l: [1.00, 1.00, 512] # summary: 392 layers, 26,299,704 parameters, 26,299,704 gradients, 93.8 GFLOPs
  x: [1.00, 1.50, 512] # summary: 392 layers, 58,993,368 parameters, 58,993,368 gradients, 209.5 GFLOPs

# YOLO26n backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
  - [-1, 2, C3k2, [256, False, 0.25]]
  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  - [-1, 2, C3k2, [512, False, 0.25]]
  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  - [-1, 2, C3k2, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  - [-1, 2, C3k2, [1024, True]]
  - [-1, 1, SPPF, [1024, 5, 3, True]] # 9
  - [-1, 2, C2PSA_TSSA, [1024]] # 10

# YOLO26n head
head:
  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
  - [-1, 2, C3k2, [512, True]] # 13

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
  - [-1, 2, C3k2, [256, True]] # 16 (P3/8-small)

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 13], 1, Concat, [1]] # cat head P4
  - [-1, 2, C3k2, [512, True]] # 19 (P4/16-medium)

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 10], 1, Concat, [1]] # cat head P5
  - [-1, 1, C3k2, [1024, True, 0.5, True]] # 22 (P5/32-large)

  - [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5)

实验

脚本

import warnings
warnings.filterwarnings('ignore')
from ultralytics import YOLO

if __name__ == '__main__':
#     修改为自己的配置文件地址
    model = YOLO('./ultralytics/cfg/models/26/yolo26-C2PSA_TSSA.yaml')
#     修改为自己的数据集地址
    model.train(data='./ultralytics/cfg/datasets/coco8.yaml',
                cache=False,
                imgsz=640,
                epochs=10,
                single_cls=False,  # 是否是单类别检测
                batch=8,
                close_mosaic=10,
                workers=0,
                optimizer='MuSGD',  
                # optimizer='SGD',
                amp=False,
                project='runs/train',
                name='yolo26-C2PSA_TSSA',
                )

结果