YOLO11首发优化：2025.9月最新注意力机制 | 低计算复杂度的 多尺度线性注意力机制（MSLA）

import torch
import torch.nn as nn
import torch.nn.functional as F

from ultralytics.nn.modules.block import C2PSA, PSABlock, C3k,Bottleneck,C3k2


class LinearAttention(nn.Module):
    def __init__(self, dim, num_heads):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads

        self.qkv = nn.Linear(dim, 3 * dim, bias=False)
        self.proj = nn.Linear(dim, dim)

    def forward(self, x):
        b, c, h, w = x.shape

        x = x.view(b, c, h * w).permute(0, 2, 1)  # (b, h*w, c)

        qkv = self.qkv(x).reshape(b, h * w, 3, self.num_heads, self.dim // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]

        key = F.softmax(k, dim=-1)
        query = F.softmax(q, dim=-2)
        context = key.transpose(-2, -1) @ v
        x = (query @ context).reshape(b, h * w, c)

        x = self.proj(x)

        x = x.permute(0, 2, 1).view(b, c, h, w)

        return x


class DepthwiseConv(nn.Module):
    def __init__(self, in_channels, kernel_size):
        super(DepthwiseConv, self).__init__()
        self.depthwise = nn.Conv2d(in_channels, in_channels, kernel_size=kernel_size, groups=in_channels,
                                   padding=kernel_size // 2)
        self.relu = nn.ReLU()

    def forward(self, x):
        residual = x
        x = self.depthwise(x)
        x = x + residual
        x = self.relu(x)
        return x


class MSLA(nn.Module):
    def __init__(self, dim, num_heads):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads

        self.dw_conv_3x3 = DepthwiseConv(dim // 4, kernel_size=3)
        self.dw_conv_5x5 = DepthwiseConv(dim // 4, kernel_size=5)
        self.dw_conv_7x7 = DepthwiseConv(dim // 4, kernel_size=7)
        self.dw_conv_9x9 = DepthwiseConv(dim // 4, kernel_size=9)

        self.linear_attention = LinearAttention(dim=dim // 4, num_heads=num_heads)

        self.final_conv = nn.Conv2d(dim, dim, 1)

        self.scale_weights = nn.Parameter(torch.ones(4), requires_grad=True)

    def forward(self, input_):
        b, n, c = input_.shape
        h = int(n ** 0.5)
        w = int(n ** 0.5)

        input_reshaped = input_.reshape([b, c, h, w])

        split_size = c // 4
        x_3x3 = input_reshaped[:, :split_size, :, :]
        x_5x5 = input_reshaped[:, split_size:2 * split_size, :, :]
        x_7x7 = input_reshaped[:, 2 * split_size:3 * split_size:, :, :]
        x_9x9 = input_reshaped[:, 3 * split_size:, :, :]

        x_3x3 = self.dw_conv_3x3(x_3x3)
        x_5x5 = self.dw_conv_5x5(x_5x5)
        x_7x7 = self.dw_conv_7x7(x_7x7)
        x_9x9 = self.dw_conv_9x9(x_9x9)

        att_3x3 = self.linear_attention(x_3x3)
        att_5x5 = self.linear_attention(x_5x5)
        att_7x7 = self.linear_attention(x_7x7)
        att_9x9 = self.linear_attention(x_9x9)

        processed_input = torch.cat([
            att_3x3 * self.scale_weights[0],
            att_5x5 * self.scale_weights[1],
            att_7x7 * self.scale_weights[2],
            att_9x9 * self.scale_weights[3]
        ], dim=1)

        final_output = self.final_conv(processed_input)

        output_reshaped = final_output.reshape(b, n, self.dim)

        return output_reshaped

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

# Ultralytics YOLO11 object detection model with P3/8 - P5/32 outputs
# Model docs: https://docs.ultralytics.com/models/yolo11
# Task docs: https://docs.ultralytics.com/tasks/detect

# Parameters
nc: 80 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.50, 0.25, 1024] # summary: 319 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
  s: [0.50, 0.50, 1024] # summary: 319 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
  m: [0.50, 1.00, 512] # summary: 409 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
  l: [1.00, 1.00, 512] # summary: 631 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
  x: [1.00, 1.50, 512] # summary: 631 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs

# YOLO11n 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_MSLA, [256, False, 0.25]]
  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  - [-1, 2, C3k2_MSLA, [512, False, 0.25]]
  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  - [-1, 2, C3k2_MSLA, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  - [-1, 2, C3k2_MSLA, [1024, True]]
  - [-1, 1, SPPF, [1024, 5]] # 9
  - [-1, 2, C2MSLA, [1024]] # 10

# YOLO11n head
head:
  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
  - [-1, 2, C3k2, [512, False]] # 13

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

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

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

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