【YOLOv12多模态融合改进】| 利用 iRMB 倒置残差移动块二次改进CGA Fusion

一、本文介绍

本文记录的是利用 iRMB 模块改进 YOLOv12 的多模态融合部分 。主要讲解如何利用一些现有的模块二次改进多模态的融合部分。

iRMB（Inverted Residual Mobile Block） 的作用在于 克服了常见模块无法同时吸收CNN 效率建模局部特征 和 利用Transformer 动态建模能力学习不同模态之间的长距离交互问题 。相比一些复杂结构或多个混合模块的方法， 能更好地权衡模型成本和精度 。

本文将其用于 CGA Fusion 模块中并进行二次创新，更好地突出不同模态的重要特征，提升模型性能。

二、iRMB注意力介绍

Rethinking Mobile Block for Efficient Attention-based Models

2.1 设计出发点

统一CNN和Transformer优势 ：从高效的Inverted Residual Block（IRB）和Transformer的有效组件出发，期望在基础设施设计层面整合两者优势，为注意力模型构建类似IRB的轻量级基础结构。
解决现有模型问题 ：当前方法存在引入复杂结构或多个混合模块的问题，不利于应用优化。希望通过重新思考IRB和Transformer组件，构建简单有效的模块。

2.2 原理

基于Meta Mobile Block（MMB） ： MMB 是通过对 MobileNetv2 中的 IRB 和 Transformer 中的核心 MHSA 及 FFN模块 重新思考并归纳抽象得到的。它以参数化的扩展比率 λ 和高效算子 F 来实例化不同模块（如IRB、MHSA、FFN），揭示了这些模块的一致本质表达。

在这里插入图片描述

遵循通用高效模型准则 ：设计遵循可用性（简单实现，不使用复杂算子，易于应用优化）、均匀性（核心模块少，降低模型复杂度，加速部署）、有效性（分类和密集预测性能好）、效率（参数和计算少，权衡精度）的准则。

2.3 结构

2.3.1 主要组成部分

从微观角度看， iRMB 由 Depth - Wise Convolution（DW - Conv） 和 改进的Expanded Window MHSA（EW - MHSA） 组成。

2.3.2 具体操作流程

首先，类似MMB的操作，使用 扩展MLP （ $MLP_{e}$ ）以输出/输入比等于λ来扩展通道维度，即 $X_{e}=MLP_{e}(X)\left(\in \mathbb{R}^{\lambda C × H × W}\right)$ 。
然后，中间算子 F 进一步增强图像特征，这里F被建模为 级联的MHSA 和 卷积 操作，即 $F(\cdot)=Conv(MHSA(\cdot))$ ，具体采用DW - Conv和EW - MHSA的组合，其中EW - MHSA计算注意力矩阵的方式为 $X(\in \mathbb{R}^{C ×H ×W})$ ，而扩展值 $x_{e}$ 用于 $V(\in \mathbb{R}^{\lambda C ×H ×W})$ 。
最后，使用收缩 $M L P$ （ $MLP_{s}$ ）以倒置的输入/输出比等于 λ 来收缩通道维度，即 $X_{s}=MLP_{s}\left(X_{f}\right)\left(\in \mathbb{R}^{C × H × W}\right)$ ，并通过 残差连接 得到最终输出 $X_{s}(\in \mathbb{R}^{C ×H ×W})$ 。

在这里插入图片描述

2.4 优势

吸收CNN和Transformer优点 ：既能吸收 CNN 的效率来 建模局部特征 ，又能利用 Transformer 的 动态建模能力学习长距离交互 。
降低模型成本
- 通过采用高效的 Window - MHSA（WMHSA） 和 Depth - Wise Convolution（DW - Conv） 并带有 跳跃连接 ，权衡了模型成本和精度。
- 设计灵活性高，如不同深度可采用不同设置，满足性能需求的同时保持结构简洁。
性能优势
- 在ImageNet - 1K数据集上进行图像分类实验， iRMB 替换标准 Transformer 结构后，在相同训练设置下能以更少的参数和计算提高性能。
- 在下游任务（如目标检测和语义分割）中，基于 iRMB 构建的 EMO模型 在多个基准测试中取得了非常有竞争力的结果，超过了当前的SoTA方法。

论文： https://arxiv.org/pdf/2301.01146.pdf
源码： https://github.com/zhangzjn/EMO

三、iRMBFusion的实现代码

iRMBFusion 的实现代码如下：

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from functools import partial
from einops import rearrange
from timm.models._efficientnet_blocks import SqueezeExcite
from timm.models.layers import DropPath

inplace = True

class LayerNorm2d(nn.Module):

    def __init__(self, normalized_shape, eps=1e-6, elementwise_affine=True):
        super().__init__()
        self.norm = nn.LayerNorm(normalized_shape, eps, elementwise_affine)

    def forward(self, x):
        x = rearrange(x, 'b c h w -> b h w c').contiguous()
        x = self.norm(x)
        x = rearrange(x, 'b h w c -> b c h w').contiguous()
        return x


def get_norm(norm_layer='in_1d'):
    eps = 1e-6
    norm_dict = {
        'none': nn.Identity,
        'in_1d': partial(nn.InstanceNorm1d, eps=eps),
        'in_2d': partial(nn.InstanceNorm2d, eps=eps),
        'in_3d': partial(nn.InstanceNorm3d, eps=eps),
        'bn_1d': partial(nn.BatchNorm1d, eps=eps),
        'bn_2d': partial(nn.BatchNorm2d, eps=eps),
        # 'bn_2d': partial(nn.SyncBatchNorm, eps=eps),
        'bn_3d': partial(nn.BatchNorm3d, eps=eps),
        'gn': partial(nn.GroupNorm, eps=eps),
        'ln_1d': partial(nn.LayerNorm, eps=eps),
        'ln_2d': partial(LayerNorm2d, eps=eps),
    }
    return norm_dict[norm_layer]


def get_act(act_layer='relu'):
    act_dict = {
        'none': nn.Identity,
        'relu': nn.ReLU,
        'relu6': nn.ReLU6,
        'silu': nn.SiLU
    }
    return act_dict[act_layer]


class ConvNormAct(nn.Module):

    def __init__(self, dim_in, dim_out, kernel_size, stride=1, dilation=1, groups=1, bias=False,
                 skip=False, norm_layer='bn_2d', act_layer='relu', inplace=True, drop_path_rate=0.):
        super(ConvNormAct, self).__init__()
        self.has_skip = skip and dim_in == dim_out
        padding = math.ceil((kernel_size - stride) / 2)
        self.conv = nn.Conv2d(dim_in, dim_out, kernel_size, stride, padding, dilation, groups, bias)
        self.norm = get_norm(norm_layer)(dim_out)
        self.act = get_act(act_layer)(inplace=inplace)
        self.drop_path = DropPath(drop_path_rate) if drop_path_rate else nn.Identity()

    def forward(self, x):
        shortcut = x
        x = self.conv(x)
        x = self.norm(x)
        x = self.act(x)
        if self.has_skip:
            x = self.drop_path(x) + shortcut
        return x



class iRMB(nn.Module):

    def __init__(self, dim_in,  norm_in=True, has_skip=True, exp_ratio=1.0, norm_layer='bn_2d',
                 act_layer='relu', v_proj=True, dw_ks=3, stride=1, dilation=1, se_ratio=0.0, dim_head=8, window_size=7,
                 attn_s=True, qkv_bias=False, attn_drop=0., drop=0., drop_path=0., v_group=False, attn_pre=False):
        super().__init__()
        dim_out = dim_in
        self.norm = get_norm(norm_layer)(dim_in) if norm_in else nn.Identity()
        dim_mid = int(dim_in * exp_ratio)
        self.has_skip = (dim_in == dim_out and stride == 1) and has_skip
        self.attn_s = attn_s
        if self.attn_s:
            assert dim_in % dim_head == 0, 'dim should be divisible by num_heads'
            self.dim_head = dim_head
            self.window_size = window_size
            self.num_head = dim_in // dim_head
            self.scale = self.dim_head ** -0.5
            self.attn_pre = attn_pre
            self.qk = ConvNormAct(dim_in, int(dim_in * 2), kernel_size=1, bias=qkv_bias, norm_layer='none',
                                  act_layer='none')
            self.v = ConvNormAct(dim_in, dim_mid, kernel_size=1, groups=self.num_head if v_group else 1, bias=qkv_bias,
                                 norm_layer='none', act_layer=act_layer, inplace=inplace)
            self.attn_drop = nn.Dropout(attn_drop)
        else:
            if v_proj:
                self.v = ConvNormAct(dim_in, dim_mid, kernel_size=1, bias=qkv_bias, norm_layer='none',
                                     act_layer=act_layer, inplace=inplace)
            else:
                self.v = nn.Identity()
        self.conv_local = ConvNormAct(dim_mid, dim_mid, kernel_size=dw_ks, stride=stride, dilation=dilation,
                                      groups=dim_mid, norm_layer='bn_2d', act_layer='silu', inplace=inplace)
        self.se = SqueezeExcite(dim_mid, rd_ratio=se_ratio, act_layer=get_act(act_layer)) if se_ratio > 0.0 else nn.Identity()

        self.proj_drop = nn.Dropout(drop)
        self.proj = ConvNormAct(dim_mid, dim_out, kernel_size=1, norm_layer='none', act_layer='none', inplace=inplace)
        self.drop_path = DropPath(drop_path) if drop_path else nn.Identity()

    def forward(self, x):
        shortcut = x
        x = self.norm(x)
        B, C, H, W = x.shape
        if self.attn_s:
            # padding
            if self.window_size <= 0:
                window_size_W, window_size_H = W, H
            else:
                window_size_W, window_size_H = self.window_size, self.window_size
            pad_l, pad_t = 0, 0
            pad_r = (window_size_W - W % window_size_W) % window_size_W
            pad_b = (window_size_H - H % window_size_H) % window_size_H
            x = F.pad(x, (pad_l, pad_r, pad_t, pad_b, 0, 0,))
            n1, n2 = (H + pad_b) // window_size_H, (W + pad_r) // window_size_W
            x = rearrange(x, 'b c (h1 n1) (w1 n2) -> (b n1 n2) c h1 w1', n1=n1, n2=n2).contiguous()
            # attention
            b, c, h, w = x.shape
            qk = self.qk(x)
            qk = rearrange(qk, 'b (qk heads dim_head) h w -> qk b heads (h w) dim_head', qk=2, heads=self.num_head,
                           dim_head=self.dim_head).contiguous()
            q, k = qk[0], qk[1]
            attn_spa = (q @ k.transpose(-2, -1)) * self.scale
            attn_spa = attn_spa.softmax(dim=-1)
            attn_spa = self.attn_drop(attn_spa)
            if self.attn_pre:
                x = rearrange(x, 'b (heads dim_head) h w -> b heads (h w) dim_head', heads=self.num_head).contiguous()
                x_spa = attn_spa @ x
                x_spa = rearrange(x_spa, 'b heads (h w) dim_head -> b (heads dim_head) h w', heads=self.num_head, h=h,
                                  w=w).contiguous()
                x_spa = self.v(x_spa)
            else:
                v = self.v(x)
                v = rearrange(v, 'b (heads dim_head) h w -> b heads (h w) dim_head', heads=self.num_head).contiguous()
                x_spa = attn_spa @ v
                x_spa = rearrange(x_spa, 'b heads (h w) dim_head -> b (heads dim_head) h w', heads=self.num_head, h=h,
                                  w=w).contiguous()
            # unpadding
            x = rearrange(x_spa, '(b n1 n2) c h1 w1 -> b c (h1 n1) (w1 n2)', n1=n1, n2=n2).contiguous()
            if pad_r > 0 or pad_b > 0:
                x = x[:, :, :H, :W].contiguous()
        else:
            x = self.v(x)

        x = x + self.se(self.conv_local(x)) if self.has_skip else self.se(self.conv_local(x))

        x = self.proj_drop(x)
        x = self.proj(x)

        x = (shortcut + self.drop_path(x)) if self.has_skip else x
        return x


class PixelAttention_CGA(nn.Module):
    def __init__(self, dim):
        super(PixelAttention_CGA, self).__init__()
        self.pa2 = nn.Conv2d(2 * dim, dim, 7, padding=3, padding_mode='reflect' ,groups=dim, bias=True)
        self.sigmoid = nn.Sigmoid()

    def forward(self, x, pattn1):
        B, C, H, W = x.shape
        x = x.unsqueeze(dim=2) # B, C, 1, H, W
        pattn1 = pattn1.unsqueeze(dim=2) # B, C, 1, H, W
        x2 = torch.cat([x, pattn1], dim=2) # B, C, 2, H, W
        x2 = rearrange(x2, 'b c t h w -> b (c t) h w')
        pattn2 = self.pa2(x2)
        pattn2 = self.sigmoid(pattn2)
        return pattn2

class iRMBFusion(nn.Module):
    def __init__(self, dim):
        super(iRMBFusion, self).__init__()
        self.cfam = iRMB(dim)
        self.pa = PixelAttention_CGA(dim)
        self.conv = nn.Conv2d(dim, dim, 1, bias=True)
        self.sigmoid = nn.Sigmoid()

    def forward(self, data):
        x, y = data
        initial = x + y
        pattn1 = self.cfam(initial)
        pattn2 = self.sigmoid(self.pa(initial, pattn1))
        result = initial + pattn2 * x + (1 - pattn2) * y
        result = self.conv(result)
        return result

四、融合步骤

5.1 修改一

① 在 ultralytics/nn/ 目录下新建 AddModules 文件夹用于存放模块代码

② 在 AddModules 文件夹下新建 iRMBFusion.py ，将 第三节 中的代码粘贴到此处

在这里插入图片描述

5.2 修改二

在 AddModules 文件夹下新建 __init__.py （已有则不用新建），在文件内导入模块： from .iRMBFusion import *

在这里插入图片描述

5.3 修改三

在 ultralytics/nn/modules/tasks.py 文件中，需要在两处位置添加各模块类名称。

首先：导入模块

在这里插入图片描述

其次：在 parse_model函数 中注册 iRMBFusion 模块

在这里插入图片描述

        elif m in {iRMBFusion}:
            c2 = ch[f[0]]
            args = [c2]

在这里插入图片描述

最后将 ultralytics/utils/torch_utils.py 中的 get_flops 函数中的 stride 指定为 640 。

在这里插入图片描述

五、yaml模型文件

5.1 中期融合⭐

📌 此模型的修方法是将原本的中期融合中的Concat融合部分换成iRMBFusion，融合骨干部分的多模态信息。

# YOLOv12 🚀, AGPL-3.0 license
# YOLOv12 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
# CFG file for YOLOv12-turbo

# Parameters
ch: 6
nc: 80 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov12n.yaml' will call yolov12.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.50, 0.25, 1024] # summary: 497 layers, 2,553,904 parameters, 2,553,888 gradients, 6.2 GFLOPs
  s: [0.50, 0.50, 1024] # summary: 497 layers, 9,127,424 parameters, 9,127,408 gradients, 19.7 GFLOPs
  m: [0.50, 1.00, 512] # summary: 533 layers, 19,670,784 parameters, 19,670,768 gradients, 60.4 GFLOPs
  l: [1.00, 1.00, 512] # summary: 895 layers, 26,506,496 parameters, 26,506,480 gradients, 83.3 GFLOPs
  x: [1.00, 1.50, 512] # summary: 895 layers, 59,414,176 parameters, 59,414,160 gradients, 185.9 GFLOPs


# YOLO12 backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, IN, []]  # 0
  - [-1, 1, Multiin, [1]]  # 1
  - [-2, 1, Multiin, [2]]  # 2

  - [1, 1, Conv,  [64, 3, 2]] # 3-P1/2
  - [-1, 1, Conv,  [128, 3, 2, 1, 2]] # 4-P2/4
  - [-1, 2, C3k2,  [256, False, 0.25]]
  - [-1, 1, Conv,  [256, 3, 2, 1, 4]] # 6-P3/8
  - [-1, 2, C3k2,  [512, False, 0.25]]
  - [-1, 1, Conv,  [512, 3, 2]] # 8-P4/16
  - [-1, 4, A2C2f, [512, True, 4]]
  - [-1, 1, Conv,  [1024, 3, 2]] # 10-P5/32
  - [-1, 4, A2C2f, [1024, True, 1]] # 11

  - [2, 1, Conv,  [64, 3, 2]] # 12-P1/2
  - [-1, 1, Conv,  [128, 3, 2, 1, 2]] # 13-P2/4
  - [-1, 2, C3k2,  [256, False, 0.25]]
  - [-1, 1, Conv,  [256, 3, 2, 1, 4]] # 15-P3/8
  - [-1, 2, C3k2,  [512, False, 0.25]]
  - [-1, 1, Conv,  [512, 3, 2]] # 17-P4/16
  - [-1, 4, A2C2f, [512, True, 4]]
  - [-1, 1, Conv,  [1024, 3, 2]] # 19-P5/32
  - [-1, 4, A2C2f, [1024, True, 1]] # 20

  - [[7, 16], 1, iRMBFusion, []]  # 21 cat backbone P3
  - [[9, 18], 1, iRMBFusion, []]  # 22 cat backbone P4
  - [[11, 20], 1, iRMBFusion, []]  # 23 cat backbone P5

# YOLO12 head
head:
  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 22], 1, Concat, [1]] # cat backbone P4
  - [-1, 2, A2C2f, [512, False, -1]] # 26

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 21], 1, Concat, [1]] # cat backbone P3
  - [-1, 2, A2C2f, [256, False, -1]] # 29

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 26], 1, Concat, [1]] # cat head P4
  - [-1, 2, A2C2f, [512, False, -1]] # 32

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 23], 1, Concat, [1]] # cat head P5
  - [-1, 2, C3k2, [1024, True]] # 35 (P5/32-large)

  - [[29, 32, 35], 1, Detect, [nc]] # Detect(P3, P4, P5)

5.2 中-后期融合⭐

📌 此模型的修方法是将原本的中-后期融合中的Concat融合部分换成iRMBFusion，融合FPN部分的多模态信息。

# YOLOv12 🚀, AGPL-3.0 license
# YOLOv12 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
# CFG file for YOLOv12-turbo

# Parameters
ch: 6
nc: 80 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov12n.yaml' will call yolov12.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.50, 0.25, 1024] # summary: 497 layers, 2,553,904 parameters, 2,553,888 gradients, 6.2 GFLOPs
  s: [0.50, 0.50, 1024] # summary: 497 layers, 9,127,424 parameters, 9,127,408 gradients, 19.7 GFLOPs
  m: [0.50, 1.00, 512] # summary: 533 layers, 19,670,784 parameters, 19,670,768 gradients, 60.4 GFLOPs
  l: [1.00, 1.00, 512] # summary: 895 layers, 26,506,496 parameters, 26,506,480 gradients, 83.3 GFLOPs
  x: [1.00, 1.50, 512] # summary: 895 layers, 59,414,176 parameters, 59,414,160 gradients, 185.9 GFLOPs


# YOLO12 backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, IN, []]  # 0
  - [-1, 1, Multiin, [1]]  # 1
  - [-2, 1, Multiin, [2]]  # 2

  - [1, 1, Conv,  [64, 3, 2]] # 3-P1/2
  - [-1, 1, Conv,  [128, 3, 2, 1, 2]] # 4-P2/4
  - [-1, 2, C3k2,  [256, False, 0.25]]
  - [-1, 1, Conv,  [256, 3, 2, 1, 4]] # 6-P3/8
  - [-1, 2, C3k2,  [512, False, 0.25]]
  - [-1, 1, Conv,  [512, 3, 2]] # 8-P4/16
  - [-1, 4, A2C2f, [512, True, 4]]
  - [-1, 1, Conv,  [1024, 3, 2]] # 10-P5/32
  - [-1, 4, A2C2f, [1024, True, 1]] # 11

  - [2, 1, Conv,  [64, 3, 2]] # 12-P1/2
  - [-1, 1, Conv,  [128, 3, 2, 1, 2]] # 13-P2/4
  - [-1, 2, C3k2,  [256, False, 0.25]]
  - [-1, 1, Conv,  [256, 3, 2, 1, 4]] # 15-P3/8
  - [-1, 2, C3k2,  [512, False, 0.25]]
  - [-1, 1, Conv,  [512, 3, 2]] # 17-P4/16
  - [-1, 4, A2C2f, [512, True, 4]]
  - [-1, 1, Conv,  [1024, 3, 2]] # 19-P5/32
  - [-1, 4, A2C2f, [1024, True, 1]] # 20

# YOLO12 head
head:
  - [11, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 9], 1, Concat, [1]] # cat backbone P4
  - [-1, 2, A2C2f, [512, False, -1]] # 23

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 7], 1, Concat, [1]] # cat backbone P3
  - [-1, 2, A2C2f, [256, False, -1]] # 26

  - [20, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 18], 1, Concat, [1]] # cat backbone P4
  - [-1, 2, A2C2f, [512, False, -1]] # 29

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 16], 1, Concat, [1]] # cat backbone P3
  - [-1, 2, A2C2f, [256, False, -1]] # 32

  - [[11, 20], 1, iRMBFusion, []]  # cat head P5  33
  - [[23, 29], 1, iRMBFusion, []]  # cat head P5  34
  - [[26, 32], 1, iRMBFusion, []]  # cat head P5  35

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 34], 1, Concat, [1]] # cat head P4
  - [-1, 2, A2C2f, [512, False, -1]] # 38

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 33], 1, Concat, [1]] # cat head P5
  - [-1, 2, C3k2, [1024, True]] # 41 (P5/32-large)

  - [[35, 38, 41], 1, Detect, [nc]] # Detect(P3, P4, P5)

5.3 后期融合⭐

📌 此模型的修方法是将原本的后期融合中的Concat融合部分换成iRMBFusion，融合颈部部分的多模态信息。

# YOLOv12 🚀, AGPL-3.0 license
# YOLOv12 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
# CFG file for YOLOv12-turbo

# Parameters
ch: 6
nc: 80 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov12n.yaml' will call yolov12.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.50, 0.25, 1024] # summary: 497 layers, 2,553,904 parameters, 2,553,888 gradients, 6.2 GFLOPs
  s: [0.50, 0.50, 1024] # summary: 497 layers, 9,127,424 parameters, 9,127,408 gradients, 19.7 GFLOPs
  m: [0.50, 1.00, 512] # summary: 533 layers, 19,670,784 parameters, 19,670,768 gradients, 60.4 GFLOPs
  l: [1.00, 1.00, 512] # summary: 895 layers, 26,506,496 parameters, 26,506,480 gradients, 83.3 GFLOPs
  x: [1.00, 1.50, 512] # summary: 895 layers, 59,414,176 parameters, 59,414,160 gradients, 185.9 GFLOPs


# YOLO12 backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, IN, []]  # 0
  - [-1, 1, Multiin, [1]]  # 1
  - [-2, 1, Multiin, [2]]  # 2

  - [1, 1, Conv,  [64, 3, 2]] # 3-P1/2
  - [-1, 1, Conv,  [128, 3, 2, 1, 2]] # 4-P2/4
  - [-1, 2, C3k2,  [256, False, 0.25]]
  - [-1, 1, Conv,  [256, 3, 2, 1, 4]] # 6-P3/8
  - [-1, 2, C3k2,  [512, False, 0.25]]
  - [-1, 1, Conv,  [512, 3, 2]] # 8-P4/16
  - [-1, 4, A2C2f, [512, True, 4]]
  - [-1, 1, Conv,  [1024, 3, 2]] # 10-P5/32
  - [-1, 4, A2C2f, [1024, True, 1]] # 11

  - [2, 1, Conv,  [64, 3, 2]] # 12-P1/2
  - [-1, 1, Conv,  [128, 3, 2, 1, 2]] # 13-P2/4
  - [-1, 2, C3k2,  [256, False, 0.25]]
  - [-1, 1, Conv,  [256, 3, 2, 1, 4]] # 15-P3/8
  - [-1, 2, C3k2,  [512, False, 0.25]]
  - [-1, 1, Conv,  [512, 3, 2]] # 17-P4/16
  - [-1, 4, A2C2f, [512, True, 4]]
  - [-1, 1, Conv,  [1024, 3, 2]] # 19-P5/32
  - [-1, 4, A2C2f, [1024, True, 1]] # 20

# YOLO12 head
head:
  - [11, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 9], 1, Concat, [1]] # cat backbone P4
  - [-1, 2, A2C2f, [512, False, -1]] # 23

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 7], 1, Concat, [1]] # cat backbone P3
  - [-1, 2, A2C2f, [256, False, -1]] # 26

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 23], 1, Concat, [1]] # cat head P4
  - [-1, 2, A2C2f, [512, False, -1]] # 29

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 11], 1, Concat, [1]] # cat head P5
  - [-1, 2, C3k2, [1024, True]] # 32 (P5/32-large)

  - [20, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 18], 1, Concat, [1]] # cat backbone P4
  - [-1, 2, A2C2f, [512, False, -1]] # 35

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 16], 1, Concat, [1]] # cat backbone P3
  - [-1, 2, A2C2f, [256, False, -1]] # 38

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 35], 1, Concat, [1]] # cat head P4
  - [-1, 2, A2C2f, [512, False, -1]] # 41

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 20], 1, Concat, [1]] # cat head P5
  - [-1, 2, C3k2, [1024, True]] # 44 (P5/32-large)

  - [[26, 38], 1, iRMBFusion, []]  # cat head P5  45
  - [[29, 41], 1, iRMBFusion, []]  # cat head P5  46
  - [[32, 44], 1, iRMBFusion, []]  # cat head P5  47

  - [[45, 46, 47], 1, Detect, [nc]] # Detect(P3, P4, P5)

六、成功运行结果

打印网络模型可以看到不同的融合层已经加入到模型中，并可以进行训练了。

YOLOv12-mid-iRMBFusion ：

YOLOv12-mid-iRMBFusion summary: 806 layers, 4,445,971 parameters, 4,445,955 gradients, 10.8 GFLOPs

                   from  n    params  module                                       arguments
  0                  -1  1         0  ultralytics.nn.AddModules.multimodal.IN      []
  1                  -1  1         0  ultralytics.nn.AddModules.multimodal.Multiin [1]
  2                  -2  1         0  ultralytics.nn.AddModules.multimodal.Multiin [2]
  3                   1  1       464  ultralytics.nn.modules.conv.Conv             [3, 16, 3, 2]
  4                  -1  1      2368  ultralytics.nn.modules.conv.Conv             [16, 32, 3, 2, 1, 2]
  5                  -1  1      6640  ultralytics.nn.modules.block.C3k2            [32, 64, 1, False, 0.25]
  6                  -1  1      9344  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2, 1, 4]
  7                  -1  1     26080  ultralytics.nn.modules.block.C3k2            [64, 128, 1, False, 0.25]
  8                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
  9                  -1  2    180864  ultralytics.nn.AddModules.A2C2f.A2C2f        [128, 128, 2, True, 4]
 10                  -1  1    295424  ultralytics.nn.modules.conv.Conv             [128, 256, 3, 2]
 11                  -1  2    689408  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 256, 2, True, 1]
 12                   2  1       464  ultralytics.nn.modules.conv.Conv             [3, 16, 3, 2]
 13                  -1  1      2368  ultralytics.nn.modules.conv.Conv             [16, 32, 3, 2, 1, 2]
 14                  -1  1      6640  ultralytics.nn.modules.block.C3k2            [32, 64, 1, False, 0.25]
 15                  -1  1      9344  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2, 1, 4]
 16                  -1  1     26080  ultralytics.nn.modules.block.C3k2            [64, 128, 1, False, 0.25]
 17                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
 18                  -1  2    180864  ultralytics.nn.AddModules.A2C2f.A2C2f        [128, 128, 2, True, 4]
 19                  -1  1    295424  ultralytics.nn.modules.conv.Conv             [128, 256, 3, 2]
 20                  -1  2    689408  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 256, 2, True, 1]
 21             [7, 16]  1     96384  ultralytics.nn.AddModules.iRMBFusion.iRMBFusion[128]
 22             [9, 18]  1     96384  ultralytics.nn.AddModules.iRMBFusion.iRMBFusion[128]
 23            [11, 20]  1    356608  ultralytics.nn.AddModules.iRMBFusion.iRMBFusion[256]
 24                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 25            [-1, 22]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 26                  -1  1     86912  ultralytics.nn.AddModules.A2C2f.A2C2f        [384, 128, 1, False, -1]
 27                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 28            [-1, 21]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 29                  -1  1     24000  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 64, 1, False, -1]
 30                  -1  1     36992  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2]
 31            [-1, 26]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 32                  -1  1     74624  ultralytics.nn.AddModules.A2C2f.A2C2f        [192, 128, 1, False, -1]
 33                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
 34            [-1, 23]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 35                  -1  1    378880  ultralytics.nn.modules.block.C3k2            [384, 256, 1, True]
 36        [29, 32, 35]  1    430867  ultralytics.nn.modules.head.Detect           [1, [64, 128, 256]]
YOLOv12-mid-iRMBFusion summary: 806 layers, 4,445,971 parameters, 4,445,955 gradients, 10.8 GFLOPs

YOLOv12-mid-to-late-iRMBFusion ：

YOLOv12-mid-to-late-CGAFusion summary: 822 layers, 4,160,628 parameters, 4,160,612 gradients, 9.8 GFLOPs

                   from  n    params  module                                       arguments
  0                  -1  1         0  ultralytics.nn.AddModules.multimodal.IN      []
  1                  -1  1         0  ultralytics.nn.AddModules.multimodal.Multiin [1]
  2                  -2  1         0  ultralytics.nn.AddModules.multimodal.Multiin [2]
  3                   1  1       464  ultralytics.nn.modules.conv.Conv             [3, 16, 3, 2]
  4                  -1  1      2368  ultralytics.nn.modules.conv.Conv             [16, 32, 3, 2, 1, 2]
  5                  -1  1      6640  ultralytics.nn.modules.block.C3k2            [32, 64, 1, False, 0.25]
  6                  -1  1      9344  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2, 1, 4]
  7                  -1  1     26080  ultralytics.nn.modules.block.C3k2            [64, 128, 1, False, 0.25]
  8                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
  9                  -1  2    180864  ultralytics.nn.AddModules.A2C2f.A2C2f        [128, 128, 2, True, 4]
 10                  -1  1    295424  ultralytics.nn.modules.conv.Conv             [128, 256, 3, 2]
 11                  -1  2    689408  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 256, 2, True, 1]
 12                   2  1       464  ultralytics.nn.modules.conv.Conv             [3, 16, 3, 2]
 13                  -1  1      2368  ultralytics.nn.modules.conv.Conv             [16, 32, 3, 2, 1, 2]
 14                  -1  1      6640  ultralytics.nn.modules.block.C3k2            [32, 64, 1, False, 0.25]
 15                  -1  1      9344  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2, 1, 4]
 16                  -1  1     26080  ultralytics.nn.modules.block.C3k2            [64, 128, 1, False, 0.25]
 17                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
 18                  -1  2    180864  ultralytics.nn.AddModules.A2C2f.A2C2f        [128, 128, 2, True, 4]
 19                  -1  1    295424  ultralytics.nn.modules.conv.Conv             [128, 256, 3, 2]
 20                  -1  2    689408  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 256, 2, True, 1]
 21                  11  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 22             [-1, 9]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 23                  -1  1     86912  ultralytics.nn.AddModules.A2C2f.A2C2f        [384, 128, 1, False, -1]
 24                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 25             [-1, 7]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 26                  -1  1     24000  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 64, 1, False, -1]
 27                  20  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 28            [-1, 18]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 29                  -1  1     86912  ultralytics.nn.AddModules.A2C2f.A2C2f        [384, 128, 1, False, -1]
 30                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 31            [-1, 16]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 32                  -1  1     24000  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 64, 1, False, -1]
 33            [11, 20]  1    107907  ultralytics.nn.AddModules.CGAFusion.CGAFusion[256]
 34            [23, 29]  1     33523  ultralytics.nn.AddModules.CGAFusion.CGAFusion[128]
 35            [26, 32]  1     11691  ultralytics.nn.AddModules.CGAFusion.CGAFusion[64]
 36                  -1  1     36992  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2]
 37            [-1, 34]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 38                  -1  1     74624  ultralytics.nn.AddModules.A2C2f.A2C2f        [192, 128, 1, False, -1]
 39                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
 40            [-1, 33]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 41                  -1  1    378880  ultralytics.nn.modules.block.C3k2            [384, 256, 1, True]
 42        [35, 38, 41]  1    430867  ultralytics.nn.modules.head.Detect           [1, [64, 128, 256]]
YOLOv12-mid-to-late-CGAFusion summary: 822 layers, 4,160,628 parameters, 4,160,612 gradients, 9.8 GFLOPs

YOLOv12-late-iRMBFusion ：

YOLOv12-late-iRMBFusion summary: 950 layers, 5,126,419 parameters, 5,126,403 gradients, 11.3 GFLOPs

                   from  n    params  module                                       arguments
  0                  -1  1         0  ultralytics.nn.AddModules.multimodal.IN      []
  1                  -1  1         0  ultralytics.nn.AddModules.multimodal.Multiin [1]
  2                  -2  1         0  ultralytics.nn.AddModules.multimodal.Multiin [2]
  3                   1  1       464  ultralytics.nn.modules.conv.Conv             [3, 16, 3, 2]
  4                  -1  1      2368  ultralytics.nn.modules.conv.Conv             [16, 32, 3, 2, 1, 2]
  5                  -1  1      6640  ultralytics.nn.modules.block.C3k2            [32, 64, 1, False, 0.25]
  6                  -1  1      9344  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2, 1, 4]
  7                  -1  1     26080  ultralytics.nn.modules.block.C3k2            [64, 128, 1, False, 0.25]
  8                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
  9                  -1  2    180864  ultralytics.nn.AddModules.A2C2f.A2C2f        [128, 128, 2, True, 4]
 10                  -1  1    295424  ultralytics.nn.modules.conv.Conv             [128, 256, 3, 2]
 11                  -1  2    689408  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 256, 2, True, 1]
 12                   2  1       464  ultralytics.nn.modules.conv.Conv             [3, 16, 3, 2]
 13                  -1  1      2368  ultralytics.nn.modules.conv.Conv             [16, 32, 3, 2, 1, 2]
 14                  -1  1      6640  ultralytics.nn.modules.block.C3k2            [32, 64, 1, False, 0.25]
 15                  -1  1      9344  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2, 1, 4]
 16                  -1  1     26080  ultralytics.nn.modules.block.C3k2            [64, 128, 1, False, 0.25]
 17                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
 18                  -1  2    180864  ultralytics.nn.AddModules.A2C2f.A2C2f        [128, 128, 2, True, 4]
 19                  -1  1    295424  ultralytics.nn.modules.conv.Conv             [128, 256, 3, 2]
 20                  -1  2    689408  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 256, 2, True, 1]
 21                  11  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 22             [-1, 9]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 23                  -1  1     86912  ultralytics.nn.AddModules.A2C2f.A2C2f        [384, 128, 1, False, -1]
 24                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 25             [-1, 7]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 26                  -1  1     24000  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 64, 1, False, -1]
 27                  -1  1     36992  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2]
 28            [-1, 23]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 29                  -1  1     74624  ultralytics.nn.AddModules.A2C2f.A2C2f        [192, 128, 1, False, -1]
 30                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
 31            [-1, 11]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 32                  -1  1    378880  ultralytics.nn.modules.block.C3k2            [384, 256, 1, True]
 33                  20  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 34            [-1, 18]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 35                  -1  1     86912  ultralytics.nn.AddModules.A2C2f.A2C2f        [384, 128, 1, False, -1]
 36                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']
 37            [-1, 16]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 38                  -1  1     24000  ultralytics.nn.AddModules.A2C2f.A2C2f        [256, 64, 1, False, -1]
 39                  -1  1     36992  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2]
 40            [-1, 35]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 41                  -1  1     74624  ultralytics.nn.AddModules.A2C2f.A2C2f        [192, 128, 1, False, -1]
 42                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]
 43            [-1, 20]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 44                  -1  1    378880  ultralytics.nn.modules.block.C3k2            [384, 256, 1, True]
 45            [26, 38]  1     27712  ultralytics.nn.AddModules.iRMBFusion.iRMBFusion[64]
 46            [29, 41]  1     96384  ultralytics.nn.AddModules.iRMBFusion.iRMBFusion[128]
 47            [32, 44]  1    356608  ultralytics.nn.AddModules.iRMBFusion.iRMBFusion[256]
 48        [45, 46, 47]  1    430867  ultralytics.nn.modules.head.Detect           [1, [64, 128, 256]]
YOLOv12-late-iRMBFusion summary: 950 layers, 5,126,419 parameters, 5,126,403 gradients, 11.3 GFLOPs