一、本文介绍

本文改进 双HS-FPN颈部结构，融合YOLOv12中的多模态特征，以优化目标检测网络模型 。

HS-FPN 借助 通道注意力机制 及独特的 多尺度融合策略 ，有效应对 目标尺寸差异及特征稀缺问题 。针对不同模态，其 利用高级特征筛选低级特征 ，增强特征表达，助力模型精准定位和识别目标， 减少因尺度变化及特征不足导致的检测误差 ，提升 YOLOv12 在多模态检测任务中的准确性与稳定性。

---

---

二、HS-FPN介绍

Accurate Leukocyte Detection Based on Deformable-DETR and Multi-Level Feature Fusion for Aiding Diagnosis of Blood Diseases

HS - FPN结构 由 特征选择模块 和 特征融合模块 组成。

• 特征选择模块中， CA模块 先处理输入 特征图 ，经 池化 、 激活函数 确定各通道权重以 过滤特征图 ， DM模块 再对不同尺度特征图降维；
• 特征融合模块中，利用 SFF机制 ， 以高级特征为权重筛选低级特征语义信息后融合 ，提升模型检测能力。

2.1 出发点

在白细胞数据集中，白细胞识别任务面临多尺度问题，不同类型白细胞直径通常有差异，相同白细胞在不同显微镜下成像大小也会不同，这使得模型难以准确识别白细胞，所以需要设计HS - FPN来实现多尺度特征融合，帮助模型捕捉更全面的白细胞特征信息。

2.2 结构原理

• 特征选择模块 ：由 CA模块 和 DM模块 组成。对于输入特征图 $f_{in}\in R^{C\times H\times W}$ ， CA模块 先进行 全局平均池化 和全 局最大池化 ，再结合结果，经 Sigmoid激活函数 确定各通道权重 $f_{CA}\in R^{C\times 1\times 1}$ ，通过与对应尺度特征图相乘得到过滤后的特征图。因不同尺度特征图通道数不同， DM模块 用1×1卷积将各尺度特征图通道数降为 256。

• 特征融合模块 ：骨干网络生成的多尺度特征图中， 高级特征语义信息丰富但目标定位粗糙，低级特征定位精确但语义信息有限 。传统直接像素求和融合有缺陷，研究中的 SFF模块 以高级特征为权重筛选低级特征中的关键语义信息 。对于输入高级特征 $f_{high}\in R^{C\times H\times W}$ 和低级特征 $f_{low}\in R^{C\times H_1\times W_1}$ ，先对高级特征用步长为2、卷积核为3 x3的 转置卷积 扩展，再用 双线性插值 统一维度得到 $f_{att}\in R^{C\times H_1\times W_1}$ ，经 CA 模块 将高级特征转为注意力权重过滤低级特征，最后融合得到 $f_{out}\in R^{C\times H_1\times W_1}$ ，其融合过程公式为 $$f_{att}=BL\left(T-Conv\left(f_{high}\right)\right)$$ $$f_{out}=f_{low}\ast CA\left(f_{att}\right)+f_{att}$$

2.3 作用

HS-FPN 能够利用通道注意力模块，以 高级语义特征为权重过滤低级特征 ，并将筛选后的特征与高级特征逐点相加，实现多尺度特征融合，从而提高模型的特征表达能力，有助于检测到细微特征，增强模型的检测能力。

论文： https://www.sciencedirect.com/science/article/abs/pii/S0010482524000015
源码： https://github.com/JustlfC03/MFDS-DETR

三、HS-FPN的实现代码

HS-FPN模块 的实现代码如下：

import torch
import torch.nn as nn

class ChannelAttention_HSFPN(nn.Module):
    def __init__(self, in_planes, ratio=4, flag=True):
        super(ChannelAttention_HSFPN, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.max_pool = nn.AdaptiveMaxPool2d(1)

        self.conv1 = nn.Conv2d(in_planes, in_planes // ratio, 1, bias=False)
        self.relu = nn.ReLU()
        self.conv2 = nn.Conv2d(in_planes // ratio, in_planes, 1, bias=False)
        self.flag = flag
        self.sigmoid = nn.Sigmoid()

        nn.init.xavier_uniform_(self.conv1.weight)
        nn.init.xavier_uniform_(self.conv2.weight)

    def forward(self, x):
        avg_out = self.conv2(self.relu(self.conv1(self.avg_pool(x))))
        max_out = self.conv2(self.relu(self.conv1(self.max_pool(x))))
        out = avg_out + max_out
        return self.sigmoid(out) * x if self.flag else self.sigmoid(out)

class Multiply(nn.Module):
    def __init__(self) -> None:
        super().__init__()

    def forward(self, x):
        return x[0] * x[1]

class Add_HSFPN(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        return torch.sum(torch.stack(x, dim=0), dim=0)

---

四、添加步骤

4.1 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 HSFPN.py ，将 第三节 中的代码粘贴到此处

4.2 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 .HSFPN import *

4.3 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

然后，在 parse_model函数 中添加 nn.Conv2d 、 ChannelAttention_HSFPN 、 Multiply 、 Add_HSFPN ：

 elif m is ChannelAttention_HSFPN:
     c2 = ch[f]
     args = [c2, *args]
 elif m is Multiply:
     c2 = ch[f[0]]
 elif m is Add_HSFPN:
     c2 = ch[f[-1]]

---

五、yaml模型文件

5.1 中期融合⭐

📌 此模型的修方法是将颈部网络换成HSFPN结构。

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

# YOLO12 head
head:
  - [-1, 1, ChannelAttention_HSFPN, [4]] # 24
  - [-1, 1, nn.Conv2d, [256, 1]] # 25
  - [-1, 1, nn.ConvTranspose2d, [256, 3, 2, 1, 1]] # 26

  - [22, 1, ChannelAttention_HSFPN, [4]] # 27
  - [-1, 1, nn.Conv2d, [256, 1]] # 28
  - [26, 1, ChannelAttention_HSFPN, [4, False]] # 29
  - [[-1, -2], 1, Multiply, []] # 30
  - [[-1, 26], 1, Add_HSFPN, []] # 31
  - [-1, 2, A2C2f, [256, False, -1]] # 32 P4/16

  - [26, 1, nn.ConvTranspose2d, [256, 3, 2, 1, 1]] # 33
  - [21, 1, ChannelAttention_HSFPN, [4]] # 34
  - [-1, 1, nn.Conv2d, [256, 1]] # 35
  - [33, 1, ChannelAttention_HSFPN, [4, False]] # 36
  - [[-1, -2], 1, Multiply, []] # 37
  - [[-1, 33], 1, Add_HSFPN, []] # 38
  - [-1, 2, C3k2, [256, True]] # 39 P3/16

  - [[25, 32, 39], 1, Detect, [nc]] # Detect(P3, P4, P5)

5.2 后期融合⭐

📌 此模型的修方法是将两个模态的颈部网络换成HSFPN结构，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, ChannelAttention_HSFPN, [4]] # 21
  - [-1, 1, nn.Conv2d, [256, 1]] # 22
  - [-1, 1, nn.ConvTranspose2d, [256, 3, 2, 1, 1]] # 23

  - [9, 1, ChannelAttention_HSFPN, [4]] # 24
  - [-1, 1, nn.Conv2d, [256, 1]] # 25
  - [23, 1, ChannelAttention_HSFPN, [4, False]] # 26
  - [[-1, -2], 1, Multiply, []] # 27
  - [[-1, 23], 1, Add_HSFPN, []] # 28
  - [-1, 2, A2C2f, [256, False, -1]] # 29 P4/16

  - [23, 1, nn.ConvTranspose2d, [256, 3, 2, 1, 1]] # 30
  - [7, 1, ChannelAttention_HSFPN, [4]] # 31
  - [-1, 1, nn.Conv2d, [256, 1]] # 32
  - [30, 1, ChannelAttention_HSFPN, [4, False]] # 33
  - [[-1, -2], 1, Multiply, []] # 34
  - [[-1, 30], 1, Add_HSFPN, []] # 35
  - [-1, 2, C3k2, [256, True]] # 36 P3/16

  - [20, 1, ChannelAttention_HSFPN, [4]] # 37
  - [-1, 1, nn.Conv2d, [256, 1]] # 38
  - [-1, 1, nn.ConvTranspose2d, [256, 3, 2, 1, 1]] # 39

  - [18, 1, ChannelAttention_HSFPN, [4]] # 40
  - [-1, 1, nn.Conv2d, [256, 1]] # 41
  - [39, 1, ChannelAttention_HSFPN, [4, False]] # 42
  - [[-1, -2], 1, Multiply, []] # 43
  - [[-1, 39], 1, Add_HSFPN, []] # 44
  - [-1, 2, A2C2f, [256, False, -1]] # 45 P4/16

  - [39, 1, nn.ConvTranspose2d, [256, 3, 2, 1, 1]] # 46
  - [16, 1, ChannelAttention_HSFPN, [4]] # 47
  - [-1, 1, nn.Conv2d, [256, 1]] # 48
  - [46, 1, ChannelAttention_HSFPN, [4, False]] # 49
  - [[-1, -2], 1, Multiply, []] # 50
  - [[-1, 46], 1, Add_HSFPN, []] # 51
  - [-1, 2, C3k2, [256, True]] # 52 P3/16

  - [[22, 38], 1, Concat, [1]]  # cat head P5  53
  - [[29, 45], 1, Concat, [1]]  # cat head P5  54
  - [[36, 52], 1, Concat, [1]]  # cat head P5  55

  - [[53, 54, 55], 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-HSFPN ：

YOLOv12-mid-HSFPN summary: 679 layers, 3,360,979 parameters, 3,360,963 gradients, 8.5 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         0  ultralytics.nn.modules.conv.Concat           [1]
 22             [9, 18]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 23            [11, 20]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 24                  -1  1    131072  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[512, 4]
 25                  -1  1     32832  torch.nn.modules.conv.Conv2d                 [512, 64, 1]
 26                  -1  1     36928  torch.nn.modules.conv.ConvTranspose2d        [64, 64, 3, 2, 1, 1]
 27                  22  1     32768  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[256, 4]
 28                  -1  1     16448  torch.nn.modules.conv.Conv2d                 [256, 64, 1]
 29                  26  1      2048  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[64, 4, False]
 30            [-1, -2]  1         0  ultralytics.nn.AddModules.HSFPN.Multiply     []
 31            [-1, 26]  1         0  ultralytics.nn.AddModules.HSFPN.Add_HSFPN    []
 32                  -1  1     17856  ultralytics.nn.AddModules.A2C2f.A2C2f        [64, 64, 1, False, -1]
 33                  26  1     36928  torch.nn.modules.conv.ConvTranspose2d        [64, 64, 3, 2, 1, 1]
 34                  21  1     32768  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[256, 4]
 35                  -1  1     16448  torch.nn.modules.conv.Conv2d                 [256, 64, 1]
 36                  33  1      2048  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[64, 4, False]
 37            [-1, -2]  1         0  ultralytics.nn.AddModules.HSFPN.Multiply     []
 38            [-1, 33]  1         0  ultralytics.nn.AddModules.HSFPN.Add_HSFPN    []
 39                  -1  1     22016  ultralytics.nn.modules.block.C3k2            [64, 64, 1, True]
 40        [25, 32, 39]  1    264211  ultralytics.nn.modules.head.Detect           [1, [64, 64, 64]]
YOLOv12-mid-HSFPN summary: 679 layers, 3,360,979 parameters, 3,360,963 gradients, 8.5 GFLOPs

YOLOv12-late-HSFPN ：

YOLOv12-late-HSFPN summary: 789 layers, 3,570,195 parameters, 3,570,179 gradients, 10.5 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     32768  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[256, 4]
 22                  -1  1     16448  torch.nn.modules.conv.Conv2d                 [256, 64, 1]
 23                  -1  1     36928  torch.nn.modules.conv.ConvTranspose2d        [64, 64, 3, 2, 1, 1]
 24                   9  1      8192  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[128, 4]
 25                  -1  1      8256  torch.nn.modules.conv.Conv2d                 [128, 64, 1]
 26                  23  1      2048  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[64, 4, False]
 27            [-1, -2]  1         0  ultralytics.nn.AddModules.HSFPN.Multiply     []
 28            [-1, 23]  1         0  ultralytics.nn.AddModules.HSFPN.Add_HSFPN    []
 29                  -1  1     17856  ultralytics.nn.AddModules.A2C2f.A2C2f        [64, 64, 1, False, -1]
 30                  23  1     36928  torch.nn.modules.conv.ConvTranspose2d        [64, 64, 3, 2, 1, 1]
 31                   7  1      8192  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[128, 4]
 32                  -1  1      8256  torch.nn.modules.conv.Conv2d                 [128, 64, 1]
 33                  30  1      2048  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[64, 4, False]
 34            [-1, -2]  1         0  ultralytics.nn.AddModules.HSFPN.Multiply     []
 35            [-1, 30]  1         0  ultralytics.nn.AddModules.HSFPN.Add_HSFPN    []
 36                  -1  1     22016  ultralytics.nn.modules.block.C3k2            [64, 64, 1, True]
 37                  20  1     32768  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[256, 4]
 38                  -1  1     16448  torch.nn.modules.conv.Conv2d                 [256, 64, 1]
 39                  -1  1     36928  torch.nn.modules.conv.ConvTranspose2d        [64, 64, 3, 2, 1, 1]
 40                  18  1      8192  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[128, 4]
 41                  -1  1      8256  torch.nn.modules.conv.Conv2d                 [128, 64, 1]
 42                  39  1      2048  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[64, 4, False]
 43            [-1, -2]  1         0  ultralytics.nn.AddModules.HSFPN.Multiply     []
 44            [-1, 39]  1         0  ultralytics.nn.AddModules.HSFPN.Add_HSFPN    []
 45                  -1  1     17856  ultralytics.nn.AddModules.A2C2f.A2C2f        [64, 64, 1, False, -1]
 46                  39  1     36928  torch.nn.modules.conv.ConvTranspose2d        [64, 64, 3, 2, 1, 1]
 47                  16  1      8192  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[128, 4]
 48                  -1  1      8256  torch.nn.modules.conv.Conv2d                 [128, 64, 1]
 49                  46  1      2048  ultralytics.nn.AddModules.HSFPN.ChannelAttention_HSFPN[64, 4, False]
 50            [-1, -2]  1         0  ultralytics.nn.AddModules.HSFPN.Multiply     []
 51            [-1, 46]  1         0  ultralytics.nn.AddModules.HSFPN.Add_HSFPN    []
 52                  -1  1     22016  ultralytics.nn.modules.block.C3k2            [64, 64, 1, True]
 53            [22, 38]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 54            [29, 45]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 55            [36, 52]  1         0  ultralytics.nn.modules.conv.Concat           [1]
 56        [53, 54, 55]  1    453715  ultralytics.nn.modules.head.Detect           [1, [128, 128, 128]]
YOLOv12-late-HSFPN summary: 789 layers, 3,570,195 parameters, 3,570,179 gradients, 10.5 GFLOPs