RT-DETR改进策略【注意力机制篇】| 2024 SCSA-CBAM 空间和通道的协同注意模块（含HGBlock二次创新）
一、本文介绍
本文记录的是
基于SCSA-CBAM注意力模块的RT-DETR目标检测改进方法研究
。现有注意力方法在空间-通道协同方面未充分挖掘其潜力，缺乏对多语义信息的充分利用来引导特征和缓解语义差异。
SCSA-CBAM注意力模块
构建一个空间-通道协同机制，
使空间注意力引导通道注意力增强综合学习，通道注意力从多语义水平调节更丰富的空间特定模式。
二、SCSA原理
SCSA ：空间注意与通道注意的协同效应研究
SCSA（Spatial and Channel Synergistic Attention）
是一种新颖的、即插即用的空间和通道协同注意力机制，其设计的原理和优势如下：
2.1 原理
-
Shared Multi - Semantic Spatial Attention（SMSA）
：
- 空间和通道分解 ：将输入X沿高度和宽度维度分解，应用全局平均池化创建两个单向1D序列结构，然后将特征集划分为K个独立的子特征，每个子特征具有C / K个通道，便于高效提取多语义空间信息。
- 轻量级卷积策略 ：在四个子特征中应用核大小为3、5、7和9的深度一维卷积，以捕获不同的语义空间结构，并使用共享卷积来对齐，解决分解特征和应用一维卷积导致的有限感受野问题。使用Group Normalization对不同语义子特征进行归一化，最后使用Sigmoid激活函数生成空间注意力。
-
Progressive Channel - wise Self - Attention（PCSA）
：
- 受ViT利用MHSA建模空间注意力中不同token之间相似性的启发，结合SMSA调制的空间先验来计算通道间相似性。
- 采用渐进压缩方法来保留和利用SMSA提取的多语义空间信息，并减少MHSA的计算成本。
- 具体实现过程包括池化、映射生成查询、键和值，进行注意力计算等。
- 协同效应 ：通过简单的串行连接集成SMSA和PCSA模块，空间注意力从每个特征中提取多语义空间信息，为通道注意力计算提供精确的空间先验；通道注意力利用整体特征图X来细化局部子特征的语义理解，缓解SMSA中多尺度卷积引起的语义差异。同时，不采用通道压缩，防止关键特征丢失。
2.2 优势
- 高效的SMSA ：利用多尺度深度共享1D卷积捕获每个特征通道的多语义空间信息，有效整合全局上下文依赖和多语义空间先验。
- PCSA缓解语义差异 ：使用SMSA计算引导的压缩空间知识来计算通道相似性和贡献，缓解空间结构中的语义差异。
- 协同效应 ：通过维度解耦、轻量级多语义引导和语义差异缓解来探索协同效应，在各种视觉任务和复杂场景中优于当前最先进的注意力机制。
-
实验验证优势
：
- 在图像分类任务中，SCSA在不同规模的网络中实现了最高的Top - 1准确率，且参数和计算复杂度较低，基于ResNet的推理速度仅次于CA，在准确性、速度和模型复杂度之间实现了较好的平衡。
- 在目标检测任务中，在各种检测器、模型大小和对象尺度上优于其他先进的注意力方法，在复杂场景（如小目标、黑暗环境和红外场景）中进一步证明了其有效性和泛化能力。
- 在分割任务中，基于多语义空间信息，在像素级任务中表现出色，显著优于其他注意力方法。
- 可视化分析 ：SCSA在相似的感受野条件下能明显关注多个关键区域，最大限度地减少关键信息丢失，为最终的下游任务提供丰富的特征信息，其协同设计在空间和通道域注意力计算中保留了关键信息，具有更优越的表示能力。
- 其他分析 ：SCSA具有更大的有效感受野，有利于网络利用丰富的上下文信息进行集体决策，从而提升性能；在计算复杂度方面，当模型宽度适当时，SCSA可以以线性复杂度进行推理；在推理吞吐量评估中，虽然SCSA比纯通道注意力略慢，但优于大多数混合注意力机制，在模型复杂性、推理速度和准确性之间实现了优化平衡。
论文： https://arxiv.org/pdf/2407.05128
源码： https://github.com/HZAI-ZJNU/SCSA
三、SCSA的实现代码
SCSA模块
的实现代码如下：
import torch
import torch.nn as nn
import typing as t
from einops import rearrange
from mmengine.model import BaseModule
from ultralytics.nn.modules.conv import LightConv
class SCSA(BaseModule):
def __init__(
self,
dim: int,
head_num: int=4,
window_size: int = 7,
group_kernel_sizes: t.List[int] = [3, 5, 7, 9],
qkv_bias: bool = False,
fuse_bn: bool = False,
norm_cfg: t.Dict = dict(type='BN'),
act_cfg: t.Dict = dict(type='ReLU'),
down_sample_mode: str = 'avg_pool',
attn_drop_ratio: float = 0.,
gate_layer: str = 'sigmoid',
):
super(SCSA, self).__init__()
self.dim = dim
self.head_num = head_num
self.head_dim = dim // head_num
self.scaler = self.head_dim ** -0.5
self.group_kernel_sizes = group_kernel_sizes
self.window_size = window_size
self.qkv_bias = qkv_bias
self.fuse_bn = fuse_bn
self.down_sample_mode = down_sample_mode
assert self.dim // 4, 'The dimension of input feature should be divisible by 4.'
self.group_chans = group_chans = self.dim // 4
self.local_dwc = nn.Conv1d(group_chans, group_chans, kernel_size=group_kernel_sizes[0],
padding=group_kernel_sizes[0] // 2, groups=group_chans)
self.global_dwc_s = nn.Conv1d(group_chans, group_chans, kernel_size=group_kernel_sizes[1],
padding=group_kernel_sizes[1] // 2, groups=group_chans)
self.global_dwc_m = nn.Conv1d(group_chans, group_chans, kernel_size=group_kernel_sizes[2],
padding=group_kernel_sizes[2] // 2, groups=group_chans)
self.global_dwc_l = nn.Conv1d(group_chans, group_chans, kernel_size=group_kernel_sizes[3],
padding=group_kernel_sizes[3] // 2, groups=group_chans)
self.sa_gate = nn.Softmax(dim=2) if gate_layer == 'softmax' else nn.Sigmoid()
self.norm_h = nn.GroupNorm(4, dim)
self.norm_w = nn.GroupNorm(4, dim)
self.conv_d = nn.Identity()
self.norm = nn.GroupNorm(1, dim)
self.q = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, bias=qkv_bias, groups=dim)
self.k = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, bias=qkv_bias, groups=dim)
self.v = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=1, bias=qkv_bias, groups=dim)
self.attn_drop = nn.Dropout(attn_drop_ratio)
self.ca_gate = nn.Softmax(dim=1) if gate_layer == 'softmax' else nn.Sigmoid()
if window_size == -1:
self.down_func = nn.AdaptiveAvgPool2d((1, 1))
else:
if down_sample_mode == 'recombination':
self.down_func = self.space_to_chans
# dimensionality reduction
self.conv_d = nn.Conv2d(in_channels=dim * window_size ** 2, out_channels=dim, kernel_size=1, bias=False)
elif down_sample_mode == 'avg_pool':
self.down_func = nn.AvgPool2d(kernel_size=(window_size, window_size), stride=window_size)
elif down_sample_mode == 'max_pool':
self.down_func = nn.MaxPool2d(kernel_size=(window_size, window_size), stride=window_size)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
The dim of x is (B, C, H, W)
"""
# Spatial attention priority calculation
b, c, h_, w_ = x.size()
# (B, C, H)
x_h = x.mean(dim=3)
l_x_h, g_x_h_s, g_x_h_m, g_x_h_l = torch.split(x_h, self.group_chans, dim=1)
# (B, C, W)
x_w = x.mean(dim=2)
l_x_w, g_x_w_s, g_x_w_m, g_x_w_l = torch.split(x_w, self.group_chans, dim=1)
x_h_attn = self.sa_gate(self.norm_h(torch.cat((
self.local_dwc(l_x_h),
self.global_dwc_s(g_x_h_s),
self.global_dwc_m(g_x_h_m),
self.global_dwc_l(g_x_h_l),
), dim=1)))
x_h_attn = x_h_attn.view(b, c, h_, 1)
x_w_attn = self.sa_gate(self.norm_w(torch.cat((
self.local_dwc(l_x_w),
self.global_dwc_s(g_x_w_s),
self.global_dwc_m(g_x_w_m),
self.global_dwc_l(g_x_w_l)
), dim=1)))
x_w_attn = x_w_attn.view(b, c, 1, w_)
x = x * x_h_attn * x_w_attn
# Channel attention based on self attention
# reduce calculations
y = self.down_func(x)
y = self.conv_d(y)
_, _, h_, w_ = y.size()
# normalization first, then reshape -> (B, H, W, C) -> (B, C, H * W) and generate q, k and v
y = self.norm(y)
q = self.q(y)
k = self.k(y)
v = self.v(y)
# (B, C, H, W) -> (B, head_num, head_dim, N)
q = rearrange(q, 'b (head_num head_dim) h w -> b head_num head_dim (h w)', head_num=int(self.head_num),
head_dim=int(self.head_dim))
k = rearrange(k, 'b (head_num head_dim) h w -> b head_num head_dim (h w)', head_num=int(self.head_num),
head_dim=int(self.head_dim))
v = rearrange(v, 'b (head_num head_dim) h w -> b head_num head_dim (h w)', head_num=int(self.head_num),
head_dim=int(self.head_dim))
# (B, head_num, head_dim, head_dim)
attn = q @ k.transpose(-2, -1) * self.scaler
attn = self.attn_drop(attn.softmax(dim=-1))
# (B, head_num, head_dim, N)
attn = attn @ v
# (B, C, H_, W_)
attn = rearrange(attn, 'b head_num head_dim (h w) -> b (head_num head_dim) h w', h=int(h_), w=int(w_))
# (B, C, 1, 1)
attn = attn.mean((2, 3), keepdim=True)
attn = self.ca_gate(attn)
return attn * x
def autopad(k, p=None, d=1): # kernel, padding, dilation
"""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] # actual kernel-size
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):
"""Standard convolution with args(ch_in, ch_out, kernel, stride, padding, groups, dilation, activation)."""
default_act = nn.SiLU() # default activation
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
"""Initialize Conv layer with given arguments including activation."""
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):
"""Apply convolution, batch normalization and activation to input tensor."""
return self.act(self.bn(self.conv(x)))
def forward_fuse(self, x):
"""Perform transposed convolution of 2D data."""
return self.act(self.conv(x))
class HGBlock_SCSA(nn.Module):
"""
HG_Block of PPHGNetV2 with 2 convolutions and LightConv.
https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/modeling/backbones/hgnet_v2.py
"""
def __init__(self, c1, cm, c2, k=3, n=6, lightconv=False, shortcut=False, act=nn.ReLU()):
"""Initializes a CSP Bottleneck with 1 convolution using specified input and output channels."""
super().__init__()
block = LightConv if lightconv else Conv
self.m = nn.ModuleList(block(c1 if i == 0 else cm, cm, k=k, act=act) for i in range(n))
self.sc = Conv(c1 + n * cm, c2 // 2, 1, 1, act=act) # squeeze conv
self.ec = Conv(c2 // 2, c2, 1, 1, act=act) # excitation conv
self.add = shortcut and c1 == c2
self.cv = SCSA(c2)
def forward(self, x):
"""Forward pass of a PPHGNetV2 backbone layer."""
y = [x]
y.extend(m(y[-1]) for m in self.m)
y = self.cv(self.ec(self.sc(torch.cat(y, 1))))
return y + x if self.add else y
四、创新模块
4.1 改进点1
模块改进方法
1️⃣：直接加入
SCSA模块
。
SCSA模块
添加后如下：
注意❗：需要声明的模块名称为：
SCSA
。
4.2 改进点2⭐
模块改进方法
2️⃣：基于
SCSA模块
的
HGBlock
。
第二种改进方法是对
RT-DETR
中的
HGBlock模块
进行改进。
SCSA
的协同设计能够在空间和通道域注意力计算中保留了关键信息，最大限度地减少关键信息丢失，使
HGBlock模块
具有更优越的表示能力。
改进代码如下：
class HGBlock_SCSA(nn.Module):
"""
HG_Block of PPHGNetV2 with 2 convolutions and LightConv.
https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/modeling/backbones/hgnet_v2.py
"""
def __init__(self, c1, cm, c2, k=3, n=6, lightconv=False, shortcut=False, act=nn.ReLU()):
"""Initializes a CSP Bottleneck with 1 convolution using specified input and output channels."""
super().__init__()
block = LightConv if lightconv else Conv
self.m = nn.ModuleList(block(c1 if i == 0 else cm, cm, k=k, act=act) for i in range(n))
self.sc = Conv(c1 + n * cm, c2 // 2, 1, 1, act=act) # squeeze conv
self.ec = Conv(c2 // 2, c2, 1, 1, act=act) # excitation conv
self.add = shortcut and c1 == c2
self.cv = SCSA(c2)
def forward(self, x):
"""Forward pass of a PPHGNetV2 backbone layer."""
y = [x]
y.extend(m(y[-1]) for m in self.m)
y = self.cv(self.ec(self.sc(torch.cat(y, 1))))
return y + x if self.add else y
注意❗：需要声明的模块名称为：
HGBlock_SCSA
。
五、添加步骤
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
SCSA.py
，将
第三节
中的代码粘贴到此处
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 .SCSA import *
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
其次：在
parse_model函数
中注册
SCSA
和
HGBlock_SCSA
模块
六、yaml模型文件
6.1 模型改进版本一
在代码配置完成后，配置模型的YAML文件。
此处以
ultralytics/cfg/models/rt-detr/rtdetr-l.yaml
为例，在同目录下创建一个用于自己数据集训练的模型文件
rtdetr-l-SCSA.yaml
。
将
rtdetr-l.yaml
中的内容复制到
rtdetr-l-SCSA.yaml
文件下，修改
nc
数量等于自己数据中目标的数量。
在骨干网络中添加
SCSA模块
，
只需要填入一个参数，通道数
。
# Ultralytics YOLO 🚀, AGPL-3.0 license
# RT-DETR-l object detection model with P3-P5 outputs. For details see https://docs.ultralytics.com/models/rtdetr
# Parameters
nc: 1 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n-cls.yaml' will call yolov8-cls.yaml with scale 'n'
# [depth, width, max_channels]
l: [1.00, 1.00, 1024]
backbone:
# [from, repeats, module, args]
- [-1, 1, HGStem, [32, 48]] # 0-P2/4
- [-1, 6, HGBlock, [48, 128, 3]] # stage 1
- [-1, 1, DWConv, [128, 3, 2, 1, False]] # 2-P3/8
- [-1, 6, HGBlock, [96, 512, 3]] # stage 2
- [-1, 1, DWConv, [512, 3, 2, 1, False]] # 4-P4/16
- [-1, 6, HGBlock, [192, 1024, 5, True, False]] # cm, c2, k, light, shortcut
- [-1, 6, HGBlock, [192, 1024, 5, True, True]]
- [-1, 6, HGBlock, [192, 1024, 5, True, True]] # stage 3
- [-1, 1, DWConv, [1024, 3, 2, 1, False]] # 8-P5/32
- [-1, 1, SCSA, [1024]] # stage 4
- [-1, 6, HGBlock, [384, 2048, 5, True, False]] # stage 4
head:
- [-1, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 10 input_proj.2
- [-1, 1, AIFI, [1024, 8]]
- [-1, 1, Conv, [256, 1, 1]] # 12, Y5, lateral_convs.0
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [7, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 14 input_proj.1
- [[-2, -1], 1, Concat, [1]]
- [-1, 3, RepC3, [256]] # 16, fpn_blocks.0
- [-1, 1, Conv, [256, 1, 1]] # 17, Y4, lateral_convs.1
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [3, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 19 input_proj.0
- [[-2, -1], 1, Concat, [1]] # cat backbone P4
- [-1, 3, RepC3, [256]] # X3 (21), fpn_blocks.1
- [-1, 1, Conv, [256, 3, 2]] # 22, downsample_convs.0
- [[-1, 18], 1, Concat, [1]] # cat Y4
- [-1, 3, RepC3, [256]] # F4 (24), pan_blocks.0
- [-1, 1, Conv, [256, 3, 2]] # 25, downsample_convs.1
- [[-1, 13], 1, Concat, [1]] # cat Y5
- [-1, 3, RepC3, [256]] # F5 (27), pan_blocks.1
- [[22, 25, 28], 1, RTDETRDecoder, [nc]] # Detect(P3, P4, P5)
6.2 模型改进版本二⭐
此处同样以
ultralytics/cfg/models/rt-detr/rtdetr-l.yaml
为例，在同目录下创建一个用于自己数据集训练的模型文件
rtdetr-l-HGBlock_SCSA.yaml
。
将
rtdetr-l.yaml
中的内容复制到
rtdetr-l-HGBlock_SCSA.yaml
文件下，修改
nc
数量等于自己数据中目标的数量。
📌 模型的修改方法是将
骨干网络
中的部分
HGBlock模块
替换成
HGBlock_SCSA模块
。
# Ultralytics YOLO 🚀, AGPL-3.0 license
# RT-DETR-l object detection model with P3-P5 outputs. For details see https://docs.ultralytics.com/models/rtdetr
# Parameters
nc: 1 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n-cls.yaml' will call yolov8-cls.yaml with scale 'n'
# [depth, width, max_channels]
l: [1.00, 1.00, 1024]
backbone:
# [from, repeats, module, args]
- [-1, 1, HGStem, [32, 48]] # 0-P2/4
- [-1, 6, HGBlock, [48, 128, 3]] # stage 1
- [-1, 1, DWConv, [128, 3, 2, 1, False]] # 2-P3/8
- [-1, 6, HGBlock, [96, 512, 3]] # stage 2
- [-1, 1, DWConv, [512, 3, 2, 1, False]] # 4-P4/16
- [-1, 6, HGBlock_SCSA, [192, 1024, 5, True, False]] # cm, c2, k, light, shortcut
- [-1, 6, HGBlock_SCSA, [192, 1024, 5, True, True]]
- [-1, 6, HGBlock_SCSA, [192, 1024, 5, True, True]] # stage 3
- [-1, 1, DWConv, [1024, 3, 2, 1, False]] # 8-P5/32
- [-1, 6, HGBlock, [384, 2048, 5, True, False]] # stage 4
head:
- [-1, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 10 input_proj.2
- [-1, 1, AIFI, [1024, 8]]
- [-1, 1, Conv, [256, 1, 1]] # 12, Y5, lateral_convs.0
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [7, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 14 input_proj.1
- [[-2, -1], 1, Concat, [1]]
- [-1, 3, RepC3, [256]] # 16, fpn_blocks.0
- [-1, 1, Conv, [256, 1, 1]] # 17, Y4, lateral_convs.1
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [3, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 19 input_proj.0
- [[-2, -1], 1, Concat, [1]] # cat backbone P4
- [-1, 3, RepC3, [256]] # X3 (21), fpn_blocks.1
- [-1, 1, Conv, [256, 3, 2]] # 22, downsample_convs.0
- [[-1, 17], 1, Concat, [1]] # cat Y4
- [-1, 3, RepC3, [256]] # F4 (24), pan_blocks.0
- [-1, 1, Conv, [256, 3, 2]] # 25, downsample_convs.1
- [[-1, 12], 1, Concat, [1]] # cat Y5
- [-1, 3, RepC3, [256]] # F5 (27), pan_blocks.1
- [[21, 24, 27], 1, RTDETRDecoder, [nc]] # Detect(P3, P4, P5)
七、成功运行结果
分别打印网络模型可以看到
SCSA
和
HGBlock_SCSA
已经加入到模型中，并可以进行训练了。
rtdetr-l-SCSA ：
rtdetr-l-SCSA summary: 697 layers, 32,824,515 parameters, 32,824,515 gradients, 108.0 GFLOPs
from n params module arguments
0 -1 1 25248 ultralytics.nn.modules.block.HGStem [3, 32, 48]
1 -1 6 155072 ultralytics.nn.modules.block.HGBlock [48, 48, 128, 3, 6]
2 -1 1 1408 ultralytics.nn.modules.conv.DWConv [128, 128, 3, 2, 1, False]
3 -1 6 839296 ultralytics.nn.modules.block.HGBlock [128, 96, 512, 3, 6]
4 -1 1 5632 ultralytics.nn.modules.conv.DWConv [512, 512, 3, 2, 1, False]
5 -1 6 1695360 ultralytics.nn.modules.block.HGBlock [512, 192, 1024, 5, 6, True, False]
6 -1 6 2055808 ultralytics.nn.modules.block.HGBlock [1024, 192, 1024, 5, 6, True, True]
7 -1 6 2055808 ultralytics.nn.modules.block.HGBlock [1024, 192, 1024, 5, 6, True, True]
8 -1 1 11264 ultralytics.nn.modules.conv.DWConv [1024, 1024, 3, 2, 1, False]
9 -1 1 16384 ultralytics.nn.AddModules.SCSA.SCSA [1024, 1024]
10 -1 6 6708480 ultralytics.nn.modules.block.HGBlock [1024, 384, 2048, 5, 6, True, False]
11 -1 1 524800 ultralytics.nn.modules.conv.Conv [2048, 256, 1, 1, None, 1, 1, False]
12 -1 1 789760 ultralytics.nn.modules.transformer.AIFI [256, 1024, 8]
13 -1 1 66048 ultralytics.nn.modules.conv.Conv [256, 256, 1, 1]
14 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
15 7 1 262656 ultralytics.nn.modules.conv.Conv [1024, 256, 1, 1, None, 1, 1, False]
16 [-2, -1] 1 0 ultralytics.nn.modules.conv.Concat [1]
17 -1 3 2232320 ultralytics.nn.modules.block.RepC3 [512, 256, 3]
18 -1 1 66048 ultralytics.nn.modules.conv.Conv [256, 256, 1, 1]
19 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
20 3 1 131584 ultralytics.nn.modules.conv.Conv [512, 256, 1, 1, None, 1, 1, False]
21 [-2, -1] 1 0 ultralytics.nn.modules.conv.Concat [1]
22 -1 3 2232320 ultralytics.nn.modules.block.RepC3 [512, 256, 3]
23 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
24 [-1, 18] 1 0 ultralytics.nn.modules.conv.Concat [1]
25 -1 3 2232320 ultralytics.nn.modules.block.RepC3 [512, 256, 3]
26 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
27 [-1, 13] 1 0 ultralytics.nn.modules.conv.Concat [1]
28 -1 3 2232320 ultralytics.nn.modules.block.RepC3 [512, 256, 3]
29 [22, 25, 28] 1 7303907 ultralytics.nn.modules.head.RTDETRDecoder [1, [256, 256, 256]]
rtdetr-l-SCSA summary: 697 layers, 32,824,515 parameters, 32,824,515 gradients, 108.0 GFLOPs
rtdetr-l-HGBlock_SCSA ：
rtdetr-l-HGBlock_SCSA summary: 730 layers, 32,857,283 parameters, 32,857,283 gradients, 108.0 GFLOPs
from n params module arguments
0 -1 1 25248 ultralytics.nn.modules.block.HGStem [3, 32, 48]
1 -1 6 155072 ultralytics.nn.modules.block.HGBlock [48, 48, 128, 3, 6]
2 -1 1 1408 ultralytics.nn.modules.conv.DWConv [128, 128, 3, 2, 1, False]
3 -1 6 839296 ultralytics.nn.modules.block.HGBlock [128, 96, 512, 3, 6]
4 -1 1 5632 ultralytics.nn.modules.conv.DWConv [512, 512, 3, 2, 1, False]
5 -1 6 1711744 ultralytics.nn.AddModules.SCSA.HGBlock_SCSA [512, 192, 1024, 5, 6, True, False]
6 -1 6 2072192 ultralytics.nn.AddModules.SCSA.HGBlock_SCSA [1024, 192, 1024, 5, 6, True, True]
7 -1 6 2072192 ultralytics.nn.AddModules.SCSA.HGBlock_SCSA [1024, 192, 1024, 5, 6, True, True]
8 -1 1 11264 ultralytics.nn.modules.conv.DWConv [1024, 1024, 3, 2, 1, False]
9 -1 6 6708480 ultralytics.nn.modules.block.HGBlock [1024, 384, 2048, 5, 6, True, False]
10 -1 1 524800 ultralytics.nn.modules.conv.Conv [2048, 256, 1, 1, None, 1, 1, False]
11 -1 1 789760 ultralytics.nn.modules.transformer.AIFI [256, 1024, 8]
12 -1 1 66048 ultralytics.nn.modules.conv.Conv [256, 256, 1, 1]
13 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
14 7 1 262656 ultralytics.nn.modules.conv.Conv [1024, 256, 1, 1, None, 1, 1, False]
15 [-2, -1] 1 0 ultralytics.nn.modules.conv.Concat [1]
16 -1 3 2232320 ultralytics.nn.modules.block.RepC3 [512, 256, 3]
17 -1 1 66048 ultralytics.nn.modules.conv.Conv [256, 256, 1, 1]
18 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
19 3 1 131584 ultralytics.nn.modules.conv.Conv [512, 256, 1, 1, None, 1, 1, False]
20 [-2, -1] 1 0 ultralytics.nn.modules.conv.Concat [1]
21 -1 3 2232320 ultralytics.nn.modules.block.RepC3 [512, 256, 3]
22 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
23 [-1, 17] 1 0 ultralytics.nn.modules.conv.Concat [1]
24 -1 3 2232320 ultralytics.nn.modules.block.RepC3 [512, 256, 3]
25 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
26 [-1, 12] 1 0 ultralytics.nn.modules.conv.Concat [1]
27 -1 3 2232320 ultralytics.nn.modules.block.RepC3 [512, 256, 3]
28 [21, 24, 27] 1 7303907 ultralytics.nn.modules.head.RTDETRDecoder [1, [256, 256, 256]]
rtdetr-l-HGBlock_SCSA summary: 730 layers, 32,857,283 parameters, 32,857,283 gradients, 108.0 GFLOPs