YOLO融合[CVPR2024]RAMiT中的下采样模块Downsizing

最新推荐文章于 2025-07-17 22:27:52 发布

今天炼丹了吗

最新推荐文章于 2025-07-17 22:27:52 发布

阅读量1k

点赞数 38

CC 4.0 BY-SA版权

文章标签： YOLO

本文链接：https://blog.youkuaiyun.com/StopAndGoyyy/article/details/149207777

YOLOv11v10v8使用教程： YOLOv11入门到入土使用教程

YOLOv11改进汇总贴：YOLOv11及自研模型更新汇总

《Reciprocal Attention Mixing Transformer for Lightweight Image Restoration》

一、模块介绍

论文链接：https://arxiv.org/pdf/2305.11474

代码链接：https://github.com/rami0205/RAMiT/tree/main

论文速览：

尽管近期许多研究在图像修复（IR）领域取得了进展，但它们往往存在参数过多的问题。另一个问题是，大多数基于 Transformer 的图像修复方法仅关注局部或全局特征，导致感受野受限或参数不足。为了解决这些问题，我们提出了一种轻量级网络——互反注意力混合 Transformer（RAMiT）。它采用了我们提出的维度互反注意力混合 Transformer（D-RAMiT）块，该块并行计算不同多头数量的二维自注意力。二维注意力相互补充对方的不足之处，然后进行混合。此外，我们引入了分层互反注意力混合（H-RAMi）层，以补偿像素级信息损失，并利用语义信息，同时保持高效的分层结构。此外，我们重新审视并修改了 MobileNet V2，以将高效卷积附加到我们提出的组件上。

总结：本文介绍其中Downsizing模块的使用方法。

⭐⭐本文二创模块仅更新于付费群中，往期免费教程可看下方链接⭐⭐

YOLOv11及自研模型更新汇总（含免费教程）文章浏览阅读366次，点赞3次，收藏4次。群文件2024/11/08日更新。,群文件2024/11/08日更新。_yolo11部署自己的数据集https://xy2668825911.blog.youkuaiyun.com/article/details/143633356

二、二创融合模块

2.1 相关代码

# https://blog.youkuaiyun.com/StopAndGoyyy?spm=1011.2124.3001.5343
# Reciprocal Attention Mixing Transformer for Lightweight Image Restoration
class MobiVari1(nn.Module):  # MobileNet v1 Variants
    def __init__(self, dim, kernel_size, stride, act=nn.LeakyReLU, out_dim=None):
        super(MobiVari1, self).__init__()
        self.dim = dim
        self.kernel_size = kernel_size
        self.out_dim = out_dim or dim

        self.dw_conv = nn.Conv2d(dim, dim, kernel_size, stride, kernel_size // 2, groups=dim)
        self.pw_conv = nn.Conv2d(dim, self.out_dim, 1, 1, 0)
        self.act = act()

    def forward(self, x):
        out = self.act(self.pw_conv(self.act(self.dw_conv(x)) + x))
        return out + x if self.dim == self.out_dim else out

    def flops(self, resolutions):
        H, W = resolutions
        flops = H * W * self.kernel_size * self.kernel_size * self.dim + H * W * 1 * 1 * self.dim * self.out_dim  # self.dw_conv + self.pw_conv
        return flops


class MobiVari2(MobiVari1):  # MobileNet v2 Variants
    def __init__(self, dim, kernel_size, stride, act=nn.LeakyReLU, out_dim=None, exp_factor=1.2, expand_groups=4):
        super(MobiVari2, self).__init__(dim, kernel_size, stride, act, out_dim)
        self.expand_groups = expand_groups
        expand_dim = int(dim * exp_factor)
        expand_dim = expand_dim + (expand_groups - expand_dim % expand_groups)
        self.expand_dim = expand_dim

        self.exp_conv = nn.Conv2d(dim, self.expand_dim, 1, 1, 0, groups=expand_groups)
        self.dw_conv = nn.Conv2d(expand_dim, expand_dim, kernel_size, stride, kernel_size // 2, groups=expand_dim)
        self.pw_conv = nn.Conv2d(expand_dim, self.out_dim, 1, 1, 0)

    def forward(self, x):
        x1 = self.act(self.exp_conv(x))
        out = self.pw_conv(self.act(self.dw_conv(x1) + x1))
        return out + x if self.dim == self.out_dim else out

    def flops(self, resolutions):
        H, W = resolutions
        flops = H * W * 1 * 1 * (self.dim // self.expand_groups) * self.expand_dim  # self.exp_conv
        flops += H * W * self.kernel_size * self.kernel_size * self.expand_dim  # self.dw_conv
        flops += H * W * 1 * 1 * self.expand_dim * self.out_dim  # self.pw_conv
        return flops


class ReshapeLayerNorm(nn.Module):
    def __init__(self, dim, norm_layer=nn.LayerNorm):
        super(ReshapeLayerNorm, self).__init__()

        self.dim = dim
        self.norm = norm_layer(dim)

    def forward(self, x):
        B, C, H, W = x.size()
        x = rearrange(x, 'b c h w -> b (h w) c')
        x = self.norm(x)
        x = rearrange(x, 'b (h w) c -> b c h w', h=H)
        return x

    def flops(self, resolutions):
        H, W = resolutions
        flops = 0
        flops += H * W * self.dim
        return flops


class Downsizing(nn.Module):
    """ Patch Merging Layer.

    Args:
        dim (int): Number of input dimension (channels).
        downsample_dim (int, optional): Number of output dimension (channels) (dim if not set).  Default: None
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self, dim, downsample_dim=None, norm_layer=ReshapeLayerNorm, mv_ver=1, mv_act=nn.LeakyReLU,
                 exp_factor=1.2, expand_groups=4):
        super(Downsizing, self).__init__()
        self.dim = dim
        self.downsample_dim = downsample_dim or dim
        self.norm = norm_layer(4 * dim)
        if mv_ver == 1:
            self.reduction = MobiVari1(4 * dim, 3, 1, act=mv_act, out_dim=self.downsample_dim)
        elif mv_ver == 2:
            self.reduction = MobiVari2(4 * dim, 3, 1, act=mv_act, out_dim=self.downsample_dim, exp_factor=exp_factor,
                                       expand_groups=expand_groups)

    def forward(self, x):
        # B, C, H, W = x.size()

        # Concat 2x2
        x0 = x[:, :, 0::2, 0::2]  # [B, C, H/2, W/2], top-left
        x1 = x[:, :, 0::2, 1::2]  # [B, C, H/2, W/2], top-right
        x2 = x[:, :, 1::2, 0::2]  # [B, C, H/2, W/2], bottom-left
        x3 = x[:, :, 1::2, 1::2]  # [B, C, H/2, W/2], bottom-right
        x = torch.cat([x0, x1, x2, x3], dim=1)  # [B, 4C, H/2, W/2]
        return self.reduction(self.norm(x))  # [B, C, H/2, W/2]

    def flops(self, resolutions):
        H, W = resolutions
        flops = self.norm.flops((H // 2, W // 2)) + self.reduction.flops((H // 2, W // 2))
        return flops

2.2更改yaml文件 (以自研模型为例)

yam文件解读：YOLO系列 “.yaml“文件解读_yolo yaml文件-优快云博客

打开更改ultralytics/cfg/models/11路径下的YOLOv11.yaml文件，替换原有模块。

# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLO11 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
# ⭐⭐Powered by https://blog.youkuaiyun.com/StopAndGoyyy,  技术指导QQ:2668825911⭐⭐

# 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: 377 layers, 2,249,525 parameters, 2,249,509 gradients, 8.7 GFLOPs/258 layers, 2,219,405 parameters, 0 gradients, 8.5 GFLOPs
  s: [0.50, 0.50, 1024] # summary: 377 layers, 8,082,389 parameters, 8,082,373 gradients, 29.8 GFLOPs/258 layers, 7,972,885 parameters, 0 gradients, 29.2 GFLOPs
  m: [0.50, 1.00, 512] # summary:  377 layers, 20,370,221 parameters, 20,370,205 gradients, 103.0 GFLOPs/258 layers, 20,153,773 parameters, 0 gradients, 101.2 GFLOPs
  l: [1.00, 1.00, 512] # summary: 521 layers, 23,648,717 parameters, 23,648,701 gradients, 124.5 GFLOPs/330 layers, 23,226,989 parameters, 0 gradients, 121.2 GFLOPs
  x: [1.00, 1.50, 512] # summary: 521 layers, 53,125,237 parameters, 53,125,221 gradients, 278.9 GFLOPs/330 layers, 52,191,589 parameters, 0 gradients, 272.1 GFLOPs

#  n: [0.33, 0.25, 1024]
#  s: [0.50, 0.50, 1024]
#  m: [0.67, 0.75, 768]
#  l: [1.00, 1.00, 512]
#  x: [1.00, 1.25, 512]
# YOLO11n backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  - [-1, 1, Downsizing, []] # 1-P2/4
  - [-1, 2, RCRep2A, [128, False, 0.25]]
  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  - [-1, 4, RCRep2A, [256, False, 0.25]]
  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  - [-1, 4, RCRep2A, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  - [-1, 2, RCRep2A, [1024, True]]
  - [-1, 1, SPPF_WD, [1024, 7]] # 9

# YOLO11n head
head:
  - [[3, 5, 7], 1, align_3In, [256, 1]] # 10
  - [[4, 6, 9], 1, align_3In, [256, 1]] # 11

  - [[-1, -2], 1, Concat, [1]] #12  cat

  - [-1, 1, RepVGGBlocks, []] #13

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]] #14
  - [[-1, 4], 1, Concat, [1]] #15 cat

  - [-1, 1, Conv, [256, 3]] # 16
  - [13, 1, Conv, [512, 3]] #17
  - [13, 1, Conv, [1024, 3, 2]] #18

  - [[16, 17, 18], 1, Detect, [nc]] # Detect(P3, P4, P5)



# ⭐⭐Powered by https://blog.youkuaiyun.com/StopAndGoyyy,  技术指导QQ:2668825911⭐⭐

2.3 修改train.py文件

创建Train脚本用于训练。

from ultralytics.models import YOLO
import os
os.environ['KMP_DUPLICATE_LIB_OK'] = 'TRUE'

if __name__ == '__main__':
    model = YOLO(model='ultralytics/cfg/models/xy_YOLO/xy_yolov1.yaml')
    # model = YOLO(model='ultralytics/cfg/models/11/yolo11l.yaml')
    model.train(data='./datasets/data.yaml', epochs=1, batch=1, device='0', imgsz=320, workers=1, cache=False,
                amp=True, mosaic=False, project='run/train', name='exp',)