学习资源站

YOLOv11改进-Conv_卷积篇-2024最新ECCV最新大感受野的小波卷积WTConv助力YOLOv11有效涨点

一、本文介绍

本文给大家带来的改进机制是一种新的 卷积层 ， 称为WTConv（小波卷积层） ，它利用 小波变换 （WT）来解决 卷积神经网络 （CNN）在实现大感受野时遇到的过度参数化问题。WTConv的主要目的是通过对输入数据的不同频率带进行处理，使CNN能够更有效地捕捉局部和全局特征，WTConv成功解决了CNN在感受野扩展中的参数膨胀问题，提供了一种更为高效、鲁棒且易于集成的卷积层解决方案，我将其用于二次创新YOLOv11中的C3k2机制可以减少一定参数量和计算量，达到一个可观的轻量化作用 （这种小波Conv对于目前的创新角度来说是非常流行的） 。

二、原理介绍

​ 官方论文地址： 官方论文地址点击此处即可跳转

官方代码地址： 官方代码地址点击此处即可跳转

​

这篇名为《用于大感受野的小波卷积》的文章提出了一种新的卷积层，称为WTConv（小波卷积层），它利用小波变换（WT）来解决卷积神经网络（CNN）在实现大感受野时遇到的过度参数化问题。WTConv的主要目的是通过对输入数据的不同频率带进行处理，使CNN能够更有效地捕捉局部和全局特征，而传统的CNN主要只能处理局部特征。

以下是文章的主要内容总结：

1. 问题背景：传统的CNN受限于卷积核的大小，难以有效捕捉全局上下文信息。尽管近年来通过增大卷积核（如视觉 Transformer ）的尝试有所进展，但这通常会导致参数数量激增， 模型 性能饱和。

2. 提出的解决方案（WTConv）：WTConv利用小波变换，通过多频率响应扩展卷积感受野，并在不同频率范围内执行小核卷积操作。通过小波分解，模型可以在更大范围内捕捉低频信息，同时避免模型的过度参数化。

3. 主要优势：

  • 参数增长缓慢：与传统方法相比，WTConv的参数数量仅随感受野大小对数级别增长，而不是平方增长。
  • 感受野扩大：WTConv通过层级小波分解，能够在不增加大量参数的情况下显著扩大CNN的感受野。
  • 形状偏差提升：WTConv层对图像中的低频信息更敏感，从而增强了CNN对形状而非纹理的响应能力。

总的来说，WTConv成功解决了CNN在感受野扩展中的参数膨胀问题，提供了一种更为高效、鲁棒且易于集成的卷积层解决方案。

三、核心代码

核心代码的使用方式看章节四！ 

  1. import torch.nn as nn
  2. from functools import partial
  3. import pywt
  4. import pywt.data
  5. import torch
  6. import torch.nn.functional as F
  9. def create_wavelet_filter(wave, in_size, out_size, type=torch.float):
  10.     w = pywt.Wavelet(wave)
  11.     dec_hi = torch.tensor(w.dec_hi[::-1], dtype=type)
  12.     dec_lo = torch.tensor(w.dec_lo[::-1], dtype=type)
  13.     dec_filters = torch.stack([dec_lo.unsqueeze(0) * dec_lo.unsqueeze(1),
  14.                                dec_lo.unsqueeze(0) * dec_hi.unsqueeze(1),
  15.                                dec_hi.unsqueeze(0) * dec_lo.unsqueeze(1),
  16.                                dec_hi.unsqueeze(0) * dec_hi.unsqueeze(1)], dim=0)
  18.     dec_filters = dec_filters[:, None].repeat(in_size, 1, 1, 1)
  20.     rec_hi = torch.tensor(w.rec_hi[::-1], dtype=type).flip(dims=[0])
  21.     rec_lo = torch.tensor(w.rec_lo[::-1], dtype=type).flip(dims=[0])
  22.     rec_filters = torch.stack([rec_lo.unsqueeze(0) * rec_lo.unsqueeze(1),
  23.                                rec_lo.unsqueeze(0) * rec_hi.unsqueeze(1),
  24.                                rec_hi.unsqueeze(0) * rec_lo.unsqueeze(1),
  25.                                rec_hi.unsqueeze(0) * rec_hi.unsqueeze(1)], dim=0)
  27.     rec_filters = rec_filters[:, None].repeat(out_size, 1, 1, 1)
  29.     return dec_filters, rec_filters
  31. def wavelet_transform(x, filters):
  32.     b, c, h, w = x.shape
  33.     pad = (filters.shape[2] // 2 - 1, filters.shape[3] // 2 - 1)
  34.     x = F.conv2d(x, filters, stride=2, groups=c, padding=pad)
  35.     x = x.reshape(b, c, 4, h // 2, w // 2)
  36.     return x
  39. def inverse_wavelet_transform(x, filters):
  40.     b, c, _, h_half, w_half = x.shape
  41.     pad = (filters.shape[2] // 2 - 1, filters.shape[3] // 2 - 1)
  42.     x = x.reshape(b, c * 4, h_half, w_half)
  43.     x = F.conv_transpose2d(x, filters, stride=2, groups=c, padding=pad)
  44.     return x
  47. class WTConv2d(nn.Module):
  48.     def __init__(self, in_channels, out_channels, kernel_size=5, stride=1, bias=True, wt_levels=1, wt_type='db1'):
  49.         super(WTConv2d, self).__init__()
  51.         assert in_channels == out_channels
  53.         self.in_channels = in_channels
  54.         self.wt_levels = wt_levels
  55.         self.stride = stride
  56.         self.dilation = 1
  58.         self.wt_filter, self.iwt_filter = create_wavelet_filter(wt_type, in_channels, in_channels, torch.float)
  59.         self.wt_filter = nn.Parameter(self.wt_filter, requires_grad=False)
  60.         self.iwt_filter = nn.Parameter(self.iwt_filter, requires_grad=False)
  62.         self.wt_function = partial(wavelet_transform, filters=self.wt_filter)
  63.         self.iwt_function = partial(inverse_wavelet_transform, filters=self.iwt_filter)
  65.         self.base_conv = nn.Conv2d(in_channels, in_channels, kernel_size, padding='same', stride=1, dilation=1,
  66.                                    groups=in_channels, bias=bias)
  67.         self.base_scale = _ScaleModule([1, in_channels, 1, 1])
  69.         self.wavelet_convs = nn.ModuleList(
  70.             [nn.Conv2d(in_channels * 4, in_channels * 4, kernel_size, padding='same', stride=1, dilation=1,
  71.                        groups=in_channels * 4, bias=False) for _ in range(self.wt_levels)]
  72.         )
  73.         self.wavelet_scale = nn.ModuleList(
  74.             [_ScaleModule([1, in_channels * 4, 1, 1], init_scale=0.1) for _ in range(self.wt_levels)]
  75.         )
  77.         if self.stride > 1:
  78.             self.stride_filter = nn.Parameter(torch.ones(in_channels, 1, 1, 1), requires_grad=False)
  79.             self.do_stride = lambda x_in: F.conv2d(x_in, self.stride_filter, bias=None, stride=self.stride,
  80.                                                    groups=in_channels)
  81.         else:
  82.             self.do_stride = None
  84.     def forward(self, x):
  86.         x_ll_in_levels = []
  87.         x_h_in_levels = []
  88.         shapes_in_levels = []
  90.         curr_x_ll = x
  92.         for i in range(self.wt_levels):
  93.             curr_shape = curr_x_ll.shape
  94.             shapes_in_levels.append(curr_shape)
  95.             if (curr_shape[2] % 2 > 0) or (curr_shape[3] % 2 > 0):
  96.                 curr_pads = (0, curr_shape[3] % 2, 0, curr_shape[2] % 2)
  97.                 curr_x_ll = F.pad(curr_x_ll, curr_pads)
  99.             curr_x = self.wt_function(curr_x_ll)
  100.             curr_x_ll = curr_x[:, :, 0, :, :]
  102.             shape_x = curr_x.shape
  103.             curr_x_tag = curr_x.reshape(shape_x[0], shape_x[1] * 4, shape_x[3], shape_x[4])
  104.             curr_x_tag = self.wavelet_scale[i](self.wavelet_convs[i](curr_x_tag))
  105.             curr_x_tag = curr_x_tag.reshape(shape_x)
  107.             x_ll_in_levels.append(curr_x_tag[:, :, 0, :, :])
  108.             x_h_in_levels.append(curr_x_tag[:, :, 1:4, :, :])
  110.         next_x_ll = 0
  112.         for i in range(self.wt_levels - 1, -1, -1):
  113.             curr_x_ll = x_ll_in_levels.pop()
  114.             curr_x_h = x_h_in_levels.pop()
  115.             curr_shape = shapes_in_levels.pop()
  117.             curr_x_ll = curr_x_ll + next_x_ll
  119.             curr_x = torch.cat([curr_x_ll.unsqueeze(2), curr_x_h], dim=2)
  120.             next_x_ll = self.iwt_function(curr_x)
  122.             next_x_ll = next_x_ll[:, :, :curr_shape[2], :curr_shape[3]]
  124.         x_tag = next_x_ll
  125.         assert len(x_ll_in_levels) == 0
  127.         x = self.base_scale(self.base_conv(x))
  128.         x = x + x_tag
  130.         if self.do_stride is not None:
  131.             x = self.do_stride(x)
  133.         return x
  136. class _ScaleModule(nn.Module):
  137.     def __init__(self, dims, init_scale=1.0, init_bias=0):
  138.         super(_ScaleModule, self).__init__()
  139.         self.dims = dims
  140.         self.weight = nn.Parameter(torch.ones(*dims) * init_scale)
  141.         self.bias = None
  143.     def forward(self, x):
  144.         return torch.mul(self.weight, x)
  148. def autopad(k, p=None, d=1):  # kernel, padding, dilation
  149.     """Pad to 'same' shape outputs."""
  150.     if d > 1:
  151.         k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k]  # actual kernel-size
  152.     if p is None:
  153.         p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
  154.     return p
  157. class Conv(nn.Module):
  158.     """Standard convolution with args(ch_in, ch_out, kernel, stride, padding, groups, dilation, activation)."""
  159.     default_act = nn.SiLU()  # default activation
  161.     def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
  162.         """Initialize Conv layer with given arguments including activation."""
  163.         super().__init__()
  164.         self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p, d), groups=g, dilation=d, bias=False)
  165.         self.bn = nn.BatchNorm2d(c2)
  166.         self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
  168.     def forward(self, x):
  169.         """Apply convolution, batch normalization and activation to input tensor."""
  170.         return self.act(self.bn(self.conv(x)))
  172.     def forward_fuse(self, x):
  173.         """Perform transposed convolution of 2D data."""
  174.         return self.act(self.conv(x))
  177. class Bottleneck(nn.Module):
  178.     """Standard bottleneck."""
  180.     def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):
  181.         """Initializes a bottleneck module with given input/output channels, shortcut option, group, kernels, and
  182.         expansion.
  183.         """
  184.         super().__init__()
  185.         c_ = int(c2 * e)  # hidden channels
  186.         self.cv1 = Conv(c1, c_, k[0], 1)
  187.         if c_ == c2:
  188.             self.cv2 = WTConv2d(c_, c2, 5, 1)
  189.         else:
  190.             self.cv2 = Conv(c_, c2, k[1], 1, g=g)
  191.         self.add = shortcut and c1 == c2
  193.     def forward(self, x):
  194.         """'forward()' applies the YOLO FPN to input data."""
  195.         return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
  198. class C2f(nn.Module):
  199.     """Faster Implementation of CSP Bottleneck with 2 convolutions."""
  201.     def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):
  202.         """Initializes a CSP bottleneck with 2 convolutions and n Bottleneck blocks for faster processing."""
  203.         super().__init__()
  204.         self.c = int(c2 * e)  # hidden channels
  205.         self.cv1 = Conv(c1, 2 * self.c, 1, 1)
  206.         self.cv2 = Conv((2 + n) * self.c, c2, 1)  # optional act=FReLU(c2)
  207.         self.m = nn.ModuleList(Bottleneck(self.c, self.c, shortcut, g, k=((3, 3), (3, 3)), e=1.0) for _ in range(n))
  209.     def forward(self, x):
  210.         """Forward pass through C2f layer."""
  211.         y = list(self.cv1(x).chunk(2, 1))
  212.         y.extend(m(y[-1]) for m in self.m)
  213.         return self.cv2(torch.cat(y, 1))
  215.     def forward_split(self, x):
  216.         """Forward pass using split() instead of chunk()."""
  217.         y = list(self.cv1(x).split((self.c, self.c), 1))
  218.         y.extend(m(y[-1]) for m in self.m)
  219.         return self.cv2(torch.cat(y, 1))
  221. class C3(nn.Module):
  222.     """CSP Bottleneck with 3 convolutions."""
  224.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
  225.         """Initialize the CSP Bottleneck with given channels, number, shortcut, groups, and expansion values."""
  226.         super().__init__()
  227.         c_ = int(c2 * e)  # hidden channels
  228.         self.cv1 = Conv(c1, c_, 1, 1)
  229.         self.cv2 = Conv(c1, c_, 1, 1)
  230.         self.cv3 = Conv(2 * c_, c2, 1)  # optional act=FReLU(c2)
  231.         self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, k=((1, 1), (3, 3)), e=1.0) for _ in range(n)))
  233.     def forward(self, x):
  234.         """Forward pass through the CSP bottleneck with 2 convolutions."""
  235.         return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))
  237. class C3k(C3):
  238.     """C3k is a CSP bottleneck module with customizable kernel sizes for feature extraction in neural networks."""
  240.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, k=3):
  241.         """Initializes the C3k module with specified channels, number of layers, and configurations."""
  242.         super().__init__(c1, c2, n, shortcut, g, e)
  243.         c_ = int(c2 * e)  # hidden channels
  244.         # self.m = nn.Sequential(*(RepBottleneck(c_, c_, shortcut, g, k=(k, k), e=1.0) for _ in range(n)))
  245.         self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, k=(k, k), e=1.0) for _ in range(n)))
  247. class C3k2_WTConv(C2f):
  248.     """Faster Implementation of CSP Bottleneck with 2 convolutions."""
  250.     def __init__(self, c1, c2, n=1, c3k=False, e=0.5, g=1, shortcut=True):
  251.         """Initializes the C3k2 module, a faster CSP Bottleneck with 2 convolutions and optional C3k blocks."""
  252.         super().__init__(c1, c2, n, shortcut, g, e)
  253.         self.m = nn.ModuleList(
  254.             C3k(self.c, self.c, 2, shortcut, g) if c3k else Bottleneck(self.c, self.c, shortcut, g) for _ in range(n)
  255.         )
  259. if __name__ == "__main__":
  260.     # Generating Sample image
  261.     image_size = (1, 64, 240, 240)
  262.     image = torch.rand(*image_size)
  264.     # Model
  265.     mobilenet_v1 = C3k2_WTConv(64, 64)
  267.     out = mobilenet_v1(image)
  268.     print(out.size())

四、手把手教你添加C3k2

4.1 修改一

第一还是建立文件，我们找到如下 ultralytics /nn/modules文件夹下建立一个目录名字呢就是'Addmodules'文件夹( 用群内的文件的话已经有了无需新建) ！然后在其内部建立一个新的py文件将核心代码复制粘贴进去即可。

​

​

4.2 修改二

第二步我们在该目录下创建一个新的py文件名字为'__init__.py'( 用群内的文件的话已经有了无需新建) ，然后在其内部导入我们的检测头如下图所示。

​​

4.3 修改三

第三步我门中到如下文件'ultralytics/nn/tasks.py'进行导入和注册我们的模块( 用群内的文件的话已经有了无需重新导入直接开始第四步即可) ！

​​

4.4 修改四

按照我的添加在parse_model里添加即可。

​​

到此就修改完成了，大家可以复制下面的yaml文件运行。

五、正式训练

5.1 yaml文件

训练信息： YOLO11-C3k2-WTConv summary: 344 layers, 2,480,347 parameters, 2,470,091 gradients, 6.3 GFLOPs

基础版本信息： YOLO11 summary: 319 layers, 2,594,715 parameters, 2,594,699 gradients, 6.5 GFLOPs 

  1. # Ultralytics YOLO 🚀, AGPL-3.0 license
  2. # YOLO11 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
  4. # Parameters
  5. nc: 80 # number of classes
  6. scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'
  7.   # [depth, width, max_channels]
  8.   n: [0.50, 0.25, 1024] # summary: 319 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
  9.   s: [0.50, 0.50, 1024] # summary: 319 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
  10.   m: [0.50, 1.00, 512] # summary: 409 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
  11.   l: [1.00, 1.00, 512] # summary: 631 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
  12.   x: [1.00, 1.50, 512] # summary: 631 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs
  14. # YOLO11n backbone
  15. backbone:
  16.   # [from, repeats, module, args]
  17.   - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  18.   - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
  19.   - [-1, 2, C3k2_WTConv, [256, False, 0.25]]
  20.   - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  21.   - [-1, 2, C3k2_WTConv, [512, False, 0.25]]
  22.   - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  23.   - [-1, 2, C3k2_WTConv, [512, True]]
  24.   - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  25.   - [-1, 2, C3k2_WTConv, [1024, True]]
  26.   - [-1, 1, SPPF, [1024, 5]] # 9
  27.   - [-1, 2, C2PSA, [1024]] # 10
  29. # YOLO11n head
  30. head:
  31.   - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  32.   - [[-1, 6], 1, Concat, [1]] # cat backbone P4
  33.   - [-1, 2, C3k2_WTConv, [512, False]] # 13
  35.   - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  36.   - [[-1, 4], 1, Concat, [1]] # cat backbone P3
  37.   - [-1, 2, C3k2_WTConv, [256, False]] # 16 (P3/8-small)
  39.   - [-1, 1, Conv, [256, 3, 2]]
  40.   - [[-1, 13], 1, Concat, [1]] # cat head P4
  41.   - [-1, 2, C3k2_WTConv, [512, False]] # 19 (P4/16-medium)
  43.   - [-1, 1, Conv, [512, 3, 2]]
  44.   - [[-1, 10], 1, Concat, [1]] # cat head P5
  45.   - [-1, 2, C3k2_WTConv, [1024, True]] # 22 (P5/32-large)
  47.   - [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5)

5.2 训练代码

大家可以创建一个py文件将我给的代码复制粘贴进去，配置好自己的文件路径即可运行。 

  1. import warnings
  2. warnings.filterwarnings('ignore')
  3. from ultralytics import YOLO
  5. if __name__ == '__main__':
  6.     model = YOLO('模型配置文件')
  7.     # 如何切换模型版本, 上面的ymal文件可以改为 yolov8s.yaml就是使用的v8s,
  8.     # 类似某个改进的yaml文件名称为yolov8-XXX.yaml那么如果想使用其它版本就把上面的名称改为yolov8l-XXX.yaml即可（改的是上面YOLO中间的名字不是配置文件的）！
  9.     # model.load('yolov8n.pt') # 是否加载预训练权重,科研不建议大家加载否则很难提升精度
  10.     model.train(data=r"C:\Users\Administrator\PycharmProjects\yolov5-master\yolov5-master\Construction Site Safety.v30-raw-images_latestversion.yolov8\data.yaml",
  11.                 # 如果大家任务是其它的'ultralytics/cfg/default.yaml'找到这里修改task可以改成detect, segment, classify, pose
  12.                 cache=False,
  13.                 imgsz=640,
  14.                 epochs=150,
  15.                 single_cls=False,  # 是否是单类别检测
  16.                 batch=16,
  17.                 close_mosaic=0,
  18.                 workers=0,
  19.                 device='0',
  20.                 optimizer='SGD', # using SGD
  21.                 # resume='runs/train/exp21/weights/last.pt', # 如过想续训就设置last.pt的地址
  22.                 amp=True,  # 如果出现训练损失为Nan可以关闭amp
  23.                 project='runs/train',
  24.                 name='exp',
  25.                 )

5.3 训练过程截图

​

五、本文总结

到此本文的正式分享内容就结束了，在这里给大家推荐我的YOLOv11改进有效涨点专栏，本专栏目前为新开的平均质量分98分，后期我会根据各种最新的前沿顶会进行论文复现，也会对一些老的改进机制进行补充，如果大家觉得本文帮助到你了，订阅本专栏，关注后续更多的更新~