4. YOLO11模型改进示例

修改1：Conv替换LAE

0）LSM-YOLO提出的轻量化自适应特征提取(LAE)模块，旨在多尺度、自适应的提取特征，全局与局部特征融合，并降低计算成本。

核心代码（fxlae.py，可复制下面代码）：

import torch import torch.nn as nn from einops import rearrange   __all__ = ['LAE', 'MSFM']     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 LAE(nn.Module):     # Light-weight Adaptive Extraction     def __init__(self, ch, group=16) -> None:         super().__init__()           self.softmax = nn.Softmax(dim=-1)         self.attention = nn.Sequential(             nn.AvgPool2d(kernel_size=3, stride=1, padding=1),             Conv(ch, ch, k=1)         )           self.ds_conv = Conv(ch, ch * 4, k=3, s=2, g=(ch // group))       def forward(self, x):         # bs, ch, 2*h, 2*w => bs, ch, h, w, 4         att = rearrange(self.attention(x), 'bs ch (s1 h) (s2 w) -> bs ch h w (s1 s2)', s1=2, s2=2)         att = self.softmax(att)           # bs, 4 * ch, h, w => bs, ch, h, w, 4         x = rearrange(self.ds_conv(x), 'bs (s ch) h w -> bs ch h w s', s=4)         x = torch.sum(x * att, dim=-1)         return x     class MatchNeck_Inner(nn.Module):     def __init__(self, channels) -> None:         super().__init__()           self.gap = nn.Sequential(             nn.AdaptiveAvgPool2d((1, 1)),             Conv(channels, channels)         )         self.pool_h = nn.AdaptiveAvgPool2d((None, 1))         self.pool_w = nn.AdaptiveAvgPool2d((1, None))         self.conv_hw = Conv(channels, channels, (3, 1))         self.conv_pool_hw = Conv(channels, channels, 1)       def forward(self, x):         _, _, h, w = x.size()         x_pool_h, x_pool_w, x_pool_ch = self.pool_h(x), self.pool_w(x).permute(0, 1, 3, 2), self.gap(x)         x_pool_hw = torch.cat([x_pool_h, x_pool_w], dim=2)         x_pool_h, x_pool_w = torch.split(x_pool_hw, [h, w], dim=2)         x_pool_hw_weight = x_pool_hw.sigmoid()         x_pool_h_weight, x_pool_w_weight = torch.split(x_pool_hw_weight, [h, w], dim=2)         x_pool_h, x_pool_w = x_pool_h * x_pool_h_weight, x_pool_w * x_pool_w_weight         x_pool_ch = x_pool_ch * torch.mean(x_pool_hw_weight, dim=2, keepdim=True)         return x * x_pool_h.sigmoid() * x_pool_w.permute(0, 1, 3, 2).sigmoid() * x_pool_ch.sigmoid()     class MatchNeck(nn.Module):     def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):         super().__init__()         c_ = int(c2 * e)  # hidden channels         self.cv1 = Conv(c1, c_, k[0], 1)         self.cv2 = Conv(c_, c2, k[1], 1, g=g)         self.add = shortcut and c1 == c2         self.MN = MatchNeck_Inner(c2)       def forward(self, x):         return x + self.MN(self.cv2(self.cv1(x))) if self.add else self.MN(self.cv2(self.cv1(x)))   class MSFM(nn.Module):     def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):         super().__init__()         self.c = int(c2 * e)  # hidden channels         self.cv1 = Conv(c1, 2 * self.c, 1, 1)         self.cv2 = Conv((2 + n) * self.c, c2, 1)         self.m = nn.ModuleList(MatchNeck(self.c, self.c, shortcut, g, k=(3, 3), e=1.0) for _ in range(n))       def forward(self, x):         y = list(self.cv1(x).chunk(2, 1))         y.extend(m(y[-1]) for m in self.m)         return self.cv2(torch.cat(y, 1))       def forward_split(self, x):         y = list(self.cv1(x).split((self.c, self.c), 1))         y.extend(m(y[-1]) for m in self.m)         return self.cv2(torch.cat(y, 1))   if __name__ == "__main__":     # Generating Sample image     image_size = (1, 64, 224, 224) image = torch.rand(*image_size)     # Model     model = LAE(64)     out = model(image)     print(out.size())

执行上面程序，可以看出：(1, 64, 224, 224)→(1, 64, 112, 112)，说明通道数没变，h、w减小1/2。

1）将fxlae.py文件放到ultralytics/nn/modules文件夹中。

2）修改ultralytics/nn/modules/__init__.py，导入LAE模块，即增加88行和99行的内容。

3）修改ultralytics/nn/tasks.py，有两处，即增加19行、1038-1039行的内容。

4）以yolo11.yaml文件为基础重新写一个yolo11-fxlae.yaml文件，位置如下：

ultralytics/cfg/models/11/yolo11-fxlae.yaml

# Parameters nc: 80 # number of classes scales: # model compound scaling constants   # [depth, width, max_channels]   n: [0.50, 0.25, 1024]   s: [0.50, 0.50, 1024]   m: [0.50, 1.00, 512]   l: [1.00, 1.00, 512]   x: [1.00, 1.50, 512] # YOLO11n-fxlae backbone backbone:   # [from, repeats, module, args]   - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2   - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4   - [-1, 2, C3k2, [256, False, 0.25]]   - [-1, 1, LAE, []] # 3-P3/8   - [-1, 2, C3k2, [512, False, 0.25]]   - [-1, 1, LAE, []] # 5-P4/16   - [-1, 2, C3k2, [512, True]]   - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32   - [-1, 2, C3k2, [1024, True]]