CNNs Meet Transformers (CMT)代码

### CNNs Meet Transformers (CMT) 的源代码实现

#### CMT 架构概述
CMT 结合了卷积神经网络(CNN)和变换器(Transformer)，旨在利用两者的优点来提升视觉识别任务的效果。这种架构通过引入局部窗口内的自注意力机制，在保持计算效率的同时增强了模型捕捉全局依赖的能力[^1]。

#### 源码结构解析
为了更好地理解如何构建一个完整的CMT模块，下面提供了一个简化版的Python实现示例：

import torch.nn as nn


class ConvolutionalStem(nn.Module):
    """定义卷积干部分"""
    
    def __init__(self, input_channels=3, output_channels=64):
        super(ConvolutionalStem, self).__init__()
        self.conv = nn.Conv2d(input_channels, output_channels, kernel_size=7, stride=2, padding=3)

    def forward(self, x):
        return self.conv(x)


class LocalAttentionBlock(nn.Module):
    """定义局部注意区块"""

    def __init__(self, dim, num_heads=8):
        super(LocalAttentionBlock, self).__init__()
        assert dim % num_heads == 0
        
        head_dim = dim // num_heads
        self.scale = head_dim ** -0.5
    
        self.qkv = nn.Linear(dim, dim * 3)
        self.proj_out = nn.Linear(dim, dim)

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, -1).permute(2, 0, 1, 3)
        q, k, v = qkv.chunk(3, dim=-1)

        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        
        out = (attn @ v).transpose(1, 2).reshape(B, N, C)
        out = self.proj_out(out)
        return out


class CMTModule(nn.Module):
    """组合成完整的CMT单元"""

    def __init__(self, stem_output_channels=64, transformer_hidden_dim=192):
        super(CMTModule, self).__init__()
        self.stem = ConvolutionalStem(output_channels=stem_output_channels)
        self.transformer_block = LocalAttentionBlock(transformer_hidden_dim)

    def forward(self, img_tensor):
        features = self.stem(img_tensor)
        b, c, h, w = features.size()
        flattened_features = features.view(b, c, -1).transpose(1, 2)
        attended_features = self.transformer_block(flattened_features)
        return attended_features

此段代码展示了如何创建一个基础版本的CMT组件，其中包括卷积特征提取层以及基于多头自注意力机制的本地化处理逻辑[^1]。