论文介绍

论文链接：https://arxiv.org/pdf/2211.11943

• 研究背景：论文指出，尽管当前研究者们通过利用大核卷积、高阶空间交互或稀疏卷积核等方法对卷积神经网络（ConvNets）的设计进行了重新思考，但如何更有效地利用卷积来构建强大的ConvNet架构仍是一个热门研究课题。
• Conv2Former提出：论文提出了一种新的卷积网络架构Conv2Former，它采用了一种简化的自注意力机制，仅使用卷积和Hadamard乘积。

创新点

• 卷积调制操作：Conv2Former的核心是卷积调制操作，它简化了自注意力机制，提高了内存效率，特别是当处理高分辨率图像时。
• 大核卷积的优势：与之前的ConvNets不同，Conv2Former可以从更大的卷积核（如11×11和21×21）中获益，这带来了更好的性能提升。

方法

• 卷积调制块：Conv2Former使用了卷积调制块，该块利用深度卷积的特征作为权重来调制值表示。
• 与自注意力和经典残差块的比较：相比自注意力，Conv2Former的方法在构建关系时更加内存高效；与经典残差块相比，由于调制操作，Conv2Former的方法可以适应输入内容。
  

模块作用

• 卷积调制块的优势：卷积调制块能够同时捕捉通道和空间维度的信息，并且比传统的自注意力机制更加高效。
• 在下游任务中的表现：Conv2Former在物体检测和语义分割等下游任务中表现出色，证明了其卷积调制块的有效性和实用性。

整体而言，Conv2Former通过提出卷积调制操作，简化了自注意力机制，提高了内存效率，并展示了在视觉识别任务中的优越性能。代码如下：

import torch
import torch.nn as nn
import torch.nn.functional as Ffrom timm.models.layers import DropPathclass MLP(nn.Module):def __init__(self, dim, mlp_ratio=4):super().__init__()self.norm = LayerNorm(dim, eps=1e-6, data_format="channels_first")self.fc1 = nn.Conv2d(dim, int(dim * mlp_ratio), 1,padding=0)self.pos = nn.Conv2d(int(dim * mlp_ratio), int(dim * mlp_ratio), 3, padding=1, groups=int(dim * mlp_ratio))self.fc2 = nn.Conv2d(int(dim * mlp_ratio), dim, 1)self.act = nn.GELU()def forward(self, x):B, C, H, W = x.shapex = self.norm(x)x = self.fc1(x)x = self.act(x)x = x + self.act(self.pos(x))x = self.fc2(x)return x
class LayerNorm(nn.Module):r""" LayerNorm that supports two data formats: channels_last (default) or channels_first.The ordering of the dimensions in the inputs. channels_last corresponds to inputs withshape (batch_size, height, width, channels) while channels_first corresponds to inputswith shape (batch_size, channels, height, width)."""def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):super().__init__()self.weight = nn.Parameter(torch.ones(normalized_shape))self.bias = nn.Parameter(torch.zeros(normalized_shape))self.eps = epsself.data_format = data_formatif self.data_format not in ["channels_last", "channels_first"]:raise NotImplementedErrorself.normalized_shape = (normalized_shape,)def forward(self, x):if self.data_format == "channels_last":return F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)elif self.data_format == "channels_first":u = x.mean(1, keepdim=True)s = (x - u).pow(2).mean(1, keepdim=True)x = (x - u) / torch.sqrt(s + self.eps)x = self.weight[:, None, None] * x + self.bias[:, None, None]return x
class SpatialAttention(nn.Module):def __init__(self, dim, kernel_size, expand_ratio=2):super().__init__()self.norm = LayerNorm(dim, eps=1e-6, data_format="channels_first")self.att = nn.Sequential(nn.Conv2d(dim, dim, 1),nn.GELU(),nn.Conv2d(dim, dim, kernel_size=kernel_size, padding=kernel_size//2, groups=dim))self.v = nn.Conv2d(dim, dim, 1)self.proj = nn.Conv2d(dim, dim, 1)def forward(self, x):B, C, H, W = x.shapex = self.norm(x)x = self.att(x) * self.v(x)x = self.proj(x)return xclass Block(nn.Module):def __init__(self, dim, kernel_size, num_head, window_size=14, mlp_ratio=4., drop_path=0.):super().__init__()self.attn = SpatialAttention(dim, kernel_size)self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()self.mlp = MLP(dim, mlp_ratio)layer_scale_init_value = 1e-6self.layer_scale_1 = nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)self.layer_scale_2 = nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)def forward(self, x):x = x + self.drop_path(self.layer_scale_1.unsqueeze(-1).unsqueeze(-1) * self.attn(x))x = x + self.drop_path(self.layer_scale_2.unsqueeze(-1).unsqueeze(-1) * self.mlp(x))return xif __name__ == '__main__':# 生成随机输入数据input_data = torch.randn(2, 64, 40, 40)mca = Block(64,3,16)output = mca(input_data)# 打印输入和输出形状print("Input size:", input_data.size())print("Output size:", output.size())

输出结果：

Input size: torch.Size([2, 64, 40, 40])
Output size: torch.Size([2, 64, 40, 40])