模塊出處

[ICLR 25 Submission] [link] UltraLightUNet: Rethinking U-shaped Network with Multi-kernel Lightweight Convolutions for Medical Image Segmentation

---

模塊名稱

Grouped Attention Gate (GAG)

---

模塊作用

輕量特征融合

---

模塊結構

---

模塊特點

• 特征融合前使用Group Conv進行處理，比標準卷積更加輕量
• 將融合得到的粗特征視為Spatial Attention Map, 并與Encoder特征相乘，從而實現名字中"Gate"的效果
• 相較于特征融合模塊，也可以視為一種利用輔助信息(Decoder)特征以增強Encoder特征的增強模塊

---

模塊代碼

import torch
import torch.nn as nn
import torch.nn.functional as Fclass GAG(nn.Module):def __init__(self, F_g, F_l, F_int, kernel_size=1, groups=1):super(GAG,self).__init__()if kernel_size == 1:groups = 1self.W_g = nn.Sequential(nn.Conv2d(F_g, F_int, kernel_size=kernel_size,stride=1,padding=kernel_size//2,groups=groups, bias=True),nn.BatchNorm2d(F_int))self.W_x = nn.Sequential(nn.Conv2d(F_l, F_int, kernel_size=kernel_size,stride=1,padding=kernel_size//2,groups=groups, bias=True),nn.BatchNorm2d(F_int))self.psi = nn.Sequential(nn.Conv2d(F_int, 1, kernel_size=1,stride=1,padding=0,bias=True),nn.BatchNorm2d(1),nn.Sigmoid())self.activation = nn.ReLU(inplace=True)def forward(self,g,x):g1 = self.W_g(g)x1 = self.W_x(x)psi = self.activation(g1+x1)psi = self.psi(psi)return x*psiif __name__ == '__main__':x1 = torch.randn([1, 64, 44, 44])x2 = torch.randn([1, 64, 44, 44])gag = GAG(F_g=64, F_l=64, F_int=64//2, kernel_size=3, groups=64//2)out = gag(x1, x2)print(out.shape)  # [1, 64, 44, 44]

---