1. SENet (Squeeze-and-Excitation Network)： 通道注意力（Channel Attention）。专注于建模通道之间的相互依赖关系，自动学习到每个通道的重要程度，然后为重要的通道赋予更大的权重。
2. CBAM (Convolutional Block Attention Module)： 通道注意力 + 空间注意力 的串联结构。认为只关注通道维度是不够的，空间位置上的信息也同样重要。CBAM依次从通道和空间两个维度计算注意力图。
3. SKNet (Selective Kernel Networks)： 动态选择不同大小的卷积核（感受野）。让网络能够根据输入信息的复杂程度，自适应地调节其感受野的大小。
4. ECANet (Efficient Channel Attention Network)： 对SENet的轻量化和改进。认为SENet中的降维操作对通道注意力预测会产生副作用，并且两个全连接层显得笨重。ECANet提出了一种不降维的、更高效的局部跨通道交互策略。
import torch import torch.nn as nn class SeNet(nn.Module): def __init__(self, inchannel, ratio=16): super(SeNet, self).__init__() self.gap = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(inchannel, inchannel // ratio, bias=False), # 从 c -> c/r nn.ReLU(inplace=True), nn.Linear(inchannel // ratio, inchannel, bias=False), # 从 c/r -> c nn.Sigmoid() ) def forward(self, x): b, c, _, _ = x.size() y = self.gap(x).view(b, c) y = self.fc(y).view(b, c, 1, 1) return x * y.expand_as(x) class CbamNet(nn.Module): def __init__(self, channels, reduction=16): super(CbamNet, self).__init__() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.max_pool = nn.AdaptiveMaxPool2d(1) self.fc1 = nn.Conv2d(channels, channels // reduction, kernel_size=1,padding=0) self.relu = nn.ReLU(inplace=True) self.fc2 = nn.Conv2d(channels // reduction, channels, kernel_size=1,padding=0) self.sigmoid_channel = nn.Sigmoid() self.conv_after_concat = nn.Conv2d(2,1,kernel_size=3,stride=1,padding=1) self.sigmoid_spatial = nn.Sigmoid() def forward(self, x): # avg全局池化+MLP avg = self.avg_pool(x) avg = self.fc1(avg) avg = self.relu(avg) avg = self.fc2(avg) # max全局池化+MLP mx = self.max_pool(x) mx = self.fc1(mx) mx = self.relu(mx) mx = self.fc2(mx) x = x * self.sigmoid_channel(avg+mx) module_input = x avg = torch.mean(x, 1, True) mx, _ = torch.max(x, 1, True) x = torch.cat((avg, mx), 1) x = self.conv_after_concat(x) x = self.sigmoid_spatial(x) x = module_input * x return x class SkNet(nn.Module): def __init__(self,inchannel,ratio=16): super(SkNet,self).__init__() self.conv3x3 = nn.Conv2d(inchannel,inchannel,kernel_size=3,dilation=1, padding=1) self.conv5x5 = nn.Conv2d(inchannel,inchannel,kernel_size=3,dilation=2, padding=2) self.avg = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(inchannel, inchannel // ratio, bias=False), # 从 c -> c/r nn.ReLU(inplace=True), nn.Linear(inchannel // ratio, inchannel*2, bias=False) # 从 c/r -> c ) self.softmax = nn.Softmax(dim=1) def forward(self,x): x1 = self.conv3x3(x) x2 = self.conv5x5(x) z = x1 + x2 B, C, _, _ = z.size() z = self.avg(z).view(B, C) z = self.fc(z) z = z.view(B, 2, C) a = z[:, 0, :].unsqueeze(-1).unsqueeze(-1) # [B, C, 1, 1] b = z[:, 1, :].unsqueeze(-1).unsqueeze(-1) x1 = x1 * a x2 = x2 * b x = x1 + x2 return x class EcaNet(nn.Module): def __init__(self,k_size=3): super(EcaNet, self).__init__() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.conv = nn.Conv1d(1, 1, kernel_size=k_size, padding=(k_size - 1) // 2, bias=False) self.sigmoid = nn.Sigmoid() def forward(self, x): y = self.avg_pool(x) y = self.conv(y.squeeze(-1).transpose(-1, -2)).transpose(-1, -2).unsqueeze(-1) y = self.sigmoid(y) return x * y.expand_as(x) se_model = SeNet(128) cbma_model = CbamNet(128,16) sk_model = SkNet(128,16) eca_model = EcaNet() x = torch.randn([1,128,200,200]) # y = cbma_model(x) torch.onnx.export(se_model,x,'se_net.onnx',opset_version=11) torch.onnx.export(cbma_model,x,'cbma_net.onnx',opset_version=11) torch.onnx.export(sk_model,x,'sk_net.onnx',opset_version=11) torch.onnx.export(eca_model,x,'eca_net.onnx',opset_version=11)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。