如何在图卷积网络上实现ECA注意力机制

Question

如何在图卷积网络上实现ECA注意力机制，下面是一个图卷积网络，基于pyg 实现

class Net(torch.nn.Module):
    def __init__(self, num_node_features, num_classes):
        super(Net, self).__init__()
        self.conv1 = pyg_nn.GCNConv(num_node_features, num_node_features)
        self.conv2 = pyg_nn.GCNConv(num_node_features, num_node_features)
        self.lin = nn.Linear(num_node_features, num_classes)

    def forward(self, data):
        x, edge_index, batch = data.x, data.edge_index, data.batch
        x = self.conv1(x, edge_index)
        x = F.relu(x)
        x = self.conv2(x, edge_index)
        x = F.relu(x)
        x = pyg_nn.global_mean_pool(x, batch)
        x = self.lin(x)
        x = F.log_softmax(x, dim=1)
        return x

Answer 1

要在图卷积网络上实现ECA注意力机制，需要在每个图卷积层之后添加一个ECA模块。下面是一个修改后的代码示例：

import torch
import torch.nn.functional as F
import torch.nn as nn
from torch_geometric.nn import GCNConv
from torch_geometric.nn import global_mean_pool

class ECA(nn.Module):
    def __init__(self, channels, gamma=2, b=1):
        super(ECA, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool1d(1)
        self.conv = nn.Conv1d(1, 1, kernel_size=gamma, padding=(gamma - 1) // 2, bias=False)
        self.sigmoid = nn.Sigmoid()
        self.b = b

    def forward(self, x):
        y = self.avg_pool(x)
        y = self.conv(y.transpose(1, 2))
        y = self.sigmoid(y)
        return x * (self.b + y.expand_as(x))

class Net(torch.nn.Module):
    def __init__(self, num_node_features, num_classes):
        super(Net, self).__init__()
        self.conv1 = GCNConv(num_node_features, num_node_features)
        self.eca1 = ECA(num_node_features)
        self.conv2 = GCNConv(num_node_features, num_node_features)
        self.eca2 = ECA(num_node_features)
        self.lin = nn.Linear(num_node_features, num_classes)

    def forward(self, data):
        x, edge_index, batch = data.x, data.edge_index, data.batch
        x = self.conv1(x, edge_index)
        x = self.eca1(x)
        x = F.relu(x)
        x = self.conv2(x, edge_index)
        x = self.eca2(x)
        x = F.relu(x)
        x = global_mean_pool(x, batch)
        x = self.lin(x)
        x = F.log_softmax(x, dim=1)
        return x

在这个修改后的代码中，我们添加了一个ECA类，它实现了ECA注意力机制。在Net类的__init__方法中，我们添加了两个ECA模块，分别在两个GCNConv层之后。在forward方法中，我们首先使用第一个GCNConv层，然后将输出传递给第一个ECA模块。然后我们使用ReLU激活函数，再使用第二个GCNConv层，然后将输出传递给第二个ECA模块。最后，我们使用全局平均池化层和线性层来生成输出。