import torch import torch.nn as nn from dgl.nn.pytorch import GraphConv, SumPooling from dgl.nn.pytorch import EdgeWeightNorm class ModifiedGCN(nn.Module): def __init__(self, input_dim, edge_dim, output_dim): super(ModifiedGCN, self).__init__() hidden_dims = [32, 32, 32, 16] self.conv1 = GraphConv(input_dim, hidden_dims[0], allow_zero_in_degree=True) self.bn1 = nn.BatchNorm1d(hidden_dims[0]) self.conv2 = GraphConv(hidden_dims[0], hidden_dims[1], allow_zero_in_degree=True) self.bn2 = nn.BatchNorm1d(hidden_dims[1]) self.conv3 = GraphConv(hidden_dims[1], hidden_dims[2], allow_zero_in_degree=True) self.bn3 = nn.BatchNorm1d(hidden_dims[2]) self.conv4 = GraphConv(hidden_dims[2], hidden_dims[3], allow_zero_in_degree=True) self.bn4 = nn.BatchNorm1d(hidden_dims[3]) self.pool = SumPooling() self.classify = nn.Linear(hidden_dims[3] + edge_dim, output_dim) # Concatenate edge features # Define activations for each layer self.elu1 = nn.ELU() self.elu2 = nn.ELU() self.elu3 = nn.ELU() self.elu4 = nn.ELU() self.edge_weight_norm = EdgeWeightNorm() def forward(self, g, features, edge_features, edge_weights): edge_weights = torch.abs(edge_weights) # Ensure all edge weights are positive edge_weights = self.edge_weight_norm(g, edge_weights) x = self.elu1(self.bn1(self.conv1(g, features, edge_weight=edge_weights))) x = self.elu2(self.bn2(self.conv2(g, x, edge_weight=edge_weights))) x = self.elu3(self.bn3(self.conv3(g, x, edge_weight=edge_weights))) x = self.elu4(self.bn4(self.conv4(g, x, edge_weight=edge_weights))) x = self.pool(g, x) # Concatenate pooled node features with edge features edge_mean = torch.mean(edge_features, dim=0).unsqueeze(0) # Add an extra dimension edge_mean = edge_mean.repeat(x.shape[0], 1) # Repeat to match the batch size x = torch.cat([x, edge_mean], dim=1) return self.classify(x)