python实现gnn算法附代码

### Python 实现图神经网络 (GNN) 算法 #### 使用 PyTorch Geometric 库构建简单 GNN 模型 为了简化开发过程并提高效率，通常会选择使用专门设计用于处理图形数据的库之一——PyTorch Geometric。下面是一个基于此库创建基础版本 Graph Convolutional Network (GCN)[^1]的例子。 ```python import torch from torch.nn import Linear, ReLU, Sequential from torch_geometric.datasets import Planetoid from torch_geometric.loader import DataLoader from torch_geometric.nn.conv import GCNConv from torch_geometric.transforms import NormalizeFeatures class SimpleGCN(torch.nn.Module): def __init__(self, input_dim, hidden_dim, output_dim): super(SimpleGCN, self).__init__() # 定义两层GCN卷积操作 self.gcn1 = GCNConv(input_dim, hidden_dim) self.act_fn = ReLU() self.fc = Linear(hidden_dim, output_dim) def forward(self, data): x, edge_index = data.x, data.edge_index # 前向传播计算 h = self.act_fn(self.gcn1(x, edge_index)) out = self.fc(h) return out if __name__ == "__main__": dataset_name = 'Cora' path = './data/' + dataset_name transform = NormalizeFeatures() # 数据预处理 dataset = Planetoid(root=path, name=dataset_name, transform=transform) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') model = SimpleGCN(dataset.num_features, 16, dataset.num_classes).to(device) optimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=5e-4) loader = DataLoader(dataset, batch_size=len(dataset), shuffle=True) for epoch in range(200): # 训练周期数 total_loss = 0 for batch in loader: batch.to(device) optimizer.zero_grad() pred = model(batch) loss = torch.nn.functional.nll_loss(pred.log_softmax(dim=-1), batch.y.squeeze()) loss.backward() optimizer.step() total_loss += float(loss.item()) * batch.num_graphs avg_loss = total_loss / len(loader.dataset) print(f'Epoch {epoch}, Loss: {avg_loss:.4f}') ``` 这段代码展示了如何利用 `torch_geometric` 创建一个简单的二层 GCN 来解决半监督节点分类问题。这里选择了 Cora 文献引用网络作为训练集，并实现了标准化特征变换以及 Adam 优化器来进行参数更新。