使用官方的GAT：https://blog.csdn.net/StarfishCu/article/details/109644271GAT学习：PyG实现GAT（使用PyG封装好的GATConv函数）（三）
别人自己实现的GAT：https://blog.csdn.net/StarfishCu/article/details/109237526GAT学习：PyG实现GAT（图注意力神经网络）网络（一）
PyG的tutorial https://colab.research.google.com/github/AntonioLonga/PytorchGeometricTutorial/blob/main/Tutorial3/Tutorial3.ipynb

PyG04-GAT - 图1

封装

  1. from torch_geometric.data import Data
  2. from torch_geometric.nn import GATConv
  3. from torch_geometric.datasets import Planetoid
  4. import torch_geometric.transforms as T
  6. import matplotlib.pyplot as plt
  8. name_data = 'Cora'
  9. dataset = Planetoid(root= '/tmp/' + name_data, name = name_data)
  10. dataset.transform = T.NormalizeFeatures()
  12. print(f"Number of Classes in {name_data}:", dataset.num_classes)
  13. print(f"Number of Node Features in {name_data}:", dataset.num_node_features)
  1. class GAT(torch.nn.Module):
  2.     def __init__(self):
  3.         super(GAT, self).__init__()
  4.         self.hid = 8
  5.         self.in_head = 8
  6.         self.out_head = 1
  9.         self.conv1 = GATConv(dataset.num_features, self.hid, heads=self.in_head, dropout=0.6)
  10.         self.conv2 = GATConv(self.hid*self.in_head, dataset.num_classes, concat=False,
  11.                              heads=self.out_head, dropout=0.6)
  13.     def forward(self, data):
  14.         x, edge_index = data.x, data.edge_index
  16.         x = F.dropout(x, p=0.6, training=self.training)
  17.         x = self.conv1(x, edge_index)
  18.         x = F.elu(x)
  19.         x = F.dropout(x, p=0.6, training=self.training)
  20.         x = self.conv2(x, edge_index)
  22.         return F.log_softmax(x, dim=1)
  1. device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
  2. device = "cpu"
  4. model = GAT().to(device)
  5. data = dataset[0].to(device)
  8. optimizer = torch.optim.Adam(model.parameters(), lr=0.005, weight_decay=5e-4)
  10. model.train()
  11. for epoch in range(1000):
  12.     model.train()
  13.     optimizer.zero_grad()
  14.     out = model(data)
  15.     loss = F.nll_loss(out[data.train_mask], data.y[data.train_mask])
  17.     if epoch%200 == 0:
  18.         print(loss)
  20.     loss.backward()
  21.     optimizer.step()
  1. model.eval()
  2. _, pred = model(data).max(dim=1)
  3. correct = float(pred[data.test_mask].eq(data.y[data.test_mask]).sum().item())
  4. acc = correct / data.test_mask.sum().item()
  5. print('Accuracy: {:.4f}'.format(acc))

自己实现