关于GNN一直垂涎已久，今日放假终于有空上手
我就简单的选择了pytorch_geometric作为工具

环境配置

升级CUDA到11.1，升级pytorch到1.9

代码上手

基本API

  1. Batch(batch=[767], edge_index=[2, 2864], ptr=[25], x=[767, 21], y=[24])
  2. 我感觉batch应该是节点的数量，即767个节点，因为他要存储下标
  3. edge index应该是边，一个边有两个节点，一共有2864条边，由于是无向图，所以应该是1432边
  4. y的话，应该是当前batch的数据量，正常的batch是32（设置的）这里是最后一个batch
  5. x：x是节点的表示，即一共767个节点，每一个节点的特征是32个

示例代码

  1. from torch_geometric.datasets import Planetoid
  2. dataset = Planetoid(root='./data/', name='Cora')
  4. import torch
  5. import torch.nn.functional as F
  6. from torch_geometric.nn import GCNConv
  8. class Net(torch.nn.Module):
  9.     def __init__(self):
  10.         super(Net, self).__init__()
  11.         self.conv1 = GCNConv(dataset.num_node_features, 16)
  12.         self.conv2 = GCNConv(16, dataset.num_classes)
  14.     def forward(self, data):
  15.         x, edge_index = data.x, data.edge_index
  17.         x = self.conv1(x, edge_index)
  18.         x = F.relu(x)
  19.         x = F.dropout(x, training=self.training)
  20.         x = self.conv2(x, edge_index)
  22.         return F.log_softmax(x, dim=1)
  24. device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
  25. model = Net().to(device)
  26. data = dataset[0].to(device)
  27. optimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=5e-4)
  29. model.train()
  30. for epoch in range(200):
  31.     optimizer.zero_grad()
  32.     out = model(data)
  33.     loss = F.nll_loss(out[data.train_mask], data.y[data.train_mask])
  34.     loss.backward()
  35.     optimizer.step()
  37. model.eval()
  38. _, pred = model(data).max(dim=1)
  39. correct = int(pred[data.test_mask].eq(data.y[data.test_mask]).sum().item())
  40. acc = correct / int(data.test_mask.sum())
  41. print('Accuracy: {:.4f}'.format(acc))

参考链接

https://pytorch-geometric.readthedocs.io/en/latest/notes/introduction.html#
https://blog.csdn.net/qq_43440798/article/details/111083989
https://blog.csdn.net/qq_39732684/article/details/104984667