GNN 代码实现图编码，要求每个图（而不是节点）都有自己的标签，尽可能让编码向量可以表征对应标签信息。给出训练、测试用例

下面是一个简单的 GNN 代码实现图编码的示例，其中每个图都有自己的标签，编码向量可以表征对应标签信息。

训练用例：

import torch
import torch.nn.functional as F
from torch_geometric.datasets import TUDataset
from torch_geometric.utils import degree
from torch_geometric.nn import GCNConv

# 加载数据集
dataset = TUDataset(root='/tmp/ENZYMES', name='ENZYMES')

# 计算每个节点的度数
data = dataset[0]
data.degrees = degree(data.edge_index[0], data.num_nodes)
data = data.to('cuda')

# 定义模型
class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = GCNConv(dataset.num_node_features, 16)
        self.conv2 = GCNConv(16, 32)
        self.lin = torch.nn.Linear(32, dataset.num_classes)

    def forward(self, x, edge_index, degrees):
        x = F.relu(self.conv1(x, edge_index, degrees))
        x = F.relu(self.conv2(x, edge_index, degrees))
        x = torch_geometric.nn.global_mean_pool(x, data.batch)
        x = self.lin(x)
        return x

model = Net().to('cuda')

# 训练模型
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
criterion = torch.nn.CrossEntropyLoss()
for epoch in range(100):
    optimizer.zero_grad()
    out = model(data.x, data.edge_index, data.degrees)
    loss = criterion(out, data.y)
    loss.backward()
    optimizer.step()

# 测试模型
model.eval()
correct = 0
for data in dataset:
    data = data.to('cuda')
    pred = model(data.x, data.edge_index, data.degrees).max(dim=1)[1]
    correct += int(pred.eq(data.y).sum())

print(f'Accuracy: {correct / len(dataset)}')

测试用例：

import torch
import torch.nn.functional as F
from torch_geometric.datasets import TUDataset
from torch_geometric.utils import degree
from torch_geometric.nn import GCNConv

# 加载数据集
dataset = TUDataset(root='/tmp/ENZYMES', name='ENZYMES')

# 计算每个节点的度数
data = dataset[0]
data.degrees = degree(data.edge_index[0], data.num_nodes)
data = data.to('cuda')

# 定义模型
class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = GCNConv(dataset.num_node_features, 16)
        self.conv2 = GCNConv(16, 32)
        self.lin = torch.nn.Linear(32, dataset.num_classes)

    def forward(self, x, edge_index, degrees):
        x = F.relu(self.conv1(x, edge_index, degrees))
        x = F.relu(self.conv2(x, edge_index, degrees))
        x = torch_geometric.nn.global_mean_pool(x, data.batch)
        x = self.lin(x)
        return x

model = Net().to('cuda')

# 加载训练好的模型
model.load_state_dict(torch.load('model.pt'))

# 测试模型
model.eval()
correct = 0
for data in dataset:
    data = data.to('cuda')
    pred = model(data.x, data.edge_index, data.degrees).max(dim=1)[1]
    correct += int(pred.eq(data.y).sum())

print(f'Accuracy: {correct / len(dataset)}')