torch-geometric

PyTorch Geometric (PyG)

概述

PyTorch Geometric 是一个基于 PyTorch 构建的库，用于开发和训练图神经网络（GNN）。在图和不规则结构上进行深度学习时可使用此技能，包括小批量处理、多 GPU 训练和几何深度学习应用。

何时使用此技能

在以下情况中应使用此技能：

• 基于图的机器学习：节点分类、图分类、链接预测
• 分子属性预测：药物发现、化学属性预测
• 社交网络分析：社区检测、影响力预测
• 引文网络：论文分类、推荐系统
• 3D 几何数据：点云、网格、分子结构
• 异构图：多类型节点和边（例如知识图谱）
• 大规模图学习：邻居采样、分布式训练

快速入门

安装

uv pip install torch_geometric

如需其他依赖项（稀疏运算、聚类）：

uv pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-${TORCH}+${CUDA}.html

创建基本图

import torch
from torch_geometric.data import Data

# Create a simple graph with 3 nodes
edge_index = torch.tensor([[0, 1, 1, 2],  # source nodes
                           [1, 0, 2, 1]], dtype=torch.long)  # target nodes
x = torch.tensor([[-1], [0], [1]], dtype=torch.float)  # node features

data = Data(x=x, edge_index=edge_index)
print(f"Nodes: {data.num_nodes}, Edges: {data.num_edges}")

加载基准数据集

from torch_geometric.datasets import Planetoid

# Load Cora citation network
dataset = Planetoid(root='/tmp/Cora', name='Cora')
data = dataset[0]  # Get the first (and only) graph

print(f"Dataset: {dataset}")
print(f"Nodes: {data.num_nodes}, Edges: {data.num_edges}")
print(f"Features: {data.num_node_features}, Classes: {dataset.num_classes}")

核心概念

数据结构

PyG 使用 torch_geometric.data.Data 类表示图，其关键属性如下：

• data.x：节点特征矩阵 [num_nodes, num_node_features]
• data.edge_index：采用 COO 格式的图连接关系 [2, num_edges]
• data.edge_attr：边特征矩阵 [num_edges, num_edge_features]（可选）
• data.y：节点或图的目标标签
• data.pos：节点空间位置 [num_nodes, num_dimensions]（可选）
• 自定义属性：可添加任意属性（例如 data.train_mask、data.batch）

重要提示：这些属性并非必需，可根据需要为 Data 对象添加自定义属性。

边索引格式

边以 COO（坐标）格式存储为 [2, num_edges] 张量：

• 第一行：源节点索引
• 第二行：目标节点索引

# Edge list: (0→1), (1→0), (1→2), (2→1)
edge_index = torch.tensor([[0, 1, 1, 2],
                           [1, 0, 2, 1]], dtype=torch.long)

小批量处理

PyG 通过创建块对角邻接矩阵来处理批次，将多个图拼接成一个较大的非连通图：

• 邻接矩阵按对角线堆叠
• 节点特征沿节点维度拼接
• batch 向量将每个节点映射到其所属的源图
• 无需填充，计算效率高

from torch_geometric.loader import DataLoader

loader = DataLoader(dataset, batch_size=32, shuffle=True)
for batch in loader:
    print(f"Batch size: {batch.num_graphs}")
    print(f"Total nodes: {batch.num_nodes}")
    # batch.batch maps nodes to graphs

构建图神经网络

消息传递范式

PyG 中的 GNN 遵循邻域聚合方案：

1. 转换节点特征
2. 沿边传播消息
3. 聚合来自邻居的消息
4. 更新节点表示

使用预构建层

PyG 提供 40 多种卷积层，常见的包括：

GCNConv（图卷积网络）：

from torch_geometric.nn import GCNConv
import torch.nn.functional as F

class GCN(torch.nn.Module):
    def __init__(self, num_features, num_classes):
        super().__init__()
        self.conv1 = GCNConv(num_features, 16)
        self.conv2 = GCNConv(16, num_classes)

    def forward(self, data):
        x, edge_index = data.x, data.edge_index
        x = self.conv1(x, edge_index)
        x = F.relu(x)
        x = F.dropout(x, training=self.training)
        x = self.conv2(x, edge_index)
        return F.log_softmax(x, dim=1)

GATConv（图注意力网络）：

from torch_geometric.nn import GATConv

class GAT(torch.nn.Module):
    def __init__(self, num_features, num_classes):
        super().__init__()
        self.conv1 = GATConv(num_features, 8, heads=8, dropout=0.6)
        self.conv2 = GATConv(8 * 8, num_classes, heads=1, concat=False, dropout=0.6)

    def forward(self, data):
        x, edge_index = data.x, data.edge_index
        x = F.dropout(x, p=0.6, training=self.training)
        x = F.elu(self.conv1(x, edge_index))
        x = F.dropout(x, p=0.6, training=self.training)
        x = self.conv2(x, edge_index)
        return F.log_softmax(x, dim=1)

GraphSAGE：

from torch_geometric.nn import SAGEConv

class GraphSAGE(torch.nn.Module):
    def __init__(self, num_features, num_classes):
        super().__init__()
        self.conv1 = SAGEConv(num_features, 64)
        self.conv2 = SAGEConv(64, num_classes)

    def forward(self, data):
        x, edge_index = data.x, data.edge_index
        x = self.conv1(x, edge_index)
        x = F.relu(x)
        x = F.dropout(x, training=self.training)
        x = self.conv2(x, edge_index)
        return F.log_softmax(x, dim=1)

自定义消息传递层

要创建自定义层，可继承 MessagePassing：

from torch_geometric.nn import MessagePassing
from torch_geometric.utils import add_self_loops, degree

class CustomConv(MessagePassing):
    def __init__(self, in_channels, out_channels):
        super().__init__(aggr='add')  # "add", "mean", or "max"
        self.lin = torch.nn.Linear(in_channels, out_channels)

    def forward(self, x, edge_index):
        # Add self-loops to adjacency matrix
        edge_index, _ = add_self_loops(edge_index, num_nodes=x.size(0))

        # Transform node features
        x = self.lin(x)

        # Compute normalization
        row, col = edge_index
        deg = degree(col, x.size(0), dtype=x.dtype)
        deg_inv_sqrt = deg.pow(-0.5)
        norm = deg_inv_sqrt[row] * deg_inv_sqrt[col]

        # Propagate messages
        return self.propagate(edge_index, x=x, norm=norm)

    def message(self, x_j, norm):
        # x_j: features of source nodes
        return norm.view(-1, 1) * x_j

关键方法：

• forward()：主要入口
• message()：构建从源节点到目标节点的消息
• aggregate()：聚合消息（通常无需重写，可设置 aggr 参数）
• update()：聚合后更新节点嵌入

变量命名约定：在张量名称后添加 _i 或 _j，即可自动将其映射到目标节点或源节点。

使用数据集

加载内置数据集

PyG 提供丰富的基准数据集：

# Citation networks (node classification)
from torch_geometric.datasets import Planetoid
dataset = Planetoid(root='/tmp/Cora', name='Cora')  # or 'CiteSeer', 'PubMed'

# Graph classification
from torch_geometric.datasets import TUDataset
dataset = TUDataset(root='/tmp/ENZYMES', name='ENZYMES')

# Molecular datasets
from torch_geometric.datasets import QM9
dataset = QM9(root='/tmp/QM9')

# Large-scale datasets
from torch_geometric.datasets import Reddit
dataset = Reddit(root='/tmp/Reddit')

完整列表请参阅 references/datasets_reference.md。

创建自定义数据集

对于可装入内存的数据集，可继承 InMemoryDataset：

from torch_geometric.data import InMemoryDataset, Data
import torch

class MyOwnDataset(InMemoryDataset):
    def __init__(self, root, transform=None, pre_transform=None):
        super().__init__(root, transform, pre_transform)
        self.load(self.processed_paths[0])

    @property
    def raw_file_names(self):
        return ['my_data.csv']  # Files needed in raw_dir

    @property
    def processed_file_names(self):
        return ['data.pt']  # Files in processed_dir

    def download(self):
        # Download raw data to self.raw_dir
        pass

    def process(self):
        # Read data, create Data objects
        data_list = []

        # Example: Create a simple graph
        edge_index = torch.tensor([[0, 1], [1, 0]], dtype=torch.long)
        x = torch.randn(2, 16)
        y = torch.tensor([0], dtype=torch.long)

        data = Data(x=x, edge_index=edge_index, y=y)
        data_list.append(data)

        # Apply pre_filter and pre_transform
        if self.pre_filter is not None:
            data_list = [d for d in data_list if self.pre_filter(d)]

        if self.pre_transform is not None:
            data_list = [self.pre_transform(d) for d in data_list]

        # Save processed data
        self.save(data_list, self.processed_paths[0])

对于无法装入内存的大型数据集，可继承 Dataset，并实现 len() 和 get(idx)。

从 CSV 加载图

import pandas as pd
import torch
from torch_geometric.data import HeteroData

# Load nodes
nodes_df = pd.read_csv('nodes.csv')
x = torch.tensor(nodes_df[['feat1', 'feat2']].values, dtype=torch.float)

# Load edges
edges_df = pd.read_csv('edges.csv')
edge_index = torch.tensor([edges_df['source'].values,
                           edges_df['target'].values], dtype=torch.long)

data = Data(x=x, edge_index=edge_index)

训练工作流

节点分类（单个图）

import torch
import torch.nn.functional as F
from torch_geometric.datasets import Planetoid

# Load dataset
dataset = Planetoid(root='/tmp/Cora', name='Cora')
data = dataset[0]

# Create model
model = GCN(dataset.num_features, dataset.num_classes)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=5e-4)

# Training
model.train()
for epoch in range(200):
    optimizer.zero_grad()
    out = model(data)
    loss = F.nll_loss(out[data.train_mask], data.y[data.train_mask])
    loss.backward()
    optimizer.step()

    if epoch % 10 == 0:
        print(f'Epoch {epoch}, Loss: {loss.item():.4f}')

# Evaluation
model.eval()
pred = model(data).argmax(dim=1)
correct = (pred[data.test_mask] == data.y[data.test_mask]).sum()
acc = int(correct) / int(data.test_mask.sum())
print(f'Test Accuracy: {acc:.4f}')

图分类（多个图）

from torch_geometric.datasets import TUDataset
from torch_geometric.loader import DataLoader
from torch_geometric.nn import global_mean_pool

class GraphClassifier(torch.nn.Module):
    def __init__(self, num_features, num_classes):
        super().__init__()
        self.conv1 = GCNConv(num_features, 64)
        self.conv2 = GCNConv(64, 64)
        self.lin = torch.nn.Linear(64, num_classes)

    def forward(self, data):
        x, edge_index, batch = data.x, data.edge_index, data.batch

        x = self.conv1(x, edge_index)
        x = F.relu(x)
        x = self.conv2(x, edge_index)
        x = F.relu(x)

        # Global pooling (aggregate node features to graph-level)
        x = global_mean_pool(x, batch)

        x = self.lin(x)
        return F.log_softmax(x, dim=1)

# Load dataset
dataset = TUDataset(root='/tmp/ENZYMES', name='ENZYMES')
loader = DataLoader(dataset, batch_size=32, shuffle=True)

model = GraphClassifier(dataset.num_features, dataset.num_classes)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

# Training
model.train()
for epoch in range(100):
    total_loss = 0
    for batch in loader:
        optimizer.zero_grad()
        out = model(batch)
        loss = F.nll_loss(out, batch.y)
        loss.backward()
        optimizer.step()
        total_loss += loss.item()

    if epoch % 10 == 0:
        print(f'Epoch {epoch}, Loss: {total_loss / len(loader):.4f}')

使用邻居采样处理大规模图

对于大型图，使用 NeighborLoader 对子图进行采样：

from torch_geometric.loader import NeighborLoader

# Create a neighbor sampler
train_loader = NeighborLoader(
    data,
    num_neighbors=[25, 10],  # Sample 25 neighbors for 1st hop, 10 for 2nd hop
    batch_size=128,
    input_nodes=data.train_mask,
)

# Training
model.train()
for batch in train_loader:
    optimizer.zero_grad()
    out = model(batch)
    # Only compute loss on seed nodes (first batch_size nodes)
    loss = F.nll_loss(out[:batch.batch_size], batch.y[:batch.batch_size])
    loss.backward()
    optimizer.step()

重要：

• 输出的子图是有向图
• 节点索引会重新编号（从 0 到 batch.num_nodes - 1）
• 仅使用种子节点的预测计算损失
• 通常无法采样超过 2-3 跳

高级功能

异构图

对于包含多种节点和边类型的图，使用 HeteroData：

from torch_geometric.data import HeteroData

data = HeteroData()

# Add node features for different types
data['paper'].x = torch.randn(100, 128)  # 100 papers with 128 features
data['author'].x = torch.randn(200, 64)  # 200 authors with 64 features

# Add edges for different types (source_type, edge_type, target_type)
data['author', 'writes', 'paper'].edge_index = torch.randint(0, 200, (2, 500))
data['paper', 'cites', 'paper'].edge_index = torch.randint(0, 100, (2, 300))

print(data)

将同构模型转换为异构模型：

from torch_geometric.nn import to_hetero

# Define homogeneous model
model = GNN(...)

# Convert to heterogeneous
model = to_hetero(model, data.metadata(), aggr='sum')

# Use as normal
out = model(data.x_dict, data.edge_index_dict)

或者使用 HeteroConv 执行自定义的、针对特定边类型的操作：

from torch_geometric.nn import HeteroConv, GCNConv, SAGEConv

class HeteroGNN(torch.nn.Module):
    def __init__(self, metadata):
        super().__init__()
        self.conv1 = HeteroConv({
            ('paper', 'cites', 'paper'): GCNConv(-1, 64),
            ('author', 'writes', 'paper'): SAGEConv((-1, -1), 64),
        }, aggr='sum')

        self.conv2 = HeteroConv({
            ('paper', 'cites', 'paper'): GCNConv(64, 32),
            ('author', 'writes', 'paper'): SAGEConv((64, 64), 32),
        }, aggr='sum')

    def forward(self, x_dict, edge_index_dict):
        x_dict = self.conv1(x_dict, edge_index_dict)
        x_dict = {key: F.relu(x) for key, x in x_dict.items()}
        x_dict = self.conv2(x_dict, edge_index_dict)
        return x_dict

变换

应用变换以修改图结构或特征：

from torch_geometric.transforms import NormalizeFeatures, AddSelfLoops, Compose

# Single transform
transform = NormalizeFeatures()
dataset = Planetoid(root='/tmp/Cora', name='Cora', transform=transform)

# Compose multiple transforms
transform = Compose([
    AddSelfLoops(),
    NormalizeFeatures(),
])
dataset = Planetoid(root='/tmp/Cora', name='Cora', transform=transform)

常见变换：

• 结构：ToUndirected、AddSelfLoops、RemoveSelfLoops、KNNGraph、RadiusGraph
• 特征：NormalizeFeatures、NormalizeScale、Center
• 采样：RandomNodeSplit、RandomLinkSplit
• 位置编码：AddLaplacianEigenvectorPE、AddRandomWalkPE

完整列表请参阅 references/transforms_reference.md。

模型可解释性

PyG 提供了可解释性工具，用于理解模型的预测结果：

from torch_geometric.explain import Explainer, GNNExplainer

# Create explainer
explainer = Explainer(
    model=model,
    algorithm=GNNExplainer(epochs=200),
    explanation_type='model',  # or 'phenomenon'
    node_mask_type='attributes',
    edge_mask_type='object',
    model_config=dict(
        mode='multiclass_classification',
        task_level='node',
        return_type='log_probs',
    ),
)

# Generate explanation for a specific node
node_idx = 10
explanation = explainer(data.x, data.edge_index, index=node_idx)

# Visualize
print(f'Node {node_idx} explanation:')
print(f'Important edges: {explanation.edge_mask.topk(5).indices}')
print(f'Important features: {explanation.node_mask[node_idx].topk(5).indices}')

池化操作

用于层次化图表示：

from torch_geometric.nn import TopKPooling, global_mean_pool

class HierarchicalGNN(torch.nn.Module):
    def __init__(self, num_features, num_classes):
        super().__init__()
        self.conv1 = GCNConv(num_features, 64)
        self.pool1 = TopKPooling(64, ratio=0.8)
        self.conv2 = GCNConv(64, 64)
        self.pool2 = TopKPooling(64, ratio=0.8)
        self.lin = torch.nn.Linear(64, num_classes)

    def forward(self, data):
        x, edge_index, batch = data.x, data.edge_index, data.batch

        x = F.relu(self.conv1(x, edge_index))
        x, edge_index, _, batch, _, _ = self.pool1(x, edge_index, None, batch)

        x = F.relu(self.conv2(x, edge_index))
        x, edge_index, _, batch, _, _ = self.pool2(x, edge_index, None, batch)

        x = global_mean_pool(x, batch)
        x = self.lin(x)
        return F.log_softmax(x, dim=1)

常见模式和最佳实践

检查图属性

# Undirected check
from torch_geometric.utils import is_undirected
print(f"Is undirected: {is_undirected(data.edge_index)}")

# Connected components
from torch_geometric.utils import connected_components
print(f"Connected components: {connected_components(data.edge_index)}")

# Contains self-loops
from torch_geometric.utils import contains_self_loops
print(f"Has self-loops: {contains_self_loops(data.edge_index)}")

GPU 训练

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)
data = data.to(device)

# For DataLoader
for batch in loader:
    batch = batch.to(device)
    # Train...

保存和加载模型

# Save
torch.save(model.state_dict(), 'model.pth')

# Load
model = GCN(num_features, num_classes)
model.load_state_dict(torch.load('model.pth'))
model.eval()

层的能力

选择层时，请考虑以下能力：

• SparseTensor：支持高效的稀疏矩阵操作
• edge_weight：处理一维边权重
• edge_attr：处理多维边特征
• Bipartite：适用于二部图（源节点和目标节点维度不同）
• Lazy：无需指定输入维度即可初始化

请参阅 references/layer_capabilities.md 中的 GNN 速查表。

资源

随附参考资料

此技能包含详细的参考文档：

• references/layers_reference.md：完整列出所有 40 多个 GNN 层，并附有说明和能力介绍
• references/datasets_reference.md：按类别整理的综合数据集目录
• references/transforms_reference.md：所有可用变换及其使用场景
• references/api_patterns.md：常见 API 模式和代码示例

脚本

​​​​scripts/ 中提供了实用脚本：

• scripts/visualize_graph.py：使用 networkx 和 matplotlib 可视化图结构
• scripts/create_gnn_template.py：为常见 GNN 架构生成样板代码
• scripts/benchmark_model.py：在标准数据集上对模型性能进行基准测试

可直接执行这些脚本，也可阅读脚本以了解实现模式。

官方资源

• 文档：https://pytorch-geometric.readthedocs.io/
• GitHub：https://github.com/pyg-team/pytorch_geometric
• 教程：https://pytorch-geometric.readthedocs.io/en/latest/get_started/introduction.html
• 示例：https://github.com/pyg-team/pytorch_geometric/tree/master/examples