目录
基于注意力的池化（Attention-based Pooling）
Variational Graph Autoencoder (VGAE)
PyTorch Geometric（PyG）是PyTorch的一个扩展库，专注于图神经网络（GNN）的实现。它提供了丰富的图数据处理工具、图神经网络层和模型。以下是对PyG库中常用方法的介绍
数据处理与转换
数据表示
PyG使用 torch_geometric.data.Data 类来表示图数据，包含节点特征 x 、边索引 edge_index 、边特征 edge_attr 等
## 数据处理与转换
# 1. 数据表示
import torch
from torch_geometric.data import Data
# 创建一个简单的图
x = torch.tensor([[1, 2], [3, 4], [5, 6]], dtype=torch.float) # 节点特征
edge_index = torch.tensor([[0, 1, 1, 2], [1, 0, 2, 1]], dtype=torch.long) # 边索引
edge_attr = torch.tensor([1.0, 2.0, 3.0, 4.0], dtype=torch.float) # 边特征
data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr)
print(data) # Data(x=[3, 2], edge_index=[2, 4], edge_attr=[4])数据加载
PyG提供了 torch_geometric.data.DataLoader 类，用于批量加载图数据
# 2. 数据加载
from torch_geometric.datasets import Planetoid
from torch_geometric.loader import DataLoader
# 加载Cora数据集
dataset = Planetoid(root='./data', name='Cora')
loader = DataLoader(dataset, batch_size=32, shuffle=True)
print(f"节点数: {data.num_nodes}") # 3
print(f"边数: {data.num_edges}") # 4
print(f"特征维度: {data.num_node_features}") # 2
print(f"类别数: {dataset.num_classes}") # 7
for batch in loader:
print(batch)
# DataBatch(x=[2708, 1433], edge_index=[2, 10556], y=[2708], train_mask=[2708], val_mask=[2708], test_mask=[2708],
# batch=[2708], ptr=[2])数据转换
特征归一化
NormalizeFeatures 是一个常用的转换方法，用于将节点特征归一化到单位范数（如 0, 1 或 -1, 1）
# 3. 数据转换
# 3.1 特征归一化
from torch_geometric.transforms import NormalizeFeatures
dataset = Planetoid(root='./data', name='Cora', transform=NormalizeFeatures())
# 查看归一化后的特征
data = dataset[0]
print(data.x)
# tensor([[0., 0., 0., ..., 0., 0., 0.],
# [0., 0., 0., ..., 0., 0., 0.],
# [0., 0., 0., ..., 0., 0., 0.],
# ...,
# [0., 0., 0., ..., 0., 0., 0.],
# [0., 0., 0., ..., 0., 0., 0.],
# [0., 0., 0., ..., 0., 0., 0.]])添加自环
AddSelfLoops 是一个转换方法，用于为图中的每个节点添加自环（即每个节点连接到自己）
# 3.2 添加自环
from torch_geometric.transforms import AddSelfLoops
dataset = Planetoid(root='./data', name='Cora', transform=AddSelfLoops())
# 查看添加自环后的边索引
data = dataset[0]
print(data.edge_index)
# tensor([[ 0, 0, 0, ..., 2705, 2706, 2707],
# [ 633, 1862, 2582, ..., 2705, 2706, 2707]])随机扰动
RandomNodeSplit 是一个转换方法，用于随机划分训练集、验证集和测试集
# 3.3 随机扰动
from torch_geometric.transforms import RandomNodeSplit
dataset = Planetoid(root='./data', name='Cora', transform=RandomNodeSplit(num_splits=10))
# 查看划分后的掩码
data = dataset[0]
print(data.train_mask)
# tensor([[False, True, True, ..., False, False, True],
# [False, False, True, ..., True, False, False],
# [False, True, False, ..., False, False, False],
# ...,
# [ True, True, True, ..., False, False, False],
# [ True, True, True, ..., False, False, True],
# [ True, True, True, ..., True, False, True]])
print(data.val_mask)
# tensor([[False, False, False, ..., False, True, False],
# [False, False, False, ..., False, False, False],
# [False, False, False, ..., False, False, True],
# ...,
# [False, False, False, ..., True, True, False],
# [False, False, False, ..., False, True, False],
# [False, False, False, ..., False, False, False]])
print(data.test_mask)
# tensor([[ True, False, False, ..., True, False, False],
# [ True, True, False, ..., False, True, True],
# [ True, False, True, ..., True, True, False],
# ...,
# [False, False, False, ..., False, False, True],
# [False, False, False, ..., True, False, False],
# [False, False, False, ..., False, True, False]])组合转换
可以将多个转换方法组合在一起，形成一个复合转换
# 3.4 组合转换
from torch_geometric.transforms import Compose, NormalizeFeatures, AddSelfLoops
# 定义一个复合转换
transform = Compose([NormalizeFeatures(), AddSelfLoops()])
# 创建一个数据集，并应用复合转换
dataset = Planetoid(root='./data', name='Cora', transform=transform)
# 查看转换后的数据
data = dataset[0]
print(data.x)
# tensor([[0., 0., 0., ..., 0., 0., 0.],
# [0., 0., 0., ..., 0., 0., 0.],
# [0., 0., 0., ..., 0., 0., 0.],
# ...,
# [0., 0., 0., ..., 0., 0., 0.],
# [0., 0., 0., ..., 0., 0., 0.],
# [0., 0., 0., ..., 0., 0., 0.]])
print(data.edge_index)
# tensor([[ 0, 0, 0, ..., 2705, 2706, 2707],
# [ 633, 1862, 2582, ..., 2705, 2706, 2707]])图神经网络层
图卷积层（GCNConv）
GCNConv是图卷积网络（GCN）的基本层
## 图神经网络层
# 1. 图卷积层 GCNConv
import torch
from torch_geometric.nn import GCNConv
class GCN(torch.nn.Module):
def __init__(self, in_channels, out_channels):
super(GCN, self).__init__()
self.conv1 = GCNConv(in_channels, 16)
self.conv2 = GCNConv(16, out_channels)
def forward(self, x, edge_index):
x = self.conv1(x, edge_index)
x = torch.relu(x)
x = self.conv2(x, edge_index)
return x
model = GCN(in_channels=dataset.num_features, out_channels=dataset.num_classes)
print(model)
# GCN(
# (conv1): GCNConv(1433, 16)
# (conv2): GCNConv(16, 7)
# )图注意力层（GATConv）
GATConv是图注意力网络（GAT）的基本层
# 2. 图注意力层 GATConv
from torch_geometric.nn import GATConv
class GAT(torch.nn.Module):
def __init__(self, in_channels, out_channels):
super(GAT, self).__init__()
self.conv1 = GATConv(in_channels, 8, heads=8, dropout=0.6)
self.conv2 = GATConv(8 * 8, out_channels, heads=1, concat=True, dropout=0.6)
def forward(self, x, edge_index):
x = torch.dropout(x, p=0.6, training=self.training)
x = self.conv1(x, edge_index)
x = torch.relu(x)
x = torch.dropout(x, p=0.6, training=self.training)
x = self.conv2(x, edge_index)
return x
model = GAT(in_channels=dataset.num_features, out_channels=dataset.num_classes)
print(model)
# GAT(
# (conv1): GATConv(1433, 8, heads=8)
# (conv2): GATConv(64, 7, heads=1)
# )池化
全局池化（Global Pooling）
全局池化将整个图的所有节点聚合为一个全局表示
全局平均池化
## 池化
# 1. 全局池化
# 1.1 全局平均池化
from torch_geometric.nn import global_mean_pool
from torch_geometric.datasets import TUDataset
from torch_geometric.loader import DataLoader
# 加载数据集
dataset = TUDataset(root='./data', name='MUTAG')
loader = DataLoader(dataset, batch_size=32, shuffle=True)
# 获取一个批次的数据
for batch in loader:
x = batch.x
batch_index = batch.batch
global_mean = global_mean_pool(x, batch_index)
print("Global Mean Pooling Result:", global_mean)
break
# tensor([[0.7647, 0.0588, 0.1176, 0.0000, 0.0588, 0.0000, 0.0000],
# [0.7500, 0.1250, 0.1250, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.6250, 0.1250, 0.2500, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.5217, 0.1739, 0.3043, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.5455, 0.2727, 0.1818, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.6400, 0.1200, 0.2400, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.6364, 0.0909, 0.2727, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.7857, 0.0714, 0.1429, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.8000, 0.0667, 0.1333, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.7647, 0.0588, 0.1176, 0.0000, 0.0000, 0.0000, 0.0588],
# [0.5000, 0.1667, 0.3333, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.7692, 0.0769, 0.1538, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.7826, 0.0435, 0.1739, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.8000, 0.0500, 0.1500, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.8696, 0.0435, 0.0870, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.7273, 0.0909, 0.1818, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.6667, 0.0833, 0.2500, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.7273, 0.0909, 0.1818, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.8125, 0.0625, 0.1250, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.8235, 0.0588, 0.1176, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.6000, 0.1000, 0.2000, 0.0000, 0.0000, 0.1000, 0.0000],
# [0.4615, 0.1538, 0.3077, 0.0000, 0.0000, 0.0769, 0.0000],
# [0.7647, 0.0588, 0.1765, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.6000, 0.0500, 0.2000, 0.0000, 0.0000, 0.1500, 0.0000],
# [0.6000, 0.1000, 0.2000, 0.1000, 0.0000, 0.0000, 0.0000],
# [0.7647, 0.0588, 0.1176, 0.0000, 0.0000, 0.0588, 0.0000],
# [0.8000, 0.0667, 0.1333, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.4615, 0.1538, 0.3077, 0.0769, 0.0000, 0.0000, 0.0000],
# [0.8696, 0.0435, 0.0870, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.8696, 0.0435, 0.0870, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.7273, 0.0909, 0.1818, 0.0000, 0.0000, 0.0000, 0.0000],
# [0.8421, 0.0526, 0.1053, 0.0000, 0.0000, 0.0000, 0.0000]])全局最大池化
# 1.2 全局最大池化
from torch_geometric.nn import global_max_pool
# 获取一个批次的数据
for batch in loader:
x = batch.x
batch_index = batch.batch
global_max = global_max_pool(x, batch_index)
print("Global Max Pooling Result:", global_max)
break全局求和池化
# 3. 全局求和池化
from torch_geometric.nn import global_add_pool
# 获取一个批次的数据
for batch in loader:
x = batch.x
batch_index = batch.batch
global_sum = global_add_pool(x, batch_index)
print("Global Sum Pooling Result:", global_sum)
break
基于注意力的池化（Attention-based Pooling）
基于注意力的池化方法通过学习节点的重要性权重来进行池化。一个常见的例子是 Set2Set 池化
# 2. 基于注意力的池化——Set2Set
from torch_geometric.nn import Set2Set
from torch_geometric.datasets import TUDataset
from torch_geometric.loader import DataLoader
# 加载数据集
dataset = TUDataset(root='./data', name='MUTAG')
loader = DataLoader(dataset, batch_size=32, shuffle=True)
# 定义 Set2Set 池化
set2set = Set2Set(in_channels=dataset.num_node_features, processing_steps=3)
# 获取一个批次的数据
for batch in loader:
x = batch.x
batch_index = batch.batch
global_set2set = set2set(x, batch_index)
print("Set2Set Pooling Result:", global_set2set)
break
# Set2Set Pooling Result: tensor([[ 0.1719, 0.0986, 0.1594, -0.0438, 0.1743, 0.1663, -0.0578, 0.8464,
# 0.0492, 0.1045, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1730, 0.0987, 0.1601, -0.0420, 0.1730, 0.1658, -0.0549, 0.8733,
# 0.0405, 0.0862, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1686, 0.0919, 0.1603, -0.0525, 0.1807, 0.1707, -0.0683, 0.7540,
# 0.0466, 0.1994, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1601, 0.1165, 0.1425, -0.0525, 0.1782, 0.1602, -0.0836, 0.6232,
# 0.2237, 0.1531, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1725, 0.0987, 0.1598, -0.0428, 0.1736, 0.1660, -0.0562, 0.8611,
# 0.0444, 0.0945, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1673, 0.1060, 0.1527, -0.0473, 0.1761, 0.1642, -0.0679, 0.7570,
# 0.1187, 0.1243, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1579, 0.0996, 0.1486, -0.0658, 0.1874, 0.1695, -0.0954, 0.5284,
# 0.1662, 0.3054, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1584, 0.0969, 0.1503, -0.0665, 0.1881, 0.1709, -0.0949, 0.5327,
# 0.1503, 0.3170, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1634, 0.0976, 0.1537, -0.0581, 0.1835, 0.1695, -0.0809, 0.6464,
# 0.1135, 0.2401, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1704, 0.0952, 0.1599, -0.0479, 0.1774, 0.1684, -0.0626, 0.8042,
# 0.0466, 0.1492, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1634, 0.1081, 0.1488, -0.0522, 0.1789, 0.1640, -0.0776, 0.6743,
# 0.1595, 0.1661, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1671, 0.0980, 0.1563, -0.0518, 0.1797, 0.1682, -0.0707, 0.7332,
# 0.0855, 0.1813, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1564, 0.1193, 0.1384, -0.0562, 0.1800, 0.1590, -0.0922, 0.5527,
# 0.2663, 0.1810, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1704, 0.0952, 0.1599, -0.0479, 0.1774, 0.1684, -0.0626, 0.8042,
# 0.0466, 0.1492, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1671, 0.0980, 0.1563, -0.0518, 0.1797, 0.1682, -0.0707, 0.7332,
# 0.0855, 0.1813, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1730, 0.0987, 0.1601, -0.0420, 0.1730, 0.1658, -0.0549, 0.8733,
# 0.0405, 0.0862, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1604, 0.0972, 0.1517, -0.0631, 0.1863, 0.1704, -0.0893, 0.5779,
# 0.1356, 0.2864, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1711, 0.0985, 0.1589, -0.0451, 0.1752, 0.1666, -0.0599, 0.8281,
# 0.0550, 0.1169, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1584, 0.1178, 0.1406, -0.0542, 0.1790, 0.1597, -0.0875, 0.5910,
# 0.2432, 0.1659, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1673, 0.1060, 0.1527, -0.0473, 0.1761, 0.1642, -0.0679, 0.7570,
# 0.1187, 0.1243, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1602, 0.1097, 0.1457, -0.0562, 0.1811, 0.1638, -0.0856, 0.6077,
# 0.1926, 0.1997, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1696, 0.1047, 0.1549, -0.0444, 0.1743, 0.1641, -0.0623, 0.8062,
# 0.0945, 0.0993, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1604, 0.0972, 0.1517, -0.0631, 0.1863, 0.1704, -0.0893, 0.5779,
# 0.1356, 0.2864, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1671, 0.0980, 0.1563, -0.0518, 0.1797, 0.1682, -0.0707, 0.7332,
# 0.0855, 0.1813, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1638, 0.0919, 0.1567, -0.0605, 0.1855, 0.1723, -0.0815, 0.6416,
# 0.0853, 0.2731, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1732, 0.1049, 0.1525, -0.0508, 0.1755, 0.1624, -0.0665, 0.7700,
# 0.0553, 0.1160, 0.0000, 0.0000, 0.0586, 0.0000],
# [ 0.1711, 0.0985, 0.1589, -0.0451, 0.1752, 0.1666, -0.0599, 0.8281,
# 0.0550, 0.1169, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1573, 0.0968, 0.1494, -0.0685, 0.1891, 0.1712, -0.0982, 0.5063,
# 0.1589, 0.3349, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1729, 0.1053, 0.1365, -0.0582, 0.1904, 0.1594, -0.0881, 0.5637,
# 0.0878, 0.1812, 0.0746, 0.0000, 0.0927, 0.0000],
# [ 0.1586, 0.1026, 0.1477, -0.0628, 0.1855, 0.1678, -0.0924, 0.5526,
# 0.1742, 0.2733, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1646, 0.1075, 0.1500, -0.0506, 0.1781, 0.1641, -0.0746, 0.6999,
# 0.1469, 0.1533, 0.0000, 0.0000, 0.0000, 0.0000],
# [ 0.1695, 0.0983, 0.1579, -0.0477, 0.1770, 0.1672, -0.0641, 0.7909,
# 0.0670, 0.1421, 0.0000, 0.0000, 0.0000, 0.0000]],
# grad_fn=<CatBackward0>)基于图的池化（Graph-based Pooling）
基于图的池化方法通过图的结构信息来进行池化。常见的方法包括 TopKPooling，通过选择重要性最高的节点来进行池化
# 3. 基于图的池化——TopKPooling
from torch_geometric.nn import TopKPooling
from torch_geometric.datasets import TUDataset
from torch_geometric.loader import DataLoader
# 加载数据集
dataset = TUDataset(root='./data', name='MUTAG')
loader = DataLoader(dataset, batch_size=32, shuffle=True)
# 定义 TopKPooling
pool = TopKPooling(in_channels=dataset.num_node_features, ratio=0.5)
# 获取一个批次的数据
for batch in loader:
x = batch.x
edge_index = batch.edge_index
batch_index = batch.batch
x, edge_index, _, batch_index, _, _ = pool(x, edge_index, batch=batch_index)
print("TopKPooling Result:", x)
break
# tensor([[-0.0000, -0.0392, -0.0000, ..., -0.0000, -0.0000, -0.0000],
# [-0.0000, -0.0392, -0.0000, ..., -0.0000, -0.0000, -0.0000],
# [-0.0000, -0.0392, -0.0000, ..., -0.0000, -0.0000, -0.0000],
# ...,
# [-0.4577, -0.0000, -0.0000, ..., -0.0000, -0.0000, -0.0000],
# [-0.4577, -0.0000, -0.0000, ..., -0.0000, -0.0000, -0.0000],
# [-0.4577, -0.0000, -0.0000, ..., -0.0000, -0.0000, -0.0000]],
# grad_fn=<MulBackward0>)层次化池化（Hierarchical Pooling）
层次化池化通过多层池化操作生成图的层次化表示。一个常见的例子是 EdgePooling，通过边的合并操作来进行池化
# 4. 层次化池化——EdgePooling
from torch_geometric.nn import EdgePooling
from torch_geometric.datasets import TUDataset
from torch_geometric.loader import DataLoader
# 加载数据集
dataset = TUDataset(root='./data', name='MUTAG')
loader = DataLoader(dataset, batch_size=32, shuffle=True)
# 定义 EdgePooling
pool = EdgePooling(in_channels=dataset.num_node_features)
# 获取一个批次的数据
for batch in loader:
x = batch.x
edge_index = batch.edge_index
batch_index = batch.batch
x, edge_index, batch_index, _ = pool(x, edge_index, batch=batch_index)
print("EdgePooling Result:", x)
break
# tensor([[0.0000, 1.5000, 1.5000, ..., 0.0000, 0.0000, 0.0000],
# [0.0000, 1.5000, 1.5000, ..., 0.0000, 0.0000, 0.0000],
# [0.0000, 1.5000, 1.5000, ..., 0.0000, 0.0000, 0.0000],
# ...,
# [0.0000, 0.0000, 1.0000, ..., 0.0000, 0.0000, 0.0000],
# [1.0000, 0.0000, 0.0000, ..., 0.0000, 0.0000, 0.0000],
# [0.0000, 0.0000, 1.0000, ..., 0.0000, 0.0000, 0.0000]],
# grad_fn=<MulBackward0>)采样
子图采样（Subgraph Sampling）
子图采样是从原始图中提取一个子图，通常用于减少计算复杂度和增强模型的泛化能力
## 采样
# 1. 子图采样
import torch
from torch_geometric.data import Data
from torch_geometric.datasets import Planetoid
from torch_geometric.utils import k_hop_subgraph
# 加载数据集
dataset = Planetoid(root='./data', name='Cora')
data = dataset[0]
# 选择一个起始节点
start_node = 0
num_hops = 2 # 采样半径
# 提取子图
sub_nodes, sub_edge_index, mapping, _ = k_hop_subgraph(start_node, num_hops, data.edge_index)
# 创建子图
sub_data = Data(x=data.x[sub_nodes], edge_index=sub_edge_index, y=data.y[sub_nodes])
print("Original Graph Nodes:", data.num_nodes) # 2708
print("Subgraph Nodes:", sub_data.num_nodes) # 8
print("Subgraph Edges:", sub_data.edge_index.shape[1]) # 20邻域采样（Neighbor Sampling）
邻域采样通过选择节点的邻居来生成子图，适用于大规模图数据
# 2. 邻域采样
from torch_geometric.datasets import Planetoid
from torch_geometric.loader import NeighborLoader
# 加载数据集
dataset = Planetoid(root='./data', name='Cora')
data = dataset[0]
# 定义 NeighborSampler
loader = NeighborLoader(
data,
num_neighbors=[10, 10], # 每层采样的邻居数量
batch_size=1024,
shuffle=True,
)
# 遍历数据加载器
for batch in loader:
print(batch)
break模型训练与评估
训练过程
## 模型训练与评估
# 1. 训练过程
import torch.nn.functional as F
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
criterion = torch.nn.CrossEntropyLoss()
def train():
model.train()
optimizer.zero_grad()
out = model(data.x, data.edge_index)
loss = criterion(out[data.train_mask], data.y[data.train_mask])
loss.backward()
optimizer.step()
return loss测试过程
# 2. 测试过程
@torch.no_grad()
def test():
model.eval()
out = model(data.x, data.edge_index)
pred = out.argmax(dim=1)
correct = int(pred[data.test_mask].eq(data.y[data.test_mask]).sum().item())
acc = correct / int(data.test_mask.sum())
return acc
for epoch in range(200):
loss = train()
acc = test()
print(f'Epoch: {epoch + 1}, Loss: {loss:.4f}, Accuracy: {acc:.4f}')异构图处理
异构图定义
## 异构图处理
# 1. 异构图定义
from torch_geometric.data import HeteroData
import torch
data = HeteroData()
# 添加两种类型节点
data['user'].x = torch.randn(4, 16) # 4个用户
data['movie'].x = torch.randn(5, 32) # 5部电影
# 添加边
data['user', 'rates', 'movie'].edge_index = torch.tensor(
[[0, 0, 1, 2, 3], [0, 2, 3, 1, 4]] # user->movie评分关系
)异构图卷积
# 2. 异构图卷积
from torch_geometric.nn import HeteroConv, GCNConv, SAGEConv
from torch_geometric.transforms import NormalizeFeatures
class HeteroGNN(torch.nn.Module):
def __init__(self, in_channels, out_channels, hidden_channels):
super().__init__()
self.conv1 = HeteroConv({
('user', 'rates', 'movie'): SAGEConv((in_channels['user'], in_channels['movie']), hidden_channels),
('movie', 'rev_rates', 'user'): GCNConv(in_channels['movie'], hidden_channels, add_self_loops=False) # 禁用自环
}, aggr='sum')
self.conv2 = HeteroConv({
('user', 'rates', 'movie'): SAGEConv((hidden_channels, hidden_channels), out_channels),
('movie', 'rev_rates', 'user'): GCNConv(hidden_channels, out_channels, add_self_loops=False) # 禁用自环
}, aggr='sum')
def forward(self, x_dict, edge_index_dict):
x_dict = self.conv1(x_dict, edge_index_dict)
x_dict = {key: torch.relu(x) for key, x in x_dict.items()}
x_dict = self.conv2(x_dict, edge_index_dict)
return x_dict
# 定义输入和输出通道数
in_channels = {'user': 16, 'movie': 32}
out_channels = 7 # 假设输出通道数为7
hidden_channels = 64 # 假设隐藏层通道数为64
# 实例化模型
model = HeteroGNN(in_channels, out_channels, hidden_channels)
print(model)
# HeteroGNN(
# (conv1): HeteroConv(num_relations=2)
# (conv2): HeteroConv(num_relations=2)
# )图生成模型
Deep Graph Infomax (DGI)
DGI 是一种无监督图表示学习方法，通过最大化局部和全局图表示之间的一致性来学习节点嵌入
## 图生成模型
# 1. Deep Graph Infomax (DGI)
from torch_geometric.nn import DeepGraphInfomax
from torch_geometric.datasets import Planetoid
from torch_geometric.nn import GCNConv
import torch.nn as nn
import torch.nn.functional as F
class Encoder(nn.Module):
def __init__(self, in_channels, hidden_channels):
super(Encoder, self).__init__()
self.conv = GCNConv(in_channels, hidden_channels)
self.prelu = nn.PReLU(hidden_channels)
def forward(self, x, edge_index):
x = self.conv(x, edge_index)
x = self.prelu(x)
return x
def corruption(x, edge_index):
return x[torch.randperm(x.size(0))], edge_index
dataset = Planetoid(root='./data', name='Cora')
data = dataset[0]
encoder = Encoder(dataset.num_features, hidden_channels=512)
model = DeepGraphInfomax(
hidden_channels=512, encoder=encoder,
summary=lambda z, *args, **kwargs: torch.sigmoid(z.mean(dim=0)),
corruption=corruption
)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
def train():
model.train()
optimizer.zero_grad()
pos_z, neg_z, summary = model(data.x, data.edge_index)
loss = model.loss(pos_z, neg_z, summary)
loss.backward()
optimizer.step()
return loss
for epoch in range(100):
loss = train()
print(f'Epoch: {epoch + 1}, Loss: {loss:.4f}')Graph Autoencoder (GAE)
GAE 是一种基于图神经网络的自编码器，用于图生成任务。它通过学习节点嵌入来重建图的邻接矩阵
# 2. Graph Autoencoder(GAE)
from torch_geometric.nn import GCNConv
from torch_geometric.nn import GAE
import torch.nn.functional as F
class Encoder(nn.Module):
def __init__(self, in_channels, out_channels):
super(Encoder, self).__init__()
self.conv1 = GCNConv(in_channels, 2 * out_channels)
self.conv2 = GCNConv(2 * out_channels, out_channels)
def forward(self, x, edge_index):
x = self.conv1(x, edge_index)
x = F.relu(x)
return self.conv2(x, edge_index)
dataset = Planetoid(root='./data', name='Cora')
data = dataset[0]
encoder = Encoder(dataset.num_features, out_channels=16)
model = GAE(encoder)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
def train():
model.train()
optimizer.zero_grad()
z = model.encode(data.x, data.edge_index)
loss = model.recon_loss(z, data.edge_index)
loss.backward()
optimizer.step()
return loss
for epoch in range(100):
loss = train()
print(f'Epoch: {epoch + 1}, Loss: {loss:.4f}')Variational Graph Autoencoder (VGAE)
VGAE 是 GAE 的变体，通过引入变分推断来学习节点嵌入的分布
# 3. Variational Graph Autoencoder(VGAE)
from torch_geometric.nn import VGAE
from torch_geometric.datasets import Planetoid
# 定义数据集
dataset = Planetoid(root='./data', name='Cora')
data = dataset[0]
class Encoder(nn.Module):
def __init__(self, in_channels, out_channels):
super(Encoder, self).__init__()
self.conv1 = GCNConv(in_channels, 2 * out_channels)
self.conv2 = GCNConv(2 * out_channels, 2 * out_channels)
def forward(self, x, edge_index):
x = self.conv1(x, edge_index)
x = F.relu(x)
x = self.conv2(x, edge_index)
mu = x[:, :x.size(1) // 2]
logstd = x[:, x.size(1) // 2:]
return mu, logstd
# 定义 Encoder
encoder = Encoder(dataset.num_features, out_channels=16)
# 定义 VGAE 模型
model = VGAE(encoder)
# 定义优化器
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
# 训练函数
def train():
model.train()
optimizer.zero_grad()
z = model.encode(data.x, data.edge_index)
loss = model.recon_loss(z, data.edge_index)
kl_loss = model.kl_loss()
loss += kl_loss
loss.backward()
optimizer.step()
return loss
# 训练模型
for epoch in range(100):
loss = train()
print(f'Epoch: {epoch + 1}, Loss: {loss:.4f}')