图卷积神经网络GCN：整图分类（含示例及代码）

关于整图分类，有篇知乎写的很好：【图分类】10分钟就学会的图分类教程，基于pytorch和dgl。下面的代码也是来者这篇知乎。

import dgl
import torch
from torch._C import device
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from dgl.data import MiniGCDataset
from dgl.nn.pytorch import GraphConv
from sklearn.metrics import accuracy_score
 
 
class Classifier(nn.Module):
    def __init__(self, in_dim, hidden_dim, n_classes):
        super(Classifier, self).__init__()
        self.conv1 = GraphConv(in_dim, hidden_dim)  # 定义第一层图卷积
        self.conv2 = GraphConv(hidden_dim, hidden_dim)  # 定义第二层图卷积
        self.classify = nn.Linear(hidden_dim, n_classes)   # 定义分类器
 
    def forward(self, g):
        """g表示批处理后的大图，N表示大图的所有节点数量，n表示图的数量 
        """
        # 为方便，我们用节点的度作为初始节点特征。对于无向图，入度 = 出度
        h = g.in_degrees().view(-1, 1).float() # [N, 1]
        # 执行图卷积和激活函数
        h = F.relu(self.conv1(g, h))  # [N, hidden_dim]
        h = F.relu(self.conv2(g, h))  # [N, hidden_dim]
        g.ndata['h'] = h    # 将特征赋予到图的节点
        # 通过平均池化每个节点的表示得到图表示
        hg = dgl.mean_nodes(g, 'h')   # [n, hidden_dim]
        return self.classify(hg)  # [n, n_classes]
 
def collate(samples):
    # 输入参数samples是一个列表
    # 列表里的每个元素是图和标签对，如[(graph1, label1), (graph2, label2), ...]
    # zip(*samples)是解压操作，解压为[(graph1, graph2, ...), (label1, label2, ...)]
    graphs, labels = map(list, zip(*samples))
    # dgl.batch 将一批图看作是具有许多互不连接的组件构成的大型图
    return dgl.batch(graphs), torch.tensor(labels, dtype=torch.long)
 
 
# 创建训练集和测试集
trainset = MiniGCDataset(2000, 10, 20)  # 生成2000个图，每个图的最小节点数>=10, 最大节点数<=20
testset = MiniGCDataset(1000, 10, 20) 
 
# 用pytorch的DataLoader和之前定义的collect函数
data_loader = DataLoader(trainset, batch_size=64, shuffle=True,
                         collate_fn=collate)
 
DEVICE = torch.device("cuda:2")
# 构造模型 
model = Classifier(1, 256, trainset.num_classes)
model.to(DEVICE)
 
# 定义分类交叉熵损失
loss_func = nn.CrossEntropyLoss()
# 定义Adam优化器
optimizer = optim.Adam(model.parameters(), lr=0.001)
 
# 模型训练
model.train()
epoch_losses = []
for epoch in range(100): 
    epoch_loss = 0
    for iter, (batchg, label) in enumerate(data_loader):
        batchg, label = batchg.to(DEVICE), label.to(DEVICE)
        prediction = model(batchg)
        loss = loss_func(prediction, label)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        epoch_loss += loss.detach().item()
    epoch_loss /= (iter + 1)
    print('Epoch {}, loss {:.4f}'.format(epoch, epoch_loss))
    epoch_losses.append(epoch_loss)
 
 
# 测试
test_loader = DataLoader(testset, batch_size=64, shuffle=False,
                         collate_fn=collate)
model.eval()
test_pred, test_label = [], []
with torch.no_grad():
    for it, (batchg, label) in enumerate(test_loader):
        batchg, label = batchg.to(DEVICE), label.to(DEVICE)
        pred = torch.softmax(model(batchg), 1)
        pred = torch.max(pred, 1)[1].view(-1)
        test_pred += pred.detach().cpu().numpy().tolist()
        test_label += label.cpu().numpy().tolist()
print("Test accuracy: ", accuracy_score(test_label, test_pred))