GCN——DGL教程

官方教程
中文版

安装DGL

内置函数和消息传递API

dgl.function

查阅需要的函数，官方教程没有注释

消息传递框架（Message Passing Paradigm）：
可以对“边” 或者 “节点” 计算，（小白，还不知道怎么运用“边”的特征）

message function消息函数：它通过将边上特征与其两端节点的特征相结合来生成消息。

update function更新函数：结合聚合后的消息和节点本身的特征来更新节点的特征。

reduce function聚合函数：聚合节点接受到的消息

图一

图二
定义一个reduce 函数，需要指定一个输入消息名称和一个输出节点特征名称。

图三

图四

import dgl
import dgl.function as fn
import torch as th
g = ... # create a DGLGraph
g.ndata['h'] = th.randn((g.num_nodes(), 10)) # each node has feature size 10
g.edata['w'] = th.randn((g.num_edges(), 1))  # each edge has feature size 1
# collect features from source nodes and aggregate them in destination nodes
g.update_all(fn.copy_u('h', 'm'), fn.sum('m', 'h_sum'))
# multiply source node features with edge weights and aggregate them in destination nodes
g.update_all(fn.u_mul_e('h', 'w', 'm'), fn.max('m', 'h_max'))
# compute edge embedding by multiplying source and destination node embeddings
g.apply_edges(fn.u_mul_v('h', 'h', 'w_new'))

解释都放出来了，理解看个人了

---

下面是灰信网的DGL介绍，带注释+我的理解。

import dgl
import dgl.function as fn
import torch as th
import torch.nn as nn
import torch.nn.functional as F
from dgl import DGLGraph
 
gcn_msg=fn.copy_src(src="h",out="m") #使用源节点数据复制为消息
gcn_reduce=fn.sum(msg="m",out="h")#按总和聚合消息，聚合邻居节点的特征
 
#定义节点的UDF（update function ） apply_nodes  他是一个完全连接层
class NodeApplyModule(nn.Module):
    #初始化
    def __init__(self,in_feats,out_feats,activation):
        super(NodeApplyModule,self).__init__()
        self.linear=nn.Linear(in_feats,out_feats)
        self.activation=activation
    #前向传播
    def forward(self,node):
        h=self.linear(node.data["h"])
        if self.activation is not None:
            h=self.activation(h)
        return {"h":h}
 
#定义GCN模块  GCN模块的本质是在所有节点上执行消息传递  然后再调用NodeApplyModule全连接层
class GCN(nn.Module):
    #初始化
    def __init__(self,in_feats,out_feats,activation):
        super(GCN,self).__init__()
        #调用全连接层模块
        self.apply_mod=NodeApplyModule(in_feats,out_feats,activation)
    #前向传播
    def forward(self,g,feature):
        g.ndata["h"]=feature#feature应该对应的整个图的特征矩阵
        g.update_all(gcn_msg,gcn_reduce)
        g.apply_nodes(func=self.apply_mod)#将更新操作应用到节点上
        
        return g.ndata.pop("h")
		##pop() 函数用于移除列表中的一个元素（默认最后一个元素），并且返回该元素的值。

#利用cora数据集搭建网络然后训练
class Net(nn.Module):
    #初始化网络参数
    def __init__(self):
        super(Net,self).__init__()
        self.gcn1=GCN(1433,16,F.relu)#第一层GCN
        self.gcn2=GCN(16,7,None)
    #前向传播
    def forward(self,g,features):
        x=self.gcn1(g,features)
        x=self.gcn2(g,x)
        return x
net=Net()
net
 
#使用DGL内置模块加载cora数据集
from dgl.data import citation_graph as citegrh
import networkx as nx
def load_cora_data():
    data = citegrh.load_cora()#加载数据集
    features=th.FloatTensor(data.features)#特征向量  张量的形式
    labels=th.LongTensor(data.labels)#所属类别
    train_mask=th.BoolTensor(data.train_mask)#那些参与训练
    test_mask=th.BoolTensor(data.test_mask)#哪些是测试集
    g=data.graph
    g.remove_edges_from(nx.selfloop_edges(g))#删除自循环的边
    g = DGLGraph(g)
    g.add_edges(g.nodes(), g.nodes()) #自连接
    return g, features, labels, train_mask, test_mask
 
g, features, labels, train_mask, test_mask=load_cora_data()
 
 
import matplotlib.pyplot as plt
nx.draw(g.to_networkx(),node_size=50,with_labels=True)
plt.show()
 
#测试模型
def evaluate(model, g, features, labels, mask):
    model.eval()#会通知所有图层您处于评估模式
    with th.no_grad():
        logits = model(g, features)
        logits = logits[mask]
        labels = labels[mask]
        _, indices = th.max(logits, dim=1)
        correct = th.sum(indices == labels)
        return correct.item() * 1.0 / len(labels)
 
#训练网络
import time
import numpy as np
g, features, labels, train_mask, test_mask = load_cora_data()
 
#定义优化器
optimizer=th.optim.Adam(net.parameters(),lr=1e-3)
dur=[]#时间
for epoch in range(200):
    print(epoch)
    if epoch>=10:
        t0=time.time()
    net.train()
    logits = net(g, features)
    logp = F.log_softmax(logits, 1)
    loss = F.nll_loss(logp[train_mask], labels[train_mask])
 
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
 
    if epoch >=10:
        dur.append(time.time() - t0)
 
    acc = evaluate(net, g, features, labels, test_mask)
    print("Epoch {:05d} | Loss {:.4f} | Test Acc {:.4f} | Time(s) {:.4f}".format(
            epoch, loss.item(), acc, np.mean(dur)))
 

第一次会下载数据集，官方加载数据与上面不同。两个都可以跑。

from dgl.data import CoraGraphDataset
def load_cora_data():
    dataset = CoraGraphDataset()
    g = dataset[0]
    features = g.ndata['feat']
    labels = g.ndata['label']
    train_mask = g.ndata['train_mask']
    test_mask = g.ndata['test_mask']
    return g, features, labels, train_mask, test_mask
g, features, labels, train_mask, test_mask=load_cora_data()
print(features.shape)
print(labels.shape)
print(train_mask.shape)
print(test_mask.shape)

Downloading C:\Users\Administrator\.dgl\cora_v2.zip from https://data.dgl.ai/dataset/cora_v2.zip...
Extracting file to C:\Users\Administrator\.dgl\cora_v2
Finished data loading and preprocessing.
  NumNodes: 2708
  NumEdges: 10556
  NumFeats: 1433
  NumClasses: 7
  NumTrainingSamples: 140
  NumValidationSamples: 500
  NumTestSamples: 1000
Done saving data into cached files.

torch.Size([2708, 1433])
torch.Size([2708])
torch.Size([2708])
torch.Size([2708])

第二个不同就是GCN网络，区别：h = g.ndata[‘h’] 和 g.ndata.pop(“h”)

class GCNLayer(nn.Module):
    def __init__(self, in_feats, out_feats):
        super(GCNLayer, self).__init__()
        self.linear = nn.Linear(in_feats, out_feats)

    def forward(self, g, feature):
        # Creating a local scope so that all the stored ndata and edata
        # (such as the `'h'` ndata below) are automatically popped out
        # when the scope exits.
        with g.local_scope():
            g.ndata['h'] = feature
            g.update_all(gcn_msg, gcn_reduce)
            h = g.ndata['h']
            return self.linear(h)

其他解说

GNN教程：DGL框架实现GCN算法！—— Datawhale 讲的内容跟上面差不多

GCN 实现3 ：代码解析 介绍数据集还行，后面模型都是截图，谁会看

空手道俱乐部（Zachary’s karate club)数据集

以下3篇都是针对这个数据集，可以扎实一下。
【1】 https://www.jianshu.com/p/c0da16b75aa3
【2】https://blog.csdn.net/qq_38234785/article/details/107984924 讲到GCN的缺点
【3】 https://zhuanlan.zhihu.com/p/93828551 苘郁蓁

这个我也不了解

为了易于理解，整个教程忽略了归一化的步骤

由于版本问题：或者看博客的评论

DGLError: DGLGraph.send is deprecated. As a replacement, use DGLGraph.apply_edges
API to compute messages as edge data. Then use DGLGraph.send_and_recv
and set the message function as dgl.function.copy_e to conduct message
aggregation.

GCNLayer需要改写下，按新的来，否则 不能运行：
import dgl.function as fn
gcn_msg=fn.copy_src(src='h',out='m')
gcn_reduce=fn.sum(msg='m',out='h')

class GCNLayer(nn.Module):
    def __init__(self, in_feats, out_feats):
        super(GCNLayer, self).__init__()
        self.linear = nn.Linear(in_feats,out_feats)

    def forward(self,g,feature):
        with g.local_scope():
            g.ndata['h'] =feature
            g.update_all(gcn_msg,gcn_reduce)
            h=g.ndata['h']
        return self.linear(h)