Code Lab - 5

        这个Colab的后半部分内容掌握的不是很好，内容上有点断层，后面有时间回看一下再补充

pip install -q git+https://github.com/snap-stanford/deepsnap.git
pip install -U -q PyDrive

import torch
import torch_geometric

---

        DeepSNAP库可以用于复杂的图形操作，如特征计算、预训练、子图提取等，尤其可以很好地处理异构图

1. 使用NetworkX分配属性 

from pylab import *
import networkx as nx
from networkx.algorithms.community import greedy_modularity_communities
import matplotlib.pyplot as plt
import copy

G = nx.karate_club_graph()
community_map = {}
for node in G.nodes(data=True):  #node[0]是节点标号，node[1]是相关数据
  if node[1]["club"] == "Mr. Hi":
    community_map[node[0]] = 0
  else:
    community_map[node[0]] = 1
# 对应颜色
node_color = []
color_map = {0: 0, 1: 1}
node_color = [color_map[community_map[node]] for node in G.nodes()]
# 画图
pos = nx.spring_layout(G)
plt.figure(figsize=(7, 7))
nx.draw(G, pos=pos, cmap=plt.get_cmap('coolwarm'), node_color=node_color)
show()


# 注意图的node时，如果标明data=True则读取数据，否则只读取编号
for node in G.nodes(data=True):
    print(node)
    print(node[0])
    print(node[1])
    break
for node in G.nodes():
    print(node)
    break

1.1 分配node type和node feature

使用nx.set_node_attributes，为节点的属性赋值

'''
对于node_type，给“Mr.Hi”俱乐部中的节点分配节点类型n0，给俱乐部“Officer”中的节点分配节点类型n1
对于node_label，给“Mr.Hi”俱乐部中的节点分配node_label_0，给俱乐部“Officer”中的节点分配node_label_1。
为每个节点分配Tensor特征向量[1，1，1，1]
'''

import torch

def assign_node_types(G, community_map):
    NodeType={} # 节点编号：属性值
    for (key,value) in community_map.items():
    if value==0:
        NodeType[k]='n0'
    else:
        NodeType[k]='n1'
    nx.set_node_attributes(G,NodeType,'node_type')

def assign_node_labels(G, community_map):
    nx.set_node_attributes(G,community_map,'node_label')

def assign_node_features(G):
    feature_vector=[1, 1, 1, 1, 1]
    nx.set_node_attributes(G,feature_vector,'node_feature')

assign_node_types(G, community_map)
assign_node_labels(G, community_map)
assign_node_features(G)

1.2 分配edge type

使用nx.set_edge_attributes

# 分配edge type
'''
“Mr.Hi”俱乐部内的边：e0
“Officer”俱乐部内的边缘：e1
俱乐部之间的边：e2
'''

def assign_edge_types(G, community_map):
  edge2attr_map={}
  for edge in G.edges():
    if G.nodes[edge[0]]['club']=='Mr. Hi' and G.nodes[edge[1]]['club']=='Mr. Hi':
      edge2attr_map[edge]='e0'
    elif G.nodes[edge[0]]['club']=='Officer' and G.nodes[edge[1]]['club']=='Officer':
      edge2attr_map[edge]='e1'
    else:
      edge2attr_map[edge]='e2'
  nx.set_edge_attributes(G,edge2attr_map,'edge_type')

  #########################################
  
assign_edge_types(G, community_map)

1.3 异构图可视化

edge_color = {}
for edge in G.edges():
    n1, n2 = edge
    edge_color[edge] = community_map[n1] if community_map[n1] == community_map[n2] else 2
    if community_map[n1] == community_map[n2] and community_map[n1] == 0:
        edge_color[edge] = 'blue'
    elif community_map[n1] == community_map[n2] and community_map[n1] == 1:
        edge_color[edge] = 'red'
    else:
        edge_color[edge] = 'green'

G_orig = copy.deepcopy(G)
nx.classes.function.set_edge_attributes(G, edge_color, name='color')
colors = nx.get_edge_attributes(G,'color').values()
labels = nx.get_node_attributes(G, 'node_type')
plt.figure(figsize=(8, 8))
nx.draw(G, pos=pos, cmap=plt.get_cmap('coolwarm'), node_color=node_color, edge_color=colors, labels=labels, font_color='white')
show()

---

2. DeepSNAP操作

2.1 将NetworkX图转换为DeepSNAP格式

from deepsnap.hetero_graph import HeteroGraph
hete = HeteroGraph(G_orig)

2.2 每种节点/边有多少个

def get_nodes_per_type(hete):
    num_nodes_n0=len(hete.node_type['n0'])
    num_nodes_n1=len(hete.node_type['n1'])
    return num_nodes_n0, num_nodes_n1

num_nodes_n0, num_nodes_n1 = get_nodes_per_type(hete)
print("Node type n0 has {} nodes".format(num_nodes_n0))
print("Node type n1 has {} nodes".format(num_nodes_n1))


def get_num_message_edges(hete):
    message_type_edges = []
    for message_type,num_edge in hete.edge_type.items():
        message_type_edges.append((message_type,len(num_edge)))
    return message_type_edges

message_type_edges = get_num_message_edges(hete)
for (message_type, num_edges) in message_type_edges:
    print("Message type {} has {} edges".format(message_type, num_edges))

2.3 切分数据集 dataset.split 以及可视化

from deepsnap.dataset import GraphDataset

dataset = GraphDataset([hete], task='node')
# Splitting the dataset
dataset_train, dataset_val, dataset_test = dataset.split(transductive=True, split_ratio=[0.4, 0.3, 0.3])
datasets = {'train': dataset_train, 'val': dataset_val, 'test': dataset_test}


def compute_dataset_split_counts(datasets):
    data_set_splits = {}
    for ds_name,ds in datasets.items():
        data_set_splits[ds_name]=ds[0].node_label_index['n0'].shape[0]+ds[0].node_label_index['n1'].shape[0]
    return data_set_splits
data_set_splits = compute_dataset_split_counts(datasets)
for dataset_name, num_nodes in data_set_splits.items():
    print("{} dataset has {} nodes".format(dataset_name, num_nodes))
    
    
titles = ['Train', 'Validation', 'Test']
for i, dataset in enumerate([dataset_train, dataset_val, dataset_test]):
    n0 = hete._convert_to_graph_index(dataset[0].node_label_index['n0'], 'n0').tolist()
    n1 = hete._convert_to_graph_index(dataset[0].node_label_index['n1'], 'n1').tolist()

    plt.figure(figsize=(7, 7))
    plt.title(titles[i])
    nx.draw(G_orig, pos=pos, node_color="grey", edge_color=colors, labels=labels, font_color='white')
    nx.draw_networkx_nodes(G_orig.subgraph(n0), pos=pos, node_color="blue")
    nx.draw_networkx_nodes(G_orig.subgraph(n1), pos=pos, node_color="red")
    show()

---

3. 异构图节点属性预测

将消息类型视为（src，relation，dst），其中消息从src传递到dst，更新节点类型b依赖于不同的消息类型relation
消息传递依赖于不同的消息类型，每个消息传递层针对一种消息类型执行消息传递和聚合,计算给定消息类型的dst节点的嵌入

3.1 Heterogeneous GNN Layer 

import copy
import torch
import deepsnap
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import torch_geometric.nn as pyg_nn

from sklearn.metrics import f1_score
from deepsnap.hetero_gnn import forward_op
from deepsnap.hetero_graph import HeteroGraph
from torch_sparse import SparseTensor, matmul

class HeteroGNNConv(pyg_nn.MessagePassing):
    def __init__(self, in_channels_src, in_channels_dst, out_channels):
        super(HeteroGNNConv, self).__init__(aggr="mean")

        self.in_channels_src = in_channels_src
        self.in_channels_dst = in_channels_dst
        self.out_channels = out_channels

        self.lin_dst=nn.Linear(in_channels_dst,out_channels)  #concat左
        self.lin_src=nn.Linear(in_channels_src,out_channels)  #concat右
        self.lin_update=nn.Linear(out_channels*2,out_channels)  #()[]

    def forward(
        self,
        node_feature_src,
        node_feature_dst,
        edge_index,
        size=None,
        res_n_id=None,
    ):
        return self.propagate(edge_index,size=size,
                              node_feature_src=node_feature_src,node_feature_dst=node_feature_dst,res_n_id=res_n_id)

    def message_and_aggregate(self, edge_index, node_feature_src):
        out = matmul(edge_index,node_feature_src,reduce=self.aggr)  
        return out

    def update(self, aggr_out, node_feature_dst, res_n_id):
        aggr_out = self.lin_src(aggr_out)
        node_feature_dst = self.lin_dst(node_feature_dst)
        concat_features = torch.cat((node_feature_dst, aggr_out),dim=-1)
        aggr_out = self.lin_update(concat_features)
        return aggr_out

3.2 Heterogeneous GNN Wrapper Layer

class HeteroGNNWrapperConv(deepsnap.hetero_gnn.HeteroConv):
    def __init__(self, convs, args, aggr="mean"):
        super(HeteroGNNWrapperConv, self).__init__(convs, None)
        self.aggr = aggr
        # Map the index and message type
        self.mapping = {}
        # A numpy array that stores the final attention probability
        self.alpha = None
        self.attn_proj = None

        if self.aggr == "attn":
            self.attn_proj = nn.Sequential(
                nn.Linear(args['hidden_size'], args['attn_size']),  #Wh+b
                nn.Tanh(),
                nn.Linear(args['attn_size'], 1, bias=False), # q_semantic_attention
            )
    
    def reset_parameters(self):
        super(HeteroConvWrapper, self).reset_parameters()
        if self.aggr == "attn":
            for layer in self.attn_proj.children():
                layer.reset_parameters()
    
    def forward(self, node_features, edge_indices):
        message_type_emb = {}
        for message_key, message_type in edge_indices.items():
            src_type, edge_type, dst_type = message_key
            node_feature_src = node_features[src_type]
            node_feature_dst = node_features[dst_type]
            edge_index = edge_indices[message_key]
            message_type_emb[message_key] = (
                self.convs[message_key](
                    node_feature_src,
                    node_feature_dst,
                    edge_index,
                )
            )
        node_emb = {dst: [] for _, _, dst in message_type_emb.keys()}
        mapping = {}        
        for (src, edge_type, dst), item in message_type_emb.items():
            mapping[len(node_emb[dst])] = (src, edge_type, dst)
            node_emb[dst].append(item)
        self.mapping = mapping
        for node_type, embs in node_emb.items():
            if len(embs) == 1:
                node_emb[node_type] = embs[0]
            else:
                node_emb[node_type] = self.aggregate(embs)
        return node_emb
    
    def aggregate(self, xs):
        if self.aggr == "mean":
            out = torch.mean(torch.stack(xs), dim=0)
            return out

        elif self.aggr == "attn":
            N = xs[0].shape[0] # Number of nodes for that node type
            M = len(xs) # Number of message types for that node type
            x = torch.cat(xs, dim=0).view(M, N, -1) # M * N * D
            z = self.attn_proj(x).view(M, N) # M * N * 1
            z = z.mean(1) # M * 1
            alpha = torch.softmax(z, dim=0) # M * 1

            # Store the attention result to self.alpha as np array
            self.alpha = alpha.view(-1).data.cpu().numpy()
  
            alpha = alpha.view(M, 1, 1)
            x = x * alpha
            return x.sum(dim=0)

3.3 Initialize Heterogeneous GNN Layers

def generate_convs(hetero_graph, conv, hidden_size, first_layer=False):
    convs = {}
    for message_type in hetero_graph.message_types:
        if first_layer is True:
            src_type = message_type[0]
            dst_type = message_type[2]
            src_size = hetero_graph.num_node_features(src_type)
            dst_size = hetero_graph.num_node_features(dst_type)
            convs[message_type] = conv(src_size,dst_size, hidden_size)
        else:
            convs[message_type] = conv(hidden_size, hidden_size, hidden_size)    
    return convs

3.4 HeteroGNN

class HeteroGNN(torch.nn.Module):
    def __init__(self, hetero_graph, args, aggr="mean"):
        super(HeteroGNN, self).__init__()

        self.aggr = aggr
        self.hidden_size = args['hidden_size']

        self.convs1 = None
        self.convs2 = None

        self.bns1 = nn.ModuleDict()
        self.bns2 = nn.ModuleDict()
        self.relus1 = nn.ModuleDict()
        self.relus2 = nn.ModuleDict()
        self.post_mps = nn.ModuleDict()

        convs1 = generate_convs(hetero_graph, HeteroGNNConv, self.hidden_size, first_layer=True)
        convs2 = generate_convs(hetero_graph, HeteroGNNConv, self.hidden_size)

        self.convs1 = HeteroGNNWrapperConv(convs1, args, aggr=self.aggr)
        self.convs2 = HeteroGNNWrapperConv(convs2, args, aggr=self.aggr)

        for node_type in hetero_graph.node_types:
            self.bns1[node_type] = torch.nn.BatchNorm1d(self.hidden_size, eps=1)
            self.bns2[node_type] = torch.nn.BatchNorm1d(self.hidden_size, eps=1)
            self.post_mps[node_type] = nn.Linear(self.hidden_size, hetero_graph.num_node_labels(node_type))
            self.relus1[node_type] = nn.LeakyReLU()
            self.relus2[node_type] = nn.LeakyReLU()
            
    def forward(self, node_feature, edge_index):
        x = node_feature
        x = self.convs1(x, edge_index)
        x = forward_op(x, self.bns1)
        x = forward_op(x, self.relus1)
        x = self.convs2(x, edge_index)
        x = forward_op(x, self.bns2)
        x = forward_op(x, self.relus2)
        x = forward_op(x, self.post_mps)
        return x

    def loss(self, preds, y, indices):
        loss = 0
        loss_func = F.cross_entropy
        for node_type in preds:
            idx = indices[node_type]
            loss += loss_func(preds[node_type][idx], y[node_type][idx])
        return loss

3.5 Dataset

print("Device: {}".format(args['device']))

# Load the data
data = torch.load("acm.pkl")

# Message types
message_type_1 = ("paper", "author", "paper")
message_type_2 = ("paper", "subject", "paper")

# Dictionary of edge indices
edge_index = {}
edge_index[message_type_1] = data['pap']
edge_index[message_type_2] = data['psp']

# Dictionary of node features
node_feature = {}
node_feature["paper"] = data['feature']

# Dictionary of node labels
node_label = {}
node_label["paper"] = data['label']

# Load the train, validation and test indices
train_idx = {"paper": data['train_idx'].to(args['device'])}
val_idx = {"paper": data['val_idx'].to(args['device'])}
test_idx = {"paper": data['test_idx'].to(args['device'])}

# Construct a deepsnap tensor backend HeteroGraph
hetero_graph = HeteroGraph(
    node_feature=node_feature,
    node_label=node_label,
    edge_index=edge_index,
    directed=True
)

print(f"ACM heterogeneous graph: {hetero_graph.num_nodes()} nodes, {hetero_graph.num_edges()} edges")

# Node feature and node label to device
for key in hetero_graph.node_feature:
    hetero_graph.node_feature[key] = hetero_graph.node_feature[key].to(args['device'])
for key in hetero_graph.node_label:
    hetero_graph.node_label[key] = hetero_graph.node_label[key].to(args['device'])

# Edge_index to sparse tensor and to device
for key in hetero_graph.edge_index:
    edge_index = hetero_graph.edge_index[key]
    adj = SparseTensor(row=edge_index[0], col=edge_index[1], sparse_sizes=(hetero_graph.num_nodes('paper'), hetero_graph.num_nodes('paper')))
    hetero_graph.edge_index[key] = adj.t().to(args['device'])
print(hetero_graph.edge_index[message_type_1])
print(hetero_graph.edge_index[message_type_2])

3.6 Training

args = {
    'device': torch.device('cuda' if torch.cuda.is_available() else 'cpu'),
    'hidden_size': 64,
    'epochs': 100,
    'weight_decay': 1e-5,
    'lr': 0.003,
    'attn_size': 32,
}

def train(model, optimizer, hetero_graph, train_idx):
    model.train()
    optimizer.zero_grad()
    preds = model(hetero_graph.node_feature, hetero_graph.edge_index)
    loss = model.loss(preds, hetero_graph.node_label, train_idx)
    loss.backward()
    optimizer.step()
    return loss.item()

def test(model, graph, indices, best_model=None, best_val=0):
    model.eval()
    accs = []
    for index in indices:
        preds = model(graph.node_feature, graph.edge_index)
        num_node_types = 0
        micro = 0
        macro = 0
        for node_type in preds:
            idx = index[node_type]
            pred = preds[node_type][idx]
            pred = pred.max(1)[1]
            label_np = graph.node_label[node_type][idx].cpu().numpy()
            pred_np = pred.cpu().numpy()
            micro = f1_score(label_np, pred_np, average='micro')
            macro = f1_score(label_np, pred_np, average='macro')
            num_node_types += 1
        micro /= num_node_types
        macro /= num_node_types
        accs.append((micro, macro))
    if accs[1][0] > best_val:
        best_val = accs[1][0]
        best_model = copy.deepcopy(model)
    return accs, best_model, best_val

3.7 Training the Mean Aggregation / Attention Aggregation

# Training the Mean Aggregation
best_model = None
best_val = 0

model = HeteroGNN(hetero_graph, args, aggr="mean").to(args['device'])
optimizer = torch.optim.Adam(model.parameters(), lr=args['lr'], weight_decay=args['weight_decay'])

for epoch in range(args['epochs']):
    loss = train(model, optimizer, hetero_graph, train_idx)
    accs, best_model, best_val = test(model, hetero_graph, [train_idx, val_idx, test_idx], best_model, best_val)
    print(
        f"Epoch {epoch + 1}: loss {round(loss, 5)}, "
        f"train micro {round(accs[0][0] * 100, 2)}%, train macro {round(accs[0][1] * 100, 2)}%, "
        f"valid micro {round(accs[1][0] * 100, 2)}%, valid macro {round(accs[1][1] * 100, 2)}%, "
        f"test micro {round(accs[2][0] * 100, 2)}%, test macro {round(accs[2][1] * 100, 2)}%"
    )
best_accs, _, _ = test(best_model, hetero_graph, [train_idx, val_idx, test_idx])
print(
    f"Best model: "
    f"train micro {round(best_accs[0][0] * 100, 2)}%, train macro {round(best_accs[0][1] * 100, 2)}%, "
    f"valid micro {round(best_accs[1][0] * 100, 2)}%, valid macro {round(best_accs[1][1] * 100, 2)}%, "
    f"test micro {round(best_accs[2][0] * 100, 2)}%, test macro {round(best_accs[2][1] * 100, 2)}%"
)

# Training the Attention Aggregation
best_model = None
best_val = 0

output_size = hetero_graph.num_node_labels('paper')
model = HeteroGNN(hetero_graph, args, aggr="attn").to(args['device'])
optimizer = torch.optim.Adam(model.parameters(), lr=args['lr'], weight_decay=args['weight_decay'])

for epoch in range(args['epochs']):
    loss = train(model, optimizer, hetero_graph, train_idx)
    accs, best_model, best_val = test(model, hetero_graph, [train_idx, val_idx, test_idx], best_model, best_val)
    print(
        f"Epoch {epoch + 1}: loss {round(loss, 5)}, "
        f"train micro {round(accs[0][0] * 100, 2)}%, train macro {round(accs[0][1] * 100, 2)}%, "
        f"valid micro {round(accs[1][0] * 100, 2)}%, valid macro {round(accs[1][1] * 100, 2)}%, "
        f"test micro {round(accs[2][0] * 100, 2)}%, test macro {round(accs[2][1] * 100, 2)}%"
    )
best_accs, _, _ = test(best_model, hetero_graph, [train_idx, val_idx, test_idx])
print(
    f"Best model: "
    f"train micro {round(best_accs[0][0] * 100, 2)}%, train macro {round(best_accs[0][1] * 100, 2)}%, "
    f"valid micro {round(best_accs[1][0] * 100, 2)}%, valid macro {round(best_accs[1][1] * 100, 2)}%, "
    f"test micro {round(best_accs[2][0] * 100, 2)}%, test macro {round(best_accs[2][1] * 100, 2)}%"
)

# Attention for each Message Type
if model.convs1.alpha is not None and model.convs2.alpha is not None:
    for idx, message_type in model.convs1.mapping.items():
        print(f"Layer 1 has attention {model.convs1.alpha[idx]} on message type {message_type}")
    for idx, message_type in model.convs2.mapping.items():
        print(f"Layer 2 has attention {model.convs2.alpha[idx]} on message type {message_type}")