graphSAGE的python实现

上一节实现了基础的GCN：https://www.cnblogs.com/xiximayou/p/13550000.html

这一节我们继续实现graphSAGE。

加载数据：load_cora.py

import numpy as np
import scipy.sparse as sp
import torch
from sklearn.preprocessing import LabelBinarizer

def normalize_adj(adjacency):
  adjacency += sp.eye(adjacency.shape[0])
  degree = np.array(adjacency.sum(1))
  d_hat = sp.diags(np.power(degree, -0.5).flatten())
  return d_hat.dot(adjacency).dot(d_hat).tocoo()

def normalize_features(features):
  return features / features.sum(1)

def load_data(path="/content/drive/My Drive/nlpdata/cora/", dataset="cora"):
    """Load citation network dataset (cora only for now)"""
    print('Loading {} dataset...'.format(dataset))

    idx_features_labels = np.genfromtxt("{}{}.content".format(path,dataset), dtype=np.dtype(str))
    features = sp.csr_matrix(idx_features_labels[:, 1:-1], dtype=np.float32)
    encode_onehot = LabelBinarizer()
    labels = encode_onehot.fit_transform(idx_features_labels[:, -1])

    # build graph
    idx = np.array(idx_features_labels[:, 0], dtype=np.int32)
    idx_map = {j: i for i, j in enumerate(idx)}
    edges_unordered = np.genfromtxt("{}{}.cites".format(path, dataset), dtype=np.int32)
    edges = np.array(list(map(idx_map.get, edges_unordered.flatten())), dtype=np.int32).reshape(edges_unordered.shape)
    adj = sp.coo_matrix((np.ones(edges.shape[0]), (edges[:, 0], edges[:, 1])), shape=(labels.shape[0], labels.shape[0]), dtype=np.float32)


    features = normalize_features(features)
    adj = normalize_adj(adj)

    idx_train = range(140)
    idx_val = range(200, 500)
    idx_test = range(500, 1500)

    features = torch.FloatTensor(np.array(features))
    labels = torch.LongTensor(np.where(labels)[1])

    num_nodes = features.shape[0]
    train_mask = np.zeros(num_nodes, dtype=np.bool)
    val_mask = np.zeros(num_nodes, dtype=np.bool)
    test_mask = np.zeros(num_nodes, dtype=np.bool)

    train_mask[idx_train] = True
    val_mask[idx_val] = True
    test_mask[idx_test] = True

    return adj, features, labels, train_mask, val_mask, test_mask
"""
adj, features, labels, train_mask, val_mask, test_mask= load_data()
print(adj.shape)
print(features.shape)
print(labels.shape)
print(train_mask.shape, val_mask.shape, test_mask.shape)
"""

采样：sampling.py

import numpy as np


def sampling(src_nodes, sample_num, neighbor_table):
    """根据源节点采样指定数量的邻居节点，注意使用的是有放回的采样；
    某个节点的邻居节点数量少于采样数量时，采样结果出现重复的节点
    
    Arguments:
        src_nodes {list, ndarray} -- 源节点列表
        sample_num {int} -- 需要采样的节点数
        neighbor_table {dict} -- 节点到其邻居节点的映射表
    
    Returns:
        np.ndarray -- 采样结果构成的列表
    """
    results = []
    for sid in src_nodes:
        # 从节点的邻居中进行有放回地进行采样
        res = np.random.choice(neighbor_table[sid], size=(sample_num, ))
        results.append(res)
    return np.asarray(results).flatten()


def multihop_sampling(src_nodes, sample_nums, neighbor_table):
    """根据源节点进行多阶采样
    
    Arguments:
        src_nodes {list, np.ndarray} -- 源节点id
        sample_nums {list of int} -- 每一阶需要采样的个数
        neighbor_table {dict} -- 节点到其邻居节点的映射
    
    Returns:
        [list of ndarray] -- 每一阶采样的结果
    """
    sampling_result = [src_nodes]
    for k, hopk_num in enumerate(sample_nums):
        hopk_result = sampling(sampling_result[k], hopk_num, neighbor_table)
        sampling_result.append(hopk_result)
    return sampling_result

建立模型：grapgsage.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init


class NeighborAggregator(nn.Module):
    def __init__(self, input_dim, output_dim, 
                 use_bias=False, aggr_method="mean"):
        """聚合节点邻居
        Args:
            input_dim: 输入特征的维度
            output_dim: 输出特征的维度
            use_bias: 是否使用偏置 (default: {False})
            aggr_method: 邻居聚合方式 (default: {mean})
        """
        super(NeighborAggregator, self).__init__()
        self.input_dim = input_dim
        self.output_dim = output_dim
        self.use_bias = use_bias
        self.aggr_method = aggr_method
        self.weight = nn.Parameter(torch.Tensor(input_dim, output_dim))
        if self.use_bias:
            self.bias = nn.Parameter(torch.Tensor(self.output_dim))
        self.reset_parameters()
    
    def reset_parameters(self):
        init.kaiming_uniform_(self.weight)
        if self.use_bias:
            init.zeros_(self.bias)

    def forward(self, neighbor_feature):
        if self.aggr_method == "mean":
            aggr_neighbor = neighbor_feature.mean(dim=1)
        elif self.aggr_method == "sum":
            aggr_neighbor = neighbor_feature.sum(dim=1)
        elif self.aggr_method == "max":
            aggr_neighbor = neighbor_feature.max(dim=1)
        else:
            raise ValueError("Unknown aggr type, expected sum, max, or mean, but got {}"
                             .format(self.aggr_method))
        
        neighbor_hidden = torch.matmul(aggr_neighbor, self.weight)
        if self.use_bias:
            neighbor_hidden += self.bias

        return neighbor_hidden

    def extra_repr(self):
        return 'in_features={}, out_features={}, aggr_method={}'.format(
            self.input_dim, self.output_dim, self.aggr_method)
    

class SageGCN(nn.Module):
    def __init__(self, input_dim, hidden_dim,
                 activation=F.relu,
                 aggr_neighbor_method="mean",
                 aggr_hidden_method="sum"):
        """SageGCN层定义
        Args:
            input_dim: 输入特征的维度
            hidden_dim: 隐层特征的维度，
                当aggr_hidden_method=sum, 输出维度为hidden_dim
                当aggr_hidden_method=concat, 输出维度为hidden_dim*2
            activation: 激活函数
            aggr_neighbor_method: 邻居特征聚合方法，["mean", "sum", "max"]
            aggr_hidden_method: 节点特征的更新方法，["sum", "concat"]
        """
        super(SageGCN, self).__init__()
        assert aggr_neighbor_method in ["mean", "sum", "max"]
        assert aggr_hidden_method in ["sum", "concat"]
        self.input_dim = input_dim
        self.hidden_dim = hidden_dim
        self.aggr_neighbor_method = aggr_neighbor_method
        self.aggr_hidden_method = aggr_hidden_method
        self.activation = activation
        self.aggregator = NeighborAggregator(input_dim, hidden_dim,
                                             aggr_method=aggr_neighbor_method)
        self.weight = nn.Parameter(torch.Tensor(input_dim, hidden_dim))
        self.reset_parameters()
    
    def reset_parameters(self):
        init.kaiming_uniform_(self.weight)

    def forward(self, src_node_features, neighbor_node_features):
        neighbor_hidden = self.aggregator(neighbor_node_features)
        self_hidden = torch.matmul(src_node_features, self.weight)
        
        if self.aggr_hidden_method == "sum":
            hidden = self_hidden + neighbor_hidden
        elif self.aggr_hidden_method == "concat":
            hidden = torch.cat([self_hidden, neighbor_hidden], dim=1)
        else:
            raise ValueError("Expected sum or concat, got {}"
                             .format(self.aggr_hidden))
        if self.activation:
            return self.activation(hidden)
        else:
            return hidden

    def extra_repr(self):
        output_dim = self.hidden_dim if self.aggr_hidden_method == "sum" else self.hidden_dim * 2
        return 'in_features={}, out_features={}, aggr_hidden_method={}'.format(
            self.input_dim, output_dim, self.aggr_hidden_method)


class GraphSage(nn.Module):
    def __init__(self, input_dim, hidden_dim,
                 num_neighbors_list):
        super(GraphSage, self).__init__()
        self.input_dim = input_dim
        self.hidden_dim = hidden_dim
        self.num_neighbors_list = num_neighbors_list
        self.num_layers = len(num_neighbors_list)
        self.gcn = nn.ModuleList()
        self.gcn.append(SageGCN(input_dim, hidden_dim[0]))
        for index in range(0, len(hidden_dim) - 2):
            self.gcn.append(SageGCN(hidden_dim[index], hidden_dim[index+1]))
        self.gcn.append(SageGCN(hidden_dim[-2], hidden_dim[-1], activation=None))

    def forward(self, node_features_list):
        hidden = node_features_list
        for l in range(self.num_layers):
            next_hidden = []
            gcn = self.gcn[l]
            for hop in range(self.num_layers - l):
                src_node_features = hidden[hop]
                src_node_num = len(src_node_features)
                neighbor_node_features = hidden[hop + 1] 
                    .view((src_node_num, self.num_neighbors_list[hop], -1))
                h = gcn(src_node_features, neighbor_node_features)
                next_hidden.append(h)
            hidden = next_hidden
        return hidden[0]

    def extra_repr(self):
        return 'in_features={}, num_neighbors_list={}'.format(
            self.input_dim, self.num_neighbors_list
        )
  

主函数：main.py

import torch

import numpy as np
import torch.nn as nn
import torch.optim as optim
from graphsage import GraphSage
from sampling import multihop_sampling
from load_cora import load_data
import pickle
import sys
sys.path.append("/content/drive/My Drive/nlpdata/cora/")

INPUT_DIM = 1433    # 输入维度
# Note: 采样的邻居阶数需要与GCN的层数保持一致
HIDDEN_DIM = [128, 7]   # 隐藏单元节点数
NUM_NEIGHBORS_LIST = [10, 10]   # 每阶采样邻居的节点数
assert len(HIDDEN_DIM) == len(NUM_NEIGHBORS_LIST)
BTACH_SIZE = 16     # 批处理大小
EPOCHS = 100
NUM_BATCH_PER_EPOCH = 20    # 每个epoch循环的批次数
LEARNING_RATE = 0.01    # 学习率
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"



adjacency, x, y, train_mask, val_mask, test_mask = load_data()
out = pickle.load(open("/content/drive/My Drive/nlpdata/cora/ind.cora.graph", "rb"), encoding="latin1")
graph = out.toarray() if hasattr(out, "toarray") else out
adjacency_dict = graph

train_index = np.where(train_mask)[0]
train_label = y[train_index]
test_index = np.where(test_mask)[0]
model = GraphSage(input_dim=INPUT_DIM, hidden_dim=HIDDEN_DIM,
                  num_neighbors_list=NUM_NEIGHBORS_LIST).to(DEVICE)
print(model)
criterion = nn.CrossEntropyLoss().to(DEVICE)
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE, weight_decay=5e-4)


def train():
    model.train()
    for e in range(EPOCHS):
        for batch in range(NUM_BATCH_PER_EPOCH):
            batch_src_index = np.random.choice(train_index, size=(BTACH_SIZE,))
            batch_src_label = train_label[batch_src_index].long().to(DEVICE)
            batch_sampling_result = multihop_sampling(batch_src_index, NUM_NEIGHBORS_LIST, adjacency_dict)
            batch_sampling_x = [x[idx].to(DEVICE) for idx in batch_sampling_result]
            batch_train_logits = model(batch_sampling_x)
            loss = criterion(batch_train_logits, batch_src_label)
            optimizer.zero_grad()
            loss.backward()  # 反向传播计算参数的梯度
            optimizer.step()  # 使用优化方法进行梯度更新
            #print("Epoch {:03d} Batch {:03d} Loss: {:.4f}".format(e, batch, loss.item()))
        test()


def test():
    model.eval()
    with torch.no_grad():
        test_sampling_result = multihop_sampling(test_index, NUM_NEIGHBORS_LIST, adjacency_dict)
        test_x = [x[idx].to(DEVICE) for idx in test_sampling_result]
        test_logits = model(test_x)
        test_label = y[test_index].long().to(DEVICE)
        predict_y = test_logits.max(1)[1]
        accuarcy = torch.eq(predict_y, test_label).float().mean().item()
        print("Test Accuracy: ", accuarcy)


if __name__ == '__main__':
    train()

运行之后：

Loading cora dataset...
GraphSage(
  in_features=1433, num_neighbors_list=[10, 10]
  (gcn): ModuleList(
    (0): SageGCN(
      in_features=1433, out_features=128, aggr_hidden_method=sum
      (aggregator): NeighborAggregator(in_features=1433, out_features=128, aggr_method=mean)
    )
    (1): SageGCN(
      in_features=128, out_features=7, aggr_hidden_method=sum
      (aggregator): NeighborAggregator(in_features=128, out_features=7, aggr_method=mean)
    )
  )
)
Epoch 000 Batch 000 Loss: 1.9266
Epoch 000 Batch 001 Loss: 1.8985
Epoch 000 Batch 002 Loss: 1.8120
Epoch 000 Batch 003 Loss: 1.6504
Epoch 000 Batch 004 Loss: 1.7163
Epoch 000 Batch 005 Loss: 1.6697
Epoch 000 Batch 006 Loss: 1.4684
Epoch 000 Batch 007 Loss: 1.2102
Epoch 000 Batch 008 Loss: 1.2884
Epoch 000 Batch 009 Loss: 1.0167
Epoch 000 Batch 010 Loss: 0.9699
Epoch 000 Batch 011 Loss: 0.9754
Epoch 000 Batch 012 Loss: 0.5973
Epoch 000 Batch 013 Loss: 0.8455
Epoch 000 Batch 014 Loss: 0.9120
Epoch 000 Batch 015 Loss: 0.6430
Epoch 000 Batch 016 Loss: 0.7662
Epoch 000 Batch 017 Loss: 0.8074
Epoch 000 Batch 018 Loss: 0.5895
Epoch 000 Batch 019 Loss: 0.4272
Test Accuracy:  0.3270000219345093......
Epoch 098 Batch 000 Loss: 0.0403
Epoch 098 Batch 001 Loss: 0.1180
Epoch 098 Batch 002 Loss: 0.0855
Epoch 098 Batch 003 Loss: 0.0344
Epoch 098 Batch 004 Loss: 0.0437
Epoch 098 Batch 005 Loss: 0.0531
Epoch 098 Batch 006 Loss: 0.0513
Epoch 098 Batch 007 Loss: 0.0787
Epoch 098 Batch 008 Loss: 0.0396
Epoch 098 Batch 009 Loss: 0.0373
Epoch 098 Batch 010 Loss: 0.0398
Epoch 098 Batch 011 Loss: 0.0337
Epoch 098 Batch 012 Loss: 0.0427
Epoch 098 Batch 013 Loss: 0.0315
Epoch 098 Batch 014 Loss: 0.0720
Epoch 098 Batch 015 Loss: 0.0827
Epoch 098 Batch 016 Loss: 0.1221
Epoch 098 Batch 017 Loss: 0.0374
Epoch 098 Batch 018 Loss: 0.0427
Epoch 098 Batch 019 Loss: 0.0373
Test Accuracy:  0.5540000200271606
Epoch 099 Batch 000 Loss: 0.0577
Epoch 099 Batch 001 Loss: 0.0373
Epoch 099 Batch 002 Loss: 0.0462
Epoch 099 Batch 003 Loss: 0.0511
Epoch 099 Batch 004 Loss: 0.0849
Epoch 099 Batch 005 Loss: 0.0571
Epoch 099 Batch 006 Loss: 0.0478
Epoch 099 Batch 007 Loss: 0.0598
Epoch 099 Batch 008 Loss: 0.0376
Epoch 099 Batch 009 Loss: 0.0414
Epoch 099 Batch 010 Loss: 0.0478
Epoch 099 Batch 011 Loss: 0.0321
Epoch 099 Batch 012 Loss: 0.1014
Epoch 099 Batch 013 Loss: 0.0617
Epoch 099 Batch 014 Loss: 0.0529
Epoch 099 Batch 015 Loss: 0.0325
Epoch 099 Batch 016 Loss: 0.0334
Epoch 099 Batch 017 Loss: 0.0432
Epoch 099 Batch 018 Loss: 0.0939
Epoch 099 Batch 019 Loss: 0.0517
Test Accuracy:  0.5260000228881836

参考：

https://github.com/FighterLYL/GraphNeuralNetwork