-
Notifications
You must be signed in to change notification settings - Fork 1
/
RF-GNN.py
182 lines (151 loc) · 7.49 KB
/
RF-GNN.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "5"
import torch
import torch.nn as nn
from RFGNNmodels import RGCN, GAT, GCN, SAGE, SGC
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
from Dataset import cresci15, Twibot20, MGTAB
from utils import sample_mask, init_weights
import numpy as np
import argparse
import time
import json
from config import Config
parser = argparse.ArgumentParser()
parser.add_argument("-dataset", type=str, help="dataset", choices=['Twibot20','MGTAB','Cresci15'])
parser.add_argument('-model', type=str, help="model", choices=['GCN', 'GAT', 'GraphSage', 'RGCN', 'SGC'])
parser.add_argument('--labelrate', type=float, default=0.1, help='labelrate')
args = parser.parse_args()
print(args)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
config_file = "./config/" + str(args.dataset) + ".ini"
config = Config(config_file)
if args.dataset == 'Twibot20':
dataset = Twibot20('Data/Twibot20')
elif args.dataset == 'MGTAB':
dataset = MGTAB('Data/MGTAB')
elif args.dataset == 'Cresci15':
dataset = cresci15('Data/Cresci15')
data = dataset[0]
if args.dataset == 'MGTAB':
data.y = data.y2
out_dim = 2
data = data.to(device)
sample_number = len(data.y)
index_select_list = (data.edge_type == 100)
relation_dict = {
0:'followers',
1:'friends'
}
relation_select_list = json.loads(config.relation_select)
relation_num = len(relation_select_list)
print('relation used:', end=' ')
for features_index in relation_select_list:
index_select_list = index_select_list + (features_index == data.edge_type)
print('{}'.format(relation_dict[features_index]), end=' ')
edge_index = data.edge_index[:, index_select_list]
edge_type = data.edge_type[index_select_list]
def main(seed):
np.random.seed(seed)
torch.manual_seed(seed)
node_id = np.arange(data.num_nodes)
np.random.shuffle(node_id)
data.n_id = torch.arange(data.num_nodes)
data.train_id = node_id[:int(data.num_nodes * args.labelrate)]
data.val_id = node_id[int(data.num_nodes * 0.1):int(data.num_nodes * 0.2)]
data.test_id = node_id[int(data.num_nodes * 0.2):]
data.train_mask = sample_mask(data.train_id, sample_number)
data.val_mask = sample_mask(data.val_id, sample_number)
data.test_mask = sample_mask(data.test_id, sample_number)
test_mask = data.test_mask
train_mask = data.train_mask
val_mask = data.val_mask
fdim = data.x.shape[1]
embedding_size = fdim
sample_size = int(sum(data.train_mask)*config.beta)
edge_count = int(edge_index.shape[1]*config.gamma)
results = torch.zeros(data.x.shape[0], out_dim).to(device)
for num in range(1, config.model_num+1):
np.random.seed(seed + num)
sub_findex = np.random.randint(0, data.x.shape[1], size = embedding_size)
sub_fmask = sub_findex > config.alpha*data.x.shape[1]
sub_sindex = np.random.randint(0, sum(data.train_mask), sample_size)
sub_eindex = np.random.randint(0, edge_index.shape[1], edge_count)
print('traning {}th model'.format(num))
if args.model == 'RGCN':
model = RGCN(embedding_size, config.hidden_dimension, out_dim, relation_num, config.dropout).to(device)
elif args.model == 'GCN':
model = GCN(embedding_size, config.hidden_dimension, out_dim, relation_num, config.dropout).to(device)
elif args.model == 'GAT':
model = GAT(embedding_size, config.hidden_dimension, out_dim, relation_num, config.dropout).to(device)
elif args.model == 'SAGE':
model = SAGE(embedding_size, config.hidden_dimension, out_dim, relation_num, config.dropout).to(device)
elif args.model == 'SGC':
model = SGC(embedding_size, config.hidden_dimension, out_dim, relation_num, config.dropout).to(device)
loss = nn.CrossEntropyLoss()
optimizer = torch.optim.AdamW(model.parameters(),
lr=config.lr, weight_decay=config.weight_decay)
model.apply(init_weights)
fdata = torch.zeros(data.x.shape).cuda()
fdata = fdata + data.x
fdata[:, sub_fmask] = 0
max_val_acc = 0
for epoch in range(config.epochs):
model.train()
output = model(fdata, data.x, edge_index[:, sub_eindex], edge_type[sub_eindex])
loss_train = loss(output[data.train_mask][sub_sindex], data.y[data.train_mask][sub_sindex])
out = output.max(1)[1].to('cpu').detach().numpy()
label = data.y.to('cpu').detach().numpy()
acc_train = accuracy_score(out[train_mask], label[train_mask])
acc_val = accuracy_score(out[val_mask], label[val_mask])
if acc_val > max_val_acc:
max_val_acc = acc_val
best_output = output
optimizer.zero_grad()
loss_train.backward(retain_graph=True)
optimizer.step()
if (epoch + 1)%100 == 0:
print('Epoch: {:04d}'.format(epoch + 1),
'loss_train: {:.4f}'.format(loss_train.item()),
'acc_train: {:.4f}'.format(acc_train.item()),
'acc_val: {:.4f}'.format(acc_val.item()))
label = data.y.to('cpu').detach().numpy()
out = best_output.max(1)[1].to('cpu').detach().numpy()
acc_test = accuracy_score(out[test_mask], label[test_mask])
f1 = f1_score(out[test_mask], label[test_mask], average='macro')
precision = precision_score(out[test_mask], label[test_mask], average='macro')
recall = recall_score(out[test_mask], label[test_mask], average='macro')
print('acc_test {:.4f}'.format(acc_test),
'f1_test: {:.4f}'.format(f1.item()),
'precision_test: {:.4f}'.format(precision.item()),
'recall_test: {:.4f}'.format(recall.item()))
results = results + best_output
results_out = results.max(1)[1].to('cpu').detach().numpy()
acc_test = accuracy_score(results_out[test_mask], label[test_mask])
f1 = f1_score(results_out[test_mask], label[test_mask], average='macro')
precision = precision_score(results_out[test_mask], label[test_mask], average='macro')
recall = recall_score(results_out[test_mask], label[test_mask], average='macro')
return acc_test, precision, recall, f1
if __name__ == "__main__":
t = time.time()
acc_list = []
precision_list = []
recall_list = []
f1_list = []
for i, seed in enumerate(json.loads(config.random_seed)):
print('traning {}th round'.format(i + 1))
acc, precision, recall, f1 = main(seed)
acc_list.append(acc * 100)
precision_list.append(precision * 100)
recall_list.append(recall * 100)
f1_list.append(f1 * 100)
print('Round:{:04d}'.format(i + 1),
'acc_test {:.4f}'.format(acc),
'f1_test: {:.4f}'.format(f1),
'precision_test: {:.4f}'.format(precision),
'recall_test: {:.4f}'.format(recall))
print('acc: {:.2f} + {:.2f}'.format(np.array(acc_list).mean(), np.std(acc_list)))
print('precision: {:.2f} + {:.2f}'.format(np.array(precision_list).mean(), np.std(precision_list)))
print('recall: {:.2f} + {:.2f}'.format(np.array(recall_list).mean(), np.std(recall_list)))
print('f1: {:.2f} + {:.2f}'.format(np.array(f1_list).mean(), np.std(f1_list)))
print('total time:', time.time() - t)