/
pretrain_self.py
200 lines (164 loc) · 7.09 KB
/
pretrain_self.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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
# coding=UTF-8
import os
import torch
import torch.nn as nn
import numpy as np
import random
import argparse
from core.model import GBNEncoder, LNClassifier
from core.dataloader import load_graph
from core.sim_metric import EDSim, SDPSim, CosSim
from core.util import get_optimizer
from core.loss import edge_mask_loss, contrastive_loss
SIM_TABLE = {
'ed': EDSim(),
'sdp': SDPSim(),
'cos': CosSim()
}
def regist_parser():
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', type=str, default='dataset')
parser.add_argument('--output_model_file', type=str, default='')
parser.add_argument('--input_model_file', type=str, default='')
parser.add_argument('--sim_metric', type=str, default='cos',
help='sim metric function,choose from list ['
'"ed"(euclidean distance),"sdp"(scaled dot product)'
',"cos"(cosine)]')
parser.add_argument('--k_hop', type=int, default=2)
parser.add_argument('--nl_weight', type=float, default=0.1)
parser.add_argument('--local', action='store_true')
parser.add_argument('--n_layer', type=int, default=3)
parser.add_argument('--dropout', type=float, default=0.1)
parser.add_argument('--bias', action='store_true')
parser.add_argument('--device', type=int, default=7)
parser.add_argument('--cpu', action='store_true', help='Ignore CUDA.')
parser.add_argument('--seed', type=int, default=1)
parser.add_argument('--n_epoch', type=int, default=100)
parser.add_argument('--optimizer', default='adam', help='Optimizer')
parser.add_argument('--lr', type=float, default=1e-3, help='Learning rate')
parser.add_argument('--decay', type=float, default=1e-3,
help='Weight decay for optimization')
parser.add_argument('--feature_type', default='glove')
parser.add_argument('--feature_dim', type=int, default=50)
parser.add_argument('--edge_feature_dim', type=int, default=5)
args = parser.parse_args()
return args
def print_settings(opt):
print('==========================Parameters==============================')
print('Dataset Dir:\t\t', opt['data_dir'])
print('Input Model Path:\t', opt['input_model_file'])
print('Output Model Path:\t', opt['output_model_file'])
print('Similarity Metric:\t', opt['sim_metric'])
print('K Hop Neighbor:\t\t', opt['k_hop'])
print('Is localized:\t\t', opt['local'])
print('GNN Layer:\t\t', opt['n_layer'])
print('Dropout:\t\t', opt['dropout'])
print('Linear Bias:\t\t', opt['bias'])
print('Random seed:\t\t', opt['seed'])
print('Learning Epoch:\t\t', opt['n_epoch'])
print('Optimizer:\t\t', opt['optimizer'])
print('Learning Rate:\t\t', opt['lr'])
print('Weight Decay:\t\t', opt['decay'])
print('Feature Type:\t\t', opt['feature_type'])
print('Feature Dim:\t\t', opt['feature_dim'])
print('Edge Feature Dim:\t', opt['edge_feature_dim'])
print('==================================================================')
def comprise_data(opt, encoder, weight):
print('loading %s......' % opt['dataset'])
pkl_path = 'graph_' + opt['feature_type'] + '.pkl'
graph_data, graph = load_graph(opt, pkl_path, pyg=True)
opt['n_class'] = len(graph.node_s.itol)
graph_data = graph_data.to(opt['device'])
graph_data.x = (graph_data.x[0].to(opt['device']),
graph_data.x[1].to(opt['device']))
d_es = graph_data.x[0].size(-1)
classifier = LNClassifier(d_es * 2, 1)
classifier.to(opt['device'])
parameters = [
{'params': [p for p in encoder.parameters() if p.requires_grad]},
{'params': [p for p in classifier.parameters() if p.requires_grad]}]
optimizer = get_optimizer(opt['optimizer'], parameters,
opt['lr'], opt['decay'])
n_epoch = opt['n_epoch'] * weight
print('loaded!')
return classifier, optimizer, graph_data, weight
def neighbor_learning_pretrain(opt, encoder, data):
sim_metric = opt['sim_metric']
es, ps = encoder(data)
loss = opt['nl_weight'] * contrastive_loss((es, ps), data, sim_metric)
return loss
def link_mask_pretrain(opt, encoder, classifier, data):
x = data.x
edge_attr = data.edge_attr
edge_index = data.edge_index
edge_size = edge_index.size(1)
neg_size = int(edge_size * 0.1)
indices = torch.randperm(edge_size, device=edge_index.device)
masked_indices = indices[:neg_size]
remain_indices = indices[neg_size:]
output = encoder(x, edge_index[:, remain_indices],
edge_attr[remain_indices])
loss = edge_mask_loss(output, data, masked_indices, classifier)
return loss
def edge_mask(opt, encoder, batch, batch_id, ite):
classifier, optimizer, data, weight = batch
total_loss = 0
for i in range(weight):
encoder.train()
optimizer.zero_grad()
loss_nl = neighbor_learning_pretrain(opt, encoder, data)
loss_lm = link_mask_pretrain(opt, encoder, classifier, data)
loss = (loss_nl + loss_lm) / weight
loss.backward()
optimizer.step()
total_loss += loss.item()
print('Ite[%d]-Batch[%d]--loss:%.5f' % (ite, batch_id, total_loss, ))
if __name__ == '__main__':
args = regist_parser()
if args.seed:
os.environ['PYTHONHASHSEED'] = str(args.seed)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# cuda device setting
if args.cpu and args.device is not None or not torch.cuda.is_available():
args.device = torch.device('cpu')
else:
args.device = torch.device('cuda:%d' % args.device)
if args.seed:
torch.cuda.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
opt = vars(args)
if opt['local']:
opt['k_hop'] = 1
print_settings(opt)
if opt['feature_type'] == 'bert':
opt['feature_dim'] = 768
device = opt['device']
opt['sim_metric'] = SIM_TABLE[opt['sim_metric'].lower()].to(device)
encoder = GBNEncoder(opt)
if opt['input_model_file']:
encoder.load_state_dict(torch.load(opt['input_model_file']+'.pth'))
encoder = encoder.to(opt['device'])
datasets = []
with open(os.path.join(opt['data_dir'], 'unsupervised_dataset.txt'), 'r') as f:
for line in f:
line = line.strip().split('\t')
assert line and line[0]
dataset = line[0]
if len(line) > 1:
weight = int(line[1])
else:
weight = 1
datasets.append((dataset, weight))
print('=================Do Pre-Train (Self-Supervised)===================')
batches = []
for dataset, weight in datasets:
opt['dataset'] = os.path.join(opt['data_dir'], dataset)
batches.append(comprise_data(opt, encoder, weight))
for ite in range(1, opt['n_epoch'] + 1):
for i, batch in enumerate(batches):
edge_mask(opt, encoder, batch, i+1, ite)
if opt['output_model_file']:
print('write model file', opt['output_model_file'], '...')
torch.save(encoder.cpu().state_dict(), opt['output_model_file']+'.pth')