-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
224 lines (209 loc) · 13.1 KB
/
main.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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @Time : 2021/9/14 4:28
# @Author : ZM7
# @File : main
# @Software: PyCharm
import torch
import os
from TKG.utils import myFloder, Collate, Logger, mkdir_if_not_exist
from rgcn import utils
from torch.utils.data import DataLoader
from TKG.load_data import load_data
from hgls import HGLS
import argparse
import time
import yaml
from yaml import SafeLoader
import datetime
import numpy as np
import dgl
import sys
from sys import exit
from TKG.utils_new import myFloder_new, collate_new
from rgcn.knowledge_graph import _read_triplets_as_list
from rgcn.utils import build_sub_graph
import warnings
import random
warnings.filterwarnings('ignore')
def set_seed(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed) # cpu
torch.cuda.manual_seed(seed) # gpu
torch.cuda.manual_seed_all(seed) # all gpus
# os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'
# torch.use_deterministic_algorithms(True)
set_seed(2018)
def inplace_relu(m):
classname = m.__class__.__name__
if classname.find('ReLU') != -1:
m.inplace=True
def test(model, all_list,total_data, test_dataset, all_ans_list_test, node_id_new, s_t, test_sid, static_graph=None):
ranks_raw, ranks_filter, mrr_raw_list, mrr_filter_list = [], [], [], []
test_losses = []
model.eval()
for test_data_list in test_dataset:
with torch.no_grad():
final_score, final_score_r, test_loss = \
model(all_list, total_data, test_data_list, node_id_new[:, test_data_list['t'][0]].to(device),
(test_data_list['t'][0] - s_t[:, test_data_list['t'][0]]).to(device), device=device, mode='test', static_graph=static_graph)
mrr_filter_snap, mrr_snap, rank_raw, rank_filter = utils.get_total_rank(test_data_list['triple'].to(device),
final_score,
all_ans_list_test[
test_data_list['t'][
0] - test_sid],
eval_bz=1000, rel_predict=0)
ranks_raw.append(rank_raw)
ranks_filter.append(rank_filter)
test_losses.append(test_loss.item())
mrr_raw, h1_raw, h3_raw, h10_raw = utils.stat_ranks(ranks_raw)
mrr_filter, h1_f, h3_f, h10_f = utils.stat_ranks(ranks_filter)
return np.mean(test_losses), [mrr_raw, h1_raw, h3_raw, h10_raw], [mrr_filter, h1_f, h3_f, h10_f]
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='HGLS')
parser.add_argument("--gpu", default='0', help="gpu")
parser.add_argument("-d", "--dataset", type=str, required=True,
help="dataset to use")
parser.add_argument("--n-layers", type=int, default=2,
help="number of propagation rounds")
parser.add_argument("--self-loop", action='store_true', default=True,
help="perform layer normalization in every layer of gcn ")
parser.add_argument("--layer-norm", action='store_true', default=False,
help="perform layer normalization in every layer of gcn ")
parser.add_argument("--relation-prediction", action='store_true', default=False,
help="add relation prediction loss")
parser.add_argument("--entity-prediction", action='store_true', default=False,
help="add entity prediction loss")
# configuration for stat training
parser.add_argument("--n-epochs", type=int, default=16,
help="number of minimum training epochs on each time step")
parser.add_argument("--lr", type=float, default=0.001,
help="learning rate")
parser.add_argument("--grad-norm", type=float, default=1.0,
help="norm to clip gradient to")
parser.add_argument("--n-hidden", type=int, default=200,
help="number of hidden units")
parser.add_argument('--k_hop', type=int, default=2, help='k_hop')
parser.add_argument("--task", type=float, default=0.7, help="weight of entity prediction task")
parser.add_argument("--short", action='store_true', default=False, help="short-term")
parser.add_argument("--long", action='store_true', default=False, help="long-term")
parser.add_argument('--gnn', default='regcn')
parser.add_argument('--fuse', default='con', help='entity fusion')
parser.add_argument('--r_fuse', default='re', help='relation fusion')
# parser.add_argument("--r_p", action='store_true', default=True, help="tkg") # 关系预测
parser.add_argument("--record", action='store_true', default=False, help="save log file")
parser.add_argument("--model_record", action='store_true', default=False, help="save model file")
# configuration for optimal parameters
parser.add_argument('--config', type=str, default='long_config.yaml')
args = parser.parse_args().__dict__ # REGCN 的参数
short_con = yaml.load(open('short_config.yaml'), Loader=SafeLoader)[args['dataset']]
long_con = yaml.load(open('long_config.yaml'), Loader=SafeLoader)[args['dataset']]
num_nodes, num_rels, train_list, valid_list, test_list, total_data, all_ans_list_test, all_ans_list_r_test, \
all_ans_list_valid, all_ans_list_r_valid, graph, node_id_new, s_t, s_f, s_l, train_sid, valid_sid, test_sid, \
total_times, time_idx = load_data(args['dataset'])
#选择环境
device = torch.device('cuda:0')
os.environ["CUDA_VISIBLE_DEVICES"] = args['gpu']
# regcn的参数补充
print(short_con["use_static"], '-------------------------------')
if short_con["use_static"]:
static_triples = np.array(_read_triplets_as_list("../bertintkg/data/" + args["dataset"] + "/e-w-graph.txt", {}, {}, load_time=False))
num_static_rels = len(np.unique(static_triples[:, 1]))
num_words = len(np.unique(static_triples[:, 2]))
static_triples[:, 2] = static_triples[:, 2] + num_nodes
static_node_id = torch.from_numpy(np.arange(num_words + num_nodes)).view(-1, 1).long().to(device)
static_graph = build_sub_graph(len(static_node_id), num_static_rels, static_triples, device)
else:
num_static_rels, num_words, static_triples, static_graph = 0, 0, [], None
short_con["num_words"] = num_words
short_con["num_static_rels"] = num_static_rels
# short_con["static_graph"] = static_graph
# HGLS的参数补充
long_con['time_length'] = len(total_data)
long_con['time_idx'] = time_idx
print(args)
print(short_con)
print(long_con)
if args['dataset'] in ['ICEWS05-15', 'ICEWS18', 'GDELT']:
print('load data from folder')
train_path = 'data/' + '_' + args['dataset'] + '/train/'
valid_path = 'data/' + '_' + args['dataset'] + '/val/'
test_path = 'data/' + '_' + args['dataset'] + '/test/'
train_set = myFloder_new(train_path, dgl.load_graphs)
val_set = myFloder_new(valid_path, dgl.load_graphs)
test_set = myFloder_new(test_path, dgl.load_graphs)
train_dataset = DataLoader(dataset=train_set, batch_size=1, collate_fn=collate_new, shuffle=True, pin_memory=True, num_workers=16)
val_dataset = DataLoader(dataset=val_set, batch_size=1, collate_fn=collate_new, shuffle=False, pin_memory=True, num_workers=3)
test_dataset = DataLoader(dataset=test_set, batch_size=1, collate_fn=collate_new, shuffle=False, pin_memory=True, num_workers=3)
else:
print('load data online')
train_set = myFloder(train_list, max_batch=100, start_id=train_sid, no_batch=True, mode='train')
val_set = myFloder(valid_list, max_batch=100, start_id=valid_sid, no_batch=True, mode='test')
test_set = myFloder(test_list, max_batch=100, start_id=test_sid, no_batch=True, mode='test')
co = Collate(num_nodes, num_rels, s_f, s_t, len(total_data), args['dataset'], long_con['encoder'], long_con['decoder'], max_length=long_con['max_length'], all=False, graph=graph, k=2)
train_dataset = DataLoader(dataset=train_set, batch_size=1, collate_fn=co.collate_rel, shuffle=True, pin_memory=True, num_workers=16)
val_dataset = DataLoader(dataset=val_set, batch_size=1, collate_fn=co.collate_rel, shuffle=False, pin_memory=True, num_workers=16)
test_dataset = DataLoader(dataset=test_set, batch_size=1, collate_fn=co.collate_rel, shuffle=False, pin_memory=True, num_workers=16)
seq_len_lis = short_con["sequence_len_lis"]
del short_con["sequence_len_lis"]
short_model = short_con["short_model"]
pe_dim_lis = short_con["pe_dim_lis"]
long_pe_dim_lis = long_con["pe_dim_lis"]
del short_con["pe_dim_lis"]
del long_con["pe_dim_lis"]
short_con["pe_dim"] = pe_dim_lis[0]
for pe_dim in long_pe_dim_lis:
print("pe_dim==>", pe_dim)
long_con["pe_dim"] = pe_dim
for seq_len in seq_len_lis:
short_con["sequence_len"] = seq_len
log_file = f'{args["dataset"]}_short_{args["short"]}_short-model_{short_model}_long_{args["long"]}_' \
f'f_{args["fuse"]}_fr_{args["r_fuse"]}_ta_{args["task"]}' \
f'_gnn1_{long_con["encoder"]}_{long_con["a_layer_num"]}_gnn2_{long_con["decoder"]}_{long_con["d_layer_num"]}' \
f'_seq_{short_con["sequence"]}_{short_con["sequence_len"]}_max_length_{long_con["max_length"]}_fil_{long_con["filter"]}_ori_{long_con["ori"]}' \
f'last_{long_con["last"]}'
if args['record']:
log_file_path = f'results/g_{args["gpu"]}_' + log_file
mkdir_if_not_exist(log_file_path)
sys.stdout = Logger(log_file_path)
print(f'Logging to {log_file_path}')
model = HGLS(graph.to(device), num_nodes, num_rels, args['n_hidden'], args['task'], args['relation_prediction'],
args['short'], args['long'], args['fuse'], args['r_fuse'], short_con, long_con, short_model=short_model).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=args['lr'], weight_decay=1e-5)
model.apply(inplace_relu)
# compute global pe
# print(len(total_data), total_data[0].shape, type(total_data[0]))
st = time.time()
all_glist = build_sub_graph(num_nodes, num_rels, np.concatenate(total_data, axis=0), device, pe_init="rw", pe_dim=20)
# print(time.time()-st)
for epoch in range(args['n_epochs']):
print('Epoch {}'.format(epoch), '_', 'Start training: ', datetime.datetime.now(),
'=============================================')
model.train()
stop = True
losses = [0]
loss_es = [0]
loss_rs = []
for train_data_list in train_dataset:
loss_e, loss_r, loss = model(all_glist,total_data, train_data_list, node_id_new[:, train_data_list['t'][0]].to(device),
(train_data_list['t'][0] - s_t[:, train_data_list['t'][0]]).to(device), device=device, mode='train', static_graph=static_graph)
losses.append(loss.item())
loss_es.append(loss_e.item())
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args['grad_norm']) # clip gradients
optimizer.step()
optimizer.zero_grad()
print('Epoch {}, loss {:.4f}'.format(epoch, np.mean(losses)), datetime.datetime.now())
print('\tStart validating: ', datetime.datetime.now())
val_result = test(model, all_glist, total_data, val_dataset, all_ans_list_valid, node_id_new, s_t, valid_sid, static_graph=static_graph)
print('\ttrain_loss:%.4f\tval_loss:%.4f\tval_Mrr_raw:%.4f\tval_Hits(raw)@1:%.4f\tval_Hits(raw)@3:%.4f\tval_Hits(raw)@10:%.4f'
'\tval_Mrr_filter:%.4f\tval_Hits(filter)@1:%.4f\tval_Hits(filter)@3:%.4f\tval_Hits(filter)@10:%.4f' %
(np.mean(losses), val_result[0], val_result[1][0], val_result[1][1], val_result[1][2], val_result[1][3],
val_result[2][0], val_result[2][1], val_result[2][2], val_result[2][3]))
print('\tStart testing: ', datetime.datetime.now())
test_result = test(model, all_glist, total_data, test_dataset, all_ans_list_test, node_id_new, s_t, test_sid, static_graph=static_graph)
print('\tval_loss:%.4f\tval_Mrr_raw:%.4f\tval_Hits(raw)@1:%.4f\tval_Hits(raw)@3:%.4f\tval_Hits(raw)@10:%.4f'
'\tval_Mrr_filter:%.4f\tval_Hits(filter)@1:%.4f\tval_Hits(filter)@3:%.4f\tval_Hits(filter)@10:%.4f' %
(test_result[0], test_result[1][0], test_result[1][1], test_result[1][2], test_result[1][3],
test_result[2][0], test_result[2][1], test_result[2][2], test_result[2][3]))