-
Notifications
You must be signed in to change notification settings - Fork 101
/
Copy pathtransH_Bernoulli_pytorch.py
380 lines (318 loc) · 15.6 KB
/
transH_Bernoulli_pytorch.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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @Date : 2018-01-07 01:00:31
# @Author : jimmy ([email protected])
# @Link : http://sdcs.sysu.edu.cn
# @Version : $Id$
import os
import torch
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
import time
import datetime
import random
from utils import *
from data import *
from evaluation import *
import loss
import model
from hyperboard import Agent
USE_CUDA = torch.cuda.is_available()
"""
The meaning of parameters:
self.dataset: Which dataset is used to train the model? Such as 'FB15k', 'WN18', etc.
self.learning_rate: Initial learning rate (lr) of the model.
self.early_stopping_round: How many times will lr decrease? If set to 0, it remains constant.
self.L1_flag: If set to True, use L1 distance as dissimilarity; else, use L2.
self.embedding_size: The embedding size of entities and relations.
self.num_batches: How many batches to train in one epoch?
self.train_times: The maximum number of epochs for training.
self.margin: The margin set for MarginLoss.
self.filter: Whether to check a generated negative sample is false negative.
self.momentum: The momentum of the optimizer.
self.optimizer: Which optimizer to use? Such as SGD, Adam, etc.
self.loss_function: Which loss function to use? Typically, we use margin loss.
self.entity_total: The number of different entities.
self.relation_total: The number of different relations.
self.batch_size: How many instances is contained in one batch?
"""
class Config(object):
def __init__(self):
self.dataset = None
self.learning_rate = 0.001
self.early_stopping_round = 0
self.L1_flag = True
self.embedding_size = 100
self.num_batches = 100
self.train_times = 1000
self.margin = 1.0
self.filter = True
self.momentum = 0.9
self.optimizer = optim.Adam
self.loss_function = loss.marginLoss
self.entity_total = 0
self.relation_total = 0
self.batch_size = 0
if __name__ == "__main__":
import argparse
argparser = argparse.ArgumentParser()
"""
The meaning of some parameters:
seed: Fix the random seed. Except for 0, which means no setting of random seed.
port: The port number used by hyperboard,
which is a demo showing training curves in real time.
You can refer to https://github.com/WarBean/hyperboard to know more.
num_processes: Number of processes used to evaluate the result.
"""
argparser.add_argument('-d', '--dataset', type=str)
argparser.add_argument('-l', '--learning_rate', type=float, default=0.001)
argparser.add_argument('-es', '--early_stopping_round', type=int, default=0)
argparser.add_argument('-L', '--L1_flag', type=int, default=1)
argparser.add_argument('-em', '--embedding_size', type=int, default=100)
argparser.add_argument('-nb', '--num_batches', type=int, default=100)
argparser.add_argument('-n', '--train_times', type=int, default=1000)
argparser.add_argument('-m', '--margin', type=float, default=1.0)
argparser.add_argument('-f', '--filter', type=int, default=1)
argparser.add_argument('-mo', '--momentum', type=float, default=0.9)
argparser.add_argument('-s', '--seed', type=int, default=0)
argparser.add_argument('-op', '--optimizer', type=int, default=1)
argparser.add_argument('-lo', '--loss_type', type=int, default=0)
argparser.add_argument('-p', '--port', type=int, default=5000)
argparser.add_argument('-np', '--num_processes', type=int, default=4)
args = argparser.parse_args()
# Start the hyperboard agent
agent = Agent(address='127.0.0.1', port=args.port)
if args.seed != 0:
torch.manual_seed(args.seed)
trainTotal, trainList, trainDict = loadTriple('./data/' + args.dataset, 'train2id.txt')
validTotal, validList, validDict = loadTriple('./data/' + args.dataset, 'valid2id.txt')
tripleTotal, tripleList, tripleDict = loadTriple('./data/' + args.dataset, 'triple2id.txt')
with open(os.path.join('./data/', args.dataset, 'head_tail_proportion.pkl'), 'rb') as fr:
tail_per_head = pickle.load(fr)
head_per_tail = pickle.load(fr)
config = Config()
config.dataset = args.dataset
config.learning_rate = args.learning_rate
config.early_stopping_round = args.early_stopping_round
if args.L1_flag == 1:
config.L1_flag = True
else:
config.L1_flag = False
config.embedding_size = args.embedding_size
config.num_batches = args.num_batches
config.train_times = args.train_times
config.margin = args.margin
if args.filter == 1:
config.filter = True
else:
config.filter = False
config.momentum = args.momentum
if args.optimizer == 0:
config.optimizer = optim.SGD
elif args.optimizer == 1:
config.optimizer = optim.Adam
elif args.optimizer == 2:
config.optimizer = optim.RMSprop
if args.loss_type == 0:
config.loss_function = loss.marginLoss
config.entity_total = getAnythingTotal('./data/' + config.dataset, 'entity2id.txt')
config.relation_total = getAnythingTotal('./data/' + config.dataset, 'relation2id.txt')
config.batch_size = trainTotal // config.num_batches
shareHyperparameters = {'dataset': args.dataset,
'learning_rate': args.learning_rate,
'early_stopping_round': args.early_stopping_round,
'L1_flag': args.L1_flag,
'embedding_size': args.embedding_size,
'margin': args.margin,
'filter': args.filter,
'momentum': args.momentum,
'seed': args.seed,
'optimizer': args.optimizer,
'loss_type': args.loss_type,
}
trainHyperparameters = shareHyperparameters.copy()
trainHyperparameters.update({'type': 'train_loss'})
validHyperparameters = shareHyperparameters.copy()
validHyperparameters.update({'type': 'valid_loss'})
hit10Hyperparameters = shareHyperparameters.copy()
hit10Hyperparameters.update({'type': 'hit10'})
meanrankHyperparameters = shareHyperparameters.copy()
meanrankHyperparameters.update({'type': 'mean_rank'})
trainCurve = agent.register(trainHyperparameters, 'train loss', overwrite=True)
validCurve = agent.register(validHyperparameters, 'valid loss', overwrite=True)
hit10Curve = agent.register(hit10Hyperparameters, 'hit@10', overwrite=True)
meanrankCurve = agent.register(meanrankHyperparameters, 'mean rank', overwrite=True)
loss_function = config.loss_function()
model = model.TransHModel(config)
if USE_CUDA:
model.cuda()
loss_function.cuda()
longTensor = torch.cuda.LongTensor
floatTensor = torch.cuda.FloatTensor
else:
longTensor = torch.LongTensor
floatTensor = torch.FloatTensor
optimizer = config.optimizer(model.parameters(), lr=config.learning_rate)
margin = autograd.Variable(floatTensor([config.margin]))
start_time = time.time()
filename = '_'.join(
['l', str(args.learning_rate),
'es', str(args.early_stopping_round),
'L', str(args.L1_flag),
'em', str(args.embedding_size),
'nb', str(args.num_batches),
'n', str(args.train_times),
'm', str(args.margin),
'f', str(args.filter),
'mo', str(args.momentum),
's', str(args.seed),
'op', str(args.optimizer),
'lo', str(args.loss_type),]) + '_TransH_Bernoulli.ckpt'
trainBatchList = getBatchList(trainList, config.num_batches)
for epoch in range(config.train_times):
total_loss = floatTensor([0.0])
random.shuffle(trainBatchList)
for batchList in trainBatchList:
if config.filter == True:
pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch = getBatch_filter_all_v2(batchList,
config.entity_total, tripleDict, tail_per_head, head_per_tail)
else:
pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch = getBatch_raw_all_v2(batchList,
config.entity_total, tail_per_head, head_per_tail)
batch_entity_set = set(pos_h_batch + pos_t_batch + neg_h_batch + neg_t_batch)
batch_relation_set = set(pos_r_batch + neg_r_batch)
batch_entity_list = list(batch_entity_set)
batch_relation_list = list(batch_relation_set)
pos_h_batch = autograd.Variable(longTensor(pos_h_batch))
pos_t_batch = autograd.Variable(longTensor(pos_t_batch))
pos_r_batch = autograd.Variable(longTensor(pos_r_batch))
neg_h_batch = autograd.Variable(longTensor(neg_h_batch))
neg_t_batch = autograd.Variable(longTensor(neg_t_batch))
neg_r_batch = autograd.Variable(longTensor(neg_r_batch))
model.zero_grad()
pos, neg = model(pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch)
if args.loss_type == 0:
losses = loss_function(pos, neg, margin)
else:
losses = loss_function(pos, neg)
ent_embeddings = model.ent_embeddings(torch.cat([pos_h_batch, pos_t_batch, neg_h_batch, neg_t_batch]))
rel_embeddings = model.rel_embeddings(torch.cat([pos_r_batch, neg_r_batch]))
norm_embeddings = model.norm_embeddings(torch.cat([pos_r_batch, neg_r_batch]))
losses += loss.orthogonalLoss(rel_embeddings, norm_embeddings)
losses = losses + loss.normLoss(ent_embeddings) + loss.normLoss(rel_embeddings)
losses.backward()
optimizer.step()
total_loss += losses.data
normalize_norm_emb = F.normalize(model.norm_embeddings.weight.data[longTensor(batch_relation_list)], p=2, dim=1)
model.norm_embeddings.weight.data[longTensor(batch_relation_list)] = normalize_norm_emb
agent.append(trainCurve, epoch, total_loss[0])
if epoch % 10 == 0:
now_time = time.time()
print(now_time - start_time)
print("Train total loss: %d %f" % (epoch, total_loss[0]))
if epoch % 10 == 0:
if config.filter == True:
pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch = getBatch_filter_random_v2(validList,
config.batch_size, config.entity_total, tripleDict, tail_per_head, head_per_tail)
else:
pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch = getBatch_raw_random_v2(validList,
config.batch_size, config.entity_total, tail_per_head, head_per_tail)
pos_h_batch = autograd.Variable(longTensor(pos_h_batch))
pos_t_batch = autograd.Variable(longTensor(pos_t_batch))
pos_r_batch = autograd.Variable(longTensor(pos_r_batch))
neg_h_batch = autograd.Variable(longTensor(neg_h_batch))
neg_t_batch = autograd.Variable(longTensor(neg_t_batch))
neg_r_batch = autograd.Variable(longTensor(neg_r_batch))
pos, neg = model(pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch)
if args.loss_type == 0:
losses = loss_function(pos, neg, margin)
else:
losses = loss_function(pos, neg)
ent_embeddings = model.ent_embeddings(torch.cat([pos_h_batch, pos_t_batch, neg_h_batch, neg_t_batch]))
rel_embeddings = model.rel_embeddings(torch.cat([pos_r_batch, neg_r_batch]))
norm_embeddings = model.norm_embeddings(torch.cat([pos_r_batch, neg_r_batch]))
losses += loss.orthogonalLoss(rel_embeddings, norm_embeddings)
losses = losses + loss.normLoss(ent_embeddings) + loss.normLoss(rel_embeddings)
print("Valid batch loss: %d %f" % (epoch, losses.data[0]))
agent.append(validCurve, epoch, losses.data[0])
if config.early_stopping_round > 0:
if epoch == 0:
ent_embeddings = model.ent_embeddings.weight.data
rel_embeddings = model.rel_embeddings.weight.data
norm_embeddings = model.norm_embeddings.weight.data
L1_flag = model.L1_flag
filter = model.filter
hit10, best_meanrank = evaluation_transH(validList, tripleDict, ent_embeddings, rel_embeddings,
norm_embeddings, L1_flag, filter, k=config.batch_size, num_processes=args.num_processes)
agent.append(hit10Curve, epoch, hit10)
agent.append(meanrankCurve, epoch, best_meanrank)
torch.save(model, os.path.join('./model/' + args.dataset, filename))
best_epoch = 0
meanrank_not_decrease_time = 0
lr_decrease_time = 0
#if USE_CUDA:
#model.cuda()
# Evaluate on validation set for every 5 epochs
elif epoch % 5 == 0:
ent_embeddings = model.ent_embeddings.weight.data
rel_embeddings = model.rel_embeddings.weight.data
norm_embeddings = model.norm_embeddings.weight.data
L1_flag = model.L1_flag
filter = model.filter
hit10, now_meanrank = evaluation_transH(validList, tripleDict, ent_embeddings, rel_embeddings,
norm_embeddings, L1_flag, filter, k=config.batch_size, num_processes=args.num_processes)
agent.append(hit10Curve, epoch, hit10)
agent.append(meanrankCurve, epoch, now_meanrank)
if now_meanrank < best_meanrank:
meanrank_not_decrease_time = 0
best_meanrank = now_meanrank
torch.save(model, os.path.join('./model/' + args.dataset, filename))
else:
meanrank_not_decrease_time += 1
# If the result hasn't improved for consecutive 5 evaluations, decrease learning rate
if meanrank_not_decrease_time == 5:
lr_decrease_time += 1
if lr_decrease_time == config.early_stopping_round:
break
else:
optimizer.param_groups[0]['lr'] *= 0.5
meanrank_not_decrease_time = 0
#if USE_CUDA:
#model.cuda()
elif (epoch + 1) % 10 == 0 or epoch == 0:
torch.save(model, os.path.join('./model/' + args.dataset, filename))
testTotal, testList, testDict = loadTriple('./data/' + args.dataset, 'test2id.txt')
oneToOneTotal, oneToOneList, oneToOneDict = loadTriple('./data/' + args.dataset, 'one_to_one.txt')
oneToManyTotal, oneToManyList, oneToManyDict = loadTriple('./data/' + args.dataset, 'one_to_many.txt')
manyToOneTotal, manyToOneList, manyToOneDict = loadTriple('./data/' + args.dataset, 'many_to_one.txt')
manyToManyTotal, manyToManyList, manyToManyDict = loadTriple('./data/' + args.dataset, 'many_to_many.txt')
ent_embeddings = model.ent_embeddings.weight.data
rel_embeddings = model.rel_embeddings.weight.data
norm_embeddings = model.norm_embeddings.weight.data
L1_flag = model.L1_flag
filter = model.filter
hit10Test, meanrankTest = evaluation_transH(testList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=0)
hit10OneToOneHead, meanrankOneToOneHead = evaluation_transH(oneToOneList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=1)
hit10OneToManyHead, meanrankOneToManyHead = evaluation_transH(oneToManyList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=1)
hit10ManyToOneHead, meanrankManyToOneHead = evaluation_transH(manyToOneList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=1)
hit10ManyToManyHead, meanrankManyToManyHead = evaluation_transH(manyToManyList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=1)
hit10OneToOneTail, meanrankOneToOneTail = evaluation_transH(oneToOneList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=2)
hit10OneToManyTail, meanrankOneToManyTail = evaluation_transH(oneToManyList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=2)
hit10ManyToOneTail, meanrankManyToOneTail = evaluation_transH(manyToOneList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=2)
hit10ManyToManyTail, meanrankManyToManyTail = evaluation_transH(manyToManyList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=2)
writeList = [filename,
'testSet', '%.6f' % hit10Test, '%.6f' % meanrankTest,
'one_to_one_head', '%.6f' % hit10OneToOneHead, '%.6f' % meanrankOneToOneHead,
'one_to_many_head', '%.6f' % hit10OneToManyHead, '%.6f' % meanrankOneToManyHead,
'many_to_one_head', '%.6f' % hit10ManyToOneHead, '%.6f' % meanrankManyToOneHead,
'many_to_many_head', '%.6f' % hit10ManyToManyHead, '%.6f' % meanrankManyToManyHead,
'one_to_one_tail', '%.6f' % hit10OneToOneTail, '%.6f' % meanrankOneToOneTail,
'one_to_many_tail', '%.6f' % hit10OneToManyTail, '%.6f' % meanrankOneToManyTail,
'many_to_one_tail', '%.6f' % hit10ManyToOneTail, '%.6f' % meanrankManyToOneTail,
'many_to_many_tail', '%.6f' % hit10ManyToManyTail, '%.6f' % meanrankManyToManyTail,]
# Write the result into file
with open(os.path.join('./result/', args.dataset + '.txt'), 'a') as fw:
fw.write('\t'.join(writeList) + '\n')