forked from DEEP-PolyU/KEQA_WSDM19
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_pred.py
238 lines (217 loc) · 10.7 KB
/
train_pred.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
import torch
import torch.nn as nn
import time
import os
import numpy as np
import random
from torchtext import data
from argparse import ArgumentParser
from embedding import EmbedVector
from sklearn.metrics.pairwise import euclidean_distances
parser = ArgumentParser(description="Training predicate vector learning")
parser.add_argument('--qa_mode', type=str, required=True, help='options are GRU, LSTM')
parser.add_argument('--embed_dim', type=int, default=250)
parser.add_argument('--no_cuda', action='store_false', help='do not use cuda', dest='cuda')
parser.add_argument('--gpu', type=int, default=0) # Use -1 for CPU
parser.add_argument('--epochs', type=int, default=30)
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--lr', type=float, default=0.0003)
parser.add_argument('--seed', type=int, default=3435)
parser.add_argument('--dev_every', type=int, default=10000)
parser.add_argument('--log_every', type=int, default=2000)
parser.add_argument('--patience', type=int, default=12)
parser.add_argument('--best_prefix', type=str, default='pred')
parser.add_argument('--output_channel', type=int, default=300)
parser.add_argument('--num_layer', type=int, default=2)
parser.add_argument('--rnn_fc_dropout', type=float, default=0.3)
parser.add_argument('--hidden_size', type=int, default=300)
parser.add_argument('--rnn_dropout', type=float, default=0.3)
parser.add_argument('--clip_gradient', type=float, default=0.6, help='gradient clipping')
parser.add_argument('--vector_cache', type=str, default="data/sq_glove300d.pt")
parser.add_argument('--weight_decay',type=float, default=0)
parser.add_argument('--fix_embed', action='store_false', dest='train_embed')
parser.add_argument('--output', type=str, default='preprocess')
args = parser.parse_args()
################## Prepare training and validation datasets ##################
pre_dic = {} # Dictionary for predicates
for line in open(os.path.join(args.output, 'relation2id.txt'), 'r'):
items = line.strip().split("\t")
pre_dic[items[0]] = int(items[1])
# Embedding for predicates
predicates_emb = torch.from_numpy(np.fromfile(os.path.join(args.output, 'predicates_emb.bin'), dtype=np.float32).reshape((len(pre_dic), args.embed_dim)))
# Set up the data for training
for filename in ['train.txt', 'valid.txt']:
outfile = open(os.path.join(args.output, 'pred_' + filename), 'w')
for line in open(os.path.join(args.output, filename), 'r'):
items = line.strip().split("\t")
if items[3] in pre_dic:
outfile.write("{}\t{}\n".format(items[5], pre_dic[items[3]]))
# pred_list.append(entities_emb[mid_dic[items[4]], :] - entities_emb[mid_dic[items[1]], :])
# token = list(compress(items[5].split(), [element == 'O' for element in items[6].split()]))
# if not token:
outfile.close()
synthetic_flag = True
if synthetic_flag:
names_map = {}
for i, line in enumerate(open(os.path.join(args.output, 'names.trimmed.txt'), 'r')):
items = line.strip().split("\t")
if len(items) != 2:
print("ERROR: line - {}".format(line))
continue
entity = items[0]
literal = items[1].strip()
if literal != "" and (names_map.get(entity) is None or len(names_map[entity].split()) > len(literal.split())):
names_map[entity] = literal
seen_fact = []
for line in open(os.path.join(args.output, 'train.txt'), 'r'):
items = line.strip().split("\t")
names_map[items[1]] = items[2]
seen_fact.append((items[1], items[3]))
seen_fact = set(seen_fact)
whereset = {'location', 'place', 'geographic', 'region', 'places'}
whoset = {'composer', 'people', 'artist', 'author', 'publisher', 'directed', 'developer', 'director', 'lyricist',
'edited', 'parents', 'instrumentalists', 'produced', 'manufacturer', 'written', 'designers', 'producer'}
outfile = open(os.path.join(args.output, 'pred_train.txt'), 'a')
for line in open(os.path.join(args.output, 'transE_valid.txt'), 'r'):
items = line.strip().split("\t")
if (items[0], items[2]) not in seen_fact and names_map.get(items[0]) is not None:
name = names_map[items[0]]
tokens = items[2].replace('.', ' ').replace('_', ' ').split()
seen = set()
clean_token = [token for token in tokens if not (token in seen or seen.add(token))]
question = 'what is the ' + ' '.join(clean_token) + ' of ' + name
for token in clean_token:
if token in whereset:
question = 'where is ' + ' '.join(clean_token) + ' of ' + name
break
elif token in whoset:
question = 'who is the ' + ' '.join(clean_token) + ' of ' + name
break
outfile.write("{}\t{}\n".format(question, pre_dic[items[2]]))
outfile.close()
del names_map, pre_dic
################## Set random seed for reproducibility ##################
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
torch.backends.cudnn.deterministic = True
if not args.cuda:
args.gpu = -1
if torch.cuda.is_available() and args.cuda:
print("Note: You are using GPU for training")
torch.cuda.set_device(args.gpu)
torch.cuda.manual_seed(args.seed)
if torch.cuda.is_available() and not args.cuda:
print("Warning: You have Cuda but not use it. You are using CPU for training.")
# Dictionary and embedding for words
if os.path.isfile(args.vector_cache):
stoi, vectors, words_dim = torch.load(args.vector_cache)
else:
print("Error: Need word embedding pt file")
exit(1)
################## Load the datasets ##################
TEXT = data.Field(lower=True)
ED = data.Field(sequential=False, use_vocab=False)
train, dev = data.TabularDataset.splits(path=args.output, train='pred_train.txt', validation='pred_valid.txt', format='tsv', fields=[('text', TEXT), ('mid', ED)])
field = [('id', None), ('sub', None), ('entity', None), ('relation', None), ('obj', None), ('text', TEXT), ('ed', None)]
test = data.TabularDataset(path=os.path.join(args.output, 'test.txt'), format='tsv', fields=field)
TEXT.build_vocab(train, dev, test)
match_embedding = 0
TEXT.vocab.vectors = torch.Tensor(len(TEXT.vocab), words_dim)
for i, token in enumerate(TEXT.vocab.itos):
wv_index = stoi.get(token, None)
if wv_index is not None:
TEXT.vocab.vectors[i] = vectors[wv_index]
match_embedding += 1
else:
TEXT.vocab.vectors[i] = torch.FloatTensor(words_dim).uniform_(-0.25, 0.25)
print("Word embedding match number {} out of {}".format(match_embedding, len(TEXT.vocab)))
del stoi, vectors
if args.cuda:
train_iter = data.Iterator(train, batch_size=args.batch_size, device=torch.device('cuda', args.gpu), train=True,
repeat=False, sort=False, shuffle=True, sort_within_batch=False)
dev_iter = data.Iterator(dev, batch_size=args.batch_size, device=torch.device('cuda', args.gpu), train=False,
repeat=False, sort=False, shuffle=False, sort_within_batch=False)
else:
train_iter = data.Iterator(train, batch_size=args.batch_size, train=True, repeat=False, sort=False, shuffle=True,
sort_within_batch=False)
dev_iter = data.Iterator(dev, batch_size=args.batch_size, train=False, repeat=False, sort=False, shuffle=False,
sort_within_batch=False)
config = args
config.words_num = len(TEXT.vocab)
config.label = args.embed_dim
config.words_dim = words_dim
model = EmbedVector(config)
model.embed.weight.data.copy_(TEXT.vocab.vectors)
if args.cuda:
modle = model.to(torch.device("cuda:{}".format(args.gpu)))
print("Shift model to GPU")
# Embedding for MID
predicates_emb = predicates_emb.cuda()
total_num = len(dev)
print(config)
print("VOCAB num",len(TEXT.vocab))
print("Train instance", len(train))
print("Dev instance", total_num)
print(model)
parameter = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adam(parameter, lr=args.lr, weight_decay=args.weight_decay)
criterion = nn.MSELoss()
early_stop = False
best_accu = 0
best_loss = total_num
iterations = 0
iters_not_improved = 0
num_dev_in_epoch = (len(train) // args.batch_size // args.dev_every) + 1
patience = args.patience * num_dev_in_epoch # for early stopping
epoch = 0
start = time.time()
print(' Time Epoch Iteration Progress (%Epoch) Loss')
log_template = ' '.join('{:>6.0f},{:>5.0f},{:>9.0f},{:>5.0f}/{:<5.0f} {:>7.0f}%,{:>8.6f}'.split(','))
while True:
if early_stop:
print("Early Stopping. Epoch: {}, Best Dev accuracy: {}, loss: {},".format(epoch, best_accu, best_loss))
break
epoch += 1
train_iter.init_epoch()
for batch_idx, batch in enumerate(train_iter):
# Batch size : (Sentence Length, Batch_size)
iterations += 1
model.train()
optimizer.zero_grad()
loss = criterion(model(batch), predicates_emb[batch.mid, :])
loss.backward()
# clip the gradient
torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip_gradient)
optimizer.step()
# evaluate performance on validation set periodically
if iterations % args.dev_every == 0:
model.eval()
dev_iter.init_epoch()
n_dev_correct = 0
dev_loss = 0
for dev_batch_idx, dev_batch in enumerate(dev_iter):
batch_size = dev_batch.text.size()[1]
answer = model(dev_batch)
learned_pred = euclidean_distances(answer.cpu().data.numpy(), predicates_emb).argmin(axis=1)
n_dev_correct += sum(dev_batch.mid.cpu().data.numpy() == learned_pred)
dev_loss += criterion(answer, predicates_emb[dev_batch.mid, :]).item() * batch_size
curr_accu = n_dev_correct / total_num
total_loss = dev_loss/total_num
print('Dev loss: {}, accuracy: {}'.format(total_loss, curr_accu))
# update model
if curr_accu > best_accu: # total_loss < best_model
best_accu = curr_accu
best_loss = total_loss
iters_not_improved = 0
# save model, delete previous 'best_snapshot' files
torch.save(model, os.path.join(args.output, args.best_prefix + '_best_model.pt'))
else:
iters_not_improved += 1
if iters_not_improved > patience:
early_stop = True
break
if iterations % args.log_every == 1:
print(log_template.format(time.time() - start,
epoch, iterations, 1 + batch_idx, len(train_iter),
100. * (1 + batch_idx) / len(train_iter), loss.item(), ' ' * 8, ' ' * 12))