-
Notifications
You must be signed in to change notification settings - Fork 5
/
train_demo.py
227 lines (212 loc) · 9.26 KB
/
train_demo.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
from fewshot_re_kit.data_loader import get_loader, get_loader_pair, get_loader_unsupervised
from fewshot_re_kit.framework import FewShotREFramework
from fewshot_re_kit.sentence_encoder import CNNSentenceEncoder, BERTSentenceEncoder, BERTPAIRSentenceEncoder
import models
from models.regrab import REGRAB
from models.d import Discriminator
import sys
import torch
from torch import optim, nn
import numpy as np
import json
import argparse
import os
import random
import pickle
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--train', default='train_wiki',
help='train file')
parser.add_argument('--val', default='val_wiki',
help='val file')
parser.add_argument('--test', default='test_wiki',
help='test file')
parser.add_argument('--adv', default=None,
help='adv file')
parser.add_argument('--trainN', default=10, type=int,
help='N in train')
parser.add_argument('--N', default=5, type=int,
help='N way')
parser.add_argument('--K', default=5, type=int,
help='K shot')
parser.add_argument('--Q', default=5, type=int,
help='Num of query per class')
parser.add_argument('--batch_size', default=4, type=int,
help='batch size')
parser.add_argument('--train_iter', default=30000, type=int,
help='num of iters in training')
parser.add_argument('--val_iter', default=1000, type=int,
help='num of iters in validation')
parser.add_argument('--test_iter', default=10000, type=int,
help='num of iters in testing')
parser.add_argument('--val_step', default=2000, type=int,
help='val after training how many iters')
parser.add_argument('--model', default='regrab',
help='model name')
parser.add_argument('--encoder', default='bert',
help='encoder: cnn or bert')
parser.add_argument('--max_length', default=128, type=int,
help='max length')
parser.add_argument('--lr', default=1e-1, type=float,
help='learning rate')
parser.add_argument('--weight_decay', default=1e-5, type=float,
help='weight decay')
parser.add_argument('--dropout', default=0.0, type=float,
help='dropout rate')
parser.add_argument('--na_rate', default=0, type=int,
help='NA rate (NA = Q * na_rate)')
parser.add_argument('--grad_iter', default=1, type=int,
help='accumulate gradient every x iterations')
parser.add_argument('--optim', default='sgd',
help='sgd / adam / bert_adam')
parser.add_argument('--hidden_size', default=230, type=int,
help='hidden size')
parser.add_argument('--load_ckpt', default=None,
help='load ckpt')
parser.add_argument('--save_ckpt', default=None,
help='save ckpt')
parser.add_argument('--fp16', action='store_true',
help='use nvidia apex fp16')
parser.add_argument('--only_test', action='store_true',
help='only test')
parser.add_argument('--pair', action='store_true',
help='use pair model')
parser.add_argument('--eps', default=0.1, type=float,
help='step size for SG-MCMC')
parser.add_argument('--temp', default=10.0, type=float,
help='temperature for softmax')
parser.add_argument('--step', default=5, type=int,
help='steps for SG-MCMC')
parser.add_argument('--smp', default=10, type=int,
help='samples for SG-MCMC')
parser.add_argument('--ratio', default=0.01, type=float,
help='decay ratio of step size for SG-MCMC')
parser.add_argument('--wtp', default=0.1, type=float,
help='weight of the prior term')
parser.add_argument('--wtn', default=1.0, type=float,
help='weight of the noise term')
parser.add_argument('--wtb', default=0.0, type=float,
help='weight of the background term')
parser.add_argument('--metric', default='dot',
help='similarity metric (dot or l2)')
parser.add_argument('--seed', default=1234, type=int,
help='random seed')
opt = parser.parse_args()
trainN = opt.trainN
N = opt.N
K = opt.K
Q = opt.Q
batch_size = opt.batch_size
model_name = opt.model
encoder_name = opt.encoder
max_length = opt.max_length
torch.manual_seed(opt.seed)
np.random.seed(opt.seed)
random.seed(opt.seed)
torch.cuda.manual_seed(opt.seed)
print("{}-way-{}-shot Few-Shot Relation Classification".format(N, K))
print("model: {}".format(model_name))
print("encoder: {}".format(encoder_name))
print("max_length: {}".format(max_length))
if encoder_name == 'cnn':
try:
glove_mat = np.load('./pretrain/glove/glove_mat.npy')
glove_word2id = json.load(open('./pretrain/glove/glove_word2id.json'))
except:
raise Exception("Cannot find glove files. Run glove/download_glove.sh to download glove files.")
sentence_encoder = CNNSentenceEncoder(
glove_mat,
glove_word2id,
max_length)
elif encoder_name == 'bert':
if opt.pair:
sentence_encoder = BERTPAIRSentenceEncoder(
'./pretrain/bert-base-uncased',
max_length)
else:
sentence_encoder = BERTSentenceEncoder(
'./pretrain/bert-base-uncased',
max_length)
else:
raise NotImplementedError
if opt.pair:
train_data_loader = get_loader_pair(opt.train, sentence_encoder,
N=trainN, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
val_data_loader = get_loader_pair(opt.val, sentence_encoder,
N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
test_data_loader = get_loader_pair(opt.test, sentence_encoder,
N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
else:
train_data_loader = get_loader(opt.train, sentence_encoder,
N=trainN, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
val_data_loader = get_loader(opt.val, sentence_encoder,
N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
test_data_loader = get_loader(opt.test, sentence_encoder,
N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
if opt.adv:
adv_data_loader = get_loader_unsupervised(opt.adv, sentence_encoder,
N=trainN, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
if opt.optim == 'sgd':
pytorch_optim = optim.SGD
elif opt.optim == 'adam':
pytorch_optim = optim.Adam
elif opt.optim == 'bert_adam':
from transformers import AdamW
pytorch_optim = AdamW
else:
raise NotImplementedError
if opt.adv:
d = Discriminator(opt.hidden_size)
framework = FewShotREFramework(train_data_loader, val_data_loader, test_data_loader, adv_data_loader, adv=opt.adv, d=d)
else:
framework = FewShotREFramework(train_data_loader, val_data_loader, test_data_loader)
prefix = '-'.join([model_name, encoder_name, opt.train, opt.val, str(N), str(K)])
if opt.adv is not None:
prefix += '-adv_' + opt.adv
if opt.na_rate != 0:
prefix += '-na{}'.format(opt.na_rate)
# Loading relation embeddings.
data = pickle.load(open('./data/embeddings.pkl', 'rb'))
rellist = data['relations']
relemb = data['embeddings']
array0 = np.zeros((1,relemb.shape[1]), dtype=relemb.dtype)
relemb = np.concatenate([array0, relemb], axis=0)
rel2id = dict([(rel, k + 1) for k, rel in enumerate(rellist)])
# Loading relation graphs.
with open('./data/graph.txt', 'r') as fi:
us, vs, ws = [], [], []
for line in fi:
items = line.strip().split('\t')
us += [rel2id[items[0]]]
vs += [rel2id[items[1]]]
ws += [float(items[2])]
index = torch.LongTensor([us, vs])
value = torch.Tensor(ws)
shape = torch.Size([len(rel2id) + 1, len(rel2id) + 1])
reladj = torch.sparse.FloatTensor(index, value, shape).cuda()
# End
model = REGRAB(sentence_encoder, hidden_size=opt.hidden_size, eps=opt.eps, temp=opt.temp, step=opt.step, smp=opt.smp, ratio=opt.ratio, wtp=opt.wtp, wtn=opt.wtn, wtb=opt.wtb, metric=opt.metric)
model.set_relemb(rel2id, relemb)
model.set_reladj(reladj)
if not os.path.exists('checkpoint'):
os.mkdir('checkpoint')
ckpt = 'checkpoint/{}.pth.tar'.format(prefix)
if opt.save_ckpt:
ckpt = opt.save_ckpt
if torch.cuda.is_available():
model.cuda()
if not opt.only_test:
if encoder_name == 'bert':
bert_optim = True
else:
bert_optim = False
framework.train(model, prefix, batch_size, trainN, N, K, Q,
pytorch_optim=pytorch_optim, load_ckpt=opt.load_ckpt, save_ckpt=ckpt,
na_rate=opt.na_rate, val_step=opt.val_step, fp16=opt.fp16, pair=opt.pair,
train_iter=opt.train_iter, val_iter=opt.val_iter, bert_optim=bert_optim)
else:
ckpt = opt.load_ckpt
acc = framework.eval(model, batch_size, N, K, Q, opt.test_iter, na_rate=opt.na_rate, ckpt=ckpt, pair=opt.pair)
print("RESULT: %.2f" % (acc * 100))
if __name__ == "__main__":
main()