forked from hsinyuan-huang/FlowQA
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_CoQA.py
323 lines (281 loc) · 14.1 KB
/
train_CoQA.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
import os
import re
import sys
import random
import string
import logging
import argparse
from shutil import copyfile
from datetime import datetime
from collections import Counter
import torch
import msgpack
import pandas as pd
import numpy as np
from QA_model.model_CoQA import QAModel
from CoQA_eval import CoQAEvaluator
from general_utils import find_best_score_and_thresh, BatchGen_CoQA
parser = argparse.ArgumentParser(
description='Train a Dialog QA model.'
)
# system
parser.add_argument('--task_name', default='CoQA')
parser.add_argument('--name', default='', help='additional name of the current run')
parser.add_argument('--log_file', default='output.log',
help='path for log file.')
parser.add_argument('--log_per_updates', type=int, default=20,
help='log model loss per x updates (mini-batches).')
parser.add_argument('--train_dir', default='CoQA/')
parser.add_argument('--dev_dir', default='CoQA/')
parser.add_argument('--answer_type_num', type=int, default=4)
parser.add_argument('--model_dir', default='models',
help='path to store saved models.')
parser.add_argument('--eval_per_epoch', type=int, default=1,
help='perform evaluation per x epoches.')
parser.add_argument('--MTLSTM_path', default='glove/MT-LSTM.pth')
parser.add_argument('--save_all', dest='save_best_only', action='store_false', help='save all models.')
parser.add_argument('--do_not_save', action='store_true', help='don\'t save any model')
parser.add_argument('--save_for_predict', action='store_true')
parser.add_argument('--seed', type=int, default=1023,
help='random seed for data shuffling, dropout, etc.')
parser.add_argument('--cuda', type=bool, default=torch.cuda.is_available(),
help='whether to use GPU acceleration.')
# training
parser.add_argument('-e', '--epoches', type=int, default=30)
parser.add_argument('-bs', '--batch_size', type=int, default=1)
parser.add_argument('-ebs', '--elmo_batch_size', type=int, default=12)
parser.add_argument('-rs', '--resume', default='',
help='previous model pathname. '
'e.g. "models/checkpoint_epoch_11.pt"')
parser.add_argument('-ro', '--resume_options', action='store_true',
help='use previous model options, ignore the cli and defaults.')
parser.add_argument('-rlr', '--reduce_lr', type=float, default=0.,
help='reduce initial (resumed) learning rate by this factor.')
parser.add_argument('-op', '--optimizer', default='adamax',
help='supported optimizer: adamax, sgd, adadelta, adam')
parser.add_argument('-gc', '--grad_clipping', type=float, default=10)
parser.add_argument('-wd', '--weight_decay', type=float, default=0)
parser.add_argument('-lr', '--learning_rate', type=float, default=0.1,
help='only applied to SGD.')
parser.add_argument('-mm', '--momentum', type=float, default=0,
help='only applied to SGD.')
parser.add_argument('-tp', '--tune_partial', type=int, default=1000,
help='finetune top-x embeddings (including <PAD>, <UNK>).')
parser.add_argument('--fix_embeddings', action='store_true',
help='if true, `tune_partial` will be ignored.')
parser.add_argument('--elmo_lambda', type=float, default=0.0)
parser.add_argument('--rationale_lambda', type=float, default=0.0)
parser.add_argument('--no_question_normalize', dest='question_normalize', action='store_false') # when set, do dialog normalize
parser.add_argument('--pretrain', default='')
# model
parser.add_argument('--explicit_dialog_ctx', type=int, default=1)
parser.add_argument('--no_dialog_flow', action='store_true')
parser.add_argument('--no_hierarchical_query', dest='do_hierarchical_query', action='store_false')
parser.add_argument('--no_prealign', dest='do_prealign', action='store_false')
parser.add_argument('--final_output_att_hidden', type=int, default=250)
parser.add_argument('--question_merge', default='linear_self_attn')
parser.add_argument('--no_ptr_update', dest='do_ptr_update', action='store_false')
parser.add_argument('--no_ptr_net_indep_attn', dest='ptr_net_indep_attn', action='store_false')
parser.add_argument('--ptr_net_attn_type', default='Bilinear', help="Attention for answer span output: Bilinear, MLP or Default")
parser.add_argument('--do_residual_rnn', dest='do_residual_rnn', action='store_true')
parser.add_argument('--do_residual_everything', dest='do_residual_everything', action='store_true')
parser.add_argument('--do_residual', dest='do_residual', action='store_true')
parser.add_argument('--rnn_layers', type=int, default=1, help="Default number of RNN layers")
parser.add_argument('--rnn_type', default='lstm',
help='supported types: rnn, gru, lstm')
parser.add_argument('--concat_rnn', dest='concat_rnn', action='store_true')
parser.add_argument('--deep_inter_att_do_similar', type=int, default=0)
parser.add_argument('--deep_att_hidden_size_per_abstr', type=int, default=250)
parser.add_argument('--hidden_size', type=int, default=125)
parser.add_argument('--self_attention_opt', type=int, default=1) # 0: no self attention
parser.add_argument('--no_elmo', dest='use_elmo', action='store_false')
parser.add_argument('--no_em', action='store_true')
parser.add_argument('--no_wemb', dest='use_wemb', action='store_false') # word embedding
parser.add_argument('--CoVe_opt', type=int, default=1) # contexualized embedding option
parser.add_argument('--no_pos', dest='use_pos', action='store_false') # pos tagging
parser.add_argument('--pos_size', type=int, default=51, help='how many kinds of POS tags.')
parser.add_argument('--pos_dim', type=int, default=12, help='the embedding dimension for POS tags.')
parser.add_argument('--no_ner', dest='use_ner', action='store_false') # named entity
parser.add_argument('--ner_size', type=int, default=19, help='how many kinds of named entity tags.')
parser.add_argument('--ner_dim', type=int, default=8, help='the embedding dimension for named entity tags.')
parser.add_argument('--prealign_hidden', type=int, default=300)
parser.add_argument('--prealign_option', type=int, default=2, help='0: No prealign, 1, 2, ...: Different options')
parser.add_argument('--no_seq_dropout', dest='do_seq_dropout', action='store_false')
parser.add_argument('--my_dropout_p', type=float, default=0.4)
parser.add_argument('--dropout_emb', type=float, default=0.4)
parser.add_argument('--max_len', type=int, default=15)
args = parser.parse_args()
if args.name != '':
args.model_dir = args.model_dir + '_' + args.name
args.log_file = os.path.dirname(args.log_file) + 'output_' + args.name + '.log'
# set model dir
model_dir = args.model_dir
os.makedirs(model_dir, exist_ok=True)
model_dir = os.path.abspath(model_dir)
# set random seed
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed_all(args.seed)
# setup logger
log = logging.getLogger(__name__)
log.setLevel(logging.DEBUG)
fh = logging.FileHandler(args.log_file)
fh.setLevel(logging.DEBUG)
ch = logging.StreamHandler(sys.stdout)
ch.setLevel(logging.INFO)
formatter = logging.Formatter(fmt='%(asctime)s %(message)s', datefmt='%m/%d/%Y %I:%M:%S')
fh.setFormatter(formatter)
ch.setFormatter(formatter)
log.addHandler(fh)
log.addHandler(ch)
def main():
log.info('[program starts.]')
opt = vars(args) # changing opt will change args
train, train_embedding, opt = load_train_data(opt)
dev, dev_embedding = load_dev_data(opt)
opt['num_features'] += args.explicit_dialog_ctx * 3 # dialog_act + previous answer
if opt['use_elmo'] == False:
opt['elmo_batch_size'] = 0
CoQAEval = CoQAEvaluator("CoQA/dev.json")
log.info('[Data loaded.]')
if args.resume:
log.info('[loading previous model...]')
checkpoint = torch.load(args.resume)
if args.resume_options:
opt = checkpoint['config']
state_dict = checkpoint['state_dict']
model = QAModel(opt, train_embedding, state_dict)
epoch_0 = checkpoint['epoch'] + 1
for i in range(checkpoint['epoch']):
random.shuffle(list(range(len(train)))) # synchronize random seed
if args.reduce_lr:
lr_decay(model.optimizer, lr_decay=args.reduce_lr)
else:
model = QAModel(opt, train_embedding)
epoch_0 = 1
if args.pretrain:
pretrain_model = torch.load(args.pretrain)
state_dict = pretrain_model['state_dict']['network']
model.get_pretrain(state_dict)
model.setup_eval_embed(dev_embedding)
log.info("[dev] Total number of params: {}".format(model.total_param))
if args.cuda:
model.cuda()
if args.resume:
batches = BatchGen_CoQA(dev, batch_size=args.batch_size, evaluation=True, gpu=args.cuda, dialog_ctx=args.explicit_dialog_ctx)
predictions = []
for batch in batches:
phrases, noans = model.predict(batch)
predictions.extend(phrases)
f1 = CoQAEval.compute_turn_score_seq(predictions)
log.info("[dev F1: {:.3f}]".format(f1))
best_val_score = f1
else:
best_val_score = 0.0
for epoch in range(epoch_0, epoch_0 + args.epoches):
log.warning('Epoch {}'.format(epoch))
# train
batches = BatchGen_CoQA(train, batch_size=args.batch_size, gpu=args.cuda, dialog_ctx=args.explicit_dialog_ctx, precompute_elmo=args.elmo_batch_size // args.batch_size)
start = datetime.now()
for i, batch in enumerate(batches):
model.update(batch)
if i % args.log_per_updates == 0:
log.info('updates[{0:6}] train loss[{1:.5f}] remaining[{2}]'.format(
model.updates, model.train_loss.avg,
str((datetime.now() - start) / (i + 1) * (len(batches) - i - 1)).split('.')[0]))
# eval
if epoch % args.eval_per_epoch == 0:
batches = BatchGen_CoQA(dev, batch_size=args.batch_size, evaluation=True, gpu=args.cuda, dialog_ctx=args.explicit_dialog_ctx, precompute_elmo=args.elmo_batch_size // args.batch_size)
predictions = []
for batch in batches:
phrases = model.predict(batch)
predictions.extend(phrases)
f1 = CoQAEval.compute_turn_score_seq(predictions)
# save
if args.save_best_only:
if f1 > best_val_score:
best_val_score = f1
model_file = os.path.join(model_dir, 'best_model.pt')
model.save(model_file, epoch)
log.info('[new best model saved.]')
else:
model_file = os.path.join(model_dir, 'checkpoint_epoch_{}.pt'.format(epoch))
model.save(model_file, epoch)
if f1 > best_val_score:
best_val_score = f1
copyfile(os.path.join(model_dir, model_file),
os.path.join(model_dir, 'best_model.pt'))
log.info('[new best model saved.]')
log.warning("Epoch {} - dev F1: {:.3f} (Best F1: {:.3f})".format(epoch, f1 * 100.0, best_val_score * 100.0))
def lr_decay(optimizer, lr_decay):
for param_group in optimizer.param_groups:
param_group['lr'] *= lr_decay
log.info('[learning rate reduced by {}]'.format(lr_decay))
return optimizer
def load_train_data(opt):
with open(os.path.join(args.train_dir, 'train_meta.msgpack'), 'rb') as f:
meta = msgpack.load(f, encoding='utf8')
embedding = torch.Tensor(meta['embedding'])
opt['vocab_size'] = embedding.size(0)
opt['embedding_dim'] = embedding.size(1)
with open(os.path.join(args.train_dir, 'train_data.msgpack'), 'rb') as f:
data = msgpack.load(f, encoding='utf8')
#data_orig = pd.read_csv(os.path.join(args.train_dir, 'train.csv'))
opt['num_features'] = len(data['context_features'][0][0])
train = {'context': list(zip(
data['context_ids'],
data['context_tags'],
data['context_ents'],
data['context'],
data['context_span'],
data['1st_question'],
data['context_tokenized'])),
'qa': list(zip(
data['question_CID'],
data['question_ids'],
data['context_features'],
data['answer_start'],
data['answer_end'],
data['rationale_start'],
data['rationale_end'],
data['answer_choice'],
data['question'],
data['answer'],
data['question_tokenized']))
}
return train, embedding, opt
def load_dev_data(opt): # can be extended to true test set
with open(os.path.join(args.dev_dir, 'dev_meta.msgpack'), 'rb') as f:
meta = msgpack.load(f, encoding='utf8')
embedding = torch.Tensor(meta['embedding'])
assert opt['embedding_dim'] == embedding.size(1)
with open(os.path.join(args.dev_dir, 'dev_data.msgpack'), 'rb') as f:
data = msgpack.load(f, encoding='utf8')
#data_orig = pd.read_csv(os.path.join(args.dev_dir, 'dev.csv'))
assert opt['num_features'] == len(data['context_features'][0][0])
dev = {'context': list(zip(
data['context_ids'],
data['context_tags'],
data['context_ents'],
data['context'],
data['context_span'],
data['1st_question'],
data['context_tokenized'])),
'qa': list(zip(
data['question_CID'],
data['question_ids'],
data['context_features'],
data['answer_start'],
data['answer_end'],
data['rationale_start'],
data['rationale_end'],
data['answer_choice'],
data['question'],
data['answer'],
data['question_tokenized']))
}
return dev, embedding
if __name__ == '__main__':
main()