forked from jiesutd/NCRFpp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_parse.py
476 lines (433 loc) · 19.6 KB
/
main_parse.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
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
# -*- coding: utf-8 -*-
# @Author: Jie
# @Date: 2017-06-15 14:11:08
# @Last Modified by: Jie Yang, Contact: [email protected]
# @Last Modified time: 2019-02-13 10:58:43
from __future__ import print_function
import time
import sys
import argparse
import random
import copy
import torch
import gc
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
from utils.metric import get_ner_fmeasure
from model.seqlabel import SeqLabel
from utils.data import Data
try:
import cPickle as pickle
except ImportError:
import pickle as pickle
seed_num = 42
random.seed(seed_num)
torch.manual_seed(seed_num)
np.random.seed(seed_num)
def data_initialization(data):
data.initial_feature_alphabets()
data.build_alphabet(data.train_dir)
data.build_alphabet(data.dev_dir)
data.build_alphabet(data.test_dir)
data.fix_alphabet()
def predict_check(pred_variable, gold_variable, mask_variable):
"""
input:
pred_variable (batch_size, sent_len): pred tag result, in numpy format
gold_variable (batch_size, sent_len): gold result variable
mask_variable (batch_size, sent_len): mask variable
"""
pred = pred_variable.cpu().data.numpy()
gold = gold_variable.cpu().data.numpy()
mask = mask_variable.cpu().data.numpy()
overlaped = (pred == gold)
right_token = np.sum(overlaped * mask)
total_token = mask.sum()
# print("right: %s, total: %s"%(right_token, total_token))
return right_token, total_token
def recover_label(pred_variable, gold_variable, mask_variable, label_alphabet, word_recover):
"""
input:
pred_variable (batch_size, sent_len): pred tag result
gold_variable (batch_size, sent_len): gold result variable
mask_variable (batch_size, sent_len): mask variable
"""
pred_variable = pred_variable[word_recover]
gold_variable = gold_variable[word_recover]
mask_variable = mask_variable[word_recover]
batch_size = gold_variable.size(0)
seq_len = gold_variable.size(1)
mask = mask_variable.cpu().data.numpy()
pred_tag = pred_variable.cpu().data.numpy()
gold_tag = gold_variable.cpu().data.numpy()
batch_size = mask.shape[0]
pred_label = []
gold_label = []
for idx in range(batch_size):
pred = [label_alphabet.get_instance(pred_tag[idx][idy]) for idy in range(seq_len) if mask[idx][idy] != 0]
gold = [label_alphabet.get_instance(gold_tag[idx][idy]) for idy in range(seq_len) if mask[idx][idy] != 0]
# print("p:",pred, pred_tag.tolist())
# print("g:", gold, gold_tag.tolist())
assert(len(pred)==len(gold))
pred_label.append(pred)
gold_label.append(gold)
return pred_label, gold_label
def recover_nbest_label(pred_variable, mask_variable, label_alphabet, word_recover):
"""
input:
pred_variable (batch_size, sent_len, nbest): pred tag result
mask_variable (batch_size, sent_len): mask variable
word_recover (batch_size)
output:
nbest_pred_label list: [batch_size, nbest, each_seq_len]
"""
# print("word recover:", word_recover.size())
# exit(0)
pred_variable = pred_variable[word_recover]
mask_variable = mask_variable[word_recover]
batch_size = pred_variable.size(0)
seq_len = pred_variable.size(1)
print(pred_variable.size())
nbest = pred_variable.size(2)
mask = mask_variable.cpu().data.numpy()
pred_tag = pred_variable.cpu().data.numpy()
batch_size = mask.shape[0]
pred_label = []
for idx in range(batch_size):
pred = []
for idz in range(nbest):
each_pred = [label_alphabet.get_instance(pred_tag[idx][idy][idz]) for idy in range(seq_len) if mask[idx][idy] != 0]
pred.append(each_pred)
pred_label.append(pred)
return pred_label
# def save_data_setting(data, save_file):
# new_data = copy.deepcopy(data)
# ## remove input instances
# new_data.train_texts = []
# new_data.dev_texts = []
# new_data.test_texts = []
# new_data.raw_texts = []
# new_data.train_Ids = []
# new_data.dev_Ids = []
# new_data.test_Ids = []
# new_data.raw_Ids = []
# ## save data settings
# with open(save_file, 'w') as fp:
# pickle.dump(new_data, fp)
# print("Data setting saved to file: ", save_file)
# def load_data_setting(save_file):
# with open(save_file, 'r') as fp:
# data = pickle.load(fp)
# print("Data setting loaded from file: ", save_file)
# data.show_data_summary()
# return data
def lr_decay(optimizer, epoch, decay_rate, init_lr):
lr = init_lr/(1+decay_rate*epoch)
print(" Learning rate is set as:", lr)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
return optimizer
def evaluate(data, model, name, nbest=None):
if name == "train":
instances = data.train_Ids
elif name == "dev":
instances = data.dev_Ids
elif name == 'test':
instances = data.test_Ids
elif name == 'raw':
instances = data.raw_Ids
else:
print("Error: wrong evaluate name,", name)
right_token = 0
whole_token = 0
nbest_pred_results = []
pred_scores = []
pred_results = []
gold_results = []
## set model in eval model
model.eval()
batch_size = data.HP_batch_size
start_time = time.time()
train_num = len(instances)
total_batch = train_num//batch_size+1
for batch_id in range(total_batch):
start = batch_id*batch_size
end = (batch_id+1)*batch_size
if end > train_num:
end = train_num
instance = instances[start:end]
if not instance:
continue
batch_word, batch_features, batch_wordlen, batch_wordrecover, batch_char, batch_charlen, batch_charrecover, batch_label, mask = batchify_with_label(instance, data.HP_gpu, True)
if nbest:
scores, nbest_tag_seq = model.decode_nbest(batch_word,batch_features, batch_wordlen, batch_char, batch_charlen, batch_charrecover, mask, nbest)
nbest_pred_result = recover_nbest_label(nbest_tag_seq, mask, data.label_alphabet, batch_wordrecover)
nbest_pred_results += nbest_pred_result
pred_scores += scores[batch_wordrecover].cpu().data.numpy().tolist()
## select the best sequence to evalurate
tag_seq = nbest_tag_seq[:,:,0]
else:
tag_seq = model(batch_word, batch_features, batch_wordlen, batch_char, batch_charlen, batch_charrecover, mask)
# print("tag:",tag_seq)
pred_label, gold_label = recover_label(tag_seq, batch_label, mask, data.label_alphabet, batch_wordrecover)
pred_results += pred_label
gold_results += gold_label
decode_time = time.time() - start_time
speed = len(instances)/decode_time
acc, p, r, f = get_ner_fmeasure(gold_results, pred_results, data.tagScheme)
if nbest:
return speed, acc, p, r, f, nbest_pred_results, pred_scores
return speed, acc, p, r, f, pred_results, pred_scores
def batchify_with_label(input_batch_list, gpu, volatile_flag=False):
"""
input: list of words, chars and labels, various length. [[words,chars, labels],[words,chars,labels],...]
words: word ids for one sentence. (batch_size, sent_len)
chars: char ids for on sentences, various length. (batch_size, sent_len, each_word_length)
output:
zero padding for word and char, with their batch length
word_seq_tensor: (batch_size, max_sent_len) Variable
word_seq_lengths: (batch_size,1) Tensor
char_seq_tensor: (batch_size*max_sent_len, max_word_len) Variable
char_seq_lengths: (batch_size*max_sent_len,1) Tensor
char_seq_recover: (batch_size*max_sent_len,1) recover char sequence order
label_seq_tensor: (batch_size, max_sent_len)
mask: (batch_size, max_sent_len)
"""
batch_size = len(input_batch_list)
words = [sent[0] for sent in input_batch_list]
features = [np.asarray(sent[1]) for sent in input_batch_list]
feature_num = len(features[0][0])
chars = [sent[2] for sent in input_batch_list]
labels = [sent[3] for sent in input_batch_list]
word_seq_lengths = torch.LongTensor(map(len, words))
max_seq_len = word_seq_lengths.max()
word_seq_tensor = autograd.Variable(torch.zeros((batch_size, max_seq_len)), volatile = volatile_flag).long()
label_seq_tensor = autograd.Variable(torch.zeros((batch_size, max_seq_len)),volatile = volatile_flag).long()
feature_seq_tensors = []
for idx in range(feature_num):
feature_seq_tensors.append(autograd.Variable(torch.zeros((batch_size, max_seq_len)),volatile = volatile_flag).long())
mask = autograd.Variable(torch.zeros((batch_size, max_seq_len)),volatile = volatile_flag).byte()
for idx, (seq, label, seqlen) in enumerate(zip(words, labels, word_seq_lengths)):
word_seq_tensor[idx, :seqlen] = torch.LongTensor(seq)
label_seq_tensor[idx, :seqlen] = torch.LongTensor(label)
mask[idx, :seqlen] = torch.Tensor([1]*seqlen)
for idy in range(feature_num):
feature_seq_tensors[idy][idx,:seqlen] = torch.LongTensor(features[idx][:,idy])
word_seq_lengths, word_perm_idx = word_seq_lengths.sort(0, descending=True)
word_seq_tensor = word_seq_tensor[word_perm_idx]
for idx in range(feature_num):
feature_seq_tensors[idx] = feature_seq_tensors[idx][word_perm_idx]
label_seq_tensor = label_seq_tensor[word_perm_idx]
mask = mask[word_perm_idx]
### deal with char
# pad_chars (batch_size, max_seq_len)
pad_chars = [chars[idx] + [[0]] * (max_seq_len-len(chars[idx])) for idx in range(len(chars))]
length_list = [map(len, pad_char) for pad_char in pad_chars]
max_word_len = max(map(max, length_list))
char_seq_tensor = autograd.Variable(torch.zeros((batch_size, max_seq_len, max_word_len)), volatile = volatile_flag).long()
char_seq_lengths = torch.LongTensor(length_list)
for idx, (seq, seqlen) in enumerate(zip(pad_chars, char_seq_lengths)):
for idy, (word, wordlen) in enumerate(zip(seq, seqlen)):
# print len(word), wordlen
char_seq_tensor[idx, idy, :wordlen] = torch.LongTensor(word)
char_seq_tensor = char_seq_tensor[word_perm_idx].view(batch_size*max_seq_len,-1)
char_seq_lengths = char_seq_lengths[word_perm_idx].view(batch_size*max_seq_len,)
char_seq_lengths, char_perm_idx = char_seq_lengths.sort(0, descending=True)
char_seq_tensor = char_seq_tensor[char_perm_idx]
_, char_seq_recover = char_perm_idx.sort(0, descending=False)
_, word_seq_recover = word_perm_idx.sort(0, descending=False)
if gpu:
word_seq_tensor = word_seq_tensor.cuda()
for idx in range(feature_num):
feature_seq_tensors[idx] = feature_seq_tensors[idx].cuda()
word_seq_lengths = word_seq_lengths.cuda()
word_seq_recover = word_seq_recover.cuda()
label_seq_tensor = label_seq_tensor.cuda()
char_seq_tensor = char_seq_tensor.cuda()
char_seq_recover = char_seq_recover.cuda()
mask = mask.cuda()
return word_seq_tensor,feature_seq_tensors, word_seq_lengths, word_seq_recover, char_seq_tensor, char_seq_lengths, char_seq_recover, label_seq_tensor, mask
def train(data):
print("Training model...")
data.show_data_summary()
save_data_name = data.model_dir +".dset"
data.save(save_data_name)
model = SeqLabel(data)
loss_function = nn.NLLLoss()
if data.optimizer.lower() == "sgd":
optimizer = optim.SGD(model.parameters(), lr=data.HP_lr, momentum=data.HP_momentum,weight_decay=data.HP_l2)
elif data.optimizer.lower() == "adagrad":
optimizer = optim.Adagrad(model.parameters(), lr=data.HP_lr, weight_decay=data.HP_l2)
elif data.optimizer.lower() == "adadelta":
optimizer = optim.Adadelta(model.parameters(), lr=data.HP_lr, weight_decay=data.HP_l2)
elif data.optimizer.lower() == "rmsprop":
optimizer = optim.RMSprop(model.parameters(), lr=data.HP_lr, weight_decay=data.HP_l2)
elif data.optimizer.lower() == "adam":
optimizer = optim.Adam(model.parameters(), lr=data.HP_lr, weight_decay=data.HP_l2)
else:
print("Optimizer illegal: %s"%(data.optimizer))
exit(0)
best_dev = -10
# data.HP_iteration = 1
## start training
for idx in range(data.HP_iteration):
epoch_start = time.time()
temp_start = epoch_start
print("Epoch: %s/%s" %(idx,data.HP_iteration))
if data.optimizer == "SGD":
optimizer = lr_decay(optimizer, idx, data.HP_lr_decay, data.HP_lr)
instance_count = 0
sample_id = 0
sample_loss = 0
total_loss = 0
right_token = 0
whole_token = 0
random.shuffle(data.train_Ids)
## set model in train model
model.train()
model.zero_grad()
batch_size = data.HP_batch_size
batch_id = 0
train_num = len(data.train_Ids)
total_batch = train_num//batch_size+1
for batch_id in range(total_batch):
start = batch_id*batch_size
end = (batch_id+1)*batch_size
if end >train_num:
end = train_num
instance = data.train_Ids[start:end]
if not instance:
continue
batch_word, batch_features, batch_wordlen, batch_wordrecover, batch_char, batch_charlen, batch_charrecover, batch_label, mask = batchify_with_label(instance, data.HP_gpu)
instance_count += 1
loss, tag_seq = model.neg_log_likelihood_loss(batch_word,batch_features, batch_wordlen, batch_char, batch_charlen, batch_charrecover, batch_label, mask)
right, whole = predict_check(tag_seq, batch_label, mask)
right_token += right
whole_token += whole
sample_loss += loss.data[0]
total_loss += loss.data[0]
if end%500 == 0:
temp_time = time.time()
temp_cost = temp_time - temp_start
temp_start = temp_time
print(" Instance: %s; Time: %.2fs; loss: %.4f; acc: %s/%s=%.4f"%(end, temp_cost, sample_loss, right_token, whole_token,(right_token+0.)/whole_token))
sys.stdout.flush()
sample_loss = 0
loss.backward()
optimizer.step()
model.zero_grad()
temp_time = time.time()
temp_cost = temp_time - temp_start
print(" Instance: %s; Time: %.2fs; loss: %.4f; acc: %s/%s=%.4f"%(end, temp_cost, sample_loss, right_token, whole_token,(right_token+0.)/whole_token))
epoch_finish = time.time()
epoch_cost = epoch_finish - epoch_start
print("Epoch: %s training finished. Time: %.2fs, speed: %.2fst/s, total loss: %s"%(idx, epoch_cost, train_num/epoch_cost, total_loss))
# continue
speed, acc, p, r, f, _,_ = evaluate(data, model, "dev")
dev_finish = time.time()
dev_cost = dev_finish - epoch_finish
if data.seg:
current_score = f
print("Dev: time: %.2fs, speed: %.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"%(dev_cost, speed, acc, p, r, f))
else:
current_score = acc
print("Dev: time: %.2fs speed: %.2fst/s; acc: %.4f"%(dev_cost, speed, acc))
if current_score > best_dev:
if data.seg:
print("Exceed previous best f score:", best_dev)
else:
print("Exceed previous best acc score:", best_dev)
model_name = data.model_dir +'.'+ str(idx) + ".model"
print("Save current best model in file:", model_name)
torch.save(model.state_dict(), model_name)
best_dev = current_score
# ## decode test
speed, acc, p, r, f, _,_ = evaluate(data, model, "test")
test_finish = time.time()
test_cost = test_finish - dev_finish
if data.seg:
print("Test: time: %.2fs, speed: %.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"%(test_cost, speed, acc, p, r, f))
else:
print("Test: time: %.2fs, speed: %.2fst/s; acc: %.4f"%(test_cost, speed, acc))
gc.collect()
def load_model_decode(data, name):
print("Load Model from file: ", data.model_dir)
model = SeqLabel(data)
## load model need consider if the model trained in GPU and load in CPU, or vice versa
# if not gpu:
# model.load_state_dict(torch.load(model_dir))
# # model.load_state_dict(torch.load(model_dir), map_location=lambda storage, loc: storage)
# # model = torch.load(model_dir, map_location=lambda storage, loc: storage)
# else:
# model.load_state_dict(torch.load(model_dir))
# # model = torch.load(model_dir)
model.load_state_dict(torch.load(data.load_model_dir))
print("Decode %s data, nbest: %s ..."%(name, data.nbest))
start_time = time.time()
speed, acc, p, r, f, pred_results, pred_scores = evaluate(data, model, name, data.nbest)
end_time = time.time()
time_cost = end_time - start_time
if data.seg:
print("%s: time:%.2fs, speed:%.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"%(name, time_cost, speed, acc, p, r, f))
else:
print("%s: time:%.2fs, speed:%.2fst/s; acc: %.4f"%(name, time_cost, speed, acc))
return pred_results, pred_scores
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Tuning with NCRF++')
parser.add_argument('--wordemb', help='Embedding for words', default='None')
parser.add_argument('--charemb', help='Embedding for chars', default='None')
parser.add_argument('--status', choices=['train', 'decode'], help='update algorithm', default='train')
parser.add_argument('--savemodel', default="data/model/saved_model.lstmcrf.")
parser.add_argument('--savedset', help='Dir of saved data setting')
parser.add_argument('--train', default="data/conll03/train.bmes")
parser.add_argument('--dev', default="data/conll03/dev.bmes" )
parser.add_argument('--test', default="data/conll03/test.bmes")
parser.add_argument('--seg', default="True")
parser.add_argument('--raw')
parser.add_argument('--loadmodel')
parser.add_argument('--output')
args = parser.parse_args()
data = Data()
data.train_dir = args.train
data.dev_dir = args.dev
data.test_dir = args.test
data.model_dir = args.savemodel
data.dset_dir = args.savedset
print("dset directory:",data.dset_dir)
status = args.status.lower()
save_model_dir = args.savemodel
data.HP_gpu = torch.cuda.is_available()
print("Seed num:",seed_num)
data.number_normalized = True
data.word_emb_dir = "../data/glove.6B.100d.txt"
if status == 'train':
print("MODEL: train")
data_initialization(data)
data.use_char = True
data.HP_batch_size = 10
data.HP_lr = 0.015
data.char_seq_feature = "CNN"
data.generate_instance('train')
data.generate_instance('dev')
data.generate_instance('test')
data.build_pretrain_emb()
train(data)
elif status == 'decode':
print("MODEL: decode")
data.load(data.dset_dir)
data.raw_dir = args.raw
data.decode_dir = args.output
data.load_model_dir = args.loadmodel
data.show_data_summary()
data.generate_instance('raw')
print("nbest: %s"%(data.nbest))
decode_results, pred_scores = load_model_decode(data, 'raw')
if data.nbest:
data.write_nbest_decoded_results(decode_results, pred_scores, 'raw')
else:
data.write_decoded_results(decode_results, 'raw')
else:
print("Invalid argument! Please use valid arguments! (train/test/decode)")