-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain.py
847 lines (733 loc) · 32.2 KB
/
main.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
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import torch.optim as optim
from torch.nn.utils.rnn import pad_sequence
from torch.nn.utils import clip_grad_norm_
from utils.data_utils import RawDataset, MorphFeaturizer, create_gender_embeddings
from utils.data_utils import Vocabulary, SeqVocabulary
from utils.metrics import accuracy
import json
import random
import re
import numpy as np
import argparse
from gensim.models import KeyedVectors
from seq2seq import Seq2Seq
from greedy_decoder import BatchSampler
from beam_decoder import BeamSampler
import matplotlib.pyplot as plt
import logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
class Vectorizer:
"""Vectorizer Class"""
def __init__(self, src_vocab_char, trg_vocab_char,
src_vocab_word, src_labels_vocab,
trg_labels_vocab, trg_gender_vocab):
"""
Args:
- src_vocab_char (SeqVocabulary): source vocab on the char level
- trg_vocab_char (SeqVocabulary): target vocab on the char level
- src_vocab_word (SeqVocabulary): source vocab on the word level
- src_labels_vocab (Vocabulary): source labels vocab on the sentence level
- trg_labels_vocab (Vocabulary): target labels vocab on the sentence level
- trg_gender_vocab (Vocabulary): target gender vocab on the sentence level
"""
self.src_vocab_char = src_vocab_char
self.trg_vocab_char = trg_vocab_char
self.src_vocab_word = src_vocab_word
self.src_labels_vocab = src_labels_vocab
self.trg_labels_vocab = trg_labels_vocab
self.trg_gender_vocab = trg_gender_vocab
@classmethod
def create_vectorizer(cls, data_examples):
"""Class method which builds the vectorizer
vocab
Args:
- data_examples: list of InputExample
Returns:
- Vectorizer object
"""
src_vocab_char = SeqVocabulary()
src_vocab_word = SeqVocabulary()
trg_vocab_char = SeqVocabulary()
src_labels_vocab = Vocabulary()
trg_labels_vocab = Vocabulary()
trg_gender_vocab = Vocabulary()
for ex in data_examples:
src = ex.src
trg = ex.trg
src_label = ex.src_label
trg_label = ex.trg_label
trg_gender = ex.trg_gender
# splitting by a regex to maintain the space
src = re.split(r'(\s+)', src)
trg = re.split(r'(\s+)', trg)
for word in src:
src_vocab_word.add_token(word)
src_vocab_char.add_many(list(word))
for word in trg:
trg_vocab_char.add_many(list(word))
src_labels_vocab.add_token(src_label)
trg_labels_vocab.add_token(trg_label)
trg_gender_vocab.add_token(trg_gender)
return cls(src_vocab_char, trg_vocab_char,
src_vocab_word, src_labels_vocab,
trg_labels_vocab, trg_gender_vocab)
def get_src_indices(self, seq):
"""Converts the source sequence chars
to indices
Args:
- seq (str): The source sequence
Returns:
- char_level_indices (list): <s> + List of chars to index mapping + </s>
- word_level_indices (list): <s> + List of words to index mapping + </s>
"""
char_level_indices = [self.src_vocab_char.sos_idx]
word_level_indices = [self.src_vocab_word.sos_idx]
seq = re.split(r'(\s+)', seq)
for word in seq:
for c in word:
char_level_indices.append(self.src_vocab_char.lookup_token(c))
word_level_indices.append(self.src_vocab_word.lookup_token(word))
word_level_indices.append(self.src_vocab_word.eos_idx)
char_level_indices.append(self.src_vocab_char.eos_idx)
assert len(word_level_indices) == len(char_level_indices)
return char_level_indices, word_level_indices
def get_trg_indices(self, seq):
"""Converts the target sequence chars
to indices
Args:
- seq (str): The target sequence
Returns:
- trg_x_indices (list): <s> + List of chars to index mapping
- trg_y_indices (list): List of chars to index mapping + </s>
"""
indices = [self.trg_vocab_char.lookup_token(t) for t in seq]
trg_x_indices = [self.trg_vocab_char.sos_idx] + indices
trg_y_indices = indices + [self.trg_vocab_char.eos_idx]
return trg_x_indices, trg_y_indices
def vectorize(self, src, trg, src_label, trg_label, trg_gender):
"""
Args:
- src (str): The source sequence
- trg (str): The target sequence
- src_label (str): The source sequence label
- trg_label (str): The target sequence label
- trg_label (str): The target sequence gender
Returns:
- vectorized_src_char (tensor): <s> + vectorized source seq on the char level + </s>
- vectorized_src_word (tensor): <s> + vectorized source seq on the word level + </s>
- vectorized_trg_x (tensor): <s> + vectorized target seq on the char level
- vectorized_trg_y (tensor): vectorized target seq on the char level + </s>
- vectorized_src_label (tensor): vectorized source label
- vectorized_trg_label (tensor): vectorized target label
- vectorized_trg_gender (tensor): vectorized target gender
"""
vectorized_src_char, vectorized_src_word = self.get_src_indices(src)
vectorized_trg_x, vectorized_trg_y = self.get_trg_indices(trg)
vectorized_src_label = self.src_labels_vocab.lookup_token(src_label)
vectorized_trg_label = self.trg_labels_vocab.lookup_token(trg_label)
vectorized_trg_gender = self.trg_gender_vocab.lookup_token(trg_gender)
return {'src_char': torch.tensor(vectorized_src_char, dtype=torch.long),
'src_word': torch.tensor(vectorized_src_word, dtype=torch.long),
'trg_x': torch.tensor(vectorized_trg_x, dtype=torch.long),
'trg_y': torch.tensor(vectorized_trg_y, dtype=torch.long),
'src_label': torch.tensor(vectorized_src_label, dtype=torch.long),
'trg_label': torch.tensor(vectorized_trg_label, dtype=torch.long),
'trg_gender': torch.tensor(vectorized_trg_gender, dtype=torch.long)
}
def to_serializable(self):
return {'src_vocab_char': self.src_vocab_char.to_serializable(),
'trg_vocab_char': self.trg_vocab_char.to_serializable(),
'src_vocab_word': self.src_vocab_word.to_serializable(),
'src_labels_vocab': self.src_labels_vocab.to_serializable(),
'trg_labels_vocab': self.trg_labels_vocab.to_serializable(),
'trg_gender_vocab': self.trg_gender_vocab.to_serializable()
}
@classmethod
def from_serializable(cls, contents):
src_vocab_char = SeqVocabulary.from_serializable(contents['src_vocab_char'])
src_vocab_word = SeqVocabulary.from_serializable(contents['src_vocab_word'])
trg_vocab_char = SeqVocabulary.from_serializable(contents['trg_vocab_char'])
src_labels_vocab = Vocabulary.from_serializable(contents['src_labels_vocab'])
trg_labels_vocab = Vocabulary.from_serializable(contents['trg_labels_vocab'])
trg_gender_vocab = Vocabulary.from_serializable(contents['trg_gender_vocab'])
return cls(src_vocab_char, trg_vocab_char,
src_vocab_word, src_labels_vocab,
trg_labels_vocab, trg_gender_vocab)
class MT_Dataset(Dataset):
"""MT Dataset as a PyTorch dataset"""
def __init__(self, raw_dataset, vectorizer):
"""
Args:
- raw_dataset (RawDataset): raw dataset object
- vectorizer (Vectorizer): vectorizer object
"""
self.vectorizer = vectorizer
self.train_examples = raw_dataset.train_examples
self.dev_examples = raw_dataset.dev_examples
self.test_examples = raw_dataset.test_examples
self.lookup_split = {'train': self.train_examples,
'dev': self.dev_examples,
'test': self.test_examples}
self.set_split('train')
def get_vectorizer(self):
return self.vectorizer
@classmethod
def load_data_and_create_vectorizer(cls, data_dir):
raw_dataset = RawDataset(data_dir)
# Note: we always create the vectorized based on the train examples
vectorizer = Vectorizer.create_vectorizer(raw_dataset.train_examples)
return cls(raw_dataset, vectorizer)
@classmethod
def load_data_and_load_vectorizer(cls, data_dir, vec_path):
raw_dataset = RawDataset(data_dir)
vectorizer = cls.load_vectorizer(vec_path)
return cls(raw_dataset, vectorizer)
@staticmethod
def load_vectorizer(vec_path):
with open(vec_path) as f:
return Vectorizer.from_serializable(json.load(f))
def save_vectorizer(self, vec_path):
with open(vec_path, 'w') as f:
return json.dump(self.vectorizer.to_serializable(), f, ensure_ascii=False)
def set_split(self, split):
self.split = split
self.split_examples = self.lookup_split[self.split]
return self.split_examples
def __getitem__(self, index):
example = self.split_examples[index]
src, trg = example.src, example.trg
src_label, trg_label = example.src_label, example.trg_label
trg_gender = example.trg_gender
vectorized = self.vectorizer.vectorize(src, trg, src_label, trg_label, trg_gender)
return vectorized
def __len__(self):
return len(self.split_examples)
class Collator:
def __init__(self, char_src_pad_idx, char_trg_pad_idx,
word_src_pad_idx):
"""
Args:
- char_src_pad_idx: source vocab padding index on the char level
- char_trg_pad_idx: target vocab padding index on the char level
- word_src_pad_idx: source vocab padding index on the word level
"""
self.char_src_pad_idx = char_src_pad_idx
self.word_src_pad_idx = word_src_pad_idx
self.char_trg_pad_idx = char_trg_pad_idx
def __call__(self, batch):
# Sorting the batch by src seqs length in descending order
sorted_batch = sorted(batch, key=lambda x: x['src_char'].shape[0], reverse=True)
src_char_seqs = [x['src_char'] for x in sorted_batch]
src_word_seqs = [x['src_word'] for x in sorted_batch]
src_seqs_labels = [x['src_label'] for x in sorted_batch]
assert len(src_word_seqs) == len(src_char_seqs)
trg_x_seqs = [x['trg_x'] for x in sorted_batch]
trg_y_seqs = [x['trg_y'] for x in sorted_batch]
trg_seqs_labels = [x['trg_label'] for x in sorted_batch]
trg_seqs_genders = [x['trg_gender'] for x in sorted_batch]
lengths = [len(seq) for seq in src_char_seqs]
padded_src_char_seqs = pad_sequence(src_char_seqs, batch_first=True, padding_value=self.char_src_pad_idx)
padded_src_word_seqs = pad_sequence(src_word_seqs, batch_first=True, padding_value=self.word_src_pad_idx)
padded_trg_x_seqs = pad_sequence(trg_x_seqs, batch_first=True, padding_value=self.char_trg_pad_idx)
padded_trg_y_seqs = pad_sequence(trg_y_seqs, batch_first=True, padding_value=self.char_trg_pad_idx)
lengths = torch.tensor(lengths, dtype=torch.long)
src_seqs_labels = torch.tensor(src_seqs_labels, dtype=torch.long)
trg_seqs_labels = torch.tensor(trg_seqs_labels, dtype=torch.long)
trg_seqs_genders = torch.tensor(trg_seqs_genders, dtype=torch.long)
return {'src_char': padded_src_char_seqs,
'src_word': padded_src_word_seqs,
'trg_x': padded_trg_x_seqs,
'trg_y': padded_trg_y_seqs,
'src_lengths': lengths,
'src_label': src_seqs_labels,
'trg_label': trg_seqs_labels,
'trg_gender': trg_seqs_genders
}
def set_seed(seed, cuda):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if cuda:
torch.cuda.manual_seed(seed)
def train(model, dataloader, optimizer, criterion, device='cpu',
teacher_forcing_prob=1, clip_grad=1.0):
model.train()
epoch_loss = 0
for batch in dataloader:
optimizer.zero_grad()
batch = {k: v.to(device) for k, v in batch.items()}
src_char = batch['src_char']
src_word = batch['src_word']
trg_x = batch['trg_x']
trg_y = batch['trg_y']
src_lengths = batch['src_lengths']
trg_gender = batch['trg_gender']
preds, attention_scores = model(char_src_seqs=src_char,
word_src_seqs=src_word,
src_seqs_lengths=src_lengths,
trg_seqs=trg_x,
trg_gender=trg_gender,
teacher_forcing_prob=teacher_forcing_prob)
# CrossEntropysLoss accepts matrices always!
# the preds must be of size (N, C) where C is the number
# of classes and N is the number of samples.
# The ground truth must be a Vector of size C!
preds = preds.contiguous().view(-1, preds.shape[-1])
trg_y = trg_y.view(-1)
loss = criterion(preds, trg_y)
epoch_loss += loss.item()
# Backprop
loss.backward()
# Gradient clipping
clip_grad_norm_(model.parameters(), max_norm=clip_grad)
# Optimizer step
optimizer.step()
return epoch_loss / len(dataloader)
def evaluate(model, dataloader, criterion, device='cpu', teacher_forcing_prob=0):
model.eval()
epoch_loss = 0
with torch.no_grad():
for batch in dataloader:
batch = {k: v.to(device) for k, v in batch.items()}
src_char = batch['src_char']
src_word = batch['src_word']
trg_x = batch['trg_x']
trg_y = batch['trg_y']
src_lengths = batch['src_lengths']
trg_gender = batch['trg_gender']
preds, attention_scores = model(char_src_seqs=src_char,
word_src_seqs=src_word,
src_seqs_lengths=src_lengths,
trg_seqs=trg_x,
trg_gender=trg_gender,
teacher_forcing_prob=teacher_forcing_prob)
# CrossEntropyLoss accepts matrices always!
# the preds must be of size (N, C) where C is the number
# of classes and N is the number of samples.
# The ground truth must be a Vector of size C!
preds = preds.contiguous().view(-1, preds.shape[-1])
trg_y = trg_y.view(-1)
loss = criterion(preds, trg_y)
epoch_loss += loss.item()
return epoch_loss / len(dataloader)
def inference(sampler, beam_sampler, dataloader, args):
output_inf_file = open(args.preds_dir + '.inf', mode='w', encoding='utf8')
output_beam_g = open(args.preds_dir + '.beam_greedy', mode='w', encoding='utf8')
output_beam = open(args.preds_dir + '.beam', mode='w', encoding='utf8')
greedy_stats = {}
beam_stats = {}
greedy_accuracy = 0
beam_accuracy = 0
for batch in dataloader:
sampler.set_batch(batch)
src = sampler.get_src_sentence(0)
trg = sampler.get_trg_sentence(0)
src_label = sampler.get_src_label(0)
trg_label = sampler.get_trg_label(0)
if args.embed_trg_gender:
trg_gender = sampler.get_trg_gender(0)
else:
trg_gender = None
translated = sampler.greedy_decode(sentence=src, trg_gender=trg_gender)
beam_trans_10 = beam_sampler.beam_decode(sentence=src, trg_gender=trg_gender, topk=1, beam_width=10, max_len=512)
beam_trans_1 = beam_sampler.beam_decode(sentence=src, trg_gender=trg_gender, topk=1, beam_width=1, max_len=512)
greedy_accuracy += accuracy(trg=trg, pred=translated)
beam_accuracy += accuracy(trg=trg, pred=beam_trans_10)
correct = 'CORRECT!' if trg == translated else 'INCORRECT!'
different_g = 'SAME!' if translated == beam_trans_1 else 'DIFF!'
different_10 = 'SAME!' if translated == beam_trans_10 else 'DIFF!'
if beam_trans_1 == trg:
greedy_stats[(src_label, trg_label, 'correct')] = 1 + greedy_stats.get((src_label, trg_label, 'correct'), 0)
else:
greedy_stats[(src_label, trg_label, 'incorrect')] = 1 + greedy_stats.get((src_label, trg_label, 'incorrect'), 0)
if beam_trans_10 == trg:
beam_stats[(src_label, trg_label, 'correct')] = 1 + beam_stats.get((src_label, trg_label, 'correct'), 0)
else:
beam_stats[(src_label, trg_label, 'incorrect')] = 1 + beam_stats.get((src_label, trg_label, 'incorrect'), 0)
output_inf_file.write(translated)
output_inf_file.write('\n')
output_beam_g.write(beam_trans_1)
output_beam_g.write('\n')
output_beam.write(beam_trans_10)
output_beam.write('\n')
logger.info(f'src:\t\t\t{src}')
logger.info(f'trg:\t\t\t{trg}')
logger.info(f'greedy:\t\t\t{translated}')
logger.info(f'beam:\t\t\t{beam_trans_10}')
logger.info(f'src label:\t\t{src_label}')
logger.info(f'trg label:\t\t{trg_label}')
if trg_gender:
logger.info(f'trg gender:\t\t{trg_gender}')
logger.info(f'res:\t\t\t{correct}')
logger.info(f'beam==greedy?:\t\t{different_10}')
logger.info('\n\n')
greedy_accuracy /= len(dataloader)
beam_accuracy /= len(dataloader)
output_inf_file.close()
output_beam_g.close()
output_beam.close()
logger.info('*******STATS*******')
assert sum([greedy_stats[x] for x in greedy_stats]) == sum([beam_stats[x] for x in beam_stats])
total_examples = sum([greedy_stats[x] for x in greedy_stats])
logger.info(f'TOTAL EXAMPLES: {total_examples}')
logger.info('\n')
correct_greedy = {(x[0], x[1]): greedy_stats[x] for x in greedy_stats if x[2] == 'correct'}
incorrect_greedy = {(x[0], x[1]): greedy_stats[x] for x in greedy_stats if x[2] == 'incorrect'}
total_correct_greedy = sum([v for k,v in correct_greedy.items()])
total_incorrect_greedy = sum([v for k, v in incorrect_greedy.items()])
logger.info('Results using greedy decoding:')
for x in correct_greedy:
logger.info(f'{x[0]}->{x[1]}')
logger.info(f'\tCorrect: {correct_greedy.get(x, 0)}\tIncorrect: {incorrect_greedy.get(x, 0)}')
logger.info(f'--------------------------------')
logger.info(f'Total Correct: {total_correct_greedy}\tTotal Incorrect: {total_incorrect_greedy}')
logger.info(f'Accuracy:\t{greedy_accuracy}')
logger.info('\n')
correct_beam = {(x[0], x[1]): beam_stats[x] for x in beam_stats if x[2] == 'correct'}
incorrect_beam = {(x[0], x[1]): beam_stats[x] for x in beam_stats if x[2] == 'incorrect'}
total_correct_beam = sum([v for k, v in correct_beam.items()])
total_incorrect_beam = sum([v for k, v in incorrect_beam.items()])
logger.info('Results using beam decoding:')
for x in correct_beam:
logger.info(f'{x[0]}->{x[1]}')
logger.info(f'\tCorrect: {correct_beam.get(x, 0)}\tIncorrect: {incorrect_beam.get(x, 0)}')
logger.info(f'--------------------------------')
logger.info(f'Total Correct: {total_correct_beam}\tTotal Incorrect: {total_incorrect_beam}')
logger.info(f'Accuracy:\t{beam_accuracy}')
def get_morph_features(args, data, word_vocab):
morph_featurizer = MorphFeaturizer(args.analyzer_db_path)
if args.reload_files:
morph_featurizer.load_morph_features(args.morph_features_path)
else:
morph_featurizer.featurize_sentences(data)
if args.cache_files:
morph_featurizer.save_morph_features(args.morph_features_path)
morph_embeddings = morph_featurizer.create_morph_embeddings(word_vocab)
return morph_embeddings
def load_fasttext_embeddings(args, vocab):
set_seed(args.seed, args.use_cuda)
fasttext_wv = KeyedVectors.load(args.fasttext_embeddings_kv_path, mmap='r')
pretrained_embeddings = torch.zeros((len(vocab), fasttext_wv.vector_size), dtype=torch.float32)
oov = 0
unks = list()
for word, index in vocab.token_to_idx.items():
if word in fasttext_wv.vocab:
pretrained_embeddings[index] = torch.tensor(fasttext_wv[word], dtype=torch.float32)
else:
oov += 1
unks.append(word)
print(f'# Vocab not in the Embeddings: {oov}', flush=True)
print(unks, flush=True)
return pretrained_embeddings
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the src and trg files."
)
parser.add_argument(
"--vectorizer_path",
default=None,
type=str,
help="The path of the saved vectorizer"
)
parser.add_argument(
"--cache_files",
action="store_true",
help="Whether to cache the vocab and the vectorizer objects or not"
)
parser.add_argument(
"--reload_files",
action="store_true",
help="Whether to reload the vocab and the vectorizer objects from a cached file"
)
parser.add_argument(
"--num_train_epochs",
default=20,
type=int,
help="Total number of training epochs to perform."
)
parser.add_argument(
"--embedding_dim",
default=32,
type=int,
help="The embedding dimensions of the model"
)
parser.add_argument(
"--trg_gender_embedding_dim",
default=0,
type=int,
help="The embedding dimensions of the target gender"
)
parser.add_argument(
"--hidd_dim",
default=64,
type=int,
help="The hidden dimensions of the model"
)
parser.add_argument(
"--num_layers",
default=1,
type=int,
help="The numbers of layers of the model"
)
parser.add_argument(
"--dropout",
default=0.0,
type=float,
help="Dropout rate."
)
parser.add_argument(
"--weight_decay",
default=0.0,
type=float,
help="Optimizer weight decay"
)
parser.add_argument(
"--learning_rate",
default=5e-4,
type=float,
help="The initial learning rate for Adam."
)
parser.add_argument(
"--clip_grad",
default=1.0,
type=float,
help="Gradient clipping norm."
)
parser.add_argument(
"--batch_size",
default=32,
type=int,
help="Batch size per GPU/CPU"
)
parser.add_argument(
"--use_cuda",
action="store_true",
help="Whether to use the gpu or not."
)
parser.add_argument(
"--seed",
default=21,
type=int,
help="Random seed."
)
parser.add_argument(
"--model_path",
type=str,
required=True,
default=None,
help="The directory of the model."
)
parser.add_argument(
"--do_train",
action="store_true",
help="Whether to run training or not."
)
parser.add_argument(
"--do_eval",
action="store_true",
help="Whether to run eval or not."
)
parser.add_argument(
"--visualize_loss",
action="store_true",
help="Whether to visualize the loss during training and evaluation."
)
parser.add_argument(
"--do_inference",
action="store_true",
help="Whether to do inference or not."
)
parser.add_argument(
"--inference_mode",
type=str,
default="dev",
help="The dataset to do inference on."
)
parser.add_argument(
"--use_morph_features",
action="store_true",
help="Whether to use morphological features or not."
)
parser.add_argument(
"--use_fasttext_embeddings",
action="store_true",
help="Whether to use fasttext embeddings or not."
)
parser.add_argument(
"--fasttext_embeddings_kv_path",
type=str,
default=None,
help="The path to the pretrained fasttext embeddings keyedvectors."
)
parser.add_argument(
"--embed_trg_gender",
action="store_true",
help="Whether to embed the target gender or not."
)
parser.add_argument(
"--one_hot_trg_gender",
action="store_true",
help="Whether to embed the target gender in a zero hot fashion."
)
parser.add_argument(
"--analyzer_db_path",
type=str,
default=None,
help="Path to the anaylzer database."
)
parser.add_argument(
"--morph_features_path",
type=str,
default=None,
help="The path of the saved morphological features."
)
parser.add_argument(
"--preds_dir",
type=str,
default=None,
help="The directory to write the translations to"
)
args = parser.parse_args()
device = torch.device('cuda' if args.use_cuda else 'cpu')
set_seed(args.seed, args.use_cuda)
if args.reload_files:
dataset = MT_Dataset.load_data_and_load_vectorizer(args.data_dir, args.vectorizer_path)
else:
dataset = MT_Dataset.load_data_and_create_vectorizer(args.data_dir)
vectorizer = dataset.get_vectorizer()
if args.cache_files:
dataset.save_vectorizer(args.vectorizer_path)
if args.use_morph_features:
# we create morph features on the src side of the
# training data
logger.info(f'Loading Gender Morph Features...')
train_src_data = [t.src for t in dataset.train_examples]
morph_embeddings = get_morph_features(args, train_src_data, vectorizer.src_vocab_word)
else:
morph_embeddings = None
if args.use_fasttext_embeddings:
logger.info(f'Loading FastText Embeddings...')
fasttext_embeddings = load_fasttext_embeddings(args, vectorizer.src_vocab_word)
else:
fasttext_embeddings = None
if args.one_hot_trg_gender:
logger.info(f'Creating one hot gender embeddings...')
gender_embeddings = create_gender_embeddings(vectorizer.trg_gender_vocab)
print(vectorizer.trg_gender_vocab.token_to_idx, flush=True)
print(gender_embeddings, flush=True)
else:
gender_embeddings = None
ENCODER_INPUT_DIM = len(vectorizer.src_vocab_char)
DECODER_INPUT_DIM = len(vectorizer.trg_vocab_char)
DECODER_OUTPUT_DIM = len(vectorizer.trg_vocab_char)
DECODER_TRG_GEN_INPUT_DIM = len(vectorizer.trg_gender_vocab)
CHAR_SRC_PAD_INDEX = vectorizer.src_vocab_char.pad_idx
WORD_SRC_PAD_INDEX = vectorizer.src_vocab_word.pad_idx
TRG_PAD_INDEX = vectorizer.trg_vocab_char.pad_idx
TRG_SOS_INDEX = vectorizer.trg_vocab_char.sos_idx
model = Seq2Seq(encoder_input_dim=ENCODER_INPUT_DIM,
encoder_embed_dim=args.embedding_dim,
encoder_hidd_dim=args.hidd_dim,
encoder_num_layers=args.num_layers,
decoder_input_dim=DECODER_INPUT_DIM,
decoder_embed_dim=args.embedding_dim,
decoder_hidd_dim=args.hidd_dim,
decoder_num_layers=args.num_layers,
decoder_output_dim=DECODER_OUTPUT_DIM,
morph_embeddings=morph_embeddings,
fasttext_embeddings=fasttext_embeddings,
embed_trg_gender=args.embed_trg_gender,
gender_input_dim=DECODER_TRG_GEN_INPUT_DIM,
gender_embed_dim=args.trg_gender_embedding_dim,
gender_embeddings=gender_embeddings,
char_src_padding_idx=CHAR_SRC_PAD_INDEX,
word_src_padding_idx=WORD_SRC_PAD_INDEX,
trg_padding_idx=TRG_PAD_INDEX,
trg_sos_idx=TRG_SOS_INDEX,
dropout=args.dropout)
# Optimizer
optimizer = optim.Adam(model.parameters(), lr=args.learning_rate, weight_decay=args.weight_decay)
# Loss function
criterion = nn.CrossEntropyLoss(ignore_index=TRG_PAD_INDEX)
# lr scheduler
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min',
patience=2, factor=0.5)
collator = Collator(char_src_pad_idx=CHAR_SRC_PAD_INDEX,
char_trg_pad_idx=TRG_PAD_INDEX,
word_src_pad_idx=WORD_SRC_PAD_INDEX)
model = model.to(device)
if args.do_train:
logger.info('Training...')
train_losses = []
dev_losses = []
best_loss = 1e10
teacher_forcing_prob = 0.3
clip_grad = args.clip_grad
set_seed(args.seed, args.use_cuda)
for epoch in range(args.num_train_epochs):
dataset.set_split('train')
dataloader = DataLoader(dataset, shuffle=True, batch_size=args.batch_size, collate_fn=collator, drop_last=True)
train_loss = train(model, dataloader, optimizer, criterion, device, teacher_forcing_prob=teacher_forcing_prob, clip_grad=clip_grad)
train_losses.append(train_loss)
dataset.set_split('dev')
dataloader = DataLoader(dataset, shuffle=True, batch_size=args.batch_size, collate_fn=collator, drop_last=True)
dev_loss = evaluate(model, dataloader, criterion, device, teacher_forcing_prob=0)
dev_losses.append(dev_loss)
#save best model
if dev_loss < best_loss:
best_loss = dev_loss
torch.save(model.state_dict(), args.model_path)
scheduler.step(dev_loss)
logger.info(f'Epoch: {(epoch + 1)}')
logger.info(f'\tTrain Loss: {train_loss:.4f} | Dev Loss: {dev_loss:.4f}')
if args.do_train and args.visualize_loss:
plt.plot(range(1, 1 + args.num_train_epochs), np.asarray(train_losses), 'b-', color='blue', label='Training')
plt.plot(range(1, 1 + args.num_train_epochs), np.asarray(dev_losses), 'b-', color='orange', label='Evaluation')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()
plt.savefig(args.model_path + '.loss.png')
if args.do_eval:
logger.info('Evaluation')
set_seed(args.seed, args.use_cuda)
dev_losses = []
for epoch in range(args.num_train_epochs):
dataset.set_split('dev')
dataloader = DataLoader(dataset, shuffle=True, batch_size=args.batch_size, collate_fn=collator)
dev_loss = evaluate(model, dataloader, criterion, device, teacher_forcing_prob=0)
dev_losses.append(dev_loss)
logger.info(f'Dev Loss: {dev_loss:.4f}')
if args.do_inference:
logger.info('Inference')
set_seed(args.seed, args.use_cuda)
model.load_state_dict(torch.load(args.model_path))
model.eval()
model = model.to(device)
dataset.set_split(args.inference_mode)
dataloader = DataLoader(dataset, batch_size=1, shuffle=False, collate_fn=collator)
sampler = BatchSampler(model=model,
src_vocab_char=vectorizer.src_vocab_char,
src_vocab_word=vectorizer.src_vocab_word,
trg_vocab_char=vectorizer.trg_vocab_char,
src_labels_vocab=vectorizer.src_labels_vocab,
trg_labels_vocab=vectorizer.trg_labels_vocab,
trg_gender_vocab=vectorizer.trg_gender_vocab)
beam_sampler = BeamSampler(model=model,
src_vocab_char=vectorizer.src_vocab_char,
src_vocab_word=vectorizer.src_vocab_word,
trg_vocab_char=vectorizer.trg_vocab_char,
src_labels_vocab=vectorizer.src_labels_vocab,
trg_labels_vocab=vectorizer.trg_labels_vocab,
trg_gender_vocab=vectorizer.trg_gender_vocab)
inference(sampler, beam_sampler, dataloader, args)
if __name__ == "__main__":
main()