-
Notifications
You must be signed in to change notification settings - Fork 35
/
train_scpn.py
executable file
·740 lines (590 loc) · 31.1 KB
/
train_scpn.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
import torch, time, argparse, os, codecs, h5py, cPickle, random
import numpy as np
from torch import nn
from torch import optim
from torch.autograd import Variable
from collections import OrderedDict
from torch.nn.utils.rnn import pad_packed_sequence as unpack
from torch.nn.utils.rnn import pack_padded_sequence as pack
from scpn_utils import *
reload(sys)
sys.setdefaultencoding('utf8')
# seq2seq w/ decoder attention
# transformation embeddings concatenated with decoder word inputs
# attention conditioned on transformation via bilinear product
class SCPN(nn.Module):
def __init__(self, d_word, d_hid, d_nt, d_trans,
len_voc, len_trans_voc, use_input_parse):
super(SCPN, self).__init__()
self.d_word = d_word
self.d_hid = d_hid
self.d_trans = d_trans
self.d_nt = d_nt + 1
self.len_voc = len_voc
self.len_trans_voc = len_trans_voc
self.use_input_parse = use_input_parse
# embeddings
self.word_embs = nn.Embedding(len_voc, d_word)
self.trans_embs = nn.Embedding(len_trans_voc, d_nt)
# lstms
if use_input_parse:
self.encoder = nn.LSTM(d_word + d_trans, d_hid, num_layers=1, bidirectional=True, batch_first=True)
else:
self.encoder = nn.LSTM(d_word, d_hid, num_layers=1, bidirectional=True, batch_first=True)
self.encoder_proj = nn.Linear(d_hid * 2, d_hid)
self.decoder = nn.LSTM(d_word + d_hid, d_hid, num_layers=2, batch_first=True)
self.trans_encoder = nn.LSTM(d_nt, d_trans, num_layers=1, batch_first=True)
self.trans_hid_init = Variable(torch.zeros(1, 1, d_trans).cuda())
self.trans_cell_init = Variable(torch.zeros(1, 1, d_trans).cuda())
self.e_hid_init = Variable(torch.zeros(2, 1, d_hid).cuda())
self.e_cell_init = Variable(torch.zeros(2, 1, d_hid).cuda())
self.d_cell_init = Variable(torch.zeros(2, 1, d_hid).cuda())
# output softmax
self.out_dense_1 = nn.Linear(d_hid * 2, d_hid)
self.out_dense_2 = nn.Linear(d_hid, len_voc)
self.att_nonlin = nn.Softmax()
self.out_nonlin = nn.LogSoftmax()
# attention params
self.att_parse_proj = nn.Linear(d_trans, d_hid)
self.att_W = nn.Parameter(torch.Tensor(d_hid, d_hid).cuda())
self.att_parse_W = nn.Parameter(torch.Tensor(d_hid, d_hid).cuda())
nn.init.xavier_uniform(self.att_parse_W.data)
nn.init.xavier_uniform(self.att_W.data)
# copy prob params
self.copy_hid_v = nn.Parameter(torch.Tensor(d_hid, 1).cuda())
self.copy_att_v = nn.Parameter(torch.Tensor(d_hid, 1).cuda())
self.copy_inp_v = nn.Parameter(torch.Tensor(d_word + d_hid, 1).cuda())
nn.init.xavier_uniform(self.copy_hid_v.data)
nn.init.xavier_uniform(self.copy_att_v.data)
nn.init.xavier_uniform(self.copy_inp_v.data)
# create matrix mask from length vector
def compute_mask(self, lengths):
max_len = torch.max(lengths)
range_row = torch.arange(0, max_len).long().cuda()[None, :].expand(lengths.size()[0], max_len)
mask = lengths[:, None].expand_as(range_row)
mask = range_row < mask
return Variable(mask.float().cuda())
# masked softmax for attention
def masked_softmax(self, vector, mask):
result = torch.nn.functional.softmax(vector)
result = result * mask
result = result / (result.sum(dim=1, keepdim=True) + 1e-13)
return result
# compute masked attention over enc hiddens with bilinear product
def compute_decoder_attention(self, hid_previous, enc_hids, in_lens):
mask = self.compute_mask(in_lens)
b_hn = hid_previous.mm(self.att_W)
scores = b_hn[:, None, :] * enc_hids
scores = torch.sum(scores, 2)
scores = self.masked_softmax(scores, mask)
return scores
# compute masked attention over parse sequence with bilinear product
def compute_transformation_attention(self, hid_previous, trans_embs, trans_lens):
mask = self.compute_mask(trans_lens)
b_hn = hid_previous.mm(self.att_parse_W)
scores = b_hn[:, None, :] * trans_embs
scores = torch.sum(scores, 2)
scores = self.masked_softmax(scores, mask)
return scores
# return encoding for an input batch
def encode_batch(self, inputs, trans, lengths):
bsz, max_len = inputs.size()
in_embs = self.word_embs(inputs)
lens, indices = torch.sort(lengths, 0, True)
# concat word embs with trans hid
if self.use_input_parse:
in_embs = torch.cat([in_embs, trans.unsqueeze(1).expand(bsz, max_len, self.d_trans)], 2)
e_hid_init = self.e_hid_init.expand(2, bsz, self.d_hid).contiguous()
e_cell_init = self.e_cell_init.expand(2, bsz, self.d_hid).contiguous()
all_hids, (enc_last_hid, _) = self.encoder(pack(in_embs[indices],
lens.tolist(), batch_first=True), (e_hid_init, e_cell_init))
_, _indices = torch.sort(indices, 0)
all_hids = unpack(all_hids, batch_first=True)[0][_indices]
all_hids = self.encoder_proj(all_hids.view(-1, self.d_hid * 2)).view(bsz, max_len, self.d_hid)
enc_last_hid = torch.cat([enc_last_hid[0], enc_last_hid[1]], 1)
enc_last_hid = self.encoder_proj(enc_last_hid)[_indices]
return all_hids, enc_last_hid
# return encoding for an input batch
def encode_transformations(self, trans, lengths, return_last=True):
bsz, _ = trans.size()
lens, indices = torch.sort(lengths, 0, True)
in_embs = self.trans_embs(trans)
t_hid_init = self.trans_hid_init.expand(1, bsz, self.d_trans).contiguous()
t_cell_init = self.trans_cell_init.expand(1, bsz, self.d_trans).contiguous()
all_hids, (enc_last_hid, _) = self.trans_encoder(pack(in_embs[indices],
lens.tolist(), batch_first=True), (t_hid_init, t_cell_init))
_, _indices = torch.sort(indices, 0)
if return_last:
return enc_last_hid.squeeze(0)[_indices]
else:
all_hids = unpack(all_hids, batch_first=True)[0]
return all_hids[_indices]
# decode one timestep
def decode_step(self, idx, prev_words, prev_hid, prev_cell,
enc_hids, trans_embs, in_sent_lens, trans_lens, bsz, max_len):
# initialize with zeros
if idx == 0:
word_input = Variable(torch.zeros(bsz, 1, self.d_word).cuda())
# get previous ground truth word embed and concat w/ transformation emb
else:
word_input = self.word_embs(prev_words)
word_input = word_input.view(bsz, 1, self.d_word)
# concatenate w/ transformation embeddings
trans_weights = self.compute_transformation_attention(prev_hid[1], trans_embs, trans_lens)
trans_ctx = torch.sum(trans_weights[:, :, None] * trans_embs, 1)
decoder_input = torch.cat([word_input, trans_ctx.unsqueeze(1)], 2)
# feed to decoder lstm
_, (hn, cn) = self.decoder(decoder_input, (prev_hid, prev_cell))
# compute attention for next time step and att weighted ave of encoder hiddens
attn_weights = self.compute_decoder_attention(hn[1], enc_hids, in_sent_lens)
attn_ctx = torch.sum(attn_weights[:, :, None] * enc_hids, 1)
# compute copy prob as function of lotsa shit
p_copy = decoder_input.squeeze(1).mm(self.copy_inp_v)
p_copy += attn_ctx.mm(self.copy_att_v)
p_copy += hn[1].mm(self.copy_hid_v)
p_copy = torch.sigmoid(p_copy).squeeze(1)
return hn, cn, attn_weights, attn_ctx, p_copy
def forward(self, inputs, outputs, in_trans, out_trans,
in_sent_lens, out_sent_lens, in_trans_lens, out_trans_lens, max_decode):
bsz, max_len = inputs.size()
# encode transformations
in_trans_hids = None
if self.use_input_parse:
in_trans_hids = self.encode_transformations(in_trans, in_trans_lens)
out_trans_hids = self.encode_transformations(out_trans, out_trans_lens, return_last=False)
out_trans_hids = self.att_parse_proj(out_trans_hids)
# encode input sentence
enc_hids, enc_last_hid = self.encode_batch(inputs, in_trans_hids, in_sent_lens)
# store decoder hiddens and attentions for copy
decoder_states = Variable(torch.zeros(max_decode, bsz, self.d_hid * 2).cuda())
decoder_copy_dists = Variable(torch.zeros(max_decode, bsz, self.len_voc).cuda())
copy_probs = Variable(torch.zeros(max_decode, bsz).cuda())
# initialize decoder hidden to last encoder hidden
hn = enc_last_hid.unsqueeze(0).expand(2, bsz, self.d_hid).contiguous()
cn = self.d_cell_init.expand(2, bsz, self.d_hid).contiguous()
# loop til max_decode, do lstm tick using previous prediction
for idx in range(max_decode):
prev_words = None
if idx > 0:
prev_words = outputs[:, idx - 1]
# concat prev word emb and trans emb and feed to decoder lstm
hn, cn, attn_weights, attn_ctx, p_copy = self.decode_step(idx, prev_words,
hn, cn, enc_hids, out_trans_hids, in_sent_lens, out_trans_lens, bsz, max_len)
# compute copy attn by scattering attn into vocab space
vocab_scores = Variable(torch.zeros(bsz, self.len_voc).cuda())
vocab_scores = vocab_scores.scatter_add_(1, inputs, attn_weights)
# store decoder hiddens and copy probs in log domain
decoder_states[idx] = torch.cat([hn[1], attn_ctx], 1)
decoder_copy_dists[idx] = torch.log(vocab_scores + 1e-20)
copy_probs[idx] = p_copy
# now do prediction over decoder states (reshape to 2d first)
decoder_states = decoder_states.transpose(0, 1).contiguous().view(-1, self.d_hid * 2)
decoder_preds = self.out_dense_1(decoder_states)
decoder_preds = self.out_dense_2(decoder_preds)
decoder_preds = self.out_nonlin(decoder_preds)
decoder_copy_dists = decoder_copy_dists.transpose(0, 1).contiguous().view(-1, self.len_voc)
# merge copy dist and pred dist using copy probs
copy_probs = copy_probs.view(-1)
final_dists = copy_probs[:, None] * decoder_copy_dists + \
(1. - copy_probs[:, None]) * decoder_preds
return final_dists
# beam search given a single input / transformation
def beam_search(self, inputs, in_trans, out_trans, in_sent_lens, in_trans_lens,
out_trans_lens, eos_idx, beam_size=4, max_steps=40):
bsz, max_len = inputs.size()
# chop input
inputs = inputs[:, :in_sent_lens[0]]
# encode transformations
in_trans_hids = None
if self.use_input_parse:
in_trans_hids = self.encode_transformations(in_trans, in_trans_lens)
out_trans_hids = self.encode_transformations(out_trans, out_trans_lens, return_last=False)
out_trans_hids = self.att_parse_proj(out_trans_hids)
# encode input sentence
enc_hids, enc_last_hid = self.encode_batch(inputs, in_trans_hids, in_sent_lens)
# initialize decoder hidden to last encoder hidden
hn = enc_last_hid.unsqueeze(0).expand(2, bsz, self.d_hid).contiguous()
cn = self.d_cell_init.expand(2, bsz, self.d_hid).contiguous()
# initialize beams
beams = [(0.0, hn, cn, [])]
nsteps = 0
# loop til max_decode, do lstm tick using previous prediction
while True:
# loop over everything in the beam
beam_candidates = []
for b in beams:
curr_prob, prev_h, prev_c, seq = b
# start with last word in sequence, if eos end the beam
if len(seq) > 0:
prev_word = seq[-1]
if prev_word == eos_idx:
beam_candidates.append(b)
continue
# load into torch var so we can do decoding
prev_word = Variable(torch.LongTensor([prev_word]).cuda())
else:
prev_word = None
# concat prev word emb and prev attn input and feed to decoder lstm
hn, cn, attn_weights, attn_ctx, p_copy = self.decode_step(len(seq), prev_word,
prev_h, prev_c, enc_hids, out_trans_hids, in_sent_lens, out_trans_lens, bsz, max_len)
# compute copy attn by scattering attn into vocab space
vocab_scores = Variable(torch.zeros(bsz, self.len_voc).cuda())
vocab_scores = vocab_scores.scatter_add_(1, inputs, attn_weights)
vocab_scores = torch.log(vocab_scores + 1e-20).squeeze()
# compute prediction over vocab for a single time step
pred_inp = torch.cat([hn[1], attn_ctx], 1)
preds = self.out_dense_1(pred_inp)
preds = self.out_dense_2(preds)
preds = self.out_nonlin(preds).squeeze()
final_preds = p_copy * vocab_scores + (1 - p_copy) * preds
# sort in descending order (log domain)
_, top_indices = torch.sort(-final_preds)
# add top n candidates
for z in range(beam_size):
word_idx = top_indices[z].data[0]
beam_candidates.append((curr_prob + final_preds[word_idx].data[0],
hn, cn, seq + [word_idx]))
beam_candidates.sort(reverse=True)
beams = beam_candidates[:beam_size]
nsteps += 1
if nsteps > max_steps:
break
return beams
# beam search given a single sentence and a batch of transformations
def batch_beam_search(self, inputs, out_trans, in_sent_lens,
out_trans_lens, eos_idx, beam_size=5, max_steps=70):
bsz, max_len = inputs.size()
# chop input
inputs = inputs[:, :in_sent_lens[0]]
# encode transformations
out_trans_hids = self.encode_transformations(out_trans, out_trans_lens, return_last=False)
out_trans_hids = self.att_parse_proj(out_trans_hids)
# encode input sentence
enc_hids, enc_last_hid = self.encode_batch(inputs, None, in_sent_lens)
# initialize decoder hidden to last encoder hidden
hn = enc_last_hid.unsqueeze(0).expand(2, bsz, self.d_hid).contiguous()
cn = self.d_cell_init
# initialize beams (dictionary of batch_idx: beam params)
beam_dict = OrderedDict()
for b_idx in range(out_trans.size()[0]):
beam_dict[b_idx] = [(0.0, hn, cn, [])]
nsteps = 0
# loop til max_decode, do lstm tick using previous prediction
while True:
# set up accumulators for predictions
# assumption: all examples have same number of beams at each timestep
prev_words = []
prev_hs = []
prev_cs = []
for b_idx in beam_dict:
beams = beam_dict[b_idx]
# loop over everything in the beam
beam_candidates = []
for b in beams:
curr_prob, prev_h, prev_c, seq = b
# start with last word in sequence, if eos end the beam
if len(seq) > 0:
prev_words.append(seq[-1])
else:
prev_words = None
prev_hs.append(prev_h)
prev_cs.append(prev_c)
# now batch decoder computations
hs = torch.cat(prev_hs, 1)
cs = torch.cat(prev_cs, 1)
num_examples = hs.size()[1]
if prev_words is not None:
prev_words = Variable(torch.from_numpy(np.array(prev_words, dtype='int32')).long().cuda())
# expand out parse states if necessary
if num_examples != out_trans_hids.size()[0]:
d1, d2, d3 = out_trans_hids.size()
rep_factor = num_examples / d1
curr_out = out_trans_hids.unsqueeze(1).expand(d1, rep_factor,
d2, d3).contiguous().view(-1, d2, d3)
curr_out_lens = out_trans_lens.unsqueeze(1).expand(d1, rep_factor).contiguous().view(-1)
else:
curr_out = out_trans_hids
curr_out_lens = out_trans_lens
# expand out inputs and encoder hiddens
_, in_len, hid_d = enc_hids.size()
curr_enc_hids = enc_hids.expand(num_examples, in_len, hid_d)
curr_enc_lens = in_sent_lens.expand(num_examples)
curr_inputs = inputs.expand(num_examples, in_sent_lens[0])
# concat prev word emb and prev attn input and feed to decoder lstm
hn, cn, attn_weights, attn_ctx, p_copy = self.decode_step(nsteps, prev_words,
hs, cs, curr_enc_hids, curr_out, curr_enc_lens, curr_out_lens, num_examples, max_len)
# compute copy attn by scattering attn into vocab space
vocab_scores = Variable(torch.zeros(num_examples, self.len_voc).cuda())
vocab_scores = vocab_scores.scatter_add_(1, curr_inputs, attn_weights)
vocab_scores = torch.log(vocab_scores + 1e-20).squeeze()
# compute prediction over vocab for a single time step
pred_inp = torch.cat([hn[1], attn_ctx], 1)
preds = self.out_dense_1(pred_inp)
preds = self.out_dense_2(preds)
preds = self.out_nonlin(preds).squeeze()
final_preds = p_copy[:, None] * vocab_scores + (1 - p_copy[:, None]) * preds
# now loop over the examples and sort each separately
for b_idx in beam_dict:
beam_candidates = []
# no words previously predicted
if num_examples == len(beam_dict):
ex_hn = hn[:,b_idx,:].unsqueeze(1)
ex_cn = cn[:,b_idx,:].unsqueeze(1)
preds = final_preds[b_idx]
_, top_indices = torch.sort(-preds)
# add top n candidates
for z in range(beam_size):
word_idx = top_indices[z].data[0]
beam_candidates.append((preds[word_idx].data[0],
ex_hn, ex_cn, [word_idx]))
beam_dict[b_idx] = beam_candidates
else:
origin_beams = beam_dict[b_idx]
start = b_idx * beam_size
end = (b_idx + 1) * beam_size
ex_hn = hn[:,start:end,:]
ex_cn = cn[:,start:end,:]
ex_preds = final_preds[start:end]
for o_idx, ob in enumerate(origin_beams):
curr_prob, _, _, seq = ob
# if one of the beams is already complete, add it to candidates
if seq[-1] == eos_idx:
beam_candidates.append(ob)
preds = ex_preds[o_idx]
curr_hn = ex_hn[:,o_idx,:]
curr_cn = ex_cn[:,o_idx,:]
_, top_indices = torch.sort(-preds)
for z in range(beam_size):
word_idx = top_indices[z].data[0]
beam_candidates.append((curr_prob + float(preds[word_idx].cpu().data[0]),
curr_hn.unsqueeze(1), curr_cn.unsqueeze(1), seq + [word_idx]))
s_inds = np.argsort([x[0] for x in beam_candidates])[::-1]
beam_candidates = [beam_candidates[x] for x in s_inds]
beam_dict[b_idx] = beam_candidates[:beam_size]
nsteps += 1
if nsteps > max_steps:
break
return beam_dict
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Syntactically Controlled Paraphrase Network')
parser.add_argument('--gpu', type=str, default='0',
help='GPU id')
parser.add_argument('--data', type=str, default='data/parsed_data.h5',
help='hdf5 location')
parser.add_argument('--vocab', type=str, default='data/parse_vocab.pkl',
help='word vocabulary')
parser.add_argument('--parse_vocab', type=str, default='data/ptb_tagset.txt',
help='tag vocabulary')
parser.add_argument('--model', type=str, default='scpn2.pt',
help='model save path')
parser.add_argument('--batch_size', type=int, default=64,
help='batch size')
parser.add_argument('--min_sent_length', type=int, default=5,
help='min number of tokens per batch')
parser.add_argument('--d_word', type=int, default=300,
help='word embedding dimension')
parser.add_argument('--d_trans', type=int, default=128,
help='transformation hidden dimension')
parser.add_argument('--d_nt', type=int, default=56,
help='nonterminal embedding dimension')
parser.add_argument('--d_hid', type=int, default=512,
help='lstm hidden dimension')
parser.add_argument('--n_epochs', type=int, default=15,
help='n_epochs')
parser.add_argument('--lr', type=float, default=0.00005,
help='learning rate')
parser.add_argument('--grad_clip', type=float, default=5.0,
help='clip if grad norm exceeds this threshold')
parser.add_argument('--save_freq', type=int, default=500,
help='how many minibatches to save model')
parser.add_argument('--lr_decay_factor', type=int, default=0.5,
help='how much to decrease LR every epoch')
parser.add_argument('--eval_mode', type=bool, default=False,
help='run beam search for some examples using a trained model')
parser.add_argument('--init_trained_model', type=int, default=0,
help='continue training a cached model')
parser.add_argument('--tree_dropout', type=float, default=0.,
help='dropout rate for dropping tree terminals')
parser.add_argument('--tree_level_dropout', type=float, default=0.,
help='dropout rate for dropping entire levels of a tree')
parser.add_argument('--short_batch_downsampling_freq', type=float, default=0.0,
help='dropout rate for dropping entire levels of a tree')
parser.add_argument('--short_batch_threshold', type=int, default=20,
help='if sentences are shorter than this, they will be downsampled')
parser.add_argument('--seed', type=int, default=1000,
help='random seed')
parser.add_argument('--use_input_parse', type=int, default=0,
help='whether or not to use the input parse')
parser.add_argument('--dev_batches', type=int, default=200,
help='how many minibatches to use for validation')
args = parser.parse_args()
batch_size = args.batch_size
d_word = args.d_word
d_hid = args.d_hid
d_trans = args.d_trans
d_nt = args.d_nt
n_epochs = args.n_epochs
lr = args.lr
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
out_file = 'models/' + args.model
# set random seed
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
# load data, word vocab, and parse vocab
h5f = h5py.File(args.data, 'r')
inp = h5f['inputs']
out = h5f['outputs']
in_parses = h5f['input_parses']
out_parses = h5f['output_parses']
in_lens = h5f['in_lengths']
out_lens = h5f['out_lengths']
vocab, rev_vocab = \
cPickle.load(open(args.vocab, 'rb'))
tag_file = codecs.open(args.parse_vocab, 'r', 'utf-8')
label_voc = {}
for idx, line in enumerate(tag_file):
line = line.strip()
if line != 'EOP':
label_voc[line] = idx
rev_label_voc = dict((v,k) for (k,v) in label_voc.iteritems())
len_voc = len(vocab)
len_parse_voc = len(label_voc)
max_decode = inp.shape[1]
minibatches = [(start, start + batch_size) \
for start in range(0, inp.shape[0], batch_size)][:-1]
random.shuffle(minibatches)
train_minibatches = minibatches[args.dev_batches:]
dev_minibatches = minibatches[:args.dev_batches]
# build network
net = SCPN(d_word, d_hid, d_nt, d_trans,
len_voc, len_parse_voc, args.use_input_parse)
net.cuda()
# load saved model if evaluating
if args.eval_mode:
saved_model = torch.load(out_file)
net.load_state_dict(saved_model['state_dict'])
train_minibatches = saved_model['trained_minibatches']
dev_batches = saved_model['dev_minibatches']
net.eval()
if args.init_trained_model:
print 'starting from cached model'
net.load_state_dict(torch.load(out_file)['state_dict'])
# ignore zero targets in loss function (they are just padding)
criterion = nn.NLLLoss(ignore_index=0)
params = net.parameters()
optimizer = optim.Adam(params, lr=lr)
if args.init_trained_model:
for param_group in optimizer.param_groups:
param_group['lr'] = param_group['lr'] * args.lr_decay_factor / 2.
print 'new LR:', param_group['lr']
for ep in range(n_epochs):
random.shuffle(train_minibatches)
ep_loss = 0.
start_time = time.time()
num_batches = 0
for b_idx, (start, end) in enumerate(train_minibatches):
# read data from hdf5
in_p = in_parses[start:end]
out_p = out_parses[start:end]
# get valid instances of transformations
z = indexify_transformations(in_p, out_p, label_voc, args)
if z == None:
continue
in_trans_np, out_trans_np, mismatch_inds, in_trans_len_np, out_trans_len_np = z
# only store valid input instances
inp_np = inp[start:end][mismatch_inds]
out_np = out[start:end][mismatch_inds]
in_len_np = in_lens[start:end][mismatch_inds]
out_len_np = out_lens[start:end][mismatch_inds]
curr_bsz = inp_np.shape[0]
# chop input based on length of last instance (for encoder efficiency)
max_in_len = int(in_len_np[-1])
inp_np = inp_np[:, :max_in_len]
# compute max output length and chop output (for decoder efficiency)
max_out_len = int(np.amax(out_len_np))
out_np = out_np[:, :max_out_len]
# sentences are too short
if max_in_len < args.min_sent_length:
continue
# downsample if input sentences are too short
if args.short_batch_downsampling_freq > 0. and max_in_len < args.short_batch_threshold:
if np.random.rand() < args.short_batch_downsampling_freq:
continue
# randomly invert 50% of batches (to remove NMT bias)
swap = random.random() > 0.5
if swap:
inp_np, out_np = out_np, inp_np
in_trans_np, out_trans_np = out_trans_np, in_trans_np
max_in_len, max_out_len = max_out_len, max_in_len
in_len_np, out_len_np = out_len_np, in_len_np
in_trans_len_np, out_trans_len_np = out_trans_len_np, in_trans_len_np
# torchify input
curr_inp = Variable(torch.from_numpy(inp_np.astype('int32')).long().cuda())
curr_out = Variable(torch.from_numpy(out_np.astype('int32')).long().cuda())
in_trans = Variable(torch.from_numpy(in_trans_np).long(98).cuda())
out_trans = Variable(torch.from_numpy(out_trans_np).long().cuda())
in_sent_lens = torch.from_numpy(in_len_np).long().cuda()
out_sent_lens = torch.from_numpy(out_len_np).long().cuda()
in_trans_lens = torch.from_numpy(in_trans_len_np).long().cuda()
out_trans_lens = torch.from_numpy(out_trans_len_np).long().cuda()
# forward prop
preds = net(curr_inp, curr_out, in_trans, out_trans,
in_sent_lens, out_sent_lens, in_trans_lens, out_trans_lens, max_out_len)
num_batches += 1
# if training, compute loss and backprop
if not args.eval_mode:
# compute masked loss
loss = criterion(preds, curr_out.view(-1))
ep_loss += loss.data[0]
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm(params, args.grad_clip)
optimizer.step()
# if training, save model and print some predictions every save_freq minibatches
# if eval, just do beam search on a few instances per minibatch
if args.eval_mode or (b_idx % args.save_freq == 0):
preds = preds.view(curr_bsz, max_out_len, -1).cpu().data.numpy()
preds = np.argmax(preds, -1)
for i in range(min(3, curr_bsz)):
# hack around beam search bug
try:
np_i = in_trans[i].cpu().data.numpy()
np_o = out_trans[i].cpu().data.numpy()
eos = np.where(out_np[i]==vocab['EOS'])[0][0]
print 'swapped:', swap
print 'input: %s' % ' '.join([rev_vocab[w] for (j,w) in enumerate(inp_np[i])\
if j < in_len_np[i]])
print 'gt output: %s' % ' '.join([rev_vocab[w] for (j,w) in enumerate(out_np[i, :eos])\
if j < out_len_np[i]])
print 'input top-level parse: %s' % ' '.join([rev_label_voc[z] for (j,z) in enumerate(np_i)\
if j < in_trans_len_np[i]])
print 'output top-level parse: %s' % ' '.join([rev_label_voc[z] for (j,z) in enumerate(np_o)\
if j < out_trans_len_np[i]])
print 'greedy: %s' % ' '.join([rev_vocab[w] for w in preds[i, :eos]])
beams = net.beam_search(curr_inp[i].unsqueeze(0), in_trans[i].unsqueeze(0),
out_trans[i].unsqueeze(0), in_sent_lens[i:i+1],
in_trans_lens[i:i+1], out_trans_lens[i:i+1], vocab['EOS'])
for beam_idx, beam in enumerate(beams):
print 'gpu beam %d, score:%f: %s' % \
(beam_idx, beam[0], ' '.join([rev_vocab[w] for w in beam[-1]]))
except:
print 'beam search error'
# print statistics about the batch
if not args.eval_mode:
print 'done with batch %d / %d in epoch %d, loss: %f, time:%d\n\n' \
% (b_idx, len(train_minibatches), ep,
ep_loss / num_batches, time.time()-start_time)
torch.save({'state_dict':net.state_dict(),
'ep_loss':ep_loss / num_batches,
'train_minibatches': train_minibatches,
'dev_minibatches': dev_minibatches,
'config_args': args}, out_file)
ep_loss = 0.
num_batches = 0
start_time = time.time()
# adjust LR for next epoch
for param_group in optimizer.param_groups:
param_group['lr'] = param_group['lr'] * args.lr_decay_factor
print 'new LR:', param_group['lr']