-
Notifications
You must be signed in to change notification settings - Fork 2
/
pretrain.py
473 lines (369 loc) · 19 KB
/
pretrain.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
import pandas as pd
import logging
import os, sys
import argparse
import random
from tqdm import tqdm, trange
import xml.etree.ElementTree as ET
from pprint import pprint
import random
import time
import numpy as np
import pickle
import torch
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
from torch.utils.data.distributed import DistributedSampler
from transformers import RobertaTokenizer
from pytorch_pretrained_bert.tokenization import BertTokenizer
from pytorch_pretrained_bert.optimization import BertAdam
from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE
from onmt.BertModules import *
from onmt.GraphBert import *
from onmt.VariationalGraphBert import *
from onmt.Utils import *
import onmt.Opt
import pdb
sys.path.append("/users4/ldu/git_clones/apex/")
from apex import amp
'''
def loss_graph(appro_matrix, true_graph, loss_fn):
appro_matrix = appro_matrix.squeeze()
true_graph = true_graph.squeeze()
assert appro_matrix.shape == true_graph.shape
L = appro_matrix.shape[1]
loss_tot = 0
for i in range(L):
for j in range(L):
if i != j:
p = appro_matrix[:,i, j].unsqueeze(1)
#p_comple = 1 - p
p_comple = appro_matrix[:,j, i].unsqueeze(1)
p = torch.cat([p, p_comple], axis=1)
#p = torch.log(p) / (1 - torch.log(p))
q = true_graph[:,i, j]
loss_tmp = loss_fn(p, q)
loss_tot = loss_tot + loss_tmp
return loss_tot
'''
#os.environ['CUDA_VISIBLE_DEVICES']="6,7"
def mask_tokens(inputs, tokenizer):
inputs_0 = inputs.clone()
""" Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
labels = inputs.clone()
# We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
probability_matrix = torch.full(labels.shape, 0.15)
#pdb.set_trace()
special_tokens_mask = [tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()]
probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
masked_indices = torch.bernoulli(probability_matrix).bool()
labels[~masked_indices] = -1 # We only compute loss on masked tokens
# 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
inputs[indices_replaced] = tokenizer.convert_tokens_to_ids(tokenizer.mask_token)
# 10% of the time, we replace masked input tokens with random word
indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
random_words = torch.randint(len(tokenizer), labels.shape, dtype=torch.long).cuda(inputs.device)
#random_words = torch.randint(len(tokenizer), labels.shape, dtype=torch.long)
inputs[indices_random] = random_words[indices_random]
# The rest of the time (10% of the time) we keep the masked input tokens unchanged
#pdb.set_trace()
return inputs, labels
def convert_examples_to_features(examples, tokenizer, max_seq_length,
baseline=False, voc=None):
features_p = []
features_q = []
if 'graph' in examples[0].keys():
has_graph = True
else:
has_graph = False
num_not_append = 0
if opt.pretrain_method == 'V':
#examples = examples[50201:] # deleting vist examples
examples = examples[:-50000] # deleting vist examples
if opt.pretrain_method == 'T':
#examples = examples[:-39590] # deleting timetravel examples
examples = examples[-50000:] # deleting timetravel examples
if opt.pretrain_method == 'R':
examples = random.sample(examples, opt.pretrain_number)
for example_index, example in enumerate(examples):
p_sents = [example['hyp1'], example['obs1'], example['hyp2']]
q_sents = example['sents']
if opt.pretrain_method == 'I':
random_example_1 = random.sample(examples, 1)
random_example_2 = random.sample(examples, 1)
random_sent_1 = random.sample(random_example_1[0]['sents'], 1)[0]
random_sent_2 = random.sample(random_example_2[0]['sents'], 1)[0]
q_sents[1] = random_sent_1
q_sents[3] = random_sent_2
if_append = True
for sents, features in zip([p_sents, q_sents], [features_p, features_q]):
if_append = True
choices_features = []
chain_tokens_tmp = []
sentence_ind_tmp = []
l_sents = [l for l in range(len(sents))]
l_sents.append(-1)
for ith_sent, sent in enumerate(sents):
sent_tokens = tokenizer.tokenize(sent)
chain_tokens_tmp.append(sent_tokens)
sentence_ind_tmp.extend([ith_sent] * (len(sent_tokens) + 1))
tokens_tmp = [["[CLS]"]] + [token + ["[SEP]"] for token in chain_tokens_tmp]
tokens_tmp[-1].pop()
tokens_tmp = [token for tokens in tokens_tmp for token in tokens]
input_ids_tmp = tokenizer.convert_tokens_to_ids(tokens_tmp)
input_mask_tmp = [1] * len(input_ids_tmp)
if (max_seq_length - len(input_ids_tmp)) >= 0:
padding = [0] * (max_seq_length - len(input_ids_tmp))
input_ids_tmp += padding
input_mask_tmp += padding
sentence_ind_tmp += [p-1 for p in padding]
else:
input_ids_tmp = input_ids_tmp[:max_seq_length]
input_mask_tmp = input_mask_tmp[:max_seq_length]
sentence_ind_tmp = sentence_ind_tmp[:max_seq_length]
if has_graph:
graph = example['graph']
else:
graph = None
if opt.pretrain_method == 'A':
graph = [np.random.rand(5, 5)]
try:
assert len(input_ids_tmp) == max_seq_length
assert len(input_mask_tmp) == max_seq_length
assert len(sentence_ind_tmp) == max_seq_length
except:
pdb.set_trace()
if set(sentence_ind_tmp) != set(l_sents[:-1]) and set(sentence_ind_tmp) != set(l_sents):
if_append = False
num_not_append += 1
print(num_not_append)
print("Too long example, id:", example_index)
choices_features.append((tokens_tmp, input_ids_tmp, input_mask_tmp, sentence_ind_tmp, graph))
answer = [0]
try:
answer[int(example['ans'])-1] = 1
except:
pdb.set_trace()
answer[example['answer']] = 1
features.append(
InputFeatures(
example_id = example_index,
choices_features = choices_features,
answer = answer
)
)
if not if_append:
features_p.pop()
features_q.pop()
assert len(features_p) == len(features_q)
#pdb.set_trace()
return [features_p, features_q]
os.environ['CUDA_VISIBLE_DEVICES'] = '1,2,3,4,5,6,7'
parser = argparse.ArgumentParser(
description='Train.py',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# onmt.opts.py
onmt.Opt.model_opts(parser)
opt = parser.parse_args()
gpu_ls = parse_gpuid(opt.gpuls)
if 'large' in opt.bert_model:
opt.train_batch_size = 12 * len(gpu_ls)
else:
opt.train_batch_size = 18 * len(gpu_ls)
random.seed(opt.seed)
np.random.seed(opt.seed)
torch.cuda.manual_seed_all(opt.seed)
wkdir = "/users4/ldu/abductive"
os.makedirs(opt.output_dir, exist_ok=True)
train_examples = None
eval_examples = None
eval_size= None
num_train_steps = None
train_examples = load_examples(os.path.join(opt.train_data_dir))[:1000]
pdb.set_trace()
num_train_steps = int(len(train_examples) / opt.train_batch_size / opt.gradient_accumulation_steps * opt.num_train_epochs)
# Prepare tokenizer
#tokenizer = torch.load(opt.bert_tokenizer)
tokenizer = RobertaTokenizer.from_pretrained(opt.bert_tokenizer)
# Prepare model
model = ini_from_pretrained(opt)
# Prepare optimizer
if opt.fp16:
param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
for n, param in model.named_parameters()]
elif opt.optimize_on_cpu:
param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
#no_decay = ['bias', 'gamma', 'beta']
#no_decay = ['gamma', 'beta']
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay_rate': opt.l2_reg},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.0}
]
t_total = num_train_steps
if opt.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
optimizer = BertAdam(optimizer_grouped_parameters,
lr=opt.learning_rate,
warmup=opt.warmup_proportion,
t_total=t_total)
# optimizer = adabound.AdaBound(optimizer_grouped_parameters, lr=opt.learning_rate, final_lr=0.1)
model.cuda(gpu_ls[0])
if 'large' in opt.bert_model:
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
else:
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
model_config = model.config
model = nn.DataParallel(model, device_ids=gpu_ls)
model.config = model_config
global_step = 0
if opt.pret:
train_features_p, train_features_q = convert_examples_to_features(
train_examples, tokenizer, opt.max_seq_length, True)
else:
train_features = train_examples
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", opt.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_example_ids = torch.tensor([train_feature_p.example_id for train_feature_p in train_features_p], dtype=torch.long)
all_input_ids_p = torch.tensor(select_field(train_features_p, 'input_ids'), dtype=torch.long)
all_attn_msks_p = torch.tensor(select_field(train_features_p, 'input_mask'), dtype=torch.long)
all_sentence_inds_p = torch.tensor(select_field(train_features_p, 'sentence_ind'), dtype=torch.long)
all_input_ids_q = torch.tensor(select_field(train_features_q, 'input_ids'), dtype=torch.long)
all_attn_msks_q = torch.tensor(select_field(train_features_q, 'input_mask'), dtype=torch.long)
all_sentence_inds_q = torch.tensor(select_field(train_features_q, 'sentence_ind'), dtype=torch.long)
all_graphs = select_field(train_features_p, 'graph') ##
all_graphs = torch.tensor(all_graphs, dtype=torch.float) ##
train_data = TensorDataset(all_example_ids, all_input_ids_p, all_attn_msks_p, all_sentence_inds_p, all_input_ids_q, all_attn_msks_q, all_sentence_inds_q, all_graphs)
if opt.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=opt.train_batch_size)
loss_fct = torch.nn.CrossEntropyLoss(ignore_index=-1)
loss_aa_fn = torch.nn.CrossEntropyLoss()
Lambda = opt.Lambda
Lambda_kl = opt.Lambda_kl
loss_aa_smooth_term = opt.loss_aa_smooth
name = parse_opt_to_name(opt)
print(name)
time_start = str(int(time.time()))[-6:]
#test_examples_all = load_examples(os.path.join(opt.test_data_dir))
#test_features_all = convert_examples_to_features(test_examples_all, tokenizer, opt.max_seq_length, True)
#for epoch in range(int(opt.num_train_epochs / 10)):
for epoch in range(opt.num_train_epochs):
print("Epoch:",epoch)
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(train_dataloader):
pdb.set_trace()
model.train()
batch = tuple(t.cuda(gpu_ls[0]) for t in batch)
'''
for both multiple choice problem and next sentence prediction,
the input is context and one of the choice.
'''
example_ids, input_ids_p, attn_msks_p, sentence_inds_p, input_ids_q, attn_msks_q, sentence_inds_q, graphs = batch
num_choices = input_ids_p.shape[1]
accurancy = None
for n in range(num_choices):
input_ids_p_tmp = input_ids_p[:,n,:]
input_ids_q_tmp = input_ids_q[:,n,:]
attn_msks_p_tmp = attn_msks_p[:,n,:]
attn_msks_q_tmp = attn_msks_q[:,n,:]
sentence_inds_p_tmp = sentence_inds_p[:,n,:]
sentence_inds_q_tmp = sentence_inds_q[:,n,:]
graphs_tmp = graphs[:,n,:]
#graphs_tmp_scaled = graphs_tmp
graphs_tmp_scaled = graphs_tmp / graphs_tmp.sum(3).unsqueeze(3)
input_ids_p_tmp_msk, labels_p = mask_tokens(input_ids_p_tmp, tokenizer)
input_ids_q_tmp_msk, _ = mask_tokens(input_ids_q_tmp, tokenizer)
if opt.model_type == 'vgb':
pred_tokens, z_p, z_q, attn_scores = model(input_ids_p = input_ids_p_tmp_msk, input_ids_q = input_ids_q_tmp_msk,
sentence_inds_p = sentence_inds_p_tmp, sentence_inds_q = sentence_inds_q_tmp) ##
elif opt.model_type == 'vgb_c':
pred_tokens, z_p, z_q, attn_scores = model(input_ids_p = input_ids_p_tmp_msk, input_ids_q = input_ids_q_tmp_msk,graph=graphs_tmp_scaled,
sentence_inds_p = sentence_inds_p_tmp, sentence_inds_q = sentence_inds_q_tmp) ##
cov = torch.ones_like(z_p).cuda(input_ids_p_tmp.device)
p = torch.distributions.normal.Normal(z_p, cov)
q = torch.distributions.normal.Normal(z_q, cov)
masked_lm_loss = loss_fct(pred_tokens.view(-1, model.config.vocab_size), input_ids_p_tmp_msk.view(-1))
#graphs_tmp_n = np.zeros(graphs_tmp.shape) + np.triu(np.ones(graphs_tmp.shape), 1)
if opt.model_type == 'vgb':
graphs_tmp_n = np.zeros(graphs_tmp.shape)
for i in range(graphs_tmp_n.shape[2] - 1):
#graphs_tmp_n[:, i, i] = 1
graphs_tmp_n[:, :, i, i+1] = 1
graphs_tmp_n = torch.LongTensor(graphs_tmp_n)
#graphs_tmp = graphs_tmp_n.cuda(gpu_ls[0])
'''
for i in range(graphs_tmp_n.shape[2]):
for j in range(graphs_tmp_n.shape[2]):
if i == j:
graphs_tmp_n[:, :, i, j] = 0.5
if (j - i) == 1:
graphs_tmp_n[:, :, i, j] = 1
if (j - i) == 2:
graphs_tmp_n[:, :, i, j] = 0.3
if (j - i) == 3:
graphs_tmp_n[:, :, i, j] = 0.1
#pdb.set_trace()
graphs_tmp_n = torch.FloatTensor(graphs_tmp_n)
graphs_tmp = graphs_tmp_n.cuda(gpu_ls[0])
#pdb.set_trace()
'''
try:
loss_aa = loss_graph(attn_scores, graphs_tmp, loss_aa_fn)
loss_kl = torch.distributions.kl.kl_divergence(p, q).sum()
#loss = masked_lm_loss
loss = masked_lm_loss + Lambda * loss_aa + Lambda_kl * loss_kl
#loss = Lambda * loss_aa + Lambda_kl * loss_kl
except:
loss = masked_lm_loss + Lambda_kl * loss_kl
if step % 20 == 0:
print("step:", step, "loss_msk_lm:", masked_lm_loss.detach().cpu().numpy(), "loss_aa:",loss_aa.detach().cpu().numpy() * Lambda, 'loss_kl:', Lambda_kl * loss_kl)
f = open(wkdir + '/records/graph_pretrained/' + name + '_' + time_start + '.csv', 'a+')
f.write(str(masked_lm_loss.detach().cpu().numpy()) + ',' + str(Lambda * loss_aa.detach().cpu().numpy()) + ',' + str(Lambda_kl * loss_kl.detach().cpu().numpy()) + '\n')
f.close()
elif opt.model_type == 'vgb_c':
loss_kl = torch.distributions.kl.kl_divergence(p, q).sum()
loss = masked_lm_loss + Lambda_kl * loss_kl
if step % 20 == 0:
print("step:", step, "loss_msk_lm:", masked_lm_loss.detach().cpu().numpy(), 'loss_kl:', Lambda_kl * loss_kl)
f = open(wkdir + '/records/graph_pretrained/' + name + '_' + time_start + '.csv', 'a+')
f.write(str(masked_lm_loss.detach().cpu().numpy()) + ',' + str(Lambda_kl * loss_kl.detach().cpu().numpy()) + '\n')
f.close()
#if step > 300:
#pdb.set_trace()
#x=zip([p.grad.norm().detach().cpu().numpy().tolist() for p in model.parameters()],model.state_dict().keys())
if opt.fp16 and opt.loss_scale != 1.0:
loss = loss * opt.loss_scale
if opt.gradient_accumulation_steps > 1:
loss = loss / opt.gradient_accumulation_steps
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
#loss.backward()
#tr_loss += loss.item()
#nb_tr_examples += input_ids.size(0)
#nb_tr_steps += 1
if (step + 1) % opt.gradient_accumulation_steps == 0:
optimizer.step()
model.zero_grad()
global_step += 1
ls = [model.config, model.state_dict()]
if opt.pretrain_method in ['R', 'V', 'T']:
torch.save(ls, wkdir + "/pretrained_models/graph_pretrained/datset/" + "e_" + str(epoch) + str(step) + name + time_start + '.pkl')
elif opt.pretrain_method == 'L':
torch.save(ls, wkdir + "/pretrained_models/graph_pretrained/lambda/" + "e_" + str(epoch) + str(step) + name + time_start + '.pkl')
else:
torch.save(ls, wkdir + "/pretrained_models/graph_pretrained/" + "e_" + str(epoch) + str(step) + name + time_start + '.pkl')
'''
if epoch == 1:
torch.save(ls, wkdir + "/ablation_models/" + "e_" + str(epoch) + str(step) + name + time_start + '.pkl')
if epoch > 1:
break
'''