-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset.py
474 lines (421 loc) · 19.9 KB
/
dataset.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
import copy
import json
import os
import math
import random
from random import random as rand
import pickle
import numpy as np
import torch
import pandas as pd
import dgl
import nltk
import jsonlines
from nltk.corpus import stopwords
from nltk.tokenize.treebank import TreebankWordDetokenizer
from ast import literal_eval
from preprocess import save_json, load_json, reverse_dict_key_val
torch.multiprocessing.set_sharing_strategy('file_system')
DEFAULT_VOCAB = ['[PAD]', '[UNK]', '[MASK]', '[SEP]', '[CLS]']
PAD_IDX, UNK_IDX, MASK_IDX, SEP_IDX, CLS_IDX = 0, 100, 103, 102, 101
NAF_TRIPLE = [UNK_IDX, UNK_IDX, UNK_IDX, UNK_IDX]
class commondataset(torch.utils.data.Dataset):
def __init__(self, args, tokenizor, data_name='train'):
assert data_name in ['train', 'test', 'valid'], "Data name should be among ['train', 'test', 'valid']."
self.args = args
self.data_name = data_name
self.ent2id = load_json(args.word2id_path)
self.rel2word = load_json(args.rel2word_path)
self.batch_size = args.batch_size
self.tokenizor = tokenizor
self.csk_dict = load_json(args.csk_dict_path)
self.max_input_len = args.max_input_len
self.max_output_len = args.max_output_len
nltk.download('stopwords')
self.stopwords = stopwords.words('english')
self.convdata = self.init_convdata()
print(f'There are {len(self.convdata)} conversations in total.', flush=True)
def _is_special(self, c):
return bool(c) and (c[0] == '[')
def decode(self, token_ids, join_str=''):
tokens = []
for i in token_ids:
t = self.tokenizor.convert_ids_to_tokens(i)
if t == '[SEP]':
tokens.append('.')
break
elif not self._is_special(t):
tokens.append(t)
tokens.append(' ')
return join_str.join(tokens)
def init_convdata(self):
print('loading conversation data', flush=True)
data_list = []
with self.args.data_path.open("r", encoding="utf-8") as fr:
for line in fr:
one_data = json.loads(line.strip())
data_list.append(one_data)
print(f"data from {self.args.data_path} size: {len(data_list)}")
return data_list
def retrieve_graph(self, words, max_triple_len=50):
key = words.lower()
if key in self.com_diction and key not in self.stopwords:
graphs = self.com_diction[key]
if len(graphs) >= max_triple_len:
graphs.sort(key=lambda t: t[3])
graphs = graphs[:max_triple_len]
ret_graphs = []
h = None
if graphs[0][0] in self.ent2id:
h = self.ent2id[graphs[0][0]]
for g in graphs:
if g[1] in self.ent2id and g[2] in self.rel2word:
t = self.ent2id[g[1]]
r = self.rel2word[g[2]]
if h is not None:
ret_graphs.append([h, t, self.ent2id[r[0]], self.ent2id[r[1]]])
if len(ret_graphs) == 0:
return [NAF_TRIPLE], False
return ret_graphs, True
else:
return [NAF_TRIPLE], False
def process_triple(self, all_triples, target_triples):
src_triple = []
target_head = []
target_tail = []
main_entity = []
for triple_list in all_triples:
processed_triple = []
for t in triple_list:
triple = self.csk_dict[t].split(', ')
head = triple[0]
rel = triple[1]
tail = triple[2]
rel = self.rel2word[rel].split()
if head in self.tokenizor.vocab and tail in self.tokenizor.vocab:
processed_triple.append(
[self.tokenizor.convert_tokens_to_ids(head), self.tokenizor.convert_tokens_to_ids(rel[0]), self.tokenizor.convert_tokens_to_ids(rel[1]), self.tokenizor.convert_tokens_to_ids(tail)])
src_triple.append(processed_triple)
entities = [p for triple in processed_triple for p in triple]
if len(entities) > 0:
entity = max(set(entities), key=entities.count)
else:
entity = None
main_entity.append(entity)
for t in target_triples:
triple = self.csk_dict[t].split(', ')
head = triple[0]
tail = triple[2]
if head in self.tokenizor.vocab and tail in self.tokenizor.vocab:
target_head.append(self.tokenizor.convert_tokens_to_ids(head))
target_tail.append(self.tokenizor.convert_tokens_to_ids(tail))
target_ent = target_head + target_tail
return src_triple, main_entity, target_ent
def make_graph(self, all_triples, main_entity, target_ent):
word2nodeid = {}
nodeid2wordid = {}
edges = []
target_node = []
embedding = []
#add a root node with 0 embeddings
embedding.append([PAD_IDX, PAD_IDX, PAD_IDX, PAD_IDX])
word2nodeid['root'] = len(word2nodeid)
nodeid2wordid[word2nodeid['root']] = 'root'
target_node.append(1)
for h_index, triples in enumerate(all_triples):
main_ent = main_entity[h_index]
if main_ent is UNK_IDX:
continue
if main_ent not in word2nodeid:
word2nodeid[main_ent] = len(word2nodeid)
e = [PAD_IDX, PAD_IDX, PAD_IDX, PAD_IDX]
embedding.append(e)
nodeid2wordid[word2nodeid[main_ent]] = main_ent
if main_ent in target_ent:
target_node.append(1)
else:
target_node.append(0)
edges.append([0, word2nodeid[main_ent]])
for t_index, t in enumerate(triples):
# define the attribute of the tail nodes and connect them to the head node
if str(t) not in word2nodeid:
word2nodeid[str(t)] = len(word2nodeid)
embedding.append(t)
nodeid2wordid[word2nodeid[str(t)]] = str(t)
if t[0] in target_ent and t[-1] in target_ent:
target_node.append(1)
else:
target_node.append(0)
edges.append([word2nodeid[main_ent], word2nodeid[str(t)]])
# target_ent = target_head+target_tail
if len(edges) > 0:
edges = torch.tensor(edges)
g = dgl.graph((edges[:, 1], edges[:, 0]), num_nodes=len(word2nodeid))
g.ndata['h'] = torch.tensor(embedding).detach()
g.ndata['_ID'] = torch.Tensor(np.array(range(len(word2nodeid)))).long()
tar = torch.tensor(target_node, dtype=torch.float32)
else:
edges = [[0, 1]]
edges = torch.tensor(edges)
g = dgl.graph((edges[:, 1], edges[:, 0]), num_nodes=2)
g.ndata['h'] = UNK_IDX * torch.ones((2, 4 * self.args.embedding_size))
g.ndata['_ID'] = torch.Tensor(np.array(range(2))).long()
tar = torch.tensor([0, 0], dtype=torch.float32)
g.ndata['tar'] = tar
g.ndata['loss'] = torch.zeros_like(tar)
g.ndata['tar_num'] = torch.zeros_like(tar)
return g
def load_obj(self, path):
with open(path, 'rb') as f:
return pickle.load(f)
def get_random_word(self):
i = random.randint(0, len(self.tokenizor.vocab) - 1)
return self.tokenizor.convert_ids_to_tokens(i)
def mask(self, src, tgt, response_target):
tokens = src + tgt
segment_ids = [4] * (len(src)) + [5] * (len(tgt))
# For masked Language Models
# the number of prediction is sometimes less than max_pred when sequence is short
effective_length = len(tgt) - 1
n_pred = min(self.args.max_pred, max(1, int(round(effective_length * self.args.mask_prob))))
# candidate positions of masked tokens
cand_pos = []
special_pos = set()
for i, tk in enumerate(tokens):
# only mask tokens_b (target sequence)
# we will mask [SEP] as an ending symbol
if (i >= len(src)) and (tk != '[CLS]'):
cand_pos.append(i)
else:
special_pos.add(i)
random.shuffle(cand_pos)
triple_list = [self.csk_dict[item].split(', ') for item in response_target if item != -1]
ent_list = [e for tri in triple_list for e in tri]
ent_pos = [i + len(src) for i, item in enumerate(tgt) if item in ent_list]
# masked_pos = set(ent_pos)
masked_pos = set()
max_cand_pos = max(cand_pos)
for pos in cand_pos:
if len(masked_pos) >= n_pred:
break
if pos in masked_pos:
continue
st_pos, end_pos = pos, pos + 1
for mp in range(st_pos, end_pos):
if (0 < mp <= max_cand_pos) and (mp not in special_pos):
masked_pos.add(mp)
else:
break
masked_pos = list(masked_pos)
if len(masked_pos) > n_pred:
random.shuffle(masked_pos)
masked_pos = masked_pos[:n_pred]
masked_tokens = [tokens[pos] for pos in masked_pos]
for pos in masked_pos:
if rand() < 0.8: # 80%
tokens[pos] = '[MASK]'
elif rand() < 0.5: # 10%
tokens[pos] = self.get_random_word()
# when n_pred < max_pred, we only calculate loss within n_pred
masked_weights = [1] * len(masked_tokens)
# Token Indexing
masked_ids = self.tokenize(masked_tokens)
# Token Indexing
input_ids = self.tokenize(tokens)
# Zero Padding
n_pad = self.max_input_len - len(input_ids)
input_ids.extend([0] * n_pad)
segment_ids.extend([0] * n_pad)
encoder_mask = torch.zeros(self.max_input_len, self.max_input_len, dtype=torch.long)
encoder_mask[:, :len(src)].fill_(1)
second_st, second_end = len(
src), len(src) + len(tgt)
_tril_matrix = torch.tril(torch.ones(
(self.max_input_len, self.max_input_len), dtype=torch.long))
encoder_mask[second_st:second_end, second_st:second_end].copy_(
_tril_matrix[:second_end - second_st, :second_end - second_st])
decoder_len = self.max_input_len + self.args.num_attention_graphs * 4
decoder_mask = torch.zeros(decoder_len, decoder_len, dtype=torch.long)
decoder_mask[:, :len(src) + 4 * self.args.num_attention_graphs].fill_(1)
second_st, second_end = len(src) + 4 * self.args.num_attention_graphs, len(src) + len(
tgt) + 4 * self.args.num_attention_graphs
_tril_matrix = torch.tril(torch.ones(
(decoder_len, decoder_len), dtype=torch.long))
decoder_mask[second_st:second_end, second_st:second_end].copy_(
_tril_matrix[:second_end - second_st, :second_end - second_st])
# Zero Padding for masked target
if self.args.max_pred > n_pred:
n_pad = self.args.max_pred - n_pred
if masked_ids is not None:
masked_ids.extend([0] * n_pad)
if masked_pos is not None:
masked_pos.extend([0] * n_pad)
if masked_weights is not None:
masked_weights.extend([0] * n_pad)
return (
torch.tensor(input_ids), torch.tensor(segment_ids), encoder_mask, decoder_mask, torch.tensor(masked_ids),
torch.tensor(masked_pos), torch.tensor(masked_weights))
def mask_eval(self, src, tgt):
tokens = src + tgt
segment_ids = [4] * (len(src)) + [5] * (len(tgt))
# For masked Language Models
# the number of prediction is sometimes less than max_pred when sequence is short
effective_length = len(tgt) - 1
# candidate positions of masked tokens
cand_pos = []
special_pos = set()
for i, tk in enumerate(tokens):
# only mask tokens_b (target sequence)
# we will mask [SEP] as an ending symbol
if (i >= len(src)) and (tk != '[CLS]'):
cand_pos.append(i)
else:
special_pos.add(i)
masked_pos = list(cand_pos)
masked_tokens = [[tokens[pos]] for pos in masked_pos]
batch_tokens = []
for pos in masked_pos:
tokens_copy = tokens.copy()
tokens_copy[pos] = '[MASK]'
batch_tokens.append(tokens_copy)
# when n_pred < max_pred, we only calculate loss within n_pred
masked_weights = [[1]] * len(batch_tokens)
# Token Indexing
masked_ids = [self.tokenize(_) for _ in masked_tokens]
# Token Indexing
input_ids = [self.tokenize(t) for t in batch_tokens]
# Zero Padding
n_pad = self.max_input_len - len(input_ids[0])
batch_ids = []
for ids in input_ids:
ids.extend([0] * n_pad)
batch_ids.append(ids)
segment_ids.extend([0] * n_pad)
encoder_mask = torch.zeros(self.max_input_len, self.max_input_len, dtype=torch.long)
encoder_mask[:, :len(src)].fill_(1)
second_st, second_end = len(
src), len(src) + len(tgt)
_tril_matrix = torch.tril(torch.ones(
(self.max_input_len, self.max_input_len), dtype=torch.long))
encoder_mask[second_st:second_end, second_st:second_end].copy_(
_tril_matrix[:second_end - second_st, :second_end - second_st])
decoder_len = self.max_input_len + self.args.num_attention_graphs * 4
decoder_mask = torch.zeros(decoder_len, decoder_len, dtype=torch.long)
decoder_mask[:, :len(src) + 4 * self.args.num_attention_graphs].fill_(1)
second_st, second_end = len(src) + 4 * self.args.num_attention_graphs, len(src) + len(
tgt) + 4 * self.args.num_attention_graphs
_tril_matrix = torch.tril(torch.ones(
(decoder_len, decoder_len), dtype=torch.long))
decoder_mask[second_st:second_end, second_st:second_end].copy_(
_tril_matrix[:second_end - second_st, :second_end - second_st])
# Zero Padding for masked target
batch_size = len(masked_pos)
masked_pos = [[pos] for pos in masked_pos]
# print(batch_ids, torch.tensor(segment_ids).repeat(batch_size), encoder_mask,decoder_mask, torch.tensor(masked_ids),masked_pos, masked_weights)
return (
torch.tensor(batch_ids), torch.tensor(segment_ids).repeat(batch_size, 1),
encoder_mask.repeat(batch_size, 1, 1),
decoder_mask.repeat(batch_size, 1, 1), torch.tensor(masked_ids),
torch.tensor(masked_pos), torch.tensor(masked_weights))
def tokenize(self, src_ids):
ids = []
for s in src_ids:
if s not in DEFAULT_VOCAB:
s = s.lower()
id = self.tokenizor.convert_tokens_to_ids(s)
ids.append(id)
return ids
def pad_1d(self, list, length, pad_idx=PAD_IDX):
""" Pad over axis 0. """
return np.pad(list, (0, length - len(list)), 'constant', constant_values=pad_idx)
def __len__(self):
return len(self.convdata)
def __getitem__(self, i):
data = self.convdata[i]
# print(data)
src, tgt, match_triples, all_triples, response_target = literal_eval(data['post']), literal_eval(
data['response']), literal_eval(data['match_triples']), literal_eval(data['all_triples']), literal_eval(
data['response_triples'])
src = TreebankWordDetokenizer().detokenize(src)
tgt = TreebankWordDetokenizer().detokenize(tgt)
src = self.tokenizor.tokenize(src)
tgt = self.tokenizor.tokenize(tgt)
if self.data_name == 'train':
src = ['[CLS]'] + src + ['[SEP]']
tgt = tgt + ['[SEP]']
src_cat_tgt, token_type, encoder_mask, decoder_mask, masked_ids, masked_pos, masked_weights = self.mask(src,
tgt,
response_target)
src_len = len(src)
tgt_len = len(tgt)
triples, main_entity, target_ent = self.process_triple(all_triples, match_triples)
graph = self.make_graph(triples, main_entity, target_ent)
else:
src = ['[CLS]'] + src + ['[SEP]']
tgt = tgt + ['[SEP]']
src_cat_tgt, token_type, encoder_mask, decoder_mask, masked_ids, masked_pos, masked_weights = self.mask_eval(
src,
tgt)
src_len = len(src)
tgt_len = len(tgt)
triples, target_head, target_tail = self.process_triple(all_triples, match_triples)
graph = self.make_graph(triples, target_head, target_tail)
graph = [copy.deepcopy(graph) for _ in range(src_cat_tgt.shape[0])]
return {'src_cat_tgt': src_cat_tgt, 'token_type': token_type, 'encoder_mask': encoder_mask,
'decoder_mask': decoder_mask,
'masked_ids': masked_ids, 'masked_pos': masked_pos, 'masked_weights': masked_weights,
'graph': graph, 'src_len': src_len, 'tgt_len': tgt_len}
def pad_2d(list, pad_idx=PAD_IDX):
ml = max([len(i) for i in list])
return np.array([i + [pad_idx] * (ml - len(i)) for i in list])
def collate_fn(batch):
src_length = torch.tensor([s['src_len'] for s in batch]) # (bsz,)
src_cat_tgt = torch.stack([s['src_cat_tgt'] for s in batch])
token_type = torch.stack([s['token_type'] for s in batch])
encoder_mask = torch.stack([s['encoder_mask'] for s in batch])
decoder_mask = torch.stack([s['decoder_mask'] for s in batch])
masked_ids = torch.stack([s['masked_ids'] for s in batch])
masked_pos = torch.stack([s['masked_pos'] for s in batch])
masked_weights = torch.stack([s['masked_weights'] for s in batch])
tgt_length = torch.tensor([s['tgt_len'] for s in batch]) # (bsz,)
graph = [s['graph'] for s in batch] # (bsz,)
batched_data = {
'src_length': src_length,
'tgt_length': tgt_length,
'src_cat_tgt': src_cat_tgt.long(),
'token_type': token_type.long(),
'encoder_mask': encoder_mask.long(),
'decoder_mask': decoder_mask.long(),
'masked_ids': masked_ids.long(),
'masked_pos': masked_pos.long(),
'masked_weights': masked_weights.long(),
'graph': graph
}
return batched_data
def get_dataloader(dataset,
batch_size=64,
shuffle=True,
num_workers=0):
data_loader = torch.utils.data.DataLoader(dataset=dataset,
batch_size=batch_size,
num_workers=num_workers,
worker_init_fn=lambda _: np.random.seed(),
pin_memory=True,
collate_fn=collate_fn,
drop_last=True,
)
return data_loader
def get_distributed_data_loader(dataset,
train_sampler,
batch_size,
num_workers=0):
data_loader = torch.utils.data.DataLoader(dataset,
sampler=train_sampler,
batch_size=batch_size,
num_workers=num_workers,
shuffle=False,
pin_memory=True,
collate_fn=collate_fn,
drop_last=True,
)
return data_loader