-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
272 lines (228 loc) · 11.4 KB
/
utils.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
import json
import numpy as np
import random
from nltk.tokenize import RegexpTokenizer
import torch
from transformers import PreTrainedTokenizerBase
def get_dict(path):
word2idx, idx2word = dict(), dict()
with open(path, encoding='utf-8') as f_in:
for line in f_in:
word = line.strip()
idx = len(word2idx)
word2idx[word] = idx
idx2word[idx] = word
return word2idx, idx2word
class Tokenizer:
def __init__(self, vocab_path):
self.word2idx, self.idx2word, self.num_token = self.preprocess(vocab_path)
def preprocess(self, path):
word2idx, idx2word, count = {}, {}, 0
word2idx['#pad#'] = count
idx2word[count] = '#pad#'
count += 1
word2idx['<unk>'] = count
idx2word[count] = '<unk>'
count += 1
with open(path, encoding='utf-8') as f:
for each in f:
word = each.strip()
if word in word2idx:
raise ValueError(word+' have already in vocab dict!')
word2idx[word] = count
idx2word[count] = word
count += 1
return word2idx, idx2word, count
def __call__(self, text):
tok_ids = []
for token in text.split():
tok_ids.append(self.word2idx.get(token, self.word2idx['<unk>']))
return tok_ids
def decode(self, text):
return ' '.join([self.idx2word[t] for t in text])
@staticmethod
def load_glove_emb(path):
pad = np.zeros(300)
unk = np.random.uniform(-1., 1., 300)
word_emb = np.load(path)
return np.vstack([pad, unk, word_emb])
class QADataset:
def __init__(self, data_path, ent2idx, rel2idx, tokenizer, batch_size, training, device,
fact_dropout=0., token_path=None, label_smooth=0., enable_entity_linking=False):
self.ent2idx = ent2idx
self.rel2idx = rel2idx
self.tokenizer = tokenizer
self.batch_size = batch_size
self.training = training
self.device = device
self.fact_dropout = fact_dropout
self.reg_tokenizer = RegexpTokenizer(r'\d{1}|\w+|[^\w\s]+') if token_path is None else None
self.label_smooth = label_smooth
self.max_seq_len = 0
data, num_omit = self.load_data(data_path, token_path)
self.enable_entity_linking = enable_entity_linking
self.num_data = len(data)
# when enable_entity_linking = False,
# full_size = num_data, (check this in log file)
# when enable_entity_linking = True,
# full_size > num_data,
# we use #num_data for training, use #full_size for evaluation
self.full_size = self.num_data + num_omit
self.max_local_entity = 0
self.global2local_maps = self.build_global2local_maps(data)
self.data_id = np.empty(self.num_data, dtype=object)
self.question = np.zeros((self.num_data, self.max_seq_len), dtype=int)
if isinstance(tokenizer, PreTrainedTokenizerBase):
self.question += self.tokenizer.pad_token_id
self.question_mask = np.zeros((self.num_data, self.max_seq_len), dtype=float)
self.topic_label = np.zeros((self.num_data, self.max_local_entity), dtype=float)
self.candidate_entity = np.zeros((self.num_data, self.max_local_entity), dtype=int) + len(ent2idx)
self.entity_mask = np.zeros((self.num_data, self.max_local_entity), dtype=float)
self.subgraph = np.empty(self.num_data, dtype=object)
self.answer_label = np.zeros((self.num_data, self.max_local_entity), dtype=float)
self.answer_list = np.empty(self.num_data, dtype=object)
self.buffer_data(data)
def load_data(self, data_path, token_path):
idx2question = {}
if token_path is not None:
with open(token_path, encoding='utf-8') as f:
for each in f:
each = json.loads(each)
processed_question = []
for tok in each['dep']:
processed_question.append(tok[0])
idx2question[each['id']] = ' '.join(processed_question)
omitted, data = [], []
with open(data_path, encoding='utf-8') as f:
for each in f:
each = json.loads(each)
if len(each['entities']) == 0:
omitted.append(each['id'])
continue
if token_path is not None:
if each['id'] not in idx2question:
raise ValueError(each['id'] + 'don\'t have tokenized question!')
each['question'] = self.tokenizer(idx2question[each['id']])
else:
each['question'] = self.tokenizer(' '.join(self.reg_tokenizer.tokenize(each['question'].lower())))
if isinstance(self.tokenizer, PreTrainedTokenizerBase):
each['question'] = each['question'].input_ids
self.max_seq_len = max(self.max_seq_len, len(each['question']))
data.append({
'id': each['id'],
'question': each['question'],
'topic entities': each['entities'], # topic entity id
'subgraph': each['subgraph']['tuples'],
'candidates': each['subgraph']['entities'], # subgraph entity id
'answers': each['answers']
})
print('Read %d data from %s' % (len(data), data_path))
print('Omit %d questions without any topic entity:\n%s' % (len(omitted), str(omitted)))
return data, len(omitted)
def build_global2local_maps(self, data):
global2local, total_local_entity = {}, 0.
for each in data:
g2l = dict()
self._add_entity_to_map(each['topic entities'], g2l)
self._add_entity_to_map(each['candidates'], g2l)
assert each['id'] not in global2local, 'Duplicate data id: ' + each['id']
global2local[each['id']] = g2l
total_local_entity += len(g2l)
self.max_local_entity = max(self.max_local_entity, len(g2l))
print('Average local entity: %.2f' % (total_local_entity / len(data)))
print('Max local entity: %d' % self.max_local_entity)
return global2local
@staticmethod
def _add_entity_to_map(entities, g2l):
for each in entities:
if each not in g2l:
g2l[each] = len(g2l)
def buffer_data(self, data):
answerable = 0
for i, each in enumerate(data):
self.data_id[i] = each['id']
g2l = self.global2local_maps[each['id']]
assert len(g2l) > 0
for j, word in enumerate(each['question']):
self.question[i, j] = word
self.question_mask[i, j] = 1
topic_ent = set([g2l[x] for x in each['topic entities']])
for ent in topic_ent:
self.topic_label[i, ent] = 1
# self.topic_label[i, ent] = 1 / len(topic_ent)
for g, l in g2l.items():
self.entity_mask[i, l] = 1
self.candidate_entity[i, l] = g
head_list, rel_list, tail_list = [], [], []
for head, rel, tail in each['subgraph']:
head_list.append(g2l[head])
tail_list.append(g2l[tail])
rel_list.append(rel)
self.subgraph[i] = (
np.array(head_list, dtype=int),
np.array(rel_list, dtype=int),
np.array(tail_list, dtype=int)
)
answer_set, local_answer_set = set(), set()
for answer in each['answers']:
answer = answer['text'] if type(answer['kb_id']) == int else answer['kb_id']
answer_id = self.ent2idx[answer]
answer_set.add(answer_id)
if answer_id in g2l:
local_answer_set.add(g2l[answer_id])
self.answer_list[i] = list(answer_set)
if len(local_answer_set) > 0:
answerable += 1
self.answer_label[i] = np.full(
self.max_local_entity, self.label_smooth/(self.max_local_entity-len(local_answer_set)), dtype=float
)
for answer in local_answer_set:
# self.answer_label[i, answer] = 1.
self.answer_label[i, answer] = (1-self.label_smooth) / len(local_answer_set)
print("There are %d / %d answerable questions" % (answerable, len(data)))
def __getitem__(self, item):
return (
self.data_id[item], self.question[item], self.question_mask[item], self.topic_label[item],
self.entity_mask[item], self.subgraph[item], self.answer_label[item], self.answer_list[item]
)
def __len__(self):
return self.num_data
def batching(self):
indices = list(range(self.num_data))
if self.training:
random.shuffle(indices)
for start_index in range(0, self.num_data, self.batch_size):
batch_indices = indices[start_index: start_index+self.batch_size]
data_id = self.data_id[batch_indices]
question = torch.tensor(self.question[batch_indices], dtype=torch.long).to(self.device)
question_mask = torch.tensor(self.question_mask[batch_indices], dtype=torch.float).to(self.device)
topic_label = torch.tensor(self.topic_label[batch_indices], dtype=torch.float).to(self.device)
candidate_entity = torch.tensor(self.candidate_entity[batch_indices], dtype=torch.long).to(self.device)
entity_mask = torch.tensor(self.entity_mask[batch_indices], dtype=torch.float).to(self.device)
subgraph = self.build_subgraph(data_id, batch_indices)
answer_label = torch.tensor(self.answer_label[batch_indices], dtype=torch.float).to(self.device)
answer_list = self.answer_list[batch_indices]
yield (data_id, question, question_mask, topic_label, candidate_entity, entity_mask, subgraph, answer_label,
answer_list)
def build_subgraph(self, data_id, batch_indices):
batch_heads, batch_tails, batch_ids = np.array([], dtype=int), np.array([], dtype=int), np.array([], dtype=int)
batch_relations = np.array([], dtype=int)
subgraph = self.subgraph[batch_indices]
for i, (head_list, rel_list, tail_list) in enumerate(subgraph):
offset = i * self.max_local_entity
fact_size = len(head_list)
if self.training:
fact_size_in_use = np.ceil(fact_size * (1-self.fact_dropout)).astype(int)
mask_index = np.random.permutation(fact_size)[:fact_size_in_use]
else:
fact_size_in_use = fact_size
mask_index = np.arange(fact_size)
batch_heads = np.append(batch_heads, head_list[mask_index]+offset)
batch_relations = np.append(batch_relations, rel_list[mask_index])
batch_tails = np.append(batch_tails, tail_list[mask_index]+offset)
batch_ids = np.append(batch_ids, np.full(fact_size_in_use, i, dtype=int))
batch_ids = torch.from_numpy(batch_ids).long().to(self.device)
batch_relations = torch.from_numpy(batch_relations).long().to(self.device)
edge_index = np.vstack([batch_heads, batch_tails])
edge_index = torch.from_numpy(edge_index).to(self.device)
return batch_ids, batch_relations, edge_index