-
Notifications
You must be signed in to change notification settings - Fork 0
/
data_utils_yelp.py
356 lines (303 loc) · 13.6 KB
/
data_utils_yelp.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
import io
import random
import numpy as np
import os
from transformers import AlbertTokenizer
import torch
import shelve
class DataUtil(object):
def __init__(self, hparams, k=8):
self.hparams = hparams
self.tokenizer = AlbertTokenizer.from_pretrained('albert-large-v2')
self.hparams.src_vocab_size = len(self.tokenizer)
self.hparams.pad_id = self.tokenizer.pad_token_id
self.hparams.unk_id = self.tokenizer.unk_token_id
self.hparams.bos_id = self.tokenizer.cls_token_id
self.hparams.eos_id = self.tokenizer.sep_token_id
self.hparams.mask_id = self.tokenizer.mask_token_id
self.hparams.pred_probs = torch.FloatTensor([hparams.word_mask, hparams.word_keep, hparams.word_rand])
if self.hparams.albert_kd and k > 1:
self.topk_db0 = shelve.open(f'{hparams.bert_dump0}/topk', 'r')
self.topk_db1 = shelve.open(f'{hparams.bert_dump1}/topk', 'r')
self.k = k
self.trg_i2w, self.trg_w2i = self._build_vocab(self.hparams.trg_vocab)
self.hparams.trg_vocab_size = len(self.trg_i2w)
#self.hparams.trg_pad_id = self.trg_w2i["<pad>"]
self.hparams.trg_pad_id = self.hparams.pad_id
print("src_vocab_size={}".format(self.hparams.src_vocab_size))
print("trg_vocab_size={}".format(self.hparams.trg_vocab_size))
if not self.hparams.decode:
self.train_size_0 = 0
self.n_train_batches_0 = 0
self.train_size_1 = 0
self.n_train_batches_1 = 0
self.train_x0, self.train_y0, _ , self.index0 = self._build_parallel(self.hparams.train_src_file0, self.hparams.train_trg_file)
self.train_size_0 = len(self.train_x0)
self.train_x1, self.train_y1, _ , self.index1 = self._build_parallel(self.hparams.train_src_file1, self.hparams.train_trg_file)
self.train_size_1 = len(self.train_x1)
self.dev_x0, self.dev_y0, _ , _ = self._build_parallel(self.hparams.dev_src_file0, self.hparams.dev_trg_file, is_train=False)
self.dev_x1, self.dev_y1, _ , _ = self._build_parallel(self.hparams.dev_src_file1, self.hparams.dev_trg_file, is_train=False)
self.dev_size_0 = len(self.dev_x0)
self.dev_size_1 = len(self.dev_x1)
self.dev_index_0 = 0
self.dev_index_1 = 0
self.reset_train()
else:
#test_src_file = os.path.join(self.hparams.data_path, self.hparams.test_src_file)
#test_trg_file = os.path.join(self.hparams.data_path, self.hparams.test_trg_file)
'''
test_src_file = self.hparams.test_src_file
test_trg_file = self.hparams.test_trg_file
self.test_x, self.test_y, _ , _ = self._build_parallel(test_src_file, test_trg_file, is_train=False)
self.test_size = len(self.test_x)
self.test_index = 0
'''
print("Class DataUtil will not be used as iterator")
def load_pretrained(self, pretrained_emb_file):
f = open(pretrained_emb_file, 'r', encoding='utf-8')
header = f.readline().split(' ')
count = int(header[0])
dim = int(header[1])
#matrix = np.zeros((len(w2i), dim), dtype=np.float32)
matrix = np.zeros((count, dim), dtype=np.float32)
#i2w = ['<pad>', '<unk>', '<s>', '<\s>']
i2w = []
#w2i = {'<pad>': 0, '<unk>':1, '<s>':2, '<\s>':3}
w2i = {}
for i in range(count):
word, vec = f.readline().split(' ', 1)
w2i[word] = len(w2i)
i2w.append(word)
matrix[i] = np.fromstring(vec, sep=' ', dtype=np.float32)
#if not word in w2i:
# print("{} no in vocab".format(word))
# continue
#matrix[w2i[word]] = np.fromstring(vec, sep=' ', dtype=np.float32)
return torch.FloatTensor(matrix), i2w, w2i
def reset_train(self):
if not self.n_train_batches_0:
self.n_train_batches_0 = (self.train_size_0 + self.hparams.batch_size - 1) // self.hparams.batch_size
if not self.n_train_batches_1:
self.n_train_batches_1 = (self.train_size_1 + self.hparams.batch_size - 1) // self.hparams.batch_size
self.train_queue_0 = np.random.permutation(self.n_train_batches_0)
self.train_index_0 = 0
self.train_queue_1 = np.random.permutation(self.n_train_batches_1)
self.train_index_1 = 0
def next_train(self):
start_index_0 = (self.train_queue_0[self.train_index_0] * self.hparams.batch_size)
end_index_0 = min(start_index_0 + self.hparams.batch_size, self.train_size_0)
start_index_1 = (self.train_queue_1[self.train_index_1] * self.hparams.batch_size)
end_index_1 = min(start_index_1 + self.hparams.batch_size, self.train_size_1)
x_train0 = self.train_x0[start_index_0:end_index_0]
x_train1 = self.train_x1[start_index_1:end_index_1]
self.train_index_0 += 1
self.train_index_1 += 1
batch_size = len(x_train0) + len(x_train1)
# pad
x_train0, _, _, _, _ = self._pad(x_train0, self.hparams.pad_id)
x_train1, _, _, _, _ = self._pad(x_train1, self.hparams.pad_id)
if (self.train_index_0 >= max(self.n_train_batches_0, self.n_train_batches_1)) or (self.train_index_1 >= max(self.n_train_batches_0, self.n_train_batches_1)):
self.reset_train()
eop = True
else:
eop = False
self.train_index_0 %= self.n_train_batches_0
self.train_index_1 %= self.n_train_batches_1
batch0, len0 = x_train0.size()
batch1, len1 = x_train1.size()
topk_logit0 = torch.zeros(batch0, len0, self.k)
topk_index0 = torch.zeros(batch0, len0, self.k , dtype = torch.long)
topk_logit1 = torch.zeros(batch1, len1, self.k)
topk_index1 = torch.zeros(batch1, len1, self.k , dtype = torch.long)
for i, j in zip(range(start_index_0, end_index_0, 1), range(end_index_0 - start_index_0)):
topk_logits0, topk_inds0 = load_topk(self.topk_db0[str(int(self.index0[i]))])
topk_logits0 = topk_logits0[:self.hparams.max_length, :self.k].float()
#print(topk_logits0.size())
#print(topk_logit0.size())
topk_inds0 = topk_inds0[:self.hparams.max_length, :self.k]
topk_logit0.data[j, :topk_logits0.size(0), :] = topk_logits0.data
topk_index0.data[j, :topk_inds0.size(0), :] = topk_inds0.data
for i, j in zip(range(start_index_1, end_index_1, 1), range(end_index_1 - start_index_1)):
topk_logits1, topk_inds1 = load_topk(self.topk_db1[str(int(self.index1[i]))])
topk_logits1 = topk_logits1[:self.hparams.max_length, :self.k].float()
#print(topk_logits1.size())
#print(topk_logit1.size())
topk_inds1 = topk_inds1[:self.hparams.max_length, :self.k]
topk_logit1.data[j, :topk_logits1.size(0), :] = topk_logits1.data
topk_index1.data[j, :topk_inds1.size(0), :] = topk_inds1.data
if torch.cuda.is_available():
topk_logit0 = topk_logit0.cuda()
topk_logit1 = topk_logit1.cuda()
topk_index0 = topk_index0.cuda()
topk_index1 = topk_index1.cuda()
return (x_train0, x_train1, topk_logit0, topk_logit1, topk_index0, topk_index1), batch_size, eop
def sample_y(self):
# first how many attrs?
attn_num = random.randint(1, (self.hparams.trg_vocab_size-1)//2)
# then select attrs
y = np.random.binomial(1, 0.5, attn_num)
y = y + np.arange(attn_num) * 2
return y.tolist()
def next_dev0(self, dev_batch_size=1, sort=True):
start_index = self.dev_index_0
end_index = min(start_index + dev_batch_size, self.dev_size_0)
batch_size = end_index - start_index
x_dev0 = self.dev_x0[start_index:end_index]
y_dev0 = self.dev_y0[start_index:end_index]
if sort:
x_dev0, y_dev0, _ = self.sort_by_xlen(x_dev0, y_dev0)
x_dev0, _, _, _, _ = self._pad(x_dev0, self.hparams.pad_id)
if end_index >= self.dev_size_0:
eop = True
self.dev_index_0 = 0
else:
eop = False
self.dev_index_0 += batch_size
return x_dev0, batch_size, eop
def next_dev1(self, dev_batch_size=1, sort=True):
start_index = self.dev_index_1
end_index = min(start_index + dev_batch_size, self.dev_size_1)
batch_size = end_index - start_index
x_dev1 = self.dev_x1[start_index:end_index]
y_dev1 = self.dev_y1[start_index:end_index]
if sort:
x_dev1, y_dev1, _ = self.sort_by_xlen(x_dev1, y_dev1)
x_dev1, _, _, _, _ = self._pad(x_dev1, self.hparams.pad_id)
if end_index >= self.dev_size_1:
eop = True
self.dev_index_1 = 0
else:
eop = False
self.dev_index_1 += batch_size
return x_dev1, batch_size, eop
def reset_test(self, test_src_file, test_trg_file):
self.test_x, self.test_y, src_len, _ = self._build_parallel(test_src_file, test_trg_file, is_train=False, insert_bos=True)
self.test_size = len(self.test_x)
self.test_index = 0
def next_test(self, test_batch_size=10):
start_index = self.test_index
end_index = min(start_index + test_batch_size, self.test_size)
batch_size = end_index - start_index
x_test = self.test_x[start_index:end_index]
y_test = self.test_y[start_index:end_index]
x_test, y_test, index = self.sort_by_xlen(x_test, y_test)
x_test, x_mask, x_count, x_len, x_pos_emb_idxs = self._pad(x_test, self.hparams.pad_id)
y_test, y_mask, y_count, y_len, y_pos_emb_idxs = self._pad(y_test, self.hparams.trg_pad_id)
y_neg = 1 - y_test
if end_index >= self.test_size:
eop = True
self.test_index = 0
else:
eop = False
self.test_index += batch_size
return x_test, x_mask, x_count, x_len, x_pos_emb_idxs, y_test, y_mask, y_count, y_len, y_pos_emb_idxs, y_neg, batch_size, eop, index
def sort_by_xlen(self, x, y, x_char_kv=None, y_char_kv=None, file_index=None, descend=True):
x = np.array(x)
y = np.array(y)
x_len = [len(i) for i in x]
index = np.argsort(x_len)
if descend:
index = index[::-1]
x, y = x[index].tolist(), y[index].tolist()
return x, y, index
def _pad(self, sentences, pad_id, char_kv=None, char_dim=None, char_sents=None):
batch_size = len(sentences)
lengths = [len(s) for s in sentences]
count = sum(lengths)
max_len = max(lengths)
padded_sentences = [s + ([pad_id]*(max_len - len(s))) for s in sentences]
mask = [[0]*len(s) + [1]*(max_len - len(s)) for s in sentences]
padded_sentences = torch.LongTensor(padded_sentences)
mask = torch.ByteTensor(mask)
pos_emb_indices = [[i+1 for i in range(len(s))] + ([0]*(max_len - len(s))) for s in sentences]
pos_emb_indices = torch.FloatTensor(pos_emb_indices)
if torch.cuda.is_available():
padded_sentences = padded_sentences.cuda()
pos_emb_indices = pos_emb_indices.cuda()
mask = mask.cuda()
return padded_sentences, mask, count, lengths, pos_emb_indices
def _build_parallel(self, src_file_name, trg_file_name, is_train=True, insert_bos = False):
print("loading parallel sentences from {} {}".format(src_file_name, trg_file_name))
with open(src_file_name, 'r', encoding='utf-8') as f:
src_lines = f.read().split('\n')
with open(trg_file_name, 'r', encoding='utf-8') as f:
trg_lines = f.read().split('\n')
src_data = []
trg_data = []
line_count = 0
skip_line_count = 0
src_unk_count = 0
trg_unk_count = 0
src_lens = []
src_unk_id = self.hparams.unk_id
for src_line, trg_line in zip(src_lines, trg_lines):
src_tokens = self.tokenizer.tokenize(src_line)
trg_tokens = trg_line.split()
if not src_tokens or not trg_tokens:
skip_line_count += 1
continue
#if is_train and not self.hparams.decode and self.hparams.max_len and (len(src_tokens) > self.hparams.max_len or len(trg_tokens) > self.hparams.max_len):
# skip_line_count += 1
# continue
src_lens.append(len(src_tokens))
if insert_bos:
src_indices, trg_indices = [self.hparams.bos_id], []
src_indices += self.tokenizer.convert_tokens_to_ids(src_tokens)
else:
src_indices, trg_indices = [], []
src_indices += self.tokenizer.convert_tokens_to_ids(src_tokens)
if len(src_indices) >= self.hparams.max_length:
src_indices = src_indices[:(self.hparams.max_length - 1)]
trg_w2i = self.trg_w2i
for trg_tok in trg_tokens:
if trg_tok not in trg_w2i:
print("trg attribute cannot have oov!")
exit(0)
else:
trg_indices.append(trg_w2i[trg_tok])
# calculate char ngram emb for trg_tok
src_indices.append(self.hparams.eos_id)
src_data.append(src_indices)
trg_data.append(trg_indices)
line_count += 1
if line_count % 10000 == 0:
print("processed {} lines".format(line_count))
index = None
if is_train:
src_data, trg_data, index = self.sort_by_xlen(src_data, trg_data, descend=False)
print("src_unk={}, trg_unk={}".format(src_unk_count, trg_unk_count))
assert len(src_data) == len(trg_data)
print("lines={}, skipped_lines={}".format(len(src_data), skip_line_count))
return src_data, trg_data, src_lens, index
def _build_vocab(self, vocab_file, max_vocab_size=None):
i2w = []
w2i = {}
i = 0
with open(vocab_file, 'r', encoding='utf-8') as f:
for line in f:
w = line.strip()
#if i == 0 and w != "<pad>":
# i2w = ['<pad>', '<unk>', '<s>', '<\s>']
# w2i = {'<pad>': 0, '<unk>':1, '<s>':2, '<\s>':3}
# i = 4
w2i[w] = i
i2w.append(w)
i += 1
if max_vocab_size and i >= max_vocab_size:
break
#if "<pad>" not in w2i:
# w2i["<pad>"] = i
# i2w.append("<pad>")
#assert i2w[self.hparams.pad_id] == '<pad>'
#assert i2w[self.hparams.unk_id] == '<unk>'
#assert i2w[self.hparams.bos_id] == '<s>'
#assert i2w[self.hparams.eos_id] == '<\s>'
#assert w2i['<pad>'] == self.hparams.pad_id
#assert w2i['<unk>'] == self.hparams.unk_id
#assert w2i['<s>'] == self.hparams.bos_id
#assert w2i['<\s>'] == self.hparams.eos_id
return i2w, w2i
def load_topk(dump):
with io.BytesIO(dump) as reader:
topk = torch.load(reader)
return topk