-
Notifications
You must be signed in to change notification settings - Fork 12
/
data.py
273 lines (232 loc) · 9.71 KB
/
data.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
import random, logging
import torch
from torch import nn
import numpy as np
from utils import move_to_device
PAD, UNK, BOS, EOS = '<PAD>', '<UNK>', '<BOS>', '<EOS>'
logger = logging.getLogger(__name__)
class Vocab(object):
def __init__(self, filename, min_occur_cnt, specials = None):
idx2token = [PAD, UNK] + (specials if specials is not None else [])
num_tot_tokens = 0
num_invocab_tokens = 0
for line in open(filename).readlines():
try:
token, cnt = line.rstrip('\n').split('\t')
cnt = int(cnt)
num_tot_tokens += cnt
except:
logger.info("(Vocab)Illegal line:", line)
if cnt >= min_occur_cnt:
idx2token.append(token)
num_invocab_tokens += cnt
self.coverage = num_invocab_tokens/num_tot_tokens
self._token2idx = dict(zip(idx2token, range(len(idx2token))))
self._idx2token = idx2token
self._padding_idx = self._token2idx[PAD]
self._unk_idx = self._token2idx[UNK]
@property
def size(self):
return len(self._idx2token)
@property
def unk_idx(self):
return self._unk_idx
@property
def padding_idx(self):
return self._padding_idx
def idx2token(self, x):
if isinstance(x, list):
return [self.idx2token(i) for i in x]
return self._idx2token[x]
def token2idx(self, x):
if isinstance(x, list):
return [self.token2idx(i) for i in x]
return self._token2idx.get(x, self.unk_idx)
def _back_to_txt_for_check(tensor, vocab, local_idx2token=None):
for bid, xs in enumerate(tensor.t().tolist()):
txt = []
for x in xs:
if x == vocab.padding_idx:
break
if x >= vocab.size:
assert local_idx2token is not None
assert local_idx2token[bid] is not None
tok = local_idx2token[bid][x]
else:
tok = vocab.idx2token(x)
txt.append(tok)
txt = ' '.join(txt)
print (txt)
print ('-'*55)
print ('='*55)
def ListsToTensor(xs, vocab=None, worddrop=0., local_vocabs=None):
pad = vocab.padding_idx if vocab else 0
def toIdx(w, i):
if vocab is None:
return w
if isinstance(w, list):
return [toIdx(_, i) for _ in w]
if random.random() < worddrop:
return vocab.unk_idx
if local_vocabs is not None:
local_vocab = local_vocabs[i]
if (local_vocab is not None) and (w in local_vocab):
return local_vocab[w]
return vocab.token2idx(w)
max_len = max(len(x) for x in xs)
ys = []
for i, x in enumerate(xs):
y = toIdx(x, i) + [pad]*(max_len-len(x))
ys.append(y)
data = np.transpose(np.array(ys))
return data
def ArraysToTensor(xs):
"list of numpy array, each has the same demonsionality"
x = np.array([ list(x.shape) for x in xs])
shape = [len(xs)] + list(x.max(axis = 0))
data = np.zeros(shape, dtype=np.int)
for i, x in enumerate(xs):
slicing_shape = list(x.shape)
slices = tuple([slice(i, i+1)]+[slice(0, x) for x in slicing_shape])
data[slices] = x
#tensor = torch.from_numpy(data).long()
return data
def batchify(data, vocabs, max_seq_len):
src_tokens = [ [BOS]+x['src_tokens'][:max_seq_len] for x in data]
tgt_tokens_in = [[BOS]+x['tgt_tokens'][:max_seq_len] for x in data]
tgt_tokens_out = [x['tgt_tokens'][:max_seq_len]+[EOS] for x in data]
src_token = ListsToTensor(src_tokens, vocabs['src'])
tgt_token_in = ListsToTensor(tgt_tokens_in, vocabs['tgt'])
tgt_token_out = ListsToTensor(tgt_tokens_out, vocabs['tgt'])
not_padding = (tgt_token_out != vocabs['tgt'].padding_idx).astype(np.int64)
tgt_lengths = np.sum(not_padding, axis=0)
tgt_num_tokens = int(np.sum(tgt_lengths))
#not_padding = (src_token != vocabs['src'].padding_idx).astype(np.int64)
#src_lengths = np.sum(not_padding, axis=0)
ret = {
'src_tokens': src_token,
#'src_lengths': src_lengths,
'tgt_tokens_in': tgt_token_in,
'tgt_tokens_out': tgt_token_out,
'tgt_num_tokens': tgt_num_tokens,
#'tgt_lengths': tgt_lengths,
'tgt_raw_sents': [x['tgt_tokens'] for x in data],
'indices': [x['index'] for x in data]
}
# only if there is some memory input
if data[0]['mem_sents']:
num_mem_sents = len(data[0]['mem_sents'])
for x in data:
assert len(x['mem_sents']) == len(x['mem_scores']) == num_mem_sents
# put all memory tokens (across the same batch) in a single list:
# No.1 memory for sentence 1, ..., No.1 memory for sentence N,
# No.2 memory for sentence 1, ...
all_mem_tokens = []
all_mem_scores = []
for i in range(num_mem_sents):
all_mem_tokens.extend([ [BOS]+x['mem_sents'][i][:max_seq_len] for x in data])
all_mem_scores.extend([x['mem_scores'][i] for x in data])
# then convert to tensors:
# all_mem_tokens -> seq_len x (num_mem_sents * bsz)
# all_mem_scores -> num_mem_sents * bsz
ret['all_mem_tokens'] = ListsToTensor(all_mem_tokens, vocabs['tgt'])
ret['all_mem_scores'] = np.array(all_mem_scores, dtype=np.float32)
# to avoid GPU OOM issue, truncate the mem to the max. length of 1.5 x src_tokens
max_mem_len = int(1.5 * src_token.shape[0])
ret['all_mem_tokens'] = ret['all_mem_tokens'][:max_mem_len,:]
#ret['retrieval_raw_sents'] = [x['mem_sents'] for x in data]
return ret
class DataLoader(object):
def __init__(self, vocabs, filename, batch_size, for_train, max_seq_len=256, rank=0, num_replica=1):
self.vocabs = vocabs
self.batch_size = batch_size
self.train = for_train
src_tokens, tgt_tokens = [], []
src_sizes, tgt_sizes = [], []
mem_sents, mem_scores = [], []
for line in open(filename).readlines()[rank::num_replica]:
try:
src, tgt, *mem = line.strip().split('\t')
except:
continue
src, tgt = src.split(), tgt.split()
#TODO please remove this part#
#if not for_train:
# if len(src)/len(tgt) > 1.5 or len(tgt)/len(src) > 1.5:
# continue
#TODO#
src_sizes.append(len(src))
tgt_sizes.append(len(tgt))
src_tokens.append(src)
tgt_tokens.append(tgt)
mem_sents.append([ref.split() for ref in mem[:-1:2]])
mem_scores.append([float(score) for score in mem[1::2]])
self.src = src_tokens
self.tgt = tgt_tokens
self.src_sizes = np.array(src_sizes)
self.tgt_sizes = np.array(tgt_sizes)
self.max_seq_len = max_seq_len
self.mem_sents = mem_sents
self.mem_scores = mem_scores
logger.info("(DataLoader rank %d) read %s file with %d paris. max src len: %d, max tgt len: %d", rank, filename, len(self.src), self.src_sizes.max(), self.tgt_sizes.max())
def __len__(self):
return len(self.src)
def __iter__(self):
if self.train:
indices = np.random.permutation(len(self))
else:
indices = np.arange(len(self))
indices = indices[np.argsort(self.tgt_sizes[indices], kind='mergesort')]
indices = indices[np.argsort(self.src_sizes[indices], kind='mergesort')]
batches = []
num_tokens, batch = 0, []
for i in indices:
num_tokens += 1 + max(self.src_sizes[i], self.tgt_sizes[i])
if num_tokens > self.batch_size:
batches.append(batch)
num_tokens, batch = 1 + max(self.src_sizes[i], self.tgt_sizes[i]), [i]
else:
batch.append(i)
if not self.train or num_tokens > self.batch_size/2:
batches.append(batch)
if self.train:
random.shuffle(batches)
for batch in batches:
data = []
for i in batch:
src_tokens = self.src[i]
tgt_tokens = self.tgt[i]
mem_sents = self.mem_sents[i]
mem_scores = self.mem_scores[i]
item = {'src_tokens':src_tokens, 'tgt_tokens':tgt_tokens, 'mem_sents':mem_sents, 'mem_scores':mem_scores, 'index':i}
data.append(item)
yield batchify(data, self.vocabs, self.max_seq_len)
def parse_config():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--src_vocab', type=str, default='es.vocab')
parser.add_argument('--tgt_vocab', type=str, default='en.vocab')
parser.add_argument('--train_data', type=str, default='dev.mem.txt')
parser.add_argument('--train_batch_size', type=int, default=4096)
return parser.parse_args()
if __name__ == '__main__':
args = parse_config()
vocabs = dict()
vocabs['src'] = Vocab(args.src_vocab, 0, [EOS])
vocabs['tgt'] = Vocab(args.tgt_vocab, 0, [BOS, EOS])
train_data = DataLoader(vocabs, args.train_data, args.train_batch_size, for_train=True)
for d in train_data:
d = move_to_device(d, torch.device('cpu'))
for k, v in d.items():
if 'raw' in k:
continue
try:
print (k, v.shape)
except:
print (k, v)
_back_to_txt_for_check(d['src_tokens'][:,5:6], vocabs['src'])
_back_to_txt_for_check(d['tgt_tokens_in'][:,5:6], vocabs['tgt'])
_back_to_txt_for_check(d['tgt_tokens_out'][:,5:6], vocabs['tgt'])
bsz = d['tgt_tokens_out'].size(1)
_back_to_txt_for_check(d['all_mem_tokens'][:,5::bsz], vocabs['tgt'])
break