-
Notifications
You must be signed in to change notification settings - Fork 0
/
data_utils.py
263 lines (216 loc) · 7.06 KB
/
data_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
import numpy as np
import collections
import os
import re
import csv
import sys
import _pickle as cPickle
from tensorflow.python.platform import gfile
import math
import json
try:
import cPickle as pickle
except ImportError:
import pickle
min_count = 3
add_count = 5
class BeamLink(object):
def __init__(self, value, h_value):
self.prev = None
self.hash_value = h_value
self.value = value
def set_prev(self, prev):
self.prev = prev
def get_path(self):
path = []
c_prev = self
while c_prev is not None:
path.append(c_prev.value)
c_prev = c_prev.prev
path.reverse()
return path
class Data(object):
def __init__(self, feats, encoder_in_idx, decoder_in, v_encoder_in, truth_captions, t_encoder_in, files):
self.length = encoder_in_idx.shape[0]
self.current = 0
self.feats = feats
self.encoder_in_idx = encoder_in_idx
self.decoder_in = decoder_in
# validation data
self.v_encoder_in = v_encoder_in
self.truth_captions = truth_captions
self.t_encoder_in = t_encoder_in
self.files = files
def next_batch(self, size):
if self.current == 0:
index = np.random.permutation(np.arange(self.length))
self.encoder_in_idx = self.encoder_in_idx[index]
self.decoder_in = self.decoder_in[index]
if self.current + size < self.length:
e_in_idx, d_in = self.encoder_in_idx[self.current:self.current+size], self.decoder_in[self.current:self.current+size,:-1]
d_out = self.decoder_in[self.current:self.current+size,1:]
e_in = self.feats[e_in_idx]
self.current += size
else:
e_in_idx, d_in = self.encoder_in_idx[self.current:], self.decoder_in[self.current:,:-1]
d_out = self.decoder_in[self.current:,1:]
e_in = self.feats[e_in_idx]
self.current = 0
return e_in, d_in, d_out
class VocabularyProcessor(object):
def __init__(self, max_document_length, vocabulary, unknown_limit=float('Inf'), drop=False):
self.max_document_length = max_document_length
self._reverse_mapping = ['<UNK>'] + vocabulary
self.make_mapping()
self.unknown_limit = unknown_limit
self.drop = drop
def make_mapping(self):
self._mapping = {}
for i, vocab in enumerate(self._reverse_mapping):
self._mapping[vocab] = i
def transform(self, raw_documents):
data = []
lengths = []
for tokens in raw_documents:
word_ids = np.ones(self.max_document_length, np.int32) * self._mapping['<EOS>']
length = 0
unknown = 0
if self.drop and len(tokens.split()) > self.max_document_length:
continue
for idx, token in enumerate(tokens.split()):
if idx >= self.max_document_length:
break
word_ids[idx] = self._mapping.get(token, 0)
length = idx
if word_ids[idx] == 0:
unknown += 1
length = length+1
if unknown <= self.unknown_limit:
data.append(word_ids)
lengths.append(length)
data = np.array(data)
lengths = np.array(lengths)
return data, lengths
# yield word_ids
def save(self, filename):
with gfile.Open(filename, 'wb') as f:
f.write(pickle.dumps(self))
@classmethod
def restore(cls, filename):
with gfile.Open(filename, 'rb') as f:
return pickle.loads(f.read())
def clean_str(string):
string = re.sub(r"\.", r"", string)
return string.strip().lower()
# load and dump the mapped train and valid's captions
def load_text_data(train_lab, prepro_train_p, vocab_path):
tlab = json.load(open(train_lab, 'r'))
vocab_dict = collections.defaultdict(int)
train_dict = {}
for caps in tlab:
train_dict[caps['id']] = ['<BOS> '+clean_str(cap)+' <EOS>' for cap in caps['caption']]
# build vocabulary
maxlen = 0
avglen = 0
total_seq = 0
for cid, captions in train_dict.items():
for caption in captions:
s_caption = caption.split()
avglen += len(s_caption)
total_seq += 1
if len(s_caption) >= maxlen:
maxlen = len(s_caption)
for word in s_caption:
vocab_dict[word] += 1
vocabulary = []
for k, v in sorted(vocab_dict.items(), key=lambda x:x[1], reverse=True):
if v >= min_count:
vocabulary.append(k)
# map sequence to its id
vocab_processor = VocabularyProcessor(max_document_length=math.ceil(avglen/total_seq)+add_count, vocabulary=vocabulary, drop=True)
t_number = 0
min_t = float('Inf')
avg_t = 0
for cid, _ in train_dict.items():
train_c_dat, lengths = vocab_processor.transform(train_dict[cid])
train_dict[cid] = {'captions':train_c_dat, 'lengths':lengths}
t_number += len(train_c_dat)
if len(train_c_dat) < min_t:
min_t = len(train_c_dat)
cPickle.dump(train_dict, open(prepro_train_p, 'wb'))
vocab_processor.save(vocab_path)
print('init sequence number: {}'.format(total_seq))
print('maximum sequence length: {}'.format(maxlen))
print('average sequence length: {}'.format(avglen/total_seq))
print('drop length: > {}'.format(math.ceil(avglen/total_seq)+add_count))
print('remaining total train number: {}'.format(t_number))
print('total video number: {}'.format(len(train_dict)))
print('minimum train number: {} per video'.format(min_t))
print('average train number: {} per video'.format(t_number//len(train_dict)))
return vocab_processor, train_dict
def load_valid(valid_dir, valid_lab):
vlab = json.load(open(valid_lab, 'r'))
paths = []
feats = []
truth_captions = []
for caps in vlab:
feat_path = os.path.join(valid_dir, caps['id']+'.npy')
paths.append(feat_path)
truth_captions.append([clean_str(cap) for cap in caps['caption']])
for path in paths:
feat = np.load(path)
feats.append(feat)
return np.array(feats, dtype='float32'), truth_captions
def load_task(task_dir):
feats = []
paths = []
files = []
for dirPath, dirNames, fileNames in os.walk(task_dir):
for f in fileNames:
paths.append(os.path.join(task_dir, f))
files.append(f)
for path in paths:
feat = np.load(path)
feats.append(feat)
return np.array(feats, dtype='float32'), files
def gen_train_data(train_dir, train_lab, train_dict):
tlab = json.load(open(train_lab, 'r'))
paths = []
feats = []
for caps in tlab:
feat_path = os.path.join(train_dir, caps['id']+'.npy')
paths.append(feat_path)
for path in paths:
feat = np.load(path)
feats.append(feat)
# here instead of using feat value, we use idx of feat to indicate the feat (faster)
encoder_in_idx = []
decoder_in = []
for idx, caps in enumerate(tlab):
for d_f in train_dict[caps['id']]['captions']:
encoder_in_idx.append(idx)
decoder_in.append(d_f)
encoder_in_idx = np.array(encoder_in_idx, dtype='int32')
decoder_in = np.array(decoder_in, dtype='float32')
return np.array(feats) ,encoder_in_idx, decoder_in
def get_unknown_word_vec(dim_size):
return np.random.uniform(-0.25, 0.25, dim_size)
def build_w2v_matrix(vocab_processor, w2v_path, vector_path, dim_size):
w2v_dict = {}
f = open(vector_path, 'r')
for line in f.readlines():
word, vec = line.strip().split(' ', 1)
w2v_dict[word] = np.loadtxt([vec], dtype='float32')
vocab_list = vocab_processor._reverse_mapping
w2v_W = np.zeros(shape=(len(vocab_list), dim_size), dtype='float32')
for i, vocab in enumerate(vocab_list):
# unknown vocab
if i == 0:
continue
else:
if vocab in w2v_dict:
w2v_W[i] = w2v_dict[vocab]
else:
w2v_W[i] = get_unknown_word_vec(dim_size)
cPickle.dump(w2v_W, open(w2v_path, 'wb'))
return w2v_W