-
Notifications
You must be signed in to change notification settings - Fork 43
/
Copy pathdataLoader.py
executable file
·278 lines (219 loc) · 9.68 KB
/
dataLoader.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
import torch.utils.data as data
from PIL import Image
import torch
import numpy as np
import h5py
import json
import pdb
import random
from misc.utils import repackage_hidden, clip_gradient, adjust_learning_rate, decode_txt
class train(data.Dataset): # torch wrapper
def __init__(self, input_img_h5, input_ques_h5, input_json, negative_sample, num_val, data_split):
print('DataLoader loading: %s' %data_split)
print('Loading image feature from %s' %input_img_h5)
if data_split == 'test':
split = 'val'
else:
split = 'train' # train and val split both corresponding to 'train'
f = json.load(open(input_json, 'r'))
self.itow = f['itow']
self.img_info = f['img_'+split]
# get the data split.
total_num = len(self.img_info)
if data_split == 'train':
s = 0
e = total_num - num_val
elif data_split == 'val':
s = total_num - num_val
e = total_num
else:
s = 0
e = total_num
self.img_info = self.img_info[s:e]
print('%s number of data: %d' %(data_split, e-s))
# load the data.
f = h5py.File(input_img_h5, 'r')
self.imgs = f['images_'+split][s:e]
f.close()
print('Loading txt from %s' %input_ques_h5)
f = h5py.File(input_ques_h5, 'r')
self.ques = f['ques_'+split][s:e]
self.ans = f['ans_'+split][s:e]
self.cap = f['cap_'+split][s:e]
self.ques_len = f['ques_len_'+split][s:e]
self.ans_len = f['ans_len_'+split][s:e]
self.cap_len = f['cap_len_'+split][s:e]
self.ans_ids = f['ans_index_'+split][s:e]
self.opt_ids = f['opt_'+split][s:e]
self.opt_list = f['opt_list_'+split][:]
self.opt_len = f['opt_len_'+split][:]
f.close()
self.ques_length = self.ques.shape[2]
self.ans_length = self.ans.shape[2]
self.his_length = self.ques_length + self.ans_length
self.vocab_size = len(self.itow)+1
print('Vocab Size: %d' % self.vocab_size)
self.split = split
self.rnd = 10
self.negative_sample = negative_sample
def __getitem__(self, index):
# get the image
img = torch.from_numpy(self.imgs[index])
# get the history
his = np.zeros((self.rnd, self.his_length))
his[0,self.his_length-self.cap_len[index]:] = self.cap[index,:self.cap_len[index]]
ques = np.zeros((self.rnd, self.ques_length))
ans = np.zeros((self.rnd, self.ans_length+1))
ans_target = np.zeros((self.rnd, self.ans_length+1))
ques_ori = np.zeros((self.rnd, self.ques_length))
opt_ans = np.zeros((self.rnd, self.negative_sample, self.ans_length+1))
ans_len = np.zeros((self.rnd))
opt_ans_len = np.zeros((self.rnd, self.negative_sample))
ans_idx = np.zeros((self.rnd))
opt_ans_idx = np.zeros((self.rnd, self.negative_sample))
for i in range(self.rnd):
# get the index
q_len = self.ques_len[index, i]
a_len = self.ans_len[index, i]
qa_len = q_len + a_len
if i+1 < self.rnd:
his[i+1, self.his_length-qa_len:self.his_length-a_len] = self.ques[index, i, :q_len]
his[i+1, self.his_length-a_len:] = self.ans[index, i, :a_len]
ques[i, self.ques_length-q_len:] = self.ques[index, i, :q_len]
ques_ori[i, :q_len] = self.ques[index, i, :q_len]
ans[i, 1:a_len+1] = self.ans[index, i, :a_len]
ans[i, 0] = self.vocab_size
ans_target[i, :a_len] = self.ans[index, i, :a_len]
ans_target[i, a_len] = self.vocab_size
ans_len[i] = self.ans_len[index, i]
opt_ids = self.opt_ids[index, i] # since python start from 0
# random select the negative samples.
ans_idx[i] = opt_ids[self.ans_ids[index, i]]
# exclude the gt index.
opt_ids = np.delete(opt_ids, ans_idx[i], 0)
random.shuffle(opt_ids)
for j in range(self.negative_sample):
ids = opt_ids[j]
opt_ans_idx[i,j] = ids
opt_len = self.opt_len[ids]
opt_ans_len[i, j] = opt_len
opt_ans[i, j, :opt_len] = self.opt_list[ids,:opt_len]
opt_ans[i, j, opt_len] = self.vocab_size
his = torch.from_numpy(his)
ques = torch.from_numpy(ques)
ans = torch.from_numpy(ans)
ans_target = torch.from_numpy(ans_target)
ques_ori = torch.from_numpy(ques_ori)
ans_len = torch.from_numpy(ans_len)
opt_ans_len = torch.from_numpy(opt_ans_len)
opt_ans = torch.from_numpy(opt_ans)
ans_idx = torch.from_numpy(ans_idx)
opt_ans_idx = torch.from_numpy(opt_ans_idx)
return img, his, ques, ans, ans_target, ans_len, ans_idx, ques_ori, \
opt_ans, opt_ans_len, opt_ans_idx
def __len__(self):
return self.ques.shape[0]
class validate(data.Dataset): # torch wrapper
def __init__(self, input_img_h5, input_ques_h5, input_json, negative_sample, num_val, data_split):
print('DataLoader loading: %s' %data_split)
print('Loading image feature from %s' %input_img_h5)
if data_split == 'test':
split = 'val'
else:
split = 'train' # train and val split both corresponding to 'train'
f = json.load(open(input_json, 'r'))
self.itow = f['itow']
self.img_info = f['img_'+split]
# get the data split.
total_num = len(self.img_info)
if data_split == 'train':
s = 0
e = total_num - num_val
elif data_split == 'val':
s = total_num - num_val
e = total_num
else:
s = 0
e = total_num
self.img_info = self.img_info[s:e]
print('%s number of data: %d' %(data_split, e-s))
# load the data.
f = h5py.File(input_img_h5, 'r')
self.imgs = f['images_'+split][s:e]
f.close()
print('Loading txt from %s' %input_ques_h5)
f = h5py.File(input_ques_h5, 'r')
self.ques = f['ques_'+split][s:e]
self.ans = f['ans_'+split][s:e]
self.cap = f['cap_'+split][s:e]
self.ques_len = f['ques_len_'+split][s:e]
self.ans_len = f['ans_len_'+split][s:e]
self.cap_len = f['cap_len_'+split][s:e]
self.ans_ids = f['ans_index_'+split][s:e]
self.opt_ids = f['opt_'+split][s:e]
self.opt_list = f['opt_list_'+split][:]
self.opt_len = f['opt_len_'+split][:]
f.close()
self.ques_length = self.ques.shape[2]
self.ans_length = self.ans.shape[2]
self.his_length = self.ques_length + self.ans_length
self.vocab_size = len(self.itow)+1
print('Vocab Size: %d' % self.vocab_size)
self.split = split
self.rnd = 10
self.negative_sample = negative_sample
def __getitem__(self, index):
# get the image
img_id = self.img_info[index]['imgId']
img = torch.from_numpy(self.imgs[index])
# get the history
his = np.zeros((self.rnd, self.his_length))
his[0,self.his_length-self.cap_len[index]:] = self.cap[index,:self.cap_len[index]]
ques = np.zeros((self.rnd, self.ques_length))
ans = np.zeros((self.rnd, self.ans_length+1))
ans_target = np.zeros((self.rnd, self.ans_length+1))
quesL = np.zeros((self.rnd, self.ques_length))
opt_ans = np.zeros((self.rnd, 100, self.ans_length+1))
ans_ids = np.zeros(self.rnd)
opt_ans_target = np.zeros((self.rnd, 100, self.ans_length+1))
ans_len = np.zeros((self.rnd))
opt_ans_len = np.zeros((self.rnd, 100))
for i in range(self.rnd):
# get the index
q_len = self.ques_len[index, i]
a_len = self.ans_len[index, i]
qa_len = q_len + a_len
if i+1 < self.rnd:
ques_ans = np.concatenate([self.ques[index, i, :q_len], self.ans[index, i, :a_len]])
his[i+1, self.his_length-qa_len:] = ques_ans
ques[i, self.ques_length-q_len:] = self.ques[index, i, :q_len]
quesL[i, :q_len] = self.ques[index, i, :q_len]
ans[i, 1:a_len+1] = self.ans[index, i, :a_len]
ans[i, 0] = self.vocab_size
ans_target[i, :a_len] = self.ans[index, i, :a_len]
ans_target[i, a_len] = self.vocab_size
ans_ids[i] = self.ans_ids[index, i] # since python start from 0
opt_ids = self.opt_ids[index, i] # since python start from 0
ans_len[i] = self.ans_len[index, i]
ans_idx = self.ans_ids[index, i]
for j, ids in enumerate(opt_ids):
opt_len = self.opt_len[ids]
opt_ans[i, j, 1:opt_len+1] = self.opt_list[ids,:opt_len]
opt_ans[i, j, 0] = self.vocab_size
opt_ans_target[i, j,:opt_len] = self.opt_list[ids,:opt_len]
opt_ans_target[i, j,opt_len] = self.vocab_size
opt_ans_len[i, j] = opt_len
opt_ans = torch.from_numpy(opt_ans)
opt_ans_target = torch.from_numpy(opt_ans_target)
ans_ids = torch.from_numpy(ans_ids)
his = torch.from_numpy(his)
ques = torch.from_numpy(ques)
ans = torch.from_numpy(ans)
ans_target = torch.from_numpy(ans_target)
quesL = torch.from_numpy(quesL)
ans_len = torch.from_numpy(ans_len)
opt_ans_len = torch.from_numpy(opt_ans_len)
return img, his, ques, ans, ans_target, quesL, opt_ans, \
opt_ans_target, ans_ids, ans_len, opt_ans_len, img_id
def __len__(self):
return self.ques.shape[0]