-
Notifications
You must be signed in to change notification settings - Fork 35
/
dataset.py
254 lines (204 loc) · 9.86 KB
/
dataset.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
# encoding=utf8
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
reload(sys)
sys.setdefaultencoding('utf8')
import json
import h5py
import os
import os.path
import numpy as np
import random
import logging
import misc.utils as utils
import torch
import torch.utils.data as data
from torch.utils.data import DataLoader, Dataset
import eval_utils
class VISTDataset(Dataset):
def __init__(self, opt):
self.mode = 'train' # by default
self.opt = opt
self.task = opt.task # option: 'story_telling', 'image_captioning'
self.data_type = {
'whole_story': False,
'split_story': True,
'caption': False
}
# open the hdf5 file
print('DataLoader loading story h5 file: ', opt.story_h5)
self.story_h5 = h5py.File(opt.story_h5, 'r', driver='core')['story']
print("story's max sentence length is ", self.story_h5.shape[1])
print('DataLoader loading story h5 file: ', opt.full_story_h5)
self.full_story_h5 = h5py.File(opt.full_story_h5, 'r', driver='core')['story']
print("full story's max sentence length is ", self.full_story_h5.shape[1])
print('DataLoader loading description h5 file: ', opt.desc_h5)
self.desc_h5 = h5py.File(opt.desc_h5, 'r', driver='core')['story']
print("caption's max sentence length is ", self.desc_h5.shape[1])
print('DataLoader loading story_line json file: ', opt.story_line_json)
self.story_line = json.load(open(opt.story_line_json))
self.id2word = self.story_line['id2words']
print("vocab[0] = ", self.id2word['0'])
print("vocab[1] = ", self.id2word['1'])
self.word2id = self.story_line['words2id']
self.vocab_size = len(self.id2word)
print('vocab size is ', self.vocab_size)
self.story_ids = {'train': [], 'val': [], 'test': []}
self.description_ids = {'train': [], 'val': [], 'test': []}
self.story_ids['train'] = self.story_line['train'].keys()
self.story_ids['val'] = self.story_line['val'].keys()
self.story_ids['test'] = self.story_line['test'].keys()
self.description_ids['train'] = self.story_line['image2caption']['train'].keys()
self.description_ids['val'] = self.story_line['image2caption']['val'].keys()
self.description_ids['test'] = self.story_line['image2caption']['test'].keys()
print('There are {} training data, {} validation data, and {} test data'.format(len(self.story_ids['train']),
len(self.story_ids['val']),
len(self.story_ids['test'])))
ref_dir = 'data/reference'
if not os.path.exists(ref_dir):
os.makedirs(ref_dir)
# write reference files for storytelling
for split in ['val', 'test']:
reference = {}
for story in self.story_line[split].values():
if story['album_id'] not in reference:
reference[story['album_id']] = [story['origin_text']]
else:
reference[story['album_id']].append(story['origin_text'])
with open(os.path.join(ref_dir, '{}_reference.json'.format(split)), 'w') as f:
json.dump(reference, f)
# write reference files for captioning
for split in ['val', 'test']:
reference = {}
for flickr_id, story in self.story_line['image2caption_original'][split].iteritems():
reference[flickr_id] = story
with open(os.path.join(ref_dir, '{}_desc_reference.json'.format(split)), 'w') as f:
json.dump(reference, f)
def __getitem__(self, index):
if self.task == 'story_telling':
story_id = self.story_ids[self.mode][index]
story = self.story_line[self.mode][story_id]
# load feature
feature_fc = np.zeros((story['length'], self.opt.feat_size), dtype='float32')
feature_conv = np.zeros((story['length'], self.opt.conv_feat_size), dtype='float32')
for i in xrange(story['length']):
# load fc feature
fc_path = os.path.join(self.opt.data_dir, 'resnet_features/fc', self.mode,
'{}.npy'.format(story['flickr_id'][i]))
feature_fc[i] = np.load(fc_path)
if self.opt.use_conv:
conv_path = os.path.join(self.opt.data_dir, 'resnet_features/conv', self.mode,
'{}.npz'.format(story['flickr_id'][i]))
feature_conv[i] = np.load(conv_path)['arr_0'].flatten()
sample = {'feature_fc': feature_fc}
if self.opt.use_conv:
sample['feature_conv'] = feature_conv
# load story
if self.data_type['whole_story']:
whole_story = self.full_story_h5[story['whole_text_index']]
sample['whole_story'] = np.int64(whole_story)
if self.data_type['split_story']:
split_story = self.story_h5[story['text_index']]
sample['split_story'] = np.int64(split_story)
# load caption
if self.data_type['caption']:
caption = []
for flickr_id in story['flickr_id']:
if flickr_id in self.story_line['image2caption'][self.mode]:
descriptions = self.story_line['image2caption'][self.mode][flickr_id]['caption']
random_idx = np.random.choice(len(descriptions), 1)[0]
caption.append(self.desc_h5[descriptions[random_idx]])
else:
caption.append(np.zeros((self.desc_h5.shape[1],), dtype='int64'))
sample['caption'] = np.asarray(caption, 'int64')
sample['index'] = np.int64(index)
return sample
elif self.task == "image_captioning":
flickr_id = self.description_ids[self.mode][index]
descriptions = self.story_line['image2caption'][self.mode][flickr_id]['caption']
random_idx = np.random.choice(len(descriptions), 1)[0]
description = descriptions[random_idx]
fc_path = os.path.join(self.opt.data_dir, 'resnet_features/fc', self.mode,
'{}{}.npy'.format(self.opt.prefix, flickr_id))
conv_path = os.path.join(self.opt.data_dir, 'resnet_features/conv', self.mode, '{}.npy'.format(flickr_id))
feature_fc = np.load(fc_path)
feature_conv = np.load(conv_path).flatten()
sample = {'feature_fc': feature_fc, 'feature_conv': feature_conv}
target = np.int64(self.desc_h5[description])
sample['whole_story'] = target
sample['mask'] = np.zeros_like(target, dtype='float32')
nonzero_num = (target != 0).sum() + 1
sample['mask'][:nonzero_num] = 1
sample['index'] = np.int64(index)
return sample
def __len__(self):
if self.task == 'story_telling':
return len(self.story_ids[self.mode])
elif self.task == 'image_captioning':
return len(self.description_ids[self.mode])
else:
raise Exception("{} task is not proper for this dataset.".format(self.task))
def train(self):
self.mode = 'train'
def val(self):
self.mode = 'val'
def test(self):
self.mode = 'test'
def set_option(self, data_type=None):
if self.task == 'story_telling':
if data_type is not None:
self.data_type = data_type
else:
pass
# logging.error("{} task is not proper for this dataset.".format(self.task))
# raise Exception("{} task is not proper for this dataset.".format(self.task))
def get_GT(self, index):
if self.task == 'story_telling':
story_id = self.story_ids[self.mode][index]
story = self.story_line[self.mode][story_id]
return story['origin_text']
elif self.task == 'image_captioning':
raise Exception("To be implemented.")
else:
raise Exception("{} task is not proper for this dataset.".format(self.task))
def get_id(self, index):
if self.task == 'story_telling':
story_id = self.story_ids[self.mode][index]
return self.story_line[self.mode][story_id]['album_id'], self.story_line[self.mode][story_id]['flickr_id']
else:
return self.description_ids[self.mode][index]
def get_all_id(self, index):
story_id = self.story_ids[self.mode][index]
return self.story_line[self.mode][story_id]['album_id'], self.story_line[self.mode][story_id]['flickr_id']
def get_vocab_size(self):
return self.vocab_size
def get_vocab(self):
return self.id2word
def get_word2id(self):
return self.word2id
def get_whole_story_length(self):
return self.full_story_h5.shape[1]
def get_story_length(self):
return self.story_h5.shape[1]
def get_caption_length(self):
return self.desc_h5.shape[1]
if __name__ == "__main__":
import sys
import os
import opts
import time
start = time.time()
opt = opts.parse_opt()
dataset = VISTDataset(opt)
print("dataset finished: ", time.time() - start)
start = time.time()
dataset.train()
train_loader = DataLoader(dataset, batch_size=64, shuffle=opt.shuffle, num_workers=8)
print("dataloader finished: ", time.time() - start)
dataset.train() # train() mode has to be called before using train_loader
for iter, batch in enumerate(train_loader):
print("enumerate: ", time.time() - start)
print(iter)
print(batch[0].size())