-
Notifications
You must be signed in to change notification settings - Fork 15
/
dataloader.py
394 lines (319 loc) · 15.3 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
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
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import json
import h5py
import os
import numpy as np
import random
import gc
import torch
import torch.utils.data as data
import multiprocessing
def get_npy_data(ix, fc_file, att_file, use_att):
if use_att == True:
return (np.load(fc_file), np.load(att_file)['feat'], ix)
else:
return (np.load(fc_file), np.zeros((1,1,1)), ix)
'''
Load data from h5 files
'''
class DataLoader():
def reset_iterator(self, split):
# if load files from directory, then reset the prefetch process
self.iterators[split] = 0
def get_vocab_size(self):
return self.vocab_size
def get_vocab(self):
vocabulary = self.ix_to_word
return vocabulary
def get_seq_length(self):
return self.seq_length
def __init__(self, opt):
self.type = 'h5'
self.opt = opt
self.att_im = getattr(opt, 'att_im', 1)
self.fc_feat_size = opt.fc_feat_size
self.att_feat_size = opt.att_feat_size
self.batch_size = opt.batch_size
self.seq_per_img = opt.seq_per_img
# load the json file which contains additional information about the dataset
print('DataLoader loading json file: ', opt.input_json)
self.info = json.load(open(self.opt.input_json))
self.ix_to_word = self.info['ix_to_word']
self.vocab_size = len(self.ix_to_word)
print('vocab size is ', self.vocab_size)
# open the hdf5 file
self.h5_fc_file = h5py.File(self.opt.input_fc_h5)
self.h5_att_file = h5py.File(self.opt.input_att_h5)
# load in the sequence data
self.h5_label_file = h5py.File(self.opt.input_label_h5, 'r', driver='core')
seq_size = self.h5_label_file['labels'].shape
self.seq_length = seq_size[1]
print('max sequence length in data is', self.seq_length)
# load the pointers in full to RAM (should be small enough)
self.label_start_ix = self.h5_label_file['label_start_ix'][:]
self.label_end_ix = self.h5_label_file['label_end_ix'][:]
self.num_images = self.label_start_ix.shape[0]
print('read %d image / features' %(self.num_images))
# separate out indexes for each of the provided splits
self.split_ix = {'train': [], 'val': [], 'test': []}
for ix in range(len(self.info['images'])):
img = self.info['images'][ix]
if img['split'] == 'train':
self.split_ix['train'].append(ix)
elif img['split'] == 'val':
self.split_ix['val'].append(ix)
elif img['split'] == 'test':
self.split_ix['test'].append(ix)
elif opt.train_only == 0: # restval
self.split_ix['train'].append(ix)
print('assigned %d images to split train' %len(self.split_ix['train']))
print('assigned %d images to split val' %len(self.split_ix['val']))
print('assigned %d images to split test' %len(self.split_ix['test']))
self.iterators = {'train': 0, 'val': 0, 'test': 0}
def get_batch(self, split, batch_size=None, seq_per_img=None):
split_ix = self.split_ix[split]
batch_size = batch_size or self.batch_size
seq_per_img = seq_per_img or self.seq_per_img
fc_batch = np.ndarray((batch_size, self.fc_feat_size), dtype = 'float32')
att_batch = np.ndarray((batch_size, 14, 14, self.att_feat_size), dtype = 'float32') if self.att_im else None
label_batch = np.zeros([batch_size * self.seq_per_img, self.seq_length + 2], dtype = 'int')
mask_batch = np.zeros([batch_size * self.seq_per_img, self.seq_length + 2], dtype = 'float32')
max_index = len(split_ix)
wrapped = False
infos = []
gts = []
for i in range(batch_size):
import time
t_start = time.time()
ri = self.iterators[split]
ri_next = ri + 1
if ri_next >= max_index:
ri_next = 0
wrapped = True
self.iterators[split] = ri_next
ix = split_ix[ri]
# fetch image
fc_batch[i] = self.h5_fc_file['fc'][ix, :]
att_batch[i] = self.h5_att_file['att'][ix, :, :, :]
# fetch the sequence labels
ix1 = self.label_start_ix[ix] - 1 #label_start_ix starts from 1
ix2 = self.label_end_ix[ix] - 1
ncap = ix2 - ix1 + 1 # number of captions available for this image
assert ncap > 0, 'an image does not have any label. this can be handled but right now isn\'t'
if ncap < self.seq_per_img:
# we need to subsample (with replacement)
seq = np.zeros([self.seq_per_img, self.seq_length], dtype = 'int')
for q in range(self.seq_per_img):
ixl = random.randint(ix1,ix2)
seq[q, :] = self.h5_label_file['labels'][ixl, :self.seq_length]
else:
ixl = random.randint(ix1, ix2 - self.seq_per_img + 1)
seq = self.h5_label_file['labels'][ixl: ixl + self.seq_per_img, :self.seq_length]
label_batch[i * self.seq_per_img : (i + 1) * self.seq_per_img, 1 : self.seq_length + 1] = seq
# Used for reward evaluation
gts_labels = self.h5_label_file['labels'][self.label_start_ix[ix] - 1: self.label_end_ix[ix]]
gts.append(gts_labels)
# record associated info as well
info_dict = {}
info_dict['ix'] = ix
info_dict['id'] = self.info['images'][ix]['id']
info_dict['file_path'] = self.info['images'][ix]['file_path']
infos.append(info_dict)
# generate mask
t_start = time.time()
nonzeros = np.array(map(lambda x: (x != 0).sum()+2, label_batch))
for ix, row in enumerate(mask_batch):
row[:nonzeros[ix]] = 1
data = {}
data['fc_feats'] = fc_batch
data['att_feats'] = att_batch
data['labels'] = label_batch
data['gts'] = gts
data['masks'] = mask_batch
data['bounds'] = {'it_pos_now': self.iterators[split], 'it_max': len(split_ix), 'wrapped': wrapped}
data['infos'] = infos
gc.collect()
return data
class DataLoader_pool(data.Dataset):
def reset_iterator(self, split):
del self._prefetch_process[split]
self._prefetch_process[split] = BlobFetcher(split, self, split=='train')
self.iterators[split] = 0
def get_vocab_size(self):
return self.vocab_size
def get_vocab(self):
return self.ix_to_word
def get_seq_length(self):
return self.seq_length
def __init__(self, opt):
self.opt = opt
self.batch_size = self.opt.batch_size
self.seq_per_img = opt.seq_per_img
self.use_att = getattr(opt, 'use_att', True)
# load the json file which contains additional information about the dataset
print('DataLoader loading json file: ', opt.input_json)
self.info = json.load(open(self.opt.input_json))
self.ix_to_word = self.info['ix_to_word']
self.vocab_size = len(self.ix_to_word)
print('vocab size is ', self.vocab_size)
# open the hdf5 file
print('DataLoader loading h5 file: ', opt.input_fc_dir, opt.input_att_dir, opt.input_label_h5)
self.h5_label_file = h5py.File(self.opt.input_label_h5, 'r', driver='core')
self.input_fc_dir = self.opt.input_fc_dir
self.input_att_dir = self.opt.input_att_dir
# load in the sequence data
seq_size = self.h5_label_file['labels'].shape
self.seq_length = seq_size[1]
print('max sequence length in data is', self.seq_length)
# load the pointers in full to RAM (should be small enough)
self.label_start_ix = self.h5_label_file['label_start_ix'][:]
self.label_end_ix = self.h5_label_file['label_end_ix'][:]
self.num_images = self.label_start_ix.shape[0]
print('read %d image features' %(self.num_images))
# separate out indexes for each of the provided splits
self.split_ix = {'train': [], 'val': [], 'test': []}
for ix in range(len(self.info['images'])):
img = self.info['images'][ix]
if img['split'] == 'train':
self.split_ix['train'].append(ix)
elif img['split'] == 'val':
self.split_ix['val'].append(ix)
elif img['split'] == 'test':
self.split_ix['test'].append(ix)
elif opt.train_only == 0: # restval
self.split_ix['train'].append(ix)
print('assigned %d images to split train' %len(self.split_ix['train']))
print('assigned %d images to split val' %len(self.split_ix['val']))
print('assigned %d images to split test' %len(self.split_ix['test']))
self.iterators = {'train': 0, 'val': 0, 'test': 0}
self._prefetch_process = {} # The three prefetch process
for split in self.iterators.keys():
self._prefetch_process[split] = BlobFetcher(split, self, split=='train')
# Terminate the child process when the parent exists
def cleanup():
print('Terminating BlobFetcher')
for split in self.iterators.keys():
del self._prefetch_process[split]
import atexit
atexit.register(cleanup)
def get_batch(self, split, batch_size=None, seq_per_img=None):
batch_size = batch_size or self.batch_size
seq_per_img = seq_per_img or self.seq_per_img
fc_batch = [] # np.ndarray((batch_size * seq_per_img, self.opt.fc_feat_size), dtype = 'float32')
att_batch = [] # np.ndarray((batch_size * seq_per_img, 14, 14, self.opt.att_feat_size), dtype = 'float32')
label_batch = np.zeros([batch_size * seq_per_img, self.seq_length + 2], dtype = 'int')
mask_batch = np.zeros([batch_size * seq_per_img, self.seq_length + 2], dtype = 'float32')
wrapped = False
infos = []
gts = []
for i in range(batch_size):
import time
t_start = time.time()
# fetch image
tmp_fc, tmp_att,\
ix, tmp_wrapped = self._prefetch_process[split].get()
fc_batch += [tmp_fc] * seq_per_img
att_batch += [tmp_att] * seq_per_img
# fetch the sequence labels
ix1 = self.label_start_ix[ix] - 1 #label_start_ix starts from 1
ix2 = self.label_end_ix[ix] - 1
ncap = ix2 - ix1 + 1 # number of captions available for this image
assert ncap > 0, 'an image does not have any label. this can be handled but right now isn\'t'
if ncap < seq_per_img:
# we need to subsample (with replacement)
seq = np.zeros([seq_per_img, self.seq_length], dtype = 'int')
for q in range(seq_per_img):
ixl = random.randint(ix1,ix2)
seq[q, :] = self.h5_label_file['labels'][ixl, :self.seq_length]
else:
ixl = random.randint(ix1, ix2 - seq_per_img + 1)
seq = self.h5_label_file['labels'][ixl: ixl + seq_per_img, :self.seq_length]
label_batch[i * seq_per_img : (i + 1) * seq_per_img, 1 : self.seq_length + 1] = seq
if tmp_wrapped:
wrapped = True
# Used for reward evaluation
gts.append(self.h5_label_file['labels'][self.label_start_ix[ix] - 1: self.label_end_ix[ix]])
# record associated info as well
info_dict = {}
info_dict['ix'] = ix
info_dict['id'] = self.info['images'][ix]['id']
info_dict['file_path'] = self.info['images'][ix]['file_path']
infos.append(info_dict)
#print(i, time.time() - t_start)
# generate mask
t_start = time.time()
nonzeros = np.array(list(map(lambda x: (x != 0).sum()+2, label_batch)))
for ix, row in enumerate(mask_batch):
row[:nonzeros[ix]] = 1
#print('mask', time.time() - t_start)
data = {}
data['fc_feats'] = np.stack(fc_batch)
data['att_feats'] = np.stack(att_batch)
data['labels'] = label_batch
data['gts'] = gts
data['masks'] = mask_batch
data['bounds'] = {'it_pos_now': self.iterators[split], 'it_max': len(self.split_ix[split]), 'wrapped': wrapped}
data['infos'] = infos
return data
# It's not coherent to make DataLoader a subclass of Dataset, but essentially, we only need to implement the following to functions,
# so that the torch.utils.data.DataLoader can load the data according the index.
# However, it's minimum change to switch to pytorch data loading.
def __getitem__(self, index):
"""This function returns a tuple that is further passed to collate_fn
"""
ix = index #self.split_ix[index]
return get_npy_data(ix, \
os.path.join(self.input_fc_dir, str(self.info['images'][ix]['id']) + '.npy'),
os.path.join(self.input_att_dir, str(self.info['images'][ix]['id']) + '.npz'),
self.use_att
)
def __len__(self):
return len(self.info['images'])
class BlobFetcher():
"""Experimental class for prefetching blobs in a separate process."""
def __init__(self, split, dataloader, if_shuffle=False):
"""
db is a list of tuples containing: imcrop_name, caption, bbox_feat of gt box, imname
"""
self.split = split
self.dataloader = dataloader
self.if_shuffle = if_shuffle
# Add more in the queue
def reset(self):
"""
Two cases:
1. not hasattr(self, 'split_loader'): Resume from previous training. Create the dataset given the saved split_ix and iterator
2. wrapped: a new epoch, the split_ix and iterator have been updated in the get_minibatch_inds already.
"""
# batch_size is 0, the merge is done in DataLoader class
self.split_loader = iter(data.DataLoader(dataset=self.dataloader,
batch_size=1,
sampler=self.dataloader.split_ix[self.split][self.dataloader.iterators[self.split]:],
shuffle=False,
pin_memory=True,
num_workers=multiprocessing.cpu_count(),
collate_fn=lambda x: x[0]))
def _get_next_minibatch_inds(self):
max_index = len(self.dataloader.split_ix[self.split])
wrapped = False
ri = self.dataloader.iterators[self.split]
ix = self.dataloader.split_ix[self.split][ri]
ri_next = ri + 1
if ri_next >= max_index:
ri_next = 0
if self.if_shuffle:
random.shuffle(self.dataloader.split_ix[self.split])
wrapped = True
self.dataloader.iterators[self.split] = ri_next
return ix, wrapped
def get(self):
if not hasattr(self, 'split_loader'):
self.reset()
ix, wrapped = self._get_next_minibatch_inds()
tmp = self.split_loader.next()
if wrapped:
self.reset()
assert tmp[2] == ix, "ix not equal"
return tmp + [wrapped]