/
create_hdf5.py
431 lines (359 loc) · 17.7 KB
/
create_hdf5.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
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
from __future__ import division
import h5py as h5
import csv
import re
import cv2
from os.path import join
import cPickle as pickle
from collections import defaultdict, Counter
import numpy as np
import glob2
import sys
'''
Creates one hdf5 file with the following properties :
# one training sample is one page
n pages total
book_ids (h_book_ids): n
page_ids (h_page_ids): n
images (h_ims): n x max_panels x 256 x 256 x 3 (or whatever your size is)
panel_mask (h_pmask): n x max_panels
text bounding box coords [bbox] (h_coords): n x max_panels x max_boxes x 4
text bbox mask [bbox_mask] (h_tmask): n x max_panels x max_boxes
text box words [words] (h_words): n x max_panels x max_boxes x max_words
text box word mask [words_mask] (h_wmask): n x max_panels x max_boxes x max_words
'''
class Create_hdf5:
def __init__(self, panels_path, ocr_file, ad_path, dims, max_panels, max_boxes, max_words, max_vocab_size, h5path):
self.panels_path = panels_path
# self.images_path = images_path
# self.n_pages = glob2.glob(join(self.images_path, '*', '*.jpg'))
# takes a really long time
self.ocr_file = ocr_file
self.dev_idx, self.test_idx, self.n_pages = self.compute_fold_starts()
self.n_pages -= self.subtract_ad_pages(ad_path)
self.dims = dims
self.h5path = h5path
self.max_panels = max_panels
self.max_boxes = max_boxes
self.max_words = max_words
self.max_vocab_size = max_vocab_size
self.dump_vocabulary('./data/comics_vocab.p')
def dump_vocabulary(self, vocab_path): # helper which dumps vocabulary in pickle
'''
Creates vocabulary
# Arguments
vocab_path : path where the vocabulary is dumped as a pickle file
'''
print 'loading ocr dict'
self.ocr_dict = defaultdict(list)
self.vcount = Counter()
self.read_ocr_file()
print 'done creating ocr dict, now filtering vocab of size %d' % \
len(self.vcount)
vocab = self.vcount.most_common(self.max_vocab_size)
self.vdict = {}
for i, (w, count) in enumerate(vocab):
self.vdict[w] = i
self.vdict['UNK'] = len(self.vdict)
self.rvdict = dict((v,k) for k,v in self.vdict.iteritems())
print 'done filtering vocab to %d words' % len(self.vdict)
pickle.dump([self.vdict, self.rvdict],
open(vocab_path, 'wb'), protocol=pickle.HIGHEST_PROTOCOL)
# use 500 books for both dev and test, remaining are for training
def compute_fold_starts(self):
raw_f = open(self.ocr_file, 'r')
f = csv.DictReader(raw_f)
# comic_no,page_no,panel_no,textbox_no,dialog_or_narration,text,x1,y1,x2,y2
# compute number of books
unique_books = Counter()
for row in f:
unique_books[int(row['comic_no'])] += 1
num_books = len(unique_books) #3929
dev_thresh = num_books - 1000
test_thresh = num_books - 500
raw_f.close()
raw_f = open(self.ocr_file, 'r')
# now count pages to find fold indices
f = csv.DictReader(raw_f)
book_count = 0
page_count = 0
prev_book = -1
prev_page = -1
dev_idx = -1
test_idx = -1
for row in f:
curr_book = int(row['comic_no'])
curr_page = int(row['page_no'])
if curr_book != prev_book:
book_count += 1
prev_book = curr_book
if prev_page != curr_page:
page_count += 1
prev_page = curr_page
if book_count == dev_thresh:
dev_idx = page_count
if book_count == test_thresh:
test_idx = page_count
return dev_idx, test_idx, page_count
def subtract_ad_pages(self, ad_path): # helper
'''
counts the number of ad pages
# Arguments :
ad_path : location of all ad pages (using a BOW classifier)
# Returns :
len(self.ad_pages) : number of ad pages
'''
ad_file = open(ad_path, 'r')
self.ad_pages = set()
for line in ad_file:
self.ad_pages.add(line.strip())
return len(self.ad_pages)
def scale_bbox(self, bbox_coords, panel_coords, new_dims):
'''
Scales text bounding box coordinates within the reference
frame of panel coordinates (between 0 and 1)
# Arguments :
bbox_coords : absolute coordinates of the text bounding box in the page
panel_coords : [0, 0, w, h] where w and h are width and height of panel
new_dims : size of the panel after saving [something like (256, 256) ]
# Returns :
u1_, v1_, u2_, v2_ which are the scaled coordinates of textbox
'''
w, h = new_dims
u1, v1, u2, v2 = bbox_coords
x1, y1, x2, y2 = panel_coords
u1_ = u1 * w / (x2 - x1)
v1_ = v1 * h / (y2 - y1)
u2_ = u2 * w / (x2 - x1)
v2_ = v2 * h / (y2 - y1)
return int(u1_), int(v1_), int(u2_), int(v2_)
def scale_textbox_coords(self, w, h, bbox_coords):
'''
Scales bounding box coordinates to between 0 and 1.
# Arguments :
w, h : panel width and height
bbox_coords : textbox_coordinates that are to be scaled
# Returns :
scaled_coords : textbox_coordinates after scaling
'''
scaled_coords = self.scale_bbox(bbox_coords, [0, 0, w, h], self.dims)
return scaled_coords
def blacken_box(self, img, coords):
'''
We blacken the textboxes in an image (so that the computer doesn't
learn to predict the answer by trying to read the box)
# Arguments :
img : panel image after resizing
coords : using coords, draw a rectangle and paint it black
# Returns :
img : image after all of its text boxes have been blackened
'''
for coord_ in coords:
u1, v1, u2, v2 = coord_.astype('uint8')
cv2.rectangle(img, (u1, v1), (u2, v2), (0, 0, 0), -1)
return img
def read_ocr_file(self, save='No'):
'''
read csv files with ocr and return a tuple containing
(comic, page no, panel no, textbox no, class (n or d),
panel img, normalized coords of box, text)
'''
with open(self.ocr_file, 'rb') as csvfile:
reader = csv.DictReader(csvfile)
for row in reader:
# ocr_list.append((comic, page, panel, textbox, text))
comic_no = row['comic_no']
page_no = row['page_no']
panel_no = row['panel_no']
textbox_no = row['textbox_no']
text = row['text']
dialog_or_narration = row['dialog_or_narration']
x1 = row['x1']
y1 = row['y1']
x2 = row['x2']
y2 = row['y2']
text = text.split()
for w in text:
self.vcount[w] += 1
self.ocr_dict[(int(comic_no), int(page_no), int(panel_no))].append(
[textbox_no, dialog_or_narration, text, x1, y1, x2, y2])
if save == 'Yes':
pickle.dump(self.ocr_dict, open('ocr_dict.p', 'w'),
protocol=pickle.HIGHEST_PROTOCOL)
def create_img_hdf5(self):
'''
# one training sample is one page
n pages total
book_ids (h_book_ids): n
page_ids (h_page_ids): n
images (h_ims): n x max_panels x 256 x 256 x 3 (or whatever your size is)
panel_mask (h_pmask): n x max_panels
text bounding box coords [bbox] (h_coords): n x max_panels x max_boxes x 4
text bbox mask [bbox_mask] (h_tmask): n x max_panels x max_boxes
text box words [words] (h_words): n x max_panels x max_boxes x max_words
text box word mask [words_mask] (h_wmask): n x max_panels x max_boxes x max_words
'''
n_train_pages = self.dev_idx
n_dev_pages = self.test_idx - self.dev_idx
# n_test_pages = self.n_pages - self.test_idx
keys = sorted(self.ocr_dict.keys()) # is in terms of panels.
# sort keys in alphabetical order
unique_pages = sorted(list(set([(key[0], key[1]) for key in keys])))
train_pages = unique_pages[:self.dev_idx]
dev_pages = unique_pages[self.dev_idx:self.test_idx]
test_pages = unique_pages[self.test_idx:]
# so, for train fold, pick all panels from keys[0] to (dev_pages[0], 0), not including
# for dev fold, pick all panels from (dev_pages[0], 0) to (test_pages[0], 0), not including
# for test fold, pick all panels from (test_pages[0], 0) to end
train_keys = []
dev_keys = []
test_keys = []
train_limit = (dev_pages[0][0], dev_pages[0][1], 0)
dev_limit = (test_pages[0][0], test_pages[0][1], 0)
for key in keys:
if key < train_limit:
train_keys.append(key)
elif key >= train_limit and key < dev_limit:
dev_keys.append(key)
elif key >= dev_limit:
test_keys.append(key)
fn_num_pages = lambda keys:len(set([(key[0], key[1]) for key in keys]))
n_train_ads = len([t for t in train_pages if '%s---%s' % (t[0], t[1]) in self.ad_pages])
n_dev_ads = len([t for t in dev_pages if '%s---%s' % (t[0], t[1]) in self.ad_pages])
n_test_ads = len([t for t in test_pages if '%s---%s' % (t[0], t[1]) in self.ad_pages])
d = {} # will help us loop over train, dev and test folds
d['train'] = {'n':fn_num_pages(train_keys)-n_train_ads, 'keys':train_keys}
d['dev'] = {'n':fn_num_pages(dev_keys)-n_dev_ads, 'keys':dev_keys}
d['test'] = {'n':fn_num_pages(test_keys)-n_test_ads, 'keys':test_keys}
detected_ads = set()
for fold in d.keys(): # loop through train, test, dev
# creating hdf5 dataset
f = h5.File(self.h5path, 'a')
h_book_ids = f.create_dataset(fold+'/book_ids',
shape=(d[fold]['n'], ), dtype=np.uint32)
h_page_ids = f.create_dataset(fold+'/page_ids',
shape=(d[fold]['n'], ), dtype=np.uint32)
h_ims = f.create_dataset(fold+'/images',
shape=(d[fold]['n'], self.max_panels, self.dims[0], self.dims[1], 3),
dtype=np.uint8)
h_pmask = f.create_dataset(fold+'/panel_mask',
shape=(d[fold]['n'], self.max_panels), dtype=np.uint8)
h_coords = f.create_dataset(fold+'/bbox',
shape=(d[fold]['n'], self.max_panels, self.max_boxes, 4),
dtype=np.uint8)
h_tmask = f.create_dataset(fold+'/bbox_mask',
shape=(d[fold]['n'], self.max_panels, self.max_boxes),
dtype=np.uint8)
h_words = f.create_dataset(fold+'/words',
shape=(d[fold]['n'], self.max_panels, self.max_boxes, self.max_words),
dtype=np.uint32)
h_wmask = f.create_dataset(fold+'/word_mask',
shape=(d[fold]['n'], self.max_panels, self.max_boxes, self.max_words),
dtype=np.uint8)
curr_page_idx = 0
prev_page = None
for i, key in enumerate(d[fold]['keys']):
comic_no, page_no, panel_no = key
if i % 5000 == 0 and i > 0:
print '%d of panels done in fold %s, %d of %d ads detected' % (i, fold, len(detected_ads),
len(self.ad_pages))
# check for ads
ad_check_id = '%s---%s' % (comic_no, page_no)
if ad_check_id in self.ad_pages:
detected_ads.add(ad_check_id)
continue
if prev_page == None:
prev_page = page_no
if page_no != prev_page:
curr_page_idx += 1
prev_page = page_no
h_book_ids[curr_page_idx] = comic_no
h_page_ids[curr_page_idx] = page_no
if panel_no >= self.max_panels:
continue
# load panel image
img = cv2.imread(join(self.panels_path, str(comic_no),
str(page_no)+'_'+str(panel_no)+'.jpg'))
h_, w_, c_ = img.shape
img = cv2.resize(img, self.dims)
panel_contents = self.ocr_dict[key]
# are there any lines in the CSV file without OCR?
if len(panel_contents) == 1 and panel_contents[0][0] == None:
# no textboxes, so nothing to blacken
h_ims[curr_page_idx, panel_no] = img
h_pmask[curr_page_idx, panel_no] = 1
else:
tcount = 0
# there are textboxes, so need to blacken
new_coords_blk_list = np.zeros((len(panel_contents), 4))
for j, textbox in enumerate(panel_contents):
# ocr_key = '%d_%d_%d_%d' % (comic_no, page_no, panel_no, j)
# if ocr_key in self.ocr_dict:
# textbox contains : [dialog_or_narration, text, x1, y1, x2, y2])
words = textbox[2]
if len(words) > 0:
dialog_or_narration = int(textbox[1])
x1, y1, x2, y2 = map(float, textbox[-4:])
new_coords = self.scale_textbox_coords(w_, h_, [x1, y1, x2, y2])
new_coords_blk_list[tcount] = new_coords
h_coords[curr_page_idx, panel_no, tcount] = new_coords
# mask is 0 if theres no textbox, 1 if its a speech bubble, 2 if its a narrative bo
h_tmask[curr_page_idx, panel_no, tcount] = dialog_or_narration
# remove unknown words
inds = [self.vdict[w] if w in self.vdict else self.vdict['UNK'] for w in words ][:self.max_words]
h_words[curr_page_idx, panel_no, tcount, :len(inds)] = inds
h_wmask[curr_page_idx, panel_no, tcount, :len(inds)] = 1
if tcount == self.max_boxes-1:
# limit the number of textboes per panel
break
tcount += 1
img = self.blacken_box(img, new_coords_blk_list)
h_ims[curr_page_idx, panel_no] = img
h_pmask[curr_page_idx, panel_no] = 1
def reconstruct_sentence(rvdict, w_, wm_):
return ' '.join([rvdict[y] for r,y in enumerate(w_) \
if wm_[r] == 1])
def preview(h5path, raw_pages_path, vocab_dict_path, fold='train'):
# open the file, display all contents of the page.
# also display the page image itself.
f = h5.File(h5path, 'r')
total_pages = len(f[fold+'/book_ids'])
samples = np.random.randint(0, high=total_pages, size=10)
for sample in samples:
comic_no = int(np.array(f[fold+'/book_ids'][sample]))
page_no = int(np.array(f[fold+'/page_ids'][sample]))
images = np.array(f[fold+'/images'][sample])
words = np.array(f[fold+'/words'][sample])
word_masks = np.array(f[fold+'/word_mask'][sample])
vdict, rvdict = pickle.load(open(vocab_dict_path, 'rb'))
print 'comic %d, page %d \n----------------' % (comic_no, page_no)
raw_page_img = cv2.imread(join(raw_pages_path, str(comic_no), str(page_no)+'.jpg'))
cv2.imshow('raw page image', cv2.resize(raw_page_img, (512, 768)))
cv2.waitKey(0)
non_empty_image_idxs = np.where(np.apply_over_axes(np.sum, images, [1, 2, 3]))[0]
for i in non_empty_image_idxs:
cv2.imshow('panel %d of %d' % (i, len(non_empty_image_idxs)),
images[i].astype('uint8'))
for j, (word, word_mask) in enumerate(zip(words[i], word_masks[i])):
print 'panel %d, dialog %d --> %s' % (i, j, reconstruct_sentence(rvdict, word, word_mask))
print
cv2.waitKey(0)
cv2.destroyAllWindows()
def main():
# images_path = '/home/varunm/Comics/data/raw_page_images/'
panels_path = './data/raw_panel_images/'
ocr_file = './data/COMICS_ocr_file.csv'
ad_path = './data/predadpages.txt' # these pages have ads according to ad filter
h5path = './data/comics.h5' # dumping h5 file at this location
dims = (224, 224) # dimensions of image
max_panels = 9
max_boxes = 3
max_words = 30
max_vocab_size = 20000
c = Create_hdf5(panels_path, ocr_file, ad_path, dims,
max_panels, max_boxes, max_words, max_vocab_size, h5path)
c.create_img_hdf5()
if __name__ == '__main__':
main()
# preview('comics.h5', '/data/raw_page_images/',
# 'comics_vocab.p', fold='dev')