-
Notifications
You must be signed in to change notification settings - Fork 20
/
Copy pathdata_loader_situation.py
403 lines (345 loc) · 17 KB
/
data_loader_situation.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
import codecs
import os
import torch
import torch.utils.data as data
from PIL import Image
from collections import defaultdict
import json
import csv
import pickle
import traceback
import numpy as np
from operator import itemgetter
from src.util import consts
from src.dataflow.numpy.anno_mapping import event_type_norm, role_name_norm
def map_to_ids(tokens, vocab):
ids = [vocab[t] if t in vocab else consts.UNK_IDX for t in tokens]
return ids
def map_to_id(token, vocab):
if token in vocab:
return vocab[token]
else:
return consts.UNK_IDX
def get_labels(class_map_file):
print('class_map_file', class_map_file)
label_name = {}
with open(class_map_file, mode='r') as csv_file:
csv_reader = csv.reader(csv_file, delimiter=',')
for row in csv_reader:
# isArgType = int(row[2])
# if isArgType:
if row[2] == '1':
label_name[row[0]] = row[1]
print('object_label', label_name)
return label_name
def load_img_object(img_id, image_dir, transform,
load_object=False, object_results=None,
object_label=None, object_detection_threshold=.1,
vocab_objlabel=None,
object_topk=50,):
try:
img_path = os.path.join(image_dir, img_id)
image = Image.open(img_path).convert('RGB')
except:
img_path = os.path.join(image_dir, img_id+'.png')
image = Image.open(img_path).convert('RGB')
if transform is not None:
image_vec = transform(image)
if load_object:
# load object detection result
bbox_entities_id = []
bbox_entities_region = []
bbox_entities_label = []
# pad the entire image to bbox vectors, in case the bbox is none
bbox_entities_id.append('0_0_0_0')
if transform is not None:
bbox_entities_region.append(image_vec)
else:
bbox_entities_region.append(image)
bbox_entities_label.append(consts.UNK_IDX)
objects = object_results[img_id]
objects_by_score = sorted(objects, key=itemgetter('score'))
count = 1
for object in objects_by_score:
if count > object_topk:
break
if object['label'] not in object_label:
# print('rejected labels', object['label'])
continue
# print('accepted labels', object_label[object['label']])
label = object_label[object['label']]
bbox = object['bbox']
score = object['score']
if score < object_detection_threshold:
continue
# transform patch to patch_vec
try:
patch = image.crop(bbox)
patch_id = '%d_%d_%d_%d' % (bbox[0], bbox[1], bbox[2], bbox[3])
bbox_entities_id.append(patch_id)
# bbox_entities_label.append(self.map_to_id(label, self.vocab_situation_noun.word2id))
bbox_entities_label.append(map_to_id(label, vocab_objlabel))
if transform is not None:
patch_vec = transform(patch)
# bbox_id, bbox_vec, label
# bbox_entities.append( (patch_id, patch_vec, label) )
bbox_entities_region.append(patch_vec)
else:
# bbox_entities.append( (patch_id, patch, label) )
bbox_entities_region.append(patch)
count += 1
except:
print('Wrong image ', img_path)
traceback.print_exc()
object_num = len(bbox_entities_region)
if load_object:
return image, image_vec, bbox_entities_id, bbox_entities_region, bbox_entities_label, object_num
else:
return image, image_vec, None, None, None, None
class ImSituDataset(data.Dataset):
def __init__(self, image_dir, vocab_situation_noun, vocab_situation_role, vocab_situation_verb,
event2id, eerole2id,
imsitu_ontology_file, imsitu_annotation_file, verb_mapping_file, object_ontology_file=None,
object_detection_pkl_file=None, object_detection_threshold=0.2, transform=None, filter_irrelevant_verbs=False,
load_object=False, filter_place=False):
self.image_dir = image_dir
self.vocab_situation_noun = vocab_situation_noun
self.vocab_situation_role = vocab_situation_role
self.vocab_situation_verb = vocab_situation_verb
self.event2id = event2id
print('self.event2id', self.event2id)
self.eerole2id = eerole2id
print('self.eerole2id', self.eerole2id)
self.filter_irrelevant_verbs = filter_irrelevant_verbs
print('************************filter_irrelevant_verbs: ', self.filter_irrelevant_verbs)
self.filter_place = filter_place
print('************************filter_place: ', self.filter_place)
self.load_object = load_object
if self.load_object:
self.object_detection_threshold = object_detection_threshold
self.object_label = get_labels(object_ontology_file)
self.object_results = pickle.load(open(object_detection_pkl_file, 'rb'))
else:
self.object_detection_threshold = 0.2
self.object_label = None
self.object_results = None
# imsitu_info = json.load(open(os.path.join(imsitu_dir, "imsitu_space.json")))
imsitu_info = json.load(open(imsitu_ontology_file))
self.nouns = imsitu_info["nouns"]
self.verbs = imsitu_info["verbs"]
self.verb_roles = self._load_verb_roles(imsitu_info["verbs"], vocab_situation_verb.word2id, vocab_situation_role.word2id)
self.role_masks = self._verb_role_mask(self.verb_roles)
self.annotation = json.load(open(imsitu_annotation_file))
self.sr_mapping_verb, self.sr_mapping_role = self.load_mapping_all(verb_mapping_file)
# print('self.sr_mapping_verb', self.sr_mapping_verb)
# print('self.sr_mapping_role', self.sr_mapping_role)
# self.img_verbs, self.img_verb_roles, self.img_verb_role_num = self.__getobjects__(image_dir)
self.img_verbs, self.img_verb_roles = self.__getobjects__(image_dir)
self.idx_imgs = self.__getids__()
self.ids = list(self.idx_imgs.keys())
self.transform = transform
# def load_mapping_verb(self, verb_mapping_file):
# verb_type_dict = dict()
# for line in open(verb_mapping_file):
# line = line.rstrip('\n')
# tabs = line.split('\t')
# verb_type_dict[tabs[0]] = tabs[1]
# return verb_type_dict
def load_mapping_all(self, verb_mapping_file):
sr_verb_mapping = defaultdict()
sr_role_mapping = defaultdict(lambda : defaultdict(str))
for line in codecs.open(verb_mapping_file, 'r', encoding='utf-8'):
line = line.rstrip('\n')
tabs = line.split('\t')
sr_verb = tabs[0]
sr_role = tabs[1]
ee_event = tabs[2]
ee_role = tabs[3]
sr_role_mapping[sr_verb][sr_role] = ee_role
sr_verb_mapping[sr_verb] = ee_event
return sr_verb_mapping, sr_role_mapping
def get_sr_mapping(self):
return self.sr_mapping_verb, self.sr_mapping_role
def _load_verb_roles(self, verb_info, verb2id, role2id):
verb_roles = defaultdict(set)
for verb in verb_info:
for role in verb_info[verb]['roles']:
if self.filter_place and 'place' == role.lower():
continue
verb_roles[verb2id[verb]].add(role2id[role])
return verb_roles
def _verb_role_mask(self, verb_roles):
# role_masks = dict()
# for verb in verb_roles:
# indexes = []
# for role in verb_roles[verb]:
# indexes.append(role2id[role])
# i = torch.LongTensor([[0]*len(indexes), indexes])
# v = torch.LongTensor([1] * len(indexes))
# role_masks[verb2id[verb]] = torch.sparse.FloatTensor(i, v, torch.Size([self.vocab_situation_role.size]))
# return role_masks
row_indexes = []
column_indexes = []
for verb in verb_roles:
for role in verb_roles[verb]:
column_indexes.append(role)
row_indexes.extend([verb] * len(verb_roles[verb]))
i = torch.LongTensor([row_indexes, column_indexes])
v = torch.LongTensor([1] * len(row_indexes))
role_masks = torch.sparse.FloatTensor(i, v, torch.Size([self.vocab_situation_verb.size, self.vocab_situation_role.size])).requires_grad_(False)
return role_masks
def get_role_mask(self):
return self.role_masks
def get_verb_role_mapping(self):
return self.verb_roles
# img -> patch & entity
def __getobjects__(self, image_dir):
img_verbs = defaultdict(lambda: str)
img_verb_roles = defaultdict(lambda : defaultdict(set))
# img_verb_role_num = defaultdict(int)
for image_id in os.listdir(image_dir):
if image_id not in self.annotation:
continue
# if self.load_object and len(self.object_results[image_id]) == 0: #(image_id not in self.object_results or ):
# # image does not have any objects
# continue
verb = self.annotation[image_id]['verb']
if self.filter_irrelevant_verbs and verb not in self.sr_mapping_verb:
continue
img_verbs[image_id] = verb.lower()
# ontology
# verb_roles[image_id] = self.verbs[verb]['order']
# role values
frames = self.annotation[image_id]['frames']
for frame in frames:
for role in frame:
role = role.lower()
if self.filter_place and 'place' == role.lower():
continue
role_value_id = frame[role]
if len(role_value_id) > 0:
role_value = self.nouns[role_value_id]['gloss']
# lower()
img_verb_roles[image_id][role].update(role_value)
# img_verb_role_num[image_id] = img_verb_role_num[image_id] + 1
return img_verbs, img_verb_roles #, img_verb_role_num
# idx -> img
def __getids__(self):
idx_img = {}
index = 0
for img in self.img_verbs:
idx_img[index] = img
index += 1
print("number of images: ", len(idx_img))
return idx_img
def __getitem__(self, index):
"""Returns one data pair (image, captions, regions and entities)."""
img_id = self.idx_imgs[index]
return self.get_img_info(img_id)
def get_img_info(self, img_id):
# print(img_id, self.img_verbs[img_id])
# try:
# print(self.sr_mapping_verb[self.img_verbs[img_id]])
# except:
# print('no ace type')
verb = map_to_id(self.img_verbs[img_id], self.vocab_situation_verb.word2id)
if self.img_verbs[img_id] in self.sr_mapping_verb:
event_str = 'B-' + event_type_norm(self.sr_mapping_verb[self.img_verbs[img_id]])
event = self.event2id[event_str]
else:
event = self.event2id[consts.O_LABEL_NAME]
verb_roles = self.img_verb_roles[img_id] # role -> role_values
roles = []
roles_ee = []
args = []
for role in verb_roles:
# keep the one has the mapping
if (self.img_verbs[img_id] in self.sr_mapping_role) and (role in self.sr_mapping_role[self.img_verbs[img_id]]):
role_ee_str = role_name_norm(self.sr_mapping_role[self.img_verbs[img_id]][role])
else:
role_ee_str = consts.ROLE_O_LABEL_NAME
for role_arg in verb_roles[role]:
# roles.append(self.make_one_hot(role, self.vocab_situation_role.word2id))
roles.append(map_to_id(role, self.vocab_situation_role.word2id))
roles_ee.append(self.eerole2id[role_ee_str]) # same vocab???
args.append(map_to_id(role_arg, self.vocab_situation_noun.word2id))
arg_num = len(args)
# roles = np.asarray(roles)
image, image_vec, bbox_entities_id, bbox_entities_region, bbox_entities_label, object_num \
= load_img_object(img_id, self.image_dir, self.transform,
self.load_object, self.object_results, self.object_label,
self.object_detection_threshold,
self.vocab_situation_noun.word2id)
if self.transform is not None:
if self.load_object:
return img_id, image_vec, verb, event, roles, roles_ee, args, bbox_entities_id, bbox_entities_region, bbox_entities_label, arg_num, object_num
else:
return img_id, image_vec, verb, event, roles, roles_ee, args, None, None, None, arg_num, None
else:
if self.load_object:
return img_id, image, verb, event, roles, roles_ee, args, bbox_entities_id, bbox_entities_region, bbox_entities_label, arg_num, object_num
else:
return img_id, image, verb, event, roles, roles_ee, args, None, None, None, arg_num, None
def __len__(self):
return len(self.ids)
# def make_one_hots(self, roles, vocab):
# role_vec = np.zeros(len(roles), len(vocab))
# for i, r in enumerate(roles):
# print(i, r, vocab[r])
# role_vec[i][vocab[r]] = 1
# # ids = [vocab[t] if t in vocab else consts.UNK_IDX ]
# return role_vec
def make_one_hot(self, role, vocab):
role_vec = [0] * len(vocab) #np.zeros(len(vocab))
# print(role)
# if role not in vocab:
# print('vocab', vocab)
role_vec[vocab[role]] = 1
return role_vec
def image_collate_fn(batch):
"""Creates mini-batch tensors from the list of tuples (img_id, captionvec, regions_tensor, entity_tensor). <-- (__getitem__ !!!)
We should build custom collate_fn rather than using default collate_fn,
because merging caption/regions/entities (including padding) is not supported in default.
"""
# Sort a data list by arg_num (descending order).
batch.sort(key=lambda x: x[-2], reverse=True)
img_id_batch, image_batch, verb_batch, event_batch, roles_batch, ee_roles_batch, args_batch, bbox_entities_id, bbox_entities_region, bbox_entities_label, arg_num_batch, object_num_batch = zip(*batch) # zip(['a', 'b', 'c'], [1, 2, 3]) =
# Merge images (from tuple of 3D tensor to 4D tensor).
image_batch = torch.stack(image_batch, 0)
# object mask
return img_id_batch, image_batch, verb_batch, event_batch, roles_batch, ee_roles_batch, args_batch, bbox_entities_id, bbox_entities_region, bbox_entities_label, arg_num_batch, object_num_batch
def unpack(batch, device, load_object=False):
img_id_batch = batch[0]
# print('img_id_batch', img_id_batch)
# print('device', device)
image_batch = batch[1].to(device)
verb_gt_batch = torch.LongTensor(batch[2]).to(device)
event_gt_batch = torch.LongTensor(batch[3]).to(device)
arg_num_batch = np.array(batch[-2])
# roles_gt_batch = torch.LongTensor(batch[3]).to(device)
# args_batch = torch.LongTensor(batch[4]).to(device)
arg_num_max = max(arg_num_batch)
# roles_gt_batch = torch.zeros([len(img_id_batch), arg_num_max, len(batch[3][0][0])]).type(torch.LongTensor).to(device)
roles_gt_batch = torch.zeros([len(img_id_batch), arg_num_max]).type(torch.LongTensor).to(device)
ee_roles_gt_batch = torch.zeros([len(img_id_batch), arg_num_max]).type(torch.LongTensor).to(device)
args_gt_batch = torch.zeros([len(img_id_batch), arg_num_max]).type(torch.LongTensor).to(device)
for batch_idx, _ in enumerate(batch[4]):
roles_gt_batch[batch_idx][:arg_num_batch[batch_idx]] = torch.LongTensor(batch[4][batch_idx])
ee_roles_gt_batch[batch_idx][:arg_num_batch[batch_idx]] = torch.LongTensor(batch[5][batch_idx])
args_gt_batch[batch_idx][:arg_num_batch[batch_idx]] = torch.LongTensor(batch[6][batch_idx])
if load_object:
bbox_entities_id = batch[7]
object_num_batch = np.array(batch[-1])
object_num_max = max(object_num_batch)
bbox_entities_region = torch.zeros(len(img_id_batch), object_num_max, 3, 224, 224).to(device)
bbox_entities_label = torch.zeros(len(img_id_batch), object_num_max).to(device)
for b_idx, _ in enumerate(batch[8]):
for obj_idx, _ in enumerate(batch[8][b_idx]):
bbox_entities_region[b_idx][obj_idx] = batch[8][b_idx][obj_idx]
bbox_entities_label[b_idx][obj_idx] = batch[9][b_idx][obj_idx]
return img_id_batch, image_batch, verb_gt_batch, event_gt_batch, roles_gt_batch, ee_roles_gt_batch, args_gt_batch, \
bbox_entities_id, bbox_entities_region, bbox_entities_label, \
arg_num_batch, object_num_batch
else:
return img_id_batch, image_batch, verb_gt_batch, event_gt_batch, roles_gt_batch, ee_roles_gt_batch, args_gt_batch, \
None, None, None, arg_num_batch, None