-
Notifications
You must be signed in to change notification settings - Fork 27
/
vicap_dataset.py
275 lines (214 loc) · 9.46 KB
/
vicap_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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
import os
import sys
import numpy as np
import pandas as pd
from PIL import Image
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import json
from pathlib import Path
from collections import Counter
import spacy
import torch
from torch.nn.utils.rnn import pad_sequence
from torch.utils.data import DataLoader, Dataset
from torch.utils.data import DataLoader, DistributedSampler, BatchSampler
from engine.utils import nested_tensor_from_tensor_list
import torchvision.transforms as T
from collections import defaultdict
from datasets.caption.transforms import *
spacy_eng = spacy.load('en_core_web_sm')
train_image_path = Path("/media/local_workspace/quang/datasets/vietcap/train-images")
test_image_path = Path("/media/local_workspace/quang/datasets/vietcap/public-test-images")
train_caption_path = "/media/local_workspace/quang/datasets/vietcap/thesis_test/train_data.json"
test_caption_path = "/media/local_workspace/quang/datasets/vietcap/thesis_test/test_data.json"
vi_caption_path = "/media/local_workspace/quang/datasets/vietcap/vi_captions.json"
def get_transform(resize_name="maxwh", size=[384, 640], randaug=False):
resize = RESIZE[resize_name](size)
if randaug:
return {
'train': Compose([resize, RandAugment(), ToTensor(), normalize()]),
'valid': Compose([resize, ToTensor(), normalize()]),
}
else:
return {
'train': Compose([resize, ToTensor(), normalize()]),
'valid': Compose([resize, ToTensor(), normalize()]),
}
class Vocabulary:
def __init__(self, freq_threshold):
self.itos = {0: "<unk>", 1: "<pad>", 2: "<bos>", 3: "<eos>"}
self.stoi = {v: k for k, v in self.itos.items()}
self.freq_threshold = freq_threshold
def __len__(self):
return len(self.itos)
@staticmethod
def tokenize(text):
return [token.text.lower() for token in spacy_eng.tokenizer(text)]
def build_vocab(self, sentence_list):
frequencies = Counter()
idx = 4
for sentence in sentence_list:
for word in self.tokenize(sentence):
frequencies[word] += 1
if frequencies[word] == self.freq_threshold:
self.stoi[word] = idx
self.itos[idx] = word
idx += 1
def numericalize(self, text):
tokenized_text = self.tokenize(text)
return [self.stoi[token] if token in self.stoi else self.stoi["<unk>"] for token in tokenized_text]
class CustomDataset(Dataset):
def __init__(self, root_dir, captions_file, vicap_file=vi_caption_path, transform=None, freq_threshold=1):
self.root_dir = root_dir
self.captions_file = captions_file
self.data = json.load(open(captions_file, "r"))
self.transform = transform
self.imgid2imgname = {entry['id']: entry['filename'] for entry in self.data['images']}
self.captions = [ann['segment_caption'] for ann in self.data['annotations']]
all_captions = json.load(open(vicap_file, "r"))
self.vocab = Vocabulary(freq_threshold)
self.vocab.build_vocab(all_captions)
def __len__(self):
return len(self.data['annotations'])
def __getitem__(self, idx):
annotation = self.data['annotations'][idx]
caption = annotation['segment_caption']
image_id = annotation["image_id"]
image_name = self.imgid2imgname[image_id]
image = Image.open(os.path.join(self.root_dir, image_name)).convert('RGB')
if self.transform is not None:
img = self.transform(image)
caption_vec = []
caption_vec += [self.vocab.stoi["<bos>"]]
caption_vec += self.vocab.numericalize(caption)
caption_vec += [self.vocab.stoi["<eos>"]]
return img, torch.tensor(caption_vec), caption
class DictDataset(CustomDataset):
def __init__(self, root_dir, captions_file, vicap_file=vi_caption_path, transform=None, freq_threshold=1):
super().__init__(root_dir, captions_file, vicap_file, transform, freq_threshold)
self.img_id_2_captions = self.img_id_2_captions()
self.img_ids = list(self.img_id_2_captions.keys())
def img_id_2_captions(self):
img_id_2_captions = defaultdict(list)
for ann in self.data['annotations']:
img_id_2_captions[ann['image_id']].append(" ".join(self.vocab.tokenize(ann['segment_caption'])))
return img_id_2_captions
def __len__(self):
return len(self.img_id_2_captions)
def __getitem__(self, idx):
img_id = self.img_ids[idx]
image_name = self.imgid2imgname[img_id]
img = Image.open(os.path.join(self.root_dir, image_name)).convert('RGB')
if self.transform is not None:
img = self.transform(img)
captions = self.img_id_2_captions[img_id]
return img, captions, img_id
class EvalCollate:
def __init__(self, pad_idx, batch_first=True, device="cuda"):
self.pad_idx = pad_idx
self.batch_first = batch_first
self.device = device
def __call__(self, batch):
imgs = [item[0] for item in batch]
imgs = nested_tensor_from_tensor_list(imgs).to(self.device)
captions = [item[1] for item in batch]
image_ids = [item[2] for item in batch]
return {'samples': imgs, 'captions': captions, 'image_ids': image_ids}
class CapsCollate:
def __init__(self, pad_idx, batch_first=True, device="cuda"):
self.pad_idx = pad_idx
self.batch_first = batch_first
self.device = device
def __call__(self, batch):
imgs = [item[0] for item in batch]
imgs = nested_tensor_from_tensor_list(imgs).to(self.device)
targets = [item[1] for item in batch]
targets = pad_sequence(targets, batch_first=self.batch_first, padding_value=self.pad_idx)
targets = targets.to(self.device)
return {'samples': imgs, 'captions': targets}
def get_datasets():
transforms = get_transform(resize_name="maxwh", size=[384, 640], randaug=False)
train_dataset = CustomDataset(
root_dir=train_image_path,
captions_file=train_caption_path,
vicap_file=vi_caption_path,
transform=transforms['train'],
)
valid_dataset = CustomDataset(
root_dir=test_image_path,
captions_file=test_caption_path,
vicap_file=vi_caption_path,
transform=transforms['valid'],
)
train_dict_dataset = DictDataset(
root_dir=train_image_path,
captions_file=train_caption_path,
vicap_file=vi_caption_path,
transform=transforms['train'],
)
valid_dict_dataset = DictDataset(
root_dir=test_image_path,
captions_file=test_caption_path,
vicap_file=vi_caption_path,
transform=transforms['valid'],
)
return {
'train': train_dataset,
'valid': valid_dataset,
'train_dict': train_dict_dataset,
'valid_dict': valid_dict_dataset,
}
def get_dataloaders(device="cuda", batch_size=8):
datasets = get_datasets()
collators = {
'train': CapsCollate(datasets['train'].vocab.stoi['<pad>'], device=device),
'valid': CapsCollate(datasets['valid'].vocab.stoi['<pad>'], device=device),
'train_dict': EvalCollate(datasets['train_dict'].vocab.stoi['<pad>'], device=device),
'valid_dict': EvalCollate(datasets['valid_dict'].vocab.stoi['<pad>'], device=device)
}
samplers = {
'train': DistributedSampler(datasets['train'], shuffle=True),
'valid': DistributedSampler(datasets['valid'], shuffle=False),
'train_dict': DistributedSampler(datasets['train_dict'], shuffle=True),
'valid_dict': DistributedSampler(datasets['valid_dict'], shuffle=False)
}
dataloaders = {k: DataLoader(datasets[k], batch_size=batch_size, num_workers=4, collate_fn=collators[k], sampler=samplers[k]) for k in datasets}
return samplers, dataloaders
if __name__ == "__main__":
# get all captions to build vocabulary
train_data = json.load(open(train_caption_path))
test_data = json.load(open(test_caption_path))
train_captions = [ann['segment_caption'] for ann in train_data['annotations']]
test_captions = [ann['segment_caption'] for ann in test_data['annotations']]
all_captions = train_captions + test_captions
with open(vi_caption_path, 'w', encoding='utf-8') as f:
json.dump(all_captions, f, indent=4, ensure_ascii=False)
# build datasets
datasets = get_datasets()
train_dataset, test_dataset = datasets['train'], datasets['valid']
# train_dataset
img, target, caption = train_dataset[0]
print(img.shape)
print(target.shape)
print(" ".join([train_dataset.vocab.itos[token] for token in target.numpy()]))
print(caption)
# test_dataset
img, captions, img_id = test_dataset[0]
print(img.shape)
print(captions)
print(img_id)
train_loader = DataLoader(train_dataset, batch_size=2, shuffle=True, collate_fn=CapsCollate(train_dataset.vocab.stoi["<pad>"]))
valid_loader = DataLoader(test_dataset, batch_size=2, shuffle=False, collate_fn=EvalCollate(train_dataset.vocab.stoi["<pad>"]))
dataloaders = {'train': train_loader, 'valid': valid_loader}
vocab = train_dataset.vocab
for batch in valid_loader:
imgs, captions = batch['samples'], batch['captions']
print(imgs.tensors.shape)
print(captions)
break
for batch in train_loader:
imgs, captions = batch['samples'], batch['captions']
print(imgs.tensors.shape)
print(captions.shape)
break