-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdata_loader_val.py
148 lines (130 loc) · 6.51 KB
/
data_loader_val.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
import nltk
import os
import torch
import torch.utils.data as data
from vocabulary import Vocabulary
from PIL import Image
from pycocotools.coco import COCO
import numpy as np
from tqdm import tqdm
import random
import json
def get_loader(transform,
mode='val',
batch_size=1,
vocab_threshold=None,
vocab_file='./vocab.pkl',
start_word="<start>",
end_word="<end>",
unk_word="<unk>",
vocab_from_file=True,
num_workers=0,
cocoapi_loc='./'):
"""Returns the data loader.
Args:
transform: Image transform.
mode: One of 'train' or 'test'.
batch_size: Batch size (if in testing mode, must have batch_size=1).
vocab_threshold: Minimum word count threshold.
vocab_file: File containing the vocabulary.
start_word: Special word denoting sentence start.
end_word: Special word denoting sentence end.
unk_word: Special word denoting unknown words.
vocab_from_file: If False, create vocab from scratch & override any existing vocab_file.
If True, load vocab from from existing vocab_file, if it exists.
num_workers: Number of subprocesses to use for data loading
cocoapi_loc: The location of the folder containing the COCO API: https://github.com/cocodataset/cocoapi
"""
assert mode is 'val', "mode must be 'val'
if vocab_from_file==False: assert mode=='val', "To generate vocab from captions file, must be in validation mode (mode='val')."
# Based on mode,obtain img_folder and annotations_file.
assert batch_size == 1 "Please change batch_size to 1 if validating your model."
assert os.path.exists(vocab_file), "Must first generate vocab.pkl from training data."
assert vocab_from_file == True "Must first generate vocab.pkl from training data."
img_folder = os.path.join(cocoapi_loc, 'cocoapi/images/val2014/')
annotations_file = os.path.join(cocoapi_loc, 'cocoapi/annotations/captions_val2014.json')
# COCO caption dataset.
dataset = CoCoDataset(transform=transform,
mode=mode,
batch_size=batch_size,
vocab_threshold=vocab_threshold,
vocab_file=vocab_file,
start_word=start_word,
end_word=end_word,
unk_word=unk_word,
annotations_file=annotations_file,
vocab_from_file=vocab_from_file,
img_folder=img_folder)
if mode == 'train':
# Randomly sample a caption length, and sample indices with that length.
indices = dataset.get_train_indices()
# Create and assign a batch sampler to retrieve a batch with the sampled indices.
initial_sampler = data.sampler.SubsetRandomSampler(indices=indices)
# data loader for COCO dataset.
data_loader = data.DataLoader(dataset=dataset,
num_workers=num_workers,
batch_sampler=data.sampler.BatchSampler(sampler=initial_sampler,
batch_size=dataset.batch_size,
drop_last=False))
else:
data_loader = data.DataLoader(dataset=dataset,
batch_size=dataset.batch_size,
shuffle=True,
num_workers=num_workers)
return data_loader
class CoCoDataset(data.Dataset):
def __init__(self, transform, mode, batch_size, vocab_threshold, vocab_file, start_word,
end_word, unk_word, annotations_file, vocab_from_file, img_folder):
self.transform = transform
self.mode = mode
self.batch_size = batch_size
self.vocab = Vocabulary(vocab_threshold, vocab_file, start_word,
end_word, unk_word, annotations_file, vocab_from_file)
self.img_folder = img_folder
if self.mode == 'train':
self.coco = COCO(annotations_file)
self.ids = list(self.coco.anns.keys())
print('Obtaining caption lengths...')
all_tokens = [nltk.tokenize.word_tokenize(str(self.coco.anns[self.ids[index]]['caption']).lower()) for index in tqdm(np.arange(len(self.ids)))]
self.caption_lengths = [len(token) for token in all_tokens]
else:
test_info = json.loads(open(annotations_file).read())
self.paths = [item['file_name'] for item in test_info['images']]
def __getitem__(self, index):
# obtain image and caption if in training mode
if self.mode == 'train':
ann_id = self.ids[index]
caption = self.coco.anns[ann_id]['caption']
img_id = self.coco.anns[ann_id]['image_id']
path = self.coco.loadImgs(img_id)[0]['file_name']
# Convert image to tensor and pre-process using transform
image = Image.open(os.path.join(self.img_folder, path)).convert('RGB')
image = self.transform(image)
# Convert caption to tensor of word ids.
tokens = nltk.tokenize.word_tokenize(str(caption).lower())
caption = []
caption.append(self.vocab(self.vocab.start_word))
caption.extend([self.vocab(token) for token in tokens])
caption.append(self.vocab(self.vocab.end_word))
caption = torch.Tensor(caption).long()
# return pre-processed image and caption tensors
return image, caption
# obtain image if in test mode
else:
path = self.paths[index]
# Convert image to tensor and pre-process using transform
PIL_image = Image.open(os.path.join(self.img_folder, path)).convert('RGB')
orig_image = np.array(PIL_image)
image = self.transform(PIL_image)
# return original image and index
return orig_image, index
def get_train_indices(self):
sel_length = np.random.choice(self.caption_lengths)
all_indices = np.where([self.caption_lengths[i] == sel_length for i in np.arange(len(self.caption_lengths))])[0]
indices = list(np.random.choice(all_indices, size=self.batch_size))
return indices
def __len__(self):
if self.mode == 'train':
return len(self.ids)
else:
return len(self.paths)