File size: 6,219 Bytes
ea1014e |
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 |
import os
from typing import List, Union
import cv2
from PIL import Image
import lmdb
import numpy as np
import pyarrow as pa
import torch
from torch.utils.data import Dataset
from torchvision.transforms import functional as F
from bert.tokenization_bert import BertTokenizer
info = {
'refcoco': {
'train': 42404,
'val': 3811,
'val-test': 3811,
'testA': 1975,
'testB': 1810
},
'refcoco+': {
'train': 42278,
'val': 3805,
'val-test': 3805,
'testA': 1975,
'testB': 1798
},
'refcocog_u': {
'train': 42226,
'val': 2573,
'val-test': 2573,
'test': 5023
},
'refcocog_g': {
'train': 44822,
'val': 5000,
'val-test': 5000
}
}
_tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
def tokenize(texts: Union[str, List[str]],
context_length: int = 77,
truncate: bool = False) -> torch.LongTensor:
"""
Returns the tokenized representation of given input string(s)
Parameters
----------
texts : Union[str, List[str]]
An input string or a list of input strings to tokenize
context_length : int
The context length to use; all CLIP models use 77 as the context length
truncate: bool
Whether to truncate the text in case its encoding is longer than the context length
Returns
-------
A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]
"""
l_mask = [0] * context_length
result = [0] * context_length
tokens = _tokenizer.encode(text=texts, add_special_tokens=True)
tokens = tokens[:context_length]
result[:len(tokens)] = tokens
l_mask[:len(tokens)] = [1]*len(tokens)
result = torch.tensor(result).unsqueeze(0)
l_mask = torch.tensor(l_mask).unsqueeze(0)
return result, l_mask
def loads_pyarrow(buf):
"""
Args:
buf: the output of `dumps`.
"""
return pa.deserialize(buf)
class RefDataset(Dataset):
def __init__(self, lmdb_dir, mask_dir, dataset, split, mode, input_size,
word_length, args):
super(RefDataset, self).__init__()
self.lmdb_dir = lmdb_dir
self.mask_dir = mask_dir
self.dataset = dataset
self.split = split
self.mode = mode
self.input_size = (input_size, input_size)
#self.mask_size = [13, 26, 52]
self.word_length = word_length
self.mean = torch.tensor([0.485, 0.456, 0.406]).reshape(3, 1, 1)
self.std = torch.tensor([0.229, 0.224, 0.225]).reshape(3, 1, 1)
self.length = info[dataset][split]
self.env = None
self.args = args
# self.coco_transforms = make_coco_transforms(mode, cautious=False)
def _init_db(self):
self.env = lmdb.open(self.lmdb_dir,
subdir=os.path.isdir(self.lmdb_dir),
readonly=True,
lock=False,
readahead=False,
meminit=False)
with self.env.begin(write=False) as txn:
self.length = loads_pyarrow(txn.get(b'__len__'))
self.keys = loads_pyarrow(txn.get(b'__keys__'))
def __len__(self):
return self.length
def __getitem__(self, index):
# Delay loading LMDB data until after initialization: https://github.com/chainer/chainermn/issues/129
if self.env is None:
self._init_db()
env = self.env
with env.begin(write=False) as txn:
byteflow = txn.get(self.keys[index])
ref = loads_pyarrow(byteflow)
# img
ori_img = cv2.imdecode(np.frombuffer(ref['img'], np.uint8),
cv2.IMREAD_COLOR)
img = cv2.cvtColor(ori_img, cv2.COLOR_BGR2RGB)
img_size = img.shape[:2]
# mask
seg_id = ref['seg_id']
mask_dir = os.path.join(self.mask_dir, str(seg_id) + '.png')
# sentences
idx = np.random.choice(ref['num_sents'])
sents = ref['sents']
# transform
# mask transform
mask = cv2.imdecode(np.frombuffer(ref['mask'], np.uint8),
cv2.IMREAD_GRAYSCALE)
mask = mask / 255.
if self.mode == 'train':
sent = sents[idx]
# sentence -> vector
img, mask, sent = self.convert(img, mask, sent, inference=False)
word_vec, pad_mask = tokenize(sent, self.word_length, True)
return img, word_vec, mask, pad_mask
elif self.mode == 'val':
# sentence -> vector
sent = sents[-1]
word_vec, pad_mask = tokenize(sent, self.word_length, True)
img, mask, sent = self.convert(img, mask, sent, inference=False)
return img, word_vec, mask, pad_mask
else:
# sentence -> vector
word_vecs = []
pad_masks = []
for sent in sents:
word_vec, pad_mask = tokenize(sent, self.word_length, True)
word_vecs.append(word_vec)
pad_masks.append(pad_mask)
img, mask, sent = self.convert(img, mask, sent, inference=True)
return ori_img, img, word_vecs, mask, pad_masks, seg_id, sents,
def convert(self, img, mask, sent, inference=False):
img = Image.fromarray(np.uint8(img))
mask = Image.fromarray(np.uint8(mask), mode="P")
img = F.resize(img, self.input_size)
if not inference:
mask = F.resize(mask, self.input_size, interpolation=Image.NEAREST)
img = F.to_tensor(img)
mask = torch.as_tensor(np.asarray(mask).copy(), dtype=torch.int64)
img = F.normalize(img, mean=self.mean, std=self.std)
return img, mask, sent
def __repr__(self):
return self.__class__.__name__ + "(" + \
f"db_path={self.lmdb_dir}, " + \
f"dataset={self.dataset}, " + \
f"split={self.split}, " + \
f"mode={self.mode}, " + \
f"input_size={self.input_size}, " + \
f"word_length={self.word_length}"
|