File size: 12,209 Bytes
1120a2f | 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 | import os
import sys
import torch.utils.data as data
import torch
import numpy as np
from PIL import Image
import pdb
import copy
import random
from random import choice
from bert.tokenization_bert import BertTokenizer
from textblob import TextBlob
import random
import h5py
from refer.refer import REFER
import json
from args import get_parser
# Dataset configuration initialization
parser = get_parser()
args = parser.parse_args()
class ReferDataset_HP_Filter(data.Dataset):
def __init__(self,
args,
image_transforms=None,
target_transforms=None,
split='train',
eval_mode=False):
self.classes = []
self.image_transforms = image_transforms
self.target_transform = target_transforms
self.split = split
self.refer = REFER(args.refer_data_root, args.dataset, args.splitBy)
self.dataset_type = args.dataset
self.max_tokens = 20
ref_ids = self.refer.getRefIds(split=self.split)
self.img_ids = self.refer.getImgIds()
all_imgs = self.refer.Imgs
self.imgs = list(all_imgs[i] for i in self.img_ids)
self.ref_ids = ref_ids
self.input_ids = []
self.input_ids_masked = []
self.attention_masks = []
self.tokenizer = BertTokenizer.from_pretrained(args.bert_tokenizer)
# for metric learning #####################
# self.ROOT = '/data2/dataset/RefCOCO/VRIS'
self.ROOT = '/data2/projects/seunghoon/VerbRIS/VerbCentric_CY/datasets/VRIS'
self.all_hp_root = "/data2/dataset/RefCOCO/refcocog/SBERT_gref_umd"
self.metric_learning = args.metric_learning
self.exclude_multiobj = args.exclude_multiobj
self.metric_mode = args.metric_mode
self.exclude_position = False
self.hp_selection = args.hp_selection
self.get_all_verbs = args.get_all_verbs
if self.metric_learning and eval_mode == False:
self.hardneg_prob = args.hn_prob
self.multi_obj_ref_ids = self._load_multi_obj_ref_ids()
self.hardpos_meta = self._load_metadata()
else:
self.hardneg_prob = 0.0
self.multi_obj_ref_ids = None
self.hardpos_meta = None
#############################################
self.eval_mode = eval_mode
for r in ref_ids:
ref = self.refer.Refs[r]
sentences_for_ref = []
sentences_for_ref_masked = []
attentions_for_ref = []
for i, (el, sent_id) in enumerate(zip(ref['sentences'], ref['sent_ids'])):
sentence_raw = el['raw']
attention_mask = [0] * self.max_tokens
padded_input_ids = [0] * self.max_tokens
padded_input_ids_masked = [0] * self.max_tokens
blob = TextBlob(sentence_raw.lower())
chara_list = blob.tags
mask_ops = []
mask_ops1 = []
for word_i, (word_now, chara) in enumerate(chara_list):
if (chara == 'NN' or chara == 'NNS') and word_i < 19 and word_now.lower():
mask_ops.append(word_i)
mask_ops1.append(word_now)
mask_ops2 = self.get_adjacent_word(mask_ops)
input_ids = self.tokenizer.encode(text=sentence_raw, add_special_tokens=True)
# truncation of tokens
input_ids = input_ids[:self.max_tokens]
padded_input_ids[:len(input_ids)] = input_ids
attention_mask[:len(input_ids)] = [1]*len(input_ids)
if len(mask_ops) == 0:
attention_remask = attention_mask
input_ids_masked = input_ids
else:
could_mask = choice(mask_ops2)
input_ids_masked = copy.deepcopy(input_ids)
for i in could_mask:
input_ids_masked[i + 1] = 0
padded_input_ids_masked[:len(input_ids_masked)] = input_ids_masked
sentences_for_ref.append(torch.tensor(padded_input_ids).unsqueeze(0))
sentences_for_ref_masked.append(torch.tensor(padded_input_ids_masked).unsqueeze(0))
attentions_for_ref.append(torch.tensor(attention_mask).unsqueeze(0))
self.input_ids.append(sentences_for_ref)
self.input_ids_masked.append(sentences_for_ref_masked)
self.attention_masks.append(attentions_for_ref)
def get_classes(self):
return self.classes
def __len__(self):
return len(self.ref_ids)
def get_adjacent_word(self, mask_list):
output_mask_list = []
length = len(mask_list)
i = 0
while i < length:
begin_pos = i
while i+1 < length and mask_list[i+1] == mask_list[i] + 1:
i += 1
end_pos = i+1
output_mask_list.append(mask_list[begin_pos:end_pos])
i = end_pos
return output_mask_list
# for metric learning #####################
###########################################
def _tokenize(self, sentence):
attention_mask = [0] * self.max_tokens
padded_input_ids = [0] * self.max_tokens
input_ids = self.tokenizer.encode(text=sentence, add_special_tokens=True)
# truncation of tokens
input_ids = input_ids[:self.max_tokens]
padded_input_ids[:len(input_ids)] = input_ids
attention_mask[:len(input_ids)] = [1]*len(input_ids)
# match shape as (1, max_tokens)
return torch.tensor(padded_input_ids).unsqueeze(0), torch.tensor(attention_mask).unsqueeze(0)
def _load_multi_obj_ref_ids(self):
# Load multi-object reference IDs based on configurations
if not self.exclude_multiobj and not self.exclude_position :
return None
elif self.exclude_position:
multiobj_path = os.path.join(self.ROOT, 'multiobj_ov2_nopos.txt')
elif self.exclude_multiobj :
multiobj_path = os.path.join(self.ROOT, 'multiobj_ov3.txt')
with open(multiobj_path, 'r') as f:
return [int(line.strip()) for line in f.readlines()]
def _load_metadata(self):
# Load metadata for hard positive verb phrases, hard negative queries
if 'refined' in self.metric_mode or 'hardneg' in self.metric_mode :
hardpos_path = '/data2/projects/seunghoon/VerbRIS/CrossVLT/hardpos_verdict_gref_v4.json'
else :
hardpos_path = os.path.join(self.ROOT, 'hardpos_verbphrase_0906upd.json')
with open(hardpos_path, 'r', encoding='utf-8') as f:
hardpos_json = json.load(f)
return hardpos_json
def _get_hardpos_verb(self, ref, seg_id, sent_idx) :
sbert_emb_size = 384
if seg_id in self.multi_obj_ref_ids:
verb_embed = torch.zeros(sbert_emb_size, dtype=torch.float32)
return '', verb_embed
# Extract metadata for hard positives if present
hardpos_dict = self.hardpos_meta.get(str(seg_id), {})
if self.hp_selection == 'strict' :
sent_id_list = list(hardpos_dict.keys())
cur_sent_id = sent_id_list[sent_idx]
cur_hardpos = hardpos_dict.get(cur_sent_id, {}).get('phrases', [])
# only implement hp_selection 'strict' mode.
if cur_hardpos :
rand_index = random.randint(0, len(cur_hardpos) - 1)
raw_verb = cur_hardpos[rand_index]
verb_embed = torch.from_numpy(self._get_hardpos_embed(seg_id, cur_sent_id, rand_index))
if self.get_all_verbs :
return raw_verb, verb_embed
else :
return cur_hardpos, verb_embed
verb_embed = torch.zeros(sbert_emb_size, dtype=torch.float32)
return '', verb_embed
def _get_hardpos_embed(self, seg_id, sent_id, rand_index):
emb_folder = os.path.join(self.all_hp_root, str(seg_id))
emb_files = sorted([f for f in os.listdir(emb_folder) if f.startswith(f"hp_{sent_id}_") and f.endswith(".npy")])
selected_emb_file = os.path.join(emb_folder, emb_files[rand_index])
return np.load(selected_emb_file)
# def _get_hardpos_verb_singlephrase(self, ref, seg_id, sent_idx) :
# if seg_id in self.multi_obj_ref_ids:
# return ''
# # Extract metadata for hard positives if present
# hardpos_dict = self.hardpos_meta.get(str(seg_id), {})
# if self.hp_selection == 'strict' :
# sent_id_list = list(hardpos_dict.keys())
# cur_hardpos = hardpos_dict.get(sent_id_list[sent_idx], {}).get('phrases', [])
# else :
# cur_hardpos = list(itertools.chain.from_iterable(hardpos_dict[sid]['phrases'] for sid in hardpos_dict))
# if cur_hardpos:
# # Assign a hard positive verb phrase if available
# raw_verb = random.choice(cur_hardpos)
# return raw_verb
# return ''
###########################################
###########################################
def __getitem__(self, index):
this_ref_id = self.ref_ids[index]
this_img_id = self.refer.getImgIds(this_ref_id)
this_img = self.refer.Imgs[this_img_id[0]]
img = Image.open(os.path.join('/data2/dataset/COCO2014/trainval2014/', this_img['file_name'])).convert("RGB")
ref = self.refer.loadRefs(this_ref_id)
if self.dataset_type == 'ref_zom':
source_type = ref[0]['source']
else:
source_type = 'one'
ref_mask = np.array(self.refer.getMask(ref[0])['mask'])
annot = np.zeros(ref_mask.shape)
annot[ref_mask == 1] = 1
annot = Image.fromarray(annot.astype(np.uint8), mode="P")
if self.image_transforms is not None:
img, target = self.image_transforms(img, annot)
if self.eval_mode:
embedding = []
embedding_masked = []
att = []
for s in range(len(self.input_ids[index])):
e = self.input_ids[index][s]
a = self.attention_masks[index][s]
embedding.append(e.unsqueeze(-1))
embedding_masked.append(e.unsqueeze(-1))
att.append(a.unsqueeze(-1))
tensor_embeddings = torch.cat(embedding, dim=-1)
tensor_embeddings_masked = torch.cat(embedding_masked, dim=-1)
attention_mask = torch.cat(att, dim=-1)
return img, target, source_type, tensor_embeddings, tensor_embeddings_masked, attention_mask
else:
# train phase
choice_sent = np.random.choice(len(self.input_ids[index]))
tensor_embeddings = self.input_ids[index][choice_sent]
tensor_embeddings_masked = self.input_ids_masked[index][choice_sent]
attention_mask = self.attention_masks[index][choice_sent]
if self.metric_learning :
pos_sent = torch.zeros_like(tensor_embeddings)
pos_attn_mask = torch.zeros_like(attention_mask)
## Only the case with hardpos_ in metric_mode
if 'hardpos_' in self.metric_mode or self.hardneg_prob == 0.0:
pos_type = 'zero'
# we always use refined mode as default.
if 'refined' in self.metric_mode :
pos_sent_picked, hardpos_embed = self._get_hardpos_verb(ref, this_ref_id, choice_sent)
if pos_sent_picked:
pos_type = 'hardpos'
pos_sent, pos_attn_mask = self._tokenize(pos_sent_picked)
return img, target, source_type, tensor_embeddings, tensor_embeddings_masked, attention_mask, pos_sent, pos_attn_mask, pos_type, hardpos_embed
return img, target, source_type, tensor_embeddings, tensor_embeddings_masked, attention_mask
|