-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdataset_own.py
241 lines (218 loc) · 11.6 KB
/
dataset_own.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
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import glob
import json
import os
import numpy as np
import random
from util_transform import *
curdir = os.path.dirname(__file__)
save_dir = os.path.join(curdir, 'util/image_retrival/')
categorys_dict = {0: '上衣', 1: '裙装', 2: '下衣', 3: '箱包', 4: '鞋子', 5: '配饰', 6: '零食', 7: '美妆', 8: '瓶饮', 9: '家具',
20: '玩具', 21: '内衣', 22: '数码硬件', 8888: '其他', 88888888: '其他'}
class ImageTextDataset(Dataset):
def __init__(self, transform_image, is_train=True, cropsize=512, data_mode = '_imageretirval', test_num=10000):
self.is_train = is_train
self.imgs_dir = '/home/senwang/dataset/image_retrival2_42w/' # '/home/senwang/dataset/image_retrival2_test2/' # '/home/admin/workspace/tbq/data/image_retrival2/
if not os.path.exists(os.path.join(self.imgs_dir, f'img_names_train{data_mode}.txt')): # 'img_names_train.txt'
print('收集图片和tag')
# self.image_paths = glob.glob(imgs_dir + '/*.webp') #
# self.image_paths = self.get_img_paths(imgs_dir)
self.image_paths = np.loadtxt(self.imgs_dir + 'image_list.txt', dtype=str)
with open(os.path.join(curdir, 'util/sales_predict/total_sp_dict_spuid2tags.json'), 'r') as file:
self.tag_data = json.load(file)
f_sale_dict = curdir + '/util/sales_predict/total_sp_dict_spuid2saletargetlife_from2016_mj_only.json'
with open(f_sale_dict, 'r') as file: # total_sp_dict_spuid2saletarget
self.target_data_mj_only = json.load(file)
# f_sale_dict = curdir + '/util/sales_predict/total_sp_dict_spuid2saletargetlife_from2016_offline_only.json'
# with open(f_sale_dict, 'r') as file: # total_sp_dict_spuid2saletarget
# self.target_data_offline_only = json.load(file) # 吊牌价线上和线下是一致的
f_spu2anchor = os.path.join(save_dir, 'fea_extract_dict_spuid2img2anchor.json')
self.spu2anchor = json.loads(open(f_spu2anchor, 'r').read())
self.image_paths.sort()
train_image_paths = self.image_paths[:-test_num]
test_image_paths = self.image_paths[-test_num:]
self.img_names_train, self.texts_train, self.anchors_train = self.get_scores_accord(train_image_paths)
self.img_names_test, self.texts_test, self.anchors_test = self.get_scores_accord(test_image_paths)
self._cleanup_early_data()
np.savetxt(os.path.join(self.imgs_dir, f'img_names_train{data_mode}.txt'), self.img_names_train, fmt='%s') # 'img_names_train.txt'
np.savetxt(os.path.join(self.imgs_dir, f'texts_train{data_mode}.txt'), self.texts_train, fmt='%s') # 'texts_train.txt'
np.savetxt(os.path.join(self.imgs_dir, f'anchors_train{data_mode}.txt'), self.anchors_train, fmt='%s') # 'anchors_train.txt'
np.savetxt(os.path.join(self.imgs_dir, f'img_names_test{data_mode}.txt'), self.img_names_test, fmt='%s') #'img_names_test.txt'
np.savetxt(os.path.join(self.imgs_dir, f'texts_test{data_mode}.txt'), self.texts_test, fmt='%s')
np.savetxt(os.path.join(self.imgs_dir, f'anchors_test{data_mode}.txt'), self.anchors_test, fmt='%s')
else:
self.img_names_train = np.loadtxt(os.path.join(self.imgs_dir, f'img_names_train{data_mode}.txt'), dtype=str)
with open(os.path.join(self.imgs_dir, f'texts_train{data_mode}.txt'), 'r') as file:
self.texts_train = file.readlines()
self.anchors_train = np.loadtxt(os.path.join(self.imgs_dir, f'anchors_train{data_mode}.txt'), dtype=int)
self.img_names_test = np.loadtxt(os.path.join(self.imgs_dir, f'img_names_test{data_mode}.txt'), dtype=str)
with open(os.path.join(self.imgs_dir, f'texts_test{data_mode}.txt'), 'r') as file:
self.texts_test = file.readlines()
self.anchors_test = np.loadtxt(os.path.join(self.imgs_dir, f'anchors_test{data_mode}.txt'), dtype=int)
# debug 用少量的数据如20个
# self.img_names_train = self.img_names_train[:20]
# self.texts_train = self.texts_train[:20]
# self.anchors_train = self.anchors_train[:20]
# self.img_names_test = self.img_names_test[:20]
# self.texts_test = self.texts_test[:20]
# self.anchors_test = self.anchors_test[:20]
self.transform_image = transform_image
self.transform_crop = Compose([
ColorJitterImgOnlyGivenBox(
brightness=0.1,
contrast=0.1,
saturation=0.1),
# RandomScale((0.6, 0.75, 1.0, 1.0, 1.25, 1.5, 1.75, 2.0)), # tbq modify from (0.75, 1.0, 1.25, 1.5, 1.75, 2.0) to (0.6, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0)
RandomCropImgOnlyGivenBox(),
HorizontalFlipImgOnlyGivenBox()
])
self.transform_crop_test = Compose([
RandomCropImgOnlyGivenBox(is_train=False),
])
def get_img_paths(self, imgs_dir, extension = '.webp'):
files = []
cnt = 0
with os.scandir(imgs_dir) as entries:
for entry in entries:
# if entry.is_file() and entry.name.endswith(extension):
files.append(entry.path)
cnt += 1
print(cnt)
print(f"Found {len(files)} files.")
return files
def _cleanup_early_data(self):
"""清理前期加载的冗余数据"""
del self.tag_data
del self.target_data_mj_only
del self.spu2anchor
del self.image_paths # 注意：应在拆分train/test后清理
# 强制垃圾回收（需import gc）
import gc
gc.collect()
def __len__(self):
if self.is_train:
return len(self.img_names_train)
else:
return len(self.img_names_test)
def __getitem__(self, idx):
def getitem(idx):
if self.is_train:
image = Image.open(os.path.join(self.imgs_dir, self.img_names_train[idx])).convert('RGB')
text = self.texts_train[idx]
### 随机变换后面不重要词条的位置
text_list = text.split(',')
text_pre = ','.join(text_list[:4])
text_suffix = text_list[4:].copy()
random.shuffle(text_suffix)
text_suffix = ','.join(text_suffix)
text = text_pre + ',' + text_suffix
box = self.anchors_train[idx]
else:
image = Image.open(os.path.join(self.imgs_dir,self.img_names_test[idx])).convert('RGB')
text = self.texts_test[idx]
box = self.anchors_test[idx]
if self.is_train:
input_data = {'im': image, 'box': box}
output_data = self.transform_crop(input_data) # 随机crop
image = output_data['im']
else:
input_data = {'im': image, 'box': box}
output_data = self.transform_crop_test(input_data)
image = output_data['im']
if self.transform_image:
image = self.transform_image(image)
return image, text
try:
image, text = getitem(idx)
except Exception as e:
print('##error:', e, self.img_names_train[idx])
image, text = getitem(0)
return image, text
def get_scores_accord(self,fnames):
texts = []
img_names = []
anchors = []
total = len(fnames)
for i, fname in enumerate(fnames):
print(f'{i}/{total}, {fname}')
try:
name = os.path.basename(fname)
spu_id = name.split('_')[0]
text_str = ''
text = self.tag_data[spu_id]
text, res = self.find_important_fea(text.copy())
text_str += res
# 获取每个spu的吊牌价格
if spu_id in self.target_data_mj_only:
text_str += '吊牌价:' + str(int(self.target_data_mj_only[spu_id]['avg_ori_price'])) + ','
text_str_other = str(text).replace('\'', '').replace('{', '').replace('}', '')
text_str += text_str_other
# 获取每个图片的anchor
img2anchor_dict = self.spu2anchor[spu_id]
anchor_multi = img2anchor_dict[os.path.basename(fname)]
for anchor in anchor_multi:
box = [int(i) for i in anchor[0].split(',')]
cls = categorys_dict[int(anchor[1])]
text_str = '商品类别:' + cls + ',' + text_str
texts.append(text_str)
img_names.append(fname)
anchors.append(box)
except Exception as e:
print('error:', fname, e)
# self.image_paths.remove(fname)
return img_names, texts, anchors
def find_important_fea(self, text):
res = ''
tmp_dict = {}
keys = list(text.keys())
for k in keys:
v = str(text[k])
v = '_'.join(list(set(v.split(',')))) # 去重， 如'无,无' -》 '无'
if '无' == v or 'nan' == v:
del text[k] # 删除无信息的元素
continue
if '品牌' == k or '商品品牌' == k:
if '品牌' in tmp_dict:
tmp_dict['品牌'].append(v)
else:
tmp_dict['品牌'] = [v]
del text[k]
elif '面料' in k or '材质' in k or '材料' in k:
if '面料' in tmp_dict:
tmp_dict['面料'].append(v)
else:
tmp_dict['面料'] = [v]
del text[k]
elif '卖点' in k:
if '卖点' in tmp_dict:
tmp_dict['卖点'].append(v)
else:
tmp_dict['卖点'] = [v]
del text[k]
elif '商品名称' in k or '品名' in k:
if '商品名称' in tmp_dict:
tmp_dict['商品名称'].append(v)
else:
tmp_dict['商品名称'] = [v]
del text[k]
elif '身高' in k or '厚薄' in k or '建议' in k or '场合' in k or '性别' in k or '重' in k or '制造' in k \
or '市场' in k or '年龄' in k or '年份' in k or '洗涤' in k or '体型' in k or '人群' in k or '支撑' in k \
or '合格' in k or '戴帽' in k or '电话' in k or '价' in k or '公司' in k or '地' in k or '货号' in k\
or '尺码' in k or '技术' in k or '安全' in k or '标准' in k or '身份' in k or '包装' in k or '修饰' in k:
del text[k]
else:
text[k] = v
if '品牌' in tmp_dict:
str_brand = '_'.join(list(set(tmp_dict['品牌']))) # 去重， 如'无,无' -》 '无'
res += '品牌:' + str_brand + ','
if '面料' in tmp_dict:
str_fabric = '_'.join(list(set(tmp_dict['面料']))) # 去重
res += '面料:' + str_fabric + ','
if '卖点' in tmp_dict:
str_sale = '_'.join(list(set(tmp_dict['卖点']))) # 去重
res += '卖点:' + str_sale + ','
if '商品名称' in tmp_dict:
str_name = '_'.join(list(set(tmp_dict['商品名称']))) # 去重
res += '商品名称:' + str_name + ','
return text, res