zooblastlbz
commited on
Create resize.py
Browse files
resize.py
ADDED
@@ -0,0 +1,93 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from io import BytesIO
|
2 |
+
|
3 |
+
import ujson
|
4 |
+
import webdataset as wds
|
5 |
+
from PIL import Image
|
6 |
+
from tqdm import tqdm
|
7 |
+
import os
|
8 |
+
import albumentations as A
|
9 |
+
import cv2
|
10 |
+
def load_text(txt: bytes):
|
11 |
+
return txt.decode()
|
12 |
+
|
13 |
+
|
14 |
+
def load_image(jpg):
|
15 |
+
return Image.open(BytesIO(jpg)).convert('RGB')
|
16 |
+
|
17 |
+
|
18 |
+
load_mapping = {
|
19 |
+
'jpg': load_image,
|
20 |
+
'txt': load_text
|
21 |
+
}
|
22 |
+
|
23 |
+
|
24 |
+
def valid_image(img):
|
25 |
+
return min(img.size) >= 64
|
26 |
+
|
27 |
+
|
28 |
+
def resize_img(img, max_size=224):
|
29 |
+
width, height = img.size
|
30 |
+
|
31 |
+
if max(width,height) > max_size:
|
32 |
+
if width < height:
|
33 |
+
new_height = max_size
|
34 |
+
new_width = int(max_size * width / height)
|
35 |
+
else:
|
36 |
+
new_width = max_size
|
37 |
+
new_height = int(max_size * height / width)
|
38 |
+
img = img.resize((new_width, new_height), Image.BICUBIC)
|
39 |
+
|
40 |
+
# Pad the image to make it square
|
41 |
+
width, height = img.size
|
42 |
+
new_img = Image.new("RGB", (max_size, max_size), (255, 255, 255))
|
43 |
+
new_img.paste(img, ((max_size - width) // 2, (max_size - height) // 2))
|
44 |
+
|
45 |
+
return new_img
|
46 |
+
|
47 |
+
|
48 |
+
|
49 |
+
|
50 |
+
def img_to_meta(img):
|
51 |
+
width, height = img.size
|
52 |
+
return {
|
53 |
+
'width': width,
|
54 |
+
'height': height
|
55 |
+
}
|
56 |
+
|
57 |
+
|
58 |
+
def get_image(img):
|
59 |
+
if not valid_image(img):
|
60 |
+
return None, None
|
61 |
+
|
62 |
+
img=resize_img(img)
|
63 |
+
img_stream = BytesIO()
|
64 |
+
img.save(img_stream, format='jpeg')
|
65 |
+
|
66 |
+
img_stream.seek(0)
|
67 |
+
return img_stream.read(), ujson.dumps(img_to_meta(img))
|
68 |
+
|
69 |
+
change_mapping = {
|
70 |
+
'jpg': get_image
|
71 |
+
}
|
72 |
+
|
73 |
+
def func(wds_dataset_str, **kwargs):
|
74 |
+
ds = wds.WebDataset(wds_dataset_str, shardshuffle=False).map_dict(**load_mapping).map_dict(**change_mapping).to_tuple(
|
75 |
+
'jpg', 'txt')
|
76 |
+
dl = wds.WebLoader(ds, batch_size=None, num_workers=48, prefetch_factor=16, **kwargs)
|
77 |
+
|
78 |
+
|
79 |
+
writer = wds.ShardWriter('target_path/%05d.tar', 10000)
|
80 |
+
for img, txt in tqdm(dl):
|
81 |
+
img_str, meta = img
|
82 |
+
if img_str is None:
|
83 |
+
continue
|
84 |
+
sample = {
|
85 |
+
'__key__': f'{writer.count:08}',
|
86 |
+
'jpg': img_str,
|
87 |
+
'txt': txt,
|
88 |
+
'json': meta
|
89 |
+
}
|
90 |
+
writer.write(sample)
|
91 |
+
|
92 |
+
if __name__ == '__main__':
|
93 |
+
func('path_to_origin_data')
|