Spaces:
Running
Running
Upload data/pretrain_dataset.py
Browse files- data/pretrain_dataset.py +59 -0
data/pretrain_dataset.py
ADDED
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
import os
|
3 |
+
import random
|
4 |
+
|
5 |
+
from torch.utils.data import Dataset
|
6 |
+
|
7 |
+
from PIL import Image
|
8 |
+
from PIL import ImageFile
|
9 |
+
ImageFile.LOAD_TRUNCATED_IMAGES = True
|
10 |
+
Image.MAX_IMAGE_PIXELS = None
|
11 |
+
|
12 |
+
from data.utils import pre_caption
|
13 |
+
import os,glob
|
14 |
+
|
15 |
+
class pretrain_dataset(Dataset):
|
16 |
+
def __init__(self, ann_file, laion_path, transform):
|
17 |
+
|
18 |
+
self.ann_pretrain = []
|
19 |
+
for f in ann_file:
|
20 |
+
print('loading '+f)
|
21 |
+
ann = json.load(open(f,'r'))
|
22 |
+
self.ann_pretrain += ann
|
23 |
+
|
24 |
+
self.laion_path = laion_path
|
25 |
+
if self.laion_path:
|
26 |
+
self.laion_files = glob.glob(os.path.join(laion_path,'*.json'))
|
27 |
+
|
28 |
+
print('loading '+self.laion_files[0])
|
29 |
+
with open(self.laion_files[0],'r') as f:
|
30 |
+
self.ann_laion = json.load(f)
|
31 |
+
|
32 |
+
self.annotation = self.ann_pretrain + self.ann_laion
|
33 |
+
else:
|
34 |
+
self.annotation = self.ann_pretrain
|
35 |
+
|
36 |
+
self.transform = transform
|
37 |
+
|
38 |
+
|
39 |
+
def reload_laion(self, epoch):
|
40 |
+
n = epoch%len(self.laion_files)
|
41 |
+
print('loading '+self.laion_files[n])
|
42 |
+
with open(self.laion_files[n],'r') as f:
|
43 |
+
self.ann_laion = json.load(f)
|
44 |
+
|
45 |
+
self.annotation = self.ann_pretrain + self.ann_laion
|
46 |
+
|
47 |
+
|
48 |
+
def __len__(self):
|
49 |
+
return len(self.annotation)
|
50 |
+
|
51 |
+
def __getitem__(self, index):
|
52 |
+
|
53 |
+
ann = self.annotation[index]
|
54 |
+
|
55 |
+
image = Image.open(ann['image']).convert('RGB')
|
56 |
+
image = self.transform(image)
|
57 |
+
caption = pre_caption(ann['caption'],30)
|
58 |
+
|
59 |
+
return image, caption
|