ARKseal commited on
Commit
24a0342
1 Parent(s): 99c6d9a

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +34 -0
README.md ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Filtered WIT, an Image-Text Dataset.
2
+ A reliable Dataset to run Image-Text models.
3
+
4
+ You can find WIT, Wikipedia Image Text Dataset, [here](https://github.com/google-research-datasets/wit)
5
+ Data was taken from [dalle-mini/wit](https://huggingface.co/datasets/dalle-mini/wit)
6
+
7
+ ## Author
8
+ - [Aarush Katta](https://github.com/ARKseal)
9
+
10
+ ## Data Structure
11
+ The data is stored as tars, containing 10,000 samples per tar.
12
+ Each tar contains a `.jpg`, `.txt`, and `.json`.
13
+ The image is stored in `.jpg`, the caption in `.txt.` and the metadata in `.json`
14
+ The preferred method to read the data is [WebDataset](https://github.com/webdataset/webdataset)
15
+ Here's an example:
16
+ ```python
17
+ import webdataset as wds
18
+
19
+ dataset = wds.WebDataset('data/00000.tar').to_tuple('txt', 'jpg', 'json')
20
+
21
+ for text, image, meta in dataset:
22
+ print(
23
+ text[:50],
24
+ image[:50],
25
+ meta[:50]
26
+ )
27
+ ```
28
+
29
+ ## Filteration
30
+ Each sample has 8 possible captions which were compared to the image using [CLIP ViT-B32](https://arxiv.org/abs/2103.00020)
31
+ The text was encoded using [multilingual CLIP text encoder](https://huggingface.co/sentence-transformers/clip-ViT-B-32-multilingual-v1)
32
+ Each possible caption was compared to the encoded image using Cosign Similarity
33
+ and kept if the sim was greater than `0.26`
34
+ Then the new caption was the filtered captions concatenated, and samples with no filtered caption were dropped.