TaiSu: A 166M Large-scale High-Quality Dataset for Chinese Vision-Language Pre-training

This paper has been accepted by NeurIPS 2022.

• paper link: https://openreview.net/pdf?id=iAxH-ikIP0I

• Dataset Construction:

1. Data collection
2. Text-based filtering
3. Image-text-retrieval-based filtering
4. Image-Captioning-based text augmentation

Since most of the original urls are expired, we decided to directly provide the images and corresponding captions. To make the download process easier, we split the image set into more than 30 parts, and the captions are gathered in a single TXT file whose format of the content is shown in .

  ID*****Web_caption*****Generated Caption
  # Empty captions are replaced  with "None"

All the data can be downloaded via the following link: https://pan.baidu.com/s/1F5aKsurZkjZie09GsseOlw?pwd=vstf

The files with the suffix of '.tgz' need first to be uncompressed to a file with the suffix of '.tar' using the command line pigz -d baidu_images*.tgz . Even though a part of the images is damaged or lost because of some reasons, you can still access the most part of TaiSu's data. Each image and its captions can be matched by the id, for example, 'img1baiducomitu1848496827104259151'.

Here is a tutorial that tells you how to download the files from BaiduCloud to your server: https://blog.csdn.net/wxplol/article/details/115283527. Hope it can help you.

Models trained on the web data of TaiSu and on the complete data of TaiSu are now availbale. Baidu cloud link：https://pan.baidu.com/s/1d3UKyQi7J4Qr1XE2j2V8og?pwd=0kjm

• Example for usage:

from models.model_infer import build_lit
from clip.clip import _transform
from utils.sp_tokenizer import SentencepieceChineseTokenizer
from PIL import Image
lit=build_lit(visual_model_path=path/to/visual/model/state_dict,txt_model_path=path/to/textual/model/state_dict)
#viusal model and textual model should be matched.
'''API:
   lit.encode_image(imgs)
   lit.encode_text(txt) '''
device = "cpu"
transform=_transform(n_px=224)
tokenizer=SentencepieceChineseTokenizer(context_length=52)
image = transform(Image.open("xxx.png")).unsqueeze(0).to(device)
texts = tokenizer.tokenize(['我爱我的家乡','xxxx']).to(device)
with torch.no_grad():
     img_emb= lit.encode_image(image)
     txt_emb=lit.encode_text(texts)
     #The embeddings should be normalized to calculate cosine similarity
     img_emb=img_emb/img_emb.norm(dim=-1,keepdim=True)
     txt_emb=txt_emb/txt_emb.norm(dim=-1,keepdim=True)
     logits=img_emb@txt_emb.t()     
   

Unless specifically labeled otherwise, these Datasets are provided to You under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License (“CC BY-NC-SA 4.0”), with the additional terms included herein. The CC BY-NC-SA 4.0 may be accessed at https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode. When You download or use the Datasets from the Website or elsewhere, You are agreeing to comply with the terms of CC BY-NC-SA 4.0, and also agreeing to the Dataset Terms. Where these Dataset Terms conflict with the terms of CC BY-NC-SA 4.0, these Dataset Terms shall prevail. We reiterate once again that this dataset is used only for non-commercial purposes such as academic research, teaching, or scientific publications. We prohibits You from using the dataset or any derivative works for commercial purposes, such as selling data or using it for commercial gain.

If any of the images belongs to you and you would like it removed, please kindly inform us, we will remove it from our dataset immediately.

Email:[email protected]

Organization: Institute of Automation, Chinese Academy of Sciences (CASIA), Beijing, China

@inproceedings{liu2022taisu,
 author = {Liu, Yulong and Zhu, Guibo and Zhu, Bin and Song, Qi and Ge, Guojing and Chen, Haoran and Qiao, GuanHui and Peng, Ru and Wu, Lingxiang and Wang, Jinqiao},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {S. Koyejo and S. Mohamed and A. Agarwal and D. Belgrave and K. Cho and A. Oh},
 pages = {16705--16717},
 publisher = {Curran Associates, Inc.},
 title = {TaiSu: A 166M Large-scale High-Quality Dataset for Chinese Vision-Language Pre-training},
 url = {https://proceedings.neurips.cc/paper_files/paper/2022/file/6a386d703b50f1cf1f61ab02a15967bb-Paper-Datasets_and_Benchmarks.pdf},
 volume = {35},
 year = {2022}
}