This repository will hold the official code of SelTDA, the self-training framework introduced in our CVPR 2023 paper "Q: How to Specialize Large Vision-Language Models to Data-Scarce VQA Tasks? A: Self-Train on Unlabeled Images!".

conda env create -f environment.yaml

• PathVQA
  • then use convert_pathvqa.py
• RSVQA
  • then use convert_rsvqa.py
• OK-VQA and A-OKVQA (use LAVIS)
  • LAVIS should automatically put them in the correct format, but if not, you can use convert_okvqa.py
• VQA Counterexamples
  • then use convert_vqa_ce.py
• AdVQA
  • then use convert_advqa.py
• VQA Rephrasings
  • then use convert_vqa_rephrasings.py

In general, the code expects that each VQA dataset is represented by a single JSON object that is a list of dictionaries. In schemas.py, we provide Pydantic models which you can use to define your own datasets or verify that the data is in the correct format.

See the examples/ directory to see examples of:

• training the teacher
  • examples/train_teacher.sh
• generating synthetic data with the teacher
  • examples/generate_synthetic_data.sh
• self-training with the synthetic data
  • examples/self_train_synthetic.sh
• evaluations
  • examples/evaluate.sh

@InProceedings{Khan_2023_CVPR,
    author    = {Khan, Zaid and BG, Vijay Kumar and Schulter, Samuel and Yu, Xiang and Fu, Yun and Chandraker, Manmohan},
    title     = {Q: How To Specialize Large Vision-Language Models to Data-Scarce VQA Tasks? A: Self-Train on Unlabeled Images!},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {15005-15015}
}

This code is heavily based on salesforce/BLIP.