G2D-Boosting-Multimodal-Learning-with-Gradient-Guided-Distillation

This is the official PyTorch implementation of the paper G²D. The paper is available at arXiv. The slides, poster, and presentation video are available here.

Accepted by: ICCV 2025

Training

Environment config

1. Python: 3.11.0
2. CUDA Version: 11.7
3. Pytorch: 2.0.1
4. Torchvision: 0.15.2

Train

1. Set the hyperparameters in config.py.
2. Train the teacher using the teacher.py script:

python teacher.py

3. Train the student using the student.py script:

python student.py

Citation

@misc{rakib2025g2dboostingmultimodallearning,
      title={G$^{2}$D: Boosting Multimodal Learning with Gradient-Guided Distillation}, 
      author={Mohammed Rakib and Arunkumar Bagavathi},
      year={2025},
      eprint={2506.21514},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.21514}, 
}