/GCAB-CFDC

Official Implementation of the paper "Exemplar-free Continual Learning of Vision Transformers via Gated Class-Attention and Cascaded Feature Drift Compensation"

Primary LanguagePython

Exemplar-free Continual Learning of Vision Transformers via Gated Class-Attention and Cascaded Feature Drift Compensation

Marco Cotogni, Fei Yang, Claudio Cusano, Andrew D. Bagdanov and Joost van de Weijer

ArXiv version of our Paper is available.

Architecture

Running code

In order to use our code, it is possible to create a conda environment using the requirements.txt file and Python 3.9.

For training the model please run:

python3 main.py --options options/data/DATASET_SCENARIO options/data/DATASET_ORDER 
options/model/DATASET_GCAB_FDC --name GCAB --data-path PATH 
--output-basedir OUTPUTDIR --compress COMPRESS_FILE.txt --report REPORT

For evaluating the trained model:

python3 main.py --options options/data/DATASET_SCENARIO options/data/DATASET_ORDER 
options/model/DATASET_GCAB_FDC --name GCAB --data-path PATH 
--output-basedir OUTPUTDIR --compress COMPRESS_FILE.txt --report REPORT 
--resume PATH_TO_THE_MODEL --eval

GCAB

Reference

If you are considering using our code or you want to cite our paper please use:

@article{cotogni2022gated,
  author = {Cotogni, Marco and Yang, Fei and Cusano, Claudio and Bagdanov, Andrew D. and van de Weijer, Joost},
  title = {Gated Class-Attention with Cascaded Feature Drift Compensation for Exemplar-free Continual Learning of Vision Transformers}, 
  journal={arXiv preprint arxiv:2211.12292},
  year={2022}
}

Credits

Our code is based on:

DyToX HAT