BIRDNN(Behavior-Imitation Based Repair of DNNs) supports alternative retraining style and fine-tuning style repair simultaneously.
Code for the paper "BIRDNN: Behavior-Imitation Based Repair of Deep Neural Networks".
The code is based on Veritex and CARE. Thank to their wonderful and clear implementation.
This tool has been confirmed to work and tested with only Python3.7.
-
Install BIRDNN pkg with pip.
python -m pip install . -
Set path to /BIRDNN under this repository.
export PYTHONPATH='<YOUR_REPO_PATH>/BIRDNN' export OPENBLAS_NUM_THREADS=1 export OMP_NUM_THREADS=1
If you want to run DNN retraining code, here is an example of code running commands:
python BIRDNN/repair/retraining.py --network 19 --alpha 0.5 --beta 0.5 --lr 0.001 --repair_neurons_num 20 --epoch 200
If you want to run DNN fine-tuning code, here is an example of code running commands:
python BIRDNN/repair/fine_tuning.py --network 19 --alpha 0.5 --repair_neurons_num 20 --lr 0.001
If you want to output the results of fine-tuning single-layer running, here is an example of code running:
python BIRDNN/repair/fine_tuning_single_layer.py --network 19 --alpha 0.5 --repair_neurons_num 20 --lr 0.001
@inproceedings{yang2022neural,
title={Neural network repair with reachability analysis},
author={Yang, Xiaodong and Yamaguchi, Tom and Tran, Hoang-Dung and Hoxha, Bardh and Johnson, Taylor T and Prokhorov, Danil},
booktitle={Formal Modeling and Analysis of Timed Systems: 20th International Conference, FORMATS 2022, Warsaw, Poland, September 13--15, 2022, Proceedings},
pages={221--236},
year={2022},
organization={Springer}
}
@inproceedings{sun2022causality,
title={Causality-based neural network repair},
author={Sun, Bing and Sun, Jun and Pham, Long H and Shi, Jie},
booktitle={Proceedings of the 44th International Conference on Software Engineering},
pages={338--349},
year={2022}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent