TorchFI is a fault injection framework build on top of Pytorch for research purposes.
git clone git@github.com:bfgoldstein/torchfi.gitWe highly recommend installing an Anaconda environment.
conda env create -f torchfi.yml
conda activate torchfi
cd ${PROJECT_PATH}
export PYTHON_PATH=$PYTHON_PATH:${PROJECT_PATH}Check out this torchFI examples using AlexNet, ResNet18 and ResNet50 model.
All pruned models to run the above experiments can be download at DeGirum and Distiller Model Zoo.
Please cite XXX in your publications if it helps your research:
@INPROCEEDINGS{goldstein2019,
Author = {Goldstein, Brunno and Srinivasan, Sudarshan and Mellempudi, Naveen K and Das, Dipankar and Santiago, Leandro and Ferreira, Victor C. and Solon, N. and Kundu, Sandip and França, Felipe M. G.},
Booktitle={2020 IEEE 11th Latin American Symposium on Circuits Systems (LASCAS)},
Title = {Reliability Evaluation of Compressed Deep Learning Models},
Year = {2020},
Keywords={resilience, soft error, transient fault, neural network, deep learning}
}
TorchFI code is released under the Apache license 2.0.