DeepVigor+; Fast and Scalable Fault Resilience Analysis Tool for Deep Neural Networks

DeepVigor+ is an innovative fault resilience analysis tool for CNNs with remarkable scalability, outperforming conventional statistical fault injection while meeting analysis accuracy requirements. With DeepVigor+, emerging CNNs can be analyzed within a few minutes.

How to run

To exploit the DeepVigor+ tool, there are multiple parameters to be set. It is executed by the following commands: python main.py --model=your_cnn --run-mode=target_run_mode --dataset=dataset --batch-size=batch_size The description of each parameter is as follows:

• model: the name and checkpoint of the pre-trained CNN model should be defined in load_model function in CNNs.py, similar to the existing CNNs in the corresponding file.
• dataset: the datasets are specified in dataset_utils.py. Existing datasets: CIFAR-10, CIFAR-100, IMAGENET.
• batch-size: it is used for dataloader while loading the datasets. recommended: 100
• run-mode: specifies the fault resilience analysis method on weights and activations separately, in both complete and sampling methods, as follows:
  • full-analysis-weight: complete analysis for all output channels of weights in convolutional layers
  • full-analysis-act: complete analysis for all output activations in convolutional layers
  • sampling-analysis-weight: sampling analysis for random output channels of weights, specified by channel-sampling
  • sampling-analysis-act: sampling analysis for random output activations, specified by channel-sampling
• channel-sampling: is specified in the case of run-mode=sampling-analysis-weight or run-mode=sampling-analysis-act, as a value in (0, 1). recommended vale: 0.1

The outputs of the tool are provided based on the specified run-mode. In the case of full-analysis, the vulnerability factors for all neurons/channels are written in separate files for each layer, in a directory: network_name-dataset_name/run_mode. Moreover, a log file is created which summarizes the results. In the case of sampling analysis, a log file is created in the corresponding directory, reporting the analysis results.

Obtained results from the tool can be exploited for fault resilience comparison and visualization, as reported in the paper.

It is highly recommended to run the tool on GPU. It is tested with python 3.10.9 and pytorch 2.1.2+cu121

citation

Check out the DeepVigor and DeepVigor+ papers. If you use or reference DeepVigor+, please cite:

@inproceedings{ahmadilivani2023deepvigor,
  title={Deepvigor: Vulnerability value ranges and factors for dnns’ reliability assessment},
  author={Ahmadilivani, Mohammad Hasan and Taheri, Mahdi and Raik, Jaan and Daneshtalab, Masoud and Jenihhin, Maksim},
  booktitle={2023 IEEE European Test Symposium (ETS)},
  pages={1--6},
  year={2023},
  organization={IEEE}
}

and

  title={DeepVigor+: Scalable and Accurate Semi-Analytical Fault Resilience Analysis for Deep Neural Network},
  author={Ahmadilivani, Mohammad Hasan and Raik, Jaan and Daneshtalab, Masoud and Jenihhin, Maksim},
  journal={arXiv preprint arXiv:2410.15742},
  year={2024}