This repository contains the PyTorch implementation of Top-k Neuron Patterns testing coverage proposed in DeepGuage paper. In this document, I have also summarized the research article as well as described its stengths, weaknesses and possible ways to extend this execullent work.
This repository contains two files:
- DeepGauge.ipynb: Google Colab file containing the Pytorch implementation.
- LeNet5.pth: This file contains the state (weights and bias) of a trained LeNet5 model.
You can either re-train a new LeNet5 model using the initial sections of DeepGauge.ipynb notebook and then compute the top-k Neuron Pattern coverage result on the MNIST test set using the Top-k Neuron Patterns section of this notebook. Note: With this approach, there might be minor changes in the coverage results compared to the currently displayed results.
Otherwise, you can directly use the pre-trained model in the notebook and start from the Top-k Neuron Patterns section to rerun and reproduce the results.