PaRoT: A Practical Framework for Robust Deep Neural Network Training
Authors: Edward W. Ayers, Francisco Eiras, Majd Hawasly, Iain Whiteside
Installing PaRoT
We recommend installing PaRoT in a virtual environment (e.g. using Anaconda). You can install it by running:
pip install .
from the main directory inside the repository.
Running Examples
The examples provided with the framework can be found inside the examples
folder of the package.
The simplest example corresponds to the training of a simple 5 layers neural network on MNIST, which showcases the ease of use of our framework. It can be launched by running:
cd examples
python3 train_simple_example.py
outputting a checkpoint which can then be loaded and tested.
In diffai_comparison.py is the code corresponding to the paper experiments
which compare PaRoT to DiffAI. Each of the cases can be run by:
python3 diffai_comparison.py --model [MODEL_ID] --domain [DOMAIN_ID] --property [PROPERTY_ID] --dataset [DATASET_ID]
where MODEL_ID can be any of the models in the paper, DOMAIN_ID can be box or hz (Hybrid Zonotope in the paper) for the
built-in domains, PROPERTY_ID can be ball, brightness or fourier for example, and DATASET_ID is
either MNIST or CIFAR10. The comparsion results are outputted to a JSON file.
Copyright 2020 FiveAI Ltd. All rights reserved. PaRoT is released under the "MIT License Agreement". Please see the LICENSE file that is included as part of this package.