Skip to content

Solving the linear programming-based neural network verification problem through Alternating Direction Method of Multipliers (ADMM).

Notifications You must be signed in to change notification settings

ShaoruChen/DeepSplit

Repository files navigation

DeepSplit: Scalable Verification of Deep Neural Networks via Operator Splitting

DeepSplit is a library for solving neural network verification problems through the Alternative Direction Method of Multipliers (ADMM). Details of this method can be found in this paper:

Installation & Usage

Run the following commands to install a conda environment:

conda create -n deepsplit python=3.10
conda activate deepsplit

# Example: install pytorch on macOS. 
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 -c pytorch

pip install auto-lirpa
conda install -c conda-forge cvxpy

pip install convex_adversarial
pip install -r requirements.txt

Run the following command to bound the output of an MLP image classifier on the MNIST dataset:

python examples/mnist/main_mnist_fc.py

About

Solving the linear programming-based neural network verification problem through Alternating Direction Method of Multipliers (ADMM).

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages