The code accompanying our TMLR 10/2022 paper: On the Paradox of Certified Training.

For a brief overview of the work, check out our blogpost.

We require the following packages:

cudatoolkit=10.1
matplotlib=3.3.2
numpy=1.19.2
pytorch==1.6.0
scikit-learn=0.23.2
tabulate=0.8.7
torchvision==0.7.0 
tqdm=4.50.2

To set up a sufficient conda environment, start by installing pytorch 1.6.0 in a python 3.6.9 environment:

$ conda create -n myenv python=3.6.9
$ conda activate myenv
$ conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch

Then, install the remaining required packages using the provided file requirements.txt:

$ pip install -r requirements.txt

To run certified training with convex relaxations invoke cert.py. In scripts/ we provide the example commands used to train all models shown in Tables 2, 3, and 4.

The training will save model snapshots to saved_models/ and the results to results/ (the folders will be automatically created).

To run with different settings, modify the examples, referring to args_factory.py which documents the command-line flags.

Additionally, we provide snapshots of the models from Table 2 in pretrained/. In scripts/table2.sh we give an example of a command used to evaluate one such snapshot.

@article{
    jovanovic2022on,
    title={On the Paradox of Certified Training},
    author={Nikola Jovanovi{\'c} and Mislav Balunovi\'{c} and Maximilian Baader and Martin Vechev},
    journal={Transactions on Machine Learning Research},
    year={2022},
    url={https://openreview.net/forum?id=atJHLVyBi8},
    note={}
}