This repository contains code for the paper 1-Lipschitz Layers Compared: Memory, Speed, and Certifiable Robustness, following an attempt to simplify it and clean it up. The original code can be found here

Instructions:

Runs scrips e.g. in the following way:

python run.py scripts/step1_measure_batch_time.py 0
...
python run.py scripts/step1_measure_batch_time.py 31

python run.py scripts/step2_print_epoch_budgets.py

python run.py scripts/step3_hp_search.py 1

Reproduce Experiments from the Paper:

Follow the pseudocode below:

For i = 0 to 31 
    >> python run.py scripts/step1_measure_batch_time.py i
    
>> python run.py scripts/step2_print_epoch_budgets.py
Update data/epoch_budgets_2h/cifar10.yaml with the printed values

For i = 0 to 31
    For _ = 1 to 16
        >> python run.py scripts/step3_hp_search.py i

>> python run.py scripts/step4_print_best_hp.py
Update data/best_hps/cifar10.yaml with the printed values

For i = 0 to 31
    >> python run.py scripts/step5_test_set_evaluation.py i

Repeat for the other datasets and other time budgets.

Requirements:

• PyTorch
• torchvision
• PyYAML
• einops