N-Activation

Code for our paper 1-Lipschitz Neural Networks are more expressive with N-Activations.
It includes code for the proposed $\mathcal{N}$-activation.

Requirements:

• Python 3.9
• PyTorch 1.12

Usage:

Use the settings_nr (see below or settings file) to specify the setting and run:

python train.py 10

The last digit of the settings number determines the activation, 0 means MaxMin and 1 for the $\mathcal{N}$-activation. The setting number also determines the dataset and method for making the convolutions 1-Lipschitz, see below.

Settings Nr	Dataset	Method
10-19	CIFAR-10	AOL
20-29	CIFAR-10	CPL
30-39	CIFAR-10	SOC
40-49	CIFAR-100	AOL
50-59	CIFAR-100	CPL
60-69	CIFAR-100	SOC
70-79	Tiny ImageNet	AOL
80-89	Tiny ImageNet	CPL
90-99	Tiny ImageNet	SOC