Skip to content

michaelklachko/NoisyNet

Repository files navigation

NoisyNet

Code for "Improving Noise Tolerance of Mixed-Signal Neural Networks" https://arxiv.org/abs/1904.01705

Dataset (4 bit CIFAR-10) can be downloaded here

To run the model with I_max = 1nA in all layers (~78%):

python noisynet.py --current 1 --act_max 5 --w_max1 0.3 --LR 0.005 --L2_1 0.0005 --L2_2 0.0002

Noise-free baseline (~88%):

python noisynet.py --L2 0.0005 --dropout 0.1 --nepochs 450

Attribution: The code in timm folder as well as efficientnet code in models folder and its supporting files are copied from Ross Wightman's https://github.com/rwightman/pytorch-image-models with some modifications and are licenced under Apache 2.0.