Selective Network Linearization for Efficient Private Inference - Published in ICML 2022

Author: Minsu Cho, Ameya Joshi, Siddharth Garg, Brandon Reagen, Chinmay Hegde

This is the official code for our paper Selective Network Linearization for Efficient Private Inference published in ICML 2022.

Setup

Basic Requirements:

1. pytorch == 1.1.0
2. torchvision == 0.12.0
3. numpy == 1.21.5

Instructions

This is the instructions for the ResNet18 on CIFAR100 with ReLU count 100k.

1. Train the ResNet18 model:

bash ./scripts/train_resnet18_c100.sh

2. Run SNL code with the saved models from Step 1.

bash ./scripts/snl_resnet18_c100_relu_100k.sh

The other examples for different relu counts are

bash ./scripts/snl_resnet18_c100_relu_25k.sh
bash ./scripts/snl_resnet18_c100_relu_50k.sh