UW-Madison-Lee-Lab/CodedInvNet

ICML2021

Coded-InvNet for Resilient Prediction Serving Systems

In this repository, we provide the source code for the Coded-InvNet method.

Here is an example of the script on MNIST, using i-ResNet-64 for the invertible network and PixPix for the fusion network, with k=2.

1. Train invertible network

python ./main.py --config_file configs/mnist.json

2. Generate fusion dataset

python ./main.py --config_file configs/mnist.json -fusion --nactors 2 -sample_fusion

3. Train fusion network

python ./main.py --config_file configs/mnist.json -fusion --nactors 2

4. Evaluate the fusion network

python ./main.py --config_file configs/mnist.json -fusion --nactors 2 -test_fusion

Credit: we make use of the code from i-ResNet, i-RevNet and Pix2Pix for our models.